[PATCH 1/2] sched/swait: allow swake_up() to return

[PATCH 1/2] sched/swait: allow swake_up() to return

[Peter Xu]

Let swake_up() to return whether any of the waiters is waked up. One use
case of it would be:

  if (swait_active(wq)) {
    swake_up(wq);
    // do something when waiter is waked up
    waked_up++;
  }

Logically it's possible that when reaching swake_up() the wait queue is
not active any more, and here doing something like waked_up++ would be
inaccurate.  To correct it, we need an atomic version of it.

With this patch, we can simply re-write it into:

  if (swake_up(wq)) {
    // do something when waiter is waked up
    waked_up++;
  }

After all we are checking swait_active() inside swake_up() too.

CC: Ingo Molnar <mingo@redhat.com>
CC: Peter Zijlstra <peterz@infradead.org>
CC: Peter Xu <peterx@redhat.com>
CC: linux-kernel@vger.kernel.org
Signed-off-by: Peter Xu <peterx@redhat.com>
---
 include/linux/swait.h |  4 ++--
 kernel/sched/swait.c  | 21 ++++++++++++++++-----
 2 files changed, 18 insertions(+), 7 deletions(-)

diff --git a/include/linux/swait.h b/include/linux/swait.h
index c1f9c62a8a50..c55171280ac8 100644
--- a/include/linux/swait.h
+++ b/include/linux/swait.h
@@ -84,9 +84,9 @@ static inline int swait_active(struct swait_queue_head *q)
 	return !list_empty(&q->task_list);
 }
 
-extern void swake_up(struct swait_queue_head *q);
+extern bool swake_up(struct swait_queue_head *q);
 extern void swake_up_all(struct swait_queue_head *q);
-extern void swake_up_locked(struct swait_queue_head *q);
+extern bool swake_up_locked(struct swait_queue_head *q);
 
 extern void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait);
 extern void prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait, int state);
diff --git a/kernel/sched/swait.c b/kernel/sched/swait.c
index 3d5610dcce11..31f8d677c690 100644
--- a/kernel/sched/swait.c
+++ b/kernel/sched/swait.c
@@ -16,29 +16,40 @@ EXPORT_SYMBOL(__init_swait_queue_head);
  * If for some reason it would return 0, that means the previously waiting
  * task is already running, so it will observe condition true (or has already).
  */
-void swake_up_locked(struct swait_queue_head *q)
+bool swake_up_locked(struct swait_queue_head *q)
 {
 	struct swait_queue *curr;
 
 	if (list_empty(&q->task_list))
-		return;
+		return false;
 
 	curr = list_first_entry(&q->task_list, typeof(*curr), task_list);
 	wake_up_process(curr->task);
 	list_del_init(&curr->task_list);
+	return true;
 }
 EXPORT_SYMBOL(swake_up_locked);
 
-void swake_up(struct swait_queue_head *q)
+/**
+ * swake_up - Wake up one process on the waiting list
+ * @q:        the waitqueue to wake up
+ *
+ * Returns true if some process is waked up, otherwise false if there
+ * is no waiter to wake up.
+ */
+bool swake_up(struct swait_queue_head *q)
 {
+	bool ret;
 	unsigned long flags;
 
 	if (!swait_active(q))
-		return;
+		return false;
 
 	raw_spin_lock_irqsave(&q->lock, flags);
-	swake_up_locked(q);
+	ret = swake_up_locked(q);
 	raw_spin_unlock_irqrestore(&q->lock, flags);
+
+	return ret;
 }
 EXPORT_SYMBOL(swake_up);
 
-- 
2.13.6

Re: [PATCH 1/2] sched/swait: allow swake_up() to return

[Paolo Bonzini]

On 09/11/2017 10:18, Peter Xu wrote:
> Let swake_up() to return whether any of the waiters is waked up. One use
> case of it would be:
> 
>   if (swait_active(wq)) {
>     swake_up(wq);
>     // do something when waiter is waked up
>     waked_up++;
>   }
> 
> Logically it's possible that when reaching swake_up() the wait queue is
> not active any more, and here doing something like waked_up++ would be
> inaccurate.  To correct it, we need an atomic version of it.
> 
> With this patch, we can simply re-write it into:
> 
>   if (swake_up(wq)) {
>     // do something when waiter is waked up
>     waked_up++;
>   }
> 
> After all we are checking swait_active() inside swake_up() too.

