dm: support bio polling

Support bio polling (REQ_POLLED) in the following approach:

1) only support io polling on normal READ/WRITE, and other abnormal IOs
still fallback to IRQ mode, so the target io (and DM's clone bio) is
exactly inside the dm io.

2) hold one refcnt on io->io_count after submitting this dm bio with
REQ_POLLED

3) support dm native bio splitting, any dm io instance associated with
current bio will be added into one list which head is bio->bi_private
which will be recovered before ending this bio

4) implement .poll_bio() callback, call bio_poll() on the single target
bio inside the dm io which is retrieved via bio->bi_bio_drv_data; call
dm_io_dec_pending() after the target io is done in .poll_bio()

5) enable QUEUE_FLAG_POLL if all underlying queues enable QUEUE_FLAG_POLL,
which is based on Jeffle's previous patch.

These changes are good for a 30-35% IOPS improvement for polled IO.

For detailed test results please see (Jens, thanks for testing!):
https://listman.redhat.com/archives/dm-devel/2022-March/049868.html
or https://marc.info/?l=linux-block&m=164684246214700&w=2

Tested-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>

1 files changed, 140 insertions, 3 deletions

diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index 454d39bc7745..d9111e17f0fc 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c
@@ -40,6 +40,13 @@
 #define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
 #define DM_COOKIE_LENGTH 24
 
+/*
+ * For REQ_POLLED fs bio, this flag is set if we link mapped underlying
+ * dm_io into one list, and reuse bio->bi_private as the list head. Before
+ * ending this fs bio, we will recover its ->bi_private.
+ */
+#define REQ_DM_POLL_LIST	REQ_DRV
+
 static const char *_name = DM_NAME;
 
 static unsigned int major = 0;
@@ -73,6 +80,7 @@ struct clone_info {
 	struct dm_io *io;
 	sector_t sector;
 	unsigned sector_count;
+	bool submit_as_polled;
 };
 
 #define DM_TARGET_IO_BIO_OFFSET (offsetof(struct dm_target_io, clone))
@@ -599,6 +607,9 @@ static struct bio *alloc_tio(struct clone_info *ci, struct dm_target *ti,
 		if (!clone)
 			return NULL;
 
+		/* REQ_DM_POLL_LIST shouldn't be inherited */
+		clone->bi_opf &= ~REQ_DM_POLL_LIST;
+
 		tio = clone_to_tio(clone);
 		tio->inside_dm_io = false;
 	}
@@ -888,8 +899,15 @@ void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
 		if (unlikely(wq_has_sleeper(&md->wait)))
 			wake_up(&md->wait);
 
-		if (io_error == BLK_STS_DM_REQUEUE)
+		if (io_error == BLK_STS_DM_REQUEUE) {
+			/*
+			 * Upper layer won't help us poll split bio, io->orig_bio
+			 * may only reflect a subset of the pre-split original,
+			 * so clear REQ_POLLED in case of requeue
+			 */
+			bio->bi_opf &= ~REQ_POLLED;
 			return;
+		}
 
 		if (bio_is_flush_with_data(bio)) {
 			/*
@@ -1441,6 +1459,47 @@ static bool __process_abnormal_io(struct clone_info *ci, struct dm_target *ti,
 }
 
 /*
+ * Reuse ->bi_private as hlist head for storing all dm_io instances
+ * associated with this bio, and this bio's bi_private needs to be
+ * stored in dm_io->data before the reuse.
+ *
+ * bio->bi_private is owned by fs or upper layer, so block layer won't
+ * touch it after splitting. Meantime it won't be changed by anyone after
+ * bio is submitted. So this reuse is safe.
+ */
+static inline struct hlist_head *dm_get_bio_hlist_head(struct bio *bio)
+{
+	return (struct hlist_head *)&bio->bi_private;
+}
+
+static void dm_queue_poll_io(struct bio *bio, struct dm_io *io)
+{
+	struct hlist_head *head = dm_get_bio_hlist_head(bio);
+
+	if (!(bio->bi_opf & REQ_DM_POLL_LIST)) {
+		bio->bi_opf |= REQ_DM_POLL_LIST;
+		/*
+		 * Save .bi_private into dm_io, so that we can reuse
+		 * .bi_private as hlist head for storing dm_io list
+		 */
+		io->data = bio->bi_private;
+
+		INIT_HLIST_HEAD(head);
+
+		/* tell block layer to poll for completion */
+		bio->bi_cookie = ~BLK_QC_T_NONE;
+	} else {
+		/*
+		 * bio recursed due to split, reuse original poll list,
+		 * and save bio->bi_private too.
+		 */
+		io->data = hlist_entry(head->first, struct dm_io, node)->data;
+	}
+
+	hlist_add_head(&io->node, head);
+}
+
+/*
  * Select the correct strategy for processing a non-flush bio.
  */
 static int __split_and_process_bio(struct clone_info *ci)
@@ -1457,6 +1516,12 @@ static int __split_and_process_bio(struct clone_info *ci)
 	if (__process_abnormal_io(ci, ti, &r))
 		return r;
 
