[PATCH] ethdev: recommend against using locks in event callbacks

[Ferruh Yigit]

On 2/1/2024 10:08 AM, Kevin Traynor wrote:
> On 01/02/2024 08:43, David Marchand wrote:
>> As described in a recent bugzilla opened against the net/iavf driver,
>> a driver may call a event callback from other calls of the ethdev API.
>>
>> Nothing guarantees in the ethdev API against such behavior.
>>
>> Add a notice against using locks in those callbacks.
>>
>> Bugzilla ID: 1337
>>
>> Signed-off-by: David Marchand <david.marchand at redhat.com>
>> ---
>>  lib/ethdev/rte_ethdev.h | 14 +++++++++++++-
>>  1 file changed, 13 insertions(+), 1 deletion(-)
>>
>> diff --git a/lib/ethdev/rte_ethdev.h b/lib/ethdev/rte_ethdev.h
>> index 21e3a21903..5c6b104fb4 100644
>> --- a/lib/ethdev/rte_ethdev.h
>> +++ b/lib/ethdev/rte_ethdev.h
>> @@ -4090,7 +4090,19 @@ enum rte_eth_event_type {
>>  	RTE_ETH_EVENT_MAX       /**< max value of this enum */
>>  };
>>  
>> -/** User application callback to be registered for interrupts. */
>> +/**
>> + * User application callback to be registered for interrupts.
>> + *
>> + * Note: there is no guarantee in the DPDK drivers that a callback won't be
>> + *       called in the middle of other parts of the ethdev API. For example,
>> + *       imagine that thread A calls rte_eth_dev_start() and as part of this
>> + *       call, a RTE_ETH_EVENT_INTR_RESET event gets generated and the
>> + *       associated callback is ran on thread A. In that example, if the
>> + *       application protects its internal data using locks before calling
>> + *       rte_eth_dev_start(), and the callback takes a same lock, a deadlock
>> + *       occurs. Because of this, it is highly recommended NOT to take locks in
>> + *       those callbacks.
>> + */
> 
> That is a good practical recommendation for an application developer.
> 
> I wonder if it should taken further so that the API formally states the
> callback MUST be non-blocking?
> 

Application still can manage the locks in a safe way, but needs to be
aware of above condition and possible deadlock.

I think above note is sufficient instead of forbidding locks in
callbacks completely.