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, 393 insertions, 0 deletions

diff --git a/include/net/fq_impl.h b/include/net/fq_impl.h
new file mode 100644
index 000000000..8aca88193
--- /dev/null
+++ b/include/net/fq_impl.h
@@ -0,0 +1,393 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (c) 2016 Qualcomm Atheros, Inc
+ *
+ * Based on net/sched/sch_fq_codel.c
+ */
+#ifndef __NET_SCHED_FQ_IMPL_H
+#define __NET_SCHED_FQ_IMPL_H
+
+#include <net/fq.h>
+
+/* functions that are embedded into includer */
+
+
+static void
+__fq_adjust_removal(struct fq *fq, struct fq_flow *flow, unsigned int packets,
+		    unsigned int bytes, unsigned int truesize)
+{
+	struct fq_tin *tin = flow->tin;
+	int idx;
+
+	tin->backlog_bytes -= bytes;
+	tin->backlog_packets -= packets;
+	flow->backlog -= bytes;
+	fq->backlog -= packets;
+	fq->memory_usage -= truesize;
+
+	if (flow->backlog)
+		return;
+
+	if (flow == &tin->default_flow) {
+		list_del_init(&tin->tin_list);
+		return;
+	}
+
+	idx = flow - fq->flows;
+	__clear_bit(idx, fq->flows_bitmap);
+}
+
+static void fq_adjust_removal(struct fq *fq,
+			      struct fq_flow *flow,
+			      struct sk_buff *skb)
+{
+	__fq_adjust_removal(fq, flow, 1, skb->len, skb->truesize);
+}
+
+static struct sk_buff *fq_flow_dequeue(struct fq *fq,
+				       struct fq_flow *flow)
+{
+	struct sk_buff *skb;
+
+	lockdep_assert_held(&fq->lock);
+
+	skb = __skb_dequeue(&flow->queue);
+	if (!skb)
+		return NULL;
+
+	fq_adjust_removal(fq, flow, skb);
+
+	return skb;
+}
+
+static int fq_flow_drop(struct fq *fq, struct fq_flow *flow,
+			fq_skb_free_t free_func)
+{
+	unsigned int packets = 0, bytes = 0, truesize = 0;
+	struct fq_tin *tin = flow->tin;
+	struct sk_buff *skb;
+	int pending;
+
+	lockdep_assert_held(&fq->lock);
+
+	pending = min_t(int, 32, skb_queue_len(&flow->queue) / 2);
+	do {
+		skb = __skb_dequeue(&flow->queue);
+		if (!skb)
+			break;
+
+		packets++;
+		bytes += skb->len;
+		truesize += skb->truesize;
+		free_func(fq, tin, flow, skb);
+	} while (packets < pending);
+
+	__fq_adjust_removal(fq, flow, packets, bytes, truesize);
+
+	return packets;
+}
+
+static struct sk_buff *fq_tin_dequeue(struct fq *fq,
+				      struct fq_tin *tin,
+				      fq_tin_dequeue_t dequeue_func)
+{
+	struct fq_flow *flow;
+	struct list_head *head;
+	struct sk_buff *skb;
+
+	lockdep_assert_held(&fq->lock);
+
+begin:
+	head = &tin->new_flows;
+	if (list_empty(head)) {
+		head = &tin->old_flows;
+		if (list_empty(head))
+			return NULL;
+	}
+
+	flow = list_first_entry(head, struct fq_flow, flowchain);
+
+	if (flow->deficit <= 0) {
+		flow->deficit += fq->quantum;
+		list_move_tail(&flow->flowchain,
+			       &tin->old_flows);
+		goto begin;
+	}
+
+	skb = dequeue_func(fq, tin, flow);
+	if (!skb) {
+		/* force a pass through old_flows to prevent starvation */
+		if ((head == &tin->new_flows) &&
+		    !list_empty(&tin->old_flows)) {
+			list_move_tail(&flow->flowchain, &tin->old_flows);
