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, 187 insertions, 0 deletions

diff --git a/drivers/md/dm-cache-policy.h b/drivers/md/dm-cache-policy.h
new file mode 100644
index 000000000..a1eedcc42
--- /dev/null
+++ b/drivers/md/dm-cache-policy.h
@@ -0,0 +1,187 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2012 Red Hat. All rights reserved.
+ *
+ * This file is released under the GPL.
+ */
+
+#ifndef DM_CACHE_POLICY_H
+#define DM_CACHE_POLICY_H
+
+#include "dm-cache-block-types.h"
+
+#include <linux/device-mapper.h>
+
+/*----------------------------------------------------------------*/
+
+/*
+ * The cache policy makes the important decisions about which blocks get to
+ * live on the faster cache device.
+ */
+enum policy_operation {
+	POLICY_PROMOTE,
+	POLICY_DEMOTE,
+	POLICY_WRITEBACK
+};
+
+/*
+ * This is the instruction passed back to the core target.
+ */
+struct policy_work {
+	enum policy_operation op;
+	dm_oblock_t oblock;
+	dm_cblock_t cblock;
+};
+
+/*
+ * The cache policy object.  It is envisaged that this structure will be
+ * embedded in a bigger, policy specific structure (ie. use container_of()).
+ */
+struct dm_cache_policy {
+	/*
+	 * Destroys this object.
+	 */
+	void (*destroy)(struct dm_cache_policy *p);
+
+	/*
+	 * Find the location of a block.
+	 *
+	 * Must not block.
+	 *
+	 * Returns 0 if in cache (cblock will be set), -ENOENT if not, < 0 for
+	 * other errors (-EWOULDBLOCK would be typical).  data_dir should be
+	 * READ or WRITE. fast_copy should be set if migrating this block would
+	 * be 'cheap' somehow (eg, discarded data). background_queued will be set
+	 * if a migration has just been queued.
+	 */
+	int (*lookup)(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock,
+		      int data_dir, bool fast_copy, bool *background_queued);
+
+	/*
+	 * Sometimes the core target can optimise a migration, eg, the
+	 * block may be discarded, or the bio may cover an entire block.
+	 * In order to optimise it needs the migration immediately though
+	 * so it knows to do something different with the bio.
+	 *
+	 * This method is optional (policy-internal will fallback to using
+	 * lookup).
+	 */
+	int (*lookup_with_work)(struct dm_cache_policy *p,
+				dm_oblock_t oblock, dm_cblock_t *cblock,
+				int data_dir, bool fast_copy,
+				struct policy_work **work);
+
+	/*
+	 * Retrieves background work.  Returns -ENODATA when there's no
+	 * background work.
+	 */
+	int (*get_background_work)(struct dm_cache_policy *p, bool idle,
+				   struct policy_work **result);
+
+	/*
+	 * You must pass in the same work pointer that you were given, not
+	 * a copy.
+	 */
+	void (*complete_background_work)(struct dm_cache_policy *p,
+					 struct policy_work *work,
+					 bool success);
+
+	void (*set_dirty)(struct dm_cache_policy *p, dm_cblock_t cblock);
+	void (*clear_dirty)(struct dm_cache_policy *p, dm_cblock_t cblock);
+
+	/*
+	 * Called when a cache target is first created.  Used to load a
+	 * mapping from the metadata device into the policy.
+	 */
+	int (*load_mapping)(struct dm_cache_policy *p, dm_oblock_t oblock,
+			    dm_cblock_t cblock, bool dirty,
+			    uint32_t hint, bool hint_valid);
+
+	/*
+	 * Drops the mapping, irrespective of whether it's clean or dirty.
+	 * Returns -ENODATA if cblock is not mapped.
+	 */
+	int (*invalidate_mapping)(struct dm_cache_policy *p, dm_cblock_t cblock);
+
+	/*
+	 * Gets the hint for a given cblock.  Called in a single threaded
+	 * context.  So no locking required.
+	 */
+	uint32_t (*get_hint)(struct dm_cache_policy *p, dm_cblock_t cblock);
+
+	/*
+	 * How full is the cache?
+	 */
+	dm_cblock_t (*residency)(struct dm_cache_policy *p);
+
+	/*
+	 * Because of where we sit in the block layer, we can be asked to
+	 * map a lot of little bios that are all in the same block (no
+	 * queue merging has occurred).  To stop the policy being fooled by
+	 * these, the core target sends regular tick() calls to the policy.
+	 * The policy should only count an entry as hit once per tick.
+	 *
+	 * This method is optional.
+	 */
+	void (*tick)(struct dm_cache_policy *p, bool can_block);
+
+	/*
+	 * Configuration.
+	 */
+	int (*emit_config_values)(struct dm_cache_policy *p, char *result,
+				  unsigned int maxlen, ssize_t *sz_ptr);
+	int (*set_config_value)(struct dm_cache_policy *p,
+				const char *key, const char *value);
+
+	void (*allow_migrations)(struct dm_cache_policy *p, bool allow);
+
+	/*
+	 * Book keeping ptr for the policy register, not for general use.
+	 */
+	void *private;
+};
+
+/*----------------------------------------------------------------*/
+
+/*
+ * We maintain a little register of the different policy types.
+ */
+#define CACHE_POLICY_NAME_SIZE 16
+#define CACHE_POLICY_VERSION_SIZE 3
+
+struct dm_cache_policy_type {
+	/* For use by the register code only. */
+	struct list_head list;
+
+	/*
+	 * Policy writers should fill in these fields.  The name field is
+	 * what gets passed on the target line to select your policy.
+	 */
+	char name[CACHE_POLICY_NAME_SIZE];
+	unsigned int version[CACHE_POLICY_VERSION_SIZE];
+
+	/*
+	 * For use by an alias dm_cache_policy_type to point to the
+	 * real dm_cache_policy_type.
+	 */
+	struct dm_cache_policy_type *real;
+
+	/*
+	 * Policies may store a hint for each cache block.
+	 * Currently the size of this hint must be 0 or 4 bytes but we
+	 * expect to relax this in future.
+	 */
+	size_t hint_size;
+
+	struct module *owner;
+	struct dm_cache_policy *(*create)(dm_cblock_t cache_size,
+					  sector_t origin_size,
+					  sector_t block_size);
+};
+
+int dm_cache_policy_register(struct dm_cache_policy_type *type);
+void dm_cache_policy_unregister(struct dm_cache_policy_type *type);
+
+/*----------------------------------------------------------------*/
+
+#endif	/* DM_CACHE_POLICY_H */