lkml.org 
[lkml]   [2007]   [Oct]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] task containersv11 add tasks file interface fix for cpusets
David wrote:
> It would probably be better to just save references to the tasks.
>
> struct cgroup_iter it;
> struct task_struct *p, **tasks;
> int i = 0;
>
> cgroup_iter_start(cs->css.cgroup, &it);
> while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
> get_task_struct(p);
> tasks[i++] = p;
> }
> cgroup_iter_end(cs->css.cgroup, &it);

Hmmm ... guess I'd have to loop over the cgroup twice, once to count
them (the 'count' field is not claimed to be accurate) and then again,
after I've kmalloc'd the tasks[] array, filling in the tasks[] array.

On a big cgroup on a big system, this could easily be thousands of
iteration loops.

And I've have to drop the css_set_lock spinlock between the two loops,
since I can't hold a spinlock while calling kmalloc.

So then I'd have to be prepared for the possibility that the second
loop found more cgroups on the list than what I counted in the
first loop.

This is doable ... indeed I've done such before, in the code that
is now known as kernel/cgroup.c:cgroup_tasks_open(). Look for how
pidarray[] is setup.

And note that that code doesn't deal with the case that more cgroups
showed up after they were counted. When supporting the reading of the
'tasks' file by user code, this is ok - it's inherently racey anyway -
so not worth trying too hard just to close the window part way.

If I need to close the window all the way, completely solving the race
condition, then I have the code in kernel/cpuset.c:update_nodemask(),
which builds an mmarray[] using two loops and some retries if newly
forked tasks are showing up too rapidly at the same time. The first of
the two loops is hidden in the cgroup_task_count() call.

That's a bunch of code, mate. If some other solution was adequate
(no worse than the current situation, which forces user space to
rewrite every pid in the tasks file back to itself if they want
a 'cpus' change to actually be applied) but took much less code,
then I'd have to give it serious consideration, as I did before.

I don't mind a bunch of code, but kernel text has to earn its keep.
I'm not yet convinced that the above page or two of somewhat fussy code
(see again the code in kernel/cpuset.c:update_nodemask() ...) has
sufficient real user value per byte of kernel text space to justify its
existence.

... by the way ... tell me again why css_set_lock is a spinlock?

I didn't think it was such a good idea to hold a spinlock while
iterating over a major list, doing lord knows what (the loops
over cgroup_iter_next() do user provided code, as in this case.)
Shouldn't that be a mutex?

Or, if there is a good reason that must remain a spinlock, then the
smallest amount of new code, and the easiest code to write, would
perhaps be adding another cgroup callback, called only by cgroup attach
() requests back to the same group. Then code that wants to do
something odd, such as cpusets, for what seems like a no-op, can do so.

--
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <pj@sgi.com> 1.925.600.0401
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2007-10-06 22:01    [from the cache]
©2003-2014 Jasper Spaans. Advertise on this site