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, 149 insertions, 0 deletions

diff --git a/arch/arc/lib/memcmp.S b/arch/arc/lib/memcmp.S
new file mode 100644
index 000000000..d6dc5e9bc
--- /dev/null
+++ b/arch/arc/lib/memcmp.S
@@ -0,0 +1,149 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
+ */
+
+#include <linux/linkage.h>
+
+#ifdef __LITTLE_ENDIAN__
+#define WORD2 r2
+#define SHIFT r3
+#else /* BIG ENDIAN */
+#define WORD2 r3
+#define SHIFT r2
+#endif
+
+ENTRY_CFI(memcmp)
+	or	r12,r0,r1
+	asl_s	r12,r12,30
+	sub	r3,r2,1
+	brls	r2,r12,.Lbytewise
+	ld	r4,[r0,0]
+	ld	r5,[r1,0]
+	lsr.f	lp_count,r3,3
+#ifdef CONFIG_ISA_ARCV2
+	/* In ARCv2 a branch can't be the last instruction in a zero overhead
+	 * loop.
+	 * So we move the branch to the start of the loop, duplicate it
+	 * after the end, and set up r12 so that the branch isn't taken
+	 *  initially.
+	 */
+	mov_s	r12,WORD2
+	lpne	.Loop_end
+	brne	WORD2,r12,.Lodd
+	ld	WORD2,[r0,4]
+#else
+	lpne	.Loop_end
+	ld_s	WORD2,[r0,4]
+#endif
+	ld_s	r12,[r1,4]
+	brne	r4,r5,.Leven
+	ld.a	r4,[r0,8]
+	ld.a	r5,[r1,8]
+#ifdef CONFIG_ISA_ARCV2
+.Loop_end:
+	brne	WORD2,r12,.Lodd
+#else
+	brne	WORD2,r12,.Lodd
+.Loop_end:
+#endif
+	asl_s	SHIFT,SHIFT,3
+	bhs_s	.Last_cmp
+	brne	r4,r5,.Leven
+	ld	r4,[r0,4]
+	ld	r5,[r1,4]
+#ifdef __LITTLE_ENDIAN__
+	nop_s
+	; one more load latency cycle
+.Last_cmp:
+	xor	r0,r4,r5
+	bset	r0,r0,SHIFT
+	sub_s	r1,r0,1
+	bic_s	r1,r1,r0
+	norm	r1,r1
+	b.d	.Leven_cmp
+	and	r1,r1,24
+.Leven:
+	xor	r0,r4,r5
+	sub_s	r1,r0,1
+	bic_s	r1,r1,r0
+	norm	r1,r1
+	; slow track insn
+	and	r1,r1,24
+.Leven_cmp:
+	asl	r2,r4,r1
+	asl	r12,r5,r1
+	lsr_s	r2,r2,1
+	lsr_s	r12,r12,1
+	j_s.d	[blink]
+	sub	r0,r2,r12
+	.balign	4
+.Lodd:
+	xor	r0,WORD2,r12
+	sub_s	r1,r0,1
+	bic_s	r1,r1,r0
+	norm	r1,r1
+	; slow track insn
+	and	r1,r1,24
+	asl_s	r2,r2,r1
+	asl_s	r12,r12,r1
+	lsr_s	r2,r2,1
+	lsr_s	r12,r12,1
+	j_s.d	[blink]
+	sub	r0,r2,r12
+#else /* BIG ENDIAN */
+.Last_cmp:
+	neg_s	SHIFT,SHIFT
+	lsr	r4,r4,SHIFT
+	lsr	r5,r5,SHIFT
+	; slow track insn
+.Leven:
+	sub.f	r0,r4,r5
+	mov.ne	r0,1
+	j_s.d	[blink]
+	bset.cs	r0,r0,31
+.Lodd:
+	cmp_s	WORD2,r12
+	mov_s	r0,1
+	j_s.d	[blink]
+	bset.cs	r0,r0,31
+#endif /* ENDIAN */
+	.balign	4
+.Lbytewise:
+	breq	r2,0,.Lnil
+	ldb	r4,[r0,0]
+	ldb	r5,[r1,0]
+	lsr.f	lp_count,r3
+#ifdef CONFIG_ISA_ARCV2
+	mov	r12,r3
+	lpne	.Lbyte_end
+	brne	r3,r12,.Lbyte_odd
+#else
+	lpne	.Lbyte_end
+#endif
+	ldb_s	r3,[r0,1]
+	ldb	r12,[r1,1]
+	brne	r4,r5,.Lbyte_even
+	ldb.a	r4,[r0,2]
+	ldb.a	r5,[r1,2]
+#ifdef CONFIG_ISA_ARCV2
+.Lbyte_end:
+	brne	r3,r12,.Lbyte_odd
+#else
+	brne	r3,r12,.Lbyte_odd
+.Lbyte_end:
+#endif
+	bcc	.Lbyte_even
+	brne	r4,r5,.Lbyte_even
+	ldb_s	r3,[r0,1]
+	ldb_s	r12,[r1,1]
+.Lbyte_odd:
+	j_s.d	[blink]
+	sub	r0,r3,r12
+.Lbyte_even:
+	j_s.d	[blink]
+	sub	r0,r4,r5
+.Lnil:
+	j_s.d	[blink]
+	mov	r0,0
+END_CFI(memcmp)