[PATCH RFC 0/2] virtio-net: interrupt related improvements

Now that we have brought the virtio overhead way down with a fast packed
ring implementation, we seem to be actually observing TCP drops
indicative of bufferbloat. So let's try to enable TSQ.  Note: it isn't
clear that the default pacing is great for the virt usecase. It's worth
trying to play with sk_pacing_shift_update to see what happens.

For this reason, and for a more important one that I didn't
have time to test it yet, sending as an RFC.

Michael S. Tsirkin (2):
  virtio-net: bql support
  virtio_net: bulk free tx skbs

 drivers/net/virtio_net.c | 29 ++++++++++++++++++++---------
 1 file changed, 20 insertions(+), 9 deletions(-)

-- 
MST

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
 drivers/net/virtio_net.c | 27 +++++++++++++++++++--------
 1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index cecfd77c9f3c..b657bde6b94b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1325,7 +1325,8 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
 	return stats.packets;
 }
 
-static void free_old_xmit_skbs(struct send_queue *sq)
+static void free_old_xmit_skbs(struct send_queue *sq, struct netdev_queue *txq,
+			       bool use_napi)
 {
 	struct sk_buff *skb;
 	unsigned int len;

@@ -1347,6 +1348,9 @@ static void free_old_xmit_skbs(struct send_queue *sq)
 	if (!packets)
 		return;
 
+	if (use_napi)
+		netdev_tx_completed_queue(txq, packets, bytes);
+
 	u64_stats_update_begin(&sq->stats.syncp);
 	sq->stats.bytes += bytes;
 	sq->stats.packets += packets;

@@ -1364,7 +1368,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
 		return;
 
 	if (__netif_tx_trylock(txq)) {
-		free_old_xmit_skbs(sq);
+		free_old_xmit_skbs(sq, txq, true);
 		__netif_tx_unlock(txq);
 	}
 

@@ -1440,7 +1444,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
 
 	__netif_tx_lock(txq, raw_smp_processor_id());
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, true);
 	__netif_tx_unlock(txq);
 
 	virtqueue_napi_complete(napi, sq->vq, 0);

@@ -1505,13 +1509,15 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 	struct send_queue *sq = &vi->sq[qnum];
 	int err;
 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !skb->xmit_more;
+	bool more = skb->xmit_more;
 	bool use_napi = sq->napi.weight;
+	unsigned int bytes = skb->len;
+	bool kick;
 
 	/* Free up any pending old buffers before queueing new ones. */
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, use_napi);
 
-	if (use_napi && kick)
+	if (use_napi && !more)
 		virtqueue_enable_cb_delayed(sq->vq);
 
 	/* timestamp packet in software */

@@ -1552,7 +1558,7 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 		if (!use_napi &&
 		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 			/* More just got used, free them then recheck. */
-			free_old_xmit_skbs(sq);
+			free_old_xmit_skbs(sq, txq, false);
 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 				netif_start_subqueue(dev, qnum);
 				virtqueue_disable_cb(sq->vq);

@@ -1560,7 +1566,12 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 		}
 	}
 
-	if (kick || netif_xmit_stopped(txq)) {
+	if (use_napi)
+		kick = __netdev_tx_sent_queue(txq, bytes, more);
+	else
+		kick = !more || netif_xmit_stopped(txq);
+
+	if (kick) {
 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
 			u64_stats_update_begin(&sq->stats.syncp);
 			sq->stats.kicks++;

-- 
MST

Use napi_consume_skb() to get bulk free.  Note that napi_consume_skb is
safe to call in a non-napi context as long as the napi_budget flag is
correct.

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
 drivers/net/virtio_net.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index b657bde6b94b..18c06322be40 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1339,7 +1339,7 @@ static void free_old_xmit_skbs(struct send_queue *sq, struct netdev_queue *txq,
 		bytes += skb->len;
 		packets++;
 
-		dev_consume_skb_any(skb);
+		napi_consume_skb(skb, use_napi);
 	}
 
 	/* Avoid overhead when no packets have been processed

-- 
MST

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the 
mode switching between napi and no napi. We should make sure when the 
packet is sent and trakced by BQL,  it should be consumed by BQL as 
well. I did it by tracking it through skb->cb.  And deal with the freeze 
by reset the BQL status. Patch attached.

But when testing with vhost-net, I don't very a stable performance, it 
was probably because we batch the used ring updating so tx interrupt may 
come randomly. We probably need to implement time bounded coalescing 
mechanism which could be configured from userspace.

Btw, maybe it's time just enable napi TX by default. I get ~10% TCP_RR 
regression on machine without APICv, (haven't found time to test APICv 
machine). But consider it was for correctness, I think it's acceptable? 
Then we can do optimization on top?


Thanks

On 2018/12/6 下午4:17, Jason Wang wrote:

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs. Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is 
the mode switching between napi and no napi. We should make sure when 
the packet is sent and trakced by BQL,  it should be consumed by BQL 
as well. I did it by tracking it through skb->cb.  And deal with the 
freeze by reset the BQL status. Patch attached.

Add the patch.

Thanks

But when testing with vhost-net, I don't very a stable performance, it 
was probably because we batch the used ring updating so tx interrupt 
may come randomly. We probably need to implement time bounded 
coalescing mechanism which could be configured from userspace.

Btw, maybe it's time just enable napi TX by default. I get ~10% TCP_RR 
regression on machine without APICv, (haven't found time to test APICv 
machine). But consider it was for correctness, I think it's 
acceptable? Then we can do optimization on top?


