On Sat, 2017-06-24 at 09:16 -0400, Jeff Layton wrote:
> On Sat, 2017-06-24 at 04:59 -0700, Christoph Hellwig wrote:
> > On Tue, Jun 20, 2017 at 01:44:44PM -0400, Jeff Layton wrote:
> > > In order to query for errors with errseq_t, you need a previously-
> > > sampled point from which to check. When you call
> > > filemap_write_and_wait_range though you don't have a struct file and so
> > > no previously-sampled value.
> >
> > So can we simply introduce variants of them that take a struct file?
> > That would be:
> >
> > a) less churn
> > b) less code
> > c) less chance to get data integrity wrong
>
> Yeah, I had that thought after I sent the reply to you earlier.
>
> The main reason I didn't do that before was that I had myself convinced
> that we needed to do the check_and_advance as late as possible in the
> fsync process, after the metadata had been written.
>
> Now that I think about it more, I think you're probably correct. As long
> as we do the check and advance at some point after doing the
> write_and_wait, we're fine here and shouldn't violate exactly once
> semantics on the fsync return.
So I have a file_write_and_wait_range now that should DTRT for this
patch.
The bigger question is -- what about more complex filesystems like
ext4? There are a couple of cases where we can return -EIO or -EROFS on
fsync before filemap_write_and_wait_range is ever called. Like this one
for instance:
if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
return -EIO;
...and the EXT4_MF_FS_ABORTED case.
Are those conditions ever recoverable, such that a later fsync could
succeed? IOW, could I do a remount or something such that the existing
fds are left open and become usable again?
If so, then we really ought to advance the errseq_t in the file when we
catch those cases as well. If we have to do that, then it probably makes
sense to leave the ext4 patch as-is.
--
Jeff Layton <jlayton@xxxxxxxxxx>
