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, 101 insertions, 0 deletions

diff --git a/arch/arm/common/vlock.S b/arch/arm/common/vlock.S
new file mode 100644
index 000000000..c5eaed5a7
--- /dev/null
+++ b/arch/arm/common/vlock.S
@@ -0,0 +1,101 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * vlock.S - simple voting lock implementation for ARM
+ *
+ * Created by:	Dave Martin, 2012-08-16
+ * Copyright:	(C) 2012-2013  Linaro Limited
+ *
+ * This algorithm is described in more detail in
+ * Documentation/arch/arm/vlocks.rst.
+ */
+
+#include <linux/linkage.h>
+#include "vlock.h"
+
+.arch armv7-a
+
+/* Select different code if voting flags  can fit in a single word. */
+#if VLOCK_VOTING_SIZE > 4
+#define FEW(x...)
+#define MANY(x...) x
+#else
+#define FEW(x...) x
+#define MANY(x...)
+#endif
+
+@ voting lock for first-man coordination
+
+.macro voting_begin rbase:req, rcpu:req, rscratch:req
+	mov	\rscratch, #1
+	strb	\rscratch, [\rbase, \rcpu]
+	dmb
+.endm
+
+.macro voting_end rbase:req, rcpu:req, rscratch:req
+	dmb
+	mov	\rscratch, #0
+	strb	\rscratch, [\rbase, \rcpu]
+	dsb	st
+	sev
+.endm
+
+/*
+ * The vlock structure must reside in Strongly-Ordered or Device memory.
+ * This implementation deliberately eliminates most of the barriers which
+ * would be required for other memory types, and assumes that independent
+ * writes to neighbouring locations within a cacheline do not interfere
+ * with one another.
+ */
+
+@ r0: lock structure base
+@ r1: CPU ID (0-based index within cluster)
+ENTRY(vlock_trylock)
+	add	r1, r1, #VLOCK_VOTING_OFFSET
+
+	voting_begin	r0, r1, r2
+
+	ldrb	r2, [r0, #VLOCK_OWNER_OFFSET]	@ check whether lock is held
+	cmp	r2, #VLOCK_OWNER_NONE
+	bne	trylock_fail			@ fail if so
+
+	@ Control dependency implies strb not observable before previous ldrb.
+
+	strb	r1, [r0, #VLOCK_OWNER_OFFSET]	@ submit my vote
+
+	voting_end	r0, r1, r2		@ implies DMB
+
+	@ Wait for the current round of voting to finish:
+
+ MANY(	mov	r3, #VLOCK_VOTING_OFFSET			)
+0:
+ MANY(	ldr	r2, [r0, r3]					)
+ FEW(	ldr	r2, [r0, #VLOCK_VOTING_OFFSET]			)
+	cmp	r2, #0
+	wfene
+	bne	0b
+ MANY(	add	r3, r3, #4					)
+ MANY(	cmp	r3, #VLOCK_VOTING_OFFSET + VLOCK_VOTING_SIZE	)
+ MANY(	bne	0b						)
+
+	@ Check who won:
+
+	dmb
+	ldrb	r2, [r0, #VLOCK_OWNER_OFFSET]
+	eor	r0, r1, r2			@ zero if I won, else nonzero
+	bx	lr
+
+trylock_fail:
+	voting_end	r0, r1, r2
+	mov	r0, #1				@ nonzero indicates that I lost
+	bx	lr
+ENDPROC(vlock_trylock)
+
+@ r0: lock structure base
+ENTRY(vlock_unlock)
+	dmb
+	mov	r1, #VLOCK_OWNER_NONE
+	strb	r1, [r0, #VLOCK_OWNER_OFFSET]
+	dsb	st
+	sev
+	bx	lr
+ENDPROC(vlock_unlock)