[PATCH V2 0/2] percpu_ref & block: reduce memory footprint of percpu_ref in fast path

[PATCH V2 0/2] percpu_ref & block: reduce memory footprint of percpu_ref in fast path

[Ming Lei]

Hi,

The 1st patch removes memory footprint of percpu_ref in fast path
from 7 words to 2 words, since it is often used in fast path and
embedded in user struct.

The 2nd patch moves .q_usage_counter to 1st cacheline of
'request_queue'.

Simple test on null_blk shows ~2% IOPS boost on one 16cores(two threads
per core) machine, dual socket/numa.

V2:
	- pass 'gfp' to kzalloc() for fixing block/027 failure reported by
	kernel test robot
	- protect percpu_ref_is_zero() with destroying percpu-refcount by
	spin lock  

Ming Lei (2):
  percpu_ref: reduce memory footprint of percpu_ref in fast path
  block: move 'q_usage_counter' into front of 'request_queue'

 drivers/infiniband/sw/rdmavt/mr.c |   2 +-
 include/linux/blkdev.h            |   3 +-
 include/linux/percpu-refcount.h   |  45 ++++-------
 lib/percpu-refcount.c             | 128 ++++++++++++++++++++++--------
 4 files changed, 115 insertions(+), 63 deletions(-)

Cc: Sagi Grimberg <sagi@grimberg.me>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Bart Van Assche <bvanassche@acm.org>
-- 
2.25.2

[PATCH V2 1/2] percpu_ref: reduce memory footprint of percpu_ref in fast path

[Ming Lei]

'struct percpu_ref' is often embedded into one user structure, and the
instance is usually referenced in fast path, however actually only
'percpu_count_ptr' is needed in fast path.

So move other fields into one new structure of 'percpu_ref_data', and
allocate it dynamically via kzalloc(), then memory footprint of
'percpu_ref' in fast path is reduced a lot and becomes suitable to put
into hot cacheline of user structure.

Cc: Sagi Grimberg <sagi@grimberg.me>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
---
 drivers/infiniband/sw/rdmavt/mr.c |   2 +-
 include/linux/percpu-refcount.h   |  45 ++++-------
 lib/percpu-refcount.c             | 128 ++++++++++++++++++++++--------
 3 files changed, 113 insertions(+), 62 deletions(-)

diff --git a/drivers/infiniband/sw/rdmavt/mr.c b/drivers/infiniband/sw/rdmavt/mr.c
index 2f7c25fea44a..8490fdb9c91e 100644
--- a/drivers/infiniband/sw/rdmavt/mr.c
+++ b/drivers/infiniband/sw/rdmavt/mr.c
@@ -499,7 +499,7 @@ static int rvt_check_refs(struct rvt_mregion *mr, const char *t)
 		rvt_pr_err(rdi,
 			   "%s timeout mr %p pd %p lkey %x refcount %ld\n",
 			   t, mr, mr->pd, mr->lkey,
-			   atomic_long_read(&mr->refcount.count));
+			   atomic_long_read(&mr->refcount.data->count));
 		rvt_get_mr(mr);
 		return -EBUSY;
 	}
diff --git a/include/linux/percpu-refcount.h b/include/linux/percpu-refcount.h
index 87d8a38bdea1..1d6ed9ca23dd 100644
--- a/include/linux/percpu-refcount.h
+++ b/include/linux/percpu-refcount.h
@@ -92,18 +92,23 @@ enum {
 	PERCPU_REF_ALLOW_REINIT	= 1 << 2,
 };
 
-struct percpu_ref {
+struct percpu_ref_data {
 	atomic_long_t		count;
-	/*
-	 * The low bit of the pointer indicates whether the ref is in percpu
-	 * mode; if set, then get/put will manipulate the atomic_t.
-	 */
-	unsigned long		percpu_count_ptr;
 	percpu_ref_func_t	*release;
 	percpu_ref_func_t	*confirm_switch;
 	bool			force_atomic:1;
 	bool			allow_reinit:1;
 	struct rcu_head		rcu;
+	struct percpu_ref	*ref;
+};
+
+struct percpu_ref {
+	/*
+	 * The low bit of the pointer indicates whether the ref is in percpu
+	 * mode; if set, then get/put will manipulate the atomic_t.
+	 */
+	unsigned long		percpu_count_ptr;
+	struct percpu_ref_data  *data;
 };
 
 int __must_check percpu_ref_init(struct percpu_ref *ref,
@@ -118,6 +123,7 @@ void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
 				 percpu_ref_func_t *confirm_kill);
 void percpu_ref_resurrect(struct percpu_ref *ref);
 void percpu_ref_reinit(struct percpu_ref *ref);
+bool percpu_ref_is_zero(struct percpu_ref *ref);
 
