On Tue, 2017-06-13 at 23:47 -0700, Christoph Hellwig wrote:

On Tue, Jun 13, 2017 at 06:24:32AM -0400, Jeff Layton wrote:

quoted

That's definitely what I want for the endgame here. My plan was to add
this flag for now, and then eventually reverse it (or drop it) once all
or most filesystems are converted.

We can do it that way from the get-go if you like. It'll mean tossing in
 a patch add this flag to all filesystems that have an fsync operation
and that use the pagecache, and then gradually remove it from them as we
convert them.

Which method do you prefer?

Please do it from the get-go.  Or in fact figure out if we can get
away without it entirely.  Moving the error reporting into ->fsync
should help greatly with that, so what's missing after that?

In this smaller set, it's only really used for DAX. In the larger patch
series I have (which needs updating on top of this), there are other
things that key off of it:

sync_file_range: ->fsync isn't called directly there, and I think we
probably want similar semantics to fsync() for it

JBD2: will try to re-set the error after clearing it with
filemap_fdatawait. That's problematic with the new infrastructure so we
need some way to avoid it.

What I think I'll do for now is add a new FS_DAX_WB_ERRSEQ flag that
will go away by the end of the series. As the need arises for a similar
flag in other areas, I'll add them then.

The overall goal is not to need these flags. It may take a bit of time
to get there though.

Thanks for the review so far!
-- 
Jeff Layton [off-list ref]