[PATCH 5/6] Makes procs file writable to move all threads by tgid at once
menage at google.com
Fri Jul 24 14:57:36 PDT 2009
On Fri, Jul 24, 2009 at 2:52 PM, Benjamin Blum<bblum at google.com> wrote:
>> I don't see how css_sets are guaranteed while cgroup_fork_mutex is not held and
>> thus does not prevent new threads from being created right now. Could you
>> elaborate on that?
> Prefetching the css sets is independent of the fork lock/race issue.
> The idea is that we build a list, kept locally, that has references on
> all the css_sets we'll need to migrate each thread to the new cgroup.
> Since ordinarily we might need to malloc a new css_set for a thread
> before moving it, and it's possible that that could fail, we need to
> do allocations for all threads to be moved before committing any of
> them. As long as we have the list of prefetched css_sets, they'll stay
> there, and at the end, we drop the extra references we took on them to
> make that guarantee when tearing down the list.
And more specifically, since only the holder of cgroup_mutex can move
a thread to a new cgroup (and hence a potentially new unique css_set),
we know that once we've run through all the threads in the
thread_group and verified that we have the appropriate pre-fetched
css_set objects for all of them, it doesn't matter if any new threads
are created - they'll share one of the pre-fetched css_sets.
More information about the Containers