[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH v2 12/19] vhost: add vhost_svq_poll
From: |
Eugenio Perez Martin |
Subject: |
Re: [PATCH v2 12/19] vhost: add vhost_svq_poll |
Date: |
Tue, 19 Jul 2022 10:42:04 +0200 |
On Tue, Jul 19, 2022 at 9:38 AM Jason Wang <jasowang@redhat.com> wrote:
>
>
> 在 2022/7/16 01:05, Eugenio Perez Martin 写道:
> > On Fri, Jul 15, 2022 at 10:48 AM Jason Wang <jasowang@redhat.com> wrote:
> >> On Fri, Jul 15, 2022 at 1:39 PM Eugenio Perez Martin
> >> <eperezma@redhat.com> wrote:
> >>> On Fri, Jul 15, 2022 at 5:59 AM Jason Wang <jasowang@redhat.com> wrote:
> >>>> On Fri, Jul 15, 2022 at 12:32 AM Eugenio Pérez <eperezma@redhat.com>
> >>>> wrote:
> >>>>> It allows the Shadow Control VirtQueue to wait for the device to use the
> >>>>> available buffers.
> >>>>>
> >>>>> Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
> >>>>> ---
> >>>>> hw/virtio/vhost-shadow-virtqueue.h | 1 +
> >>>>> hw/virtio/vhost-shadow-virtqueue.c | 22 ++++++++++++++++++++++
> >>>>> 2 files changed, 23 insertions(+)
> >>>>>
> >>>>> diff --git a/hw/virtio/vhost-shadow-virtqueue.h
> >>>>> b/hw/virtio/vhost-shadow-virtqueue.h
> >>>>> index 1692541cbb..b5c6e3b3b4 100644
> >>>>> --- a/hw/virtio/vhost-shadow-virtqueue.h
> >>>>> +++ b/hw/virtio/vhost-shadow-virtqueue.h
> >>>>> @@ -89,6 +89,7 @@ void vhost_svq_push_elem(VhostShadowVirtqueue *svq,
> >>>>> const SVQElement *elem,
> >>>>> int vhost_svq_add(VhostShadowVirtqueue *svq, const struct iovec
> >>>>> *out_sg,
> >>>>> size_t out_num, const struct iovec *in_sg, size_t
> >>>>> in_num,
> >>>>> SVQElement *elem);
> >>>>> +size_t vhost_svq_poll(VhostShadowVirtqueue *svq);
> >>>>>
> >>>>> void vhost_svq_set_svq_kick_fd(VhostShadowVirtqueue *svq, int
> >>>>> svq_kick_fd);
> >>>>> void vhost_svq_set_svq_call_fd(VhostShadowVirtqueue *svq, int
> >>>>> call_fd);
> >>>>> diff --git a/hw/virtio/vhost-shadow-virtqueue.c
> >>>>> b/hw/virtio/vhost-shadow-virtqueue.c
> >>>>> index 5244896358..31a267f721 100644
> >>>>> --- a/hw/virtio/vhost-shadow-virtqueue.c
> >>>>> +++ b/hw/virtio/vhost-shadow-virtqueue.c
> >>>>> @@ -486,6 +486,28 @@ static void vhost_svq_flush(VhostShadowVirtqueue
> >>>>> *svq,
> >>>>> } while (!vhost_svq_enable_notification(svq));
> >>>>> }
> >>>>>
> >>>>> +/**
> >>>>> + * Poll the SVQ for one device used buffer.
> >>>>> + *
> >>>>> + * This function race with main event loop SVQ polling, so extra
> >>>>> + * synchronization is needed.
> >>>>> + *
> >>>>> + * Return the length written by the device.
> >>>>> + */
> >>>>> +size_t vhost_svq_poll(VhostShadowVirtqueue *svq)
> >>>>> +{
> >>>>> + do {
> >>>>> + uint32_t len;
> >>>>> + SVQElement *elem = vhost_svq_get_buf(svq, &len);
> >>>>> + if (elem) {
> >>>>> + return len;
> >>>>> + }
> >>>>> +
> >>>>> + /* Make sure we read new used_idx */
> >>>>> + smp_rmb();
> >>>> There's already one smp_rmb(0 in vhost_svq_get_buf(). So this seems
> >>>> useless?
> >>>>
> >>> That rmb is after checking for new entries with (vq->last_used_idx !=
> >>> svq->shadow_used_idx) , to avoid reordering used_idx read with the
> >>> actual used entry. So my understanding is
> >>> that the compiler is free to skip that check within the while loop.
> >> What do you mean by "that check" here?
> >>
> > The first check of (presumably cached) last_used_idx !=
> > shadow_used_idx. Right before the SVQ's vring fetch of
> > svq->vring.used->idx.
> >
> >>> Maybe the right solution is to add it in vhost_svq_more_used after the
> >>> condition (vq->last_used_idx != svq->shadow_used_idx) is false?
