[dpdk-dev] [PATCH v4 3/5] net/ice: fix max mtu size packets with vlan tag cannot be received by default

[Ananyev, Konstantin]

> >> On 10/20/2020 10:07 AM, Ananyev, Konstantin wrote:
> >>>
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> testpmd will initialize default max packet length to 1518 which
> >>>>>>>>>>>>>>>> doesn't include vlan tag size in ether overheader. Once, send the
> >>>>>>>>>>>>>>>> max mtu length packet with vlan tag, the max packet length will
> >>>>>>>>>>>>>>>> exceed 1518 that will cause packets dropped directly from NIC hw
> >>>>>>>>>>>> side.
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> ice can support dual vlan tags that need more 8 bytes for max
> >>>>>>>>>>>>>>>> packet size, so, configures the correct max packet size in
> >>>>>>>>>>>>>>>> dev_config
> >>>>>>>>>>>>> ops.
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> Fixes: 50cc9d2a6e9d ("net/ice: fix max frame size")
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> Signed-off-by: SteveX Yang <stevex.yang at intel.com>
> >>>>>>>>>>>>>>>> ---
> >>>>>>>>>>>>>>>>       drivers/net/ice/ice_ethdev.c | 11 +++++++++++
> >>>>>>>>>>>>>>>>       1 file changed, 11 insertions(+)
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> diff --git a/drivers/net/ice/ice_ethdev.c
> >>>>>>>>>>>>>>>> b/drivers/net/ice/ice_ethdev.c index
> >>>>>>>>>>>>>>>> cfd357b05..6b7098444 100644
> >>>>>>>>>>>>>>>> --- a/drivers/net/ice/ice_ethdev.c
> >>>>>>>>>>>>>>>> +++ b/drivers/net/ice/ice_ethdev.c
> >>>>>>>>>>>>>>>> @@ -3146,6 +3146,7 @@ ice_dev_configure(struct rte_eth_dev
> >>>>>>>>>> *dev)
> >>>>>>>>>>>>>>>> struct ice_adapter *ad =
> >>>>>>>>>>>>>>>> ICE_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
> >>>>>>>>>>>>>>>>       struct ice_pf *pf =
> >>>>>>>>>>>>>>>> ICE_DEV_PRIVATE_TO_PF(dev->data->dev_private);
> >>>>>>>>>>>>>>>> +uint32_t frame_size = dev->data->mtu + ICE_ETH_OVERHEAD;
> >>>>>>>>>>>>>>>>       int ret;
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>>       /* Initialize to TRUE. If any of Rx queues doesn't meet the @@
> >>>>>>>>>>>>>>>> -3157,6
> >>>>>>>>>>>>>>>> +3158,16 @@ ice_dev_configure(struct rte_eth_dev *dev)
> >>>>>>>>>>>>>>>>       if (dev->data->dev_conf.rxmode.mq_mode &
> >>>>>>>>>> ETH_MQ_RX_RSS_FLAG)
> >>>>>>>>>>>>>>>> dev->data->dev_conf.rxmode.offloads |=
> >>>>>>>>>>>>> DEV_RX_OFFLOAD_RSS_HASH;
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> +/**
> >>>>>>>>>>>>>>>> + * Considering QinQ packet, max frame size should be equal or
> >>>>>>>>>>>>>>>> + * larger than total size of MTU and Ether overhead.
> >>>>>>>>>>>>>>>> + */
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> +if (frame_size > dev->data->dev_conf.rxmode.max_rx_pkt_len) {
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>> Why we need this check?
> >>>>>>>>>>>>>>> Can we just call ice_mtu_set directly
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> I think that without that check we can silently overwrite provided
> >>>>>>>>>>>>>> by user dev_conf.rxmode.max_rx_pkt_len value.
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> OK, I see
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> But still have one question
> >>>>>>>>>>>>> dev->data->mtu is initialized to 1518 as default , but if
> >>>>>>>>>>>>> dev->data->application set
> >>>>>>>>>>>>> dev_conf.rxmode.max_rx_pkt_len = 1000 in dev_configure.
> >>>>>>>>>>>>> does that mean we will still will set mtu to 1518, is this expected?
> >>>>>>>>>>>>>
> >>>>>>>>>>>>
> >>>>>>>>>>>> max_rx_pkt_len should be larger than mtu at least, so we should raise
> >>>>>>>>>>>> the max_rx_pkt_len (e.g.:1518) to hold expected mtu value (e.g.: 1500).
