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, 75 insertions, 0 deletions

diff --git a/Documentation/filesystems/dnotify.rst b/Documentation/filesystems/dnotify.rst
new file mode 100644
index 000000000..a28a1f9ef
--- /dev/null
+++ b/Documentation/filesystems/dnotify.rst
@@ -0,0 +1,75 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+============================
+Linux Directory Notification
+============================
+
+	   Stephen Rothwell <sfr@canb.auug.org.au>
+
+The intention of directory notification is to allow user applications
+to be notified when a directory, or any of the files in it, are changed.
+The basic mechanism involves the application registering for notification
+on a directory using a fcntl(2) call and the notifications themselves
+being delivered using signals.
+
+The application decides which "events" it wants to be notified about.
+The currently defined events are:
+
+	=========	=====================================================
+	DN_ACCESS	A file in the directory was accessed (read)
+	DN_MODIFY	A file in the directory was modified (write,truncate)
+	DN_CREATE	A file was created in the directory
+	DN_DELETE	A file was unlinked from directory
+	DN_RENAME	A file in the directory was renamed
+	DN_ATTRIB	A file in the directory had its attributes
+			changed (chmod,chown)
+	=========	=====================================================
+
+Usually, the application must reregister after each notification, but
+if DN_MULTISHOT is or'ed with the event mask, then the registration will
+remain until explicitly removed (by registering for no events).
+
+By default, SIGIO will be delivered to the process and no other useful
+information.  However, if the F_SETSIG fcntl(2) call is used to let the
+kernel know which signal to deliver, a siginfo structure will be passed to
+the signal handler and the si_fd member of that structure will contain the
+file descriptor associated with the directory in which the event occurred.
+
+Preferably the application will choose one of the real time signals
+(SIGRTMIN + <n>) so that the notifications may be queued.  This is
+especially important if DN_MULTISHOT is specified.  Note that SIGRTMIN
+is often blocked, so it is better to use (at least) SIGRTMIN + 1.
+
+Implementation expectations (features and bugs :-))
+---------------------------------------------------
+
+The notification should work for any local access to files even if the
+actual file system is on a remote server.  This implies that remote
+access to files served by local user mode servers should be notified.
+Also, remote accesses to files served by a local kernel NFS server should
+be notified.
+
+In order to make the impact on the file system code as small as possible,
+the problem of hard links to files has been ignored.  So if a file (x)
+exists in two directories (a and b) then a change to the file using the
+name "a/x" should be notified to a program expecting notifications on
+directory "a", but will not be notified to one expecting notifications on
+directory "b".
+
+Also, files that are unlinked, will still cause notifications in the
+last directory that they were linked to.
+
+Configuration
+-------------
+
+Dnotify is controlled via the CONFIG_DNOTIFY configuration option.  When
+disabled, fcntl(fd, F_NOTIFY, ...) will return -EINVAL.
+
+Example
+-------
+See tools/testing/selftests/filesystems/dnotify_test.c for an example.
+
+NOTE
+----
+Beginning with Linux 2.6.13, dnotify has been replaced by inotify.
+See Documentation/filesystems/inotify.rst for more information on it.