[PATCH 0/3] cgroup: improve cgroup_file_notify() scalability

[PATCH 0/3] cgroup: improve cgroup_file_notify() scalability

[Shakeel Butt]

Jakub Kicinski reported that after the commit d929525c2e30 ("memcg:
net: track network throttling due to memcg memory pressure"), in
Meta fleet, the network intensive workloads started seeing extensive
spinlock contention in the rx path in the function
cgroup_file_notify().

cgroup_file_notify() uses a single global spinlock to protect the
cfile->kn pointer across all cgroups.  On systems with many cgroups
under memory pressure, this becomes a bottleneck as multiple CPUs in
reclaim call __memcg_memory_event(), which walks the cgroup hierarchy
calling cgroup_file_notify() at each level -- all serialized on the
same lock.

This series reduces the lock hold time by moving kernfs_notify()
outside the critical section, adds lockless fast-path checks to avoid
the lock entirely in common cases (torn-down files and rate-limited
bursts), and replaces the global lock with a per-cgroup_file lock to
eliminate cross-cgroup contention.

Shakeel Butt (3):
  cgroup: reduce cgroup_file_kn_lock hold time in cgroup_file_notify()
  cgroup: add lockless fast-path checks to cgroup_file_notify()
  cgroup: replace global cgroup_file_kn_lock with per-cgroup_file lock

 include/linux/cgroup-defs.h |  1 +
 kernel/cgroup/cgroup.c      | 50 +++++++++++++++++++++----------------
 2 files changed, 29 insertions(+), 22 deletions(-)

-- 
2.47.3

[PATCH 1/3] cgroup: reduce cgroup_file_kn_lock hold time in cgroup_file_notify()

[Shakeel Butt]

cgroup_file_notify() calls kernfs_notify() while holding the global
cgroup_file_kn_lock.  kernfs_notify() does non-trivial work including
wake_up_interruptible() and acquisition of a second global spinlock
(kernfs_notify_lock), inflating the hold time.

Take a kernfs_get() reference under the lock and call kernfs_notify()
after dropping it, following the pattern from cgroup_file_show().

Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Reported-by: Jakub Kicinski <kuba@kernel.org>
---
 kernel/cgroup/cgroup.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
index be1d71dda317..33282c7d71e4 100644
--- a/kernel/cgroup/cgroup.c
+++ b/kernel/cgroup/cgroup.c
@@ -4687,6 +4687,7 @@ int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
 void cgroup_file_notify(struct cgroup_file *cfile)
 {
 	unsigned long flags;
+	struct kernfs_node *kn = NULL;
 
 	spin_lock_irqsave(&cgroup_file_kn_lock, flags);
 	if (cfile->kn) {
@@ -4696,11 +4697,17 @@ void cgroup_file_notify(struct cgroup_file *cfile)
 		if (time_in_range(jiffies, last, next)) {
 			timer_reduce(&cfile->notify_timer, next);
 		} else {
-			kernfs_notify(cfile->kn);
+			kn = cfile->kn;
+			kernfs_get(kn);
 			cfile->notified_at = jiffies;
 		}
 	}
 	spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
+
+	if (kn) {
+		kernfs_notify(kn);
+		kernfs_put(kn);
+	}
 }
 EXPORT_SYMBOL_GPL(cgroup_file_notify);
 
-- 
2.47.3

[PATCH 2/3] cgroup: add lockless fast-path checks to cgroup_file_notify()

[Shakeel Butt]

Add two lockless checks before acquiring the lock:

1. READ_ONCE(cfile->kn) NULL check to skip torn-down files.
2. READ_ONCE(cfile->notified_at) check to skip when within the
   rate-limit window (~10ms).

Both checks have safe error directions -- a stale read can only cause
unnecessary lock acquisition, never a missed notification.  Annotate
all write sites with WRITE_ONCE() to pair with the lockless readers.

The trade-off is that trailing timer_reduce() calls during bursts are
skipped, so the deferred notification that delivers the final state
may be lost.  This is acceptable for the primary callers like
__memcg_memory_event() where events keep arriving.

Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Reported-by: Jakub Kicinski <kuba@kernel.org>
---
 kernel/cgroup/cgroup.c | 21 ++++++++++++++-------
 1 file changed, 14 insertions(+), 7 deletions(-)

diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
index 33282c7d71e4..5473ebd0f6c1 100644
--- a/kernel/cgroup/cgroup.c
+++ b/kernel/cgroup/cgroup.c
@@ -1749,7 +1749,7 @@ static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
 		struct cgroup_file *cfile = (void *)css + cft->file_offset;
 
 		spin_lock_irq(&cgroup_file_kn_lock);
-		cfile->kn = NULL;
+		WRITE_ONCE(cfile->kn, NULL);
 		spin_unlock_irq(&cgroup_file_kn_lock);
 
 		timer_delete_sync(&cfile->notify_timer);
@@ -4430,7 +4430,7 @@ static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
 		timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
 
 		spin_lock_irq(&cgroup_file_kn_lock);
-		cfile->kn = kn;
+		WRITE_ONCE(cfile->kn, kn);
 		spin_unlock_irq(&cgroup_file_kn_lock);
 	}
 
@@ -4686,20 +4686,27 @@ int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
  */
 void cgroup_file_notify(struct cgroup_file *cfile)
 {
-	unsigned long flags;
+	unsigned long flags, last, next;
 	struct kernfs_node *kn = NULL;
 
+	if (!READ_ONCE(cfile->kn))
+		return;
+
+	last = READ_ONCE(cfile->notified_at);
+	if (time_before_eq(jiffies, last + CGROUP_FILE_NOTIFY_MIN_INTV))
+		return;
+
 	spin_lock_irqsave(&cgroup_file_kn_lock, flags);
 	if (cfile->kn) {
-		unsigned long last = cfile->notified_at;
-		unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
+		last = cfile->notified_at;
+		next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
 
-		if (time_in_range(jiffies, last, next)) {
+		if (time_before_eq(jiffies, next)) {
 			timer_reduce(&cfile->notify_timer, next);
 		} else {
 			kn = cfile->kn;
 			kernfs_get(kn);
-			cfile->notified_at = jiffies;
+			WRITE_ONCE(cfile->notified_at, jiffies);
 		}
 	}
 	spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
-- 
2.47.3

[PATCH 3/3] cgroup: replace global cgroup_file_kn_lock with per-cgroup_file lock

[Shakeel Butt]

Replace the global cgroup_file_kn_lock with a per-cgroup_file spinlock
to eliminate cross-cgroup contention as it is not really protecting
data shared between different cgroups.

The lock is initialized in cgroup_add_file() alongside timer_setup().
No lock acquisition is needed during initialization since the cgroup
directory is being populated under cgroup_mutex and no concurrent
accessors exist at that point.

Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Reported-by: Jakub Kicinski <kuba@kernel.org>
---
 include/linux/cgroup-defs.h |  1 +
 kernel/cgroup/cgroup.c      | 24 ++++++++----------------
 2 files changed, 9 insertions(+), 16 deletions(-)

diff --git a/include/linux/cgroup-defs.h b/include/linux/cgroup-defs.h
index bb92f5c169ca..ba26b5d05ce3 100644
--- a/include/linux/cgroup-defs.h
+++ b/include/linux/cgroup-defs.h
@@ -167,6 +167,7 @@ struct cgroup_file {
 	struct kernfs_node *kn;
 	unsigned long notified_at;
 	struct timer_list notify_timer;
+	spinlock_t lock;
 };
 
 /*
diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
index 5473ebd0f6c1..b502acad3c5c 100644
--- a/kernel/cgroup/cgroup.c
+++ b/kernel/cgroup/cgroup.c
@@ -107,12 +107,6 @@ static bool cgroup_debug __read_mostly;
  */
 static DEFINE_SPINLOCK(cgroup_idr_lock);
 
-/*
- * Protects cgroup_file->kn for !self csses.  It synchronizes notifications
- * against file removal/re-creation across css hiding.
- */
-static DEFINE_SPINLOCK(cgroup_file_kn_lock);
-
 DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem);
 
 #define cgroup_assert_mutex_or_rcu_locked()				\
@@ -1748,9 +1742,9 @@ static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
 		struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
 		struct cgroup_file *cfile = (void *)css + cft->file_offset;
 
-		spin_lock_irq(&cgroup_file_kn_lock);
+		spin_lock_irq(&cfile->lock);
 		WRITE_ONCE(cfile->kn, NULL);
-		spin_unlock_irq(&cgroup_file_kn_lock);
+		spin_unlock_irq(&cfile->lock);
 
 		timer_delete_sync(&cfile->notify_timer);
 	}
@@ -4428,10 +4422,8 @@ static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
 		struct cgroup_file *cfile = (void *)css + cft->file_offset;
 
 		timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
-
-		spin_lock_irq(&cgroup_file_kn_lock);
-		WRITE_ONCE(cfile->kn, kn);
-		spin_unlock_irq(&cgroup_file_kn_lock);
+		spin_lock_init(&cfile->lock);
+		cfile->kn = kn;
 	}
 
 	return 0;
@@ -4696,7 +4688,7 @@ void cgroup_file_notify(struct cgroup_file *cfile)
 	if (time_before_eq(jiffies, last + CGROUP_FILE_NOTIFY_MIN_INTV))
 		return;
 
-	spin_lock_irqsave(&cgroup_file_kn_lock, flags);
+	spin_lock_irqsave(&cfile->lock, flags);
 	if (cfile->kn) {
 		last = cfile->notified_at;
 		next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
@@ -4709,7 +4701,7 @@ void cgroup_file_notify(struct cgroup_file *cfile)
 			WRITE_ONCE(cfile->notified_at, jiffies);
 		}
 	}
-	spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
+	spin_unlock_irqrestore(&cfile->lock, flags);
 
 	if (kn) {
 		kernfs_notify(kn);
@@ -4727,10 +4719,10 @@ void cgroup_file_show(struct cgroup_file *cfile, bool show)
 {
 	struct kernfs_node *kn;
 
-	spin_lock_irq(&cgroup_file_kn_lock);
+	spin_lock_irq(&cfile->lock);
 	kn = cfile->kn;
 	kernfs_get(kn);
-	spin_unlock_irq(&cgroup_file_kn_lock);
+	spin_unlock_irq(&cfile->lock);
 
 	if (kn)
 		kernfs_show(kn, show);
-- 
2.47.3