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, 200 insertions, 0 deletions

diff --git a/tools/testing/radix-tree/regression1.c b/tools/testing/radix-tree/regression1.c
new file mode 100644
index 000000000..63f468bf8
--- /dev/null
+++ b/tools/testing/radix-tree/regression1.c
@@ -0,0 +1,200 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Regression1
+ * Description:
+ * Salman Qazi describes the following radix-tree bug:
+ *
+ * In the following case, we get can get a deadlock:
+ *
+ * 0.  The radix tree contains two items, one has the index 0.
+ * 1.  The reader (in this case find_get_pages) takes the rcu_read_lock.
+ * 2.  The reader acquires slot(s) for item(s) including the index 0 item.
+ * 3.  The non-zero index item is deleted, and as a consequence the other item
+ *     is moved to the root of the tree. The place where it used to be is queued
+ *     for deletion after the readers finish.
+ * 3b. The zero item is deleted, removing it from the direct slot, it remains in
+ *     the rcu-delayed indirect node.
+ * 4.  The reader looks at the index 0 slot, and finds that the page has 0 ref
+ *     count
+ * 5.  The reader looks at it again, hoping that the item will either be freed
+ *     or the ref count will increase. This never happens, as the slot it is
+ *     looking at will never be updated. Also, this slot can never be reclaimed
+ *     because the reader is holding rcu_read_lock and is in an infinite loop.
+ *
+ * The fix is to re-use the same "indirect" pointer case that requires a slot
+ * lookup retry into a general "retry the lookup" bit.
+ *
+ * Running:
+ * This test should run to completion in a few seconds. The above bug would
+ * cause it to hang indefinitely.
+ *
+ * Upstream commit:
+ * Not yet
+ */
+#include <linux/kernel.h>
+#include <linux/gfp.h>
+#include <linux/slab.h>
+#include <linux/radix-tree.h>
+#include <linux/rcupdate.h>
+#include <stdlib.h>
+#include <pthread.h>
+#include <stdio.h>
+#include <assert.h>
+
+#include "regression.h"
+
+static RADIX_TREE(mt_tree, GFP_KERNEL);
+
+struct page {
+	pthread_mutex_t lock;
+	struct rcu_head rcu;
+	int count;
+	unsigned long index;
+};
+
+static struct page *page_alloc(int index)
+{
+	struct page *p;
+	p = malloc(sizeof(struct page));
+	p->count = 1;
+	p->index = index;
+	pthread_mutex_init(&p->lock, NULL);
+
+	return p;
+}
+
+static void page_rcu_free(struct rcu_head *rcu)
+{
+	struct page *p = container_of(rcu, struct page, rcu);
+	assert(!p->count);
+	pthread_mutex_destroy(&p->lock);
+	free(p);
+}
+
+static void page_free(struct page *p)
+{
+	call_rcu(&p->rcu, page_rcu_free);
+}
+
+static unsigned find_get_pages(unsigned long start,
+			    unsigned int nr_pages, struct page **pages)
+{
+	XA_STATE(xas, &mt_tree, start);
+	struct page *page;
+	unsigned int ret = 0;
+
+	rcu_read_lock();
+	xas_for_each(&xas, page, ULONG_MAX) {
+		if (xas_retry(&xas, page))
+			continue;
+
+		pthread_mutex_lock(&page->lock);
+		if (!page->count)
+			goto unlock;
+
+		/* don't actually update page refcount */
+		pthread_mutex_unlock(&page->lock);
+
+		/* Has the page moved? */
+		if (unlikely(page != xas_reload(&xas)))
+			goto put_page;
+
+		pages[ret] = page;
+		ret++;
+		continue;
+unlock:
+		pthread_mutex_unlock(&page->lock);
+put_page:
+		xas_reset(&xas);
+	}
+	rcu_read_unlock();
+	return ret;
+}
+
+static pthread_barrier_t worker_barrier;
+
+static void *regression1_fn(void *arg)
+{
+	rcu_register_thread();
+
+	if (pthread_barrier_wait(&worker_barrier) ==
+			PTHREAD_BARRIER_SERIAL_THREAD) {
+		int j;
+
+		for (j = 0; j < 1000000; j++) {
+			struct page *p;
+
+			p = page_alloc(0);
+			xa_lock(&mt_tree);
+			radix_tree_insert(&mt_tree, 0, p);
+			xa_unlock(&mt_tree);
+
+			p = page_alloc(1);
+			xa_lock(&mt_tree);
+			radix_tree_insert(&mt_tree, 1, p);
+			xa_unlock(&mt_tree);
+
+			xa_lock(&mt_tree);
+			p = radix_tree_delete(&mt_tree, 1);
+			pthread_mutex_lock(&p->lock);
+			p->count--;
+			pthread_mutex_unlock(&p->lock);
+			xa_unlock(&mt_tree);
+			page_free(p);
+
+			xa_lock(&mt_tree);
+			p = radix_tree_delete(&mt_tree, 0);
+			pthread_mutex_lock(&p->lock);
+			p->count--;
+			pthread_mutex_unlock(&p->lock);
+			xa_unlock(&mt_tree);
+			page_free(p);
+		}
+	} else {
+		int j;
+
+		for (j = 0; j < 100000000; j++) {
+			struct page *pages[10];
+
+			find_get_pages(0, 10, pages);
+		}
+	}
+
+	rcu_unregister_thread();
+
+	return NULL;
+}
+
+static pthread_t *threads;
+void regression1_test(void)
+{
+	int nr_threads;
+	int i;
+	long arg;
+
+	/* Regression #1 */
+	printv(1, "running regression test 1, should finish in under a minute\n");
+	nr_threads = 2;
+	pthread_barrier_init(&worker_barrier, NULL, nr_threads);
+
+	threads = malloc(nr_threads * sizeof(*threads));
+
+	for (i = 0; i < nr_threads; i++) {
+		arg = i;
+		if (pthread_create(&threads[i], NULL, regression1_fn, (void *)arg)) {
+			perror("pthread_create");
+			exit(1);
+		}
+	}
+
+	for (i = 0; i < nr_threads; i++) {
+		if (pthread_join(threads[i], NULL)) {
+			perror("pthread_join");
+			exit(1);
+		}
+	}
+
+	free(threads);
+
+	printv(1, "regression test 1, done\n");
+}