Skip to contentSkip to conversation

blk-core: fix queue stuck on attempt to submit request from unplug

PATCH In review Roman Pen 2 participants 3 messages Sep 27, 2015 – Sep 28, 2015

Subsystems (2)
lore
Share & tools
lore

Roman Pen

author
#0 lore raw

PATCH 1/1 blk-core: fix queue stuck on attempt to submit request from unplug

In case of several stacked block devices, which both were inited by blk_init_queue call, you can catch the queue stuck, if first device in stack makes bio submit being in a flush of a plug list.

Let's consider this regular scenario taking readahead into account (readahead.c:read_pages):

1. Start plug
2. Read pages in loop
3. Finish plug

This example generates backtrace as follows:

1. blk_start_plug
2. generic_make_request

        q->make_request_fn
        [blk_queue_bio]
            if (current->plug)
                list_add_tail(&req->queuelist, &plug->list);
3. blk_finish_plug
        blk_flush_plug_list
            queue_unplugged
                __blk_run_queue
                    XXX_request_fn [some request handler of block device]
                        generic_make_request
                            q->make_request_fn
                            [blk_queue_bio]
                                if (current->plug)
                                    list_add_tail(&req->queuelist, &plug->list);

So the problem is, that on step 3. XXX_request_fn makes another request, which again will be put to plug list, because plug is till active, thus new request will be stuck forever in the queue.

How to fix?
Do flush plug list till it becomes empty.

---
 block/blk-core.c | 10 ++++++++++
 block/blk-mq.c   | 13 +++++++++++++
 2 files changed, 23 insertions(+)
block/blk-core.c +10 -0
index 2eb722d..36b3bd2 100644
@@ -3151,6 +3151,7 @@ void blk_flush_plug_list(struct blk_plug *plug, bool from_schedule)
 	LIST_HEAD(list);
 	unsigned int depth;
 
+repeat:
 	flush_plug_callbacks(plug, from_schedule);
 
 	if (!list_empty(&plug->mq_list))
@@ -3212,6 +3213,15 @@ void blk_flush_plug_list(struct blk_plug *plug, bool from_schedule)
 		queue_unplugged(q, depth, from_schedule);
 
 	local_irq_restore(flags);
+
+	/*
+	 * We have to repeat the whole flush till list becomes
+	 * empty, because underlying block device can submit
+	 * another bio which again will be put to plug list.
+	 * To avoid stuck of these subsequent bios in the queue
+	 * we have to flush till the end.
+	 */
+	goto repeat;
 }
 
 void blk_finish_plug(struct blk_plug *plug)
block/blk-mq.c +13 -0
index f2d67b4..ced83eb 100644
@@ -1091,6 +1091,10 @@ void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
 	LIST_HEAD(ctx_list);
 	unsigned int depth;
 
+repeat:
+	if (list_empty(&plug->mq_list))
+		return;
+
 	list_splice_init(&plug->mq_list, &list);
 
 	list_sort(NULL, &list, plug_ctx_cmp);
@@ -1127,6 +1131,15 @@ void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
 		blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
 				       from_schedule);
 	}
+
+	/*
+	 * We have to repeat the whole flush till list becomes
+	 * empty, because underlying block device can submit
+	 * another bio which again will be put to plug list.
+	 * To avoid stuck of these subsequent bios in the queue
+	 * we have to flush till the end.
+	 */
+	goto repeat;
 }
 
 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
Signature · 1 line
2.5.1
maintainer
#1 lore raw
Quoted 30 lines · Roman Pen

In case of several stacked block devices, which both were inited by blk_init_queue call, you can catch the queue stuck, if first device in stack makes bio submit being in a flush of a plug list.

Let's consider this regular scenario taking readahead into account (readahead.c:read_pages):

1. Start plug
2. Read pages in loop
3. Finish plug

This example generates backtrace as follows:

1. blk_start_plug
2. generic_make_request

         q->make_request_fn
         [blk_queue_bio]
             if (current->plug)
                 list_add_tail(&req->queuelist, &plug->list);
3. blk_finish_plug
         blk_flush_plug_list
             queue_unplugged
                 __blk_run_queue
                     XXX_request_fn [some request handler of block device]
                         generic_make_request
                             q->make_request_fn
                             [blk_queue_bio]
                                 if (current->plug)
                                     list_add_tail(&req->queuelist, &plug->list);
Quoted 7 lines

So the problem is, that on step 3. XXX_request_fn makes another request, which again will be put to plug list, because plug is till active, thus new request will be stuck forever in the queue.

How to fix?
Do flush plug list till it becomes empty.

What devices submit IO from request_fn? Seems like something that would be better handled out of a make_request_fn hook for that device, which would also avoid this corner case.

Signature · 1 line
Jens Axboe
#2 lore raw
Quoted 38 lines · Jens Axboe
Quoted 37 lines · Roman Pen

In case of several stacked block devices, which both were inited by blk_init_queue call, you can catch the queue stuck, if first device in stack makes bio submit being in a flush of a plug list.

Let's consider this regular scenario taking readahead into account (readahead.c:read_pages):

1. Start plug
2. Read pages in loop
3. Finish plug

This example generates backtrace as follows:

1. blk_start_plug
2. generic_make_request

         q->make_request_fn
         [blk_queue_bio]
             if (current->plug)
                 list_add_tail(&req->queuelist, &plug->list);
3. blk_finish_plug
         blk_flush_plug_list
             queue_unplugged
                 __blk_run_queue
                     XXX_request_fn [some request handler of block device]
                         generic_make_request
                             q->make_request_fn
                             [blk_queue_bio]
                                 if (current->plug)
                                     list_add_tail(&req->queuelist,

&plug->list);

So the problem is, that on step 3. XXX_request_fn makes another request, which again will be put to plug list, because plug is till active, thus new request will be stuck forever in the queue.

Quoted 7 lines
Quoted 2 lines

How to fix?
Do flush plug list till it becomes empty.

What devices submit IO from request_fn? Seems like something that would be better handled out of a make_request_fn hook for that device, which would also avoid this corner case.

Practically, this stuck was observed while doing some experiments with md, which was switched from bio mode to request mode.

But what drawbacks has this loop-till-empty? Because it seems that those corner cases with plugging can be easily fixed. Or I am missing something?

Also e.g. flush_plug_callbacks() does exactly the same:

while (!list_empty(&plug->cb_list)) {
      list_splice_init(&plug->cb_list, &callbacks);

      while (!list_empty(&callbacks)) {
       [...]
      }
}

--
Roman

Quoted 5 lines

--
Jens Axboe