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, 110 insertions, 0 deletions

diff --git a/drivers/md/dm-vdo/funnel-queue.h b/drivers/md/dm-vdo/funnel-queue.h
new file mode 100644
index 000000000..bde0f1def
--- /dev/null
+++ b/drivers/md/dm-vdo/funnel-queue.h
@@ -0,0 +1,110 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright 2023 Red Hat
+ */
+
+#ifndef VDO_FUNNEL_QUEUE_H
+#define VDO_FUNNEL_QUEUE_H
+
+#include <linux/atomic.h>
+#include <linux/cache.h>
+
+/*
+ * A funnel queue is a simple (almost) lock-free queue that accepts entries from multiple threads
+ * (multi-producer) and delivers them to a single thread (single-consumer). "Funnel" is an attempt
+ * to evoke the image of requests from more than one producer being "funneled down" to a single
+ * consumer.
+ *
+ * This is an unsynchronized but thread-safe data structure when used as intended. There is no
+ * mechanism to ensure that only one thread is consuming from the queue. If more than one thread
+ * attempts to consume from the queue, the resulting behavior is undefined. Clients must not
+ * directly access or manipulate the internals of the queue, which are only exposed for the purpose
+ * of allowing the very simple enqueue operation to be inlined.
+ *
+ * The implementation requires that a funnel_queue_entry structure (a link pointer) is embedded in
+ * the queue entries, and pointers to those structures are used exclusively by the queue. No macros
+ * are defined to template the queue, so the offset of the funnel_queue_entry in the records placed
+ * in the queue must all be the same so the client can derive their structure pointer from the
+ * entry pointer returned by vdo_funnel_queue_poll().
+ *
+ * Callers are wholly responsible for allocating and freeing the entries. Entries may be freed as
+ * soon as they are returned since this queue is not susceptible to the "ABA problem" present in
+ * many lock-free data structures. The queue is dynamically allocated to ensure cache-line
+ * alignment, but no other dynamic allocation is used.
+ *
+ * The algorithm is not actually 100% lock-free. There is a single point in vdo_funnel_queue_put()
+ * at which a preempted producer will prevent the consumers from seeing items added to the queue by
+ * later producers, and only if the queue is short enough or the consumer fast enough for it to
+ * reach what was the end of the queue at the time of the preemption.
+ *
+ * The consumer function, vdo_funnel_queue_poll(), will return NULL when the queue is empty. To
+ * wait for data to consume, spin (if safe) or combine the queue with a struct event_count to
+ * signal the presence of new entries.
+ */
+
+/* This queue link structure must be embedded in client entries. */
+struct funnel_queue_entry {
+	/* The next (newer) entry in the queue. */
+	struct funnel_queue_entry *next;
+};
+
+/*
+ * The dynamically allocated queue structure, which is allocated on a cache line boundary so the
+ * producer and consumer fields in the structure will land on separate cache lines. This should be
+ * consider opaque but it is exposed here so vdo_funnel_queue_put() can be inlined.
+ */
+struct __aligned(L1_CACHE_BYTES) funnel_queue {
+	/*
+	 * The producers' end of the queue, an atomically exchanged pointer that will never be
+	 * NULL.
+	 */
+	struct funnel_queue_entry *newest;
+
+	/* The consumer's end of the queue, which is owned by the consumer and never NULL. */
+	struct funnel_queue_entry *oldest __aligned(L1_CACHE_BYTES);
+
+	/* A dummy entry used to provide the non-NULL invariants above. */
+	struct funnel_queue_entry stub;
+};
+
+int __must_check vdo_make_funnel_queue(struct funnel_queue **queue_ptr);
+
+void vdo_free_funnel_queue(struct funnel_queue *queue);
+
+/*
+ * Put an entry on the end of the queue.
+ *
+ * The entry pointer must be to the struct funnel_queue_entry embedded in the caller's data
+ * structure. The caller must be able to derive the address of the start of their data structure
+ * from the pointer that passed in here, so every entry in the queue must have the struct
+ * funnel_queue_entry at the same offset within the client's structure.
+ */
+static inline void vdo_funnel_queue_put(struct funnel_queue *queue,
+					struct funnel_queue_entry *entry)
+{
+	struct funnel_queue_entry *previous;
+
+	/*
+	 * Barrier requirements: All stores relating to the entry ("next" pointer, containing data
+	 * structure fields) must happen before the previous->next store making it visible to the
+	 * consumer. Also, the entry's "next" field initialization to NULL must happen before any
+	 * other producer threads can see the entry (the xchg) and try to update the "next" field.
+	 *
+	 * xchg implements a full barrier.
+	 */
+	WRITE_ONCE(entry->next, NULL);
+	previous = xchg(&queue->newest, entry);
+	/*
+	 * Preemptions between these two statements hide the rest of the queue from the consumer,
+	 * preventing consumption until the following assignment runs.
+	 */
+	WRITE_ONCE(previous->next, entry);
+}
+
+struct funnel_queue_entry *__must_check vdo_funnel_queue_poll(struct funnel_queue *queue);
+
+bool __must_check vdo_is_funnel_queue_empty(struct funnel_queue *queue);
+
+bool __must_check vdo_is_funnel_queue_idle(struct funnel_queue *queue);
+
+#endif /* VDO_FUNNEL_QUEUE_H */