From 4617721c502b2ddaa4e324e86da4997edf738fa5 Mon Sep 17 00:00:00 2001 From: Steven Rostedt Date: Wed, 2 Sep 2026 09:55:01 -0400 Subject: ftrace: Synchronize the initialization of ftrace_ops There's some internal state that ftrace_ops needs to have set, but since it can be declared outside of the ftrace.c code, it calls ftrace_ops_init() on the ops in every global function. The issue is that if two tasks call it on the same ops at the same time it is possible to have the initialization of one corrupt the initialization of the other call. Create a ops_mutex to use to synchronize every initialization of the ftrace_ops. The mutex is taken within checking the ftrace_ops flag that states it was initializied but the flag is checked again after the mutex has been taken. Checking first outside the mutex allows it to shortcut having to take the mutex. But then the check needs to be done again after the mute is taken in case of races. Cc: stable@vger.kernel.org Link: https://patch.msgid.link/20260902095501.6b59af20@gandalf.local.home Fixes: f04f24fb7e48d ("ftrace, kprobes: Fix a deadlock on ftrace_regex_lock") Reported-by: sashiko-bot@kernel.org Close: https://lore.kernel.org/all/20260829025528.49A831F000E9@smtp.kernel.org/ Signed-off-by: Steven Rostedt --- fs/ocfs2/dlm/Makefile | 5 + fs/ocfs2/dlm/dlmapi.h | 201 +++ fs/ocfs2/dlm/dlmast.c | 475 ++++++ fs/ocfs2/dlm/dlmcommon.h | 1132 ++++++++++++++ fs/ocfs2/dlm/dlmconvert.c | 555 +++++++ fs/ocfs2/dlm/dlmconvert.h | 18 + fs/ocfs2/dlm/dlmdebug.c | 844 +++++++++++ fs/ocfs2/dlm/dlmdebug.h | 49 + fs/ocfs2/dlm/dlmdomain.c | 2371 +++++++++++++++++++++++++++++ fs/ocfs2/dlm/dlmdomain.h | 41 + fs/ocfs2/dlm/dlmlock.c | 743 +++++++++ fs/ocfs2/dlm/dlmmaster.c | 3557 ++++++++++++++++++++++++++++++++++++++++++++ fs/ocfs2/dlm/dlmrecovery.c | 2962 ++++++++++++++++++++++++++++++++++++ fs/ocfs2/dlm/dlmthread.c | 809 ++++++++++ fs/ocfs2/dlm/dlmunlock.c | 695 +++++++++ 15 files changed, 14457 insertions(+) create mode 100644 fs/ocfs2/dlm/Makefile create mode 100644 fs/ocfs2/dlm/dlmapi.h create mode 100644 fs/ocfs2/dlm/dlmast.c create mode 100644 fs/ocfs2/dlm/dlmcommon.h create mode 100644 fs/ocfs2/dlm/dlmconvert.c create mode 100644 fs/ocfs2/dlm/dlmconvert.h create mode 100644 fs/ocfs2/dlm/dlmdebug.c create mode 100644 fs/ocfs2/dlm/dlmdebug.h create mode 100644 fs/ocfs2/dlm/dlmdomain.c create mode 100644 fs/ocfs2/dlm/dlmdomain.h create mode 100644 fs/ocfs2/dlm/dlmlock.c create mode 100644 fs/ocfs2/dlm/dlmmaster.c create mode 100644 fs/ocfs2/dlm/dlmrecovery.c create mode 100644 fs/ocfs2/dlm/dlmthread.c create mode 100644 fs/ocfs2/dlm/dlmunlock.c (limited to 'fs/ocfs2/dlm') diff --git a/fs/ocfs2/dlm/Makefile b/fs/ocfs2/dlm/Makefile new file mode 100644 index 000000000..5e700b45d --- /dev/null +++ b/fs/ocfs2/dlm/Makefile @@ -0,0 +1,5 @@ +# SPDX-License-Identifier: GPL-2.0-only +obj-$(CONFIG_OCFS2_FS_O2CB) += ocfs2_dlm.o + +ocfs2_dlm-objs := dlmdomain.o dlmdebug.o dlmthread.o dlmrecovery.o \ + dlmmaster.o dlmast.o dlmconvert.o dlmlock.o dlmunlock.o diff --git a/fs/ocfs2/dlm/dlmapi.h b/fs/ocfs2/dlm/dlmapi.h new file mode 100644 index 000000000..1969db8ff --- /dev/null +++ b/fs/ocfs2/dlm/dlmapi.h @@ -0,0 +1,201 @@ +/* SPDX-License-Identifier: GPL-2.0-or-later */ +/* + * dlmapi.h + * + * externally exported dlm interfaces + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + +#ifndef DLMAPI_H +#define DLMAPI_H + +struct dlm_lock; +struct dlm_ctxt; + +/* NOTE: changes made to this enum should be reflected in dlmdebug.c */ +enum dlm_status { + DLM_NORMAL = 0, /* 0: request in progress */ + DLM_GRANTED, /* 1: request granted */ + DLM_DENIED, /* 2: request denied */ + DLM_DENIED_NOLOCKS, /* 3: request denied, out of system resources */ + DLM_WORKING, /* 4: async request in progress */ + DLM_BLOCKED, /* 5: lock request blocked */ + DLM_BLOCKED_ORPHAN, /* 6: lock request blocked by a orphan lock*/ + DLM_DENIED_GRACE_PERIOD, /* 7: topological change in progress */ + DLM_SYSERR, /* 8: system error */ + DLM_NOSUPPORT, /* 9: unsupported */ + DLM_CANCELGRANT, /* 10: can't cancel convert: already granted */ + DLM_IVLOCKID, /* 11: bad lockid */ + DLM_SYNC, /* 12: synchronous request granted */ + DLM_BADTYPE, /* 13: bad resource type */ + DLM_BADRESOURCE, /* 14: bad resource handle */ + DLM_MAXHANDLES, /* 15: no more resource handles */ + DLM_NOCLINFO, /* 16: can't contact cluster manager */ + DLM_NOLOCKMGR, /* 17: can't contact lock manager */ + DLM_NOPURGED, /* 18: can't contact purge daemon */ + DLM_BADARGS, /* 19: bad api args */ + DLM_VOID, /* 20: no status */ + DLM_NOTQUEUED, /* 21: NOQUEUE was specified and request failed */ + DLM_IVBUFLEN, /* 22: invalid resource name length */ + DLM_CVTUNGRANT, /* 23: attempted to convert ungranted lock */ + DLM_BADPARAM, /* 24: invalid lock mode specified */ + DLM_VALNOTVALID, /* 25: value block has been invalidated */ + DLM_REJECTED, /* 26: request rejected, unrecognized client */ + DLM_ABORT, /* 27: blocked lock request cancelled */ + DLM_CANCEL, /* 28: conversion request cancelled */ + DLM_IVRESHANDLE, /* 29: invalid resource handle */ + DLM_DEADLOCK, /* 30: deadlock recovery refused this request */ + DLM_DENIED_NOASTS, /* 31: failed to allocate AST */ + DLM_FORWARD, /* 32: request must wait for primary's response */ + DLM_TIMEOUT, /* 33: timeout value for lock has expired */ + DLM_IVGROUPID, /* 34: invalid group specification */ + DLM_VERS_CONFLICT, /* 35: version conflicts prevent request handling */ + DLM_BAD_DEVICE_PATH, /* 36: Locks device does not exist or path wrong */ + DLM_NO_DEVICE_PERMISSION, /* 37: Client has insufficient pers for device */ + DLM_NO_CONTROL_DEVICE, /* 38: Cannot set options on opened device */ + + DLM_RECOVERING, /* 39: extension, allows caller to fail a lock + request if it is being recovered */ + DLM_MIGRATING, /* 40: extension, allows caller to fail a lock + request if it is being migrated */ + DLM_MAXSTATS, /* 41: upper limit for return code validation */ +}; + +/* for pretty-printing dlm_status error names */ +const char *dlm_errname(enum dlm_status err); + +/* Eventually the DLM will use standard errno values, but in the + * meantime this lets us track dlm errors as they bubble up. When we + * bring its error reporting into line with the rest of the stack, + * these can just be replaced with calls to mlog_errno. */ +#define dlm_error(st) do { \ + if ((st) != DLM_RECOVERING && \ + (st) != DLM_MIGRATING && \ + (st) != DLM_FORWARD) \ + mlog(ML_ERROR, "dlm status = %s\n", dlm_errname((st))); \ +} while (0) + +#define DLM_LKSB_UNUSED1 0x01 +#define DLM_LKSB_PUT_LVB 0x02 +#define DLM_LKSB_GET_LVB 0x04 +#define DLM_LKSB_UNUSED2 0x08 +#define DLM_LKSB_UNUSED3 0x10 +#define DLM_LKSB_UNUSED4 0x20 +#define DLM_LKSB_UNUSED5 0x40 +#define DLM_LKSB_UNUSED6 0x80 + +#define DLM_LVB_LEN 64 + +/* Callers are only allowed access to the lvb and status members of + * this struct. */ +struct dlm_lockstatus { + enum dlm_status status; + u32 flags; + struct dlm_lock *lockid; + char lvb[DLM_LVB_LEN]; +}; + +/* Valid lock modes. */ +#define LKM_IVMODE (-1) /* invalid mode */ +#define LKM_NLMODE 0 /* null lock */ +#define LKM_CRMODE 1 /* concurrent read unsupported */ +#define LKM_CWMODE 2 /* concurrent write unsupported */ +#define LKM_PRMODE 3 /* protected read */ +#define LKM_PWMODE 4 /* protected write unsupported */ +#define LKM_EXMODE 5 /* exclusive */ +#define LKM_MAXMODE 5 +#define LKM_MODEMASK 0xff + +/* Flags passed to dlmlock and dlmunlock: + * reserved: flags used by the "real" dlm + * only a few are supported by this dlm + * (U) = unsupported by ocfs2 dlm */ +#define LKM_ORPHAN 0x00000010 /* this lock is orphanable (U) */ +#define LKM_PARENTABLE 0x00000020 /* this lock was orphaned (U) */ +#define LKM_BLOCK 0x00000040 /* blocking lock request (U) */ +#define LKM_LOCAL 0x00000080 /* local lock request */ +#define LKM_VALBLK 0x00000100 /* lock value block request */ +#define LKM_NOQUEUE 0x00000200 /* non blocking request */ +#define LKM_CONVERT 0x00000400 /* conversion request */ +#define LKM_NODLCKWT 0x00000800 /* this lock won't deadlock (U) */ +#define LKM_UNLOCK 0x00001000 /* deallocate this lock */ +#define LKM_CANCEL 0x00002000 /* cancel conversion request */ +#define LKM_DEQALL 0x00004000 /* remove all locks held by proc (U) */ +#define LKM_INVVALBLK 0x00008000 /* invalidate lock value block */ +#define LKM_SYNCSTS 0x00010000 /* return synchronous status if poss (U) */ +#define LKM_TIMEOUT 0x00020000 /* lock request contains timeout (U) */ +#define LKM_SNGLDLCK 0x00040000 /* request can self-deadlock (U) */ +#define LKM_FINDLOCAL 0x00080000 /* find local lock request (U) */ +#define LKM_PROC_OWNED 0x00100000 /* owned by process, not group (U) */ +#define LKM_XID 0x00200000 /* use transaction id for deadlock (U) */ +#define LKM_XID_CONFLICT 0x00400000 /* do not allow lock inheritance (U) */ +#define LKM_FORCE 0x00800000 /* force unlock flag */ +#define LKM_REVVALBLK 0x01000000 /* temporary solution: re-validate + lock value block (U) */ +/* unused */ +#define LKM_UNUSED1 0x00000001 /* unused */ +#define LKM_UNUSED2 0x00000002 /* unused */ +#define LKM_UNUSED3 0x00000004 /* unused */ +#define LKM_UNUSED4 0x00000008 /* unused */ +#define LKM_UNUSED5 0x02000000 /* unused */ +#define LKM_UNUSED6 0x04000000 /* unused */ +#define LKM_UNUSED7 0x08000000 /* unused */ + +/* ocfs2 extensions: internal only + * should never be used by caller */ +#define LKM_MIGRATION 0x10000000 /* extension: lockres is to be migrated + to another node */ +#define LKM_PUT_LVB 0x20000000 /* extension: lvb is being passed + should be applied to lockres */ +#define LKM_GET_LVB 0x40000000 /* extension: lvb should be copied + from lockres when lock is granted */ +#define LKM_RECOVERY 0x80000000 /* extension: flag for recovery lock + used to avoid recovery rwsem */ + + +typedef void (dlm_astlockfunc_t)(void *); +typedef void (dlm_bastlockfunc_t)(void *, int); +typedef void (dlm_astunlockfunc_t)(void *, enum dlm_status); + +enum dlm_status dlmlock(struct dlm_ctxt *dlm, + int mode, + struct dlm_lockstatus *lksb, + int flags, + const char *name, + int namelen, + dlm_astlockfunc_t *ast, + void *data, + dlm_bastlockfunc_t *bast); + +enum dlm_status dlmunlock(struct dlm_ctxt *dlm, + struct dlm_lockstatus *lksb, + int flags, + dlm_astunlockfunc_t *unlockast, + void *data); + +struct dlm_protocol_version { + u8 pv_major; + u8 pv_minor; +}; +struct dlm_ctxt * dlm_register_domain(const char *domain, u32 key, + struct dlm_protocol_version *fs_proto); + +void dlm_unregister_domain(struct dlm_ctxt *dlm); + +void dlm_print_one_lock(struct dlm_lock *lockid); + +typedef void (dlm_eviction_func)(int, void *); +struct dlm_eviction_cb { + struct list_head ec_item; + dlm_eviction_func *ec_func; + void *ec_data; +}; +void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb, + dlm_eviction_func *f, + void *data); +void dlm_register_eviction_cb(struct dlm_ctxt *dlm, + struct dlm_eviction_cb *cb); +void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb); + +#endif /* DLMAPI_H */ diff --git a/fs/ocfs2/dlm/dlmast.c b/fs/ocfs2/dlm/dlmast.c new file mode 100644 index 000000000..c681ba957 --- /dev/null +++ b/fs/ocfs2/dlm/dlmast.c @@ -0,0 +1,475 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * dlmast.c + * + * AST and BAST functionality for local and remote nodes + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + + +#include "../cluster/heartbeat.h" +#include "../cluster/nodemanager.h" +#include "../cluster/tcp.h" + +#include "dlmapi.h" +#include "dlmcommon.h" + +#define MLOG_MASK_PREFIX ML_DLM +#include "../cluster/masklog.h" + +static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, + struct dlm_lock *lock); +static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock); + +/* Should be called as an ast gets queued to see if the new + * lock level will obsolete a pending bast. + * For example, if dlm_thread queued a bast for an EX lock that + * was blocking another EX, but before sending the bast the + * lock owner downconverted to NL, the bast is now obsolete. + * Only the ast should be sent. + * This is needed because the lock and convert paths can queue + * asts out-of-band (not waiting for dlm_thread) in order to + * allow for LKM_NOQUEUE to get immediate responses. */ +static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock) +{ + assert_spin_locked(&dlm->ast_lock); + assert_spin_locked(&lock->spinlock); + + if (lock->ml.highest_blocked == LKM_IVMODE) + return 0; + BUG_ON(lock->ml.highest_blocked == LKM_NLMODE); + + if (lock->bast_pending && + list_empty(&lock->bast_list)) + /* old bast already sent, ok */ + return 0; + + if (lock->ml.type == LKM_EXMODE) + /* EX blocks anything left, any bast still valid */ + return 0; + else if (lock->ml.type == LKM_NLMODE) + /* NL blocks nothing, no reason to send any bast, cancel it */ + return 1; + else if (lock->ml.highest_blocked != LKM_EXMODE) + /* PR only blocks EX */ + return 1; + + return 0; +} + +void __dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock) +{ + struct dlm_lock_resource *res; + + BUG_ON(!dlm); + BUG_ON(!lock); + + res = lock->lockres; + + assert_spin_locked(&dlm->ast_lock); + + if (!list_empty(&lock->ast_list)) { + mlog(ML_ERROR, "%s: res %.*s, lock %u:%llu, " + "AST list not empty, pending %d, newlevel %d\n", + dlm->name, res->lockname.len, res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), + lock->ast_pending, lock->ml.type); + BUG(); + } + if (lock->ast_pending) + mlog(0, "%s: res %.*s, lock %u:%llu, AST getting flushed\n", + dlm->name, res->lockname.len, res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie))); + + /* putting lock on list, add a ref */ + dlm_lock_get(lock); + spin_lock(&lock->spinlock); + + /* check to see if this ast obsoletes the bast */ + if (dlm_should_cancel_bast(dlm, lock)) { + mlog(0, "%s: res %.*s, lock %u:%llu, Cancelling BAST\n", + dlm->name, res->lockname.len, res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie))); + lock->bast_pending = 0; + list_del_init(&lock->bast_list); + lock->ml.highest_blocked = LKM_IVMODE; + /* removing lock from list, remove a ref. guaranteed + * this won't be the last ref because of the get above, + * so res->spinlock will not be taken here */ + dlm_lock_put(lock); + /* free up the reserved bast that we are cancelling. + * guaranteed that this will not be the last reserved + * ast because *both* an ast and a bast were reserved + * to get to this point. the res->spinlock will not be + * taken here */ + dlm_lockres_release_ast(dlm, res); + } + list_add_tail(&lock->ast_list, &dlm->pending_asts); + lock->ast_pending = 1; + spin_unlock(&lock->spinlock); +} + +void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock) +{ + BUG_ON(!dlm); + BUG_ON(!lock); + + spin_lock(&dlm->ast_lock); + __dlm_queue_ast(dlm, lock); + spin_unlock(&dlm->ast_lock); +} + + +void __dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock) +{ + struct dlm_lock_resource *res; + + BUG_ON(!dlm); + BUG_ON(!lock); + + assert_spin_locked(&dlm->ast_lock); + + res = lock->lockres; + + BUG_ON(!list_empty(&lock->bast_list)); + if (lock->bast_pending) + mlog(0, "%s: res %.*s, lock %u:%llu, BAST getting flushed\n", + dlm->name, res->lockname.len, res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie))); + + /* putting lock on list, add a ref */ + dlm_lock_get(lock); + spin_lock(&lock->spinlock); + list_add_tail(&lock->bast_list, &dlm->pending_basts); + lock->bast_pending = 1; + spin_unlock(&lock->spinlock); +} + +static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, + struct dlm_lock *lock) +{ + struct dlm_lockstatus *lksb = lock->lksb; + BUG_ON(!lksb); + + /* only updates if this node masters the lockres */ + spin_lock(&res->spinlock); + if (res->owner == dlm->node_num) { + /* check the lksb flags for the direction */ + if (lksb->flags & DLM_LKSB_GET_LVB) { + mlog(0, "getting lvb from lockres for %s node\n", + lock->ml.node == dlm->node_num ? "master" : + "remote"); + memcpy(lksb->lvb, res->lvb, DLM_LVB_LEN); + } + /* Do nothing for lvb put requests - they should be done in + * place when the lock is downconverted - otherwise we risk + * racing gets and puts which could result in old lvb data + * being propagated. We leave the put flag set and clear it + * here. In the future we might want to clear it at the time + * the put is actually done. + */ + } + spin_unlock(&res->spinlock); + + /* reset any lvb flags on the lksb */ + lksb->flags &= ~(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB); +} + +void dlm_do_local_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, + struct dlm_lock *lock) +{ + dlm_astlockfunc_t *fn; + + mlog(0, "%s: res %.*s, lock %u:%llu, Local AST\n", dlm->name, + res->lockname.len, res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie))); + + fn = lock->ast; + BUG_ON(lock->ml.node != dlm->node_num); + + dlm_update_lvb(dlm, res, lock); + (*fn)(lock->astdata); +} + + +int dlm_do_remote_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, + struct dlm_lock *lock) +{ + int ret; + struct dlm_lockstatus *lksb; + int lksbflags; + + mlog(0, "%s: res %.*s, lock %u:%llu, Remote AST\n", dlm->name, + res->lockname.len, res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie))); + + lksb = lock->lksb; + BUG_ON(lock->ml.node == dlm->node_num); + + lksbflags = lksb->flags; + dlm_update_lvb(dlm, res, lock); + + /* lock request came from another node + * go do the ast over there */ + ret = dlm_send_proxy_ast(dlm, res, lock, lksbflags); + return ret; +} + +void dlm_do_local_bast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, + struct dlm_lock *lock, int blocked_type) +{ + dlm_bastlockfunc_t *fn = lock->bast; + + BUG_ON(lock->ml.node != dlm->node_num); + + mlog(0, "%s: res %.*s, lock %u:%llu, Local BAST, blocked %d\n", + dlm->name, res->lockname.len, res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), + blocked_type); + + (*fn)(lock->astdata, blocked_type); +} + + + +int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + int ret; + unsigned int locklen; + struct dlm_ctxt *dlm = data; + struct dlm_lock_resource *res = NULL; + struct dlm_lock *lock = NULL; + struct dlm_proxy_ast *past = (struct dlm_proxy_ast *) msg->buf; + char *name; + struct list_head *head = NULL; + __be64 cookie; + u32 flags; + u8 node; + + if (!dlm_grab(dlm)) { + dlm_error(DLM_REJECTED); + return DLM_REJECTED; + } + + mlog_bug_on_msg(!dlm_domain_fully_joined(dlm), + "Domain %s not fully joined!\n", dlm->name); + + name = past->name; + locklen = past->namelen; + cookie = past->cookie; + flags = be32_to_cpu(past->flags); + node = past->node_idx; + + if (locklen > DLM_LOCKID_NAME_MAX) { + ret = DLM_IVBUFLEN; + mlog(ML_ERROR, "Invalid name length (%d) in proxy ast " + "handler!\n", locklen); + goto leave; + } + + if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) == + (LKM_PUT_LVB|LKM_GET_LVB)) { + mlog(ML_ERROR, "Both PUT and GET lvb specified, (0x%x)\n", + flags); + ret = DLM_BADARGS; + goto leave; + } + + mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" : + (flags & LKM_GET_LVB ? "get lvb" : "none")); + + mlog(0, "type=%d, blocked_type=%d\n", past->type, past->blocked_type); + + if (past->type != DLM_AST && + past->type != DLM_BAST) { + mlog(ML_ERROR, "Unknown ast type! %d, cookie=%u:%llu" + "name=%.*s, node=%u\n", past->type, + dlm_get_lock_cookie_node(be64_to_cpu(cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(cookie)), + locklen, name, node); + ret = DLM_IVLOCKID; + goto leave; + } + + res = dlm_lookup_lockres(dlm, name, locklen); + if (!res) { + mlog(0, "Got %sast for unknown lockres! cookie=%u:%llu, " + "name=%.*s, node=%u\n", (past->type == DLM_AST ? "" : "b"), + dlm_get_lock_cookie_node(be64_to_cpu(cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(cookie)), + locklen, name, node); + ret = DLM_IVLOCKID; + goto leave; + } + + /* cannot get a proxy ast message if this node owns it */ + BUG_ON(res->owner == dlm->node_num); + + mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len, + res->lockname.name); + + spin_lock(&res->spinlock); + if (res->state & DLM_LOCK_RES_RECOVERING) { + mlog(0, "Responding with DLM_RECOVERING!\n"); + ret = DLM_RECOVERING; + goto unlock_out; + } + if (res->state & DLM_LOCK_RES_MIGRATING) { + mlog(0, "Responding with DLM_MIGRATING!\n"); + ret = DLM_MIGRATING; + goto unlock_out; + } + /* try convert queue for both ast/bast */ + head = &res->converting; + lock = NULL; + list_for_each_entry(lock, head, list) { + if (lock->ml.cookie == cookie) + goto do_ast; + } + + /* if not on convert, try blocked for ast, granted for bast */ + if (past->type == DLM_AST) + head = &res->blocked; + else + head = &res->granted; + + list_for_each_entry(lock, head, list) { + /* if lock is found but unlock is pending ignore the bast */ + if (lock->ml.cookie == cookie) { + if (lock->unlock_pending) + break; + goto do_ast; + } + } + + mlog(0, "Got %sast for unknown lock! cookie=%u:%llu, name=%.*s, " + "node=%u\n", past->type == DLM_AST ? "" : "b", + dlm_get_lock_cookie_node(be64_to_cpu(cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(cookie)), + locklen, name, node); + + ret = DLM_NORMAL; +unlock_out: + spin_unlock(&res->spinlock); + goto leave; + +do_ast: + ret = DLM_NORMAL; + if (past->type == DLM_AST) { + /* do not alter lock refcount. switching lists. */ + list_move_tail(&lock->list, &res->granted); + mlog(0, "%s: res %.*s, lock %u:%llu, Granted type %d => %d\n", + dlm->name, res->lockname.len, res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(cookie)), + lock->ml.type, lock->ml.convert_type); + + if (lock->ml.convert_type != LKM_IVMODE) { + lock->ml.type = lock->ml.convert_type; + lock->ml.convert_type = LKM_IVMODE; + } else { + // should already be there.... + } + + lock->lksb->status = DLM_NORMAL; + + /* if we requested the lvb, fetch it into our lksb now */ + if (flags & LKM_GET_LVB) { + BUG_ON(!(lock->lksb->flags & DLM_LKSB_GET_LVB)); + memcpy(lock->lksb->lvb, past->lvb, DLM_LVB_LEN); + } + } + spin_unlock(&res->spinlock); + + if (past->type == DLM_AST) + dlm_do_local_ast(dlm, res, lock); + else + dlm_do_local_bast(dlm, res, lock, past->blocked_type); + +leave: + if (res) + dlm_lockres_put(res); + + dlm_put(dlm); + return ret; +} + + + +int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, + struct dlm_lock *lock, int msg_type, + int blocked_type, int flags) +{ + int ret = 0; + struct dlm_proxy_ast past; + struct kvec vec[2]; + size_t veclen = 1; + int status; + + mlog(0, "%s: res %.*s, to %u, type %d, blocked_type %d\n", dlm->name, + res->lockname.len, res->lockname.name, lock->ml.node, msg_type, + blocked_type); + + memset(&past, 0, sizeof(struct dlm_proxy_ast)); + past.node_idx = dlm->node_num; + past.type = msg_type; + past.blocked_type = blocked_type; + past.namelen = res->lockname.len; + memcpy(past.name, res->lockname.name, past.namelen); + past.cookie = lock->ml.cookie; + + vec[0].iov_len = sizeof(struct dlm_proxy_ast); + vec[0].iov_base = &past; + if (flags & DLM_LKSB_GET_LVB) { + be32_add_cpu(&past.flags, LKM_GET_LVB); + vec[1].iov_len = DLM_LVB_LEN; + vec[1].iov_base = lock->lksb->lvb; + veclen++; + } + + ret = o2net_send_message_vec(DLM_PROXY_AST_MSG, dlm->key, vec, veclen, + lock->ml.node, &status); + if (ret < 0) + mlog(ML_ERROR, "%s: res %.*s, error %d send AST to node %u\n", + dlm->name, res->lockname.len, res->lockname.name, ret, + lock->ml.node); + else { + if (status == DLM_RECOVERING) { + mlog(ML_ERROR, "sent AST to node %u, it thinks this " + "node is dead!\n", lock->ml.node); + BUG(); + } else if (status == DLM_MIGRATING) { + mlog(ML_ERROR, "sent AST to node %u, it returned " + "DLM_MIGRATING!\n", lock->ml.node); + BUG(); + } else if (status != DLM_NORMAL && status != DLM_IVLOCKID) { + mlog(ML_ERROR, "AST to node %u returned %d!\n", + lock->ml.node, status); + /* ignore it */ + } + ret = 0; + } + return ret; +} diff --git a/fs/ocfs2/dlm/dlmcommon.h b/fs/ocfs2/dlm/dlmcommon.h new file mode 100644 index 000000000..20f790a47 --- /dev/null +++ b/fs/ocfs2/dlm/dlmcommon.h @@ -0,0 +1,1132 @@ +/* SPDX-License-Identifier: GPL-2.0-or-later */ +/* + * dlmcommon.h + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + +#ifndef DLMCOMMON_H +#define DLMCOMMON_H + +#include + +#define DLM_HB_NODE_DOWN_PRI (0xf000000) +#define DLM_HB_NODE_UP_PRI (0x8000000) + +#define DLM_LOCKID_NAME_MAX 32 + +#define DLM_LOCK_RES_OWNER_UNKNOWN O2NM_MAX_NODES + +#define DLM_HASH_SIZE_DEFAULT (1 << 17) +#if DLM_HASH_SIZE_DEFAULT < PAGE_SIZE +# define DLM_HASH_PAGES 1 +#else +# define DLM_HASH_PAGES (DLM_HASH_SIZE_DEFAULT / PAGE_SIZE) +#endif +#define DLM_BUCKETS_PER_PAGE (PAGE_SIZE / sizeof(struct hlist_head)) +#define DLM_HASH_BUCKETS (DLM_HASH_PAGES * DLM_BUCKETS_PER_PAGE) + +/* Intended to make it easier for us to switch out hash functions */ +#define dlm_lockid_hash(_n, _l) full_name_hash(NULL, _n, _l) + +enum dlm_mle_type { + DLM_MLE_BLOCK = 0, + DLM_MLE_MASTER = 1, + DLM_MLE_MIGRATION = 2, + DLM_MLE_NUM_TYPES = 3, +}; + +struct dlm_master_list_entry { + struct hlist_node master_hash_node; + struct list_head hb_events; + struct dlm_ctxt *dlm; + spinlock_t spinlock; + wait_queue_head_t wq; + atomic_t woken; + struct kref mle_refs; + int inuse; + unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; + unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; + unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; + unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; + u8 master; + u8 new_master; + enum dlm_mle_type type; + struct o2hb_callback_func mle_hb_up; + struct o2hb_callback_func mle_hb_down; + struct dlm_lock_resource *mleres; + unsigned char mname[DLM_LOCKID_NAME_MAX]; + unsigned int mnamelen; + unsigned int mnamehash; +}; + +enum dlm_ast_type { + DLM_AST = 0, + DLM_BAST = 1, + DLM_ASTUNLOCK = 2, +}; + + +#define LKM_VALID_FLAGS (LKM_VALBLK | LKM_CONVERT | LKM_UNLOCK | \ + LKM_CANCEL | LKM_INVVALBLK | LKM_FORCE | \ + LKM_RECOVERY | LKM_LOCAL | LKM_NOQUEUE) + +#define DLM_RECOVERY_LOCK_NAME "$RECOVERY" +#define DLM_RECOVERY_LOCK_NAME_LEN 9 + +static inline int dlm_is_recovery_lock(const char *lock_name, int name_len) +{ + if (name_len == DLM_RECOVERY_LOCK_NAME_LEN && + memcmp(lock_name, DLM_RECOVERY_LOCK_NAME, name_len)==0) + return 1; + return 0; +} + +#define DLM_RECO_STATE_ACTIVE 0x0001 +#define DLM_RECO_STATE_FINALIZE 0x0002 + +struct dlm_recovery_ctxt +{ + struct list_head resources; + struct list_head node_data; + u8 new_master; + u8 dead_node; + u16 state; + unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; + wait_queue_head_t event; +}; + +enum dlm_ctxt_state { + DLM_CTXT_NEW = 0, + DLM_CTXT_JOINED = 1, + DLM_CTXT_IN_SHUTDOWN = 2, + DLM_CTXT_LEAVING = 3, +}; + +struct dlm_ctxt +{ + struct list_head list; + struct hlist_head **lockres_hash; + struct list_head dirty_list; + struct list_head purge_list; + struct list_head pending_asts; + struct list_head pending_basts; + struct list_head tracking_list; + unsigned int purge_count; + spinlock_t spinlock; + spinlock_t ast_lock; + spinlock_t track_lock; + char *name; + u8 node_num; + u32 key; + u8 joining_node; + u8 migrate_done; /* set to 1 means node has migrated all lock resources */ + wait_queue_head_t dlm_join_events; + unsigned long live_nodes_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; + unsigned long domain_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; + unsigned long exit_domain_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; + unsigned long recovery_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; + struct dlm_recovery_ctxt reco; + spinlock_t master_lock; + struct hlist_head **master_hash; + struct list_head mle_hb_events; + + /* these give a really vague idea of the system load */ + atomic_t mle_tot_count[DLM_MLE_NUM_TYPES]; + atomic_t mle_cur_count[DLM_MLE_NUM_TYPES]; + atomic_t res_tot_count; + atomic_t res_cur_count; + + struct dentry *dlm_debugfs_subroot; + + /* NOTE: Next three are protected by dlm_domain_lock */ + struct kref dlm_refs; + enum dlm_ctxt_state dlm_state; + unsigned int num_joins; + + struct o2hb_callback_func dlm_hb_up; + struct o2hb_callback_func dlm_hb_down; + struct task_struct *dlm_thread_task; + struct task_struct *dlm_reco_thread_task; + struct workqueue_struct *dlm_worker; + wait_queue_head_t dlm_thread_wq; + wait_queue_head_t dlm_reco_thread_wq; + wait_queue_head_t ast_wq; + wait_queue_head_t migration_wq; + + struct work_struct dispatched_work; + struct list_head work_list; + spinlock_t work_lock; + struct list_head dlm_domain_handlers; + struct list_head dlm_eviction_callbacks; + + /* The filesystem specifies this at domain registration. We + * cache it here to know what to tell other nodes. */ + struct dlm_protocol_version fs_locking_proto; + /* This is the inter-dlm communication version */ + struct dlm_protocol_version dlm_locking_proto; +}; + +static inline struct hlist_head *dlm_lockres_hash(struct dlm_ctxt *dlm, unsigned i) +{ + return dlm->lockres_hash[(i / DLM_BUCKETS_PER_PAGE) % DLM_HASH_PAGES] + (i % DLM_BUCKETS_PER_PAGE); +} + +static inline struct hlist_head *dlm_master_hash(struct dlm_ctxt *dlm, + unsigned i) +{ + return dlm->master_hash[(i / DLM_BUCKETS_PER_PAGE) % DLM_HASH_PAGES] + + (i % DLM_BUCKETS_PER_PAGE); +} + +/* these keventd work queue items are for less-frequently + * called functions that cannot be directly called from the + * net message handlers for some reason, usually because + * they need to send net messages of their own. */ +void dlm_dispatch_work(struct work_struct *work); + +struct dlm_lock_resource; +struct dlm_work_item; + +typedef void (dlm_workfunc_t)(struct dlm_work_item *, void *); + +struct dlm_request_all_locks_priv +{ + u8 reco_master; + u8 dead_node; +}; + +struct dlm_mig_lockres_priv +{ + struct dlm_lock_resource *lockres; + u8 real_master; + u8 extra_ref; +}; + +struct dlm_assert_master_priv +{ + struct dlm_lock_resource *lockres; + u8 request_from; + u32 flags; + unsigned ignore_higher:1; +}; + +struct dlm_deref_lockres_priv +{ + struct dlm_lock_resource *deref_res; + u8 deref_node; +}; + +struct dlm_work_item +{ + struct list_head list; + dlm_workfunc_t *func; + struct dlm_ctxt *dlm; + void *data; + union { + struct dlm_request_all_locks_priv ral; + struct dlm_mig_lockres_priv ml; + struct dlm_assert_master_priv am; + struct dlm_deref_lockres_priv dl; + } u; +}; + +static inline void dlm_init_work_item(struct dlm_ctxt *dlm, + struct dlm_work_item *i, + dlm_workfunc_t *f, void *data) +{ + memset(i, 0, sizeof(*i)); + i->func = f; + INIT_LIST_HEAD(&i->list); + i->data = data; + i->dlm = dlm; /* must have already done a dlm_grab on this! */ +} + + + +static inline void __dlm_set_joining_node(struct dlm_ctxt *dlm, + u8 node) +{ + assert_spin_locked(&dlm->spinlock); + + dlm->joining_node = node; + wake_up(&dlm->dlm_join_events); +} + +#define DLM_LOCK_RES_UNINITED 0x00000001 +#define DLM_LOCK_RES_RECOVERING 0x00000002 +#define DLM_LOCK_RES_READY 0x00000004 +#define DLM_LOCK_RES_DIRTY 0x00000008 +#define DLM_LOCK_RES_IN_PROGRESS 0x00000010 +#define DLM_LOCK_RES_MIGRATING 0x00000020 +#define DLM_LOCK_RES_DROPPING_REF 0x00000040 +#define DLM_LOCK_RES_BLOCK_DIRTY 0x00001000 +#define DLM_LOCK_RES_SETREF_INPROG 0x00002000 +#define DLM_LOCK_RES_RECOVERY_WAITING 0x00004000 + +/* max milliseconds to wait to sync up a network failure with a node death */ +#define DLM_NODE_DEATH_WAIT_MAX (5 * 1000) + +#define DLM_PURGE_INTERVAL_MS (8 * 1000) + +struct dlm_lock_resource +{ + /* WARNING: Please see the comment in dlm_init_lockres before + * adding fields here. */ + struct hlist_node hash_node; + struct qstr lockname; + struct kref refs; + + /* + * Please keep granted, converting, and blocked in this order, + * as some funcs want to iterate over all lists. + * + * All four lists are protected by the hash's reference. + */ + struct list_head granted; + struct list_head converting; + struct list_head blocked; + struct list_head purge; + + /* + * These two lists require you to hold an additional reference + * while they are on the list. + */ + struct list_head dirty; + struct list_head recovering; // dlm_recovery_ctxt.resources list + + /* Added during init and removed during release */ + struct list_head tracking; /* dlm->tracking_list */ + + /* unused lock resources have their last_used stamped and are + * put on a list for the dlm thread to run. */ + unsigned long last_used; + + struct dlm_ctxt *dlm; + + unsigned migration_pending:1; + atomic_t asts_reserved; + spinlock_t spinlock; + wait_queue_head_t wq; + u8 owner; //node which owns the lock resource, or unknown + u16 state; + char lvb[DLM_LVB_LEN]; + unsigned int inflight_locks; + unsigned int inflight_assert_workers; + unsigned long refmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; +}; + +struct dlm_migratable_lock +{ + __be64 cookie; + + /* these 3 are just padding for the in-memory structure, but + * list and flags are actually used when sent over the wire */ + __be16 pad1; + u8 list; // 0=granted, 1=converting, 2=blocked + u8 flags; + + s8 type; + s8 convert_type; + s8 highest_blocked; + u8 node; +}; // 16 bytes + +struct dlm_lock +{ + struct dlm_migratable_lock ml; + + struct list_head list; + struct list_head ast_list; + struct list_head bast_list; + struct dlm_lock_resource *lockres; + spinlock_t spinlock; + struct kref lock_refs; + + // ast and bast must be callable while holding a spinlock! + dlm_astlockfunc_t *ast; + dlm_bastlockfunc_t *bast; + void *astdata; + struct dlm_lockstatus *lksb; + unsigned ast_pending:1, + bast_pending:1, + convert_pending:1, + lock_pending:1, + cancel_pending:1, + unlock_pending:1, + lksb_kernel_allocated:1; +}; + +enum dlm_lockres_list { + DLM_GRANTED_LIST = 0, + DLM_CONVERTING_LIST = 1, + DLM_BLOCKED_LIST = 2, +}; + +static inline int dlm_lvb_is_empty(char *lvb) +{ + int i; + for (i=0; igranted; + else if (idx == DLM_CONVERTING_LIST) + ret = &res->converting; + else if (idx == DLM_BLOCKED_LIST) + ret = &res->blocked; + else + BUG(); + return ret; +} + + + + +struct dlm_node_iter +{ + unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; + int curnode; +}; + + +enum { + DLM_MASTER_REQUEST_MSG = 500, + DLM_UNUSED_MSG1 = 501, + DLM_ASSERT_MASTER_MSG = 502, + DLM_CREATE_LOCK_MSG = 503, + DLM_CONVERT_LOCK_MSG = 504, + DLM_PROXY_AST_MSG = 505, + DLM_UNLOCK_LOCK_MSG = 506, + DLM_DEREF_LOCKRES_MSG = 507, + DLM_MIGRATE_REQUEST_MSG = 508, + DLM_MIG_LOCKRES_MSG = 509, + DLM_QUERY_JOIN_MSG = 510, + DLM_ASSERT_JOINED_MSG = 511, + DLM_CANCEL_JOIN_MSG = 512, + DLM_EXIT_DOMAIN_MSG = 513, + DLM_MASTER_REQUERY_MSG = 514, + DLM_LOCK_REQUEST_MSG = 515, + DLM_RECO_DATA_DONE_MSG = 516, + DLM_BEGIN_RECO_MSG = 517, + DLM_FINALIZE_RECO_MSG = 518, + DLM_QUERY_REGION = 519, + DLM_QUERY_NODEINFO = 520, + DLM_BEGIN_EXIT_DOMAIN_MSG = 521, + DLM_DEREF_LOCKRES_DONE = 522, +}; + +struct dlm_reco_node_data +{ + int state; + u8 node_num; + struct list_head list; +}; + +enum { + DLM_RECO_NODE_DATA_DEAD = -1, + DLM_RECO_NODE_DATA_INIT = 0, + DLM_RECO_NODE_DATA_REQUESTING = 1, + DLM_RECO_NODE_DATA_REQUESTED = 2, + DLM_RECO_NODE_DATA_RECEIVING = 3, + DLM_RECO_NODE_DATA_DONE = 4, + DLM_RECO_NODE_DATA_FINALIZE_SENT = 5, +}; + + +enum { + DLM_MASTER_RESP_NO = 0, + DLM_MASTER_RESP_YES = 1, + DLM_MASTER_RESP_MAYBE = 2, + DLM_MASTER_RESP_ERROR = 3, +}; + + +struct dlm_master_request +{ + u8 node_idx; + u8 namelen; + __be16 pad1; + __be32 flags; + + u8 name[O2NM_MAX_NAME_LEN]; +}; + +#define DLM_ASSERT_RESPONSE_REASSERT 0x00000001 +#define DLM_ASSERT_RESPONSE_MASTERY_REF 0x00000002 + +#define DLM_ASSERT_MASTER_MLE_CLEANUP 0x00000001 +#define DLM_ASSERT_MASTER_REQUERY 0x00000002 +#define DLM_ASSERT_MASTER_FINISH_MIGRATION 0x00000004 +struct dlm_assert_master +{ + u8 node_idx; + u8 namelen; + __be16 pad1; + __be32 flags; + + u8 name[O2NM_MAX_NAME_LEN]; +}; + +#define DLM_MIGRATE_RESPONSE_MASTERY_REF 0x00000001 + +struct dlm_migrate_request +{ + u8 master; + u8 new_master; + u8 namelen; + u8 pad1; + __be32 pad2; + u8 name[O2NM_MAX_NAME_LEN]; +}; + +struct dlm_master_requery +{ + u8 pad1; + u8 pad2; + u8 node_idx; + u8 namelen; + __be32 pad3; + u8 name[O2NM_MAX_NAME_LEN]; +}; + +#define DLM_MRES_RECOVERY 0x01 +#define DLM_MRES_MIGRATION 0x02 +#define DLM_MRES_ALL_DONE 0x04 + +/* + * We would like to get one whole lockres into a single network + * message whenever possible. Generally speaking, there will be + * at most one dlm_lock on a lockres for each node in the cluster, + * plus (infrequently) any additional locks coming in from userdlm. + * + * struct _dlm_lockres_page + * { + * dlm_migratable_lockres mres; + * dlm_migratable_lock ml[DLM_MAX_MIGRATABLE_LOCKS]; + * u8 pad[DLM_MIG_LOCKRES_RESERVED]; + * }; + * + * from ../cluster/tcp.h + * O2NET_MAX_PAYLOAD_BYTES (4096 - sizeof(net_msg)) + * (roughly 4080 bytes) + * and sizeof(dlm_migratable_lockres) = 112 bytes + * and sizeof(dlm_migratable_lock) = 16 bytes + * + * Choosing DLM_MAX_MIGRATABLE_LOCKS=240 and + * DLM_MIG_LOCKRES_RESERVED=128 means we have this: + * + * (DLM_MAX_MIGRATABLE_LOCKS * sizeof(dlm_migratable_lock)) + + * sizeof(dlm_migratable_lockres) + DLM_MIG_LOCKRES_RESERVED = + * NET_MAX_PAYLOAD_BYTES + * (240 * 16) + 112 + 128 = 4080 + * + * So a lockres would need more than 240 locks before it would + * use more than one network packet to recover. Not too bad. + */ +#define DLM_MAX_MIGRATABLE_LOCKS 240 + +struct dlm_migratable_lockres +{ + u8 master; + u8 lockname_len; + u8 num_locks; // locks sent in this structure + u8 flags; + __be32 total_locks; // locks to be sent for this migration cookie + __be64 mig_cookie; // cookie for this lockres migration + // or zero if not needed + // 16 bytes + u8 lockname[DLM_LOCKID_NAME_MAX]; + // 48 bytes + u8 lvb[DLM_LVB_LEN]; + // 112 bytes + struct dlm_migratable_lock ml[]; // 16 bytes each, begins at byte 112 +}; +#define DLM_MIG_LOCKRES_MAX_LEN \ + (sizeof(struct dlm_migratable_lockres) + \ + (sizeof(struct dlm_migratable_lock) * \ + DLM_MAX_MIGRATABLE_LOCKS) ) + +/* from above, 128 bytes + * for some undetermined future use */ +#define DLM_MIG_LOCKRES_RESERVED (O2NET_MAX_PAYLOAD_BYTES - \ + DLM_MIG_LOCKRES_MAX_LEN) + +struct dlm_create_lock +{ + __be64 cookie; + + __be32 flags; + u8 pad1; + u8 node_idx; + s8 requested_type; + u8 namelen; + + u8 name[O2NM_MAX_NAME_LEN]; +}; + +struct dlm_convert_lock +{ + __be64 cookie; + + __be32 flags; + u8 pad1; + u8 node_idx; + s8 requested_type; + u8 namelen; + + u8 name[O2NM_MAX_NAME_LEN]; + + s8 lvb[]; +}; +#define DLM_CONVERT_LOCK_MAX_LEN (sizeof(struct dlm_convert_lock)+DLM_LVB_LEN) + +struct dlm_unlock_lock +{ + __be64 cookie; + + __be32 flags; + __be16 pad1; + u8 node_idx; + u8 namelen; + + u8 name[O2NM_MAX_NAME_LEN]; + + s8 lvb[]; +}; +#define DLM_UNLOCK_LOCK_MAX_LEN (sizeof(struct dlm_unlock_lock)+DLM_LVB_LEN) + +struct dlm_proxy_ast +{ + __be64 cookie; + + __be32 flags; + u8 node_idx; + u8 type; + u8 blocked_type; + u8 namelen; + + u8 name[O2NM_MAX_NAME_LEN]; + + s8 lvb[]; +}; +#define DLM_PROXY_AST_MAX_LEN (sizeof(struct dlm_proxy_ast)+DLM_LVB_LEN) + +#define DLM_MOD_KEY (0x666c6172) +enum dlm_query_join_response_code { + JOIN_DISALLOW = 0, + JOIN_OK = 1, + JOIN_OK_NO_MAP = 2, + JOIN_PROTOCOL_MISMATCH = 3, +}; + +struct dlm_query_join_packet { + u8 code; /* Response code. dlm_minor and fs_minor + are only valid if this is JOIN_OK */ + u8 dlm_minor; /* The minor version of the protocol the + dlm is speaking. */ + u8 fs_minor; /* The minor version of the protocol the + filesystem is speaking. */ + u8 reserved; +}; + +union dlm_query_join_response { + __be32 intval; + struct dlm_query_join_packet packet; +}; + +struct dlm_lock_request +{ + u8 node_idx; + u8 dead_node; + __be16 pad1; + __be32 pad2; +}; + +struct dlm_reco_data_done +{ + u8 node_idx; + u8 dead_node; + __be16 pad1; + __be32 pad2; + + /* unused for now */ + /* eventually we can use this to attempt + * lvb recovery based on each node's info */ + u8 reco_lvb[DLM_LVB_LEN]; +}; + +struct dlm_begin_reco +{ + u8 node_idx; + u8 dead_node; + __be16 pad1; + __be32 pad2; +}; + +struct dlm_query_join_request +{ + u8 node_idx; + u8 pad1[2]; + u8 name_len; + struct dlm_protocol_version dlm_proto; + struct dlm_protocol_version fs_proto; + u8 domain[O2NM_MAX_NAME_LEN]; + u8 node_map[BITS_TO_BYTES(O2NM_MAX_NODES)]; +}; + +struct dlm_assert_joined +{ + u8 node_idx; + u8 pad1[2]; + u8 name_len; + u8 domain[O2NM_MAX_NAME_LEN]; +}; + +struct dlm_cancel_join +{ + u8 node_idx; + u8 pad1[2]; + u8 name_len; + u8 domain[O2NM_MAX_NAME_LEN]; +}; + +struct dlm_query_region { + u8 qr_node; + u8 qr_numregions; + u8 qr_namelen; + u8 pad1; + u8 qr_domain[O2NM_MAX_NAME_LEN]; + u8 qr_regions[O2HB_MAX_REGION_NAME_LEN * O2NM_MAX_REGIONS]; +}; + +struct dlm_node_info { + u8 ni_nodenum; + u8 pad1; + __be16 ni_ipv4_port; + __be32 ni_ipv4_address; +}; + +struct dlm_query_nodeinfo { + u8 qn_nodenum; + u8 qn_numnodes; + u8 qn_namelen; + u8 pad1; + u8 qn_domain[O2NM_MAX_NAME_LEN]; + struct dlm_node_info qn_nodes[O2NM_MAX_NODES]; +}; + +struct dlm_exit_domain +{ + u8 node_idx; + u8 pad1[3]; +}; + +struct dlm_finalize_reco +{ + u8 node_idx; + u8 dead_node; + u8 flags; + u8 pad1; + __be32 pad2; +}; + +struct dlm_deref_lockres +{ + u32 pad1; + u16 pad2; + u8 node_idx; + u8 namelen; + + u8 name[O2NM_MAX_NAME_LEN]; +}; + +enum { + DLM_DEREF_RESPONSE_DONE = 0, + DLM_DEREF_RESPONSE_INPROG = 1, +}; + +struct dlm_deref_lockres_done { + u32 pad1; + u16 pad2; + u8 node_idx; + u8 namelen; + + u8 name[O2NM_MAX_NAME_LEN]; +}; + +static inline enum dlm_status +__dlm_lockres_state_to_status(struct dlm_lock_resource *res) +{ + enum dlm_status status = DLM_NORMAL; + + assert_spin_locked(&res->spinlock); + + if (res->state & (DLM_LOCK_RES_RECOVERING| + DLM_LOCK_RES_RECOVERY_WAITING)) + status = DLM_RECOVERING; + else if (res->state & DLM_LOCK_RES_MIGRATING) + status = DLM_MIGRATING; + else if (res->state & DLM_LOCK_RES_IN_PROGRESS) + status = DLM_FORWARD; + + return status; +} + +static inline u8 dlm_get_lock_cookie_node(u64 cookie) +{ + u8 ret; + cookie >>= 56; + ret = (u8)(cookie & 0xffULL); + return ret; +} + +static inline unsigned long long dlm_get_lock_cookie_seq(u64 cookie) +{ + unsigned long long ret; + ret = ((unsigned long long)cookie) & 0x00ffffffffffffffULL; + return ret; +} + +struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie, + struct dlm_lockstatus *lksb); +void dlm_lock_get(struct dlm_lock *lock); +void dlm_lock_put(struct dlm_lock *lock); + +void dlm_lock_attach_lockres(struct dlm_lock *lock, + struct dlm_lock_resource *res); + +int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); + +void dlm_revert_pending_convert(struct dlm_lock_resource *res, + struct dlm_lock *lock); +void dlm_revert_pending_lock(struct dlm_lock_resource *res, + struct dlm_lock *lock); + +int dlm_unlock_lock_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +void dlm_commit_pending_cancel(struct dlm_lock_resource *res, + struct dlm_lock *lock); +void dlm_commit_pending_unlock(struct dlm_lock_resource *res, + struct dlm_lock *lock); + +int dlm_launch_thread(struct dlm_ctxt *dlm); +void dlm_complete_thread(struct dlm_ctxt *dlm); +int dlm_launch_recovery_thread(struct dlm_ctxt *dlm); +void dlm_complete_recovery_thread(struct dlm_ctxt *dlm); +void dlm_wait_for_recovery(struct dlm_ctxt *dlm); +void dlm_kick_recovery_thread(struct dlm_ctxt *dlm); +int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node); +void dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout); +void dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout); + +void dlm_put(struct dlm_ctxt *dlm); +struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm); +int dlm_domain_fully_joined(struct dlm_ctxt *dlm); + +void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); +void dlm_lockres_calc_usage(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); +static inline void dlm_lockres_get(struct dlm_lock_resource *res) +{ + /* This is called on every lookup, so it might be worth + * inlining. */ + kref_get(&res->refs); +} +void dlm_lockres_put(struct dlm_lock_resource *res); +void __dlm_unhash_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res); +void __dlm_insert_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res); +struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm, + const char *name, + unsigned int len, + unsigned int hash); +struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm, + const char *name, + unsigned int len, + unsigned int hash); +struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm, + const char *name, + unsigned int len); + +int dlm_is_host_down(int errno); + +struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm, + const char *lockid, + int namelen, + int flags); +struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm, + const char *name, + unsigned int namelen); + +void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, int bit); +void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, int bit); + +void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); +void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); + +void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); + +void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock); +void __dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock); +void __dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock); +void dlm_do_local_ast(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock); +int dlm_do_remote_ast(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock); +void dlm_do_local_bast(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + int blocked_type); +int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + int msg_type, + int blocked_type, int flags); +static inline int dlm_send_proxy_bast(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + int blocked_type) +{ + return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_BAST, + blocked_type, 0); +} + +static inline int dlm_send_proxy_ast(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + int flags) +{ + return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_AST, + 0, flags); +} + +void dlm_print_one_lock_resource(struct dlm_lock_resource *res); +void __dlm_print_one_lock_resource(struct dlm_lock_resource *res); + +void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res); +void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res); + + +void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data); +void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data); + +int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res); +int dlm_finish_migration(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + u8 old_master); +void dlm_lockres_release_ast(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); +void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res); + +int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +void dlm_assert_master_post_handler(int status, void *data, void *ret_data); +int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_deref_lockres_done_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, + u8 nodenum, u8 *real_master); + +void __dlm_do_purge_lockres(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); + +int dlm_dispatch_assert_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + int ignore_higher, + u8 request_from, + u32 flags); + + +int dlm_send_one_lockres(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_migratable_lockres *mres, + u8 send_to, + u8 flags); +void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); + +/* will exit holding res->spinlock, but may drop in function */ +void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags); + +/* will exit holding res->spinlock, but may drop in function */ +static inline void __dlm_wait_on_lockres(struct dlm_lock_resource *res) +{ + __dlm_wait_on_lockres_flags(res, (DLM_LOCK_RES_IN_PROGRESS| + DLM_LOCK_RES_RECOVERING| + DLM_LOCK_RES_RECOVERY_WAITING| + DLM_LOCK_RES_MIGRATING)); +} + +void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle); +void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle); + +/* create/destroy slab caches */ +int dlm_init_master_caches(void); +void dlm_destroy_master_caches(void); + +int dlm_init_lock_cache(void); +void dlm_destroy_lock_cache(void); + +int dlm_init_mle_cache(void); +void dlm_destroy_mle_cache(void); + +void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up); +int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); +void dlm_clean_master_list(struct dlm_ctxt *dlm, + u8 dead_node); +void dlm_force_free_mles(struct dlm_ctxt *dlm); +int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock); +int __dlm_lockres_has_locks(struct dlm_lock_resource *res); +int __dlm_lockres_unused(struct dlm_lock_resource *res); + +static inline const char * dlm_lock_mode_name(int mode) +{ + switch (mode) { + case LKM_EXMODE: + return "EX"; + case LKM_PRMODE: + return "PR"; + case LKM_NLMODE: + return "NL"; + } + return "UNKNOWN"; +} + + +static inline int dlm_lock_compatible(int existing, int request) +{ + /* NO_LOCK compatible with all */ + if (request == LKM_NLMODE || + existing == LKM_NLMODE) + return 1; + + /* EX incompatible with all non-NO_LOCK */ + if (request == LKM_EXMODE) + return 0; + + /* request must be PR, which is compatible with PR */ + if (existing == LKM_PRMODE) + return 1; + + return 0; +} + +static inline int dlm_lock_on_list(struct list_head *head, + struct dlm_lock *lock) +{ + struct dlm_lock *tmplock; + + list_for_each_entry(tmplock, head, list) { + if (tmplock == lock) + return 1; + } + return 0; +} + + +static inline enum dlm_status dlm_err_to_dlm_status(int err) +{ + enum dlm_status ret; + if (err == -ENOMEM) + ret = DLM_SYSERR; + else if (err == -ETIMEDOUT || o2net_link_down(err, NULL)) + ret = DLM_NOLOCKMGR; + else if (err == -EINVAL) + ret = DLM_BADPARAM; + else if (err == -ENAMETOOLONG) + ret = DLM_IVBUFLEN; + else + ret = DLM_BADARGS; + return ret; +} + + +static inline void dlm_node_iter_init(unsigned long *map, + struct dlm_node_iter *iter) +{ + bitmap_copy(iter->node_map, map, O2NM_MAX_NODES); + iter->curnode = -1; +} + +static inline int dlm_node_iter_next(struct dlm_node_iter *iter) +{ + int bit; + bit = find_next_bit(iter->node_map, O2NM_MAX_NODES, iter->curnode+1); + if (bit >= O2NM_MAX_NODES) { + iter->curnode = O2NM_MAX_NODES; + return -ENOENT; + } + iter->curnode = bit; + return bit; +} + +static inline void dlm_set_lockres_owner(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + u8 owner) +{ + assert_spin_locked(&res->spinlock); + + res->owner = owner; +} + +static inline void dlm_change_lockres_owner(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + u8 owner) +{ + assert_spin_locked(&res->spinlock); + + if (owner != res->owner) + dlm_set_lockres_owner(dlm, res, owner); +} + +#endif /* DLMCOMMON_H */ diff --git a/fs/ocfs2/dlm/dlmconvert.c b/fs/ocfs2/dlm/dlmconvert.c new file mode 100644 index 000000000..450d46eef --- /dev/null +++ b/fs/ocfs2/dlm/dlmconvert.c @@ -0,0 +1,555 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * dlmconvert.c + * + * underlying calls for lock conversion + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + + +#include "../cluster/heartbeat.h" +#include "../cluster/nodemanager.h" +#include "../cluster/tcp.h" + +#include "dlmapi.h" +#include "dlmcommon.h" + +#include "dlmconvert.h" + +#define MLOG_MASK_PREFIX ML_DLM +#include "../cluster/masklog.h" + +/* NOTE: __dlmconvert_master is the only function in here that + * needs a spinlock held on entry (res->spinlock) and it is the + * only one that holds a lock on exit (res->spinlock). + * All other functions in here need no locks and drop all of + * the locks that they acquire. */ +static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags, + int type, int *call_ast, + int *kick_thread); +static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags, int type); + +/* + * this is only called directly by dlmlock(), and only when the + * local node is the owner of the lockres + * locking: + * caller needs: none + * taken: takes and drops res->spinlock + * held on exit: none + * returns: see __dlmconvert_master + */ +enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags, int type) +{ + int call_ast = 0, kick_thread = 0; + enum dlm_status status; + + spin_lock(&res->spinlock); + /* we are not in a network handler, this is fine */ + __dlm_wait_on_lockres(res); + __dlm_lockres_reserve_ast(res); + res->state |= DLM_LOCK_RES_IN_PROGRESS; + + status = __dlmconvert_master(dlm, res, lock, flags, type, + &call_ast, &kick_thread); + + res->state &= ~DLM_LOCK_RES_IN_PROGRESS; + spin_unlock(&res->spinlock); + wake_up(&res->wq); + if (status != DLM_NORMAL && status != DLM_NOTQUEUED) + dlm_error(status); + + /* either queue the ast or release it */ + if (call_ast) + dlm_queue_ast(dlm, lock); + else + dlm_lockres_release_ast(dlm, res); + + if (kick_thread) + dlm_kick_thread(dlm, res); + + return status; +} + +/* performs lock conversion at the lockres master site + * locking: + * caller needs: res->spinlock + * taken: takes and drops lock->spinlock + * held on exit: res->spinlock + * returns: DLM_NORMAL, DLM_NOTQUEUED, DLM_DENIED + * call_ast: whether ast should be called for this lock + * kick_thread: whether dlm_kick_thread should be called + */ +static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags, + int type, int *call_ast, + int *kick_thread) +{ + enum dlm_status status = DLM_NORMAL; + struct dlm_lock *tmplock=NULL; + + assert_spin_locked(&res->spinlock); + + mlog(0, "type=%d, convert_type=%d, new convert_type=%d\n", + lock->ml.type, lock->ml.convert_type, type); + + spin_lock(&lock->spinlock); + + /* already converting? */ + if (lock->ml.convert_type != LKM_IVMODE) { + mlog(ML_ERROR, "attempted to convert a lock with a lock " + "conversion pending\n"); + status = DLM_DENIED; + goto unlock_exit; + } + + /* must be on grant queue to convert */ + if (!dlm_lock_on_list(&res->granted, lock)) { + mlog(ML_ERROR, "attempted to convert a lock not on grant " + "queue\n"); + status = DLM_DENIED; + goto unlock_exit; + } + + if (flags & LKM_VALBLK) { + switch (lock->ml.type) { + case LKM_EXMODE: + /* EX + LKM_VALBLK + convert == set lvb */ + mlog(0, "will set lvb: converting %s->%s\n", + dlm_lock_mode_name(lock->ml.type), + dlm_lock_mode_name(type)); + lock->lksb->flags |= DLM_LKSB_PUT_LVB; + break; + case LKM_PRMODE: + case LKM_NLMODE: + /* refetch if new level is not NL */ + if (type > LKM_NLMODE) { + mlog(0, "will fetch new value into " + "lvb: converting %s->%s\n", + dlm_lock_mode_name(lock->ml.type), + dlm_lock_mode_name(type)); + lock->lksb->flags |= DLM_LKSB_GET_LVB; + } else { + mlog(0, "will NOT fetch new value " + "into lvb: converting %s->%s\n", + dlm_lock_mode_name(lock->ml.type), + dlm_lock_mode_name(type)); + flags &= ~(LKM_VALBLK); + } + break; + } + } + + + /* in-place downconvert? */ + if (type <= lock->ml.type) + goto grant; + + /* upconvert from here on */ + status = DLM_NORMAL; + list_for_each_entry(tmplock, &res->granted, list) { + if (tmplock == lock) + continue; + if (!dlm_lock_compatible(tmplock->ml.type, type)) + goto switch_queues; + } + + list_for_each_entry(tmplock, &res->converting, list) { + if (!dlm_lock_compatible(tmplock->ml.type, type)) + goto switch_queues; + /* existing conversion requests take precedence */ + if (!dlm_lock_compatible(tmplock->ml.convert_type, type)) + goto switch_queues; + } + + /* fall thru to grant */ + +grant: + mlog(0, "res %.