[dpdk-dev] [PATCH] examples: remove Rx checksum offload

Jerin Jacob jerin.jacob at caviumnetworks.com
Mon Jul 30 16:40:33 CEST 2018


-----Original Message-----
> Date: Mon, 30 Jul 2018 14:12:12 +0000
> From: "Ananyev, Konstantin" <konstantin.ananyev at intel.com>
> To: Jerin Jacob <jerin.jacob at caviumnetworks.com>
> CC: Thomas Monjalon <thomas at monjalon.net>, "dev at dpdk.org" <dev at dpdk.org>,
>  "Yigit, Ferruh" <ferruh.yigit at intel.com>, "shahafs at mellanox.com"
>  <shahafs at mellanox.com>
> Subject: RE: [dpdk-dev] [PATCH] examples: remove Rx checksum offload
> 
> 
> >
> > -----Original Message-----
> > > Date: Mon, 30 Jul 2018 11:00:02 +0000
> > > From: "Ananyev, Konstantin" <konstantin.ananyev at intel.com>
> > > To: Thomas Monjalon <thomas at monjalon.net>, Jerin Jacob
> > >  <jerin.jacob at caviumnetworks.com>
> > > CC: "dev at dpdk.org" <dev at dpdk.org>, "Yigit, Ferruh"
> > >  <ferruh.yigit at intel.com>, "shahafs at mellanox.com" <shahafs at mellanox.com>
> > > Subject: RE: [dpdk-dev] [PATCH] examples: remove Rx checksum offload
> > >
> > > External Email
> > >
> > > > -----Original Message-----
> > > > From: Thomas Monjalon [mailto:thomas at monjalon.net]
> > > > Sent: Monday, July 30, 2018 10:51 AM
> > > > To: Jerin Jacob <jerin.jacob at caviumnetworks.com>; Ananyev, Konstantin <konstantin.ananyev at intel.com>
> > > > Cc: dev at dpdk.org; Yigit, Ferruh <ferruh.yigit at intel.com>; shahafs at mellanox.com
> > > > Subject: Re: [dpdk-dev] [PATCH] examples: remove Rx checksum offload
> > > >
> > > > 30/07/2018 11:35, Jerin Jacob:
> > > > > From: "Ananyev, Konstantin" <konstantin.ananyev at intel.com>
> > > > > > >
> > > > > > > As of now, application does not check PKT_RX_*_CKSUM_* flags per
> > > > > > > packet, so it does not matter DEV_RX_OFFLOAD_CHECKSUM enabled or not.
> > > > > > >
> > > > > > > Removing DEV_RX_OFFLOAD_CHECKSUM offload so that driver can save a few
> > > > > > > cycles if possible.
> > > > > >
> > > > > > Personally, I'd move in other direction: keep RX checksum offload and add
> > > > > > checks inside sample apps to handle (drop) packets with invalid checksum.
> > > > >
> > > > > OK. Till someones add the DROP logic in application, Can we take
> > > > > this patch? Because there is no point in enabling DEV_RX_OFFLOAD_CHECKSUM
> > > > > without DROP or any meaning full action in application.
> > >
> > > Probably, but at least it gives users a right estimation how long the proper
> > > RX/TX routine would take.
> >
> > For estimation, application can add any flag they want in local setup.
> > It does not need to be upstream with out feature complete.
> >
> > > From other side what the point to disable these flags now, if we know that
> >
> > At least nicvf Rx routines are crafted based DEV_RX_OFFLOAD_CHECKSUM
> > flags. If driver Rx routine crafted such case it will be useful.
> >
> > > we are doing wrong thing and will have to re-enable them again in future?
> >
> > But it is not correct now either. Right?
> 
> Yes, right now invalid cksum information is simply ignored.
> As you pointed - some PMD select RX routine based on checksum offload flags.
> Yes, removing these flags might produce better performance numbers.
> But from my perspective - it would be an artificial and temporary improvement,
> as for l3fwd like apps we'll need to revert it back and add code to drop invalid packets.

IMO, It is OK get a performance hit when do that support in l3fwd. There
is no harm in removing the DEV_RX_OFFLOAD_CHECKSUM flag for now and it
is correct from application perspective.(you are enabling an offload when
you are using it, else don't enable it. I believe, this was philosophy for
enabling Rx/Tx offloads)

Since it is going in circles, I leave decision to ethdev maintainers.

> Konstantin
> 
> >
> > >
> > > >
> > > > If there is no patch sent to use this offload on August 1st,
> > > > then I will apply this patch to remove the offload request.
> > > >
> > >
> > > Isn't it too late to do such things right now?
> > > We are in RC3 stage and doesn't look like a critical issue.
> >
> > Yes. We can add it when have we proper fix. Currently, it signaling a wrong
> > interpretation to application.
> >
> >
> > > Konstantin
> > >
> > >


More information about the dev mailing list