Thanks

quoted

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
  drivers/net/virtio_net.c | 27 +++++++++++++++++++--------
  1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index cecfd77c9f3c..b657bde6b94b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1325,7 +1325,8 @@ static int virtnet_receive(struct receive_queue 

*rq, int budget,
      return stats.packets;
  }
  -static void free_old_xmit_skbs(struct send_queue *sq)
+static void free_old_xmit_skbs(struct send_queue *sq, struct 
netdev_queue *txq,
+                   bool use_napi)
  {
      struct sk_buff *skb;
      unsigned int len;

@@ -1347,6 +1348,9 @@ static void free_old_xmit_skbs(struct 

send_queue *sq)
      if (!packets)
          return;
  +    if (use_napi)
+        netdev_tx_completed_queue(txq, packets, bytes);
+
      u64_stats_update_begin(&sq->stats.syncp);
      sq->stats.bytes += bytes;
      sq->stats.packets += packets;

@@ -1364,7 +1368,7 @@ static void virtnet_poll_cleantx(struct 

receive_queue *rq)
          return;
        if (__netif_tx_trylock(txq)) {
-        free_old_xmit_skbs(sq);
+        free_old_xmit_skbs(sq, txq, true);
          __netif_tx_unlock(txq);
      }
  @@ -1440,7 +1444,7 @@ static int virtnet_poll_tx(struct napi_struct 
*napi, int budget)
      struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, 
vq2txq(sq->vq));
        __netif_tx_lock(txq, raw_smp_processor_id());
-    free_old_xmit_skbs(sq);
+    free_old_xmit_skbs(sq, txq, true);
      __netif_tx_unlock(txq);
        virtqueue_napi_complete(napi, sq->vq, 0);

@@ -1505,13 +1509,15 @@ static netdev_tx_t start_xmit(struct sk_buff 

*skb, struct net_device *dev)
      struct send_queue *sq = &vi->sq[qnum];
      int err;
      struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-    bool kick = !skb->xmit_more;
+    bool more = skb->xmit_more;
      bool use_napi = sq->napi.weight;
+    unsigned int bytes = skb->len;
+    bool kick;
        /* Free up any pending old buffers before queueing new ones. */
-    free_old_xmit_skbs(sq);
+    free_old_xmit_skbs(sq, txq, use_napi);
  -    if (use_napi && kick)
+    if (use_napi && !more)
          virtqueue_enable_cb_delayed(sq->vq);
        /* timestamp packet in software */

@@ -1552,7 +1558,7 @@ static netdev_tx_t start_xmit(struct sk_buff 

*skb, struct net_device *dev)
          if (!use_napi &&
              unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
              /* More just got used, free them then recheck. */
-            free_old_xmit_skbs(sq);
+            free_old_xmit_skbs(sq, txq, false);
              if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
                  netif_start_subqueue(dev, qnum);
                  virtqueue_disable_cb(sq->vq);

@@ -1560,7 +1566,12 @@ static netdev_tx_t start_xmit(struct sk_buff 

*skb, struct net_device *dev)
          }
      }
  -    if (kick || netif_xmit_stopped(txq)) {
+    if (use_napi)
+        kick = __netdev_tx_sent_queue(txq, bytes, more);
+    else
+        kick = !more || netif_xmit_stopped(txq);
+
+    if (kick) {
          if (virtqueue_kick_prepare(sq->vq) && 
virtqueue_notify(sq->vq)) {
              u64_stats_update_begin(&sq->stats.syncp);
              sq->stats.kicks++;

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well. I did it by
tracking it through skb->cb.  And deal with the freeze by reset the BQL
status. Patch attached.

But when testing with vhost-net, I don't very a stable performance, it was
probably because we batch the used ring updating so tx interrupt may come
randomly. We probably need to implement time bounded coalescing mechanism
which could be configured from userspace.

Btw, maybe it's time just enable napi TX by default. I get ~10% TCP_RR
regression on machine without APICv, (haven't found time to test APICv
machine). But consider it was for correctness, I think it's acceptable? Then
we can do optimization on top?


Thanks

I don't see how it's for correctness to be frank.
What if we just do the bulk free? Does that fix the regression?

quoted

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
  drivers/net/virtio_net.c | 27 +++++++++++++++++++--------
  1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index cecfd77c9f3c..b657bde6b94b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1325,7 +1325,8 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
  	return stats.packets;
  }
-static void free_old_xmit_skbs(struct send_queue *sq)
+static void free_old_xmit_skbs(struct send_queue *sq, struct netdev_queue *txq,
+			       bool use_napi)
  {
  	struct sk_buff *skb;
  	unsigned int len;

@@ -1347,6 +1348,9 @@ static void free_old_xmit_skbs(struct send_queue *sq)
  	if (!packets)
  		return;
+	if (use_napi)
+		netdev_tx_completed_queue(txq, packets, bytes);
+
  	u64_stats_update_begin(&sq->stats.syncp);
  	sq->stats.bytes += bytes;
  	sq->stats.packets += packets;

@@ -1364,7 +1368,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
  		return;
  	if (__netif_tx_trylock(txq)) {
-		free_old_xmit_skbs(sq);
+		free_old_xmit_skbs(sq, txq, true);
  		__netif_tx_unlock(txq);
  	}

@@ -1440,7 +1444,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
  	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
  	__netif_tx_lock(txq, raw_smp_processor_id());
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, true);
  	__netif_tx_unlock(txq);
  	virtqueue_napi_complete(napi, sq->vq, 0);

@@ -1505,13 +1509,15 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
  	struct send_queue *sq = &vi->sq[qnum];
  	int err;
  	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !skb->xmit_more;
+	bool more = skb->xmit_more;
  	bool use_napi = sq->napi.weight;
+	unsigned int bytes = skb->len;
+	bool kick;
  	/* Free up any pending old buffers before queueing new ones. */
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, use_napi);
-	if (use_napi && kick)
+	if (use_napi && !more)
  		virtqueue_enable_cb_delayed(sq->vq);
  	/* timestamp packet in software */

@@ -1552,7 +1558,7 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
  		if (!use_napi &&
  		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
  			/* More just got used, free them then recheck. */
-			free_old_xmit_skbs(sq);
+			free_old_xmit_skbs(sq, txq, false);
  			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
  				netif_start_subqueue(dev, qnum);
  				virtqueue_disable_cb(sq->vq);

@@ -1560,7 +1566,12 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
  		}
  	}
-	if (kick || netif_xmit_stopped(txq)) {
+	if (use_napi)
+		kick = __netdev_tx_sent_queue(txq, bytes, more);
+	else
+		kick = !more || netif_xmit_stopped(txq);
+
+	if (kick) {
  		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
  			u64_stats_update_begin(&sq->stats.syncp);
  			sq->stats.kicks++;

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well. I did it by
tracking it through skb->cb.  And deal with the freeze by reset the BQL
status. Patch attached.

