On ti, 2015-02-24 at 15:58 +0100, Bjørn Mork wrote:

This fixes a bug where hibernation completes, but the system
fails to power off after the image has been saved.

Bisection lead to commit da2bc1b9db33 ("drm/i915: add poweroff_late
handler") which added a .poweroff_late hook pointing to the same
function as the .freeze_late hook, without any justification or
explanation why this would be correct, except "for consistency".

The assumption that this "makes the power off handling identical to
the S3 suspend and S4 freeze handling" is completely bogus. As clearly
documented in Documentation/power/devices.txt, the poweroff* hooks
are part of the hibernation API along with the freeze* hooks.  The
phases involved in hibernation are:

   prepare, freeze, freeze_late, freeze_noirq, thaw_noirq, thaw_early,
   thaw, complete, prepare, poweroff, poweroff_late, poweroff_noirq

With the above sequence in mind, it should be fairly obvious that you
cannot save registers and disable the device in both the freeze_late
and poweroff_late phases.  Or as Documentation/power/devices.txt
explains it:

 The poweroff, poweroff_late and poweroff_noirq callbacks should do essentially
 the same things as the suspend, suspend_late and suspend_noirq callbacks,
 respectively.  The only notable difference is that they need not store the
 device register values, because the registers should already have been stored
 during the freeze, freeze_late or freeze_noirq phases.

The "consistency" between suspend and hibernate was already in
place, with freeze_late pointing to the same function as suspend_late.
There is no need for any poweroff_late hook in this driver.

The poweroff handlers undo the actions of the thaw handlers. As the
original commit stated saving the registers is not needed there, but
it's also not a big overhead and there should be no problem doing it. We
are planning to optimize the hibernation sequence by for example not
shutting down the display between freeze and thaw, and also getting rid
of unnecessary steps at the power off phase. But before that we want to
actually unify things rather than having special cases, as maintaining
the special code paths caused already quite a lot of problems for us so
far.

Reverting the commit may hide some other issue, so before doing that
could you try the following patch:
http://lists.freedesktop.org/archives/intel-gfx/2015-February/060529.html

If with that you still get the hang could you try on top of that the
patch below, first having only pci_set_power_state uncommented, then
both pci_set_power_state and pci_disable_device uncommented?

Thanks,
Imre

diff --git a/drivers/gpu/drm/i915/i915_drv.c

b/drivers/gpu/drm/i915/i915_drv.c
index 4badb23..dc91d4b 100644

--- a/drivers/gpu/drm/i915/i915_drv.c
+++ b/drivers/gpu/drm/i915/i915_drv.c

@@ -968,6 +968,23 @@ static int i915_pm_suspend_late(struct device *dev)
 	return i915_drm_suspend_late(drm_dev);
 }
 
+static int i915_pm_poweroff_late(struct device *dev)
+{
+	struct drm_device *drm_dev = dev_to_i915(dev)->dev;
+	struct drm_i915_private *dev_priv = drm_dev->dev_private;
+	int ret;
+
+	ret = intel_suspend_complete(dev_priv);
+
+	if (ret)
+		return ret;
+
+	pci_disable_device(drm_dev->pdev);
+//	pci_set_power_state(drm_dev->pdev, PCI_D3hot);
+
+	return 0;
+}
+
 static int i915_pm_resume_early(struct device *dev)
 {
 	struct drm_device *drm_dev = dev_to_i915(dev)->dev;

@@ -1535,7 +1552,7 @@ static const struct dev_pm_ops i915_pm_ops = {
 	.thaw_early = i915_pm_resume_early,
 	.thaw = i915_pm_resume,
 	.poweroff = i915_pm_suspend,
-	.poweroff_late = i915_pm_suspend_late,
+	.poweroff_late = i915_pm_poweroff_late,
 	.restore_early = i915_pm_resume_early,
 	.restore = i915_pm_resume,