sched/cache: Keep nr_pref_llc_running in the runnable domain, to fix LLC mis-scheduling buggrafted

alb_break_llc() decides whether to break LLC preference during active
load balance. It does so by testing that every runnable fair task on the
source rq prefers its LLC:

	env->src_rq->nr_pref_llc_running == env->src_rq->cfs.h_nr_runnable

But the two counters cover different sets. nr_pref_llc_running is updated
in account_llc_enqueue()/account_llc_dequeue(), next to cfs_rq->nr_queued,
so it follows queued tasks. h_nr_runnable is updated in set_delayed()/
clear_delayed() and drops delay-dequeued tasks.

So under DELAY_DEQUEUE, a preferring task that goes to sleep stays counted
in nr_pref_llc_running while h_nr_runnable falls. The equality then breaks,
alb_break_llc() returns false, and active balance is free to pull a task
off its preferred LLC. Active balance only moves runnable tasks, and this
is the only LLC check it consults: once the stopper runs, LBF_ACTIVE_LB
skips the per-task test in can_migrate_task(). The runnable set is the one
we want.

Fix it on the counter side. A task should be counted in
nr_pref_llc_running exactly while it is both queued on its preferred LLC
(pref_llc_queued) and runnable (!sched_delayed). Define that membership
once in task_pref_llc_runnable(), and adjust the counter only through
pref_llc_running_inc()/pref_llc_running_dec() from the four sites that
change either input: account_llc_enqueue(), account_llc_dequeue(),
set_delayed() and clear_delayed(). Gating every update on the same
predicate keeps the delay, wake and dequeue paths from double-counting
or underflowing; see the comments at those sites for the ordering.

nr_llc_running and sd->llc_counts are not touched and stay on queued
semantics.

Fixes: 714059f79ff0 ("sched/cache: Handle moving single tasks to/from their preferred LLC")
Closes: https://lore.kernel.org/lkml/20260827135000.735138-1-zhanxusheng@xiaomi.com/
Reported-by: Zhan Xusheng <zhanxusheng@xiaomi.com>
Suggested-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kayra Cizmeci <kayracizmeci@gmail.com>
Cc: <stable@kernel.org> # v7.2.x
Link: https://patch.msgid.link/06af61afedac32e6477f57feb4d658f6c411c3af.1790035273.git.tim.c.chen@linux.intel.com