*s, granting %s lock\n", res->lockname.len, + res->lockname.name, dlm_lock_mode_name(type)); + /* immediately grant the new lock type */ + lock->lksb->status = DLM_NORMAL; + if (lock->ml.node == dlm->node_num) + mlog(0, "doing in-place convert for nonlocal lock\n"); + lock->ml.type = type; + if (lock->lksb->flags & DLM_LKSB_PUT_LVB) + memcpy(res->lvb, lock->lksb->lvb, DLM_LVB_LEN); + + /* + * Move the lock to the tail because it may be the only lock which has + * an invalid lvb. + */ + list_move_tail(&lock->list, &res->granted); + + status = DLM_NORMAL; + *call_ast = 1; + goto unlock_exit; + +switch_queues: + if (flags & LKM_NOQUEUE) { + mlog(0, "failed to convert NOQUEUE lock %.*s from " + "%d to %d...\n", res->lockname.len, res->lockname.name, + lock->ml.type, type); + status = DLM_NOTQUEUED; + goto unlock_exit; + } + mlog(0, "res %.*s, queueing...\n", res->lockname.len, + res->lockname.name); + + lock->ml.convert_type = type; + /* do not alter lock refcount. switching lists. */ + list_move_tail(&lock->list, &res->converting); + +unlock_exit: + spin_unlock(&lock->spinlock); + if (status == DLM_DENIED) { + __dlm_print_one_lock_resource(res); + } + if (status == DLM_NORMAL) + *kick_thread = 1; + return status; +} + +void dlm_revert_pending_convert(struct dlm_lock_resource *res, + struct dlm_lock *lock) +{ + /* do not alter lock refcount. switching lists. */ + list_move_tail(&lock->list, &res->granted); + lock->ml.convert_type = LKM_IVMODE; + lock->lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB); +} + +/* messages the master site to do lock conversion + * locking: + * caller needs: none + * taken: takes and drops res->spinlock, uses DLM_LOCK_RES_IN_PROGRESS + * held on exit: none + * returns: DLM_NORMAL, DLM_RECOVERING, status from remote node + */ +enum dlm_status dlmconvert_remote(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags, int type) +{ + enum dlm_status status; + + mlog(0, "type=%d, convert_type=%d, busy=%d\n", lock->ml.type, + lock->ml.convert_type, res->state & DLM_LOCK_RES_IN_PROGRESS); + + spin_lock(&res->spinlock); + if (res->state & DLM_LOCK_RES_RECOVERING) { + mlog(0, "bailing out early since res is RECOVERING " + "on secondary queue\n"); + /* __dlm_print_one_lock_resource(res); */ + status = DLM_RECOVERING; + goto bail; + } + /* will exit this call with spinlock held */ + __dlm_wait_on_lockres(res); + + if (lock->ml.convert_type != LKM_IVMODE) { + __dlm_print_one_lock_resource(res); + mlog(ML_ERROR, "converting a remote lock that is already " + "converting! (cookie=%u:%llu, conv=%d)\n", + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), + lock->ml.convert_type); + status = DLM_DENIED; + goto bail; + } + + if (lock->ml.type == type && lock->ml.convert_type == LKM_IVMODE) { + mlog(0, "last convert request returned DLM_RECOVERING, but " + "owner has already queued and sent ast to me. res %.*s, " + "(cookie=%u:%llu, type=%d, conv=%d)\n", + res->lockname.len, res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), + lock->ml.type, lock->ml.convert_type); + status = DLM_NORMAL; + goto bail; + } + + res->state |= DLM_LOCK_RES_IN_PROGRESS; + /* move lock to local convert queue */ + /* do not alter lock refcount. switching lists. */ + list_move_tail(&lock->list, &res->converting); + lock->convert_pending = 1; + lock->ml.convert_type = type; + + if (flags & LKM_VALBLK) { + if (lock->ml.type == LKM_EXMODE) { + flags |= LKM_PUT_LVB; + lock->lksb->flags |= DLM_LKSB_PUT_LVB; + } else { + if (lock->ml.convert_type == LKM_NLMODE) + flags &= ~LKM_VALBLK; + else { + flags |= LKM_GET_LVB; + lock->lksb->flags |= DLM_LKSB_GET_LVB; + } + } + } + spin_unlock(&res->spinlock); + + /* no locks held here. + * need to wait for a reply as to whether it got queued or not. */ + status = dlm_send_remote_convert_request(dlm, res, lock, flags, type); + + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_IN_PROGRESS; + /* if it failed, move it back to granted queue. + * if master returns DLM_NORMAL and then down before sending ast, + * it may have already been moved to granted queue, reset to + * DLM_RECOVERING and retry convert */ + if (status != DLM_NORMAL) { + if (status != DLM_NOTQUEUED) + dlm_error(status); + dlm_revert_pending_convert(res, lock); + } else if (!lock->convert_pending) { + mlog(0, "%s: res %.*s, owner died and lock has been moved back " + "to granted list, retry convert.\n", + dlm->name, res->lockname.len, res->lockname.name); + status = DLM_RECOVERING; + } + + lock->convert_pending = 0; +bail: + spin_unlock(&res->spinlock); + + /* TODO: should this be a wake_one? */ + /* wake up any IN_PROGRESS waiters */ + wake_up(&res->wq); + + return status; +} + +/* sends DLM_CONVERT_LOCK_MSG to master site + * locking: + * caller needs: none + * taken: none + * held on exit: none + * returns: DLM_NOLOCKMGR, status from remote node + */ +static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags, int type) +{ + struct dlm_convert_lock convert; + int tmpret; + enum dlm_status ret; + int status = 0; + struct kvec vec[2]; + size_t veclen = 1; + + mlog(0, "%.*s\n", res->lockname.len, res->lockname.name); + + memset(&convert, 0, sizeof(struct dlm_convert_lock)); + convert.node_idx = dlm->node_num; + convert.requested_type = type; + convert.cookie = lock->ml.cookie; + convert.namelen = res->lockname.len; + convert.flags = cpu_to_be32(flags); + memcpy(convert.name, res->lockname.name, convert.namelen); + + vec[0].iov_len = sizeof(struct dlm_convert_lock); + vec[0].iov_base = &convert; + + if (flags & LKM_PUT_LVB) { + /* extra data to send if we are updating lvb */ + vec[1].iov_len = DLM_LVB_LEN; + vec[1].iov_base = lock->lksb->lvb; + veclen++; + } + + tmpret = o2net_send_message_vec(DLM_CONVERT_LOCK_MSG, dlm->key, + vec, veclen, res->owner, &status); + if (tmpret >= 0) { + // successfully sent and received + ret = status; // this is already a dlm_status + if (ret == DLM_RECOVERING) { + mlog(0, "node %u returned DLM_RECOVERING from convert " + "message!\n", res->owner); + } else if (ret == DLM_MIGRATING) { + mlog(0, "node %u returned DLM_MIGRATING from convert " + "message!\n", res->owner); + } else if (ret == DLM_FORWARD) { + mlog(0, "node %u returned DLM_FORWARD from convert " + "message!\n", res->owner); + } else if (ret != DLM_NORMAL && ret != DLM_NOTQUEUED) + dlm_error(ret); + } else { + mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to " + "node %u\n", tmpret, DLM_CONVERT_LOCK_MSG, dlm->key, + res->owner); + if (dlm_is_host_down(tmpret)) { + /* instead of logging the same network error over + * and over, sleep here and wait for the heartbeat + * to notice the node is dead. times out after 5s. */ + dlm_wait_for_node_death(dlm, res->owner, + DLM_NODE_DEATH_WAIT_MAX); + ret = DLM_RECOVERING; + mlog(0, "node %u died so returning DLM_RECOVERING " + "from convert message!\n", res->owner); + } else { + ret = dlm_err_to_dlm_status(tmpret); + } + } + + return ret; +} + +/* handler for DLM_CONVERT_LOCK_MSG on master site + * locking: + * caller needs: none + * taken: takes and drop res->spinlock + * held on exit: none + * returns: DLM_NORMAL, DLM_IVLOCKID, DLM_BADARGS, + * status from __dlmconvert_master + */ +int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_convert_lock *cnv = (struct dlm_convert_lock *)msg->buf; + struct dlm_lock_resource *res = NULL; + struct dlm_lock *lock = NULL; + struct dlm_lock *tmp_lock; + struct dlm_lockstatus *lksb; + enum dlm_status status = DLM_NORMAL; + u32 flags; + int call_ast = 0, kick_thread = 0, ast_reserved = 0, wake = 0; + + if (!dlm_grab(dlm)) { + dlm_error(DLM_REJECTED); + return DLM_REJECTED; + } + + mlog_bug_on_msg(!dlm_domain_fully_joined(dlm), + "Domain %s not fully joined!\n", dlm->name); + + if (cnv->namelen > DLM_LOCKID_NAME_MAX) { + status = DLM_IVBUFLEN; + dlm_error(status); + goto leave; + } + + flags = be32_to_cpu(cnv->flags); + + if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) == + (LKM_PUT_LVB|LKM_GET_LVB)) { + mlog(ML_ERROR, "both PUT and GET lvb specified\n"); + status = DLM_BADARGS; + goto leave; + } + + mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" : + (flags & LKM_GET_LVB ? "get lvb" : "none")); + + status = DLM_IVLOCKID; + res = dlm_lookup_lockres(dlm, cnv->name, cnv->namelen); + if (!res) { + dlm_error(status); + goto leave; + } + + spin_lock(&res->spinlock); + status = __dlm_lockres_state_to_status(res); + if (status != DLM_NORMAL) { + spin_unlock(&res->spinlock); + dlm_error(status); + goto leave; + } + list_for_each_entry(tmp_lock, &res->granted, list) { + if (tmp_lock->ml.cookie == cnv->cookie && + tmp_lock->ml.node == cnv->node_idx) { + lock = tmp_lock; + dlm_lock_get(lock); + break; + } + } + spin_unlock(&res->spinlock); + if (!lock) { + status = DLM_IVLOCKID; + mlog(ML_ERROR, "did not find lock to convert on grant queue! " + "cookie=%u:%llu\n", + dlm_get_lock_cookie_node(be64_to_cpu(cnv->cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(cnv->cookie))); + dlm_print_one_lock_resource(res); + goto leave; + } + + /* found the lock */ + lksb = lock->lksb; + + /* see if caller needed to get/put lvb */ + if (flags & LKM_PUT_LVB) { + BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB)); + lksb->flags |= DLM_LKSB_PUT_LVB; + memcpy(&lksb->lvb[0], &cnv->lvb[0], DLM_LVB_LEN); + } else if (flags & LKM_GET_LVB) { + BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB)); + lksb->flags |= DLM_LKSB_GET_LVB; + } + + spin_lock(&res->spinlock); + status = __dlm_lockres_state_to_status(res); + if (status == DLM_NORMAL) { + __dlm_lockres_reserve_ast(res); + ast_reserved = 1; + res->state |= DLM_LOCK_RES_IN_PROGRESS; + status = __dlmconvert_master(dlm, res, lock, flags, + cnv->requested_type, + &call_ast, &kick_thread); + res->state &= ~DLM_LOCK_RES_IN_PROGRESS; + wake = 1; + } + spin_unlock(&res->spinlock); + if (wake) + wake_up(&res->wq); + + if (status != DLM_NORMAL) { + if (status != DLM_NOTQUEUED) + dlm_error(status); + lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB); + } + +leave: + if (lock) + dlm_lock_put(lock); + + /* either queue the ast or release it, if reserved */ + if (call_ast) + dlm_queue_ast(dlm, lock); + else if (ast_reserved) + dlm_lockres_release_ast(dlm, res); + + if (kick_thread) + dlm_kick_thread(dlm, res); + + if (res) + dlm_lockres_put(res); + + dlm_put(dlm); + + return status; +} diff --git a/fs/ocfs2/dlm/dlmconvert.h b/fs/ocfs2/dlm/dlmconvert.h new file mode 100644 index 000000000..1f3717165 --- /dev/null +++ b/fs/ocfs2/dlm/dlmconvert.h @@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: GPL-2.0-or-later */ +/* + * dlmconvert.h + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + +#ifndef DLMCONVERT_H +#define DLMCONVERT_H + +enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags, int type); +enum dlm_status dlmconvert_remote(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags, int type); + +#endif diff --git a/fs/ocfs2/dlm/dlmdebug.c b/fs/ocfs2/dlm/dlmdebug.c new file mode 100644 index 000000000..ce6d301cb --- /dev/null +++ b/fs/ocfs2/dlm/dlmdebug.c @@ -0,0 +1,844 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * dlmdebug.c + * + * debug functionality for the dlm + * + * Copyright (C) 2004, 2008 Oracle. All rights reserved. + */ + +#include +#include +#include +#include +#include +#include +#include +#include + +#include "../cluster/heartbeat.h" +#include "../cluster/nodemanager.h" +#include "../cluster/tcp.h" + +#include "dlmapi.h" +#include "dlmcommon.h" +#include "dlmdomain.h" +#include "dlmdebug.h" + +#define MLOG_MASK_PREFIX ML_DLM +#include "../cluster/masklog.h" + +static int stringify_lockname(const char *lockname, int locklen, char *buf, + int len); + +void dlm_print_one_lock_resource(struct dlm_lock_resource *res) +{ + spin_lock(&res->spinlock); + __dlm_print_one_lock_resource(res); + spin_unlock(&res->spinlock); +} + +static void dlm_print_lockres_refmap(struct dlm_lock_resource *res) +{ + int bit; + assert_spin_locked(&res->spinlock); + + printk(" refmap nodes: [ "); + bit = 0; + while (1) { + bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit); + if (bit >= O2NM_MAX_NODES) + break; + printk("%u ", bit); + bit++; + } + printk("], inflight=%u\n", res->inflight_locks); +} + +static void __dlm_print_lock(struct dlm_lock *lock) +{ + spin_lock(&lock->spinlock); + + printk(" type=%d, conv=%d, node=%u, cookie=%u:%llu, " + "ref=%u, ast=(empty=%c,pend=%c), bast=(empty=%c,pend=%c), " + "pending=(conv=%c,lock=%c,cancel=%c,unlock=%c)\n", + lock->ml.type, lock->ml.convert_type, lock->ml.node, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), + kref_read(&lock->lock_refs), + (list_empty(&lock->ast_list) ? 'y' : 'n'), + (lock->ast_pending ? 'y' : 'n'), + (list_empty(&lock->bast_list) ? 'y' : 'n'), + (lock->bast_pending ? 'y' : 'n'), + (lock->convert_pending ? 'y' : 'n'), + (lock->lock_pending ? 'y' : 'n'), + (lock->cancel_pending ? 'y' : 'n'), + (lock->unlock_pending ? 'y' : 'n')); + + spin_unlock(&lock->spinlock); +} + +void __dlm_print_one_lock_resource(struct dlm_lock_resource *res) +{ + struct dlm_lock *lock; + char buf[DLM_LOCKID_NAME_MAX]; + + assert_spin_locked(&res->spinlock); + + stringify_lockname(res->lockname.name, res->lockname.len, + buf, sizeof(buf)); + printk("lockres: %s, owner=%u, state=%u\n", + buf, res->owner, res->state); + printk(" last used: %lu, refcnt: %u, on purge list: %s\n", + res->last_used, kref_read(&res->refs), + str_no_yes(list_empty(&res->purge))); + printk(" on dirty list: %s, on reco list: %s, " + "migrating pending: %s\n", + str_no_yes(list_empty(&res->dirty)), + str_no_yes(list_empty(&res->recovering)), + str_yes_no(res->migration_pending)); + printk(" inflight locks: %d, asts reserved: %d\n", + res->inflight_locks, atomic_read(&res->asts_reserved)); + dlm_print_lockres_refmap(res); + printk(" granted queue:\n"); + list_for_each_entry(lock, &res->granted, list) { + __dlm_print_lock(lock); + } + printk(" converting queue:\n"); + list_for_each_entry(lock, &res->converting, list) { + __dlm_print_lock(lock); + } + printk(" blocked queue:\n"); + list_for_each_entry(lock, &res->blocked, list) { + __dlm_print_lock(lock); + } +} + +void dlm_print_one_lock(struct dlm_lock *lockid) +{ + dlm_print_one_lock_resource(lockid->lockres); +} +EXPORT_SYMBOL_GPL(dlm_print_one_lock); + +static const char *dlm_errnames[] = { + [DLM_NORMAL] = "DLM_NORMAL", + [DLM_GRANTED] = "DLM_GRANTED", + [DLM_DENIED] = "DLM_DENIED", + [DLM_DENIED_NOLOCKS] = "DLM_DENIED_NOLOCKS", + [DLM_WORKING] = "DLM_WORKING", + [DLM_BLOCKED] = "DLM_BLOCKED", + [DLM_BLOCKED_ORPHAN] = "DLM_BLOCKED_ORPHAN", + [DLM_DENIED_GRACE_PERIOD] = "DLM_DENIED_GRACE_PERIOD", + [DLM_SYSERR] = "DLM_SYSERR", + [DLM_NOSUPPORT] = "DLM_NOSUPPORT", + [DLM_CANCELGRANT] = "DLM_CANCELGRANT", + [DLM_IVLOCKID] = "DLM_IVLOCKID", + [DLM_SYNC] = "DLM_SYNC", + [DLM_BADTYPE] = "DLM_BADTYPE", + [DLM_BADRESOURCE] = "DLM_BADRESOURCE", + [DLM_MAXHANDLES] = "DLM_MAXHANDLES", + [DLM_NOCLINFO] = "DLM_NOCLINFO", + [DLM_NOLOCKMGR] = "DLM_NOLOCKMGR", + [DLM_NOPURGED] = "DLM_NOPURGED", + [DLM_BADARGS] = "DLM_BADARGS", + [DLM_VOID] = "DLM_VOID", + [DLM_NOTQUEUED] = "DLM_NOTQUEUED", + [DLM_IVBUFLEN] = "DLM_IVBUFLEN", + [DLM_CVTUNGRANT] = "DLM_CVTUNGRANT", + [DLM_BADPARAM] = "DLM_BADPARAM", + [DLM_VALNOTVALID] = "DLM_VALNOTVALID", + [DLM_REJECTED] = "DLM_REJECTED", + [DLM_ABORT] = "DLM_ABORT", + [DLM_CANCEL] = "DLM_CANCEL", + [DLM_IVRESHANDLE] = "DLM_IVRESHANDLE", + [DLM_DEADLOCK] = "DLM_DEADLOCK", + [DLM_DENIED_NOASTS] = "DLM_DENIED_NOASTS", + [DLM_FORWARD] = "DLM_FORWARD", + [DLM_TIMEOUT] = "DLM_TIMEOUT", + [DLM_IVGROUPID] = "DLM_IVGROUPID", + [DLM_VERS_CONFLICT] = "DLM_VERS_CONFLICT", + [DLM_BAD_DEVICE_PATH] = "DLM_BAD_DEVICE_PATH", + [DLM_NO_DEVICE_PERMISSION] = "DLM_NO_DEVICE_PERMISSION", + [DLM_NO_CONTROL_DEVICE ] = "DLM_NO_CONTROL_DEVICE ", + [DLM_RECOVERING] = "DLM_RECOVERING", + [DLM_MIGRATING] = "DLM_MIGRATING", + [DLM_MAXSTATS] = "DLM_MAXSTATS", +}; + +const char *dlm_errname(enum dlm_status err) +{ + if (err >= DLM_MAXSTATS || err < 0) + return dlm_errnames[DLM_MAXSTATS]; + return dlm_errnames[err]; +} +EXPORT_SYMBOL_GPL(dlm_errname); + +/* NOTE: This function converts a lockname into a string. It uses knowledge + * of the format of the lockname that should be outside the purview of the dlm. + * We are adding only to make dlm debugging slightly easier. + * + * For more on lockname formats, please refer to dlmglue.c and ocfs2_lockid.h. + */ +static int stringify_lockname(const char *lockname, int locklen, char *buf, + int len) +{ + int out = 0; + __be64 inode_blkno_be; + +#define OCFS2_DENTRY_LOCK_INO_START 18 + if (*lockname == 'N') { + memcpy((__be64 *)&inode_blkno_be, + (char *)&lockname[OCFS2_DENTRY_LOCK_INO_START], + sizeof(__be64)); + out += scnprintf(buf + out, len - out, "%.*s%08x", + OCFS2_DENTRY_LOCK_INO_START - 1, lockname, + (unsigned int)be64_to_cpu(inode_blkno_be)); + } else + out += scnprintf(buf + out, len - out, "%.*s", + locklen, lockname); + return out; +} + +static int stringify_nodemap(unsigned long *nodemap, int maxnodes, + char *buf, int len) +{ + int out = 0; + int i = -1; + + while ((i = find_next_bit(nodemap, maxnodes, i + 1)) < maxnodes) + out += scnprintf(buf + out, len - out, "%d ", i); + + return out; +} + +static int dump_mle(struct dlm_master_list_entry *mle, char *buf, int len) +{ + int out = 0; + char *mle_type; + + if (mle->type == DLM_MLE_BLOCK) + mle_type = "BLK"; + else if (mle->type == DLM_MLE_MASTER) + mle_type = "MAS"; + else + mle_type = "MIG"; + + out += stringify_lockname(mle->mname, mle->mnamelen, buf + out, len - out); + out += scnprintf(buf + out, len - out, + "\t%3s\tmas=%3u\tnew=%3u\tevt=%1d\tuse=%1d\tref=%3d\n", + mle_type, mle->master, mle->new_master, + !list_empty(&mle->hb_events), + !!mle->inuse, + kref_read(&mle->mle_refs)); + + out += scnprintf(buf + out, len - out, "Maybe="); + out += stringify_nodemap(mle->maybe_map, O2NM_MAX_NODES, + buf + out, len - out); + out += scnprintf(buf + out, len - out, "\n"); + + out += scnprintf(buf + out, len - out, "Vote="); + out += stringify_nodemap(mle->vote_map, O2NM_MAX_NODES, + buf + out, len - out); + out += scnprintf(buf + out, len - out, "\n"); + + out += scnprintf(buf + out, len - out, "Response="); + out += stringify_nodemap(mle->response_map, O2NM_MAX_NODES, + buf + out, len - out); + out += scnprintf(buf + out, len - out, "\n"); + + out += scnprintf(buf + out, len - out, "Node="); + out += stringify_nodemap(mle->node_map, O2NM_MAX_NODES, + buf + out, len - out); + out += scnprintf(buf + out, len - out, "\n"); + + out += scnprintf(buf + out, len - out, "\n"); + + return out; +} + +void dlm_print_one_mle(struct dlm_master_list_entry *mle) +{ + char *buf; + + buf = kzalloc(PAGE_SIZE, GFP_ATOMIC); + if (buf) { + dump_mle(mle, buf, PAGE_SIZE - 1); + kfree(buf); + } +} + +#ifdef CONFIG_DEBUG_FS + +static struct dentry *dlm_debugfs_root; + +#define DLM_DEBUGFS_DIR "o2dlm" +#define DLM_DEBUGFS_DLM_STATE "dlm_state" +#define DLM_DEBUGFS_LOCKING_STATE "locking_state" +#define DLM_DEBUGFS_MLE_STATE "mle_state" +#define DLM_DEBUGFS_PURGE_LIST "purge_list" + +/* begin - utils funcs */ +static int debug_release(struct inode *inode, struct file *file) +{ + kfree(file->private_data); + return 0; +} + +static ssize_t debug_read(struct file *file, char __user *buf, + size_t nbytes, loff_t *ppos) +{ + return simple_read_from_buffer(buf, nbytes, ppos, file->private_data, + i_size_read(file->f_mapping->host)); +} +/* end - util funcs */ + +/* begin - purge list funcs */ +static int debug_purgelist_print(struct dlm_ctxt *dlm, char *buf, int len) +{ + struct dlm_lock_resource *res; + int out = 0; + unsigned long total = 0; + + out += scnprintf(buf + out, len - out, + "Dumping Purgelist for Domain: %s\n", dlm->name); + + spin_lock(&dlm->spinlock); + list_for_each_entry(res, &dlm->purge_list, purge) { + ++total; + if (len - out < 100) + continue; + spin_lock(&res->spinlock); + out += stringify_lockname(res->lockname.name, + res->lockname.len, + buf + out, len - out); + out += scnprintf(buf + out, len - out, "\t%ld\n", + (jiffies - res->last_used)/HZ); + spin_unlock(&res->spinlock); + } + spin_unlock(&dlm->spinlock); + + out += scnprintf(buf + out, len - out, "Total on list: %lu\n", total); + + return out; +} + +static int debug_purgelist_open(struct inode *inode, struct file *file) +{ + struct dlm_ctxt *dlm = inode->i_private; + char *buf = NULL; + + buf = kzalloc(PAGE_SIZE, GFP_NOFS); + if (!buf) + return -ENOMEM; + + i_size_write(inode, debug_purgelist_print(dlm, buf, PAGE_SIZE - 1)); + + file->private_data = buf; + + return 0; +} + +static const struct file_operations debug_purgelist_fops = { + .open = debug_purgelist_open, + .release = debug_release, + .read = debug_read, + .llseek = generic_file_llseek, +}; +/* end - purge list funcs */ + +/* begin - debug mle funcs */ +static int debug_mle_print(struct dlm_ctxt *dlm, char *buf, int len) +{ + struct dlm_master_list_entry *mle; + struct hlist_head *bucket; + int i, out = 0; + unsigned long total = 0, longest = 0, bucket_count = 0; + + out += scnprintf(buf + out, len - out, + "Dumping MLEs for Domain: %s\n", dlm->name); + + spin_lock(&dlm->master_lock); + for (i = 0; i < DLM_HASH_BUCKETS; i++) { + bucket = dlm_master_hash(dlm, i); + hlist_for_each_entry(mle, bucket, master_hash_node) { + ++total; + ++bucket_count; + if (len - out < 200) + continue; + out += dump_mle(mle, buf + out, len - out); + } + longest = max(longest, bucket_count); + bucket_count = 0; + } + spin_unlock(&dlm->master_lock); + + out += scnprintf(buf + out, len - out, + "Total: %lu, Longest: %lu\n", total, longest); + return out; +} + +static int debug_mle_open(struct inode *inode, struct file *file) +{ + struct dlm_ctxt *dlm = inode->i_private; + char *buf = NULL; + + buf = kzalloc(PAGE_SIZE, GFP_NOFS); + if (!buf) + return -ENOMEM; + + i_size_write(inode, debug_mle_print(dlm, buf, PAGE_SIZE - 1)); + + file->private_data = buf; + + return 0; +} + +static const struct file_operations debug_mle_fops = { + .open = debug_mle_open, + .release = debug_release, + .read = debug_read, + .llseek = generic_file_llseek, +}; + +/* end - debug mle funcs */ + +/* begin - debug lockres funcs */ +static int dump_lock(struct dlm_lock *lock, int list_type, char *buf, int len) +{ + int out; + +#define DEBUG_LOCK_VERSION 1 + spin_lock(&lock->spinlock); + out = scnprintf(buf, len, "LOCK:%d,%d,%d,%d,%d,%d:%lld,%d,%d,%d,%d,%d," + "%d,%d,%d,%d\n", + DEBUG_LOCK_VERSION, + list_type, lock->ml.type, lock->ml.convert_type, + lock->ml.node, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), + !list_empty(&lock->ast_list), + !list_empty(&lock->bast_list), + lock->ast_pending, lock->bast_pending, + lock->convert_pending, lock->lock_pending, + lock->cancel_pending, lock->unlock_pending, + kref_read(&lock->lock_refs)); + spin_unlock(&lock->spinlock); + + return out; +} + +static int dump_lockres(struct dlm_lock_resource *res, char *buf, int len) +{ + struct dlm_lock *lock; + int i; + int out = 0; + + out += scnprintf(buf + out, len - out, "NAME:"); + out += stringify_lockname(res->lockname.name, res->lockname.len, + buf + out, len - out); + out += scnprintf(buf + out, len - out, "\n"); + +#define DEBUG_LRES_VERSION 1 + out += scnprintf(buf + out, len - out, + "LRES:%d,%d,%d,%ld,%d,%d,%d,%d,%d,%d,%d\n", + DEBUG_LRES_VERSION, + res->owner, res->state, res->last_used, + !list_empty(&res->purge), + !list_empty(&res->dirty), + !list_empty(&res->recovering), + res->inflight_locks, res->migration_pending, + atomic_read(&res->asts_reserved), + kref_read(&res->refs)); + + /* refmap */ + out += scnprintf(buf + out, len - out, "RMAP:"); + out += stringify_nodemap(res->refmap, O2NM_MAX_NODES, + buf + out, len - out); + out += scnprintf(buf + out, len - out, "\n"); + + /* lvb */ + out += scnprintf(buf + out, len - out, "LVBX:"); + for (i = 0; i < DLM_LVB_LEN; i++) + out += scnprintf(buf + out, len - out, + "%02x", (unsigned char)res->lvb[i]); + out += scnprintf(buf + out, len - out, "\n"); + + /* granted */ + list_for_each_entry(lock, &res->granted, list) + out += dump_lock(lock, 0, buf + out, len - out); + + /* converting */ + list_for_each_entry(lock, &res->converting, list) + out += dump_lock(lock, 1, buf + out, len - out); + + /* blocked */ + list_for_each_entry(lock, &res->blocked, list) + out += dump_lock(lock, 2, buf + out, len - out); + + out += scnprintf(buf + out, len - out, "\n"); + + return out; +} + +static void *lockres_seq_start(struct seq_file *m, loff_t *pos) +{ + struct debug_lockres *dl = m->private; + struct dlm_ctxt *dlm = dl->dl_ctxt; + struct dlm_lock_resource *oldres = dl->dl_res; + struct dlm_lock_resource *res = NULL, *iter; + struct list_head *track_list; + + spin_lock(&dlm->track_lock); + if (oldres) + track_list = &oldres->tracking; + else { + track_list = &dlm->tracking_list; + if (list_empty(track_list)) { + dl = NULL; + spin_unlock(&dlm->track_lock); + goto bail; + } + } + + list_for_each_entry(iter, track_list, tracking) { + if (&iter->tracking != &dlm->tracking_list) { + dlm_lockres_get(iter); + res = iter; + } + break; + } + spin_unlock(&dlm->track_lock); + + if (oldres) + dlm_lockres_put(oldres); + + dl->dl_res = res; + + if (res) { + spin_lock(&res->spinlock); + dump_lockres(res, dl->dl_buf, dl->dl_len - 1); + spin_unlock(&res->spinlock); + } else + dl = NULL; + +bail: + /* passed to seq_show */ + return dl; +} + +static void lockres_seq_stop(struct seq_file *m, void *v) +{ +} + +static void *lockres_seq_next(struct seq_file *m, void *v, loff_t *pos) +{ + return NULL; +} + +static int lockres_seq_show(struct seq_file *s, void *v) +{ + struct debug_lockres *dl = (struct debug_lockres *)v; + + seq_printf(s, "%s", dl->dl_buf); + + return 0; +} + +static const struct seq_operations debug_lockres_ops = { + .start = lockres_seq_start, + .stop = lockres_seq_stop, + .next = lockres_seq_next, + .show = lockres_seq_show, +}; + +static int debug_lockres_open(struct inode *inode, struct file *file) +{ + struct dlm_ctxt *dlm = inode->i_private; + struct debug_lockres *dl; + void *buf; + int status = -ENOMEM; + + buf = kmalloc(PAGE_SIZE, GFP_KERNEL); + if (!buf) + goto bail; + + dl = __seq_open_private(file, &debug_lockres_ops, sizeof(*dl)); + if (!dl) + goto bailfree; + + dl->dl_len = PAGE_SIZE; + dl->dl_buf = buf; + + /* ->release uses dl_ctxt after open, so it needs a real pin. */ + dl->dl_ctxt = dlm_grab(dlm); + if (!dl->dl_ctxt) { + status = -ENOENT; + goto bailseq; + } + + return 0; + +bailseq: + seq_release_private(inode, file); +bailfree: + kfree(buf); +bail: + if (status != -ENOENT) + mlog_errno(status); + return status; +} + +static int debug_lockres_release(struct inode *inode, struct file *file) +{ + struct seq_file *seq = file->private_data; + struct debug_lockres *dl = (struct debug_lockres *)seq->private; + + if (dl->dl_res) + dlm_lockres_put(dl->dl_res); + dlm_put(dl->dl_ctxt); + kfree(dl->dl_buf); + return seq_release_private(inode, file); +} + +static const struct file_operations debug_lockres_fops = { + .open = debug_lockres_open, + .release = debug_lockres_release, + .read = seq_read, + .llseek = seq_lseek, +}; +/* end - debug lockres funcs */ + +/* begin - debug state funcs */ +static int debug_state_print(struct dlm_ctxt *dlm, char *buf, int len) +{ + int out = 0; + struct dlm_reco_node_data *node; + char *state; + int cur_mles = 0, tot_mles = 0; + int i; + + spin_lock(&dlm->spinlock); + + switch (dlm->dlm_state) { + case DLM_CTXT_NEW: + state = "NEW"; break; + case DLM_CTXT_JOINED: + state = "JOINED"; break; + case DLM_CTXT_IN_SHUTDOWN: + state = "SHUTDOWN"; break; + case DLM_CTXT_LEAVING: + state = "LEAVING"; break; + default: + state = "UNKNOWN"; break; + } + + /* Domain: xxxxxxxxxx Key: 0xdfbac769 */ + out += scnprintf(buf + out, len - out, + "Domain: %s Key: 0x%08x Protocol: %d.%d\n", + dlm->name, dlm->key, dlm->dlm_locking_proto.pv_major, + dlm->dlm_locking_proto.pv_minor); + + /* Thread Pid: xxx Node: xxx State: xxxxx */ + out += scnprintf(buf + out, len - out, + "Thread Pid: %d Node: %d State: %s\n", + task_pid_nr(dlm->dlm_thread_task), dlm->node_num, state); + + /* Number of Joins: xxx Joining Node: xxx */ + out += scnprintf(buf + out, len - out, + "Number of Joins: %d Joining Node: %d\n", + dlm->num_joins, dlm->joining_node); + + /* Domain Map: xx xx xx */ + out += scnprintf(buf + out, len - out, "Domain Map: "); + out += stringify_nodemap(dlm->domain_map, O2NM_MAX_NODES, + buf + out, len - out); + out += scnprintf(buf + out, len - out, "\n"); + + /* Exit Domain Map: xx xx xx */ + out += scnprintf(buf + out, len - out, "Exit Domain Map: "); + out += stringify_nodemap(dlm->exit_domain_map, O2NM_MAX_NODES, + buf + out, len - out); + out += scnprintf(buf + out, len - out, "\n"); + + /* Live Map: xx xx xx */ + out += scnprintf(buf + out, len - out, "Live Map: "); + out += stringify_nodemap(dlm->live_nodes_map, O2NM_MAX_NODES, + buf + out, len - out); + out += scnprintf(buf + out, len - out, "\n"); + + /* Lock Resources: xxx (xxx) */ + out += scnprintf(buf + out, len - out, + "Lock Resources: %d (%d)\n", + atomic_read(&dlm->res_cur_count), + atomic_read(&dlm->res_tot_count)); + + for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) + tot_mles += atomic_read(&dlm->mle_tot_count[i]); + + for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) + cur_mles += atomic_read(&dlm->mle_cur_count[i]); + + /* MLEs: xxx (xxx) */ + out += scnprintf(buf + out, len - out, + "MLEs: %d (%d)\n", cur_mles, tot_mles); + + /* Blocking: xxx (xxx) */ + out += scnprintf(buf + out, len - out, + " Blocking: %d (%d)\n", + atomic_read(&dlm->mle_cur_count[DLM_MLE_BLOCK]), + atomic_read(&dlm->mle_tot_count[DLM_MLE_BLOCK])); + + /* Mastery: xxx (xxx) */ + out += scnprintf(buf + out, len - out, + " Mastery: %d (%d)\n", + atomic_read(&dlm->mle_cur_count[DLM_MLE_MASTER]), + atomic_read(&dlm->mle_tot_count[DLM_MLE_MASTER])); + + /* Migration: xxx (xxx) */ + out += scnprintf(buf + out, len - out, + " Migration: %d (%d)\n", + atomic_read(&dlm->mle_cur_count[DLM_MLE_MIGRATION]), + atomic_read(&dlm->mle_tot_count[DLM_MLE_MIGRATION])); + + /* Lists: Dirty=Empty Purge=InUse PendingASTs=Empty ... */ + out += scnprintf(buf + out, len - out, + "Lists: Dirty=%s Purge=%s PendingASTs=%s " + "PendingBASTs=%s\n", + (list_empty(&dlm->dirty_list) ? "Empty" : "InUse"), + (list_empty(&dlm->purge_list) ? "Empty" : "InUse"), + (list_empty(&dlm->pending_asts) ? "Empty" : "InUse"), + (list_empty(&dlm->pending_basts) ? "Empty" : "InUse")); + + /* Purge Count: xxx Refs: xxx */ + out += scnprintf(buf + out, len - out, + "Purge Count: %d Refs: %d\n", dlm->purge_count, + kref_read(&dlm->dlm_refs)); + + /* Dead Node: xxx */ + out += scnprintf(buf + out, len - out, + "Dead Node: %d\n", dlm->reco.dead_node); + + /* What about DLM_RECO_STATE_FINALIZE? */ + if (dlm->reco.state == DLM_RECO_STATE_ACTIVE) + state = "ACTIVE"; + else + state = "INACTIVE"; + + /* Recovery Pid: xxxx Master: xxx State: xxxx */ + out += scnprintf(buf + out, len - out, + "Recovery Pid: %d Master: %d State: %s\n", + task_pid_nr(dlm->dlm_reco_thread_task), + dlm->reco.new_master, state); + + /* Recovery Map: xx xx */ + out += scnprintf(buf + out, len - out, "Recovery Map: "); + out += stringify_nodemap(dlm->recovery_map, O2NM_MAX_NODES, + buf + out, len - out); + out += scnprintf(buf + out, len - out, "\n"); + + /* Recovery Node State: */ + out += scnprintf(buf + out, len - out, "Recovery Node State:\n"); + list_for_each_entry(node, &dlm->reco.node_data, list) { + switch (node->state) { + case DLM_RECO_NODE_DATA_INIT: + state = "INIT"; + break; + case DLM_RECO_NODE_DATA_REQUESTING: + state = "REQUESTING"; + break; + case DLM_RECO_NODE_DATA_DEAD: + state = "DEAD"; + break; + case DLM_RECO_NODE_DATA_RECEIVING: + state = "RECEIVING"; + break; + case DLM_RECO_NODE_DATA_REQUESTED: + state = "REQUESTED"; + break; + case DLM_RECO_NODE_DATA_DONE: + state = "DONE"; + break; + case DLM_RECO_NODE_DATA_FINALIZE_SENT: + state = "FINALIZE-SENT"; + break; + default: + state = "BAD"; + break; + } + out += scnprintf(buf + out, len - out, "\t%u - %s\n", + node->node_num, state); + } + + spin_unlock(&dlm->spinlock); + + return out; +} + +static int debug_state_open(struct inode *inode, struct file *file) +{ + struct dlm_ctxt *dlm = inode->i_private; + char *buf = NULL; + + buf = kzalloc(PAGE_SIZE, GFP_NOFS); + if (!buf) + return -ENOMEM; + + i_size_write(inode, debug_state_print(dlm, buf, PAGE_SIZE - 1)); + + file->private_data = buf; + + return 0; +} + +static const struct file_operations debug_state_fops = { + .open = debug_state_open, + .release = debug_release, + .read = debug_read, + .llseek = generic_file_llseek, +}; +/* end - debug state funcs */ + +/* files in subroot */ +void dlm_debug_init(struct dlm_ctxt *dlm) +{ + /* for dumping dlm_ctxt */ + debugfs_create_file(DLM_DEBUGFS_DLM_STATE, S_IFREG|S_IRUSR, + dlm->dlm_debugfs_subroot, dlm, &debug_state_fops); + + /* for dumping lockres */ + debugfs_create_file(DLM_DEBUGFS_LOCKING_STATE, S_IFREG|S_IRUSR, + dlm->dlm_debugfs_subroot, dlm, &debug_lockres_fops); + + /* for dumping mles */ + debugfs_create_file(DLM_DEBUGFS_MLE_STATE, S_IFREG|S_IRUSR, + dlm->dlm_debugfs_subroot, dlm, &debug_mle_fops); + + /* for dumping lockres on the purge list */ + debugfs_create_file(DLM_DEBUGFS_PURGE_LIST, S_IFREG|S_IRUSR, + dlm->dlm_debugfs_subroot, dlm, + &debug_purgelist_fops); +} + +/* subroot - domain dir */ +void dlm_create_debugfs_subroot(struct dlm_ctxt *dlm) +{ + dlm->dlm_debugfs_subroot = debugfs_create_dir(dlm->name, + dlm_debugfs_root); +} + +void dlm_destroy_debugfs_subroot(struct dlm_ctxt *dlm) +{ + debugfs_remove_recursive(dlm->dlm_debugfs_subroot); +} + +/* debugfs root */ +void dlm_create_debugfs_root(void) +{ + dlm_debugfs_root = debugfs_create_dir(DLM_DEBUGFS_DIR, NULL); +} + +void dlm_destroy_debugfs_root(void) +{ + debugfs_remove(dlm_debugfs_root); +} +#endif /* CONFIG_DEBUG_FS */ diff --git a/fs/ocfs2/dlm/dlmdebug.h b/fs/ocfs2/dlm/dlmdebug.h new file mode 100644 index 000000000..e08f7357e --- /dev/null +++ b/fs/ocfs2/dlm/dlmdebug.h @@ -0,0 +1,49 @@ +/* SPDX-License-Identifier: GPL-2.0-or-later */ +/* + * dlmdebug.h + * + * Copyright (C) 2008 Oracle. All rights reserved. + */ + +#ifndef DLMDEBUG_H +#define DLMDEBUG_H + +void dlm_print_one_mle(struct dlm_master_list_entry *mle); + +#ifdef CONFIG_DEBUG_FS + +struct debug_lockres { + int dl_len; + char *dl_buf; + struct dlm_ctxt *dl_ctxt; + struct dlm_lock_resource *dl_res; +}; + +void dlm_debug_init(struct dlm_ctxt *dlm); + +void dlm_create_debugfs_subroot(struct dlm_ctxt *dlm); +void dlm_destroy_debugfs_subroot(struct dlm_ctxt *dlm); + +void dlm_create_debugfs_root(void); +void dlm_destroy_debugfs_root(void); + +#else + +static inline void dlm_debug_init(struct dlm_ctxt *dlm) +{ +} +static inline void dlm_create_debugfs_subroot(struct dlm_ctxt *dlm) +{ +} +static inline void dlm_destroy_debugfs_subroot(struct dlm_ctxt *dlm) +{ +} +static inline void dlm_create_debugfs_root(void) +{ +} +static inline void dlm_destroy_debugfs_root(void) +{ +} + +#endif /* CONFIG_DEBUG_FS */ +#endif /* DLMDEBUG_H */ diff --git a/fs/ocfs2/dlm/dlmdomain.c b/fs/ocfs2/dlm/dlmdomain.c new file mode 100644 index 000000000..97bb9400e --- /dev/null +++ b/fs/ocfs2/dlm/dlmdomain.c @@ -0,0 +1,2371 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * dlmdomain.c + * + * defines domain join / leave apis + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "../cluster/heartbeat.h" +#include "../cluster/nodemanager.h" +#include "../cluster/tcp.h" + +#include "dlmapi.h" +#include "dlmcommon.h" +#include "dlmdomain.h" +#include "dlmdebug.h" + +#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN) +#include "../cluster/masklog.h" + +/* + * ocfs2 node maps are array of long int, which limits to send them freely + * across the wire due to endianness issues. To workaround this, we convert + * long ints to byte arrays. Following 3 routines are helper functions to + * set/test/copy bits within those array of bytes + */ +static inline void byte_set_bit(u8 nr, u8 map[]) +{ + map[nr >> 3] |= (1UL << (nr & 7)); +} + +static inline int byte_test_bit(u8 nr, u8 map[]) +{ + return ((1UL << (nr & 7)) & (map[nr >> 3])) != 0; +} + +static inline void byte_copymap(u8 dmap[], unsigned long smap[], + unsigned int sz) +{ + unsigned int nn; + + if (!sz) + return; + + memset(dmap, 0, ((sz + 7) >> 3)); + for (nn = 0 ; nn < sz; nn++) + if (test_bit(nn, smap)) + byte_set_bit(nn, dmap); +} + +static void dlm_free_pagevec(void **vec, int pages) +{ + while (pages--) + kfree(vec[pages]); + kfree(vec); +} + +static void **dlm_alloc_pagevec(int pages) +{ + void **vec = kmalloc_array(pages, sizeof(void *), GFP_KERNEL); + int i; + + if (!vec) + return NULL; + + for (i = 0; i < pages; i++) { + vec[i] = kmalloc(PAGE_SIZE, GFP_KERNEL); + if (!vec[i]) + goto out_free; + } + + mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n", + pages, (unsigned long)DLM_HASH_PAGES, + (unsigned long)DLM_BUCKETS_PER_PAGE); + return vec; +out_free: + dlm_free_pagevec(vec, i); + return NULL; +} + +/* + * + * spinlock lock ordering: if multiple locks are needed, obey this ordering: + * dlm_domain_lock + * struct dlm_ctxt->spinlock + * struct dlm_lock_resource->spinlock + * struct dlm_ctxt->master_lock + * struct dlm_ctxt->ast_lock + * dlm_master_list_entry->spinlock + * dlm_lock->spinlock + * + */ + +DEFINE_SPINLOCK(dlm_domain_lock); +LIST_HEAD(dlm_domains); +static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events); + +/* + * The supported protocol version for DLM communication. Running domains + * will have a negotiated version with the same major number and a minor + * number equal or smaller. The dlm_ctxt->dlm_locking_proto field should + * be used to determine what a running domain is actually using. + * + * New in version 1.1: + * - Message DLM_QUERY_REGION added to support global heartbeat + * - Message DLM_QUERY_NODEINFO added to allow online node removes + * New in version 1.2: + * - Message DLM_BEGIN_EXIT_DOMAIN_MSG added to mark start of exit domain + * New in version 1.3: + * - Message DLM_DEREF_LOCKRES_DONE added to inform non-master that the + * refmap is cleared + */ +static const struct dlm_protocol_version dlm_protocol = { + .pv_major = 1, + .pv_minor = 3, +}; + +#define DLM_DOMAIN_BACKOFF_MS 200 + +static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +static int dlm_query_region_handler(struct o2net_msg *msg, u32 len, + void *data, void **ret_data); +static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data); +static int dlm_protocol_compare(struct dlm_protocol_version *existing, + struct dlm_protocol_version *request); + +static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm); + +void __dlm_unhash_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) +{ + if (hlist_unhashed(&res->hash_node)) + return; + + mlog(0, "%s: Unhash res %.*s\n", dlm->name, res->lockname.len, + res->lockname.name); + hlist_del_init(&res->hash_node); + dlm_lockres_put(res); +} + +void __dlm_insert_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) +{ + struct hlist_head *bucket; + + assert_spin_locked(&dlm->spinlock); + + bucket = dlm_lockres_hash(dlm, res->lockname.hash); + + /* get a reference for our hashtable */ + dlm_lockres_get(res); + + hlist_add_head(&res->hash_node, bucket); + + mlog(0, "%s: Hash res %.*s\n", dlm->name, res->lockname.len, + res->lockname.name); +} + +struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm, + const char *name, + unsigned int len, + unsigned int hash) +{ + struct hlist_head *bucket; + struct dlm_lock_resource *res; + + mlog(0, "%.*s\n", len, name); + + assert_spin_locked(&dlm->spinlock); + + bucket = dlm_lockres_hash(dlm, hash); + + hlist_for_each_entry(res, bucket, hash_node) { + if (res->lockname.name[0] != name[0]) + continue; + if (unlikely(res->lockname.len != len)) + continue; + if (memcmp(res->lockname.name + 1, name + 1, len - 1)) + continue; + dlm_lockres_get(res); + return res; + } + return NULL; +} + +/* intended to be called by functions which do not care about lock + * resources which are being purged (most net _handler functions). + * this will return NULL for any lock resource which is found but + * currently in the process of dropping its mastery reference. + * use __dlm_lookup_lockres_full when you need the lock resource + * regardless (e.g. dlm_get_lock_resource) */ +struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm, + const char *name, + unsigned int len, + unsigned int hash) +{ + struct dlm_lock_resource *res = NULL; + + mlog(0, "%.