On Wed, Feb 23 2022, Jason Gunthorpe <jgg@xxxxxxxxxx> wrote: > On Wed, Feb 23, 2022 at 06:06:13PM +0100, Cornelia Huck wrote: >> On Sun, Feb 20 2022, Yishai Hadas <yishaih@xxxxxxxxxx> wrote: >> > +/* >> > + * Indicates the device can support the migration API through >> > + * VFIO_DEVICE_FEATURE_MIG_DEVICE_STATE. If present flags must be non-zero and >> > + * VFIO_DEVICE_FEATURE_MIG_DEVICE_STATE is supported. The RUNNING and >> >> I'm having trouble parsing this. I think what it tries to say is that at >> least one of the flags defined below must be set? >> >> > + * ERROR states are always supported if this GET succeeds. >> >> What about the following instead: >> >> "Indicates device support for the migration API through >> VFIO_DEVICE_FEATURE_MIG_DEVICE_STATE. If present, the RUNNING and ERROR >> states are always supported. Support for additional states is indicated >> via the flags field; at least one of the flags defined below must be >> set." > > Almost, 'at least VFIO_MIGRATION_STOP_COPY must be set' It feels a bit odd to split the mandatory states between a base layer (RUNNING/ERROR) and the ones governed by VFIO_MIGRATION_STOP_COPY. Do we want to keep the possibility of a future implementation that does not use the semantics indicated by VFIO_MIGRATION_STOP_COPY? If yes, it should be "one of the flags" and the flags that require VFIO_MIGRATION_STOP_COPY to be set as well need to note that dependency. If not, we should explicitly tag VFIO_MIGRATION_STOP_COPY as mandatory (so that the flag's special status is obvious.)
