Re: [dpdk-dev] [PATCH 1/5] ethdev: add API to negotiate support for Rx meta information

[Jerin Jacob <jerinjacobk@gmail.com>]

On Thu, Sep 2, 2021 at 7:54 PM Ivan Malov <ivan.malov@oktetlabs.ru> wrote:
>
> Per-packet meta information (flag, mark and the likes) might
> be expensive to deliver in terms of small packet performance.
> If the features are not enabled by default, enabling them at
> short notice (for example, when a flow rule with action MARK
> gets created) without traffic disruption may not be possible.
>
> Letting applications request delivery of Rx meta information
> during initialisation can solve the said problem efficiently.
>
> Technically, that could be accomplished by defining new bits
> in DEV_RX_OFFLOAD namespace, but the ability to extract meta
> data cannot be considered an offload on its own. For example,
> Rx checksumming is an offload, while mark delivery is not as
> it needs an external party, a flow rule with action MARK, to
> hold the value and trigger mark insertion in the first place.
>
> With this in mind, add a means to let applications negotiate
> adapter support for the very delivery of Rx meta information.

Good stuff.


>
> Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
> Reviewed-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
> ---
>  lib/ethdev/ethdev_driver.h | 19 +++++++++++
>  lib/ethdev/rte_ethdev.c    | 25 +++++++++++++++
>  lib/ethdev/rte_ethdev.h    | 66 ++++++++++++++++++++++++++++++++++++++
>  lib/ethdev/version.map     |  3 ++
>  4 files changed, 113 insertions(+)
>
> diff --git a/lib/ethdev/ethdev_driver.h b/lib/ethdev/ethdev_driver.h
> index 40e474aa7e..3e29555fc7 100644
> --- a/lib/ethdev/ethdev_driver.h
> +++ b/lib/ethdev/ethdev_driver.h
> @@ -789,6 +789,22 @@ typedef int (*eth_get_monitor_addr_t)(void *rxq,
>  typedef int (*eth_representor_info_get_t)(struct rte_eth_dev *dev,
>         struct rte_eth_representor_info *info);
>
> +/**
> + * @internal
> + * Negotiate support for specific fractions of Rx meta information.
> + *
> + * @param[in] dev
> + *   Port (ethdev) handle
> + *
> + * @param[inout] features

All the params are in by default, I think, only [out] only needed.

> + *   Feature selection buffer
> + *
> + * @return
> + *   Negative errno value on error, zero otherwise
> + */
> +typedef int (*eth_negotiate_rx_meta_t)(struct rte_eth_dev *dev,
> +                                      uint64_t *features);
> +
>  /**
>   * @internal A structure containing the functions exported by an Ethernet driver.
>   */
> @@ -949,6 +965,9 @@ struct eth_dev_ops {
>
>         eth_representor_info_get_t representor_info_get;
>         /**< Get representor info. */
> +
> +       eth_negotiate_rx_meta_t negotiate_rx_meta;
> +       /**< Negotiate support for specific fractions of Rx meta information. */
>  };
>
>  /**
> diff --git a/lib/ethdev/rte_ethdev.c b/lib/ethdev/rte_ethdev.c
> index 9d95cd11e1..821450cbf9 100644
> --- a/lib/ethdev/rte_ethdev.c
> +++ b/lib/ethdev/rte_ethdev.c
> @@ -6311,6 +6311,31 @@ rte_eth_representor_info_get(uint16_t port_id,
>         return eth_err(port_id, (*dev->dev_ops->representor_info_get)(dev, info));
>  }
>
> +int
> +rte_eth_negotiate_rx_meta(uint16_t port_id, uint64_t *features)
> +{
> +       struct rte_eth_dev *dev;
> +
> +       RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, -ENODEV);
> +       dev = &rte_eth_devices[port_id];
> +
> +       if (dev->data->dev_configured != 0) {
> +               RTE_ETHDEV_LOG(ERR,
> +                       "The port (id=%"PRIu16") is already configured\n",
> +                       port_id);
> +               return -EBUSY;
> +       }
> +
> +       if (features == NULL) {
> +               RTE_ETHDEV_LOG(ERR, "Invalid features (NULL)\n");
> +               return -EINVAL;
> +       }
> +
> +       RTE_FUNC_PTR_OR_ERR_RET(*dev->dev_ops->negotiate_rx_meta, -ENOTSUP);
> +       return eth_err(port_id,
> +                      (*dev->dev_ops->negotiate_rx_meta)(dev, features));
> +}
> +
>  RTE_LOG_REGISTER_DEFAULT(rte_eth_dev_logtype, INFO);
>
>  RTE_INIT(ethdev_init_telemetry)
> diff --git a/lib/ethdev/rte_ethdev.h b/lib/ethdev/rte_ethdev.h
> index d2b27c351f..ac4d164aa8 100644
> --- a/lib/ethdev/rte_ethdev.h
> +++ b/lib/ethdev/rte_ethdev.h
> @@ -4888,6 +4888,72 @@ __rte_experimental
>  int rte_eth_representor_info_get(uint16_t port_id,
>                                  struct rte_eth_representor_info *info);
>
> +/**
> + * The ethdev will be able to detect flagged packets provided that
> + * there are active flow rules comprising the corresponding action.
> + */
> +#define RTE_ETH_RX_META_USER_FLAG (UINT64_C(1) << 0)

