Michael S. Tsirkin
2023-Mar-24 05:59 UTC
[External] Re: [PATCH] virtio_ring: Suppress tx interrupt when napi_tx disable
On Fri, Mar 24, 2023 at 11:41:12AM +0800, Jason Wang wrote:> On Thu, Mar 23, 2023 at 4:01?PM ?? <huangjie.albert at bytedance.com> wrote: > > > > Jason Wang <jasowang at redhat.com> ?2023?3?22??? 10:37??? > > > > > > On Tue, Mar 21, 2023 at 5:00?PM Albert Huang > > > <huangjie.albert at bytedance.com> wrote: > > > > > > > > From: "huangjie.albert" <huangjie.albert at bytedance.com> > > > > > > > > fix commit 8d622d21d248 ("virtio: fix up virtio_disable_cb") > > > > > > > > if we disable the napi_tx. when we triger a tx interrupt, the > > > > > > typo should be "trigger" > > > > > > > OK, thanks for this. I will correct it in the next version > > > > > > vq->event_triggered will be set to true. It will no longer be > > > > set to false. Unless we explicitly call virtqueue_enable_cb_delayed > > > > or virtqueue_enable_cb_prepare > > > > > > > > if we disable the napi_tx, It will only be called when the tx ring > > > > buffer is relatively small: > > > > virtio_net->start_xmit: > > > > if (sq->vq->num_free < 2+MAX_SKB_FRAGS) { > > > > netif_stop_subqueue(dev, qnum); > > > > if (!use_napi && > > > > unlikely(!virtqueue_enable_cb_delayed(sq->vq))) { > > > > /* More just got used, free them then recheck. */ > > > > free_old_xmit_skbs(sq, false); > > > > if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) { > > > > netif_start_subqueue(dev, qnum); > > > > virtqueue_disable_cb(sq->vq); > > > > } > > > > > > The code example here is out of date, make sure your tree has this: > > > > also, I will correct it in the next version?this is from kernel 5.15. > > > > > > > > commit d71ebe8114b4bf622804b810f5e274069060a174 > > > Author: Jason Wang <jasowang at redhat.com> > > > Date: Tue Jan 17 11:47:07 2023 +0800 > > > > > > virtio-net: correctly enable callback during start_xmit > > > > > > > } > > > > } > > > > Because event_triggered is true.Therefore, VRING_AVAIL_F_NO_INTERRUPT or > > > > VRING_PACKED_EVENT_FLAG_DISABLE will not be set.So we update > > > > vring_used_event(&vq->split.vring) or vq->packed.vring.driver->off_wrap > > > > every time we call virtqueue_get_buf_ctx.This will bring more interruptions. > > > > > > Can you please post how to test with the performance numbers? > > > > > > > iperf3 tcp stream: > > vm1 -----------------> vm2 > > vm2 just receive tcp data stream from vm1, and send the ack to vm1, > > there are so > > many tx interruptions in vm2. > > > > but without event_triggered there are just a few tx interruptions. > > > > > > > > > > if event_triggered is set to true, do not update vring_used_event(&vq->split.vring) > > > > or vq->packed.vring.driver->off_wrap > > > > > > > > Signed-off-by: huangjie.albert <huangjie.albert at bytedance.com> > > > > --- > > > > drivers/virtio/virtio_ring.c | 6 ++++-- > > > > 1 file changed, 4 insertions(+), 2 deletions(-) > > > > > > > > diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c > > > > index 307e139cb11d..f486cccadbeb 100644 > > > > --- a/drivers/virtio/virtio_ring.c > > > > +++ b/drivers/virtio/virtio_ring.c > > > > @@ -795,7 +795,8 @@ static void *virtqueue_get_buf_ctx_split(struct virtqueue *_vq, > > > > /* If we expect an interrupt for the next entry, tell host > > > > * by writing event index and flush out the write before > > > > * the read in the next get_buf call. */ > > > > - if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) > > > > + if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) > > > > + && (vq->event_triggered == false)) > > > > > > I'm not sure this can work, when event_triggered is true it means > > > we've got an interrupt, in this case if we want another interrupt for > > > the next entry, we should update used_event otherwise we will lose > > > that interrupt? > > > > > > Thanks > > > > Normally, if we receive an interrupt, we should disable the interrupt > > in the interrupt callback handler. > > So the problem is: > > 1) event_triggered was set to true in vring_interrupt() > > 2) after this nothing will happen for virtqueue_disable_cb() so > VRING_AVAIL_F_NO_INTERRUPT is not set in avail_flags_shadow > 3) virtqueue_get_buf_ctx_split() will still think the cb is enabled > then it tries to publish new eventOh. Good point! I think when I wrote up 8d622d21d248 ("virtio: fix up virtio_disable_cb") I missed this corner case.