But when testing with vhost-net, I don't very a stable performance,

So how about increasing TSQ pacing shift then?

it was
probably because we batch the used ring updating so tx interrupt may come
randomly. We probably need to implement time bounded coalescing mechanism
which could be configured from userspace.

I don't think it's reasonable to expect userspace to be that smart ...
Why do we need time bounded? used ring is always updated when ring
becomes empty.

Btw, maybe it's time just enable napi TX by default. I get ~10% TCP_RR
regression on machine without APICv, (haven't found time to test APICv
machine). But consider it was for correctness, I think it's acceptable? Then
we can do optimization on top?


Thanks

quoted

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
  drivers/net/virtio_net.c | 27 +++++++++++++++++++--------
  1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index cecfd77c9f3c..b657bde6b94b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1325,7 +1325,8 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
  	return stats.packets;
  }
-static void free_old_xmit_skbs(struct send_queue *sq)
+static void free_old_xmit_skbs(struct send_queue *sq, struct netdev_queue *txq,
+			       bool use_napi)
  {
  	struct sk_buff *skb;
  	unsigned int len;

@@ -1347,6 +1348,9 @@ static void free_old_xmit_skbs(struct send_queue *sq)
  	if (!packets)
  		return;
+	if (use_napi)
+		netdev_tx_completed_queue(txq, packets, bytes);
+
  	u64_stats_update_begin(&sq->stats.syncp);
  	sq->stats.bytes += bytes;
  	sq->stats.packets += packets;

@@ -1364,7 +1368,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
  		return;
  	if (__netif_tx_trylock(txq)) {
-		free_old_xmit_skbs(sq);
+		free_old_xmit_skbs(sq, txq, true);
  		__netif_tx_unlock(txq);
  	}

@@ -1440,7 +1444,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
  	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
  	__netif_tx_lock(txq, raw_smp_processor_id());
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, true);
  	__netif_tx_unlock(txq);
  	virtqueue_napi_complete(napi, sq->vq, 0);

@@ -1505,13 +1509,15 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
  	struct send_queue *sq = &vi->sq[qnum];
  	int err;
  	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !skb->xmit_more;
+	bool more = skb->xmit_more;
  	bool use_napi = sq->napi.weight;
+	unsigned int bytes = skb->len;
+	bool kick;
  	/* Free up any pending old buffers before queueing new ones. */
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, use_napi);
-	if (use_napi && kick)
+	if (use_napi && !more)
  		virtqueue_enable_cb_delayed(sq->vq);
  	/* timestamp packet in software */

@@ -1552,7 +1558,7 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
  		if (!use_napi &&
  		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
  			/* More just got used, free them then recheck. */
-			free_old_xmit_skbs(sq);
+			free_old_xmit_skbs(sq, txq, false);
  			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
  				netif_start_subqueue(dev, qnum);
  				virtqueue_disable_cb(sq->vq);

@@ -1560,7 +1566,12 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
  		}
  	}
-	if (kick || netif_xmit_stopped(txq)) {
+	if (use_napi)
+		kick = __netdev_tx_sent_queue(txq, bytes, more);
+	else
+		kick = !more || netif_xmit_stopped(txq);
+
+	if (kick) {
  		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
  			u64_stats_update_begin(&sq->stats.syncp);
  			sq->stats.kicks++;

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well.

I just went over the patch again and I don't understand this comment.
This patch only enabled BQL with tx napi.

Thus there's no mode switching.

What did I miss?

I did it by
tracking it through skb->cb.  And deal with the freeze by reset the BQL
status. Patch attached.

But when testing with vhost-net, I don't very a stable performance, it was
probably because we batch the used ring updating so tx interrupt may come
randomly. We probably need to implement time bounded coalescing mechanism
which could be configured from userspace.

Btw, maybe it's time just enable napi TX by default. I get ~10% TCP_RR
regression on machine without APICv, (haven't found time to test APICv
machine). But consider it was for correctness, I think it's acceptable? Then
we can do optimization on top?


Thanks

quoted

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
  drivers/net/virtio_net.c | 27 +++++++++++++++++++--------
  1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index cecfd77c9f3c..b657bde6b94b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1325,7 +1325,8 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
  	return stats.packets;
  }
-static void free_old_xmit_skbs(struct send_queue *sq)
+static void free_old_xmit_skbs(struct send_queue *sq, struct netdev_queue *txq,
+			       bool use_napi)
  {
  	struct sk_buff *skb;
  	unsigned int len;

@@ -1347,6 +1348,9 @@ static void free_old_xmit_skbs(struct send_queue *sq)
  	if (!packets)
  		return;
+	if (use_napi)
+		netdev_tx_completed_queue(txq, packets, bytes);
+
  	u64_stats_update_begin(&sq->stats.syncp);
  	sq->stats.bytes += bytes;
  	sq->stats.packets += packets;

@@ -1364,7 +1368,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
  		return;
  	if (__netif_tx_trylock(txq)) {
-		free_old_xmit_skbs(sq);
+		free_old_xmit_skbs(sq, txq, true);
  		__netif_tx_unlock(txq);
  	}

@@ -1440,7 +1444,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
  	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
  	__netif_tx_lock(txq, raw_smp_processor_id());
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, true);
  	__netif_tx_unlock(txq);
  	virtqueue_napi_complete(napi, sq->vq, 0);

@@ -1505,13 +1509,15 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
  	struct send_queue *sq = &vi->sq[qnum];
  	int err;
  	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !skb->xmit_more;
+	bool more = skb->xmit_more;
  	bool use_napi = sq->napi.weight;
+	unsigned int bytes = skb->len;
+	bool kick;
  	/* Free up any pending old buffers before queueing new ones. */
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, use_napi);
-	if (use_napi && kick)
+	if (use_napi && !more)
  		virtqueue_enable_cb_delayed(sq->vq);
  	/* timestamp packet in software */

