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, 116 insertions, 0 deletions

diff --git a/arch/riscv/lib/delay.c b/arch/riscv/lib/delay.c
new file mode 100644
index 000000000..bd9b38256
--- /dev/null
+++ b/arch/riscv/lib/delay.c
@@ -0,0 +1,116 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright (C) 2012 Regents of the University of California
+ */
+
+#include <linux/delay.h>
+#include <linux/math.h>
+#include <linux/param.h>
+#include <linux/types.h>
+#include <linux/export.h>
+
+#include <asm/processor.h>
+
+/*
+ * This is copies from arch/arm/include/asm/delay.h
+ *
+ * Loop (or tick) based delay:
+ *
+ * loops = loops_per_jiffy * jiffies_per_sec * delay_us / us_per_sec
+ *
+ * where:
+ *
+ * jiffies_per_sec = HZ
+ * us_per_sec = 1000000
+ *
+ * Therefore the constant part is HZ / 1000000 which is a small
+ * fractional number. To make this usable with integer math, we
+ * scale up this constant by 2^31, perform the actual multiplication,
+ * and scale the result back down by 2^31 with a simple shift:
+ *
+ * loops = (loops_per_jiffy * delay_us * UDELAY_MULT) >> 31
+ *
+ * where:
+ *
+ * UDELAY_MULT = 2^31 * HZ / 1000000
+ *             = (2^31 / 1000000) * HZ
+ *             = 2147.483648 * HZ
+ *             = 2147 * HZ + 483648 * HZ / 1000000
+ *
+ * 31 is the biggest scale shift value that won't overflow 32 bits for
+ * delay_us * UDELAY_MULT assuming HZ <= 1000 and delay_us <= 2000.
+ */
+#define MAX_UDELAY_US	2000
+#define MAX_UDELAY_HZ	1000
+#define UDELAY_MULT	(2147UL * HZ + 483648UL * HZ / 1000000UL)
+#define UDELAY_SHIFT	31
+
+#if HZ > MAX_UDELAY_HZ
+#error "HZ > MAX_UDELAY_HZ"
+#endif
+
+/*
+ * RISC-V supports both UDELAY and NDELAY.  This is largely the same as above,
+ * but with different constants.  I added 10 bits to the shift to get this, but
+ * the result is that I need a 64-bit multiply, which is slow on 32-bit
+ * platforms.
+ *
+ * NDELAY_MULT = 2^41 * HZ / 1000000000
+ *             = (2^41 / 1000000000) * HZ
+ *             = 2199.02325555 * HZ
+ *             = 2199 * HZ + 23255550 * HZ / 1000000000
+ *
+ * The maximum here is to avoid 64-bit overflow, but it isn't checked as it
+ * won't happen.
+ */
+#define MAX_NDELAY_NS   (1ULL << 42)
+#define MAX_NDELAY_HZ	MAX_UDELAY_HZ
+#define NDELAY_MULT	((unsigned long long)(2199ULL * HZ + 23255550ULL * HZ / 1000000000ULL))
+#define NDELAY_SHIFT	41
+
+#if HZ > MAX_NDELAY_HZ
+#error "HZ > MAX_NDELAY_HZ"
+#endif
+
+void __delay(unsigned long cycles)
+{
+	u64 t0 = get_cycles();
+
+	while ((unsigned long)(get_cycles() - t0) < cycles)
+		cpu_relax();
+}
+EXPORT_SYMBOL(__delay);
+
+void udelay(unsigned long usecs)
+{
+	u64 ucycles = (u64)usecs * lpj_fine * UDELAY_MULT;
+	u64 n;
+
+	if (unlikely(usecs > MAX_UDELAY_US)) {
+		n = (u64)usecs * riscv_timebase;
+		do_div(n, 1000000);
+
+		__delay(n);
+		return;
+	}
+
+	__delay(ucycles >> UDELAY_SHIFT);
+}
+EXPORT_SYMBOL(udelay);
+
+void ndelay(unsigned long nsecs)
+{
+	/*
+	 * This doesn't bother checking for overflow, as it won't happen (it's
+	 * an hour) of delay.
+	 */
+	unsigned long long ncycles = nsecs * lpj_fine * NDELAY_MULT;
+	__delay(ncycles >> NDELAY_SHIFT);
+}
+EXPORT_SYMBOL(ndelay);
+
+bool delay_read_timer(unsigned long *timer_val)
+{
+	*timer_val = get_cycles();
+	return true;
+}