Re: [PATCH] rcu: Use cpus_read_lock() while looking at cpu_online_mask

[Boqun Feng <boqun.feng@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 3584 bytes --]

Hi, Sebastian

On Mon, Oct 15, 2018 at 04:42:17PM +0200, Sebastian Andrzej Siewior wrote:
> On 2018-10-13 06:48:13 [-0700], Paul E. McKenney wrote:
> > 
> > My concern would be that it would queue it by default for the current
> > CPU, which would serialize the processing, losing the concurrency of
> > grace-period initialization.  But that was a long time ago, and perhaps
> > workqueues have changed. 
> 
> but the code here is always using the first CPU of a NUMA node or did I
> miss something?
> 

The thing is the original way is to pick one CPU for a *RCU* node to
run the grace-period work, but with your proposal, if a RCU node is
smaller than a NUMA node (having fewer CPUs), we could end up having two
grace-period works running on one CPU. I think that's Paul's concern.

Regards,
Boqun

> > So, have you tried using rcuperf to test the
> > update performance on a large system?
> 
> Something like this:
> | tools/testing/selftests/rcutorture/bin/kvm.sh --torture rcuperf --configs TREE
> | ----Start batch 1: Mon Oct 15 12:46:13 CEST 2018
> | TREE 142: Starting build. …
> | …
> | Average grace-period duration: 19952.7 microseconds
> | Minimum grace-period duration: 9004.51
> | 50th percentile grace-period duration: 19998.3
> | 90th percentile grace-period duration: 24994.4
> | 99th percentile grace-period duration: 30002.1
> | Maximum grace-period duration: 42998.2
> | Grace periods: 6560 Batches: 209 Ratio: 31.3876
> | Computed from rcuperf printk output.
> | Completed in 27 vs. 1800
> | 
> | Average grace-period duration: 18700 microseconds
> | Minimum grace-period duration: 7069.2
> | 50th percentile grace-period duration: 18987.5
> | 90th percentile grace-period duration: 22997
> | 99th percentile grace-period duration: 28944.7
> | Maximum grace-period duration: 36994.5
> | Grace periods: 6551 Batches: 209 Ratio: 31.3445
> | Computed from rcuperf printk output.
> | Completed in 27 vs. 1800
> 
> two runs patched followed by two runs without the patched:
> | Average grace-period duration: 19423.3 microseconds
> | Minimum grace-period duration: 8006.93
> | 50th percentile grace-period duration: 19002.8
> | 90th percentile grace-period duration: 23997.5
> | 99th percentile grace-period duration: 29995.7
> | Maximum grace-period duration: 37997.9
> | Grace periods: 6526 Batches: 208 Ratio: 31.375
> | Computed from rcuperf printk output.
> | Completed in 27 vs. 1800
> | 
> | Average grace-period duration: 18822.4 microseconds
> | Minimum grace-period duration: 8348.15
> | 50th percentile grace-period duration: 18996.9
> | 90th percentile grace-period duration: 23000
> | 99th percentile grace-period duration: 27999.5
> | Maximum grace-period duration: 39001.9
> | Grace periods: 6540 Batches: 209 Ratio: 31.2919
> | Computed from rcuperf printk output.
> | Completed in 27 vs. 1800
> 
> I think difference might come from cpufreq on the host. But in general,
> this is what you have been asking for or did you want to see something
> on the host (or an additional argument to the script)?
> 
> > If this change does not impact performance on an rcuperf test, why not
> > send me a formal patch with Signed-off-by and commit log (including
> > performance testing results)?  I will then apply it, it will be exposed
> > to 0day and eventually -next testing, and if no problems arise, it will go
> > to mainline, perhaps as soon as the merge window after the upcoming one.
> > 
> > Fair enough?
> sure.
> 
> > 							Thanx, Paul
> > 
> 
> Sebastian

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]