lkml.org 
[lkml]   [2006]   [Apr]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch 1/1] sys_sync_file_range()
On Mon, Apr 03 2006, Andrew Morton wrote:
> Jens Axboe <axboe@suse.de> wrote:
> >
> > On Mon, Apr 03 2006, Neil Brown wrote:
> > > On Friday March 31, nathans@sgi.com wrote:
> > > > On Thu, Mar 30, 2006 at 06:58:46PM +1100, Neil Brown wrote:
> > > > > On Wednesday March 29, akpm@osdl.org wrote:
> > > > > > Remove the recently-added LINUX_FADV_ASYNC_WRITE and LINUX_FADV_WRITE_WAIT
> > > > > > fadvise() additions, do it in a new sys_sync_file_range() syscall
> > > > > > instead.
> > > > >
> > > > > Hmmm... any chance this could be split into a sys_sync_file_range and
> > > > > a vfs_sync_file_range which takes a 'struct file*' and does less (or
> > > > > no) sanity checking, so I can call it from nfsd?
> > > > >
> > > > > Currently I implement COMMIT (which has a range) with a by messing
> > > > > around with filemap_fdatawrite and filemap_fdatawait (ignoring the
> > > > > range) and I'd rather than a vfs helper.
> > > >
> > > > I'm not 100% sure, but it looks like the PF_SYNCWRITE process flag
> > > > should be set on the nfsd's while they're doing that, which doesn't
> > > > seem to be happening atm. Looks like a couple of the IO schedulers
> > > > will make use of that knowledge now. All the more reason for a VFS
> > > > helper here I guess. ;)
> > >
> > > PF_SYNCWRITE? What's that???
> > >
> > > (find | xargs grep ...)
> > > Oh. The block device schedulers like to know if a request is sync or
> > > async (and all reads are assumed to be sync) - which is reasonable -
> > > and so have a per-task flag to tell them - which isn't (IMO).
> > >
> > > md/raid (particularly raid5) often does the write from a different
> > > process than generated the original request, so that will break
> > > completely.
> >
> > I don't think any disagrees with you, the sync-write process flag is
> > indeed an atrocious beast...
>
> Yeah. PF_SYNCWRITE was a performance tweak for the anticipatory scheduler.
> As cfq is using it as well now (hopefully to good effect) I guess it could
> be formalised more.

Yup, both 'as' and 'cfq' would prefer to just look at a SYNC bio flag
instead. But the logic itself is definitely needed.

> > > What is wrong with a bio flag I wonder....
> >
> > Nothing, in fact I would love for it to be changed. I'm sure such a
> > patch would be accepted with open arms! :-)
>
> I think once someone starts coding it, they'll become a big fan of
> PF_SYNCWRITE...

They might not become a big fan, but they'll surely appreciate the
simplicity of it :-)

> For the page writeback functions it's probably possible to use
> writeback_control.sync_mode=WB_SYNC_ALL as a trigger, propagate that into
> the IO layer. Maybe that'll always be sufficient - it's hard to tell. The
> writeback paths are twisty and deep...
>
> Then again, using WB_SYNC_ALL as a hint that this process will be waiting
> for this writeout to complete is a bit hacky too - it doesn't _really_ mean
> that - it just means that I/O should be _started_ against all relevant
> dirty data.
>
> Good luck ;)

It's not a hard problem, but it will definitely cost a little sweat to
go through. I'm sure Neil could pull it off, the question is more if he
wants to :-)

--
Jens Axboe

-
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/

\
 
 \ /
  Last update: 2006-04-03 16:33    [from the cache]
©2003-2011 Jasper Spaans