[Intel-gfx] [PATCH v2 02/12] drm/fb-helper: Avoid race with DRM userspace

Daniel Vetter daniel at ffwll.ch
Wed Apr 17 13:24:00 UTC 2019


On Tue, Apr 16, 2019 at 08:46:24PM +0200, Noralf Trønnes wrote:
> 
> 
> Den 16.04.2019 09.59, skrev Daniel Vetter:
> > On Sun, Apr 07, 2019 at 06:52:33PM +0200, Noralf Trønnes wrote:
> >> drm_fb_helper_is_bound() is used to check if DRM userspace is in control.
> >> This is done by looking at the fb on the primary plane. By the time
> >> fb-helper gets around to committing, it's possible that the facts have
> >> changed.
> >>
> >> Avoid this race by holding the drm_device->master_mutex lock while
> >> committing. When DRM userspace does its first open, it will now wait
> >> until fb-helper is done. The helper will stay away if there's a master.
> >>
> >> Locking rule: Always take the fb-helper lock first.
> >>
> >> v2:
> >> - Remove drm_fb_helper_is_bound() (Daniel Vetter)
> >> - No need to check fb_helper->dev->master in
> >>   drm_fb_helper_single_fb_probe(), restore_fbdev_mode() has the check.
> >>
> >> Suggested-by: Daniel Vetter <daniel.vetter at ffwll.ch>
> >> Signed-off-by: Noralf Trønnes <noralf at tronnes.org>
> >> ---
> >>  drivers/gpu/drm/drm_auth.c      | 20 ++++++++
> >>  drivers/gpu/drm/drm_fb_helper.c | 90 ++++++++++++++++-----------------
> >>  drivers/gpu/drm/drm_internal.h  |  2 +
> >>  3 files changed, 67 insertions(+), 45 deletions(-)
> >>
> >> diff --git a/drivers/gpu/drm/drm_auth.c b/drivers/gpu/drm/drm_auth.c
> >> index 1669c42c40ed..db199807b7dc 100644
> >> --- a/drivers/gpu/drm/drm_auth.c
> >> +++ b/drivers/gpu/drm/drm_auth.c
> >> @@ -368,3 +368,23 @@ void drm_master_put(struct drm_master **master)
> >>  	*master = NULL;
> >>  }
> >>  EXPORT_SYMBOL(drm_master_put);
> >> +
> >> +/* Used by drm_client and drm_fb_helper */
> >> +bool drm_master_internal_acquire(struct drm_device *dev)
> >> +{
> >> +	mutex_lock(&dev->master_mutex);
> >> +	if (dev->master) {
> >> +		mutex_unlock(&dev->master_mutex);
> >> +		return false;
> >> +	}
> >> +
> >> +	return true;
> >> +}
> >> +EXPORT_SYMBOL(drm_master_internal_acquire);
> >> +
> >> +/* Used by drm_client and drm_fb_helper */
> >> +void drm_master_internal_release(struct drm_device *dev)
> >> +{
> >> +	mutex_unlock(&dev->master_mutex);
> >> +}
> >> +EXPORT_SYMBOL(drm_master_internal_release);
> >> diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
> >> index 84791dd4a90d..a6be09ae899b 100644
> >> --- a/drivers/gpu/drm/drm_fb_helper.c
> >> +++ b/drivers/gpu/drm/drm_fb_helper.c
> >> @@ -44,6 +44,7 @@
> >>  
> >>  #include "drm_crtc_internal.h"
> >>  #include "drm_crtc_helper_internal.h"
> >> +#include "drm_internal.h"
> >>  
> >>  static bool drm_fbdev_emulation = true;
> >>  module_param_named(fbdev_emulation, drm_fbdev_emulation, bool, 0600);
> >> @@ -509,7 +510,7 @@ static int restore_fbdev_mode_legacy(struct drm_fb_helper *fb_helper)
> >>  	return ret;
> >>  }
> >>  
> >> -static int restore_fbdev_mode(struct drm_fb_helper *fb_helper)
> >> +static int restore_fbdev_mode_force(struct drm_fb_helper *fb_helper)
> > 
> > Bikeshed: usually the function variant that's run with locks already taken
> > is called _locked or has a __ prefix. _force feels a bit misplaced.
> 
> This isn't a _locked function in the usual sense, it is: apply modeset
> even if there is a DRM master. So we are _forcing a modeset on a
> possibly unexpecting DRM userspace. To me a _locked function would imply
> that the caller _must_ take a lock in order to use it.

