[PATCH 0/2] vhost_worker fixes

[PATCH 0/2] vhost_worker fixes

[Bart Van Assche]

There are two patches in this series:
* A patch that reduces vhost_work_flush() latency on busy systems.
* A patch that avoids that vhost_work_flush() returns early.

I found these issues via code inspection.

[PATCH 1/2] vhost: Reduce vhost_work_flush() wakeup latency

[Bart Van Assche]

If the TIF_NEED_RESCHED task flag is set, wake up any vhost_work_flush()
waiters before rescheduling instead of after rescheduling.

Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Asias He <asias@redhat.com>
---
 drivers/vhost/vhost.c |   42 +++++++++++++++++++-----------------------
 1 file changed, 19 insertions(+), 23 deletions(-)

diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index e58cf00..e7ffc10 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -201,47 +201,43 @@ static void vhost_vq_reset(struct vhost_dev *dev,
 static int vhost_worker(void *data)
 {
 	struct vhost_dev *dev = data;
-	struct vhost_work *work = NULL;
-	unsigned uninitialized_var(seq);
+	struct vhost_work *work;
+	unsigned seq;
 	mm_segment_t oldfs = get_fs();
 
 	set_fs(USER_DS);
 	use_mm(dev->mm);
 
-	for (;;) {
+	spin_lock_irq(&dev->work_lock);
+	while (!kthread_should_stop()) {
 		/* mb paired w/ kthread_stop */
 		set_current_state(TASK_INTERRUPTIBLE);
-
-		spin_lock_irq(&dev->work_lock);
-		if (work) {
-			work->done_seq = seq;
-			if (work->flushing)
-				wake_up_all(&work->done);
-		}
-
-		if (kthread_should_stop()) {
-			spin_unlock_irq(&dev->work_lock);
-			__set_current_state(TASK_RUNNING);
-			break;
-		}
 		if (!list_empty(&dev->work_list)) {
 			work = list_first_entry(&dev->work_list,
 						struct vhost_work, node);
 			list_del_init(&work->node);
 			seq = work->queue_seq;
-		} else
-			work = NULL;
-		spin_unlock_irq(&dev->work_lock);
+			spin_unlock_irq(&dev->work_lock);
 
-		if (work) {
 			__set_current_state(TASK_RUNNING);
 			work->fn(work);
-			if (need_resched())
-				schedule();
-		} else
+
+			spin_lock_irq(&dev->work_lock);
+			work->done_seq = seq;
+			if (work->flushing)
+				wake_up_all(&work->done);
+		}
+		if (list_empty(&dev->work_list) || need_resched()) {
+			spin_unlock_irq(&dev->work_lock);
+
 			schedule();
 
+			spin_lock_irq(&dev->work_lock);
+		}
 	}
+	spin_unlock_irq(&dev->work_lock);
+
+	__set_current_state(TASK_RUNNING);
 	unuse_mm(dev->mm);
 	set_fs(oldfs);
 	return 0;
-- 
1.7.10.4

Re: [PATCH 1/2] vhost: Reduce vhost_work_flush() wakeup latency

[Michael S. Tsirkin]

On Wed, Aug 14, 2013 at 09:02:32AM +0200, Bart Van Assche wrote:
> If the TIF_NEED_RESCHED task flag is set, wake up any vhost_work_flush()
> waiters before rescheduling instead of after rescheduling.
> 
> Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> Cc: Michael S. Tsirkin <mst@redhat.com>
> Cc: Asias He <asias@redhat.com>

Why exactly? It's not like flush needs to be extra fast ...

