Two small validation relaxations for cgroup v2, making it easier to
manage the hierarchy without the current pain points. Both changes
work for both mems and cpus (but I have no NUMA machine to test mems).
Hopefully the patches has an ok description about what change they
provide, and why they are helpful.
Odin Ugedal (2):
cgroup2: cpuset: Disable subset validation on set
cgroup2: cpuset: Always allow setting empty cpuset
kernel/cgroup/cpuset.c | 17 ++++++++++-------
1 file changed, 10 insertions(+), 7 deletions(-)
--
2.31.0
Due to how cpuset works in v2, there is no need to check if direct
children are subsets of the new cpuset. In v2, the effective cpuset
of a cgroup is the intersection between the effective value of the parent
and the cpuset.cpus value of the cgroup, with a fallback to the effective
parent value in case the intersection is an empty set.
Therefore, in v2, it is allowed to set cpuset.cpus to a a value that is
not a subset of the parents effective value, resulting in inheriting the
effective cpuset from the parent. Therefore the validation when updating
the parent cpuset (in this case) is not necessary, and can be disabled.
Example:
/sys/fs/cgroup/A (cpus=1-2,cpus.effective=1-2)
/sys/fs/cgroup/A/B (cpus="", cpus.effective=1-2)
Setting cpus to 3-4 is possible for A/B
/sys/fs/cgroup/A (cpus=1-2,cpus.effective=1-2)
/sys/fs/cgroup/A/B (cpus=3-4,cpus.effective=1-2)
Setting cpus to 1 for A would result in an -EBUSY error, but
with this patch it will work as expected:
/sys/fs/cgroup/A (cpus=1, cpus.effective=1)
/sys/fs/cgroup/A/B (cpus=3-4,cpus.effective=1)
This also applies in a similar manner on cpuset.mems.
Signed-off-by: Odin Ugedal <[email protected]>
---
kernel/cgroup/cpuset.c | 10 ++++++----
1 file changed, 6 insertions(+), 4 deletions(-)
diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
index 5258b68153e0..f543c4c6084a 100644
--- a/kernel/cgroup/cpuset.c
+++ b/kernel/cgroup/cpuset.c
@@ -572,11 +572,13 @@ static int validate_change(struct cpuset *cur, struct cpuset *trial)
rcu_read_lock();
- /* Each of our child cpusets must be a subset of us */
+ /* On legacy hierarchy, each of our child cpusets must be a subset of us */
ret = -EBUSY;
- cpuset_for_each_child(c, css, cur)
- if (!is_cpuset_subset(c, trial))
- goto out;
+ if (!is_in_v2_mode()) {
+ cpuset_for_each_child(c, css, cur)
+ if (!is_cpuset_subset(c, trial))
+ goto out;
+ }
/* Remaining checks don't apply to root cpuset */
ret = 0;
--
2.31.0
On Tue, Apr 13, 2021 at 11:02:33AM +0200, Odin Ugedal wrote:
> Two small validation relaxations for cgroup v2, making it easier to
> manage the hierarchy without the current pain points. Both changes
> work for both mems and cpus (but I have no NUMA machine to test mems).
>
> Hopefully the patches has an ok description about what change they
> provide, and why they are helpful.
I'm generally in favor of removing configuration constraints but let's hear
what Li thinks.
Thanks.
--
tejun