]> git.ipfire.org Git - thirdparty/libvirt.git/commit
vircgroup: Fix virCgroupKillRecursive() wrt nested controllers
authorMichal Privoznik <mprivozn@redhat.com>
Fri, 16 Apr 2021 14:39:14 +0000 (16:39 +0200)
committerMichal Privoznik <mprivozn@redhat.com>
Mon, 19 Apr 2021 09:21:40 +0000 (11:21 +0200)
commitea7d0ca37cce76e1327945c4864b996d7fd6d2e6
tree431cca28b23720e3279b3420eadcdffecbc92383
parenta0815484b1068e6a0612ba9871237bcfc5710f42
vircgroup: Fix virCgroupKillRecursive() wrt nested controllers

I've encountered the following bug, but only on Gentoo with
systemd and CGroupsV2. I've started an LXC container successfully
but destroying it reported the following error:

  error: Failed to destroy domain 'amd64'
  error: internal error: failed to get cgroup backend for 'pathOfController'

Debugging showed, that CGroup hierarchy is full of surprises:

/sys/fs/cgroup/machine.slice/machine-lxc\x2d861\x2damd64.scope/
└── libvirt
    ├── dev-hugepages.mount
    ├── dev-mqueue.mount
    ├── init.scope
    ├── sys-fs-fuse-connections.mount
    ├── sys-kernel-config.mount
    ├── sys-kernel-debug.mount
    ├── sys-kernel-tracing.mount
    ├── system.slice
    │   ├── console-getty.service
    │   ├── dbus.service
    │   ├── system-getty.slice
    │   ├── system-modprobe.slice
    │   ├── systemd-journald.service
    │   ├── systemd-logind.service
    │   └── tmp.mount
    └── user.slice

For comparison, here's the same container on recent Rawhide:

/sys/fs/cgroup/machine.slice/machine-lxc\x2d13550\x2damd64.scope/
└── libvirt

Anyway, those nested directories should not be a problem, because
virCgroupKillRecursiveInternal() removes them recursively, right?
Sort of. The function really does remove nested directories, but
it assumes that every directory has the same controller as the
rest. Just take a look at virCgroupV2KillRecursive() - it gets
'Any' controller (the first one it found in ".scope") and then
passes it to virCgroupKillRecursiveInternal().

This assumption is not true though. The controllers found in
".scope" are the following:

  cpuset cpu io memory pids

while "libvirt" has fewer:

  cpuset cpu io memory

Up until now it's not problem, because of how we order
controllers internally - "cpu" is the first and thus picking
"Any" controller returns just that. But the rest of directories
has no controllers, their "cgroup.controllers" is just empty.

What fixes the bug is dropping @controller argument from
virCgroupKillRecursiveInternal() and letting each iteration work
pick its own controller.

Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Pavel Hrdina <phrdina@redhat.com>
src/util/vircgroup.c
src/util/vircgrouppriv.h
src/util/vircgroupv1.c
src/util/vircgroupv2.c