Re: generic/068 crash on 5.18-rc2?

[Matthew Wilcox <willy@infradead.org>]

On Mon, May 02, 2022 at 08:25:34PM -0700, Darrick J. Wong wrote:
> On Mon, May 02, 2022 at 08:20:00AM -0400, Brian Foster wrote:
> > On Sat, Apr 30, 2022 at 10:40:31PM +0100, Matthew Wilcox wrote:
> > > On Sat, Apr 30, 2022 at 04:44:07AM +0100, Matthew Wilcox wrote:
> > > > (I do not love this, have not even compiled it; it's late.  We may be
> > > > better off just storing next_folio inside the folio_iter).
> > > 
> > > Does anyone have a preference for fixing this between Option A:
> > > 
> > 
> > After seeing the trace in my previous mail and several thousand
> > successful iterations of the test hack, I had reworked it into this
> > (which survived weekend testing until it ran into some other XFS problem
> > that looks unrelated):
> > 
> > diff --git a/include/linux/bio.h b/include/linux/bio.h
> > index 278cc81cc1e7..aa820e09978e 100644
> > --- a/include/linux/bio.h
> > +++ b/include/linux/bio.h
> > @@ -269,6 +269,7 @@ struct folio_iter {
> >  	size_t offset;
> >  	size_t length;
> >  	/* private: for use by the iterator */
> > +	struct folio *_next;
> >  	size_t _seg_count;
> >  	int _i;
> >  };
> > @@ -279,6 +280,7 @@ static inline void bio_first_folio(struct folio_iter *fi, struct bio *bio,
> >  	struct bio_vec *bvec = bio_first_bvec_all(bio) + i;
> >  
> >  	fi->folio = page_folio(bvec->bv_page);
> > +	fi->_next = folio_next(fi->folio);
> >  	fi->offset = bvec->bv_offset +
> >  			PAGE_SIZE * (bvec->bv_page - &fi->folio->page);
> >  	fi->_seg_count = bvec->bv_len;
> > @@ -290,13 +292,15 @@ static inline void bio_next_folio(struct folio_iter *fi, struct bio *bio)
> >  {
> >  	fi->_seg_count -= fi->length;
> >  	if (fi->_seg_count) {
> > -		fi->folio = folio_next(fi->folio);
> > +		fi->folio = fi->_next;
> > +		fi->_next = folio_next(fi->folio);
> >  		fi->offset = 0;
> >  		fi->length = min(folio_size(fi->folio), fi->_seg_count);
> >  	} else if (fi->_i + 1 < bio->bi_vcnt) {
> >  		bio_first_folio(fi, bio, fi->_i + 1);
> >  	} else {
> >  		fi->folio = NULL;
> > +		fi->_next = NULL;
> >  	}
> >  }
> > 
> > So FWIW, that is just to say that I find option A to be cleaner and more
> > readable.
> 
> Me too.  I'll queue up the usual nightly tests with that patch added and
> we'll see how that does.

I've just pushed essentially that patch to my for-next tree in case
anybody does any testing with that.  I'll give it a couple of days
before creating a folio-5.18f tag and asking Linus to pull the first two
commits on

git://git.infradead.org/users/willy/pagecache.git for-next

That is, commits

1a4c97e2dd5b ("block: Do not call folio_next() on an unreferenced folio")
095099da208b ("mm/readahead: Fix readahead with large folios")

(more than happy to update anything about those patches)