On Thu, Aug 12, 2021 at 8:45 AM Jason Wang <jasow...@redhat.com> wrote:
>
> On Thu, Aug 12, 2021 at 2:43 PM Eugenio Perez Martin
> <epere...@redhat.com> wrote:
> >
> > On Thu, Aug 12, 2021 at 8:16 AM Jason Wang <jasow...@redhat.com> wrote:
> > >
> > > On Thu, Aug 12, 2021 at 12:32 AM Eugenio Pérez <epere...@redhat.com>
> > > wrote:
> > > >
> > > > With the introduction of the batch hinting, meaningless batches can be
> > > > created with no IOTLB updates if the memory region was skipped by
> > > > vhost_vdpa_listener_skipped_section. This is the case of host notifiers
> > > > memory regions, device un/realize, and others. This causes the vdpa
> > > > device to receive dma mapping settings with no changes, a possibly
> > > > expensive operation for nothing.
> > > >
> > > > To avoid that, VHOST_IOTLB_BATCH_BEGIN hint is delayed until we have a
> > > > meaningful (not skipped section) mapping or unmapping operation, and
> > > > VHOST_IOTLB_BATCH_END is not written unless at least one of _UPDATE /
> > > > _INVALIDATE has been issued.
> > > >
> > > > Signed-off-by: Eugenio Pérez <epere...@redhat.com>
> > > > ---
> > > > include/hw/virtio/vhost-vdpa.h | 1 +
> > > > hw/virtio/vhost-vdpa.c | 50 ++++++++++++++++++++++++++--------
> > > > 2 files changed, 39 insertions(+), 12 deletions(-)
> > > >
> > > > diff --git a/include/hw/virtio/vhost-vdpa.h
> > > > b/include/hw/virtio/vhost-vdpa.h
> > > > index e98e327f12..6538572a6f 100644
> > > > --- a/include/hw/virtio/vhost-vdpa.h
> > > > +++ b/include/hw/virtio/vhost-vdpa.h
> > > > @@ -23,6 +23,7 @@ typedef struct vhost_vdpa {
> > > > int device_fd;
> > > > int index;
> > > > uint32_t msg_type;
> > > > + size_t n_mr_updated;
> > > > MemoryListener listener;
> > > > struct vhost_dev *dev;
> > > > VhostVDPAHostNotifier notifier[VIRTIO_QUEUE_MAX];
> > > > diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
> > > > index 6ce94a1f4d..512fa18d68 100644
> > > > --- a/hw/virtio/vhost-vdpa.c
> > > > +++ b/hw/virtio/vhost-vdpa.c
> > > > @@ -89,19 +89,13 @@ static int vhost_vdpa_dma_unmap(struct vhost_vdpa
> > > > *v, hwaddr iova,
> > > > return ret;
> > > > }
> > > >
> > > > -static void vhost_vdpa_listener_begin(MemoryListener *listener)
> > > > +static void vhost_vdpa_listener_begin_batch(struct vhost_vdpa *v)
> > > > {
> > > > - struct vhost_vdpa *v = container_of(listener, struct vhost_vdpa,
> > > > listener);
> > > > - struct vhost_dev *dev = v->dev;
> > > > - struct vhost_msg_v2 msg = {};
> > > > int fd = v->device_fd;
> > > > -
> > > > - if (!(dev->backend_cap & (0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH)))
> > > > {
> > > > - return;
> > > > - }
> > > > -
> > > > - msg.type = v->msg_type;
> > > > - msg.iotlb.type = VHOST_IOTLB_BATCH_BEGIN;
> > > > + struct vhost_msg_v2 msg = {
> > > > + .type = v->msg_type,
> > > > + .iotlb.type = VHOST_IOTLB_BATCH_BEGIN,
> > > > + };
> > > >
> > > > if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) {
> > > > error_report("failed to write, fd=%d, errno=%d (%s)",
> > > > @@ -109,6 +103,25 @@ static void
> > > > vhost_vdpa_listener_begin(MemoryListener *listener)
> > > > }
> > > > }
> > > >
> > > > +static bool vhost_vdpa_iotlb_batch_is_started(const struct vhost_vdpa
> > > > *v)
> > > > +{
> > > > + return v->n_mr_updated != 0;
> > > > +}
> > > > +
> > > > +static void vhost_vdpa_iotlb_batch_begin_once(struct vhost_vdpa *v)
> > > > +{
> > > > + if (!vhost_vdpa_iotlb_batch_is_started(v)) {
> > > > + vhost_vdpa_listener_begin_batch(v);
> > > > + }
> > > > +
> > > > + v->n_mr_updated++;
> > > > +}
> > > > +
> > > > +static void vhost_vdpa_iotlb_batch_reset(struct vhost_vdpa *v)
