[dpdk-dev] [PATCH v3 1/7] net/mlx4: remove error flows from Tx fast path

Adrien Mazarguil adrien.mazarguil at 6wind.com
Tue Oct 31 11:16:58 CET 2017


Hi Matan,

On Mon, Oct 30, 2017 at 06:11:31PM +0000, Matan Azrad wrote:
> Hi Adrien
> 
> > -----Original Message-----
> > From: Adrien Mazarguil [mailto:adrien.mazarguil at 6wind.com]
> > Sent: Monday, October 30, 2017 4:23 PM
> > To: Matan Azrad <matan at mellanox.com>
> > Cc: dev at dpdk.org; Ophir Munk <ophirmu at mellanox.com>
> > Subject: Re: [PATCH v3 1/7] net/mlx4: remove error flows from Tx fast path
> > 
> > On Mon, Oct 30, 2017 at 10:07:23AM +0000, Matan Azrad wrote:
> > > Move unnecessary error flows to DEBUG mode.
> > >
> > > Signed-off-by: Matan Azrad <matan at mellanox.com>
> > > Acked-by: Adrien Mazarguil <adrien.mazarguil at 6wind.com>
> > 
> > I missed a couple of details while reviewing the original version, the first one
> > being mlx4_post_send()'s return value is still documented as updating
> > rte_errno in case of error, it's not the case anymore after this patch.
> > 
> Good attention, Will be fixed in next version.
> 
> > Please see below for the other one:
> > 
> > > ---
> > >  drivers/net/mlx4/mlx4_rxtx.c | 16 ++++++----------
> > >  1 file changed, 6 insertions(+), 10 deletions(-)
> > >
> > > diff --git a/drivers/net/mlx4/mlx4_rxtx.c
> > > b/drivers/net/mlx4/mlx4_rxtx.c
> > <snip>
> > >  /**
> > > @@ -510,8 +508,6 @@ struct pv {
> > >  	assert(max <= elts_n);
> > >  	/* Always leave one free entry in the ring. */
> > >  	--max;
> > > -	if (max == 0)
> > > -		return 0;
> > >  	if (max > pkts_n)
> > >  		max = pkts_n;
> > >  	for (i = 0; (i != max); ++i) {
> > 
> > While minor, this change has nothing to do with this patch, right?
> > 
> Yes you right, maybe it can be merged in patch 4/7.
>  
> > I think it can slightly degrade an application performance as it removes the
> > guarantee that subsequent code only needs to be run if there is at least one
> > packet to process in case the TX ring is constantly full (SW faster than HW).
> >
> 
> In case the TX ring is full, the loop condition should fail in the start and then return with 0  because the packet counter is 0.(more 2 checks)
> Since this case are less common (in my opinion) than at least 1 free space in ring, we can prevent this unnecessary check for all these common cases.    
>   
> Are you sure the 2 extra check important for performance in this empty case? Doesn't the application will call us again? 

No, I don't think they're important to performance, like the changes from
patch 4/7, I'm not certain they actually make any difference. My suggestion
was mainly to leave it alone because of that. It's OK if you want to keep
and move it to 4/7.

-- 
Adrien Mazarguil
6WIND


More information about the dev mailing list