@@ -570,13 +570,15 @@ static void dm_end_io_acct(struct dm_io *io)
dm_io_acct(io, true);
}
-static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio)
+static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio, gfp_t gfp_mask)
{
struct dm_io *io;
struct dm_target_io *tio;
struct bio *clone;
- clone = bio_alloc_clone(NULL, bio, GFP_NOIO, &md->mempools->io_bs);
+ clone = bio_alloc_clone(NULL, bio, gfp_mask, &md->mempools->io_bs);
+ if (unlikely(!clone))
+ return NULL;
tio = clone_to_tio(clone);
tio->flags = 0;
dm_tio_set_flag(tio, DM_TIO_INSIDE_DM_IO);
@@ -1714,10 +1716,6 @@ static blk_status_t __split_and_process_bio(struct clone_info *ci)
if (unlikely(!ti))
return BLK_STS_IOERR;
- if (unlikely((ci->bio->bi_opf & REQ_NOWAIT) != 0) &&
- unlikely(!dm_target_supports_nowait(ti->type)))
- return BLK_STS_NOTSUPP;
-
if (unlikely(ci->is_abnormal_io))
return __process_abnormal_io(ci, ti);
@@ -1729,7 +1727,17 @@ static blk_status_t __split_and_process_bio(struct clone_info *ci)
len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
setup_split_accounting(ci, len);
- clone = alloc_tio(ci, ti, 0, &len, GFP_NOIO);
+
+ if (unlikely(ci->bio->bi_opf & REQ_NOWAIT)) {
+ if (unlikely(!dm_target_supports_nowait(ti->type)))
+ return BLK_STS_NOTSUPP;
+
+ clone = alloc_tio(ci, ti, 0, &len, GFP_NOWAIT);
+ if (unlikely(!clone))
+ return BLK_STS_AGAIN;
+ } else {
+ clone = alloc_tio(ci, ti, 0, &len, GFP_NOIO);
+ }
__map_bio(clone);
ci->sector += len;
@@ -1738,11 +1746,11 @@ static blk_status_t __split_and_process_bio(struct clone_info *ci)
return BLK_STS_OK;
}
-static void init_clone_info(struct clone_info *ci, struct mapped_device *md,
+static void init_clone_info(struct clone_info *ci, struct dm_io *io,
struct dm_table *map, struct bio *bio, bool is_abnormal)
{
ci->map = map;
- ci->io = alloc_io(md, bio);
+ ci->io = io;
ci->bio = bio;
ci->is_abnormal_io = is_abnormal;
ci->submit_as_polled = false;
@@ -1764,7 +1772,7 @@ static void dm_split_and_process_bio(struct mapped_device *md,
struct clone_info ci;
struct dm_io *io;
blk_status_t error = BLK_STS_OK;
- bool is_abnormal;
+ bool is_abnormal, is_preflush = !!(bio->bi_opf & REQ_PREFLUSH);
is_abnormal = is_abnormal_io(bio);
if (unlikely(is_abnormal)) {
@@ -1777,10 +1785,26 @@ static void dm_split_and_process_bio(struct mapped_device *md,
return;
}
- init_clone_info(&ci, md, map, bio, is_abnormal);
- io = ci.io;
+ if (unlikely(bio->bi_opf & REQ_NOWAIT)) {
+ /* Only support nowait for normal IO */
+ if (unlikely(is_preflush || is_abnormal)) {
+ bio->bi_status = BLK_STS_NOTSUPP;
+ bio_endio(bio);
+ return;
+ }
- if (bio->bi_opf & REQ_PREFLUSH) {
+ io = alloc_io(md, bio, GFP_NOWAIT);
+ if (unlikely(!io)) {
+ /* Unable to do anything without dm_io. */
+ bio_wouldblock_error(bio);
+ return;
+ }
+ } else {
+ io = alloc_io(md, bio, GFP_NOIO);
+ }
+ init_clone_info(&ci, io, map, bio, is_abnormal);
+
+ if (is_preflush) {
__send_empty_flush(&ci);
/* dm_io_complete submits any data associated with flush */
goto out;
Update DM core's normal IO submission to allocate required memory using GFP_NOWAIT if REQ_NOWAIT is set. Tested with simple test provided in commit a9ce385344f916 ("dm: don't attempt to queue IO under RCU protection") that was enhanced to check error codes. Also tested using fio's pvsync2 with nowait=1. But testing with induced GFP_NOWAIT allocation failures wasn't performed (yet). Signed-off-by: Mike Snitzer <snitzer@kernel.org> --- drivers/md/dm.c | 50 ++++++++++++++++++++++++++++++++++++------------- 1 file changed, 37 insertions(+), 13 deletions(-) v3: - followed Mikulas's suggestion of only supporting NOWAIT for normal IO - changed attribution to me since basically all the code was rewritten - removed dm_io_rewind() changes thanks to Ming's reminder that DM's requeue won't block IO submission because it is called from worker (so it using GFP_NOIO is fine).