On Tue, Jul 29, 2025 at 03:01:28PM -0400, Tony Battersby wrote:
> Yes, you understand correctly. The test creates a number of sequential
> writes, and this patch cuts the stream of sequential bios on the stripe
> boundaries rather than letting the bios span stripes, so that MD doesn't
> have to do extra work for writes that cross the boundary. I am actually
> working on an out-of-tree RAID driver that benefits hugely from this
> because it doesn't have the complexity of the MD caching layer. But
> benchmarks showed that MD benefited from it (slightly) also, so I
> figured it was worth submitting.
>
> The problem with using iomap_can_add_to_ioend() is that it returns
> true/false, whereas sometimes it is necessary to add some of the folio
> to the current bio and the rest to a new bio.
Hm. Maybe something like this would be more clear?
(contents and indeed name of iomap_should_split_ioend() very much TBD)
diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index 9f541c05103b..429890fb7763 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -1684,6 +1684,7 @@ static int iomap_add_to_ioend(struct iomap_writepage_ctx *wpc,
struct iomap_folio_state *ifs = folio->private;
size_t poff = offset_in_folio(folio, pos);
unsigned int ioend_flags = 0;
+ unsigned thislen;
int error;
if (wpc->iomap.type == IOMAP_UNWRITTEN)
@@ -1704,8 +1705,16 @@ static int iomap_add_to_ioend(struct iomap_writepage_ctx *wpc,
ioend_flags);
}
- if (!bio_add_folio(&wpc->ioend->io_bio, folio, len, poff))
+ thislen = iomap_should_split_ioend(wpc, pos, len);
+
+ if (!bio_add_folio(&wpc->ioend->io_bio, folio, thislen, poff))
+ goto new_ioend;
+ if (thislen < len) {
+ pos += thislen;
+ len -= thislen;
+ wbc_account_cgroup_owner(wbc, folio, thislen);
goto new_ioend;
+ }
if (ifs)
atomic_add(len, &ifs->write_bytes_pending);
