On 06/02/2024 20:33, Ferruh Yigit wrote:
> 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.march...@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.
>
Just to explain a bit more, if you look at the original issue in the
Bugzilla [0], I think there was an assumption that
rte_eth_dev_configure() would not block or deadlock with the
eal-intr-thread. So then it was assumed that waiting for the lock in the
callback was ok, because rte_eth_dev_configure() would return and
callback would obtain the lock.
So i'm showing that in this example the lack of a guarantee or clarity
or bad assumption about the behavior of rte_eth_dev_configure() made it
difficult for an app developer to know if their locks were safe or not.
That's why I was thinking about something more formal.
> I think above note is sufficient instead of forbidding locks in
> callbacks completely.
>
In the end the difference between "highly recommended NOT to" and "must
not" is not much and either way is probably enough to scare someone
enough to avoid them.
[0] https://bugs.dpdk.org/show_bug.cgi?id=1337#c0
