diff mbox series

[10/14] SUNRPC: change svc_pool_wake_idle_thread() to return nothing.

Message ID 168966228866.11075.18415964365983945832.stgit@noble.brown (mailing list archive)
State New, archived
Headers show
Series Refactor SUNRPC svc thread code, and use llist | expand

Commit Message

NeilBrown July 18, 2023, 6:38 a.m. UTC
No callers of svc_pool_wake_idle_thread() care which thread was woken -
except one that wants to trace the wakeup.  For now we drop that
tracepoint.

One caller wants to know if anything was woken to set SP_CONGESTED, so
set that inside the function instead.

Now svc_pool_wake_idle_thread() can "return" void.

Signed-off-by: NeilBrown <neilb@suse.de>
---
 include/linux/sunrpc/svc.h |    2 +-
 net/sunrpc/svc.c           |   13 +++++--------
 net/sunrpc/svc_xprt.c      |   18 +++---------------
 3 files changed, 9 insertions(+), 24 deletions(-)

Comments

Chuck Lever July 18, 2023, 1:58 p.m. UTC | #1
On Tue, Jul 18, 2023 at 04:38:08PM +1000, NeilBrown wrote:
> No callers of svc_pool_wake_idle_thread() care which thread was woken -
> except one that wants to trace the wakeup.  For now we drop that
> tracepoint.

That's an important tracepoint, IMO.

It might be better to have svc_pool_wake_idle_thread() return void
right from it's introduction, and move the tracepoint into that
function. I can do that and respin if you agree.


