[PATCH 1/8] drm/gem: Fix race in drm_gem_handle_create_tail()

[PATCH 1/8] drm/gem: Fix race in drm_gem_handle_create_tail()

[Simona Vetter]

Object creation is a careful dance where we must guarantee that the
object is fully constructed before it is visible to other threads, and
GEM buffer objects are no difference.

Final publishing happens by calling drm_gem_handle_create(). After
that the only allowed thing to do is call drm_gem_object_put() because
a concurrent call to the GEM_CLOSE ioctl with a correctly guessed id
(which is trivial since we have a linear allocator) can already tear
down the object again.

Luckily most drivers get this right, the very few exceptions I've
pinged the relevant maintainers for. Unfortunately we also need
drm_gem_handle_create() when creating additional handles for an
already existing object (e.g. GETFB ioctl or the various bo import
ioctl), and hence we cannot have a drm_gem_handle_create_and_put() as
the only exported function to stop these issues from happening.

Now unfortunately the implementation of drm_gem_handle_create() isn't
living up to standards: It does correctly finishe object
initialization at the global level, and hence is safe against a
concurrent tear down. But it also sets up the file-private aspects of
the handle, and that part goes wrong: We fully register the object in
the drm_file.object_idr before calling drm_vma_node_allow() or
obj->funcs->open, which opens up races against concurrent removal of
that handle in drm_gem_handle_delete().

Fix this with the usual two-stage approach of first reserving the
handle id, and then only registering the object after we've completed
the file-private setup.

Jacek reported this with a testcase of concurrently calling GEM_CLOSE
on a freshly-created object (which also destroys the object), but it
should be possible to hit this with just additional handles created
through import or GETFB without completed destroying the underlying
object with the concurrent GEM_CLOSE ioctl calls.

Note that the close-side of this race was fixed in f6cd7daecff5 ("drm:
Release driver references to handle before making it available
again"), which means a cool 9 years have passed until someone noticed
that we need to make this symmetry or there's still gaps left :-/
Without the 2-stage close approach we'd still have a race, therefore
that's an integral part of this bugfix.

More importantly, this means we can have NULL pointers behind
allocated id in our drm_file.object_idr. We need to check for that
now:

- drm_gem_handle_delete() checks for ERR_OR_NULL already

- drm_gem.c:object_lookup() also chekcs for NULL

- drm_gem_release() should never be called if there's another thread
  still existing that could call into an IOCTL that creates a new
  handle, so cannot race. For paranoia I added a NULL check to
  drm_gem_object_release_handle() though.

- most drivers (etnaviv, i915, msm) are find because they use
  idr_find, which maps both ENOENT and NULL to NULL.

- vmgfx is already broken vmw_debugfs_gem_info_show() because NULL
  pointers might exist due to drm_gem_handle_delete(). This needs a
  separate patch. This is because idr_for_each_entry terminates on the
  first NULL entry and so might not iterate over everything.

- similar for amd in amdgpu_debugfs_gem_info_show() and
  amdgpu_gem_force_release(). The latter is really questionable though
  since it's a best effort hack and there's no way to close all the
  races. Needs separate patches.

- xe is really broken because it not uses idr_for_each_entry() but
  also drops the drm_file.table_lock, which can wreak the idr iterator
  state if you're unlucky enough. Maybe another reason to look into
  the drm fdinfo memory stats instead of hand-rolling too much.

- drm_show_memory_stats() is also broken since it uses
  idr_for_each_entry. But since that's a preexisting bug I'll follow
  up with a separate patch.

Reported-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Cc: stable@vger.kernel.org
Cc: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Maxime Ripard <mripard@kernel.org>
Cc: Thomas Zimmermann <tzimmermann@suse.de>
Cc: David Airlie <airlied@gmail.com>
Cc: Simona Vetter <simona@ffwll.ch>
Signed-off-by: Simona Vetter <simona.vetter@intel.com>
Signed-off-by: Simona Vetter <simona.vetter@ffwll.ch>
---
 drivers/gpu/drm/drm_gem.c | 10 +++++++++-
 include/drm/drm_file.h    |  3 +++
 2 files changed, 12 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
index 1e659d2660f7..e4e20dda47b1 100644
--- a/drivers/gpu/drm/drm_gem.c
+++ b/drivers/gpu/drm/drm_gem.c
@@ -279,6 +279,9 @@ drm_gem_object_release_handle(int id, void *ptr, void *data)
 	struct drm_file *file_priv = data;
 	struct drm_gem_object *obj = ptr;
 
+	if (WARN_ON(!data))
+		return 0;
+
 	if (obj->funcs->close)
 		obj->funcs->close(obj, file_priv);
 
@@ -399,7 +402,7 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
 	idr_preload(GFP_KERNEL);
 	spin_lock(&file_priv->table_lock);
 