Hm, good point. See my comments later on, I'm not sure the "forcing" is
really what we want. I think lastclose would be much better served if it
also checks for master status.

> But no big deal, I can rename it _locked if that reads better. After a
> few years of reading kernel code I've come to appreciate the consistency
> in how things are done and named. Every time things are different it
> slows down my internal logic/pattern parser.

If you agree with the entire "we should _force" then I think the __ prefix
would fit. It's the general "beware, this is special/internal" annotation.
My _locked suggestion was under the assumption that there's never a real
case where we want to do the unprotected modeset for an internal
drm_client.

> >>  {
> >>  	struct drm_device *dev = fb_helper->dev;
> >>  
> >> @@ -519,6 +520,21 @@ static int restore_fbdev_mode(struct drm_fb_helper *fb_helper)
> >>  		return restore_fbdev_mode_legacy(fb_helper);
> >>  }
> >>  
> >> +static int restore_fbdev_mode(struct drm_fb_helper *fb_helper)
> >> +{
> >> +	struct drm_device *dev = fb_helper->dev;
> >> +	int ret;
> >> +
> >> +	if (!drm_master_internal_acquire(dev))
> >> +		return -EBUSY;
> >> +
> >> +	ret = restore_fbdev_mode_force(fb_helper);
> >> +
> >> +	drm_master_internal_release(dev);
> >> +
> >> +	return ret;
> >> +}
> >> +
> >>  /**
> >>   * drm_fb_helper_restore_fbdev_mode_unlocked - restore fbdev configuration
> >>   * @fb_helper: driver-allocated fbdev helper, can be NULL
> >> @@ -556,34 +572,6 @@ int drm_fb_helper_restore_fbdev_mode_unlocked(struct drm_fb_helper *fb_helper)
> >>  }
> >>  EXPORT_SYMBOL(drm_fb_helper_restore_fbdev_mode_unlocked);
> >>  
> >> -static bool drm_fb_helper_is_bound(struct drm_fb_helper *fb_helper)
> >> -{
> >> -	struct drm_device *dev = fb_helper->dev;
> >> -	struct drm_crtc *crtc;
> >> -	int bound = 0, crtcs_bound = 0;
> >> -
> >> -	/*
> >> -	 * Sometimes user space wants everything disabled, so don't steal the
> >> -	 * display if there's a master.
> >> -	 */
> >> -	if (READ_ONCE(dev->master))
> >> -		return false;
> >> -
> >> -	drm_for_each_crtc(crtc, dev) {
> >> -		drm_modeset_lock(&crtc->mutex, NULL);
> >> -		if (crtc->primary->fb)
> >> -			crtcs_bound++;
> >> -		if (crtc->primary->fb == fb_helper->fb)
> >> -			bound++;
> >> -		drm_modeset_unlock(&crtc->mutex);
> >> -	}
> >> -
> >> -	if (bound < crtcs_bound)
> >> -		return false;
> >> -
> >> -	return true;
> >> -}
> >> -
> >>  #ifdef CONFIG_MAGIC_SYSRQ
> >>  /*
> >>   * restore fbcon display for all kms driver's using this helper, used for sysrq
> >> @@ -604,7 +592,7 @@ static bool drm_fb_helper_force_kernel_mode(void)
> >>  			continue;
> >>  
> >>  		mutex_lock(&helper->lock);
> >> -		ret = restore_fbdev_mode(helper);
> >> +		ret = restore_fbdev_mode_force(helper);
> > 
> > I'd leave this as-is, because:
> > a) I'm too lazy to review the locking of our open/close calls to convince
> > myself that lastclose can't race with the next open
> > b) it won't hurt
> > c) leaves the door open to easily make our open/close more concurrent in
> > the future
> > 
> 
> I'm not actually changing anything here, it's restore_fbdev_mode() that
> has changed, it now bails out if there's a DRM master. If we don't
> change this to _force/_locked, then sysrq won't work if there's a DRM
> master. Which kind of defeats the whole idea of this 'give me fbcon
> right now' functionality doesn't it?

Uh, I was blind, I thought this was for lastclose. With this one here I
think _force makes sense. I still think that the main drm_client_modeset
interface should be the normal one which checks for master for you.

