[PATCH] blk-mq: don't leak preempt counter/q_usage_counter when allocating rq failed

[PATCH] blk-mq: don't leak preempt counter/q_usage_counter when allocating rq failed

[Ming Lei]

From: Ming Lei <minlei@redhat.com>

When blk_mq_get_request() failed, preempt counter isn't
released, and blk_mq_make_request() doesn't release the counter
too.

This patch fixes the issue, and makes sure that preempt counter
is only held if rq is allocated successfully. The same policy is
applied on .q_usage_counter too.

Signed-off-by: Ming Lei <minlei@redhat.com>
---
 block/blk-mq.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 041f7b7fa0d6..4f2561412f13 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -301,11 +301,14 @@ static struct request *blk_mq_get_request(struct request_queue *q,
 	struct elevator_queue *e = q->elevator;
 	struct request *rq;
 	unsigned int tag;
+	bool get_cpu = false;
 
 	blk_queue_enter_live(q);
 	data->q = q;
-	if (likely(!data->ctx))
+	if (likely(!data->ctx)) {
 		data->ctx = blk_mq_get_ctx(q);
+		get_cpu = true;
+	}
 	if (likely(!data->hctx))
 		data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
 	if (op & REQ_NOWAIT)
@@ -324,6 +327,8 @@ static struct request *blk_mq_get_request(struct request_queue *q,
 
 	tag = blk_mq_get_tag(data);
 	if (tag == BLK_MQ_TAG_FAIL) {
+		if (get_cpu)
+			blk_mq_put_ctx(data->ctx);
 		blk_queue_exit(q);
 		return NULL;
 	}
@@ -356,12 +361,12 @@ struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 
 	rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 
-	blk_mq_put_ctx(alloc_data.ctx);
-	blk_queue_exit(q);
-
 	if (!rq)
 		return ERR_PTR(-EWOULDBLOCK);
 
+	blk_mq_put_ctx(alloc_data.ctx);
+	blk_queue_exit(q);
+
 	rq->__data_len = 0;
 	rq->__sector = (sector_t) -1;
 	rq->bio = rq->biotail = NULL;
@@ -407,11 +412,11 @@ struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 
 	rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 
-	blk_queue_exit(q);
-
 	if (!rq)
 		return ERR_PTR(-EWOULDBLOCK);
 
+	blk_queue_exit(q);
+
 	return rq;
 }
 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
-- 
2.9.4

Re: [PATCH] blk-mq: don't leak preempt counter/q_usage_counter when allocating rq failed

[Ming Lei]

On Fri, Jul 14, 2017 at 04:41:36PM +0800, Ming Lei wrote:
> From: Ming Lei <minlei@redhat.com>
> 
> When blk_mq_get_request() failed, preempt counter isn't
> released, and blk_mq_make_request() doesn't release the counter
> too.
> 
> This patch fixes the issue, and makes sure that preempt counter
> is only held if rq is allocated successfully. The same policy is
> applied on .q_usage_counter too.

Hi Jens,

Given NOWAIT has been merged, could you consider this patch for
v4.13?

Thanks,
Ming

Re: [PATCH] blk-mq: don't leak preempt counter/q_usage_counter when allocating rq failed

[Jens Axboe]

On 07/14/2017 02:41 AM, Ming Lei wrote:
> From: Ming Lei <minlei@redhat.com>
> 
> When blk_mq_get_request() failed, preempt counter isn't
> released, and blk_mq_make_request() doesn't release the counter
> too.
> 
> This patch fixes the issue, and makes sure that preempt counter
> is only held if rq is allocated successfully. The same policy is
> applied on .q_usage_counter too.

Can you replace the 'get_cpu' bool with just a ctx, and change
the logic to put it if set? I think that would be cleaner to read,
generally I hate bool 'do_something' variables that are set. It's
much cleaner to have:

if (likely(!data->ctx))
	data->ctx = local_ctx = blk_mq_get_ctx(q);

and have the put case be

if (local_ctx)
	blk_mq_put_ctx(local_ctx);

Either that, or at least just have blk_mq_get_request() do:

drop_ctx = data->ctx == NULL;

instead. The 'get_cpu' naming is confusing, we're worried about dropping
the sw queue here, the fact that it's related to get/put_cpu() need not
bubble up here.

-- 
Jens Axboe

Re: [PATCH] blk-mq: don't leak preempt counter/q_usage_counter when allocating rq failed

[Ming Lei]

On Tue, Aug 01, 2017 at 09:18:23AM -0600, Jens Axboe wrote:
> On 07/14/2017 02:41 AM, Ming Lei wrote:
> > From: Ming Lei <minlei@redhat.com>
> > 
> > When blk_mq_get_request() failed, preempt counter isn't
> > released, and blk_mq_make_request() doesn't release the counter
> > too.
> > 
> > This patch fixes the issue, and makes sure that preempt counter
> > is only held if rq is allocated successfully. The same policy is
> > applied on .q_usage_counter too.
> 
> Can you replace the 'get_cpu' bool with just a ctx, and change
> the logic to put it if set? I think that would be cleaner to read,
> generally I hate bool 'do_something' variables that are set. It's
> much cleaner to have:
> 
> if (likely(!data->ctx))
> 	data->ctx = local_ctx = blk_mq_get_ctx(q);
> 
> and have the put case be
> 
> if (local_ctx)
> 	blk_mq_put_ctx(local_ctx);
> 
> Either that, or at least just have blk_mq_get_request() do:
> 
> drop_ctx = data->ctx == NULL;
> 
> instead. The 'get_cpu' naming is confusing, we're worried about dropping
> the sw queue here, the fact that it's related to get/put_cpu() need not
> bubble up here.

Good point, V2 is sent out with this change. Thanks your suggestion!

-- 
Ming