*s\n", len, name); + + assert_spin_locked(&dlm->spinlock); + + res = __dlm_lookup_lockres_full(dlm, name, len, hash); + if (res) { + spin_lock(&res->spinlock); + if (res->state & DLM_LOCK_RES_DROPPING_REF) { + spin_unlock(&res->spinlock); + dlm_lockres_put(res); + return NULL; + } + spin_unlock(&res->spinlock); + } + + return res; +} + +struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm, + const char *name, + unsigned int len) +{ + struct dlm_lock_resource *res; + unsigned int hash = dlm_lockid_hash(name, len); + + spin_lock(&dlm->spinlock); + res = __dlm_lookup_lockres(dlm, name, len, hash); + spin_unlock(&dlm->spinlock); + return res; +} + +static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len) +{ + struct dlm_ctxt *tmp; + + assert_spin_locked(&dlm_domain_lock); + + /* tmp->name here is always NULL terminated, + * but domain may not be! */ + list_for_each_entry(tmp, &dlm_domains, list) { + if (strlen(tmp->name) == len && + memcmp(tmp->name, domain, len)==0) + return tmp; + } + + return NULL; +} + +/* For null terminated domain strings ONLY */ +static struct dlm_ctxt * __dlm_lookup_domain(const char *domain) +{ + assert_spin_locked(&dlm_domain_lock); + + return __dlm_lookup_domain_full(domain, strlen(domain)); +} + + +/* returns true on one of two conditions: + * 1) the domain does not exist + * 2) the domain exists and it's state is "joined" */ +static int dlm_wait_on_domain_helper(const char *domain) +{ + int ret = 0; + struct dlm_ctxt *tmp = NULL; + + spin_lock(&dlm_domain_lock); + + tmp = __dlm_lookup_domain(domain); + if (!tmp) + ret = 1; + else if (tmp->dlm_state == DLM_CTXT_JOINED) + ret = 1; + + spin_unlock(&dlm_domain_lock); + return ret; +} + +static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm) +{ + dlm_destroy_debugfs_subroot(dlm); + + if (dlm->lockres_hash) + dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES); + + if (dlm->master_hash) + dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES); + + kfree(dlm->name); + kfree(dlm); +} + +/* A little strange - this function will be called while holding + * dlm_domain_lock and is expected to be holding it on the way out. We + * will however drop and reacquire it multiple times */ +static void dlm_ctxt_release(struct kref *kref) +{ + struct dlm_ctxt *dlm; + + dlm = container_of(kref, struct dlm_ctxt, dlm_refs); + + BUG_ON(dlm->num_joins); + BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED); + + /* we may still be in the list if we hit an error during join. */ + list_del_init(&dlm->list); + + spin_unlock(&dlm_domain_lock); + + mlog(0, "freeing memory from domain %s\n", dlm->name); + + wake_up(&dlm_domain_events); + + dlm_free_ctxt_mem(dlm); + + spin_lock(&dlm_domain_lock); +} + +void dlm_put(struct dlm_ctxt *dlm) +{ + spin_lock(&dlm_domain_lock); + kref_put(&dlm->dlm_refs, dlm_ctxt_release); + spin_unlock(&dlm_domain_lock); +} + +static void __dlm_get(struct dlm_ctxt *dlm) +{ + kref_get(&dlm->dlm_refs); +} + +/* given a questionable reference to a dlm object, gets a reference if + * it can find it in the list, otherwise returns NULL in which case + * you shouldn't trust your pointer. */ +struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm) +{ + struct dlm_ctxt *target; + struct dlm_ctxt *ret = NULL; + + spin_lock(&dlm_domain_lock); + + list_for_each_entry(target, &dlm_domains, list) { + if (target == dlm) { + __dlm_get(target); + ret = target; + break; + } + } + + spin_unlock(&dlm_domain_lock); + + return ret; +} + +int dlm_domain_fully_joined(struct dlm_ctxt *dlm) +{ + int ret; + + spin_lock(&dlm_domain_lock); + ret = (dlm->dlm_state == DLM_CTXT_JOINED) || + (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN); + spin_unlock(&dlm_domain_lock); + + return ret; +} + +static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm) +{ + if (dlm->dlm_worker) { + destroy_workqueue(dlm->dlm_worker); + dlm->dlm_worker = NULL; + } +} + +static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm) +{ + dlm_unregister_domain_handlers(dlm); + dlm_complete_thread(dlm); + dlm_complete_recovery_thread(dlm); + dlm_destroy_dlm_worker(dlm); + + /* We've left the domain. Now we can take ourselves out of the + * list and allow the kref stuff to help us free the + * memory. */ + spin_lock(&dlm_domain_lock); + list_del_init(&dlm->list); + spin_unlock(&dlm_domain_lock); + + /* Wake up anyone waiting for us to remove this domain */ + wake_up(&dlm_domain_events); +} + +static int dlm_migrate_all_locks(struct dlm_ctxt *dlm) +{ + int i, num, n, ret = 0; + struct dlm_lock_resource *res; + struct hlist_node *iter; + struct hlist_head *bucket; + int dropped; + + mlog(0, "Migrating locks from domain %s\n", dlm->name); + + num = 0; + spin_lock(&dlm->spinlock); + for (i = 0; i < DLM_HASH_BUCKETS; i++) { +redo_bucket: + n = 0; + bucket = dlm_lockres_hash(dlm, i); + iter = bucket->first; + while (iter) { + n++; + res = hlist_entry(iter, struct dlm_lock_resource, + hash_node); + dlm_lockres_get(res); + /* migrate, if necessary. this will drop the dlm + * spinlock and retake it if it does migration. */ + dropped = dlm_empty_lockres(dlm, res); + + spin_lock(&res->spinlock); + if (dropped) + __dlm_lockres_calc_usage(dlm, res); + else + iter = res->hash_node.next; + spin_unlock(&res->spinlock); + + dlm_lockres_put(res); + + if (dropped) { + cond_resched_lock(&dlm->spinlock); + goto redo_bucket; + } + } + cond_resched_lock(&dlm->spinlock); + num += n; + } + + if (!num) { + if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) { + mlog(0, "%s: perhaps there are more lock resources " + "need to be migrated after dlm recovery\n", dlm->name); + ret = -EAGAIN; + } else { + mlog(0, "%s: we won't do dlm recovery after migrating " + "all lock resources\n", dlm->name); + dlm->migrate_done = 1; + } + } + + spin_unlock(&dlm->spinlock); + wake_up(&dlm->dlm_thread_wq); + + /* let the dlm thread take care of purging, keep scanning until + * nothing remains in the hash */ + if (num) { + mlog(0, "%s: %d lock resources in hash last pass\n", + dlm->name, num); + ret = -EAGAIN; + } + mlog(0, "DONE Migrating locks from domain %s\n", dlm->name); + return ret; +} + +static int dlm_no_joining_node(struct dlm_ctxt *dlm) +{ + int ret; + + spin_lock(&dlm->spinlock); + ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN; + spin_unlock(&dlm->spinlock); + + return ret; +} + +static int dlm_begin_exit_domain_handler(struct o2net_msg *msg, u32 len, + void *data, void **ret_data) +{ + struct dlm_ctxt *dlm = data; + unsigned int node; + struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf; + + if (!dlm_grab(dlm)) + return 0; + + node = exit_msg->node_idx; + mlog(0, "%s: Node %u sent a begin exit domain message\n", dlm->name, node); + + spin_lock(&dlm->spinlock); + set_bit(node, dlm->exit_domain_map); + spin_unlock(&dlm->spinlock); + + dlm_put(dlm); + + return 0; +} + +static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm) +{ + /* Yikes, a double spinlock! I need domain_lock for the dlm + * state and the dlm spinlock for join state... Sorry! */ +again: + spin_lock(&dlm_domain_lock); + spin_lock(&dlm->spinlock); + + if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) { + mlog(0, "Node %d is joining, we wait on it.\n", + dlm->joining_node); + spin_unlock(&dlm->spinlock); + spin_unlock(&dlm_domain_lock); + + wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm)); + goto again; + } + + dlm->dlm_state = DLM_CTXT_LEAVING; + spin_unlock(&dlm->spinlock); + spin_unlock(&dlm_domain_lock); +} + +static void __dlm_print_nodes(struct dlm_ctxt *dlm) +{ + int node = -1, num = 0; + + assert_spin_locked(&dlm->spinlock); + + printk("( "); + while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, + node + 1)) < O2NM_MAX_NODES) { + printk("%d ", node); + ++num; + } + printk(") %u nodes\n", num); +} + +static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + unsigned int node; + struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf; + + mlog(0, "%p %u %p", msg, len, data); + + if (!dlm_grab(dlm)) + return 0; + + node = exit_msg->node_idx; + + spin_lock(&dlm->spinlock); + clear_bit(node, dlm->domain_map); + clear_bit(node, dlm->exit_domain_map); + printk(KERN_NOTICE "o2dlm: Node %u leaves domain %s ", node, dlm->name); + __dlm_print_nodes(dlm); + + /* notify anything attached to the heartbeat events */ + dlm_hb_event_notify_attached(dlm, node, 0); + + spin_unlock(&dlm->spinlock); + + dlm_put(dlm); + + return 0; +} + +static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, u32 msg_type, + unsigned int node) +{ + int status; + struct dlm_exit_domain leave_msg; + + mlog(0, "%s: Sending domain exit message %u to node %u\n", dlm->name, + msg_type, node); + + memset(&leave_msg, 0, sizeof(leave_msg)); + leave_msg.node_idx = dlm->node_num; + + status = o2net_send_message(msg_type, dlm->key, &leave_msg, + sizeof(leave_msg), node, NULL); + if (status < 0) + mlog(ML_ERROR, "Error %d sending domain exit message %u " + "to node %u on domain %s\n", status, msg_type, node, + dlm->name); + + return status; +} + +static void dlm_begin_exit_domain(struct dlm_ctxt *dlm) +{ + int node = -1; + + /* Support for begin exit domain was added in 1.2 */ + if (dlm->dlm_locking_proto.pv_major == 1 && + dlm->dlm_locking_proto.pv_minor < 2) + return; + + /* + * Unlike DLM_EXIT_DOMAIN_MSG, DLM_BEGIN_EXIT_DOMAIN_MSG is purely + * informational. Meaning if a node does not receive the message, + * so be it. + */ + spin_lock(&dlm->spinlock); + while (1) { + node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, node + 1); + if (node >= O2NM_MAX_NODES) + break; + if (node == dlm->node_num) + continue; + + spin_unlock(&dlm->spinlock); + dlm_send_one_domain_exit(dlm, DLM_BEGIN_EXIT_DOMAIN_MSG, node); + spin_lock(&dlm->spinlock); + } + spin_unlock(&dlm->spinlock); +} + +static void dlm_leave_domain(struct dlm_ctxt *dlm) +{ + int node, clear_node, status; + + /* At this point we've migrated away all our locks and won't + * accept mastership of new ones. The dlm is responsible for + * almost nothing now. We make sure not to confuse any joining + * nodes and then commence shutdown procedure. */ + + spin_lock(&dlm->spinlock); + /* Clear ourselves from the domain map */ + clear_bit(dlm->node_num, dlm->domain_map); + while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, + 0)) < O2NM_MAX_NODES) { + /* Drop the dlm spinlock. This is safe wrt the domain_map. + * -nodes cannot be added now as the + * query_join_handlers knows to respond with OK_NO_MAP + * -we catch the right network errors if a node is + * removed from the map while we're sending him the + * exit message. */ + spin_unlock(&dlm->spinlock); + + clear_node = 1; + + status = dlm_send_one_domain_exit(dlm, DLM_EXIT_DOMAIN_MSG, + node); + if (status < 0 && + status != -ENOPROTOOPT && + status != -ENOTCONN) { + mlog(ML_NOTICE, "Error %d sending domain exit message " + "to node %d\n", status, node); + + /* Not sure what to do here but lets sleep for + * a bit in case this was a transient + * error... */ + msleep(DLM_DOMAIN_BACKOFF_MS); + clear_node = 0; + } + + spin_lock(&dlm->spinlock); + /* If we're not clearing the node bit then we intend + * to loop back around to try again. */ + if (clear_node) + clear_bit(node, dlm->domain_map); + } + spin_unlock(&dlm->spinlock); +} + +void dlm_unregister_domain(struct dlm_ctxt *dlm) +{ + int leave = 0; + struct dlm_lock_resource *res; + + spin_lock(&dlm_domain_lock); + BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED); + BUG_ON(!dlm->num_joins); + + dlm->num_joins--; + if (!dlm->num_joins) { + /* We mark it "in shutdown" now so new register + * requests wait until we've completely left the + * domain. Don't use DLM_CTXT_LEAVING yet as we still + * want new domain joins to communicate with us at + * least until we've completed migration of our + * resources. */ + dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN; + leave = 1; + } + spin_unlock(&dlm_domain_lock); + + if (leave) { + mlog(0, "shutting down domain %s\n", dlm->name); + dlm_begin_exit_domain(dlm); + + /* We changed dlm state, notify the thread */ + dlm_kick_thread(dlm, NULL); + + while (dlm_migrate_all_locks(dlm)) { + /* Give dlm_thread time to purge the lockres' */ + msleep(500); + mlog(0, "%s: more migration to do\n", dlm->name); + } + + /* This list should be empty. If not, print remaining lockres */ + if (!list_empty(&dlm->tracking_list)) { + mlog(ML_ERROR, "Following lockres' are still on the " + "tracking list:\n"); + list_for_each_entry(res, &dlm->tracking_list, tracking) + dlm_print_one_lock_resource(res); + } + + dlm_mark_domain_leaving(dlm); + dlm_leave_domain(dlm); + printk(KERN_NOTICE "o2dlm: Leaving domain %s\n", dlm->name); + dlm_force_free_mles(dlm); + dlm_complete_dlm_shutdown(dlm); + } + dlm_put(dlm); +} +EXPORT_SYMBOL_GPL(dlm_unregister_domain); + +static int dlm_query_join_proto_check(char *proto_type, int node, + struct dlm_protocol_version *ours, + struct dlm_protocol_version *request) +{ + int rc; + struct dlm_protocol_version proto = *request; + + if (!dlm_protocol_compare(ours, &proto)) { + mlog(0, + "node %u wanted to join with %s locking protocol " + "%u.%u, we respond with %u.%u\n", + node, proto_type, + request->pv_major, + request->pv_minor, + proto.pv_major, proto.pv_minor); + request->pv_minor = proto.pv_minor; + rc = 0; + } else { + mlog(ML_NOTICE, + "Node %u wanted to join with %s locking " + "protocol %u.%u, but we have %u.%u, disallowing\n", + node, proto_type, + request->pv_major, + request->pv_minor, + ours->pv_major, + ours->pv_minor); + rc = 1; + } + + return rc; +} + +/* + * struct dlm_query_join_packet is made up of four one-byte fields. They + * are effectively in big-endian order already. However, little-endian + * machines swap them before putting the packet on the wire (because + * query_join's response is a status, and that status is treated as a u32 + * on the wire). Thus, a big-endian and little-endian machines will treat + * this structure differently. + * + * The solution is to have little-endian machines swap the structure when + * converting from the structure to the u32 representation. This will + * result in the structure having the correct format on the wire no matter + * the host endian format. + */ +static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet, + u32 *wire) +{ + union dlm_query_join_response response; + + response.packet = *packet; + *wire = be32_to_cpu(response.intval); +} + +static void dlm_query_join_wire_to_packet(u32 wire, + struct dlm_query_join_packet *packet) +{ + union dlm_query_join_response response; + + response.intval = cpu_to_be32(wire); + *packet = response.packet; +} + +static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_query_join_request *query; + struct dlm_query_join_packet packet = { + .code = JOIN_DISALLOW, + }; + struct dlm_ctxt *dlm = NULL; + u32 response; + u8 nodenum; + + query = (struct dlm_query_join_request *) msg->buf; + + mlog(0, "node %u wants to join domain %s\n", query->node_idx, + query->domain); + + /* + * If heartbeat doesn't consider the node live, tell it + * to back off and try again. This gives heartbeat a chance + * to catch up. + */ + if (!o2hb_check_node_heartbeating_no_sem(query->node_idx)) { + mlog(0, "node %u is not in our live map yet\n", + query->node_idx); + + packet.code = JOIN_DISALLOW; + goto respond; + } + + packet.code = JOIN_OK_NO_MAP; + + spin_lock(&dlm_domain_lock); + dlm = __dlm_lookup_domain_full(query->domain, query->name_len); + if (!dlm) + goto unlock_respond; + + /* + * There is a small window where the joining node may not see the + * node(s) that just left but still part of the cluster. DISALLOW + * join request if joining node has different node map. + */ + nodenum=0; + while (nodenum < O2NM_MAX_NODES) { + if (test_bit(nodenum, dlm->domain_map)) { + if (!byte_test_bit(nodenum, query->node_map)) { + mlog(0, "disallow join as node %u does not " + "have node %u in its nodemap\n", + query->node_idx, nodenum); + packet.code = JOIN_DISALLOW; + goto unlock_respond; + } + } + nodenum++; + } + + /* Once the dlm ctxt is marked as leaving then we don't want + * to be put in someone's domain map. + * Also, explicitly disallow joining at certain troublesome + * times (ie. during recovery). */ + if (dlm->dlm_state != DLM_CTXT_LEAVING) { + int bit = query->node_idx; + spin_lock(&dlm->spinlock); + + if (dlm->dlm_state == DLM_CTXT_NEW && + dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) { + /*If this is a brand new context and we + * haven't started our join process yet, then + * the other node won the race. */ + packet.code = JOIN_OK_NO_MAP; + } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) { + /* Disallow parallel joins. */ + packet.code = JOIN_DISALLOW; + } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) { + mlog(0, "node %u trying to join, but recovery " + "is ongoing.\n", bit); + packet.code = JOIN_DISALLOW; + } else if (test_bit(bit, dlm->recovery_map)) { + mlog(0, "node %u trying to join, but it " + "still needs recovery.\n", bit); + packet.code = JOIN_DISALLOW; + } else if (test_bit(bit, dlm->domain_map)) { + mlog(0, "node %u trying to join, but it " + "is still in the domain! needs recovery?\n", + bit); + packet.code = JOIN_DISALLOW; + } else { + /* Alright we're fully a part of this domain + * so we keep some state as to who's joining + * and indicate to him that needs to be fixed + * up. */ + + /* Make sure we speak compatible locking protocols. */ + if (dlm_query_join_proto_check("DLM", bit, + &dlm->dlm_locking_proto, + &query->dlm_proto)) { + packet.code = JOIN_PROTOCOL_MISMATCH; + } else if (dlm_query_join_proto_check("fs", bit, + &dlm->fs_locking_proto, + &query->fs_proto)) { + packet.code = JOIN_PROTOCOL_MISMATCH; + } else { + packet.dlm_minor = query->dlm_proto.pv_minor; + packet.fs_minor = query->fs_proto.pv_minor; + packet.code = JOIN_OK; + __dlm_set_joining_node(dlm, query->node_idx); + } + } + + spin_unlock(&dlm->spinlock); + } +unlock_respond: + spin_unlock(&dlm_domain_lock); + +respond: + mlog(0, "We respond with %u\n", packet.code); + + dlm_query_join_packet_to_wire(&packet, &response); + return response; +} + +static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_assert_joined *assert; + struct dlm_ctxt *dlm = NULL; + + assert = (struct dlm_assert_joined *) msg->buf; + + mlog(0, "node %u asserts join on domain %s\n", assert->node_idx, + assert->domain); + + spin_lock(&dlm_domain_lock); + dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len); + /* XXX should we consider no dlm ctxt an error? */ + if (dlm) { + spin_lock(&dlm->spinlock); + + /* Alright, this node has officially joined our + * domain. Set him in the map and clean up our + * leftover join state. */ + BUG_ON(dlm->joining_node != assert->node_idx); + + if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) { + mlog(0, "dlm recovery is ongoing, disallow join\n"); + spin_unlock(&dlm->spinlock); + spin_unlock(&dlm_domain_lock); + return -EAGAIN; + } + + set_bit(assert->node_idx, dlm->domain_map); + clear_bit(assert->node_idx, dlm->exit_domain_map); + __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); + + printk(KERN_NOTICE "o2dlm: Node %u joins domain %s ", + assert->node_idx, dlm->name); + __dlm_print_nodes(dlm); + + /* notify anything attached to the heartbeat events */ + dlm_hb_event_notify_attached(dlm, assert->node_idx, 1); + + spin_unlock(&dlm->spinlock); + } + spin_unlock(&dlm_domain_lock); + + return 0; +} + +static int dlm_match_regions(struct dlm_ctxt *dlm, + struct dlm_query_region *qr, + char *local, int locallen) +{ + char *remote = qr->qr_regions; + char *l, *r; + int localnr, i, j, foundit; + int status = 0; + + if (!o2hb_global_heartbeat_active()) { + if (qr->qr_numregions) { + mlog(ML_ERROR, "Domain %s: Joining node %d has global " + "heartbeat enabled but local node %d does not\n", + qr->qr_domain, qr->qr_node, dlm->node_num); + status = -EINVAL; + } + goto bail; + } + + if (o2hb_global_heartbeat_active() && !qr->qr_numregions) { + mlog(ML_ERROR, "Domain %s: Local node %d has global " + "heartbeat enabled but joining node %d does not\n", + qr->qr_domain, dlm->node_num, qr->qr_node); + status = -EINVAL; + goto bail; + } + + if (qr->qr_numregions > O2NM_MAX_REGIONS) { + mlog(ML_ERROR, "Domain %s: Joining node %d has invalid " + "number of heartbeat regions %u\n", + qr->qr_domain, qr->qr_node, qr->qr_numregions); + status = -EINVAL; + goto bail; + } + + r = remote; + for (i = 0; i < qr->qr_numregions; ++i) { + mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, r); + r += O2HB_MAX_REGION_NAME_LEN; + } + + localnr = min(O2NM_MAX_REGIONS, locallen/O2HB_MAX_REGION_NAME_LEN); + localnr = o2hb_get_all_regions(local, (u8)localnr); + + /* compare local regions with remote */ + l = local; + for (i = 0; i < localnr; ++i) { + foundit = 0; + r = remote; + for (j = 0; j < qr->qr_numregions; ++j) { + if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) { + foundit = 1; + break; + } + r += O2HB_MAX_REGION_NAME_LEN; + } + if (!foundit) { + status = -EINVAL; + mlog(ML_ERROR, "Domain %s: Region '%.*s' registered " + "in local node %d but not in joining node %d\n", + qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, l, + dlm->node_num, qr->qr_node); + goto bail; + } + l += O2HB_MAX_REGION_NAME_LEN; + } + + /* compare remote with local regions */ + r = remote; + for (i = 0; i < qr->qr_numregions; ++i) { + foundit = 0; + l = local; + for (j = 0; j < localnr; ++j) { + if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) { + foundit = 1; + break; + } + l += O2HB_MAX_REGION_NAME_LEN; + } + if (!foundit) { + status = -EINVAL; + mlog(ML_ERROR, "Domain %s: Region '%.*s' registered " + "in joining node %d but not in local node %d\n", + qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, r, + qr->qr_node, dlm->node_num); + goto bail; + } + r += O2HB_MAX_REGION_NAME_LEN; + } + +bail: + return status; +} + +static int dlm_send_regions(struct dlm_ctxt *dlm, unsigned long *node_map) +{ + struct dlm_query_region *qr = NULL; + int status, ret = 0, i; + char *p; + + if (find_first_bit(node_map, O2NM_MAX_NODES) >= O2NM_MAX_NODES) + goto bail; + + qr = kzalloc_obj(struct dlm_query_region); + if (!qr) { + ret = -ENOMEM; + mlog_errno(ret); + goto bail; + } + + qr->qr_node = dlm->node_num; + qr->qr_namelen = strlen(dlm->name); + memcpy(qr->qr_domain, dlm->name, qr->qr_namelen); + /* if local hb, the numregions will be zero */ + if (o2hb_global_heartbeat_active()) + qr->qr_numregions = o2hb_get_all_regions(qr->qr_regions, + O2NM_MAX_REGIONS); + + p = qr->qr_regions; + for (i = 0; i < qr->qr_numregions; ++i, p += O2HB_MAX_REGION_NAME_LEN) + mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, p); + + i = -1; + while ((i = find_next_bit(node_map, O2NM_MAX_NODES, + i + 1)) < O2NM_MAX_NODES) { + if (i == dlm->node_num) + continue; + + mlog(0, "Sending regions to node %d\n", i); + + ret = o2net_send_message(DLM_QUERY_REGION, DLM_MOD_KEY, qr, + sizeof(struct dlm_query_region), + i, &status); + if (ret >= 0) + ret = status; + if (ret) { + mlog(ML_ERROR, "Region mismatch %d, node %d\n", + ret, i); + break; + } + } + +bail: + kfree(qr); + return ret; +} + +static int dlm_query_region_handler(struct o2net_msg *msg, u32 len, + void *data, void **ret_data) +{ + struct dlm_query_region *qr; + struct dlm_ctxt *dlm = NULL; + char *local = NULL; + int status = 0; + + qr = (struct dlm_query_region *) msg->buf; + + mlog(0, "Node %u queries hb regions on domain %s\n", qr->qr_node, + qr->qr_domain); + + /* buffer used in dlm_match_regions() */ + local = kmalloc(sizeof(qr->qr_regions), GFP_KERNEL); + if (!local) + return -ENOMEM; + + status = -EINVAL; + + spin_lock(&dlm_domain_lock); + dlm = __dlm_lookup_domain_full(qr->qr_domain, qr->qr_namelen); + if (!dlm) { + mlog(ML_ERROR, "Node %d queried hb regions on domain %s " + "before join domain\n", qr->qr_node, qr->qr_domain); + goto out_domain_lock; + } + + spin_lock(&dlm->spinlock); + if (dlm->joining_node != qr->qr_node) { + mlog(ML_ERROR, "Node %d queried hb regions on domain %s " + "but joining node is %d\n", qr->qr_node, qr->qr_domain, + dlm->joining_node); + goto out_dlm_lock; + } + + /* Support for global heartbeat was added in 1.1 */ + if (dlm->dlm_locking_proto.pv_major == 1 && + dlm->dlm_locking_proto.pv_minor == 0) { + mlog(ML_ERROR, "Node %d queried hb regions on domain %s " + "but active dlm protocol is %d.%d\n", qr->qr_node, + qr->qr_domain, dlm->dlm_locking_proto.pv_major, + dlm->dlm_locking_proto.pv_minor); + goto out_dlm_lock; + } + + status = dlm_match_regions(dlm, qr, local, sizeof(qr->qr_regions)); + +out_dlm_lock: + spin_unlock(&dlm->spinlock); + +out_domain_lock: + spin_unlock(&dlm_domain_lock); + + kfree(local); + + return status; +} + +static int dlm_match_nodes(struct dlm_ctxt *dlm, struct dlm_query_nodeinfo *qn) +{ + struct o2nm_node *local; + struct dlm_node_info *remote; + int i, j; + int status = 0; + + for (j = 0; j < qn->qn_numnodes; ++j) + mlog(0, "Node %3d, %pI4:%u\n", qn->qn_nodes[j].ni_nodenum, + &(qn->qn_nodes[j].ni_ipv4_address), + ntohs(qn->qn_nodes[j].ni_ipv4_port)); + + for (i = 0; i < O2NM_MAX_NODES && !status; ++i) { + local = o2nm_get_node_by_num(i); + remote = NULL; + for (j = 0; j < qn->qn_numnodes; ++j) { + if (qn->qn_nodes[j].ni_nodenum == i) { + remote = &(qn->qn_nodes[j]); + break; + } + } + + if (!local && !remote) + continue; + + if ((local && !remote) || (!local && remote)) + status = -EINVAL; + + if (!status && + ((remote->ni_nodenum != local->nd_num) || + (remote->ni_ipv4_port != local->nd_ipv4_port) || + (remote->ni_ipv4_address != local->nd_ipv4_address))) + status = -EINVAL; + + if (status) { + if (remote && !local) + mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) " + "registered in joining node %d but not in " + "local node %d\n", qn->qn_domain, + remote->ni_nodenum, + &(remote->ni_ipv4_address), + ntohs(remote->ni_ipv4_port), + qn->qn_nodenum, dlm->node_num); + if (local && !remote) + mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) " + "registered in local node %d but not in " + "joining node %d\n", qn->qn_domain, + local->nd_num, &(local->nd_ipv4_address), + ntohs(local->nd_ipv4_port), + dlm->node_num, qn->qn_nodenum); + BUG_ON((!local && !remote)); + } + + if (local) + o2nm_node_put(local); + } + + return status; +} + +static int dlm_send_nodeinfo(struct dlm_ctxt *dlm, unsigned long *node_map) +{ + struct dlm_query_nodeinfo *qn = NULL; + struct o2nm_node *node; + int ret = 0, status, count, i; + + if (find_first_bit(node_map, O2NM_MAX_NODES) >= O2NM_MAX_NODES) + goto bail; + + qn = kzalloc_obj(struct dlm_query_nodeinfo); + if (!qn) { + ret = -ENOMEM; + mlog_errno(ret); + goto bail; + } + + for (i = 0, count = 0; i < O2NM_MAX_NODES; ++i) { + node = o2nm_get_node_by_num(i); + if (!node) + continue; + qn->qn_nodes[count].ni_nodenum = node->nd_num; + qn->qn_nodes[count].ni_ipv4_port = node->nd_ipv4_port; + qn->qn_nodes[count].ni_ipv4_address = node->nd_ipv4_address; + mlog(0, "Node %3d, %pI4:%u\n", node->nd_num, + &(node->nd_ipv4_address), ntohs(node->nd_ipv4_port)); + ++count; + o2nm_node_put(node); + } + + qn->qn_nodenum = dlm->node_num; + qn->qn_numnodes = count; + qn->qn_namelen = strlen(dlm->name); + memcpy(qn->qn_domain, dlm->name, qn->qn_namelen); + + i = -1; + while ((i = find_next_bit(node_map, O2NM_MAX_NODES, + i + 1)) < O2NM_MAX_NODES) { + if (i == dlm->node_num) + continue; + + mlog(0, "Sending nodeinfo to node %d\n", i); + + ret = o2net_send_message(DLM_QUERY_NODEINFO, DLM_MOD_KEY, + qn, sizeof(struct dlm_query_nodeinfo), + i, &status); + if (ret >= 0) + ret = status; + if (ret) { + mlog(ML_ERROR, "node mismatch %d, node %d\n", ret, i); + break; + } + } + +bail: + kfree(qn); + return ret; +} + +static int dlm_query_nodeinfo_handler(struct o2net_msg *msg, u32 len, + void *data, void **ret_data) +{ + struct dlm_query_nodeinfo *qn; + struct dlm_ctxt *dlm = NULL; + int status = -EINVAL; + + qn = (struct dlm_query_nodeinfo *) msg->buf; + + mlog(0, "Node %u queries nodes on domain %s\n", qn->qn_nodenum, + qn->qn_domain); + + spin_lock(&dlm_domain_lock); + dlm = __dlm_lookup_domain_full(qn->qn_domain, qn->qn_namelen); + if (!dlm) { + mlog(ML_ERROR, "Node %d queried nodes on domain %s before " + "join domain\n", qn->qn_nodenum, qn->qn_domain); + goto bail; + } + + spin_lock(&dlm->spinlock); + if (dlm->joining_node != qn->qn_nodenum) { + mlog(ML_ERROR, "Node %d queried nodes on domain %s but " + "joining node is %d\n", qn->qn_nodenum, qn->qn_domain, + dlm->joining_node); + goto unlock; + } + + /* Support for node query was added in 1.1 */ + if (dlm->dlm_locking_proto.pv_major == 1 && + dlm->dlm_locking_proto.pv_minor == 0) { + mlog(ML_ERROR, "Node %d queried nodes on domain %s " + "but active dlm protocol is %d.%d\n", qn->qn_nodenum, + qn->qn_domain, dlm->dlm_locking_proto.pv_major, + dlm->dlm_locking_proto.pv_minor); + goto unlock; + } + + status = dlm_match_nodes(dlm, qn); + +unlock: + spin_unlock(&dlm->spinlock); +bail: + spin_unlock(&dlm_domain_lock); + + return status; +} + +static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_cancel_join *cancel; + struct dlm_ctxt *dlm = NULL; + + cancel = (struct dlm_cancel_join *) msg->buf; + + mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx, + cancel->domain); + + spin_lock(&dlm_domain_lock); + dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len); + + if (dlm) { + spin_lock(&dlm->spinlock); + + /* Yikes, this guy wants to cancel his join. No + * problem, we simply cleanup our join state. */ + BUG_ON(dlm->joining_node != cancel->node_idx); + __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); + + spin_unlock(&dlm->spinlock); + } + spin_unlock(&dlm_domain_lock); + + return 0; +} + +static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm, + unsigned int node) +{ + int status; + struct dlm_cancel_join cancel_msg; + + memset(&cancel_msg, 0, sizeof(cancel_msg)); + cancel_msg.node_idx = dlm->node_num; + cancel_msg.name_len = strlen(dlm->name); + memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len); + + status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY, + &cancel_msg, sizeof(cancel_msg), node, + NULL); + if (status < 0) { + mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to " + "node %u\n", status, DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY, + node); + goto bail; + } + +bail: + return status; +} + +/* map_size should be in bytes. */ +static int dlm_send_join_cancels(struct dlm_ctxt *dlm, + unsigned long *node_map, + unsigned int map_size) +{ + int status, tmpstat; + int node; + + if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) * + sizeof(unsigned long))) { + mlog(ML_ERROR, + "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n", + map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES)); + return -EINVAL; + } + + status = 0; + node = -1; + while ((node = find_next_bit(node_map, O2NM_MAX_NODES, + node + 1)) < O2NM_MAX_NODES) { + if (node == dlm->node_num) + continue; + + tmpstat = dlm_send_one_join_cancel(dlm, node); + if (tmpstat) { + mlog(ML_ERROR, "Error return %d cancelling join on " + "node %d\n", tmpstat, node); + if (!status) + status = tmpstat; + } + } + + if (status) + mlog_errno(status); + return status; +} + +static int dlm_request_join(struct dlm_ctxt *dlm, + int node, + enum dlm_query_join_response_code *response) +{ + int status; + struct dlm_query_join_request join_msg; + struct dlm_query_join_packet packet; + u32 join_resp; + + mlog(0, "querying node %d\n", node); + + memset(&join_msg, 0, sizeof(join_msg)); + join_msg.node_idx = dlm->node_num; + join_msg.name_len = strlen(dlm->name); + memcpy(join_msg.domain, dlm->name, join_msg.name_len); + join_msg.dlm_proto = dlm->dlm_locking_proto; + join_msg.fs_proto = dlm->fs_locking_proto; + + /* copy live node map to join message */ + byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES); + + status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg, + sizeof(join_msg), node, &join_resp); + if (status < 0 && status != -ENOPROTOOPT) { + mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to " + "node %u\n", status, DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, + node); + goto bail; + } + dlm_query_join_wire_to_packet(join_resp, &packet); + + /* -ENOPROTOOPT from the net code means the other side isn't + listening for our message type -- that's fine, it means + his dlm isn't up, so we can consider him a 'yes' but not + joined into the domain. */ + if (status == -ENOPROTOOPT) { + status = 0; + *response = JOIN_OK_NO_MAP; + } else { + *response = packet.code; + switch (packet.code) { + case JOIN_DISALLOW: + case JOIN_OK_NO_MAP: + break; + case JOIN_PROTOCOL_MISMATCH: + mlog(ML_NOTICE, + "This node requested DLM locking protocol %u.%u and " + "filesystem locking protocol %u.%u. At least one of " + "the protocol versions on node %d is not compatible, " + "disconnecting\n", + dlm->dlm_locking_proto.pv_major, + dlm->dlm_locking_proto.pv_minor, + dlm->fs_locking_proto.pv_major, + dlm->fs_locking_proto.pv_minor, + node); + status = -EPROTO; + break; + case JOIN_OK: + /* Use the same locking protocol as the remote node */ + dlm->dlm_locking_proto.pv_minor = packet.dlm_minor; + dlm->fs_locking_proto.pv_minor = packet.fs_minor; + mlog(0, + "Node %d responds JOIN_OK with DLM locking protocol " + "%u.%u and fs locking protocol %u.%u\n", + node, + dlm->dlm_locking_proto.pv_major, + dlm->dlm_locking_proto.pv_minor, + dlm->fs_locking_proto.pv_major, + dlm->fs_locking_proto.pv_minor); + break; + default: + status = -EINVAL; + mlog(ML_ERROR, "invalid response %d from node %u\n", + packet.code, node); + /* Reset response to JOIN_DISALLOW */ + *response = JOIN_DISALLOW; + break; + } + } + + mlog(0, "status %d, node %d response is %d\n", status, node, + *response); + +bail: + return status; +} + +static int dlm_send_one_join_assert(struct dlm_ctxt *dlm, + unsigned int node) +{ + int status; + int ret; + struct dlm_assert_joined assert_msg; + + mlog(0, "Sending join assert to node %u\n", node); + + memset(&assert_msg, 0, sizeof(assert_msg)); + assert_msg.node_idx = dlm->node_num; + assert_msg.name_len = strlen(dlm->name); + memcpy(assert_msg.domain, dlm->name, assert_msg.name_len); + + status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY, + &assert_msg, sizeof(assert_msg), node, + &ret); + if (status < 0) + mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to " + "node %u\n", status, DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY, + node); + else + status = ret; + + return status; +} + +static void dlm_send_join_asserts(struct dlm_ctxt *dlm, + unsigned long *node_map) +{ + int status, node, live; + + node = -1; + while ((node = find_next_bit(node_map, O2NM_MAX_NODES, + node + 1)) < O2NM_MAX_NODES) { + if (node == dlm->node_num) + continue; + + do { + /* It is very important that this message be + * received so we spin until either the node + * has died or it gets the message. */ + status = dlm_send_one_join_assert(dlm, node); + + spin_lock(&dlm->spinlock); + live = test_bit(node, dlm->live_nodes_map); + spin_unlock(&dlm->spinlock); + + if (status) { + mlog(ML_ERROR, "Error return %d asserting " + "join on node %d\n", status, node); + + /* give us some time between errors... */ + if (live) + msleep(DLM_DOMAIN_BACKOFF_MS); + } + } while (status && live); + } +} + +struct domain_join_ctxt { + unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; + unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; +}; + +static int dlm_should_restart_join(struct dlm_ctxt *dlm, + struct domain_join_ctxt *ctxt, + enum dlm_query_join_response_code response) +{ + int ret; + + if (response == JOIN_DISALLOW) { + mlog(0, "Latest response of disallow -- should restart\n"); + return 1; + } + + spin_lock(&dlm->spinlock); + /* For now, we restart the process if the node maps have + * changed at all */ + ret = !bitmap_equal(ctxt->live_map, dlm->live_nodes_map, + O2NM_MAX_NODES); + spin_unlock(&dlm->spinlock); + + if (ret) + mlog(0, "Node maps changed -- should restart\n"); + + return ret; +} + +static int dlm_try_to_join_domain(struct dlm_ctxt *dlm) +{ + int status = 0, tmpstat, node; + struct domain_join_ctxt *ctxt; + enum dlm_query_join_response_code response = JOIN_DISALLOW; + + mlog(0, "%p", dlm); + + ctxt = kzalloc_obj(*ctxt); + if (!ctxt) { + status = -ENOMEM; + mlog_errno(status); + goto bail; + } + + /* group sem locking should work for us here -- we're already + * registered for heartbeat events so filling this should be + * atomic wrt getting those handlers called. */ + o2hb_fill_node_map(dlm->live_nodes_map, O2NM_MAX_NODES); + + spin_lock(&dlm->spinlock); + bitmap_copy(ctxt->live_map, dlm->live_nodes_map, O2NM_MAX_NODES); + __dlm_set_joining_node(dlm, dlm->node_num); + spin_unlock(&dlm->spinlock); + + node = -1; + while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES, + node + 1)) < O2NM_MAX_NODES) { + if (node == dlm->node_num) + continue; + + status = dlm_request_join(dlm, node, &response); + if (status < 0) { + mlog_errno(status); + goto bail; + } + + /* Ok, either we got a response or the node doesn't have a + * dlm up. */ + if (response == JOIN_OK) + set_bit(node, ctxt->yes_resp_map); + + if (dlm_should_restart_join(dlm, ctxt, response)) { + status = -EAGAIN; + goto bail; + } + } + + mlog(0, "Yay, done querying nodes!\n"); + + /* Yay, everyone agree's we can join the domain. My domain is + * comprised of all nodes who were put in the + * yes_resp_map. Copy that into our domain map and send a join + * assert message to clean up everyone elses state. */ + spin_lock(&dlm->spinlock); + bitmap_copy(dlm->domain_map, ctxt->yes_resp_map, O2NM_MAX_NODES); + set_bit(dlm->node_num, dlm->domain_map); + spin_unlock(&dlm->spinlock); + + /* Support for global heartbeat and node info was added in 1.1 */ + if (dlm->dlm_locking_proto.pv_major > 1 || + dlm->dlm_locking_proto.pv_minor > 0) { + status = dlm_send_nodeinfo(dlm, ctxt->yes_resp_map); + if (status) { + mlog_errno(status); + goto bail; + } + status = dlm_send_regions(dlm, ctxt->yes_resp_map); + if (status) { + mlog_errno(status); + goto bail; + } + } + + dlm_send_join_asserts(dlm, ctxt->yes_resp_map); + + /* Joined state *must* be set before the joining node + * information, otherwise the query_join handler may read no + * current joiner but a state of NEW and tell joining nodes + * we're not in the domain. */ + spin_lock(&dlm_domain_lock); + dlm->dlm_state = DLM_CTXT_JOINED; + dlm->num_joins++; + spin_unlock(&dlm_domain_lock); + +bail: + spin_lock(&dlm->spinlock); + __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); + if (!status) { + printk(KERN_NOTICE "o2dlm: Joining domain %s ", dlm->name); + __dlm_print_nodes(dlm); + } + spin_unlock(&dlm->spinlock); + + if (ctxt) { + /* Do we need to send a cancel message to any nodes? */ + if (status < 0) { + tmpstat = dlm_send_join_cancels(dlm, + ctxt->yes_resp_map, + sizeof(ctxt->yes_resp_map)); + if (tmpstat < 0) + mlog_errno(tmpstat); + } + kfree(ctxt); + } + + mlog(0, "returning %d\n", status); + return status; +} + +static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm) +{ + o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_up); + o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_down); + o2net_unregister_handler_list(&dlm->dlm_domain_handlers); +} + +static int dlm_register_domain_handlers(struct dlm_ctxt *dlm) +{ + int status; + + mlog(0, "registering handlers.\n"); + + o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB, + dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI); + o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB, + dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI); + + status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_down); + if (status) + goto bail; + + status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_up); + if (status) + goto bail; + + status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key, + sizeof(struct dlm_master_request), + dlm_master_request_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key, + sizeof(struct dlm_assert_master), + dlm_assert_master_handler, + dlm, dlm_assert_master_post_handler, + &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key, + sizeof(struct dlm_create_lock), + dlm_create_lock_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key, + DLM_CONVERT_LOCK_MAX_LEN, + dlm_convert_lock_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key, + DLM_UNLOCK_LOCK_MAX_LEN, + dlm_unlock_lock_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key, + DLM_PROXY_AST_MAX_LEN, + dlm_proxy_ast_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key, + sizeof(struct dlm_exit_domain), + dlm_exit_domain_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key, + sizeof(struct dlm_deref_lockres), + dlm_deref_lockres_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key, + sizeof(struct dlm_migrate_request), + dlm_migrate_request_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key, + DLM_MIG_LOCKRES_MAX_LEN, + dlm_mig_lockres_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key, + sizeof(struct dlm_master_requery), + dlm_master_requery_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key, + sizeof(struct dlm_lock_request), + dlm_request_all_locks_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key, + sizeof(struct dlm_reco_data_done), + dlm_reco_data_done_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key, + sizeof(struct dlm_begin_reco), + dlm_begin_reco_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key, + sizeof(struct dlm_finalize_reco), + dlm_finalize_reco_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_BEGIN_EXIT_DOMAIN_MSG, dlm->key, + sizeof(struct dlm_exit_domain), + dlm_begin_exit_domain_handler, + dlm, NULL, &dlm->dlm_domain_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_DEREF_LOCKRES_DONE, dlm->key, + sizeof(struct dlm_deref_lockres_done), + dlm_deref_lockres_done_handler, + dlm, NULL, &dlm->dlm_domain_handlers); +bail: + if (status) + dlm_unregister_domain_handlers(dlm); + + return status; +} + +static int dlm_join_domain(struct dlm_ctxt *dlm) +{ + int status; + unsigned int backoff; + unsigned int total_backoff = 0; + char wq_name[O2NM_MAX_NAME_LEN]; + + BUG_ON(!dlm); + + mlog(0, "Join domain %s\n", dlm->name); + + status = dlm_register_domain_handlers(dlm); + if (status) { + mlog_errno(status); + goto bail; + } + + status = dlm_launch_thread(dlm); + if (status < 0) { + mlog_errno(status); + goto bail; + } + + status = dlm_launch_recovery_thread(dlm); + if (status < 0) { + mlog_errno(status); + goto bail; + } + + dlm_debug_init(dlm); + + snprintf(wq_name, O2NM_MAX_NAME_LEN, "dlm_wq-%s", dlm->name); + dlm->dlm_worker = alloc_workqueue(wq_name, WQ_MEM_RECLAIM | WQ_PERCPU, + 0); + if (!dlm->dlm_worker) { + status = -ENOMEM; + mlog_errno(status); + goto bail; + } + + do { + status = dlm_try_to_join_domain(dlm); + + /* If we're racing another node to the join, then we + * need to back off temporarily and let them + * complete. */ +#define DLM_JOIN_TIMEOUT_MSECS 90000 + if (status == -EAGAIN) { + if (signal_pending(current)) { + status = -ERESTARTSYS; + goto bail; + } + + if (total_backoff > DLM_JOIN_TIMEOUT_MSECS) { + status = -ERESTARTSYS; + mlog(ML_NOTICE, "Timed out joining dlm domain " + "%s after %u msecs\n", dlm->name, + total_backoff); + goto bail; + } + + /* + * After you! + * No, after you! + * I insist! + * But you first! + * ... + */ + backoff = (unsigned int)(jiffies & 0x3); + backoff *= DLM_DOMAIN_BACKOFF_MS; + total_backoff += backoff; + mlog(0, "backoff %d\n", backoff); + msleep(backoff); + } + } while (status == -EAGAIN); + + if (status < 0) { + mlog_errno(status); + goto bail; + } + + status = 0; +bail: + wake_up(&dlm_domain_events); + + if (status) { + dlm_unregister_domain_handlers(dlm); + dlm_complete_thread(dlm); + dlm_complete_recovery_thread(dlm); + dlm_destroy_dlm_worker(dlm); + } + + return status; +} + +static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain, + u32 key) +{ + int i; + int ret; + struct dlm_ctxt *dlm = NULL; + + dlm = kzalloc_obj(*dlm); + if (!dlm) { + ret = -ENOMEM; + mlog_errno(ret); + goto leave; + } + + dlm->name = kstrdup(domain, GFP_KERNEL); + if (dlm->name == NULL) { + ret = -ENOMEM; + mlog_errno(ret); + goto leave; + } + + dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES); + if (!dlm->lockres_hash) { + ret = -ENOMEM; + mlog_errno(ret); + goto leave; + } + + for (i = 0; i < DLM_HASH_BUCKETS; i++) + INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i)); + + dlm->master_hash = (struct hlist_head **) + dlm_alloc_pagevec(DLM_HASH_PAGES); + if (!dlm->master_hash) { + ret = -ENOMEM; + mlog_errno(ret); + goto leave; + } + + for (i = 0; i < DLM_HASH_BUCKETS; i++) + INIT_HLIST_HEAD(dlm_master_hash(dlm, i)); + + dlm->key = key; + dlm->node_num = o2nm_this_node(); + + dlm_create_debugfs_subroot(dlm); + + spin_lock_init(&dlm->spinlock); + spin_lock_init(&dlm->master_lock); + spin_lock_init(&dlm->ast_lock); + spin_lock_init(&dlm->track_lock); + INIT_LIST_HEAD(&dlm->list); + INIT_LIST_HEAD(&dlm->dirty_list); + INIT_LIST_HEAD(&dlm->reco.resources); + INIT_LIST_HEAD(&dlm->reco.node_data); + INIT_LIST_HEAD(&dlm->purge_list); + INIT_LIST_HEAD(&dlm->dlm_domain_handlers); + INIT_LIST_HEAD(&dlm->tracking_list); + dlm->reco.state = 0; + + INIT_LIST_HEAD(&dlm->pending_asts); + INIT_LIST_HEAD(&dlm->pending_basts); + + mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n", + dlm->recovery_map, &(dlm->recovery_map[0])); + + bitmap_zero(dlm->recovery_map, O2NM_MAX_NODES); + bitmap_zero(dlm->live_nodes_map, O2NM_MAX_NODES); + bitmap_zero(dlm->domain_map, O2NM_MAX_NODES); + + dlm->dlm_thread_task = NULL; + dlm->dlm_reco_thread_task = NULL; + dlm->dlm_worker = NULL; + init_waitqueue_head(&dlm->dlm_thread_wq); + init_waitqueue_head(&dlm->dlm_reco_thread_wq); + init_waitqueue_head(&dlm->reco.event); + init_waitqueue_head(&dlm->ast_wq); + init_waitqueue_head(&dlm->migration_wq); + INIT_LIST_HEAD(&dlm->mle_hb_events); + + dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN; + init_waitqueue_head(&dlm->dlm_join_events); + + dlm->migrate_done = 0; + + dlm->reco.new_master = O2NM_INVALID_NODE_NUM; + dlm->reco.dead_node = O2NM_INVALID_NODE_NUM; + + atomic_set(&dlm->res_tot_count, 0); + atomic_set(&dlm->res_cur_count, 0); + for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) { + atomic_set(&dlm->mle_tot_count[i], 0); + atomic_set(&dlm->mle_cur_count[i], 0); + } + + spin_lock_init(&dlm->work_lock); + INIT_LIST_HEAD(&dlm->work_list); + INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work); + + kref_init(&dlm->dlm_refs); + dlm->dlm_state = DLM_CTXT_NEW; + + INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks); + + mlog(0, "context init: refcount %u\n", + kref_read(&dlm->dlm_refs)); + + ret = 0; +leave: + if (ret < 0 && dlm) { + if (dlm->master_hash) + dlm_free_pagevec((void **)dlm->master_hash, + DLM_HASH_PAGES); + + if (dlm->lockres_hash) + dlm_free_pagevec((void **)dlm->lockres_hash, + DLM_HASH_PAGES); + + kfree(dlm->name); + kfree(dlm); + dlm = NULL; + } + return dlm; +} + +/* + * Compare a requested locking protocol version against the current one. + * + * If the major numbers are different, they are incompatible. + * If the current minor is greater than the request, they are incompatible. + * If the current minor is less than or equal to the request, they are + * compatible, and the requester should run at the current minor version. + */ +static int dlm_protocol_compare(struct dlm_protocol_version *existing, + struct dlm_protocol_version *request) +{ + if (existing->pv_major != request->pv_major) + return 1; + + if (existing->pv_minor > request->pv_minor) + return 1; + + if (existing->pv_minor < request->pv_minor) + request->pv_minor = existing->pv_minor; + + return 0; +} + +/* + * dlm_register_domain: one-time setup per "domain". + * + * The filesystem passes in the requested locking version via proto. + * If registration was successful, proto will contain the negotiated + * locking protocol. + */ +struct dlm_ctxt * dlm_register_domain(const char *domain, + u32 key, + struct dlm_protocol_version *fs_proto) +{ + int ret; + struct dlm_ctxt *dlm = NULL; + struct dlm_ctxt *new_ctxt = NULL; + + if (strlen(domain) >= O2NM_MAX_NAME_LEN) { + ret = -ENAMETOOLONG; + mlog(ML_ERROR, "domain name length too long\n"); + goto leave; + } + + mlog(0, "register called for domain \"%s\"\n", domain); + +retry: + dlm = NULL; + if (signal_pending(current)) { + ret = -ERESTARTSYS; + mlog_errno(ret); + goto leave; + } + + spin_lock(&dlm_domain_lock); + + dlm = __dlm_lookup_domain(domain); + if (dlm) { + if (dlm->dlm_state != DLM_CTXT_JOINED) { + spin_unlock(&dlm_domain_lock); + + mlog(0, "This ctxt is not joined yet!