[dpdk-dev] [PATCH] vhost: Clean iotlb cache on vring stop

[Eugenio Perez Martin]

On Tue, Aug 31, 2021 at 4:02 AM Xia, Chenbo <chenbo.xia at intel.com> wrote:
>
> Hi Eugenio,
>
> > -----Original Message-----
> > From: Eugenio Perez Martin <eperezma at redhat.com>
> > Sent: Tuesday, August 31, 2021 2:10 AM
> > To: Xia, Chenbo <chenbo.xia at intel.com>
> > Cc: Maxime Coquelin <maxime.coquelin at redhat.com>; dev at dpdk.org; Pei Zhang
> > <pezhang at redhat.com>; Jason Wang <jasowang at redhat.com>
> > Subject: Re: [PATCH] vhost: Clean iotlb cache on vring stop
> >
> > On Mon, Aug 30, 2021 at 1:58 PM Xia, Chenbo <chenbo.xia at intel.com> wrote:
> > >
> > > Hi Eugenio,
> > >
> >
> > Hi Chenbo,
> >
> > > > -----Original Message-----
> > > > From: Eugenio Pérez <eperezma at redhat.com>
> > > > Sent: Saturday, August 28, 2021 12:13 AM
> > > > To: Maxime Coquelin <maxime.coquelin at redhat.com>; Xia, Chenbo
> > > > <chenbo.xia at intel.com>
> > > > Cc: dev at dpdk.org; Pei Zhang <pezhang at redhat.com>; Jason Wang
> > > > <jasowang at redhat.com>
> > > > Subject: [PATCH] vhost: Clean iotlb cache on vring stop
> > >
> > > Clean -> clean
> > >
> >
> > Is that something I need to send a new revision for, or is it ok to
> > apply on the maintainer side?
> >
> > > >
> > > > Old IOVA cache entries are left when there is a change on virtio driver
> > > > in VM. In case that all these old entries have iova addresses lesser
> > > > than new iova entries, vhost code will need to iterate all the cache to
> > > > find the new ones. In case of just a new iova entry needed for the new
> > > > translations, this condition will last forever.
> > > >
> > > > This has been observed in virtio-net to testpmd's vfio-pci driver
> > > > transition, reducing the performance from more than 10Mpps to less than
> > > > 0.07Mpps if the hugepage address was higher than the networking
> > > > buffers. Since all new buffers are contained in this new gigantic page,
> > > > vhost needs to scan IOTLB_CACHE_SIZE - 1 for each translation at worst.
> > >
> > > I'm curious why QEMU will not invalidate iotlb when virtio-net driver is
> > removed
> > > (dma region should be unmapped).
> > >
> >
> > I'm going to investigate this more, but qemu iommu notifier callback
> > (vhost_iommu_unmap_notify) is never called through all the test. Also,
> > guest kernel code calls dma_unmap_page for each buffer and vqs, but it
> > never generates an iotlb flush.
> >
> > Or do you mean that qemu should also flush all iotlb entries on vhost
> > device stop?
>
> I think as you said, the driver unmapped all entries, so theoretically the iotlb
> entries should all be invalidated. The iotlb invalidation should be triggered by
> iommu entry unmapping. Make sense to you?
>

It totally makes sense. I just wanted to make sure at what level you
meant the unmap should happen: If you meant qemu, then we need more
code. If you meant the virtio-net/ring kernel driver, then there is a
bug/misunderstanding on what is happening.

I think the iotlb unmapping is not happening because the entries are
still valid, although not needed with vfio/guest's testpmd. If that
were not true we would see way more errors with virtio-net+iommu
because of invalid translations:

* virtio-net adds a buffer to virtio_ring. virtio-ring dma map does
not generate a iotlb map: Qemu will wait for each device to request
the missing pages.
* The device will miss the translation: It requests it, and saves it
in the list for the future.
* virtio-ring is done with the buffer: It unmaps them. After a while,
kernel needs to dma_map that page again. This new mapping does not
generate a page miss as far as I see.
* Since this new mapping is already on the device's iotlb cache, the
device will not issue a page miss. That means the page must be mapped
to the same GPA as before: If not, the network would fail.

With this assumption, it is useless to iotlb-unmap them: The device
will need to request them again if they're cleared from its cache. But
all of this is my guess and I could be wrong :).

> >
> > > And since the perf drop is huge, why not cc to stable and add fix tag?
> > >
> >
> > I was not sure if it was worth it to backport, but I would say that
> > the issue can be reproduced with enough bad luck. Since translations
> > have always been saved in a linked list:
> >
> > Fixes: d012d1f293f4 ("vhost: add IOTLB helper functions")
> >
> > Same question as before, if no changes to the code are needed for the
> > patch, do I need to send a second revision?
>
> I think it sounds a bug to me. So no matter how often it could be reproduced, it's
> worth to backport. I suggest to send a v2 with the title fixed and Fixes & cc
> stable tag added so that it will show in stable mailing list and save Maxime's effort
> when applying.
>

Sure, I will do it.

Thanks!

> Thanks,
> Chenbo
>
> >
> > Thanks!
> >
> >
> > > Thanks,
> > > Chenbo
> > >
> > > >
> > > > Signed-off-by: Eugenio Pérez <eperezma at redhat.com>
> > > > Reported-by: Pei Zhang <pezhang at redhat.com>
> > > > ---
> > > >  lib/vhost/vhost_user.c | 2 ++
> > > >  1 file changed, 2 insertions(+)
> > > >
> > > > diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c
> > > > index 29a4c9af60..7de48f5333 100644
> > > > --- a/lib/vhost/vhost_user.c
> > > > +++ b/lib/vhost/vhost_user.c
> > > > @@ -2113,6 +2113,8 @@ vhost_user_get_vring_base(struct virtio_net **pdev,
> > > >       msg->size = sizeof(msg->payload.state);
> > > >       msg->fd_num = 0;
> > > >
> > > > +     vhost_user_iotlb_flush_all(vq);
> > > > +
> > > >       vring_invalidate(dev, vq);
> > > >
> > > >       return RTE_VHOST_MSG_RESULT_REPLY;
> > > > --
> > > > 2.27.0
> > >
>