Re: [Qemu-devel] [RFC] O_EXCL or not open block device

[NeilBrown <neilb@suse.de>]

[-- Attachment #1: Type: text/plain, Size: 2453 bytes --]

On Thu, 12 Sep 2013 16:27:50 +0200 Kevin Wolf <kwolf@redhat.com> wrote:

> Am 12.09.2013 um 15:58 hat Stefan Hajnoczi geschrieben:
> > On Thu, Sep 12, 2013 at 01:27:32PM +0200, Jack Wang wrote:
> > > Hi all,
> > > 
> > > We're using qemu export md-raid to guest OS, and we saw deadlock on
> > > MD(which is already fixed by Neil), please see thread below:
> > > > http://marc.info/?l=linux-raid&m=137894040228125&w=2
> > > 
> > > As Neil suggested it would be good for userspace applications to call
> > > open() with O_EXCL flag, to avoid such MD hanging problems at the begining.
> > > 
> > > And we checked qemu, it looks it doesn't include O_EXCL flag when open
> > > block device.
> > > 
> > > After search in the mail list we found there are a similar discussion:
> > > 
> > > https://lists.nongnu.org/archive/html/qemu-devel/2011-04/msg00722.html
> > > 
> > > > To O_EXCL or not to O_EXCL open host_cdrom
> > > 
> > > Which looks prefer enable O_EXCL, but I checked latest qemu tree, there
> > > no such code, could anyone give comments on this?
> > 
> > Continuing from that discussion, I think the shared CD-ROM case is
> > something we must avoid breaking.  It worked in the past so it shouldn't
> > break in a new QEMU version.
> > 
> > Kevin: Do you think we should add an option to the host_device
> > BlockDriver that sets the O_EXCL open flag?  That way users and new
> > libvirt can use O_EXCL for host block devices.
> > 
> > The simpler alternative is to always use O_EXCL for non-CDROM host
> > devices.  Simple patch, no configuration required, but it means we
> > continue to lack O_EXCL on CD-ROMs.
> 
> I'm not sure why O_EXCL would be correct on generic block devices when
> it's wrong on CD-ROMs. I think it's in fact more likely that other
> devices are shared, as backing files.

O_EXCL make sense if you or someone else might change the contents.
If the device is read-only, or is uniformly treated as read-only, then O_EXCL
is unnecessary and could definitely get in the way.
I would be nice if we had O_SHARED and O_EXCL, but unfortunately we don't....
I wonder if O_EXCL|O_RDONLY could be treated as a shared lock....  Maybe one
day.

> 
> Adding an option is certainly possible, but what would the default be?
> If O_EXCL is off by default, would anyone actually use it?

I would suggest the default would be to use O_EXCL iff opening with WRITE
access.

NeilBrown

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 828 bytes --]