Better subject:

sched/swait: make swake_up() return whether there were any waiters

I like this patch.

> 
> CC: Ingo Molnar <mingo@redhat.com>
> CC: Peter Zijlstra <peterz@infradead.org>
> CC: Peter Xu <peterx@redhat.com>
> CC: linux-kernel@vger.kernel.org
> Signed-off-by: Peter Xu <peterx@redhat.com>
> ---
>  include/linux/swait.h |  4 ++--
>  kernel/sched/swait.c  | 21 ++++++++++++++++-----
>  2 files changed, 18 insertions(+), 7 deletions(-)
> 
> diff --git a/include/linux/swait.h b/include/linux/swait.h
> index c1f9c62a8a50..c55171280ac8 100644
> --- a/include/linux/swait.h
> +++ b/include/linux/swait.h
> @@ -84,9 +84,9 @@ static inline int swait_active(struct swait_queue_head *q)
>  	return !list_empty(&q->task_list);
>  }
>  
> -extern void swake_up(struct swait_queue_head *q);
> +extern bool swake_up(struct swait_queue_head *q);
>  extern void swake_up_all(struct swait_queue_head *q);
> -extern void swake_up_locked(struct swait_queue_head *q);
> +extern bool swake_up_locked(struct swait_queue_head *q);
>  
>  extern void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait);
>  extern void prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait, int state);
> diff --git a/kernel/sched/swait.c b/kernel/sched/swait.c
> index 3d5610dcce11..31f8d677c690 100644
> --- a/kernel/sched/swait.c
> +++ b/kernel/sched/swait.c
> @@ -16,29 +16,40 @@ EXPORT_SYMBOL(__init_swait_queue_head);
>   * If for some reason it would return 0, that means the previously waiting
>   * task is already running, so it will observe condition true (or has already).
>   */
> -void swake_up_locked(struct swait_queue_head *q)
> +bool swake_up_locked(struct swait_queue_head *q)
>  {
>  	struct swait_queue *curr;
>  
>  	if (list_empty(&q->task_list))
> -		return;
> +		return false;
>  
>  	curr = list_first_entry(&q->task_list, typeof(*curr), task_list);
>  	wake_up_process(curr->task);
>  	list_del_init(&curr->task_list);
> +	return true;
>  }
>  EXPORT_SYMBOL(swake_up_locked);
>  
> -void swake_up(struct swait_queue_head *q)
> +/**
> + * swake_up - Wake up one process on the waiting list
> + * @q:        the waitqueue to wake up
> + *
> + * Returns true if some process is waked up, otherwise false if there
> + * is no waiter to wake up.
> + */
> +bool swake_up(struct swait_queue_head *q)
>  {
> +	bool ret;
>  	unsigned long flags;
>  
>  	if (!swait_active(q))
> -		return;
> +		return false;
>  
>  	raw_spin_lock_irqsave(&q->lock, flags);
> -	swake_up_locked(q);
> +	ret = swake_up_locked(q);
>  	raw_spin_unlock_irqrestore(&q->lock, flags);
> +
> +	return ret;
>  }
>  EXPORT_SYMBOL(swake_up);
>  
> 

Re: [PATCH 1/2] sched/swait: allow swake_up() to return

[Peter Zijlstra]

On Thu, Nov 09, 2017 at 05:18:53PM +0800, Peter Xu wrote:
> Let swake_up() to return whether any of the waiters is waked up. One use
> case of it would be:
> 
>   if (swait_active(wq)) {
>     swake_up(wq);
>     // do something when waiter is waked up
>     waked_up++;
>   }

The word is 'woken', and no that doesn't work. All it says is that there
was a waiter, not that you were to one to wake it. Another concurrent
wakeup might have done so.

> 
> Logically it's possible that when reaching swake_up() the wait queue is
> not active any more, and here doing something like waked_up++ would be
> inaccurate.  To correct it, we need an atomic version of it.
> 
> With this patch, we can simply re-write it into:
> 
>   if (swake_up(wq)) {
>     // do something when waiter is waked up
>     waked_up++;
>   }
> 
> After all we are checking swait_active() inside swake_up() too.