\n"); + wait_event_interruptible(dlm_domain_events, + dlm_wait_on_domain_helper( + domain)); + goto retry; + } + + if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) { + spin_unlock(&dlm_domain_lock); + mlog(ML_ERROR, + "Requested locking protocol version is not " + "compatible with already registered domain " + "\"%s\"\n", domain); + ret = -EPROTO; + goto leave; + } + + __dlm_get(dlm); + dlm->num_joins++; + + spin_unlock(&dlm_domain_lock); + + ret = 0; + goto leave; + } + + /* doesn't exist */ + if (!new_ctxt) { + spin_unlock(&dlm_domain_lock); + + new_ctxt = dlm_alloc_ctxt(domain, key); + if (new_ctxt) + goto retry; + + ret = -ENOMEM; + mlog_errno(ret); + goto leave; + } + + /* a little variable switch-a-roo here... */ + dlm = new_ctxt; + new_ctxt = NULL; + + /* add the new domain */ + list_add_tail(&dlm->list, &dlm_domains); + spin_unlock(&dlm_domain_lock); + + /* + * Pass the locking protocol version into the join. If the join + * succeeds, it will have the negotiated protocol set. + */ + dlm->dlm_locking_proto = dlm_protocol; + dlm->fs_locking_proto = *fs_proto; + + ret = dlm_join_domain(dlm); + if (ret) { + mlog_errno(ret); + dlm_put(dlm); + goto leave; + } + + /* Tell the caller what locking protocol we negotiated */ + *fs_proto = dlm->fs_locking_proto; + + ret = 0; +leave: + if (new_ctxt) + dlm_free_ctxt_mem(new_ctxt); + + if (ret < 0) + dlm = ERR_PTR(ret); + + return dlm; +} +EXPORT_SYMBOL_GPL(dlm_register_domain); + +static LIST_HEAD(dlm_join_handlers); + +static void dlm_unregister_net_handlers(void) +{ + o2net_unregister_handler_list(&dlm_join_handlers); +} + +static int dlm_register_net_handlers(void) +{ + int status = 0; + + status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, + sizeof(struct dlm_query_join_request), + dlm_query_join_handler, + NULL, NULL, &dlm_join_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY, + sizeof(struct dlm_assert_joined), + dlm_assert_joined_handler, + NULL, NULL, &dlm_join_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY, + sizeof(struct dlm_cancel_join), + dlm_cancel_join_handler, + NULL, NULL, &dlm_join_handlers); + if (status) + goto bail; + + status = o2net_register_handler(DLM_QUERY_REGION, DLM_MOD_KEY, + sizeof(struct dlm_query_region), + dlm_query_region_handler, + NULL, NULL, &dlm_join_handlers); + + if (status) + goto bail; + + status = o2net_register_handler(DLM_QUERY_NODEINFO, DLM_MOD_KEY, + sizeof(struct dlm_query_nodeinfo), + dlm_query_nodeinfo_handler, + NULL, NULL, &dlm_join_handlers); +bail: + if (status < 0) + dlm_unregister_net_handlers(); + + return status; +} + +/* Domain eviction callback handling. + * + * The file system requires notification of node death *before* the + * dlm completes it's recovery work, otherwise it may be able to + * acquire locks on resources requiring recovery. Since the dlm can + * evict a node from it's domain *before* heartbeat fires, a similar + * mechanism is required. */ + +/* Eviction is not expected to happen often, so a per-domain lock is + * not necessary. Eviction callbacks are allowed to sleep for short + * periods of time. */ +static DECLARE_RWSEM(dlm_callback_sem); + +void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm, + int node_num) +{ + struct dlm_eviction_cb *cb; + + down_read(&dlm_callback_sem); + list_for_each_entry(cb, &dlm->dlm_eviction_callbacks, ec_item) { + cb->ec_func(node_num, cb->ec_data); + } + up_read(&dlm_callback_sem); +} + +void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb, + dlm_eviction_func *f, + void *data) +{ + INIT_LIST_HEAD(&cb->ec_item); + cb->ec_func = f; + cb->ec_data = data; +} +EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb); + +void dlm_register_eviction_cb(struct dlm_ctxt *dlm, + struct dlm_eviction_cb *cb) +{ + down_write(&dlm_callback_sem); + list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks); + up_write(&dlm_callback_sem); +} +EXPORT_SYMBOL_GPL(dlm_register_eviction_cb); + +void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb) +{ + down_write(&dlm_callback_sem); + list_del_init(&cb->ec_item); + up_write(&dlm_callback_sem); +} +EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb); + +static int __init dlm_init(void) +{ + int status; + + status = dlm_init_mle_cache(); + if (status) { + mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n"); + goto error; + } + + status = dlm_init_master_caches(); + if (status) { + mlog(ML_ERROR, "Could not create o2dlm_lockres and " + "o2dlm_lockname slabcaches\n"); + goto error; + } + + status = dlm_init_lock_cache(); + if (status) { + mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n"); + goto error; + } + + status = dlm_register_net_handlers(); + if (status) { + mlog(ML_ERROR, "Unable to register network handlers\n"); + goto error; + } + + dlm_create_debugfs_root(); + + return 0; +error: + dlm_unregister_net_handlers(); + dlm_destroy_lock_cache(); + dlm_destroy_master_caches(); + dlm_destroy_mle_cache(); + return -1; +} + +static void __exit dlm_exit (void) +{ + dlm_destroy_debugfs_root(); + dlm_unregister_net_handlers(); + dlm_destroy_lock_cache(); + dlm_destroy_master_caches(); + dlm_destroy_mle_cache(); +} + +MODULE_AUTHOR("Oracle"); +MODULE_LICENSE("GPL"); +MODULE_DESCRIPTION("OCFS2 Distributed Lock Management"); + +module_init(dlm_init); +module_exit(dlm_exit); diff --git a/fs/ocfs2/dlm/dlmdomain.h b/fs/ocfs2/dlm/dlmdomain.h new file mode 100644 index 000000000..815abe30a --- /dev/null +++ b/fs/ocfs2/dlm/dlmdomain.h @@ -0,0 +1,41 @@ +/* SPDX-License-Identifier: GPL-2.0-or-later */ +/* + * dlmdomain.h + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + +#ifndef DLMDOMAIN_H +#define DLMDOMAIN_H + +extern spinlock_t dlm_domain_lock; +extern struct list_head dlm_domains; + +static inline int dlm_joined(struct dlm_ctxt *dlm) +{ + int ret = 0; + + spin_lock(&dlm_domain_lock); + if (dlm->dlm_state == DLM_CTXT_JOINED) + ret = 1; + spin_unlock(&dlm_domain_lock); + + return ret; +} + +static inline int dlm_shutting_down(struct dlm_ctxt *dlm) +{ + int ret = 0; + + spin_lock(&dlm_domain_lock); + if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN) + ret = 1; + spin_unlock(&dlm_domain_lock); + + return ret; +} + +void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm, + int node_num); + +#endif diff --git a/fs/ocfs2/dlm/dlmlock.c b/fs/ocfs2/dlm/dlmlock.c new file mode 100644 index 000000000..46120eb75 --- /dev/null +++ b/fs/ocfs2/dlm/dlmlock.c @@ -0,0 +1,743 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * dlmlock.c + * + * underlying calls for lock creation + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + + +#include "../cluster/heartbeat.h" +#include "../cluster/nodemanager.h" +#include "../cluster/tcp.h" + +#include "dlmapi.h" +#include "dlmcommon.h" + +#include "dlmconvert.h" + +#define MLOG_MASK_PREFIX ML_DLM +#include "../cluster/masklog.h" + +static struct kmem_cache *dlm_lock_cache; + +static DEFINE_SPINLOCK(dlm_cookie_lock); +static u64 dlm_next_cookie = 1; + +static enum dlm_status dlm_send_remote_lock_request(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags); +static void dlm_init_lock(struct dlm_lock *newlock, int type, + u8 node, u64 cookie); +static void dlm_lock_release(struct kref *kref); +static void dlm_lock_detach_lockres(struct dlm_lock *lock); + +int dlm_init_lock_cache(void) +{ + dlm_lock_cache = kmem_cache_create("o2dlm_lock", + sizeof(struct dlm_lock), + 0, SLAB_HWCACHE_ALIGN, NULL); + if (dlm_lock_cache == NULL) + return -ENOMEM; + return 0; +} + +void dlm_destroy_lock_cache(void) +{ + kmem_cache_destroy(dlm_lock_cache); +} + +/* Tell us whether we can grant a new lock request. + * locking: + * caller needs: res->spinlock + * taken: none + * held on exit: none + * returns: 1 if the lock can be granted, 0 otherwise. + */ +static int dlm_can_grant_new_lock(struct dlm_lock_resource *res, + struct dlm_lock *lock) +{ + struct dlm_lock *tmplock; + + list_for_each_entry(tmplock, &res->granted, list) { + if (!dlm_lock_compatible(tmplock->ml.type, lock->ml.type)) + return 0; + } + + list_for_each_entry(tmplock, &res->converting, list) { + if (!dlm_lock_compatible(tmplock->ml.type, lock->ml.type)) + return 0; + if (!dlm_lock_compatible(tmplock->ml.convert_type, + lock->ml.type)) + return 0; + } + + return 1; +} + +/* performs lock creation at the lockres master site + * locking: + * caller needs: none + * taken: takes and drops res->spinlock + * held on exit: none + * returns: DLM_NORMAL, DLM_NOTQUEUED + */ +static enum dlm_status dlmlock_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags) +{ + int call_ast = 0, kick_thread = 0; + enum dlm_status status = DLM_NORMAL; + + mlog(0, "type=%d\n", lock->ml.type); + + spin_lock(&res->spinlock); + /* if called from dlm_create_lock_handler, need to + * ensure it will not sleep in dlm_wait_on_lockres */ + status = __dlm_lockres_state_to_status(res); + if (status != DLM_NORMAL && + lock->ml.node != dlm->node_num) { + /* erf. state changed after lock was dropped. */ + spin_unlock(&res->spinlock); + dlm_error(status); + return status; + } + __dlm_wait_on_lockres(res); + __dlm_lockres_reserve_ast(res); + + if (dlm_can_grant_new_lock(res, lock)) { + mlog(0, "I can grant this lock right away\n"); + /* got it right away */ + lock->lksb->status = DLM_NORMAL; + status = DLM_NORMAL; + dlm_lock_get(lock); + list_add_tail(&lock->list, &res->granted); + + /* for the recovery lock, we can't allow the ast + * to be queued since the dlmthread is already + * frozen. but the recovery lock is always locked + * with LKM_NOQUEUE so we do not need the ast in + * this special case */ + if (!dlm_is_recovery_lock(res->lockname.name, + res->lockname.len)) { + kick_thread = 1; + call_ast = 1; + } else { + mlog(0, "%s: returning DLM_NORMAL to " + "node %u for reco lock\n", dlm->name, + lock->ml.node); + } + } else { + /* for NOQUEUE request, unless we get the + * lock right away, return DLM_NOTQUEUED */ + if (flags & LKM_NOQUEUE) { + status = DLM_NOTQUEUED; + if (dlm_is_recovery_lock(res->lockname.name, + res->lockname.len)) { + mlog(0, "%s: returning NOTQUEUED to " + "node %u for reco lock\n", dlm->name, + lock->ml.node); + } + } else { + status = DLM_NORMAL; + dlm_lock_get(lock); + list_add_tail(&lock->list, &res->blocked); + kick_thread = 1; + } + } + + spin_unlock(&res->spinlock); + wake_up(&res->wq); + + /* either queue the ast or release it */ + if (call_ast) + dlm_queue_ast(dlm, lock); + else + dlm_lockres_release_ast(dlm, res); + + dlm_lockres_calc_usage(dlm, res); + if (kick_thread) + dlm_kick_thread(dlm, res); + + return status; +} + +void dlm_revert_pending_lock(struct dlm_lock_resource *res, + struct dlm_lock *lock) +{ + /* remove from local queue if it failed */ + list_del_init(&lock->list); + lock->lksb->flags &= ~DLM_LKSB_GET_LVB; +} + + +/* + * locking: + * caller needs: none + * taken: takes and drops res->spinlock + * held on exit: none + * returns: DLM_DENIED, DLM_RECOVERING, or net status + */ +static enum dlm_status dlmlock_remote(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags) +{ + enum dlm_status status = DLM_DENIED; + int lockres_changed = 1; + + mlog(0, "type=%d, lockres %.*s, flags = 0x%x\n", + lock->ml.type, res->lockname.len, + res->lockname.name, flags); + + /* + * Wait if resource is getting recovered, remastered, etc. + * If the resource was remastered and new owner is self, then exit. + */ + spin_lock(&res->spinlock); + __dlm_wait_on_lockres(res); + if (res->owner == dlm->node_num) { + spin_unlock(&res->spinlock); + return DLM_RECOVERING; + } + res->state |= DLM_LOCK_RES_IN_PROGRESS; + + /* add lock to local (secondary) queue */ + dlm_lock_get(lock); + list_add_tail(&lock->list, &res->blocked); + lock->lock_pending = 1; + spin_unlock(&res->spinlock); + + /* spec seems to say that you will get DLM_NORMAL when the lock + * has been queued, meaning we need to wait for a reply here. */ + status = dlm_send_remote_lock_request(dlm, res, lock, flags); + + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_IN_PROGRESS; + lock->lock_pending = 0; + if (status != DLM_NORMAL) { + if (status == DLM_RECOVERING && + dlm_is_recovery_lock(res->lockname.name, + res->lockname.len)) { + /* recovery lock was mastered by dead node. + * we need to have calc_usage shoot down this + * lockres and completely remaster it. */ + mlog(0, "%s: recovery lock was owned by " + "dead node %u, remaster it now.\n", + dlm->name, res->owner); + } else if (status != DLM_NOTQUEUED) { + /* + * DO NOT call calc_usage, as this would unhash + * the remote lockres before we ever get to use + * it. treat as if we never made any change to + * the lockres. + */ + lockres_changed = 0; + dlm_error(status); + } + dlm_revert_pending_lock(res, lock); + dlm_lock_put(lock); + } else if (dlm_is_recovery_lock(res->lockname.name, + res->lockname.len)) { + /* special case for the $RECOVERY lock. + * there will never be an AST delivered to put + * this lock on the proper secondary queue + * (granted), so do it manually. */ + mlog(0, "%s: $RECOVERY lock for this node (%u) is " + "mastered by %u; got lock, manually granting (no ast)\n", + dlm->name, dlm->node_num, res->owner); + list_move_tail(&lock->list, &res->granted); + } + spin_unlock(&res->spinlock); + + if (lockres_changed) + dlm_lockres_calc_usage(dlm, res); + + wake_up(&res->wq); + return status; +} + + +/* for remote lock creation. + * locking: + * caller needs: none, but need res->state & DLM_LOCK_RES_IN_PROGRESS + * taken: none + * held on exit: none + * returns: DLM_NOLOCKMGR, or net status + */ +static enum dlm_status dlm_send_remote_lock_request(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, int flags) +{ + struct dlm_create_lock create; + int tmpret, status = 0; + enum dlm_status ret; + + memset(&create, 0, sizeof(create)); + create.node_idx = dlm->node_num; + create.requested_type = lock->ml.type; + create.cookie = lock->ml.cookie; + create.namelen = res->lockname.len; + create.flags = cpu_to_be32(flags); + memcpy(create.name, res->lockname.name, create.namelen); + + tmpret = o2net_send_message(DLM_CREATE_LOCK_MSG, dlm->key, &create, + sizeof(create), res->owner, &status); + if (tmpret >= 0) { + ret = status; + if (ret == DLM_REJECTED) { + mlog(ML_ERROR, "%s: res %.*s, Stale lockres no longer " + "owned by node %u. That node is coming back up " + "currently.\n", dlm->name, create.namelen, + create.name, res->owner); + dlm_print_one_lock_resource(res); + BUG(); + } + } else { + mlog(ML_ERROR, "%s: res %.*s, Error %d send CREATE LOCK to " + "node %u\n", dlm->name, create.namelen, create.name, + tmpret, res->owner); + if (dlm_is_host_down(tmpret)) + ret = DLM_RECOVERING; + else + ret = dlm_err_to_dlm_status(tmpret); + } + + return ret; +} + +void dlm_lock_get(struct dlm_lock *lock) +{ + kref_get(&lock->lock_refs); +} + +void dlm_lock_put(struct dlm_lock *lock) +{ + kref_put(&lock->lock_refs, dlm_lock_release); +} + +static void dlm_lock_release(struct kref *kref) +{ + struct dlm_lock *lock; + + lock = container_of(kref, struct dlm_lock, lock_refs); + + BUG_ON(!list_empty(&lock->list)); + BUG_ON(!list_empty(&lock->ast_list)); + BUG_ON(!list_empty(&lock->bast_list)); + BUG_ON(lock->ast_pending); + BUG_ON(lock->bast_pending); + + dlm_lock_detach_lockres(lock); + + if (lock->lksb_kernel_allocated) { + mlog(0, "freeing kernel-allocated lksb\n"); + kfree(lock->lksb); + } + kmem_cache_free(dlm_lock_cache, lock); +} + +/* associate a lock with it's lockres, getting a ref on the lockres */ +void dlm_lock_attach_lockres(struct dlm_lock *lock, + struct dlm_lock_resource *res) +{ + dlm_lockres_get(res); + lock->lockres = res; +} + +/* drop ref on lockres, if there is still one associated with lock */ +static void dlm_lock_detach_lockres(struct dlm_lock *lock) +{ + struct dlm_lock_resource *res; + + res = lock->lockres; + if (res) { + lock->lockres = NULL; + mlog(0, "removing lock's lockres reference\n"); + dlm_lockres_put(res); + } +} + +static void dlm_init_lock(struct dlm_lock *newlock, int type, + u8 node, u64 cookie) +{ + INIT_LIST_HEAD(&newlock->list); + INIT_LIST_HEAD(&newlock->ast_list); + INIT_LIST_HEAD(&newlock->bast_list); + spin_lock_init(&newlock->spinlock); + newlock->ml.type = type; + newlock->ml.convert_type = LKM_IVMODE; + newlock->ml.highest_blocked = LKM_IVMODE; + newlock->ml.node = node; + newlock->ml.pad1 = 0; + newlock->ml.list = 0; + newlock->ml.flags = 0; + newlock->ast = NULL; + newlock->bast = NULL; + newlock->astdata = NULL; + newlock->ml.cookie = cpu_to_be64(cookie); + newlock->ast_pending = 0; + newlock->bast_pending = 0; + newlock->convert_pending = 0; + newlock->lock_pending = 0; + newlock->unlock_pending = 0; + newlock->cancel_pending = 0; + newlock->lksb_kernel_allocated = 0; + + kref_init(&newlock->lock_refs); +} + +struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie, + struct dlm_lockstatus *lksb) +{ + struct dlm_lock *lock; + int kernel_allocated = 0; + + lock = kmem_cache_zalloc(dlm_lock_cache, GFP_NOFS); + if (!lock) + return NULL; + + if (!lksb) { + /* zero memory only if kernel-allocated */ + lksb = kzalloc_obj(*lksb, GFP_NOFS); + if (!lksb) { + kmem_cache_free(dlm_lock_cache, lock); + return NULL; + } + kernel_allocated = 1; + } + + dlm_init_lock(lock, type, node, cookie); + if (kernel_allocated) + lock->lksb_kernel_allocated = 1; + lock->lksb = lksb; + lksb->lockid = lock; + return lock; +} + +/* handler for lock creation net message + * locking: + * caller needs: none + * taken: takes and drops res->spinlock + * held on exit: none + * returns: DLM_NORMAL, DLM_SYSERR, DLM_IVLOCKID, DLM_NOTQUEUED + */ +int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_create_lock *create = (struct dlm_create_lock *)msg->buf; + struct dlm_lock_resource *res = NULL; + struct dlm_lock *newlock = NULL; + struct dlm_lockstatus *lksb = NULL; + enum dlm_status status = DLM_NORMAL; + char *name; + unsigned int namelen; + + BUG_ON(!dlm); + + if (!dlm_grab(dlm)) + return DLM_REJECTED; + + name = create->name; + namelen = create->namelen; + status = DLM_REJECTED; + if (!dlm_domain_fully_joined(dlm)) { + mlog(ML_ERROR, "Domain %s not fully joined, but node %u is " + "sending a create_lock message for lock %.*s!\n", + dlm->name, create->node_idx, namelen, name); + dlm_error(status); + goto leave; + } + + status = DLM_IVBUFLEN; + if (namelen > DLM_LOCKID_NAME_MAX) { + dlm_error(status); + goto leave; + } + + status = DLM_SYSERR; + newlock = dlm_new_lock(create->requested_type, + create->node_idx, + be64_to_cpu(create->cookie), NULL); + if (!newlock) { + dlm_error(status); + goto leave; + } + + lksb = newlock->lksb; + + if (be32_to_cpu(create->flags) & LKM_GET_LVB) { + lksb->flags |= DLM_LKSB_GET_LVB; + mlog(0, "set DLM_LKSB_GET_LVB flag\n"); + } + + status = DLM_IVLOCKID; + res = dlm_lookup_lockres(dlm, name, namelen); + if (!res) { + dlm_error(status); + goto leave; + } + + spin_lock(&res->spinlock); + status = __dlm_lockres_state_to_status(res); + spin_unlock(&res->spinlock); + + if (status != DLM_NORMAL) { + mlog(0, "lockres recovering/migrating/in-progress\n"); + goto leave; + } + + dlm_lock_attach_lockres(newlock, res); + + status = dlmlock_master(dlm, res, newlock, be32_to_cpu(create->flags)); +leave: + if (status != DLM_NORMAL) + if (newlock) + dlm_lock_put(newlock); + + if (res) + dlm_lockres_put(res); + + dlm_put(dlm); + + return status; +} + + +/* fetch next node-local (u8 nodenum + u56 cookie) into u64 */ +static inline void dlm_get_next_cookie(u8 node_num, u64 *cookie) +{ + u64 tmpnode = node_num; + + /* shift single byte of node num into top 8 bits */ + tmpnode <<= 56; + + spin_lock(&dlm_cookie_lock); + *cookie = (dlm_next_cookie | tmpnode); + if (++dlm_next_cookie & 0xff00000000000000ull) { + mlog(0, "This node's cookie will now wrap!\n"); + dlm_next_cookie = 1; + } + spin_unlock(&dlm_cookie_lock); +} + +enum dlm_status dlmlock(struct dlm_ctxt *dlm, int mode, + struct dlm_lockstatus *lksb, int flags, + const char *name, int namelen, dlm_astlockfunc_t *ast, + void *data, dlm_bastlockfunc_t *bast) +{ + enum dlm_status status; + struct dlm_lock_resource *res = NULL; + struct dlm_lock *lock = NULL; + int convert = 0, recovery = 0; + + /* yes this function is a mess. + * TODO: clean this up. lots of common code in the + * lock and convert paths, especially in the retry blocks */ + if (!lksb) { + dlm_error(DLM_BADARGS); + return DLM_BADARGS; + } + + status = DLM_BADPARAM; + if (mode != LKM_EXMODE && mode != LKM_PRMODE && mode != LKM_NLMODE) { + dlm_error(status); + goto error; + } + + if (flags & ~LKM_VALID_FLAGS) { + dlm_error(status); + goto error; + } + + convert = (flags & LKM_CONVERT); + recovery = (flags & LKM_RECOVERY); + + if (recovery && + (!dlm_is_recovery_lock(name, namelen) || convert) ) { + dlm_error(status); + goto error; + } + if (convert && (flags & LKM_LOCAL)) { + mlog(ML_ERROR, "strange LOCAL convert request!\n"); + goto error; + } + + if (convert) { + /* CONVERT request */ + + /* if converting, must pass in a valid dlm_lock */ + lock = lksb->lockid; + if (!lock) { + mlog(ML_ERROR, "NULL lock pointer in convert " + "request\n"); + goto error; + } + + res = lock->lockres; + if (!res) { + mlog(ML_ERROR, "NULL lockres pointer in convert " + "request\n"); + goto error; + } + dlm_lockres_get(res); + + /* XXX: for ocfs2 purposes, the ast/bast/astdata/lksb are + * static after the original lock call. convert requests will + * ensure that everything is the same, or return DLM_BADARGS. + * this means that DLM_DENIED_NOASTS will never be returned. + */ + if (lock->lksb != lksb || lock->ast != ast || + lock->bast != bast || lock->astdata != data) { + status = DLM_BADARGS; + mlog(ML_ERROR, "new args: lksb=%p, ast=%p, bast=%p, " + "astdata=%p\n", lksb, ast, bast, data); + mlog(ML_ERROR, "orig args: lksb=%p, ast=%p, bast=%p, " + "astdata=%p\n", lock->lksb, lock->ast, + lock->bast, lock->astdata); + goto error; + } +retry_convert: + dlm_wait_for_recovery(dlm); + + if (res->owner == dlm->node_num) + status = dlmconvert_master(dlm, res, lock, flags, mode); + else + status = dlmconvert_remote(dlm, res, lock, flags, mode); + if (status == DLM_RECOVERING || status == DLM_MIGRATING || + status == DLM_FORWARD) { + /* for now, see how this works without sleeping + * and just retry right away. I suspect the reco + * or migration will complete fast enough that + * no waiting will be necessary */ + mlog(0, "retrying convert with migration/recovery/" + "in-progress\n"); + msleep(100); + goto retry_convert; + } + } else { + u64 tmpcookie; + + /* LOCK request */ + status = DLM_BADARGS; + if (!name) { + dlm_error(status); + goto error; + } + + status = DLM_IVBUFLEN; + if (namelen > DLM_LOCKID_NAME_MAX || namelen < 1) { + dlm_error(status); + goto error; + } + + dlm_get_next_cookie(dlm->node_num, &tmpcookie); + lock = dlm_new_lock(mode, dlm->node_num, tmpcookie, lksb); + if (!lock) { + dlm_error(status); + goto error; + } + + if (!recovery) + dlm_wait_for_recovery(dlm); + + /* find or create the lock resource */ + res = dlm_get_lock_resource(dlm, name, namelen, flags); + if (!res) { + status = DLM_IVLOCKID; + dlm_error(status); + goto error; + } + + mlog(0, "type=%d, flags = 0x%x\n", mode, flags); + mlog(0, "creating lock: lock=%p res=%p\n", lock, res); + + dlm_lock_attach_lockres(lock, res); + lock->ast = ast; + lock->bast = bast; + lock->astdata = data; + +retry_lock: + if (flags & LKM_VALBLK) { + mlog(0, "LKM_VALBLK passed by caller\n"); + + /* LVB requests for non PR, PW or EX locks are + * ignored. */ + if (mode < LKM_PRMODE) + flags &= ~LKM_VALBLK; + else { + flags |= LKM_GET_LVB; + lock->lksb->flags |= DLM_LKSB_GET_LVB; + } + } + + if (res->owner == dlm->node_num) + status = dlmlock_master(dlm, res, lock, flags); + else + status = dlmlock_remote(dlm, res, lock, flags); + + if (status == DLM_RECOVERING || status == DLM_MIGRATING || + status == DLM_FORWARD) { + msleep(100); + if (recovery) { + if (status != DLM_RECOVERING) + goto retry_lock; + /* wait to see the node go down, then + * drop down and allow the lockres to + * get cleaned up. need to remaster. */ + dlm_wait_for_node_death(dlm, res->owner, + DLM_NODE_DEATH_WAIT_MAX); + } else { + dlm_wait_for_recovery(dlm); + goto retry_lock; + } + } + + /* Inflight taken in dlm_get_lock_resource() is dropped here */ + spin_lock(&res->spinlock); + dlm_lockres_drop_inflight_ref(dlm, res); + spin_unlock(&res->spinlock); + + dlm_lockres_calc_usage(dlm, res); + dlm_kick_thread(dlm, res); + + if (status != DLM_NORMAL) { + lock->lksb->flags &= ~DLM_LKSB_GET_LVB; + if (status != DLM_NOTQUEUED) + dlm_error(status); + goto error; + } + } + +error: + if (status != DLM_NORMAL) { + if (lock && !convert) + dlm_lock_put(lock); + // this is kind of unnecessary + lksb->status = status; + } + + /* put lockres ref from the convert path + * or from dlm_get_lock_resource */ + if (res) + dlm_lockres_put(res); + + return status; +} +EXPORT_SYMBOL_GPL(dlmlock); diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c new file mode 100644 index 000000000..612969867 --- /dev/null +++ b/fs/ocfs2/dlm/dlmmaster.c @@ -0,0 +1,3557 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * dlmmod.c + * + * standalone DLM module + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "../cluster/heartbeat.h" +#include "../cluster/nodemanager.h" +#include "../cluster/tcp.h" + +#include "dlmapi.h" +#include "dlmcommon.h" +#include "dlmdomain.h" +#include "dlmdebug.h" + +#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER) +#include "../cluster/masklog.h" + +static void dlm_mle_node_down(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle, + struct o2nm_node *node, + int idx); +static void dlm_mle_node_up(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle, + struct o2nm_node *node, + int idx); + +static void dlm_assert_master_worker(struct dlm_work_item *item, void *data); +static int dlm_do_assert_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + void *nodemap, u32 flags); +static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data); + +static inline int dlm_mle_equal(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle, + const char *name, + unsigned int namelen) +{ + if (dlm != mle->dlm) + return 0; + + if (namelen != mle->mnamelen || + memcmp(name, mle->mname, namelen) != 0) + return 0; + + return 1; +} + +static struct kmem_cache *dlm_lockres_cache; +static struct kmem_cache *dlm_lockname_cache; +static struct kmem_cache *dlm_mle_cache; + +static void dlm_mle_release(struct kref *kref); +static void dlm_init_mle(struct dlm_master_list_entry *mle, + enum dlm_mle_type type, + struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + const char *name, + unsigned int namelen); +static void dlm_put_mle(struct dlm_master_list_entry *mle); +static void __dlm_put_mle(struct dlm_master_list_entry *mle); +static int dlm_find_mle(struct dlm_ctxt *dlm, + struct dlm_master_list_entry **mle, + char *name, unsigned int namelen); + +static int dlm_do_master_request(struct dlm_lock_resource *res, + struct dlm_master_list_entry *mle, int to); + + +static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_master_list_entry *mle, + int *blocked); +static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_master_list_entry *mle, + int blocked); +static int dlm_add_migration_mle(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_master_list_entry *mle, + struct dlm_master_list_entry **oldmle, + const char *name, unsigned int namelen, + u8 new_master, u8 master); + +static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); +static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); +static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + u8 target); +static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res); + + +int dlm_is_host_down(int errno) +{ + switch (errno) { + case -EBADF: + case -ECONNREFUSED: + case -ENOTCONN: + case -ECONNRESET: + case -EPIPE: + case -EHOSTDOWN: + case -EHOSTUNREACH: + case -ETIMEDOUT: + case -ECONNABORTED: + case -ENETDOWN: + case -ENETUNREACH: + case -ENETRESET: + case -ESHUTDOWN: + case -ENOPROTOOPT: + case -EINVAL: /* if returned from our tcp code, + this means there is no socket */ + return 1; + } + return 0; +} + + +/* + * MASTER LIST FUNCTIONS + */ + + +/* + * regarding master list entries and heartbeat callbacks: + * + * in order to avoid sleeping and allocation that occurs in + * heartbeat, master list entries are simply attached to the + * dlm's established heartbeat callbacks. the mle is attached + * when it is created, and since the dlm->spinlock is held at + * that time, any heartbeat event will be properly discovered + * by the mle. the mle needs to be detached from the + * dlm->mle_hb_events list as soon as heartbeat events are no + * longer useful to the mle, and before the mle is freed. + * + * as a general rule, heartbeat events are no longer needed by + * the mle once an "answer" regarding the lock master has been + * received. + */ +static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle) +{ + assert_spin_locked(&dlm->spinlock); + + list_add_tail(&mle->hb_events, &dlm->mle_hb_events); +} + + +static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle) +{ + if (!list_empty(&mle->hb_events)) + list_del_init(&mle->hb_events); +} + + +static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle) +{ + spin_lock(&dlm->spinlock); + __dlm_mle_detach_hb_events(dlm, mle); + spin_unlock(&dlm->spinlock); +} + +static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle) +{ + struct dlm_ctxt *dlm; + dlm = mle->dlm; + + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&dlm->master_lock); + mle->inuse++; + kref_get(&mle->mle_refs); +} + +static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle) +{ + struct dlm_ctxt *dlm; + dlm = mle->dlm; + + spin_lock(&dlm->spinlock); + spin_lock(&dlm->master_lock); + mle->inuse--; + __dlm_put_mle(mle); + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); + +} + +/* remove from list and free */ +static void __dlm_put_mle(struct dlm_master_list_entry *mle) +{ + struct dlm_ctxt *dlm; + dlm = mle->dlm; + + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&dlm->master_lock); + if (!kref_read(&mle->mle_refs)) { + /* this may or may not crash, but who cares. + * it's a BUG. */ + mlog(ML_ERROR, "bad mle: %p\n", mle); + dlm_print_one_mle(mle); + BUG(); + } else + kref_put(&mle->mle_refs, dlm_mle_release); +} + + +/* must not have any spinlocks coming in */ +static void dlm_put_mle(struct dlm_master_list_entry *mle) +{ + struct dlm_ctxt *dlm; + dlm = mle->dlm; + + spin_lock(&dlm->spinlock); + spin_lock(&dlm->master_lock); + __dlm_put_mle(mle); + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); +} + +static inline void dlm_get_mle(struct dlm_master_list_entry *mle) +{ + kref_get(&mle->mle_refs); +} + +static void dlm_init_mle(struct dlm_master_list_entry *mle, + enum dlm_mle_type type, + struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + const char *name, + unsigned int namelen) +{ + assert_spin_locked(&dlm->spinlock); + + mle->dlm = dlm; + mle->type = type; + INIT_HLIST_NODE(&mle->master_hash_node); + INIT_LIST_HEAD(&mle->hb_events); + bitmap_zero(mle->maybe_map, O2NM_MAX_NODES); + spin_lock_init(&mle->spinlock); + init_waitqueue_head(&mle->wq); + atomic_set(&mle->woken, 0); + kref_init(&mle->mle_refs); + bitmap_zero(mle->response_map, O2NM_MAX_NODES); + mle->master = O2NM_MAX_NODES; + mle->new_master = O2NM_MAX_NODES; + mle->inuse = 0; + + BUG_ON(mle->type != DLM_MLE_BLOCK && + mle->type != DLM_MLE_MASTER && + mle->type != DLM_MLE_MIGRATION); + + if (mle->type == DLM_MLE_MASTER) { + BUG_ON(!res); + mle->mleres = res; + memcpy(mle->mname, res->lockname.name, res->lockname.len); + mle->mnamelen = res->lockname.len; + mle->mnamehash = res->lockname.hash; + } else { + BUG_ON(!name); + mle->mleres = NULL; + memcpy(mle->mname, name, namelen); + mle->mnamelen = namelen; + mle->mnamehash = dlm_lockid_hash(name, namelen); + } + + atomic_inc(&dlm->mle_tot_count[mle->type]); + atomic_inc(&dlm->mle_cur_count[mle->type]); + + /* copy off the node_map and register hb callbacks on our copy */ + bitmap_copy(mle->node_map, dlm->domain_map, O2NM_MAX_NODES); + bitmap_copy(mle->vote_map, dlm->domain_map, O2NM_MAX_NODES); + clear_bit(dlm->node_num, mle->vote_map); + clear_bit(dlm->node_num, mle->node_map); + + /* attach the mle to the domain node up/down events */ + __dlm_mle_attach_hb_events(dlm, mle); +} + +void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle) +{ + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&dlm->master_lock); + + if (!hlist_unhashed(&mle->master_hash_node)) + hlist_del_init(&mle->master_hash_node); +} + +void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle) +{ + struct hlist_head *bucket; + + assert_spin_locked(&dlm->master_lock); + + bucket = dlm_master_hash(dlm, mle->mnamehash); + hlist_add_head(&mle->master_hash_node, bucket); +} + +/* returns 1 if found, 0 if not */ +static int dlm_find_mle(struct dlm_ctxt *dlm, + struct dlm_master_list_entry **mle, + char *name, unsigned int namelen) +{ + struct dlm_master_list_entry *tmpmle; + struct hlist_head *bucket; + unsigned int hash; + + assert_spin_locked(&dlm->master_lock); + + hash = dlm_lockid_hash(name, namelen); + bucket = dlm_master_hash(dlm, hash); + hlist_for_each_entry(tmpmle, bucket, master_hash_node) { + if (!dlm_mle_equal(dlm, tmpmle, name, namelen)) + continue; + dlm_get_mle(tmpmle); + *mle = tmpmle; + return 1; + } + return 0; +} + +void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up) +{ + struct dlm_master_list_entry *mle; + + assert_spin_locked(&dlm->spinlock); + + list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) { + if (node_up) + dlm_mle_node_up(dlm, mle, NULL, idx); + else + dlm_mle_node_down(dlm, mle, NULL, idx); + } +} + +static void dlm_mle_node_down(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle, + struct o2nm_node *node, int idx) +{ + spin_lock(&mle->spinlock); + + if (!test_bit(idx, mle->node_map)) + mlog(0, "node %u already removed from nodemap!\n", idx); + else + clear_bit(idx, mle->node_map); + + spin_unlock(&mle->spinlock); +} + +static void dlm_mle_node_up(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle, + struct o2nm_node *node, int idx) +{ + spin_lock(&mle->spinlock); + + if (test_bit(idx, mle->node_map)) + mlog(0, "node %u already in node map!\n", idx); + else + set_bit(idx, mle->node_map); + + spin_unlock(&mle->spinlock); +} + + +int dlm_init_mle_cache(void) +{ + dlm_mle_cache = kmem_cache_create("o2dlm_mle", + sizeof(struct dlm_master_list_entry), + 0, SLAB_HWCACHE_ALIGN, + NULL); + if (dlm_mle_cache == NULL) + return -ENOMEM; + return 0; +} + +void dlm_destroy_mle_cache(void) +{ + kmem_cache_destroy(dlm_mle_cache); +} + +static void dlm_mle_release(struct kref *kref) +{ + struct dlm_master_list_entry *mle; + struct dlm_ctxt *dlm; + + mle = container_of(kref, struct dlm_master_list_entry, mle_refs); + dlm = mle->dlm; + + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&dlm->master_lock); + + mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname, + mle->type); + + /* remove from list if not already */ + __dlm_unlink_mle(dlm, mle); + + /* detach the mle from the domain node up/down events */ + __dlm_mle_detach_hb_events(dlm, mle); + + atomic_dec(&dlm->mle_cur_count[mle->type]); + + /* NOTE: kfree under spinlock here. + * if this is bad, we can move this to a freelist. */ + kmem_cache_free(dlm_mle_cache, mle); +} + + +/* + * LOCK RESOURCE FUNCTIONS + */ + +int dlm_init_master_caches(void) +{ + dlm_lockres_cache = kmem_cache_create("o2dlm_lockres", + sizeof(struct dlm_lock_resource), + 0, SLAB_HWCACHE_ALIGN, NULL); + if (!dlm_lockres_cache) + goto bail; + + dlm_lockname_cache = kmem_cache_create("o2dlm_lockname", + DLM_LOCKID_NAME_MAX, 0, + SLAB_HWCACHE_ALIGN, NULL); + if (!dlm_lockname_cache) + goto bail; + + return 0; +bail: + dlm_destroy_master_caches(); + return -ENOMEM; +} + +void dlm_destroy_master_caches(void) +{ + kmem_cache_destroy(dlm_lockname_cache); + dlm_lockname_cache = NULL; + + kmem_cache_destroy(dlm_lockres_cache); + dlm_lockres_cache = NULL; +} + +static void dlm_lockres_release(struct kref *kref) +{ + struct dlm_lock_resource *res; + struct dlm_ctxt *dlm; + + res = container_of(kref, struct dlm_lock_resource, refs); + dlm = res->dlm; + + /* This should not happen -- all lockres' have a name + * associated with them at init time. */ + BUG_ON(!res->lockname.name); + + mlog(0, "destroying lockres %.*s\n", res->lockname.len, + res->lockname.name); + + atomic_dec(&dlm->res_cur_count); + + if (!hlist_unhashed(&res->hash_node) || + !list_empty(&res->granted) || + !list_empty(&res->converting) || + !list_empty(&res->blocked) || + !list_empty(&res->dirty) || + !list_empty(&res->recovering) || + !list_empty(&res->purge)) { + mlog(ML_ERROR, + "Going to BUG for resource %.*s." + " We're on a list! [%c%c%c%c%c%c%c]\n", + res->lockname.len, res->lockname.name, + !hlist_unhashed(&res->hash_node) ? 'H' : ' ', + !list_empty(&res->granted) ? 'G' : ' ', + !list_empty(&res->converting) ? 'C' : ' ', + !list_empty(&res->blocked) ? 'B' : ' ', + !list_empty(&res->dirty) ? 'D' : ' ', + !list_empty(&res->recovering) ? 'R' : ' ', + !list_empty(&res->purge) ? 'P' : ' '); + + dlm_print_one_lock_resource(res); + } + + /* By the time we're ready to blow this guy away, we shouldn't + * be on any lists. */ + BUG_ON(!hlist_unhashed(&res->hash_node)); + BUG_ON(!list_empty(&res->granted)); + BUG_ON(!list_empty(&res->converting)); + BUG_ON(!list_empty(&res->blocked)); + BUG_ON(!list_empty(&res->dirty)); + BUG_ON(!list_empty(&res->recovering)); + BUG_ON(!list_empty(&res->purge)); + + kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name); + + kmem_cache_free(dlm_lockres_cache, res); +} + +void dlm_lockres_put(struct dlm_lock_resource *res) +{ + kref_put(&res->refs, dlm_lockres_release); +} + +static void dlm_init_lockres(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + const char *name, unsigned int namelen) +{ + char *qname; + + /* If we memset here, we lose our reference to the kmalloc'd + * res->lockname.name, so be sure to init every field + * correctly! */ + + qname = (char *) res->lockname.name; + memcpy(qname, name, namelen); + + res->lockname.len = namelen; + res->lockname.hash = dlm_lockid_hash(name, namelen); + + init_waitqueue_head(&res->wq); + spin_lock_init(&res->spinlock); + INIT_HLIST_NODE(&res->hash_node); + INIT_LIST_HEAD(&res->granted); + INIT_LIST_HEAD(&res->converting); + INIT_LIST_HEAD(&res->blocked); + INIT_LIST_HEAD(&res->dirty); + INIT_LIST_HEAD(&res->recovering); + INIT_LIST_HEAD(&res->purge); + INIT_LIST_HEAD(&res->tracking); + atomic_set(&res->asts_reserved, 0); + res->migration_pending = 0; + res->inflight_locks = 0; + res->inflight_assert_workers = 0; + + res->dlm = dlm; + + kref_init(&res->refs); + + atomic_inc(&dlm->res_tot_count); + atomic_inc(&dlm->res_cur_count); + + /* just for consistency */ + spin_lock(&res->spinlock); + dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN); + spin_unlock(&res->spinlock); + + res->state = DLM_LOCK_RES_IN_PROGRESS; + + res->last_used = 0; + + spin_lock(&dlm->track_lock); + list_add_tail(&res->tracking, &dlm->tracking_list); + spin_unlock(&dlm->track_lock); + + memset(res->lvb, 0, DLM_LVB_LEN); + bitmap_zero(res->refmap, O2NM_MAX_NODES); +} + +struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm, + const char *name, + unsigned int namelen) +{ + struct dlm_lock_resource *res = NULL; + + res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS); + if (!res) + goto error; + + res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS); + if (!res->lockname.name) + goto error; + + dlm_init_lockres(dlm, res, name, namelen); + return res; + +error: + if (res) + kmem_cache_free(dlm_lockres_cache, res); + return NULL; +} + +void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, int bit) +{ + assert_spin_locked(&res->spinlock); + + mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len, + res->lockname.name, bit, __builtin_return_address(0)); + + set_bit(bit, res->refmap); +} + +void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, int bit) +{ + assert_spin_locked(&res->spinlock); + + mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len, + res->lockname.name, bit, __builtin_return_address(0)); + + clear_bit(bit, res->refmap); +} + +static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + res->inflight_locks++; + + mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name, + res->lockname.len, res->lockname.name, res->inflight_locks, + __builtin_return_address(0)); +} + +void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + assert_spin_locked(&res->spinlock); + __dlm_lockres_grab_inflight_ref(dlm, res); +} + +void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + assert_spin_locked(&res->spinlock); + + BUG_ON(res->inflight_locks == 0); + + res->inflight_locks--; + + mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name, + res->lockname.len, res->lockname.name, res->inflight_locks, + __builtin_return_address(0)); + + wake_up(&res->wq); +} + +void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + assert_spin_locked(&res->spinlock); + res->inflight_assert_workers++; + mlog(0, "%s:%.*s: inflight assert worker++: now %u\n", + dlm->name, res->lockname.len, res->lockname.name, + res->inflight_assert_workers); +} + +static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + assert_spin_locked(&res->spinlock); + BUG_ON(res->inflight_assert_workers == 0); + res->inflight_assert_workers--; + mlog(0, "%s:%.*s: inflight assert worker--: now %u\n", + dlm->name, res->lockname.len, res->lockname.name, + res->inflight_assert_workers); +} + +static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + spin_lock(&res->spinlock); + __dlm_lockres_drop_inflight_worker(dlm, res); + spin_unlock(&res->spinlock); +} + +/* + * lookup a lock resource by name. + * may already exist in the hashtable. + * lockid is null terminated + * + * if not, allocate enough for the lockres and for + * the temporary structure used in doing the mastering. + * + * also, do a lookup in the dlm->master_list to see + * if another node has begun mastering the same lock. + * if so, there should be a block entry in there + * for this name, and we should *not* attempt to master + * the lock here. need to wait around for that node + * to assert_master (or die). + * + */ +struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm, + const char *lockid, + int namelen, + int flags) +{ + struct dlm_lock_resource *tmpres=NULL, *res=NULL; + struct dlm_master_list_entry *mle = NULL; + struct dlm_master_list_entry *alloc_mle = NULL; + int blocked = 0; + int ret, nodenum; + struct dlm_node_iter iter; + unsigned int hash; + int tries = 0; + int bit, wait_on_recovery = 0; + + BUG_ON(!lockid); + + hash = dlm_lockid_hash(lockid, namelen); + + mlog(0, "get lockres %s (len %d)\n", lockid, namelen); + +lookup: + spin_lock(&dlm->spinlock); + tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash); + if (tmpres) { + spin_unlock(&dlm->spinlock); + spin_lock(&tmpres->spinlock); + + /* + * Right after dlm spinlock was released, dlm_thread could have + * purged the lockres. Check if lockres got unhashed. If so + * start over. + */ + if (hlist_unhashed(&tmpres->hash_node)) { + spin_unlock(&tmpres->spinlock); + dlm_lockres_put(tmpres); + tmpres = NULL; + goto lookup; + } + + /* Wait on the thread that is mastering the resource */ + if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { + __dlm_wait_on_lockres(tmpres); + BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN); + spin_unlock(&tmpres->spinlock); + dlm_lockres_put(tmpres); + tmpres = NULL; + goto lookup; + } + + /* Wait on the resource purge to complete before continuing */ + if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) { + BUG_ON(tmpres->owner == dlm->node_num); + __dlm_wait_on_lockres_flags(tmpres, + DLM_LOCK_RES_DROPPING_REF); + spin_unlock(&tmpres->spinlock); + dlm_lockres_put(tmpres); + tmpres = NULL; + goto lookup; + } + + /* Grab inflight ref to pin the resource */ + dlm_lockres_grab_inflight_ref(dlm, tmpres); + + spin_unlock(&tmpres->spinlock); + if (res) { + spin_lock(&dlm->track_lock); + if (!list_empty(&res->tracking)) + list_del_init(&res->tracking); + else + mlog(ML_ERROR, "Resource %.*s not " + "on the Tracking list\n", + res->lockname.len, + res->lockname.name); + spin_unlock(&dlm->track_lock); + dlm_lockres_put(res); + } + res = tmpres; + goto leave; + } + + if (!res) { + spin_unlock(&dlm->spinlock); + mlog(0, "allocating a new resource\n"); + /* nothing found and we need to allocate one. */ + alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); + if (!alloc_mle) + goto leave; + res = dlm_new_lockres(dlm, lockid, namelen); + if (!res) + goto leave; + goto lookup; + } + + mlog(0, "no lockres found, allocated our own: %p\n", res); + + if (flags & LKM_LOCAL) { + /* caller knows it's safe to assume it's not mastered elsewhere + * DONE! return right away */ + spin_lock(&res->spinlock); + dlm_change_lockres_owner(dlm, res, dlm->node_num); + __dlm_insert_lockres(dlm, res); + dlm_lockres_grab_inflight_ref(dlm, res); + spin_unlock(&res->spinlock); + spin_unlock(&dlm->spinlock); + /* lockres still marked IN_PROGRESS */ + goto wake_waiters; + } + + /* check master list to see if another node has started mastering it */ + spin_lock(&dlm->master_lock); + + /* if we found a block, wait for lock to be mastered by another node */ + blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen); + if (blocked) { + int mig; + if (mle->type == DLM_MLE_MASTER) { + mlog(ML_ERROR, "master entry for nonexistent lock!\n"); + BUG(); + } + mig = (mle->type == DLM_MLE_MIGRATION); + /* if there is a migration in progress, let the migration + * finish before continuing. we can wait for the absence + * of the MIGRATION mle: either the migrate finished or + * one of the nodes died and the mle was cleaned up. + * if there is a BLOCK here, but it already has a master + * set, we are too late. the master does not have a ref + * for us in the refmap. detach the mle and drop it. + * either way, go back to the top and start over. */ + if (mig || mle->master != O2NM_MAX_NODES) { + BUG_ON(mig && mle->master == dlm->node_num); + /* we arrived too late. the master does not + * have a ref for us. retry. */ + mlog(0, "%s:%.*s: late on %s\n", + dlm->name, namelen, lockid, + mig ? "MIGRATION" : "BLOCK"); + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); + + /* master is known, detach */ + if (!mig) + dlm_mle_detach_hb_events(dlm, mle); + dlm_put_mle(mle); + mle = NULL; + /* this is lame, but we can't wait on either + * the mle or lockres waitqueue here */ + if (mig) + msleep(100); + goto lookup; + } + } else { + /* go ahead and try to master lock on this node */ + mle = alloc_mle; + /* make sure this does not get freed below */ + alloc_mle = NULL; + dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0); + set_bit(dlm->node_num, mle->maybe_map); + __dlm_insert_mle(dlm, mle); + + /* still holding the dlm spinlock, check the recovery map + * to see if there are any nodes that still need to be + * considered. these will not appear in the mle nodemap + * but they might own this lockres. wait on them. */ + bit = find_first_bit(dlm->recovery_map, O2NM_MAX_NODES); + if (bit < O2NM_MAX_NODES) { + mlog(0, "%s: res %.*s, At least one node (%d) " + "to recover before lock mastery can begin\n", + dlm->name, namelen, (char *)lockid, bit); + wait_on_recovery = 1; + } + } + + /* at this point there is either a DLM_MLE_BLOCK or a + * DLM_MLE_MASTER on the master list, so it's safe to add the + * lockres to the hashtable. anyone who finds the lock will + * still have to wait on the IN_PROGRESS. */ + + /* finally add the lockres to its hash bucket */ + __dlm_insert_lockres(dlm, res); + + /* since this lockres is new it doesn't not require the spinlock */ + __dlm_lockres_grab_inflight_ref(dlm, res); + + /* get an extra ref on the mle in case this is a BLOCK + * if so, the creator of the BLOCK may try to put the last + * ref at this time in the assert master handler, so we + * need an extra one to keep from a bad ptr deref. */ + dlm_get_mle_inuse(mle); + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); + +redo_request: + while (wait_on_recovery) { + /* any cluster changes that occurred after dropping the + * dlm spinlock would be detectable be a change on the mle, + * so we only need to clear out the recovery map once. */ + if (dlm_is_recovery_lock(lockid, namelen)) { + mlog(0, "%s: Recovery map is not empty, but must " + "master $RECOVERY lock now\n", dlm->name); + if (!dlm_pre_master_reco_lockres(dlm, res)) + wait_on_recovery = 0; + else { + mlog(0, "%s: waiting 500ms for heartbeat state " + "change\n", dlm->name); + msleep(500); + } + continue; + } + + dlm_kick_recovery_thread(dlm); + msleep(1000); + dlm_wait_for_recovery(dlm); + + spin_lock(&dlm->spinlock); + bit = find_first_bit(dlm->recovery_map, O2NM_MAX_NODES); + if (bit < O2NM_MAX_NODES) { + mlog(0, "%s: res %.*s, At least one node (%d) " + "to recover before lock mastery can begin\n", + dlm->name, namelen, (char *)lockid, bit); + wait_on_recovery = 1; + } else + wait_on_recovery = 0; + spin_unlock(&dlm->spinlock); + + if (wait_on_recovery) + dlm_wait_for_node_recovery(dlm, bit, 10000); + } + + /* must wait for lock to be mastered elsewhere */ + if (blocked) + goto wait; + + dlm_node_iter_init(mle->vote_map, &iter); + while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { + ret = dlm_do_master_request(res, mle, nodenum); + if (ret < 0) + mlog_errno(ret); + if (mle->master != O2NM_MAX_NODES) { + /* found a master ! */ + if (mle->master <= nodenum) + break; + /* if our master request has not reached the master + * yet, keep going until it does. this is how the + * master will know that asserts are needed back to + * the lower nodes. */ + mlog(0, "%s: res %.*s, Requests only up to %u but " + "master is %u, keep going\n", dlm->name, namelen, + lockid, nodenum, mle->master); + } + } + +wait: + /* keep going until the response map includes all nodes */ + ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked); + if (ret < 0) { + wait_on_recovery = 1; + mlog(0, "%s: res %.*s, Node map changed, redo the master " + "request now, blocked=%d\n", dlm->name, res->lockname.len, + res->lockname.name, blocked); + if (++tries > 20) { + mlog(ML_ERROR, "%s: res %.*s, Spinning on " + "dlm_wait_for_lock_mastery, blocked = %d\n", + dlm->name, res->lockname.len, + res->lockname.name, blocked); + dlm_print_one_lock_resource(res); + dlm_print_one_mle(mle); + tries = 0; + } + goto redo_request; + } + + mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len, + res->lockname.name, res->owner); + /* make sure we never continue without this */ + BUG_ON(res->owner == O2NM_MAX_NODES); + + /* master is known, detach if not already detached */ + dlm_mle_detach_hb_events(dlm, mle); + dlm_put_mle(mle); + /* put the extra ref */ + dlm_put_mle_inuse(mle); + +wake_waiters: + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_IN_PROGRESS; + spin_unlock(&res->spinlock); + wake_up(&res->wq); + +leave: + /* need to free the unused mle */ + if (alloc_mle) + kmem_cache_free(dlm_mle_cache, alloc_mle); + + return res; +} + + +#define DLM_MASTERY_TIMEOUT_MS 5000 + +static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_master_list_entry *mle, + int *blocked) +{ + u8 m; + int ret, bit; + int map_changed, voting_done; + int assert, sleep; + +recheck: + ret = 0; + assert = 0; + + /* check if another node has already become the owner */ + spin_lock(&res->spinlock); + if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) { + mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name, + res->lockname.len, res->lockname.name, res->owner); + spin_unlock(&res->spinlock); + /* this will cause the master to re-assert across + * the whole cluster, freeing up mles */ + if (res->owner != dlm->node_num) { + ret = dlm_do_master_request(res, mle, res->owner); + if (ret < 0) { + /* give recovery a chance to run */ + mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret); + msleep(500); + goto recheck; + } + } + ret = 0; + goto leave; + } + spin_unlock(&res->spinlock); + + spin_lock(&mle->spinlock); + m = mle->master; + map_changed = !bitmap_equal(mle->vote_map, mle->node_map, + O2NM_MAX_NODES); + voting_done = bitmap_equal(mle->vote_map, mle->response_map, + O2NM_MAX_NODES); + + /* restart if we hit any errors */ + if (map_changed) { + int b; + mlog(0, "%s: %.*s: node map changed, restarting\n", + dlm->name, res->lockname.len, res->lockname.name); + ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked); + b = (mle->type == DLM_MLE_BLOCK); + if ((*blocked && !b) || (!*blocked && b)) { + mlog(0, "%s:%.*s: status change: old=%d new=%d\n", + dlm->name, res->lockname.len, res->lockname.name, + *blocked, b); + *blocked = b; + } + spin_unlock(&mle->spinlock); + if (ret < 0) { + mlog_errno(ret); + goto leave; + } + mlog(0, "%s:%.*s: restart lock mastery succeeded, " + "rechecking now\n", dlm->name, res->lockname.len, + res->lockname.name); + goto recheck; + } else { + if (!voting_done) { + mlog(0, "map not changed and voting not done " + "for %s:%.*s\n", dlm->name, res->lockname.len, + res->lockname.name); + } + } + + if (m != O2NM_MAX_NODES) { + /* another node has done an assert! + * all done! */ + sleep = 0; + } else { + sleep = 1; + /* have all nodes responded? */ + if (voting_done && !*blocked) { + bit = find_first_bit(mle->maybe_map, O2NM_MAX_NODES); + if (dlm->node_num <= bit) { + /* my node number is lowest. + * now tell other nodes that I am + * mastering this. */ + mle->master = dlm->node_num; + /* ref was grabbed in get_lock_resource + * will be dropped in dlmlock_master */ + assert = 1; + sleep = 0; + } + /* if voting is done, but we have not received + * an assert master yet, we must sleep */ + } + } + + spin_unlock(&mle->spinlock); + + /* sleep if we haven't finished voting yet */ + if (sleep) { + unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS); + atomic_set(&mle->woken, 0); + (void)wait_event_timeout(mle->wq, + (atomic_read(&mle->woken) == 1), + timeo); + if (res->owner == O2NM_MAX_NODES) { + mlog(0, "%s:%.*s: waiting again\n", dlm->name, + res->lockname.len, res->lockname.name); + goto recheck; + } + mlog(0, "done waiting, master is %u\n", res->owner); + ret = 0; + goto leave; + } + + ret = 0; /* done */ + if (assert) { + m = dlm->node_num; + mlog(0, "about to master %.*s here, this=%u\n", + res->lockname.len, res->lockname.name, m); + ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0); + if (ret) { + /* This is a failure in the network path, + * not in the response to the assert_master + * (any nonzero response is a BUG on this node). + * Most likely a socket just got disconnected + * due to node death. */ + mlog_errno(ret); + } + /* no longer need to restart lock mastery. + * all living nodes have been contacted. */ + ret = 0; + } + + /* set the lockres owner */ + spin_lock(&res->spinlock); + /* mastery reference obtained either during + * assert_master_handler or in get_lock_resource */ + dlm_change_lockres_owner(dlm, res, m); + spin_unlock(&res->spinlock); + +leave: + return ret; +} + +struct dlm_bitmap_diff_iter +{ + int curnode; + unsigned long *orig_bm; + unsigned long *cur_bm; + unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)]; +}; + +enum dlm_node_state_change +{ + NODE_DOWN = -1, + NODE_NO_CHANGE = 0, + NODE_UP +}; + +static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter, + unsigned long *orig_bm, + unsigned long *cur_bm) +{ + unsigned long p1, p2; + int i; + + iter->curnode = -1; + iter->orig_bm = orig_bm; + iter->cur_bm = cur_bm; + + for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) { + p1 = *(iter->orig_bm + i); + p2 = *(iter->cur_bm + i); + iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1); + } +} + +static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter, + enum dlm_node_state_change *state) +{ + int bit; + + if (iter->curnode >= O2NM_MAX_NODES) + return -ENOENT; + + bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES, + iter->curnode+1); + if (bit >= O2NM_MAX_NODES) { + iter->curnode = O2NM_MAX_NODES; + return -ENOENT; + } + + /* if it was there in the original then this node died */ + if (test_bit(bit, iter->orig_bm)) + *state = NODE_DOWN; + else + *state = NODE_UP; + + iter->curnode = bit; + return bit; +} + + +static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_master_list_entry *mle, + int blocked) +{ + struct dlm_bitmap_diff_iter bdi; + enum dlm_node_state_change sc; + int node; + int ret = 0; + + mlog(0, "something happened such that the " + "master process may need to be restarted!\n"); + + assert_spin_locked(&mle->spinlock); + + dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map); + node = dlm_bitmap_diff_iter_next(&bdi, &sc); + while (node >= 0) { + if (sc == NODE_UP) { + /* a node came up. clear any old vote from + * the response map and set it in the vote map + * then restart the mastery. */ + mlog(ML_NOTICE, "node %d up while restarting\n", node); + + /* redo the master request, but only for the new node */ + mlog(0, "sending request to new node\n"); + clear_bit(node, mle->response_map); + set_bit(node, mle->vote_map); + } else { + mlog(ML_ERROR, "node down! %d\n", node); + if (blocked) { + int lowest = find_first_bit(mle->maybe_map, + O2NM_MAX_NODES); + + /* act like it was never there */ + clear_bit(node, mle->maybe_map); + + if (node == lowest) { + mlog(0, "expected master %u died" + " while this node was blocked " + "waiting on it!\n", node); + lowest = find_next_bit(mle->maybe_map, + O2NM_MAX_NODES, + lowest+1); + if (lowest < O2NM_MAX_NODES) { + mlog(0, "%s:%.*s:still " + "blocked. waiting on %u " + "now\n", dlm->name, + res->lockname.len, + res->lockname.name, + lowest); + } else { + /* mle is an MLE_BLOCK, but + * there is now nothing left to + * block on. we need to return + * all the way back out and try + * again with an MLE_MASTER. + * dlm_do_local_recovery_cleanup + * has already run, so the mle + * refcount is ok */ + mlog(0, "%s:%.*s: no " + "longer blocking. try to " + "master this here\n", + dlm->name, + res->lockname.len, + res->lockname.name); + mle->type = DLM_MLE_MASTER; + mle->mleres = res; + } + } + } + + /* now blank out everything, as if we had never + * contacted anyone */ + bitmap_zero(mle->maybe_map, O2NM_MAX_NODES); + bitmap_zero(mle->response_map, O2NM_MAX_NODES); + /* reset the vote_map to the current node_map */ + bitmap_copy(mle->vote_map, mle->node_map, + O2NM_MAX_NODES); + /* put myself into the maybe map */ + if (mle->type != DLM_MLE_BLOCK) + set_bit(dlm->node_num, mle->maybe_map); + } + ret = -EAGAIN; + node = dlm_bitmap_diff_iter_next(&bdi, &sc); + } + return ret; +} + + +/* + * DLM_MASTER_REQUEST_MSG + * + * returns: 0 on success, + * -errno on a network error + * + * on error, the caller should assume the target node is "dead" + * + */ + +static int dlm_do_master_request(struct dlm_lock_resource *res, + struct dlm_master_list_entry *mle, int to) +{ + struct dlm_ctxt *dlm = mle->dlm; + struct dlm_master_request request; + int ret, response=0, resend; + + memset(&request, 0, sizeof(request)); + request.node_idx = dlm->node_num; + + BUG_ON(mle->type == DLM_MLE_MIGRATION); + + request.namelen = (u8)mle->mnamelen; + memcpy(request.name, mle->mname, request.namelen); + +again: + ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request, + sizeof(request), to, &response); + if (ret < 0) { + if (ret == -ESRCH) { + /* should never happen */ + mlog(ML_ERROR, "TCP stack not ready!\n"); + BUG(); + } else if (ret == -EINVAL) { + mlog(ML_ERROR, "bad args passed to o2net!\n"); + BUG(); + } else if (ret == -ENOMEM) { + mlog(ML_ERROR, "out of memory while trying to send " + "network message! retrying\n"); + /* this is totally crude */ + msleep(50); + goto again; + } else if (!dlm_is_host_down(ret)) { + /* not a network error. bad. */ + mlog_errno(ret); + mlog(ML_ERROR, "unhandled error!"); + BUG(); + } + /* all other errors should be network errors, + * and likely indicate node death */ + mlog(ML_ERROR, "link to %d went down!\n", to); + goto out; + } + + ret = 0; + resend = 0; + spin_lock(&mle->spinlock); + switch (response) { + case DLM_MASTER_RESP_YES: + set_bit(to, mle->response_map); + mlog(0, "node %u is the master, response=YES\n", to); + mlog(0, "%s:%.*s: master node %u now knows I have a " + "reference\n", dlm->name, res->lockname.len, + res->lockname.name, to); + mle->master = to; + break; + case DLM_MASTER_RESP_NO: + mlog(0, "node %u not master, response=NO\n", to); + set_bit(to, mle->response_map); + break; + case DLM_MASTER_RESP_MAYBE: + mlog(0, "node %u not master, response=MAYBE\n", to); + set_bit(to, mle->response_map); + set_bit(to, mle->maybe_map); + break; + case DLM_MASTER_RESP_ERROR: + mlog(0, "node %u hit an error, resending\n", to); + resend = 1; + response = 0; + break; + default: + mlog(ML_ERROR, "bad response! %u\n", response); + BUG(); + } + spin_unlock(&mle->spinlock); + if (resend) { + /* this is also totally crude */ + msleep(50); + goto again; + } + +out: + return ret; +} + +/* + * locks that can be taken here: + * dlm->spinlock + * res->spinlock + * mle->spinlock + * dlm->master_list + * + * if possible, TRIM THIS DOWN!!! + */ +int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + u8 response = DLM_MASTER_RESP_MAYBE; + struct dlm_ctxt *dlm = data; + struct dlm_lock_resource *res = NULL; + struct dlm_master_request *request = (struct dlm_master_request *) msg->buf; + struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL; + char *name; + unsigned int namelen, hash; + int found, ret; + int set_maybe; + int dispatch_assert = 0; + int dispatched = 0; + + if (!dlm_grab(dlm)) + return DLM_MASTER_RESP_NO; + + if (!dlm_domain_fully_joined(dlm)) { + response = DLM_MASTER_RESP_NO; + goto send_response; + } + + name = request->name; + namelen = request->namelen; + hash = dlm_lockid_hash(name, namelen); + + if (namelen > DLM_LOCKID_NAME_MAX) { + response = DLM_IVBUFLEN; + goto send_response; + } + +way_up_top: + spin_lock(&dlm->spinlock); + res = __dlm_lookup_lockres(dlm, name, namelen, hash); + if (res) { + spin_unlock(&dlm->spinlock); + + /* take care of the easy cases up front */ + spin_lock(&res->spinlock); + + /* + * Right after dlm spinlock was released, dlm_thread could have + * purged the lockres. Check if lockres got unhashed. If so + * start over. + */ + if (hlist_unhashed(&res->hash_node)) { + spin_unlock(&res->spinlock); + dlm_lockres_put(res); + goto way_up_top; + } + + if (res->state & (DLM_LOCK_RES_RECOVERING| + DLM_LOCK_RES_MIGRATING)) { + spin_unlock(&res->spinlock); + mlog(0, "returning DLM_MASTER_RESP_ERROR since res is " + "being recovered/migrated\n"); + response = DLM_MASTER_RESP_ERROR; + if (mle) + kmem_cache_free(dlm_mle_cache, mle); + goto send_response; + } + + if (res->owner == dlm->node_num) { + dlm_lockres_set_refmap_bit(dlm, res, request->node_idx); + spin_unlock(&res->spinlock); + response = DLM_MASTER_RESP_YES; + if (mle) + kmem_cache_free(dlm_mle_cache, mle); + + /* this node is the owner. + * there is some extra work that needs to + * happen now. the requesting node has + * caused all nodes up to this one to + * create mles. this node now needs to + * go back and clean those up. */ + dispatch_assert = 1; + goto send_response; + } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) { + spin_unlock(&res->spinlock); + response = DLM_MASTER_RESP_NO; + if (mle) + kmem_cache_free(dlm_mle_cache, mle); + goto send_response; + } + + /* ok, there is no owner. either this node is + * being blocked, or it is actively trying to + * master this lock. */ + if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) { + mlog(ML_ERROR, "lock with no owner should be " + "in-progress!\n"); + BUG(); + } + + spin_lock(&dlm->master_lock); + found = dlm_find_mle(dlm, &tmpmle, name, namelen); + if (!found) { + mlog(ML_ERROR, "no mle found for this lock!\n"); + BUG(); + } + set_maybe = 1; + spin_lock(&tmpmle->spinlock); + if (tmpmle->type == DLM_MLE_BLOCK) { + response = DLM_MASTER_RESP_NO; + } else if (tmpmle->type == DLM_MLE_MIGRATION) { + mlog(0, "node %u is master, but trying to migrate to " + "node %u.\n", tmpmle->master, tmpmle->new_master); + if (tmpmle->master == dlm->node_num) { + mlog(ML_ERROR, "no owner on lockres, but this " + "node is trying to migrate it to %u?!\n", + tmpmle->new_master); + BUG(); + } else { + /* the real master can respond on its own */ + response = DLM_MASTER_RESP_NO; + } + } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) { + set_maybe = 0; + if (tmpmle->master == dlm->node_num) { + response = DLM_MASTER_RESP_YES; + /* this node will be the owner. + * go back and clean the mles on any + * other nodes */ + dispatch_assert = 1; + dlm_lockres_set_refmap_bit(dlm, res, + request->node_idx); + } else + response = DLM_MASTER_RESP_NO; + } else { + response = DLM_MASTER_RESP_MAYBE; + } + if (set_maybe) + set_bit(request->node_idx, tmpmle->maybe_map); + spin_unlock(&tmpmle->spinlock); + + spin_unlock(&dlm->master_lock); + spin_unlock(&res->spinlock); + + /* keep the mle attached to heartbeat events */ + dlm_put_mle(tmpmle); + if (mle) + kmem_cache_free(dlm_mle_cache, mle); + goto send_response; + } + + /* + * lockres doesn't exist on this node + * if there is an MLE_BLOCK, return NO + * if there is an MLE_MASTER, return MAYBE + * otherwise, add an MLE_BLOCK, return NO + */ + spin_lock(&dlm->master_lock); + found = dlm_find_mle(dlm, &tmpmle, name, namelen); + if (!found) { + /* this lockid has never been seen on this node yet */ + if (!mle) { + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); + + mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); + if (!mle) { + response = DLM_MASTER_RESP_ERROR; + mlog_errno(-ENOMEM); + goto send_response; + } + goto way_up_top; + } + + dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen); + set_bit(request->node_idx, mle->maybe_map); + __dlm_insert_mle(dlm, mle); + response = DLM_MASTER_RESP_NO; + } else { + spin_lock(&tmpmle->spinlock); + if (tmpmle->master == dlm->node_num) { + mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n"); + BUG(); + } + if (tmpmle->type == DLM_MLE_BLOCK) + response = DLM_MASTER_RESP_NO; + else if (tmpmle->type == DLM_MLE_MIGRATION) { + mlog(0, "migration mle was found (%u->%u)\n", + tmpmle->master, tmpmle->new_master); + /* real master can respond on its own */ + response = DLM_MASTER_RESP_NO; + } else + response = DLM_MASTER_RESP_MAYBE; + set_bit(request->node_idx, tmpmle->maybe_map); + spin_unlock(&tmpmle->spinlock); + } + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); + + if (found) { + /* keep the mle attached to heartbeat events */ + dlm_put_mle(tmpmle); + } +send_response: + /* + * __dlm_lookup_lockres() grabbed a reference to this lockres. + * The reference is released by dlm_assert_master_worker() under + * the call to dlm_dispatch_assert_master(). If + * dlm_assert_master_worker() isn't called, we drop it here. + */ + if (dispatch_assert) { + mlog(0, "%u is the owner of %.*s, cleaning everyone else\n", + dlm->node_num, res->lockname.len, res->lockname.name); + spin_lock(&res->spinlock); + ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx, + DLM_ASSERT_MASTER_MLE_CLEANUP); + if (ret < 0) { + mlog(ML_ERROR, "failed to dispatch assert master work\n"); + response = DLM_MASTER_RESP_ERROR; + spin_unlock(&res->spinlock); + dlm_lockres_put(res); + } else { + dispatched = 1; + __dlm_lockres_grab_inflight_worker(dlm, res); + spin_unlock(&res->spinlock); + } + } else { + if (res) + dlm_lockres_put(res); + } + + if (!dispatched) + dlm_put(dlm); + return response; +} + +/* + * DLM_ASSERT_MASTER_MSG + */ + + +/* + * NOTE: this can be used for debugging + * can periodically run all locks owned by this node + * and re-assert across the cluster... + */ +static int dlm_do_assert_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + void *nodemap, u32 flags) +{ + struct dlm_assert_master assert; + int to, tmpret; + struct dlm_node_iter iter; + int ret = 0; + int reassert; + const char *lockname = res->lockname.name; + unsigned int namelen = res->lockname.len; + + BUG_ON(namelen > O2NM_MAX_NAME_LEN); + + spin_lock(&res->spinlock); + res->state |= DLM_LOCK_RES_SETREF_INPROG; + spin_unlock(&res->spinlock); + +again: + reassert = 0; + + /* note that if this nodemap is empty, it returns 0 */ + dlm_node_iter_init(nodemap, &iter); + while ((to = dlm_node_iter_next(&iter)) >= 0) { + int r = 0; + struct dlm_master_list_entry *mle = NULL; + + mlog(0, "sending assert master to %d (%.*s)\n", to, + namelen, lockname); + memset(&assert, 0, sizeof(assert)); + assert.node_idx = dlm->node_num; + assert.namelen = namelen; + memcpy(assert.name, lockname, namelen); + assert.flags = cpu_to_be32(flags); + + tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key, + &assert, sizeof(assert), to, &r); + if (tmpret < 0) { + mlog(ML_ERROR, "Error %d when sending message %u (key " + "0x%x) to node %u\n", tmpret, + DLM_ASSERT_MASTER_MSG, dlm->key, to); + if (!dlm_is_host_down(tmpret)) { + mlog(ML_ERROR, "unhandled error=%d!\n", tmpret); + BUG(); + } + /* a node died. finish out the rest of the nodes. */ + mlog(0, "link to %d went down!\n", to); + /* any nonzero status return will do */ + ret = tmpret; + r = 0; + } else if (r < 0) { + /* ok, something horribly messed. kill thyself. */ + mlog(ML_ERROR,"during assert master of %.*s to %u, " + "got %d.\n", namelen, lockname, to, r); + spin_lock(&dlm->spinlock); + spin_lock(&dlm->master_lock); + if (dlm_find_mle(dlm, &mle, (char *)lockname, + namelen)) { + dlm_print_one_mle(mle); + __dlm_put_mle(mle); + } + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); + BUG(); + } + + if (r & DLM_ASSERT_RESPONSE_REASSERT && + !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) { + mlog(ML_ERROR, "%.*s: very strange, " + "master MLE but no lockres on %u\n", + namelen, lockname, to); + } + + if (r & DLM_ASSERT_RESPONSE_REASSERT) { + mlog(0, "%.*s: node %u create mles on other " + "nodes and requests a re-assert\n", + namelen, lockname, to); + reassert = 1; + } + if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) { + mlog(0, "%.*s: node %u has a reference to this " + "lockres, set the bit in the refmap\n", + namelen, lockname, to); + spin_lock(&res->spinlock); + dlm_lockres_set_refmap_bit(dlm, res, to); + spin_unlock(&res->spinlock); + } + } + + if (reassert) + goto again; + + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_SETREF_INPROG; + spin_unlock(&res->spinlock); + wake_up(&res->wq); + + return ret; +} + +/* + * locks that can be taken here: + * dlm->spinlock + * res->spinlock + * mle->spinlock + * dlm->master_list + * + * if possible, TRIM THIS DOWN!!! + */ +int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_master_list_entry *mle = NULL; + struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf; + struct dlm_lock_resource *res = NULL; + char *name; + unsigned int namelen, hash; + u32 flags; + int master_request = 0, have_lockres_ref = 0; + int ret = 0; + + if (!dlm_grab(dlm)) + return 0; + + name = assert->name; + namelen = assert->namelen; + hash = dlm_lockid_hash(name, namelen); + flags = be32_to_cpu(assert->flags); + + if (namelen > DLM_LOCKID_NAME_MAX) { + mlog(ML_ERROR, "Invalid name length!"); + goto done; + } + + spin_lock(&dlm->spinlock); + + if (flags) + mlog(0, "assert_master with flags: %u\n", flags); + + /* find the MLE */ + spin_lock(&dlm->master_lock); + if (!dlm_find_mle(dlm, &mle, name, namelen)) { + /* not an error, could be master just re-asserting */ + mlog(0, "just got an assert_master from %u, but no " + "MLE for it! (%.*s)\n", assert->node_idx, + namelen, name); + } else { + int bit = find_first_bit(mle->maybe_map, O2NM_MAX_NODES); + if (bit >= O2NM_MAX_NODES) { + /* not necessarily an error, though less likely. + * could be master just re-asserting. */ + mlog(0, "no bits set in the maybe_map, but %u " + "is asserting! (%.*s)\n", assert->node_idx, + namelen, name); + } else if (bit != assert->node_idx) { + if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) { + mlog(0, "master %u was found, %u should " + "back off\n", assert->node_idx, bit); + } else { + /* with the fix for bug 569, a higher node + * number winning the mastery will respond + * YES to mastery requests, but this node + * had no way of knowing. let it pass. */ + mlog(0, "%u is the lowest node, " + "%u is asserting. (%.*s) %u must " + "have begun after %u won.\n", bit, + assert->node_idx, namelen, name, bit, + assert->node_idx); + } + } + if (mle->type == DLM_MLE_MIGRATION) { + if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) { + mlog(0, "%s:%.*s: got cleanup assert" + " from %u for migration\n", + dlm->name, namelen, name, + assert->node_idx); + } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) { + mlog(0, "%s:%.*s: got unrelated assert" + " from %u for migration, ignoring\n", + dlm->name, namelen, name, + assert->node_idx); + __dlm_put_mle(mle); + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); + goto done; + } + } + } + spin_unlock(&dlm->master_lock); + + /* ok everything checks out with the MLE + * now check to see if there is a lockres */ + res = __dlm_lookup_lockres(dlm, name, namelen, hash); + if (res) { + spin_lock(&res->spinlock); + if (res->state & DLM_LOCK_RES_RECOVERING) { + mlog(ML_ERROR, "%u asserting but %.*s is " + "RECOVERING!\n", assert->node_idx, namelen, name); + goto kill; + } + if (!mle) { + if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN && + res->owner != assert->node_idx) { + mlog(ML_ERROR, "DIE! Mastery assert from %u, " + "but current owner is %u! (%.*s)\n", + assert->node_idx, res->owner, namelen, + name); + __dlm_print_one_lock_resource(res); + BUG(); + } + } else if (mle->type != DLM_MLE_MIGRATION) { + if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) { + /* owner is just re-asserting */ + if (res->owner == assert->node_idx) { + mlog(0, "owner %u re-asserting on " + "lock %.*s\n", assert->node_idx, + namelen, name); + goto ok; + } + mlog(ML_ERROR, "got assert_master from " + "node %u, but %u is the owner! " + "(%.*s)\n", assert->node_idx, + res->owner, namelen, name); + goto kill; + } + if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) { + mlog(ML_ERROR, "got assert from %u, but lock " + "with no owner should be " + "in-progress! (%.*s)\n", + assert->node_idx, + namelen, name); + goto kill; + } + } else /* mle->type == DLM_MLE_MIGRATION */ { + /* should only be getting an assert from new master */ + if (assert->node_idx != mle->new_master) { + mlog(ML_ERROR, "got assert from %u, but " + "new master is %u, and old master " + "was %u (%.*s)\n", + assert->node_idx, mle->new_master, + mle->master, namelen, name); + goto kill; + } + + } +ok: + spin_unlock(&res->spinlock); + } + + if (mle) { + int extra_ref = 0; + int nn = -1; + int rr, err = 0; + + spin_lock(&mle->spinlock); + if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION) + extra_ref = 1; + else { + /* MASTER mle: if any bits set in the response map + * then the calling node needs to re-assert to clear + * up nodes that this node contacted */ + while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES, + nn+1)) < O2NM_MAX_NODES) { + if (nn != dlm->node_num && nn != assert->node_idx) { + master_request = 1; + break; + } + } + } + mle->master = assert->node_idx; + atomic_set(&mle->woken, 1); + wake_up(&mle->wq); + spin_unlock(&mle->spinlock); + + if (res) { + int wake = 0; + spin_lock(&res->spinlock); + if (mle->type == DLM_MLE_MIGRATION) { + mlog(0, "finishing off migration of lockres %.*s, " + "from %u to %u\n", + res->lockname.len, res->lockname.name, + dlm->node_num, mle->new_master); + res->state &= ~DLM_LOCK_RES_MIGRATING; + wake = 1; + dlm_change_lockres_owner(dlm, res, mle->new_master); + BUG_ON(res->state & DLM_LOCK_RES_DIRTY); + } else { + dlm_change_lockres_owner(dlm, res, mle->master); + } + spin_unlock(&res->spinlock); + have_lockres_ref = 1; + if (wake) + wake_up(&res->wq); + } + + /* master is known, detach if not already detached. + * ensures that only one assert_master call will happen + * on this mle. */ + spin_lock(&dlm->master_lock); + + rr = kref_read(&mle->mle_refs); + if (mle->inuse > 0) { + if (extra_ref && rr < 3) + err = 1; + else if (!extra_ref && rr < 2) + err = 1; + } else { + if (extra_ref && rr < 2) + err = 1; + else if (!extra_ref && rr < 1) + err = 1; + } + if (err) { + mlog(ML_ERROR, "%s:%.*s: got assert master from %u " + "that will mess up this node, refs=%d, extra=%d, " + "inuse=%d\n", dlm->name, namelen, name, + assert->node_idx, rr, extra_ref, mle->inuse); + dlm_print_one_mle(mle); + } + __dlm_unlink_mle(dlm, mle); + __dlm_mle_detach_hb_events(dlm, mle); + __dlm_put_mle(mle); + if (extra_ref) { + /* the assert master message now balances the extra + * ref given by the master / migration request message. + * if this is the last put, it will be removed + * from the list. */ + __dlm_put_mle(mle); + } + spin_unlock(&dlm->master_lock); + } else if (res) { + if (res->owner != assert->node_idx) { + mlog(0, "assert_master from %u, but current " + "owner is %u (%.*s), no mle\n", assert->node_idx, + res->owner, namelen, name); + } + } + spin_unlock(&dlm->spinlock); + +done: + ret = 0; + if (res) { + spin_lock(&res->spinlock); + res->state |= DLM_LOCK_RES_SETREF_INPROG; + spin_unlock(&res->spinlock); + *ret_data = (void *)res; + } + dlm_put(dlm); + if (master_request) { + mlog(0, "need to tell master to reassert\n"); + /* positive. negative would shoot down the node. */ + ret |= DLM_ASSERT_RESPONSE_REASSERT; + if (!have_lockres_ref) { + mlog(ML_ERROR, "strange, got assert from %u, MASTER " + "mle present here for %s:%.*s, but no lockres!\n", + assert->node_idx, dlm->name, namelen, name); + } + } + if (have_lockres_ref) { + /* let the master know we have a reference to the lockres */ + ret |= DLM_ASSERT_RESPONSE_MASTERY_REF; + mlog(0, "%s:%.*s: got assert from %u, need a ref\n", + dlm->name, namelen, name, assert->node_idx); + } + return ret; + +kill: + /* kill the caller! */ + mlog(ML_ERROR, "Bad message received from another node. Dumping state " + "and killing the other node now! This node is OK and can continue.\n"); + __dlm_print_one_lock_resource(res); + spin_unlock(&res->spinlock); + spin_lock(&dlm->master_lock); + if (mle) + __dlm_put_mle(mle); + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); + *ret_data = (void *)res; + dlm_put(dlm); + return -EINVAL; +} + +void dlm_assert_master_post_handler(int status, void *data, void *ret_data) +{ + struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data; + + if (ret_data) { + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_SETREF_INPROG; + spin_unlock(&res->spinlock); + wake_up(&res->wq); + dlm_lockres_put(res); + } + return; +} + +int dlm_dispatch_assert_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + int ignore_higher, u8 request_from, u32 flags) +{ + struct dlm_work_item *item; + item = kzalloc_obj(*item, GFP_ATOMIC); + if (!item) + return -ENOMEM; + + + /* queue up work for dlm_assert_master_worker */ + dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL); + item->u.am.lockres = res; /* already have a ref */ + /* can optionally ignore node numbers higher than this node */ + item->u.am.ignore_higher = ignore_higher; + item->u.am.request_from = request_from; + item->u.am.flags = flags; + + if (ignore_higher) + mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len, + res->lockname.name); + + spin_lock(&dlm->work_lock); + list_add_tail(&item->list, &dlm->work_list); + spin_unlock(&dlm->work_lock); + + queue_work(dlm->dlm_worker, &dlm->dispatched_work); + return 0; +} + +static void dlm_assert_master_worker(struct dlm_work_item *item, void *data) +{ + struct dlm_ctxt *dlm = data; + int ret = 0; + struct dlm_lock_resource *res; + unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)]; + int ignore_higher; + int bit; + u8 request_from; + u32 flags; + + dlm = item->dlm; + res = item->u.am.lockres; + ignore_higher = item->u.am.ignore_higher; + request_from = item->u.am.request_from; + flags = item->u.am.flags; + + spin_lock(&dlm->spinlock); + bitmap_copy(nodemap, dlm->domain_map, O2NM_MAX_NODES); + spin_unlock(&dlm->spinlock); + + clear_bit(dlm->node_num, nodemap); + if (ignore_higher) { + /* if is this just to clear up mles for nodes below + * this node, do not send the message to the original + * caller or any node number higher than this */ + clear_bit(request_from, nodemap); + bit = dlm->node_num; + while (1) { + bit = find_next_bit(nodemap, O2NM_MAX_NODES, + bit+1); + if (bit >= O2NM_MAX_NODES) + break; + clear_bit(bit, nodemap); + } + } + + /* + * If we're migrating this lock to someone else, we are no + * longer allowed to assert out own mastery. OTOH, we need to + * prevent migration from starting while we're still asserting + * our dominance. The reserved ast delays migration. + */ + spin_lock(&res->spinlock); + if (res->state & DLM_LOCK_RES_MIGRATING) { + mlog(0, "Someone asked us to assert mastery, but we're " + "in the middle of migration. Skipping assert, " + "the new master will handle that.\n"); + spin_unlock(&res->spinlock); + goto put; + } else + __dlm_lockres_reserve_ast(res); + spin_unlock(&res->spinlock); + + /* this call now finishes out the nodemap + * even if one or more nodes die */ + mlog(0, "worker about to master %.*s here, this=%u\n", + res->lockname.len, res->lockname.name, dlm->node_num); + ret = dlm_do_assert_master(dlm, res, nodemap, flags); + if (ret < 0) { + /* no need to restart, we are done */ + if (!dlm_is_host_down(ret)) + mlog_errno(ret); + } + + /* Ok, we've asserted ourselves. Let's let migration start. */ + dlm_lockres_release_ast(dlm, res); + +put: + dlm_lockres_drop_inflight_worker(dlm, res); + + dlm_lockres_put(res); + + mlog(0, "finished with dlm_assert_master_worker\n"); +} + +/* SPECIAL CASE for the $RECOVERY lock used by the recovery thread. + * We cannot wait for node recovery to complete to begin mastering this + * lockres because this lockres is used to kick off recovery! ;-) + * So, do a pre-check on all living nodes to see if any of those nodes + * think that $RECOVERY is currently mastered by a dead node. If so, + * we wait a short time to allow that node to get notified by its own + * heartbeat stack, then check again. All $RECOVERY lock resources + * mastered by dead nodes are purged when the heartbeat callback is + * fired, so we can know for sure that it is safe to continue once + * the node returns a live node or no node. */ +static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + struct dlm_node_iter iter; + int nodenum; + int ret = 0; + u8 master = DLM_LOCK_RES_OWNER_UNKNOWN; + + spin_lock(&dlm->spinlock); + dlm_node_iter_init(dlm->domain_map, &iter); + spin_unlock(&dlm->spinlock); + + while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { + /* do not send to self */ + if (nodenum == dlm->node_num) + continue; + ret = dlm_do_master_requery(dlm, res, nodenum, &master); + if (ret < 0) { + mlog_errno(ret); + if (!dlm_is_host_down(ret)) + BUG(); + /* host is down, so answer for that node would be + * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */ + ret = 0; + } + + if (master != DLM_LOCK_RES_OWNER_UNKNOWN) { + /* check to see if this master is in the recovery map */ + spin_lock(&dlm->spinlock); + if (test_bit(master, dlm->recovery_map)) { + mlog(ML_NOTICE, "%s: node %u has not seen " + "node %u go down yet, and thinks the " + "dead node is mastering the recovery " + "lock. must wait.\n", dlm->name, + nodenum, master); + ret = -EAGAIN; + } + spin_unlock(&dlm->spinlock); + mlog(0, "%s: reco lock master is %u\n", dlm->name, + master); + break; + } + } + return ret; +} + +/* + * DLM_DEREF_LOCKRES_MSG + */ + +int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) +{ + struct dlm_deref_lockres deref; + int ret = 0, r; + const char *lockname; + unsigned int namelen; + + lockname = res->lockname.name; + namelen = res->lockname.len; + BUG_ON(namelen > O2NM_MAX_NAME_LEN); + + memset(&deref, 0, sizeof(deref)); + deref.node_idx = dlm->node_num; + deref.namelen = namelen; + memcpy(deref.name, lockname, namelen); + + ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key, + &deref, sizeof(deref), res->owner, &r); + if (ret < 0) + mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n", + dlm->name, namelen, lockname, ret, res->owner); + else if (r < 0) { + /* BAD. other node says I did not have a ref. */ + mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n", + dlm->name, namelen, lockname, res->owner, r); + dlm_print_one_lock_resource(res); + if (r == -ENOMEM) + BUG(); + } else + ret = r; + + return ret; +} + +int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf; + struct dlm_lock_resource *res = NULL; + char *name; + unsigned int namelen; + int ret = -EINVAL; + u8 node; + unsigned int hash; + struct dlm_work_item *item; + int cleared = 0; + int dispatch = 0; + + if (!dlm_grab(dlm)) + return 0; + + name = deref->name; + namelen = deref->namelen; + node = deref->node_idx; + + if (namelen > DLM_LOCKID_NAME_MAX) { + mlog(ML_ERROR, "Invalid name length!"); + goto done; + } + if (deref->node_idx >= O2NM_MAX_NODES) { + mlog(ML_ERROR, "Invalid node number: %u\n", node); + goto done; + } + + hash = dlm_lockid_hash(name, namelen); + + spin_lock(&dlm->spinlock); + res = __dlm_lookup_lockres_full(dlm, name, namelen, hash); + if (!res) { + spin_unlock(&dlm->spinlock); + mlog(ML_ERROR, "%s:%.*s: bad lockres name\n", + dlm->name, namelen, name); + goto done; + } + spin_unlock(&dlm->spinlock); + + spin_lock(&res->spinlock); + if (res->state & DLM_LOCK_RES_SETREF_INPROG) + dispatch = 1; + else { + BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF); + if (test_bit(node, res->refmap)) { + dlm_lockres_clear_refmap_bit(dlm, res, node); + cleared = 1; + } + } + spin_unlock(&res->spinlock); + + if (!dispatch) { + if (cleared) + dlm_lockres_calc_usage(dlm, res); + else { + mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref " + "but it is already dropped!\n", dlm->name, + res->lockname.len, res->lockname.name, node); + dlm_print_one_lock_resource(res); + } + ret = DLM_DEREF_RESPONSE_DONE; + goto done; + } + + item = kzalloc_obj(*item, GFP_NOFS); + if (!item) { + ret = -ENOMEM; + mlog_errno(ret); + goto done; + } + + dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL); + item->u.dl.deref_res = res; + item->u.dl.deref_node = node; + + spin_lock(&dlm->work_lock); + list_add_tail(&item->list, &dlm->work_list); + spin_unlock(&dlm->work_lock); + + queue_work(dlm->dlm_worker, &dlm->dispatched_work); + return DLM_DEREF_RESPONSE_INPROG; + +done: + if (res) + dlm_lockres_put(res); + dlm_put(dlm); + + return ret; +} + +int dlm_deref_lockres_done_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_deref_lockres_done *deref + = (struct dlm_deref_lockres_done *)msg->buf; + struct dlm_lock_resource *res = NULL; + char *name; + unsigned int namelen; + int ret = -EINVAL; + u8 node; + unsigned int hash; + + if (!dlm_grab(dlm)) + return 0; + + name = deref->name; + namelen = deref->namelen; + node = deref->node_idx; + + if (namelen > DLM_LOCKID_NAME_MAX) { + mlog(ML_ERROR, "Invalid name length!"); + goto done; + } + if (deref->node_idx >= O2NM_MAX_NODES) { + mlog(ML_ERROR, "Invalid node number: %u\n", node); + goto done; + } + + hash = dlm_lockid_hash(name, namelen); + + spin_lock(&dlm->spinlock); + res = __dlm_lookup_lockres_full(dlm, name, namelen, hash); + if (!res) { + spin_unlock(&dlm->spinlock); + mlog(ML_ERROR, "%s:%.*s: bad lockres name\n", + dlm->name, namelen, name); + goto done; + } + + spin_lock(&res->spinlock); + if (!(res->state & DLM_LOCK_RES_DROPPING_REF)) { + spin_unlock(&res->spinlock); + spin_unlock(&dlm->spinlock); + mlog(ML_NOTICE, "%s:%.*s: node %u sends deref done " + "but it is already derefed!\n", dlm->name, + res->lockname.len, res->lockname.name, node); + ret = 0; + goto done; + } + + __dlm_do_purge_lockres(dlm, res); + spin_unlock(&res->spinlock); + wake_up(&res->wq); + + spin_unlock(&dlm->spinlock); + + ret = 0; +done: + if (res) + dlm_lockres_put(res); + dlm_put(dlm); + return ret; +} + +static void dlm_drop_lockres_ref_done(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, u8 node) +{ + struct dlm_deref_lockres_done deref; + int ret = 0, r; + const char *lockname; + unsigned int namelen; + + lockname = res->lockname.name; + namelen = res->lockname.len; + BUG_ON(namelen > O2NM_MAX_NAME_LEN); + + memset(&deref, 0, sizeof(deref)); + deref.node_idx = dlm->node_num; + deref.namelen = namelen; + memcpy(deref.name, lockname, namelen); + + ret = o2net_send_message(DLM_DEREF_LOCKRES_DONE, dlm->key, + &deref, sizeof(deref), node, &r); + if (ret < 0) { + mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF DONE " + " to node %u\n", dlm->name, namelen, + lockname, ret, node); + } else if (r < 0) { + /* ignore the error */ + mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n", + dlm->name, namelen, lockname, node, r); + dlm_print_one_lock_resource(res); + } +} + +static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data) +{ + struct dlm_ctxt *dlm; + struct dlm_lock_resource *res; + u8 node; + u8 cleared = 0; + + dlm = item->dlm; + res = item->u.dl.deref_res; + node = item->u.dl.deref_node; + + spin_lock(&res->spinlock); + BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF); + __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG); + if (test_bit(node, res->refmap)) { + dlm_lockres_clear_refmap_bit(dlm, res, node); + cleared = 1; + } + spin_unlock(&res->spinlock); + + dlm_drop_lockres_ref_done(dlm, res, node); + + if (cleared) { + mlog(0, "%s:%.*s node %u ref dropped in dispatch\n", + dlm->name, res->lockname.len, res->lockname.name, node); + dlm_lockres_calc_usage(dlm, res); + } else { + mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref " + "but it is already dropped!\n", dlm->name, + res->lockname.len, res->lockname.name, node); + dlm_print_one_lock_resource(res); + } + + dlm_lockres_put(res); +} + +/* + * A migratable resource is one that is : + * 1. locally mastered, and, + * 2. zero local locks, and, + * 3. one or more non-local locks, or, one or more references + * Returns 1 if yes, 0 if not. + */ +static int dlm_is_lockres_migratable(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + enum dlm_lockres_list idx; + int nonlocal = 0, node_ref; + struct list_head *queue; + struct dlm_lock *lock; + u64 cookie; + + assert_spin_locked(&res->spinlock); + + /* delay migration when the lockres is in MIGRATING state */ + if (res->state & DLM_LOCK_RES_MIGRATING) + return 0; + + /* delay migration when the lockres is in RECOCERING state */ + if (res->state & (DLM_LOCK_RES_RECOVERING| + DLM_LOCK_RES_RECOVERY_WAITING)) + return 0; + + if (res->owner != dlm->node_num) + return 0; + + for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) { + queue = dlm_list_idx_to_ptr(res, idx); + list_for_each_entry(lock, queue, list) { + if (lock->ml.node != dlm->node_num) { + nonlocal++; + continue; + } + cookie = be64_to_cpu(lock->ml.cookie); + mlog(0, "%s: Not migratable res %.*s, lock %u:%llu on " + "%s list\n", dlm->name, res->lockname.len, + res->lockname.name, + dlm_get_lock_cookie_node(cookie), + dlm_get_lock_cookie_seq(cookie), + dlm_list_in_text(idx)); + return 0; + } + } + + if (!nonlocal) { + node_ref = find_first_bit(res->refmap, O2NM_MAX_NODES); + if (node_ref >= O2NM_MAX_NODES) + return 0; + } + + mlog(0, "%s: res %.*s, Migratable\n", dlm->name, res->lockname.len, + res->lockname.name); + + return 1; +} + +/* + * DLM_MIGRATE_LOCKRES + */ + + +static int dlm_migrate_lockres(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, u8 target) +{ + struct dlm_master_list_entry *mle = NULL; + struct dlm_master_list_entry *oldmle = NULL; + struct dlm_migratable_lockres *mres = NULL; + int ret = 0; + const char *name; + unsigned int namelen; + int mle_added = 0; + int wake = 0; + + if (!dlm_grab(dlm)) + return -EINVAL; + + name = res->lockname.name; + namelen = res->lockname.len; + + mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name, + target); + + /* preallocate up front. if this fails, abort */ + ret = -ENOMEM; + mres = kmalloc(PAGE_SIZE, GFP_NOFS); + if (!mres) { + mlog_errno(ret); + goto leave; + } + + mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); + if (!mle) { + mlog_errno(ret); + goto leave; + } + ret = 0; + + /* + * clear any existing master requests and + * add the migration mle to the list + */ + spin_lock(&dlm->spinlock); + spin_lock(&dlm->master_lock); + ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name, + namelen, target, dlm->node_num); + /* get an extra reference on the mle. + * otherwise the assert_master from the new + * master will destroy this. + */ + if (ret != -EEXIST) + dlm_get_mle_inuse(mle); + + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); + + if (ret == -EEXIST) { + mlog(0, "another process is already migrating it\n"); + goto fail; + } + mle_added = 1; + + /* + * set the MIGRATING flag and flush asts + * if we fail after this we need to re-dirty the lockres + */ + if (dlm_mark_lockres_migrating(dlm, res, target) < 0) { + mlog(ML_ERROR, "tried to migrate %.*s to %u, but " + "the target went down.\n", res->lockname.len, + res->lockname.name, target); + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_MIGRATING; + wake = 1; + spin_unlock(&res->spinlock); + ret = -EINVAL; + } + +fail: + if (ret != -EEXIST && oldmle) { + /* master is known, detach if not already detached */ + dlm_mle_detach_hb_events(dlm, oldmle); + dlm_put_mle(oldmle); + } + + if (ret < 0) { + if (mle_added) { + dlm_mle_detach_hb_events(dlm, mle); + dlm_put_mle(mle); + dlm_put_mle_inuse(mle); + } else if (mle) { + kmem_cache_free(dlm_mle_cache, mle); + mle = NULL; + } + goto leave; + } + + /* + * at this point, we have a migration target, an mle + * in the master list, and the MIGRATING flag set on + * the lockres + */ + + /* now that remote nodes are spinning on the MIGRATING flag, + * ensure that all assert_master work is flushed. */ + flush_workqueue(dlm->dlm_worker); + + /* notify new node and send all lock state */ + /* call send_one_lockres with migration flag. + * this serves as notice to the target node that a + * migration is starting. */ + ret = dlm_send_one_lockres(dlm, res, mres, target, + DLM_MRES_MIGRATION); + + if (ret < 0) { + mlog(0, "migration to node %u failed with %d\n", + target, ret); + /* migration failed, detach and clean up mle */ + dlm_mle_detach_hb_events(dlm, mle); + dlm_put_mle(mle); + dlm_put_mle_inuse(mle); + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_MIGRATING; + wake = 1; + spin_unlock(&res->spinlock); + if (dlm_is_host_down(ret)) + dlm_wait_for_node_death(dlm, target, + DLM_NODE_DEATH_WAIT_MAX); + goto leave; + } + + /* at this point, the target sends a message to all nodes, + * (using dlm_do_migrate_request). this node is skipped since + * we had to put an mle in the list to begin the process. this + * node now waits for target to do an assert master. this node + * will be the last one notified, ensuring that the migration + * is complete everywhere. if the target dies while this is + * going on, some nodes could potentially see the target as the + * master, so it is important that my recovery finds the migration + * mle and sets the master to UNKNOWN. */ + + + /* wait for new node to assert master */ + while (1) { + ret = wait_event_interruptible_timeout(mle->wq, + (atomic_read(&mle->woken) == 1), + msecs_to_jiffies(5000)); + + if (ret >= 0) { + if (atomic_read(&mle->woken) == 1 || + res->owner == target) + break; + + mlog(0, "%s:%.*s: timed out during migration\n", + dlm->name, res->lockname.len, res->lockname.name); + /* avoid hang during shutdown when migrating lockres + * to a node which also goes down */ + if (dlm_is_node_dead(dlm, target)) { + mlog(0, "%s:%.*s: expected migration " + "target %u is no longer up, restarting\n", + dlm->name, res->lockname.len, + res->lockname.name, target); + ret = -EINVAL; + /* migration failed, detach and clean up mle */ + dlm_mle_detach_hb_events(dlm, mle); + dlm_put_mle(mle); + dlm_put_mle_inuse(mle); + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_MIGRATING; + wake = 1; + spin_unlock(&res->spinlock); + goto leave; + } + } else + mlog(0, "%s:%.*s: caught signal during migration\n", + dlm->name, res->lockname.len, res->lockname.name); + } + + /* all done, set the owner, clear the flag */ + spin_lock(&res->spinlock); + dlm_set_lockres_owner(dlm, res, target); + res->state &= ~DLM_LOCK_RES_MIGRATING; + dlm_remove_nonlocal_locks(dlm, res); + spin_unlock(&res->spinlock); + wake_up(&res->wq); + + /* master is known, detach if not already detached */ + dlm_mle_detach_hb_events(dlm, mle); + dlm_put_mle_inuse(mle); + ret = 0; + + dlm_lockres_calc_usage(dlm, res); + +leave: + /* re-dirty the lockres if we failed */ + if (ret < 0) + dlm_kick_thread(dlm, res); + + /* wake up waiters if the MIGRATING flag got set + * but migration failed */ + if (wake) + wake_up(&res->wq); + + kfree(mres); + + dlm_put(dlm); + + mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen, + name, target, ret); + return ret; +} + +/* + * Should be called only after beginning the domain leave process. + * There should not be any remaining locks on nonlocal lock resources, + * and there should be no local locks left on locally mastered resources. + * + * Called with the dlm spinlock held, may drop it to do migration, but + * will re-acquire before exit. + * + * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped + */ +int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) + __must_hold(&dlm->spinlock) +{ + int ret; + int lock_dropped = 0; + u8 target = O2NM_MAX_NODES; + + assert_spin_locked(&dlm->spinlock); + + spin_lock(&res->spinlock); + if (dlm_is_lockres_migratable(dlm, res)) + target = dlm_pick_migration_target(dlm, res); + spin_unlock(&res->spinlock); + + if (target == O2NM_MAX_NODES) + goto leave; + + /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */ + spin_unlock(&dlm->spinlock); + lock_dropped = 1; + ret = dlm_migrate_lockres(dlm, res, target); + if (ret) + mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n", + dlm->name, res->lockname.len, res->lockname.name, + target, ret); + spin_lock(&dlm->spinlock); +leave: + return lock_dropped; +} + +int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock) +{ + int ret; + spin_lock(&dlm->ast_lock); + spin_lock(&lock->spinlock); + ret = (list_empty(&lock->bast_list) && !lock->bast_pending); + spin_unlock(&lock->spinlock); + spin_unlock(&dlm->ast_lock); + return ret; +} + +static int dlm_migration_can_proceed(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + u8 mig_target) +{ + int can_proceed; + spin_lock(&res->spinlock); + can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING); + spin_unlock(&res->spinlock); + + /* target has died, so make the caller break out of the + * wait_event, but caller must recheck the domain_map */ + spin_lock(&dlm->spinlock); + if (!test_bit(mig_target, dlm->domain_map)) + can_proceed = 1; + spin_unlock(&dlm->spinlock); + return can_proceed; +} + +static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + int ret; + spin_lock(&res->spinlock); + ret = !!