[libvirt] [PATCH] qemu: Correctly label migration TCP socket passed to qemu

[Eric Blake]

On 08/25/2011 08:30 AM, Jiri Denemark wrote:
>>> -        VIR_FORCE_CLOSE(spec->dest.fd.qemu);
>>
>> Unconditional, changed to...
>>
>>> +        if (virSecurityManagerSetImageFDLabel(driver->securityManager, vm,
>>> +                                              spec->dest.fd.qemu) == 0) {
>>> +            ret = qemuMonitorMigrateToFd(priv->mon, migrate_flags,
>>> +                                         spec->dest.fd.qemu);
>>> +            VIR_FORCE_CLOSE(spec->dest.fd.qemu);
>>
>> ...conditional.  Oops - you leak the fd if changing the label fails.
>
> No, there's no leak here. The fd is closed by the caller in any case. Note,
> we don't close it here if we fail before we ever get to
> qemuMonitorMigrateToFd. This additional close after qemuMonitorMigrateToFd()
> is just for better feeling in case migration started to avoid having an fd
> open which doesn't technically belong to us anymore for the whole time of
> migration.

Ah, makes sense.  Until a child process is using the fd, it doesn't 
matter if we close early, and you've convinced me that the caller mops 
up in the cases where we don't close early.  Meanwhile, once a child is 
using the fd, then we don't want to get in the way (especially important 
on pipes, where having multiple processes open as writers on the same 
pipe can prevent readers from seeing EOF), so only closing early after 
spawning the child makes sense.  So,

ACK to this patch.

-- 
Eric Blake   eblake at redhat.com    +1-801-349-2682
Libvirt virtualization library http://libvirt.org