Re: [PATCH 1/3] ptr_ring: batch ring zeroing

2 messages, 2 authors, 2017-05-10 · open the first message on its own page

Re: [PATCH 1/3] ptr_ring: batch ring zeroing

From: Jesper Dangaard Brouer <hidden>
Date: 2017-05-10 09:18:22

On Tue, 9 May 2017 16:33:14 +0300
"Michael S. Tsirkin" [off-list ref] wrote:
On Sat, Apr 08, 2017 at 02:14:08PM +0200, Jesper Dangaard Brouer wrote:
quoted
On Fri, 7 Apr 2017 08:49:57 +0300
"Michael S. Tsirkin" [off-list ref] wrote:
  
quoted
A known weakness in ptr_ring design is that it does not handle well the
situation when ring is almost full: as entries are consumed they are
immediately used again by the producer, so consumer and producer are
writing to a shared cache line.

To fix this, add batching to consume calls: as entries are
consumed do not write NULL into the ring until we get
a multiple (in current implementation 2x) of cache lines
away from the producer. At that point, write them all out.

We do the write out in the reverse order to keep
producer from sharing cache with consumer for as long
as possible.

Writeout also triggers when ring wraps around - there's
no special reason to do this but it helps keep the code
a bit simpler.

What should we do if getting away from producer by 2 cache lines
would mean we are keeping the ring moe than half empty?
Maybe we should reduce the batching in this case,
current patch simply reduces the batching.

Notes:
- it is no longer true that a call to consume guarantees
  that the following call to produce will succeed.
  No users seem to assume that.
- batching can also in theory reduce the signalling rate:
  users that would previously send interrups to the producer
  to wake it up after consuming each entry would now only
  need to do this once in a batch.
  Doing this would be easy by returning a flag to the caller.
  No users seem to do signalling on consume yet so this was not
  implemented yet.

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---

Jason, I am curious whether the following gives you some of
the performance boost that you see with vhost batching
patches. Is vhost batching on top still helpful?

 include/linux/ptr_ring.h | 63 +++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 54 insertions(+), 9 deletions(-)
diff --git a/include/linux/ptr_ring.h b/include/linux/ptr_ring.h
index 6c70444..6b2e0dd 100644
--- a/include/linux/ptr_ring.h
+++ b/include/linux/ptr_ring.h
@@ -34,11 +34,13 @@
 struct ptr_ring {
 	int producer ____cacheline_aligned_in_smp;
 	spinlock_t producer_lock;
-	int consumer ____cacheline_aligned_in_smp;
+	int consumer_head ____cacheline_aligned_in_smp; /* next valid entry */
+	int consumer_tail; /* next entry to invalidate */
 	spinlock_t consumer_lock;
 	/* Shared consumer/producer data */
 	/* Read-only by both the producer and the consumer */
 	int size ____cacheline_aligned_in_smp; /* max entries in queue */
+	int batch; /* number of entries to consume in a batch */
 	void **queue;
 };
 
@@ -170,7 +172,7 @@ static inline int ptr_ring_produce_bh(struct ptr_ring *r, void *ptr)
 static inline void *__ptr_ring_peek(struct ptr_ring *r)
 {
 	if (likely(r->size))
-		return r->queue[r->consumer];
+		return r->queue[r->consumer_head];
 	return NULL;
 }
 
@@ -231,9 +233,38 @@ static inline bool ptr_ring_empty_bh(struct ptr_ring *r)
 /* Must only be called after __ptr_ring_peek returned !NULL */
 static inline void __ptr_ring_discard_one(struct ptr_ring *r)
 {
-	r->queue[r->consumer++] = NULL;
-	if (unlikely(r->consumer >= r->size))
-		r->consumer = 0;
+	/* Fundamentally, what we want to do is update consumer
+	 * index and zero out the entry so producer can reuse it.
+	 * Doing it naively at each consume would be as simple as:
+	 *       r->queue[r->consumer++] = NULL;
+	 *       if (unlikely(r->consumer >= r->size))
+	 *               r->consumer = 0;
+	 * but that is suboptimal when the ring is full as producer is writing
+	 * out new entries in the same cache line.  Defer these updates until a
+	 * batch of entries has been consumed.
+	 */
+	int head = r->consumer_head++;
+
+	/* Once we have processed enough entries invalidate them in
+	 * the ring all at once so producer can reuse their space in the ring.
+	 * We also do this when we reach end of the ring - not mandatory
+	 * but helps keep the implementation simple.
+	 */
+	if (unlikely(r->consumer_head - r->consumer_tail >= r->batch ||
+		     r->consumer_head >= r->size)) {
+		/* Zero out entries in the reverse order: this way we touch the
+		 * cache line that producer might currently be reading the last;
+		 * producer won't make progress and touch other cache lines
+		 * besides the first one until we write out all entries.
+		 */
+		while (likely(head >= r->consumer_tail))
+			r->queue[head--] = NULL;
+		r->consumer_tail = r->consumer_head;
+	}
+	if (unlikely(r->consumer_head >= r->size)) {
+		r->consumer_head = 0;
+		r->consumer_tail = 0;
+	}
 }  
I love this idea.  Reviewed and discussed the idea in-person with MST
during netdevconf[1] at this laptop.  I promised I will also run it
through my micro-benchmarking[2] once I return home (hint ptr_ring gets
used in network stack as skb_array).  
I'm merging this through my tree. Any objections?
I just did the micro-benchmark evaluation I promised and everything
looks good, so no objections from me.