We're not in fact; you've been staring at old code; see commit:

  35a2897c2a30 ("sched/wait: Remove the lockless swait_active() check in swake_up*()")


Also, you're changing the interface relative to the regular wait
interface. The two should be similar wherever possible.

Re: [PATCH 1/2] sched/swait: allow swake_up() to return

[Peter Xu]

On Thu, Nov 09, 2017 at 11:23:03AM +0100, Peter Zijlstra wrote:
> On Thu, Nov 09, 2017 at 05:18:53PM +0800, Peter Xu wrote:
> > Let swake_up() to return whether any of the waiters is waked up. One use
> > case of it would be:
> > 
> >   if (swait_active(wq)) {
> >     swake_up(wq);
> >     // do something when waiter is waked up
> >     waked_up++;
> >   }
> 
> The word is 'woken', and no that doesn't work. All it says is that there
> was a waiter, not that you were to one to wake it. Another concurrent
> wakeup might have done so.

Yes.  Or IIUC the waiter can be calling finish_swait() somehow so it
removed itself from the list before being woken.

> 
> > 
> > Logically it's possible that when reaching swake_up() the wait queue is
> > not active any more, and here doing something like waked_up++ would be
> > inaccurate.  To correct it, we need an atomic version of it.
> > 
> > With this patch, we can simply re-write it into:
> > 
> >   if (swake_up(wq)) {
> >     // do something when waiter is waked up
> >     waked_up++;
> >   }
> > 
> > After all we are checking swait_active() inside swake_up() too.
> 
> We're not in fact; you've been staring at old code; see commit:
> 
>   35a2897c2a30 ("sched/wait: Remove the lockless swait_active() check in swake_up*()")

I thought the tree was new enough, but obviously I was wrong...
Thanks for the pointer.

> 
> 
> Also, you're changing the interface relative to the regular wait
> interface. The two should be similar wherever possible.

Indeed.

I came to this when reading kvm_vcpu_wake_up(), so that only affects
some statistic which may not be that critical.  However I don't know
whether there would be any other real use case that we would like to
know exactly whether a call to [s]wake_up() has really done something
or just returned with a NOP.

Anyway, please let me know if you think the same change to wake_up()
would be meaningful, otherwise I can drop this patch and post another
KVM-only one to clean up the redundant callers of swait_active(),
since even if we dropped that list check in 35a2897c2a30, we'll do
that again in swake_up_locked().

And after knowing 35a2897c2a30, I do think that calling swait_active()
before swake_up() is not good since that call is without a lock as
well, just like what can happen before 35a2897c2a30.

(I am not 100% sure whether I fully understand the problem mentioned
 in 35a2897c2a30, but I think it's the memory barrier in the
 lock/unlock that matters.)

Thanks,

-- 
Peter Xu

Re: [PATCH 1/2] sched/swait: allow swake_up() to return

[Peter Xu]

On Thu, Nov 09, 2017 at 11:06:53AM +0100, Paolo Bonzini wrote:
> On 09/11/2017 10:18, Peter Xu wrote:
> > Let swake_up() to return whether any of the waiters is waked up. One use
> > case of it would be:
> > 
> >   if (swait_active(wq)) {
> >     swake_up(wq);
> >     // do something when waiter is waked up
> >     waked_up++;
> >   }
> > 
> > Logically it's possible that when reaching swake_up() the wait queue is
> > not active any more, and here doing something like waked_up++ would be
> > inaccurate.  To correct it, we need an atomic version of it.
> > 
> > With this patch, we can simply re-write it into:
> > 
> >   if (swake_up(wq)) {
> >     // do something when waiter is waked up
> >     waked_up++;
> >   }
> > 
> > After all we are checking swait_active() inside swake_up() too.
> 
> Better subject:
> 
> sched/swait: make swake_up() return whether there were any waiters
> 
> I like this patch.

