Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1200659 > unrolled thread

Re: [PATCH] user_ns: use correct check for single-threadedness

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: [PATCH] user_ns: use correct check for single-threadedness Ingo Molnar <mingo@kernel.org> - 2015-08-05 13:40 +0200
    Re: [PATCH] user_ns: use correct check for single-threadedness "Kirill A. Shutemov" <kirill@shutemov.name> - 2015-08-05 14:00 +0200
      Re: [PATCH] user_ns: use correct check for single-threadedness Ricky Zhou <ricky@rzhou.org> - 2015-08-05 15:20 +0200

#1200659 — Re: [PATCH] user_ns: use correct check for single-threadedness

* Kirill A. Shutemov <kirill@shutemov.name> wrote:

> On Tue, Jul 28, 2015 at 02:35:04PM -0700, Andrew Morton wrote:
> > On Tue, 28 Jul 2015 10:15:00 -0700 Kees Cook <keescook@chromium.org> wrote:
> > 
> > > From: Ricky Zhou <rickyz@chromium.org>
> > > 
> > > Checking mm_users > 1 does not mean a process is multithreaded. For
> > > example, reading /proc/PID/maps temporarily increments mm_users, allowing
> > > other processes to (accidentally) interfere with unshare() calls.
> > > 
> > > This fixes observed failures of unshare(CLONE_NEWUSER) incorrectly
> > > returning EINVAL if another processes happened to be simultaneously
> > > reading the maps file.
> > 
> > Yikes.  current_is_single_threaded() is expensive.  Are we sure this
> > isn't going to kill someone's workload?
> 
> It's expensive only if mm_users > 1. We will go to for_each_process() only
> if somebody outside of the process grabs mm_users references (like reading
> /proc/PID/maps). Or if it called it from multithreaded application.

It's considerably expensive:

        rcu_read_lock();
        for_each_process(p) {
                do {
	...
                } while_each_thread(p, t);
        }


'only' if it's multi-threaded, i.e. when some workload cares so much about 
performance that it uses multiple threads?

Can you see the contradiction there?

Thanks,

	Ingo
--
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/

[toc] | [next] | [standalone]

#1200665

On Wed, Aug 05, 2015 at 01:38:54PM +0200, Ingo Molnar wrote:
> 
> * Kirill A. Shutemov <kirill@shutemov.name> wrote:
> 
> > On Tue, Jul 28, 2015 at 02:35:04PM -0700, Andrew Morton wrote:
> > > On Tue, 28 Jul 2015 10:15:00 -0700 Kees Cook <keescook@chromium.org> wrote:
> > > 
> > > > From: Ricky Zhou <rickyz@chromium.org>
> > > > 
> > > > Checking mm_users > 1 does not mean a process is multithreaded. For
> > > > example, reading /proc/PID/maps temporarily increments mm_users, allowing
> > > > other processes to (accidentally) interfere with unshare() calls.
> > > > 
> > > > This fixes observed failures of unshare(CLONE_NEWUSER) incorrectly
> > > > returning EINVAL if another processes happened to be simultaneously
> > > > reading the maps file.
> > > 
> > > Yikes.  current_is_single_threaded() is expensive.  Are we sure this
> > > isn't going to kill someone's workload?
> > 
> > It's expensive only if mm_users > 1. We will go to for_each_process() only
> > if somebody outside of the process grabs mm_users references (like reading
> > /proc/PID/maps). Or if it called it from multithreaded application.
> 
> It's considerably expensive:
> 
>         rcu_read_lock();
>         for_each_process(p) {
>                 do {
> 	...
>                 } while_each_thread(p, t);
>         }
> 
> 
> 'only' if it's multi-threaded, i.e. when some workload cares so much about 
> performance that it uses multiple threads?
> 
> Can you see the contradiction there?

I can. man 2 unshare:

	CLONE_NEWUSER requires that the calling process  is  not  threaded;

The workload cares so much about performance that it ignores API
requirements. Some slow down looks like a fair price to me.

-- 
 Kirill A. Shutemov
--
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/

[toc] | [prev] | [next] | [standalone]

#1200717

On Wed, Aug 5, 2015 at 4:53 AM, Kirill A. Shutemov <kirill@shutemov.name> wrote:
>> 'only' if it's multi-threaded, i.e. when some workload cares so much about
>> performance that it uses multiple threads?
>>
>> Can you see the contradiction there?
>
> I can. man 2 unshare:
>
>         CLONE_NEWUSER requires that the calling process  is  not  threaded;
>
> The workload cares so much about performance that it ignores API
> requirements. Some slow down looks like a fair price to me.
To be fair, the entire reason for this patch is that the slow path
(mm_users > 1) can happen even when the process is single-threaded. I
was concerned about how expensive current_is_single_threaded looked as
well, but didn't see any lightweight alternatives short of adding a
field to mm_struct. Do folks think it's worth going down that route
instead?

Thanks,
Ricky
--
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/

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web