[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH v2] Only call stat() when adding a disk if we expect a device to exist.

On 26/04/13 12:17, Ian Campbell wrote:
On Fri, 2013-04-26 at 12:07 +0100, Dave Scott wrote:
On 26/04/13 11:55, Ian Campbell wrote:
On Wed, 2013-04-24 at 12:56 +0100, Ian Campbell wrote:
On Tue, 2013-04-23 at 15:08 +0100, Roger Pau Monne wrote:
On 23/04/13 11:59, David Scott wrote:
We consider calling stat() a helpful error check in the following
circumstances only:
   1. the disk backend type must be PHYsical
   2. the disk backend domain must be the same as the running libxl
   3. there must not be a hotplug script because this would imply that
      the device won't be created until after the hotplug script has

With this fix, it is possible to use qemu's built-in block drivers
such as ceph/rbd, with a xl config disk spec like this:


Signed-off-by: David Scott <dave.scott@xxxxxxxxxxxxx>
Acked-by: Ian Campbell <ian.campbell@xxxxxxxxxx>

Acked-by: Roger Pau Monnà <roger.pau@xxxxxxxxxx>

Applied, thanks.

This patch had the slightly unfortunate impact of causing libxl to not
use blkback for actual block devices present in dom0, causing it to fall
back to tap/qemu.

I'm having a look now...

Thanks (and sorry!)

Looking at the code again, is it because bypassing the stat() in the
UNKNOWN case where in fact it is a block device, leaves a.stab
uninitialised which then confuses

          if (libxl__try_phy_backend(a->stab.st_mode))
              return backend;

in disk_try_backend?

That's my working theory at the minute.

Perhaps a better patch would do the stat() anyway and only worry about
the failure if (disk->backend = ... PHY && disk->backend_domid = ... etc)

I'm not sure why we can't do the libxl__try_phy_backend at the same time
as the stat, all it does is abstract the difference away between Linux
and BSD duplicating a subset of the existing test after the stat...

That's a good point -- we could do only one (platform specific) set of S_IS{BLK,REG} checks if we merged them together.

How about a refactoring along the lines of the attached (built but not yet tested) patch? The changes are:

1. remove the cached stat results in disk_try_backend_args
2. in disk_try_backend(... PHY) we
   * fail if not RAW or empty (as before)
   * assume phy is ok if we have a hotplug script (as before)
* if we're in LIBXL_TOOLSTACK_DOMID, stat() the disk and call libxl__try_backend * if we're not in LIBXL_TOOLSTACK_DOMID then assume phy is ok (basically we can't tell for sure)
2. in libxl__device_disk_set_backend we
   * keep the CDROM sanity check
   * don't do any stat()ing
   * immediately call disk_try_backend (probing if UNKNOWN as before)

Does that make sense? I'll do a bit of testing now.


Attachment: libxl-device-stat.patch
Description: Text Data

Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.