Re: [dpdk-dev] [RFC 2/2] librte_ether: add new fields to rte_eth_dev_info struct

[Thomas Monjalon <thomas.monjalon@6wind.com>]

2016-04-14 10:44, Reshma Pattan:
> --- a/lib/librte_ether/rte_ethdev.h
> +++ b/lib/librte_ether/rte_ethdev.h
> @@ -908,6 +908,9 @@ struct rte_eth_dev_info {
>  	struct rte_eth_desc_lim rx_desc_lim;  /**< RX descriptors limits */
>  	struct rte_eth_desc_lim tx_desc_lim;  /**< TX descriptors limits */
>  	uint32_t speed_capa;  /**< Supported speeds bitmap (ETH_LINK_SPEED_). */
> +	/** number of queues configured by software*/
> +	uint16_t nb_rx_queues; /**< Number of RX queues. */
> +	uint16_t nb_tx_queues; /**< Number of TX queues. */
>  };

I think the ethdev design is strange for these structures.
struct rte_eth_dev is internal to be used inside the ethdev API
or by the drivers.
It contains struct rte_eth_dev_data which can be of interest for
the application, except the dev_private part (which could be
directly in struct rte_eth_dev).

So the global question is: how to share the device data with the
application?
Instead of giving a pointer or a copy of struct rte_eth_dev_data,
we have some different accessors:
	- rte_eth_dev_info_get() with a specific struct rte_eth_dev_info
which gathers a lot of info, not only from struct rte_eth_dev_data
	- rte_eth_macaddr_get()
	- rte_eth_dev_socket_id()
	- rte_eth_link_get() which is more than an accessor

I think having some specialized accessors is good.
But the rte_eth_dev_info_get() looks like to be a big request
without precise goal and going to break ABI really often.
There are some queues informations, some (not so precise)
offload capabilities, some steering (RSS/VMDq) informations,
the default configuration of some Intel NIC thresholds,
the speed capabilities, etc.

Shouldn't we try to streamline this API?