diff mbox

[1/1] null_blk: fix handling of BLKPREP_DEFER case

Message ID CAJN_NGbXH0A8i9OF5Y95hsgCqdrE-CU5dRBsa6iNTKQPz+g+oA@mail.gmail.com (mailing list archive)
State New, archived
Headers show

Commit Message

Dmitry Krivenok April 19, 2015, 6:19 p.m. UTC
When we fail to allocate new cmd in null_rq_prep_fn we return BLKPREP_DEFER
which is not handled properly. In single-queue mode of null_blk the following
command hangs forever in io_schedule():
$ dd if=/dev/nullb0 of=/dev/null bs=8M count=5000 iflag=direct

The reason is that when 64 commands have been allocated, the 65th command
allocation will fail due to missing free tag. The request, however, will be
kept in the queue which will never be started again (unless you run another
command that does I/O to /dev/nullb0).

This small patch tries to solve the issue by stopping the queue when we
detect that all tags were exhausted and starting it again when we free the tag.

I've verified that the command mentioned above doesn't hang anymore and also
made sure that null_blk with my change survives fio-based stress tests.

Signed-off-by: Dmitry V. Krivenok <krivenok.dmitry@gmail.com>
---
 drivers/block/null_blk.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)


@@ -341,6 +352,7 @@ static int null_rq_prep_fn(struct request_queue
*q, struct request *req)
 static void null_request_fn(struct request_queue *q)
 {
     struct request *rq;
+    struct nullb_queue *nq = nullb_to_queue(q->queuedata);

     while ((rq = blk_fetch_request(q)) != NULL) {
         struct nullb_cmd *cmd = rq->special;
@@ -349,6 +361,9 @@ static void null_request_fn(struct request_queue *q)
         null_handle_cmd(cmd);
         spin_lock_irq(q->queue_lock);
     }
+    if(nq->no_cmds) {
+        blk_stop_queue(q);
+    }
 }

 static int null_queue_rq(struct blk_mq_hw_ctx *hctx,
@@ -430,6 +445,7 @@ static int setup_commands(struct nullb_queue *nq)
     if (!nq->cmds)
         return -ENOMEM;

+    nq->no_cmds = false;
     tag_size = ALIGN(nq->queue_depth, BITS_PER_LONG) / BITS_PER_LONG;
     nq->tag_map = kzalloc(tag_size * sizeof(unsigned long), GFP_KERNEL);
     if (!nq->tag_map) {

Comments

Dmitry Krivenok April 27, 2015, 4:52 p.m. UTC | #1
Folks, any comments on this patch?