I'll see how PeterZ would like me to do next, or I can drop this patch
and send another clean up which is part of patch 2.  Thanks for the
positive feedback and commenting. :-)

-- 
Peter Xu

Re: [PATCH 1/2] sched/swait: allow swake_up() to return

[Peter Zijlstra]

On Fri, Nov 10, 2017 at 03:10:17PM +0800, Peter Xu wrote:
> I came to this when reading kvm_vcpu_wake_up(), so that only affects
> some statistic which may not be that critical.  However I don't know
> whether there would be any other real use case that we would like to
> know exactly whether a call to [s]wake_up() has really done something
> or just returned with a NOP.
> 
> Anyway, please let me know if you think the same change to wake_up()
> would be meaningful, otherwise I can drop this patch and post another
> KVM-only one to clean up the redundant callers of swait_active(),
> since even if we dropped that list check in 35a2897c2a30, we'll do
> that again in swake_up_locked().

See commits:

  8cd641e3c7cb ("sched/wait: Add swq_has_sleeper()")
  5e0018b3e39e ("kvm: Serialize wq active checks in kvm_vcpu_wake_up()")


In any case, I don't think we want the change you propose. The numbers
don't mean much and there's no point in making all the callers in the
kernel slower for it.

Re: [PATCH 1/2] sched/swait: allow swake_up() to return

[Peter Xu]

On Fri, Nov 10, 2017 at 09:05:20AM +0100, Peter Zijlstra wrote:
> On Fri, Nov 10, 2017 at 03:10:17PM +0800, Peter Xu wrote:
> > I came to this when reading kvm_vcpu_wake_up(), so that only affects
> > some statistic which may not be that critical.  However I don't know
> > whether there would be any other real use case that we would like to
> > know exactly whether a call to [s]wake_up() has really done something
> > or just returned with a NOP.
> > 
> > Anyway, please let me know if you think the same change to wake_up()
> > would be meaningful, otherwise I can drop this patch and post another
> > KVM-only one to clean up the redundant callers of swait_active(),
> > since even if we dropped that list check in 35a2897c2a30, we'll do
> > that again in swake_up_locked().
> 
> See commits:
> 
>   8cd641e3c7cb ("sched/wait: Add swq_has_sleeper()")
>   5e0018b3e39e ("kvm: Serialize wq active checks in kvm_vcpu_wake_up()")
> 
> 
> In any case, I don't think we want the change you propose. The numbers
> don't mean much and there's no point in making all the callers in the
> kernel slower for it.

I see. And also we can introduce a new API for that if really needed.

I'll repost with KVM only changes.  Thanks for reviewing.

-- 
Peter Xu

Re: [PATCH 1/2] sched/swait: allow swake_up() to return

[Peter Xu]

On Mon, Nov 13, 2017 at 11:33:43AM +0800, Peter Xu wrote:
> On Fri, Nov 10, 2017 at 09:05:20AM +0100, Peter Zijlstra wrote:
> > On Fri, Nov 10, 2017 at 03:10:17PM +0800, Peter Xu wrote:
> > > I came to this when reading kvm_vcpu_wake_up(), so that only affects
> > > some statistic which may not be that critical.  However I don't know
> > > whether there would be any other real use case that we would like to
> > > know exactly whether a call to [s]wake_up() has really done something
> > > or just returned with a NOP.
> > > 
> > > Anyway, please let me know if you think the same change to wake_up()
> > > would be meaningful, otherwise I can drop this patch and post another
> > > KVM-only one to clean up the redundant callers of swait_active(),
> > > since even if we dropped that list check in 35a2897c2a30, we'll do
> > > that again in swake_up_locked().
> > 
> > See commits:
> > 
> >   8cd641e3c7cb ("sched/wait: Add swq_has_sleeper()")
> >   5e0018b3e39e ("kvm: Serialize wq active checks in kvm_vcpu_wake_up()")
> > 
> > 
> > In any case, I don't think we want the change you propose. The numbers
> > don't mean much and there's no point in making all the callers in the
> > kernel slower for it.
> 
> I see. And also we can introduce a new API for that if really needed.
> 
> I'll repost with KVM only changes.  Thanks for reviewing.

Wait... I see that https://lkml.org/lkml/2017/9/5/622 seems to have
fixed all the occurences.  So I'll drop the series.  Thanks,

-- 
Peter Xu