svn commit: r341578 - head/sys/dev/mlx5/mlx5_en

[Hans Petter Selasky]

On 2019-12-17 18:14, Andrey V. Elsukov wrote:
> On 13.12.2019 17:27, Hans Petter Selasky wrote:
>> On 2019-12-13 14:40, Andrey V. Elsukov wrote:
>>> On 05.12.2018 17:20, Slava Shwartsman wrote:
>>>> Author: slavash
>>>> Date: Wed Dec  5 14:20:57 2018
>>>> New Revision: 341578
>>>> URL: https://svnweb.freebsd.org/changeset/base/341578
>>>>
>>>> Log:
>>>>     mlx5en: Remove the DRBR and associated logic in the transmit path.
>>>>        The hardware queues are deep enough currently and using the
>>>> DRBR and associated
>>>>     callbacks only leads to more task switching in the TX path. The is
>>>> also a race
>>>>     setting the queue_state which can lead to hung TX rings.
>>>
>>> JFYI. We have compared the same router+firewall workloads on the host
>>> with this change and before, and I can say, that without DRBR on TX now
>>> we constantly have several percents of packets drops due to ENOBUFS
>>> error from mlx5e_xmit().
>>>
>>
>> Have you tried to tune the TX/RX parameters?
>>
>> Especially the tx_queue_size .
> 
> We use the following settings:
> % sysctl dev.mce.4.conf. | grep que
> dev.mce.4.conf.rx_queue_size: 16384
> dev.mce.4.conf.tx_queue_size: 16384
> dev.mce.4.conf.rx_queue_size_max: 16384
> dev.mce.4.conf.tx_queue_size_max: 16384
> 
> Also, previously I have patched MLX5E_SQ_TX_QUEUE_SIZE value up to 16384.

Hi,

What about the other parameters. Did you tune any of those?

At what rate does this happen?

Can you send me the full dev.mce.4 sysctl tree off-list?

--HPS