Re: [PATCH 00/13] mvneta Buffer Management and enhancements

[Florian Fainelli <f.fainelli@gmail.com>]

(no top posting please)

On 02/12/15 00:26, Marcin Wojtas wrote:
> Hi Florian,
> 
> Can you please describe in more details, what would you expect from
> such special abstraction layer regarding buffer managers? I'd like to
> understand more of your expectations and evaluate possible work.

Well, something along these lines:

- have the ability to register a particular pool (location + number of
buffers) in a way that is relatively device agnostic (initialization
would of course be device specific)

- provide a set of buffer management APIs like those you proposed below,
and have some generic code that leverages what
drivers/net/ethernet/sun/niu.c does for instance

- introduce a netdev_alloc_skb_from_pool() or something like that which
would limit the amount of code to change in your network driver to
benefit from that feature so based

I am sure David would be able to suggest more detailed API.

> 
> Best regards,
> Marcin
> 
> 2015-11-30 17:25 GMT+01:00 David Miller <davem@redhat.com>:
>> From: Marcin Wojtas <mw@semihalf.com>
>> Date: Mon, 30 Nov 2015 15:13:22 +0100
>>
>>> What kind of abstraction and helpers do you mean? Some kind of API
>>> (e.g. bm_alloc_buffer, bm_initialize_ring bm_put_buffer,
>>> bm_get_buffer), which would be used by platform drivers (and specific
>>> aplications if one wants to develop on top of the kernel)?
>>>
>>> In general, what is your top-view of such solution and its cooperation
>>> with the drivers?
>>
>> The tricky parts involved have to do with allocating pages for the
>> buffer pools and minimizing the number of atomic refcounting
>> operations on those pages for for the puts and gets, particularly
>> around buffer replenish runs.
>>
>> For example, if you're allocating a page for a buffer pool the device
>> will chop into N (for any N < PAGE_SIZE) byte pieces, you can
>> eliminate many atomic operations.


-- 
Florian