> One caller wants to know if anything was woken to set SP_CONGESTED, so
> set that inside the function instead.
> 
> Now svc_pool_wake_idle_thread() can "return" void.
> 
> Signed-off-by: NeilBrown <neilb@suse.de>
> ---
>  include/linux/sunrpc/svc.h |    2 +-
>  net/sunrpc/svc.c           |   13 +++++--------
>  net/sunrpc/svc_xprt.c      |   18 +++---------------
>  3 files changed, 9 insertions(+), 24 deletions(-)
> 
> diff --git a/include/linux/sunrpc/svc.h b/include/linux/sunrpc/svc.h
> index ea3ce1315416..b39c613fbe06 100644
> --- a/include/linux/sunrpc/svc.h
> +++ b/include/linux/sunrpc/svc.h
> @@ -454,7 +454,7 @@ int		   svc_register(const struct svc_serv *, struct net *, const int,
>  
>  void		   svc_wake_up(struct svc_serv *);
>  void		   svc_reserve(struct svc_rqst *rqstp, int space);
> -struct svc_rqst	  *svc_pool_wake_idle_thread(struct svc_serv *serv,
> +void		   svc_pool_wake_idle_thread(struct svc_serv *serv,
>  					     struct svc_pool *pool);
>  struct svc_pool   *svc_pool_for_cpu(struct svc_serv *serv);
>  char *		   svc_print_addr(struct svc_rqst *, char *, size_t);
> diff --git a/net/sunrpc/svc.c b/net/sunrpc/svc.c
> index b18175ef74ec..fd49e7b12c94 100644
> --- a/net/sunrpc/svc.c
> +++ b/net/sunrpc/svc.c
> @@ -696,13 +696,10 @@ svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
>   * @serv: RPC service
>   * @pool: service thread pool
>   *
> - * Returns an idle service thread (now marked BUSY), or NULL
> - * if no service threads are available. Finding an idle service
> - * thread and marking it BUSY is atomic with respect to other
> - * calls to svc_pool_wake_idle_thread().
> + * Wake an idle thread if there is one, else mark the pool as congested.
>   */
> -struct svc_rqst *svc_pool_wake_idle_thread(struct svc_serv *serv,
> -					   struct svc_pool *pool)
> +void svc_pool_wake_idle_thread(struct svc_serv *serv,
> +			       struct svc_pool *pool)
>  {
>  	struct svc_rqst	*rqstp;
>  
> @@ -715,13 +712,13 @@ struct svc_rqst *svc_pool_wake_idle_thread(struct svc_serv *serv,
>  		wake_up_process(rqstp->rq_task);
>  		rcu_read_unlock();
>  		percpu_counter_inc(&pool->sp_threads_woken);
> -		return rqstp;
> +		return;
>  	}
>  	rcu_read_unlock();
>  
>  	trace_svc_pool_starved(serv, pool);
>  	percpu_counter_inc(&pool->sp_threads_starved);
> -	return NULL;
> +	set_bit(SP_CONGESTED, &pool->sp_flags);
>  }
>  EXPORT_SYMBOL_GPL(svc_pool_wake_idle_thread);
>  
> diff --git a/net/sunrpc/svc_xprt.c b/net/sunrpc/svc_xprt.c
> index 948605e7043b..964c97dbb36c 100644
> --- a/net/sunrpc/svc_xprt.c
> +++ b/net/sunrpc/svc_xprt.c
> @@ -456,7 +456,6 @@ static bool svc_xprt_ready(struct svc_xprt *xprt)
>   */
>  void svc_xprt_enqueue(struct svc_xprt *xprt)
>  {
> -	struct svc_rqst *rqstp;
>  	struct svc_pool *pool;
>  
>  	if (!svc_xprt_ready(xprt))
> @@ -477,13 +476,7 @@ void svc_xprt_enqueue(struct svc_xprt *xprt)
>  	list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
>  	spin_unlock_bh(&pool->sp_lock);
>  
> -	rqstp = svc_pool_wake_idle_thread(xprt->xpt_server, pool);
> -	if (!rqstp) {
> -		set_bit(SP_CONGESTED, &pool->sp_flags);
> -		return;
> -	}
> -
> -	trace_svc_xprt_enqueue(xprt, rqstp);
> +	svc_pool_wake_idle_thread(xprt->xpt_server, pool);
>  }
>  EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
>  
> @@ -587,14 +580,9 @@ static void svc_xprt_release(struct svc_rqst *rqstp)
>  void svc_wake_up(struct svc_serv *serv)
>  {
>  	struct svc_pool *pool = &serv->sv_pools[0];
> -	struct svc_rqst *rqstp;
>  
> -	rqstp = svc_pool_wake_idle_thread(serv, pool);
> -	if (!rqstp) {
> -		set_bit(SP_TASK_PENDING, &pool->sp_flags);
> -		smp_wmb();
> -		return;
> -	}
> +	set_bit(SP_TASK_PENDING, &pool->sp_flags);
> +	svc_pool_wake_idle_thread(serv, pool);
>  }
>  EXPORT_SYMBOL_GPL(svc_wake_up);
>  
> 
>
NeilBrown July 19, 2023, 1:16 a.m. UTC | #2
On Tue, 18 Jul 2023, Chuck Lever wrote:
> On Tue, Jul 18, 2023 at 04:38:08PM +1000, NeilBrown wrote:
> > No callers of svc_pool_wake_idle_thread() care which thread was woken -
> > except one that wants to trace the wakeup.  For now we drop that
> > tracepoint.
> 
> That's an important tracepoint, IMO.
> 
> It might be better to have svc_pool_wake_idle_thread() return void
> right from it's introduction, and move the tracepoint into that
> function. I can do that and respin if you agree.

Mostly I agree.

It isn't clear to me how you would handle trace_svc_xprt_enqueue(),
as there would be no code that can see both the trigger xprt, and the
woken rqst.

I also wonder if having the trace point when the wake-up is requested
makes any sense, as there is no guarantee that thread with handle that
xprt.

Maybe the trace point should report when the xprt is dequeued.  i.e.
maybe trace_svc_pool_awoken() should report the pid, and we could have
trace_svc_xprt_enqueue() only report the xprt, not the rqst.

