[PATCH 0/2] blk-mq: fix schedule from atomic context

[PATCH 0/2] blk-mq: fix schedule from atomic context

[Ming Lei]

Hi Jens,

These two patches fix the problem of schedule from atomic context
which is introduced in bitmap based allocation.

Thanks,
--
Ming Lei

[PATCH 1/2] blk-mq: move blk_mq_get_ctx/blk_mq_put_ctx to mq private header

[Ming Lei]

The blk-mq tag code need these helpers.

Signed-off-by: Ming Lei <tom.leiming@gmail.com>
---
 block/blk-mq.c |   22 ----------------------
 block/blk-mq.h |   22 ++++++++++++++++++++++
 2 files changed, 22 insertions(+), 22 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 0f5879c..b9230c5 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -33,28 +33,6 @@ static LIST_HEAD(all_q_list);
 
 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
 
-static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
-					   unsigned int cpu)
-{
-	return per_cpu_ptr(q->queue_ctx, cpu);
-}
-
-/*
- * This assumes per-cpu software queueing queues. They could be per-node
- * as well, for instance. For now this is hardcoded as-is. Note that we don't
- * care about preemption, since we know the ctx's are persistent. This does
- * mean that we can't rely on ctx always matching the currently running CPU.
- */
-static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
-{
-	return __blk_mq_get_ctx(q, get_cpu());
-}
-
-static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
-{
-	put_cpu();
-}
-
 /*
  * Check if any of the ctx's have pending work in this hardware queue
  */
diff --git a/block/blk-mq.h b/block/blk-mq.h
index de7b3bb..57a7968 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -69,4 +69,26 @@ struct blk_align_bitmap {
 	unsigned long depth;
 } ____cacheline_aligned_in_smp;
 
+static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
+					   unsigned int cpu)
+{
+	return per_cpu_ptr(q->queue_ctx, cpu);
+}
+
+/*
+ * This assumes per-cpu software queueing queues. They could be per-node
+ * as well, for instance. For now this is hardcoded as-is. Note that we don't
+ * care about preemption, since we know the ctx's are persistent. This does
+ * mean that we can't rely on ctx always matching the currently running CPU.
+ */
+static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
+{
+	return __blk_mq_get_ctx(q, get_cpu());
+}
+
+static inline void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
+{
+	put_cpu();
+}
+
 #endif
-- 
1.7.9.5

[PATCH 2/2] blk-mq: fix schedule from atomic context

[Ming Lei]

blk_mq_put_ctx() has to be called before io_schedule() in
bt_get().

This patch fixes the problem by taking similar approach from
percpu_ida allocation for the situation.

Signed-off-by: Ming Lei <tom.leiming@gmail.com>
---
 block/blk-mq-tag.c |   48 +++++++++++++++++++++++++++++++-----------------
 block/blk-mq-tag.h |    2 +-
 block/blk-mq.c     |   36 +++++++++++++++++++++++-------------
 block/blk-mq.h     |   23 +++++++++++++++++++++++
 4 files changed, 78 insertions(+), 31 deletions(-)

diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index d90c4ae..1aab39f 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -221,8 +221,10 @@ static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
 	return bs;
 }
 
-static int bt_get(struct blk_mq_bitmap_tags *bt, struct blk_mq_hw_ctx *hctx,
-		  unsigned int *last_tag, gfp_t gfp)
+static int bt_get(struct blk_mq_alloc_data *data,
+		struct blk_mq_bitmap_tags *bt,
+		struct blk_mq_hw_ctx *hctx,
+		unsigned int *last_tag)
 {
 	struct bt_wait_state *bs;
 	DEFINE_WAIT(wait);
@@ -232,7 +234,7 @@ static int bt_get(struct blk_mq_bitmap_tags *bt, struct blk_mq_hw_ctx *hctx,
 	if (tag != -1)
 		return tag;
 
-	if (!(gfp & __GFP_WAIT))
+	if (!(data->gfp & __GFP_WAIT))
 		return -1;
 
 	bs = bt_wait_ptr(bt, hctx);
@@ -249,50 +251,62 @@ static int bt_get(struct blk_mq_bitmap_tags *bt, struct blk_mq_hw_ctx *hctx,
 		if (was_empty)
 			atomic_set(&bs->wait_cnt, bt->wake_cnt);
 
