[Intel-wired-lan] [PATCH 00/11] XDP unaligned chunk placement support

[Jonathan Lemon <jonathan.lemon@gmail.com>]

On 2 Jul 2019, at 2:27, Richardson, Bruce wrote:

>> -----Original Message-----
>> From: Jakub Kicinski [mailto:jakub.kicinski at netronome.com]
>> Sent: Monday, July 1, 2019 10:20 PM
>> To: Laatz, Kevin <kevin.laatz@intel.com>
>> Cc: Jonathan Lemon <jonathan.lemon@gmail.com>; 
>> netdev at vger.kernel.org;
>> ast at kernel.org; daniel at iogearbox.net; Topel, Bjorn
>> <bjorn.topel@intel.com>; Karlsson, Magnus 
>> <magnus.karlsson@intel.com>;
>> bpf at vger.kernel.org; intel-wired-lan at lists.osuosl.org; Richardson, 
>> Bruce
>> <bruce.richardson@intel.com>; Loftus, Ciara <ciara.loftus@intel.com>
>> Subject: Re: [PATCH 00/11] XDP unaligned chunk placement support
>>
>> On Mon, 1 Jul 2019 15:44:29 +0100, Laatz, Kevin wrote:
>>> On 28/06/2019 21:29, Jonathan Lemon wrote:
>>>> On 28 Jun 2019, at 9:19, Laatz, Kevin wrote:
>>>>> On 27/06/2019 22:25, Jakub Kicinski wrote:
>>>>>> I think that's very limiting.? What is the challenge in 
>>>>>> providing
>>>>>> aligned addresses, exactly?
>>>>> The challenges are two-fold:
>>>>> 1) it prevents using arbitrary buffer sizes, which will be an 
>>>>> issue
>>>>> supporting e.g. jumbo frames in future.
>>>>> 2) higher level user-space frameworks which may want to use 
>>>>> AF_XDP,
>>>>> such as DPDK, do not currently support having buffers with 'fixed'
>>>>> alignment.
>>>>> ??? The reason that DPDK uses arbitrary placement is that:
>>>>> ??? ??? - it would stop things working on certain NICs which 
>>>>> need
>>>>> the actual writable space specified in units of 1k - therefore we
>>>>> need 2k
>>>>> + metadata space.
>>>>> ??? ??? - we place padding between buffers to avoid 
>>>>> constantly
>>>>> hitting the same memory channels when accessing memory.
>>>>> ??? ??? - it allows the application to choose the actual 
>>>>> buffer
>>>>> size it wants to use.
>>>>> ??? We make use of the above to allow us to speed up processing
>>>>> significantly and also reduce the packet buffer memory size.
>>>>>
>>>>> ??? Not having arbitrary buffer alignment also means an AF_XDP
>>>>> driver for DPDK cannot be a drop-in replacement for existing
>>>>> drivers in those frameworks. Even with a new capability to allow 
>>>>> an
>>>>> arbitrary buffer alignment, existing apps will need to be modified
>>>>> to use that new capability.
>>>>
>>>> Since all buffers in the umem are the same chunk size, the original
>>>> buffer address can be recalculated with some multiply/shift math.
>>>> However, this is more expensive than just a mask operation.
>>>
>>> Yes, we can do this.
>>
>> That'd be best, can DPDK reasonably guarantee the slicing is uniform?
>> E.g. it's not desperate buffer pools with different bases?
>
> It's generally uniform, but handling the crossing of (huge)page 
> boundaries
> complicates things a bit. Therefore I think the final option below
> is best as it avoids any such problems.
>
>>
>>> Another option we have is to add a socket option for querying the
>>> metadata length from the driver (assuming it doesn't vary per 
>>> packet).
>>> We can use that information to get back to the original address 
>>> using
>>> subtraction.
>>
>> Unfortunately the metadata depends on the packet and how much info 
>> the
>> device was able to extract.  So it's variable length.
>>
>>> Alternatively, we can change the Rx descriptor format to include the
>>> metadata length. We could do this in a couple of ways, for example,
>>> rather than returning the address as the start of the packet, 
>>> instead
>>> return the buffer address that was passed in, and adding another
>>> 16-bit field to specify the start of packet offset with that buffer.
>>> If using another 16-bits of the descriptor space is not desirable, 
>>> an
>>> alternative could be to limit umem sizes to e.g. 2^48 bits (256
>>> terabytes should be enough, right :-) ) and use the remaining 16 
>>> bits
>>> of the address as a packet offset. Other variations on these 
>>> approach
>>> are obviously possible too.
>>
>> Seems reasonable to me..
>
> I think this is probably the best solution, and also has the advantage 
> that
> a buffer retains its base address the full way through the cycle of Rx 
> and Tx.

I like this as well - it also has the advantage that drivers can keep
performing adjustments on the handle, which ends up just modifying the
offset.
-- 
Jonathan