[PATCH v3 1/3] ethdev: introduce pool sort capability

[Ferruh Yigit]

On 9/13/2022 9:06 AM, Andrew Rybchenko wrote:
> On 9/2/22 10:00, Hanumanth Pothula wrote:
>> This patch adds support for the pool sort capability.
> 
> "Add support for serveral (?) mbuf pools per Rx queue."
> 
> I dislike the word "sort" in summary and the feature
> description. IMHO it is too restrictive for intended behaviour.
> 
> The key feature here is just support for more than one mbuf
> pool per Rx queue. That's it. Everything else should be out
> of scope of the definiteion.
> 

ack, and author already agreed to update it as 'MULTIPLE_POOL',
perhaps should we say 'MULTIPLE_MEMPOOL' ?

> If buffers from many pools are provided, the hardware may do
> whatever it wants with it. Use smaller buffers for small
> packets and bigger for big. Use bigger buffers for small
> packets if there is no small buffers available. Use big plus
> small buffer if Rx scatter is enabled and a packet fits in
> such combination. And so so on.
> 
> I.e. the feature should be orthogoal to Rx scatter.
> Rx scatter just says if driver/application allows to chain
> mbufs to receive a packet. If Rx scatter is disabled,
> a packet must be delivered in a single mbuf (either big or
> small). If Rx scatter is enable, a packet may be delivered
> using a chain of mbufs obtained from provided pools (either
> just one or many if several pools are supported).
> 
> Ideally the feature should be orthogonal to buffer split as
> well. I.e. provide many pools for different segments.
> May be it is an overkill to provide pools A and B for the first
> segment and C and D for the second. It could be limitted to the
> last segment only. If so, we need separate strcture (not
> rte_eth_rxseg) to pass many pools. IMHO, an array of mempools
> is sufficient - similar to Rx queue configuration.
> I.e. no extra length since data length may be derived from
> mempool element size.
> 
>> Some of the HW has support for choosing memory pools based on the
>> packet's size. The pool sort capability allows PMD to choose a
>> memory pool based on the packet's length.
>>
>> This is often useful for saving the memory where the application
>> can create a different pool to steer the specific size of the
>> packet, thus enabling effective use of memory.
>>
>> For example, let's say HW has a capability of three pools,
>>   - pool-1 size is 2K
>>   - pool-2 size is > 2K and < 4K
>>   - pool-3 size is > 4K
>> Here,
>>          pool-1 can accommodate packets with sizes < 2K
>>          pool-2 can accommodate packets with sizes > 2K and < 4K
>>          pool-3 can accommodate packets with sizes > 4K
>>
>> With pool sort capability enabled in SW, an application may create
>> three pools of different sizes and send them to PMD. Allowing PMD
>> to program HW based on packet lengths. So that packets with less
>> than 2K are received on pool-1, packets with lengths between 2K
>> and 4K are received on pool-2 and finally packets greater than 4K
>> are received on pool-3.
>>
>> The following two capabilities are added to the rte_eth_rxseg_capa
>> structure,
>> 1. pool_sort --> tells pool sort capability is supported by HW.
>> 2. max_npool --> max number of pools supported by HW.
>>
>> Defined new structure rte_eth_rxseg_sort, to be used only when pool
>> sort capability is present. If required this may be extended further
>> to support more configurations.
>>
>> Signed-off-by: Hanumanth Pothula <hpothula at marvell.com>
>>
>> v3:
>>   - Implemented Pool Sort capability as new Rx offload capability,
>>     RTE_ETH_RX_OFFLOAD_BUFFER_SORT.
>> v2:
>>   - Along with spec changes, uploading testpmd and driver changes.
>> ---
> 
> [snip]
>