lkml.org 
[lkml]   [2007]   [Dec]   [20]   [last100]   RSS Feed
Views: [more markup]  [less markup]  [headers]  [forward] 
 
Messages in this thread
/
FromNick Piggin <>
SubjectRe: [PATCH 24/28] AFS: Add a function to excise a rejected write from the pagecache [try #2]
DateFri, 21 Dec 2007 12:11:14 +1100
On Friday 21 December 2007 05:49, David Howells wrote:
> Nick Piggin <nickpiggin@yahoo.com.au> wrote:
> > > Nick Piggin <nickpiggin@yahoo.com.au> wrote:
> > > > This reintroduces the fault vs truncate race window, which must be
> > > > fixed.
> > >
> > > Hmmm...  perhaps.
> >
> > What do you mean by perhaps?
>
> I mean 'perhaps'.  I'm not sure I remember what the race was, so I can't
> evaluate whether or not the same race crops up in AFS too.  So: can you
> describe the race please.

It's in the changelogs.


> > No, you could do writeback caching but disallow read of dirty data.
>
> Someone might already have read-access via mmap at the point someone
> attempts to write to an mmapped region.  That means that I'd have to revoke
> the read access of the first someone before letting the write take place.
>
> Does NFS do this?

I don't know. But yeah I guess it's tricky. Still, if I was writing a
filesystem, I'd focus on behavioural niceness/correctness above
performance, but maybe that's not realistic.


> > > > But otherwise I guess if you really want to discard the dirty data
> > > > after a failed writeback attempt, what's wrong with just
> > > > invalidate_inode_pages2?
> > >
> > > Erm...  Because it deadlocks?
> >
> > Why don't you call it after calling end_page_writeback?
>
> Because then there can be a race over who gets to flush the dead write.

You should solve it in your filesystem.


> Actually, this may no longer be a problem with your write_begin() changes.
> I'll need to have another look at those.
>
> Besides, I don't agree that invalidate_inode_pages2() is necessarily the
> right way to do things.

Why?


\
 
 \ /
  Last update: 2007-12-21 02:15    [from the cache]
©2003-2008