Thanks,
NeilBrown
Chuck Lever July 19, 2023, 1:12 p.m. UTC | #3
> On Jul 18, 2023, at 9:16 PM, NeilBrown <neilb@suse.de> wrote:
> 
> On Tue, 18 Jul 2023, Chuck Lever wrote:
>> On Tue, Jul 18, 2023 at 04:38:08PM +1000, NeilBrown wrote:
>>> No callers of svc_pool_wake_idle_thread() care which thread was woken -
>>> except one that wants to trace the wakeup.  For now we drop that
>>> tracepoint.
>> 
>> That's an important tracepoint, IMO.
>> 
>> It might be better to have svc_pool_wake_idle_thread() return void
>> right from it's introduction, and move the tracepoint into that
>> function. I can do that and respin if you agree.
> 
> Mostly I agree.
> 
> It isn't clear to me how you would handle trace_svc_xprt_enqueue(),
> as there would be no code that can see both the trigger xprt, and the
> woken rqst.
> 
> I also wonder if having the trace point when the wake-up is requested
> makes any sense, as there is no guarantee that thread with handle that
> xprt.
> 
> Maybe the trace point should report when the xprt is dequeued.  i.e.
> maybe trace_svc_pool_awoken() should report the pid, and we could have
> trace_svc_xprt_enqueue() only report the xprt, not the rqst.

I'll come up with something that rearranges the tracepoints so that
svc_pool_wake_idle_thread() can return void.

svc_pool_wake_idle_thread() can save the waker's PID in svc_rqst
somewhere, for example. The dequeue tracepoint can then report that
(if it's still interesting when we're all done with this work).


--
Chuck Lever
NeilBrown July 19, 2023, 11:20 p.m. UTC | #4
On Wed, 19 Jul 2023, Chuck Lever III wrote:
> 
> > On Jul 18, 2023, at 9:16 PM, NeilBrown <neilb@suse.de> wrote:
> > 
> > On Tue, 18 Jul 2023, Chuck Lever wrote:
> >> On Tue, Jul 18, 2023 at 04:38:08PM +1000, NeilBrown wrote:
> >>> No callers of svc_pool_wake_idle_thread() care which thread was woken -
> >>> except one that wants to trace the wakeup.  For now we drop that
> >>> tracepoint.
> >> 
> >> That's an important tracepoint, IMO.
> >> 
> >> It might be better to have svc_pool_wake_idle_thread() return void
> >> right from it's introduction, and move the tracepoint into that
> >> function. I can do that and respin if you agree.
> > 
> > Mostly I agree.
> > 
> > It isn't clear to me how you would handle trace_svc_xprt_enqueue(),
> > as there would be no code that can see both the trigger xprt, and the
> > woken rqst.
> > 
> > I also wonder if having the trace point when the wake-up is requested
> > makes any sense, as there is no guarantee that thread with handle that
> > xprt.
> > 
> > Maybe the trace point should report when the xprt is dequeued.  i.e.
> > maybe trace_svc_pool_awoken() should report the pid, and we could have
> > trace_svc_xprt_enqueue() only report the xprt, not the rqst.
> 
> I'll come up with something that rearranges the tracepoints so that
> svc_pool_wake_idle_thread() can return void.

My current draft code has svc_pool_wake_idle_thread() returning bool -
if it found something to wake up - purely for logging.
I think it is worth logging whether an event triggered a wake up or not,
and which event did that.  I'm less you that the pid is relevant.  But
as you say - this will probably become clearer as the code settles down.

