lkml.org 
[lkml]   [2009]   [Jan]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: 2.6.29 -mm merge plans
On Tue, Jan 06, 2009 at 06:24:18PM -0500, Christoph Hellwig wrote:
> On Tue, Jan 06, 2009 at 03:11:31PM -0800, Andrew Morton wrote:
> > > I'm not sure this is a good idea. Concurrent syncs are a bad idea
> > > to start with and we should just synchronyze do_sync completely.
> > > sync_filesystems as one of the main components of do_sync already
> > > is synchronized in that way, and taking that to a higher level would
> > > get rid of all the worries about concurrent syncs.
> >
> > Yes, single-threading sys_sync() would fix the problem which that patch
> > addresses.
> >
> > However there are a lot of performance and correctness issues around
> > sys_sync()-versus-fsync(), etc for which such a simple fix won't be
> > acceptable.
>
> fsync should really not much interac with sync at that level. While
> they both end up at same primitives at the lowest level those aren't
> the ones we're trying to protect against. I'm currently in the process
> of a major rework of sys_sync/do_sync to make it work properly for
> modern filesystems and the global synchronization was one of the first
> things I did..
>
> So if you have any workloads where that causes a problem please send
> them my way. Not that I can really thing of them, given the global
> nature of sys_sync I can't see any benefit of doing multiple of these
> in parallel.

I can't see a problem with putting a global mutex around sys_sync (almost
by definition, any app in the last 10+ years that calls sys_sync is not
performance critical).

But this patch fixes a correctness problem, so I think it is OK to go
upstream now.



\
 
 \ /
  Last update: 2009-01-07 02:17    [W:0.101 / U:0.028 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site