@@ -1552,7 +1558,7 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
  		if (!use_napi &&
  		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
  			/* More just got used, free them then recheck. */
-			free_old_xmit_skbs(sq);
+			free_old_xmit_skbs(sq, txq, false);
  			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
  				netif_start_subqueue(dev, qnum);
  				virtqueue_disable_cb(sq->vq);

@@ -1560,7 +1566,12 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
  		}
  	}
-	if (kick || netif_xmit_stopped(txq)) {
+	if (use_napi)
+		kick = __netdev_tx_sent_queue(txq, bytes, more);
+	else
+		kick = !more || netif_xmit_stopped(txq);
+
+	if (kick) {
  		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
  			u64_stats_update_begin(&sq->stats.syncp);
  			sq->stats.kicks++;

On 2018/12/26 下午11:15, Michael S. Tsirkin wrote:

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

quoted

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well. I did it by
tracking it through skb->cb.  And deal with the freeze by reset the BQL
status. Patch attached.

But when testing with vhost-net, I don't very a stable performance, it was
probably because we batch the used ring updating so tx interrupt may come
randomly. We probably need to implement time bounded coalescing mechanism
which could be configured from userspace.

Btw, maybe it's time just enable napi TX by default. I get ~10% TCP_RR
regression on machine without APICv, (haven't found time to test APICv
machine). But consider it was for correctness, I think it's acceptable? Then
we can do optimization on top?


Thanks

I don't see how it's for correctness to be frank.

Socket accounting is wrong in the case. This should be a bug in fact.

What if we just do the bulk free? Does that fix the regression?

I can test it.

quoted

quoted

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
   drivers/net/virtio_net.c | 27 +++++++++++++++++++--------
   1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index cecfd77c9f3c..b657bde6b94b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1325,7 +1325,8 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
   	return stats.packets;
   }
-static void free_old_xmit_skbs(struct send_queue *sq)
+static void free_old_xmit_skbs(struct send_queue *sq, struct netdev_queue *txq,
+			       bool use_napi)
   {
   	struct sk_buff *skb;
   	unsigned int len;

@@ -1347,6 +1348,9 @@ static void free_old_xmit_skbs(struct send_queue *sq)
   	if (!packets)
   		return;
+	if (use_napi)
+		netdev_tx_completed_queue(txq, packets, bytes);
+
   	u64_stats_update_begin(&sq->stats.syncp);
   	sq->stats.bytes += bytes;
   	sq->stats.packets += packets;

@@ -1364,7 +1368,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
   		return;
   	if (__netif_tx_trylock(txq)) {
-		free_old_xmit_skbs(sq);
+		free_old_xmit_skbs(sq, txq, true);
   		__netif_tx_unlock(txq);
   	}

@@ -1440,7 +1444,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
   	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
   	__netif_tx_lock(txq, raw_smp_processor_id());
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, true);
   	__netif_tx_unlock(txq);
   	virtqueue_napi_complete(napi, sq->vq, 0);

@@ -1505,13 +1509,15 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   	struct send_queue *sq = &vi->sq[qnum];
   	int err;
   	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !skb->xmit_more;
+	bool more = skb->xmit_more;
   	bool use_napi = sq->napi.weight;
+	unsigned int bytes = skb->len;
+	bool kick;
   	/* Free up any pending old buffers before queueing new ones. */
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, use_napi);
-	if (use_napi && kick)
+	if (use_napi && !more)
   		virtqueue_enable_cb_delayed(sq->vq);
   	/* timestamp packet in software */

@@ -1552,7 +1558,7 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   		if (!use_napi &&
   		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
   			/* More just got used, free them then recheck. */
-			free_old_xmit_skbs(sq);
+			free_old_xmit_skbs(sq, txq, false);
   			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
   				netif_start_subqueue(dev, qnum);
   				virtqueue_disable_cb(sq->vq);

@@ -1560,7 +1566,12 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   		}
   	}
-	if (kick || netif_xmit_stopped(txq)) {
+	if (use_napi)
+		kick = __netdev_tx_sent_queue(txq, bytes, more);
+	else
+		kick = !more || netif_xmit_stopped(txq);
+
+	if (kick) {
   		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
   			u64_stats_update_begin(&sq->stats.syncp);
   			sq->stats.kicks++;

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

On 2018/12/26 下午11:19, Michael S. Tsirkin wrote:

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

quoted

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well. I did it by
tracking it through skb->cb.  And deal with the freeze by reset the BQL
status. Patch attached.

But when testing with vhost-net, I don't very a stable performance,

So how about increasing TSQ pacing shift then?

I can test this. But changing default TCP value is much more than a 
virtio-net specific thing.

quoted

it was
probably because we batch the used ring updating so tx interrupt may come
randomly. We probably need to implement time bounded coalescing mechanism
which could be configured from userspace.

I don't think it's reasonable to expect userspace to be that smart ...
Why do we need time bounded? used ring is always updated when ring
becomes empty.

We don't add used when means BQL may not see the consumed packet in 
time. And the delay varies based on the workload since we count packets 
not bytes or time before doing the batched updating.

Thanks

quoted

Btw, maybe it's time just enable napi TX by default. I get ~10% TCP_RR
regression on machine without APICv, (haven't found time to test APICv
machine). But consider it was for correctness, I think it's acceptable? Then
we can do optimization on top?