Thanks,
NeilBrown


> 
> svc_pool_wake_idle_thread() can save the waker's PID in svc_rqst
> somewhere, for example. The dequeue tracepoint can then report that
> (if it's still interesting when we're all done with this work).
> 
> 
> --
> Chuck Lever
> 
> 
>
Chuck Lever July 19, 2023, 11:44 p.m. UTC | #5
> On Jul 19, 2023, at 7:20 PM, NeilBrown <neilb@suse.de> wrote:
> 
> On Wed, 19 Jul 2023, Chuck Lever III wrote:
>> 
>>> On Jul 18, 2023, at 9:16 PM, NeilBrown <neilb@suse.de> wrote:
>>> 
>>> On Tue, 18 Jul 2023, Chuck Lever wrote:
>>>> On Tue, Jul 18, 2023 at 04:38:08PM +1000, NeilBrown wrote:
>>>>> No callers of svc_pool_wake_idle_thread() care which thread was woken -
>>>>> except one that wants to trace the wakeup.  For now we drop that
>>>>> tracepoint.
>>>> 
>>>> That's an important tracepoint, IMO.
>>>> 
>>>> It might be better to have svc_pool_wake_idle_thread() return void
>>>> right from it's introduction, and move the tracepoint into that
>>>> function. I can do that and respin if you agree.
>>> 
>>> Mostly I agree.
>>> 
>>> It isn't clear to me how you would handle trace_svc_xprt_enqueue(),
>>> as there would be no code that can see both the trigger xprt, and the
>>> woken rqst.
>>> 
>>> I also wonder if having the trace point when the wake-up is requested
>>> makes any sense, as there is no guarantee that thread with handle that
>>> xprt.
>>> 
>>> Maybe the trace point should report when the xprt is dequeued.  i.e.
>>> maybe trace_svc_pool_awoken() should report the pid, and we could have
>>> trace_svc_xprt_enqueue() only report the xprt, not the rqst.
>> 
>> I'll come up with something that rearranges the tracepoints so that
>> svc_pool_wake_idle_thread() can return void.
> 
> My current draft code has svc_pool_wake_idle_thread() returning bool -
> if it found something to wake up - purely for logging.

This is also where I have ended up. I'll post an update probably tomorrow
my time. Too much other stuff going on to finish it today.


> I think it is worth logging whether an event triggered a wake up or not,
> and which event did that.

Agreed. I have some experimental code that records _RET_IP_ of the caller
of svc_xprt_enqueue(), but again it's questionable whether that is of
long term value.


> I'm less you that the pid is relevant.  But
> as you say - this will probably become clearer as the code settles down.
> 
> Thanks,
> NeilBrown
> 
> 
>> 
>> svc_pool_wake_idle_thread() can save the waker's PID in svc_rqst
>> somewhere, for example. The dequeue tracepoint can then report that
>> (if it's still interesting when we're all done with this work).
>> 
>> 
>> --
>> Chuck Lever


--
Chuck Lever
Chuck Lever July 20, 2023, 2:29 p.m. UTC | #6
> On Jul 19, 2023, at 7:44 PM, Chuck Lever III <chuck.lever@oracle.com> wrote:
> 
> 
> 
>> On Jul 19, 2023, at 7:20 PM, NeilBrown <neilb@suse.de> wrote:
>> 
>> On Wed, 19 Jul 2023, Chuck Lever III wrote:
>>> 
>>>> On Jul 18, 2023, at 9:16 PM, NeilBrown <neilb@suse.de> wrote:
>>>> 
>>>> On Tue, 18 Jul 2023, Chuck Lever wrote:
>>>>> On Tue, Jul 18, 2023 at 04:38:08PM +1000, NeilBrown wrote:
>>>>>> No callers of svc_pool_wake_idle_thread() care which thread was woken -
>>>>>> except one that wants to trace the wakeup.  For now we drop that
>>>>>> tracepoint.
>>>>> 
>>>>> That's an important tracepoint, IMO.
>>>>> 
>>>>> It might be better to have svc_pool_wake_idle_thread() return void
>>>>> right from it's introduction, and move the tracepoint into that
>>>>> function. I can do that and respin if you agree.
>>>> 
>>>> Mostly I agree.
>>>> 
>>>> It isn't clear to me how you would handle trace_svc_xprt_enqueue(),
>>>> as there would be no code that can see both the trigger xprt, and the
>>>> woken rqst.
>>>> 
>>>> I also wonder if having the trace point when the wake-up is requested
>>>> makes any sense, as there is no guarantee that thread with handle that
>>>> xprt.
>>>> 
>>>> Maybe the trace point should report when the xprt is dequeued.  i.e.
>>>> maybe trace_svc_pool_awoken() should report the pid, and we could have
>>>> trace_svc_xprt_enqueue() only report the xprt, not the rqst.
>>> 
>>> I'll come up with something that rearranges the tracepoints so that
>>> svc_pool_wake_idle_thread() can return void.
>> 
>> My current draft code has svc_pool_wake_idle_thread() returning bool -
>> if it found something to wake up - purely for logging.
> 
> This is also where I have ended up. I'll post an update probably tomorrow
> my time. Too much other stuff going on to finish it today.

