Re: [RESEND] [PATCH] lseek: remove i_mutex

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 06 Feb 2009 09:20:30 +0900
Hisashi Hifumi <hifumi.hisashi@oss.ntt.co.jp> wrote:

> 
> At 05:05 09/02/06, Andrew Morton wrote:
> >On Tue, 03 Feb 2009 17:04:40 +0900
> >Hisashi Hifumi <hifumi.hisashi@oss.ntt.co.jp> wrote:
> >
> >> I removed i_mutex from generic_file_llseek.
> >> I think that the reason of protecting lseek with i_mutex is just
> >> touching i_size atomically.
> >> 
> >> So I introduce i_size_read here so i_mutex is no longer needed.
> >> 
> >> Following patch removes i_mutex from generic_file_llseek, and deletes 
> >> generic_file_llseek_nolock totally.
> >> 
> >> Currently there is i_mutex contention not only around lseek, but also 
> >fsync or write.
> >> So,  I think we can mitigate i_mutex contention between fsync lseek and 
> >write by
> >> removing i_mutex.
> >
> >Prior to this change, generic_file_llseek() modified file->f_pos
> >atomically with respect to other i_mutex holders.
> >
> >After this change, it doesn't.
> 
> Hi Andrew.
> 
> Even before this change is applied, file->f_pos access is not atomic.
> sys_read change f_pos value through file_pos_write without i_mutex.

I know.  That's why I specified "with respect to other i_mutex holders".

This patch makes things worse.

At very very minimum the changelog should explain that this patch makes
things worse, and demonstrate why this is justifiable.

> I think seqlock is needed to make f_pos access atomic.

Maybe.  Or atomic64_t, or spinlocking, or i_mutex, or something else.