Re: [PATCH v3 2/2] ethdev: allow pmd to advertise pool handle

[Olivier MATZ <olivier.matz@6wind.com>]

Hi Santosh,

On Tue, Aug 15, 2017 at 01:37:17PM +0530, Santosh Shukla wrote:
> Now that dpdk supports more than one mempool drivers and
> each mempool driver works best for specific PMD, example:
> - sw ring based mempool for Intel PMD drivers
> - dpaa2 HW mempool manager for dpaa2 PMD driver.
> - fpa HW mempool manager for Octeontx PMD driver.
> 
> Application like to know `preferred mempool vs PMD driver`
> information in advance before port setup.
> 
> Introducing rte_eth_dev_get_preferred_pool_ops() API,
> which allows PMD driver to advertise their pool capability to application.
> 
> Application side programing sequence would be:
> 
> char pref_mempool[RTE_MEMPOOL_OPS_NAMESIZE];
> rte_eth_dev_get_preferred_pool_ops(ethdev_port_id, pref_mempoolx /*out*/);
> rte_mempool_create_empty();
> rte_mempool_set_ops_byname( , pref_memppol, );
> rte_mempool_populate_default();
> 
> Signed-off-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
> ---
> v2 --> v3:
> - Updated version.map entry to DPDK_v17.11.
> 
> v1 --> v2:
> - Renamed _get_preferred_pool to _get_preferred_pool_ops().
> Per v1 review feedback, Olivier suggested to rename api
> to rte_eth_dev_pool_ops_supported(), considering that 2nd param
> for that api will return pool handle 'priority' for that port.
> However, per v1 [1], we're opting for approach 1) where
> ethdev API returns _preferred_ pool handle to application and Its upto
> application to decide on policy - whether application wants to create
> pool with received preferred pool handle or not. For more discussion details
> on this topic refer [1].

Well, I still think it would be more flexible to have an API like
 rte_eth_dev_pool_ops_supported(uint8_t port_id, const char *pool)

It supports the easy case (= one preferred mempool) without much pain,
and provides a more precise manner to describe what is supported or not
by the driver. Example: "pmd_foo" prefers "mempool_foo" (best perf), but
also supporst "mempool_stack" and "mempool_ring", but "mempool_bar"
won't work at all.

Having only one preferred pool_ops also prevents from smoothly renaming
a pool (supporting both during some time) or to have 2 names for
different variants of the same pool_ops (ex: ring_mp_mc, ring_sp_sc).

But if the users (I guess at least Cavium and NXP) are happy with
what you propose, I'm fine with it.

> --- a/lib/librte_ether/rte_ethdev.c
> +++ b/lib/librte_ether/rte_ethdev.c
> @@ -3409,3 +3409,21 @@ rte_eth_dev_adjust_nb_rx_tx_desc(uint8_t port_id,
>  
>  	return 0;
>  }
> +
> +int
> +rte_eth_dev_get_preferred_pool_ops(uint8_t port_id, char *pool)
> +{
> +	struct rte_eth_dev *dev;
> +	const char *tmp;
> +
> +	RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, -ENODEV);
> +
> +	dev = &rte_eth_devices[port_id];
> +
> +	if (*dev->dev_ops->get_preferred_pool_ops == NULL) {
> +		tmp = rte_eal_mbuf_default_mempool_ops();
> +		snprintf(pool, RTE_MBUF_POOL_OPS_NAMESIZE, "%s", tmp);
> +		return 0;
> +	}
> +	return (*dev->dev_ops->get_preferred_pool_ops)(dev, pool);
> +}

I think adding the length of the pool buffer to the function arguments
would be better: only documenting that the length is
RTE_MBUF_POOL_OPS_NAMESIZE looks a bit weak to me, because if one day it
changes to another value, the users of the function may not notice it
(no ABI/API change).


One more comment: it would be helpful to have one user of this API in
the example apps or testpmd.

Olivier