Pushed to https://git.kernel.org/pub/scm/linux/kernel/git/cel/linux.git
in branch topic-sunrpc-thread-scheduling


>> I think it is worth logging whether an event triggered a wake up or not,
>> and which event did that.
> 
> Agreed. I have some experimental code that records _RET_IP_ of the caller
> of svc_xprt_enqueue(), but again it's questionable whether that is of
> long term value.
> 
> 
>> I'm less you that the pid is relevant.  But
>> as you say - this will probably become clearer as the code settles down.
>> 
>> Thanks,
>> NeilBrown
>> 
>> 
>>> 
>>> svc_pool_wake_idle_thread() can save the waker's PID in svc_rqst
>>> somewhere, for example. The dequeue tracepoint can then report that
>>> (if it's still interesting when we're all done with this work).
>>> 
>>> 
>>> --
>>> Chuck Lever
> 
> 
> --
> Chuck Lever


--
Chuck Lever
NeilBrown July 24, 2023, 12:03 a.m. UTC | #7
On Fri, 21 Jul 2023, Chuck Lever III wrote:
> 
> > On Jul 19, 2023, at 7:44 PM, Chuck Lever III <chuck.lever@oracle.com> wrote:
> > 
> > 
> > 
> >> On Jul 19, 2023, at 7:20 PM, NeilBrown <neilb@suse.de> wrote:
> >> 
> >> On Wed, 19 Jul 2023, Chuck Lever III wrote:
> >>> 
> >>>> On Jul 18, 2023, at 9:16 PM, NeilBrown <neilb@suse.de> wrote:
> >>>> 
> >>>> On Tue, 18 Jul 2023, Chuck Lever wrote:
> >>>>> On Tue, Jul 18, 2023 at 04:38:08PM +1000, NeilBrown wrote:
> >>>>>> No callers of svc_pool_wake_idle_thread() care which thread was woken -
> >>>>>> except one that wants to trace the wakeup.  For now we drop that
> >>>>>> tracepoint.
> >>>>> 
> >>>>> That's an important tracepoint, IMO.
> >>>>> 
> >>>>> It might be better to have svc_pool_wake_idle_thread() return void
> >>>>> right from it's introduction, and move the tracepoint into that
> >>>>> function. I can do that and respin if you agree.
> >>>> 
> >>>> Mostly I agree.
> >>>> 
> >>>> It isn't clear to me how you would handle trace_svc_xprt_enqueue(),
> >>>> as there would be no code that can see both the trigger xprt, and the
> >>>> woken rqst.
> >>>> 
> >>>> I also wonder if having the trace point when the wake-up is requested
> >>>> makes any sense, as there is no guarantee that thread with handle that
> >>>> xprt.
> >>>> 
> >>>> Maybe the trace point should report when the xprt is dequeued.  i.e.
> >>>> maybe trace_svc_pool_awoken() should report the pid, and we could have
> >>>> trace_svc_xprt_enqueue() only report the xprt, not the rqst.
> >>> 
> >>> I'll come up with something that rearranges the tracepoints so that
> >>> svc_pool_wake_idle_thread() can return void.
> >> 
> >> My current draft code has svc_pool_wake_idle_thread() returning bool -
> >> if it found something to wake up - purely for logging.
> > 
> > This is also where I have ended up. I'll post an update probably tomorrow
> > my time. Too much other stuff going on to finish it today.
> 
> Pushed to https://git.kernel.org/pub/scm/linux/kernel/git/cel/linux.git
> in branch topic-sunrpc-thread-scheduling
> 

