On 07/28/2011 12:07 PM, Matthias Bolte wrote:
2011/7/27 Matthias Bolte <matthias.bolte(a)googlemail.com>:
> doRemoteClose doesn't free the virNetClientPtr and this creates a
> 260kb leak per remote connection. This happens because
> virNetClientFree doesn't remove the last ref, because virNetClientNew
> creates the virNetClientPtr with a refcount of 2.
>
The memory leak I saw was due to virsh calling
virEventRegisterDefaultImpl, but not calling virEventRunDefaultImpl.
Because the event loop is initialized, the call to
virNetSocketAddIOCallback in virNetClientNew succeeds. But virsh not
driving the event loop results in not removing callbacks that were
marked for deletion. Finally this leaks the virNetClientPtr using in
the remote driver. I used a virsh -c qemu:///system to test.
I was able fix this by calling virEventRunDefaultImpl after
virConnectClose in virsh. But I don't think that this is the correct
general solution.
Where do we stand with 0.9.4? Is this something where we need the
general fix before the release, or is just the virsh hack to call
virEventRunDefaultImpl good enough for the release, or is this problem
not severe enough and we can release as-is?
To me the general problem seems to be the entanglement between the
virNetClientPtr refcounting and the event loop. I'm not sure how to
improve this in general. Maybe should have a thread calling
virEventRunDefaultImpl as the comment on virEventRunDefaultImpl
suggest.
I'm hoping danpb will be able to chime in soon.
--
Eric Blake eblake(a)redhat.com +1-801-349-2682
Libvirt virtualization library
http://libvirt.org