[PATCH] sbitmap: only queue kyber's wait callback if not already active

[PATCH] sbitmap: only queue kyber's wait callback if not already active

[David Jeffery]

Under heavy loads where the kyber I/O scheduler hits the token limits for
its scheduling domains, kyber can become stuck.  When active requests
complete, kyber may not be woken up leaving the I/O requests in kyber
stuck.

This stuck state is due to a race condition with kyber and the sbitmap
functions it uses to run a callback when enough requests have completed.
The running of a sbt_wait callback can race with the attempt to insert the
sbt_wait.  Since sbitmap_del_wait_queue removes the sbt_wait from the list
first then sets the sbq field to NULL, kyber can see the item as not on a
list but the call to sbitmap_add_wait_queue will see sbq as non-NULL. This
results in the sbt_wait being inserted onto the wait list but ws_active
doesn't get incremented.  So the sbitmap queue does not know there is a
waiter on a wait list.

Since sbitmap doesn't think there is a waiter, kyber may never be
informed that there are domain tokens available and the I/O never advances.
With the sbt_wait on a wait list, kyber believes it has an active waiter
so cannot insert a new waiter when reaching the domain's full state.

This race can be fixed by only adding the sbt_wait to the queue if the
sbq field is NULL.  If sbq is not NULL, there is already an action active
which will trigger the re-running of kyber.  Let it run and add the
sbt_wait to the wait list if still needing to wait.

Signed-off-by: David Jeffery <djeffery@redhat.com>
Reported-by: John Pittman <jpittman@redhat.com>
Tested-by: John Pittman <jpittman@redhat.com>
---

This bug was reliably being triggered on several test systems.  With the
fix, the tests no longer fail.

 sbitmap.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/lib/sbitmap.c b/lib/sbitmap.c
index 33feec8989f1..af88d1346dd7 100644
--- a/lib/sbitmap.c
+++ b/lib/sbitmap.c
@@ -650,8 +650,8 @@ void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
 	if (!sbq_wait->sbq) {
 		sbq_wait->sbq = sbq;
 		atomic_inc(&sbq->ws_active);
+		add_wait_queue(&ws->wait, &sbq_wait->wait);
 	}
-	add_wait_queue(&ws->wait, &sbq_wait->wait);
 }
 EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
 

Re: [PATCH] sbitmap: only queue kyber's wait callback if not already active

[Jens Axboe]

On 12/17/19 9:00 AM, David Jeffery wrote:
> Under heavy loads where the kyber I/O scheduler hits the token limits for
> its scheduling domains, kyber can become stuck.  When active requests
> complete, kyber may not be woken up leaving the I/O requests in kyber
> stuck.
> 
> This stuck state is due to a race condition with kyber and the sbitmap
> functions it uses to run a callback when enough requests have completed.
> The running of a sbt_wait callback can race with the attempt to insert the
> sbt_wait.  Since sbitmap_del_wait_queue removes the sbt_wait from the list
> first then sets the sbq field to NULL, kyber can see the item as not on a
> list but the call to sbitmap_add_wait_queue will see sbq as non-NULL. This
> results in the sbt_wait being inserted onto the wait list but ws_active
> doesn't get incremented.  So the sbitmap queue does not know there is a
> waiter on a wait list.
> 
> Since sbitmap doesn't think there is a waiter, kyber may never be
> informed that there are domain tokens available and the I/O never advances.
> With the sbt_wait on a wait list, kyber believes it has an active waiter
> so cannot insert a new waiter when reaching the domain's full state.
> 
> This race can be fixed by only adding the sbt_wait to the queue if the
> sbq field is NULL.  If sbq is not NULL, there is already an action active
> which will trigger the re-running of kyber.  Let it run and add the
> sbt_wait to the wait list if still needing to wait.
> 
> Signed-off-by: David Jeffery <djeffery@redhat.com>
> Reported-by: John Pittman <jpittman@redhat.com>
> Tested-by: John Pittman <jpittman@redhat.com>
> ---
> 
> This bug was reliably being triggered on several test systems.  With the
> fix, the tests no longer fail.
> 
>  sbitmap.c |    2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/lib/sbitmap.c b/lib/sbitmap.c
> index 33feec8989f1..af88d1346dd7 100644
> --- a/lib/sbitmap.c
> +++ b/lib/sbitmap.c
> @@ -650,8 +650,8 @@ void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
>  	if (!sbq_wait->sbq) {
>  		sbq_wait->sbq = sbq;
>  		atomic_inc(&sbq->ws_active);
> +		add_wait_queue(&ws->wait, &sbq_wait->wait);
>  	}
> -	add_wait_queue(&ws->wait, &sbq_wait->wait);
>  }
>  EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);

This looks good to me, waiting for Omar to take a look (CC'ed).

-- 
Jens Axboe

Re: [PATCH] sbitmap: only queue kyber's wait callback if not already active

[Omar Sandoval]

On Tue, Dec 17, 2019 at 11:00:24AM -0500, David Jeffery wrote:
> Under heavy loads where the kyber I/O scheduler hits the token limits for
> its scheduling domains, kyber can become stuck.  When active requests
> complete, kyber may not be woken up leaving the I/O requests in kyber
> stuck.
> 
> This stuck state is due to a race condition with kyber and the sbitmap
> functions it uses to run a callback when enough requests have completed.
> The running of a sbt_wait callback can race with the attempt to insert the
> sbt_wait.  Since sbitmap_del_wait_queue removes the sbt_wait from the list
> first then sets the sbq field to NULL, kyber can see the item as not on a
> list but the call to sbitmap_add_wait_queue will see sbq as non-NULL. This
> results in the sbt_wait being inserted onto the wait list but ws_active
> doesn't get incremented.  So the sbitmap queue does not know there is a
> waiter on a wait list.
> 
> Since sbitmap doesn't think there is a waiter, kyber may never be
> informed that there are domain tokens available and the I/O never advances.
> With the sbt_wait on a wait list, kyber believes it has an active waiter
> so cannot insert a new waiter when reaching the domain's full state.
> 
> This race can be fixed by only adding the sbt_wait to the queue if the
> sbq field is NULL.  If sbq is not NULL, there is already an action active
> which will trigger the re-running of kyber.  Let it run and add the
> sbt_wait to the wait list if still needing to wait.

So the race here is:

Thread 1                        Thread 2
kyber_domain_wake
  sbitmap_del_wait_queue
    list_del_init
    atomic_dec sbq->ws_active
                                kyber_get_domain_token
                                  list_empty_careful
                                  sbitmap_add_wait_queue
                                    if (!sqb_wait->sb) // false
                                    add_wait_queue
    sbq_wait->sbq = NULL

Now sbq_wait->sbq == NULL, sbq->ws_active = 0, and
!list_empty(domain_wait), so sbq_wake_ptr returns NULL and
sbitmap_queue_wake_up does nothing.

I get the feeling that sbitmap_{add,del}_wait_queue need some memory
barriers... But ignoring that, this fix seems right.

Reviewed-by: Omar Sandoval <osandov@fb.com>

P.S. s/sbt_wait/sbq_wait/g in the commit message.