[Intel-gfx] [PATCH 1/2] drm/dp_helper: retry on -ETIMEDOUT in drm_dp_dpcd_access()

Lyude cpaul at redhat.com
Thu Mar 17 15:40:44 UTC 2016

Some sinks need some time during the process of resuming the system from
sleep before they're ready to handle transactions. While it would be
nice if they responded with NACKs in these scenarios, this isn't always
the case as a few sinks will just timeout on all of the transactions
they receive.

This patch was originally intended to be a workaround for a very
mysterious bug on the T560, where any monitors connected to the dock
would fail to turn back on after resume. When resuming the laptop, it
appears that there's a short period of time where we're unable to
complete any aux transactions, as they all immediately timeout. The only
machine I'm able to reproduce this on is the T560 as other production
Skylake models seem to be fine. The period during which AUX transactions
fail appears to be around 22ms long. AFAIK, the dock for the T560 never
actually turns off, the only difference is that it's in SST mode at the
start of the resume process, so it's unclear as to why it would need so
much time to come back up.

There's been a discussion on this issue going on for a while on the
intel-gfx mailing list about this that has, in addition to including
developers from Intel, also had the correspondence of one of the
hardware engineers for Intel:


We've already looked into a couple of possible explanations for the

- Calling intel_dp_mst_resume() before right fix.
  intel_runtime_pm_enable_interrupts(). This was the first fix I tried,
  and while it worked it definitely wasn't the right fix. This worked
  because DP aux transactions don't actually require interrupts to work:

	static uint32_t
	intel_dp_aux_wait_done(struct intel_dp *intel_dp, bool has_aux_irq)
		struct intel_digital_port *intel_dig_port = dp_to_dig_port(intel_dp);
		struct drm_device *dev = intel_dig_port->base.base.dev;
		struct drm_i915_private *dev_priv = dev->dev_private;
		i915_reg_t ch_ctl = intel_dp->aux_ch_ctl_reg;
		uint32_t status;
		bool done;

	#define C (((status = I915_READ_NOTRACE(ch_ctl)) & DP_AUX_CH_CTL_SEND_BUSY) == 0)
		if (has_aux_irq)
			done = wait_event_timeout(dev_priv->gmbus_wait_queue, C,
			done = wait_for_atomic(C, 10) == 0;
		if (!done)
			DRM_ERROR("dp aux hw did not signal timeout (has irq: %i)!\n",
	#undef C

		return status;

  When there's no interrupts enabled, we end up timing out on the
  wait_event_timeout() call, which causes us to check the DP status
  register once to see if the transaction was successful or not. Since
  this adds a 10ms delay to each aux transaction, it ends up adding a
  long enough delay to the resume process for aux transactions to become
  functional again. This gave us the illusion that enabling interrupts
  had something to do with making things work again, and put me on the
  wrong track for a while.

- Interrupts occurring when we try to perform the aux transactions
  required to put the dock back into MST mode. This isn't the problem,
  as the only interrupts I've observed that come during this timeout
  period are from the snd_hda_intel driver, and disabling that driver
  doesn't appear to change the behavior at all.

- Skylake's PSR block causing issues by performing aux transactions
  while we try to bring the dock out of MST mode. Disabling PSR through
  i915's command line options doesn't seem to change the behavior
  either, nor does preventing the DMC firmware from being loaded.

Since this investigation went on for about 2 weeks, we decided it would
be better for the time being to just workaround this issue until we find
a better fix. This patch adds some behavior we want in
drm_dp_dpcd_access() anyway, since DP aux transactions aren't exactly
robust and this will probably fix quite a few other issues with DP MST
hardware not responding in time. Plus, this is something we already do
in the i915 driver with intel_dp_dpcd_read_wake(), except that that
function is somewhat of a hack and DRM helpers can't make use of it.

			    Changes since v1

- Patch has been reworked to take the retry logic out of
  intel_dp_mst_resume() and into drm_dp_dpcd_access(), based off a
  suggestion from Daniel Vetter

- Commit message is much longer and gives a better description of the
  issue this was originally intended to workaround.

Signed-off-by: Lyude <cpaul at redhat.com>
 drivers/gpu/drm/drm_dp_helper.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/drm_dp_helper.c b/drivers/gpu/drm/drm_dp_helper.c
index 9535c5b..38dd969 100644
--- a/drivers/gpu/drm/drm_dp_helper.c
+++ b/drivers/gpu/drm/drm_dp_helper.c
@@ -198,7 +198,7 @@ static int drm_dp_dpcd_access(struct drm_dp_aux *aux, u8 request,
 		err = aux->transfer(aux, &msg);
 		if (err < 0) {
-			if (err == -EBUSY)
+			if (err == -EBUSY || err == -ETIMEDOUT)
 			return err;

More information about the Intel-gfx mailing list