 /**
  * percpu_ref_kill - drop the initial ref
@@ -191,7 +197,7 @@ static inline void percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr)
 	if (__ref_is_percpu(ref, &percpu_count))
 		this_cpu_add(*percpu_count, nr);
 	else
-		atomic_long_add(nr, &ref->count);
+		atomic_long_add(nr, &ref->data->count);
 
 	rcu_read_unlock();
 }
@@ -231,7 +237,7 @@ static inline bool percpu_ref_tryget_many(struct percpu_ref *ref,
 		this_cpu_add(*percpu_count, nr);
 		ret = true;
 	} else {
-		ret = atomic_long_add_unless(&ref->count, nr, 0);
+		ret = atomic_long_add_unless(&ref->data->count, nr, 0);
 	}
 
 	rcu_read_unlock();
@@ -279,7 +285,7 @@ static inline bool percpu_ref_tryget_live(struct percpu_ref *ref)
 		this_cpu_inc(*percpu_count);
 		ret = true;
 	} else if (!(ref->percpu_count_ptr & __PERCPU_REF_DEAD)) {
-		ret = atomic_long_inc_not_zero(&ref->count);
+		ret = atomic_long_inc_not_zero(&ref->data->count);
 	}
 
 	rcu_read_unlock();
@@ -305,8 +311,8 @@ static inline void percpu_ref_put_many(struct percpu_ref *ref, unsigned long nr)
 
 	if (__ref_is_percpu(ref, &percpu_count))
 		this_cpu_sub(*percpu_count, nr);
-	else if (unlikely(atomic_long_sub_and_test(nr, &ref->count)))
-		ref->release(ref);
+	else if (unlikely(atomic_long_sub_and_test(nr, &ref->data->count)))
+		ref->data->release(ref);
 
 	rcu_read_unlock();
 }
@@ -339,21 +345,4 @@ static inline bool percpu_ref_is_dying(struct percpu_ref *ref)
 	return ref->percpu_count_ptr & __PERCPU_REF_DEAD;
 }
 
-/**
- * percpu_ref_is_zero - test whether a percpu refcount reached zero
- * @ref: percpu_ref to test
- *
- * Returns %true if @ref reached zero.
- *
- * This function is safe to call as long as @ref is between init and exit.
- */
-static inline bool percpu_ref_is_zero(struct percpu_ref *ref)
-{
-	unsigned long __percpu *percpu_count;
-
-	if (__ref_is_percpu(ref, &percpu_count))
-		return false;
-	return !atomic_long_read(&ref->count);
-}
-
 #endif
diff --git a/lib/percpu-refcount.c b/lib/percpu-refcount.c
index 0ba686b8fe57..e2931784ae47 100644
--- a/lib/percpu-refcount.c
+++ b/lib/percpu-refcount.c
@@ -4,6 +4,7 @@
 #include <linux/kernel.h>
 #include <linux/sched.h>
 #include <linux/wait.h>
+#include <linux/slab.h>
 #include <linux/percpu-refcount.h>
 
 /*
@@ -64,18 +65,25 @@ int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release,
 	size_t align = max_t(size_t, 1 << __PERCPU_REF_FLAG_BITS,
 			     __alignof__(unsigned long));
 	unsigned long start_count = 0;
+	struct percpu_ref_data *data;
 
 	ref->percpu_count_ptr = (unsigned long)
 		__alloc_percpu_gfp(sizeof(unsigned long), align, gfp);
 	if (!ref->percpu_count_ptr)
 		return -ENOMEM;
 
