Re: [PATCH 1/7] rcu/segcblist: Prevent useless GP start if no CBs to accelerate

[Joel Fernandes <joel@joelfernandes.org>]

On Thu, Jun 18, 2020 at 04:09:34PM -0700, Paul E. McKenney wrote:
> On Thu, Jun 18, 2020 at 03:11:19PM -0700, Paul E. McKenney wrote:
> > On Thu, Jun 18, 2020 at 04:29:49PM -0400, Joel Fernandes (Google) wrote:
> > 
> > First, this looks like a very nice optimization, thank you!

Thanks!

> > > Cc: urezki@gmail.com
> > > Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
> 
> As discussed over IRC, I updated the patch as shown below.  Does that
> work for you?

Yes, that works for me. Thanks!

 - Joel


> 							Thanx, Paul
> 
> ------------------------------------------------------------------------
> 
> commit ec037e1f438074eb16fd68a63d699fc419c9ba0c
> Author: Joel Fernandes (Google) <joel@joelfernandes.org>
> Date:   Thu Jun 18 16:29:49 2020 -0400
> 
>     rcu/segcblist: Prevent useless GP start if no CBs to accelerate
>     
>     The rcu_segcblist_accelerate() function returns true iff it is necessary
>     to request another grace period.  A tracing session showed that this
>     function unnecessarily requests grace periods.
>     
>     For exmaple, consider the following sequence of events:
>     1. Callbacks are queued only on the NEXT segment of CPU A's callback list.
>     2. CPU A runs RCU_SOFTIRQ, accelerating these callbacks from NEXT to WAIT.
>     3. Thus rcu_segcblist_accelerate() returns true, requesting grace period N.
>     4. RCU's grace-period kthread wakes up on CPU B and starts grace period N.
>     4. CPU A notices the new grace period and invokes RCU_SOFTIRQ.
>     5. CPU A's RCU_SOFTIRQ again invokes rcu_segcblist_accelerate(), but
>        there are no new callbacks.  However, rcu_segcblist_accelerate()
>        nevertheless (uselessly) requests a new grace period N+1.
>     
>     This extra grace period results in additional lock contention and also
>     additional wakeups, all for no good reason.
>     
>     This commit therefore adds a check to rcu_segcblist_accelerate() that
>     prevents the return of true when there are no new callbacks.
>     
>     This change reduces the number of grace periods (GPs) and wakeups in each
>     of eleven five-second rcutorture runs as follows:
>     
>     +----+-------------------+-------------------+
>     | #  | Number of GPs     | Number of Wakeups |
>     +====+=========+=========+=========+=========+
>     | 1  | With    | Without | With    | Without |
>     +----+---------+---------+---------+---------+
>     | 2  |      75 |      89 |     113 |     119 |
>     +----+---------+---------+---------+---------+
>     | 3  |      62 |      91 |     105 |     123 |
>     +----+---------+---------+---------+---------+
>     | 4  |      60 |      79 |      98 |     110 |
>     +----+---------+---------+---------+---------+
>     | 5  |      63 |      79 |      99 |     112 |
>     +----+---------+---------+---------+---------+
>     | 6  |      57 |      89 |      96 |     123 |
>     +----+---------+---------+---------+---------+
>     | 7  |      64 |      85 |      97 |     118 |
>     +----+---------+---------+---------+---------+
>     | 8  |      58 |      83 |      98 |     113 |
>     +----+---------+---------+---------+---------+
>     | 9  |      57 |      77 |      89 |     104 |
>     +----+---------+---------+---------+---------+
>     | 10 |      66 |      82 |      98 |     119 |
>     +----+---------+---------+---------+---------+
>     | 11 |      52 |      82 |      83 |     117 |
>     +----+---------+---------+---------+---------+
>     
>     The reduction in the number of wakeups ranges from 5% to 40%.
>     
>     Cc: urezki@gmail.com
>     [ paulmck: Rework commit log and comment. ]
>     Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
>     Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
> 
> diff --git a/kernel/rcu/rcu_segcblist.c b/kernel/rcu/rcu_segcblist.c
> index 9a0f661..2d2a6b6b9 100644
> --- a/kernel/rcu/rcu_segcblist.c
> +++ b/kernel/rcu/rcu_segcblist.c
> @@ -475,8 +475,16 @@ bool rcu_segcblist_accelerate(struct rcu_segcblist *rsclp, unsigned long seq)
>  	 * Also advance to the oldest segment of callbacks whose
>  	 * ->gp_seq[] completion is at or after that passed in via "seq",
>  	 * skipping any empty segments.
> +	 *
> +	 * Note that segment "i" (and any lower-numbered segments
> +	 * containing older callbacks) will be unaffected, and their
> +	 * grace-period numbers remain unchanged.  For example, if i ==
> +	 * WAIT_TAIL, then neither WAIT_TAIL nor DONE_TAIL will be touched.
> +	 * Instead, the CBs in NEXT_TAIL will be merged with those in
> +	 * NEXT_READY_TAIL and the grace-period number of NEXT_READY_TAIL
> +	 * would be updated.  NEXT_TAIL would then be empty.
>  	 */
> -	if (++i >= RCU_NEXT_TAIL)
> +	if (rcu_segcblist_restempty(rsclp, i) || ++i >= RCU_NEXT_TAIL)
>  		return false;
>  
>  	/*