I think, we need to add @see note in rte_flow API where this is requested.


> +
> +/**
> + * The ethdev will manage to see mark IDs in packets provided that
> + * there are active flow rules comprising the corresponding action.
> + */
> +#define RTE_ETH_RX_META_USER_MARK (UINT64_C(1) << 1)

Same as above

> +
> +/**
> + * The ethdev will be able to identify partially offloaded packets
> + * and process rte_flow_get_restore_info() invocations accordingly
> + * provided that there're so-called "tunnel_set" flow rules in use.
> + */
> +#define RTE_ETH_RX_META_TUNNEL_ID (UINT64_C(1) << 2)

Same as above


> +
> +/**
> + * @warning
> + * @b EXPERIMENTAL: this API may change without prior notice
> + *
> + * Negotiate support for specific fractions of Rx meta information.
> + *
> + * This function has to be invoked as early as possible, precisely,
> + * before the first rte_eth_dev_configure() invocation, to let the
> + * PMD make preparations which might be hard to do on later stages.
> + *
> + * The negotiation process is assumed to be carried out as follows:
> + *
> + * - the application composes a mask of preferred Rx meta features
> + *   intending to enable at least some of them and invokes the API;
> + *
> + * - the ethdev driver reports back the optimal (from its point of
> + *   view) subset of the initial feature set thus agreeing to make
> + *   those comprising the subset simultaneously available to users;
> + *
> + * - should the application find the result unsatisfactory, it may
> + *   come up with another pick of preferred features and try again;
> + *
> + * - the application can pass zero to clear the negotiation result;
> + *
> + * - the last negotiated result takes effect upon the ethdev start.
> + *
> + * If the method itself is unsupported by the PMD, the application
> + * may just ignore that and proceed with the rest of configuration
> + * procedure intending to simply try using the features it prefers.
> + *
> + * @param[in] port_id
> + *   Port (ethdev) identifier

See above comment.

> + *
> + * @param[inout] features
> + *   Feature selection buffer
> + *
> + * @return
> + *   - (-EBUSY) if the port can't handle this in its current state;
> + *   - (-ENOTSUP) if the method itself is not supported by the PMD;
> + *   - (-ENODEV) if *port_id* is invalid;
> + *   - (-EINVAL) if *features* is NULL;
> + *   - (-EIO) if the device is removed;
> + *   - (0) on success
> + */
> +__rte_experimental
> +int rte_eth_negotiate_rx_meta(uint16_t port_id, uint64_t *features);

# I would prefer to call it as rte_eth_rx_meta_negotiate as
action/verb can come the last API to
to have a proper namespace.

# IMO, We need to update the release notes for this behavior change.

# Is any of the in-tree examples or test application needs to update
to adaptor this behavior change

> +
>  #include <rte_ethdev_core.h>
>
>  /**
> diff --git a/lib/ethdev/version.map b/lib/ethdev/version.map
> index 3eece75b72..e390e5718c 100644
> --- a/lib/ethdev/version.map
> +++ b/lib/ethdev/version.map
> @@ -249,6 +249,9 @@ EXPERIMENTAL {
>         rte_mtr_meter_policy_delete;
>         rte_mtr_meter_policy_update;
>         rte_mtr_meter_policy_validate;
> +
> +       # added in 21.11
> +       rte_eth_negotiate_rx_meta;
>  };
>
>  INTERNAL {
> --
> 2.20.1
>