[dpdk-dev] [PATCH v4 1/2] eventdev: add device stop flush callback

[Eads, Gage]

> -----Original Message-----
> From: Van Haaren, Harry
> Sent: Friday, March 23, 2018 11:57 AM
> To: Eads, Gage <gage.eads at intel.com>; dev at dpdk.org
> Cc: jerin.jacob at caviumnetworks.com; hemant.agrawal at nxp.com; Richardson,
> Bruce <bruce.richardson at intel.com>; santosh.shukla at caviumnetworks.com;
> nipun.gupta at nxp.com
> Subject: RE: [PATCH v4 1/2] eventdev: add device stop flush callback
> 
> > From: Eads, Gage
> > Sent: Tuesday, March 20, 2018 2:13 PM
> > To: dev at dpdk.org
> > Cc: jerin.jacob at caviumnetworks.com; Van Haaren, Harry
> > <harry.van.haaren at intel.com>; hemant.agrawal at nxp.com; Richardson,
> > Bruce <bruce.richardson at intel.com>; santosh.shukla at caviumnetworks.com;
> > nipun.gupta at nxp.com
> > Subject: [PATCH v4 1/2] eventdev: add device stop flush callback
> >
> > When an event device is stopped, it drains all event queues. These
> > events may contain pointers, so to prevent memory leaks eventdev now
> > supports a user-provided flush callback that is called during the queue drain
> process.
> > This callback is stored in process memory, so the callback must be
> > registered by any process that may call rte_event_dev_stop().
> >
> > This commit also clarifies the behavior of rte_event_dev_stop().
> >
> > This follows this mailing list discussion:
> > http://dpdk.org/ml/archives/dev/2018-January/087484.html
> >
> > Signed-off-by: Gage Eads <gage.eads at intel.com>
> 
> <snip most of the code - looks good!>
> 
> >  /**
> > - * Stop an event device. The device can be restarted with a call to
> > - * rte_event_dev_start()
> > + * Stop an event device.
> > + *
> > + * This function causes all queued events to be drained. While
> > + draining
> > events
> > + * out of the device, this function calls the user-provided flush
> > + callback
> > + * (if one was registered) once per event.
> > + *
> > + * This function does not drain events from event ports; the
> > + application is
> > + * responsible for flushing events from all ports before stopping the
> > device.
> 
> 
> Question about how an application is expected to correctly cleanup all the
> events here. Note in particular the last part: "application is responsible for
> flushing events from all ports **BEFORE** stopping the device".
> 
> Given the event device is still running, how can the application be sure it has
> flushed all the events (from the dequeue side in particular)?
> 

Appreciate the feedback -- good points all around.

I was expecting that the application would unlink queues from the ports, and then dequeue until each port has no events. However, there are PMDs for which runtime port link/unlink is not supported, so I see that this is not a viable approach. Plus, this adds the application burden that you describe below.

> 
> In order to drain all events from the ports, I was expecting the following:
> 
> // stop scheduling new events to worker cores
> rte_event_dev_stop()
> ---> callback gets called for each event
> 
> // to dequeue events from each port, and app cleans them up?
> FOR_EACH_PORT( rte_event_dev_dequeue(..., port_id, ...) )
> 
> 
> I'd like to avoid the dequeue-each-port() approach in application, as it adds extra
> burden to clean up correctly...

Agreed, but for a different reason: that approach means we'd have to change the documented eventdev behavior. rte_eventdev.h states that the "schedule, enqueue and dequeue functions should not be invoked when the device is stopped," and this patch reiterates that in the rte_event_dev_stop() documentation ("Threads that continue to enqueue/dequeue while the device is stopped, or being stopped, will result in undefined behavior"). Since a PMD's stop cleanup code could just be repeated calls to a PMD's dequeue code, allowing applications to dequeue simultaneously could be troublesome.

> 
> What if we say that dequeue() returns zero after stop() (leaving events possibly
> in the port-dequeue side SW buffers), and these events which were about to be
> dequeued by the worker core are also passed to the dev_stop_flush callback?

I'd prefer to have dequeue-while-stopped be unsupported, so we don't need an additional check or synchronization in the datapath, but passing the events in a port to the callback should work (for the sw PMD, at least). How does that sound?

Thanks,
Gage