On Mon, Jul 31, 2023 at 07:16:01PM -0400, Chuck Lever wrote:
> On Mon, Jul 31, 2023 at 04:48:33PM +1000, NeilBrown wrote:
> > svc_get_next_xprt() does a lot more than just get an xprt. It also
> > decides if it needs to sleep, depending not only on the availability of
> > xprts, but also on the need to exit or handle external work
> > (SP_TASK_PENDING).
> >
> > So rename it to svc_rqst_wait_and_dequeue_work(), don't return the xprt
> > (which can easily be found in rqstp->rq_xprt), and restructure to make a
> > clear separation between waiting and dequeueing.
>
> For me, the most valuable part of this patch is the last part here:
> refactoring the dequeue and the wait, and deduplicating the dequeue.
>
>
> > All the scheduling-related code like try_to_freeze() and
> > kthread_should_stop() is moved into svc_rqst_wait_and_dequeue_work().
> >
> > Rather than calling svc_xprt_dequeue() twice (before and after deciding
> > to wait), it now calls rqst_should_sleep() twice. If the first fails,
> > we skip all the waiting code completely. In the waiting code we call
> > again after setting the task state in case we missed a wake-up.
> >
> > We now only have one call to try_to_freeze() and one call to
> > svc_xprt_dequeue(). We still have two calls to kthread_should_stop() -
> > one in rqst_should_sleep() to avoid sleeping, and one afterwards to
> > avoid dequeueing any work (it previously came after dequeueing which
> > doesn't seem right).
> >
> > Signed-off-by: NeilBrown <neilb@xxxxxxx>
> > ---
> > net/sunrpc/svc_xprt.c | 62 +++++++++++++++++++++----------------------
> > 1 file changed, 31 insertions(+), 31 deletions(-)
> >
> > diff --git a/net/sunrpc/svc_xprt.c b/net/sunrpc/svc_xprt.c
> > index 380fb3caea4c..67f2b34cb8e4 100644
> > --- a/net/sunrpc/svc_xprt.c
> > +++ b/net/sunrpc/svc_xprt.c
> > @@ -722,47 +722,51 @@ rqst_should_sleep(struct svc_rqst *rqstp)
> > return true;
> > }
> >
> > -static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp)
> > +static void svc_rqst_wait_and_dequeue_work(struct svc_rqst *rqstp)
>
> It would be simpler to follow if you renamed this function once
> (here), and changed directly from returning struct svc_xprt to
> returning bool.
>
>
> > {
> > struct svc_pool *pool = rqstp->rq_pool;
> > + bool slept = false;
> >
> > /* rq_xprt should be clear on entry */
> > WARN_ON_ONCE(rqstp->rq_xprt);
> >
> > - rqstp->rq_xprt = svc_xprt_dequeue(pool);
> > - if (rqstp->rq_xprt) {
> > - trace_svc_pool_polled(rqstp);
> > - goto out_found;
> > + if (rqst_should_sleep(rqstp)) {
> > + set_current_state(TASK_IDLE);
> > + smp_mb__before_atomic();
> > + clear_bit(SP_CONGESTED, &pool->sp_flags);
> > + clear_bit(RQ_BUSY, &rqstp->rq_flags);
> > + smp_mb__after_atomic();
> > +
> > + /* Need to test again after setting task state */
>
> This comment isn't illuminating. It needs to explain the "need to
> test again".
>
>
> > + if (likely(rqst_should_sleep(rqstp))) {
>
> Is likely() still needed here?
>
>
> > + schedule();
> > + slept = true;
> > + } else {
> > + __set_current_state(TASK_RUNNING);
> > + cond_resched();
>
> This makes me happy. Only call cond_resched() if we didn't sleep.
>
>
> > + }
> > + set_bit(RQ_BUSY, &rqstp->rq_flags);
> > + smp_mb__after_atomic();
> > }
> > -
> > - set_current_state(TASK_IDLE);
> > - smp_mb__before_atomic();
> > - clear_bit(SP_CONGESTED, &pool->sp_flags);
> > - clear_bit(RQ_BUSY, &rqstp->rq_flags);
> > - smp_mb__after_atomic();
> > -
> > - if (likely(rqst_should_sleep(rqstp)))
> > - schedule();
> > - else
> > - __set_current_state(TASK_RUNNING);
> > -
> > try_to_freeze();
> >
> > - set_bit(RQ_BUSY, &rqstp->rq_flags);
> > - smp_mb__after_atomic();
> > + if (kthread_should_stop())
> > + return;
> > +
> > clear_bit(SP_TASK_PENDING, &pool->sp_flags);
> > rqstp->rq_xprt = svc_xprt_dequeue(pool);
> > if (rqstp->rq_xprt) {
> > - trace_svc_pool_awoken(rqstp);
> > + if (slept)
> > + trace_svc_pool_awoken(rqstp);
> > + else
> > + trace_svc_pool_polled(rqstp);
>
> Again, it would perhaps be better if we rearranged this code first,
> and then added tracepoints. This is ... well, ugly.
I've dropped the three tracepoint patches and pushed out the changes
to topic-sunrpc-thread-scheduling . We can circle back to adding
tracepoints once this code has settled.
> > goto out_found;
> > }
> >
> > - if (kthread_should_stop())
> > - return NULL;
> > - percpu_counter_inc(&pool->sp_threads_no_work);
> > - return NULL;
> > + if (slept)
> > + percpu_counter_inc(&pool->sp_threads_no_work);
> > + return;
> > out_found:
> > - clear_bit(SP_TASK_PENDING, &pool->sp_flags);
> > /* Normally we will wait up to 5 seconds for any required
> > * cache information to be provided.
> > */
> > @@ -770,7 +774,6 @@ static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp)
> > rqstp->rq_chandle.thread_wait = 5*HZ;
> > else
> > rqstp->rq_chandle.thread_wait = 1*HZ;
> > - return rqstp->rq_xprt;
> > }
> >
> > static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
> > @@ -854,12 +857,9 @@ void svc_recv(struct svc_rqst *rqstp)
> > if (!svc_alloc_arg(rqstp))
> > goto out;
> >
> > - try_to_freeze();
> > - cond_resched();
> > - if (kthread_should_stop())
> > - goto out;
> > + svc_rqst_wait_and_dequeue_work(rqstp);
> >
> > - xprt = svc_get_next_xprt(rqstp);
> > + xprt = rqstp->rq_xprt;
> > if (!xprt)
> > goto out;
> >
> > --
> > 2.40.1
> >
>
> --
> Chuck Lever
--
Chuck Lever
