From a5f7a5bb3b7f28ba7e4fa246775b29a0e5537255 Mon Sep 17 00:00:00 2001 From: Jinke Han Date: Tue, 8 Sep 2026 15:37:42 +0800 Subject: x86/kprobes: Fix crash when probing CS CALL instructions When using eBPF to probe CS CALL instructions within a function, a crash can be triggered. The eBPF tool probes offset 257 of the __hrtimer_run_queues() function: <__hrtimer_run_queues+249>: nopl 0x0(%rax,%rax,1) <__hrtimer_run_queues+254>: mov %r14,%rdi <__hrtimer_run_queues+257>: cs call <__x86_indirect_thunk_r12> <__hrtimer_run_queues+263>: mov %eax,%r12d <__hrtimer_run_queues+266>: xchg %ax,%ax <__hrtimer_run_queues+268>: mov %r13,%rdi Which triggers this crash: BUG: unable to handle page fault for address: 00000000000f41c9 #PF: supervisor write access in kernel mode #PF: error_code(0x0002) - not-present page PGD 0 P4D 0 Oops: 0002 [#1] SMP NOPTI CPU: 1 PID: 0 Comm: swapper/1 Kdump: loaded Tainted: P RIP: 0010:__hrtimer_run_queues+0x106/0x230 Note that __hrtimer_run_queues+0x106 is __hrtimer_run_queues+262, which is at the 6th byte of the above CS CALL instruction. Since the CS CALL instruction occupies 6 bytes, the exception occurred in the middle of that call instruction. The root cause is that when using eBPF tools to probe in the middle of a function, a kprobe with INT3 is used as the underlying implementation. During single-step emulation of the original CALL instruction, int3_emulate_call() assumes that the probed CALL instruction is 5 bytes long. However, the actual CS-prefixed CALL instruction occupies 6 bytes, so it constructs an incorrect exception return address. When the CPU returns from the kprobe handler, the next instruction to be executed is at the address of the last byte of that CS CALL instruction. Coincidentally, starting from that address, the CPU fetches and decodes a completely different instruction, which ultimately triggers a kernel crash. Fix the issue by using the actual instruction length obtained from the instruction decoder when constructing the exception return address, rather than relying on the hardcoded CALL_INSN_SIZE macro. [ mingo: Refined the changelog ] Fixes: 6256e668b7af ("x86/kprobes: Use int3 instead of debug trap for single-step") Suggested-by: Masami Hiramatsu (Google) Signed-off-by: Jinke Han Signed-off-by: Ingo Molnar Reviewed-by: Masami Hiramatsu (Google) Acked-by: Yafang Shao Acked-by: Borislav Petkov Cc: Peter Zijlstra Link: https://patch.msgid.link/20260908073742.GA10517@didi-ThinkCentre-M920t-N000 --- net/sched/ematch.c | 550 +++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 550 insertions(+) create mode 100644 net/sched/ematch.c (limited to 'net/sched/ematch.c') diff --git a/net/sched/ematch.c b/net/sched/ematch.c new file mode 100644 index 000000000..5c1235e60 --- /dev/null +++ b/net/sched/ematch.c @@ -0,0 +1,550 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * net/sched/ematch.c Extended Match API + * + * Authors: Thomas Graf + * + * ========================================================================== + * + * An extended match (ematch) is a small classification tool not worth + * writing a full classifier for. Ematches can be interconnected to form + * a logic expression and get attached to classifiers to extend their + * functionatlity. + * + * The userspace part transforms the logic expressions into an array + * consisting of multiple sequences of interconnected ematches separated + * by markers. Precedence is implemented by a special ematch kind + * referencing