Re: [PATCH v2 04/12] mm/filemap: Add mapping_seek_hole_data

[Matthew Wilcox]

On Tue, Sep 29, 2020 at 10:46:53AM +0200, Jan Kara wrote:
> On Mon 14-09-20 14:00:34, Matthew Wilcox (Oracle) wrote:
> > Rewrite shmem_seek_hole_data() and move it to filemap.c.
> > 
> > +	rcu_read_lock();
> > +	while ((page = xas_find_get_entry(&xas, max, XA_PRESENT))) {
> > +		loff_t pos = xas.xa_index * PAGE_SIZE;
> 
> OK, but for ordinary filesystems this could be problematic because of
> exceptional entries?

For ordinary filesystems, I have this queued up on top:

http://git.infradead.org/users/willy/pagecache.git/commitdiff/02c740b215bab901f95a560759b3bd906648da08

which handles exceptional entries.  It treats shadow/swap/DAX entries
the same -- there's definitely data there, it's just not in a struct
page right now.

> Also for shmem you've dropped the PageUptodate check which I'm not sure is
> safe?

That was unintentional.  I did run xfstests against this patch (just did
it again ... it passes), so I suspect it doesn't create a !Uptodate page.
I'll see if I can enhance the existing xfstests to catch this case.

The patch I link to above also doesn't handle !Uptodate pages on shmem
filesystems the same way that the current code does.  So ... on top of
this patch, I propose doing this:

@@ -2416,6 +2416,14 @@ generic_file_read_iter(struct kiocb *iocb, struct iov_ite
r *iter)
 }
 EXPORT_SYMBOL(generic_file_read_iter);
 
+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))
+               return seek_data ? start : end;
+       return seek_data ? end : start;
+}
+
 static inline
 unsigned int seek_page_size(struct xa_state *xas, struct page *page)
 {
@@ -2463,10 +2471,10 @@ loff_t mapping_seek_hole_data(struct address_space *mapping, loff_t start,
                        start = pos;
                }
 
-               if (seek_data)
+               pos += seek_page_size(&xas, page);
+               start = page_seek_hole_data(page, start, pos, seek_data);
+               if (start < pos)
                        goto unlock;
-
-               start = pos + seek_page_size(&xas, page);
        }
        rcu_read_unlock();
 

... and then rebasing the other patch on top of this works out nicely.

Here's the result:
http://git.infradead.org/users/willy/pagecache.git/commitdiff/9eb3f496b7cdcdcae83026e861e148f46921c367
http://git.infradead.org/users/willy/pagecache.git/commitdiff/7d93274088f0872d849a906d783dc260bee106b9