(res->state & DLM_LOCK_RES_DIRTY); + spin_unlock(&res->spinlock); + return ret; +} + + +static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + u8 target) +{ + int ret = 0; + + mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n", + res->lockname.len, res->lockname.name, dlm->node_num, + target); + /* need to set MIGRATING flag on lockres. this is done by + * ensuring that all asts have been flushed for this lockres. */ + spin_lock(&res->spinlock); + BUG_ON(res->migration_pending); + res->migration_pending = 1; + /* strategy is to reserve an extra ast then release + * it below, letting the release do all of the work */ + __dlm_lockres_reserve_ast(res); + spin_unlock(&res->spinlock); + + /* now flush all the pending asts */ + dlm_kick_thread(dlm, res); + /* before waiting on DIRTY, block processes which may + * try to dirty the lockres before MIGRATING is set */ + spin_lock(&res->spinlock); + BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY); + res->state |= DLM_LOCK_RES_BLOCK_DIRTY; + spin_unlock(&res->spinlock); + /* now wait on any pending asts and the DIRTY state */ + wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res)); + dlm_lockres_release_ast(dlm, res); + + mlog(0, "about to wait on migration_wq, dirty=%s\n", + str_yes_no(res->state & DLM_LOCK_RES_DIRTY)); + /* if the extra ref we just put was the final one, this + * will pass thru immediately. otherwise, we need to wait + * for the last ast to finish. */ +again: + ret = wait_event_interruptible_timeout(dlm->migration_wq, + dlm_migration_can_proceed(dlm, res, target), + msecs_to_jiffies(1000)); + if (ret < 0) { + mlog(0, "woken again: migrating? %s, dead? %s\n", + str_yes_no(res->state & DLM_LOCK_RES_MIGRATING), + str_no_yes(test_bit(target, dlm->domain_map))); + } else { + mlog(0, "all is well: migrating? %s, dead? %s\n", + str_yes_no(res->state & DLM_LOCK_RES_MIGRATING), + str_no_yes(test_bit(target, dlm->domain_map))); + } + if (!dlm_migration_can_proceed(dlm, res, target)) { + mlog(0, "trying again...\n"); + goto again; + } + + ret = 0; + /* did the target go down or die? */ + spin_lock(&dlm->spinlock); + if (!test_bit(target, dlm->domain_map)) { + mlog(ML_ERROR, "aha. migration target %u just went down\n", + target); + ret = -EHOSTDOWN; + } + spin_unlock(&dlm->spinlock); + + /* + * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for + * another try; otherwise, we are sure the MIGRATING state is there, + * drop the unneeded state which blocked threads trying to DIRTY + */ + spin_lock(&res->spinlock); + BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY)); + res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY; + if (!ret) + BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING)); + else + res->migration_pending = 0; + spin_unlock(&res->spinlock); + + /* + * at this point: + * + * o the DLM_LOCK_RES_MIGRATING flag is set if target not down + * o there are no pending asts on this lockres + * o all processes trying to reserve an ast on this + * lockres must wait for the MIGRATING flag to clear + */ + return ret; +} + +/* last step in the migration process. + * original master calls this to free all of the dlm_lock + * structures that used to be for other nodes. */ +static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + struct list_head *queue = &res->granted; + int i, bit; + struct dlm_lock *lock, *next; + + assert_spin_locked(&res->spinlock); + + BUG_ON(res->owner == dlm->node_num); + + for (i=0; i<3; i++) { + list_for_each_entry_safe(lock, next, queue, list) { + if (lock->ml.node != dlm->node_num) { + mlog(0, "putting lock for node %u\n", + lock->ml.node); + /* be extra careful */ + BUG_ON(!list_empty(&lock->ast_list)); + BUG_ON(!list_empty(&lock->bast_list)); + BUG_ON(lock->ast_pending); + BUG_ON(lock->bast_pending); + dlm_lockres_clear_refmap_bit(dlm, res, + lock->ml.node); + list_del_init(&lock->list); + dlm_lock_put(lock); + /* In a normal unlock, we would have added a + * DLM_UNLOCK_FREE_LOCK action. Force it. */ + dlm_lock_put(lock); + } + } + queue++; + } + bit = 0; + while (1) { + bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit); + if (bit >= O2NM_MAX_NODES) + break; + /* do not clear the local node reference, if there is a + * process holding this, let it drop the ref itself */ + if (bit != dlm->node_num) { + mlog(0, "%s:%.*s: node %u had a ref to this " + "migrating lockres, clearing\n", dlm->name, + res->lockname.len, res->lockname.name, bit); + dlm_lockres_clear_refmap_bit(dlm, res, bit); + } + bit++; + } +} + +/* + * Pick a node to migrate the lock resource to. This function selects a + * potential target based first on the locks and then on refmap. It skips + * nodes that are in the process of exiting the domain. + */ +static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + enum dlm_lockres_list idx; + struct list_head *queue; + struct dlm_lock *lock; + int noderef; + u8 nodenum = O2NM_MAX_NODES; + + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&res->spinlock); + + /* Go through all the locks */ + for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) { + queue = dlm_list_idx_to_ptr(res, idx); + list_for_each_entry(lock, queue, list) { + if (lock->ml.node == dlm->node_num) + continue; + if (test_bit(lock->ml.node, dlm->exit_domain_map)) + continue; + nodenum = lock->ml.node; + goto bail; + } + } + + /* Go thru the refmap */ + noderef = -1; + while (1) { + noderef = find_next_bit(res->refmap, O2NM_MAX_NODES, + noderef + 1); + if (noderef >= O2NM_MAX_NODES) + break; + if (noderef == dlm->node_num) + continue; + if (test_bit(noderef, dlm->exit_domain_map)) + continue; + nodenum = noderef; + goto bail; + } + +bail: + return nodenum; +} + +/* this is called by the new master once all lockres + * data has been received */ +static int dlm_do_migrate_request(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + u8 master, u8 new_master, + struct dlm_node_iter *iter) +{ + struct dlm_migrate_request migrate; + int ret, skip, status = 0; + int nodenum; + + memset(&migrate, 0, sizeof(migrate)); + migrate.namelen = res->lockname.len; + memcpy(migrate.name, res->lockname.name, migrate.namelen); + migrate.new_master = new_master; + migrate.master = master; + + ret = 0; + + /* send message to all nodes, except the master and myself */ + while ((nodenum = dlm_node_iter_next(iter)) >= 0) { + if (nodenum == master || + nodenum == new_master) + continue; + + /* We could race exit domain. If exited, skip. */ + spin_lock(&dlm->spinlock); + skip = (!test_bit(nodenum, dlm->domain_map)); + spin_unlock(&dlm->spinlock); + if (skip) { + clear_bit(nodenum, iter->node_map); + continue; + } + + ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key, + &migrate, sizeof(migrate), nodenum, + &status); + if (ret < 0) { + mlog(ML_ERROR, "%s: res %.*s, Error %d send " + "MIGRATE_REQUEST to node %u\n", dlm->name, + migrate.namelen, migrate.name, ret, nodenum); + if (!dlm_is_host_down(ret)) { + mlog(ML_ERROR, "unhandled error=%d!\n", ret); + BUG(); + } + clear_bit(nodenum, iter->node_map); + ret = 0; + } else if (status < 0) { + mlog(0, "migrate request (node %u) returned %d!\n", + nodenum, status); + ret = status; + } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) { + /* during the migration request we short-circuited + * the mastery of the lockres. make sure we have + * a mastery ref for nodenum */ + mlog(0, "%s:%.*s: need ref for node %u\n", + dlm->name, res->lockname.len, res->lockname.name, + nodenum); + spin_lock(&res->spinlock); + dlm_lockres_set_refmap_bit(dlm, res, nodenum); + spin_unlock(&res->spinlock); + } + } + + if (ret < 0) + mlog_errno(ret); + + mlog(0, "returning ret=%d\n", ret); + return ret; +} + + +/* if there is an existing mle for this lockres, we now know who the master is. + * (the one who sent us *this* message) we can clear it up right away. + * since the process that put the mle on the list still has a reference to it, + * we can unhash it now, set the master and wake the process. as a result, + * we will have no mle in the list to start with. now we can add an mle for + * the migration and this should be the only one found for those scanning the + * list. */ +int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_lock_resource *res = NULL; + struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf; + struct dlm_master_list_entry *mle = NULL, *oldmle = NULL; + const char *name; + unsigned int namelen, hash; + int ret = 0; + + if (!dlm_grab(dlm)) + return 0; + + name = migrate->name; + namelen = migrate->namelen; + if (namelen > DLM_LOCKID_NAME_MAX) { + mlog(ML_ERROR, "%s: invalid name length %u in migrate request\n", + dlm->name, namelen); + ret = -EINVAL; + goto leave; + } + hash = dlm_lockid_hash(name, namelen); + + /* preallocate.. if this fails, abort */ + mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); + + if (!mle) { + ret = -ENOMEM; + goto leave; + } + + /* check for pre-existing lock */ + spin_lock(&dlm->spinlock); + res = __dlm_lookup_lockres(dlm, name, namelen, hash); + if (res) { + spin_lock(&res->spinlock); + if (res->state & DLM_LOCK_RES_RECOVERING) { + /* if all is working ok, this can only mean that we got + * a migrate request from a node that we now see as + * dead. what can we do here? drop it to the floor? */ + spin_unlock(&res->spinlock); + mlog(ML_ERROR, "Got a migrate request, but the " + "lockres is marked as recovering!"); + kmem_cache_free(dlm_mle_cache, mle); + ret = -EINVAL; /* need a better solution */ + goto unlock; + } + res->state |= DLM_LOCK_RES_MIGRATING; + spin_unlock(&res->spinlock); + } + + spin_lock(&dlm->master_lock); + /* ignore status. only nonzero status would BUG. */ + ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, + name, namelen, + migrate->new_master, + migrate->master); + + if (ret < 0) + kmem_cache_free(dlm_mle_cache, mle); + + spin_unlock(&dlm->master_lock); +unlock: + spin_unlock(&dlm->spinlock); + + if (oldmle) { + /* master is known, detach if not already detached */ + dlm_mle_detach_hb_events(dlm, oldmle); + dlm_put_mle(oldmle); + } + + if (res) + dlm_lockres_put(res); +leave: + dlm_put(dlm); + return ret; +} + +/* must be holding dlm->spinlock and dlm->master_lock + * when adding a migration mle, we can clear any other mles + * in the master list because we know with certainty that + * the master is "master". so we remove any old mle from + * the list after setting it's master field, and then add + * the new migration mle. this way we can hold with the rule + * of having only one mle for a given lock name at all times. */ +static int dlm_add_migration_mle(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_master_list_entry *mle, + struct dlm_master_list_entry **oldmle, + const char *name, unsigned int namelen, + u8 new_master, u8 master) +{ + int found; + int ret = 0; + + *oldmle = NULL; + + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&dlm->master_lock); + + /* caller is responsible for any ref taken here on oldmle */ + found = dlm_find_mle(dlm, oldmle, (char *)name, namelen); + if (found) { + struct dlm_master_list_entry *tmp = *oldmle; + spin_lock(&tmp->spinlock); + if (tmp->type == DLM_MLE_MIGRATION) { + if (master == dlm->node_num) { + /* ah another process raced me to it */ + mlog(0, "tried to migrate %.*s, but some " + "process beat me to it\n", + namelen, name); + spin_unlock(&tmp->spinlock); + return -EEXIST; + } else { + /* bad. 2 NODES are trying to migrate! */ + mlog(ML_ERROR, "migration error mle: " + "master=%u new_master=%u // request: " + "master=%u new_master=%u // " + "lockres=%.*s\n", + tmp->master, tmp->new_master, + master, new_master, + namelen, name); + BUG(); + } + } else { + /* this is essentially what assert_master does */ + tmp->master = master; + atomic_set(&tmp->woken, 1); + wake_up(&tmp->wq); + /* remove it so that only one mle will be found */ + __dlm_unlink_mle(dlm, tmp); + __dlm_mle_detach_hb_events(dlm, tmp); + if (tmp->type == DLM_MLE_MASTER) { + ret = DLM_MIGRATE_RESPONSE_MASTERY_REF; + mlog(0, "%s:%.*s: master=%u, newmaster=%u, " + "telling master to get ref " + "for cleared out mle during " + "migration\n", dlm->name, + namelen, name, master, + new_master); + } + } + spin_unlock(&tmp->spinlock); + } + + /* now add a migration mle to the tail of the list */ + dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen); + mle->new_master = new_master; + /* the new master will be sending an assert master for this. + * at that point we will get the refmap reference */ + mle->master = master; + /* do this for consistency with other mle types */ + set_bit(new_master, mle->maybe_map); + __dlm_insert_mle(dlm, mle); + + return ret; +} + +/* + * Sets the owner of the lockres, associated to the mle, to UNKNOWN + */ +static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle) +{ + struct dlm_lock_resource *res; + + /* Find the lockres associated to the mle and set its owner to UNK */ + res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen, + mle->mnamehash); + if (res) { + spin_unlock(&dlm->master_lock); + + /* move lockres onto recovery list */ + spin_lock(&res->spinlock); + dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN); + dlm_move_lockres_to_recovery_list(dlm, res); + spin_unlock(&res->spinlock); + dlm_lockres_put(res); + + /* about to get rid of mle, detach from heartbeat */ + __dlm_mle_detach_hb_events(dlm, mle); + + /* dump the mle */ + spin_lock(&dlm->master_lock); + __dlm_put_mle(mle); + spin_unlock(&dlm->master_lock); + } + + return res; +} + +static void dlm_clean_migration_mle(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle) +{ + __dlm_mle_detach_hb_events(dlm, mle); + + spin_lock(&mle->spinlock); + __dlm_unlink_mle(dlm, mle); + atomic_set(&mle->woken, 1); + spin_unlock(&mle->spinlock); + + wake_up(&mle->wq); +} + +static void dlm_clean_block_mle(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle, u8 dead_node) +{ + int bit; + + BUG_ON(mle->type != DLM_MLE_BLOCK); + + spin_lock(&mle->spinlock); + bit = find_first_bit(mle->maybe_map, O2NM_MAX_NODES); + if (bit != dead_node) { + mlog(0, "mle found, but dead node %u would not have been " + "master\n", dead_node); + spin_unlock(&mle->spinlock); + } else { + /* Must drop the refcount by one since the assert_master will + * never arrive. This may result in the mle being unlinked and + * freed, but there may still be a process waiting in the + * dlmlock path which is fine. */ + mlog(0, "node %u was expected master\n", dead_node); + atomic_set(&mle->woken, 1); + spin_unlock(&mle->spinlock); + wake_up(&mle->wq); + + /* Do not need events any longer, so detach from heartbeat */ + __dlm_mle_detach_hb_events(dlm, mle); + __dlm_put_mle(mle); + } +} + +void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node) +{ + struct dlm_master_list_entry *mle; + struct dlm_lock_resource *res; + struct hlist_head *bucket; + struct hlist_node *tmp; + unsigned int i; + + mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node); +top: + assert_spin_locked(&dlm->spinlock); + + /* clean the master list */ + spin_lock(&dlm->master_lock); + for (i = 0; i < DLM_HASH_BUCKETS; i++) { + bucket = dlm_master_hash(dlm, i); + hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) { + BUG_ON(mle->type != DLM_MLE_BLOCK && + mle->type != DLM_MLE_MASTER && + mle->type != DLM_MLE_MIGRATION); + + /* MASTER mles are initiated locally. The waiting + * process will notice the node map change shortly. + * Let that happen as normal. */ + if (mle->type == DLM_MLE_MASTER) + continue; + + /* BLOCK mles are initiated by other nodes. Need to + * clean up if the dead node would have been the + * master. */ + if (mle->type == DLM_MLE_BLOCK) { + dlm_clean_block_mle(dlm, mle, dead_node); + continue; + } + + /* Everything else is a MIGRATION mle */ + + /* The rule for MIGRATION mles is that the master + * becomes UNKNOWN if *either* the original or the new + * master dies. All UNKNOWN lockres' are sent to + * whichever node becomes the recovery master. The new + * master is responsible for determining if there is + * still a master for this lockres, or if he needs to + * take over mastery. Either way, this node should + * expect another message to resolve this. */ + + if (mle->master != dead_node && + mle->new_master != dead_node) + continue; + + if (mle->new_master == dead_node && mle->inuse) { + mlog(ML_NOTICE, "%s: target %u died during " + "migration from %u, the MLE is " + "still keep used, ignore it!\n", + dlm->name, dead_node, + mle->master); + continue; + } + + /* If we have reached this point, this mle needs to be + * removed from the list and freed. */ + dlm_clean_migration_mle(dlm, mle); + + mlog(0, "%s: node %u died during migration from " + "%u to %u!\n", dlm->name, dead_node, mle->master, + mle->new_master); + + /* If we find a lockres associated with the mle, we've + * hit this rare case that messes up our lock ordering. + * If so, we need to drop the master lock so that we can + * take the lockres lock, meaning that we will have to + * restart from the head of list. */ + res = dlm_reset_mleres_owner(dlm, mle); + if (res) + /* restart */ + goto top; + + /* This may be the last reference */ + __dlm_put_mle(mle); + } + } + spin_unlock(&dlm->master_lock); +} + +int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, + u8 old_master) +{ + struct dlm_node_iter iter; + int ret = 0; + + spin_lock(&dlm->spinlock); + dlm_node_iter_init(dlm->domain_map, &iter); + clear_bit(old_master, iter.node_map); + clear_bit(dlm->node_num, iter.node_map); + spin_unlock(&dlm->spinlock); + + /* ownership of the lockres is changing. account for the + * mastery reference here since old_master will briefly have + * a reference after the migration completes */ + spin_lock(&res->spinlock); + dlm_lockres_set_refmap_bit(dlm, res, old_master); + spin_unlock(&res->spinlock); + + mlog(0, "now time to do a migrate request to other nodes\n"); + ret = dlm_do_migrate_request(dlm, res, old_master, + dlm->node_num, &iter); + if (ret < 0) { + mlog_errno(ret); + goto leave; + } + + mlog(0, "doing assert master of %.*s to all except the original node\n", + res->lockname.len, res->lockname.name); + /* this call now finishes out the nodemap + * even if one or more nodes die */ + ret = dlm_do_assert_master(dlm, res, iter.node_map, + DLM_ASSERT_MASTER_FINISH_MIGRATION); + if (ret < 0) { + /* no longer need to retry. all living nodes contacted. */ + mlog_errno(ret); + ret = 0; + } + + bitmap_zero(iter.node_map, O2NM_MAX_NODES); + set_bit(old_master, iter.node_map); + mlog(0, "doing assert master of %.*s back to %u\n", + res->lockname.len, res->lockname.name, old_master); + ret = dlm_do_assert_master(dlm, res, iter.node_map, + DLM_ASSERT_MASTER_FINISH_MIGRATION); + if (ret < 0) { + mlog(0, "assert master to original master failed " + "with %d.\n", ret); + /* the only nonzero status here would be because of + * a dead original node. we're done. */ + ret = 0; + } + + /* all done, set the owner, clear the flag */ + spin_lock(&res->spinlock); + dlm_set_lockres_owner(dlm, res, dlm->node_num); + res->state &= ~DLM_LOCK_RES_MIGRATING; + spin_unlock(&res->spinlock); + /* re-dirty it on the new master */ + dlm_kick_thread(dlm, res); + wake_up(&res->wq); +leave: + return ret; +} + +/* + * LOCKRES AST REFCOUNT + * this is integral to migration + */ + +/* for future intent to call an ast, reserve one ahead of time. + * this should be called only after waiting on the lockres + * with dlm_wait_on_lockres, and while still holding the + * spinlock after the call. */ +void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res) +{ + assert_spin_locked(&res->spinlock); + if (res->state & DLM_LOCK_RES_MIGRATING) { + __dlm_print_one_lock_resource(res); + } + BUG_ON(res->state & DLM_LOCK_RES_MIGRATING); + + atomic_inc(&res->asts_reserved); +} + +/* + * used to drop the reserved ast, either because it went unused, + * or because the ast/bast was actually called. + * + * also, if there is a pending migration on this lockres, + * and this was the last pending ast on the lockres, + * atomically set the MIGRATING flag before we drop the lock. + * this is how we ensure that migration can proceed with no + * asts in progress. note that it is ok if the state of the + * queues is such that a lock should be granted in the future + * or that a bast should be fired, because the new master will + * shuffle the lists on this lockres as soon as it is migrated. + */ +void dlm_lockres_release_ast(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock)) + return; + + if (!res->migration_pending) { + spin_unlock(&res->spinlock); + return; + } + + BUG_ON(res->state & DLM_LOCK_RES_MIGRATING); + res->migration_pending = 0; + res->state |= DLM_LOCK_RES_MIGRATING; + spin_unlock(&res->spinlock); + wake_up(&res->wq); + wake_up(&dlm->migration_wq); +} + +void dlm_force_free_mles(struct dlm_ctxt *dlm) +{ + int i; + struct hlist_head *bucket; + struct dlm_master_list_entry *mle; + struct hlist_node *tmp; + + /* + * We notified all other nodes that we are exiting the domain and + * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still + * around we force free them and wake any processes that are waiting + * on the mles + */ + spin_lock(&dlm->spinlock); + spin_lock(&dlm->master_lock); + + BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING); + BUG_ON((find_first_bit(dlm->domain_map, O2NM_MAX_NODES) < O2NM_MAX_NODES)); + + for (i = 0; i < DLM_HASH_BUCKETS; i++) { + bucket = dlm_master_hash(dlm, i); + hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) { + if (mle->type != DLM_MLE_BLOCK) { + mlog(ML_ERROR, "bad mle: %p\n", mle); + dlm_print_one_mle(mle); + } + atomic_set(&mle->woken, 1); + wake_up(&mle->wq); + + __dlm_unlink_mle(dlm, mle); + __dlm_mle_detach_hb_events(dlm, mle); + __dlm_put_mle(mle); + } + } + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); +} diff --git a/fs/ocfs2/dlm/dlmrecovery.c b/fs/ocfs2/dlm/dlmrecovery.c new file mode 100644 index 000000000..9d4a2695b --- /dev/null +++ b/fs/ocfs2/dlm/dlmrecovery.c @@ -0,0 +1,2962 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * dlmrecovery.c + * + * recovery stuff + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "../cluster/heartbeat.h" +#include "../cluster/nodemanager.h" +#include "../cluster/tcp.h" + +#include "dlmapi.h" +#include "dlmcommon.h" +#include "dlmdomain.h" + +#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY) +#include "../cluster/masklog.h" + +static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node); + +static int dlm_recovery_thread(void *data); +static int dlm_do_recovery(struct dlm_ctxt *dlm); + +static int dlm_pick_recovery_master(struct dlm_ctxt *dlm); +static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node); +static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node); +static int dlm_request_all_locks(struct dlm_ctxt *dlm, + u8 request_from, u8 dead_node); +static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm); + +static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res); +static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres, + const char *lockname, int namelen, + int total_locks, u64 cookie, + u8 flags, u8 master); +static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm, + struct dlm_migratable_lockres *mres, + u8 send_to, + struct dlm_lock_resource *res, + int total_locks); +static int dlm_process_recovery_data(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_migratable_lockres *mres); +static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm); +static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, + u8 dead_node, u8 send_to); +static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node); +static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm, + struct list_head *list, u8 dead_node); +static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm, + u8 dead_node, u8 new_master); +static void dlm_reco_ast(void *astdata); +static void dlm_reco_bast(void *astdata, int blocked_type); +static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st); +static void dlm_request_all_locks_worker(struct dlm_work_item *item, + void *data); +static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data); +static int dlm_lockres_master_requery(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + u8 *real_master); + +static u64 dlm_get_next_mig_cookie(void); + +static DEFINE_SPINLOCK(dlm_reco_state_lock); +static DEFINE_SPINLOCK(dlm_mig_cookie_lock); +static u64 dlm_mig_cookie = 1; + +static u64 dlm_get_next_mig_cookie(void) +{ + u64 c; + spin_lock(&dlm_mig_cookie_lock); + c = dlm_mig_cookie; + if (dlm_mig_cookie == (~0ULL)) + dlm_mig_cookie = 1; + else + dlm_mig_cookie++; + spin_unlock(&dlm_mig_cookie_lock); + return c; +} + +static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm, + u8 dead_node) +{ + assert_spin_locked(&dlm->spinlock); + if (dlm->reco.dead_node != dead_node) + mlog(0, "%s: changing dead_node from %u to %u\n", + dlm->name, dlm->reco.dead_node, dead_node); + dlm->reco.dead_node = dead_node; +} + +static inline void dlm_set_reco_master(struct dlm_ctxt *dlm, + u8 master) +{ + assert_spin_locked(&dlm->spinlock); + mlog(0, "%s: changing new_master from %u to %u\n", + dlm->name, dlm->reco.new_master, master); + dlm->reco.new_master = master; +} + +static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm) +{ + assert_spin_locked(&dlm->spinlock); + clear_bit(dlm->reco.dead_node, dlm->recovery_map); + dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); + dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM); +} + +/* Worker function used during recovery. */ +void dlm_dispatch_work(struct work_struct *work) +{ + struct dlm_ctxt *dlm = + container_of(work, struct dlm_ctxt, dispatched_work); + LIST_HEAD(tmp_list); + struct dlm_work_item *item, *next; + dlm_workfunc_t *workfunc; + int tot=0; + + spin_lock(&dlm->work_lock); + list_splice_init(&dlm->work_list, &tmp_list); + spin_unlock(&dlm->work_lock); + + list_for_each_entry(item, &tmp_list, list) { + tot++; + } + mlog(0, "%s: work thread has %d work items\n", dlm->name, tot); + + list_for_each_entry_safe(item, next, &tmp_list, list) { + workfunc = item->func; + list_del_init(&item->list); + + /* already have ref on dlm to avoid having + * it disappear. just double-check. */ + BUG_ON(item->dlm != dlm); + + /* this is allowed to sleep and + * call network stuff */ + workfunc(item, item->data); + + dlm_put(dlm); + kfree(item); + } +} + +/* + * RECOVERY THREAD + */ + +void dlm_kick_recovery_thread(struct dlm_ctxt *dlm) +{ + /* wake the recovery thread + * this will wake the reco thread in one of three places + * 1) sleeping with no recovery happening + * 2) sleeping with recovery mastered elsewhere + * 3) recovery mastered here, waiting on reco data */ + + wake_up(&dlm->dlm_reco_thread_wq); +} + +/* Launch the recovery thread */ +int dlm_launch_recovery_thread(struct dlm_ctxt *dlm) +{ + mlog(0, "starting dlm recovery thread...\n"); + + dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm, + "dlm_reco-%s", dlm->name); + if (IS_ERR(dlm->dlm_reco_thread_task)) { + mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task)); + dlm->dlm_reco_thread_task = NULL; + return -EINVAL; + } + + return 0; +} + +void dlm_complete_recovery_thread(struct dlm_ctxt *dlm) +{ + if (dlm->dlm_reco_thread_task) { + mlog(0, "waiting for dlm recovery thread to exit\n"); + kthread_stop(dlm->dlm_reco_thread_task); + dlm->dlm_reco_thread_task = NULL; + } +} + + + +/* + * this is lame, but here's how recovery works... + * 1) all recovery threads cluster wide will work on recovering + * ONE node at a time + * 2) negotiate who will take over all the locks for the dead node. + * that's right... ALL the locks. + * 3) once a new master is chosen, everyone scans all locks + * and moves aside those mastered by the dead guy + * 4) each of these locks should be locked until recovery is done + * 5) the new master collects up all of secondary lock queue info + * one lock at a time, forcing each node to communicate back + * before continuing + * 6) each secondary lock queue responds with the full known lock info + * 7) once the new master has run all its locks, it sends a ALLDONE! + * message to everyone + * 8) upon receiving this message, the secondary queue node unlocks + * and responds to the ALLDONE + * 9) once the new master gets responses from everyone, he unlocks + * everything and recovery for this dead node is done + *10) go back to 2) while there are still dead nodes + * + */ + +static void dlm_print_reco_node_status(struct dlm_ctxt *dlm) +{ + struct dlm_reco_node_data *ndata; + struct dlm_lock_resource *res; + + mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n", + dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), + dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive", + dlm->reco.dead_node, dlm->reco.new_master); + + list_for_each_entry(ndata, &dlm->reco.node_data, list) { + char *st = "unknown"; + switch (ndata->state) { + case DLM_RECO_NODE_DATA_INIT: + st = "init"; + break; + case DLM_RECO_NODE_DATA_REQUESTING: + st = "requesting"; + break; + case DLM_RECO_NODE_DATA_DEAD: + st = "dead"; + break; + case DLM_RECO_NODE_DATA_RECEIVING: + st = "receiving"; + break; + case DLM_RECO_NODE_DATA_REQUESTED: + st = "requested"; + break; + case DLM_RECO_NODE_DATA_DONE: + st = "done"; + break; + case DLM_RECO_NODE_DATA_FINALIZE_SENT: + st = "finalize-sent"; + break; + default: + st = "bad"; + break; + } + mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n", + dlm->name, ndata->node_num, st); + } + list_for_each_entry(res, &dlm->reco.resources, recovering) { + mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n", + dlm->name, res->lockname.len, res->lockname.name); + } +} + +#define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000) + +static int dlm_recovery_thread(void *data) +{ + int status; + struct dlm_ctxt *dlm = data; + unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS); + + mlog(0, "dlm thread running for %s...\n", dlm->name); + + while (!kthread_should_stop()) { + if (dlm_domain_fully_joined(dlm)) { + status = dlm_do_recovery(dlm); + if (status == -EAGAIN) { + /* do not sleep, recheck immediately. */ + continue; + } + if (status < 0) + mlog_errno(status); + } + + wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq, + kthread_should_stop(), + timeout); + } + + mlog(0, "quitting DLM recovery thread\n"); + return 0; +} + +/* returns true when the recovery master has contacted us */ +static int dlm_reco_master_ready(struct dlm_ctxt *dlm) +{ + int ready; + spin_lock(&dlm->spinlock); + ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM); + spin_unlock(&dlm->spinlock); + return ready; +} + +/* returns true if node is no longer in the domain + * could be dead or just not joined */ +int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node) +{ + int dead; + spin_lock(&dlm->spinlock); + dead = !test_bit(node, dlm->domain_map); + spin_unlock(&dlm->spinlock); + return dead; +} + +/* returns true if node is no longer in the domain + * could be dead or just not joined */ +static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node) +{ + int recovered; + spin_lock(&dlm->spinlock); + recovered = !test_bit(node, dlm->recovery_map); + spin_unlock(&dlm->spinlock); + return recovered; +} + + +void dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout) +{ + if (dlm_is_node_dead(dlm, node)) + return; + + printk(KERN_NOTICE "o2dlm: Waiting on the death of node %u in " + "domain %s\n", node, dlm->name); + + if (timeout) + wait_event_timeout(dlm->dlm_reco_thread_wq, + dlm_is_node_dead(dlm, node), + msecs_to_jiffies(timeout)); + else + wait_event(dlm->dlm_reco_thread_wq, + dlm_is_node_dead(dlm, node)); +} + +void dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout) +{ + if (dlm_is_node_recovered(dlm, node)) + return; + + printk(KERN_NOTICE "o2dlm: Waiting on the recovery of node %u in " + "domain %s\n", node, dlm->name); + + if (timeout) + wait_event_timeout(dlm->dlm_reco_thread_wq, + dlm_is_node_recovered(dlm, node), + msecs_to_jiffies(timeout)); + else + wait_event(dlm->dlm_reco_thread_wq, + dlm_is_node_recovered(dlm, node)); +} + +/* callers of the top-level api calls (dlmlock/dlmunlock) should + * block on the dlm->reco.event when recovery is in progress. + * the dlm recovery thread will set this state when it begins + * recovering a dead node (as the new master or not) and clear + * the state and wake as soon as all affected lock resources have + * been marked with the RECOVERY flag */ +static int dlm_in_recovery(struct dlm_ctxt *dlm) +{ + int in_recovery; + spin_lock(&dlm->spinlock); + in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE); + spin_unlock(&dlm->spinlock); + return in_recovery; +} + + +void dlm_wait_for_recovery(struct dlm_ctxt *dlm) +{ + if (dlm_in_recovery(dlm)) { + mlog(0, "%s: reco thread %d in recovery: " + "state=%d, master=%u, dead=%u\n", + dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), + dlm->reco.state, dlm->reco.new_master, + dlm->reco.dead_node); + } + wait_event(dlm->reco.event, !dlm_in_recovery(dlm)); +} + +static void dlm_begin_recovery(struct dlm_ctxt *dlm) +{ + assert_spin_locked(&dlm->spinlock); + BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE); + printk(KERN_NOTICE "o2dlm: Begin recovery on domain %s for node %u\n", + dlm->name, dlm->reco.dead_node); + dlm->reco.state |= DLM_RECO_STATE_ACTIVE; +} + +static void dlm_end_recovery(struct dlm_ctxt *dlm) +{ + spin_lock(&dlm->spinlock); + BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE)); + dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE; + spin_unlock(&dlm->spinlock); + printk(KERN_NOTICE "o2dlm: End recovery on domain %s\n", dlm->name); + wake_up(&dlm->reco.event); +} + +static void dlm_print_recovery_master(struct dlm_ctxt *dlm) +{ + printk(KERN_NOTICE "o2dlm: Node %u (%s) is the Recovery Master for the " + "dead node %u in domain %s\n", dlm->reco.new_master, + (dlm->node_num == dlm->reco.new_master ? "me" : "he"), + dlm->reco.dead_node, dlm->name); +} + +static int dlm_do_recovery(struct dlm_ctxt *dlm) +{ + int status = 0; + int ret; + + spin_lock(&dlm->spinlock); + + if (dlm->migrate_done) { + mlog(0, "%s: no need do recovery after migrating all " + "lock resources\n", dlm->name); + spin_unlock(&dlm->spinlock); + return 0; + } + + /* check to see if the new master has died */ + if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM && + test_bit(dlm->reco.new_master, dlm->recovery_map)) { + mlog(0, "new master %u died while recovering %u!\n", + dlm->reco.new_master, dlm->reco.dead_node); + /* unset the new_master, leave dead_node */ + dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM); + } + + /* select a target to recover */ + if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { + int bit; + + bit = find_first_bit(dlm->recovery_map, O2NM_MAX_NODES); + if (bit >= O2NM_MAX_NODES || bit < 0) + dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); + else + dlm_set_reco_dead_node(dlm, bit); + } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) { + /* BUG? */ + mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n", + dlm->reco.dead_node); + dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); + } + + if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { + spin_unlock(&dlm->spinlock); + /* return to main thread loop and sleep. */ + return 0; + } + mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n", + dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), + dlm->reco.dead_node); + + /* take write barrier */ + /* (stops the list reshuffling thread, proxy ast handling) */ + dlm_begin_recovery(dlm); + + spin_unlock(&dlm->spinlock); + + if (dlm->reco.new_master == dlm->node_num) + goto master_here; + + if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) { + /* choose a new master, returns 0 if this node + * is the master, -EEXIST if it's another node. + * this does not return until a new master is chosen + * or recovery completes entirely. */ + ret = dlm_pick_recovery_master(dlm); + if (!ret) { + /* already notified everyone. go. */ + goto master_here; + } + mlog(0, "another node will master this recovery session.\n"); + } + + dlm_print_recovery_master(dlm); + + /* it is safe to start everything back up here + * because all of the dead node's lock resources + * have been marked as in-recovery */ + dlm_end_recovery(dlm); + + /* sleep out in main dlm_recovery_thread loop. */ + return 0; + +master_here: + dlm_print_recovery_master(dlm); + + status = dlm_remaster_locks(dlm, dlm->reco.dead_node); + if (status < 0) { + /* we should never hit this anymore */ + mlog(ML_ERROR, "%s: Error %d remastering locks for node %u, " + "retrying.\n", dlm->name, status, dlm->reco.dead_node); + /* yield a bit to allow any final network messages + * to get handled on remaining nodes */ + msleep(100); + } else { + /* success! see if any other nodes need recovery */ + mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n", + dlm->name, dlm->reco.dead_node, dlm->node_num); + spin_lock(&dlm->spinlock); + __dlm_reset_recovery(dlm); + dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; + spin_unlock(&dlm->spinlock); + } + dlm_end_recovery(dlm); + + /* continue and look for another dead node */ + return -EAGAIN; +} + +static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node) +{ + int status = 0; + struct dlm_reco_node_data *ndata; + int all_nodes_done; + int destroy = 0; + int pass = 0; + + do { + /* we have become recovery master. there is no escaping + * this, so just keep trying until we get it. */ + status = dlm_init_recovery_area(dlm, dead_node); + if (status < 0) { + mlog(ML_ERROR, "%s: failed to alloc recovery area, " + "retrying\n", dlm->name); + msleep(1000); + } + } while (status != 0); + + /* safe to access the node data list without a lock, since this + * process is the only one to change the list */ + list_for_each_entry(ndata, &dlm->reco.node_data, list) { + BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT); + ndata->state = DLM_RECO_NODE_DATA_REQUESTING; + + mlog(0, "%s: Requesting lock info from node %u\n", dlm->name, + ndata->node_num); + + if (ndata->node_num == dlm->node_num) { + ndata->state = DLM_RECO_NODE_DATA_DONE; + continue; + } + + do { + status = dlm_request_all_locks(dlm, ndata->node_num, + dead_node); + if (status < 0) { + mlog_errno(status); + if (dlm_is_host_down(status)) { + /* node died, ignore it for recovery */ + status = 0; + ndata->state = DLM_RECO_NODE_DATA_DEAD; + /* wait for the domain map to catch up + * with the network state. */ + wait_event_timeout(dlm->dlm_reco_thread_wq, + dlm_is_node_dead(dlm, + ndata->node_num), + msecs_to_jiffies(1000)); + mlog(0, "waited 1 sec for %u, " + "dead? %s\n", ndata->node_num, + str_yes_no(dlm_is_node_dead(dlm, ndata->node_num))); + } else { + /* -ENOMEM on the other node */ + mlog(0, "%s: node %u returned " + "%d during recovery, retrying " + "after a short wait\n", + dlm->name, ndata->node_num, + status); + msleep(100); + } + } + } while (status != 0); + + spin_lock(&dlm_reco_state_lock); + switch (ndata->state) { + case DLM_RECO_NODE_DATA_INIT: + case DLM_RECO_NODE_DATA_FINALIZE_SENT: + case DLM_RECO_NODE_DATA_REQUESTED: + BUG(); + break; + case DLM_RECO_NODE_DATA_DEAD: + mlog(0, "node %u died after requesting " + "recovery info for node %u\n", + ndata->node_num, dead_node); + /* fine. don't need this node's info. + * continue without it. */ + break; + case DLM_RECO_NODE_DATA_REQUESTING: + ndata->state = DLM_RECO_NODE_DATA_REQUESTED; + mlog(0, "now receiving recovery data from " + "node %u for dead node %u\n", + ndata->node_num, dead_node); + break; + case DLM_RECO_NODE_DATA_RECEIVING: + mlog(0, "already receiving recovery data from " + "node %u for dead node %u\n", + ndata->node_num, dead_node); + break; + case DLM_RECO_NODE_DATA_DONE: + mlog(0, "already DONE receiving recovery data " + "from node %u for dead node %u\n", + ndata->node_num, dead_node); + break; + } + spin_unlock(&dlm_reco_state_lock); + } + + mlog(0, "%s: Done requesting all lock info\n", dlm->name); + + /* nodes should be sending reco data now + * just need to wait */ + + while (1) { + /* check all the nodes now to see if we are + * done, or if anyone died */ + all_nodes_done = 1; + spin_lock(&dlm_reco_state_lock); + list_for_each_entry(ndata, &dlm->reco.node_data, list) { + mlog(0, "checking recovery state of node %u\n", + ndata->node_num); + switch (ndata->state) { + case DLM_RECO_NODE_DATA_INIT: + case DLM_RECO_NODE_DATA_REQUESTING: + mlog(ML_ERROR, "bad ndata state for " + "node %u: state=%d\n", + ndata->node_num, ndata->state); + BUG(); + break; + case DLM_RECO_NODE_DATA_DEAD: + mlog(0, "node %u died after " + "requesting recovery info for " + "node %u\n", ndata->node_num, + dead_node); + break; + case DLM_RECO_NODE_DATA_RECEIVING: + case DLM_RECO_NODE_DATA_REQUESTED: + mlog(0, "%s: node %u still in state %s\n", + dlm->name, ndata->node_num, + ndata->state==DLM_RECO_NODE_DATA_RECEIVING ? + "receiving" : "requested"); + all_nodes_done = 0; + break; + case DLM_RECO_NODE_DATA_DONE: + mlog(0, "%s: node %u state is done\n", + dlm->name, ndata->node_num); + break; + case DLM_RECO_NODE_DATA_FINALIZE_SENT: + mlog(0, "%s: node %u state is finalize\n", + dlm->name, ndata->node_num); + break; + } + } + spin_unlock(&dlm_reco_state_lock); + + mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass, + str_yes_no(all_nodes_done)); + if (all_nodes_done) { + int ret; + + /* Set this flag on recovery master to avoid + * a new recovery for another dead node start + * before the recovery is not done. That may + * cause recovery hung.*/ + spin_lock(&dlm->spinlock); + dlm->reco.state |= DLM_RECO_STATE_FINALIZE; + spin_unlock(&dlm->spinlock); + + /* all nodes are now in DLM_RECO_NODE_DATA_DONE state + * just send a finalize message to everyone and + * clean up */ + mlog(0, "all nodes are done! send finalize\n"); + ret = dlm_send_finalize_reco_message(dlm); + if (ret < 0) + mlog_errno(ret); + + spin_lock(&dlm->spinlock); + dlm_finish_local_lockres_recovery(dlm, dead_node, + dlm->node_num); + spin_unlock(&dlm->spinlock); + mlog(0, "should be done with recovery!\n"); + + mlog(0, "finishing recovery of %s at %lu, " + "dead=%u, this=%u, new=%u\n", dlm->name, + jiffies, dlm->reco.dead_node, + dlm->node_num, dlm->reco.new_master); + destroy = 1; + status = 0; + /* rescan everything marked dirty along the way */ + dlm_kick_thread(dlm, NULL); + break; + } + /* wait to be signalled, with periodic timeout + * to check for node death */ + wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq, + kthread_should_stop(), + msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS)); + + } + + if (destroy) + dlm_destroy_recovery_area(dlm); + + return status; +} + +static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node) +{ + int num=0; + struct dlm_reco_node_data *ndata; + + spin_lock(&dlm->spinlock); + bitmap_copy(dlm->reco.node_map, dlm->domain_map, O2NM_MAX_NODES); + /* nodes can only be removed (by dying) after dropping + * this lock, and death will be trapped later, so this should do */ + spin_unlock(&dlm->spinlock); + + while (1) { + num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num); + if (num >= O2NM_MAX_NODES) { + break; + } + BUG_ON(num == dead_node); + + ndata = kzalloc_obj(*ndata, GFP_NOFS); + if (!ndata) { + dlm_destroy_recovery_area(dlm); + return -ENOMEM; + } + ndata->node_num = num; + ndata->state = DLM_RECO_NODE_DATA_INIT; + spin_lock(&dlm_reco_state_lock); + list_add_tail(&ndata->list, &dlm->reco.node_data); + spin_unlock(&dlm_reco_state_lock); + num++; + } + + return 0; +} + +static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm) +{ + struct dlm_reco_node_data *ndata, *next; + LIST_HEAD(tmplist); + + spin_lock(&dlm_reco_state_lock); + list_splice_init(&dlm->reco.node_data, &tmplist); + spin_unlock(&dlm_reco_state_lock); + + list_for_each_entry_safe(ndata, next, &tmplist, list) { + list_del_init(&ndata->list); + kfree(ndata); + } +} + +static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from, + u8 dead_node) +{ + struct dlm_lock_request lr; + int ret; + int status; + + mlog(0, "\n"); + + + mlog(0, "dlm_request_all_locks: dead node is %u, sending request " + "to %u\n", dead_node, request_from); + + memset(&lr, 0, sizeof(lr)); + lr.node_idx = dlm->node_num; + lr.dead_node = dead_node; + + // send message + ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key, + &lr, sizeof(lr), request_from, &status); + + /* negative status is handled by caller */ + if (ret < 0) + mlog(ML_ERROR, "%s: Error %d send LOCK_REQUEST to node %u " + "to recover dead node %u\n", dlm->name, ret, + request_from, dead_node); + else + ret = status; + // return from here, then + // sleep until all received or error + return ret; + +} + +int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf; + char *buf = NULL; + struct dlm_work_item *item = NULL; + + if (!dlm_grab(dlm)) + return -EINVAL; + + if (lr->dead_node != dlm->reco.dead_node) { + mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local " + "dead_node is %u\n", dlm->name, lr->node_idx, + lr->dead_node, dlm->reco.dead_node); + dlm_print_reco_node_status(dlm); + /* this is a hack */ + dlm_put(dlm); + return -ENOMEM; + } + BUG_ON(lr->dead_node != dlm->reco.dead_node); + + item = kzalloc_obj(*item, GFP_NOFS); + if (!item) { + dlm_put(dlm); + return -ENOMEM; + } + + /* this will get freed by dlm_request_all_locks_worker */ + buf = kmalloc(PAGE_SIZE, GFP_NOFS); + if (!buf) { + kfree(item); + dlm_put(dlm); + return -ENOMEM; + } + + /* queue up work for dlm_request_all_locks_worker */ + dlm_grab(dlm); /* get an extra ref for the work item */ + dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf); + item->u.ral.reco_master = lr->node_idx; + item->u.ral.dead_node = lr->dead_node; + spin_lock(&dlm->work_lock); + list_add_tail(&item->list, &dlm->work_list); + spin_unlock(&dlm->work_lock); + queue_work(dlm->dlm_worker, &dlm->dispatched_work); + + dlm_put(dlm); + return 0; +} + +static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data) +{ + struct dlm_migratable_lockres *mres; + struct dlm_lock_resource *res; + struct dlm_ctxt *dlm; + LIST_HEAD(resources); + int ret; + u8 dead_node, reco_master; + int skip_all_done = 0; + + dlm = item->dlm; + dead_node = item->u.ral.dead_node; + reco_master = item->u.ral.reco_master; + mres = (struct dlm_migratable_lockres *)data; + + mlog(0, "%s: recovery worker started, dead=%u, master=%u\n", + dlm->name, dead_node, reco_master); + + if (dead_node != dlm->reco.dead_node || + reco_master != dlm->reco.new_master) { + /* worker could have been created before the recovery master + * died. if so, do not continue, but do not error. */ + if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) { + mlog(ML_NOTICE, "%s: will not send recovery state, " + "recovery master %u died, thread=(dead=%u,mas=%u)" + " current=(dead=%u,mas=%u)\n", dlm->name, + reco_master, dead_node, reco_master, + dlm->reco.dead_node, dlm->reco.new_master); + } else { + mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, " + "master=%u), request(dead=%u, master=%u)\n", + dlm->name, dlm->reco.dead_node, + dlm->reco.new_master, dead_node, reco_master); + } + goto leave; + } + + /* lock resources should have already been moved to the + * dlm->reco.resources list. now move items from that list + * to a temp list if the dead owner matches. note that the + * whole cluster recovers only one node at a time, so we + * can safely move UNKNOWN lock resources for each recovery + * session. */ + dlm_move_reco_locks_to_list(dlm, &resources, dead_node); + + /* now we can begin blasting lockreses without the dlm lock */ + + /* any errors returned will be due to the new_master dying, + * the dlm_reco_thread should detect this */ + list_for_each_entry(res, &resources, recovering) { + ret = dlm_send_one_lockres(dlm, res, mres, reco_master, + DLM_MRES_RECOVERY); + if (ret < 0) { + mlog(ML_ERROR, "%s: node %u went down while sending " + "recovery state for dead node %u, ret=%d\n", dlm->name, + reco_master, dead_node, ret); + skip_all_done = 1; + break; + } + } + + /* move the resources back to the list */ + spin_lock(&dlm->spinlock); + list_splice_init(&resources, &dlm->reco.resources); + spin_unlock(&dlm->spinlock); + + if (!skip_all_done) { + ret = dlm_send_all_done_msg(dlm, dead_node, reco_master); + if (ret < 0) { + mlog(ML_ERROR, "%s: node %u went down while sending " + "recovery all-done for dead node %u, ret=%d\n", + dlm->name, reco_master, dead_node, ret); + } + } +leave: + kfree(data); +} + + +static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to) +{ + int ret, tmpret; + struct dlm_reco_data_done done_msg; + + memset(&done_msg, 0, sizeof(done_msg)); + done_msg.node_idx = dlm->node_num; + done_msg.dead_node = dead_node; + mlog(0, "sending DATA DONE message to %u, " + "my node=%u, dead node=%u\n", send_to, done_msg.node_idx, + done_msg.dead_node); + + ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg, + sizeof(done_msg), send_to, &tmpret); + if (ret < 0) { + mlog(ML_ERROR, "%s: Error %d send RECO_DATA_DONE to node %u " + "to recover dead node %u\n", dlm->name, ret, send_to, + dead_node); + if (!dlm_is_host_down(ret)) { + BUG(); + } + } else + ret = tmpret; + return ret; +} + + +int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf; + struct dlm_reco_node_data *ndata = NULL; + int ret = -EINVAL; + + if (!dlm_grab(dlm)) + return -EINVAL; + + mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, " + "node_idx=%u, this node=%u\n", done->dead_node, + dlm->reco.dead_node, done->node_idx, dlm->node_num); + + mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node), + "Got DATA DONE: dead_node=%u, reco.dead_node=%u, " + "node_idx=%u, this node=%u\n", done->dead_node, + dlm->reco.dead_node, done->node_idx, dlm->node_num); + + spin_lock(&dlm_reco_state_lock); + list_for_each_entry(ndata, &dlm->reco.node_data, list) { + if (ndata->node_num != done->node_idx) + continue; + + switch (ndata->state) { + /* should have moved beyond INIT but not to FINALIZE yet */ + case DLM_RECO_NODE_DATA_INIT: + case DLM_RECO_NODE_DATA_DEAD: + case DLM_RECO_NODE_DATA_FINALIZE_SENT: + mlog(ML_ERROR, "bad ndata state for node %u:" + " state=%d\n", ndata->node_num, + ndata->state); + BUG(); + break; + /* these states are possible at this point, anywhere along + * the line of recovery */ + case DLM_RECO_NODE_DATA_DONE: + case DLM_RECO_NODE_DATA_RECEIVING: + case DLM_RECO_NODE_DATA_REQUESTED: + case DLM_RECO_NODE_DATA_REQUESTING: + mlog(0, "node %u is DONE sending " + "recovery data!