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, 166 insertions, 0 deletions

diff --git a/arch/alpha/lib/memcpy.c b/arch/alpha/lib/memcpy.c
new file mode 100644
index 000000000..78b6850a9
--- /dev/null
+++ b/arch/alpha/lib/memcpy.c
@@ -0,0 +1,166 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ *  linux/arch/alpha/lib/memcpy.c
+ *
+ *  Copyright (C) 1995  Linus Torvalds
+ */
+
+/*
+ * This is a reasonably optimized memcpy() routine.
+ */
+
+/*
+ * Note that the C code is written to be optimized into good assembly. However,
+ * at this point gcc is unable to sanely compile "if (n >= 0)", resulting in a
+ * explicit compare against 0 (instead of just using the proper "blt reg, xx" or
+ * "bge reg, xx"). I hope alpha-gcc will be fixed to notice this eventually..
+ */
+
+#include <linux/types.h>
+#include <linux/export.h>
+#include <linux/string.h>
+
+/*
+ * This should be done in one go with ldq_u*2/mask/stq_u. Do it
+ * with a macro so that we can fix it up later..
+ */
+#define ALIGN_DEST_TO8_UP(d,s,n) \
+	while (d & 7) { \
+		if (n <= 0) return; \
+		n--; \
+		*(char *) d = *(char *) s; \
+		d++; s++; \
+	}
+#define ALIGN_DEST_TO8_DN(d,s,n) \
+	while (d & 7) { \
+		if (n <= 0) return; \
+		n--; \
+		d--; s--; \
+		*(char *) d = *(char *) s; \
+	}
+
+/*
+ * This should similarly be done with ldq_u*2/mask/stq. The destination
+ * is aligned, but we don't fill in a full quad-word
+ */
+#define DO_REST_UP(d,s,n) \
+	while (n > 0) { \
+		n--; \
+		*(char *) d = *(char *) s; \
+		d++; s++; \
+	}
+#define DO_REST_DN(d,s,n) \
+	while (n > 0) { \
+		n--; \
+		d--; s--; \
+		*(char *) d = *(char *) s; \
+	}
+
+/*
+ * This should be done with ldq/mask/stq. The source and destination are
+ * aligned, but we don't fill in a full quad-word
+ */
+#define DO_REST_ALIGNED_UP(d,s,n) DO_REST_UP(d,s,n)
+#define DO_REST_ALIGNED_DN(d,s,n) DO_REST_DN(d,s,n)
+
+/*
+ * This does unaligned memory copies. We want to avoid storing to
+ * an unaligned address, as that would do a read-modify-write cycle.
+ * We also want to avoid double-reading the unaligned reads.
+ *
+ * Note the ordering to try to avoid load (and address generation) latencies.
+ */
+static inline void __memcpy_unaligned_up (unsigned long d, unsigned long s,
+					  long n)
+{
+	ALIGN_DEST_TO8_UP(d,s,n);
+	n -= 8;			/* to avoid compare against 8 in the loop */
+	if (n >= 0) {
+		unsigned long low_word, high_word;
+		__asm__("ldq_u %0,%1":"=r" (low_word):"m" (*(unsigned long *) s));
+		do {
+			unsigned long tmp;
+			__asm__("ldq_u %0,%1":"=r" (high_word):"m" (*(unsigned long *)(s+8)));
+			n -= 8;
+			__asm__("extql %1,%2,%0"
+				:"=r" (low_word)
+				:"r" (low_word), "r" (s));
+			__asm__("extqh %1,%2,%0"
+				:"=r" (tmp)
+				:"r" (high_word), "r" (s));
+			s += 8;
+			*(unsigned long *) d = low_word | tmp;
+			d += 8;
+			low_word = high_word;
+		} while (n >= 0);
+	}
+	n += 8;
+	DO_REST_UP(d,s,n);
+}
+
+static inline void __memcpy_unaligned_dn (unsigned long d, unsigned long s,
+					  long n)
+{
+	/* I don't understand AXP assembler well enough for this. -Tim */
+	s += n;
+	d += n;
+	while (n--)
+		* (char *) --d = * (char *) --s;
+}
+
+/*
+ * Hmm.. Strange. The __asm__ here is there to make gcc use an integer register
+ * for the load-store. I don't know why, but it would seem that using a floating
+ * point register for the move seems to slow things down (very small difference,
+ * though).
+ *
+ * Note the ordering to try to avoid load (and address generation) latencies.
+ */
+static inline void __memcpy_aligned_up (unsigned long d, unsigned long s,
+					long n)
+{
+	ALIGN_DEST_TO8_UP(d,s,n);
+	n -= 8;
+	while (n >= 0) {
+		unsigned long tmp;
+		__asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
+		n -= 8;
+		s += 8;
+		*(unsigned long *) d = tmp;
+		d += 8;
+	}
+	n += 8;
+	DO_REST_ALIGNED_UP(d,s,n);
+}
+static inline void __memcpy_aligned_dn (unsigned long d, unsigned long s,
+					long n)
+{
+	s += n;
+	d += n;
+	ALIGN_DEST_TO8_DN(d,s,n);
+	n -= 8;
+	while (n >= 0) {
+		unsigned long tmp;
+		s -= 8;
+		__asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
+		n -= 8;
+		d -= 8;
+		*(unsigned long *) d = tmp;
+	}
+	n += 8;
+	DO_REST_ALIGNED_DN(d,s,n);
+}
+
+#undef memcpy
+
+void * memcpy(void * dest, const void *src, size_t n)
+{
+	if (!(((unsigned long) dest ^ (unsigned long) src) & 7)) {
+		__memcpy_aligned_up ((unsigned long) dest, (unsigned long) src,
+				     n);
+		return dest;
+	}
+	__memcpy_unaligned_up ((unsigned long) dest, (unsigned long) src, n);
+	return dest;
+}
+EXPORT_SYMBOL(memcpy);