-	ref->force_atomic = flags & PERCPU_REF_INIT_ATOMIC;
-	ref->allow_reinit = flags & PERCPU_REF_ALLOW_REINIT;
+	data = kzalloc(sizeof(*ref->data), gfp);
+	if (!data) {
+		free_percpu((void __percpu *)ref->percpu_count_ptr);
+		return -ENOMEM;
+	}
+
+	data->force_atomic = flags & PERCPU_REF_INIT_ATOMIC;
+	data->allow_reinit = flags & PERCPU_REF_ALLOW_REINIT;
 
 	if (flags & (PERCPU_REF_INIT_ATOMIC | PERCPU_REF_INIT_DEAD)) {
 		ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
-		ref->allow_reinit = true;
+		data->allow_reinit = true;
 	} else {
 		start_count += PERCPU_COUNT_BIAS;
 	}
@@ -85,14 +93,28 @@ int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release,
 	else
 		start_count++;
 
-	atomic_long_set(&ref->count, start_count);
+	atomic_long_set(&data->count, start_count);
 
-	ref->release = release;
-	ref->confirm_switch = NULL;
+	data->release = release;
+	data->confirm_switch = NULL;
+	data->ref = ref;
+	ref->data = data;
 	return 0;
 }
 EXPORT_SYMBOL_GPL(percpu_ref_init);
 
+static void __percpu_ref_exit(struct percpu_ref *ref)
+{
+	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
+
+	if (percpu_count) {
+		/* non-NULL confirm_switch indicates switching in progress */
+		WARN_ON_ONCE(ref->data->confirm_switch);
+		free_percpu(percpu_count);
+		ref->percpu_count_ptr = __PERCPU_REF_ATOMIC_DEAD;
+	}
+}
+
 /**
  * percpu_ref_exit - undo percpu_ref_init()
  * @ref: percpu_ref to exit
@@ -105,27 +127,33 @@ EXPORT_SYMBOL_GPL(percpu_ref_init);
  */
 void percpu_ref_exit(struct percpu_ref *ref)
 {
-	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
+	struct percpu_ref_data *data = ref->data;
+	unsigned long flags;
 
-	if (percpu_count) {
-		/* non-NULL confirm_switch indicates switching in progress */
-		WARN_ON_ONCE(ref->confirm_switch);
-		free_percpu(percpu_count);
-		ref->percpu_count_ptr = __PERCPU_REF_ATOMIC_DEAD;
-	}
+	__percpu_ref_exit(ref);
+
+	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
+	ref->percpu_count_ptr |= atomic_long_read(&ref->data->count) <<
+		__PERCPU_REF_FLAG_BITS;
+	ref->data = NULL;
+	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
+
+	kfree(data);
 }
 EXPORT_SYMBOL_GPL(percpu_ref_exit);
 
 static void percpu_ref_call_confirm_rcu(struct rcu_head *rcu)
 {
-	struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu);
+	struct percpu_ref_data *data = container_of(rcu,
+			struct percpu_ref_data, rcu);
+	struct percpu_ref *ref = data->ref;
 
-	ref->confirm_switch(ref);
-	ref->confirm_switch = NULL;
+	data->confirm_switch(ref);
+	data->confirm_switch = NULL;
 	wake_up_all(&percpu_ref_switch_waitq);
 
-	if (!ref->allow_reinit)
-		percpu_ref_exit(ref);
+	if (!data->allow_reinit)
+		__percpu_ref_exit(ref);
 
 	/* drop ref from percpu_ref_switch_to_atomic() */
 	percpu_ref_put(ref);
@@ -133,7 +161,9 @@ static void percpu_ref_call_confirm_rcu(struct rcu_head *rcu)
 
 static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu)
 {
-	struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu);
+	struct percpu_ref_data *data = container_of(rcu,
+			struct percpu_ref_data, rcu);
+	struct percpu_ref *ref = data->ref;
 	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
 	unsigned long count = 0;
 	int cpu;
@@ -142,7 +172,7 @@ static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu)
 		count += *per_cpu_ptr(percpu_count, cpu);
 
 	pr_debug("global %lu percpu %lu\n",
-		 atomic_long_read(&ref->count), count);
+		 atomic_long_read(&data->count), count);
 
 	/*
 	 * It's crucial that we sum the percpu counters _before_ adding the sum
@@ -156,11 +186,11 @@ static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu)
 	 * reaching 0 before we add the percpu counts. But doing it at the same
 	 * time is equivalent and saves us atomic operations:
 	 */
-	atomic_long_add((long)count - PERCPU_COUNT_BIAS, &ref->count);
+	atomic_long_add((long)count - PERCPU_COUNT_BIAS, &data->count);
 
