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, 85 insertions, 0 deletions

diff --git a/Documentation/dev-tools/sparse.rst b/Documentation/dev-tools/sparse.rst
new file mode 100644
index 000000000..37b201708
--- /dev/null
+++ b/Documentation/dev-tools/sparse.rst
@@ -0,0 +1,85 @@
+.. Copyright 2004 Linus Torvalds
+.. Copyright 2004 Pavel Machek <pavel@ucw.cz>
+.. Copyright 2006 Bob Copeland <me@bobcopeland.com>
+
+Sparse
+======
+
+Sparse is a semantic checker for C programs; it can be used to find a
+number of potential problems with kernel code.  See
+https://lwn.net/Articles/689907/ for an overview of sparse; this document
+contains some kernel-specific sparse information.
+More information on sparse, mainly about its internals, can be found in
+its official pages at https://sparse.docs.kernel.org.
+
+
+Using sparse for typechecking
+-----------------------------
+
+"__bitwise" is a type attribute, so you have to do something like this::
+
+        typedef int __bitwise pm_request_t;
+
+        enum pm_request {
+                PM_SUSPEND = (__force pm_request_t) 1,
+                PM_RESUME = (__force pm_request_t) 2
+        };
+
+which makes PM_SUSPEND and PM_RESUME "bitwise" integers (the "__force" is
+there because sparse will complain about casting to/from a bitwise type,
+but in this case we really _do_ want to force the conversion). And because
+the enum values are all the same type, now "enum pm_request" will be that
+type too.
+
+And with gcc, all the "__bitwise"/"__force stuff" goes away, and it all
+ends up looking just like integers to gcc.
+
+Quite frankly, you don't need the enum there. The above all really just
+boils down to one special "int __bitwise" type.
+
+So the simpler way is to just do::
+
+        typedef int __bitwise pm_request_t;
+
+        #define PM_SUSPEND ((__force pm_request_t) 1)
+        #define PM_RESUME ((__force pm_request_t) 2)
+
+and you now have all the infrastructure needed for strict typechecking.
+
+One small note: the constant integer "0" is special. You can use a
+constant zero as a bitwise integer type without sparse ever complaining.
+This is because "bitwise" (as the name implies) was designed for making
+sure that bitwise types don't get mixed up (little-endian vs big-endian
+vs cpu-endian vs whatever), and there the constant "0" really _is_
+special.
+
+Getting sparse
+--------------
+
+You can get tarballs of the latest released versions from:
+https://www.kernel.org/pub/software/devel/sparse/dist/
+
+Alternatively, you can get snapshots of the latest development version
+of sparse using git to clone::
+
+        git://git.kernel.org/pub/scm/devel/sparse/sparse.git
+
+Once you have it, just do::
+
+        make
+        make install
+
+as a regular user, and it will install sparse in your ~/bin directory.
+
+Using sparse
+------------
+
+Do a kernel make with "make C=1" to run sparse on all the C files that get
+recompiled, or use "make C=2" to run sparse on the files whether they need to
+be recompiled or not.  The latter is a fast way to check the whole tree if you
+have already built it.
+
+The optional make variable CF can be used to pass arguments to sparse.  The
+build system passes -Wbitwise to sparse automatically.
+
+Note that sparse defines the __CHECKER__ preprocessor symbol.