+	/*
+	 * Only support bio polling for normal IO, and the target io is
+	 * exactly inside the dm_io instance (verified in dm_poll_dm_io)
+	 */
+	ci->submit_as_polled = ci->bio->bi_opf & REQ_POLLED;
+
 	len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
 	clone = alloc_tio(ci, ti, 0, &len, GFP_NOIO);
 	__map_bio(clone);
@@ -1473,6 +1538,7 @@ static void init_clone_info(struct clone_info *ci, struct mapped_device *md,
 	ci->map = map;
 	ci->io = alloc_io(md, bio);
 	ci->bio = bio;
+	ci->submit_as_polled = false;
 	ci->sector = bio->bi_iter.bi_sector;
 	ci->sector_count = bio_sectors(bio);
 
@@ -1522,8 +1588,17 @@ out:
 	if (ci.io->start_io_acct)
 		dm_start_io_acct(ci.io, NULL);
 
-	/* drop the extra reference count */
-	dm_io_dec_pending(ci.io, errno_to_blk_status(error));
+	/*
+	 * Drop the extra reference count for non-POLLED bio, and hold one
+	 * reference for POLLED bio, which will be released in dm_poll_bio
+	 *
+	 * Add every dm_io instance into the hlist_head which is stored in
+	 * bio->bi_private, so that dm_poll_bio can poll them all.
+	 */
+	if (error || !ci.submit_as_polled)
+		dm_io_dec_pending(ci.io, errno_to_blk_status(error));
+	else
+		dm_queue_poll_io(bio, ci.io);
 }
 
 static void dm_submit_bio(struct bio *bio)
@@ -1558,6 +1633,67 @@ out:
 	dm_put_live_table(md, srcu_idx);
 }
 
+static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
+			  unsigned int flags)
+{
+	WARN_ON_ONCE(!io->tio.inside_dm_io);
+
+	/* don't poll if the mapped io is done */
+	if (atomic_read(&io->io_count) > 1)
+		bio_poll(&io->tio.clone, iob, flags);
+
+	/* bio_poll holds the last reference */
+	return atomic_read(&io->io_count) == 1;
+}
+
+static int dm_poll_bio(struct bio *bio, struct io_comp_batch *iob,
+		       unsigned int flags)
+{
+	struct hlist_head *head = dm_get_bio_hlist_head(bio);
+	struct hlist_head tmp = HLIST_HEAD_INIT;
+	struct hlist_node *next;
+	struct dm_io *io;
+
+	/* Only poll normal bio which was marked as REQ_DM_POLL_LIST */
+	if (!(bio->bi_opf & REQ_DM_POLL_LIST))
+		return 0;
+
+	WARN_ON_ONCE(hlist_empty(head));
+
+	hlist_move_list(head, &tmp);
+
+	/*
+	 * Restore .bi_private before possibly completing dm_io.
+	 *
+	 * bio_poll() is only possible once @bio has been completely
+	 * submitted via submit_bio_noacct()'s depth-first submission.
+	 * So there is no dm_queue_poll_io() race associated with
+	 * clearing REQ_DM_POLL_LIST here.
+	 */
+	bio->bi_opf &= ~REQ_DM_POLL_LIST;
+	bio->bi_private = hlist_entry(tmp.first, struct dm_io, node)->data;
+
+	hlist_for_each_entry_safe(io, next, &tmp, node) {
+		if (dm_poll_dm_io(io, iob, flags)) {
+			hlist_del_init(&io->node);
+			/*
+			 * clone_endio() has already occurred, so passing
+			 * error as 0 here doesn't override io->status
+			 */
+			dm_io_dec_pending(io, 0);
+		}
+	}
+
+	/* Not done? */
+	if (!hlist_empty(&tmp)) {
+		bio->bi_opf |= REQ_DM_POLL_LIST;
+		/* Reset bio->bi_private to dm_io list head */
+		hlist_move_list(&tmp, head);
+		return 0;
+	}
+	return 1;
+}
+
 /*-----------------------------------------------------------------
  * An IDR is used to keep track of allocated minor numbers.
  *---------------------------------------------------------------*/
@@ -2983,6 +3119,7 @@ static const struct pr_ops dm_pr_ops = {
 
 static const struct block_device_operations dm_blk_dops = {
 	.submit_bio = dm_submit_bio,
+	.poll_bio = dm_poll_bio,
 	.open = dm_blk_open,
 	.release = dm_blk_close,
 	.ioctl = dm_blk_ioctl,