blk-mq: add new API of blk_mq_hctx_set_fq_lock_class

flush_end_io() may be called recursively from some driver, such as
nvme-loop, so lockdep may complain 'possible recursive locking'.
Commit b3c6a59("block: Fix a lockdep complaint triggered by
request queue flushing") tried to address this issue by assigning
dynamically allocated per-flush-queue lock class. This solution
adds synchronize_rcu() for each hctx's release handler, and causes
horrible SCSI MQ probe delay(more than half an hour on megaraid sas).

Add new API of blk_mq_hctx_set_fq_lock_class() for these drivers, so
we just need to use driver specific lock class for avoiding the
lockdep warning of 'possible recursive locking'.

Tested-by: Kashyap Desai <[email protected]>
Reported-by: Qian Cai <[email protected]>
Cc: Sumit Saxena <[email protected]>
Cc: John Garry <[email protected]>
Cc: Kashyap Desai <[email protected]>
Cc: Bart Van Assche <[email protected]>
Cc: Hannes Reinecke <[email protected]>
Signed-off-by: Ming Lei <[email protected]>
Reviewed-by: Hannes Reinecke <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

Author: Ming Lei

block/blk-flush.c

@@ -490,3 +490,28 @@ void blk_free_flush_queue(struct blk_flush_queue *fq)
 	kfree(fq->flush_rq);
 	kfree(fq);
 }
+
+/*
+ * Allow driver to set its own lock class to fq->mq_flush_lock for
+ * avoiding lockdep complaint.
+ *
+ * flush_end_io() may be called recursively from some driver, such as
+ * nvme-loop, so lockdep may complain 'possible recursive locking' because
+ * all 'struct blk_flush_queue' instance share same mq_flush_lock lock class
+ * key. We need to assign different lock class for these driver's
+ * fq->mq_flush_lock for avoiding the lockdep warning.
+ *
+ * Use dynamically allocated lock class key for each 'blk_flush_queue'
+ * instance is over-kill, and more worse it introduces horrible boot delay
+ * issue because synchronize_rcu() is implied in lockdep_unregister_key which
+ * is called for each hctx release. SCSI probing may synchronously create and
+ * destroy lots of MQ request_queues for non-existent devices, and some robot
+ * test kernel always enable lockdep option. It is observed that more than half
+ * an hour is taken during SCSI MQ probe with per-fq lock class.
+ */
+void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,
+		struct lock_class_key *key)
+{
+	lockdep_set_class(&hctx->fq->mq_flush_lock, key);
+}
+EXPORT_SYMBOL_GPL(blk_mq_hctx_set_fq_lock_class);

include/linux/blk-mq.h

@@ -5,6 +5,7 @@
 #include <linux/blkdev.h>
 #include <linux/sbitmap.h>
 #include <linux/srcu.h>
+#include <linux/lockdep.h>
 
 struct blk_mq_tags;
 struct blk_flush_queue;
@@ -594,5 +595,7 @@ static inline void blk_mq_cleanup_rq(struct request *rq)
 }
 
 blk_qc_t blk_mq_submit_bio(struct bio *bio);
+void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,
+		struct lock_class_key *key);
 
 #endif