block/mq: don't complete requests via IPI

The IPI runs in hardirq context and there are sleeping locks. This patch
moves the completion into a workqueue.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>

4 files changed, 30 insertions, 1 deletions

diff --git a/block/blk-core.c b/block/blk-core.c
index cff0a60ee200..b76d196d6759 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -189,6 +189,9 @@ void blk_rq_init(struct request_queue *q, struct request *rq)
 
 	INIT_LIST_HEAD(&rq->queuelist);
 	INIT_LIST_HEAD(&rq->timeout_list);
+#ifdef CONFIG_PREEMPT_RT_FULL
+	INIT_WORK(&rq->work, __blk_mq_complete_request_remote_work);
+#endif
 	rq->cpu = -1;
 	rq->q = q;
 	rq->__sector = (sector_t) -1;
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 262a9787031d..3074f1f39efa 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -320,6 +320,9 @@ static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 	rq->extra_len = 0;
 	rq->__deadline = 0;
 
+#ifdef CONFIG_PREEMPT_RT_FULL
+	INIT_WORK(&rq->work, __blk_mq_complete_request_remote_work);
+#endif
 	INIT_LIST_HEAD(&rq->timeout_list);
 	rq->timeout = 0;
 
@@ -547,12 +550,24 @@ void blk_mq_end_request(struct request *rq, blk_status_t error)
 }
 EXPORT_SYMBOL(blk_mq_end_request);
 
+#ifdef CONFIG_PREEMPT_RT_FULL
+
+void __blk_mq_complete_request_remote_work(struct work_struct *work)
+{
+	struct request *rq = container_of(work, struct request, work);
+
+	rq->q->softirq_done_fn(rq);
+}
+
+#else
+
 static void __blk_mq_complete_request_remote(void *data)
 {
 	struct request *rq = data;
 
 	rq->q->softirq_done_fn(rq);
 }
+#endif
 
 static void __blk_mq_complete_request(struct request *rq)
 {
@@ -575,10 +590,18 @@ static void __blk_mq_complete_request(struct request *rq)
 		shared = cpus_share_cache(cpu, ctx->cpu);
 
 	if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
+#ifdef CONFIG_PREEMPT_RT_FULL
+		/*
+		 * We could force QUEUE_FLAG_SAME_FORCE then we would not get in
+		 * here. But we could try to invoke it one the CPU like this.
+		 */
+		schedule_work_on(ctx->cpu, &rq->work);
+#else
 		rq->csd.func = __blk_mq_complete_request_remote;
 		rq->csd.info = rq;
 		rq->csd.flags = 0;
 		smp_call_function_single_async(ctx->cpu, &rq->csd);
+#endif
 	} else {
 		rq->q->softirq_done_fn(rq);
 	}
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 1da59c16f637..04c15b5ca76c 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -249,7 +249,7 @@ static inline u16 blk_mq_unique_tag_to_tag(u32 unique_tag)
 	return unique_tag & BLK_MQ_UNIQUE_TAG_MASK;
 }
 
-
+void __blk_mq_complete_request_remote_work(struct work_struct *work);
 int blk_mq_request_started(struct request *rq);
 void blk_mq_start_request(struct request *rq);
 void blk_mq_end_request(struct request *rq, blk_status_t error);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 6980014357d4..f93ae914abda 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -149,6 +149,9 @@ enum mq_rq_state {
  */
 struct request {
 	struct request_queue *q;
+#ifdef CONFIG_PREEMPT_RT_FULL
+	struct work_struct work;
+#endif
 	struct blk_mq_ctx *mq_ctx;
 
 	int cpu;