Re: [dpdk-dev] [PATCH v6 4/6] ethdev: adjust APIs removal error report

[Thomas Monjalon <thomas@monjalon.net>]

19/01/2018 19:13, Ferruh Yigit:
> On 1/19/2018 5:54 PM, Thomas Monjalon wrote:
> > 19/01/2018 17:19, Ferruh Yigit:
> >> On 1/18/2018 6:10 PM, Matan Azrad wrote:
> >>> From: Ferruh Yigit, Thursday, January 18, 2018 7:31 PM
> >>>> This patch updates *all* ethdev public APIs to add if device is removed
> >>>> check?
> >>>
> >>> Yes.
> >>>
> >>>> And each check goes to ethdev is_removed() dev_ops to ask if dev is
> >>>> removed.
> >>> Probably, if the REMOVED state setted in will not call device is_remove.
> >>>
> >>>> These must be better way of doing this, am I missing something.
> >>>
> >>> Suggest.
> >>
> >> With a silly analogy, this is like a blind person asking each time if he is dead
> >> before talking to other person.
> >>
> >> At first glance I can think of a kind of watchdog timer can be implemented in
> >> ethdev layer. It provides periodic checks and if device is dead it calls the
> >> registered user callback function.
> >>
> >> This method presented as synchronous method but not triggered from side where
> >> event happens, I mean not triggered from PMD but from application.
> >> So does application doing polling continuously if device is dead?
> >> Or if application is relying this patch to add a check in each API, what happens
> >> if device removed during data processing, will app rely on asynchronous method?
> > 
> > We cannot put a mutex on hardware removal :)
> > So we have to live with errors du to removal.
> > If we are trying to configure a removed device,
> > the error will be not related to the root cause.
> > This patch is just trying to improve the situation by returning
> > an appropriate error code if removal can be detected.
> > Note: the check is run only if there is an error
> > and if the removal is not already detected.
> > 
> > If I understand well, you prefer relying only on asynchronous
> > hotplug events? Even if they come really late?
> 
> I think asynchronous hotplug events are better approach, but if they are late I
> understand you need a solution.
> 
> I assume issue is when device is removed, application can make API calls until
> it detects the removal.
> 
> For that case what do you think instead of ethdev abstraction layer doing a
> translation, define a specific error type and return it from PMDs to indicate
> that error is related to the missing device and application will know device is
> no more there? And consistently use that error type in all PMDs.

Yes it is also a good solution.
I think Matan proposed to integrate it in ethdev to avoid duplicating
the same mechanism in every PMDs.

> Or what do you think above suggested flexible watchdog timer implementation in
> ethdev, so applications may be informed about missing device faster.

I think the watchdog would compete with hotplug events.
So I prefer not integrating one more asynchronous mechanism
for the same purpose.
If we want polling, it can be an option to enable in EAL hotplug.