On 08/02/2011 01:29 PM, Eric Blake wrote:
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?
Laine and I chatted some more on IRC:
<eblake> laine: should I push both patches now (or even squash
together), or delay the second patch till after the release?
<laine> eblake: I don't see a problem with the second patch, I think I
like it better. Not knowing the history of the race you mentioned and
whether or not this code still touches it, I wouldn't single-handledly
say to squash the two together, but if you're comfortable with it, then
it looks good to me.
<eblake> I'll update the commit message of the second; the race was
first solved in 6a1f5f5 by making the parent not unlink() if
libvirt_iohelper will also be open()ing the file
<eblake> but solving it in that manner required passing delete=true all
the way through the fdstream code
<eblake> later, I pushed commit 1eb66479, which taught libvirt_iostream
to operate on an inherited fd instead of calling open()
<eblake> and in so doing, the child no longer needs to unlink()
<eblake> so we no longer have the unlink() in parent prior to
open()/unlink() in the child, and can get rid of the extra parameter
Given that, I went ahead and updated the commit comment, then pushed
both patches separately, to make it into 0.9.4.
Revert 6a1f5f568f8. Now that libvirt_iohelper takes fds by
inheritance rather than by open() (commit 1eb66479), there is
no longer a race where the parent can unlink() a file prior to
the iohelper open()ing the same file. From there, it makes
more sense to have the callers both create and unlink, rather
than the caller create and the stream unlink, since the latter
was only needed when iohelper had to do the unlink.
--
Eric Blake eblake@xxxxxxxxxx +1-801-349-2682
Libvirt virtualization library http://libvirt.org
--
libvir-list mailing list
libvir-list@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/libvir-list