[PATCH v2 0/3] block: skip the blkcg walk in blk_cgroup_congested() when nothing is throttled

[PATCH v2 0/3] block: skip the blkcg walk in blk_cgroup_congested() when nothing is throttled

[Usama Arif]

blk_cgroup_congested() walks the current task's blkcg ancestor chain on every
readahead decision and, once swap is in use, on every anonymous and shmem
folio allocation.  The answer is almost always "no", but finding that out
costs two loads per level on two cold cache lines, plus an out-of-line
kthread_blkcg() and an RCU read-side pair.  On a fleet profile of hosts
running containers with 5-10 level hierarchies it costs about as much as all
of mutex_lock(), 99.4% of it under __folio_throttle_swaprate().

Patch 3 gates the walk on a global count of blkcgs with a non-zero
congestion_count, so the common case is a load and a predicted branch.

That only works if the count is correctly maintained, currently two teardown
paths can leave a blkcg permanently marked congested.  Today that only hurts
tasks in the affected cgroup, but it hurts them for the life of the cgroup -
readahead cut to a single page, async readahead skipped, and a throttle
scheduled on every anonymous folio allocation.  With a global gate it would
cost every other task on the machine the walk as well.  Patches 1 and 2 fix
those two paths and stand on their own as bugfixes; patch 3 depends on them.

v1 -> v2 (Tejun):
- Rename blkcg_congested_blkcgs to blkcg_nr_congested to make it clear
  that the global tracks a count rather than a boolean.
- Warn if blkcg_css_free() finds a residual congestion_count, while still
  dropping its contribution so it cannot disable the fast path permanently.
- Use atomic_dec_and_test() for the congestion_count 1 -> 0 transition.

Usama Arif (3):
  blk-iolatency: clear delay state when freeing policy data
  blk-iocost: clear delay state when freeing policy data
  block: skip blkcg walk in blk_cgroup_congested() when nothing
    throttled

 block/blk-cgroup.c         | 15 ++++++++++++++-
 block/blk-cgroup.h         | 25 +++++++++++++++++++++----
 block/blk-iocost.c         |  7 +++++++
 block/blk-iolatency.c      |  9 +++++++++
 include/linux/blk-cgroup.h | 23 ++++++++++++++++++++++-
 5 files changed, 73 insertions(+), 6 deletions(-)

-- 
2.53.0-Meta

[PATCH v2 1/3] blk-iolatency: clear delay state when freeing policy data

[Usama Arif]

io.latency can throttle a group which has no latency target of its own.
When a sibling misses its target, check_scale_change() scales down its
peers, and a peer that reaches queue depth one gets blkcg_use_delay()
called on it on every further scale-down, even with min_lat_nsec == 0.

iolatency_pd_offline() resets the target through
iolatency_set_min_lat_nsec(), which clears the delay only on a nonzero
to zero transition, so it never clears such a peer.  Freeing the policy
data then leaves blkg->use_delay set and blkcg->congestion_count
elevated with nothing left that can drop it.

blk_cgroup_congested() then returns true for every task in that cgroup
and its descendants for as long as the cgroup lives: page_cache_sync_ra()
cuts readahead to a single page, page_cache_async_ra() skips it
altogether, and __folio_throttle_swaprate() takes swap_avail_lock and
schedules a throttle on anonymous folio allocation.

Clear the delay in iolatency_pd_free().  By then bio-held blkg
references have drained, or the queue is frozen for policy
deactivation, so check_scale_change() cannot re-arm it.  The free
callback can also see policy data which was never attached to a blkg,
hence the pd->blkg check.

