CA-416516: vm.slice/cgroup.procs write operation gets EBUSY #6650
Conversation
see cgroups v2 "no internal processes" rule if cgroup.subtree_control is not empty, and we attach a pid to cgroup.procs, kernel would return EBUSY Signed-off-by: Chunjie Zhu <chunjie.zhu@cloud.com>
|
add @rosslagerwall @MarkSymsCtx @DeliZhangX @stephenchengCloud @liulinC to review |
|
It looks like there is something wrong with my account. I cannot add reviewers. |
|
Thanks @BengangY . |
| # into cgroup.procs, kernel would return EBUSY | ||
| cgroup_slice_dir = os.path.join("/sys/fs/cgroup", cgroup_slice) | ||
| qemu_dm_dir = os.path.join(cgroup_slice_dir, "qemu-dm") | ||
| if not os.path.exists(qemu_dm_dir): |
There was a problem hiding this comment.
os.makedirs(qemu_dm_dir, exist_ok=True) ? https://www.w3schools.com/python/ref_os_makedirs.asp
| # if cgroup.subtree_control is not empty, and we attach a pid | ||
| # into cgroup.procs, kernel would return EBUSY | ||
| cgroup_slice_dir = os.path.join("/sys/fs/cgroup", cgroup_slice) | ||
| qemu_dm_dir = os.path.join(cgroup_slice_dir, "qemu-dm") |
There was a problem hiding this comment.
can be one line: os.path.join("/sys/fs/cgroup", cgroup_slice，"qemu-dm")
There was a problem hiding this comment.
May consider to create a common sub slice e.g. "vm.slice/vm_common.slice" for attaching any existing processes including https://github.com/xapi-project/sm/blob/master/libs/sm/blktap2.py#L765
There was a problem hiding this comment.
For XS9 we now have a new enough systemd that we can create per domain slices.
There was a problem hiding this comment.
@MarkSymsCtx I tried your idea. I cannot implement your idea.
Option 1. Let systemd manages qemu-dm process automatically. It does not work. Qemu-dm is not same as Qemu-dp, the startup arguments of qemu-dm is complicated, and qemu-wrapper does some xenstore operation. We cannot prepare a qemu-dm@.service configuration which can replace qemu-wrapper.
Option 2. In qemu-wrapper, I create /sys/fs/cgroup/vm.slice/qemu-dm-$pid for each qemu-dm process, this is ok. Then I have to add xenopsd hook scripts to remove the folder when vm is destroyed. In vm-pre-shutdown script, cgdelete cannot delete the qemu-dm-$pid folder. In vm-post-destroy, it does not have domid parameter, we cannot get qemu-dm process id, so we cannot delete the qemu-dm-$pid folder.
I will merge my commit.
6 participants
see cgroups v2 "no internal processes" rule
if cgroup.subtree_control is not empty, and we attach a pid
to cgroup.procs, kernel would return EBUSY