Re: [dpdk-dev] [PATCH] ethdev: claim device reset as async

[Thomas Monjalon <thomas@monjalon.net>]

Hi,

You forgot to Cc Andrew, co-maintainer of ethdev.

20/03/2019 05:54, Qi Zhang:
> Device reset should be implemented in an async way since it is
> possible to be invoked in interrupt thread and sometimes to reset a
> device need to wait for some dependency, for example, a VF expects for
> PF ready or a NIC function as part of a SOC wait for the whole system
> reset complete, and all these time-consuming tasks will block the
> interrupt thread.
> The patch rename rte_eth_dev_reset to rte_eth_dev_reset_async and
> rework the implementation. It will spawn a new thread which will call
> ops->dev_reset, and when finished it will raise the event
> RTE_ETH_EVENT_RESET_COMPLETE. The application should always wait for
> this event before it continues to configure and restart the device.
> 
> Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
> ---
> - * When this function is called, it first stops the port and then calls the
> - * PMD specific dev_uninit( ) and dev_init( ) to return the port to initial
> - * state, in which no Tx and Rx queues are setup, as if the port has been
> - * reset and not started. The port keeps the port id it had before the
> - * function call.
> - *
> - * After calling rte_eth_dev_reset( ), the application should use
> - * rte_eth_dev_configure( ), rte_eth_rx_queue_setup( ),
> - * rte_eth_tx_queue_setup( ), and rte_eth_dev_start( )
> - * to reconfigure the device as appropriate.
> - *
> - * Note: To avoid unexpected behavior, the application should stop calling
> - * Tx and Rx functions before calling rte_eth_dev_reset( ). For thread
> - * safety, all these controlling functions should be called from the same
> - * thread.
> + * @note
> + * Device reset may have the dependency, for example, a VF reset expects
> + * PF ready, or a NIC function as a part of a SOC need to wait for other
> + * parts of the system be ready, these are time-consuming tasks and will
> + * block current thread.
> + *
> + * As the name, rte_eth_dev_reset_async is an async API, it will spwan a
> + * new thread to call ops->dev_reset, once it is finished, it will raise
> + * the RTE_ETH_EVENT_RESET_COMPLETE event to notify application.  That makes
> + * things easy for an application that want to reset the device from the
> + * interrupt thread since typically a RTE_ETH_EVENT_INTR_RESET handler is
> + * invoked in interrupt thread. The typical implementation of ops->dev_reset
> + * will do some hardware reset operations through calling dev_uninit() and
> + * dev_init().
> + *
> + * Application should not assume device reset is finished after
> + * rte_eth_dev_reset_async return, it should always wait for a
> + * RTE_ETH_EVENT_RESET_COMPLETE event and check the reset result.
> + * If reset success, application should call rte_eth_dev_configure( ),
> + * rte_eth_rx_queue_setup( ), rte_eth_tx_queue_setup( ),
> + * and rte_eth_dev_start( ) to reconfigure the device as appropriate.
> + *
> + * @Note
> + * To avoid unexpected behavior, the application should stop calling
> + * Tx and Rx functions before calling rte_eth_dev_reset_async( ).
>   *
>   * @param port_id
>   *   The port identifier of the Ethernet device.
> @@ -1880,12 +1892,10 @@ void rte_eth_dev_close(uint16_t port_id);
>   *   - (0) if successful.
>   *   - (-EINVAL) if port identifier is invalid.
>   *   - (-ENOTSUP) if hardware doesn't support this function.
> - *   - (-EPERM) if not ran from the primary process.
> - *   - (-EIO) if re-initialisation failed or device is removed.
>   *   - (-ENOMEM) if the reset failed due to OOM.
> - *   - (-EAGAIN) if the reset temporarily failed and should be retried later.
> + *   - (<0) other errors from low level driver.
>   */
> -int rte_eth_dev_reset(uint16_t port_id);
> +int rte_eth_dev_reset_async(uint16_t port_id);

Sorry I didn't check whether this API is better or not,
but I know it cannot be accepted before proposing a deprecation notice.
Perhaps you may keep the old API and just add the new one.

Honestly, I never really agreed with the purpose of the reset API.
So making it async or not, I have no real opinion...
... but spawning a new thread at each function call, I feel it is bad.