Fixes: d70675121546 ("block: introduce blk-iolatency io controller")
Signed-off-by: Usama Arif <usama.arif@linux.dev>
---
 block/blk-iolatency.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/block/blk-iolatency.c b/block/blk-iolatency.c
index 9eb69010c34e6..2caa79a008ad1 100644
--- a/block/blk-iolatency.c
+++ b/block/blk-iolatency.c
@@ -1043,6 +1043,15 @@ static void iolat_release(struct rcu_head *rcu)
 
 static void iolatency_pd_free(struct blkg_policy_data *pd)
 {
+	struct blkcg_gq *blkg = pd_to_blkg(pd);
+
+	/*
+	 * Groups throttled as collateral have min_lat_nsec == 0, so
+	 * iolatency_pd_offline() leaves their delay set.  Drop it here, where
+	 * no in-flight bio can re-arm it via check_scale_change().
+	 */
+	if (blkg)
+		blkcg_clear_delay(blkg);
 	call_rcu(&pd->rcu_head, iolat_release);
 }
 
-- 
2.53.0-Meta

[PATCH v2 2/3] blk-iocost: clear delay state when freeing policy data

[Usama Arif]

iocg_kick_delay() turns sufficiently large debt into an explicit
block-cgroup delay with blkcg_set_delay(), setting blkg->use_delay to
-1 and incrementing blkcg->congestion_count.  Clearing it again depends
on iocg_kick_delay() running from the period timer, the waitq timer or
the issue path.

ioc_pd_free() removes the iocg from active_iocgs and cancels its waitq
timer, and no further bios can arrive, so once it has run nothing is
left which can reduce the debt and clear the delay.  The blkcg stays
marked congested for the rest of its life.

blk_cgroup_congested() then returns true for every task in that cgroup
and its descendants: page_cache_sync_ra() cuts readahead to a single
page, page_cache_async_ra() skips it altogether, and
__folio_throttle_swaprate() takes swap_avail_lock and schedules a
throttle on anonymous folio allocation.

Clear it explicitly, after the list removal and the synchronous
hrtimer_cancel() so that neither timer processing nor an I/O path can
re-arm it.  The free callback can also see policy data which was never
attached to a blkg, hence the pd->blkg check.

Fixes: 7caa47151ab2 ("blkcg: implement blk-iocost")
Signed-off-by: Usama Arif <usama.arif@linux.dev>
---
 block/blk-iocost.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/block/blk-iocost.c b/block/blk-iocost.c
index b60625613e095..64b92aa3e5d48 100644
--- a/block/blk-iocost.c
+++ b/block/blk-iocost.c
@@ -3063,6 +3063,7 @@ static void iocg_release(struct rcu_head *rcu)
 static void ioc_pd_free(struct blkg_policy_data *pd)
 {
 	struct ioc_gq *iocg = pd_to_iocg(pd);
+	struct blkcg_gq *blkg = pd_to_blkg(pd);
 	struct ioc *ioc = iocg->ioc;
 	unsigned long flags;
 
@@ -3085,6 +3086,12 @@ static void ioc_pd_free(struct blkg_policy_data *pd)
 		hrtimer_cancel(&iocg->waitq_timer);
 	}
 
+	/* off ->active_iocgs and timer gone, so nothing can re-arm the delay */
+	iocg->delay = 0;
+	iocg->indelay_since = 0;
+	if (blkg)
+		blkcg_clear_delay(blkg);
+
 	call_rcu(&pd->rcu_head, iocg_release);
 }
 
-- 
2.53.0-Meta

[PATCH v2 3/3] block: skip blkcg walk in blk_cgroup_congested() when nothing throttled

[Usama Arif]

blk_cgroup_congested() walks the current task's blkcg ancestor chain on
every readahead decision and, once swap is in use, on every anonymous and
shmem folio allocation.  The answer is almost always "no", but finding that
out costs two loads per level on two cold cache lines, plus an out-of-line
kthread_blkcg() and an RCU read-side pair.  On a fleet profile of hosts
running containers with 5-10 level hierarchies it costs about as much as
all of mutex_lock(), 99.4% of it under __folio_throttle_swaprate().

