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, 139 insertions, 0 deletions

diff --git a/arch/arm64/lib/memcmp.S b/arch/arm64/lib/memcmp.S
new file mode 100644
index 000000000..a5ccf2c55
--- /dev/null
+++ b/arch/arm64/lib/memcmp.S
@@ -0,0 +1,139 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (c) 2013-2021, Arm Limited.
+ *
+ * Adapted from the original at:
+ * https://github.com/ARM-software/optimized-routines/blob/e823e3abf5f89ecb/string/aarch64/memcmp.S
+ */
+
+#include <linux/linkage.h>
+#include <asm/assembler.h>
+
+/* Assumptions:
+ *
+ * ARMv8-a, AArch64, unaligned accesses.
+ */
+
+#define L(label) .L ## label
+
+/* Parameters and result.  */
+#define src1		x0
+#define src2		x1
+#define limit		x2
+#define result		w0
+
+/* Internal variables.  */
+#define data1		x3
+#define data1w		w3
+#define data1h		x4
+#define data2		x5
+#define data2w		w5
+#define data2h		x6
+#define tmp1		x7
+#define tmp2		x8
+
+SYM_FUNC_START(__pi_memcmp)
+	subs	limit, limit, 8
+	b.lo	L(less8)
+
+	ldr	data1, [src1], 8
+	ldr	data2, [src2], 8
+	cmp	data1, data2
+	b.ne	L(return)
+
+	subs	limit, limit, 8
+	b.gt	L(more16)
+
+	ldr	data1, [src1, limit]
+	ldr	data2, [src2, limit]
+	b	L(return)
+
+L(more16):
+	ldr	data1, [src1], 8
+	ldr	data2, [src2], 8
+	cmp	data1, data2
+	bne	L(return)
+
+	/* Jump directly to comparing the last 16 bytes for 32 byte (or less)
+	   strings.  */
+	subs	limit, limit, 16
+	b.ls	L(last_bytes)
+
+	/* We overlap loads between 0-32 bytes at either side of SRC1 when we
+	   try to align, so limit it only to strings larger than 128 bytes.  */
+	cmp	limit, 96
+	b.ls	L(loop16)
+
+	/* Align src1 and adjust src2 with bytes not yet done.  */
+	and	tmp1, src1, 15
+	add	limit, limit, tmp1
+	sub	src1, src1, tmp1
+	sub	src2, src2, tmp1
+
+	/* Loop performing 16 bytes per iteration using aligned src1.
+	   Limit is pre-decremented by 16 and must be larger than zero.
+	   Exit if <= 16 bytes left to do or if the data is not equal.  */
+	.p2align 4
+L(loop16):
+	ldp	data1, data1h, [src1], 16
+	ldp	data2, data2h, [src2], 16
+	subs	limit, limit, 16
+	ccmp	data1, data2, 0, hi
+	ccmp	data1h, data2h, 0, eq
+	b.eq	L(loop16)
+
+	cmp	data1, data2
+	bne	L(return)
+	mov	data1, data1h
+	mov	data2, data2h
+	cmp	data1, data2
+	bne	L(return)
+
+	/* Compare last 1-16 bytes using unaligned access.  */
+L(last_bytes):
+	add	src1, src1, limit
+	add	src2, src2, limit
+	ldp	data1, data1h, [src1]
+	ldp	data2, data2h, [src2]
+	cmp	data1, data2
+	bne	L(return)
+	mov	data1, data1h
+	mov	data2, data2h
+	cmp	data1, data2
+
+	/* Compare data bytes and set return value to 0, -1 or 1.  */
+L(return):
+#ifndef __AARCH64EB__
+	rev	data1, data1
+	rev	data2, data2
+#endif
+	cmp	data1, data2
+L(ret_eq):
+	cset	result, ne
+	cneg	result, result, lo
+	ret
+
+	.p2align 4
+	/* Compare up to 8 bytes.  Limit is [-8..-1].  */
+L(less8):
+	adds	limit, limit, 4
+	b.lo	L(less4)
+	ldr	data1w, [src1], 4
+	ldr	data2w, [src2], 4
+	cmp	data1w, data2w
+	b.ne	L(return)
+	sub	limit, limit, 4
+L(less4):
+	adds	limit, limit, 4
+	beq	L(ret_eq)
+L(byte_loop):
+	ldrb	data1w, [src1], 1
+	ldrb	data2w, [src2], 1
+	subs	limit, limit, 1
+	ccmp	data1w, data2w, 0, ne	/* NZCV = 0b0000.  */
+	b.eq	L(byte_loop)
+	sub	result, data1w, data2w
+	ret
+SYM_FUNC_END(__pi_memcmp)
+SYM_FUNC_ALIAS_WEAK(memcmp, __pi_memcmp)
+EXPORT_SYMBOL_NOKASAN(memcmp)