a sequence beyond the marker of the current sequence + * causing the current position in the sequence to be pushed onto a stack + * to allow the current position to be overwritten by the position referenced + * in the special ematch. Matching continues in the new sequence until a + * marker is reached causing the position to be restored from the stack. + * + * Example: + * A AND (B1 OR B2) AND C AND D + * + * ------->-PUSH------- + * -->-- / -->-- \ -->-- + * / \ / / \ \ / \ + * +-------+-------+-------+-------+-------+--------+ + * | A AND | B AND | C AND | D END | B1 OR | B2 END | + * +-------+-------+-------+-------+-------+--------+ + * \ / + * --------<-POP--------- + * + * where B is a virtual ematch referencing to sequence starting with B1. + * + * ========================================================================== + * + * How to write an ematch in 60 seconds + * ------------------------------------ + * + * 1) Provide a matcher function: + * static int my_match(struct sk_buff *skb, struct tcf_ematch *m, + * struct tcf_pkt_info *info) + * { + * struct mydata *d = (struct mydata *) m->data; + * + * if (...matching goes here...) + * return 1; + * else + * return 0; + * } + * + * 2) Fill out a struct tcf_ematch_ops: + * static struct tcf_ematch_ops my_ops = { + * .kind = unique id, + * .datalen = sizeof(struct mydata), + * .match = my_match, + * .owner = THIS_MODULE, + * }; + * + * 3) Register/Unregister your ematch: + * static int __init init_my_ematch(void) + * { + * return tcf_em_register(&my_ops); + * } + * + * static void __exit exit_my_ematch(void) + * { + * tcf_em_unregister(&my_ops); + * } + * + * module_init(init_my_ematch); + * module_exit(exit_my_ematch); + * + * 4) By now you should have two more seconds left, barely enough to + * open up a beer to watch the compilation going. + */ + +#include +#include +#include +#include +#include +#include +#include +#include + +static LIST_HEAD(ematch_ops); +static DEFINE_RWLOCK(ematch_mod_lock); + +static struct tcf_ematch_ops *tcf_em_lookup(u16 kind) +{ + struct tcf_ematch_ops *e = NULL; + + read_lock(&ematch_mod_lock); + list_for_each_entry(e, &ematch_ops, link) { + if (kind == e->kind) { + if (!try_module_get(e->owner)) + e = NULL; + read_unlock(&ematch_mod_lock); + return e; + } + } + read_unlock(&ematch_mod_lock); + + return NULL; +} + +/** + * tcf_em_register - register an extended match + * + * @ops: ematch operations lookup table + * + * This function must be called by ematches to announce their presence. + * The given @ops must have kind set to a unique identifier and the + * callback match() must be implemented. All other callbacks are optional + * and a fallback implementation is used instead. + * + * Returns -EEXISTS if an ematch of the same kind has already registered. + */ +int tcf_em_register(struct tcf_ematch_ops *ops) +{ + int err = -EEXIST; + struct tcf_ematch_ops *e; + + if (ops->match == NULL) + return -EINVAL; + + write_lock(&ematch_mod_lock); + list_for_each_entry(e, &ematch_ops, link) + if (ops->kind == e->kind) + goto errout; + + list_add_tail(&ops->link, &ematch_ops); + err = 0; +errout: + write_unlock(&ematch_mod_lock); + return err; +} +EXPORT_SYMBOL(tcf_em_register); + +/** + * tcf_em_unregister - unregister and extended match + * + * @ops: ematch operations lookup table + * + * This function must be called by ematches to announce their disappearance + * for examples when the module gets unloaded. The @ops parameter must be + * the same as