On Mon, 28 Feb 2022 16:29:19 -0400
Jason Gunthorpe <jgg@xxxxxxxxxx> wrote:
> On Mon, Feb 28, 2022 at 01:16:14PM -0700, Alex Williamson wrote:
> > On Mon, 28 Feb 2022 14:05:20 -0400
> > Jason Gunthorpe <jgg@xxxxxxxxxx> wrote:
> >
> > > On Mon, Feb 28, 2022 at 06:01:44PM +0000, Shameerali Kolothum Thodi wrote:
> > >
> > > > +static long hisi_acc_vf_save_unl_ioctl(struct file *filp,
> > > > + unsigned int cmd, unsigned long arg)
> > > > +{
> > > > + struct hisi_acc_vf_migration_file *migf = filp->private_data;
> > > > + loff_t *pos = &filp->f_pos;
> > > > + struct vfio_device_mig_precopy precopy;
> > > > + unsigned long minsz;
> > > > +
> > > > + if (cmd != VFIO_DEVICE_MIG_PRECOPY)
> > > > + return -EINVAL;
> > >
> > > ENOTTY
> > >
> > > > +
> > > > + minsz = offsetofend(struct vfio_device_mig_precopy, dirty_bytes);
> > > > +
> > > > + if (copy_from_user(&precopy, (void __user *)arg, minsz))
> > > > + return -EFAULT;
> > > > + if (precopy.argsz < minsz)
> > > > + return -EINVAL;
> > > > +
> > > > + mutex_lock(&migf->lock);
> > > > + if (*pos > migf->total_length) {
> > > > + mutex_unlock(&migf->lock);
> > > > + return -EINVAL;
> > > > + }
> > > > +
> > > > + precopy.dirty_bytes = 0;
> > > > + precopy.initial_bytes = migf->total_length - *pos;
> > > > + mutex_unlock(&migf->lock);
> > > > + return copy_to_user((void __user *)arg, &precopy, minsz) ? -EFAULT : 0;
> > > > +}
> > >
> > > Yes
> > >
> > > And I noticed this didn't include the ENOMSG handling, read() should
> > > return ENOMSG when it reaches EOS for the pre-copy:
> > >
> > > + * During pre-copy the migration data FD has a temporary "end of stream" that is
> > > + * reached when both initial_bytes and dirty_byte are zero. For instance, this
> > > + * may indicate that the device is idle and not currently dirtying any internal
> > > + * state. When read() is done on this temporary end of stream the kernel driver
> > > + * should return ENOMSG from read(). Userspace can wait for more data (which may
> > > + * never come) by using poll.
> >
> > I'm confused by your previous reply that the use of curr_state should
> > be eliminated, isn't this ioctl only valid while the device is in the
> > PRE_COPY or PRE_COPY_P2P states? Otherwise the STOP_COPY state would
> > have some expectation to be able to use this ioctl for devices
> > supporting PRE_COPY.
>
> I think it is fine to keep working on stop copy, though the
> implementation here isn't quite right for that..
>
> if (migf->total_length > QM_MATCH_SIZE)
> precopy.dirty_bytes = migf->total_length - QM_MATCH_SIZE - *pos;
> else
> precopy.dity_bytes = 0;
>
> if (*pos < QM_MATCH_SIZE)
> precopy.initial_bytes = QM_MATCH_SIZE - *pos;
> else
> precopy.initial_Bytes = 0;
>
> Unless you think we should block it.
What's the meaning of initial_bytes and dirty_bytes while in STOP_COPY?
It seems like these become meaningless and if so, why shouldn't the
ioctl simply return -EINVAL if the device state doesn't match the
window where it's useful?
> > I'd like to see the uapi clarify exactly what states allow this
> > ioctl and define the behavior of the ioctl when transitioning out of
> > those states with an open data_fd, ie. is it defined to return an
> > -errno once in STOP_COPY? Thanks,
>
> The ioctl is on the data_fd, so it should follow all the normal rules
> of the data_fd just like read() - ie all ioctls/read/write fails when
> teh state is moved outside one where the data_fd is valid.
>
> That looks like another issue with the above, it doesn't chck
> migf->disabled.
>
> Should we add another sentence about this?
Right, of course the ioctl goes away when the data_fd is invalid, the
question is more that we've created this PRE_COPY_* specific ioctl and
what does it mean to call it when not in a device state where the
data_fd is still valid but this ioctl is really not. We should
specify how the driver is intended to respond to this ioctl in
STOP_COPY. Thanks,
Alex
