RFR: 8253435: Cgroup: 'stomping of _mount_path' crash if manually mounted cpusets exist [v2]

[Volker Simonis]

On Thu, 24 Sep 2020 19:16:51 GMT, Bob Vandette <bobv at openjdk.org> wrote:

> I was expecting to see some logic in this "else if" section that recorded the first occurance but did the validation on
> the second pass (cg_infos[CPUSET_IDX]._mount_path != NULL). When this situation is detected, we accept the mount with
> the /sys/fs/cgroup.  ```
>     } else if (strcmp(token, "cpuset") == 0) {
>       assert(cg_infos[CPUSET_IDX]._mount_path == NULL, "stomping of _mount_path");
>       cg_infos[CPUSET_IDX]._mount_path = os::strdup(tmpmount);
>       cg_infos[CPUSET_IDX]._root_mount_path = os::strdup(tmproot);
>       cg_infos[CPUSET_IDX]._data_complete = true;
> ```

You're right, the logic to ignore a `cpusets` entry should be moved into the `else if (strcmp(token, "cpuset") == 0)`
and I've done that in the new version of my patch.

However, the problem is that we don't know if the first or the second occurrence of `cpusets´ is *the right one*. I'm
also not sure if the Cgroup controllers are ALWAYS mounted under `/sys/fs/cgroup` (see my answer to Severin's
comments). If you can confirm that that's really the case, we could go with the simple solution proposed by Severin.

Otherwise, my current solution tries to be conservative and does not assume a predefined mount point for Cgroup
controllers. Instead it records the mount point of the first controller out of `cpu`, `cpuacct` and `memory` and checks
that all the other controllers from this set are mounted under the same location. In addition, it only accepts
`cpusets` if it is under the same mount point like the other controllers or under `/sys/fs/cgroup` if no other
controller has been seen before.

-------------

PR: https://git.openjdk.java.net/jdk/pull/295