> >>>>>>>>>>>
> >>>>>>>>>>> Ok, this describe the problem more general and better to replace exist
> >>>>>>>>>> code comment and commit log for easy understanding.
> >>>>>>>>>>> Please send a new version for reword
> >>>>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> I didn't really get this set.
> >>>>>>>>>>
> >>>>>>>>>> Application explicitly sets 'max_rx_pkt_len' to '1518', and a frame bigger than
> >>>>>>>>>> this size is dropped.
> >>>>>>>>>
> >>>>>>>>> Sure, it is normal case for dropping oversize data.
> >>>>>>>>>
> >>>>>>>>>> Isn't this what should be, why we are trying to overwrite user configuration
> >>>>>>>>>> in PMD to prevent this?
> >>>>>>>>>>
> >>>>>>>>>
> >>>>>>>>> But it is a confliction that application/user sets mtu & max_rx_pkt_len at the same time.
> >>>>>>>>> This fix will make a decision when confliction occurred.
> >>>>>>>>> MTU value will come from user operation (e.g.: port config mtu 0 1500) directly,
> >>>>>>>>> so, the max_rx_pkt_len will resize itself to adapt expected MTU value if its size is smaller than MTU + Ether overhead.
> >>>>>>>>>
> >>>>>>>>>> During eth_dev allocation, mtu set to default '1500', by ethdev layer.
> >>>>>>>>>> And testpmd sets 'max_rx_pkt_len' by default to '1518'.
> >>>>>>>>>> I think Qi's concern above is valid, what is user set 'max_rx_pkt_len' to '1000'
> >>>>>>>>>> and mean it? PMD will not honor the user config.
> >>>>>>>>>
> >>>>>>>>> I'm not sure when set 'mtu' to '1500' and 'max_rx_pkt_len' to '1000', what's the behavior expected?
> >>>>>>>>> If still keep the 'max_rx_pkt_len' value, that means the larger 'mtu' will be invalid.
> >>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> Why not simply increase the default 'max_rx_pkt_len' in testpmd?
> >>>>>>>>>>
> >>>>>>>>> The default 'max_rx_pkt_len' has been initialized to generical value (1518) and default 'mtu' is '1500' in testpmd,
> >>>>>>>>> But it isn't suitable to those NIC drivers which Ether overhead is larger than 18. (e.g.: ice, i40e) if 'mtu' value is preferable.
> >>>>>>>>>
> >>>>>>>>>> And I guess even better what we need is to tell to the application what the
> >>>>>>>>>> frame overhead PMD accepts.
> >>>>>>>>>> So the application can set proper 'max_rx_pkt_len' value per port for a
> >>>>>>>>>> given/requested MTU value.
> >>>>>>>>>> @Ian, cc'ed, was complaining almost same thing years ago, these PMD
> >>>>>>>>>> overhead macros and 'max_mtu'/'min_mtu' added because of that, perhaps
> >>>>>>>>>> he has a solution now?
> >>>>>>>>
> >>>>>>>>     From my perspective the main problem here:
> >>>>>>>> We have 2 different variables for nearly the same thing:
> >>>>>>>> rte_eth_dev_data.mtu and rte_eth_dev_data.dev_conf.max_rx_pkt_len.
> >>>>>>>> and 2 different API to update them: dev_mtu_set() and dev_configure().
> >>>>>>>
> >>>>>>> According API 'max_rx_pkt_len' is 'Only used if JUMBO_FRAME enabled'
> >>>>>>> Although not sure that is practically what is done for all drivers.
> >>>>>>
> >>>>>> I think most of Intel PMDs use it unconditionally.
> >>>>>>
> >>>>>>>
> >>>>>>>> And inside majority of Intel PMDs we don't keep these 2 variables in sync:
> >>>>>>>> - mtu_set() will update both variables.
> >>>>>>>> - dev_configure() will update only max_rx_pkt_len, but will keep mtu intact.
> >>>>>>>>
> >>>>>>>> This patch fixes this inconsistency, which I think is a good thing.
> >>>>>>>> Though yes, it introduces change in behaviour.