Thanks

quoted

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
   drivers/net/virtio_net.c | 27 +++++++++++++++++++--------
   1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index cecfd77c9f3c..b657bde6b94b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1325,7 +1325,8 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
   	return stats.packets;
   }
-static void free_old_xmit_skbs(struct send_queue *sq)
+static void free_old_xmit_skbs(struct send_queue *sq, struct netdev_queue *txq,
+			       bool use_napi)
   {
   	struct sk_buff *skb;
   	unsigned int len;

@@ -1347,6 +1348,9 @@ static void free_old_xmit_skbs(struct send_queue *sq)
   	if (!packets)
   		return;
+	if (use_napi)
+		netdev_tx_completed_queue(txq, packets, bytes);
+
   	u64_stats_update_begin(&sq->stats.syncp);
   	sq->stats.bytes += bytes;
   	sq->stats.packets += packets;

@@ -1364,7 +1368,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
   		return;
   	if (__netif_tx_trylock(txq)) {
-		free_old_xmit_skbs(sq);
+		free_old_xmit_skbs(sq, txq, true);
   		__netif_tx_unlock(txq);
   	}

@@ -1440,7 +1444,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
   	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
   	__netif_tx_lock(txq, raw_smp_processor_id());
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, true);
   	__netif_tx_unlock(txq);
   	virtqueue_napi_complete(napi, sq->vq, 0);

@@ -1505,13 +1509,15 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   	struct send_queue *sq = &vi->sq[qnum];
   	int err;
   	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !skb->xmit_more;
+	bool more = skb->xmit_more;
   	bool use_napi = sq->napi.weight;
+	unsigned int bytes = skb->len;
+	bool kick;
   	/* Free up any pending old buffers before queueing new ones. */
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, use_napi);
-	if (use_napi && kick)
+	if (use_napi && !more)
   		virtqueue_enable_cb_delayed(sq->vq);
   	/* timestamp packet in software */

@@ -1552,7 +1558,7 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   		if (!use_napi &&
   		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
   			/* More just got used, free them then recheck. */
-			free_old_xmit_skbs(sq);
+			free_old_xmit_skbs(sq, txq, false);
   			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
   				netif_start_subqueue(dev, qnum);
   				virtqueue_disable_cb(sq->vq);

@@ -1560,7 +1566,12 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   		}
   	}
-	if (kick || netif_xmit_stopped(txq)) {
+	if (use_napi)
+		kick = __netdev_tx_sent_queue(txq, bytes, more);
+	else
+		kick = !more || netif_xmit_stopped(txq);
+
+	if (kick) {
   		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
   			u64_stats_update_begin(&sq->stats.syncp);
   			sq->stats.kicks++;

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

On 2018/12/26 下午11:22, Michael S. Tsirkin wrote:

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

quoted

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well.

I just went over the patch again and I don't understand this comment.
This patch only enabled BQL with tx napi.

Thus there's no mode switching.

What did I miss?

Consider the case:


TX NAPI is disabled:

send N packets

turn TX NAPI on:

get tx interrupt

BQL try to consume those packets when lead WARN for dql.


Thanks

quoted

I did it by
tracking it through skb->cb.  And deal with the freeze by reset the BQL
status. Patch attached.

But when testing with vhost-net, I don't very a stable performance, it was
probably because we batch the used ring updating so tx interrupt may come
randomly. We probably need to implement time bounded coalescing mechanism
which could be configured from userspace.

Btw, maybe it's time just enable napi TX by default. I get ~10% TCP_RR
regression on machine without APICv, (haven't found time to test APICv
machine). But consider it was for correctness, I think it's acceptable? Then
we can do optimization on top?


Thanks

quoted

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
   drivers/net/virtio_net.c | 27 +++++++++++++++++++--------
   1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index cecfd77c9f3c..b657bde6b94b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1325,7 +1325,8 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
   	return stats.packets;
   }
-static void free_old_xmit_skbs(struct send_queue *sq)
+static void free_old_xmit_skbs(struct send_queue *sq, struct netdev_queue *txq,
+			       bool use_napi)
   {
   	struct sk_buff *skb;
   	unsigned int len;

@@ -1347,6 +1348,9 @@ static void free_old_xmit_skbs(struct send_queue *sq)
   	if (!packets)
   		return;
+	if (use_napi)
+		netdev_tx_completed_queue(txq, packets, bytes);
+
   	u64_stats_update_begin(&sq->stats.syncp);
   	sq->stats.bytes += bytes;
   	sq->stats.packets += packets;

@@ -1364,7 +1368,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
   		return;
   	if (__netif_tx_trylock(txq)) {
-		free_old_xmit_skbs(sq);
+		free_old_xmit_skbs(sq, txq, true);
   		__netif_tx_unlock(txq);
   	}

@@ -1440,7 +1444,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
   	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
   	__netif_tx_lock(txq, raw_smp_processor_id());
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, true);
   	__netif_tx_unlock(txq);
   	virtqueue_napi_complete(napi, sq->vq, 0);

@@ -1505,13 +1509,15 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   	struct send_queue *sq = &vi->sq[qnum];
   	int err;
   	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !skb->xmit_more;
+	bool more = skb->xmit_more;
   	bool use_napi = sq->napi.weight;
+	unsigned int bytes = skb->len;
+	bool kick;
   	/* Free up any pending old buffers before queueing new ones. */
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, use_napi);
-	if (use_napi && kick)
+	if (use_napi && !more)
   		virtqueue_enable_cb_delayed(sq->vq);
   	/* timestamp packet in software */

@@ -1552,7 +1558,7 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   		if (!use_napi &&
   		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
   			/* More just got used, free them then recheck. */
-			free_old_xmit_skbs(sq);
+			free_old_xmit_skbs(sq, txq, false);
   			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
   				netif_start_subqueue(dev, qnum);
   				virtqueue_disable_cb(sq->vq);

@@ -1560,7 +1566,12 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   		}
   	}
-	if (kick || netif_xmit_stopped(txq)) {
+	if (use_napi)
+		kick = __netdev_tx_sent_queue(txq, bytes, more);
+	else
+		kick = !more || netif_xmit_stopped(txq);
+
+	if (kick) {
   		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
   			u64_stats_update_begin(&sq->stats.syncp);
   			sq->stats.kicks++;

On Thu, Dec 27, 2018 at 06:00:36PM +0800, Jason Wang wrote:

On 2018/12/26 下午11:19, Michael S. Tsirkin wrote:

quoted

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

quoted

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well. I did it by
tracking it through skb->cb.  And deal with the freeze by reset the BQL
status. Patch attached.

But when testing with vhost-net, I don't very a stable performance,

So how about increasing TSQ pacing shift then?

I can test this. But changing default TCP value is much more than a
virtio-net specific thing.