> >> I'm not sure I get the goal of the smp_rmb() here. What barrier does it
> >> pair?
> >>
> > It pairs with the actual device memory write.
> >
> > Note that I'm worried about compiler optimizations or reordering
> > causing an infinite loop, not that the memory is updated properly.
> >
> > Let's say we just returned from vhost_svq_add, and avail_idx -
> > used_idx > 0. The device still did not update SVQ vring used idx, and
> > qemu is very fast so it completes a full call of vhost_svq_get_buf
> > before the device is able to increment the used index. We can trace
> > the full vhost_svq_get_buf without a memory barrier.
> >
> > If the compiler inlines enough and we delete the new smp_rmb barrier,
> > this is what it sees:
> >
> > size_t vhost_svq_poll(VhostShadowVirtqueue *svq)
> > {
> > do {
> > more_used = false
> > // The next conditional returns false
> > if (svq->last_used_idx != svq->shadow_used_idx) {
> > goto useful;
> > }
> >
> > svq->shadow_used_idx = cpu_to_le16(svq->vring.used->idx);
> >
> > // next conditional returns false too
> > if (!(svq->last_used_idx != svq->shadow_used_idx))
> > continue;
> >
> > useful:
> > // actual code to handle new used buffer
> > break;
> > }
> > }
> > }
> >
> > And qemu does not need to read again none of the variables since
> > nothing modifies them within the loop before "useful" tag
> > (svq->vring.used->idx, svq->last_used_idx, svq->shadow_used_idx). So
> > it could freely rewrite as:
> >
> > size_t vhost_svq_poll(VhostShadowVirtqueue *svq) {
> > if (svq->last_used_idx == svq->shadow_used_idx &&
> > svq->last_used_idx == svq->vring.used->idx) {
> > for (;;);
> > }
> > }
> >
> > That's why I think the right place for the mb is right after caller
> > code see (potentially cached) last_used_idx == shadow_used_idx, and it
> > needs to read a value paired with the "device's mb" in the SVQ vring.
>
>
> I think you need "volatile" instead of the memory barriers.
The kernel's READ_ONCE implementation uses a volatile casting but also
a memory read barrier after it. I guess it's because the compiler can
reorder non-volatile accesses around volatile ones. In the proposed
code, that barrier is provided by the vhost_svq_more_used caller
(vhost_svq_get_buf). I think no other caller should need it.
Thanks!
> If I
> understand correctly, you want the load from the memory instead of the
> registers here.
>
> Thanks
>
>
> >
> > We didn't have that problem before, since we clear event_notifier
> > right before the do{}while(), and event loop should hit a memory
> > barrier in the next select / poll / read / whatever syscall to check
> > that event notifier fd is set again.
> >
> >> Since we are in the busy loop, we will read the for new used_idx for
> >> sure,
> > I'm not so sure of that, but maybe I've missed something.
> >
> > I'm sending v3 with this comment pending, so we can iterate faster.
> >
> > Thanks!
> >
> >> and we can't forecast when the used_idx is committed to memory.
> >>
>
- [PATCH v2 09/19] vhost: Track number of descs in SVQElement, (continued)
- [PATCH v2 09/19] vhost: Track number of descs in SVQElement, Eugenio Pérez, 2022/07/14
- [PATCH v2 05/19] vhost: Move vhost_svq_kick call to vhost_svq_add, Eugenio Pérez, 2022/07/14
- [PATCH v2 12/19] vhost: add vhost_svq_poll, Eugenio Pérez, 2022/07/14
- Re: [PATCH v2 12/19] vhost: add vhost_svq_poll, Jason Wang, 2022/07/14
- Re: [PATCH v2 12/19] vhost: add vhost_svq_poll, Eugenio Perez Martin, 2022/07/15
- Re: [PATCH v2 12/19] vhost: add vhost_svq_poll, Jason Wang, 2022/07/15
- Re: [PATCH v2 12/19] vhost: add vhost_svq_poll, Eugenio Perez Martin, 2022/07/15
- Re: [PATCH v2 12/19] vhost: add vhost_svq_poll, Jason Wang, 2022/07/19
- Re: [PATCH v2 12/19] vhost: add vhost_svq_poll,
Eugenio Perez Martin <=
- Re: [PATCH v2 12/19] vhost: add vhost_svq_poll, Jason Wang, 2022/07/19
- Re: [PATCH v2 12/19] vhost: add vhost_svq_poll, Eugenio Perez Martin, 2022/07/19
[PATCH v2 13/19] vhost: Add svq avail_handler callback, Eugenio Pérez, 2022/07/14
[PATCH v2 14/19] vdpa: Export vhost_vdpa_dma_map and unmap calls, Eugenio Pérez, 2022/07/14
[PATCH v2 15/19] vdpa: manual forward CVQ buffers, Eugenio Pérez, 2022/07/14
[PATCH v2 11/19] vhost: Expose vhost_svq_add, Eugenio Pérez, 2022/07/14