Re: [PATCH 2/2] vringh: fetch used_idx from vring at vringh_init_iotlb

From: Eugenio Perez Martin
Date: Tue Jan 31 2023 - 03:00:10 EST


On Tue, Jan 31, 2023 at 4:16 AM Jason Wang <jasowang@xxxxxxxxxx> wrote:
>
> On Tue, Jan 31, 2023 at 12:39 AM Eugenio Perez Martin
> <eperezma@xxxxxxxxxx> wrote:
> >
> > On Sun, Jan 29, 2023 at 7:01 AM Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > >
> > > On Thu, Jan 19, 2023 at 4:11 PM Eugenio Perez Martin
> > > <eperezma@xxxxxxxxxx> wrote:
> > > >
> > > > On Thu, Jan 19, 2023 at 4:20 AM Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Thu, Jan 19, 2023 at 12:44 AM Eugenio Pérez <eperezma@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > Starting from an used_idx different than 0 is needed in use cases like
> > > > > > virtual machine migration. Not doing so and letting the caller set an
> > > > > > avail idx different than 0 causes destination device to try to use old
> > > > > > buffers that source driver already recover and are not available
> > > > > > anymore.
> > > > > >
> > > > > > While callers like vdpa_sim set avail_idx directly it does not set
> > > > > > used_idx. Instead of let the caller do the assignment, fetch it from
> > > > > > the guest at initialization like vhost-kernel do.
> > > > > >
> > > > > > To perform the same at vring_kernel_init and vring_user_init is left for
> > > > > > the future.
> > > > > >
> > > > > > Signed-off-by: Eugenio Pérez <eperezma@xxxxxxxxxx>
> > > > > > ---
> > > > > > drivers/vhost/vringh.c | 25 +++++++++++++++++++++++--
> > > > > > 1 file changed, 23 insertions(+), 2 deletions(-)
> > > > > >
> > > > > > diff --git a/drivers/vhost/vringh.c b/drivers/vhost/vringh.c
> > > > > > index 33eb941fcf15..0eed825197f2 100644
> > > > > > --- a/drivers/vhost/vringh.c
> > > > > > +++ b/drivers/vhost/vringh.c
> > > > > > @@ -1301,6 +1301,17 @@ static inline int putused_iotlb(const struct vringh *vrh,
> > > > > > return 0;
> > > > > > }
> > > > > >
> > > > > > +/**
> > > > > > + * vringh_update_used_idx - fetch used idx from driver's used split vring
> > > > > > + * @vrh: The vring.
> > > > > > + *
> > > > > > + * Returns -errno or 0.
> > > > > > + */
> > > > > > +static inline int vringh_update_used_idx(struct vringh *vrh)
> > > > > > +{
> > > > > > + return getu16_iotlb(vrh, &vrh->last_used_idx, &vrh->vring.used->idx);
> > > > > > +}
> > > > > > +
> > > > > > /**
> > > > > > * vringh_init_iotlb - initialize a vringh for a ring with IOTLB.
> > > > > > * @vrh: the vringh to initialize.
> > > > > > @@ -1319,8 +1330,18 @@ int vringh_init_iotlb(struct vringh *vrh, u64 features,
> > > > > > struct vring_avail *avail,
> > > > > > struct vring_used *used)
> > > > > > {
> > > > >
> > > > > While at this, I wonder if it's better to have a dedicated parameter
> > > > > for last_avail_idx?
> > > > >
> > > >
> > > > I also had that thought. To directly assign last_avail_idx is not a
> > > > specially elegant API IMO.
> > > >
> > > > Maybe expose a way to fetch used_idx from device vring and pass
> > > > used_idx as parameter too?
> > >
> > > If I was not wrong, we can start from last_avail_idx, for used_idx it
> > > is only needed for inflight descriptors which might require other
> > > APIs?
> > >
> > > (All the current vDPA user of vringh is doing in order processing)
> > >
> >
> > That was actually my first attempt and it works equally well for the
> > moment, but it diverges from vhost-kernel behavior for little benefit.
> >
> > To assign both values at set_vring_base mean that if vDPA introduces
> > an (hypothetical) VHOST_VDPA_F_INFLIGHT backend feature in the future,
> > the initialization process would vary a lot:
> > * Without that feature, the used_idx starts with 0, and the avail one
> > is 0 or whatever value the user set with vring_set_base.
> > * With that feature, the device will read guest's used_idx as
> > vhost-kernel? We would enable a new ioctl to set it, or expand
> > set_base to include used_idx, effectively diverting from vhost-kernel?
>
> Adding Longpeng who is looking at this.
>