Gate the walk on a global count of blkcgs with a non-zero
congestion_count.  The counter only moves on the 0 <-> 1 transitions of
each blkcg's congestion_count, so the extra atomic stays in the throttle
arm/disarm paths and never appears in steady state. When something is
throttled the counter is non-zero and the walk runs as before.

Signed-off-by: Usama Arif <usama.arif@linux.dev>
---
 block/blk-cgroup.c         | 15 ++++++++++++++-
 block/blk-cgroup.h         | 25 +++++++++++++++++++++----
 include/linux/blk-cgroup.h | 23 ++++++++++++++++++++++-
 3 files changed, 57 insertions(+), 6 deletions(-)

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 354637f3b158c..2b5c29434e426 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -53,6 +53,9 @@ EXPORT_SYMBOL_GPL(blkcg_root);
 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
 EXPORT_SYMBOL_GPL(blkcg_root_css);
 
+/* number of blkcgs with a non-zero congestion_count */
+atomic_t blkcg_nr_congested __read_mostly = ATOMIC_INIT(0);
+
 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
 
 static LIST_HEAD(all_blkcgs);		/* protected by blkcg_pol_mutex */
@@ -1350,6 +1353,16 @@ static void blkcg_css_free(struct cgroup_subsys_state *css)
 	struct blkcg *blkcg = css_to_blkcg(css);
 	int i;
 
+	/*
+	 * Every blkg holds a reference on this css and drops any delay it
+	 * still has from pd_free_fn(), so this is expected to be zero.  Should
+	 * a policy ever leave one behind, drop it here rather than let it pin
+	 * blkcg_nr_congested and disable the fast path for the rest of the
+	 * boot.  Nothing can race with us at this point.
+	 */
+	if (WARN_ON_ONCE(atomic_xchg(&blkcg->congestion_count, 0) > 0))
+		atomic_dec(&blkcg_nr_congested);
+
 	mutex_lock(&blkcg_pol_mutex);
 
 	list_del(&blkcg->all_blkcgs_node);
@@ -2228,7 +2241,7 @@ void blk_cgroup_bio_start(struct bio *bio)
 	put_cpu();
 }
 
-bool blk_cgroup_congested(void)
+bool __blk_cgroup_congested(void)
 {
 	struct blkcg *blkcg;
 	bool ret = false;
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index 615390f751aa2..e67c698391294 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -375,12 +375,29 @@ static inline void blkg_put(struct blkcg_gq *blkg)
 		if (((d_blkg) = blkg_lookup(css_to_blkcg(pos_css),	\
 					    (p_blkg)->q)))
 
+/*
+ * blkcg_nr_congested gates the hierarchy walk in blk_cgroup_congested().
+ * These two helpers keep it in step with each blkcg's congestion_count in
+ * normal operation; blkcg_css_free() drops a residual count as a backstop.
+ */
+static inline void blkcg_inc_congestion_count(struct blkcg *blkcg)
+{
+	if (atomic_inc_return(&blkcg->congestion_count) == 1)
+		atomic_inc(&blkcg_nr_congested);
+}
+
+static inline void blkcg_dec_congestion_count(struct blkcg *blkcg)
+{
+	if (atomic_dec_and_test(&blkcg->congestion_count))
+		atomic_dec(&blkcg_nr_congested);
+}
+
 static inline void blkcg_use_delay(struct blkcg_gq *blkg)
 {
 	if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
 		return;
 	if (atomic_add_return(1, &blkg->use_delay) == 1)
-		atomic_inc(&blkg->blkcg->congestion_count);
+		blkcg_inc_congestion_count(blkg->blkcg);
 }
 
 static inline int blkcg_unuse_delay(struct blkcg_gq *blkg)
@@ -405,7 +422,7 @@ static inline int blkcg_unuse_delay(struct blkcg_gq *blkg)
 	if (old == 0)
 		return 0;
 	if (old == 1)
-		atomic_dec(&blkg->blkcg->congestion_count);
+		blkcg_dec_congestion_count(blkg->blkcg);
 	return 1;
 }
 