> ---
>  drivers/vhost/vhost.c |   42 +++++++++++++++++++-----------------------
>  1 file changed, 19 insertions(+), 23 deletions(-)
> 
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index e58cf00..e7ffc10 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -201,47 +201,43 @@ static void vhost_vq_reset(struct vhost_dev *dev,
>  static int vhost_worker(void *data)
>  {
>  	struct vhost_dev *dev = data;
> -	struct vhost_work *work = NULL;
> -	unsigned uninitialized_var(seq);
> +	struct vhost_work *work;
> +	unsigned seq;
>  	mm_segment_t oldfs = get_fs();
>  
>  	set_fs(USER_DS);
>  	use_mm(dev->mm);
>  
> -	for (;;) {
> +	spin_lock_irq(&dev->work_lock);
> +	while (!kthread_should_stop()) {
>  		/* mb paired w/ kthread_stop */
>  		set_current_state(TASK_INTERRUPTIBLE);
> -
> -		spin_lock_irq(&dev->work_lock);
> -		if (work) {
> -			work->done_seq = seq;
> -			if (work->flushing)
> -				wake_up_all(&work->done);
> -		}
> -
> -		if (kthread_should_stop()) {
> -			spin_unlock_irq(&dev->work_lock);
> -			__set_current_state(TASK_RUNNING);
> -			break;
> -		}
>  		if (!list_empty(&dev->work_list)) {
>  			work = list_first_entry(&dev->work_list,
>  						struct vhost_work, node);
>  			list_del_init(&work->node);
>  			seq = work->queue_seq;
> -		} else
> -			work = NULL;
> -		spin_unlock_irq(&dev->work_lock);
> +			spin_unlock_irq(&dev->work_lock);
>  
> -		if (work) {
>  			__set_current_state(TASK_RUNNING);
>  			work->fn(work);
> -			if (need_resched())
> -				schedule();
> -		} else
> +
> +			spin_lock_irq(&dev->work_lock);
> +			work->done_seq = seq;
> +			if (work->flushing)
> +				wake_up_all(&work->done);
> +		}
> +		if (list_empty(&dev->work_list) || need_resched()) {
> +			spin_unlock_irq(&dev->work_lock);
> +
>  			schedule();
>  
> +			spin_lock_irq(&dev->work_lock);
> +		}
>  	}
> +	spin_unlock_irq(&dev->work_lock);
> +
> +	__set_current_state(TASK_RUNNING);
>  	unuse_mm(dev->mm);
>  	set_fs(oldfs);
>  	return 0;
> -- 
> 1.7.10.4

Re: [PATCH 1/2] vhost: Reduce vhost_work_flush() wakeup latency

[Bart Van Assche]

On 08/14/13 13:37, Michael S. Tsirkin wrote:
> On Wed, Aug 14, 2013 at 09:02:32AM +0200, Bart Van Assche wrote:
>> If the TIF_NEED_RESCHED task flag is set, wake up any vhost_work_flush()
>> waiters before rescheduling instead of after rescheduling.
>>
>> Signed-off-by: Bart Van Assche <bvanassche@acm.org>
>> Cc: Michael S. Tsirkin <mst@redhat.com>
>> Cc: Asias He <asias@redhat.com>
>
> Why exactly? It's not like flush needs to be extra fast ...

I'm not worried about how fast a flush is processed either. But I think 
this patch is a nice cleanup of the vhost_worker() function. It 
eliminates the uninitialized_var() construct and moves the assignment of 
"done_seq" below the read of "queue_seq".

Bart.

Re: [PATCH 1/2] vhost: Reduce vhost_work_flush() wakeup latency

[Michael S. Tsirkin]

On Wed, Aug 14, 2013 at 05:22:36PM +0200, Bart Van Assche wrote:
> On 08/14/13 13:37, Michael S. Tsirkin wrote:
> >On Wed, Aug 14, 2013 at 09:02:32AM +0200, Bart Van Assche wrote:
> >>If the TIF_NEED_RESCHED task flag is set, wake up any vhost_work_flush()
> >>waiters before rescheduling instead of after rescheduling.
> >>
> >>Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> >>Cc: Michael S. Tsirkin <mst@redhat.com>
> >>Cc: Asias He <asias@redhat.com>
> >
> >Why exactly? It's not like flush needs to be extra fast ...
> 
> I'm not worried about how fast a flush is processed either. But I
> think this patch is a nice cleanup of the vhost_worker() function.
> It eliminates the uninitialized_var() construct and moves the
> assignment of "done_seq" below the read of "queue_seq".
> 
> Bart.

I'm not worried about uninitialized_var - it's just a compiler bug.
done_seq below read is nice, but this is at the cost of sprinkling
lock/unlock, lock/unlock all over the code.  Maybe we can come up with
something that doesn't have this property.

-- 
MST

Re: [PATCH 1/2] vhost: Reduce vhost_work_flush() wakeup latency

[Asias He]

On Wed, Aug 14, 2013 at 08:58:25PM +0300, Michael S. Tsirkin wrote:
> On Wed, Aug 14, 2013 at 05:22:36PM +0200, Bart Van Assche wrote:
> > On 08/14/13 13:37, Michael S. Tsirkin wrote:
> > >On Wed, Aug 14, 2013 at 09:02:32AM +0200, Bart Van Assche wrote:
> > >>If the TIF_NEED_RESCHED task flag is set, wake up any vhost_work_flush()
> > >>waiters before rescheduling instead of after rescheduling.
> > >>
> > >>Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> > >>Cc: Michael S. Tsirkin <mst@redhat.com>
> > >>Cc: Asias He <asias@redhat.com>
> > >
> > >Why exactly? It's not like flush needs to be extra fast ...
> > 
> > I'm not worried about how fast a flush is processed either. But I
> > think this patch is a nice cleanup of the vhost_worker() function.
> > It eliminates the uninitialized_var() construct and moves the
> > assignment of "done_seq" below the read of "queue_seq".
> > 
> > Bart.
> 
> I'm not worried about uninitialized_var - it's just a compiler bug.
> done_seq below read is nice, but this is at the cost of sprinkling
> lock/unlock, lock/unlock all over the code.  Maybe we can come up with
> something that doesn't have this property.