> >>>>>>>>
> >>>>>>>> Let say the code:
> >>>>>>>> rte_eth_dev_set_mtu(port, 1500);
> >>>>>>>> dev_conf.max_rx_pkt_len = 1000;
> >>>>>>>> rte_eth_dev_configure(port, 1, 1, &dev_conf);
> >>>>>>>>
> >>>>>>>
> >>>>>>> 'rte_eth_dev_configure()' is one of the first APIs called, it is called before
> >>>>>>> 'rte_eth_dev_set_mtu().
> >>>>>>
> >>>>>> Usually yes.
> >>>>>> But you can still do sometimes later: dev_mtu_set(); ...; dev_stop(); dev_configure(); dev_start();
> >>>>>>
> >>>>>>>
> >>>>>>> When 'rte_eth_dev_configure()' is called, MTU is set to '1500' by default by
> >>>>>>> ethdev layer, so it is not user configuration, but 'max_rx_pkt_len' is.
> >>>>>>
> >>>>>> See above.
> >>>>>> PMD doesn't know where this MTU value came from (default ethdev value or user specified value)
> >>>>>> and probably it shouldn't care.
> >>>>>>
> >>>>>>>
> >>>>>>> And later, when 'rte_eth_dev_set_mtu()' is called, but MTU and 'max_rx_pkt_len'
> >>>>>>> are updated (mostly).
> >>>>>>
> >>>>>> Yes, in mtu_set() we update both.
> >>>>>> But we don't update MTU in dev_configure(), only max_rx_pkt_len.
> >>>>>> That what this patch tries to fix (as I understand it).
> >>>>>
> >>>>> To be more precise - it doesn't change MTU value in dev_configure(),
> >>>>> but instead doesn't allow max_rx_pkt_len to become smaller
> >>>>> then MTU + OVERHEAD.
> >>>>> Probably changing MTU value instead is a better choice.
> >>>>>
> >>>>
> >>>> +1 to change mtu for this case.
> >>>> And this is what happens in practice when there is no 'rte_eth_dev_set_mtu()'
> >>>> call, since PMD is using ('max_rx_pkt_len' - OVERHEAD) to set MTU.
> >>>
> >>> Hmm, I don't see that happens within Intel PMDs.
> >>> As I can read the code: if user never call mtu_set(), then MTU value is left intact.
> >>>
> >>
> >> I was checking ice,
> >> in 'ice_dev_start()', 'rxmode.max_rx_pkt_len' is used to configure the device.
> >
> > Yes, I am not arguing with that.
> > What I am saying - dev_config() doesn't update MTU based on max_rx_pkt_len.
> > While it probably should.
> >
> 
> Yes 'dev_configure()' doesn't update the 'dev->data->mtu' and 'max_rx_pkt_len' &
> 'dev->data->mtu' may diverge there.
> 
> I think best place to update 'dev->data->mtu' is where the device is actually
> updated, but to prevent the diversion above we can update 'dev->data->mtu' in
> ethdev layer, in 'rte_eth_dev_configure()' based on 'max_rx_pkt_len', will it work?

I think - yes.
At least, I don't foresee any implications with that.

> 
> Only concern I see is if user reads the MTU ('rte_eth_dev_get_mtu()') after
> 'rte_eth_dev_configure()' but before device configured, user will get the wrong
> value, I guess that problem was already there but changing default value may
> make it more visible.
> 
> >>
> >>>> But this won't solve the problem Steve is trying to solve.
> >>>
> >>> You mean we still need to update test-pmd code to calculate max_rx_pkt_len
> >>> properly for default mtu value?
> >>>
> >>
> >> Yes.
> >> Because target of this set is able to receive packets with payload size
> >> 'RTE_ETHER_MTU', if MTU is updated according to the provided 'max_rx_pkt_len',
> >> device still won't able to receive those packets.
> >
> > Agree.
> >
> >>
> >>>>>>>
> >>>>>>>
> >>>>>>>> Before the patch will result:
> >>>>>>>> mtu==1500, max_rx_pkt_len=1000;  //out of sync looks wrong to me
> >>>>>>>>
> >>>>>>>> After the patch:
> >>>>>>>> mtu=1500, max_rx_ptk_len=1518; // in sync, change in behaviour.
> >>>>>>>>
> >>>>>>>> If you think we need to preserve current behaviour,
> >>>>>>>> then I suppose the easiest thing would be to change dev_config() code
> >>>>>>>> to update mtu value based on max_rx_pkt_len.