> This makes me think about whether or not we really need > event_triggered. The assumption in the virtqueue_disable_cb() seems > wrong: > > /* If device triggered an event already it won't trigger one again: > * no need to disable. > */ > if (vq->event_triggered) > return; > > This is wrong if there's no event index support.I don't get it. how does this get triggered? You are talking about device without event index? Here's code from vring_interrupt(): /* Just a hint for performance: so it's ok that this can be racy! */ if (vq->event) vq->event_triggered = true;> And the > event_triggered is somehow duplicated with the > VRING_AVAIL_F_NO_INTERRUPT in the case of event index. The correct fix > might be: > > 1) remove event_triggered > 2) set VRING_AVAIL_F_NO_INTERRUPT in avail_flags_shadow in > vring_interrrupt if event index is supported > > ? > > ThanksI am not sure all this is right and I'd rather we focused performance/correctness and cleanups separately.> > > But because of the introduction of event_triggered, here, > > virtqueue_get_buf_ctx_split cannot be recognized > > that the interrupt has been turned off. > > > > if we want another interrupt for the next entry, We should probably > > call virtqueue_enable_cb? > > > > Thanks > > > > > > > > > virtio_store_mb(vq->weak_barriers, > > > > &vring_used_event(&vq->split.vring), > > > > cpu_to_virtio16(_vq->vdev, vq->last_used_idx)); > > > > @@ -1529,7 +1530,8 @@ static void *virtqueue_get_buf_ctx_packed(struct virtqueue *_vq, > > > > * by writing event index and flush out the write before > > > > * the read in the next get_buf call. > > > > */ > > > > - if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DESC) > > > > + if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DESC > > > > + && (vq->event_triggered == false)) > > > > virtio_store_mb(vq->weak_barriers, > > > > &vq->packed.vring.driver->off_wrap, > > > > cpu_to_le16(vq->last_used_idx)); > > > > -- > > > > 2.31.1 > > > > > > > > >
Jason Wang
2023-Mar-24 06:32 UTC
[External] Re: [PATCH] virtio_ring: Suppress tx interrupt when napi_tx disable
On Fri, Mar 24, 2023 at 1:59?PM Michael S. Tsirkin <mst at redhat.com> wrote:> > On Fri, Mar 24, 2023 at 11:41:12AM +0800, Jason Wang wrote: > > On Thu, Mar 23, 2023 at 4:01?PM ?? <huangjie.albert at bytedance.com> wrote: > > > > > > Jason Wang <jasowang at redhat.com> ?2023?3?22??? 10:37??? > > > > > > > > On Tue, Mar 21, 2023 at 5:00?PM Albert Huang > > > > <huangjie.albert at bytedance.com> wrote: > > > > > > > > > > From: "huangjie.albert" <huangjie.albert at bytedance.com> > > > > > > > > > > fix commit 8d622d21d248 ("virtio: fix up virtio_disable_cb") > > > > > > > > > > if we disable the napi_tx. when we triger a tx interrupt, the > > > > > > > > typo should be "trigger" > > > > > > > > > > OK, thanks for this. I will correct it in the next version > > > > > > > > vq->event_triggered will be set to true. It will no longer be > > > > > set to false. Unless we explicitly call virtqueue_enable_cb_delayed > > > > > or virtqueue_enable_cb_prepare > > > > > > > > > > if we disable the napi_tx, It will only be called when the tx ring > > > > > buffer is relatively small: > > > > > virtio_net->start_xmit: > > > > > if (sq->vq->num_free < 2+MAX_SKB_FRAGS) { > > > > > netif_stop_subqueue(dev, qnum); > > > > > if (!use_napi && > > > > > unlikely(!virtqueue_enable_cb_delayed(sq->vq))) { > > > > > /* More just got used, free them then recheck. */ > > > > > free_old_xmit_skbs(sq, false); > > > > > if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) { > > > > > netif_start_subqueue(dev, qnum); > > > > > virtqueue_disable_cb(sq->vq); > > > > > } > > > > > > > > The code example here is out of date, make sure your tree has this: > > > > > > also, I will correct it in the next version?this is from kernel 5.15. > > > > > > > > > > > commit d71ebe8114b4bf622804b810f5e274069060a174 > > > > Author: Jason Wang <jasowang at redhat.com> > > > > Date: Tue Jan 17 11:47:07 2023 +0800 > > > > > > > > virtio-net: correctly enable callback during start_xmit > > > > > > > > > } > > > > > } > > > > > Because event_triggered is true.Therefore, VRING_AVAIL_F_NO_INTERRUPT or > > > > > VRING_PACKED_EVENT_FLAG_DISABLE will not be set.So we update > > > > > vring_used_event(&vq->split.vring) or vq->packed.vring.driver->off_wrap > > > > > every time we call virtqueue_get_buf_ctx.This will bring more interruptions. > > > > > > > > Can you please post how to test with the performance numbers? > > > > > > > > > > iperf3 tcp stream: > > > vm1 -----------------> vm2 > > > vm2 just receive tcp data stream from vm1, and send the ack to vm1, > > > there are so > > > many tx