The extra locking introduced here does not look good to me neither.

> -- 
> MST

-- 
Asias

Re: [PATCH 1/2] vhost: Reduce vhost_work_flush() wakeup latency

[Bart Van Assche]

On 08/15/13 03:30, Asias He wrote:
> On Wed, Aug 14, 2013 at 08:58:25PM +0300, Michael S. Tsirkin wrote:
>> On Wed, Aug 14, 2013 at 05:22:36PM +0200, Bart Van Assche wrote:
>>> On 08/14/13 13:37, Michael S. Tsirkin wrote:
>>>> On Wed, Aug 14, 2013 at 09:02:32AM +0200, Bart Van Assche wrote:
>>>>> If the TIF_NEED_RESCHED task flag is set, wake up any vhost_work_flush()
>>>>> waiters before rescheduling instead of after rescheduling.
>>>>>
>>>>> Signed-off-by: Bart Van Assche <bvanassche@acm.org>
>>>>> Cc: Michael S. Tsirkin <mst@redhat.com>
>>>>> Cc: Asias He <asias@redhat.com>
>>>>
>>>> Why exactly? It's not like flush needs to be extra fast ...
>>>
>>> I'm not worried about how fast a flush is processed either. But I
>>> think this patch is a nice cleanup of the vhost_worker() function.
>>> It eliminates the uninitialized_var() construct and moves the
>>> assignment of "done_seq" below the read of "queue_seq".
>>>
>>> Bart.
>>
>> I'm not worried about uninitialized_var - it's just a compiler bug.
>> done_seq below read is nice, but this is at the cost of sprinkling
>> lock/unlock, lock/unlock all over the code.  Maybe we can come up with
>> something that doesn't have this property.
>
> The extra locking introduced here does not look good to me neither.

Please note that although there are additional locking statements in the 
code, there are no additional locking calls at runtime.

Bart.

Re: [PATCH 1/2] vhost: Reduce vhost_work_flush() wakeup latency

[Michael S. Tsirkin]

On Thu, Aug 15, 2013 at 08:13:14AM +0200, Bart Van Assche wrote:
> On 08/15/13 03:30, Asias He wrote:
> >On Wed, Aug 14, 2013 at 08:58:25PM +0300, Michael S. Tsirkin wrote:
> >>On Wed, Aug 14, 2013 at 05:22:36PM +0200, Bart Van Assche wrote:
> >>>On 08/14/13 13:37, Michael S. Tsirkin wrote:
> >>>>On Wed, Aug 14, 2013 at 09:02:32AM +0200, Bart Van Assche wrote:
> >>>>>If the TIF_NEED_RESCHED task flag is set, wake up any vhost_work_flush()
> >>>>>waiters before rescheduling instead of after rescheduling.
> >>>>>
> >>>>>Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> >>>>>Cc: Michael S. Tsirkin <mst@redhat.com>
> >>>>>Cc: Asias He <asias@redhat.com>
> >>>>
> >>>>Why exactly? It's not like flush needs to be extra fast ...
> >>>
> >>>I'm not worried about how fast a flush is processed either. But I
> >>>think this patch is a nice cleanup of the vhost_worker() function.
> >>>It eliminates the uninitialized_var() construct and moves the
> >>>assignment of "done_seq" below the read of "queue_seq".
> >>>
> >>>Bart.
> >>
> >>I'm not worried about uninitialized_var - it's just a compiler bug.
> >>done_seq below read is nice, but this is at the cost of sprinkling
> >>lock/unlock, lock/unlock all over the code.  Maybe we can come up with
> >>something that doesn't have this property.
> >
> >The extra locking introduced here does not look good to me neither.
> 
> Please note that although there are additional locking statements in
> the code, there are no additional locking calls at runtime.
> 
> Bart.

That's true but this is supposed to be a cleanup.

-- 
MST

[PATCH 2/2] vhost: Avoid that vhost_work_flush() returns early

[Bart Van Assche]

If two or more items are queued on dev->work_list before
vhost_worker() starts processing these then the value of
work->done_seq will be set to the sequence number of a work item
that has not yet been processed. Avoid this by letting
vhost_worker() count the number of items that have already been
processed.

Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Asias He <asias@redhat.com>
---
 drivers/vhost/vhost.c |    5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index e7ffc10..11d668a 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -202,7 +202,7 @@ static int vhost_worker(void *data)
 {
 	struct vhost_dev *dev = data;
 	struct vhost_work *work;
-	unsigned seq;
+	unsigned seq = 0;
 	mm_segment_t oldfs = get_fs();
 
 	set_fs(USER_DS);
@@ -216,14 +216,13 @@ static int vhost_worker(void *data)
 			work = list_first_entry(&dev->work_list,
 						struct vhost_work, node);
 			list_del_init(&work->node);
-			seq = work->queue_seq;
 			spin_unlock_irq(&dev->work_lock);
 
 			__set_current_state(TASK_RUNNING);
 			work->fn(work);
 
 			spin_lock_irq(&dev->work_lock);
-			work->done_seq = seq;
+			work->done_seq = ++seq;
 			if (work->flushing)
 				wake_up_all(&work->done);
 		}
-- 
1.7.10.4

Re: [PATCH 2/2] vhost: Avoid that vhost_work_flush() returns early

[Michael S. Tsirkin]

On Wed, Aug 14, 2013 at 09:03:28AM +0200, Bart Van Assche wrote:
> If two or more items are queued on dev->work_list before
> vhost_worker() starts processing these then the value of
> work->done_seq will be set to the sequence number of a work item
> that has not yet been processed. Avoid this by letting
> vhost_worker() count the number of items that have already been
> processed.
> 
> Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> Cc: Michael S. Tsirkin <mst@redhat.com>
> Cc: Asias He <asias@redhat.com>

I'm confused by this explanation.
done_seq is set here:
                if (work) {
                        work->done_seq = seq;
                        if (work->flushing)
                                wake_up_all(&work->done);
                }

and work is set here:

                if (!list_empty(&dev->work_list)) {
                        work = list_first_entry(&dev->work_list,
                                                struct vhost_work, node);
                        list_del_init(&work->node);
                        seq = work->queue_seq;
                }

this work is processed on the next line:

                if (work) {
                        __set_current_state(TASK_RUNNING);
                        work->fn(work);
                        if (need_resched())
                                schedule();
                }

so how do we end up with a sequence of a work item
that isn't processed?


> ---
>  drivers/vhost/vhost.c |    5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index e7ffc10..11d668a 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -202,7 +202,7 @@ static int vhost_worker(void *data)
>  {
>  	struct vhost_dev *dev = data;
>  	struct vhost_work *work;
> -	unsigned seq;
> +	unsigned seq = 0;
>  	mm_segment_t oldfs = get_fs();
>  
>  	set_fs(USER_DS);
> @@ -216,14 +216,13 @@ static int vhost_worker(void *data)
>  			work = list_first_entry(&dev->work_list,
>  						struct vhost_work, node);
>  			list_del_init(&work->node);
> -			seq = work->queue_seq;
>  			spin_unlock_irq(&dev->work_lock);
>  
>  			__set_current_state(TASK_RUNNING);
>  			work->fn(work);
>  
>  			spin_lock_irq(&dev->work_lock);
> -			work->done_seq = seq;
> +			work->done_seq = ++seq;
>  			if (work->flushing)
>  				wake_up_all(&work->done);
>  		}
> -- 
> 1.7.10.4

Re: [PATCH 2/2] vhost: Avoid that vhost_work_flush() returns early

[Bart Van Assche]

On 08/14/13 13:46, Michael S. Tsirkin wrote:
> I'm confused by this explanation.
> done_seq is set here:
>                  if (work) {
>                          work->done_seq = seq;
>                          if (work->flushing)
>                                  wake_up_all(&work->done);
>                  }
>
> and work is set here:
>
>                  if (!list_empty(&dev->work_list)) {
>                          work = list_first_entry(&dev->work_list,
>                                                  struct vhost_work, node);
>                          list_del_init(&work->node);
>                          seq = work->queue_seq;
>                  }
>
> this work is processed on the next line:
>
>                  if (work) {
>                          __set_current_state(TASK_RUNNING);
>                          work->fn(work);
>                          if (need_resched())
>                                  schedule();
>                  }
>
> so how do we end up with a sequence of a work item
> that isn't processed?

I was wondering what would happen if a work item got requeued before 
done_seq is set. I think you are right and this should work fine.

Bart.

Re: [PATCH 0/2] vhost_worker fixes

[Michael S. Tsirkin]

On Wed, Aug 14, 2013 at 09:01:27AM +0200, Bart Van Assche wrote:
> There are two patches in this series:
> * A patch that reduces vhost_work_flush() latency on busy systems.
> * A patch that avoids that vhost_work_flush() returns early.
> 
> I found these issues via code inspection.

Interesting. What makes you think vhost_work_flush
latency matters?

Please copy all relevant mailing lists in the future:
you can find them in the MAINTAINERS file.