> >>>>>>>> I.E: dev_configure {...; mtu_set(max_rx_pkt_len - OVERHEAD); ...}
> >>>>>>>> So the code snippet above will result:
> >>>>>>>> mtu=982,max_rx_pkt_len=1000;
> >>>>>>>>
> >>>>>>>
> >>>>>>> The 'max_rx_ptk_len' is annoyance for a long time, what do you think to just
> >>>>>>> drop it?
> >>>>>>>
> >>>>>>> By default device will be up with default MTU (1500), later
> >>>>>>> 'rte_eth_dev_set_mtu' can be used to set the MTU, no frame size setting at all.
> >>>>>>>
> >>>>>>> Will this work?
> >>>>>>
> >>>>>> I think it might, but that's a big change, probably too risky at that stage...
> >>>>>>
> >>>>
> >>>> Defintely, I was thinking for 21.11. Let me send a deprecation notice and see
> >>>> what happens.
> >>>>
> >>>>>>
> >>>>>>>
> >>>>>>>
> >>>>>>> And for short term, for above Intel PMDs, there must be a place this
> >>>>>>> 'max_rx_pkt_len' value taken into account (mostly 'start()' dev_ops), that
> >>>>>>> function can be updated to take 'max_rx_pkt_len' only if JUMBO_FRAME set,
> >>>>>>> otherwise use the 'MTU' value.
> >>>>>>
> >>>>>> Even if we'll use max_rx_pkt_len only when if JUMBO_FRAME is set,
> >>>>>> I think we still need to keep max_rx_pkt_len and MTU values in sync.
> >>>>>>
> >>>>>>>
> >>>>>>> Without 'start()' updated the current logic won't work after stop & start anyway.
> >>>>>>>
> >>>>>>>
> >>>>>>>>
> >>>>>>>>
> >>>>>>>>
> >>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> And why this same thing can't happen to other PMDs? If this is a problem for
> >>>>>>>>>> all PMDs, we should solve in other level, not for only some PMDs.
> >>>>>>>>>>
> >>>>>>>>> No, all PMDs exist the same issue, another proposal:
> >>>>>>>>>      -  rte_ethdev provides the unique resize 'max_rx_pkt_len' in rte_eth_dev_configure();
> >>>>>>>>>      - provide the uniform API for fetching the NIC's supported Ether Overhead size;
> >>>>>>>>> Is it feasible?
> >>>>>>>>>
> >>>>>>>>>>>
> >>>>>>>>>>>> Generally, the mtu value can be adjustable from user (e.g.: ip link
> >>>>>>>>>>>> set ens801f0 mtu 1400), hence, we just adjust the max_rx_pkt_len to
> >>>>>>>>>>>> satisfy mtu requirement.
> >>>>>>>>>>>>
> >>>>>>>>>>>>> Should we just call ice_mtu_set(dev, dev_conf.rxmode.max_rx_pkt_len)
> >>>>>>>>>>>>> here?
> >>>>>>>>>>>> ice_mtu_set(dev, mtu) will append ether overhead to
> >>>>>>>>>>>> frame_size/max_rx_pkt_len, so we need pass the mtu value as the 2nd
> >>>>>>>>>>>> parameter, or not the max_rx_pkt_len.
> >>>>>>>>>>>>
> >>>>>>>>>>>>>
> >>>>>>>>>>>>>
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>>> And please remove above comment, since ether overhead is already
> >>>>>>>>>>>>>> considered in ice_mtu_set.
> >>>>>>>>>>>> Ether overhead is already considered in ice_mtu_set, but it also
> >>>>>>>>>>>> should be considered as the adjustment condition that if ice_mtu_set
> >>>>>>>>>> need be invoked.
> >>>>>>>>>>>> So, it perhaps should remain this comment before this if() condition.
> >>>>>>>>>>>>
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> +ret = ice_mtu_set(dev, dev->data->mtu); if (ret != 0) return
> >>>>>>>>>>>>>>>> +ret; }
> >>>>>>>>>>>>>>>> +
> >>>>>>>>>>>>>>>>       ret = ice_init_rss(pf);
> >>>>>>>>>>>>>>>>       if (ret) {
> >>>>>>>>>>>>>>>>       PMD_DRV_LOG(ERR, "Failed to enable rss for PF");
> >>>>>>>>>>>>>>>> --
> >>>>>>>>>>>>>>>> 2.17.1
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>
> >>>>>>>>>>>>
> >>>>>>>>>>>
> >>>>>>>>
> >>>>>
> >>>
> >