public inbox for netdev@vger.kernel.org 
 help / color / mirror / Atom feed
From: Jesper Dangaard Brouer <brouer@redhat•com>
To: Eric Dumazet <eric.dumazet@gmail•com>
Cc: Rick Jones <rick.jones2@hpe•com>,
	netdev@vger•kernel.org, Saeed Mahameed <saeedm@mellanox•com>,
	Tariq Toukan <tariqt@mellanox•com>,
	Achiad Shochat <achiad@mellanox•com>,
	brouer@redhat•com
Subject: Re: [WIP] net+mlx4: auto doorbell
Date: Wed, 30 Nov 2016 20:17:11 +0100	[thread overview]
Message-ID: <20161130201711.2f353a76@redhat.com> (raw)
In-Reply-To: <1480521386.18162.189.camel@edumazet-glaptop3.roam.corp.google.com>

On Wed, 30 Nov 2016 07:56:26 -0800
Eric Dumazet <eric.dumazet@gmail•com> wrote:

> On Wed, 2016-11-30 at 12:38 +0100, Jesper Dangaard Brouer wrote:
> > I've played with a somewhat similar patch (from Achiad Shochat) for
> > mlx5 (attached).  While it gives huge improvements, the problem I ran
> > into was that; TX performance became a function of the TX completion
> > time/interrupt and could easily be throttled if configured too
> > high/slow.
> > 
> > Can your patch be affected by this too?  
> 
> Like all TX business, you should know this Jesper.
> No need to constantly remind us something very well known.

Don't take is as critique Eric.  I was hoping your patch would have
solved this issue of being sensitive to TX completion adjustments.  You
usually have good solutions for difficult issues. I basically rejected
Achiad's approach/patch because it was too sensitive to these kind of
adjustments.


> > On Mon, 28 Nov 2016 22:58:36 -0800 Eric Dumazet <eric.dumazet@gmail•com> wrote:
[...]
> > 
> > These +75% number is pktgen without "burst", and definitely show that
> > your patch activate xmit_more.
> > What is the pps performance number when using pktgen "burst" option?  
> 
> About the same really. About all packets now get the xmit_more effect.

Perfect!

> > [...]  
> > > diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
> > > index 4b597dca5c52..affebb435679 100644
> > > --- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
> > > +++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c  
> > [...]  
> > > -static inline bool mlx4_en_is_tx_ring_full(struct mlx4_en_tx_ring *ring)
> > > +static inline bool mlx4_en_is_tx_ring_full(const struct mlx4_en_tx_ring *ring)
> > >  {
> > > -	return ring->prod - ring->cons > ring->full_size;
> > > +	return READ_ONCE(ring->prod) - READ_ONCE(ring->cons) > ring->full_size;
> > >  }  
> > [...]
> >   
> > > @@ -1033,6 +1058,14 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
> > >  	}
> > >  	send_doorbell = !skb->xmit_more || netif_xmit_stopped(ring->tx_queue);
> > >  
> > > +	/* Doorbell avoidance : We can omit doorbell if we know a TX completion
> > > +	 * will happen shortly.
> > > +	 */
> > > +	if (send_doorbell &&
> > > +	    dev->doorbell_opt &&
> > > +	    (s32)(READ_ONCE(ring->prod_bell) - READ_ONCE(ring->ncons)) > 0)  
> > 
> > It would be nice with a function call with an appropriate name, instead
> > of an open-coded queue size check.  I'm also confused by the "ncons" name.
> >   
> > > +		send_doorbell = false;
> > > +  
> > [...]
> >   
> > > diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
> > > index 574bcbb1b38f..c3fd0deda198 100644
> > > --- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
> > > +++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
> > > @@ -280,6 +280,7 @@ struct mlx4_en_tx_ring {
> > >  	 */
> > >  	u32			last_nr_txbb;
> > >  	u32			cons;
> > > +	u32			ncons;  
> > 
> > Maybe we can find a better name than "ncons" ?  
> 
> Thats because 'cons' in this driver really means 'old cons' 
> 
> and new cons = old cons + last_nr_txbb;

It was not clear to me that "n" meant "new".  And also not clear that
this drive have an issue of "cons" (consumer) is tracking "old" cons.

  
> > >  	unsigned long		wake_queue;
> > >  	struct netdev_queue	*tx_queue;
> > >  	u32			(*free_tx_desc)(struct mlx4_en_priv *priv,
> > > @@ -290,6 +291,7 @@ struct mlx4_en_tx_ring {
> > >  
> > >  	/* cache line used and dirtied in mlx4_en_xmit() */
> > >  	u32			prod ____cacheline_aligned_in_smp;
> > > +	u32			prod_bell;  
> > 
> > Good descriptive variable name.