\n", + ndata->node_num); + + ndata->state = DLM_RECO_NODE_DATA_DONE; + ret = 0; + break; + } + } + spin_unlock(&dlm_reco_state_lock); + + /* wake the recovery thread, some node is done */ + if (!ret) + dlm_kick_recovery_thread(dlm); + + if (ret < 0) + mlog(ML_ERROR, "failed to find recovery node data for node " + "%u\n", done->node_idx); + dlm_put(dlm); + + mlog(0, "leaving reco data done handler, ret=%d\n", ret); + return ret; +} + +static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm, + struct list_head *list, + u8 dead_node) +{ + struct dlm_lock_resource *res, *next; + struct dlm_lock *lock; + + spin_lock(&dlm->spinlock); + list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) { + /* always prune any $RECOVERY entries for dead nodes, + * otherwise hangs can occur during later recovery */ + if (dlm_is_recovery_lock(res->lockname.name, + res->lockname.len)) { + spin_lock(&res->spinlock); + list_for_each_entry(lock, &res->granted, list) { + if (lock->ml.node == dead_node) { + mlog(0, "AHA! there was " + "a $RECOVERY lock for dead " + "node %u (%s)!\n", + dead_node, dlm->name); + list_del_init(&lock->list); + dlm_lock_put(lock); + /* Can't schedule DLM_UNLOCK_FREE_LOCK + * - do manually */ + dlm_lock_put(lock); + break; + } + } + spin_unlock(&res->spinlock); + continue; + } + + if (res->owner == dead_node) { + mlog(0, "found lockres owned by dead node while " + "doing recovery for node %u. sending it.\n", + dead_node); + list_move_tail(&res->recovering, list); + } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { + mlog(0, "found UNKNOWN owner while doing recovery " + "for node %u. sending it.\n", dead_node); + list_move_tail(&res->recovering, list); + } + } + spin_unlock(&dlm->spinlock); +} + +static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res) +{ + int total_locks = 0; + struct list_head *iter, *queue = &res->granted; + int i; + + for (i=0; i<3; i++) { + list_for_each(iter, queue) + total_locks++; + queue++; + } + return total_locks; +} + + +static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm, + struct dlm_migratable_lockres *mres, + u8 send_to, + struct dlm_lock_resource *res, + int total_locks) +{ + u64 mig_cookie = be64_to_cpu(mres->mig_cookie); + int mres_total_locks = be32_to_cpu(mres->total_locks); + int ret = 0, status = 0; + u8 orig_flags = mres->flags, + orig_master = mres->master; + + BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS); + if (!mres->num_locks) + return 0; + + /* add an all-done flag if we reached the last lock */ + orig_flags = mres->flags; + BUG_ON(total_locks > mres_total_locks); + if (total_locks == mres_total_locks) + mres->flags |= DLM_MRES_ALL_DONE; + + mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n", + dlm->name, res->lockname.len, res->lockname.name, + orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery", + send_to); + + /* send it */ + ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres, + struct_size(mres, ml, mres->num_locks), + send_to, &status); + if (ret < 0) { + /* XXX: negative status is not handled. + * this will end up killing this node. */ + mlog(ML_ERROR, "%s: res %.*s, Error %d send MIG_LOCKRES to " + "node %u (%s)\n", dlm->name, mres->lockname_len, + mres->lockname, ret, send_to, + (orig_flags & DLM_MRES_MIGRATION ? + "migration" : "recovery")); + } else { + /* might get an -ENOMEM back here */ + ret = status; + if (ret < 0) { + mlog_errno(ret); + + if (ret == -EFAULT) { + mlog(ML_ERROR, "node %u told me to kill " + "myself!\n", send_to); + BUG(); + } + } + } + + /* zero and reinit the message buffer */ + dlm_init_migratable_lockres(mres, res->lockname.name, + res->lockname.len, mres_total_locks, + mig_cookie, orig_flags, orig_master); + return ret; +} + +static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres, + const char *lockname, int namelen, + int total_locks, u64 cookie, + u8 flags, u8 master) +{ + /* mres here is one full page */ + clear_page(mres); + mres->lockname_len = namelen; + memcpy(mres->lockname, lockname, namelen); + mres->num_locks = 0; + mres->total_locks = cpu_to_be32(total_locks); + mres->mig_cookie = cpu_to_be64(cookie); + mres->flags = flags; + mres->master = master; +} + +static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock, + struct dlm_migratable_lockres *mres, + int queue) +{ + if (!lock->lksb) + return; + + /* Ignore lvb in all locks in the blocked list */ + if (queue == DLM_BLOCKED_LIST) + return; + + /* Only consider lvbs in locks with granted EX or PR lock levels */ + if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE) + return; + + if (dlm_lvb_is_empty(mres->lvb)) { + memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN); + return; + } + + /* Ensure the lvb copied for migration matches in other valid locks */ + if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN)) + return; + + mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, " + "node=%u\n", + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), + lock->lockres->lockname.len, lock->lockres->lockname.name, + lock->ml.node); + dlm_print_one_lock_resource(lock->lockres); + BUG(); +} + +/* returns 1 if this lock fills the network structure, + * 0 otherwise */ +static int dlm_add_lock_to_array(struct dlm_lock *lock, + struct dlm_migratable_lockres *mres, int queue) +{ + struct dlm_migratable_lock *ml; + int lock_num = mres->num_locks; + + ml = &(mres->ml[lock_num]); + ml->cookie = lock->ml.cookie; + ml->type = lock->ml.type; + ml->convert_type = lock->ml.convert_type; + ml->highest_blocked = lock->ml.highest_blocked; + ml->list = queue; + if (lock->lksb) { + ml->flags = lock->lksb->flags; + dlm_prepare_lvb_for_migration(lock, mres, queue); + } + ml->node = lock->ml.node; + mres->num_locks++; + /* we reached the max, send this network message */ + if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS) + return 1; + return 0; +} + +static void dlm_add_dummy_lock(struct dlm_ctxt *dlm, + struct dlm_migratable_lockres *mres) +{ + struct dlm_lock dummy; + memset(&dummy, 0, sizeof(dummy)); + dummy.ml.cookie = 0; + dummy.ml.type = LKM_IVMODE; + dummy.ml.convert_type = LKM_IVMODE; + dummy.ml.highest_blocked = LKM_IVMODE; + dummy.lksb = NULL; + dummy.ml.node = dlm->node_num; + dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST); +} + +static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm, + struct dlm_migratable_lock *ml, + u8 *nodenum) +{ + if (unlikely(ml->cookie == 0 && + ml->type == LKM_IVMODE && + ml->convert_type == LKM_IVMODE && + ml->highest_blocked == LKM_IVMODE && + ml->list == DLM_BLOCKED_LIST)) { + *nodenum = ml->node; + return 1; + } + return 0; +} + +int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, + struct dlm_migratable_lockres *mres, + u8 send_to, u8 flags) +{ + struct list_head *queue; + int total_locks, i; + u64 mig_cookie = 0; + struct dlm_lock *lock; + int ret = 0; + + BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); + + mlog(0, "sending to %u\n", send_to); + + total_locks = dlm_num_locks_in_lockres(res); + if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) { + /* rare, but possible */ + mlog(0, "argh. lockres has %d locks. this will " + "require more than one network packet to " + "migrate\n", total_locks); + mig_cookie = dlm_get_next_mig_cookie(); + } + + dlm_init_migratable_lockres(mres, res->lockname.name, + res->lockname.len, total_locks, + mig_cookie, flags, res->owner); + + total_locks = 0; + for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) { + queue = dlm_list_idx_to_ptr(res, i); + list_for_each_entry(lock, queue, list) { + /* add another lock. */ + total_locks++; + if (!dlm_add_lock_to_array(lock, mres, i)) + continue; + + /* this filled the lock message, + * we must send it immediately. */ + ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, + res, total_locks); + if (ret < 0) + goto error; + } + } + if (total_locks == 0) { + /* send a dummy lock to indicate a mastery reference only */ + mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n", + dlm->name, res->lockname.len, res->lockname.name, + send_to, flags & DLM_MRES_RECOVERY ? "recovery" : + "migration"); + dlm_add_dummy_lock(dlm, mres); + } + /* flush any remaining locks */ + ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks); + if (ret < 0) + goto error; + return ret; + +error: + mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n", + dlm->name, ret); + if (!dlm_is_host_down(ret)) + BUG(); + mlog(0, "%s: node %u went down while sending %s " + "lockres %.*s\n", dlm->name, send_to, + flags & DLM_MRES_RECOVERY ? "recovery" : "migration", + res->lockname.len, res->lockname.name); + return ret; +} + + + +/* + * this message will contain no more than one page worth of + * recovery data, and it will work on only one lockres. + * there may be many locks in this page, and we may need to wait + * for additional packets to complete all the locks (rare, but + * possible). + */ +/* + * NOTE: the allocation error cases here are scary + * we really cannot afford to fail an alloc in recovery + * do we spin? returning an error only delays the problem really + */ + +int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_migratable_lockres *mres = + (struct dlm_migratable_lockres *)msg->buf; + int ret = 0; + u8 real_master; + u8 extra_refs = 0; + char *buf = NULL; + struct dlm_work_item *item = NULL; + struct dlm_lock_resource *res = NULL; + unsigned int hash; + + if (!dlm_grab(dlm)) + return -EINVAL; + + if (mres->lockname_len > DLM_LOCKID_NAME_MAX || + mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS || + be16_to_cpu(msg->data_len) < struct_size(mres, ml, mres->num_locks)) { + mlog(ML_ERROR, "%s: invalid lockres migration message from %u\n", + dlm->name, mres->master); + dlm_put(dlm); + return -EINVAL; + } + + if (!dlm_joined(dlm)) { + mlog(ML_ERROR, "Domain %s not joined! " + "lockres %.*s, master %u\n", + dlm->name, mres->lockname_len, + mres->lockname, mres->master); + dlm_put(dlm); + return -EINVAL; + } + + BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); + + real_master = mres->master; + if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { + /* cannot migrate a lockres with no master */ + BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); + } + + mlog(0, "%s message received from node %u\n", + (mres->flags & DLM_MRES_RECOVERY) ? + "recovery" : "migration", mres->master); + if (mres->flags & DLM_MRES_ALL_DONE) + mlog(0, "all done flag. all lockres data received!\n"); + + ret = -ENOMEM; + buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS); + item = kzalloc_obj(*item, GFP_NOFS); + if (!buf || !item) + goto leave; + + /* lookup the lock to see if we have a secondary queue for this + * already... just add the locks in and this will have its owner + * and RECOVERY flag changed when it completes. */ + hash = dlm_lockid_hash(mres->lockname, mres->lockname_len); + spin_lock(&dlm->spinlock); + res = __dlm_lookup_lockres_full(dlm, mres->lockname, mres->lockname_len, + hash); + if (res) { + /* this will get a ref on res */ + /* mark it as recovering/migrating and hash it */ + spin_lock(&res->spinlock); + if (res->state & DLM_LOCK_RES_DROPPING_REF) { + mlog(0, "%s: node is attempting to migrate " + "lockres %.*s, but marked as dropping " + " ref!\n", dlm->name, + mres->lockname_len, mres->lockname); + ret = -EINVAL; + spin_unlock(&res->spinlock); + spin_unlock(&dlm->spinlock); + dlm_lockres_put(res); + goto leave; + } + + if (mres->flags & DLM_MRES_RECOVERY) { + res->state |= DLM_LOCK_RES_RECOVERING; + } else { + if (res->state & DLM_LOCK_RES_MIGRATING) { + /* this is at least the second + * lockres message */ + mlog(0, "lock %.*s is already migrating\n", + mres->lockname_len, + mres->lockname); + } else if (res->state & DLM_LOCK_RES_RECOVERING) { + /* caller should BUG */ + mlog(ML_ERROR, "node is attempting to migrate " + "lock %.*s, but marked as recovering!\n", + mres->lockname_len, mres->lockname); + ret = -EFAULT; + spin_unlock(&res->spinlock); + spin_unlock(&dlm->spinlock); + dlm_lockres_put(res); + goto leave; + } + res->state |= DLM_LOCK_RES_MIGRATING; + } + spin_unlock(&res->spinlock); + spin_unlock(&dlm->spinlock); + } else { + spin_unlock(&dlm->spinlock); + /* need to allocate, just like if it was + * mastered here normally */ + res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len); + if (!res) + goto leave; + + /* to match the ref that we would have gotten if + * dlm_lookup_lockres had succeeded */ + dlm_lockres_get(res); + + /* mark it as recovering/migrating and hash it */ + if (mres->flags & DLM_MRES_RECOVERY) + res->state |= DLM_LOCK_RES_RECOVERING; + else + res->state |= DLM_LOCK_RES_MIGRATING; + + spin_lock(&dlm->spinlock); + __dlm_insert_lockres(dlm, res); + spin_unlock(&dlm->spinlock); + + /* Add an extra ref for this lock-less lockres lest the + * dlm_thread purges it before we get the chance to add + * locks to it */ + dlm_lockres_get(res); + + /* There are three refs that need to be put. + * 1. Taken above. + * 2. kref_init in dlm_new_lockres()->dlm_init_lockres(). + * 3. dlm_lookup_lockres() + * The first one is handled at the end of this function. The + * other two are handled in the worker thread after locks have + * been attached. Yes, we don't wait for purge time to match + * kref_init. The lockres will still have at least one ref + * added because it is in the hash __dlm_insert_lockres() */ + extra_refs++; + + /* now that the new lockres is inserted, + * make it usable by other processes */ + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_IN_PROGRESS; + spin_unlock(&res->spinlock); + wake_up(&res->wq); + } + + /* at this point we have allocated everything we need, + * and we have a hashed lockres with an extra ref and + * the proper res->state flags. */ + ret = 0; + spin_lock(&res->spinlock); + /* drop this either when master requery finds a different master + * or when a lock is added by the recovery worker */ + dlm_lockres_grab_inflight_ref(dlm, res); + if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) { + /* migration cannot have an unknown master */ + BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); + mlog(0, "recovery has passed me a lockres with an " + "unknown owner.. will need to requery: " + "%.*s\n", mres->lockname_len, mres->lockname); + } else { + /* take a reference now to pin the lockres, drop it + * when locks are added in the worker */ + dlm_change_lockres_owner(dlm, res, dlm->node_num); + } + spin_unlock(&res->spinlock); + + /* queue up work for dlm_mig_lockres_worker */ + dlm_grab(dlm); /* get an extra ref for the work item */ + memcpy(buf, msg->buf, be16_to_cpu(msg->data_len)); /* copy the whole message */ + dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf); + item->u.ml.lockres = res; /* already have a ref */ + item->u.ml.real_master = real_master; + item->u.ml.extra_ref = extra_refs; + spin_lock(&dlm->work_lock); + list_add_tail(&item->list, &dlm->work_list); + spin_unlock(&dlm->work_lock); + queue_work(dlm->dlm_worker, &dlm->dispatched_work); + +leave: + /* One extra ref taken needs to be put here */ + if (extra_refs) + dlm_lockres_put(res); + + dlm_put(dlm); + if (ret < 0) { + kfree(buf); + kfree(item); + mlog_errno(ret); + } + + return ret; +} + + +static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data) +{ + struct dlm_ctxt *dlm; + struct dlm_migratable_lockres *mres; + int ret = 0; + struct dlm_lock_resource *res; + u8 real_master; + u8 extra_ref; + + dlm = item->dlm; + mres = (struct dlm_migratable_lockres *)data; + + res = item->u.ml.lockres; + real_master = item->u.ml.real_master; + extra_ref = item->u.ml.extra_ref; + + if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { + /* this case is super-rare. only occurs if + * node death happens during migration. */ +again: + ret = dlm_lockres_master_requery(dlm, res, &real_master); + if (ret < 0) { + mlog(0, "dlm_lockres_master_requery ret=%d\n", + ret); + goto again; + } + if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { + mlog(0, "lockres %.*s not claimed. " + "this node will take it.\n", + res->lockname.len, res->lockname.name); + } else { + spin_lock(&res->spinlock); + dlm_lockres_drop_inflight_ref(dlm, res); + spin_unlock(&res->spinlock); + mlog(0, "master needs to respond to sender " + "that node %u still owns %.*s\n", + real_master, res->lockname.len, + res->lockname.name); + /* cannot touch this lockres */ + goto leave; + } + } + + ret = dlm_process_recovery_data(dlm, res, mres); + if (ret < 0) + mlog(0, "dlm_process_recovery_data returned %d\n", ret); + else + mlog(0, "dlm_process_recovery_data succeeded\n"); + + if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) == + (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) { + ret = dlm_finish_migration(dlm, res, mres->master); + if (ret < 0) + mlog_errno(ret); + } + +leave: + /* See comment in dlm_mig_lockres_handler() */ + if (res) { + if (extra_ref) + dlm_lockres_put(res); + dlm_lockres_put(res); + } + kfree(data); +} + + + +static int dlm_lockres_master_requery(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + u8 *real_master) +{ + struct dlm_node_iter iter; + int nodenum; + int ret = 0; + + *real_master = DLM_LOCK_RES_OWNER_UNKNOWN; + + /* we only reach here if one of the two nodes in a + * migration died while the migration was in progress. + * at this point we need to requery the master. we + * know that the new_master got as far as creating + * an mle on at least one node, but we do not know + * if any nodes had actually cleared the mle and set + * the master to the new_master. the old master + * is supposed to set the owner to UNKNOWN in the + * event of a new_master death, so the only possible + * responses that we can get from nodes here are + * that the master is new_master, or that the master + * is UNKNOWN. + * if all nodes come back with UNKNOWN then we know + * the lock needs remastering here. + * if any node comes back with a valid master, check + * to see if that master is the one that we are + * recovering. if so, then the new_master died and + * we need to remaster this lock. if not, then the + * new_master survived and that node will respond to + * other nodes about the owner. + * if there is an owner, this node needs to dump this + * lockres and alert the sender that this lockres + * was rejected. */ + spin_lock(&dlm->spinlock); + dlm_node_iter_init(dlm->domain_map, &iter); + spin_unlock(&dlm->spinlock); + + while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { + /* do not send to self */ + if (nodenum == dlm->node_num) + continue; + ret = dlm_do_master_requery(dlm, res, nodenum, real_master); + if (ret < 0) { + mlog_errno(ret); + if (!dlm_is_host_down(ret)) + BUG(); + /* host is down, so answer for that node would be + * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */ + } + if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) { + mlog(0, "lock master is %u\n", *real_master); + break; + } + } + return ret; +} + + +int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, + u8 nodenum, u8 *real_master) +{ + int ret; + struct dlm_master_requery req; + int status = DLM_LOCK_RES_OWNER_UNKNOWN; + + memset(&req, 0, sizeof(req)); + req.node_idx = dlm->node_num; + req.namelen = res->lockname.len; + memcpy(req.name, res->lockname.name, res->lockname.len); + +resend: + ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key, + &req, sizeof(req), nodenum, &status); + if (ret < 0) + mlog(ML_ERROR, "Error %d when sending message %u (key " + "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG, + dlm->key, nodenum); + else if (status == -ENOMEM) { + mlog_errno(status); + msleep(50); + goto resend; + } else { + BUG_ON(status < 0); + BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN); + *real_master = (u8) (status & 0xff); + mlog(0, "node %u responded to master requery with %u\n", + nodenum, *real_master); + ret = 0; + } + return ret; +} + + +/* this function cannot error, so unless the sending + * or receiving of the message failed, the owner can + * be trusted */ +int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf; + struct dlm_lock_resource *res = NULL; + unsigned int hash; + int master = DLM_LOCK_RES_OWNER_UNKNOWN; + u32 flags = DLM_ASSERT_MASTER_REQUERY; + int dispatched = 0; + + if (!dlm_grab(dlm)) { + /* since the domain has gone away on this + * node, the proper response is UNKNOWN */ + return master; + } + + hash = dlm_lockid_hash(req->name, req->namelen); + + spin_lock(&dlm->spinlock); + res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash); + if (res) { + spin_lock(&res->spinlock); + master = res->owner; + if (master == dlm->node_num) { + int ret = dlm_dispatch_assert_master(dlm, res, + 0, 0, flags); + if (ret < 0) { + mlog_errno(ret); + spin_unlock(&res->spinlock); + dlm_lockres_put(res); + spin_unlock(&dlm->spinlock); + dlm_put(dlm); + /* sender will take care of this and retry */ + return ret; + } else { + dispatched = 1; + __dlm_lockres_grab_inflight_worker(dlm, res); + spin_unlock(&res->spinlock); + } + } else { + /* put.. in case we are not the master */ + spin_unlock(&res->spinlock); + dlm_lockres_put(res); + } + } + spin_unlock(&dlm->spinlock); + + if (!dispatched) + dlm_put(dlm); + return master; +} + +static inline struct list_head * +dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num) +{ + struct list_head *ret; + BUG_ON(list_num < 0); + BUG_ON(list_num > 2); + ret = &(res->granted); + ret += list_num; + return ret; +} +/* TODO: do ast flush business + * TODO: do MIGRATING and RECOVERING spinning + */ + +/* +* NOTE about in-flight requests during migration: +* +* Before attempting the migrate, the master has marked the lockres as +* MIGRATING and then flushed all of its pending ASTS. So any in-flight +* requests either got queued before the MIGRATING flag got set, in which +* case the lock data will reflect the change and a return message is on +* the way, or the request failed to get in before MIGRATING got set. In +* this case, the caller will be told to spin and wait for the MIGRATING +* flag to be dropped, then recheck the master. +* This holds true for the convert, cancel and unlock cases, and since lvb +* updates are tied to these same messages, it applies to lvb updates as +* well. For the lock case, there is no way a lock can be on the master +* queue and not be on the secondary queue since the lock is always added +* locally first. This means that the new target node will never be sent +* a lock that he doesn't already have on the list. +* In total, this means that the local lock is correct and should not be +* updated to match the one sent by the master. Any messages sent back +* from the master before the MIGRATING flag will bring the lock properly +* up-to-date, and the change will be ordered properly for the waiter. +* We will *not* attempt to modify the lock underneath the waiter. +*/ + +static int dlm_process_recovery_data(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_migratable_lockres *mres) +{ + struct dlm_migratable_lock *ml; + struct list_head *queue, *iter; + struct list_head *tmpq = NULL; + struct dlm_lock *newlock = NULL; + struct dlm_lockstatus *lksb = NULL; + int ret = 0; + int i, j, bad; + struct dlm_lock *lock; + u8 from = O2NM_MAX_NODES; + __be64 c; + + mlog(0, "running %d locks for this lockres\n", mres->num_locks); + for (i=0; inum_locks; i++) { + ml = &(mres->ml[i]); + + if (dlm_is_dummy_lock(dlm, ml, &from)) { + /* placeholder, just need to set the refmap bit */ + BUG_ON(mres->num_locks != 1); + mlog(0, "%s:%.*s: dummy lock for %u\n", + dlm->name, mres->lockname_len, mres->lockname, + from); + spin_lock(&res->spinlock); + dlm_lockres_set_refmap_bit(dlm, res, from); + spin_unlock(&res->spinlock); + break; + } + BUG_ON(ml->highest_blocked != LKM_IVMODE); + newlock = NULL; + lksb = NULL; + + queue = dlm_list_num_to_pointer(res, ml->list); + tmpq = NULL; + + /* if the lock is for the local node it needs to + * be moved to the proper location within the queue. + * do not allocate a new lock structure. */ + if (ml->node == dlm->node_num) { + /* MIGRATION ONLY! */ + BUG_ON(!(mres->flags & DLM_MRES_MIGRATION)); + + lock = NULL; + spin_lock(&res->spinlock); + for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) { + tmpq = dlm_list_idx_to_ptr(res, j); + list_for_each(iter, tmpq) { + lock = list_entry(iter, + struct dlm_lock, list); + if (lock->ml.cookie == ml->cookie) + break; + lock = NULL; + } + if (lock) + break; + } + + /* lock is always created locally first, and + * destroyed locally last. it must be on the list */ + if (!lock) { + c = ml->cookie; + mlog(ML_ERROR, "Could not find local lock " + "with cookie %u:%llu, node %u, " + "list %u, flags 0x%x, type %d, " + "conv %d, highest blocked %d\n", + dlm_get_lock_cookie_node(be64_to_cpu(c)), + dlm_get_lock_cookie_seq(be64_to_cpu(c)), + ml->node, ml->list, ml->flags, ml->type, + ml->convert_type, ml->highest_blocked); + __dlm_print_one_lock_resource(res); + BUG(); + } + + if (lock->ml.node != ml->node) { + c = lock->ml.cookie; + mlog(ML_ERROR, "Mismatched node# in lock " + "cookie %u:%llu, name %.*s, node %u\n", + dlm_get_lock_cookie_node(be64_to_cpu(c)), + dlm_get_lock_cookie_seq(be64_to_cpu(c)), + res->lockname.len, res->lockname.name, + lock->ml.node); + c = ml->cookie; + mlog(ML_ERROR, "Migrate lock cookie %u:%llu, " + "node %u, list %u, flags 0x%x, type %d, " + "conv %d, highest blocked %d\n", + dlm_get_lock_cookie_node(be64_to_cpu(c)), + dlm_get_lock_cookie_seq(be64_to_cpu(c)), + ml->node, ml->list, ml->flags, ml->type, + ml->convert_type, ml->highest_blocked); + __dlm_print_one_lock_resource(res); + BUG(); + } + + if (tmpq != queue) { + c = ml->cookie; + mlog(0, "Lock cookie %u:%llu was on list %u " + "instead of list %u for %.*s\n", + dlm_get_lock_cookie_node(be64_to_cpu(c)), + dlm_get_lock_cookie_seq(be64_to_cpu(c)), + j, ml->list, res->lockname.len, + res->lockname.name); + __dlm_print_one_lock_resource(res); + spin_unlock(&res->spinlock); + continue; + } + + /* see NOTE above about why we do not update + * to match the master here */ + + /* move the lock to its proper place */ + /* do not alter lock refcount. switching lists. */ + list_move_tail(&lock->list, queue); + spin_unlock(&res->spinlock); + + mlog(0, "just reordered a local lock!\n"); + continue; + } + + /* lock is for another node. */ + newlock = dlm_new_lock(ml->type, ml->node, + be64_to_cpu(ml->cookie), NULL); + if (!newlock) { + ret = -ENOMEM; + goto leave; + } + lksb = newlock->lksb; + dlm_lock_attach_lockres(newlock, res); + + if (ml->convert_type != LKM_IVMODE) { + BUG_ON(queue != &res->converting); + newlock->ml.convert_type = ml->convert_type; + } + lksb->flags |= (ml->flags & + (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB)); + + if (ml->type == LKM_NLMODE) + goto skip_lvb; + + /* + * If the lock is in the blocked list it can't have a valid lvb, + * so skip it + */ + if (ml->list == DLM_BLOCKED_LIST) + goto skip_lvb; + + if (!dlm_lvb_is_empty(mres->lvb)) { + if (lksb->flags & DLM_LKSB_PUT_LVB) { + /* other node was trying to update + * lvb when node died. recreate the + * lksb with the updated lvb. */ + memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN); + /* the lock resource lvb update must happen + * NOW, before the spinlock is dropped. + * we no longer wait for the AST to update + * the lvb. */ + memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); + } else { + /* otherwise, the node is sending its + * most recent valid lvb info */ + BUG_ON(ml->type != LKM_EXMODE && + ml->type != LKM_PRMODE); + if (!dlm_lvb_is_empty(res->lvb) && + (ml->type == LKM_EXMODE || + memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) { + int i; + mlog(ML_ERROR, "%s:%.*s: received bad " + "lvb! type=%d\n", dlm->name, + res->lockname.len, + res->lockname.name, ml->type); + printk("lockres lvb=["); + for (i=0; ilvb[i]); + printk("]\nmigrated lvb=["); + for (i=0; ilvb[i]); + printk("]\n"); + dlm_print_one_lock_resource(res); + BUG(); + } + memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); + } + } +skip_lvb: + + /* NOTE: + * wrt lock queue ordering and recovery: + * 1. order of locks on granted queue is + * meaningless. + * 2. order of locks on converting queue is + * LOST with the node death. sorry charlie. + * 3. order of locks on the blocked queue is + * also LOST. + * order of locks does not affect integrity, it + * just means that a lock request may get pushed + * back in line as a result of the node death. + * also note that for a given node the lock order + * for its secondary queue locks is preserved + * relative to each other, but clearly *not* + * preserved relative to locks from other nodes. + */ + bad = 0; + spin_lock(&res->spinlock); + list_for_each_entry(lock, queue, list) { + if (lock->ml.cookie == ml->cookie) { + c = lock->ml.cookie; + mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already " + "exists on this lockres!\n", dlm->name, + res->lockname.len, res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(c)), + dlm_get_lock_cookie_seq(be64_to_cpu(c))); + + mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, " + "node=%u, cookie=%u:%llu, queue=%d\n", + ml->type, ml->convert_type, ml->node, + dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)), + ml->list); + + __dlm_print_one_lock_resource(res); + bad = 1; + break; + } + } + if (!bad) { + dlm_lock_get(newlock); + if (mres->flags & DLM_MRES_RECOVERY && + ml->list == DLM_CONVERTING_LIST && + newlock->ml.type > + newlock->ml.convert_type) { + /* newlock is doing downconvert, add it to the + * head of converting list */ + list_add(&newlock->list, queue); + } else + list_add_tail(&newlock->list, queue); + mlog(0, "%s:%.*s: added lock for node %u, " + "setting refmap bit\n", dlm->name, + res->lockname.len, res->lockname.name, ml->node); + dlm_lockres_set_refmap_bit(dlm, res, ml->node); + } + spin_unlock(&res->spinlock); + } + mlog(0, "done running all the locks\n"); + +leave: + /* balance the ref taken when the work was queued */ + spin_lock(&res->spinlock); + dlm_lockres_drop_inflight_ref(dlm, res); + spin_unlock(&res->spinlock); + + if (ret < 0) + mlog_errno(ret); + + return ret; +} + +void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + int i; + struct list_head *queue; + struct dlm_lock *lock, *next; + + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&res->spinlock); + res->state |= DLM_LOCK_RES_RECOVERING; + if (!list_empty(&res->recovering)) { + mlog(0, + "Recovering res %s:%.*s, is already on recovery list!\n", + dlm->name, res->lockname.len, res->lockname.name); + list_del_init(&res->recovering); + dlm_lockres_put(res); + } + /* We need to hold a reference while on the recovery list */ + dlm_lockres_get(res); + list_add_tail(&res->recovering, &dlm->reco.resources); + + /* find any pending locks and put them back on proper list */ + for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) { + queue = dlm_list_idx_to_ptr(res, i); + list_for_each_entry_safe(lock, next, queue, list) { + dlm_lock_get(lock); + if (lock->convert_pending) { + /* move converting lock back to granted */ + mlog(0, "node died with convert pending " + "on %.*s. move back to granted list.\n", + res->lockname.len, res->lockname.name); + dlm_revert_pending_convert(res, lock); + lock->convert_pending = 0; + } else if (lock->lock_pending) { + /* remove pending lock requests completely */ + BUG_ON(i != DLM_BLOCKED_LIST); + mlog(0, "node died with lock pending " + "on %.*s. remove from blocked list and skip.\n", + res->lockname.len, res->lockname.name); + /* lock will be floating until ref in + * dlmlock_remote is freed after the network + * call returns. ok for it to not be on any + * list since no ast can be called + * (the master is dead). */ + dlm_revert_pending_lock(res, lock); + lock->lock_pending = 0; + } else if (lock->unlock_pending) { + /* if an unlock was in progress, treat as + * if this had completed successfully + * before sending this lock state to the + * new master. note that the dlm_unlock + * call is still responsible for calling + * the unlockast. that will happen after + * the network call times out. for now, + * just move lists to prepare the new + * recovery master. */ + BUG_ON(i != DLM_GRANTED_LIST); + mlog(0, "node died with unlock pending " + "on %.*s. remove from blocked list and skip.\n", + res->lockname.len, res->lockname.name); + dlm_commit_pending_unlock(res, lock); + lock->unlock_pending = 0; + } else if (lock->cancel_pending) { + /* if a cancel was in progress, treat as + * if this had completed successfully + * before sending this lock state to the + * new master */ + BUG_ON(i != DLM_CONVERTING_LIST); + mlog(0, "node died with cancel pending " + "on %.*s. move back to granted list.\n", + res->lockname.len, res->lockname.name); + dlm_commit_pending_cancel(res, lock); + lock->cancel_pending = 0; + } + dlm_lock_put(lock); + } + } +} + + + +/* removes all recovered locks from the recovery list. + * sets the res->owner to the new master. + * unsets the RECOVERY flag and wakes waiters. */ +static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm, + u8 dead_node, u8 new_master) +{ + int i; + struct hlist_head *bucket; + struct dlm_lock_resource *res, *next; + + assert_spin_locked(&dlm->spinlock); + + list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) { + if (res->owner == dead_node) { + mlog(0, "%s: res %.*s, Changing owner from %u to %u\n", + dlm->name, res->lockname.len, res->lockname.name, + res->owner, new_master); + list_del_init(&res->recovering); + spin_lock(&res->spinlock); + /* new_master has our reference from + * the lock state sent during recovery */ + dlm_change_lockres_owner(dlm, res, new_master); + res->state &= ~DLM_LOCK_RES_RECOVERING; + if (__dlm_lockres_has_locks(res)) + __dlm_dirty_lockres(dlm, res); + spin_unlock(&res->spinlock); + wake_up(&res->wq); + dlm_lockres_put(res); + } + } + + /* this will become unnecessary eventually, but + * for now we need to run the whole hash, clear + * the RECOVERING state and set the owner + * if necessary */ + for (i = 0; i < DLM_HASH_BUCKETS; i++) { + bucket = dlm_lockres_hash(dlm, i); + hlist_for_each_entry(res, bucket, hash_node) { + if (res->state & DLM_LOCK_RES_RECOVERY_WAITING) { + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_RECOVERY_WAITING; + spin_unlock(&res->spinlock); + wake_up(&res->wq); + } + + if (!(res->state & DLM_LOCK_RES_RECOVERING)) + continue; + + if (res->owner != dead_node && + res->owner != dlm->node_num) + continue; + + if (!list_empty(&res->recovering)) { + list_del_init(&res->recovering); + dlm_lockres_put(res); + } + + /* new_master has our reference from + * the lock state sent during recovery */ + mlog(0, "%s: res %.*s, Changing owner from %u to %u\n", + dlm->name, res->lockname.len, res->lockname.name, + res->owner, new_master); + spin_lock(&res->spinlock); + dlm_change_lockres_owner(dlm, res, new_master); + res->state &= ~DLM_LOCK_RES_RECOVERING; + if (__dlm_lockres_has_locks(res)) + __dlm_dirty_lockres(dlm, res); + spin_unlock(&res->spinlock); + wake_up(&res->wq); + } + } +} + +static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local) +{ + if (local) { + if (lock->ml.type != LKM_EXMODE && + lock->ml.type != LKM_PRMODE) + return 1; + } else if (lock->ml.type == LKM_EXMODE) + return 1; + return 0; +} + +static void dlm_revalidate_lvb(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, u8 dead_node) +{ + struct list_head *queue; + struct dlm_lock *lock; + int blank_lvb = 0, local = 0; + int i; + u8 search_node; + + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&res->spinlock); + + if (res->owner == dlm->node_num) + /* if this node owned the lockres, and if the dead node + * had an EX when he died, blank out the lvb */ + search_node = dead_node; + else { + /* if this is a secondary lockres, and we had no EX or PR + * locks granted, we can no longer trust the lvb */ + search_node = dlm->node_num; + local = 1; /* check local state for valid lvb */ + } + + for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) { + queue = dlm_list_idx_to_ptr(res, i); + list_for_each_entry(lock, queue, list) { + if (lock->ml.node == search_node) { + if (dlm_lvb_needs_invalidation(lock, local)) { + /* zero the lksb lvb and lockres lvb */ + blank_lvb = 1; + memset(lock->lksb->lvb, 0, DLM_LVB_LEN); + } + } + } + } + + if (blank_lvb) { + mlog(0, "clearing %.*s lvb, dead node %u had EX\n", + res->lockname.len, res->lockname.name, dead_node); + memset(res->lvb, 0, DLM_LVB_LEN); + } +} + +static void dlm_free_dead_locks(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, u8 dead_node) +{ + struct dlm_lock *lock, *next; + unsigned int freed = 0; + + /* this node is the lockres master: + * 1) remove any stale locks for the dead node + * 2) if the dead node had an EX when he died, blank out the lvb + */ + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&res->spinlock); + + /* We do two dlm_lock_put(). One for removing from list and the other is + * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */ + + /* TODO: check pending_asts, pending_basts here */ + list_for_each_entry_safe(lock, next, &res->granted, list) { + if (lock->ml.node == dead_node) { + list_del_init(&lock->list); + dlm_lock_put(lock); + /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ + dlm_lock_put(lock); + freed++; + } + } + list_for_each_entry_safe(lock, next, &res->converting, list) { + if (lock->ml.node == dead_node) { + list_del_init(&lock->list); + dlm_lock_put(lock); + /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ + dlm_lock_put(lock); + freed++; + } + } + list_for_each_entry_safe(lock, next, &res->blocked, list) { + if (lock->ml.node == dead_node) { + list_del_init(&lock->list); + dlm_lock_put(lock); + /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ + dlm_lock_put(lock); + freed++; + } + } + + if (freed) { + mlog(0, "%s:%.*s: freed %u locks for dead node %u, " + "dropping ref from lockres\n", dlm->name, + res->lockname.len, res->lockname.name, freed, dead_node); + if(!test_bit(dead_node, res->refmap)) { + mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, " + "but ref was not set\n", dlm->name, + res->lockname.len, res->lockname.name, freed, dead_node); + __dlm_print_one_lock_resource(res); + } + res->state |= DLM_LOCK_RES_RECOVERY_WAITING; + dlm_lockres_clear_refmap_bit(dlm, res, dead_node); + } else if (test_bit(dead_node, res->refmap)) { + mlog(0, "%s:%.*s: dead node %u had a ref, but had " + "no locks and had not purged before dying\n", dlm->name, + res->lockname.len, res->lockname.name, dead_node); + dlm_lockres_clear_refmap_bit(dlm, res, dead_node); + } + + /* do not kick thread yet */ + __dlm_dirty_lockres(dlm, res); +} + +static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node) +{ + struct dlm_lock_resource *res; + int i; + struct hlist_head *bucket; + struct hlist_node *tmp; + struct dlm_lock *lock; + + + /* purge any stale mles */ + dlm_clean_master_list(dlm, dead_node); + + /* + * now clean up all lock resources. there are two rules: + * + * 1) if the dead node was the master, move the lockres + * to the recovering list. set the RECOVERING flag. + * this lockres needs to be cleaned up before it can + * be used further. + * + * 2) if this node was the master, remove all locks from + * each of the lockres queues that were owned by the + * dead node. once recovery finishes, the dlm thread + * can be kicked again to see if any ASTs or BASTs + * need to be fired as a result. + */ + for (i = 0; i < DLM_HASH_BUCKETS; i++) { + bucket = dlm_lockres_hash(dlm, i); + hlist_for_each_entry_safe(res, tmp, bucket, hash_node) { + /* always prune any $RECOVERY entries for dead nodes, + * otherwise hangs can occur during later recovery */ + if (dlm_is_recovery_lock(res->lockname.name, + res->lockname.len)) { + spin_lock(&res->spinlock); + list_for_each_entry(lock, &res->granted, list) { + if (lock->ml.node == dead_node) { + mlog(0, "AHA! there was " + "a $RECOVERY lock for dead " + "node %u (%s)!\n", + dead_node, dlm->name); + list_del_init(&lock->list); + dlm_lock_put(lock); + /* Can't schedule + * DLM_UNLOCK_FREE_LOCK + * - do manually */ + dlm_lock_put(lock); + break; + } + } + + if ((res->owner == dead_node) && + (res->state & DLM_LOCK_RES_DROPPING_REF)) { + dlm_lockres_get(res); + __dlm_do_purge_lockres(dlm, res); + spin_unlock(&res->spinlock); + wake_up(&res->wq); + dlm_lockres_put(res); + continue; + } else if (res->owner == dlm->node_num) + dlm_lockres_clear_refmap_bit(dlm, res, dead_node); + spin_unlock(&res->spinlock); + continue; + } + spin_lock(&res->spinlock); + /* zero the lvb if necessary */ + dlm_revalidate_lvb(dlm, res, dead_node); + if (res->owner == dead_node) { + if (res->state & DLM_LOCK_RES_DROPPING_REF) { + mlog(0, "%s:%.*s: owned by " + "dead node %u, this node was " + "dropping its ref when master died. " + "continue, purging the lockres.\n", + dlm->name, res->lockname.len, + res->lockname.name, dead_node); + dlm_lockres_get(res); + __dlm_do_purge_lockres(dlm, res); + spin_unlock(&res->spinlock); + wake_up(&res->wq); + dlm_lockres_put(res); + continue; + } + dlm_move_lockres_to_recovery_list(dlm, res); + } else if (res->owner == dlm->node_num) { + dlm_free_dead_locks(dlm, res, dead_node); + __dlm_lockres_calc_usage(dlm, res); + } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { + if (test_bit(dead_node, res->refmap)) { + mlog(0, "%s:%.*s: dead node %u had a ref, but had " + "no locks and had not purged before dying\n", + dlm->name, res->lockname.len, + res->lockname.name, dead_node); + dlm_lockres_clear_refmap_bit(dlm, res, dead_node); + } + } + spin_unlock(&res->spinlock); + } + } + +} + +static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx) +{ + assert_spin_locked(&dlm->spinlock); + + if (dlm->reco.new_master == idx) { + mlog(0, "%s: recovery master %d just died\n", + dlm->name, idx); + if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { + /* finalize1 was reached, so it is safe to clear + * the new_master and dead_node. that recovery + * is complete. */ + mlog(0, "%s: dead master %d had reached " + "finalize1 state, clearing\n", dlm->name, idx); + dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; + __dlm_reset_recovery(dlm); + } + } + + /* Clean up join state on node death. */ + if (dlm->joining_node == idx) { + mlog(0, "Clearing join state for node %u\n", idx); + __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); + } + + /* check to see if the node is already considered dead */ + if (!test_bit(idx, dlm->live_nodes_map)) { + mlog(0, "for domain %s, node %d is already dead. " + "another node likely did recovery already.\n", + dlm->name, idx); + return; + } + + /* check to see if we do not care about this node */ + if (!test_bit(idx, dlm->domain_map)) { + /* This also catches the case that we get a node down + * but haven't joined the domain yet. */ + mlog(0, "node %u already removed from domain!\n", idx); + return; + } + + clear_bit(idx, dlm->live_nodes_map); + + /* make sure local cleanup occurs before the heartbeat events */ + if (!test_bit(idx, dlm->recovery_map)) + dlm_do_local_recovery_cleanup(dlm, idx); + + /* notify anything attached to the heartbeat events */ + dlm_hb_event_notify_attached(dlm, idx, 0); + + mlog(0, "node %u being removed from domain map!\n", idx); + clear_bit(idx, dlm->domain_map); + clear_bit(idx, dlm->exit_domain_map); + /* wake up migration waiters if a node goes down. + * perhaps later we can genericize this for other waiters. */ + wake_up(&dlm->migration_wq); + + set_bit(idx, dlm->recovery_map); +} + +void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data) +{ + struct dlm_ctxt *dlm = data; + + if (!dlm_grab(dlm)) + return; + + /* + * This will notify any dlm users that a node in our domain + * went away without notifying us first. + */ + if (test_bit(idx, dlm->domain_map)) + dlm_fire_domain_eviction_callbacks(dlm, idx); + + spin_lock(&dlm->spinlock); + __dlm_hb_node_down(dlm, idx); + spin_unlock(&dlm->spinlock); + + dlm_put(dlm); +} + +void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data) +{ + struct dlm_ctxt *dlm = data; + + if (!dlm_grab(dlm)) + return; + + spin_lock(&dlm->spinlock); + set_bit(idx, dlm->live_nodes_map); + /* do NOT notify mle attached to the heartbeat events. + * new nodes are not interesting in mastery until joined. */ + spin_unlock(&dlm->spinlock); + + dlm_put(dlm); +} + +static void dlm_reco_ast(void *astdata) +{ + struct dlm_ctxt *dlm = astdata; + mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n", + dlm->node_num, dlm->name); +} +static void dlm_reco_bast(void *astdata, int blocked_type) +{ + struct dlm_ctxt *dlm = astdata; + mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n", + dlm->node_num, dlm->name); +} +static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st) +{ + mlog(0, "unlockast for recovery lock fired!\n"); +} + +/* + * dlm_pick_recovery_master will continually attempt to use + * dlmlock() on the special "$RECOVERY" lockres with the + * LKM_NOQUEUE flag to get an EX. every thread that enters + * this function on each node racing to become the recovery + * master will not stop attempting this until either: + * a) this node gets the EX (and becomes the recovery master), + * or b) dlm->reco.new_master gets set to some nodenum + * != O2NM_INVALID_NODE_NUM (another node will do the reco). + * so each time a recovery master is needed, the entire cluster + * will sync at this point. if the new master dies, that will + * be detected in dlm_do_recovery */ +static int dlm_pick_recovery_master(struct dlm_ctxt *dlm) +{ + enum dlm_status ret; + struct dlm_lockstatus lksb; + int status = -EINVAL; + + mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n", + dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num); +again: + memset(&lksb, 0, sizeof(lksb)); + + ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY, + DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN, + dlm_reco_ast, dlm, dlm_reco_bast); + + mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n", + dlm->name, ret, lksb.status); + + if (ret == DLM_NORMAL) { + mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n", + dlm->name, dlm->node_num); + + /* got the EX lock. check to see if another node + * just became the reco master */ + if (dlm_reco_master_ready(dlm)) { + mlog(0, "%s: got reco EX lock, but %u will " + "do the recovery\n", dlm->name, + dlm->reco.new_master); + status = -EEXIST; + } else { + status = 0; + + /* see if recovery was already finished elsewhere */ + spin_lock(&dlm->spinlock); + if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { + status = -EINVAL; + mlog(0, "%s: got reco EX lock, but " + "node got recovered already\n", dlm->name); + if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { + mlog(ML_ERROR, "%s: new master is %u " + "but no dead node!