-	ret = idr_alloc(&file_priv->object_idr, obj, 1, 0, GFP_NOWAIT);
+	ret = idr_alloc(&file_priv->object_idr, NULL, 1, 0, GFP_NOWAIT);
 
 	spin_unlock(&file_priv->table_lock);
 	idr_preload_end();
@@ -420,6 +423,11 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
 			goto err_revoke;
 	}
 
+	/* mirrors drm_gem_handle_delete to avoid races */
+	spin_lock(&file_priv->table_lock);
+	obj = idr_replace(&file_priv->object_idr, obj, handle);
+	WARN_ON(obj != NULL);
+	spin_unlock(&file_priv->table_lock);
 	*handlep = handle;
 	return 0;
 
diff --git a/include/drm/drm_file.h b/include/drm/drm_file.h
index 5c3b2aa3e69d..d344d41e6cfe 100644
--- a/include/drm/drm_file.h
+++ b/include/drm/drm_file.h
@@ -300,6 +300,9 @@ struct drm_file {
 	 *
 	 * Mapping of mm object handles to object pointers. Used by the GEM
 	 * subsystem. Protected by @table_lock.
+	 *
+	 * Note that allocated entries might be NULL as a transient state when
+	 * creating or deleting a handle.
 	 */
 	struct idr object_idr;
 
-- 
2.49.0

[PATCH 2/8] drm/fdinfo: Switch to idr_for_each() in drm_show_memory_stats()

[Simona Vetter]

Unlike idr_for_each_entry(), which terminates on the first NULL entry,
idr_for_each passes them through. This fixes potential issues with the
idr walk terminating prematurely due to transient NULL entries the
exist when creating and destroying a handle.

Note that transient NULL pointers in drm_file.object_idr have been a
thing since f6cd7daecff5 ("drm: Release driver references to handle
before making it available again"), this is a really old issue.

Aside from temporarily inconsistent fdinfo statistic there's no other
impact of this issue.

Fixes: 686b21b5f6ca ("drm: Add fdinfo memory stats")
Cc: Rob Clark <robdclark@chromium.org>
Cc: Emil Velikov <emil.l.velikov@gmail.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: <stable@vger.kernel.org> # v6.5+
Signed-off-by: Simona Vetter <simona.vetter@intel.com>
Signed-off-by: Simona Vetter <simona.vetter@ffwll.ch>
---
 drivers/gpu/drm/drm_file.c | 95 ++++++++++++++++++++++----------------
 1 file changed, 55 insertions(+), 40 deletions(-)

diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c
index 246cf845e2c9..428a4eb85e94 100644
--- a/drivers/gpu/drm/drm_file.c
+++ b/drivers/gpu/drm/drm_file.c
@@ -892,6 +892,58 @@ void drm_print_memory_stats(struct drm_printer *p,
 }
 EXPORT_SYMBOL(drm_print_memory_stats);
 
+struct drm_bo_print_data {
+	struct drm_memory_stats status;
+	enum drm_gem_object_status supported_status;
+};
+
+static int
+drm_bo_memory_stats(int id, void *ptr, void *data)
+{
+	struct drm_bo_print_data *drm_data;
+	struct drm_gem_object *obj = ptr;
+	enum drm_gem_object_status s = 0;
+	size_t add_size;
+
+	if (!obj)
+		return 0;
+
+	add_size = (obj->funcs && obj->funcs->rss) ?
+		obj->funcs->rss(obj) : obj->size;
+
+	if (obj->funcs && obj->funcs->status) {
+		s = obj->funcs->status(obj);
+		drm_data->supported_status |= s;
+	}
+
+	if (drm_gem_object_is_shared_for_memory_stats(obj))
+		drm_data->status.shared += obj->size;
+	else
+		drm_data->status.private += obj->size;
+
+	if (s & DRM_GEM_OBJECT_RESIDENT) {
+		drm_data->status.resident += add_size;
+	} else {
+		/* If already purged or not yet backed by pages, don't
+		 * count it as purgeable:
+		 */
+		s &= ~DRM_GEM_OBJECT_PURGEABLE;
+	}
+
+	if (!dma_resv_test_signaled(obj->resv, dma_resv_usage_rw(true))) {
+		drm_data->status.active += add_size;
+		drm_data->supported_status |= DRM_GEM_OBJECT_ACTIVE;
+
+		/* If still active, don't count as purgeable: */
+		s &= ~DRM_GEM_OBJECT_PURGEABLE;
+	}
+
+	if (s & DRM_GEM_OBJECT_PURGEABLE)
+		drm_data->status.purgeable += add_size;
+
+	return 0;
+}
+
 /**
  * drm_show_memory_stats - Helper to collect and show standard fdinfo memory stats
  * @p: the printer to print output to
@@ -902,50 +954,13 @@ EXPORT_SYMBOL(drm_print_memory_stats);
  */
 void drm_show_memory_stats(struct drm_printer *p, struct drm_file *file)
 {
-	struct drm_gem_object *obj;
-	struct drm_memory_stats status = {};
-	enum drm_gem_object_status supported_status = 0;
-	int id;
+	struct drm_bo_print_data data = {};
 