-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

  reply	other threads:[~2016-11-30 19:17 UTC|newest]

Thread overview: 62+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2014-09-03 14:59 High perf top ip_idents_reserve doing netperf UDP_STREAM Jesper Dangaard Brouer
2014-09-03 15:17 ` Eric Dumazet
2016-11-16 12:16   ` Netperf UDP issue with connected sockets Jesper Dangaard Brouer
2016-11-16 17:46     ` Rick Jones
2016-11-16 22:40       ` Jesper Dangaard Brouer
2016-11-16 22:50         ` Rick Jones
2016-11-17  0:34         ` Eric Dumazet
2016-11-17  8:16           ` Jesper Dangaard Brouer
2016-11-17 13:20             ` Eric Dumazet
2016-11-17 13:42               ` Jesper Dangaard Brouer
2016-11-17 14:17                 ` Eric Dumazet
2016-11-17 14:57                   ` Jesper Dangaard Brouer
2016-11-17 16:21                     ` Eric Dumazet
2016-11-17 18:30                       ` Jesper Dangaard Brouer
2016-11-17 18:51                         ` Eric Dumazet
2016-11-17 21:19                           ` Jesper Dangaard Brouer
2016-11-17 21:44                             ` Eric Dumazet
2016-11-17 23:08                               ` Rick Jones
2016-11-18  0:37                                 ` Julian Anastasov
2016-11-18  0:42                                   ` Rick Jones
2016-11-18 17:12                               ` Jesper Dangaard Brouer
2016-11-21 16:03                           ` Jesper Dangaard Brouer
2016-11-21 18:10                             ` Eric Dumazet
2016-11-29  6:58                               ` [WIP] net+mlx4: auto doorbell Eric Dumazet
2016-11-30 11:38                                 ` Jesper Dangaard Brouer
2016-11-30 15:56                                   ` Eric Dumazet
2016-11-30 19:17                                     ` Jesper Dangaard Brouer [this message]
2016-11-30 19:30                                       ` Eric Dumazet
2016-11-30 22:30                                         ` Jesper Dangaard Brouer
2016-11-30 22:40                                           ` Eric Dumazet
2016-12-01  0:27                                         ` Eric Dumazet
2016-12-01  1:16                                           ` Tom Herbert
2016-12-01  2:32                                             ` Eric Dumazet
2016-12-01  2:50                                               ` Eric Dumazet
2016-12-02 18:16                                                 ` Eric Dumazet
2016-12-01  5:03                                               ` Tom Herbert
2016-12-01 19:24                                                 ` Willem de Bruijn
2016-11-30 13:50                                 ` Saeed Mahameed
2016-11-30 15:44                                   ` Eric Dumazet
2016-11-30 16:27                                     ` Saeed Mahameed
2016-11-30 17:28                                       ` Eric Dumazet
2016-12-01 12:05                                       ` Jesper Dangaard Brouer
2016-12-01 14:24                                         ` Eric Dumazet
2016-12-01 16:04                                           ` Jesper Dangaard Brouer
2016-12-01 17:04                                             ` Eric Dumazet
2016-12-01 19:17                                               ` Jesper Dangaard Brouer
2016-12-01 20:11                                                 ` Eric Dumazet
2016-12-01 20:20                                               ` David Miller
2016-12-01 22:10                                                 ` Eric Dumazet
2016-12-02 14:23                                               ` Eric Dumazet
2016-12-01 21:32                                 ` Alexander Duyck
2016-12-01 22:04                                   ` Eric Dumazet
2016-11-17 17:34                     ` Netperf UDP issue with connected sockets David Laight
2016-11-17 22:39                       ` Alexander Duyck
2016-11-17 17:42             ` Rick Jones
2016-11-28 18:33             ` Rick Jones
2016-11-28 18:40               ` Rick Jones
2016-11-30 10:43               ` Jesper Dangaard Brouer
2016-11-30 17:42                 ` Rick Jones
2016-11-30 18:11                   ` David Miller
  -- strict thread matches above, loose matches on Subject: below --
2016-11-30  7:28 [WIP] net+mlx4: auto doorbell Alexei Starovoitov
2016-11-30 15:50 ` Eric Dumazet

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20161130201711.2f353a76@redhat.com \
    --to=brouer@redhat$(echo .)com \
    --cc=achiad@mellanox$(echo .)com \
    --cc=eric.dumazet@gmail$(echo .)com \
    --cc=netdev@vger$(echo .)kernel.org \
    --cc=rick.jones2@hpe$(echo .)com \
    --cc=saeedm@mellanox$(echo .)com \
    --cc=tariqt@mellanox$(echo .)com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox