Re: [PATCH] aio: propogate post-EIOCBQUEUED errors to completion event

[Trond Myklebust <trond.myklebust@fys.uio.no>]

On Tue, 2007-02-20 at 11:08 -0500, Chris Mason wrote:
> On Tue, Feb 20, 2007 at 11:01:50AM -0500, Trond Myklebust wrote:
> > On Mon, 2007-02-19 at 19:21 -0500, Benjamin LaHaise wrote:
> > > On Mon, Feb 19, 2007 at 04:50:48PM -0500, Chris Mason wrote:
> > > > aio is not responsible for this particular synchronization.  Those fixes
> > > > (if we make them) should come from other places.  The patch is important
> > > > to get aio error handling right.
> > > > 
> > > > I would argue that one common cause of the EIO is userland
> > > > error (mmap concurrent with O_DIRECT), and EIO is the correct answer.
> > > 
> > > I disagree.  That means that using the pagecache to synchronize things like 
> > > the proposed online defragmentation will occasionally make O_DIRECT users 
> > > fail.  O_DIRECT doesn't prevent the sysadmin from copying files or other 
> > > page cache uses, which implies that generating an error in these cases is 
> > > horrifically broken.  If only root could do it, I wouldn't complain, but 
> > > this would seem to imply that user vs root holes still exist.
> > 
> > We don't try to resolve "conflicting" writes between ordinary mmap() and
> > write(), so why should we be doing it for mmap and O_DIRECT?
> > 
> > mmap() is designed to violate the ordinary mutex locks for write(), so
> > if a conflict arises, whether it be with O_DIRECT or ordinary writes
> > then it is a case of "last writer wins".
> 
> There are some strange O_DIRECT corner cases in here such that the 'last
> writer' may actually be a 'last reader' and winning can mean have a copy
> of the page in page cache older than the copy on disk.

As long as it is marked dirty so that it eventually gets synced to disk,
it shouldn't matter.

> One option is to have invalidate_inode_pages2_range continue if it can't
> toss a page but still return something that O_DIRECT ignores (living
> with the race), but it looks like I can make a launder_page op that does
> the right thing.  I'll give it a shot.

I already sent in a patch to do that last week.

Cheers
  Trond