> >>  		if (ret)
> >>  			error = true;
> >>  		mutex_unlock(&helper->lock);
> >> @@ -663,20 +651,22 @@ static void dpms_legacy(struct drm_fb_helper *fb_helper, int dpms_mode)
> >>  static void drm_fb_helper_dpms(struct fb_info *info, int dpms_mode)
> >>  {
> >>  	struct drm_fb_helper *fb_helper = info->par;
> >> +	struct drm_device *dev = fb_helper->dev;
> >>  
> >>  	/*
> >>  	 * For each CRTC in this fb, turn the connectors on/off.
> >>  	 */
> >>  	mutex_lock(&fb_helper->lock);
> >> -	if (!drm_fb_helper_is_bound(fb_helper)) {
> >> -		mutex_unlock(&fb_helper->lock);
> >> -		return;
> >> -	}
> >> +	if (!drm_master_internal_acquire(dev))
> >> +		goto unlock;
> >>  
> >> -	if (drm_drv_uses_atomic_modeset(fb_helper->dev))
> >> +	if (drm_drv_uses_atomic_modeset(dev))
> >>  		restore_fbdev_mode_atomic(fb_helper, dpms_mode == DRM_MODE_DPMS_ON);
> >>  	else
> >>  		dpms_legacy(fb_helper, dpms_mode);
> >> +
> >> +	drm_master_internal_release(dev);
> >> +unlock:
> >>  	mutex_unlock(&fb_helper->lock);
> >>  }
> >>  
> >> @@ -1509,6 +1499,7 @@ static int setcmap_atomic(struct fb_cmap *cmap, struct fb_info *info)
> >>  int drm_fb_helper_setcmap(struct fb_cmap *cmap, struct fb_info *info)
> >>  {
> >>  	struct drm_fb_helper *fb_helper = info->par;
> >> +	struct drm_device *dev = fb_helper->dev;
> >>  	int ret;
> >>  
> >>  	if (oops_in_progress)
> >> @@ -1516,9 +1507,9 @@ int drm_fb_helper_setcmap(struct fb_cmap *cmap, struct fb_info *info)
> >>  
> >>  	mutex_lock(&fb_helper->lock);
> >>  
> >> -	if (!drm_fb_helper_is_bound(fb_helper)) {
> >> +	if (!drm_master_internal_acquire(dev)) {
> >>  		ret = -EBUSY;
> >> -		goto out;
> >> +		goto unlock;
> >>  	}
> >>  
> >>  	if (info->fix.visual == FB_VISUAL_TRUECOLOR)
> >> @@ -1528,7 +1519,8 @@ int drm_fb_helper_setcmap(struct fb_cmap *cmap, struct fb_info *info)
> >>  	else
> >>  		ret = setcmap_legacy(cmap, info);
> >>  
> >> -out:
> >> +	drm_master_internal_release(dev);
> >> +unlock:
> >>  	mutex_unlock(&fb_helper->lock);
> >>  
> >>  	return ret;
> >> @@ -1548,12 +1540,13 @@ int drm_fb_helper_ioctl(struct fb_info *info, unsigned int cmd,
> >>  			unsigned long arg)
> >>  {
> >>  	struct drm_fb_helper *fb_helper = info->par;
> >> +	struct drm_device *dev = fb_helper->dev;
> >>  	struct drm_mode_set *mode_set;
> >>  	struct drm_crtc *crtc;
> >>  	int ret = 0;
> >>  
> >>  	mutex_lock(&fb_helper->lock);
> >> -	if (!drm_fb_helper_is_bound(fb_helper)) {
> >> +	if (!drm_master_internal_acquire(dev)) {
> >>  		ret = -EBUSY;
> >>  		goto unlock;
> >>  	}
> >> @@ -1591,11 +1584,12 @@ int drm_fb_helper_ioctl(struct fb_info *info, unsigned int cmd,
> >>  		}
> >>  
> >>  		ret = 0;
> >> -		goto unlock;
> >> +		break;
> >>  	default:
> >>  		ret = -ENOTTY;
> >>  	}
> >>  
> >> +	drm_master_internal_release(dev);
> >>  unlock:
> >>  	mutex_unlock(&fb_helper->lock);
> >>  	return ret;
> >> @@ -1847,15 +1841,18 @@ int drm_fb_helper_pan_display(struct fb_var_screeninfo *var,
> >>  		return -EBUSY;
> >>  
> >>  	mutex_lock(&fb_helper->lock);
> >> -	if (!drm_fb_helper_is_bound(fb_helper)) {
> >> -		mutex_unlock(&fb_helper->lock);
> >> -		return -EBUSY;
> >> +	if (!drm_master_internal_acquire(dev)) {
> >> +		ret = -EBUSY;
> >> +		goto unlock;
> >>  	}
> >>  
> >>  	if (drm_drv_uses_atomic_modeset(dev))
> >>  		ret = pan_display_atomic(var, info);
> >>  	else
> >>  		ret = pan_display_legacy(var, info);
> >> +
> >> +	drm_master_internal_release(dev);
> >> +unlock:
> >>  	mutex_unlock(&fb_helper->lock);
> >>  
> >>  	return ret;
> >> @@ -2014,7 +2011,7 @@ static int drm_fb_helper_single_fb_probe(struct drm_fb_helper *fb_helper,
> >>  		DRM_INFO("Cannot find any crtc or sizes\n");
> >>  
> >>  		/* First time: disable all crtc's.. */
> >> -		if (!fb_helper->deferred_setup && !READ_ONCE(fb_helper->dev->master))
> >> +		if (!fb_helper->deferred_setup)
> >>  			restore_fbdev_mode(fb_helper);
> > 
> > I think we need to return the errno here, since without that the higher
> > levels won't reprobe correctly. Plus we need to remap -EBUSY to -EAGAIN
> > (or change the check in __drm_fb_helper_initial_config_and_unlock to also
> > retry on -EBUSY).
> > 
> 
> I don't think so, because -EAGAIN is returned unconditionally on the
> line below. The restore_fbdev_mode() call is just to disable the outputs
> _if_ there's no DRM master, which that function now checks for.

