Re: [PATCH 1/2] block: Avoid invoking blk_run_queue() recursively

[Bart Van Assche <bvanassche@acm.org>]

On 02/22/13 20:01, Jens Axboe wrote:
> That'd be fine. I agree with Tejun that just returning is error prone
> and could cause hard-to-debug hangs or stalls. So something ala
>
>          if (blk_queue(dead))
>                  return;
>          else if (q->request_fn_active) {
>                  blk_delay_queue(q, 0);
>                  return;
>          }
>
> would work. Alternatively, you could mark the queue as needing a re-run,
> so any existing runner of it would notice and clear this flag and re-run
> the queue. But that's somewhat more fragile, and since it isn't a
> hot/performance path, I suspect the simple "just re-run me soon" is good
> enough.

I'm worried that in the device mapper the approach using 
blk_delay_queue() could trigger a queue run after the final dm_put(). 
And I'm also afraid that that approach could trigger several needless 
context switches before q->request_fn_active finally drops to zero. So 
the approach with the "need re-run" flag seems preferable to me. But 
maybe I'm overlooking something ?

Bart.