Re: [PATCH v3 04/12] mm/filemap: Add mapping_seek_hole_data

From: Jan Kara
Date: Mon Oct 26 2020 - 10:49:45 EST


On Mon 26-10-20 12:17:27, Matthew Wilcox wrote:
> On Mon, Oct 26, 2020 at 11:48:06AM +0100, Jan Kara wrote:
> > > +static inline loff_t page_seek_hole_data(struct page *page,
> > > + loff_t start, loff_t end, bool seek_data)
> > > +{
> > > + if (xa_is_value(page) || PageUptodate(page))
> >
> > Please add a comment here that this is currently tmpfs specific treating
> > exceptional entries as swapped out pages and thus data. It took me quite a
> > while to figure this out. You can remove the comment later when it is no
> > longer true...
>
> But it's not tmpfs specific. If the value entry is a DAX entry, there's
> data here, and if the value entry is a shadow entry, there's data here
> too. Not that it should be called for either of those cases because the
> filesystem should know, but a value entry always means there's data here.

Good point but for shadow entries I'm not convinced - we do have page cache
pages instantiated by reads from holes. When they get evicted, we have a
shadow entry there but it is still a hole. Actually, similarly we can have
zeroed page over an unwritten extent and that should still count as a hole
IMO. Only once the page becomes dirty, it should be treated as data. This
looks like a bug even in the current page_seek_hole_data() helper:

# fallocate -l 4096 testfile
# xfs_io -x -c "seek -h 0" testfile
Whence Result
HOLE 0
# dd if=testfile bs=4096 count=1 of=/dev/null
# xfs_io -x -c "seek -h 0" testfile
Whence Result
HOLE 4096

Which is indeed a bit weird result... But we seem to be pretty consistent
in this behavior for quite some time. I'll send an email to fs folks about
this.

> > > + return seek_data ? start : end;
> > > + return seek_data ? end : start;
> > > +}
> > > +
> > > +static inline
> > > +unsigned int seek_page_size(struct xa_state *xas, struct page *page)
> > > +{
> > > + if (xa_is_value(page))
> > > + return PAGE_SIZE << xa_get_order(xas->xa, xas->xa_index);
> > > + return thp_size(page);
> > > +}
> > > +
> > > +/**
> > > + * mapping_seek_hole_data - Seek for SEEK_DATA / SEEK_HOLE in the page cache.
> > > + * @mapping: Address space to search.
> > > + * @start: First byte to consider.
> > > + * @end: Limit of search (exclusive).
> > > + * @whence: Either SEEK_HOLE or SEEK_DATA.
> > > + *
> > > + * If the page cache knows which blocks contain holes and which blocks
> > > + * contain data, your filesystem can use this function to implement
> > > + * SEEK_HOLE and SEEK_DATA. This is useful for filesystems which are
> > > + * entirely memory-based such as tmpfs, and filesystems which support
> > > + * unwritten extents.
> > > + *
> > > + * Return: The requested offset on successs, or -ENXIO if @whence specifies
> > > + * SEEK_DATA and there is no data after @start. There is an implicit hole
> > > + * after @end - 1, so SEEK_HOLE returns @end if all the bytes between @start
> > > + * and @end contain data.
> > > + */
> > > +loff_t mapping_seek_hole_data(struct address_space *mapping, loff_t start,
> > > + loff_t end, int whence)
> > > +{
> > > + XA_STATE(xas, &mapping->i_pages, start >> PAGE_SHIFT);
> > > + pgoff_t max = (end - 1) / PAGE_SIZE;
> > > + bool seek_data = (whence == SEEK_DATA);
> > > + struct page *page;
> > > +
> > > + if (end <= start)
> > > + return -ENXIO;
> > > +
> > > + rcu_read_lock();
> > > + while ((page = xas_find_get_entry(&xas, max, XA_PRESENT))) {
> > > + loff_t pos = xas.xa_index * PAGE_SIZE;
> > > +
> > > + if (start < pos) {
> > > + if (!seek_data)
> > > + goto unlock;
> > > + start = pos;
> > > + }
> > > +
> > > + pos += seek_page_size(&xas, page);
> > > + start = page_seek_hole_data(page, start, pos, seek_data);
> > > + if (start < pos)
> > > + goto unlock;
> >
> > Uh, I was staring at this function for half an hour but I still couldn't
> > convince myself that it is correct in all the corner cases. Maybe I'm dumb
> > but I'd wish this was more intuitive (and I have to say that the original
> > tmpfs function is much more obviously correct to me). It would more
> > understandable for me if we had a code like:
> >
> > if (page_seek_match(page, seek_data))
> > goto unlock;
> >
> > which would be just the condition in page_seek_hole_data(). Honestly at the
> > moment I fail to see why you bother with 'pos' in the above four lines at
> > all.
>
> So this?
>
> static bool page_seek_match(struct page *page, bool seek_data)
> {
> /* Swap, shadow & DAX entries all represent data */
> if (xa_is_value(page) || PageUptodate(page))
> return seek_data;
> return !seek_data;
> }
>
> ...
>
> if (page_seek_match(page, seek_data))
> goto unlock;
> start = pos + seek_page_size(&xas, page);
>
> The function makes more sense when page_seek_hole_data() gains the
> ability to look at sub-page uptodate status and it needs to return
> where in the page the data (or hole) starts. But that can be delayed
> for the later patch.

Yeah, this looks much more comprehensible for me. Thanks!

> With those changes,
>
> Ran: generic/285 generic/286 generic/436 generic/445 generic/448 generic/490 generic/539
> Passed all 7 tests
>
> > BTW I suspect that this loop forgets to release the page reference it has got
> > when doing SEEK_HOLE.
>
> You're right. I need a put_page() at the end of the loop. Also true
> for the case where we find a !Uptodate page when doing SEEK_DATA.

Right.

Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR