[dpdk-dev] Survey for final decision about per-port offload API

[Ananyev, Konstantin]

> -----Original Message-----
> From: Yigit, Ferruh
> Sent: Tuesday, April 24, 2018 1:28 PM
> To: Ananyev, Konstantin <konstantin.ananyev at intel.com>; Thomas Monjalon <thomas at monjalon.net>; dev at dpdk.org
> Cc: Ajit Khaparde <ajit.khaparde at broadcom.com>; Jerin Jacob <jerin.jacob at caviumnetworks.com>; Shijith Thotton
> <shijith.thotton at cavium.com>; Santosh Shukla <santosh.shukla at caviumnetworks.com>; Rahul Lakkireddy
> <rahul.lakkireddy at chelsio.com>; John Daley <johndale at cisco.com>; Lu, Wenzhuo <wenzhuo.lu at intel.com>; Xing, Beilei
> <beilei.xing at intel.com>; Zhang, Qi Z <qi.z.zhang at intel.com>; Wu, Jingjing <jingjing.wu at intel.com>; Adrien Mazarguil
> <adrien.mazarguil at 6wind.com>; Nelio Laranjeiro <nelio.laranjeiro at 6wind.com>; Yongseok Koh <yskoh at mellanox.com>; Shahaf Shuler
> <shahafs at mellanox.com>; Tomasz Duszynski <tdu at semihalf.com>; Jianbo Liu <jianbo.liu at arm.com>; Alejandro Lucero
> <alejandro.lucero at netronome.com>; Hemant Agrawal <hemant.agrawal at nxp.com>; Shreyansh Jain <shreyansh.jain at nxp.com>; Harish
> Patil <harish.patil at cavium.com>; Rasesh Mody <rasesh.mody at cavium.com>; Andrew Rybchenko <arybchenko at solarflare.com>;
> Shrikrishna Khare <skhare at vmware.com>; Maxime Coquelin <maxime.coquelin at redhat.com>; Legacy, Allain (Wind River)
> <allain.legacy at windriver.com>; Richardson, Bruce <bruce.richardson at intel.com>; Gaetan Rivet <gaetan.rivet at 6wind.com>; Olivier Matz
> <olivier.matz at 6wind.com>
> Subject: Re: [dpdk-dev] Survey for final decision about per-port offload API
> 
> On 4/24/2018 12:08 PM, Ananyev, Konstantin wrote:
> > Hi Ferruh,
> >
> >>
> >> On 3/30/2018 2:47 PM, Thomas Monjalon wrote:
> >>> There are some discussions about a specific part of the offload API:
> >>> 	"To enable per-port offload, the offload should be set on both
> >>> 	device configuration and queue setup."
> >>>
> >>> It means the application must repeat the port offload flags
> >>> in rte_eth_conf.[rt]xmode.offloads and rte_eth_[rt]xconf.offloads,
> >>> when calling respectively rte_eth_dev_configure() and
> >>> rte_eth_[rt]x_queue_setup for each queue.
> >>>
> >>> The PMD must check if there is mismatch, i.e. a port offload not
> >>> repeated in queue setup.
> >>> There is a proposal to do this check at ethdev level:
> >>> 	http://dpdk.org/ml/archives/dev/2018-March/094023.html
> >>>
> >>> It was also proposed to relax the API and allow "forgetting" port
> >>> offloads in queue offloads:
> >>> 	http://dpdk.org/ml/archives/dev/2018-March/092978.html
> >>>
> >>> It would mean the offloads applied to a queue result of OR operation:
> >>> 	rte_eth_conf.[rt]xmode.offloads | rte_eth_[rt]xconf.offloads
> >>>
> >>> 1/ Do you agree with above API change?
> >>
> >> There is a detail of ability to disabling queue level offloads in queue_setup()
> >> function, I want to discuss here.
> >>
> >> Prolog:
> >> port level offload: An offload only can be applied port level, to all queues.
> >> queue level offload: An offload can be applied into individual queues of the port
> >>
> >> PMD reports port offload capability: port level offload + queue level offload
> >> PMD reports queue offload capability: queue level offload
> >>
> >>
> >> Above suggested change to API:
> >> - Application will be limited in configure() to set only an offload within "port
> >> offload capability"
> >> - Application will be limited in queue_setup() to set only an offload within
> >> "queue offload capability"
> >>
> >>
> >> This doesn't say much about disabling an offload in queue_setup(), as a rule:
> >> - An "port level offload" can't be disabled in queue_setup()
> >>
> >>
> >> There are two cases of disable:
> >> 1- Disabling a "queue level offload" enabled queue_setup() previously
> >> 2- Disabling a "queue level offload" enabled in configure()
> >>
> >> If second is not supported, to disable the offload, applications should
> >> stop->re-configure()->re-queue_setup()->start the port. But having this
> >> capability makes the offloading parameters more confusing for applications.
> >>
> >>
> >> I suggest adding disable support to fist one but not second one.
> >
> > Not sure why to introduce such limitation?
> 
> Not supporting second one?
> 
> To differentiate disable request for that case is harder. How can we say to
> disable a "queue level offloads" enabled by configure()?
> 
> It may be by setting these offloads in queue_setup() as well and when any
> offload is missing in queue_setup() it can be taken as disable request. This
> forces applications to duplicate/set "queue level offloads" enabled by
> configure() in the queue_setup() function by default.
> 
> This is an option, but my concern was to this may be harder to manage by
> applications.
> An application will have to remove "port level offload" from port_offloads and
> feed it into each queue_setup().
> 
> Also this is closer to existing API but not same, the difference is
> queue_setup() doesn't get "port level offload"
> 
> We can go with this one if there is a requirement for it.
> 
> And if we prefer to go with this option, perhaps we can think twice about
> changing exiting API because this will be very close the existing API. Only
> logically it is not correct to force application to set some offloads in
> queue_setup() for the PMD that doesn't support queue offload at all, this can be
> handled in PMD, and saves us of all the trouble of the change.

I suppose both ways are possible - though if we don't allow user to disable queue-specific
offload on particular queue, we would end up with most users just not specifying
any queue-specific offloads at configure() at all - just to have an ability to disable it in future
for particular queue.

Konstantin