[dpdk-dev] [RFC 0/7] hide eth dev related structures

[Ferruh Yigit]

On 8/26/2021 1:37 PM, Jerin Jacob wrote:
> On Fri, Aug 20, 2021 at 9:59 PM Konstantin Ananyev
> <konstantin.ananyev at intel.com> wrote:
>>
>> NOTE: This is just an RFC to start further discussion and collect the feedback.
>> Due to significant amount of work, changes required are applied only to two
>> PMDs so far: net/i40e and net/ice.
>> So to build it you'll need to add:
>> -Denable_drivers='common/*,mempool/*,net/ice,net/i40e'
>> to your config options.
> 
>>
>> That approach was selected to avoid(/minimize) possible performance losses.
>>
>> So far I done only limited amount functional and performance testing.
>> Didn't spot any functional problems, and performance numbers
>> remains the same before and after the patch on my box (testpmd, macswap fwd).
> 
> 
> Based on testing on octeonxt2. We see some regression in testpmd and
> bit on l3fwd too.
> 
> Without patch: 73.5mpps/core in testpmd iofwd
> With out patch: 72 5mpps/core in testpmd iofwd
> 

So roughly 1.3% drop.

> Based on my understanding it is due to additional indirection.
> 

What is the additional indirection? I can see all dereferences area same.

> My suggestion to fix the problem by:
> Removing the additional `data` redirection and pull callback function
> pointers back
> and keep rest as opaque as done in the existing patch like [1]
> 
> I don't believe this has any real implication on future ABI stability
> as we will not be adding
> any new item in rte_eth_fp in any way as new features can be added in slowpath
> rte_eth_dev as mentioned in the patch.
> 
> [2] is the patch of doing the same as I don't see any performance
> regression after [2].
> 

Only there is an additional check after Konstantin's patch (in both
'rte_eth_rx_burst()' & 'rte_eth_tx_burst()'):
"
if (port_id >= RTE_MAX_ETHPORTS)
        return -EINVAL;
"

Which I can see removed again in your patch [2] which fixes the regression. I
wonder if this can be reason of restoring the performance drop, can you please
test original patch by just removing the 'RTE_MAX_ETHPORTS' check?

Thanks,
ferruh

> 
> [1]
> - struct rte_eth_burst_api {
> - struct rte_eth_fp {
> + void *data;
>   rte_eth_rx_burst_t rx_pkt_burst;
>   /**< PMD receive function. */
>   rte_eth_tx_burst_t tx_pkt_burst;
> @@ -85,8 +100,19 @@ struct rte_eth_burst_api {
>   /**< Check the status of a Rx descriptor. */
>   rte_eth_tx_descriptor_status_t tx_descriptor_status;
>   /**< Check the status of a Tx descriptor. */
> + /**
> + * User-supplied functions called from rx_burst to post-process
> + * received packets before passing them to the user
> + */
> + struct rte_eth_rxtx_callback
> + *post_rx_burst_cbs[RTE_MAX_QUEUES_PER_PORT];
> + /**
> + * User-supplied functions called from tx_burst to pre-process
> + * received packets before passing them to the driver for transmission.
> + */
> + struct rte_eth_rxtx_callback *pre_tx_burst_cbs[RTE_MAX_QUEUES_PER_PORT];
>   uintptr_t reserved[2];
> -} __rte_cache_min_aligned;
> +} __rte_cache_aligned;
> 
> [2]
> https://pastebin.com/CuqkrCW4
>