[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CALzJLG8wEHMEg-ixYPyoOCOQMcDVq03eVPo-7TSk0dR7mHM=Bg@mail.gmail.com>
Date: Wed, 5 Sep 2018 22:24:31 -0700
From: Saeed Mahameed <saeedm@....mellanox.co.il>
To: Konstantin Khlebnikov <khlebnikov@...dex-team.ru>
Cc: Tariq Toukan <tariqt@...lanox.com>,
Linux Netdev List <netdev@...r.kernel.org>,
Saeed Mahameed <saeedm@...lanox.com>,
Gal Pressman <galp@...lanox.com>,
Or Gerlitz <ogerlitz@...lanox.com>,
"David S. Miller" <davem@...emloft.net>
Subject: Re: [PATCH RFC] net/mlx5_en: switch to Toeplitz RSS hash by default
On Sun, Sep 2, 2018 at 2:55 AM, Konstantin Khlebnikov
<khlebnikov@...dex-team.ru> wrote:
> On 02.09.2018 12:29, Tariq Toukan wrote:
>>
>>
>>
>> On 31/08/2018 2:29 PM, Konstantin Khlebnikov wrote:
>>>
>>> XOR (MLX5_RX_HASH_FN_INVERTED_XOR8) gives only 8 bits.
>>> It seems not enough for RFS. All other drivers use toeplitz.
>>>
>>> Driver mlx4_en uses Toeplitz by default and warns if hash XOR is used
>>> together with NETIF_F_RXHASH (enabled by default too): "Enabling both
>>> XOR Hash function and RX Hashing can limit RPS functionality".
>>>
>>> XOR is default in mlx5_en since commit 2be6967cdbc9
>>> ("net/mlx5e: Support ETH_RSS_HASH_XOR").
>>>
>>> Hash function could be set via ethtool. But it would be nice to have
>>> single standard for drivers or proper description why this one is
>>> special.
>>>
>>> Signed-off-by: Konstantin Khlebnikov <khlebnikov@...dex-team.ru>
>>> ---
>>
>>
>> Hi Konstantin,
>>
>> Thanks for the patch.
>>
>> I understand the motivation.
>>
>> This change affects the default out-of-the-box behavior and requires a
>> full performance cycle. We'll run performance regression tomorrow, results
>> should be ready by EOW.
>> > I'll update.
>
>
> Ok, thank you.
>
> The only mention I've found in your documentation
> http://www.mellanox.com/related-docs/prod_software/Mellanox_EN_for_Linux_User_Manual_v4_4.pdf
>
> is
> ---
> 1.1.10 RSS Support
> 1.1.10.1 RSS Hash Function
> The device has the ability to use XOR as the RSS distribution function,
> instead of the default
> Toplitz function.
> The XOR function can be better distributed among driver's receive queues in
> small number of
> streams, where it distributes each TCP/UDP stream to a different queue.
> ---
>
> So Toeplitz is supposed to be default hash function for all versions of
> drivers and hardware.
>
> Also XOR8 seems vulnerable for ddos - hash is predictable, no random\secret
> vector, only 8 bits.
> So, it's easy to route all flows into one point. As we got it by accident.
>
> Moreover, in kernel 4.4.y hash switch via ethtool is broken and does not
> work =)
>
is it broken in mlx5 only or for the whole kernel ?
If it is mlx5 then this might be the reason:
commit 2d75b2bc8a8c0ce5567a6ecef52e194d117efe3f
net/mlx5e: Add ethtool RSS configuration options
was submitted to kernel 4.3
and an important fix for hash function change was submitted to 4.5:
commit bdfc028de1b3cd59490d5413a5c87b0fa50040c2
Author: Tariq Toukan <tariqt@...lanox.com>
Date: Mon Feb 29 21:17:12 2016 +0200
net/mlx5e: Fix ethtool RX hash func configuration change
We should modify TIRs explicitly to apply the new RSS configuration.
The light ndo close/open calls do not "refresh" them.
Fixes: 2d75b2bc8a8c ('net/mlx5e: Add ethtool RSS configuration options')
>>
>> Regards,
>> Tariq
Powered by blists - more mailing lists