blk-mq: drain I/O when all CPUs in a hctx are offline

Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup
up queue mapping. Thomas mentioned the following point[1]:

"That was the constraint of managed interrupts from the very beginning:

The driver/subsystem has to quiesce the interrupt line and the associated
queue _before_ it gets shutdown in CPU unplug and not fiddle with it
until it's restarted by the core when the CPU is plugged in again."

However, current blk-mq implementation doesn't quiesce hw queue before
the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a
cpuhp state handled after the CPU is down, so there isn't any chance to
quiesce the hctx before shutting down the CPU.

Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs
where the last CPU goes away, and wait for completion of in-flight
requests. This guarantees that there is no inflight I/O before shutting
down the managed IRQ.

Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need
to wait for completion of in-flight requests from these drivers to avoid
a potential dead-lock. It is safe to do this for stacking drivers as those
do not use interrupts at all and their I/O completions are triggered by
underlying devices I/O completion.

[1] https://lore.kernel.org/linux-block/[e-mail]/

[hch: different retry mechanism, merged two patches, minor cleanups]

Signed-off-by: Ming Lei <[e-mail]>
Signed-off-by: Christoph Hellwig <[e-mail]>
Reviewed-by: Hannes Reinecke <[e-mail]>
Reviewed-by: Daniel Wagner <[e-mail]>
Signed-off-by: Jens Axboe <[e-mail]>

block/blk-mq-debugfs.c

···
 	HCTX_STATE_NAME(STOPPED),
 	HCTX_STATE_NAME(TAG_ACTIVE),
 	HCTX_STATE_NAME(SCHED_RESTART),
+	HCTX_STATE_NAME(INACTIVE),
 };
 #undef HCTX_STATE_NAME
 
···
 	HCTX_FLAG_NAME(TAG_SHARED),
 	HCTX_FLAG_NAME(BLOCKING),
 	HCTX_FLAG_NAME(NO_SCHED),
+	HCTX_FLAG_NAME(STACKING),
 };
 #undef HCTX_FLAG_NAME
 

block/blk-mq-tag.c

···
 	sbitmap_finish_wait(bt, ws, &wait);
 
 found_tag:
+	/*
+	 * Give up this allocation if the hctx is inactive.  The caller will
+	 * retry on an active hctx.
+	 */
+	if (unlikely(test_bit(BLK_MQ_S_INACTIVE, &data->hctx->state))) {
+		blk_mq_put_tag(tags, data->ctx, tag + tag_offset);
+		return BLK_MQ_NO_TAG;
+	}
 	return tag + tag_offset;
 }
 

block/blk-mq.c

···
 			e->type->ops.limit_depth(data->cmd_flags, data);
 	}
 
+retry:
 	data->ctx = blk_mq_get_ctx(q);
 	data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
 	if (!(data->flags & BLK_MQ_REQ_INTERNAL))
 		blk_mq_tag_busy(data->hctx);
 
+	/*
+	 * Waiting allocations only fail because of an inactive hctx.  In that
+	 * case just retry the hctx assignment and tag allocation as CPU hotplug
+	 * should have migrated us to an online CPU by now.
+	 */
 	tag = blk_mq_get_tag(data);
-	if (tag == BLK_MQ_NO_TAG)
-		return NULL;
+	if (tag == BLK_MQ_NO_TAG) {
+		if (data->flags & BLK_MQ_REQ_NOWAIT)
+			return NULL;
+
+		/*
+		 * Give up the CPU and sleep for a random short time to ensure
+		 * that thread using a realtime scheduling class are migrated
+		 * off the the CPU, and thus off the hctx that is going away.
+		 */
+		msleep(3);
+		goto retry;
+	}
 	return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
 }
 
