[dpdk-dev] [PATCH v7 1/4] ethdev: support device reset and recovery events

[Ferruh Yigit]

On 2/1/2022 1:09 PM, Kalesh Anakkur Purayil wrote:
> Thank you Ferruh for the review. Please see inline.
> 
> On Tue, Feb 1, 2022 at 5:41 PM Ferruh Yigit <ferruh.yigit at intel.com <mailto:ferruh.yigit at intel.com>> wrote:
> 
>     On 1/28/2022 12:48 PM, Kalesh A P wrote:
>      > From: Kalesh AP <kalesh-anakkur.purayil at broadcom.com <mailto:kalesh-anakkur.purayil at broadcom.com>>
>      >
>      > Adding support for the device reset and recovery events in the
>      > rte_eth_event framework. FW error and FW reset conditions would be
>      > managed internally by the PMD without needing application intervention.
>      > In such cases, PMD would need reset/recovery events to notify application
>      > that PMD is undergoing a reset.
>      >
>      > While most of the recovery process is transparent to the application since
>      > most of the driver ensures recovery from FW reset or FW error conditions,
>      > the application will have to reprogram any flows which were offloaded to
>      > the underlying hardware.
>      >
>      > Signed-off-by: Kalesh AP <kalesh-anakkur.purayil at broadcom.com <mailto:kalesh-anakkur.purayil at broadcom.com>>
>      > Signed-off-by: Somnath Kotur <somnath.kotur at broadcom.com <mailto:somnath.kotur at broadcom.com>>
>      > Reviewed-by: Ajit Khaparde <ajit.khaparde at broadcom.com <mailto:ajit.khaparde at broadcom.com>>
> 
>     More developer cc'ed.
> 
>      > ---
>      >   doc/guides/prog_guide/poll_mode_drv.rst | 24 ++++++++++++++++++++++++
>      >   lib/ethdev/rte_ethdev.h                 | 18 ++++++++++++++++++
>      >   2 files changed, 42 insertions(+)
>      >
>      > diff --git a/doc/guides/prog_guide/poll_mode_drv.rst b/doc/guides/prog_guide/poll_mode_drv.rst
>      > index 6831289..9ecc0e4 100644
>      > --- a/doc/guides/prog_guide/poll_mode_drv.rst
>      > +++ b/doc/guides/prog_guide/poll_mode_drv.rst
>      > @@ -623,3 +623,27 @@ by application.
>      >   The PMD itself should not call rte_eth_dev_reset(). The PMD can trigger
>      >   the application to handle reset event. It is duty of application to
>      >   handle all synchronization before it calls rte_eth_dev_reset().
>      > +
>      > +Error recovery support
>      > +~~~~~~~~~~~~~~~~~~~~~~
>      > +
>      > +When the PMD detects a FW reset or error condition, it may try to recover
>      > +from the error without needing the application intervention. In such cases,
>      > +PMD would need events to notify the application that it is undergoing
>      > +an error recovery.
>      > +
>      > +The PMD should trigger RTE_ETH_EVENT_ERR_RECOVERING event to notify the
>      > +application that PMD detected a FW reset or FW error condition. PMD may
>      > +try to recover from the error by itself. Data path may be quiesced and
>      > +control path operations may fail during the recovery period. The application
>      > +should stop polling till it receives RTE_ETH_EVENT_RECOVERED event from the PMD.
>      > +
> 
>     Between the time FW error occurred and the application receive the RECOVERING event,
>     datapath will continue to poll and application may call control APIs, so the event
>     really is not solving the issue and driver somehow should be sure this won't crash
>     the application, in that case not sure about the benefit of this event.
> 
> [Kalesh]: As soon as the driver detects a FW dead or reset condition, it sets the fastpath pointers to dummy functions. This will prevent the crash. All control path operations would fail with -EBUSY. This change is already there in bnxt PMD. This event is a notification to the application that the PMD is recovering from a FW error condition so that it can stop polling and issue control path operations.
> 

This was my point indeed. PMD can't rely on this event and should take actions (for both
datapath and control path) to prevent possible crash/error. If that is the case event
doesn't have that much benefit since PMD already has to protect itself.

> 
>      > +The PMD should trigger RTE_ETH_EVENT_RECOVERED event to notify the application
>      > +that the it has recovered from the error condition. PMD re-configures the port
>      > +to the state prior to the error condition. Control path and data path are up now.
>      > +Since the device has undergone a reset, flow rules offloaded prior to reset
>      > +may be lost and the application should recreate the rules again.
>      > +
> 
>     I think the most difficult part here is clarify what application should do
>     when this event received consistent for all devices, "flow rules may be lost"
>     looks very vague to me.
>     Unless it is not clear for application what to do when this event is received,
>     it is not that useful or it will be specific to some PMDs. And I can see it is
>     hard to clarify this but perhaps we can define a set of common behavior.
>     Not sure what others are thinking.
> 
> [Kalesh]: Sure, let's wait for others' opinions as well.
> 
> 
>      > +The PMD should trigger RTE_ETH_EVENT_INTR_RMV event to notify the application
>      > +that it has failed to recover from the error condition. The device may not be
>      > +usable anymore.
>      > diff --git a/lib/ethdev/rte_ethdev.h b/lib/ethdev/rte_ethdev.h
>      > index 147cc1c..a46819f 100644
>      > --- a/lib/ethdev/rte_ethdev.h
>      > +++ b/lib/ethdev/rte_ethdev.h
>      > @@ -3818,6 +3818,24 @@ enum rte_eth_event_type {
>      >       RTE_ETH_EVENT_DESTROY,  /**< port is released */
>      >       RTE_ETH_EVENT_IPSEC,    /**< IPsec offload related event */
>      >       RTE_ETH_EVENT_FLOW_AGED,/**< New aged-out flows is detected */
>      > +     RTE_ETH_EVENT_ERR_RECOVERING,
>      > +                     /**< port recovering from an error
>      > +                      *
>      > +                      * PMD detected a FW reset or error condition.
>      > +                      * PMD will try to recover from the error.
>      > +                      * Data path may be quiesced and Control path operations
>      > +                      * may fail at this time.
>      > +                      */
> 
>     Please put multi line comments before enum, Andrew did a set of cleanups for these.
> 
> [Kalesh]: Sure, will do.
> 
> 
>      > +     RTE_ETH_EVENT_RECOVERED,
>      > +                     /**< port recovered from an error
>      > +                      *
>      > +                      * PMD has recovered from the error condition.
>      > +                      * Control path and Data path are up now.
>      > +                      * PMD re-configures the port to the state prior to the error.
>      > +                      * Since the device has undergone a reset, flow rules
>      > +                      * offloaded prior to reset may be lost and
>      > +                      * the application should recreate the rules again.
>      > +                      */
>      >       RTE_ETH_EVENT_MAX       /**< max value of this enum */
>      >   };
>      >
> 
> 
> 
> -- 
> Regards,
> Kalesh A P