the one used for registration. + * + * Returns -ENOENT if no matching ematch was found. + */ +void tcf_em_unregister(struct tcf_ematch_ops *ops) +{ + write_lock(&ematch_mod_lock); + list_del(&ops->link); + write_unlock(&ematch_mod_lock); +} +EXPORT_SYMBOL(tcf_em_unregister); + +static inline struct tcf_ematch *tcf_em_get_match(struct tcf_ematch_tree *tree, + int index) +{ + return &tree->matches[index]; +} + + +static int tcf_em_validate(struct tcf_proto *tp, + struct tcf_ematch_tree_hdr *tree_hdr, + struct tcf_ematch *em, struct nlattr *nla, int idx) +{ + int err = -EINVAL; + struct tcf_ematch_hdr *em_hdr = nla_data(nla); + int data_len = nla_len(nla) - sizeof(*em_hdr); + void *data = (void *) em_hdr + sizeof(*em_hdr); + struct net *net = tp->chain->block->net; + + if (!TCF_EM_REL_VALID(em_hdr->flags)) + goto errout; + + if (em_hdr->kind == TCF_EM_CONTAINER) { + /* Special ematch called "container", carries an index + * referencing an external ematch sequence. + */ + u32 ref; + + if (data_len < sizeof(ref)) + goto errout; + ref = *(u32 *) data; + + if (ref >= tree_hdr->nmatches) + goto errout; + + /* We do not allow backward jumps to avoid loops and jumps + * to our own position are of course illegal. + */ + if (ref <= idx) + goto errout; + + + em->data = ref; + } else { + /* Note: This lookup will increase the module refcnt + * of the ematch module referenced. In case of a failure, + * a destroy function is called by the underlying layer + * which automatically releases the reference again, therefore + * the module MUST not be given back under any circumstances + * here. Be aware, the destroy function assumes that the + * module is held if the ops field is non zero. + */ + em->ops = tcf_em_lookup(em_hdr->kind); + + if (em->ops == NULL) { + err = -ENOENT; +#ifdef CONFIG_MODULES + __rtnl_unlock(); + request_module("ematch-kind-%u", em_hdr->kind); + rtnl_lock(); + em->ops = tcf_em_lookup(em_hdr->kind); + if (em->ops) { + /* We dropped the RTNL mutex in order to + * perform the module load. Tell the caller + * to replay the request. + */ + module_put(em->ops->owner); + em->ops = NULL; + err = -EAGAIN; + } +#endif + goto errout; + } + + /* ematch module provides expected length of data, so we + * can do a basic sanity check. + */ + if (em->ops->datalen && data_len < em->ops->datalen) + goto errout; + + if (em->ops->change) { + err = -EINVAL; + if (em_hdr->flags & TCF_EM_SIMPLE) + goto errout; + err = em->ops->change(net, data, data_len, em); + if (err < 0) + goto errout; + } else if (data_len > 0) { + /* ematch module doesn't provide an own change + * procedure and expects us to allocate and copy + * the ematch data. + * + * TCF_EM_SIMPLE may be specified stating that the + * data only consists of a u32 integer and the module + * does not expected a memory reference but rather + * the value carried. + */ + if (em_hdr->flags & TCF_EM_SIMPLE) { + if (em->ops->datalen > 0) + goto errout; + if (data_len < sizeof(u32)) + goto errout; + em->data = *(u32 *) data; + } else { + void *v = kmemdup(data, data_len, GFP_KERNEL); + if (v == NULL) { + err = -ENOBUFS; + goto errout; + } + em->data = (unsigned long) v; + } + em->datalen = data_len; + } + } + + em->matchid = em_hdr->matchid; + em->flags = em_hdr->flags; + em->net = net; + + err = 0; +errout: + return err; +} + +static const struct nla_policy em_policy[TCA_EMATCH_TREE_MAX + 1] = { + [TCA_EMATCH_TREE_HDR] = { .len = sizeof(struct tcf_ematch_tree_hdr) }, + [TCA_EMATCH_TREE_LIST] = { .type = NLA_NESTED }, +}; + +/** + * tcf_em_tree_validate - validate ematch config TLV and build ematch tree + * + * @tp: classifier kind handle + * @nla: ematch tree configuration TLV + * @tree: destination ematch tree variable to store the resulting + * ematch tree. + * + * This function validates the given configuration TLV @nla and builds an + * ematch tree in @tree. The resulting tree must later be copied into + * the private classifier data using tcf_em_tree_change(). You MUST NOT + * provide the ematch tree variable of the private classifier data directly, + * the changes would not be locked properly. + * + * Returns a negative error code if the configuration TLV contains errors. + */ +int tcf_em_tree_validate(struct tcf_proto *tp, struct nlattr *nla, + struct tcf_ematch_tree *tree) +{ + int idx, list_len, matches_len, err; + struct nlattr *tb[TCA_EMATCH_TREE_MAX + 1]; + struct nlattr *rt_match, *rt_hdr, *rt_list; + struct tcf_ematch_tree_hdr *tree_hdr; + struct tcf_ematch *em; + + memset(tree, 0, sizeof(*tree)); + if (!nla) + return 0; + + err = nla_parse_nested_deprecated(tb, TCA_EMATCH_TREE_MAX, nla, + em_policy, NULL); + if (err < 0) + goto errout; + + err = -EINVAL; + rt_hdr = tb[TCA_EMATCH_TREE_HDR]; + rt_list = tb[TCA_EMATCH_TREE_LIST]; + + if (rt_hdr == NULL || rt_list == NULL) + goto errout; + + tree_hdr = nla_data(rt_hdr); + memcpy(&tree->hdr, tree_hdr, sizeof(*tree_hdr)); + + rt_match = nla_data(rt_list); + list_len = nla_len(rt_list); + matches_len = tree_hdr->nmatches * sizeof(*em); + + tree->matches = kzalloc(matches_len, GFP_KERNEL); + if (tree->matches == NULL) + goto errout; + + /* We do not use nla_parse_nested here because the maximum + * number of attributes is unknown. This saves us the allocation + * for a tb buffer which would serve no purpose at all. + * + * The array of rt attributes is parsed in the order as they are + * provided, their type must be incremental from 1 to n. Even + * if it does not serve any real purpose, a failure of sticking + * to this policy will result in parsing failure. + */ + for (idx = 0; nla_ok(rt_match, list_len); idx++) { + err = -EINVAL; + + if (rt_match->nla_type != (idx + 1)) + goto errout_abort; + + if (idx >= tree_hdr->nmatches) + goto errout_abort; + + if (nla_len(rt_match) < sizeof(struct tcf_ematch_hdr)) + goto errout_abort; + + em = tcf_em_get_match(tree, idx); + + err = tcf_em_validate(tp, tree_hdr, em, rt_match, idx); + if (err < 0) + goto errout_abort; + + rt_match = nla_next(rt_match, &list_len); + } + + /* Check if the number of matches provided by userspace actually + * complies with the array of matches. The number was used for + * the validation of references and a mismatch could lead to + * undefined references during the matching process. + */ + if (idx != tree_hdr->nmatches) { + err = -EINVAL; + goto errout_abort; + } + + err = 0; +errout: + return err; + +errout_abort: + tcf_em_tree_destroy(tree); + return err; +} +EXPORT_SYMBOL(tcf_em_tree_validate); + +/** + * tcf_em_tree_destroy - destroy an ematch tree + * + * @tree: ematch tree to be deleted + * + * This functions destroys an ematch tree previously created by + * tcf_em_tree_validate()/tcf_em_tree_change(). You must ensure that + * the ematch tree is not in use before calling this function. + */ +void tcf_em_tree_destroy(struct tcf_ematch_tree *tree) +{ + int i; + + if (tree->matches == NULL) + return; + + for (i = 0; i < tree->hdr.nmatches; i++) { + struct tcf_ematch *em = tcf_em_get_match(tree, i); + + if (em->ops) { + if (em->ops->destroy) + em->ops->destroy(em); + else if (!tcf_em_is_simple(em)) + kfree((void *) em->data); + module_put(em->ops->owner); + } + } + + tree->hdr.nmatches = 0; + kfree(tree->matches); + tree->matches = NULL; +} +EXPORT_SYMBOL(tcf_em_tree_destroy); + +/** + * tcf_em_tree_dump - dump ematch tree into a rtnl message + * + * @skb: skb holding the rtnl message + * @tree: ematch tree to be dumped + * @tlv: TLV type to be used to encapsulate the tree + * + * This function dumps a ematch tree into a rtnl message. It is valid to + * call this function while the ematch tree is in use. + * + * Returns -1 if the skb tailroom is insufficient. + */ +int tcf_em_tree_dump(struct sk_buff *skb, struct tcf_ematch_tree *tree, int tlv) +{ + int i; + u8 *tail; + struct nlattr *top_start; + struct nlattr *list_start; + + top_start = nla_nest_start_noflag(skb, tlv); + if (top_start == NULL) + goto nla_put_failure; + + if (nla_put(skb, TCA_EMATCH_TREE_HDR, sizeof(tree->hdr), &tree->hdr)) + goto nla_put_failure; + + list_start = nla_nest_start_noflag(skb, TCA_EMATCH_TREE_LIST); + if (list_start == NULL) + goto nla_put_failure; + + tail = skb_tail_pointer(skb); + for (i = 0; i < tree->hdr.nmatches; i++) { + struct nlattr *match_start = (struct nlattr *)tail; + struct tcf_ematch *em = tcf_em_get_match(tree, i); + struct tcf_ematch_hdr em_hdr = { + .kind = em->ops ? em->ops->kind : TCF_EM_CONTAINER, + .matchid = em->matchid, + .flags = em->flags + }; + + if (nla_put(skb, i + 1, sizeof(em_hdr), &em_hdr)) + goto nla_put_failure; + + if (em->ops && em->ops->dump) { + if (em->ops->dump(skb, em) < 0) + goto nla_put_failure; + } else if (tcf_em_is_container(em) || tcf_em_is_simple(em)) { + u32 u = em->data; + nla_put_nohdr(skb, sizeof(u), &u); + } else if (em->datalen > 0) + nla_put_nohdr(skb, em->datalen, (void *) em->data); + + tail = skb_tail_pointer(skb); + match_start->nla_len = tail - (u8 *)match_start; + } + + nla_nest_end(skb, list_start); + nla_nest_end(skb, top_start); + + return 0; + +nla_put_failure: + return -1; +} +EXPORT_SYMBOL(tcf_em_tree_dump); + +static inline int tcf_em_match(struct sk_buff *skb, struct tcf_ematch *em, + struct tcf_pkt_info *info) +{ + int r = em->ops->match(skb, em, info); + + return tcf_em_is_inverted(em) ? !r : r; +} + +/* Do not use this function directly, use tcf_em_tree_match instead */ +int __tcf_em_tree_match(struct sk_buff *skb, struct tcf_ematch_tree *tree, + struct tcf_pkt_info *info) +{ + int stackp = 0, match_idx = 0, res = 0; + struct tcf_ematch *cur_match; + int stack[CONFIG_NET_EMATCH_STACK]; + +proceed: + while (match_idx < tree->hdr.nmatches) { + cur_match = tcf_em_get_match(tree, match_idx); + + if (tcf_em_is_container(cur_match)) { + if (unlikely(stackp >= CONFIG_NET_EMATCH_STACK)) + goto stack_overflow; + + stack[stackp++] = match_idx; + match_idx = cur_match->data; + goto proceed; + } + + res = tcf_em_match(skb, cur_match, info); + + if (tcf_em_early_end(cur_match, res)) + break; + + match_idx++; + } + +pop_stack: + if (stackp > 0) { + match_idx = stack[--stackp]; + cur_match = tcf_em_get_match(tree, match_idx); + + if (tcf_em_is_inverted(cur_match)) + res = !res; + + if (tcf_em_early_end(cur_match, res)) { + goto pop_stack; + } else { + match_idx++; + goto proceed; + } + } + + return res; + +stack_overflow: + net_warn_ratelimited("tc ematch: local stack overflow, increase NET_EMATCH_STACK\n"); + return -1; +} +EXPORT_SYMBOL(__tcf_em_tree_match); -- cgit v1.3.1