Sorry, I'll CC longpeng2@xxxxxxxxxx in future series like this one.

> We can leave this to the caller to decide.
>
> Btw, a question, at which case the device used index does not equal to
> the used index in the vring when the device is suspended? I think the
> correct way is to flush the pending used indexes before suspending.
> Otherwise we need an API to get pending used indices?
>
> >
> > To me the wisest option is to move this with vhost-kernel. Maybe we
> > need to add a feature bit to know that the hypervisor can trust the
> > device will do "the right thing" (VHOST_VDPA_F_FETCH_USED_AT_ENABLE?),
> > but we should keep it orthogonal to inflight descriptor migration in
> > my opinion.
>
> I think we need to understand if there are any other possible use
> cases for setting used idx other than inflight stuff.
>

Answering this and the previous comment, I cannot think in any case
outside of inflight migration. I'm just trying to avoid different
behavior between vhost-kernel and vhost-vdpa, and to make features as
orthogonal as possible.

> >
> > Having said that, I'm totally ok to do it otherwise (or to expand the
> > patch message if needed).
>
> I tend to do that in another series (not mix with the fixes).
>
> >
> > > >
> > > > > > - return vringh_init_kern(vrh, features, num, weak_barriers,
> > > > > > - desc, avail, used);
> > > > > > + int r = vringh_init_kern(vrh, features, num, weak_barriers, desc,
> > > > > > + avail, used);
> > > > > > +
> > > > > > + if (r != 0)
> > > > > > + return r;
> > > > > > +
> > > > > > + /* Consider the ring not initialized */
> > > > > > + if ((void *)desc == used)
> > > > > > + return 0;
> > > > >
> > > > > I don't understand when we can get this (actually it should be a bug
> > > > > of the caller).
> > > > >
> > > >
> > > > You can see it in vdpasim_vq_reset.
> > > >
> > > > Note that to consider desc == 0 to be an uninitialized ring is a bug
> > > > IMO. QEMU considers it that way also, but the standard does not forbid
> > > > any ring to be at address 0. Especially if we use vIOMMU.
> > > >
> > > > So I think the best way to know if we can use the vringh is either
> > > > this way, or provide an explicit "initialized" boolean attribute.
> > > > Maybe a new "bool is_initialized(vrh)" is enough, if we don't want to
> > > > add new attributes.
> > >
> > > I wonder if we can avoid this in the simulator level instead of the
> > > vringh (anyhow it only exposes a vringh_init_xxx() helper now).
> > >
> >
> > In my opinion that is a mistake if other drivers will use it to
> > implement the emulated control virtqueue. And it requires more
> > changes. But it is doable for sure.
>
> The problem is, there's no reset API in vringh, that's why you need to
> do if ((void *)desc == used) which depends on behaviour of the vringh
> user.
>

That's a very good point indeed.

> So I think we should either:
>
> 1) move that check in vdpa_sim (since it's not guaranteed that all the
> vringh users will make desc equal to used during reset)
>
> or
>
> 2) introduce a vringh_reset_xxx()
>
> 1) seems a good step for -stable.
>

We can go to 1 for sure. So let's set last_used_idx at
vdpasim_set_vq_state, same value as last_avail_idx, and stash it at
vdpasim_queue_ready.

Do I need to resend the previous patch in this series?

Do we need to offer a new feature flag indicating we will set used_idx
with avail_idx?

Thanks!