Thanks.
One little thing to fix.
In
    SUNRPC: Report when no service thread is available.
you now have

-	return false;
+
+	trace_svc_pool_starved(serv, pool);
+	percpu_counter_inc(&pool->sp_threads_starved);
+	return NULL;

That should still be "return false" at the end;

NeilBrown
NeilBrown July 24, 2023, 4:44 a.m. UTC | #8
On Fri, 21 Jul 2023, Chuck Lever III wrote:
> 
> > On Jul 19, 2023, at 7:44 PM, Chuck Lever III <chuck.lever@oracle.com> wrote:
> > 
> > 
> > 
> >> On Jul 19, 2023, at 7:20 PM, NeilBrown <neilb@suse.de> wrote:
> >> 
> >> On Wed, 19 Jul 2023, Chuck Lever III wrote:
> >>> 
> >>>> On Jul 18, 2023, at 9:16 PM, NeilBrown <neilb@suse.de> wrote:
> >>>> 
> >>>> On Tue, 18 Jul 2023, Chuck Lever wrote:
> >>>>> On Tue, Jul 18, 2023 at 04:38:08PM +1000, NeilBrown wrote:
> >>>>>> No callers of svc_pool_wake_idle_thread() care which thread was woken -
> >>>>>> except one that wants to trace the wakeup.  For now we drop that
> >>>>>> tracepoint.
> >>>>> 
> >>>>> That's an important tracepoint, IMO.
> >>>>> 
> >>>>> It might be better to have svc_pool_wake_idle_thread() return void
> >>>>> right from it's introduction, and move the tracepoint into that
> >>>>> function. I can do that and respin if you agree.
> >>>> 
> >>>> Mostly I agree.
> >>>> 
> >>>> It isn't clear to me how you would handle trace_svc_xprt_enqueue(),
> >>>> as there would be no code that can see both the trigger xprt, and the
> >>>> woken rqst.
> >>>> 
> >>>> I also wonder if having the trace point when the wake-up is requested
> >>>> makes any sense, as there is no guarantee that thread with handle that
> >>>> xprt.
> >>>> 
> >>>> Maybe the trace point should report when the xprt is dequeued.  i.e.
> >>>> maybe trace_svc_pool_awoken() should report the pid, and we could have
> >>>> trace_svc_xprt_enqueue() only report the xprt, not the rqst.
> >>> 
> >>> I'll come up with something that rearranges the tracepoints so that
> >>> svc_pool_wake_idle_thread() can return void.
> >> 
> >> My current draft code has svc_pool_wake_idle_thread() returning bool -
> >> if it found something to wake up - purely for logging.
> > 
> > This is also where I have ended up. I'll post an update probably tomorrow
> > my time. Too much other stuff going on to finish it today.
> 
> Pushed to https://git.kernel.org/pub/scm/linux/kernel/git/cel/linux.git
> in branch topic-sunrpc-thread-scheduling

Another thing.
You have made svc_pool_wake_idle_thread() return, but for different
reasons than me.

I wanted bool so I could trace a wake up due to enqueuing an xprt
differently from a wakeup due to a call to svc_wake_up().  I thought the
difference might be important.

You have it returning a bool so that:
- in one case you can set SP_CONGESTED - but that can be safely set
  inside svc_pool_wake_idle_thread()
- in another case so SP_TASK_PENDING can be set.  But I think it is
  best to set that anyway, and clear it when svc_recv() wakes up.

So maybe it can return void after all.

