Re: [PATCH] clear PageError bit in msync & fsync

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

On Tue, 09 Nov 2010 16:15:52 -0500
Rik van Riel <riel@redhat.com> wrote:

> On 11/09/2010 04:07 PM, Ted Ts'o wrote:
> > On Tue, Nov 09, 2010 at 02:33:29PM -0500, Rik van Riel wrote:
> >>
> >> There are essentially two possibilities:
> >> 1) the VM can potentially be filled up with uncleanable dirty pages, or
> >> 2) pages that hit an IO error are left in a clean state, so they can
> >>     be reclaimed under memory pressure
> >>
> >> Alternative 1 could cause the entire system to deadlock, while
> >> option 2 puts the onus on userland apps to rewrite the data
> >> from a failed msync/fsync.
> >>
> >> Currently the VM has behaviour #2 which is preserved with my
> >> patch.
> >>
> >> The only difference with my patch is, we won't keep returning
> >> -EIO on subsequent, error free, msync or fsync calls to files
> >> that had an IO error at some previous point in the past.
> >
> > Do we guarantee that the application will get EIO at least once?  I
> > thought there were issues where the error bit could get lost if the
> > page writeback was triggered by sync() run by a third-party
> > application.
> 
> There is no such guarantee in the current kernel, either
> with or without my patch.
> 
> A third application calling fsync or msync can get the
> EIO cleared, so the application that did the write does
> not see it.

yup.  It's a userspace bug, really.  Although that bug might be
expressed as "userspace didn't know about linux-specific EIO
behaviour".

> The VM could also reclaim the PageError page due to
> memory pressure, so the application calling fsync or
> msync does not see it.

That would be a kernel bug, methinks.  The page's end_io handler should
set the address_space's AS_EIO flag (see mpage_end_io_write()), to be
later returned to (and cleared by) the fsync/msync caller.

It wouldn't surprise me if lots of end_io handlers got that wrong.

> I see no good way in which we could guarantee that
> every process calling msync or fsync on a file that
> had an IO error in the past gets EIO once - at least,
> not without every one of them always getting EIO on
> the file even after the IO path is good again...

Yes, there's no obviously good design here.

And a lot of these problems also apply to ENOSPC, and an ENOSPC
condition most certainly does magically fix itself up in real time...