Re: [Qemu-devel] [PATCH 3/6] vhost: Update rings information for IOTLB earlier

["Michael S. Tsirkin" <mst@redhat.com>]

On Fri, May 19, 2017 at 11:48:52AM +0200, Maxime Coquelin wrote:
> 
> 
> On 05/18/2017 05:24 PM, Michael S. Tsirkin wrote:
> > On Thu, May 18, 2017 at 04:45:23PM +0200, Maxime Coquelin wrote:
> > > Hi Michael,
> > > 
> > > On 05/18/2017 09:35 AM, Maxime Coquelin wrote:
> > > > 
> > > > 
> > > > On 05/17/2017 06:41 PM, Michael S. Tsirkin wrote:
> > > > > On Fri, May 12, 2017 at 01:21:18PM +0200, Maxime Coquelin wrote:
> > > > > > 
> > > > > > On 05/11/2017 07:33 PM, Michael S. Tsirkin wrote:
> > > > > > > On Thu, May 11, 2017 at 02:32:43PM +0200, Maxime Coquelin wrote:
> > > > > > > > Vhost-kernel backend need to receive IOTLB entries for rings
> > > > > > > > information early, but vhost-user need the same information
> > > > > > > > earlier, before VHOST_USER_SET_VRING_ADDR is sent.
> > > > > > > Weird. What does VHOST_USER_SET_VRING_ADDR have to do with it?
> > > > > > > 
> > > > > > > According to
> > > > > > >      Starting and stopping rings
> > > > > > > in vhost user spec, vhost user does not access
> > > > > > > anything until ring is started and enabled.
> > > > > > > 
> > > > > > > 
> > > > > > > > This patch also trigger IOTLB miss for all rings informations
> > > > > > > > for robustness, even if in practice these adresses are on the
> > > > > > > > same page.
> > > > > > Actually, the DPDK vhost-user backend is compliant with the spec,
> > > > > > but when handling VHOST_USER_SET_VRING_ADDR request, it translates the
> > > > > > guest addresses into backend VAs, and check they are valid. I
> > > > > > will make the
> > > > > > commit message clearer about this in next revision.
> > > > > > 
> > > > > > The check could be done later, for example when the ring are started,
> > > > > > but it wouldn't change the need to trigger a miss at some point.
> > > > > I think it should be done later, yes. As long as ring is not
> > > > > started addresses should not be interpreted.
> > > > > 
> > > > 
> > > > Ok, then I'll move these addresses translations in the
> > > > VHOST_USER_SET_VRING_KICK handler.
> > > s/VHOST_USER_SET_VRING_KICK/VHOST_USER_SET_VRING_ENABLE/
> > 
> > Note that when protocol features are off ring is started in
> > enabled state, but iommu requires protocol features.
> 
> OK, I will take care of this.
> 
> Note that currently in DPDK, the ring is created in enabled state,
> so it is enabled as soon as started even with protocol features.
> I have done the patch to fix this, will be posted with the patch that
> do the ring addresses translations only when starting/enabling the ring.
> 
> Also, note that disabling VHOST_USER_F_PROTOCOL_FEATURES with latest
> DPDK and QEMU seems broken. I'll add this to my todo list to understand
> where the problem is, but this is lower priority.

And hopefully add a unit test without this so we don't break
it in the future.

> > > I just looked at implementing this change, but I'm not convinced this is
> > > the right thing to do.
> > > 
> > > On backend side, it means saving temporarily the vhost_vring_addr struct
> > > into the vq struct, and moving all what is done currently in
> > > SET_VRING_ADDR handler to SET_VRING_ENABLE one.
> > 
> > Yes, and this is consistent with what the kernel does.
> > 
> > > My understanding of the "Starting and stopping rings" chapter of the
> > > spec is that the ring must not be processed as long as not started and
> > > enabled, not that the addresses passed should not be checked/translated
> > > as it is done today both in DPDK and libvhost-user.
> > > 
> > > If the addresses are invalid, isn't it better to know as soon as
> > > possible?
> > > 
> > > Cheers,
> > > Maxime
> > 
> > There could be valid reasons to set an invalid address temporarily.
> > For example to make sure connection is reset.
> 
> Ok.
> 
> Thanks,
> Maxime