On Fri, Sep 06, 2019 at 10:52:41PM +0200, Andreas Gruenbacher wrote: > Hi, > > I've just fixed a mmap write vs. truncate consistency issue on gfs on > filesystems with a block size smaller that the page size [1]. > > It turns out that the same problem exists between mmap write and hole > punching, and since xfstests doesn't seem to cover that, AFAIA, fsx exercises it pretty often. Certainly it's found problems with XFS in the past w.r.t. these operations. > I've written a > new test [2]. I suspect that what we really want is a test that runs _test_generic_punch using mmap rather than pwrite... > Ext4 and xfs both pass that test; they both apparently > mark the pages that have a hole punched in them as read-only so that > page_mkwrite is called before those pages can be written to again. XFS invalidates the range being hole punched (see xfs_flush_unmap_range() under XFS_MMAPLOCK_EXCL, which means any attempt to fault that page back in will block on the MMAPLOCK until the hole punch finishes. > gfs2 fails that: for some reason, the partially block-mapped pages are > not marked read-only on gfs2, and so page_mkwrite is not called for the > partially block-mapped pages, and the hole is not filled in correctly. > > The attached patch fixes the problem, but this really doesn't look right > as neither ext4 nor xfs require this kind of hack. So what am I > overlooking, how does this work on ext4 and xfs? XFS uses XFS_MMAPLOCK_* to serialise page faults against extent manipulations (shift, hole punch, truncate, swap, etc) and ext4 uses a similar locking mechanism to do the same thing. Fundamentally, the page cache does not provide the necessary mechanisms to detect and prevent invalidation races inside EOF.... > > Signed-off-by: Andreas Gruenbacher <agruenba@xxxxxxxxxx> > --- > fs/gfs2/bmap.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/fs/gfs2/bmap.c b/fs/gfs2/bmap.c > index 9ef543dd38e2..e677e813be4c 100644 > --- a/fs/gfs2/bmap.c > +++ b/fs/gfs2/bmap.c > @@ -2475,6 +2475,13 @@ int __gfs2_punch_hole(struct file *file, loff_t offset, loff_t length) > if (error) > goto out; > } > + /* > + * If the first or last page partially lies in the hole, mark > + * the page read-only so that memory-mapped writes will trigger > + * page_mkwrite. > + */ > + pagecache_isize_extended(inode, offset, inode->i_size); > + pagecache_isize_extended(inode, offset + length, inode->i_size); See xfs_flush_unmap_range(), which is run under XFS_MMAPLOCK_EXCL to serialise against incoming page faults... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