Thanks,
NeilBrown
Chuck Lever July 25, 2023, 1:49 p.m. UTC | #9
> On Jul 24, 2023, at 12:44 AM, NeilBrown <neilb@suse.de> wrote:
> 
> On Fri, 21 Jul 2023, Chuck Lever III wrote:
>> 
>>> On Jul 19, 2023, at 7:44 PM, Chuck Lever III <chuck.lever@oracle.com> wrote:
>>> 
>>> 
>>> 
>>>> On Jul 19, 2023, at 7:20 PM, NeilBrown <neilb@suse.de> wrote:
>>>> 
>>>> On Wed, 19 Jul 2023, Chuck Lever III wrote:
>>>>> 
>>>>>> On Jul 18, 2023, at 9:16 PM, NeilBrown <neilb@suse.de> wrote:
>>>>>> 
>>>>>> On Tue, 18 Jul 2023, Chuck Lever wrote:
>>>>>>> On Tue, Jul 18, 2023 at 04:38:08PM +1000, NeilBrown wrote:
>>>>>>>> No callers of svc_pool_wake_idle_thread() care which thread was woken -
>>>>>>>> except one that wants to trace the wakeup.  For now we drop that
>>>>>>>> tracepoint.
>>>>>>> 
>>>>>>> That's an important tracepoint, IMO.
>>>>>>> 
>>>>>>> It might be better to have svc_pool_wake_idle_thread() return void
>>>>>>> right from it's introduction, and move the tracepoint into that
>>>>>>> function. I can do that and respin if you agree.
>>>>>> 
>>>>>> Mostly I agree.
>>>>>> 
>>>>>> It isn't clear to me how you would handle trace_svc_xprt_enqueue(),
>>>>>> as there would be no code that can see both the trigger xprt, and the
>>>>>> woken rqst.
>>>>>> 
>>>>>> I also wonder if having the trace point when the wake-up is requested
>>>>>> makes any sense, as there is no guarantee that thread with handle that
>>>>>> xprt.
>>>>>> 
>>>>>> Maybe the trace point should report when the xprt is dequeued.  i.e.
>>>>>> maybe trace_svc_pool_awoken() should report the pid, and we could have
>>>>>> trace_svc_xprt_enqueue() only report the xprt, not the rqst.
>>>>> 
>>>>> I'll come up with something that rearranges the tracepoints so that
>>>>> svc_pool_wake_idle_thread() can return void.
>>>> 
>>>> My current draft code has svc_pool_wake_idle_thread() returning bool -
>>>> if it found something to wake up - purely for logging.
>>> 
>>> This is also where I have ended up. I'll post an update probably tomorrow
>>> my time. Too much other stuff going on to finish it today.
>> 
>> Pushed to https://git.kernel.org/pub/scm/linux/kernel/git/cel/linux.git
>> in branch topic-sunrpc-thread-scheduling
> 
> Another thing.
> You have made svc_pool_wake_idle_thread() return, but for different
> reasons than me.
> 
> I wanted bool so I could trace a wake up due to enqueuing an xprt
> differently from a wakeup due to a call to svc_wake_up().  I thought the
> difference might be important.
> 
> You have it returning a bool so that:
> - in one case you can set SP_CONGESTED - but that can be safely set
>  inside svc_pool_wake_idle_thread()

So, set CONGESTED whenever an idle thread cannot be found, no matter
the caller.


> - in another case so SP_TASK_PENDING can be set.  But I think it is
>  best to set that anyway, and clear it when svc_recv() wakes up.

Maybe that should be done in a separate patch. Can you give it a try?


> So maybe it can return void after all.