Well same logic as wifi applies. Unpredictable latencies related
to radio in one case, to host scheduler in the other.

quoted

quoted

it was
probably because we batch the used ring updating so tx interrupt may come
randomly. We probably need to implement time bounded coalescing mechanism
which could be configured from userspace.

I don't think it's reasonable to expect userspace to be that smart ...
Why do we need time bounded? used ring is always updated when ring
becomes empty.

We don't add used when means BQL may not see the consumed packet in time.
And the delay varies based on the workload since we count packets not bytes
or time before doing the batched updating.

Thanks

Sorry I still don't get it.
When nothing is outstanding then we do update the used.
So if BQL stops userspace from sending packets then
we get an interrupt and packets start flowing again.

It might be suboptimal, we might need to tune it but I doubt running
timers is a solution, timer interrupts cause VM exits.

quoted

quoted

Btw, maybe it's time just enable napi TX by default. I get ~10% TCP_RR
regression on machine without APICv, (haven't found time to test APICv
machine). But consider it was for correctness, I think it's acceptable? Then
we can do optimization on top?


Thanks

quoted

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
   drivers/net/virtio_net.c | 27 +++++++++++++++++++--------
   1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index cecfd77c9f3c..b657bde6b94b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1325,7 +1325,8 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
   	return stats.packets;
   }
-static void free_old_xmit_skbs(struct send_queue *sq)
+static void free_old_xmit_skbs(struct send_queue *sq, struct netdev_queue *txq,
+			       bool use_napi)
   {
   	struct sk_buff *skb;
   	unsigned int len;

@@ -1347,6 +1348,9 @@ static void free_old_xmit_skbs(struct send_queue *sq)
   	if (!packets)
   		return;
+	if (use_napi)
+		netdev_tx_completed_queue(txq, packets, bytes);
+
   	u64_stats_update_begin(&sq->stats.syncp);
   	sq->stats.bytes += bytes;
   	sq->stats.packets += packets;

@@ -1364,7 +1368,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
   		return;
   	if (__netif_tx_trylock(txq)) {
-		free_old_xmit_skbs(sq);
+		free_old_xmit_skbs(sq, txq, true);
   		__netif_tx_unlock(txq);
   	}

@@ -1440,7 +1444,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
   	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
   	__netif_tx_lock(txq, raw_smp_processor_id());
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, true);
   	__netif_tx_unlock(txq);
   	virtqueue_napi_complete(napi, sq->vq, 0);

@@ -1505,13 +1509,15 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   	struct send_queue *sq = &vi->sq[qnum];
   	int err;
   	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !skb->xmit_more;
+	bool more = skb->xmit_more;
   	bool use_napi = sq->napi.weight;
+	unsigned int bytes = skb->len;
+	bool kick;
   	/* Free up any pending old buffers before queueing new ones. */
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, use_napi);
-	if (use_napi && kick)
+	if (use_napi && !more)
   		virtqueue_enable_cb_delayed(sq->vq);
   	/* timestamp packet in software */

@@ -1552,7 +1558,7 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   		if (!use_napi &&
   		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
   			/* More just got used, free them then recheck. */
-			free_old_xmit_skbs(sq);
+			free_old_xmit_skbs(sq, txq, false);
   			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
   				netif_start_subqueue(dev, qnum);
   				virtqueue_disable_cb(sq->vq);

@@ -1560,7 +1566,12 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   		}
   	}
-	if (kick || netif_xmit_stopped(txq)) {
+	if (use_napi)
+		kick = __netdev_tx_sent_queue(txq, bytes, more);
+	else
+		kick = !more || netif_xmit_stopped(txq);
+
+	if (kick) {
   		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
   			u64_stats_update_begin(&sq->stats.syncp);
   			sq->stats.kicks++;

On Thu, Dec 27, 2018 at 06:04:53PM +0800, Jason Wang wrote:

On 2018/12/26 下午11:22, Michael S. Tsirkin wrote:

quoted

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

quoted

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well.

I just went over the patch again and I don't understand this comment.
This patch only enabled BQL with tx napi.

Thus there's no mode switching.

What did I miss?

Consider the case:


TX NAPI is disabled:

send N packets

turn TX NAPI on:

get tx interrupt

BQL try to consume those packets when lead WARN for dql.


Thanks

Can one really switch tx napi on and off? How?
While root can change the napi_tx module parameter, I don't think
that has any effect outside device probe time. What did I miss?

quoted

quoted

I did it by
tracking it through skb->cb.  And deal with the freeze by reset the BQL
status. Patch attached.

But when testing with vhost-net, I don't very a stable performance, it was
probably because we batch the used ring updating so tx interrupt may come
randomly. We probably need to implement time bounded coalescing mechanism
which could be configured from userspace.

Btw, maybe it's time just enable napi TX by default. I get ~10% TCP_RR
regression on machine without APICv, (haven't found time to test APICv
machine). But consider it was for correctness, I think it's acceptable? Then
we can do optimization on top?


Thanks

quoted

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
   drivers/net/virtio_net.c | 27 +++++++++++++++++++--------
   1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index cecfd77c9f3c..b657bde6b94b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1325,7 +1325,8 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
   	return stats.packets;
   }
-static void free_old_xmit_skbs(struct send_queue *sq)
+static void free_old_xmit_skbs(struct send_queue *sq, struct netdev_queue *txq,
+			       bool use_napi)
   {
   	struct sk_buff *skb;
   	unsigned int len;

@@ -1347,6 +1348,9 @@ static void free_old_xmit_skbs(struct send_queue *sq)
   	if (!packets)
   		return;
+	if (use_napi)
+		netdev_tx_completed_queue(txq, packets, bytes);
+
   	u64_stats_update_begin(&sq->stats.syncp);
   	sq->stats.bytes += bytes;
   	sq->stats.packets += packets;

@@ -1364,7 +1368,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
   		return;
   	if (__netif_tx_trylock(txq)) {
-		free_old_xmit_skbs(sq);
+		free_old_xmit_skbs(sq, txq, true);
   		__netif_tx_unlock(txq);
   	}