1 files changed, 126 insertions, 0 deletions

diff --git a/Documentation/filesystems/fuse/uapi/fuse-uapi-io-uring.rst b/Documentation/filesystems/fuse/uapi/fuse-uapi-io-uring.rst
new file mode 100644
index 000000000..8367be7ea
--- /dev/null
+++ b/Documentation/filesystems/fuse/uapi/fuse-uapi-io-uring.rst
@@ -0,0 +1,126 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+=====================================
+FUSE-over-io-uring uapi documentation
+=====================================
+
+Commands
+========
+
+``enum fuse_uring_cmd``:
+
+``FUSE_IO_URING_CMD_ADD_QUEUE``
+  Create a queue identified by ``fuse_uring_cmd_req.qid``. Queue-wide
+  options are passed in ``fuse_uring_cmd_req.flags``:
+
+  ``FUSE_URING_ZERO_COPY``
+    Enable zero-copy on this queue. Requires ``CAP_SYS_ADMIN`` and a buffer
+    pool, which is added separately via ``ADD_BUFPOOL`` before registering
+    entries (see `Zero-copy`_).
+
+``FUSE_IO_URING_CMD_ADD_BUFPOOL``
+  Register the payload buffer pool for an existing queue. The server provides
+  a single contiguous region in ``fuse_uring_cmd_req.bufpool.uaddr`` /
+  ``.len``. This command must be issued after ``ADD_QUEUE`` and before
+  registering any payload-carrying entries on that queue.
+  ``fuse_uring_cmd_req.flags`` must be 0. Submitting this command with
+  ``IORING_URING_CMD_FIXED`` marks the pool as registered, which avoids per
+  i/o pinning/unpinning and mapping overhead (see `Buffer pools`_).
+
+``FUSE_IO_URING_CMD_REGISTER``
+  Register a ring entry (a long-lived SQE that carries the request header
+  iovec). For a zero-copy queue, ``fuse_uring_cmd_req.ent_zero_copy_buf_index``
+  indicates the reserved registered buffer table slot this entry uses for
+  zero-copy (see `Zero-copy`_).
+
+``FUSE_IO_URING_CMD_COMMIT_AND_FETCH``
+  Commit the reply for a completed request and fetch the next one. The
+  request is identified by ``fuse_uring_cmd_req.commit_id`` (the value the
+  kernel reported in ``fuse_uring_ent_in_out.commit_id``).
+
+Structures
+==========
+
+``struct fuse_uring_cmd_req`` (80-byte SQE command area):
+
+============================  ==================================================
+Field                         Meaning
+============================  ==================================================
+``flags``                     Command-specific flags (see each command).
+``commit_id``                 Request id, for ``COMMIT_AND_FETCH``.
+``qid``                       Queue index.
+``bufpool.uaddr``             Pool base address, for ``ADD_BUFPOOL``.
+``bufpool.len``               Pool length in bytes, for ``ADD_BUFPOOL``.
+``bufpool.reserved``          Must be 0, for ``ADD_BUFPOOL``.
+``ent_zero_copy_buf_index``   Per-entry zero-copy slot, for ``REGISTER``.
+============================  ==================================================
+
+``struct fuse_uring_ent_in_out`` (reported by the kernel per request):
+
+============================  ==================================================
+Field                         Meaning
+============================  ==================================================
+``flags``                     ``FUSE_URING_ENT_ZERO_COPY`` if zero-copied.
+``commit_id``                 Id to echo back in ``COMMIT_AND_FETCH``.
+``payload_sz``                Total payload size in bytes (see `Zero-copy`_).
+``offset``                    Payload buffer offset within the pool.
+============================  ==================================================
+
+Buffer pools
+============
+Setup:
+
+* Issue ``ADD_QUEUE`` for the qid.
+* Issue ``ADD_BUFPOOL`` with ``bufpool.uaddr`` and ``bufpool.len`` pointing
+  at the region.
+* Register entries with ``REGISTER``.
+
+For every request that has a payload, the kernel reports where the payload
+lives in ``struct fuse_uring_ent_in_out`` (part of
+``struct fuse_uring_req_header``):
+
+``offset``
+  Byte offset, within the pool region, for this request's payload buffer.
+  The server adds this to the pool base address to locate the payload.
+
+``payload_sz``
+  Number of payload bytes for this request.
+
+To use registered buffers, the server registers the pool region with io_uring
+and submits ``ADD_BUFPOOL`` with ``IORING_URING_CMD_FIXED`` set in
+``sqe->uring_cmd_flags`` and the index of the registered bufpool in
+``sqe->buf_index``. Every SQE the server submits afterwards must follow the
+same fixed-buffer protocol, carrying ``IORING_URING_CMD_FIXED`` and that same
+``sqe->buf_index``. The same registered buffer can be reused for the server's
+backing-store I/O as well (e.g. ``IORING_OP_READ_FIXED`` /
+``IORING_OP_WRITE_FIXED``).
+
+Zero-copy
+=========
+Requirements:
+
+* The server must be privileged (``CAP_SYS_ADMIN``).
+* A zero-copy queue: ``ADD_QUEUE`` with the ``FUSE_URING_ZERO_COPY`` flag set.
+* A buffer pool: ``ADD_BUFPOOL``.
+* For each entry, ``REGISTER`` with ``ent_zero_copy_buf_index`` set to the
+  index this entry uses in the server's io_uring registered-buffer table.
+  This is where the kernel registers the request's pages for the server to
+  access (it is separate from the payload pool). On a non-zero-copy queue this
+  field must be 0.
+
+Zero-copy is selected per open file. The server sets the open-file flag in
+the ``FUSE_OPEN`` / ``FUSE_CREATE`` reply:
+
+``FOPEN_IO_URING_ZERO_COPY``
+  Reads/writes on this open file should use zero-copy.
+
+For a request that is zero-copied, the kernel sets ``FUSE_URING_ENT_ZERO_COPY``
+in ``fuse_uring_ent_in_out.flags`` and places the request's pages at the
+entry's ``ent_zero_copy_buf_index``. The server then issues
+``IORING_OP_READ_FIXED`` / ``IORING_OP_WRITE_FIXED`` against that index to
+transfer the data directly to/from the client's pages.
+
+For such a request, ``payload_sz`` includes the zero-copied page bytes
+(transferred via the registered buffer at ``ent_zero_copy_buf_index``). Any
+non-page-backed args (e.g. op headers) are still copied through the pool
+payload buffer at ``offset``.