On 7/29/25 15:17, Matthew Wilcox wrote:
> 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);
How is this? Does ioend_flags need to be recomputed (particularly
IOMAP_IOEND_BOUNDARY) when processing the remainder of the folio?
diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index fb4519158f3a..0967e6fd62a1 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -1669,6 +1669,39 @@ static bool iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t pos,
return true;
}
+static unsigned int iomap_should_split_ioend(struct iomap_writepage_ctx *wpc,
+ loff_t pos, unsigned int len)
+{
+ struct queue_limits *lim = bdev_limits(wpc->iomap.bdev);
+
+ if ((lim->features & BLK_FEAT_RAID_PARTIAL_STRIPES_EXPENSIVE) &&
+ !(wpc->iomap.flags & IOMAP_F_ANON_WRITE)) {
+ unsigned int io_align = lim->io_opt >> SECTOR_SHIFT;
+
+ /* Split sequential writes along io_align boundaries. */
+ if (io_align) {
+ sector_t lba = bio_end_sector(&wpc->ioend->io_bio);
+ unsigned int mod = lba % io_align;
+ unsigned int max_len;
+
+ /*
+ * If the end sector is already aligned and the bio is
+ * nonempty, then start a new bio for the remainder.
+ */
+ if (!mod && wpc->ioend->io_bio.bi_iter.bi_size)
+ return 0;
+
+ /*
+ * Clip the end of the bio to the alignment boundary.
+ */
+ max_len = (io_align - mod) << SECTOR_SHIFT;
+ if (len > max_len)
+ len = max_len;
+ }
+ }
+ return len;
+}
+
/*
* Test to see if we have an existing ioend structure that we could append to
* first; otherwise finish off the current ioend and start another.
@@ -1688,6 +1721,7 @@ 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 int thislen;
int error;
if (wpc->iomap.type == IOMAP_UNWRITTEN)
@@ -1708,11 +1742,14 @@ 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 (!thislen)
+ goto new_ioend;
+ if (!bio_add_folio(&wpc->ioend->io_bio, folio, thislen, poff))
goto new_ioend;
if (ifs)
- atomic_add(len, &ifs->write_bytes_pending);
+ atomic_add(thislen, &ifs->write_bytes_pending);
/*
* Clamp io_offset and io_size to the incore EOF so that ondisk
@@ -1755,11 +1792,18 @@ static int iomap_add_to_ioend(struct iomap_writepage_ctx *wpc,
* Note that this defeats the ability to chain the ioends of
* appending writes.
*/
- wpc->ioend->io_size += len;
+ wpc->ioend->io_size += thislen;
if (wpc->ioend->io_offset + wpc->ioend->io_size > end_pos)
wpc->ioend->io_size = end_pos - wpc->ioend->io_offset;
- wbc_account_cgroup_owner(wbc, folio, len);
+ wbc_account_cgroup_owner(wbc, folio, thislen);
+
+ if (thislen < len) {
+ pos += thislen;
+ len -= thislen;
+ goto new_ioend;
+ }
+
return 0;
}
--
2.43.0