@@ -1440,7 +1444,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
   	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
   	__netif_tx_lock(txq, raw_smp_processor_id());
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, true);
   	__netif_tx_unlock(txq);
   	virtqueue_napi_complete(napi, sq->vq, 0);

@@ -1505,13 +1509,15 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   	struct send_queue *sq = &vi->sq[qnum];
   	int err;
   	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !skb->xmit_more;
+	bool more = skb->xmit_more;
   	bool use_napi = sq->napi.weight;
+	unsigned int bytes = skb->len;
+	bool kick;
   	/* Free up any pending old buffers before queueing new ones. */
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, use_napi);
-	if (use_napi && kick)
+	if (use_napi && !more)
   		virtqueue_enable_cb_delayed(sq->vq);
   	/* timestamp packet in software */

@@ -1552,7 +1558,7 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   		if (!use_napi &&
   		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
   			/* More just got used, free them then recheck. */
-			free_old_xmit_skbs(sq);
+			free_old_xmit_skbs(sq, txq, false);
   			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
   				netif_start_subqueue(dev, qnum);
   				virtqueue_disable_cb(sq->vq);

@@ -1560,7 +1566,12 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
   		}
   	}
-	if (kick || netif_xmit_stopped(txq)) {
+	if (use_napi)
+		kick = __netdev_tx_sent_queue(txq, bytes, more);
+	else
+		kick = !more || netif_xmit_stopped(txq);
+
+	if (kick) {
   		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
   			u64_stats_update_begin(&sq->stats.syncp);
   			sq->stats.kicks++;

On 2018/12/31 上午2:45, Michael S. Tsirkin wrote:

On Thu, Dec 27, 2018 at 06:00:36PM +0800, Jason Wang wrote:

quoted

On 2018/12/26 下午11:19, Michael S. Tsirkin wrote:

quoted

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

quoted

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well. I did it by
tracking it through skb->cb.  And deal with the freeze by reset the BQL
status. Patch attached.

But when testing with vhost-net, I don't very a stable performance,

So how about increasing TSQ pacing shift then?

I can test this. But changing default TCP value is much more than a
virtio-net specific thing.

Well same logic as wifi applies. Unpredictable latencies related
to radio in one case, to host scheduler in the other.

quoted

quoted

quoted

it was
probably because we batch the used ring updating so tx interrupt may come
randomly. We probably need to implement time bounded coalescing mechanism
which could be configured from userspace.

I don't think it's reasonable to expect userspace to be that smart ...
Why do we need time bounded? used ring is always updated when ring
becomes empty.

We don't add used when means BQL may not see the consumed packet in time.
And the delay varies based on the workload since we count packets not bytes
or time before doing the batched updating.

Thanks

Sorry I still don't get it.
When nothing is outstanding then we do update the used.
So if BQL stops userspace from sending packets then
we get an interrupt and packets start flowing again.

Yes, but how about the cases of multiple flows. That's where I see 
unstable results.

It might be suboptimal, we might need to tune it but I doubt running
timers is a solution, timer interrupts cause VM exits.

Probably not a timer but a time counter (or event byte counter) in vhost 
to add used and signal guest if it exceeds a value instead of waiting 
the number of packets.


Thanks

quoted

quoted

quoted

Btw, maybe it's time just enable napi TX by default. I get ~10% TCP_RR
regression on machine without APICv, (haven't found time to test APICv
machine). But consider it was for correctness, I think it's acceptable? Then
we can do optimization on top?


Thanks

quoted

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
---
    drivers/net/virtio_net.c | 27 +++++++++++++++++++--------
    1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index cecfd77c9f3c..b657bde6b94b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1325,7 +1325,8 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
    	return stats.packets;
    }
-static void free_old_xmit_skbs(struct send_queue *sq)
+static void free_old_xmit_skbs(struct send_queue *sq, struct netdev_queue *txq,
+			       bool use_napi)
    {
    	struct sk_buff *skb;
    	unsigned int len;

@@ -1347,6 +1348,9 @@ static void free_old_xmit_skbs(struct send_queue *sq)
    	if (!packets)
    		return;
+	if (use_napi)
+		netdev_tx_completed_queue(txq, packets, bytes);
+
    	u64_stats_update_begin(&sq->stats.syncp);
    	sq->stats.bytes += bytes;
    	sq->stats.packets += packets;

@@ -1364,7 +1368,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
    		return;
    	if (__netif_tx_trylock(txq)) {
-		free_old_xmit_skbs(sq);
+		free_old_xmit_skbs(sq, txq, true);
    		__netif_tx_unlock(txq);
    	}

@@ -1440,7 +1444,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
    	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
    	__netif_tx_lock(txq, raw_smp_processor_id());
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, true);
    	__netif_tx_unlock(txq);
    	virtqueue_napi_complete(napi, sq->vq, 0);

@@ -1505,13 +1509,15 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
    	struct send_queue *sq = &vi->sq[qnum];
    	int err;
    	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !skb->xmit_more;
+	bool more = skb->xmit_more;
    	bool use_napi = sq->napi.weight;
+	unsigned int bytes = skb->len;
+	bool kick;
    	/* Free up any pending old buffers before queueing new ones. */
-	free_old_xmit_skbs(sq);
+	free_old_xmit_skbs(sq, txq, use_napi);
-	if (use_napi && kick)
+	if (use_napi && !more)
    		virtqueue_enable_cb_delayed(sq->vq);
    	/* timestamp packet in software */

@@ -1552,7 +1558,7 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
    		if (!use_napi &&
    		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
    			/* More just got used, free them then recheck. */
-			free_old_xmit_skbs(sq);
+			free_old_xmit_skbs(sq, txq, false);
    			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
    				netif_start_subqueue(dev, qnum);
    				virtqueue_disable_cb(sq->vq);

@@ -1560,7 +1566,12 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
    		}
    	}
-	if (kick || netif_xmit_stopped(txq)) {
+	if (use_napi)
+		kick = __netdev_tx_sent_queue(txq, bytes, more);
+	else
+		kick = !more || netif_xmit_stopped(txq);
+
+	if (kick) {
    		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
    			u64_stats_update_begin(&sq->stats.syncp);
    			sq->stats.kicks++;

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

On 2018/12/31 上午2:48, Michael S. Tsirkin wrote:

On Thu, Dec 27, 2018 at 06:04:53PM +0800, Jason Wang wrote:

quoted

On 2018/12/26 下午11:22, Michael S. Tsirkin wrote:

quoted

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

quoted

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well.