···
 	return -ENOMEM;
 }
 
+struct rq_iter_data {
+	struct blk_mq_hw_ctx *hctx;
+	bool has_rq;
+};
+
+static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
+{
+	struct rq_iter_data *iter_data = data;
+
+	if (rq->mq_hctx != iter_data->hctx)
+		return true;
+	iter_data->has_rq = true;
+	return false;
+}
+
+static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
+{
+	struct blk_mq_tags *tags = hctx->sched_tags ?
+			hctx->sched_tags : hctx->tags;
+	struct rq_iter_data data = {
+		.hctx	= hctx,
+	};
+
+	blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
+	return data.has_rq;
+}
+
+static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
+		struct blk_mq_hw_ctx *hctx)
+{
+	if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
+		return false;
+	if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
+		return false;
+	return true;
+}
+
+static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
+{
+	struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
+			struct blk_mq_hw_ctx, cpuhp_online);
+
+	if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
+	    !blk_mq_last_cpu_in_hctx(cpu, hctx))
+		return 0;
+
+	/*
+	 * Prevent new request from being allocated on the current hctx.
+	 *
+	 * The smp_mb__after_atomic() Pairs with the implied barrier in
+	 * test_and_set_bit_lock in sbitmap_get().  Ensures the inactive flag is
+	 * seen once we return from the tag allocator.
+	 */
+	set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
+	smp_mb__after_atomic();
+
+	/*
+	 * Try to grab a reference to the queue and wait for any outstanding
+	 * requests.  If we could not grab a reference the queue has been
+	 * frozen and there are no requests.
+	 */
+	if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
+		while (blk_mq_hctx_has_requests(hctx))
+			msleep(5);
+		percpu_ref_put(&hctx->queue->q_usage_counter);
+	}
+
+	return 0;
+}
+
+static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
+{
+	struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
+			struct blk_mq_hw_ctx, cpuhp_online);
+
+	if (cpumask_test_cpu(cpu, hctx->cpumask))
+		clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
+	return 0;
+}
+
 /*
  * 'cpu' is going away. splice any existing rq_list entries from this
  * software queue to the hw queue dispatch list, and ensure that it
···
 	enum hctx_type type;
 
 	hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
+	if (!cpumask_test_cpu(cpu, hctx->cpumask))
+		return 0;
+
 	ctx = __blk_mq_get_ctx(hctx->queue, cpu);
 	type = hctx->type;
 
···
 
 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
 {
+	if (!(hctx->flags & BLK_MQ_F_STACKING))
+		cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
+						    &hctx->cpuhp_online);
 	cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
 					    &hctx->cpuhp_dead);
 }
···
 {
 	hctx->queue_num = hctx_idx;
 
+	if (!(hctx->flags & BLK_MQ_F_STACKING))
+		cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
+				&hctx->cpuhp_online);
 	cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
 
 	hctx->tags = set->tags[hctx_idx];
···
 {
 	cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
 				blk_mq_hctx_notify_dead);
+	cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
+				blk_mq_hctx_notify_online,
+				blk_mq_hctx_notify_offline);
 	return 0;
 }
 subsys_initcall(blk_mq_init);

drivers/block/loop.c

···
 	lo->tag_set.queue_depth = 128;
 	lo->tag_set.numa_node = NUMA_NO_NODE;
 	lo->tag_set.cmd_size = sizeof(struct loop_cmd);
-	lo->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
+	lo->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_STACKING;
 	lo->tag_set.driver_data = lo;
 
 	err = blk_mq_alloc_tag_set(&lo->tag_set);

drivers/md/dm-rq.c

···
 	md->tag_set->ops = &dm_mq_ops;
 	md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
 	md->tag_set->numa_node = md->numa_node_id;
-	md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
+	md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_STACKING;
 	md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
 	md->tag_set->driver_data = md;
 

include/linux/blk-mq.h

···
 	 */
 	atomic_t		nr_active;
 
+	/** @cpuhp_online: List to store request if CPU is going to die */
+	struct hlist_node	cpuhp_online;
 	/** @cpuhp_dead: List to store request if some CPU die. */
 	struct hlist_node	cpuhp_dead;
 	/** @kobj: Kernel object for sysfs. */
···
 enum {
 	BLK_MQ_F_SHOULD_MERGE	= 1 << 0,
 	BLK_MQ_F_TAG_SHARED	= 1 << 1,
+	/*
+	 * Set when this device requires underlying blk-mq device for
+	 * completing IO:
+	 */
+	BLK_MQ_F_STACKING	= 1 << 2,
 	BLK_MQ_F_BLOCKING	= 1 << 5,
 	BLK_MQ_F_NO_SCHED	= 1 << 6,
 	BLK_MQ_F_ALLOC_POLICY_START_BIT = 8,
···
 	BLK_MQ_S_STOPPED	= 0,
 	BLK_MQ_S_TAG_ACTIVE	= 1,
 	BLK_MQ_S_SCHED_RESTART	= 2,
+
+	/* hw queue is inactive after all its CPUs become offline */
+	BLK_MQ_S_INACTIVE	= 3,
 
 	BLK_MQ_MAX_DEPTH	= 10240,
 

include/linux/cpuhotplug.h

···
 	CPUHP_AP_SMPBOOT_THREADS,
 	CPUHP_AP_X86_VDSO_VMA_ONLINE,
 	CPUHP_AP_IRQ_AFFINITY_ONLINE,
+	CPUHP_AP_BLK_MQ_ONLINE,
 	CPUHP_AP_ARM_MVEBU_SYNC_CLOCKS,
 	CPUHP_AP_X86_INTEL_EPB_ONLINE,
 	CPUHP_AP_PERF_ONLINE,