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, 168 insertions, 0 deletions

diff --git a/net/rds/ib_ring.c b/net/rds/ib_ring.c
new file mode 100644
index 000000000..006b2e441
--- /dev/null
+++ b/net/rds/ib_ring.c
@@ -0,0 +1,168 @@
+/*
+ * Copyright (c) 2006 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+
+#include "rds.h"
+#include "ib.h"
+
+/*
+ * Locking for IB rings.
+ * We assume that allocation is always protected by a mutex
+ * in the caller (this is a valid assumption for the current
+ * implementation).
+ *
+ * Freeing always happens in an interrupt, and hence only
+ * races with allocations, but not with other free()s.
+ *
+ * The interaction between allocation and freeing is that
+ * the alloc code has to determine the number of free entries.
+ * To this end, we maintain two counters; an allocation counter
+ * and a free counter. Both are allowed to run freely, and wrap
+ * around.
+ * The number of used entries is always (alloc_ctr - free_ctr) % NR.
+ *
+ * The current implementation makes free_ctr atomic. When the
+ * caller finds an allocation fails, it should set an "alloc fail"
+ * bit and retry the allocation. The "alloc fail" bit essentially tells
+ * the CQ completion handlers to wake it up after freeing some
+ * more entries.
+ */
+
+/*
+ * This only happens on shutdown.
+ */
+DECLARE_WAIT_QUEUE_HEAD(rds_ib_ring_empty_wait);
+
+void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr)
+{
+	memset(ring, 0, sizeof(*ring));
+	ring->w_nr = nr;
+	rdsdebug("ring %p nr %u\n", ring, ring->w_nr);
+}
+
+static inline u32 __rds_ib_ring_used(struct rds_ib_work_ring *ring)
+{
+	u32 diff;
+
+	/* This assumes that atomic_t has at least as many bits as u32 */
+	diff = ring->w_alloc_ctr - (u32) atomic_read(&ring->w_free_ctr);
+	BUG_ON(diff > ring->w_nr);
+
+	return diff;
+}
+
+void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr)
+{
+	/* We only ever get called from the connection setup code,
+	 * prior to creating the QP. */
+	BUG_ON(__rds_ib_ring_used(ring));
+	ring->w_nr = nr;
+}
+
+static int __rds_ib_ring_empty(struct rds_ib_work_ring *ring)
+{
+	return __rds_ib_ring_used(ring) == 0;
+}
+
+u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos)
+{
+	u32 ret = 0, avail;
+
+	avail = ring->w_nr - __rds_ib_ring_used(ring);
+
+	rdsdebug("ring %p val %u next %u free %u\n", ring, val,
+		 ring->w_alloc_ptr, avail);
+
+	if (val && avail) {
+		ret = min(val, avail);
+		*pos = ring->w_alloc_ptr;
+
+		ring->w_alloc_ptr = (ring->w_alloc_ptr + ret) % ring->w_nr;
+		ring->w_alloc_ctr += ret;
+	}
+
+	return ret;
+}
+
+void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val)
+{
+	ring->w_free_ptr = (ring->w_free_ptr + val) % ring->w_nr;
+	atomic_add(val, &ring->w_free_ctr);
+
+	if (__rds_ib_ring_empty(ring) &&
+	    waitqueue_active(&rds_ib_ring_empty_wait))
+		wake_up(&rds_ib_ring_empty_wait);
+}
+
+void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val)
+{
+	ring->w_alloc_ptr = (ring->w_alloc_ptr - val) % ring->w_nr;
+	ring->w_alloc_ctr -= val;
+}
+
+int rds_ib_ring_empty(struct rds_ib_work_ring *ring)
+{
+	return __rds_ib_ring_empty(ring);
+}
+
+int rds_ib_ring_low(struct rds_ib_work_ring *ring)
+{
+	return __rds_ib_ring_used(ring) <= (ring->w_nr >> 1);
+}
+
+/*
+ * returns the oldest allocated ring entry.  This will be the next one
+ * freed.  This can't be called if there are none allocated.
+ */
+u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring)
+{
+	return ring->w_free_ptr;
+}
+
+/*
+ * returns the number of completed work requests.
+ */
+
+u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest)
+{
+	u32 ret;
+
+	if (oldest <= (unsigned long long)wr_id)
+		ret = (unsigned long long)wr_id - oldest + 1;
+	else
+		ret = ring->w_nr - oldest + (unsigned long long)wr_id + 1;
+
+	rdsdebug("ring %p ret %u wr_id %u oldest %u\n", ring, ret,
+		 wr_id, oldest);
+	return ret;
+}