On Wed, 03/09 09:22, Paolo Bonzini wrote: > > > On 09/03/2016 09:00, Fam Zheng wrote: > >> > On 09/03/2016 04:35, Fam Zheng wrote: > >>>>> > >> > enum BdrvTrackedRequestType { > >>>>> > >> > BDRV_TRACKED_READ, > >>>>> > >> > BDRV_TRACKED_WRITE, > >>>>> > >> > - BDRV_TRACKED_FLUSH, > >>>>> > >> > - BDRV_TRACKED_IOCTL, > >>>>> > >> > BDRV_TRACKED_DISCARD, > >>> > > Okay, so flush and ioctl are not needed, but why is discard different? > >> > > >> > Discard can modify the contents of the device, so I think it's safer to > >> > serialize it against RMW and copy-on-read operations. > > Okay, that makes sense, but ioctl like SG_IO can also modify content, no? > > If you use SG_IO you shouldn't use RMW (because scsi_read_complete traps > READ CAPACITY and sets the host block size as the guest block size) or > copy-on-read (because raw has no backing file). > > Besides, BDRV_TRACKED_IOCTL didn't include sector_num/nr_sectors > operations so it didn't provide serialization. >
Yes, that's right. Thanks. Fam