[libvirt] [PATCH 2/4] Avoid create/unlink with block devs used for QEMU save

[Daniel P. Berrange]

On Thu, Apr 22, 2010 at 11:35:17AM +0100, Daniel P. Berrange wrote:
> On Wed, Apr 21, 2010 at 02:49:42PM -0600, Eric Blake wrote:
> > On 04/21/2010 10:56 AM, Daniel P. Berrange wrote:
> > > It is possible to use block devices with domain save/restore. Upon
> > > failure QEMU unlinks the path being saved to. This isn't good when
> > > it is a block device !
> > > 
> > > +    struct stat sb;
> > > +    int is_bdev = 0;
> > 
> > Should this be bool instead of int?
> > 
> > >  
> > > +    /* path might be a pre-existing block dev, in which case
> > > +     * we need to skip the create step, and also avoid unlink
> > > +     * in the failure case */
> > > +    if (stat(path, &sb) < 0) {
> > > +        if (errno != ENOENT) {
> > > +            virReportSystemError(errno, _("unable to access %s"), path);
> > > +            goto endjob;
> > > +        } else {
> > > +            is_bdev = 0;
> > > +        }
> > > +    } else {
> > > +        is_bdev = S_ISBLK(sb.st_mode);
> > 
> > although if you use bool, you'd have to explicitly compare against 0
> > here, based on the gnulib restraints on stdbool.h.  Also, what about
> > other non-regular files?  A directory will fail on the open, but what
> > about named fifos (on the other hand, is anyone crazy enough to try a
> > non-seekable file as the file to open?).  So I'm wondering if it's
> > better to check for S_ISREG, changing is_bdev to is_reg, and changing
> > the cleanup logic to only attempt unlink() if is_reg, rather than
> > skipping if is_bdev.
> > 
> > > +    }
> > > +
> > > +
> > >      /* Setup hook data needed by virFileOperation hook function */
> > >      hdata.dom = dom;
> > >      hdata.path = path;
> > > @@ -4849,7 +4866,8 @@ static int qemudDomainSaveFlag(virDomainPtr dom, const char *path,
> > >      /* Write header to file, followed by XML */
> > >  
> > >      /* First try creating the file as root */
> > > -    if ((rc = virFileOperation(path, O_CREAT|O_TRUNC|O_WRONLY,
> > > +    if (!is_bdev &&
> > > +        (rc = virFileOperation(path, O_CREAT|O_TRUNC|O_WRONLY,
> > 
> > This is racy.  You stat() prior to open()ing.  Wouldn't it be better to
> > open(), then fstat()?
> 
> That would require significant changes to the virFileOperation code
> which is not something I think is worth the effort here. 

Actually, the whole thing is racy regardless of what we do, since QEMU itself
will spawn another program which will re-open this file.

To avoid a race, we'd need to switch to QEMU's  'fd:' based migration and
pass the pre-opened file handle to it

Daniel
-- 
|: Red Hat, Engineering, London    -o-   http://people.redhat.com/berrange/ :|
|: http://libvirt.org -o- http://virt-manager.org -o- http://deltacloud.org :|
|: http://autobuild.org        -o-         http://search.cpan.org/~danberr/ :|
|: GnuPG: 7D3B9505  -o-   F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|