--
Chuck Lever
diff mbox series

Patch

diff --git a/include/linux/sunrpc/svc.h b/include/linux/sunrpc/svc.h
index ea3ce1315416..b39c613fbe06 100644
--- a/include/linux/sunrpc/svc.h
+++ b/include/linux/sunrpc/svc.h
@@ -454,7 +454,7 @@  int		   svc_register(const struct svc_serv *, struct net *, const int,
 
 void		   svc_wake_up(struct svc_serv *);
 void		   svc_reserve(struct svc_rqst *rqstp, int space);
-struct svc_rqst	  *svc_pool_wake_idle_thread(struct svc_serv *serv,
+void		   svc_pool_wake_idle_thread(struct svc_serv *serv,
 					     struct svc_pool *pool);
 struct svc_pool   *svc_pool_for_cpu(struct svc_serv *serv);
 char *		   svc_print_addr(struct svc_rqst *, char *, size_t);
diff --git a/net/sunrpc/svc.c b/net/sunrpc/svc.c
index b18175ef74ec..fd49e7b12c94 100644
--- a/net/sunrpc/svc.c
+++ b/net/sunrpc/svc.c
@@ -696,13 +696,10 @@  svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
  * @serv: RPC service
  * @pool: service thread pool
  *
- * Returns an idle service thread (now marked BUSY), or NULL
- * if no service threads are available. Finding an idle service
- * thread and marking it BUSY is atomic with respect to other
- * calls to svc_pool_wake_idle_thread().
+ * Wake an idle thread if there is one, else mark the pool as congested.
  */
-struct svc_rqst *svc_pool_wake_idle_thread(struct svc_serv *serv,
-					   struct svc_pool *pool)
+void svc_pool_wake_idle_thread(struct svc_serv *serv,
+			       struct svc_pool *pool)
 {
 	struct svc_rqst	*rqstp;
 
@@ -715,13 +712,13 @@  struct svc_rqst *svc_pool_wake_idle_thread(struct svc_serv *serv,
 		wake_up_process(rqstp->rq_task);
 		rcu_read_unlock();
 		percpu_counter_inc(&pool->sp_threads_woken);
-		return rqstp;
+		return;
 	}
 	rcu_read_unlock();
 
 	trace_svc_pool_starved(serv, pool);
 	percpu_counter_inc(&pool->sp_threads_starved);
-	return NULL;
+	set_bit(SP_CONGESTED, &pool->sp_flags);
 }
 EXPORT_SYMBOL_GPL(svc_pool_wake_idle_thread);
 
diff --git a/net/sunrpc/svc_xprt.c b/net/sunrpc/svc_xprt.c
index 948605e7043b..964c97dbb36c 100644
--- a/net/sunrpc/svc_xprt.c
+++ b/net/sunrpc/svc_xprt.c
@@ -456,7 +456,6 @@  static bool svc_xprt_ready(struct svc_xprt *xprt)
  */
 void svc_xprt_enqueue(struct svc_xprt *xprt)
 {
-	struct svc_rqst *rqstp;
 	struct svc_pool *pool;
 
 	if (!svc_xprt_ready(xprt))
@@ -477,13 +476,7 @@  void svc_xprt_enqueue(struct svc_xprt *xprt)
 	list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
 	spin_unlock_bh(&pool->sp_lock);
 
-	rqstp = svc_pool_wake_idle_thread(xprt->xpt_server, pool);
-	if (!rqstp) {
-		set_bit(SP_CONGESTED, &pool->sp_flags);
-		return;
-	}
-
-	trace_svc_xprt_enqueue(xprt, rqstp);
+	svc_pool_wake_idle_thread(xprt->xpt_server, pool);
 }
 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
 
@@ -587,14 +580,9 @@  static void svc_xprt_release(struct svc_rqst *rqstp)
 void svc_wake_up(struct svc_serv *serv)
 {
 	struct svc_pool *pool = &serv->sv_pools[0];
-	struct svc_rqst *rqstp;
 
-	rqstp = svc_pool_wake_idle_thread(serv, pool);
-	if (!rqstp) {
-		set_bit(SP_TASK_PENDING, &pool->sp_flags);
-		smp_wmb();
-		return;
-	}
+	set_bit(SP_TASK_PENDING, &pool->sp_flags);
+	svc_pool_wake_idle_thread(serv, pool);
 }
 EXPORT_SYMBOL_GPL(svc_wake_up);