Am 28.09.2015 um 19:32 hat Max Reitz geschrieben: > On 28.09.2015 05:29, Jeff Cody wrote: > > During mirror, if the target device does not have support zero > > initialization, a mirror may result in a corrupt image. > > > > For instance, on mirror to a host device with format = raw, whatever > > random data is on the target device will still be there for unallocated > > sectors. > > > > This is because during the mirror, we set the dirty bitmap to copy only > > sectors allocated above 'base'. In the case of target devices where we > > cannot assume unallocated sectors will be read as zeroes, we need to > > explicitely zero out this data. > > > > In order to avoid zeroing out all sectors of the target device prior to > > mirroring, we do zeroing as part of the block job. A second dirty > > bitmap cache is created, to track sectors that are unallocated above > > 'base'. These sectors are then checked for status of BDRV_BLOCK_ZERO > > on the target - if they are not, then zeroes are explicitly written. > > > > This only occurs under two conditions: > > > > 1. 'mode' != "existing" > > 2. bdrv_has_zero_init(target) == NULL > > > > We perform the mirroring through mirror_iteration() as before, except > > in two passes. If the above two conditions are met, the first pass > > is using the bitmap tracking unallocated sectors, to write the needed > > zeroes. Then, the second pass is performed, to mirror the actual data > > as before. > > > > If the above two conditions are not met, then the first pass is skipped, > > and only the second pass (the one with the actual data) is performed. > > > > Signed-off-by: Jeff Cody <jc...@redhat.com>
> > @@ -737,6 +787,7 @@ static void mirror_start_job(BlockDriverState *bs, > > BlockDriverState *target, > > return; > > } > > > > + s->zero_unallocated = !existing && !bdrv_has_zero_init(target); > > I think this should be set only if we're doing a full mirror operation. > For instance, I could do a none, top or incremental mirror to a new > qcow2 file, which would give it a backing file, obviously. You're lucky > in that qcow2 claims to always have zero initialization, when this is in > fact not true (someone's ought to fix that...): With a backing file, an > overlay file just cannot have zero initialization, it's impossible > (well, unless the backing file is completely zero). bdrv_has_zero_init() takes care of that, in theory. The "problem" here is that the target is opened with BDRV_O_NO_BACKING, so the block layer doesn't consider this an image with a backing file. Is there anything better than bs->backing_hd that we could check? Kevin
pgpOwODfO_1yQ.pgp
Description: PGP signature