Re: [dpdk-dev] [PATCH v4 1/3] ethdev: support API to set max LRO packet size

[Matan Azrad <matan@mellanox.com>]

Hi Konstantin

From: Ananyev, Konstantin
> Sent: Friday, November 8, 2019 6:29 PM
> To: Dekel Peled <dekelp@mellanox.com>; Matan Azrad
> <matan@mellanox.com>; Yigit, Ferruh <ferruh.yigit@intel.com>; Mcnamara,
> John <john.mcnamara@intel.com>; Kovacevic, Marko
> <marko.kovacevic@intel.com>; nhorman@tuxdriver.com;
> ajit.khaparde@broadcom.com; somnath.kotur@broadcom.com; Burakov,
> Anatoly <anatoly.burakov@intel.com>; xuanziyang2@huawei.com;
> cloud.wangxiaoyun@huawei.com; zhouguoyang@huawei.com; Lu, Wenzhuo
> <wenzhuo.lu@intel.com>; Shahaf Shuler <shahafs@mellanox.com>; Slava
> Ovsiienko <viacheslavo@mellanox.com>; rmody@marvell.com;
> shshaikh@marvell.com; maxime.coquelin@redhat.com; Bie, Tiwei
> <tiwei.bie@intel.com>; Wang, Zhihong <zhihong.wang@intel.com>;
> yongwang@vmware.com; Thomas Monjalon <thomas@monjalon.net>;
> arybchenko@solarflare.com; Wu, Jingjing <jingjing.wu@intel.com>;
> Iremonger, Bernard <bernard.iremonger@intel.com>
> Cc: dev@dpdk.org
> Subject: RE: [dpdk-dev] [PATCH v4 1/3] ethdev: support API to set max LRO
> packet size
> 
> 
> > >
> > >
> > > > > > > >>>> On 11/7/2019 12:35 PM, Dekel Peled wrote:
> > > > > > > >>>>> @@ -1266,6 +1286,18 @@ struct rte_eth_dev *
> > > > > > > >>>>>
> > > > > > > >>>> 	RTE_ETHER_MAX_LEN;
> > > > > > > >>>>>  	}
> > > > > > > >>>>>
> > > > > > > >>>>> +	/*
> > > > > > > >>>>> +	 * If LRO is enabled, check that the maximum
> > > aggregated
> > > > > > > packet
> > > > > > > >>>>> +	 * size is supported by the configured device.
> > > > > > > >>>>> +	 */
> > > > > > > >>>>> +	if (dev_conf->rxmode.offloads &
> > > > > > > DEV_RX_OFFLOAD_TCP_LRO) {
> > > > > > > >>>>> +		ret = check_lro_pkt_size(
> > > > > > > >>>>> +				port_id, dev_conf-
> > > > > > > >>>>> rxmode.max_lro_pkt_size,
> > > > > > > >>>>> +				dev_info.max_lro_pkt_size);
> > > > > > > >>>>> +		if (ret != 0)
> > > > > > > >>>>> +			goto rollback;
> > > > > > > >>>>> +	}
> > > > > > > >>>>> +
> > > > > > > >>>>
> > > > > > > >>>> This check forces applications that enable LRO to
> > > > > > > >>>> provide
> > > > > > > >> 'max_lro_pkt_size'
> > > > > > > >>>> config value.
> > > > > > > >>>
> > > > > > > >>> Yes.(we can break an API, we noticed it)
> > > > > > > >>
> > > > > > > >> I am not talking about API/ABI breakage, that part is OK.
> > > > > > > >> With this check, if the application requested LRO offload
> > > > > > > >> but not provided 'max_lro_pkt_size' value, device
> > > > > > > >> configuration will
> > > fail.
> > > > > > > >>
> > > > > > > > Yes
> > > > > > > >> Can there be a case application is good with whatever the
> > > > > > > >> PMD can support as max?
> > > > > > > > Yes can be - you know, we can do everything we want but it
> > > > > > > > is better to be
> > > > > > > consistent:
> > > > > > > > Due to the fact of Max rx pkt len field is mandatory for
> > > > > > > > JUMBO offload, max
> > > > > > > lro pkt len should be mandatory for LRO offload.
> > > > > > > >
> > > > > > > > So your question is actually why both, non-lro packets and
> > > > > > > > LRO packets max
> > > > > > > size are mandatory...
> > > > > > > >
> > > > > > > >
> > > > > > > > I think it should be important values for net applications
> > > management.
> > > > > > > > Also good for mbuf size managements.
> > > > > > > >
> > > > > > > >>>
> > > > > > > >>>> - Why it is mandatory now, how it was working before if
> > > > > > > >>>> it is mandatory value?
> > > > > > > >>>
> > > > > > > >>> It is the same as max_rx_pkt_len which is mandatory for
> > > > > > > >>> jumbo frame
> > > > > > > >> offload.
> > > > > > > >>> So now, when the user configures a LRO offload he must
> > > > > > > >>> to set max lro pkt
> > > > > > > >> len.
> > > > > > > >>> We don't want to confuse the user here with the max rx
> > > > > > > >>> pkt len
> > > > > > > >> configurations and behaviors, they should be with same logic.
> > > > > > > >>>
> > > > > > > >>> This parameter defines well the LRO behavior.
> > > > > > > >>> Before this, each PMD took its own interpretation to
> > > > > > > >>> what should be the
> > > > > > > >> maximum size for LRO aggregated packets.
> > > > > > > >>> Now, the user must say what is his intension, and the
> > > > > > > >>> ethdev can limit it
> > > > > > > >> according to the device capability.
> > > > > > > >>> By this way, also, the PMD can organize\optimize its
> > > > > > > >>> data-path
> > > more.
> > > > > > > >>> Also, the application can create different mempools for
> > > > > > > >>> LRO queues to
> > > > > > > >> allow bigger packet receiving for LRO traffic.
> > > > > > > >>>