-	WARN_ONCE(atomic_long_read(&ref->count) <= 0,
+	WARN_ONCE(atomic_long_read(&data->count) <= 0,
 		  "percpu ref (%ps) <= 0 (%ld) after switching to atomic",
-		  ref->release, atomic_long_read(&ref->count));
+		  data->release, atomic_long_read(&data->count));
 
 	/* @ref is viewed as dead on all CPUs, send out switch confirmation */
 	percpu_ref_call_confirm_rcu(rcu);
@@ -186,10 +216,11 @@ static void __percpu_ref_switch_to_atomic(struct percpu_ref *ref,
 	 * Non-NULL ->confirm_switch is used to indicate that switching is
 	 * in progress.  Use noop one if unspecified.
 	 */
-	ref->confirm_switch = confirm_switch ?: percpu_ref_noop_confirm_switch;
+	ref->data->confirm_switch = confirm_switch ?:
+		percpu_ref_noop_confirm_switch;
 
 	percpu_ref_get(ref);	/* put after confirmation */
-	call_rcu(&ref->rcu, percpu_ref_switch_to_atomic_rcu);
+	call_rcu(&ref->data->rcu, percpu_ref_switch_to_atomic_rcu);
 }
 
 static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref)
@@ -202,10 +233,10 @@ static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref)
 	if (!(ref->percpu_count_ptr & __PERCPU_REF_ATOMIC))
 		return;
 
-	if (WARN_ON_ONCE(!ref->allow_reinit))
+	if (WARN_ON_ONCE(!ref->data->allow_reinit))
 		return;
 
-	atomic_long_add(PERCPU_COUNT_BIAS, &ref->count);
+	atomic_long_add(PERCPU_COUNT_BIAS, &ref->data->count);
 
 	/*
 	 * Restore per-cpu operation.  smp_store_release() is paired
@@ -223,6 +254,8 @@ static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref)
 static void __percpu_ref_switch_mode(struct percpu_ref *ref,
 				     percpu_ref_func_t *confirm_switch)
 {
+	struct percpu_ref_data *data = ref->data;
+
 	lockdep_assert_held(&percpu_ref_switch_lock);
 
 	/*
@@ -230,10 +263,10 @@ static void __percpu_ref_switch_mode(struct percpu_ref *ref,
 	 * its completion.  If the caller ensures that ATOMIC switching
 	 * isn't in progress, this function can be called from any context.
 	 */
-	wait_event_lock_irq(percpu_ref_switch_waitq, !ref->confirm_switch,
+	wait_event_lock_irq(percpu_ref_switch_waitq, !data->confirm_switch,
 			    percpu_ref_switch_lock);
 
-	if (ref->force_atomic || (ref->percpu_count_ptr & __PERCPU_REF_DEAD))
+	if (data->force_atomic || (ref->percpu_count_ptr & __PERCPU_REF_DEAD))
 		__percpu_ref_switch_to_atomic(ref, confirm_switch);
 	else
 		__percpu_ref_switch_to_percpu(ref);
@@ -266,7 +299,7 @@ void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
 
 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
 
-	ref->force_atomic = true;
+	ref->data->force_atomic = true;
 	__percpu_ref_switch_mode(ref, confirm_switch);
 
 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
@@ -284,7 +317,7 @@ EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic);
 void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref)
 {
 	percpu_ref_switch_to_atomic(ref, NULL);
-	wait_event(percpu_ref_switch_waitq, !ref->confirm_switch);
+	wait_event(percpu_ref_switch_waitq, !ref->data->confirm_switch);
 }
 EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic_sync);
 
@@ -312,7 +345,7 @@ void percpu_ref_switch_to_percpu(struct percpu_ref *ref)
 
 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
 
-	ref->force_atomic = false;
+	ref->data->force_atomic = false;
 	__percpu_ref_switch_mode(ref, NULL);
 
 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
@@ -344,7 +377,8 @@ void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
 
 	WARN_ONCE(ref->percpu_count_ptr & __PERCPU_REF_DEAD,
-		  "%s called more than once on %ps!", __func__, ref->release);
+		  "%s called more than once on %ps!", __func__,
+		  ref->data->release);
 
 	ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
 	__percpu_ref_switch_mode(ref, confirm_kill);
@@ -354,6 +388,34 @@ void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
 }
 EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm);
 