Hm right, my logic parser went *boink* I think. Looks good on 2nd reading.
-Daniel
> 
> Noralf.
> 
> >>  		return -EAGAIN;
> >>  	}
> >> @@ -2842,6 +2839,7 @@ EXPORT_SYMBOL(drm_fb_helper_initial_config);
> >>   */
> >>  int drm_fb_helper_hotplug_event(struct drm_fb_helper *fb_helper)
> >>  {
> >> +	struct drm_device *dev = fb_helper->dev;
> >>  	int err = 0;
> >>  
> >>  	if (!drm_fbdev_emulation || !fb_helper)
> >> @@ -2854,12 +2852,14 @@ int drm_fb_helper_hotplug_event(struct drm_fb_helper *fb_helper)
> >>  		return err;
> >>  	}
> >>  
> >> -	if (!fb_helper->fb || !drm_fb_helper_is_bound(fb_helper)) {
> >> +	if (!fb_helper->fb || !drm_master_internal_acquire(dev)) {
> >>  		fb_helper->delayed_hotplug = true;
> >>  		mutex_unlock(&fb_helper->lock);
> >>  		return err;
> >>  	}
> >>  
> >> +	drm_master_internal_release(dev);
> >> +
> >>  	DRM_DEBUG_KMS("\n");
> >>  
> >>  	drm_setup_crtcs(fb_helper, fb_helper->fb->width, fb_helper->fb->height);
> >> diff --git a/drivers/gpu/drm/drm_internal.h b/drivers/gpu/drm/drm_internal.h
> >> index d9a483a5fce0..3ee97c9998a2 100644
> >> --- a/drivers/gpu/drm/drm_internal.h
> >> +++ b/drivers/gpu/drm/drm_internal.h
> >> @@ -91,6 +91,8 @@ int drm_dropmaster_ioctl(struct drm_device *dev, void *data,
> >>  			 struct drm_file *file_priv);
> >>  int drm_master_open(struct drm_file *file_priv);
> >>  void drm_master_release(struct drm_file *file_priv);
> >> +bool drm_master_internal_acquire(struct drm_device *dev);
> >> +void drm_master_internal_release(struct drm_device *dev);
> >>  
> >>  /* drm_sysfs.c */
> >>  extern struct class *drm_class;
> > 
> > With the nits addressed:
> > 
> > Reviewed-by: Daniel Vetter <daniel.vetter at ffwll.ch>
> > 
> >> -- 
> >> 2.20.1
> >>
> > 

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


More information about the Intel-gfx mailing list