Re: [RFC, PATCH] CLONE_NEWIPC and exit_group()

["Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>]

[-- Attachment #1: Type: text/plain, Size: 1085 bytes --]

On Tue, Jun 26, 2012 at 05:04:57PM +0000, Serge E. Hallyn wrote:
> Quoting Kirill A. Shutemov (kirill.shutemov@linux.intel.com):
> > Hi,
> > 
> > Patch to move kern_unmount() out of exit_group() code path is below.
> > Dmitry, could you check if it's beneficial for your use-case?
> 
> Hi,
> 
> sorry, I don't seem to have the thread handy for contest.  What is the
> point of this?  The work being moved was not being done under lock,
> so what is this meant to gain?

It's basically addition to this patch (tested with the patch applied):

http://thread.gmane.org/gmane.linux.kernel.cifs/6347/focus=23929

Some context: Dmitry has workload which run a lot of short-living tasks in
sandboxed environment. He noticed that exit_group() syscall of the last
process in IPC namespace is a bottleneck.

The bottleneck was mainly due rcu_barrier() in kern_umount(). It's fixed
by patch in the link (Andrew took it in -mm).

But probably having kern_umount() in exit_group() code path is not a good
idea from scalability point of view?..

-- 
 Kirill A. Shutemov

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 836 bytes --]