On Fri, Nov 09, 2018 at 03:42:51PM -0800, Andrew Morton wrote:
> On Mon, 5 Nov 2018 11:36:13 -0500 Brian Foster <bfoster@xxxxxxxxxx> wrote:
>
> > write_cache_pages() currently breaks out of the writepage loop in
> > the event of a ->writepage() error. This causes problems for
> > integrity writeback on XFS
>
> For the uninitiated, please define the term "integrity writeback".
> Quite carefully ;) I'm not sure what it actually means. grepping
> fs/xfs for "integrity" doesn't reveal anything.
>
> <reads the code>
>
> OK, it appears the term means "to sync data to disk" as opposed to
> "periodic dirty memory cleaning". I guess we don't have particularly
> well-established terms for the two concepts.
>
Indeed. The intent is basically to describe any writeback that is
intended to persist data and so so before returning (i.e., fsync(),
etc.). That was the best term I came across to describe it ("integrity
sync" is used in some of the existing comments), but I can try to be
more descriptive in the commit log.
> > in the event of a persistent error as XFS
> > expects to process every dirty+delalloc page such that it can
> > discard delalloc blocks when real block allocation fails. Failure
> > to handle all delalloc pages leaves the filesystem in an
> > inconsistent state if the integrity writeback happens to be due to
> > an unmount, for example.
> >
> > Update write_cache_pages() to continue processing pages for
> > integrity writeback regardless of ->writepage() errors. Save the
> > first encountered error and return it once complete. This
> > facilitates XFS or any other fs that expects integrity writeback to
> > process the entire set of dirty pages regardless of errors.
> > Background writeback continues to exit on the first error
> > encountered.
> >
> > ...
> >
> > --- a/mm/page-writeback.c
> > +++ b/mm/page-writeback.c
> > @@ -2156,6 +2156,7 @@ int write_cache_pages(struct address_space *mapping,
> > {
> > int ret = 0;
> > int done = 0;
> > + int error;
> > struct pagevec pvec;
> > int nr_pages;
> > pgoff_t uninitialized_var(writeback_index);
> > @@ -2236,25 +2237,29 @@ int write_cache_pages(struct address_space *mapping,
> > goto continue_unlock;
> >
> > trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
> > - ret = (*writepage)(page, wbc, data);
> > - if (unlikely(ret)) {
> > - if (ret == AOP_WRITEPAGE_ACTIVATE) {
> > + error = (*writepage)(page, wbc, data);
> > + if (unlikely(error)) {
> > + if (error == AOP_WRITEPAGE_ACTIVATE) {
> > unlock_page(page);
> > - ret = 0;
> > - } else {
> > + error = 0;
> > + } else if (wbc->sync_mode != WB_SYNC_ALL &&
> > + !wbc->for_sync) {
>
> And here we're determining that it is not a sync-data-to-disk
> operation, hence it must be a clean-dirty-pages operation.
>
> This isn't very well-controlled, is it? It's an inference which was
> put together by examining current callers, I assume?
>
Yeah, sort of. Some of the comments do already explain how WB_SYNC_ALL
refers to "integrity" writeback/sync (above write_cache_pages(), for
example). The ->for_sync thing is more of an inference based on its use
in __writeback_single_inode() and the comment where it is defined.
> It would be good if we could force callers to be explicit about their
> intent here. But I'm not sure that adding a new writeback_sync_mode is
> the way to do this.
>
The more I look at it, however, I think I could probably drop the
for_sync bit here. It appears only be used for a special case of
WB_SYNC_ALL that isn't relevant to this patch, so it only serves to
complicate in this context.
I'm not sure if you had more in mind beyond that..? There are a lot of
knobs on the wbc in general. It might be interesting to see if some of
that could be cleaned up to factor out some of those seemingly bolt-on
knobs, but I'd have to stare at the code more and think about it. Even
still, any such change is probably better as a follow on patch to this
one, which is intended to be an isolated bug fix. Thoughts on any of
that is appreciated.
> At a minimum it would be good to have careful comments in here
> explaining what is going on, justifying the above inference, explaining
> the xfs requirement (hopefully in a way which isn't xfs-specific).
>
Ok, I can add a comment here that covers such details. Thanks for the
feedback.
Brian
> > /*
> > - * done_index is set past this page,
> > - * so media errors will not choke
> > + * done_index is set past this page, so
> > + * media errors will not choke
> > * background writeout for the entire
> > * file. This has consequences for
> > * range_cyclic semantics (ie. it may
> > * not be suitable for data integrity
> > * writeout).
> > */
>
