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, 151 insertions, 0 deletions

diff --git a/drivers/md/dm-cache-metadata.h b/drivers/md/dm-cache-metadata.h
new file mode 100644
index 000000000..91f8706b4
--- /dev/null
+++ b/drivers/md/dm-cache-metadata.h
@@ -0,0 +1,151 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2012 Red Hat, Inc.
+ *
+ * This file is released under the GPL.
+ */
+
+#ifndef DM_CACHE_METADATA_H
+#define DM_CACHE_METADATA_H
+
+#include "dm-cache-block-types.h"
+#include "dm-cache-policy-internal.h"
+#include "persistent-data/dm-space-map-metadata.h"
+
+/*----------------------------------------------------------------*/
+
+#define DM_CACHE_METADATA_BLOCK_SIZE DM_SM_METADATA_BLOCK_SIZE
+
+/* FIXME: remove this restriction */
+/*
+ * The metadata device is currently limited in size.
+ */
+#define DM_CACHE_METADATA_MAX_SECTORS DM_SM_METADATA_MAX_SECTORS
+
+/*
+ * A metadata device larger than 16GB triggers a warning.
+ */
+#define DM_CACHE_METADATA_MAX_SECTORS_WARNING (16 * (1024 * 1024 * 1024 >> SECTOR_SHIFT))
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Ext[234]-style compat feature flags.
+ *
+ * A new feature which old metadata will still be compatible with should
+ * define a DM_CACHE_FEATURE_COMPAT_* flag (rarely useful).
+ *
+ * A new feature that is not compatible with old code should define a
+ * DM_CACHE_FEATURE_INCOMPAT_* flag and guard the relevant code with
+ * that flag.
+ *
+ * A new feature that is not compatible with old code accessing the
+ * metadata RDWR should define a DM_CACHE_FEATURE_RO_COMPAT_* flag and
+ * guard the relevant code with that flag.
+ *
+ * As these various flags are defined they should be added to the
+ * following masks.
+ */
+
+#define DM_CACHE_FEATURE_COMPAT_SUPP	  0UL
+#define DM_CACHE_FEATURE_COMPAT_RO_SUPP	  0UL
+#define DM_CACHE_FEATURE_INCOMPAT_SUPP	  0UL
+
+struct dm_cache_metadata;
+
+/*
+ * Reopens or creates a new, empty metadata volume.  Returns an ERR_PTR on
+ * failure.  If reopening then features must match.
+ */
+struct dm_cache_metadata *dm_cache_metadata_open(struct block_device *bdev,
+						 sector_t data_block_size,
+						 bool may_format_device,
+						 size_t policy_hint_size,
+						 unsigned int metadata_version);
+
+void dm_cache_metadata_close(struct dm_cache_metadata *cmd);
+
+/*
+ * The metadata needs to know how many cache blocks there are.  We don't
+ * care about the origin, assuming the core target is giving us valid
+ * origin blocks to map to.
+ */
+int dm_cache_resize(struct dm_cache_metadata *cmd, dm_cblock_t new_cache_size);
+
+int dm_cache_discard_bitset_resize(struct dm_cache_metadata *cmd,
+				   sector_t discard_block_size,
+				   dm_dblock_t new_nr_entries);
+
+typedef int (*load_discard_fn)(void *context, sector_t discard_block_size,
+			       dm_dblock_t dblock, bool discarded);
+int dm_cache_load_discards(struct dm_cache_metadata *cmd,
+			   load_discard_fn fn, void *context);
+
+int dm_cache_set_discard(struct dm_cache_metadata *cmd, dm_dblock_t dblock, bool discard);
+
+int dm_cache_remove_mapping(struct dm_cache_metadata *cmd, dm_cblock_t cblock);
+int dm_cache_insert_mapping(struct dm_cache_metadata *cmd, dm_cblock_t cblock, dm_oblock_t oblock);
+int dm_cache_changed_this_transaction(struct dm_cache_metadata *cmd);
+
+typedef int (*load_mapping_fn)(void *context, dm_oblock_t oblock,
+			       dm_cblock_t cblock, bool dirty,
+			       uint32_t hint, bool hint_valid);
+int dm_cache_load_mappings(struct dm_cache_metadata *cmd,
+			   struct dm_cache_policy *policy,
+			   load_mapping_fn fn,
+			   void *context);
+
+int dm_cache_set_dirty_bits(struct dm_cache_metadata *cmd,
+			    unsigned int nr_bits, unsigned long *bits);
+
+struct dm_cache_statistics {
+	uint32_t read_hits;
+	uint32_t read_misses;
+	uint32_t write_hits;
+	uint32_t write_misses;
+};
+
+void dm_cache_metadata_get_stats(struct dm_cache_metadata *cmd,
+				 struct dm_cache_statistics *stats);
+
+/*
+ * 'void' because it's no big deal if it fails.
+ */
+void dm_cache_metadata_set_stats(struct dm_cache_metadata *cmd,
+				 struct dm_cache_statistics *stats);
+
+int dm_cache_commit(struct dm_cache_metadata *cmd, bool clean_shutdown);
+
+int dm_cache_get_free_metadata_block_count(struct dm_cache_metadata *cmd,
+					   dm_block_t *result);
+
+int dm_cache_get_metadata_dev_size(struct dm_cache_metadata *cmd,
+				   dm_block_t *result);
+
+/*
+ * The policy is invited to save a 32bit hint value for every cblock (eg,
+ * for a hit count).  These are stored against the policy name.  If
+ * policies are changed, then hints will be lost.  If the machine crashes,
+ * hints will be lost.
+ *
+ * The hints are indexed by the cblock, but many policies will not
+ * necessarily have a fast way of accessing efficiently via cblock.  So
+ * rather than querying the policy for each cblock, we let it walk its data
+ * structures and fill in the hints in whatever order it wishes.
+ */
+int dm_cache_write_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *p);
+
+int dm_cache_metadata_needs_check(struct dm_cache_metadata *cmd, bool *result);
+int dm_cache_metadata_set_needs_check(struct dm_cache_metadata *cmd);
+void dm_cache_metadata_set_read_only(struct dm_cache_metadata *cmd);
+void dm_cache_metadata_set_read_write(struct dm_cache_metadata *cmd);
+int dm_cache_metadata_abort(struct dm_cache_metadata *cmd);
+
+/*
+ * Query method.  Was the metadata cleanly shut down when opened?
+ */
+int dm_cache_metadata_clean_when_opened(struct dm_cache_metadata *cmd, bool *result);
+
+/*----------------------------------------------------------------*/
+
+#endif /* DM_CACHE_METADATA_H */