> > > > > > > >>>> - What happens if PMD doesn't provide
> > > > > > > >>>> 'max_lro_pkt_size', so it is
> > > > > '0'?
> > > > > > > >>> Yes, you can see the feature description Dekel added.
> > > > > > > >>> This patch also updates all the PMDs support an LRO for
> > > > > > > >>> non-0
> > > value.
> > > > > > > >>
> > > > > > > >> Of course I can see the updates Matan, my point is "What
> > > > > > > >> happens if PMD doesn't provide 'max_lro_pkt_size'",
> > > > > > > >> 1) There is no check for it right, so it is acceptable?
> > > > > > > >
> > > > > > > > There is check.
> > > > > > > > If the capability is 0, any non-zero configuration will fail.
> > > > > > > >
> > > > > > > >> 2) Are we making this filed mandatory to provide for
> > > > > > > >> PMDs, it is easy to make new fields mandatory for PMDs
> > > > > > > >> but is this really
> > > > > necessary?
> > > > > > > >
> > > > > > > > Yes, for consistence.
> > > > > > > >
> > > > > > > >>>
> > > > > > > >>> as same as max rx pkt len, no?
> > > > > > > >>>
> > > > > > > >>>> - What do you think setting 'max_lro_pkt_size' config
> > > > > > > >>>> value to what PMD provided if application doesn't provide
> it?
> > > > > > > >>> Same answers as above.
> > > > > > > >>>
> > > > > > > >>
> > > > > > > >> If application doesn't care the value, as it has been
> > > > > > > >> till now, and not provided explicit 'max_lro_pkt_size',
> > > > > > > >> why not ethdev level use the value provided by PMD instead
> of failing?
> > > > > > > >
> > > > > > > > Again, same question we can ask on max rx pkt len.
> > > > > > > >
> > > > > > > > Looks like the packet size is very important value which
> > > > > > > > should be set by
> > > > > > > the application.
> > > > > > > >
> > > > > > > > Previous applications have no option to configure it, so
> > > > > > > > they haven't
> > > > > > > configure it, (probably cover it somehow) I think it is our
> > > > > > > miss to supply this info.
> > > > > > > >
> > > > > > > > Let's do it in same way as we do max rx pkt len (as this
> > > > > > > > patch main
> > > idea).
> > > > > > > > Later, we can change both to other meaning.
> > > > > > > >
> > > > > > >
> > > > > > > I think it is not a good reason to introduce a new mandatory
> > > > > > > config option for application because of 'max_rx_pkt_len' does it.
> > > > > >
> > > > > > It is mandatory only if LRO offload is configured.
> > > > >
> > > > > So max_rx_pkt_len will remain max size of one packet, while
> > > > > max_lro_len will be max accumulate size for each LRO session?
> > > > >
> > > >
> > > > Yes.
> > > >
> > > > > BTW, I think that for ixgbe max lro is RTE_IPV4_MAX_PKT_LEN.
> > > >
> > > > Please see my change in drivers/net/ixgbe/ixgbe_ethdev.c.
> > > > Change to RTE_IPV4_MAX_PKT_LEN?
> > > >
> > > > > ixgbe_vf, as I remember, doesn’t support LRO at all.
> > > >
> > > > Please see my change in drivers/net/ixgbe/ixgbe_vf_representor.c
> > > > Remove it?
> > >
> > > Yes, please for both.
> >
> > Will change in v5.
> >
> > >
> > > >
> > > > >
> > > > > >
> > > > > > > Will it work, if:
> > > > > > > - If application doesn't provide this value, use the PMD max
> > > > > >
> > > > > > May cause a problem if the mbuf size is not enough for the PMD
> > > maximum.
> > > > >
> > > > > Another question, what will happen if PMD will ignore that value
> > > > > and will generate packets bigger then requested?
> > > >
> > > > PMD should use this value and not ignore it.
> > >
> > > Hmm, ok but this patch updates mxl driver only...
> > > I suppose you expect other PMD maintainers to do the job for their
> > > PMDs, right?
> > > If so, are they aware (and agree) for this new hard requirement and
> > > changes required?
> > > Again what PMD should do if it can't support exact value?
> > > Let say user asked max_lro_size=20KB but PMD can do only 16KB or
> 24KB?
> > > Should it fail, or round to smallest, or ...?
> > >
> > > Actually I wonder, should it really be a hard requirement or more
> > > like a guidance to PMD?
> > > Why app needs and *exact* value for LRO size?
> >
> > The exact value should be configured to HW as LRO session limit.
> 
> But if the HW can't support this exact value, see the example above?
> In fact, shouldn't we allow PMD to forbid user to configure max LRO size?
> Let say if in dev_info max_lro_size==0, then PMD doesn't support LRO size
> configuration at all.
> That way PMDs who do support LRO, but don't want to (can't to) support
> configurable LRO size will stay untouched.

Each HW should support packet size limitation no matter if it is LRO packet or not:
How does the PMD limit the packet size for max rx packet len conf?
How does the PMD limit the packet size for the mbuf size?