+		blk_mq_put_ctx(data->ctx);
+
 		io_schedule();
+
+		data->ctx = blk_mq_get_ctx(data->q);
+		data->hctx = data->q->mq_ops->map_queue(data->q,
+				data->ctx->cpu);
+		if (data->reserved) {
+			bt = &data->hctx->tags->breserved_tags;
+		} else {
+			last_tag = &data->ctx->last_tag;
+			hctx = data->hctx;
+			bt = &hctx->tags->bitmap_tags;
+		}
+		finish_wait(&bs->wait, &wait);
+		bs = bt_wait_ptr(bt, hctx);
 	} while (1);
 
 	finish_wait(&bs->wait, &wait);
 	return tag;
 }
 
-static unsigned int __blk_mq_get_tag(struct blk_mq_tags *tags,
-				     struct blk_mq_hw_ctx *hctx,
-				     unsigned int *last_tag, gfp_t gfp)
+static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
 {
 	int tag;
 
-	tag = bt_get(&tags->bitmap_tags, hctx, last_tag, gfp);
+	tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
+			&data->ctx->last_tag);
 	if (tag >= 0)
-		return tag + tags->nr_reserved_tags;
+		return tag + data->hctx->tags->nr_reserved_tags;
 
 	return BLK_MQ_TAG_FAIL;
 }
 
-static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_tags *tags,
-					      gfp_t gfp)
+static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
 {
 	int tag, zero = 0;
 
-	if (unlikely(!tags->nr_reserved_tags)) {
+	if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
 		WARN_ON_ONCE(1);
 		return BLK_MQ_TAG_FAIL;
 	}
 
-	tag = bt_get(&tags->breserved_tags, NULL, &zero, gfp);
+	tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero);
 	if (tag < 0)
 		return BLK_MQ_TAG_FAIL;
 
 	return tag;
 }
 
-unsigned int blk_mq_get_tag(struct blk_mq_hw_ctx *hctx, unsigned int *last_tag,
-			    gfp_t gfp, bool reserved)
+unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
 {
-	if (!reserved)
-		return __blk_mq_get_tag(hctx->tags, hctx, last_tag, gfp);
+	if (!data->reserved)
+		return __blk_mq_get_tag(data);
 
-	return __blk_mq_get_reserved_tag(hctx->tags, gfp);
+	return __blk_mq_get_reserved_tag(data);
 }
 
 static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
diff --git a/block/blk-mq-tag.h b/block/blk-mq-tag.h
index c959de5..98696a6 100644
--- a/block/blk-mq-tag.h
+++ b/block/blk-mq-tag.h
@@ -48,7 +48,7 @@ struct blk_mq_tags {
 extern struct blk_mq_tags *blk_mq_init_tags(unsigned int nr_tags, unsigned int reserved_tags, int node);
 extern void blk_mq_free_tags(struct blk_mq_tags *tags);
 
-extern unsigned int blk_mq_get_tag(struct blk_mq_hw_ctx *hctx, unsigned int *last_tag, gfp_t gfp, bool reserved);
+extern unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data);
 extern void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag, unsigned int *last_tag);
 extern bool blk_mq_has_free_tags(struct blk_mq_tags *tags);
 extern ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page);
diff --git a/block/blk-mq.c b/block/blk-mq.c
index b9230c5..43eb315 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -210,24 +210,23 @@ static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 }
 
 static struct request *
-__blk_mq_alloc_request(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
-		struct blk_mq_ctx *ctx, int rw, gfp_t gfp, bool reserved)
+__blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 {
 	struct request *rq;
 	unsigned int tag;
 
-	tag = blk_mq_get_tag(hctx, &ctx->last_tag, gfp, reserved);
+	tag = blk_mq_get_tag(data);
 	if (tag != BLK_MQ_TAG_FAIL) {
-		rq = hctx->tags->rqs[tag];
+		rq = data->hctx->tags->rqs[tag];
 
 		rq->cmd_flags = 0;
-		if (blk_mq_tag_busy(hctx)) {
+		if (blk_mq_tag_busy(data->hctx)) {
 			rq->cmd_flags = REQ_MQ_INFLIGHT;
-			atomic_inc(&hctx->nr_active);
+			atomic_inc(&data->hctx->nr_active);
 		}
 
 		rq->tag = tag;
-		blk_mq_rq_ctx_init(q, ctx, rq, rw);
+		blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 		return rq;
 	}
 
