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, 67 insertions, 0 deletions

diff --git a/arch/arc/lib/strcpy-700.S b/arch/arc/lib/strcpy-700.S
new file mode 100644
index 000000000..6e2294d13
--- /dev/null
+++ b/arch/arc/lib/strcpy-700.S
@@ -0,0 +1,67 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
+ */
+
+/* If dst and src are 4 byte aligned, copy 8 bytes at a time.
+   If the src is 4, but not 8 byte aligned, we first read 4 bytes to get
+   it 8 byte aligned.  Thus, we can do a little read-ahead, without
+   dereferencing a cache line that we should not touch.
+   Note that short and long instructions have been scheduled to avoid
+   branch stalls.
+   The beq_s to r3z could be made unaligned & long to avoid a stall
+   there, but the it is not likely to be taken often, and it
+   would also be likey to cost an unaligned mispredict at the next call.  */
+
+#include <linux/linkage.h>
+
+ENTRY_CFI(strcpy)
+	or	r2,r0,r1
+	bmsk_s	r2,r2,1
+	brne.d	r2,0,charloop
+	mov_s	r10,r0
+	ld_s	r3,[r1,0]
+	mov	r8,0x01010101
+	bbit0.d	r1,2,loop_start
+	ror	r12,r8
+	sub	r2,r3,r8
+	bic_s	r2,r2,r3
+	tst_s	r2,r12
+	bne	r3z
+	mov_s	r4,r3
+	.balign 4
+loop:
+	ld.a	r3,[r1,4]
+	st.ab	r4,[r10,4]
+loop_start:
+	ld.a	r4,[r1,4]
+	sub	r2,r3,r8
+	bic_s	r2,r2,r3
+	tst_s	r2,r12
+	bne_s	r3z
+	st.ab	r3,[r10,4]
+	sub	r2,r4,r8
+	bic	r2,r2,r4
+	tst	r2,r12
+	beq	loop
+	mov_s	r3,r4
+#ifdef __LITTLE_ENDIAN__
+r3z:	bmsk.f	r1,r3,7
+	lsr_s	r3,r3,8
+#else
+r3z:	lsr.f	r1,r3,24
+	asl_s	r3,r3,8
+#endif
+	bne.d	r3z
+	stb.ab	r1,[r10,1]
+	j_s	[blink]
+
+	.balign	4
+charloop:
+	ldb.ab	r3,[r1,1]
+
+
+	brne.d	r3,0,charloop
+	stb.ab	r3,[r10,1]
+	j	[blink]
+END_CFI(strcpy)