On 08/02/2011 01:18 PM, Laine Stump wrote:
On 08/02/2011 01:31 PM, Eric Blake wrote:
> Revert 6a1f5f568f8. Now that libvirt_iohelper no longer has a
> race where it can open() a file after the parent process has
> unlink()d the same file, it makes more sense to make the callers
> both create and unlink, rather than the caller create and the
> stream unlink.
I wasn't paying attention to the messages/patches related to the race
condition you reference,
Commit 1eb66479 plugged the race; commit 6a1f5f5 introduced the race in
the first place.
The problem was that if we use libvirt_iohelper, and the child process
calls open(), but the parent process calls unlink() before the child
process gets to run very far, then the child process will fail to
open(). But by changing fdstream to pass the fd to libvirt-iohelper by
fd inheritance instead of by name-wise open() calls, there is no longer
an open() race, so we can once again unlink() in the parent.
but this (caller creates and unlinks)
definitely seems cleaner than the other way. Beyond that, the patch
seems to be correct. ACK.
Should this go in for 0.9.4, or am I correct in deferring it until after
the release?
--
Eric Blake eblake(a)redhat.com +1-801-349-2682
Libvirt virtualization library
http://libvirt.org