On Wed, 2018-04-04 at 22:31 +0300, Ville Syrjälä wrote:
On Wed, Apr 04, 2018 at 02:59:09PM -0400, Lyude Paul wrote:
On Wed, 2018-04-04 at 21:53 +0300, Ville Syrjälä wrote:
On Wed, Apr 04, 2018 at 02:37:41PM -0400, Lyude Paul wrote:
On Wed, 2018-04-04 at 18:34 +0300, Ville Syrjälä wrote:
On Mon, Apr 02, 2018 at 05:26:16PM -0400, Lyude Paul wrote:
While enabling/disabling DPMS before link training with MST hubs is perfectly valid; unfortunately disabling DPMS results in some devices disabling their AUX CH block as well. For SST this isn't as much of a problem, but for MST we need to be able to continue handling aux transactions even when none of the sinks are turned on since it's possible for us to have a single atomic commit which results in disabling each downstream sink, followed by subsequently re- enabling each sink.
If we don't do this, we'll end up stalling any pending ESI interrupts from the sink for up to 1ms. Unfortunately, dropping ESIs during this timespan makes it so that link fallback retraining for MST (which I will be submitting to the ML shortly) fails due to the channel EQ failure interrupts potentially getting dropped. Additionally, when performing a modeset that brings the hub status's link status from bad -> good having ESIs disabled for that long causes us to miss the hub's response to us trying to start link training as well.
Since any sink with MST is going to support DisplayPort 1.2 anyway, save us the hassle of trying to wait until the sink comes back up and just never shut the aux block down.
Changes since v2:
- Fix patch name, no functional changes
Signed-off-by: Lyude Paul lyude@redhat.com Cc: Laura Abbott labbott@redhat.com Cc: Dhinakaran Pandiyan dhinakaran.pandiyan@intel.com Cc: Ville Syrjälä ville.syrjala@linux.intel.com Cc: stable@vger.kernel.org Fixes: ad260ab32a4d9 ("drm/i915/dp: Write to SET_POWER dpcd to enable MST hub.")
drivers/gpu/drm/i915/intel_dp.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/i915/intel_dp.c b/drivers/gpu/drm/i915/intel_dp.c index 62f82c4298ac..0479c377981b 100644 --- a/drivers/gpu/drm/i915/intel_dp.c +++ b/drivers/gpu/drm/i915/intel_dp.c @@ -2589,11 +2589,13 @@ void intel_dp_sink_dpms(struct intel_dp *intel_dp, int mode) return; if (mode != DRM_MODE_DPMS_ON) {
unsigned char data = intel_dp->is_mst ?DP_SET_POWER_D3_AUX_ON : DP_SET_POWER_D3;This smells like a workaround for an actual bug somewhere. Why exactly is the slower wakeup or the AUX block a problem for MST but not for SST when the link training is exactly the same for SST and MST?
I actually thought about this but I still think this is the appropriate fix. So; the real reason for the wakeup not being a problem with SST is that for DPMS on with SST, we actually do a wait to make sure that the hub is ready before continuing. And yes: I'm fairly sure SST does actually have around the same wakeup time that MST does, but with the wait we do it doesn't reallhy make a difference. With MST, we could do this but there's a few reasons I don't think we should:
- We don't need to. D3_AUX_ON is a part of the 1.2 spec, so any hub
that has MST is going to be guaranteed to have this.
- Turning off the aux block means that there's a high chance we're
going to miss ESIs from sinks
And how exactly do we lose irqs? The hub/whatever throws the up req msgs away if we don't read them within some really short time?
That's my hypothesis at least. I'm betting that on the fact that when I was implementing MST retraining before we put the intel_dp_check_mst_status() (or whatever it's called) into the dig workqueue, getting the sink to go down and come back up was a lot more unreliable whenever I introduced anything that would block the esi handler for longer then a very brief period of time (say, 50-100ms?). I've seen some notes elsewhere too that seemed to imply for SST, things were pretty sensitive to irq latency (line 1050, intel_dp.c) so it wouldn't be terribly surprising if it's the same for MST. At the very least, now that we've got the ESI handler running in the dig worker things seem to have gotten a /lot/ more reliable now that we can basically go the whole modeset without blocking the ESI handler for very long.
Hmm. OK, so the spec seems to be saying that we have 100ms to read the UP_REQ/DOWN_REPLY msg after the IRQ_HPD. That's still a lot more than the 1ms max allowed wakeup time. Looks like there's a extended wakeup time request/grant mechanism now, but without the explicit grant (which we don't do) the 1ms still holds.
mm, that is true. There were definitely interrupts getting dropped though with this patch, although it was rather rare. The other thing too (and this was a lot easier to reproduce) was that when we do a modeset that lowers the link rate:
* check modeset * commit disables so we can reprogram the vcpi (puts sink into D3) * commit enables * actually send request to put sink into D0 * Go to start link training... * *TIMEOUT on response from hub* * continue modeset... * commit finishes * short period of time after the commit finishes, we receive the response from the hub where we tried to start link training
This being said: I think it may actually be a good idea for us to consider waiting for the hub anyway when we're turning it on like we do with SST, since we could be waking it up from a D3 state we didn't set ourselves anyway, but either way I think since we can at least benefit from not having to deal with the wakeup time at all in modesets like this (and it's not really problematic to other hubs), we probably should do that as well with or without a busy wait.
- It's faster to keep the aux block on anyway
- if (downstream_hpd_needs_d0(intel_dp)) return;
ret = drm_dp_dpcd_writeb(&intel_dp->aux,DP_SET_POWER,
DP_SET_POWER_D3);
ret = drm_dp_dpcd_writeb(&intel_dp->aux,DP_SET_POWER, data); } else { struct intel_lspcon *lspcon = dp_to_lspcon(intel_dp); -- 2.14.3
-- Cheers, Lyude Paul
-- Cheers, Lyude Paul