Re: Kernel 4.19 network performance - forwarding/routing normal users traffic

[Aaron Lu <aaron.lu@intel.com>]

On Sat, Nov 03, 2018 at 01:53:25PM +0100, Jesper Dangaard Brouer wrote:
> 
> On Fri, 2 Nov 2018 22:20:24 +0800 Aaron Lu <aaron.lu@intel.com> wrote:
> 
> > On Fri, Nov 02, 2018 at 12:40:37PM +0100, Jesper Dangaard Brouer wrote:
> > > On Fri, 2 Nov 2018 13:23:56 +0800
> > > Aaron Lu <aaron.lu@intel.com> wrote:
> > >   
> > > > On Thu, Nov 01, 2018 at 08:23:19PM +0000, Saeed Mahameed wrote:  
> > > > > On Thu, 2018-11-01 at 23:27 +0800, Aaron Lu wrote:    
> > > > > > On Thu, Nov 01, 2018 at 10:22:13AM +0100, Jesper Dangaard Brouer
> > > > > > wrote:
> > > > > > ... ...    
> > > > > > > Section copied out:
> > > > > > > 
> > > > > > >   mlx5e_poll_tx_cq
> > > > > > >   |          
> > > > > > >    --16.34%--napi_consume_skb
> > > > > > >              |          
> > > > > > >              |--12.65%--__free_pages_ok
> > > > > > >              |          |          
> > > > > > >              |           --11.86%--free_one_page
> > > > > > >              |                     |          
> > > > > > >              |                     |--10.10%
> > > > > > > --queued_spin_lock_slowpath
> > > > > > >              |                     |          
> > > > > > >              |                      --0.65%--_raw_spin_lock    
> > > > > > 
> > > > > > This callchain looks like it is freeing higher order pages than order
> > > > > > 0:
> > > > > > __free_pages_ok is only called for pages whose order are bigger than
> > > > > > 0.    
> > > > > 
> > > > > mlx5 rx uses only order 0 pages, so i don't know where these high order
> > > > > tx SKBs are coming from..     
> > > > 
> > > > Perhaps here:
> > > > __netdev_alloc_skb(), __napi_alloc_skb(), __netdev_alloc_frag() and
> > > > __napi_alloc_frag() will all call page_frag_alloc(), which will use
> > > > __page_frag_cache_refill() to get an order 3 page if possible, or fall
> > > > back to an order 0 page if order 3 page is not available.
> > > > 
> > > > I'm not sure if your workload will use the above code path though.  
> > > 
> > > TL;DR: this is order-0 pages (code-walk trough proof below)
> > > 
> > > To Aaron, the network stack *can* call __free_pages_ok() with order-0
> > > pages, via:
> > > 
> > > static void skb_free_head(struct sk_buff *skb)
> > > {
> > > 	unsigned char *head = skb->head;
> > > 
> > > 	if (skb->head_frag)
> > > 		skb_free_frag(head);
> > > 	else
> > > 		kfree(head);
> > > }
> > > 
> > > static inline void skb_free_frag(void *addr)
> > > {
> > > 	page_frag_free(addr);
> > > }
> > > 
> > > /*
> > >  * Frees a page fragment allocated out of either a compound or order 0 page.
> > >  */
> > > void page_frag_free(void *addr)
> > > {
> > > 	struct page *page = virt_to_head_page(addr);
> > > 
> > > 	if (unlikely(put_page_testzero(page)))
> > > 		__free_pages_ok(page, compound_order(page));
> > > }
> > > EXPORT_SYMBOL(page_frag_free);  
> > 
> > I think here is a problem - order 0 pages are freed directly to buddy,
> > bypassing per-cpu-pages. This might be the reason lock contention
> > appeared on free path. 
> 
> OMG - you just found a significant issue with the network stacks
> interaction with the page allocator!  This explains why I could not get
> the PCP (Per-Cpu-Pages) system to have good performance, in my
> performance networking benchmarks. As we are basically only using the
> alloc side of PCP, and not the free side.

Exactly.

>  We have spend years adding different driver level recycle tricks to
> avoid this code path getting activated, exactly because it is rather
> slow and problematic that we hit this zone->lock.

I can see when this code path is hit, it causes unnecessary taking of
zone lock for order-0 pages and cause lock contention.

> 
> > Can someone apply below diff and see if lock contention is gone?
> 
> I have also applied and tested this patch, and yes the lock contention
> is gone.  As mentioned is it rather difficult to hit this code path, as
> the driver page recycle mechanism tries to hide/avoid it, but mlx5 +
> page_pool + CPU-map recycling have a known weakness that bypass the
> driver page recycle scheme (that I've not fixed yet).  I observed a 7%
> speedup for this micro benchmark.

Good to know this, I will prepare a formal patch.
  
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index e2ef1c17942f..65c0ae13215a 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -4554,8 +4554,14 @@ void page_frag_free(void *addr)
> >  {
> >  	struct page *page = virt_to_head_page(addr);
> >  
> > -	if (unlikely(put_page_testzero(page)))
> > -		__free_pages_ok(page, compound_order(page));
> > +	if (unlikely(put_page_testzero(page))) {
> > +		unsigned int order = compound_order(page);
> > +
> > +		if (order == 0)
> > +			free_unref_page(page);
> > +		else
> > +			__free_pages_ok(page, order);
> > +	}
> >  }
> >  EXPORT_SYMBOL(page_frag_free);
> 
> Thank you Aaron for spotting this!!!

Which is impossible without your analysis :-)