lkml.org 
[lkml]   [2017]   [Apr]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v3 10/20] fuse: set mapping error in writepage_locked when it fails
On Tue 25-04-17 06:35:13, Jeff Layton wrote:
> On Tue, 2017-04-25 at 10:17 +0200, Jan Kara wrote:
> > On Mon 24-04-17 13:14:36, Jeff Layton wrote:
> > > On Mon, 2017-04-24 at 18:04 +0200, Jan Kara wrote:
> > > > On Mon 24-04-17 09:22:49, Jeff Layton wrote:
> > > > > This ensures that we see errors on fsync when writeback fails.
> > > > >
> > > > > Signed-off-by: Jeff Layton <jlayton@redhat.com>
> > > >
> > > > Hum, but do we really want to clobber mapping errors with temporary stuff
> > > > like ENOMEM? Or do you want to handle that in mapping_set_error?
> > > >
> > >
> > > Right now we don't really have such a thing as temporary errors in the
> > > writeback codepath. If you return an error here, the data doesn't stay
> > > dirty or anything, and I think we want to ensure that that gets reported
> > > via fsync.
> > >
> > > I'd like to see us add better handling for retryable errors for stuff
> > > like ENOMEM or EAGAIN. I think this is the first step toward that
> > > though. Once we have more consistent handling of writeback errors in
> > > general, then we can start doing more interesting things with retryable
> > > errors.
> > >
> > > So yeah, I this is the right thing to do for now.
> >
> > OK, fair enough. And question number 2):
> >
> > Who is actually responsible for setting the error in the mapping when error
> > happens inside ->writepage()? Is it the ->writepage() callback or the
> > caller of ->writepage()? Or something else? Currently it seems to be a
> > strange mix (e.g. mm/page-writeback.c: __writepage() calls
> > mapping_set_error() when ->writepage() returns error) so I'd like to
> > understand what's the plan and have that recorded in the changelogs.
> >
>
> That's an excellent question.
>
> I think we probably want the writepage/launder_page operations to call
> mapping_set_error. That makes it possible for filesystems (e.g. NFS) to
> handle their own error tracking and reporting without using the new
> infrastructure. If they never call mapping_set_error then we'll always
> just return whatever their ->fsync operation returns on an fsync.

OK, makes sense. It is also in line with what you did for DAX, 9p, or here
for FUSE. So feel free to add:

Reviewed-by: Jan Kara <jack@suse.cz>

for this patch but please also add a sentense that ->writepage() is
responsible for calling mapping_set_error() if it fails and page is not
redirtied to the changelogs of patches changing writepage handlers.

> I'll make another pass through the tree and see whether we have some
> mapping_set_error calls that should be removed, and will flesh out
> vfs.txt to state this. Maybe that file needs a whole section on
> writeback error reporting? Hmmm...

I think it would be nice to have all the logic described in one place. So
+1 from me.

> That probably also means that I should drop patch 8 from this series
> (mm: ensure that we set mapping error if writeout fails), since that
> should be happening in writepage already.

Yes.

Honza
--
Jan Kara <jack@suse.com>
SUSE Labs, CR

\
 
 \ /
  Last update: 2017-04-25 13:20    [W:0.088 / U:0.720 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site