+		} else {
+			list_del_init(&flow->flowchain);
+			flow->tin = NULL;
+		}
+		goto begin;
+	}
+
+	flow->deficit -= skb->len;
+	tin->tx_bytes += skb->len;
+	tin->tx_packets++;
+
+	return skb;
+}
+
+static u32 fq_flow_idx(struct fq *fq, struct sk_buff *skb)
+{
+	u32 hash = skb_get_hash(skb);
+
+	return reciprocal_scale(hash, fq->flows_cnt);
+}
+
+static struct fq_flow *fq_flow_classify(struct fq *fq,
+					struct fq_tin *tin, u32 idx,
+					struct sk_buff *skb)
+{
+	struct fq_flow *flow;
+
+	lockdep_assert_held(&fq->lock);
+
+	flow = &fq->flows[idx];
+	if (flow->tin && flow->tin != tin) {
+		flow = &tin->default_flow;
+		tin->collisions++;
+		fq->collisions++;
+	}
+
+	if (!flow->tin)
+		tin->flows++;
+
+	return flow;
+}
+
+static struct fq_flow *fq_find_fattest_flow(struct fq *fq)
+{
+	struct fq_tin *tin;
+	struct fq_flow *flow = NULL;
+	u32 len = 0;
+	int i;
+
+	for_each_set_bit(i, fq->flows_bitmap, fq->flows_cnt) {
+		struct fq_flow *cur = &fq->flows[i];
+		unsigned int cur_len;
+
+		cur_len = cur->backlog;
+		if (cur_len <= len)
+			continue;
+
+		flow = cur;
+		len = cur_len;
+	}
+
+	list_for_each_entry(tin, &fq->tin_backlog, tin_list) {
+		unsigned int cur_len = tin->default_flow.backlog;
+
+		if (cur_len <= len)
+			continue;
+
+		flow = &tin->default_flow;
+		len = cur_len;
+	}
+
+	return flow;
+}
+
+static void fq_tin_enqueue(struct fq *fq,
+			   struct fq_tin *tin, u32 idx,
+			   struct sk_buff *skb,
+			   fq_skb_free_t free_func)
+{
+	struct fq_flow *flow;
+	struct sk_buff *next;
+	bool oom;
+
+	lockdep_assert_held(&fq->lock);
+
+	flow = fq_flow_classify(fq, tin, idx, skb);
+
+	if (!flow->backlog) {
+		if (flow != &tin->default_flow)
+			__set_bit(idx, fq->flows_bitmap);
+		else if (list_empty(&tin->tin_list))
+			list_add(&tin->tin_list, &fq->tin_backlog);
+	}
+
+	flow->tin = tin;
+	skb_list_walk_safe(skb, skb, next) {
+		skb_mark_not_on_list(skb);
+		flow->backlog += skb->len;
+		tin->backlog_bytes += skb->len;
+		tin->backlog_packets++;
+		fq->memory_usage += skb->truesize;
+		fq->backlog++;
+		__skb_queue_tail(&flow->queue, skb);
+	}
+
+	if (list_empty(&flow->flowchain)) {
+		flow->deficit = fq->quantum;
+		list_add_tail(&flow->flowchain,
+			      &tin->new_flows);
+	}
+
+	oom = (fq->memory_usage > fq->memory_limit);
+	while (fq->backlog > fq->limit || oom) {
+		flow = fq_find_fattest_flow(fq);
+		if (!flow)
+			return;
+
+		if (!fq_flow_drop(fq, flow, free_func))
+			return;
+
+		flow->tin->overlimit++;
+		fq->overlimit++;
+		if (oom) {
+			fq->overmemory++;
+			oom = (fq->memory_usage > fq->memory_limit);
+		}
+	}
+}
+
+static void fq_flow_filter(struct fq *fq,
+			   struct fq_flow *flow,
+			   fq_skb_filter_t filter_func,
+			   void *filter_data,
+			   fq_skb_free_t free_func)
+{
+	struct fq_tin *tin = flow->tin;
+	struct sk_buff *skb, *tmp;
+
+	lockdep_assert_held(&fq->lock);
+
+	skb_queue_walk_safe(&flow->queue, skb, tmp) {
