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

[Kevin Wolf]

Am 07.03.2013 um 09:53 hat Peter Lieven geschrieben:
> On 06.03.2013 21:39, Paolo Bonzini wrote:
> >Il 06/03/2013 20:03, Peter Lieven ha scritto:
> >>Am 06.03.2013 19:48, schrieb Jeff Cody:
> >>>On Wed, Mar 06, 2013 at 07:31:51PM +0100, Paolo Bonzini wrote:
> >>>>Il 06/03/2013 19:14, Jeff Cody ha scritto:
> >>>>>QCOW breaks with it using a normal raw posix file as a device.  As a
> >>>>>test: qemu-img create -f qcow test.qcow 5G.  Now run qemu with that
> >>>>>drive mounted, and try to partition and format it.  QEMU now asserts.
> >>>>>
> >>>>>The nicety of being able to using truncate during a write call,
> >>>>>especially for VHDX (which can have relatively large block/cluster
> >>>>>sizes), so to grow the file sparsely in a dynamically allocated file.
> >>>>
> >>>>Perhaps we need two APIs, "truncate" and "revalidate".
> >>>>
> >>>>Truncate should be a no-op if (!bs->growable).
> >>>>
> >>>>Revalidate could be called by the block_resize monitor command with no
> >>>>size specified.
> >>>>
> >>>>Paolo
> >>>
> >>>I think that is a good solution.  Is it better to have "truncate" and
> >>>"revalidate", or "truncate" and "grow", with grow being a subset of
> >>>truncate, with fewer restrictions?  There may still be operations
> >>>where it is OK to grow a file, but not OK to shrink it.
> >>
> >>Or as a first step:
> >>
> >>a) Call brdv_drain_all() only if the device is shrinked (independently of 
> >>!bs->growable)
> >>b) Call brdv_drain_all() inside iscsi_truncate() because it is a special 
> >>requirement there
> >>c) Fix the value of bs->growable for all drivers
> >
> >Let's start from (c).  bdrv_file_open sets bs->growable = 1.  I think it
> >should be removed and only the file protocol should set it.
> >
> >Then we can add bdrv_revalidate and, for block_resize, call
> >bdrv_revalidate+bdrv_truncate.  For bs->growable = 0 &&
> >!bs->drv->bdrv_truncate, bdrv_truncate can just check that the actual
> >size is the same or bigger as the one requested, and fail otherwise.
> >
> >Paolo
> >
> 
> Regarding brd_drain_all(). Is the fix right to call it only on device shrink?
> In this case it has to be added to iscsi_truncate as well.

The real fix would bdrv_drain(bs). I hope we're not too far away from
that today, even though we're not quite there.

Kevin