Thanks,
Dmitry

On Sun, Apr 19, 2015 at 9:19 PM, Dmitry Krivenok
<krivenok.dmitry@gmail.com> wrote:
> When we fail to allocate new cmd in null_rq_prep_fn we return BLKPREP_DEFER
> which is not handled properly. In single-queue mode of null_blk the following
> command hangs forever in io_schedule():
> $ dd if=/dev/nullb0 of=/dev/null bs=8M count=5000 iflag=direct
>
> The reason is that when 64 commands have been allocated, the 65th command
> allocation will fail due to missing free tag. The request, however, will be
> kept in the queue which will never be started again (unless you run another
> command that does I/O to /dev/nullb0).
>
> This small patch tries to solve the issue by stopping the queue when we
> detect that all tags were exhausted and starting it again when we free the tag.
>
> I've verified that the command mentioned above doesn't hang anymore and also
> made sure that null_blk with my change survives fio-based stress tests.
>
> Signed-off-by: Dmitry V. Krivenok <krivenok.dmitry@gmail.com>
> ---
>  drivers/block/null_blk.c | 16 ++++++++++++++++
>  1 file changed, 16 insertions(+)
>
> diff --git a/drivers/block/null_blk.c b/drivers/block/null_blk.c
> index 65cd61a..4ac684b 100644
> --- a/drivers/block/null_blk.c
> +++ b/drivers/block/null_blk.c
> @@ -25,6 +25,7 @@ struct nullb_queue {
>      unsigned int queue_depth;
>
>      struct nullb_cmd *cmds;
> +    bool no_cmds;
>  };
>
>  struct nullb {
> @@ -171,6 +172,13 @@ static unsigned int get_tag(struct nullb_queue *nq)
>  static void free_cmd(struct nullb_cmd *cmd)
>  {
>      put_tag(cmd->nq, cmd->tag);
> +    if(cmd->nq->no_cmds) {
> +        unsigned long flags;
> +        cmd->nq->no_cmds = false;
> +        spin_lock_irqsave(cmd->rq->q->queue_lock, flags);
> +        blk_start_queue(cmd->rq->q);
> +        spin_unlock_irqrestore(cmd->rq->q->queue_lock, flags);
> +    }
>  }
>
>  static struct nullb_cmd *__alloc_cmd(struct nullb_queue *nq)
> @@ -195,6 +203,9 @@ static struct nullb_cmd *alloc_cmd(struct
> nullb_queue *nq, int can_wait)
>      DEFINE_WAIT(wait);
>
>      cmd = __alloc_cmd(nq);
> +    if (!cmd && !can_wait) {
> +        nq->no_cmds = true;
> +    }
>      if (cmd || !can_wait)
>          return cmd;
>
> @@ -341,6 +352,7 @@ static int null_rq_prep_fn(struct request_queue
> *q, struct request *req)
>  static void null_request_fn(struct request_queue *q)
>  {
>      struct request *rq;
> +    struct nullb_queue *nq = nullb_to_queue(q->queuedata);
>
>      while ((rq = blk_fetch_request(q)) != NULL) {
>          struct nullb_cmd *cmd = rq->special;
> @@ -349,6 +361,9 @@ static void null_request_fn(struct request_queue *q)
>          null_handle_cmd(cmd);
>          spin_lock_irq(q->queue_lock);
>      }
> +    if(nq->no_cmds) {
> +        blk_stop_queue(q);
> +    }
>  }
>
>  static int null_queue_rq(struct blk_mq_hw_ctx *hctx,
> @@ -430,6 +445,7 @@ static int setup_commands(struct nullb_queue *nq)
>      if (!nq->cmds)
>          return -ENOMEM;
>
> +    nq->no_cmds = false;
>      tag_size = ALIGN(nq->queue_depth, BITS_PER_LONG) / BITS_PER_LONG;
>      nq->tag_map = kzalloc(tag_size * sizeof(unsigned long), GFP_KERNEL);
>      if (!nq->tag_map) {
> --
> 2.3.5
Jens Axboe April 27, 2015, 5:25 p.m. UTC | #2
On 04/19/2015 12:19 PM, Dmitry Krivenok wrote:
> When we fail to allocate new cmd in null_rq_prep_fn we return BLKPREP_DEFER
> which is not handled properly. In single-queue mode of null_blk the following
> command hangs forever in io_schedule():
> $ dd if=/dev/nullb0 of=/dev/null bs=8M count=5000 iflag=direct
>
> The reason is that when 64 commands have been allocated, the 65th command
> allocation will fail due to missing free tag. The request, however, will be
> kept in the queue which will never be started again (unless you run another
> command that does I/O to /dev/nullb0).
>
> This small patch tries to solve the issue by stopping the queue when we
> detect that all tags were exhausted and starting it again when we free the tag.
>
> I've verified that the command mentioned above doesn't hang anymore and also
> made sure that null_blk with my change survives fio-based stress tests.

You are right, legacy request_fn mode has a bug there. I'd get rid of 
the no_cmds bool, though. If we fail allocating a command in alloc_cmd() 
and we're in NULL_Q_RQ mode, stop the queue. In free_cmd(), again if 
we're in NULL_Q_RQ_MODE, check the stopped flag and start the queue if 
it is set.
diff mbox

Patch

diff --git a/drivers/block/null_blk.c b/drivers/block/null_blk.c
index 65cd61a..4ac684b 100644
--- a/drivers/block/null_blk.c
+++ b/drivers/block/null_blk.c
@@ -25,6 +25,7 @@  struct nullb_queue {
     unsigned int queue_depth;

     struct nullb_cmd *cmds;
+    bool no_cmds;
 };

 struct nullb {
@@ -171,6 +172,13 @@  static unsigned int get_tag(struct nullb_queue *nq)
 static void free_cmd(struct nullb_cmd *cmd)
 {
     put_tag(cmd->nq, cmd->tag);
+    if(cmd->nq->no_cmds) {
+        unsigned long flags;
+        cmd->nq->no_cmds = false;
+        spin_lock_irqsave(cmd->rq->q->queue_lock, flags);
+        blk_start_queue(cmd->rq->q);
+        spin_unlock_irqrestore(cmd->rq->q->queue_lock, flags);
+    }
 }

 static struct nullb_cmd *__alloc_cmd(struct nullb_queue *nq)
@@ -195,6 +203,9 @@  static struct nullb_cmd *alloc_cmd(struct
nullb_queue *nq, int can_wait)
     DEFINE_WAIT(wait);

     cmd = __alloc_cmd(nq);
+    if (!cmd && !can_wait) {
+        nq->no_cmds = true;
+    }
     if (cmd || !can_wait)
         return cmd;