[PATCH 5/6] Makes procs file writable to move all threads by tgid at once

[Matt Helsley]

On Fri, Jul 24, 2009 at 09:01:46AM -0700, Paul Menage wrote:
> On Fri, Jul 24, 2009 at 8:50 AM, Matt Helsley<matthltc at us.ibm.com> wrote:
> >
> > There is much ado about not taking additional "global locks" in fork()
> > paths.
> >
> > * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
> > * (usually) take cgroup_mutex.  These are the two most performance
> > * critical pieces of code here.
> > ...
> >
> > and as I recall cgroup_fork() doesn't ever take cgroup_mutex because it is
> > so performance critical.
> 
> cgroup_mutex is a much bigger and heavier mutex than the new rwsem
> being introduced in this patch. It's sort of the BKL of cgroups,
> although where possible I'm encouraging use of finer-grained
> alternatives (such as subsystem-specific locks, the per-hierarchy
> lock, etc).
> 
> > Assuming the above comments in kernel/cgroup.c
> > are correct then this patch adds a performance regression by introducing a
> > global mutex in the fork path, doesn't it?
> 
> Yes, although to what magnitude isn't clear.

OK.

> Alternatives that we looked at were:
> 
> - add a clone_rwsem to task_struct, and require that a clone operation
> that's adding to the same thread group take a read lock on the
> leader's clone_rwsem; then the effect would be localised to a single
> process; but for a system that has one big multi-threaded server on
> it, the effect would still be similar to a global lock

Except most processes aren't big multi-threaded servers and not everyone
runs such processes. They'll experience the overhead of a global lock
when they don't have to. Again, it's a question of magnitudes we don't
know I think.

> 
> - move the housekeeping done by cgroup_fork() inside the tasklist_lock
> critical section in do_fork(); then cgroup_attach_proc() can rely on
> the existing global tasklist_lock to provide the necessary
> synchronization, rather than introducing a second global lock; the
> downside is that it slightly increases the size of the section where
> tasklist_lock is held for write.

Well, I imagine holding tasklist_lock is worse than cgroup_mutex in some
ways since it's used even more widely. Makes sense not to use it here..

Cheers,
	-Matt Helsley