John Fastabend recently posted a RFC patchset for removing the qdisc
lock.  The main purpose is to separate enqueue'ers and dequeue'ers from
serializing on the same lock.  But we need a new queue implementation
that avoids the false-sharing between enq+deq.

This is why John choose to use ptr_ring, changed (pfifo_fast) qdisc to
use this ptr_ring.  I think this patch might help his overload testing,
as my theory is that he is hitting false-sharing on the queue, due to
the queue always being full.

quoted
Reviewed-by: Jesper Dangaard Brouer <redacted>

[1] http://netdevconf.org/2.1/
[2] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/lib/skb_array_bench01.c
If you like can also add my:

Tested-by: Jesper Dangaard Brouer <redacted>

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

$ modprobe skb_array_test01 
$ dmesg 
[73228.381497] skb_array_test01: Loaded
[73228.381498] skb_array_test01: PASSED - basic_init_and_cleanup()
[73228.381498] skb_array_test01: PASSED - basic_add_and_remove_object()
[73228.381505] skb_array_test01: PASSED - test_queue_full_condition()
[73228.381505] skb_array_test01: PASSED - test_queue_empty_condition()
[73228.381510] skb_array_test01: PASSED - test_queue_resize()

Re: [PATCH 1/3] ptr_ring: batch ring zeroing

From: "Michael S. Tsirkin" <mst@redhat.com>
Date: 2017-05-10 12:20:36

On Wed, May 10, 2017 at 11:18:13AM +0200, Jesper Dangaard Brouer wrote:
On Tue, 9 May 2017 16:33:14 +0300
"Michael S. Tsirkin" [off-list ref] wrote:
quoted
On Sat, Apr 08, 2017 at 02:14:08PM +0200, Jesper Dangaard Brouer wrote:
quoted
On Fri, 7 Apr 2017 08:49:57 +0300
"Michael S. Tsirkin" [off-list ref] wrote:
  
quoted
A known weakness in ptr_ring design is that it does not handle well the
situation when ring is almost full: as entries are consumed they are
immediately used again by the producer, so consumer and producer are
writing to a shared cache line.

To fix this, add batching to consume calls: as entries are
consumed do not write NULL into the ring until we get
a multiple (in current implementation 2x) of cache lines
away from the producer. At that point, write them all out.

We do the write out in the reverse order to keep
producer from sharing cache with consumer for as long
as possible.

Writeout also triggers when ring wraps around - there's
no special reason to do this but it helps keep the code
a bit simpler.

What should we do if getting away from producer by 2 cache lines
would mean we are keeping the ring moe than half empty?
Maybe we should reduce the batching in this case,
current patch simply reduces the batching.

Notes:
- it is no longer true that a call to consume guarantees
  that the following call to produce will succeed.
  No users seem to assume that.
- batching can also in theory reduce the signalling rate:
  users that would previously send interrups to the producer
  to wake it up after consuming each entry would now only
  need to do this once in a batch.
  Doing this would be easy by returning a flag to the caller.
  No users seem to do signalling on consume yet so this was not
  implemented yet.