\n", + dlm->name, dlm->reco.new_master); + BUG(); + } + } + spin_unlock(&dlm->spinlock); + } + + /* if this node has actually become the recovery master, + * set the master and send the messages to begin recovery */ + if (!status) { + mlog(0, "%s: dead=%u, this=%u, sending " + "begin_reco now\n", dlm->name, + dlm->reco.dead_node, dlm->node_num); + status = dlm_send_begin_reco_message(dlm, + dlm->reco.dead_node); + /* this always succeeds */ + BUG_ON(status); + + /* set the new_master to this node */ + spin_lock(&dlm->spinlock); + dlm_set_reco_master(dlm, dlm->node_num); + spin_unlock(&dlm->spinlock); + } + + /* recovery lock is a special case. ast will not get fired, + * so just go ahead and unlock it. */ + ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm); + if (ret == DLM_DENIED) { + mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n"); + ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm); + } + if (ret != DLM_NORMAL) { + /* this would really suck. this could only happen + * if there was a network error during the unlock + * because of node death. this means the unlock + * is actually "done" and the lock structure is + * even freed. we can continue, but only + * because this specific lock name is special. */ + mlog(ML_ERROR, "dlmunlock returned %d\n", ret); + } + } else if (ret == DLM_NOTQUEUED) { + mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n", + dlm->name, dlm->node_num); + /* another node is master. wait on + * reco.new_master != O2NM_INVALID_NODE_NUM + * for at most one second */ + wait_event_timeout(dlm->dlm_reco_thread_wq, + dlm_reco_master_ready(dlm), + msecs_to_jiffies(1000)); + if (!dlm_reco_master_ready(dlm)) { + mlog(0, "%s: reco master taking a while\n", + dlm->name); + goto again; + } + /* another node has informed this one that it is reco master */ + mlog(0, "%s: reco master %u is ready to recover %u\n", + dlm->name, dlm->reco.new_master, dlm->reco.dead_node); + status = -EEXIST; + } else if (ret == DLM_RECOVERING) { + mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n", + dlm->name, dlm->node_num); + goto again; + } else { + struct dlm_lock_resource *res; + + /* dlmlock returned something other than NOTQUEUED or NORMAL */ + mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), " + "lksb.status=%s\n", dlm->name, dlm_errname(ret), + dlm_errname(lksb.status)); + res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, + DLM_RECOVERY_LOCK_NAME_LEN); + if (res) { + dlm_print_one_lock_resource(res); + dlm_lockres_put(res); + } else { + mlog(ML_ERROR, "recovery lock not found\n"); + } + BUG(); + } + + return status; +} + +static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node) +{ + struct dlm_begin_reco br; + int ret = 0; + struct dlm_node_iter iter; + int nodenum; + int status; + + mlog(0, "%s: dead node is %u\n", dlm->name, dead_node); + + spin_lock(&dlm->spinlock); + dlm_node_iter_init(dlm->domain_map, &iter); + spin_unlock(&dlm->spinlock); + + clear_bit(dead_node, iter.node_map); + + memset(&br, 0, sizeof(br)); + br.node_idx = dlm->node_num; + br.dead_node = dead_node; + + while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { + ret = 0; + if (nodenum == dead_node) { + mlog(0, "not sending begin reco to dead node " + "%u\n", dead_node); + continue; + } + if (nodenum == dlm->node_num) { + mlog(0, "not sending begin reco to self\n"); + continue; + } +retry: + mlog(0, "attempting to send begin reco msg to %d\n", + nodenum); + ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key, + &br, sizeof(br), nodenum, &status); + /* negative status is handled ok by caller here */ + if (ret >= 0) + ret = status; + if (dlm_is_host_down(ret)) { + /* node is down. not involved in recovery + * so just keep going */ + mlog(ML_NOTICE, "%s: node %u was down when sending " + "begin reco msg (%d)\n", dlm->name, nodenum, ret); + ret = 0; + } + + /* + * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8, + * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN. + * We are handling both for compatibility reasons. + */ + if (ret == -EAGAIN || ret == EAGAIN) { + mlog(0, "%s: trying to start recovery of node " + "%u, but node %u is waiting for last recovery " + "to complete, backoff for a bit\n", dlm->name, + dead_node, nodenum); + msleep(100); + goto retry; + } + if (ret < 0) { + struct dlm_lock_resource *res; + + /* this is now a serious problem, possibly ENOMEM + * in the network stack. must retry */ + mlog_errno(ret); + mlog(ML_ERROR, "begin reco of dlm %s to node %u " + "returned %d\n", dlm->name, nodenum, ret); + res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, + DLM_RECOVERY_LOCK_NAME_LEN); + if (res) { + dlm_print_one_lock_resource(res); + dlm_lockres_put(res); + } else { + mlog(ML_ERROR, "recovery lock not found\n"); + } + /* sleep for a bit in hopes that we can avoid + * another ENOMEM */ + msleep(100); + goto retry; + } + } + + return ret; +} + +int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf; + + /* ok to return 0, domain has gone away */ + if (!dlm_grab(dlm)) + return 0; + + spin_lock(&dlm->spinlock); + if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { + mlog(0, "%s: node %u wants to recover node %u (%u:%u) " + "but this node is in finalize state, waiting on finalize2\n", + dlm->name, br->node_idx, br->dead_node, + dlm->reco.dead_node, dlm->reco.new_master); + spin_unlock(&dlm->spinlock); + dlm_put(dlm); + return -EAGAIN; + } + spin_unlock(&dlm->spinlock); + + mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n", + dlm->name, br->node_idx, br->dead_node, + dlm->reco.dead_node, dlm->reco.new_master); + + dlm_fire_domain_eviction_callbacks(dlm, br->dead_node); + + spin_lock(&dlm->spinlock); + if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { + if (test_bit(dlm->reco.new_master, dlm->recovery_map)) { + mlog(0, "%s: new_master %u died, changing " + "to %u\n", dlm->name, dlm->reco.new_master, + br->node_idx); + } else { + mlog(0, "%s: new_master %u NOT DEAD, changing " + "to %u\n", dlm->name, dlm->reco.new_master, + br->node_idx); + /* may not have seen the new master as dead yet */ + } + } + if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) { + mlog(ML_NOTICE, "%s: dead_node previously set to %u, " + "node %u changing it to %u\n", dlm->name, + dlm->reco.dead_node, br->node_idx, br->dead_node); + } + dlm_set_reco_master(dlm, br->node_idx); + dlm_set_reco_dead_node(dlm, br->dead_node); + if (!test_bit(br->dead_node, dlm->recovery_map)) { + mlog(0, "recovery master %u sees %u as dead, but this " + "node has not yet. marking %u as dead\n", + br->node_idx, br->dead_node, br->dead_node); + if (!test_bit(br->dead_node, dlm->domain_map) || + !test_bit(br->dead_node, dlm->live_nodes_map)) + mlog(0, "%u not in domain/live_nodes map " + "so setting it in reco map manually\n", + br->dead_node); + /* force the recovery cleanup in __dlm_hb_node_down + * both of these will be cleared in a moment */ + set_bit(br->dead_node, dlm->domain_map); + set_bit(br->dead_node, dlm->live_nodes_map); + __dlm_hb_node_down(dlm, br->dead_node); + } + spin_unlock(&dlm->spinlock); + + dlm_kick_recovery_thread(dlm); + + mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n", + dlm->name, br->node_idx, br->dead_node, + dlm->reco.dead_node, dlm->reco.new_master); + + dlm_put(dlm); + return 0; +} + +#define DLM_FINALIZE_STAGE2 0x01 +static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm) +{ + int ret = 0; + struct dlm_finalize_reco fr; + struct dlm_node_iter iter; + int nodenum; + int status; + int stage = 1; + + mlog(0, "finishing recovery for node %s:%u, " + "stage %d\n", dlm->name, dlm->reco.dead_node, stage); + + spin_lock(&dlm->spinlock); + dlm_node_iter_init(dlm->domain_map, &iter); + spin_unlock(&dlm->spinlock); + +stage2: + memset(&fr, 0, sizeof(fr)); + fr.node_idx = dlm->node_num; + fr.dead_node = dlm->reco.dead_node; + if (stage == 2) + fr.flags |= DLM_FINALIZE_STAGE2; + + while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { + if (nodenum == dlm->node_num) + continue; + ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key, + &fr, sizeof(fr), nodenum, &status); + if (ret >= 0) + ret = status; + if (ret < 0) { + mlog(ML_ERROR, "Error %d when sending message %u (key " + "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG, + dlm->key, nodenum); + if (dlm_is_host_down(ret)) { + /* this has no effect on this recovery + * session, so set the status to zero to + * finish out the last recovery */ + mlog(ML_ERROR, "node %u went down after this " + "node finished recovery.\n", nodenum); + ret = 0; + continue; + } + break; + } + } + if (stage == 1) { + /* reset the node_iter back to the top and send finalize2 */ + iter.curnode = -1; + stage = 2; + goto stage2; + } + + return ret; +} + +int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf; + int stage = 1; + + /* ok to return 0, domain has gone away */ + if (!dlm_grab(dlm)) + return 0; + + if (fr->flags & DLM_FINALIZE_STAGE2) + stage = 2; + + mlog(0, "%s: node %u finalizing recovery stage%d of " + "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage, + fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master); + + spin_lock(&dlm->spinlock); + + if (dlm->reco.new_master != fr->node_idx) { + mlog(ML_ERROR, "node %u sent recovery finalize msg, but node " + "%u is supposed to be the new master, dead=%u\n", + fr->node_idx, dlm->reco.new_master, fr->dead_node); + BUG(); + } + if (dlm->reco.dead_node != fr->dead_node) { + mlog(ML_ERROR, "node %u sent recovery finalize msg for dead " + "node %u, but node %u is supposed to be dead\n", + fr->node_idx, fr->dead_node, dlm->reco.dead_node); + BUG(); + } + + switch (stage) { + case 1: + dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx); + if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { + mlog(ML_ERROR, "%s: received finalize1 from " + "new master %u for dead node %u, but " + "this node has already received it!\n", + dlm->name, fr->node_idx, fr->dead_node); + dlm_print_reco_node_status(dlm); + BUG(); + } + dlm->reco.state |= DLM_RECO_STATE_FINALIZE; + spin_unlock(&dlm->spinlock); + break; + case 2: + if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) { + mlog(ML_ERROR, "%s: received finalize2 from " + "new master %u for dead node %u, but " + "this node did not have finalize1!\n", + dlm->name, fr->node_idx, fr->dead_node); + dlm_print_reco_node_status(dlm); + BUG(); + } + dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; + __dlm_reset_recovery(dlm); + spin_unlock(&dlm->spinlock); + dlm_kick_recovery_thread(dlm); + break; + } + + mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n", + dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master); + + dlm_put(dlm); + return 0; +} diff --git a/fs/ocfs2/dlm/dlmthread.c b/fs/ocfs2/dlm/dlmthread.c new file mode 100644 index 000000000..eedf07ca2 --- /dev/null +++ b/fs/ocfs2/dlm/dlmthread.c @@ -0,0 +1,809 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * dlmthread.c + * + * standalone DLM module + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + + +#include "../cluster/heartbeat.h" +#include "../cluster/nodemanager.h" +#include "../cluster/tcp.h" + +#include "dlmapi.h" +#include "dlmcommon.h" +#include "dlmdomain.h" + +#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD) +#include "../cluster/masklog.h" + +static int dlm_thread(void *data); +static void dlm_flush_asts(struct dlm_ctxt *dlm); + +/* will exit holding res->spinlock, but may drop in function */ +/* waits until flags are cleared on res->state */ +void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags) +{ + DECLARE_WAITQUEUE(wait, current); + + assert_spin_locked(&res->spinlock); + + add_wait_queue(&res->wq, &wait); +repeat: + set_current_state(TASK_UNINTERRUPTIBLE); + if (res->state & flags) { + spin_unlock(&res->spinlock); + schedule(); + spin_lock(&res->spinlock); + goto repeat; + } + remove_wait_queue(&res->wq, &wait); + __set_current_state(TASK_RUNNING); +} + +int __dlm_lockres_has_locks(struct dlm_lock_resource *res) +{ + if (list_empty(&res->granted) && + list_empty(&res->converting) && + list_empty(&res->blocked)) + return 0; + return 1; +} + +/* "unused": the lockres has no locks, is not on the dirty list, + * has no inflight locks (in the gap between mastery and acquiring + * the first lock), and has no bits in its refmap. + * truly ready to be freed. */ +int __dlm_lockres_unused(struct dlm_lock_resource *res) +{ + int bit; + + assert_spin_locked(&res->spinlock); + + if (__dlm_lockres_has_locks(res)) + return 0; + + /* Locks are in the process of being created */ + if (res->inflight_locks) + return 0; + + if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY) + return 0; + + if (res->state & (DLM_LOCK_RES_RECOVERING| + DLM_LOCK_RES_RECOVERY_WAITING)) + return 0; + + /* Another node has this resource with this node as the master */ + bit = find_first_bit(res->refmap, O2NM_MAX_NODES); + if (bit < O2NM_MAX_NODES) + return 0; + + return 1; +} + + +/* Call whenever you may have added or deleted something from one of + * the lockres queue's. This will figure out whether it belongs on the + * unused list or not and does the appropriate thing. */ +void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&res->spinlock); + + if (__dlm_lockres_unused(res)){ + if (list_empty(&res->purge)) { + mlog(0, "%s: Adding res %.*s to purge list\n", + dlm->name, res->lockname.len, res->lockname.name); + + res->last_used = jiffies; + dlm_lockres_get(res); + list_add_tail(&res->purge, &dlm->purge_list); + dlm->purge_count++; + } + } else if (!list_empty(&res->purge)) { + mlog(0, "%s: Removing res %.*s from purge list\n", + dlm->name, res->lockname.len, res->lockname.name); + + list_del_init(&res->purge); + dlm_lockres_put(res); + dlm->purge_count--; + } +} + +void dlm_lockres_calc_usage(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + spin_lock(&dlm->spinlock); + spin_lock(&res->spinlock); + + __dlm_lockres_calc_usage(dlm, res); + + spin_unlock(&res->spinlock); + spin_unlock(&dlm->spinlock); +} + +/* + * Do the real purge work: + * unhash the lockres, and + * clear flag DLM_LOCK_RES_DROPPING_REF. + * It requires dlm and lockres spinlock to be taken. + */ +void __dlm_do_purge_lockres(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&res->spinlock); + + if (!list_empty(&res->purge)) { + mlog(0, "%s: Removing res %.*s from purgelist\n", + dlm->name, res->lockname.len, res->lockname.name); + list_del_init(&res->purge); + dlm_lockres_put(res); + dlm->purge_count--; + } + + if (!__dlm_lockres_unused(res)) { + mlog(ML_ERROR, "%s: res %.*s in use after deref\n", + dlm->name, res->lockname.len, res->lockname.name); + __dlm_print_one_lock_resource(res); + BUG(); + } + + __dlm_unhash_lockres(dlm, res); + + spin_lock(&dlm->track_lock); + if (!list_empty(&res->tracking)) + list_del_init(&res->tracking); + else { + mlog(ML_ERROR, "%s: Resource %.*s not on the Tracking list\n", + dlm->name, res->lockname.len, res->lockname.name); + __dlm_print_one_lock_resource(res); + } + spin_unlock(&dlm->track_lock); + + /* + * lockres is not in the hash now. drop the flag and wake up + * any processes waiting in dlm_get_lock_resource. + */ + res->state &= ~DLM_LOCK_RES_DROPPING_REF; +} + +static void dlm_purge_lockres(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + int master; + int ret = 0; + + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&res->spinlock); + + master = (res->owner == dlm->node_num); + + mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name, + res->lockname.len, res->lockname.name, master); + + if (!master) { + if (res->state & DLM_LOCK_RES_DROPPING_REF) { + mlog(ML_NOTICE, "%s: res %.*s already in DLM_LOCK_RES_DROPPING_REF state\n", + dlm->name, res->lockname.len, res->lockname.name); + spin_unlock(&res->spinlock); + return; + } + + res->state |= DLM_LOCK_RES_DROPPING_REF; + /* drop spinlock... retake below */ + spin_unlock(&res->spinlock); + spin_unlock(&dlm->spinlock); + + spin_lock(&res->spinlock); + /* This ensures that clear refmap is sent after the set */ + __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG); + spin_unlock(&res->spinlock); + + /* clear our bit from the master's refmap, ignore errors */ + ret = dlm_drop_lockres_ref(dlm, res); + if (ret < 0) { + if (!dlm_is_host_down(ret)) + BUG(); + } + spin_lock(&dlm->spinlock); + spin_lock(&res->spinlock); + } + + if (!list_empty(&res->purge)) { + mlog(0, "%s: Removing res %.*s from purgelist, master %d\n", + dlm->name, res->lockname.len, res->lockname.name, master); + list_del_init(&res->purge); + dlm_lockres_put(res); + dlm->purge_count--; + } + + if (!master && ret == DLM_DEREF_RESPONSE_INPROG) { + mlog(0, "%s: deref %.*s in progress\n", + dlm->name, res->lockname.len, res->lockname.name); + spin_unlock(&res->spinlock); + return; + } + + if (!__dlm_lockres_unused(res)) { + mlog(ML_ERROR, "%s: res %.*s in use after deref\n", + dlm->name, res->lockname.len, res->lockname.name); + __dlm_print_one_lock_resource(res); + BUG(); + } + + __dlm_unhash_lockres(dlm, res); + + spin_lock(&dlm->track_lock); + if (!list_empty(&res->tracking)) + list_del_init(&res->tracking); + else { + mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", + res->lockname.len, res->lockname.name); + __dlm_print_one_lock_resource(res); + } + spin_unlock(&dlm->track_lock); + + /* lockres is not in the hash now. drop the flag and wake up + * any processes waiting in dlm_get_lock_resource. */ + if (!master) { + res->state &= ~DLM_LOCK_RES_DROPPING_REF; + spin_unlock(&res->spinlock); + wake_up(&res->wq); + } else + spin_unlock(&res->spinlock); +} + +static void dlm_run_purge_list(struct dlm_ctxt *dlm, + int purge_now) +{ + unsigned int run_max, unused; + unsigned long purge_jiffies; + struct dlm_lock_resource *lockres; + + spin_lock(&dlm->spinlock); + run_max = dlm->purge_count; + + while(run_max && !list_empty(&dlm->purge_list)) { + run_max--; + + lockres = list_entry(dlm->purge_list.next, + struct dlm_lock_resource, purge); + + spin_lock(&lockres->spinlock); + + purge_jiffies = lockres->last_used + + msecs_to_jiffies(DLM_PURGE_INTERVAL_MS); + + /* Make sure that we want to be processing this guy at + * this time. */ + if (!purge_now && time_after(purge_jiffies, jiffies)) { + /* Since resources are added to the purge list + * in tail order, we can stop at the first + * unpurgable resource -- anyone added after + * him will have a greater last_used value */ + spin_unlock(&lockres->spinlock); + break; + } + + /* Status of the lockres *might* change so double + * check. If the lockres is unused, holding the dlm + * spinlock will prevent people from getting and more + * refs on it. */ + unused = __dlm_lockres_unused(lockres); + if (!unused || + (lockres->state & DLM_LOCK_RES_MIGRATING) || + (lockres->inflight_assert_workers != 0)) { + mlog(0, "%s: res %.*s is in use or being remastered, " + "used %d, state %d, assert master workers %u\n", + dlm->name, lockres->lockname.len, + lockres->lockname.name, + !unused, lockres->state, + lockres->inflight_assert_workers); + list_move_tail(&lockres->purge, &dlm->purge_list); + spin_unlock(&lockres->spinlock); + continue; + } + + dlm_lockres_get(lockres); + + dlm_purge_lockres(dlm, lockres); + + dlm_lockres_put(lockres); + + /* Avoid adding any scheduling latencies */ + cond_resched_lock(&dlm->spinlock); + } + + spin_unlock(&dlm->spinlock); +} + +static void dlm_shuffle_lists(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + struct dlm_lock *lock, *target; + int can_grant = 1; + + /* + * Because this function is called with the lockres + * spinlock, and because we know that it is not migrating/ + * recovering/in-progress, it is fine to reserve asts and + * basts right before queueing them all throughout + */ + assert_spin_locked(&dlm->ast_lock); + assert_spin_locked(&res->spinlock); + BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING| + DLM_LOCK_RES_RECOVERING| + DLM_LOCK_RES_IN_PROGRESS))); + +converting: + if (list_empty(&res->converting)) + goto blocked; + mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name, + res->lockname.len, res->lockname.name); + + target = list_entry(res->converting.next, struct dlm_lock, list); + if (target->ml.convert_type == LKM_IVMODE) { + mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n", + dlm->name, res->lockname.len, res->lockname.name); + BUG(); + } + list_for_each_entry(lock, &res->granted, list) { + if (lock==target) + continue; + if (!dlm_lock_compatible(lock->ml.type, + target->ml.convert_type)) { + can_grant = 0; + /* queue the BAST if not already */ + if (lock->ml.highest_blocked == LKM_IVMODE) { + __dlm_lockres_reserve_ast(res); + __dlm_queue_bast(dlm, lock); + } + /* update the highest_blocked if needed */ + if (lock->ml.highest_blocked < target->ml.convert_type) + lock->ml.highest_blocked = + target->ml.convert_type; + } + } + + list_for_each_entry(lock, &res->converting, list) { + if (lock==target) + continue; + if (!dlm_lock_compatible(lock->ml.type, + target->ml.convert_type)) { + can_grant = 0; + if (lock->ml.highest_blocked == LKM_IVMODE) { + __dlm_lockres_reserve_ast(res); + __dlm_queue_bast(dlm, lock); + } + if (lock->ml.highest_blocked < target->ml.convert_type) + lock->ml.highest_blocked = + target->ml.convert_type; + } + } + + /* we can convert the lock */ + if (can_grant) { + spin_lock(&target->spinlock); + BUG_ON(target->ml.highest_blocked != LKM_IVMODE); + + mlog(0, "%s: res %.*s, AST for Converting lock %u:%llu, type " + "%d => %d, node %u\n", dlm->name, res->lockname.len, + res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)), + target->ml.type, + target->ml.convert_type, target->ml.node); + + target->ml.type = target->ml.convert_type; + target->ml.convert_type = LKM_IVMODE; + list_move_tail(&target->list, &res->granted); + + BUG_ON(!target->lksb); + target->lksb->status = DLM_NORMAL; + + spin_unlock(&target->spinlock); + + __dlm_lockres_reserve_ast(res); + __dlm_queue_ast(dlm, target); + /* go back and check for more */ + goto converting; + } + +blocked: + if (list_empty(&res->blocked)) + goto leave; + target = list_entry(res->blocked.next, struct dlm_lock, list); + + list_for_each_entry(lock, &res->granted, list) { + if (lock==target) + continue; + if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) { + can_grant = 0; + if (lock->ml.highest_blocked == LKM_IVMODE) { + __dlm_lockres_reserve_ast(res); + __dlm_queue_bast(dlm, lock); + } + if (lock->ml.highest_blocked < target->ml.type) + lock->ml.highest_blocked = target->ml.type; + } + } + + list_for_each_entry(lock, &res->converting, list) { + if (lock==target) + continue; + if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) { + can_grant = 0; + if (lock->ml.highest_blocked == LKM_IVMODE) { + __dlm_lockres_reserve_ast(res); + __dlm_queue_bast(dlm, lock); + } + if (lock->ml.highest_blocked < target->ml.type) + lock->ml.highest_blocked = target->ml.type; + } + } + + /* we can grant the blocked lock (only + * possible if converting list empty) */ + if (can_grant) { + spin_lock(&target->spinlock); + BUG_ON(target->ml.highest_blocked != LKM_IVMODE); + + mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, " + "node %u\n", dlm->name, res->lockname.len, + res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)), + target->ml.type, target->ml.node); + + /* target->ml.type is already correct */ + list_move_tail(&target->list, &res->granted); + + BUG_ON(!target->lksb); + target->lksb->status = DLM_NORMAL; + + spin_unlock(&target->spinlock); + + __dlm_lockres_reserve_ast(res); + __dlm_queue_ast(dlm, target); + /* go back and check for more */ + goto converting; + } + +leave: + return; +} + +/* must have NO locks when calling this with res !=NULL * */ +void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) +{ + if (res) { + spin_lock(&dlm->spinlock); + spin_lock(&res->spinlock); + __dlm_dirty_lockres(dlm, res); + spin_unlock(&res->spinlock); + spin_unlock(&dlm->spinlock); + } + wake_up(&dlm->dlm_thread_wq); +} + +void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) +{ + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&res->spinlock); + + /* don't shuffle secondary queues */ + if (res->owner == dlm->node_num) { + if (res->state & (DLM_LOCK_RES_MIGRATING | + DLM_LOCK_RES_BLOCK_DIRTY)) + return; + + if (list_empty(&res->dirty)) { + /* ref for dirty_list */ + dlm_lockres_get(res); + list_add_tail(&res->dirty, &dlm->dirty_list); + res->state |= DLM_LOCK_RES_DIRTY; + } + } + + mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len, + res->lockname.name); +} + + +/* Launch the NM thread for the mounted volume */ +int dlm_launch_thread(struct dlm_ctxt *dlm) +{ + mlog(0, "Starting dlm_thread...\n"); + + dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm-%s", + dlm->name); + if (IS_ERR(dlm->dlm_thread_task)) { + mlog_errno(PTR_ERR(dlm->dlm_thread_task)); + dlm->dlm_thread_task = NULL; + return -EINVAL; + } + + return 0; +} + +void dlm_complete_thread(struct dlm_ctxt *dlm) +{ + if (dlm->dlm_thread_task) { + mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n"); + kthread_stop(dlm->dlm_thread_task); + dlm->dlm_thread_task = NULL; + } +} + +static int dlm_dirty_list_empty(struct dlm_ctxt *dlm) +{ + int empty; + + spin_lock(&dlm->spinlock); + empty = list_empty(&dlm->dirty_list); + spin_unlock(&dlm->spinlock); + + return empty; +} + +static void dlm_flush_asts(struct dlm_ctxt *dlm) +{ + int ret; + struct dlm_lock *lock; + struct dlm_lock_resource *res; + u8 hi; + + spin_lock(&dlm->ast_lock); + while (!list_empty(&dlm->pending_asts)) { + lock = list_entry(dlm->pending_asts.next, + struct dlm_lock, ast_list); + /* get an extra ref on lock */ + dlm_lock_get(lock); + res = lock->lockres; + mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, " + "node %u\n", dlm->name, res->lockname.len, + res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), + lock->ml.type, lock->ml.node); + + BUG_ON(!lock->ast_pending); + + /* remove from list (including ref) */ + list_del_init(&lock->ast_list); + dlm_lock_put(lock); + spin_unlock(&dlm->ast_lock); + + if (lock->ml.node != dlm->node_num) { + ret = dlm_do_remote_ast(dlm, res, lock); + if (ret < 0) + mlog_errno(ret); + } else + dlm_do_local_ast(dlm, res, lock); + + spin_lock(&dlm->ast_lock); + + /* possible that another ast was queued while + * we were delivering the last one */ + if (!list_empty(&lock->ast_list)) { + mlog(0, "%s: res %.*s, AST queued while flushing last " + "one\n", dlm->name, res->lockname.len, + res->lockname.name); + } else + lock->ast_pending = 0; + + /* drop the extra ref. + * this may drop it completely. */ + dlm_lock_put(lock); + dlm_lockres_release_ast(dlm, res); + } + + while (!list_empty(&dlm->pending_basts)) { + lock = list_entry(dlm->pending_basts.next, + struct dlm_lock, bast_list); + /* get an extra ref on lock */ + dlm_lock_get(lock); + res = lock->lockres; + + BUG_ON(!lock->bast_pending); + + /* get the highest blocked lock, and reset */ + spin_lock(&lock->spinlock); + BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE); + hi = lock->ml.highest_blocked; + lock->ml.highest_blocked = LKM_IVMODE; + spin_unlock(&lock->spinlock); + + /* remove from list (including ref) */ + list_del_init(&lock->bast_list); + dlm_lock_put(lock); + spin_unlock(&dlm->ast_lock); + + mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, " + "blocked %d, node %u\n", + dlm->name, res->lockname.len, res->lockname.name, + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), + hi, lock->ml.node); + + if (lock->ml.node != dlm->node_num) { + ret = dlm_send_proxy_bast(dlm, res, lock, hi); + if (ret < 0) + mlog_errno(ret); + } else + dlm_do_local_bast(dlm, res, lock, hi); + + spin_lock(&dlm->ast_lock); + + /* possible that another bast was queued while + * we were delivering the last one */ + if (!list_empty(&lock->bast_list)) { + mlog(0, "%s: res %.*s, BAST queued while flushing last " + "one\n", dlm->name, res->lockname.len, + res->lockname.name); + } else + lock->bast_pending = 0; + + /* drop the extra ref. + * this may drop it completely. */ + dlm_lock_put(lock); + dlm_lockres_release_ast(dlm, res); + } + wake_up(&dlm->ast_wq); + spin_unlock(&dlm->ast_lock); +} + + +#define DLM_THREAD_TIMEOUT_MS (4 * 1000) +#define DLM_THREAD_MAX_DIRTY 100 + +static int dlm_thread(void *data) +{ + struct dlm_lock_resource *res; + struct dlm_ctxt *dlm = data; + unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS); + + mlog(0, "dlm thread running for %s...\n", dlm->name); + + while (!kthread_should_stop()) { + int n = DLM_THREAD_MAX_DIRTY; + + /* dlm_shutting_down is very point-in-time, but that + * doesn't matter as we'll just loop back around if we + * get false on the leading edge of a state + * transition. */ + dlm_run_purge_list(dlm, dlm_shutting_down(dlm)); + + /* We really don't want to hold dlm->spinlock while + * calling dlm_shuffle_lists on each lockres that + * needs to have its queues adjusted and AST/BASTs + * run. So let's pull each entry off the dirty_list + * and drop dlm->spinlock ASAP. Once off the list, + * res->spinlock needs to be taken again to protect + * the queues while calling dlm_shuffle_lists. */ + spin_lock(&dlm->spinlock); + while (!list_empty(&dlm->dirty_list)) { + int delay = 0; + res = list_entry(dlm->dirty_list.next, + struct dlm_lock_resource, dirty); + + /* peel a lockres off, remove it from the list, + * unset the dirty flag and drop the dlm lock */ + BUG_ON(!res); + dlm_lockres_get(res); + + spin_lock(&res->spinlock); + /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */ + list_del_init(&res->dirty); + spin_unlock(&res->spinlock); + spin_unlock(&dlm->spinlock); + /* Drop dirty_list ref */ + dlm_lockres_put(res); + + /* lockres can be re-dirtied/re-added to the + * dirty_list in this gap, but that is ok */ + + spin_lock(&dlm->ast_lock); + spin_lock(&res->spinlock); + if (res->owner != dlm->node_num) { + __dlm_print_one_lock_resource(res); + mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d," + " dirty %d\n", dlm->name, + !!(res->state & DLM_LOCK_RES_IN_PROGRESS), + !!(res->state & DLM_LOCK_RES_MIGRATING), + !!(res->state & DLM_LOCK_RES_RECOVERING), + !!(res->state & DLM_LOCK_RES_DIRTY)); + } + BUG_ON(res->owner != dlm->node_num); + + /* it is now ok to move lockreses in these states + * to the dirty list, assuming that they will only be + * dirty for a short while. */ + BUG_ON(res->state & DLM_LOCK_RES_MIGRATING); + if (res->state & (DLM_LOCK_RES_IN_PROGRESS | + DLM_LOCK_RES_RECOVERING | + DLM_LOCK_RES_RECOVERY_WAITING)) { + /* move it to the tail and keep going */ + res->state &= ~DLM_LOCK_RES_DIRTY; + spin_unlock(&res->spinlock); + spin_unlock(&dlm->ast_lock); + mlog(0, "%s: res %.*s, inprogress, delay list " + "shuffle, state %d\n", dlm->name, + res->lockname.len, res->lockname.name, + res->state); + delay = 1; + goto in_progress; + } + + /* at this point the lockres is not migrating/ + * recovering/in-progress. we have the lockres + * spinlock and do NOT have the dlm lock. + * safe to reserve/queue asts and run the lists. */ + + /* called while holding lockres lock */ + dlm_shuffle_lists(dlm, res); + res->state &= ~DLM_LOCK_RES_DIRTY; + spin_unlock(&res->spinlock); + spin_unlock(&dlm->ast_lock); + + dlm_lockres_calc_usage(dlm, res); + +in_progress: + + spin_lock(&dlm->spinlock); + /* if the lock was in-progress, stick + * it on the back of the list */ + if (delay) { + spin_lock(&res->spinlock); + __dlm_dirty_lockres(dlm, res); + spin_unlock(&res->spinlock); + } + dlm_lockres_put(res); + + /* unlikely, but we may need to give time to + * other tasks */ + if (!--n) { + mlog(0, "%s: Throttling dlm thread\n", + dlm->name); + break; + } + } + + spin_unlock(&dlm->spinlock); + dlm_flush_asts(dlm); + + /* yield and continue right away if there is more work to do */ + if (!n) { + cond_resched(); + continue; + } + + wait_event_interruptible_timeout(dlm->dlm_thread_wq, + !dlm_dirty_list_empty(dlm) || + kthread_should_stop(), + timeout); + } + + mlog(0, "quitting DLM thread\n"); + return 0; +} diff --git a/fs/ocfs2/dlm/dlmunlock.c b/fs/ocfs2/dlm/dlmunlock.c new file mode 100644 index 000000000..7318e4794 --- /dev/null +++ b/fs/ocfs2/dlm/dlmunlock.c @@ -0,0 +1,695 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * dlmunlock.c + * + * underlying calls for unlocking locks + * + * Copyright (C) 2004 Oracle. All rights reserved. + */ + + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "../cluster/heartbeat.h" +#include "../cluster/nodemanager.h" +#include "../cluster/tcp.h" + +#include "dlmapi.h" +#include "dlmcommon.h" + +#define MLOG_MASK_PREFIX ML_DLM +#include "../cluster/masklog.h" + +#define DLM_UNLOCK_FREE_LOCK 0x00000001 +#define DLM_UNLOCK_CALL_AST 0x00000002 +#define DLM_UNLOCK_REMOVE_LOCK 0x00000004 +#define DLM_UNLOCK_REGRANT_LOCK 0x00000008 +#define DLM_UNLOCK_CLEAR_CONVERT_TYPE 0x00000010 + + +static enum dlm_status dlm_get_cancel_actions(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + struct dlm_lockstatus *lksb, + int *actions); +static enum dlm_status dlm_get_unlock_actions(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + struct dlm_lockstatus *lksb, + int *actions); + +static enum dlm_status dlm_send_remote_unlock_request(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + struct dlm_lockstatus *lksb, + int flags, + u8 owner); + + +/* + * according to the spec: + * http://opendlm.sourceforge.net/cvsmirror/opendlm/docs/dlmbook_final.pdf + * + * flags & LKM_CANCEL != 0: must be converting or blocked + * flags & LKM_CANCEL == 0: must be granted + * + * So to unlock a converting lock, you must first cancel the + * convert (passing LKM_CANCEL in flags), then call the unlock + * again (with no LKM_CANCEL in flags). + */ + + +/* + * locking: + * caller needs: none + * taken: res->spinlock and lock->spinlock taken and dropped + * held on exit: none + * returns: DLM_NORMAL, DLM_NOLOCKMGR, status from network + * all callers should have taken an extra ref on lock coming in + */ +static enum dlm_status dlmunlock_common(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + struct dlm_lockstatus *lksb, + int flags, int *call_ast, + int master_node) +{ + enum dlm_status status; + int actions = 0; + int in_use; + u8 owner; + int recovery_wait = 0; + + mlog(0, "master_node = %d, valblk = %d\n", master_node, + flags & LKM_VALBLK); + + if (master_node) + BUG_ON(res->owner != dlm->node_num); + else + BUG_ON(res->owner == dlm->node_num); + + spin_lock(&dlm->ast_lock); + /* We want to be sure that we're not freeing a lock + * that still has AST's pending... */ + in_use = !list_empty(&lock->ast_list); + spin_unlock(&dlm->ast_lock); + if (in_use && !(flags & LKM_CANCEL)) { + mlog(ML_ERROR, "lockres %.*s: Someone is calling dlmunlock " + "while waiting for an ast!", res->lockname.len, + res->lockname.name); + return DLM_BADPARAM; + } + + spin_lock(&res->spinlock); + if (res->state & DLM_LOCK_RES_IN_PROGRESS) { + if (master_node && !(flags & LKM_CANCEL)) { + mlog(ML_ERROR, "lockres in progress!\n"); + spin_unlock(&res->spinlock); + return DLM_FORWARD; + } + /* ok for this to sleep if not in a network handler */ + __dlm_wait_on_lockres(res); + res->state |= DLM_LOCK_RES_IN_PROGRESS; + } + spin_lock(&lock->spinlock); + + if (res->state & DLM_LOCK_RES_RECOVERING) { + status = DLM_RECOVERING; + goto leave; + } + + if (res->state & DLM_LOCK_RES_MIGRATING) { + status = DLM_MIGRATING; + goto leave; + } + + /* see above for what the spec says about + * LKM_CANCEL and the lock queue state */ + if (flags & LKM_CANCEL) + status = dlm_get_cancel_actions(dlm, res, lock, lksb, &actions); + else + status = dlm_get_unlock_actions(dlm, res, lock, lksb, &actions); + + if (status != DLM_NORMAL && (status != DLM_CANCELGRANT || !master_node)) + goto leave; + + /* By now this has been masked out of cancel requests. */ + if (flags & LKM_VALBLK) { + /* make the final update to the lvb */ + if (master_node) + memcpy(res->lvb, lksb->lvb, DLM_LVB_LEN); + else + flags |= LKM_PUT_LVB; /* let the send function + * handle it. */ + } + + if (!master_node) { + owner = res->owner; + /* drop locks and send message */ + if (flags & LKM_CANCEL) + lock->cancel_pending = 1; + else + lock->unlock_pending = 1; + spin_unlock(&lock->spinlock); + spin_unlock(&res->spinlock); + status = dlm_send_remote_unlock_request(dlm, res, lock, lksb, + flags, owner); + spin_lock(&res->spinlock); + spin_lock(&lock->spinlock); + /* if the master told us the lock was already granted, + * let the ast handle all of these actions */ + if (status == DLM_CANCELGRANT) { + actions &= ~(DLM_UNLOCK_REMOVE_LOCK| + DLM_UNLOCK_REGRANT_LOCK| + DLM_UNLOCK_CLEAR_CONVERT_TYPE); + } else if (status == DLM_RECOVERING || + status == DLM_MIGRATING || + status == DLM_FORWARD || + status == DLM_NOLOCKMGR + ) { + /* must clear the actions because this unlock + * is about to be retried. cannot free or do + * any list manipulation. */ + mlog(0, "%s:%.*s: clearing actions, %s\n", + dlm->name, res->lockname.len, + res->lockname.name, + status==DLM_RECOVERING?"recovering": + (status==DLM_MIGRATING?"migrating": + (status == DLM_FORWARD ? "forward" : + "nolockmanager"))); + actions = 0; + } + if (flags & LKM_CANCEL) + lock->cancel_pending = 0; + else { + if (!lock->unlock_pending) + recovery_wait = 1; + else + lock->unlock_pending = 0; + } + } + + /* get an extra ref on lock. if we are just switching + * lists here, we dont want the lock to go away. */ + dlm_lock_get(lock); + + if (actions & DLM_UNLOCK_REMOVE_LOCK) { + list_del_init(&lock->list); + dlm_lock_put(lock); + } + if (actions & DLM_UNLOCK_REGRANT_LOCK) { + dlm_lock_get(lock); + list_add_tail(&lock->list, &res->granted); + } + if (actions & DLM_UNLOCK_CLEAR_CONVERT_TYPE) { + mlog(0, "clearing convert_type at %smaster node\n", + master_node ? "" : "non-"); + lock->ml.convert_type = LKM_IVMODE; + } + + /* remove the extra ref on lock */ + dlm_lock_put(lock); + +leave: + res->state &= ~DLM_LOCK_RES_IN_PROGRESS; + if (!dlm_lock_on_list(&res->converting, lock)) + BUG_ON(lock->ml.convert_type != LKM_IVMODE); + else + BUG_ON(lock->ml.convert_type == LKM_IVMODE); + spin_unlock(&lock->spinlock); + spin_unlock(&res->spinlock); + wake_up(&res->wq); + + if (recovery_wait) { + spin_lock(&res->spinlock); + /* Unlock request will directly succeed after owner dies, + * and the lock is already removed from grant list. We have to + * wait for RECOVERING done or we miss the chance to purge it + * since the removement is much faster than RECOVERING proc. + */ + __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_RECOVERING); + spin_unlock(&res->spinlock); + } + + /* let the caller's final dlm_lock_put handle the actual kfree */ + if (actions & DLM_UNLOCK_FREE_LOCK) { + /* this should always be coupled with list removal */ + BUG_ON(!(actions & DLM_UNLOCK_REMOVE_LOCK)); + mlog(0, "lock %u:%llu should be gone now! refs=%d\n", + dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), + kref_read(&lock->lock_refs)-1); + dlm_lock_put(lock); + } + if (actions & DLM_UNLOCK_CALL_AST) + *call_ast = 1; + + /* if cancel or unlock succeeded, lvb work is done */ + if (status == DLM_NORMAL) + lksb->flags &= ~(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB); + + return status; +} + +void dlm_commit_pending_unlock(struct dlm_lock_resource *res, + struct dlm_lock *lock) +{ + /* leave DLM_LKSB_PUT_LVB on the lksb so any final + * update of the lvb will be sent to the new master */ + list_del_init(&lock->list); +} + +void dlm_commit_pending_cancel(struct dlm_lock_resource *res, + struct dlm_lock *lock) +{ + list_move_tail(&lock->list, &res->granted); + lock->ml.convert_type = LKM_IVMODE; +} + + +static inline enum dlm_status dlmunlock_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + struct dlm_lockstatus *lksb, + int flags, + int *call_ast) +{ + return dlmunlock_common(dlm, res, lock, lksb, flags, call_ast, 1); +} + +static inline enum dlm_status dlmunlock_remote(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + struct dlm_lockstatus *lksb, + int flags, int *call_ast) +{ + return dlmunlock_common(dlm, res, lock, lksb, flags, call_ast, 0); +} + +/* + * locking: + * caller needs: none + * taken: none + * held on exit: none + * returns: DLM_NORMAL, DLM_NOLOCKMGR, status from network + */ +static enum dlm_status dlm_send_remote_unlock_request(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + struct dlm_lockstatus *lksb, + int flags, + u8 owner) +{ + struct dlm_unlock_lock unlock; + int tmpret; + enum dlm_status ret; + int status = 0; + struct kvec vec[2]; + size_t veclen = 1; + + mlog(0, "%.*s\n", res->lockname.len, res->lockname.name); + + if (owner == dlm->node_num) { + /* ended up trying to contact ourself. this means + * that the lockres had been remote but became local + * via a migration. just retry it, now as local */ + mlog(0, "%s:%.*s: this node became the master due to a " + "migration, re-evaluate now\n", dlm->name, + res->lockname.len, res->lockname.name); + return DLM_FORWARD; + } + + memset(&unlock, 0, sizeof(unlock)); + unlock.node_idx = dlm->node_num; + unlock.flags = cpu_to_be32(flags); + unlock.cookie = lock->ml.cookie; + unlock.namelen = res->lockname.len; + memcpy(unlock.name, res->lockname.name, unlock.namelen); + + vec[0].iov_len = sizeof(struct dlm_unlock_lock); + vec[0].iov_base = &unlock; + + if (flags & LKM_PUT_LVB) { + /* extra data to send if we are updating lvb */ + vec[1].iov_len = DLM_LVB_LEN; + vec[1].iov_base = lock->lksb->lvb; + veclen++; + } + + tmpret = o2net_send_message_vec(DLM_UNLOCK_LOCK_MSG, dlm->key, + vec, veclen, owner, &status); + if (tmpret >= 0) { + // successfully sent and received + if (status == DLM_FORWARD) + mlog(0, "master was in-progress. retry\n"); + ret = status; + } else { + mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to " + "node %u\n", tmpret, DLM_UNLOCK_LOCK_MSG, dlm->key, owner); + if (dlm_is_host_down(tmpret)) { + /* NOTE: this seems strange, but it is what we want. + * when the master goes down during a cancel or + * unlock, the recovery code completes the operation + * as if the master had not died, then passes the + * updated state to the recovery master. this thread + * just needs to finish out the operation and call + * the unlockast. */ + if (dlm_is_node_dead(dlm, owner)) + ret = DLM_NORMAL; + else + ret = DLM_NOLOCKMGR; + } else { + /* something bad. this will BUG in ocfs2 */ + ret = dlm_err_to_dlm_status(tmpret); + } + } + + return ret; +} + +/* + * locking: + * caller needs: none + * taken: takes and drops res->spinlock + * held on exit: none + * returns: DLM_NORMAL, DLM_BADARGS, DLM_IVLOCKID, + * return value from dlmunlock_master + */ +int dlm_unlock_lock_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) +{ + struct dlm_ctxt *dlm = data; + struct dlm_unlock_lock *unlock = (struct dlm_unlock_lock *)msg->buf; + struct dlm_lock_resource *res = NULL; + struct dlm_lock *lock = NULL, *iter; + enum dlm_status status = DLM_NORMAL; + int i; + struct dlm_lockstatus *lksb = NULL; + int ignore; + u32 flags; + struct list_head *queue; + + flags = be32_to_cpu(unlock->flags); + + if (flags & LKM_GET_LVB) { + mlog(ML_ERROR, "bad args! GET_LVB specified on unlock!\n"); + return DLM_BADARGS; + } + + if ((flags & (LKM_PUT_LVB|LKM_CANCEL)) == (LKM_PUT_LVB|LKM_CANCEL)) { + mlog(ML_ERROR, "bad args! cannot modify lvb on a CANCEL " + "request!\n"); + return DLM_BADARGS; + } + + if (unlock->namelen > DLM_LOCKID_NAME_MAX) { + mlog(ML_ERROR, "Invalid name length in unlock handler!\n"); + return DLM_IVBUFLEN; + } + + if (!dlm_grab(dlm)) + return DLM_FORWARD; + + mlog_bug_on_msg(!dlm_domain_fully_joined(dlm), + "Domain %s not fully joined!\n", dlm->name); + + mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" : "none"); + + res = dlm_lookup_lockres(dlm, unlock->name, unlock->namelen); + if (!res) { + /* We assume here that a no lock resource simply means + * it was migrated away and destroyed before the other + * node could detect it. */ + mlog(0, "returning DLM_FORWARD -- res no longer exists\n"); + status = DLM_FORWARD; + goto not_found; + } + + queue=&res->granted; + spin_lock(&res->spinlock); + if (res->state & DLM_LOCK_RES_RECOVERING) { + spin_unlock(&res->spinlock); + mlog(0, "returning DLM_RECOVERING\n"); + status = DLM_RECOVERING; + goto leave; + } + + if (res->state & DLM_LOCK_RES_MIGRATING) { + spin_unlock(&res->spinlock); + mlog(0, "returning DLM_MIGRATING\n"); + status = DLM_MIGRATING; + goto leave; + } + + if (res->owner != dlm->node_num) { + spin_unlock(&res->spinlock); + mlog(0, "returning DLM_FORWARD -- not master\n"); + status = DLM_FORWARD; + goto leave; + } + + for (i=0; i<3; i++) { + list_for_each_entry(iter, queue, list) { + if (iter->ml.cookie == unlock->cookie && + iter->ml.node == unlock->node_idx) { + dlm_lock_get(iter); + lock = iter; + break; + } + } + if (lock) + break; + /* scan granted -> converting -> blocked queues */ + queue++; + } + spin_unlock(&res->spinlock); + if (!lock) { + status = DLM_IVLOCKID; + goto not_found; + } + + /* lock was found on queue */ + lksb = lock->lksb; + if (flags & (LKM_VALBLK|LKM_PUT_LVB) && + lock->ml.type != LKM_EXMODE) + flags &= ~(LKM_VALBLK|LKM_PUT_LVB); + + /* unlockast only called on originating node */ + if (flags & LKM_PUT_LVB) { + lksb->flags |= DLM_LKSB_PUT_LVB; + memcpy(&lksb->lvb[0], &unlock->lvb[0], DLM_LVB_LEN); + } + + /* if this is in-progress, propagate the DLM_FORWARD + * all the way back out */ + status = dlmunlock_master(dlm, res, lock, lksb, flags, &ignore); + if (status == DLM_FORWARD) + mlog(0, "lockres is in progress\n"); + + if (flags & LKM_PUT_LVB) + lksb->flags &= ~DLM_LKSB_PUT_LVB; + + dlm_lockres_calc_usage(dlm, res); + dlm_kick_thread(dlm, res); + +not_found: + if (!lock) + mlog(ML_ERROR, "failed to find lock to unlock! " + "cookie=%u:%llu\n", + dlm_get_lock_cookie_node(be64_to_cpu(unlock->cookie)), + dlm_get_lock_cookie_seq(be64_to_cpu(unlock->cookie))); + else + dlm_lock_put(lock); + +leave: + if (res) + dlm_lockres_put(res); + + dlm_put(dlm); + + return status; +} + + +static enum dlm_status dlm_get_cancel_actions(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + struct dlm_lockstatus *lksb, + int *actions) +{ + enum dlm_status status; + + if (dlm_lock_on_list(&res->blocked, lock)) { + /* cancel this outright */ + status = DLM_NORMAL; + *actions = (DLM_UNLOCK_CALL_AST | + DLM_UNLOCK_REMOVE_LOCK); + } else if (dlm_lock_on_list(&res->converting, lock)) { + /* cancel the request, put back on granted */ + status = DLM_NORMAL; + *actions = (DLM_UNLOCK_CALL_AST | + DLM_UNLOCK_REMOVE_LOCK | + DLM_UNLOCK_REGRANT_LOCK | + DLM_UNLOCK_CLEAR_CONVERT_TYPE); + } else if (dlm_lock_on_list(&res->granted, lock)) { + /* too late, already granted. */ + status = DLM_CANCELGRANT; + *actions = DLM_UNLOCK_CALL_AST; + } else { + mlog(ML_ERROR, "lock to cancel is not on any list!\n"); + status = DLM_IVLOCKID; + *actions = 0; + } + return status; +} + +static enum dlm_status dlm_get_unlock_actions(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + struct dlm_lock *lock, + struct dlm_lockstatus *lksb, + int *actions) +{ + enum dlm_status status; + + /* unlock request */ + if (!dlm_lock_on_list(&res->granted, lock)) { + status = DLM_DENIED; + dlm_error(status); + *actions = 0; + } else { + /* unlock granted lock */ + status = DLM_NORMAL; + *actions = (DLM_UNLOCK_FREE_LOCK | + DLM_UNLOCK_CALL_AST | + DLM_UNLOCK_REMOVE_LOCK); + } + return status; +} + +/* there seems to be no point in doing this async + * since (even for the remote case) there is really + * no work to queue up... so just do it and fire the + * unlockast by hand when done... */ +enum dlm_status dlmunlock(struct dlm_ctxt *dlm, struct dlm_lockstatus *lksb, + int flags, dlm_astunlockfunc_t *unlockast, void *data) +{ + enum dlm_status status; + struct dlm_lock_resource *res; + struct dlm_lock *lock = NULL; + int call_ast, is_master; + + if (!lksb) { + dlm_error(DLM_BADARGS); + return DLM_BADARGS; + } + + if (flags & ~(LKM_CANCEL | LKM_VALBLK | LKM_INVVALBLK)) { + dlm_error(DLM_BADPARAM); + return DLM_BADPARAM; + } + + if ((flags & (LKM_VALBLK | LKM_CANCEL)) == (LKM_VALBLK | LKM_CANCEL)) { + mlog(0, "VALBLK given with CANCEL: ignoring VALBLK\n"); + flags &= ~LKM_VALBLK; + } + + if (!lksb->lockid || !lksb->lockid->lockres) { + dlm_error(DLM_BADPARAM); + return DLM_BADPARAM; + } + + lock = lksb->lockid; + BUG_ON(!lock); + dlm_lock_get(lock); + + res = lock->lockres; + BUG_ON(!res); + dlm_lockres_get(res); +retry: + call_ast = 0; + /* need to retry up here because owner may have changed */ + mlog(0, "lock=%p res=%p\n", lock, res); + + spin_lock(&res->spinlock); + is_master = (res->owner == dlm->node_num); + if (flags & LKM_VALBLK && lock->ml.type != LKM_EXMODE) + flags &= ~LKM_VALBLK; + spin_unlock(&res->spinlock); + + if (is_master) { + status = dlmunlock_master(dlm, res, lock, lksb, flags, + &call_ast); + mlog(0, "done calling dlmunlock_master: returned %d, " + "call_ast is %d\n", status, call_ast); + } else { + status = dlmunlock_remote(dlm, res, lock, lksb, flags, + &call_ast); + mlog(0, "done calling dlmunlock_remote: returned %d, " + "call_ast is %d\n", status, call_ast); + } + + if (status == DLM_RECOVERING || + status == DLM_MIGRATING || + status == DLM_FORWARD || + status == DLM_NOLOCKMGR) { + + /* We want to go away for a tiny bit to allow recovery + * / migration to complete on this resource. I don't + * know of any wait queue we could sleep on as this + * may be happening on another node. Perhaps the + * proper solution is to queue up requests on the + * other end? */ + + /* do we want to yield(); ?? */ + msleep(50); + + mlog(0, "retrying unlock due to pending recovery/" + "migration/in-progress/reconnect\n"); + goto retry; + } + + if (call_ast) { + mlog(0, "calling unlockast(%p, %d)\n", data, status); + if (is_master) { + /* it is possible that there is one last bast + * pending. make sure it is flushed, then + * call the unlockast. + * not an issue if this is a mastered remotely, + * since this lock has been removed from the + * lockres queues and cannot be found. */ + dlm_kick_thread(dlm, NULL); + wait_event(dlm->ast_wq, + dlm_lock_basts_flushed(dlm, lock)); + } + (*unlockast)(data, status); + } + + if (status == DLM_CANCELGRANT) + status = DLM_NORMAL; + + if (status == DLM_NORMAL) { + mlog(0, "kicking the thread\n"); + dlm_kick_thread(dlm, res); + } else + dlm_error(status); + + dlm_lockres_calc_usage(dlm, res); + dlm_lockres_put(res); + dlm_lock_put(lock); + + mlog(0, "returning status=%d!\n", status); + return status; +} +EXPORT_SYMBOL_GPL(dlmunlock); + -- cgit v1.3.1