Re: [RFC] eventpoll: Move a kmem_cache_alloc and kmem_cache_free

[Nathan Zimmer <nzimmer@sgi.com>]

On Mon, Sep 23, 2013 at 11:17:39AM -0400, Jason Baron wrote:
> On 09/19/2013 12:37 PM, Nathan Zimmer wrote:
> > On 09/18/2013 02:09 PM, Jason Baron wrote:
> >> On 09/13/2013 11:54 AM, Nathan Zimmer wrote:
> >>> We noticed some scaling issue in the SPECjbb benchmark.  Running perf
> >>> we found that the it was spending lots of time in SYS_epoll_ctl.
> >>> In particular it is holding the epmutex.
> >>> This patch helps by moving out the kmem_cache_alloc and kmem_cache_free out
> >>> from under the lock.  It improves throughput by around 15% on 16 sockets.
> >>>
> >>> While this patch should be fine as it is there are probably is more things
> >>> that can be done out side the lock, like wakeup_source_unregister, but I am
> >>> not familar with the area and I don't know of many tests.  I did find the
> >>> one posted by Jason Baron at https://lkml.org/lkml/2011/2/25/297.
> >>>
> >>> Any thoughts?
> >>>
> >> Hi,
> >>
> >> Intersting - I think its also possible to completely drop taking the
> >> 'epmutex' for EPOLL_CTL_DEL by using rcu, and restricting it on add
> >> to more 'complex' topologies. That is when we have an epoll descriptor
> >> that doesn't nest with other epoll descriptors, we don't need the
> >> global 'epmutex' either. Any chance you can re-run with this? Its a bit
> >> hacky, but we can clean it up if it makes sense.
> >>
> >> Thanks,
> >>
> >> -Jason
> >>
> > That is working GREAT.  It is scaling to 16 jobs quite well.
> > I will have to grab a larger machine( to see what the new scaling curve 
> > will be.
> > 
> 
> Cool. Any specific numbers would be helpful for the changelog in support of these
> changes. Also, I think the move the alloc/free out of from under the locks still
> might be nice, since we are still taking the per-ep lock in most cases. If you
> want I can roll those too into a patch series for this when I resubmit.
> 
> Also, if you're still testing I have a small additional optimization on top of the
> prior patch:
> 
> diff --git a/fs/eventpoll.c b/fs/eventpoll.c
> index d98105d..d967fd7 100644
> --- a/fs/eventpoll.c
> +++ b/fs/eventpoll.c
> @@ -857,7 +857,7 @@ static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
>  	struct eventpoll *ep = file->private_data;
>  	struct readyevents_params params;
>  
> -	params.locked = ((wait->_qproc == ep_ptable_queue_proc) ? 1 : 0);
> +	params.locked = ((wait && (wait->_qproc == ep_ptable_queue_proc)) ? 1 : 0);
>  	params.ep = ep;
>  
>  	/* Insert inside our poll wait queue */
> @@ -1907,7 +1907,6 @@ SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
>  			} else
>  				list_add(&tf.file->f_tfile_llink, &tfile_check_list);
>  			mutex_lock_nested(&ep->mtx, 0);
> -			ep->type = EVENTPOLL_COMPLEX;
>  			if (is_file_epoll(tf.file)) {
>  				mutex_lock_nested(&(((struct eventpoll *)tf.file->private_data)->mtx), 1);
>  				((struct eventpoll *)tf.file->private_data)->type = EVENTPOLL_COMPLEX;
> 
> 
> Thanks,
> 
> -Jason
> 
> 
> 

I just finished up run some tests over the weekend.
The specjbb benchmark went from only scaling 10-12 sockets to scaling to over
32 sockets.  I don't have an exact point where it stops scaling but it is under
64 sockets, the size of the machine I had handy.  perf seems to indicate the
problems are elsewhere, but I will have to rerun and grap some more data.

Nate