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---

Jason, I am curious whether the following gives you some of
the performance boost that you see with vhost batching
patches. Is vhost batching on top still helpful?

 include/linux/ptr_ring.h | 63 +++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 54 insertions(+), 9 deletions(-)
diff --git a/include/linux/ptr_ring.h b/include/linux/ptr_ring.h
index 6c70444..6b2e0dd 100644
--- a/include/linux/ptr_ring.h
+++ b/include/linux/ptr_ring.h
@@ -34,11 +34,13 @@
 struct ptr_ring {
 	int producer ____cacheline_aligned_in_smp;
 	spinlock_t producer_lock;
-	int consumer ____cacheline_aligned_in_smp;
+	int consumer_head ____cacheline_aligned_in_smp; /* next valid entry */
+	int consumer_tail; /* next entry to invalidate */
 	spinlock_t consumer_lock;
 	/* Shared consumer/producer data */
 	/* Read-only by both the producer and the consumer */
 	int size ____cacheline_aligned_in_smp; /* max entries in queue */
+	int batch; /* number of entries to consume in a batch */
 	void **queue;
 };
 
@@ -170,7 +172,7 @@ static inline int ptr_ring_produce_bh(struct ptr_ring *r, void *ptr)
 static inline void *__ptr_ring_peek(struct ptr_ring *r)
 {
 	if (likely(r->size))
-		return r->queue[r->consumer];
+		return r->queue[r->consumer_head];
 	return NULL;
 }
 
@@ -231,9 +233,38 @@ static inline bool ptr_ring_empty_bh(struct ptr_ring *r)
 /* Must only be called after __ptr_ring_peek returned !NULL */
 static inline void __ptr_ring_discard_one(struct ptr_ring *r)
 {
-	r->queue[r->consumer++] = NULL;
-	if (unlikely(r->consumer >= r->size))
-		r->consumer = 0;
+	/* Fundamentally, what we want to do is update consumer
+	 * index and zero out the entry so producer can reuse it.
+	 * Doing it naively at each consume would be as simple as:
+	 *       r->queue[r->consumer++] = NULL;
+	 *       if (unlikely(r->consumer >= r->size))
+	 *               r->consumer = 0;
+	 * but that is suboptimal when the ring is full as producer is writing
+	 * out new entries in the same cache line.  Defer these updates until a
+	 * batch of entries has been consumed.
+	 */
+	int head = r->consumer_head++;
+
+	/* Once we have processed enough entries invalidate them in
+	 * the ring all at once so producer can reuse their space in the ring.
+	 * We also do this when we reach end of the ring - not mandatory
+	 * but helps keep the implementation simple.
+	 */
+	if (unlikely(r->consumer_head - r->consumer_tail >= r->batch ||
+		     r->consumer_head >= r->size)) {
+		/* Zero out entries in the reverse order: this way we touch the
+		 * cache line that producer might currently be reading the last;
+		 * producer won't make progress and touch other cache lines
+		 * besides the first one until we write out all entries.
+		 */
+		while (likely(head >= r->consumer_tail))
+			r->queue[head--] = NULL;
+		r->consumer_tail = r->consumer_head;
+	}
+	if (unlikely(r->consumer_head >= r->size)) {
+		r->consumer_head = 0;
+		r->consumer_tail = 0;
+	}
 }  
I love this idea.  Reviewed and discussed the idea in-person with MST
during netdevconf[1] at this laptop.  I promised I will also run it
through my micro-benchmarking[2] once I return home (hint ptr_ring gets
used in network stack as skb_array).  
I'm merging this through my tree. Any objections?
I just did the micro-benchmark evaluation I promised and everything
looks good, so no objections from me.

John Fastabend recently posted a RFC patchset for removing the qdisc
lock.  The main purpose is to separate enqueue'ers and dequeue'ers from
serializing on the same lock.  But we need a new queue implementation
that avoids the false-sharing between enq+deq.

This is why John choose to use ptr_ring, changed (pfifo_fast) qdisc to
use this ptr_ring.  I think this patch might help his overload testing,
as my theory is that he is hitting false-sharing on the queue, due to
the queue always being full.

quoted
quoted
Reviewed-by: Jesper Dangaard Brouer <redacted>

[1] http://netdevconf.org/2.1/
[2] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/lib/skb_array_bench01.c
If you like can also add my:

Tested-by: Jesper Dangaard Brouer <redacted>
I pushed it out already unfortunately so I can't attach that.
Sorry.  Thanks a lot for the testing!
-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

$ modprobe skb_array_test01 
$ dmesg 
[73228.381497] skb_array_test01: Loaded
[73228.381498] skb_array_test01: PASSED - basic_init_and_cleanup()
[73228.381498] skb_array_test01: PASSED - basic_add_and_remove_object()
[73228.381505] skb_array_test01: PASSED - test_queue_full_condition()
[73228.381505] skb_array_test01: PASSED - test_queue_empty_condition()
[73228.381510] skb_array_test01: PASSED - test_queue_resize()
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help