I'll explain what I did and why.

In a first diff I tried to imitate what kettenis@ did for amdgpu:

  https://cvsweb.openbsd.org/diff/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_drv.c?rev=1.68&prev=1.67

I have no amdgpu machine, I have no way to know if the problem solved by
that diff is the same.  The only I can say for sure is that calling
memset() or rasops_show_screen() *before* drm_client_dev_restore() (as
kettenis did for amdgpu) didn't work in my intel machine in the case I
described.  Besides, obviously, limiting them to S4 didn't solve the
issue for S3.

In a discussion I had with jsg@ in private I was asked to not use
memset() with S3 and S0, I think I get why.  Anyways, after trying this
and that, at some point I realized that memset wasn't necessary (at
least for my case), rasops_show_screen() was enough.

As far as my knowledge goes, I'm no expert, and the little I've
investigated, I'll explain the possible reason why these resets don't
work if you put them before drm_client_dev_restore().  I'll appreciate
to be corrected if I'm wrong in the following argument.

If you take a look to i915_driver.c, you'll see that besides
inteldrm_activate there are two more functions where rasops_show_screen
is called:

void
inteldrm_doswitch(void *v)
{
	struct drm_i915_private *dev_priv = v;
	struct drm_device *dev = &dev_priv->drm;
	struct rasops_info *ri = &dev_priv->ro;

	rasops_show_screen(ri, dev_priv->switchcookie, 0, NULL, NULL);
	drm_client_dev_restore(dev);

	if (dev_priv->switchcb)
		(*dev_priv->switchcb)(dev_priv->switchcbarg, 0, 0);
}

void
inteldrm_enter_ddb(void *v, void *cookie)
{
	struct drm_i915_private *dev_priv = v;
	struct drm_device *dev = &dev_priv->drm;
	struct rasops_info *ri = &dev_priv->ro;

	if (cookie == ri->ri_active)
		return;

	rasops_show_screen(ri, cookie, 0, NULL, NULL);
	drm_client_dev_restore(dev);
}

As you see, in both cases it's called *before* drm_client_dev_restore().
But, in those cases we're switching between active states, where the GPU
translation tables and power are already there.  Not the case in
DVACT_WAKEUP in inteldrm_activate(), the hardware is coming from an
uninitialized state.  If we call rasops_show_screen before
drm_client_dev_restore during resume, we risk writing to the framebuffer
before the DRM client has fully restablished the correct hardware plane
and GTT state.  This causes a desync between the rasops backing store
and the physical VRAM.


-- 
Walter