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, 86 insertions, 0 deletions

diff --git a/Documentation/driver-api/sync_file.rst b/Documentation/driver-api/sync_file.rst
new file mode 100644
index 000000000..496fb2c3b
--- /dev/null
+++ b/Documentation/driver-api/sync_file.rst
@@ -0,0 +1,86 @@
+===================
+Sync File API Guide
+===================
+
+:Author: Gustavo Padovan <gustavo at padovan dot org>
+
+This document serves as a guide for device drivers writers on what the
+sync_file API is, and how drivers can support it. Sync file is the carrier of
+the fences(struct dma_fence) that are needed to synchronize between drivers or
+across process boundaries.
+
+The sync_file API is meant to be used to send and receive fence information
+to/from userspace. It enables userspace to do explicit fencing, where instead
+of attaching a fence to the buffer a producer driver (such as a GPU or V4L
+driver) sends the fence related to the buffer to userspace via a sync_file.
+
+The sync_file then can be sent to the consumer (DRM driver for example), that
+will not use the buffer for anything before the fence(s) signals, i.e., the
+driver that issued the fence is not using/processing the buffer anymore, so it
+signals that the buffer is ready to use. And vice-versa for the consumer ->
+producer part of the cycle.
+
+Sync files allows userspace awareness on buffer sharing synchronization between
+drivers.
+
+Sync file was originally added in the Android kernel but current Linux Desktop
+can benefit a lot from it.
+
+in-fences and out-fences
+------------------------
+
+Sync files can go either to or from userspace. When a sync_file is sent from
+the driver to userspace we call the fences it contains 'out-fences'. They are
+related to a buffer that the driver is processing or is going to process, so
+the driver creates an out-fence to be able to notify, through
+dma_fence_signal(), when it has finished using (or processing) that buffer.
+Out-fences are fences that the driver creates.
+
+On the other hand if the driver receives fence(s) through a sync_file from
+userspace we call these fence(s) 'in-fences'. Receiving in-fences means that
+we need to wait for the fence(s) to signal before using any buffer related to
+the in-fences.
+
+Creating Sync Files
+-------------------
+
+When a driver needs to send an out-fence userspace it creates a sync_file.
+
+Interface::
+
+	struct sync_file *sync_file_create(struct dma_fence *fence);
+
+The caller pass the out-fence and gets back the sync_file. That is just the
+first step, next it needs to install an fd on sync_file->file. So it gets an
+fd::
+
+	fd = get_unused_fd_flags(O_CLOEXEC);
+
+and installs it on sync_file->file::
+
+	fd_install(fd, sync_file->file);
+
+The sync_file fd now can be sent to userspace.
+
+If the creation process fail, or the sync_file needs to be released by any
+other reason fput(sync_file->file) should be used.
+
+Receiving Sync Files from Userspace
+-----------------------------------
+
+When userspace needs to send an in-fence to the driver it passes file descriptor
+of the Sync File to the kernel. The kernel can then retrieve the fences
+from it.
+
+Interface::
+
+	struct dma_fence *sync_file_get_fence(int fd);
+
+
+The returned reference is owned by the caller and must be disposed of
+afterwards using dma_fence_put(). In case of error, a NULL is returned instead.
+
+References:
+
+1. struct sync_file in include/linux/sync_file.h
+2. All interfaces mentioned above defined in include/linux/sync_file.h