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, 216 insertions, 0 deletions

diff --git a/drivers/md/persistent-data/dm-btree.h b/drivers/md/persistent-data/dm-btree.h
new file mode 100644
index 000000000..1b92acd78
--- /dev/null
+++ b/drivers/md/persistent-data/dm-btree.h
@@ -0,0 +1,216 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2011 Red Hat, Inc.
+ *
+ * This file is released under the GPL.
+ */
+#ifndef _LINUX_DM_BTREE_H
+#define _LINUX_DM_BTREE_H
+
+#include "dm-block-manager.h"
+
+struct dm_transaction_manager;
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Annotations used to check on-disk metadata is handled as little-endian.
+ */
+#ifdef __CHECKER__
+#  define __dm_written_to_disk(x) __releases(x)
+#  define __dm_reads_from_disk(x) __acquires(x)
+#  define __dm_bless_for_disk(x) __acquire(x)
+#  define __dm_unbless_for_disk(x) __release(x)
+#else
+#  define __dm_written_to_disk(x)
+#  define __dm_reads_from_disk(x)
+#  define __dm_bless_for_disk(x)
+#  define __dm_unbless_for_disk(x)
+#endif
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Manipulates hierarchical B+ trees with 64-bit keys and arbitrary-sized
+ * values.
+ */
+
+/*
+ * Information about the values stored within the btree.
+ */
+struct dm_btree_value_type {
+	void *context;
+
+	/*
+	 * The size in bytes of each value.
+	 */
+	uint32_t size;
+
+	/*
+	 * Any of these methods can be safely set to NULL if you do not
+	 * need the corresponding feature.
+	 */
+
+	/*
+	 * The btree is making a duplicate of a run of values, for instance
+	 * because previously-shared btree nodes have now diverged.
+	 * @value argument is the new copy that the copy function may modify.
+	 * (Probably it just wants to increment a reference count
+	 * somewhere.) This method is _not_ called for insertion of a new
+	 * value: It is assumed the ref count is already 1.
+	 */
+	void (*inc)(void *context, const void *value, unsigned int count);
+
+	/*
+	 * These values are being deleted.  The btree takes care of freeing
+	 * the memory pointed to by @value.  Often the del function just
+	 * needs to decrement a reference counts somewhere.
+	 */
+	void (*dec)(void *context, const void *value, unsigned int count);
+
+	/*
+	 * A test for equality between two values.  When a value is
+	 * overwritten with a new one, the old one has the dec method
+	 * called _unless_ the new and old value are deemed equal.
+	 */
+	int (*equal)(void *context, const void *value1, const void *value2);
+};
+
+/*
+ * The shape and contents of a btree.
+ */
+struct dm_btree_info {
+	struct dm_transaction_manager *tm;
+
+	/*
+	 * Number of nested btrees. (Not the depth of a single tree.)
+	 */
+	unsigned int levels;
+	struct dm_btree_value_type value_type;
+};
+
+/*
+ * Set up an empty tree.  O(1).
+ */
+int dm_btree_empty(struct dm_btree_info *info, dm_block_t *root);
+
+/*
+ * Delete a tree.  O(n) - this is the slow one!  It can also block, so
+ * please don't call it on an IO path.
+ */
+int dm_btree_del(struct dm_btree_info *info, dm_block_t root);
+
+/*
+ * All the lookup functions return -ENODATA if the key cannot be found.
+ */
+
+/*
+ * Tries to find a key that matches exactly.  O(ln(n))
+ */
+int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
+		    uint64_t *keys, void *value_le);
+
+/*
+ * Tries to find the first key where the bottom level key is >= to that
+ * given.  Useful for skipping empty sections of the btree.
+ */
+int dm_btree_lookup_next(struct dm_btree_info *info, dm_block_t root,
+			 uint64_t *keys, uint64_t *rkey, void *value_le);
+
+/*
+ * Insertion (or overwrite an existing value).  O(ln(n))
+ */
+int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
+		    uint64_t *keys, void *value, dm_block_t *new_root)
+	__dm_written_to_disk(value);
+
+/*
+ * A variant of insert that indicates whether it actually inserted or just
+ * overwrote.  Useful if you're keeping track of the number of entries in a
+ * tree.
+ */
+int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
+			   uint64_t *keys, void *value, dm_block_t *new_root,
+			   int *inserted)
+			   __dm_written_to_disk(value);
+
+/*
+ * Remove a key if present.  This doesn't remove empty sub trees.  Normally
+ * subtrees represent a separate entity, like a snapshot map, so this is
+ * correct behaviour.  O(ln(n)).
+ */
+int dm_btree_remove(struct dm_btree_info *info, dm_block_t root,
+		    uint64_t *keys, dm_block_t *new_root);
+
+/*
+ * Removes a _contiguous_ run of values starting from 'keys' and not
+ * reaching keys2 (where keys2 is keys with the final key replaced with
+ * 'end_key').  'end_key' is the one-past-the-end value.  'keys' may be
+ * altered.
+ */
+int dm_btree_remove_leaves(struct dm_btree_info *info, dm_block_t root,
+			   uint64_t *keys, uint64_t end_key,
+			   dm_block_t *new_root, unsigned int *nr_removed);
+
+/*
+ * Returns < 0 on failure.  Otherwise the number of key entries that have
+ * been filled out.  Remember trees can have zero entries, and as such have
+ * no lowest key.
+ */
+int dm_btree_find_lowest_key(struct dm_btree_info *info, dm_block_t root,
+			     uint64_t *result_keys);
+
+/*
+ * Returns < 0 on failure.  Otherwise the number of key entries that have
+ * been filled out.  Remember trees can have zero entries, and as such have
+ * no highest key.
+ */
+int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
+			      uint64_t *result_keys);
+
+/*
+ * Iterate through the a btree, calling fn() on each entry.
+ * It only works for single level trees and is internally recursive, so
+ * monitor stack usage carefully.
+ */
+int dm_btree_walk(struct dm_btree_info *info, dm_block_t root,
+		  int (*fn)(void *context, uint64_t *keys, void *leaf),
+		  void *context);
+
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Cursor API.  This does not follow the rolling lock convention.  Since we
+ * know the order that values are required we can issue prefetches to speed
+ * up iteration.  Use on a single level btree only.
+ */
+#define DM_BTREE_CURSOR_MAX_DEPTH 16
+
+struct cursor_node {
+	struct dm_block *b;
+	unsigned int index;
+};
+
+struct dm_btree_cursor {
+	struct dm_btree_info *info;
+	dm_block_t root;
+
+	bool prefetch_leaves;
+	unsigned int depth;
+	struct cursor_node nodes[DM_BTREE_CURSOR_MAX_DEPTH];
+};
+
+/*
+ * Creates a fresh cursor.  If prefetch_leaves is set then it is assumed
+ * the btree contains block indexes that will be prefetched.  The cursor is
+ * quite large, so you probably don't want to put it on the stack.
+ */
+int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
+			  bool prefetch_leaves, struct dm_btree_cursor *c);
+void dm_btree_cursor_end(struct dm_btree_cursor *c);
+int dm_btree_cursor_next(struct dm_btree_cursor *c);
+int dm_btree_cursor_skip(struct dm_btree_cursor *c, uint32_t count);
+int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le);
+
+#endif	/* _LINUX_DM_BTREE_H */