Re: [patch 12/12] mm: correctly synchronize rss-counters at exit/exec

[Linus Torvalds]

On Thu, Jun 7, 2012 at 6:05 PM, Markus Trippelsdorf
<markus@xxxxxxxxxxxxxxx> wrote:
>
> You've somehow merged the wrong patch.
> The correct version can be found here:
> http://marc.info/?l=linux-kernel&m=133848759505805

It looks like Andrew sent me a bad version.

However, that patch you point at isn't good *either*.

It does totally insane things in xacct_add_tsk(). You can't call
"sync_mm_rss(mm)" on somebody elses mm, yet that is exactly what it
does (and you can't pass in another thread pointer either, since the
whole point of the per-thread counters is that they don't have locking
and aren't atomic, so you can't read them from any other context than
"current").

The thing is, the *only* point where it makes sense to sync the rss
pointers is when you detach the mm from the current thread. And
possibly at "fork()" time, *before* you duplicate the "struct
task_struct" and pollute the new one with stale rss counter values
from the old one.

So doing sync_mm_rss() in xacct_add_tsk() is crazy. Doing it
*anywhere* where mm is not clearly "current->mm" is wrong. If there is
a "get_task_mm()" or similar nearby, it's wrong, it's crap, and it
shouldn't be done.

Oleg, please rescue me? Your patch looks much closer to sane, but it's
not quite there..

          Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/