I just went over the patch again and I don't understand this comment.
This patch only enabled BQL with tx napi.

Thus there's no mode switching.

What did I miss?

Consider the case:


TX NAPI is disabled:

send N packets

turn TX NAPI on:

get tx interrupt

BQL try to consume those packets when lead WARN for dql.


Thanks

Can one really switch tx napi on and off? How?
While root can change the napi_tx module parameter, I don't think
that has any effect outside device probe time. What did I miss?

We support switch the mode through ethtool recently. See

commit 0c465be183c7c57a26446df6ea96d8676b865f92
Author: Jason Wang [off-list ref]
Date:   Tue Oct 9 10:06:26 2018 +0800

     virtio_net: ethtool tx napi configuration

     Implement ethtool .set_coalesce (-C) and .get_coalesce (-c) handlers.
     Interrupt moderation is currently not supported, so these accept and
     display the default settings of 0 usec and 1 frame.

     Toggle tx napi through setting tx-frames. So as to not interfere
     with possible future interrupt moderation, value 1 means tx napi while
     value 0 means not.

     Only allow the switching when device is down for simplicity.

     Link: https://patchwork.ozlabs.org/patch/948149/
     Suggested-by: Jason Wang [off-list ref]
     Signed-off-by: Willem de Bruijn [off-list ref]
     Signed-off-by: Jason Wang [off-list ref]
     Signed-off-by: David S. Miller [off-list ref]

Thanks

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

On Wed, Jan 02, 2019 at 11:30:11AM +0800, Jason Wang wrote:

On 2018/12/31 上午2:48, Michael S. Tsirkin wrote:

quoted

On Thu, Dec 27, 2018 at 06:04:53PM +0800, Jason Wang wrote:

quoted

On 2018/12/26 下午11:22, Michael S. Tsirkin wrote:

quoted

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

quoted

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well.

I just went over the patch again and I don't understand this comment.
This patch only enabled BQL with tx napi.

Thus there's no mode switching.

What did I miss?

Consider the case:


TX NAPI is disabled:

send N packets

turn TX NAPI on:

get tx interrupt

BQL try to consume those packets when lead WARN for dql.


Thanks

Can one really switch tx napi on and off? How?
While root can change the napi_tx module parameter, I don't think
that has any effect outside device probe time. What did I miss?

We support switch the mode through ethtool recently. See

commit 0c465be183c7c57a26446df6ea96d8676b865f92
Author: Jason Wang [off-list ref]
Date:   Tue Oct 9 10:06:26 2018 +0800

    virtio_net: ethtool tx napi configuration

    Implement ethtool .set_coalesce (-C) and .get_coalesce (-c) handlers.
    Interrupt moderation is currently not supported, so these accept and
    display the default settings of 0 usec and 1 frame.

    Toggle tx napi through setting tx-frames. So as to not interfere
    with possible future interrupt moderation, value 1 means tx napi while
    value 0 means not.

    Only allow the switching when device is down for simplicity.

    Link: https://patchwork.ozlabs.org/patch/948149/
    Suggested-by: Jason Wang [off-list ref]
    Signed-off-by: Willem de Bruijn [off-list ref]
    Signed-off-by: Jason Wang [off-list ref]
    Signed-off-by: David S. Miller [off-list ref]

Thanks

It's disabled when device is up - isn't that enough?

-- 
MST
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

On Wed, Jan 02, 2019 at 11:28:43AM +0800, Jason Wang wrote:

On 2018/12/31 上午2:45, Michael S. Tsirkin wrote:

quoted

On Thu, Dec 27, 2018 at 06:00:36PM +0800, Jason Wang wrote:

quoted

On 2018/12/26 下午11:19, Michael S. Tsirkin wrote:

quoted

On Thu, Dec 06, 2018 at 04:17:36PM +0800, Jason Wang wrote:

quoted

On 2018/12/6 上午6:54, Michael S. Tsirkin wrote:

quoted

When use_napi is set, let's enable BQLs.  Note: some of the issues are
similar to wifi.  It's worth considering whether something similar to
commit 36148c2bbfbe ("mac80211: Adjust TSQ pacing shift") might be
benefitial.

I've played a similar patch several days before. The tricky part is the mode
switching between napi and no napi. We should make sure when the packet is
sent and trakced by BQL,  it should be consumed by BQL as well. I did it by
tracking it through skb->cb.  And deal with the freeze by reset the BQL
status. Patch attached.

But when testing with vhost-net, I don't very a stable performance,

So how about increasing TSQ pacing shift then?

I can test this. But changing default TCP value is much more than a
virtio-net specific thing.

Well same logic as wifi applies. Unpredictable latencies related
to radio in one case, to host scheduler in the other.

quoted

quoted

quoted

it was
probably because we batch the used ring updating so tx interrupt may come
randomly. We probably need to implement time bounded coalescing mechanism
which could be configured from userspace.

I don't think it's reasonable to expect userspace to be that smart ...
Why do we need time bounded? used ring is always updated when ring
becomes empty.

We don't add used when means BQL may not see the consumed packet in time.
And the delay varies based on the workload since we count packets not bytes
or time before doing the batched updating.

Thanks

Sorry I still don't get it.
When nothing is outstanding then we do update the used.
So if BQL stops userspace from sending packets then
we get an interrupt and packets start flowing again.

Yes, but how about the cases of multiple flows. That's where I see unstable
results.

quoted

It might be suboptimal, we might need to tune it but I doubt running
timers is a solution, timer interrupts cause VM exits.

Probably not a timer but a time counter (or event byte counter) in vhost to
add used and signal guest if it exceeds a value instead of waiting the
number of packets.


Thanks