Date:   Mon, 28 Oct 2019 22:57:41 +0000
From:   Saeed Mahameed <saeedm@...lanox.com>
To:     "brouer@...hat.com" <brouer@...hat.com>
CC:     "jonathan.lemon@...il.com" <jonathan.lemon@...il.com>,
        "davem@...emloft.net" <davem@...emloft.net>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        "ilias.apalodimas@...aro.org" <ilias.apalodimas@...aro.org>
Subject: Re: [PATCH net-nex V2 2/3] page_pool: Don't recycle non-reusable
 pages

On Fri, 2019-10-25 at 15:33 +0200, Jesper Dangaard Brouer wrote:
> On Wed, 23 Oct 2019 19:37:00 +0000
> Saeed Mahameed <saeedm@...lanox.com> wrote:
> 
> > A page is NOT reusable when at least one of the following is true:
> > 1) allocated when system was under some pressure.
> > (page_is_pfmemalloc)
> > 2) belongs to a different NUMA node than pool->p.nid.
> > 
> > To update pool->p.nid users should call page_pool_update_nid().
> > 
> > Holding on to such pages in the pool will hurt the consumer
> > performance
> > when the pool migrates to a different numa node.
> > 
> > Performance testing:
> > XDP drop/tx rate and TCP single/multi stream, on mlx5 driver
> > while migrating rx ring irq from close to far numa:
> > 
> > mlx5 internal page cache was locally disabled to get pure page pool
> > results.
> 
> Could you show us the code that disable the local page cache?
> 

here you go, very simple patch, just avoid calling mlx5 internal cache
API and avoid dma mapping/unmapping.

https://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux.git/commit/?h=topic/page-pool-numa&id=aa7345f62cad62cf19fbf2dee2a8b15de34b33e3

> 
> > CPU: Intel(R) Xeon(R) CPU E5-2603 v4 @ 1.70GHz
> > NIC: Mellanox Technologies MT27700 Family [ConnectX-4] (100G)
> > 
> > XDP Drop/TX single core:
> > NUMA  | XDP  | Before    | After
> > ---------------------------------------
> > Close | Drop | 11   Mpps | 10.9 Mpps
> > Far   | Drop | 4.4  Mpps | 5.8  Mpps
> > 
> > Close | TX   | 6.5 Mpps  | 6.5 Mpps
> > Far   | TX   | 3.5 Mpps  | 4  Mpps
> > 
> > Improvement is about 30% drop packet rate, 15% tx packet rate for
> > numa
> > far test.
> > No degradation for numa close tests.
> > 
> > TCP single/multi cpu/stream:
> > NUMA  | #cpu | Before  | After
> > --------------------------------------
> > Close | 1    | 18 Gbps | 18 Gbps
> > Far   | 1    | 15 Gbps | 18 Gbps
> > Close | 12   | 80 Gbps | 80 Gbps
> > Far   | 12   | 68 Gbps | 80 Gbps
> > 
> > In all test cases we see improvement for the far numa case, and no
> > impact on the close numa case.
> > 
> > The impact of adding a check per page is very negligible, and shows
> > no
> > performance degradation whatsoever, also functionality wise it
> > seems more
> > correct and more robust for page pool to verify when pages should
> > be
> > recycled, since page pool can't guarantee where pages are coming
> > from.
> > 
> > Signed-off-by: Saeed Mahameed <saeedm@...lanox.com>
> > Acked-by: Jonathan Lemon <jonathan.lemon@...il.com>
> > ---
> >  net/core/page_pool.c | 14 +++++++++++++-
> >  1 file changed, 13 insertions(+), 1 deletion(-)
> > 
> > diff --git a/net/core/page_pool.c b/net/core/page_pool.c
> > index 953af6d414fb..73e4173c4dce 100644
> > --- a/net/core/page_pool.c
> > +++ b/net/core/page_pool.c
> > @@ -283,6 +283,17 @@ static bool __page_pool_recycle_direct(struct
> > page *page,
> >  	return true;
> >  }
> >  
> > +/* page is NOT reusable when:
> > + * 1) allocated when system is under some pressure.
> > (page_is_pfmemalloc)
> > + * 2) belongs to a different NUMA node than pool->p.nid.
> > + *
> > + * To update pool->p.nid users must call page_pool_update_nid.
> > + */
> > +static bool pool_page_reusable(struct page_pool *pool, struct page
> > *page)
> > +{
> > +	return !page_is_pfmemalloc(page) && page_to_nid(page) == pool-
> > >p.nid;
> > +}
> > +
> >  void __page_pool_put_page(struct page_pool *pool,
> >  			  struct page *page, bool allow_direct)
> >  {
> > @@ -292,7 +303,8 @@ void __page_pool_put_page(struct page_pool
> > *pool,
> >  	 *
> >  	 * refcnt == 1 means page_pool owns page, and can recycle it.
> >  	 */
> > -	if (likely(page_ref_count(page) == 1)) {
> > +	if (likely(page_ref_count(page) == 1 &&
> > +		   pool_page_reusable(pool, page))) {
> 
> I'm afraid that we are slowly chipping away the performance benefit
> with these incremental changes, adding more checks. We have an
> extreme
> performance use-case with XDP_DROP, where we want drivers to use this
> code path to hit __page_pool_recycle_direct(), that is a simple array
> update (protected under NAPI) into pool->alloc.cache[].
> 
> To preserve this hot-path, you could instead flush pool-
> >alloc.cache[]
> in the call page_pool_update_nid().  And move the
> pool_page_reusable()
> check into __page_pool_recycle_into_ring().  (Below added the '>>'
> with
> remaining code to make this easier to see)
> 

Flush simply won't work, what about pages that are currently in use by
the driver HW, and will be returned to the cache only after the flush ?
who will flush those pages ? and flushing the HW queue will be too
heavy !

As Jonathan pointed out, all drivers are doing this check today, all we
need is to simply move this to the cache and take it out of the drivers
responsibility. 

for the pfmemalloc, we can make sure the page pool will never allow
such pages to be allocated in first place, but this is very risky and
can break many "kernel in emergency" features.

I think what Jonathan and I did here, is the best course of action.

> 
> >  		/* Read barrier done in page_ref_count / READ_ONCE */
> >  
> >  		if (allow_direct && in_serving_softirq())
> > > 			if (__page_pool_recycle_direct(page, pool))
> > > 				return;
> > > 
> > > 		if (!__page_pool_recycle_into_ring(pool, page)) {
> > > 			/* Cache full, fallback to free pages */
> > > 			__page_pool_return_page(pool, page);
> > > 		}
> > > 		return;
> > > 	}
> > > 	/* Fallback/non-XDP mode: API user have elevated refcnt.
> 
> 

Powered by blists - more mailing lists