> > > > +{
> > > > + v->n_mr_updated = 0;
> > > > +}
> > > > +
> > > > static void vhost_vdpa_listener_commit(MemoryListener *listener)
> > > > {
> > > > struct vhost_vdpa *v = container_of(listener, struct vhost_vdpa,
> > > > listener);
> > > > @@ -120,6 +133,10 @@ static void
> > > > vhost_vdpa_listener_commit(MemoryListener *listener)
> > > > return;
> > > > }
> > > >
> > > > + if (vhost_vdpa_iotlb_batch_is_started(v)) {
> > > > + return;
> > > > + }
> > > > +
> > > > msg.type = v->msg_type;
> > > > msg.iotlb.type = VHOST_IOTLB_BATCH_END;
> > > >
> > > > @@ -127,6 +144,8 @@ static void
> > > > vhost_vdpa_listener_commit(MemoryListener *listener)
> > > > error_report("failed to write, fd=%d, errno=%d (%s)",
> > > > fd, errno, strerror(errno));
> > > > }
> > > > +
> > > > + vhost_vdpa_iotlb_batch_reset(v);
> > > > }
> > > >
> > > > static void vhost_vdpa_listener_region_add(MemoryListener *listener,
> > > > @@ -170,6 +189,10 @@ static void
> > > > vhost_vdpa_listener_region_add(MemoryListener *listener,
> > > >
> > > > llsize = int128_sub(llend, int128_make64(iova));
> > > >
> > > > + if (v->dev->backend_cap & (0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH))
> > > > {
> > >
> > > Let's move this in to vhost_vdpa_iotlb_batch_begin_once()?
> > >
> >
> > Sure
> >
> > > > + vhost_vdpa_iotlb_batch_begin_once(v);
> > > > + }
> > > > +
> > > > ret = vhost_vdpa_dma_map(v, iova, int128_get64(llsize),
> > > > vaddr, section->readonly);
> > > > if (ret) {
> > > > @@ -221,6 +244,10 @@ static void
> > > > vhost_vdpa_listener_region_del(MemoryListener *listener,
> > > >
> > > > llsize = int128_sub(llend, int128_make64(iova));
> > > >
> > > > + if (v->dev->backend_cap & (0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH))
> > > > {
> > > > + vhost_vdpa_iotlb_batch_begin_once(v);
> > > > + }
> > > > +
> > >
> > > Do we need to check vhost_vdpa_iotlb_batch_is_started() in the .commit?
> > >
> >
> > I don't follow you here. It's that comment in this position of the
> > patch for a reason?
> >
> > That checking is the one that allows qemu to skip the IOTLB_END write.
>
> So it looks to me the BATCH_END will be sent unconditionally in this patch?
>
> If yes, will this become a problem?
>
Ok so actually the check in .commit needs to be the opposite,
!vhost_vdpa_iotlb_batch_is_started. In my fast testing, the host
notifiers memory regions were the ones actually sending the guest's
memory regions.
Once that testing is negated, it shouldn't send any batches that have
been skipped:
* listener_begin it's not called unconditionally anymore,
* IOTLB_BATCH_BEGIN is only written once in region add/del *if*
section is not skipped and other tests like (iova + len > iova).
* IOTLB_BATCH_END is only written if IOTLB_BATCH_BEGIN has been sent,
that should make vhost_vdpa_iotlb_batch_is_started return true.
Sending the next version with the mentioned changes as v3, please let
me know if I've missed something else.
Thanks!
> Thanks
>
> >
> > > Others look good.
> > >
> > > Thanks
> > >
> > > > ret = vhost_vdpa_dma_unmap(v, iova, int128_get64(llsize));
> > > > if (ret) {
> > > > error_report("vhost_vdpa dma unmap error!");
> > > > @@ -234,7 +261,6 @@ static void
> > > > vhost_vdpa_listener_region_del(MemoryListener *listener,
> > > > * depends on the addnop().
> > > > */
> > > > static const MemoryListener vhost_vdpa_memory_listener = {
> > > > - .begin = vhost_vdpa_listener_begin,
> > > > .commit = vhost_vdpa_listener_commit,
> > > > .region_add = vhost_vdpa_listener_region_add,
> > > > .region_del = vhost_vdpa_listener_region_del,
> > > > --
> > > > 2.27.0
> > > >
> > >
> >
>