+		if (!filter_func(fq, tin, flow, skb, filter_data))
+			continue;
+
+		__skb_unlink(skb, &flow->queue);
+		fq_adjust_removal(fq, flow, skb);
+		free_func(fq, tin, flow, skb);
+	}
+}
+
+static void fq_tin_filter(struct fq *fq,
+			  struct fq_tin *tin,
+			  fq_skb_filter_t filter_func,
+			  void *filter_data,
+			  fq_skb_free_t free_func)
+{
+	struct fq_flow *flow;
+
+	lockdep_assert_held(&fq->lock);
+
+	list_for_each_entry(flow, &tin->new_flows, flowchain)
+		fq_flow_filter(fq, flow, filter_func, filter_data, free_func);
+	list_for_each_entry(flow, &tin->old_flows, flowchain)
+		fq_flow_filter(fq, flow, filter_func, filter_data, free_func);
+}
+
+static void fq_flow_reset(struct fq *fq,
+			  struct fq_flow *flow,
+			  fq_skb_free_t free_func)
+{
+	struct fq_tin *tin = flow->tin;
+	struct sk_buff *skb;
+
+	while ((skb = fq_flow_dequeue(fq, flow)))
+		free_func(fq, tin, flow, skb);
+
+	if (!list_empty(&flow->flowchain)) {
+		list_del_init(&flow->flowchain);
+		if (list_empty(&tin->new_flows) &&
+		    list_empty(&tin->old_flows))
+			list_del_init(&tin->tin_list);
+	}
+
+	flow->tin = NULL;
+
+	WARN_ON_ONCE(flow->backlog);
+}
+
+static void fq_tin_reset(struct fq *fq,
+			 struct fq_tin *tin,
+			 fq_skb_free_t free_func)
+{
+	struct list_head *head;
+	struct fq_flow *flow;
+
+	for (;;) {
+		head = &tin->new_flows;
+		if (list_empty(head)) {
+			head = &tin->old_flows;
+			if (list_empty(head))
+				break;
+		}
+
+		flow = list_first_entry(head, struct fq_flow, flowchain);
+		fq_flow_reset(fq, flow, free_func);
+	}
+
+	WARN_ON_ONCE(!list_empty(&tin->tin_list));
+	WARN_ON_ONCE(tin->backlog_bytes);
+	WARN_ON_ONCE(tin->backlog_packets);
+}
+
+static void fq_flow_init(struct fq_flow *flow)
+{
+	INIT_LIST_HEAD(&flow->flowchain);
+	__skb_queue_head_init(&flow->queue);
+}
+
+static void fq_tin_init(struct fq_tin *tin)
+{
+	INIT_LIST_HEAD(&tin->new_flows);
+	INIT_LIST_HEAD(&tin->old_flows);
+	INIT_LIST_HEAD(&tin->tin_list);
+	fq_flow_init(&tin->default_flow);
+}
+
+static int fq_init(struct fq *fq, int flows_cnt)
+{
+	int i;
+
+	memset(fq, 0, sizeof(fq[0]));
+	spin_lock_init(&fq->lock);
+	INIT_LIST_HEAD(&fq->tin_backlog);
+	fq->flows_cnt = max_t(u32, flows_cnt, 1);
+	fq->quantum = 300;
+	fq->limit = 8192;
+	fq->memory_limit = 16 << 20; /* 16 MBytes */
+
+	fq->flows = kvzalloc_objs(fq->flows[0], fq->flows_cnt);
+	if (!fq->flows)
+		return -ENOMEM;
+
+	fq->flows_bitmap = bitmap_zalloc(fq->flows_cnt, GFP_KERNEL);
+	if (!fq->flows_bitmap) {
+		kvfree(fq->flows);
+		fq->flows = NULL;
+		return -ENOMEM;
+	}
+
+	for (i = 0; i < fq->flows_cnt; i++)
+		fq_flow_init(&fq->flows[i]);
+
+	return 0;
+}
+
+static void fq_reset(struct fq *fq,
+		     fq_skb_free_t free_func)
+{
+	int i;
+
+	for (i = 0; i < fq->flows_cnt; i++)
+		fq_flow_reset(fq, &fq->flows[i], free_func);
+
+	kvfree(fq->flows);
+	fq->flows = NULL;
+
+	bitmap_free(fq->flows_bitmap);
+	fq->flows_bitmap = NULL;
+}
+
+#endif