On Tue, Mar 01, 2022 at 05:37:17AM -0800, Christoph Hellwig wrote:

quoted

-			hctxs[i] = blk_mq_alloc_and_init_hctx(set, q, i,
-					old_node);
-			WARN_ON_ONCE(!hctxs[i]);
+			WARN_ON_ONCE(!blk_mq_alloc_and_init_hctx(set, q, i,
+						old_node));

Please avoid doing the actual work inside a WARN_ON statement.

OK.

quoted

 
 	for (; j < end; j++) {
-		struct blk_mq_hw_ctx *hctx = hctxs[j];
+		struct blk_mq_hw_ctx *hctx = blk_mq_get_hctx(q, j);
 
-		if (hctx) {
+		if (hctx)
 			blk_mq_exit_hctx(q, set, hctx, j);
-			hctxs[j] = NULL;
-		}
 	}

Instead of a for loop that does xa_loads repeatedly this can just
use xa_for_each_range.  Same for a bunch of other loops like that,
e.g. in blk_mq_unregister_dev or the __blk_mq_register_dev failure
path.

quoted

@@ -919,12 +919,12 @@ static inline void *blk_mq_rq_to_pdu(struct request *rq)
 static inline struct blk_mq_hw_ctx *blk_mq_get_hctx(struct request_queue *q,
 		unsigned int hctx_idx)
 {
-	return q->queue_hw_ctx[hctx_idx];
+	return xa_load(&q->hctx_table, hctx_idx);
 }
 
 #define queue_for_each_hw_ctx(q, hctx, i)				\
 	for ((i) = 0; (i) < (q)->nr_hw_queues &&			\
-	     ({ hctx = blk_mq_get_hctx((q), (i)); 1; }); (i)++)
+	     (hctx = blk_mq_get_hctx((q), (i))); (i)++)

This should be using a xa_for_each loop.

I did considered xa_for_each(), but it requires rcu read lock.

Also queue_for_each_hw_ctx() is supposed to not run in fast path,
meantime xa_load() is lightweight enough too, so repeated xa_load()
is fine here.


Thanks,
Ming