[Intel-gfx] [PATCH v3] drm/i915: Pin the ifbdev for the info->system_base GGTT mmapping

[Jesse Barnes]

On 11/20/2015 06:34 AM, Chris Wilson wrote:
> A long time ago (before 3.14) we relied on a permanent pinning of the
> ifbdev to lock the fb in place inside the GGTT. However, the
> introduction of stealing the BIOS framebuffer and reusing its address in
> the GGTT for the fbdev has muddied waters and we use an inherited fb.
> However, the inherited fb is only pinned whilst it is active and we no
> longer have an explicit pin for the info->system_base mmapping used by
> the fbdev. The result is that after some aperture pressure the fbdev may
> be evicted, but we continue to write the fbcon into the same GGTT
> address - overwriting anything else that may be put into that offset.
> The effect is most pronounced across suspend/resume as
> intel_fbdev_set_suspend() does a full clear over the whole scanout.
> 
> v2: Only unpin the intel_fb is we allocate it. If we inherit the fb from
> the BIOS, we do not own the pinned vma (except for the reference we add
> in this patch for our access via info->screen_base).
> 
> v3: Finish balancing the vma pinning for the normal !preallocated case.
> 
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: "Goel, Akash" <akash.goel at intel.com>
> Cc: Daniel Vetter <daniel.vetter at ffwll.ch>
> Cc: Jesse Barnes <jbarnes at virtuousgeek.org>
> Cc: stable at vger.kernel.org
> ---
>  drivers/gpu/drm/i915/intel_fbdev.c | 23 +++++++++++++++++++++++
>  1 file changed, 23 insertions(+)
> 
> diff --git a/drivers/gpu/drm/i915/intel_fbdev.c b/drivers/gpu/drm/i915/intel_fbdev.c
> index 7ccde58f8c98..7a415fe31299 100644
> --- a/drivers/gpu/drm/i915/intel_fbdev.c
> +++ b/drivers/gpu/drm/i915/intel_fbdev.c
> @@ -225,6 +225,16 @@ static int intelfb_create(struct drm_fb_helper *helper,
>  
>  	mutex_lock(&dev->struct_mutex);
>  
> +	/* The fb constructor will have already pinned us (or inherited a
> +	 * GGTT region from the BIOS) suitable for a scanout, so
> +	 * this should just be a no-op and increment the pin count for the
> +	 * fbdev mmapping. It does have a useful side-effect of validating
> +	 * the pin for fbdev's use via a GGTT mmapping.
> +	 */
> +	ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE);
> +	if (ret)
> +		goto out_unlock;
> +
>  	info = drm_fb_helper_alloc_fbi(helper);
>  	if (IS_ERR(info)) {
>  		DRM_ERROR("Failed to allocate fb_info\n");
> @@ -279,6 +289,12 @@ static int intelfb_create(struct drm_fb_helper *helper,
>  		      fb->width, fb->height,
>  		      i915_gem_obj_ggtt_offset(obj), obj);
>  
> +	/* We pin the vma for our access through info->screen_base, so
> +	 * we can drop the pin we took if we created the intel_fb.
> +	 */
> +	if (!prealloc)
> +		i915_gem_object_ggtt_unpin(obj);
> +
>  	mutex_unlock(&dev->struct_mutex);
>  	vga_switcheroo_client_fb_set(dev->pdev, info);
>  	return 0;
> @@ -286,7 +302,12 @@ static int intelfb_create(struct drm_fb_helper *helper,
>  out_destroy_fbi:
>  	drm_fb_helper_release_fbi(helper);
>  out_unpin:
> +	/* Once for info->screen_base mmaping... */
>  	i915_gem_object_ggtt_unpin(obj);
> +out_unlock:
> +	if (!prealloc)
> +		/* ...and once for the intel_fb */
> +		i915_gem_object_ggtt_unpin(obj);
>  	mutex_unlock(&dev->struct_mutex);
>  	return ret;
>  }
> @@ -524,6 +545,8 @@ static const struct drm_fb_helper_funcs intel_fb_helper_funcs = {
>  static void intel_fbdev_destroy(struct drm_device *dev,
>  				struct intel_fbdev *ifbdev)
>  {
> +	/* Release the pinning for the info->screen_base mmaping. */
> +	i915_gem_object_ggtt_unpin(ifbdev->fb->obj);
>  
>  	drm_fb_helper_unregister_fbi(&ifbdev->helper);
>  	drm_fb_helper_release_fbi(&ifbdev->helper);
> 

Now you're making me look at the pin/unpin handling...  Could probably
make the prealloc vs non-prealloc cases a bit clearer, but it looks
correct.  In the prealloc case we need the additional pin, since
create_stolen_for_preallocated just pins the pages and doesn't up the
pin count, right?  But in the non-prealloc case we'll have done a
regular fb alloc, which does a pin & fence, so we can drop the extra pin
count.  And I think the page unpin is already taken care of?  ISTR bugs
there when we first landed the initial plane allocation stuff.

Reviewed-by: Jesse Barnes <jbarnes at virtuousgeek.org>