@@ -424,7 +441,7 @@ static inline void blkcg_set_delay(struct blkcg_gq *blkg, u64 delay)
 
 	/* We only want 1 person setting the congestion count for this blkg. */
 	if (!old && atomic_try_cmpxchg(&blkg->use_delay, &old, -1))
-		atomic_inc(&blkg->blkcg->congestion_count);
+		blkcg_inc_congestion_count(blkg->blkcg);
 
 	atomic64_set(&blkg->delay_nsec, delay);
 }
@@ -441,7 +458,7 @@ static inline void blkcg_clear_delay(struct blkcg_gq *blkg)
 
 	/* We only want 1 person clearing the congestion count for this blkg. */
 	if (old && atomic_try_cmpxchg(&blkg->use_delay, &old, 0))
-		atomic_dec(&blkg->blkcg->congestion_count);
+		blkcg_dec_congestion_count(blkg->blkcg);
 }
 
 /**
diff --git a/include/linux/blk-cgroup.h b/include/linux/blk-cgroup.h
index dd5841a42c331..58abde49f8c59 100644
--- a/include/linux/blk-cgroup.h
+++ b/include/linux/blk-cgroup.h
@@ -14,6 +14,8 @@
  * 	              Nauman Rafique <nauman@google.com>
  */
 
+#include <linux/atomic.h>
+#include <linux/compiler.h>
 #include <linux/types.h>
 
 struct bio;
@@ -24,10 +26,29 @@ struct gendisk;
 
 #ifdef CONFIG_BLK_CGROUP
 extern struct cgroup_subsys_state * const blkcg_root_css;
+extern atomic_t blkcg_nr_congested;
 
 void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay);
 void blkcg_maybe_throttle_current(void);
-bool blk_cgroup_congested(void);
+bool __blk_cgroup_congested(void);
+
+/**
+ * blk_cgroup_congested - is the current task in a throttled blkcg?
+ *
+ * Called from mm hot paths where the answer is almost always false, so keep
+ * that case to a load and a branch and only walk the hierarchy out of line
+ * when something in the system really is throttled.
+ *
+ * Return: %true if the current task's blkcg or any of its ancestors is
+ * throttled, %false otherwise.
+ */
+static inline bool blk_cgroup_congested(void)
+{
+	if (likely(!atomic_read(&blkcg_nr_congested)))
+		return false;
+	return __blk_cgroup_congested();
+}
+
 void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css);
 void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css);
 struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css);
-- 
2.53.0-Meta

Re: [PATCH v2 0/3] block: skip the blkcg walk in blk_cgroup_congested() when nothing is throttled

[Tejun Heo]

On Fri, Aug 14, 2026 at 09:56:36AM -0700, Usama Arif wrote:
> blk_cgroup_congested() walks the current task's blkcg ancestor chain on every
> readahead decision and, once swap is in use, on every anonymous and shmem
> folio allocation.  The answer is almost always "no", but finding that out
> costs two loads per level on two cold cache lines, plus an out-of-line
> kthread_blkcg() and an RCU read-side pair.  On a fleet profile of hosts
> running containers with 5-10 level hierarchies it costs about as much as all
> of mutex_lock(), 99.4% of it under __folio_throttle_swaprate().
> 
> Patch 3 gates the walk on a global count of blkcgs with a non-zero
> congestion_count, so the common case is a load and a predicted branch.
> 
> That only works if the count is correctly maintained, currently two teardown
> paths can leave a blkcg permanently marked congested.  Today that only hurts
> tasks in the affected cgroup, but it hurts them for the life of the cgroup -
> readahead cut to a single page, async readahead skipped, and a throttle
> scheduled on every anonymous folio allocation.  With a global gate it would
> cost every other task on the machine the walk as well.  Patches 1 and 2 fix
> those two paths and stand on their own as bugfixes; patch 3 depends on them.

For the series,

Acked-by: Tejun Heo <tj@kernel.org>

Thanks.

-- 
tejun