interruptions in vm2. > > > > > > but without event_triggered there are just a few tx interruptions. > > > > > > > > > > > > > if event_triggered is set to true, do not update vring_used_event(&vq->split.vring) > > > > > or vq->packed.vring.driver->off_wrap > > > > > > > > > > Signed-off-by: huangjie.albert <huangjie.albert at bytedance.com> > > > > > --- > > > > > drivers/virtio/virtio_ring.c | 6 ++++-- > > > > > 1 file changed, 4 insertions(+), 2 deletions(-) > > > > > > > > > > diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c > > > > > index 307e139cb11d..f486cccadbeb 100644 > > > > > --- a/drivers/virtio/virtio_ring.c > > > > > +++ b/drivers/virtio/virtio_ring.c > > > > > @@ -795,7 +795,8 @@ static void *virtqueue_get_buf_ctx_split(struct virtqueue *_vq, > > > > > /* If we expect an interrupt for the next entry, tell host > > > > > * by writing event index and flush out the write before > > > > > * the read in the next get_buf call. */ > > > > > - if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) > > > > > + if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) > > > > > + && (vq->event_triggered == false)) > > > > > > > > I'm not sure this can work, when event_triggered is true it means > > > > we've got an interrupt, in this case if we want another interrupt for > > > > the next entry, we should update used_event otherwise we will lose > > > > that interrupt? > > > > > > > > Thanks > > > > > > Normally, if we receive an interrupt, we should disable the interrupt > > > in the interrupt callback handler. > > > > So the problem is: > > > > 1) event_triggered was set to true in vring_interrupt() > > > > 2) after this nothing will happen for virtqueue_disable_cb() so > > VRING_AVAIL_F_NO_INTERRUPT is not set in avail_flags_shadow > > 3) virtqueue_get_buf_ctx_split() will still think the cb is enabled > > then it tries to publish new event > > Oh. Good point! I think when I wrote up > 8d622d21d248 ("virtio: fix up virtio_disable_cb") > I missed this corner case. > > > > > This makes me think about whether or not we really need > > event_triggered. The assumption in the virtqueue_disable_cb() seems > > wrong: > > > > /* If device triggered an event already it won't trigger one again: > > * no need to disable. > > */ > > if (vq->event_triggered) > > return; > > > > This is wrong if there's no event index support. > > > I don't get it. how does this get triggered? > > You are talking about device without event index? > Here's code from vring_interrupt(): > > /* Just a hint for performance: so it's ok that this can be racy! */ > if (vq->event) > vq->event_triggered = true;But we have the following in virtqueue_disable_cb(): /* If device triggered an event already it won't trigger one again: * no need to disable. */ if (vq->event_triggered) return; if (vq->packed_ring) virtqueue_disable_cb_packed(_vq); else virtqueue_disable_cb_split(_vq); This means, without an event index, we don't set avail flags. So the interrupt is not disabled actually in this case. Thanks> > > > > > And the > > event_triggered is somehow duplicated with the > > VRING_AVAIL_F_NO_INTERRUPT in the case of event index. The correct fix > > might be: > > > > 1) remove event_triggered > > 2) set VRING_AVAIL_F_NO_INTERRUPT in avail_flags_shadow in > > vring_interrrupt if event index is supported > > > > ? > > > > Thanks > > I am not sure all this is right and I'd rather we focused > performance/correctness and cleanups separately. > > > > > > > > > But because of the introduction of event_triggered, here, > > > virtqueue_get_buf_ctx_split cannot be recognized > > > that the interrupt has been turned off. > > > > > > if we want another interrupt for the next entry, We should probably > > > call virtqueue_enable_cb? > > > > > > Thanks > > > > > > > > > > > > virtio_store_mb(vq->weak_barriers, > > > > > &vring_used_event(&vq->split.vring), > > > > > cpu_to_virtio16(_vq->vdev, vq->last_used_idx)); > > > > > @@ -1529,7 +1530,8 @@ static void *virtqueue_get_buf_ctx_packed(struct virtqueue *_vq, > > > > > * by writing event index and flush out the write before > > > > > * the read in the next get_buf call. > > > > > */ > > > > > - if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DESC) > > > > > + if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DESC > > > > > + && (vq->event_triggered == false)) > > > > > virtio_store_mb(vq->weak_barriers, > > > > > &vq->packed.vring.driver->off_wrap, > > > > > cpu_to_le16(vq->last_used_idx)); > > > > > -- > > > > > 2.31.1 > > > > > > > > > > > > >