Re: [Qemu-devel] [PATCH] block: Flush image after open

[Avi Kivity <avi@redhat.com>]

On 03/21/2011 03:02 PM, Kevin Wolf wrote:
> Am 21.03.2011 13:23, schrieb Avi Kivity:
> >  On 03/09/2011 07:38 PM, Anthony Liguori wrote:
> >>  On 03/09/2011 11:27 AM, Christoph Hellwig wrote:
> >>>  On Wed, Mar 09, 2011 at 05:15:53PM +0100, Kevin Wolf wrote:
> >>>>  Quoting the bug report:
> >>>>
> >>>>       qemu ensures that guest writes and qemu metadata writes hit the
> >>>>  disk
> >>>>       when necessary to prevent data corruption. However, if an image
> >>>>  was
> >>>>       in host pagecache prior to starting qemu, for example after
> >>>>  running
> >>>>       qemu-img convert, then nothing prevents writes from reaching the
> >>>>       disk out of order, potentially causing corruption.
> >>>>
> >>>>  I'm not entirely sure if there is a realistic case where we would get
> >>>>  corruption, but it's probably a case of better safe than sorry.
> >>>  Except for SCSI with ordered tags (which we don't support) there are not
> >>>  ordering guarantees in the storage protocols, and as such the above
> >>>  explanation
> >>>  doesn't make any sense at all.
> >>
> >>  Even if there was, a guest shouldn't be relying on the ordering of a
> >>  write that comes from a non-guest.
> >>
> >>  I don't understand the failure scenario here.
> >
> >     $ cp x.img y.img
> >     $ qemu -drive file=y.img,cache=writeback
> >  <read something from disk, send it over the network>
> >  <no guest flushes>
> >  <host crash>
> >
> >  The guest may expect that any or none of its writes hit the disk, but
> >  that anything that it read from the disk, stays there.
>
> Is it true for real hardware? Consider a reboot, you could still have
> some data in a volatile disk write cache if the OS that ran before the
> reboot hasn't flushed it.

That's if RESET doesn't flush the cache.  It's probably false for fc or 
iscsi, but possibly true for IDE.

But it can't happen for a single-boot host, or a dual boot host with no 
shared partitions.

-- 
error compiling committee.c: too many arguments to function