Re: [PATCH] nbd: server: Report holes for raw images

[Kevin Wolf <kwolf@redhat.com>]

Am 25.02.2021 um 19:50 hat Vladimir Sementsov-Ogievskiy geschrieben:
> 19.02.2021 19:58, Eric Blake wrote:
> > On 2/19/21 10:42 AM, Eric Blake wrote:
> > 
> > > > To me, data=false looks compatible with NBD_STATE_HOLE. From user point
> > > > of view, getting same results from qemu-nbd and qemu-img is more
> > > > important than being more correct about allocation status.
> > > 
> > > More to the point, here is our inconsistency:
> > > 
> > > In nbd/server.c, we turn !BDRV_BLOCK_ALLOCATED into NBD_STATE_HOLE
> > > 
> > > In block/nbd.c, we turn !NBD_STATE_HOLE into BDRV_BLOCK_DATA
> > > 
> > > The fact that we are not doing a round-trip conversion means that one of
> > > the two places is wrong.  And your argument that the server side is
> > > wrong makes sense to me.
> > 
> > In fact, when I went back and researched when this was introduced (see
> > commit e7b1948d51 in 2018), we may have been aware of the inconsistency
> > between client and server, but didn't make up our minds at the time:
> > https://lists.gnu.org/archive/html/qemu-devel/2018-03/msg03465.html
> > "? Hm, don't remember, what we decided about DATA/HOLE flags mapping.."
> > 
> > > 
> > > I'll wait a few days for any other reviewer commentary before taking
> > > this through my NBD tree.
> > > 
> > 
> 
> 
> I can add the following.
> 
> First, link to my research of block_status in Qemu:
> https://lists.gnu.org/archive/html/qemu-devel/2020-04/msg05136.html
> 
> And about HOLE and ZERO..
> 
> As I've noted in the research above, SCSI may return HOLE & !ZERO:
> 
> from SCSI: Logical Block Provisioning Read Zeros (LBPRZ) bit 1     If
> the logical block provisioning read zeros (LBPRZ) bit is set to one,
> then, for an unmapped LBA specified by a read operation, the
> deviceserver shall send user data with all bits set to zero to the
> data-in buffer.  0     If the TPRZ bit is set to zero, then, for an
> unmapped LBA specified by a read operation, the device server may send
> user data with all bitsset to any value to the data-in buffer.
> 
> So we can have an unmapped area that can be read as any random data.
> Same thing can be said about null-co driver with read-zeroes=false
> 
> Also, qcow2 support ALLOCATED ZERO clusters which reads as zero but
> data is allocated - they are reasonable to report as ZERO & !HOLE
> 
> And of-course UNALLOCATED ZERO clusters in qcow2 and lseek-holes are
> reasonable to report as ZERO & HOLE,  because they reads as zero and
> "future writes to that area may cause fragmentation or encounter an
> NBD_ENOSPC"..
> 
> So, all combination are reasonable, we just need to fix Qemu NBD
> server to report correct statuses in all these cases.
> 
> It seems that ZERO/HOLE specification is a lot more reasonable than
> what we have with ZERO/DATA/ALLOCATED in Qemu, and may be true way is
> move internal block_status to use NBD terms.

Is there not a 1:1 correspondence between our internal flags and the NBD
ones? ZERO is exactly the same, and HOLE is the inversion of DATA.

ALLOCATED is important internally when finding the node in a backing
file chain that actually defines the content, but for a user it doesn't
make a difference. This is why it isn't exposed in NBD.

So I think both QEMU and NBD use the flags that make sense in the
respective context.

Kevin