[dpdk-dev] [PATCH v1] examples/l3fwd: fix jumbo packet drop issue

[Ferruh Yigit]

On 8/13/2021 9:42 AM, Ferruh Yigit wrote:
> On 8/13/2021 5:48 AM, Rohit Raj wrote:
>>
>>
>>> -----Original Message-----
>>> From: Ferruh Yigit <ferruh.yigit at intel.com>
>>> Sent: Wednesday, August 11, 2021 5:01 PM
>>> To: Rohit Raj <rohit.raj at nxp.com>
>>> Cc: dev at dpdk.org; Nipun Gupta <nipun.gupta at nxp.com>; Hemant Agrawal
>>> <hemant.agrawal at nxp.com>; Sachin Saxena <sachin.saxena at nxp.com>;
>>> Vanshika Shukla <vanshika.shukla at nxp.com>
>>> Subject: Re: [dpdk-dev] [PATCH v1] examples/l3fwd: fix jumbo packet drop
>>> issue
>>>
>>> On 7/27/2021 10:25 AM, rohit.raj at nxp.com wrote:
>>>> From: Rohit Raj <rohit.raj at nxp.com>
>>>>
>>>> l3fwd uses mbufs with 2KB data size. If we enable jumbo packets, it is
>>>> not able to store packets with size greater than 2KB, hence these
>>>> packets are dropped.
>>>>
>>>> This patch fixes this issue by enabling scatter for jumbo packet, if
>>>> it is supported by NIC.
>>>>
>>>> If scatter is not supported by NIC and max jumbo packet length is
>>>> greater than default mbuf data size, then application exits with
>>>> proper error message.
>>>>
>>>> Fixes: f68aad7904f ("examples/l3fwd: update")
>>>>
>>>> Signed-off-by: Rohit Raj <rohit.raj at nxp.com>
>>>> Signed-off-by: Sachin Saxena <sachin.saxena at nxp.com>
>>>> Signed-off-by: Vanshika Shukla <vanshika.shukla at nxp.com>
>>>> ---
>>>>   examples/l3fwd/main.c | 14 ++++++++++++++
>>>>   1 file changed, 14 insertions(+)
>>>>
>>>> diff --git a/examples/l3fwd/main.c b/examples/l3fwd/main.c index
>>>> 4cb800aa15..6aaaa8ecb5 100644
>>>> --- a/examples/l3fwd/main.c
>>>> +++ b/examples/l3fwd/main.c
>>>> @@ -1035,6 +1035,20 @@ l3fwd_poll_resource_setup(void)
>>>>                                "Error during getting device (port %u) info: %s\n",
>>>>                                portid, strerror(-ret));
>>>>
>>>> +             /* Enable Receive side SCATTER, if supported by NIC,
>>>> +              * when jumbo packet is enabled.
>>>> +              */
>>>> +             if (local_port_conf.rxmode.offloads &
>>>> +                             DEV_RX_OFFLOAD_JUMBO_FRAME){
>>>> +                     if (dev_info.rx_offload_capa & DEV_RX_OFFLOAD_SCATTER)
>>>> +                             local_port_conf.rxmode.offloads |=
>>>> +                                             DEV_RX_OFFLOAD_SCATTER;
>>>> +                     else if (local_port_conf.rxmode.max_rx_pkt_len >
>>>> +                                     RTE_MBUF_DEFAULT_DATAROOM)
>>>> +                             rte_exit(EXIT_FAILURE,
>>>> +                                     "Max packet length greater than
>>>> + default MBUF size\n");
>>>
>>> This is a configuration set by application. So application is failing itself because
>>> of configuration it sets, seems odd.
>>>
>>> I guess the jumbo frame can be enabled when user provides '--enable-jumbo'
>>> argument. What do you think adding above check where that argument is
>>> parsed.
>>
>> We need to check if hardware supports Rx scatter. To do this, we need to get Rx
>> capability. Hence, this is a better place to perform this check.
>>
> 
> Got it.
> 
> There is already a patch [1] that I am trying to update the 'max_rx_pkt_len'
> usage and remove OFFLOAD_JUMBO_FRAME flag [2] etc.. Those touch to the same
> piece of code, can you please check them?
> 
> After those patches, your update can be put into 'config_port_max_pkt_len()' I
> think. Can you be OK to wait them to merged first and make you patch on top of it?
> 
> [1]
> https://patches.dpdk.org/project/dpdk/patch/20210722172113.3236450-1-ferruh.yigit@intel.com/
> 
> [2]
> https://patches.dpdk.org/project/dpdk/patch/20210722172113.3236450-4-ferruh.yigit@intel.com/
> 

Hi Rohit,

Can you please check this patch on top of latest next-net?

Since 'DEV_RX_OFFLOAD_JUMBO_FRAME' is removed, only 'DEV_RX_OFFLOAD_SCATTER' capability
check can be sufficient.

And still same comment on 'RTE_MBUF_DEFAULT_DATAROOM' check, it is application's
responsibility to allocate the mempool for mbufs, so instead of checking the
mbuf buffer size and fail, application can allocate mempool size based on
user provided frame size value if driver doesn't support scattered Rx.


Thanks,
ferruh

>>>
>>> Btw, no need to enable scattered Rx if the packets fits into buffer, so above
>>> check can be done slightly different:
>>>
>>> if (max_rx_pkt_len > buffer_size)
>>>          if (OFFLOAD_SCATTER supported)
>>>                  enable OFFLOAD_SCATTER
>>>          else
>>>                  fail
>>
>> Agreed. I will do this in next version of the patch.
>>
>>>
>>>
>>>> +             }
>>>> +
>>>>                if (dev_info.tx_offload_capa & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
>>>>                        local_port_conf.txmode.offloads |=
>>>>                                DEV_TX_OFFLOAD_MBUF_FAST_FREE;
>>>>
>>
>