Re: [Qemu-devel] [PATCH 1/7] block: only force IO completion in .bdrv_truncate if we are shrinking

[Jeff Cody <jcody@redhat.com>]

On Wed, Mar 06, 2013 at 06:50:56PM +0100, Peter Lieven wrote:
> Am 06.03.2013 15:47, schrieb Jeff Cody:
> > Commit 9a665b2b made bdrv_truncate() call bdrv_drain_all(), but this breaks
> > QCOW images, as well other future image formats (such as VHDX) that may call
> > bdrv_truncate(bs->file) from within a read/write operation.  For example, QCOW
> > will cause an assert, due to tracked_requests not being empty (since the
> > read/write that called bdrv_truncate() is still in progress).
> > 
> > This modifies the check to only force the bdrv_drain_all() call if the BDS to
> > be truncated is not growable, or if we are shrinking the BDS.  Otherwise, if we
> > are just growing the file, allow it to happen without forcing a call to
> > bdrv_drain_all().
> 
> This will actually break iscsi_truncate(). Paolo requested that the iscsi_tuncate command
> should use sync I/O for rereading the LUN capacity. Using sync I/O is only possible if
> there are no async requests in-flight.
> 
> Looking at the source I have not found a place where bs->growable is set to 0 for any
> block driver, maybe I miss something. Having bs->growable for iSCSI would also be ok.
> 
> Shouldn't it be possible to call bdrv_drain_all() any time? There are other places
> where this is called. One I have in mind is e.g. if you cancel an ongoing block migration.
> 

That is a good point - what happens to QCOW now, if there is a block
job in progress (e.g. block-commit, block-stream, etc...)?  I would
imagine -EBUSY gets thrown, since bdrv_truncate() checks
bdrv_in_use().

> 
> > 
> > Signed-off-by: Jeff Cody <jcody@redhat.com>
> > ---
> >  block.c | 10 ++++++++--
> >  1 file changed, 8 insertions(+), 2 deletions(-)
> > 
> > diff --git a/block.c b/block.c
> > index 124a9eb..dce844c 100644
> > --- a/block.c
> > +++ b/block.c
> > @@ -2450,8 +2450,14 @@ int bdrv_truncate(BlockDriverState *bs, int64_t offset)
> >      if (bdrv_in_use(bs))
> >          return -EBUSY;
> >  
> > -    /* There better not be any in-flight IOs when we truncate the device. */
> > -    bdrv_drain_all();
> > +    /* Don't force a drain if we are just growing the file - otherwise,
> > +     * using bdrv_truncate() from within a block driver in a read/write
> > +     * operation will cause an assert, because bdrv_drain_all() will assert if
> > +     * a tracked_request is still in progress. */
> > +    if (!bs->growable || offset < bdrv_getlength(bs)) {
> > +        /* There better not be any in-flight IOs when we truncate the device. */
> > +        bdrv_drain_all();
> > +    }
> >  
> >      ret = drv->bdrv_truncate(bs, offset);
> >      if (ret == 0) {
> > 
>