> From: Jason Gunthorpe <jgg@xxxxxxxxxx>
> Sent: Friday, April 19, 2024 10:11 PM
>
> On Fri, Apr 19, 2024 at 05:23:30AM +0000, Tian, Kevin wrote:
> > > From: Alex Williamson <alex.williamson@xxxxxxxxxx>
> > > Sent: Friday, April 19, 2024 6:55 AM
> > >
> > > On Wed, 17 Apr 2024 22:31:41 +0800
> > > Xin Zeng <xin.zeng@xxxxxxxxx> wrote:
> > >
> > > > +
> > > > + /*
> > > > + * As the device is not capable of just stopping P2P DMAs, suspend
> > > the
> > > > + * device completely once any of the P2P states are reached.
> > > > + * On the opposite direction, resume the device after transiting from
> > > > + * the P2P state.
> > > > + */
> > > > + if ((cur == VFIO_DEVICE_STATE_RUNNING && new ==
> > > VFIO_DEVICE_STATE_RUNNING_P2P) ||
> > > > + (cur == VFIO_DEVICE_STATE_PRE_COPY && new ==
> > > VFIO_DEVICE_STATE_PRE_COPY_P2P)) {
> > > > + ret = qat_vfmig_suspend(qat_vdev->mdev);
> > > > + if (ret)
> > > > + return ERR_PTR(ret);
> > > > + return NULL;
> > > > + }
> > >
> > > This doesn't appear to be a valid way to support P2P, the P2P states
> > > are defined as running states. The guest driver may legitimately
> > > access and modify the device state during P2P states.
> >
> > yes it's a conceptual violation of the definition of the P2P states.
>
> It depends what suspend actually does.
>
> Like if it halts all queues and keeps them halted, while still
> allowing queue head/tail pointer updats then it would be a fine
> implementation for P2P.
Yes that really depends. e.g. a queue accepting direct stores (MOVDIR64B)
for work submission may have problem if that store is simply abandoned
when the queue is disabled. ENQCMD is possibly OK as unaccepted store
will get a retry indicator to software so nothing is lost.
I'll let Xin confirm on the QAT implementation (for all device registers).
If it is like Jason's example then we should provide a clear comment
clarifying that doing suspend at RUNNING_P2P is safe for QAT as the
device MMIO interface still works according to the definition of RUNNING
and no request is lost (either from CPU or peer). There is nothing to stop
from RUNNING_P2P to STOP because the device doesn't execute any
request to further change the internal state other than MMIO.
>
> > > Should this device be advertising support for P2P?
> >
> > Jason suggests all new migration drivers must support P2P state.
> > In an old discussion [1]
>
> I did? I don't think that is what the link says..
Not this link which I roughly remember was a follow-up to your earlier
comment.
But I cannot find that statement so probably my memory was bad. The
closest one is:
https://lore.kernel.org/intel-wired-lan/ZJMH3DF7nJ+OG9BJ@xxxxxxxx/#t
but it just talks about lacking of P2P is a problem similar to you replied below:
>
> We've been saying for a while that devices should try hard to
> implement P2P because if they don't then multi VFIO VMM's won't work
> and people will be unhappy..
yes but we have to admit that existing devices may not meet this requirement. 😊
