On 05/08/2012 12:42 PM, Laine Stump wrote:
On 05/07/2012 07:00 PM, Eric Blake wrote:
> It turns out that when cgroups are enabled, the use of a block device
> for a snapshot target was failing with EPERM due to libvirt failing
> to add the block device to the cgroup whitelist. See also
>
https://bugzilla.redhat.com/show_bug.cgi?id=810200
>
> * src/qemu/qemu_driver.c
> (qemuDomainSnapshotCreateSingleDiskActive)
> (qemuDomainSnapshotUndoSingleDiskActive): Account for cgroup.
> (qemuDomainSnapshotCreateDiskActive): Update caller.
> ---
> @@ -9948,8 +9949,15 @@ qemuDomainSnapshotCreateSingleDiskActive(struct qemud_driver
*driver,
>
> if (virDomainLockDiskAttach(driver->lockManager, vm, disk) < 0)
> goto cleanup;
> + if (cgroup && qemuSetupDiskCgroup(driver, vm, cgroup, disk) < 0) {
As far as I can see, you can't get to here if cgroup is NULL.
cgroup can be NULL; see below...
Technically there's nothing wrong with checking it, but isn't
it
unnecessary?
No, it's necessary (unless we fix qemuSetupDiskCgroup to be a no-op when
cgroup is NULL).
> @@ -10084,6 +10095,7 @@
qemuDomainSnapshotCreateDiskActive(virConnectPtr conn,
> int thaw = 0; /* 1 if freeze succeeded, -1 if freeze failed */
> bool atomic = (flags & VIR_DOMAIN_SNAPSHOT_CREATE_ATOMIC) != 0;
> bool reuse = (flags & VIR_DOMAIN_SNAPSHOT_CREATE_REUSE_EXT) != 0;
> + virCgroupPtr cgroup = NULL;
>
> if (qemuDomainObjBeginJobWithDriver(driver, vm, QEMU_JOB_MODIFY) < 0)
> return -1;
> @@ -10094,6 +10106,14 @@ qemuDomainSnapshotCreateDiskActive(virConnectPtr conn,
> goto endjob;
> }
>
> + if (qemuCgroupControllerActive(driver, VIR_CGROUP_CONTROLLER_DEVICES)
&&
> + virCgroupForDomain(driver->cgroup, vm->def->name, &cgroup, 0))
{
> + qemuReportError(VIR_ERR_INTERNAL_ERROR,
> + _("Unable to find cgroup for %s"),
> + vm->def->name);
> + goto endjob;
> + }
> +
By the time you get here, you're guaranteed the cgroup is non-NULL, right?
No. If qemuCgroupControllerActive() returns false (possible if your
qemu.conf left "devices" out of cgroup_controllers, or if you don't have
cgroups mounted in the kernel at all), then we never call
virCgroupForDomain to populate cgroup with a non-NULL value.
Also, looking beyond the context provided by git - virCGroupFree() is
called in cleanup, which is higher up in the function than end_job, but
there are 2 occurrences of goto end_job past where you get the cgroup -
if you encountered an error and took either of those goto's, you would
end up leaking cgroup.
Eek; I'll fix that and provide a v2.
> @@ -10156,7 +10176,7 @@
qemuDomainSnapshotCreateDiskActive(virConnectPtr conn,
> }
> }
>
> - ret = qemuDomainSnapshotCreateSingleDiskActive(driver, vm,
> + ret = qemuDomainSnapshotCreateSingleDiskActive(driver, vm, cgroup,
So back to the other topic - this is the only place this function is
called, and you're guaranteed that cgroup will be non-NULL, so I don't
think qemuDomainSnapshotCreateSingleDiskActive() actually needs to check
for NULL cgroup.
Calling with NULL cgroup is permissible.
> @@ -10214,6 +10234,8 @@
qemuDomainSnapshotCreateDiskActive(virConnectPtr conn,
> }
>
> cleanup:
> + if (cgroup)
> + virCgroupFree(&cgroup);
I think this either needs to be moved down to end_job, or the to goto
end_job's I mentioned above need to goto cleanup (I think I prefer the
former).
cleanup falls through to endjob - we have to use cleanup if we pause the
cpus or if we modify state that needs saving. But since we do indeed
have a 'goto endjob' that occurs after cgroup might be set, we do indeed
need to move the virCgroupFree() into the endjob label.
--
Eric Blake eblake(a)redhat.com +1-919-301-3266
Libvirt virtualization library
http://libvirt.org