[Intel-gfx] [RFC v2 5/8] drm/fb-helper: Don't restore if fbdev is not in use

[Daniel Vetter]

On Wed, Jan 03, 2018 at 11:21:07PM +0100, Noralf Trønnes wrote:
> Keep track of fbdev users and only restore fbdev in
> drm_fb_helper_restore_fbdev_mode_unlocked() when in use. This avoids
> fbdev being restored in drm_driver.last_close when nothing uses it.
> Additionally fbdev is turned off when the last user is closing.
> fbcon is a user in this context.
> 
> Signed-off-by: Noralf Trønnes <noralf at tronnes.org>

> ---
>  drivers/gpu/drm/drm_fb_helper.c | 15 +++++++++++++++
>  include/drm/drm_fb_helper.h     | 14 ++++++++++++++
>  2 files changed, 29 insertions(+)
> 
> diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
> index 2c6adf1d80c2..f9dcc7a5761f 100644
> --- a/drivers/gpu/drm/drm_fb_helper.c
> +++ b/drivers/gpu/drm/drm_fb_helper.c
> @@ -522,6 +522,9 @@ int drm_fb_helper_restore_fbdev_mode_unlocked(struct drm_fb_helper *fb_helper)
>  	if (READ_ONCE(fb_helper->deferred_setup))
>  		return 0;
>  
> +	if (!atomic_read(&fb_helper->open_count))
> +		return 0;
> +
>  	mutex_lock(&fb_helper->lock);
>  	ret = restore_fbdev_mode(fb_helper);
>  
> @@ -781,6 +784,7 @@ void drm_fb_helper_prepare(struct drm_device *dev, struct drm_fb_helper *helper,
>  	INIT_WORK(&helper->resume_work, drm_fb_helper_resume_worker);
>  	INIT_WORK(&helper->dirty_work, drm_fb_helper_dirty_work);
>  	helper->dirty_clip.x1 = helper->dirty_clip.y1 = ~0;
> +	atomic_set(&helper->open_count, 1);
>  	mutex_init(&helper->lock);
>  	helper->funcs = funcs;
>  	helper->dev = dev;
> @@ -1212,6 +1216,7 @@ EXPORT_SYMBOL(drm_fb_helper_cfb_imageblit);
>   * @info: fbdev registered by the helper
>   * @user: 1=userspace, 0=fbcon
>   *
> + * Increase fbdev use count.
>   * If &fb_ops is wrapped in a library, pin the driver module.
>   */
>  int drm_fb_helper_fb_open(struct fb_info *info, int user)
> @@ -1224,6 +1229,8 @@ int drm_fb_helper_fb_open(struct fb_info *info, int user)
>  			return -ENODEV;
>  	}
>  
> +	atomic_inc(&fb_helper->open_count);
> +
>  	return 0;
>  }
>  EXPORT_SYMBOL(drm_fb_helper_fb_open);
> @@ -1233,6 +1240,7 @@ EXPORT_SYMBOL(drm_fb_helper_fb_open);
>   * @info: fbdev registered by the helper
>   * @user: 1=userspace, 0=fbcon
>   *
> + * Decrease fbdev use count and turn off if there are no users left.
>   * If &fb_ops is wrapped in a library, unpin the driver module.
>   */
>  int drm_fb_helper_fb_release(struct fb_info *info, int user)
> @@ -1240,6 +1248,10 @@ int drm_fb_helper_fb_release(struct fb_info *info, int user)
>  	struct drm_fb_helper *fb_helper = info->par;
>  	struct drm_device *dev = fb_helper->dev;
>  
> +	if (atomic_dec_and_test(&fb_helper->open_count) &&
> +	    !drm_dev_is_unplugged(fb_helper->dev))
> +		drm_fb_helper_blank(FB_BLANK_POWERDOWN, info);
> +
>  	if (info->fbops->owner != dev->driver->fops->owner)
>  		module_put(dev->driver->fops->owner);
>  
> @@ -1936,6 +1948,9 @@ static int drm_fb_helper_single_fb_probe(struct drm_fb_helper *fb_helper,
>  	if (ret < 0)
>  		return ret;
>  
> +	if (fb_helper->fbdev->fbops->fb_open == drm_fb_helper_fb_open)
> +		atomic_set(&fb_helper->open_count, 0);
> +
>  	strcpy(fb_helper->fb->comm, "[fbcon]");
>  	return 0;
>  }
> diff --git a/include/drm/drm_fb_helper.h b/include/drm/drm_fb_helper.h
> index a593f01ff69e..16d8773b60e3 100644
> --- a/include/drm/drm_fb_helper.h
> +++ b/include/drm/drm_fb_helper.h
> @@ -232,6 +232,20 @@ struct drm_fb_helper {
>  	 * See also: @deferred_setup
>  	 */
>  	int preferred_bpp;
> +
> +	/**
> +	 * @open_count:
> +	 *
> +	 * Keeps track of fbdev use to know when to not restore fbdev.
> +	 *
> +	 * Drivers that use drm_fb_helper_fb_open() as their \.fb_open
> +	 * callback will get an initial value of 0 and get restore based on
> +	 * actual use. Others will get an initial value of 1 which means that
> +	 * fbdev will always be restored. Drivers that call
> +	 * drm_fb_helper_fb_open() in their \.fb_open, thus needs to set the
> +	 * initial value to 0 themselves.
> +	 */
> +	atomic_t open_count;

Do we really need an atomic_t here? Especially the "disable stuff on last
fb_close" would still be race without more locking, and from a quick look
at the fbdev core code we're protected with fb_info->lock. I think an
unsigned is perfectly fine here. And if we really need atomics, then
refcount_t (for the safety checks that provides).

With that changed:

Reviewed-by: Daniel Vetter <daniel.vetter at ffwll.ch>

But before we merge this we definitely need to roll out fb_open/close to
all drivers. But I guess that's why the RFC in your series.

On that: There's some intersting runtime pm stuff for
nouveau/amd/radeon. I guess we'll need to keep that. Plus udl has some
defio fun :-/
-Daniel

>  };
>  
>  /**
> -- 
> 2.14.2
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch