On Tue, Jul 15, 2025 at 07:37:33AM -0700, Darrick J. Wong wrote:
> On Tue, Jul 15, 2025 at 08:36:54AM -0400, Brian Foster wrote:
> > On Mon, Jul 14, 2025 at 10:34:17PM -0700, Darrick J. Wong wrote:
> > > On Mon, Jul 14, 2025 at 04:41:21PM -0400, Brian Foster wrote:
> > > > iomap_zero_range() optimizes the partial eof block zeroing use case
> > > > by force zeroing if the mapping is dirty. This is to avoid frequent
> > > > flushing on file extending workloads, which hurts performance.
> > > >
> > > > Now that the folio batch mechanism provides a more generic solution
> > > > and is used by the only real zero range user (XFS), this isolated
> > > > optimization is no longer needed. Remove the unnecessary code and
> > > > let callers use the folio batch or fall back to flushing by default.
> > > >
> > > > Signed-off-by: Brian Foster <bfoster@xxxxxxxxxx>
> > > > Reviewed-by: Christoph Hellwig <hch@xxxxxx>
> > >
> > > Heh, I was staring at this last Friday chasing fuse+iomap bugs in
> > > fallocate zerorange and straining to remember what this does.
> > > Is this chunk still needed if the ->iomap_begin implementation doesn't
> > > (or forgets to) grab the folio batch for iomap?
> > >
> >
> > No, the hunk removed by this patch is just an optimization. The fallback
> > code here flushes the range if it's dirty and retries the lookup (i.e.
> > picking up unwritten conversions that were pending via dirty pagecache).
> > That flush logic caused a performance regression in a particular
> > workload, so this was introduced to mitigate that regression by just
> > doing the zeroing for the first block or so if the folio is dirty. [1]
> >
> > The reason for removing it is more just for maintainability. XFS is
> > really the only user here and it is changing over to the more generic
> > batch mechanism, which effectively provides the same optimization, so
> > this basically becomes dead/duplicate code. If an fs doesn't use the
> > batch mechanism it will just fall back to the flush and retry approach,
> > which can be slower but is functionally correct.
>
> Oh ok thanks for the reminder.
> Reviewed-by: "Darrick J. Wong" <djwong@xxxxxxxxxx>
>
> > > My bug turned out to be a bug in my fuse+iomap design -- with the way
> > > iomap_zero_range does things, you have to flush+unmap, punch the range
> > > and zero the range. If you punch and realloc the range and *then* try
> > > to zero the range, the new unwritten extents cause iomap to miss dirty
> > > pages that fuse should've unmapped. Ooops.
> > >
> >
> > I don't quite follow. How do you mean it misses dirty pages?
>
> Oops, I misspoke, the folios were clean. Let's say the pagecache is
> sparsely populated with some folios for written space:
>
> -------fffff-------fffffff
> wwwwwwwwwwwwwwwwwwwwwwwwww
>
> Now you tell it to go zero range the middle. fuse's fallocate code
> issues the upcall to userspace, whch changes some mappings:
>
> -------fffff-------fffffff
> wwwwwuuuuuuuuuuuwwwwwwwwww
>
> Only after the upcall returns does the kernel try to do the pagecache
> zeroing. Unfortunately, the mapping changed to unwritten so
> iomap_zero_range doesn't see the "fffff" and leaves its contents intact.
>
Ah, interesting. So presumably the fuse fs is not doing any cache
managment, and this creates an unexpected inconsistency between
pagecache and block state.
So what's the solution to this for fuse+iomap? Invalidate the cache
range before or after the callback or something?
Brian
> (Note: Non-iomap fuse defers everything to the fuse server so this isn't
> a problem if the fuse server does all the zeroing itself.)
>
> --D
>
> > Brian
> >
> > [1] Details described in the commit log of fde4c4c3ec1c ("iomap: elide
> > flush from partial eof zero range").
> >
> > > --D
> > >
> > > > ---
> > > > fs/iomap/buffered-io.c | 24 ------------------------
> > > > 1 file changed, 24 deletions(-)
> > > >
> > > > diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
> > > > index 194e3cc0857f..d2bbed692c06 100644
> > > > --- a/fs/iomap/buffered-io.c
> > > > +++ b/fs/iomap/buffered-io.c
> > > > @@ -1484,33 +1484,9 @@ iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
> > > > .private = private,
> > > > };
> > > > struct address_space *mapping = inode->i_mapping;
> > > > - unsigned int blocksize = i_blocksize(inode);
> > > > - unsigned int off = pos & (blocksize - 1);
> > > > - loff_t plen = min_t(loff_t, len, blocksize - off);
> > > > int ret;
> > > > bool range_dirty;
> > > >
> > > > - /*
> > > > - * Zero range can skip mappings that are zero on disk so long as
> > > > - * pagecache is clean. If pagecache was dirty prior to zero range, the
> > > > - * mapping converts on writeback completion and so must be zeroed.
> > > > - *
> > > > - * The simplest way to deal with this across a range is to flush
> > > > - * pagecache and process the updated mappings. To avoid excessive
> > > > - * flushing on partial eof zeroing, special case it to zero the
> > > > - * unaligned start portion if already dirty in pagecache.
> > > > - */
> > > > - if (!iter.fbatch && off &&
> > > > - filemap_range_needs_writeback(mapping, pos, pos + plen - 1)) {
> > > > - iter.len = plen;
> > > > - while ((ret = iomap_iter(&iter, ops)) > 0)
> > > > - iter.status = iomap_zero_iter(&iter, did_zero);
> > > > -
> > > > - iter.len = len - (iter.pos - pos);
> > > > - if (ret || !iter.len)
> > > > - return ret;
> > > > - }
> > > > -
> > > > /*
> > > > * To avoid an unconditional flush, check pagecache state and only flush
> > > > * if dirty and the fs returns a mapping that might convert on
> > > > --
> > > > 2.50.0
> > > >
> > > >
> > >
> >
> >
>