+/**
+ * percpu_ref_is_zero - test whether a percpu refcount reached zero
+ * @ref: percpu_ref to test
+ *
+ * Returns %true if @ref reached zero.
+ *
+ * This function is safe to call as long as @ref is between init and exit.
+ */
+bool percpu_ref_is_zero(struct percpu_ref *ref)
+{
+	unsigned long __percpu *percpu_count;
+	unsigned long count, flags;
+
+	if (__ref_is_percpu(ref, &percpu_count))
+		return false;
+
+	/* protect us from being destroyed */
+	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
+	if (ref->data)
+		count = atomic_long_read(&ref->data->count);
+	else
+		count = ref->percpu_count_ptr >> __PERCPU_REF_FLAG_BITS;
+	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
+
+	return count == 0;
+}
+EXPORT_SYMBOL_GPL(percpu_ref_is_zero);
+
 /**
  * percpu_ref_reinit - re-initialize a percpu refcount
  * @ref: perpcu_ref to re-initialize
-- 
2.25.2

Re: [PATCH V2 1/2] percpu_ref: reduce memory footprint of percpu_ref in fast path

[Christoph Hellwig]

On Wed, Sep 02, 2020 at 08:26:42PM +0800, Ming Lei wrote:
> 'struct percpu_ref' is often embedded into one user structure, and the
> instance is usually referenced in fast path, however actually only
> 'percpu_count_ptr' is needed in fast path.
> 
> So move other fields into one new structure of 'percpu_ref_data', and
> allocate it dynamically via kzalloc(), then memory footprint of
> 'percpu_ref' in fast path is reduced a lot and becomes suitable to put
> into hot cacheline of user structure.

This looks good:

Reviewed-by: Christoph Hellwig <hch@lst.de>

[PATCH V2 2/2] block: move 'q_usage_counter' into front of 'request_queue'

[Ming Lei]

The field of 'q_usage_counter' is always fetched in fast path of every
block driver, and move it into front of 'request_queue', so it can be
fetched into 1st cacheline of 'request_queue' instance.

Cc: Sagi Grimberg <sagi@grimberg.me>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
---
 include/linux/blkdev.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index d0d61bc81615..7575fa0aae6e 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -397,6 +397,8 @@ struct request_queue {
 	struct request		*last_merge;
 	struct elevator_queue	*elevator;
 
+	struct percpu_ref	q_usage_counter;
+
 	struct blk_queue_stats	*stats;
 	struct rq_qos		*rq_qos;
 
@@ -567,7 +569,6 @@ struct request_queue {
 	 * percpu_ref_kill() and percpu_ref_reinit().
 	 */
 	struct mutex		mq_freeze_lock;
-	struct percpu_ref	q_usage_counter;
 
 	struct blk_mq_tag_set	*tag_set;
 	struct list_head	tag_set_list;
-- 
2.25.2

Re: [PATCH V2 2/2] block: move 'q_usage_counter' into front of 'request_queue'

[Christoph Hellwig]

On Wed, Sep 02, 2020 at 08:26:43PM +0800, Ming Lei wrote:
> The field of 'q_usage_counter' is always fetched in fast path of every
> block driver, and move it into front of 'request_queue', so it can be
> fetched into 1st cacheline of 'request_queue' instance.

Looks good,

Reviewed-by: Christoph Hellwig <hch@lst.de>

Re: [PATCH V2 0/2] percpu_ref & block: reduce memory footprint of percpu_ref in fast path

[Jens Axboe]

On 9/2/20 6:26 AM, Ming Lei wrote:
> Hi,
> 
> The 1st patch removes memory footprint of percpu_ref in fast path
> from 7 words to 2 words, since it is often used in fast path and
> embedded in user struct.
> 
> The 2nd patch moves .q_usage_counter to 1st cacheline of
> 'request_queue'.
> 
> Simple test on null_blk shows ~2% IOPS boost on one 16cores(two threads
> per core) machine, dual socket/numa.
> 
> V2:
> 	- pass 'gfp' to kzalloc() for fixing block/027 failure reported by
> 	kernel test robot
> 	- protect percpu_ref_is_zero() with destroying percpu-refcount by
> 	spin lock  
> 
> Ming Lei (2):
>   percpu_ref: reduce memory footprint of percpu_ref in fast path
>   block: move 'q_usage_counter' into front of 'request_queue'

Applied, thanks.

-- 
Jens Axboe