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, 93 insertions, 0 deletions

diff --git a/arch/arc/lib/strcmp.S b/arch/arc/lib/strcmp.S
new file mode 100644
index 000000000..b20c98fb3
--- /dev/null
+++ b/arch/arc/lib/strcmp.S
@@ -0,0 +1,93 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
+ */
+
+/* This is optimized primarily for the ARC700.
+   It would be possible to speed up the loops by one cycle / word
+   respective one cycle / byte by forcing double source 1 alignment, unrolling
+   by a factor of two, and speculatively loading the second word / byte of
+   source 1; however, that would increase the overhead for loop setup / finish,
+   and strcmp might often terminate early.  */
+
+#include <linux/linkage.h>
+
+ENTRY_CFI(strcmp)
+	or	r2,r0,r1
+	bmsk_s	r2,r2,1
+	brne	r2,0,.Lcharloop
+	mov_s	r12,0x01010101
+	ror	r5,r12
+.Lwordloop:
+	ld.ab	r2,[r0,4]
+	ld.ab	r3,[r1,4]
+	nop_s
+	sub	r4,r2,r12
+	bic	r4,r4,r2
+	and	r4,r4,r5
+	brne	r4,0,.Lfound0
+	breq	r2,r3,.Lwordloop
+#ifdef	__LITTLE_ENDIAN__
+	xor	r0,r2,r3	; mask for difference
+	sub_s	r1,r0,1
+	bic_s	r0,r0,r1	; mask for least significant difference bit
+	sub	r1,r5,r0
+	xor	r0,r5,r1	; mask for least significant difference byte
+	and_s	r2,r2,r0
+	and_s	r3,r3,r0
+#endif /* LITTLE ENDIAN */
+	cmp_s	r2,r3
+	mov_s	r0,1
+	j_s.d	[blink]
+	bset.lo	r0,r0,31
+
+	.balign	4
+#ifdef __LITTLE_ENDIAN__
+.Lfound0:
+	xor	r0,r2,r3	; mask for difference
+	or	r0,r0,r4	; or in zero indicator
+	sub_s	r1,r0,1
+	bic_s	r0,r0,r1	; mask for least significant difference bit
+	sub	r1,r5,r0
+	xor	r0,r5,r1	; mask for least significant difference byte
+	and_s	r2,r2,r0
+	and_s	r3,r3,r0
+	sub.f	r0,r2,r3
+	mov.hi	r0,1
+	j_s.d	[blink]
+	bset.lo	r0,r0,31
+#else /* BIG ENDIAN */
+	/* The zero-detection above can mis-detect 0x01 bytes as zeroes
+	   because of carry-propagateion from a lower significant zero byte.
+	   We can compensate for this by checking that bit0 is zero.
+	   This compensation is not necessary in the step where we
+	   get a low estimate for r2, because in any affected bytes
+	   we already have 0x00 or 0x01, which will remain unchanged
+	   when bit 7 is cleared.  */
+	.balign	4
+.Lfound0:
+	lsr	r0,r4,8
+	lsr_s	r1,r2
+	bic_s	r2,r2,r0	; get low estimate for r2 and get ...
+	bic_s	r0,r0,r1	; <this is the adjusted mask for zeros>
+	or_s	r3,r3,r0	; ... high estimate r3 so that r2 > r3 will ...
+	cmp_s	r3,r2		; ... be independent of trailing garbage
+	or_s	r2,r2,r0	; likewise for r3 > r2
+	bic_s	r3,r3,r0
+	rlc	r0,0		; r0 := r2 > r3 ? 1 : 0
+	cmp_s	r2,r3
+	j_s.d	[blink]
+	bset.lo	r0,r0,31
+#endif /* ENDIAN */
+
+	.balign	4
+.Lcharloop:
+	ldb.ab	r2,[r0,1]
+	ldb.ab	r3,[r1,1]
+	nop_s
+	breq	r2,0,.Lcmpend
+	breq	r2,r3,.Lcharloop
+.Lcmpend:
+	j_s.d	[blink]
+	sub	r0,r2,r3
+END_CFI(strcmp)