[igt-dev] [PATCH i-g-t] tests/chamelium: Add test for hotplug workaround

Fri Mar 15 02:46:21 UTC 2019

On Fri, Mar 15, 2019 at 03:27:08AM +0200, Imre Deak wrote:
> On Fri, Mar 15, 2019 at 02:00:41AM +0200, Souza, Jose wrote:
> [...]
> > > > > > +
> > > > > > +	/*
> > > > > > +	 * Give some time to kernel try to process hotplug but
> > > > > > it
> > > > > > should fail
> > > > > > +	 */
> > > > > > +	igt_hotplug_detected(mon, FAST_HOTPLUG_SEC_TIMEOUT);
> > > > > > +	status = connector_status_get(data, port);
> > > > > > +	igt_assert(status == DRM_MODE_DISCONNECTED);
> > > > > 
> > > > > Hm, how does a second disconnect event get signaled here? After
> > > > > the
> > > > > previous hotplug event above where the state was disconnected
> > > > > already
> > > > > there
> > > > > shouldn't have been any change to the state, hence there
> > > > > shouldn't be
> > > > > any event sent by the driver.
> > > > 
> > > > It is not signaled, that I why there is not igt_assert() on this
> > > > igt_hotplug_detected() but call it will poll/sleep for the time we
> > > > want.
> > > 
> > > But then I don't see how it will work. The sequence is:
> > > 
> > > <connector is in disconnected state, corresponding event delivered>
> > > 1. disable DDC
> > > 2. generate a plug event
> > > 3. wait for the plug event delivery with 1 sec timeout
> > > 4. re-enable DDC
> > > 5. wait for the plug event delivery (that should be triggered by the
> > > new
> > >    retry logic in the driver)
> > > 
> > > Since after 2. DDC is disabled the driver hotplug handler will
> > > conclude
> > > that the connector is still disconnected and hence doesn't generate
> > > any
> > > hotplug event. B/c of this 3. will time out after 1 sec.
> > > 
> > > So in 4. we'll re-enable DDC only after 1 sec after the plug event
> > > (interrupt) generated in 2. Since the retry in the driver happens
> > > after
> > > 1 sec from the plug interrupt as well the retry processing could
> > > easily
> > > race with the DDC re-enabling in 4. and thus the detection could
> > > fail.
> > 
> > You are not taking in the account the time that kernel will take to
> > process that, I measured just the time spend in the hotplug() hook on
> > my ICL.
> > 
> > [185950.212037] [drm:intel_encoder_hotplug [i915]] [CONNECTOR:196:DP-1] 
> > status updated from connected to disconnected
> > [185950.212109] [drm:i915_hotplug_work_func [i915]]     hotplug()
> > took=673123725 nsec(673 msec) ret=1
> > 
> > [185950.956536] [drm:intel_encoder_hotplug [i915]] [CONNECTOR:196:DP-1] 
> > status updated from disconnected to connected
> > [185950.957068] [drm:i915_hotplug_work_func [i915]]     hotplug()
> > took=60222955 nsec(60 msec) ret=1
> > 
> > 
> > [185953.480877] [drm:drm_dp_dpcd_access] Too many retries, giving up.
> > First error: -110
> > [185953.480969] [drm:intel_encoder_hotplug [i915]] [CONNECTOR:196:DP-1] 
> > status updated from connected to disconnected
> > [185953.481038] [drm:i915_hotplug_work_func [i915]]     hotplug()
> > took=672309486 nsec(672 msec) ret=1
> > 	
> > More than half of a second retrying until it gives up and change/keep
> > the status to disconnected.
> 
> 670msec for detection to fail sounds strange, where is that coming from?
> AUX reads should time out much earlier even with all the retries.

Ah, could be the 4msec (max) AUX HW timeout on ICL. That with retries
adds up to 5*32*4ms = 640msec about what you have measured.

But up to SKL the AUX HW timeout is only 1.6msec giving a 256msec delay,
so the hotplug retry could happen as soon as ~1.3sec after the plug event.

So I guess a 1 sec delay/poll at step 3. is ok along with some
explanation about the duration like the above calculation. I think it
could still be possible for this 1 sec delay/poll to last longer than
1.3sec (due to scheduling) in which case the detection would fail on
some platforms. So I'd still add the time measurement between step 2.
and 4. as described below and rerun the test if it was > 1.2sec and
detection failed.

> > 
> > So in my rough estimation:
> > t 0s = IGT disables DDC and do the hotplug(1 and 2 from your sequence)
> > t 0.7s = kernel gives up and keep connector as disconnected
> > t 1s = IGT read connector as disconnected and enables DCC(3 and 4 from
> > your sequence)
> > t 1.7 = kernel try to probe again
> > t 1.8 = kernel probe and mark connector as connected
> > t 2s = IGT read connector as connected(5 from your sequence)
> > 
> > Maybe to avoid test failures the second timeout should be bigger
> > 
> > > 
> > > Since I don't see a good way to ensure that we re-enable DDC after
> > > the
> > > first detection cycle ran (but not too late missing the retry cycle)
> > > I
> > > would rather suggest a simple wait at 3., let's say 500msec. With
> > > that
> > > things should always work. We may not always exercise the driver's
> > > retry
> > > path if there was a long scheduling delay, but that's unlikely.
> > > 
> > > However a scheduling delay after 2. and before 4. could cause a
> > > detection failure. To avoid that I'd also check the elapsed time
> > > starting from right before 2. until right after 4. and run the
> > > sequence
> > > again if the elapsed time was too close to 1sec (and hence detection
> > > possibly failed because of the race described above).
> > > 
> > > > > > +
> > > > > > +	/*
> > > > > > +	 * Enable the DDC line and the kernel workaround should
> > > > > > reprobe
> > > > > > and
> > > > > > +	 * report as connected
> > > > > > +	 */
> > > > > > +	chamelium_port_set_ddc_state(data->chamelium, port,
> > > > > > true);
> > > > > > +	igt_assert(chamelium_port_get_ddc_state(data-
> > > > > > >chamelium,
> > > > > > port));
> > > > > > +	igt_assert(igt_hotplug_detected(mon,
> > > > > > FAST_HOTPLUG_SEC_TIMEOUT));
> > > > > > +	status = connector_status_get(data, port);
> > > > > > +	igt_assert(status == DRM_MODE_CONNECTED);
> > > > > > +}
> > > > > > +
> > > > > >  static void
> > > > > >  test_edid_read(data_t *data, struct chamelium_port *port,
> > > > > >  	       int edid_id, const unsigned char *edid)
> > > > > > @@ -1308,6 +1375,9 @@ igt_main
> > > > > >  
> > > > > >  		connector_subtest("dp-frame-dump", DisplayPort)
> > > > > >  			test_display_frame_dump(&data, port);
> > > > > > +
> > > > > > +		connector_subtest("dp-late-aux-wa",
> > > > > > DisplayPort)
> > > > > > +			test_late_aux_wa(&data, port);
> > > 
> > > This also needs the retry logic to be added for DP ports on old
> > > platforms (intel_dp_hotplug() in the driver).
> > > 
> > > > > >  	}
> > > > > >  
> > > > > >  	igt_subtest_group {
> > > > > > -- 
> > > > > > 2.21.0
> > > > > > 
> > > 
> > > 
> 
> 