@@ -240,22 +239,27 @@ struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 	struct blk_mq_ctx *ctx;
 	struct blk_mq_hw_ctx *hctx;
 	struct request *rq;
+	struct blk_mq_alloc_data alloc_data;
 
 	if (blk_mq_queue_enter(q))
 		return NULL;
 
 	ctx = blk_mq_get_ctx(q);
 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
+	blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
+			reserved, ctx, hctx);
 
-	rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp & ~__GFP_WAIT,
-				    reserved);
+	rq = __blk_mq_alloc_request(&alloc_data, rw);
 	if (!rq && (gfp & __GFP_WAIT)) {
 		__blk_mq_run_hw_queue(hctx);
 		blk_mq_put_ctx(ctx);
 
 		ctx = blk_mq_get_ctx(q);
 		hctx = q->mq_ops->map_queue(q, ctx->cpu);
-		rq =  __blk_mq_alloc_request(q, hctx, ctx, rw, gfp, reserved);
+		blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
+				hctx);
+		rq =  __blk_mq_alloc_request(&alloc_data, rw);
+		ctx = alloc_data.ctx;
 	}
 	blk_mq_put_ctx(ctx);
 	return rq;
@@ -1136,6 +1140,7 @@ static struct request *blk_mq_map_request(struct request_queue *q,
 	struct blk_mq_ctx *ctx;
 	struct request *rq;
 	int rw = bio_data_dir(bio);
+	struct blk_mq_alloc_data alloc_data;
 
 	if (unlikely(blk_mq_queue_enter(q))) {
 		bio_endio(bio, -EIO);
@@ -1149,7 +1154,9 @@ static struct request *blk_mq_map_request(struct request_queue *q,
 		rw |= REQ_SYNC;
 
 	trace_block_getrq(q, bio, rw);
-	rq = __blk_mq_alloc_request(q, hctx, ctx, rw, GFP_ATOMIC, false);
+	blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
+			hctx);
+	rq = __blk_mq_alloc_request(&alloc_data, rw);
 	if (unlikely(!rq)) {
 		__blk_mq_run_hw_queue(hctx);
 		blk_mq_put_ctx(ctx);
@@ -1157,8 +1164,11 @@ static struct request *blk_mq_map_request(struct request_queue *q,
 
 		ctx = blk_mq_get_ctx(q);
 		hctx = q->mq_ops->map_queue(q, ctx->cpu);
-		rq = __blk_mq_alloc_request(q, hctx, ctx, rw,
-					    __GFP_WAIT|GFP_ATOMIC, false);
+		blk_mq_set_alloc_data(&alloc_data, q,
+				__GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
+		rq = __blk_mq_alloc_request(&alloc_data, rw);
+		ctx = alloc_data.ctx;
+		hctx = alloc_data.hctx;
 	}
 
 	hctx->queued++;
diff --git a/block/blk-mq.h b/block/blk-mq.h
index 57a7968..2646088 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -91,4 +91,27 @@ static inline void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
 	put_cpu();
 }
 
+struct blk_mq_alloc_data {
+	/* input parameter */
+	struct request_queue *q;
+	gfp_t gfp;
+	bool reserved;
+
+	/* input & output parameter */
+	struct blk_mq_ctx *ctx;
+	struct blk_mq_hw_ctx *hctx;
+};
+
+static inline void blk_mq_set_alloc_data(struct blk_mq_alloc_data *data,
+		struct request_queue *q, gfp_t gfp, bool reserved,
+		struct blk_mq_ctx *ctx,
+		struct blk_mq_hw_ctx *hctx)
+{
+	data->q = q;
+	data->gfp = gfp;
+	data->reserved = reserved;
+	data->ctx = ctx;
+	data->hctx = hctx;
+}
+
 #endif
-- 
1.7.9.5

Re: [PATCH 2/2] blk-mq: fix schedule from atomic context

[Jens Axboe]

On 2014-05-31 10:43, Ming Lei wrote:
> blk_mq_put_ctx() has to be called before io_schedule() in
> bt_get().
>
> This patch fixes the problem by taking similar approach from
> percpu_ida allocation for the situation.

Thanks, I've applied these to for post-merge integration. I do think we 
can clean up further around 'data' for bt_get(), for instance, and 
possibly in other places too.

-- 
Jens Axboe