[PATCH 0/3 rfc] Fix nvme-tcp and nvme-rdma controller reset hangs

[Christoph Hellwig]

What about something like this?

diff --git a/block/blk-core.c b/block/blk-core.c
index fc60ff20849738..4344f3c9058282 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -792,7 +792,7 @@ static inline blk_status_t blk_check_zone_append(struct request_queue *q,
 	return BLK_STS_OK;
 }
 
-static noinline_for_stack bool submit_bio_checks(struct bio *bio)
+noinline_for_stack bool submit_bio_checks(struct bio *bio)
 {
 	struct block_device *bdev = bio->bi_bdev;
 	struct request_queue *q = bdev->bd_disk->queue;
diff --git a/block/blk-mq.c b/block/blk-mq.c
index d4d7c1caa43966..4ff85692843b49 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -2286,6 +2286,43 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 	return BLK_QC_T_NONE;
 }
 
+/**
+ * blk_mq_submit_bio_direct - hand a bio directly to the driver for I/O
+ * @bio:  The bio describing the location in memory and on the device.
+ *
+ * This function behaves similar to submit_bio_noacct(), but does never waits
+ * for the queue to be unfreozen, instead it return false and lets the caller
+ * deal with the fallout.  It also does not protect against recursion and thus
+ * must only be used if the called driver is known to be blk-mq based.
+ */
+bool blk_mq_submit_bio_direct(struct bio *bio, blk_qc_t *qc)
+{
+	struct gendisk *disk = bio->bi_bdev->bd_disk;
+	struct request_queue *q = disk->queue;
+
+	if (WARN_ON_ONCE(!current->bio_list) ||
+	    WARN_ON_ONCE(disk->fops->submit_bio)) {
+		bio_io_error(bio);
+		goto fail;
+	}
+	if (!submit_bio_checks(bio))
+		goto fail;
+
+	if (unlikely(blk_queue_enter(q, BLK_MQ_REQ_NOWAIT)))
+		return false;
+	if (!blk_crypto_bio_prep(&bio))
+		goto fail_queue_exit;
+	*qc = blk_mq_submit_bio(bio);
+	return true;
+
+fail_queue_exit:
+	blk_queue_exit(disk->queue);
+fail:
+	*qc = BLK_QC_T_NONE;
+	return true;
+}
+EXPORT_SYMBOL_GPL(blk_mq_submit_bio_direct);
+
 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
 		     unsigned int hctx_idx)
 {
diff --git a/block/blk.h b/block/blk.h
index 3b53e44b967e4e..c4c66b2a9ffb19 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -221,6 +221,7 @@ ssize_t part_timeout_show(struct device *, struct device_attribute *, char *);
 ssize_t part_timeout_store(struct device *, struct device_attribute *,
 				const char *, size_t);
 
+bool submit_bio_checks(struct bio *bio);
 void __blk_queue_split(struct bio **bio, unsigned int *nr_segs);
 int ll_back_merge_fn(struct request *req, struct bio *bio,
 		unsigned int nr_segs);
diff --git a/drivers/nvme/host/multipath.c b/drivers/nvme/host/multipath.c
index a1d476e1ac020f..92adebfaf86fd1 100644
--- a/drivers/nvme/host/multipath.c
+++ b/drivers/nvme/host/multipath.c
@@ -309,6 +309,7 @@ blk_qc_t nvme_ns_head_submit_bio(struct bio *bio)
 	 */
 	blk_queue_split(&bio);
 
+retry:
 	srcu_idx = srcu_read_lock(&head->srcu);
 	ns = nvme_find_path(head);
 	if (likely(ns)) {
@@ -316,7 +317,12 @@ blk_qc_t nvme_ns_head_submit_bio(struct bio *bio)
 		bio->bi_opf |= REQ_NVME_MPATH;
 		trace_block_bio_remap(bio, disk_devt(ns->head->disk),
 				      bio->bi_iter.bi_sector);
-		ret = submit_bio_noacct(bio);
+
+		if (!blk_mq_submit_bio_direct(bio, &ret)) {
+			nvme_mpath_clear_current_path(ns);
+			srcu_read_unlock(&head->srcu, srcu_idx);
+			goto retry;
+		}
 	} else if (nvme_available_path(head)) {
 		dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
 
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 2c473c9b899089..6804f397106ada 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -615,6 +615,7 @@ static inline void blk_rq_bio_prep(struct request *rq, struct bio *bio,
 }
 
 blk_qc_t blk_mq_submit_bio(struct bio *bio);
+bool blk_mq_submit_bio_direct(struct bio *bio, blk_qc_t *qc);
 void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,
 		struct lock_class_key *key);