 	spin_lock(&file->table_lock);
-	idr_for_each_entry (&file->object_idr, obj, id) {
-		enum drm_gem_object_status s = 0;
-		size_t add_size = (obj->funcs && obj->funcs->rss) ?
-			obj->funcs->rss(obj) : obj->size;
-
-		if (obj->funcs && obj->funcs->status) {
-			s = obj->funcs->status(obj);
-			supported_status |= s;
-		}
-
-		if (drm_gem_object_is_shared_for_memory_stats(obj))
-			status.shared += obj->size;
-		else
-			status.private += obj->size;
-
-		if (s & DRM_GEM_OBJECT_RESIDENT) {
-			status.resident += add_size;
-		} else {
-			/* If already purged or not yet backed by pages, don't
-			 * count it as purgeable:
-			 */
-			s &= ~DRM_GEM_OBJECT_PURGEABLE;
-		}
-
-		if (!dma_resv_test_signaled(obj->resv, dma_resv_usage_rw(true))) {
-			status.active += add_size;
-			supported_status |= DRM_GEM_OBJECT_ACTIVE;
-
-			/* If still active, don't count as purgeable: */
-			s &= ~DRM_GEM_OBJECT_PURGEABLE;
-		}
-
-		if (s & DRM_GEM_OBJECT_PURGEABLE)
-			status.purgeable += add_size;
-	}
+	idr_for_each(&file->object_idr, &drm_bo_memory_stats, &data);
 	spin_unlock(&file->table_lock);
 
-	drm_print_memory_stats(p, &status, supported_status, "memory");
+	drm_print_memory_stats(p, &data.status, data.supported_status, "memory");
 }
 EXPORT_SYMBOL(drm_show_memory_stats);
 
-- 
2.49.0

Re: [PATCH 2/8] drm/fdinfo: Switch to idr_for_each() in drm_show_memory_stats()

[Simona Vetter]

On Wed, May 28, 2025 at 11:13:00AM +0200, Simona Vetter wrote:
> Unlike idr_for_each_entry(), which terminates on the first NULL entry,
> idr_for_each passes them through. This fixes potential issues with the
> idr walk terminating prematurely due to transient NULL entries the
> exist when creating and destroying a handle.
> 
> Note that transient NULL pointers in drm_file.object_idr have been a
> thing since f6cd7daecff5 ("drm: Release driver references to handle
> before making it available again"), this is a really old issue.
> 
> Aside from temporarily inconsistent fdinfo statistic there's no other
> impact of this issue.
> 
> Fixes: 686b21b5f6ca ("drm: Add fdinfo memory stats")
> Cc: Rob Clark <robdclark@chromium.org>
> Cc: Emil Velikov <emil.l.velikov@gmail.com>
> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> Cc: <stable@vger.kernel.org> # v6.5+
> Signed-off-by: Simona Vetter <simona.vetter@intel.com>
> Signed-off-by: Simona Vetter <simona.vetter@ffwll.ch>

Ok I screwed up reading idr_for_each_entry() respectively
idr_get_next_ul() big time, it already copes with NULL entries entirely
fine.

Mea culpa.
-Sima

> ---
>  drivers/gpu/drm/drm_file.c | 95 ++++++++++++++++++++++----------------
>  1 file changed, 55 insertions(+), 40 deletions(-)
> 
> diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c
> index 246cf845e2c9..428a4eb85e94 100644
> --- a/drivers/gpu/drm/drm_file.c
> +++ b/drivers/gpu/drm/drm_file.c
> @@ -892,6 +892,58 @@ void drm_print_memory_stats(struct drm_printer *p,
>  }
>  EXPORT_SYMBOL(drm_print_memory_stats);
>  
> +struct drm_bo_print_data {
> +	struct drm_memory_stats status;
> +	enum drm_gem_object_status supported_status;
> +};
> +
> +static int
> +drm_bo_memory_stats(int id, void *ptr, void *data)
> +{
> +	struct drm_bo_print_data *drm_data;
> +	struct drm_gem_object *obj = ptr;
> +	enum drm_gem_object_status s = 0;
> +	size_t add_size;
> +
> +	if (!obj)
> +		return 0;
> +
> +	add_size = (obj->funcs && obj->funcs->rss) ?
> +		obj->funcs->rss(obj) : obj->size;
> +
> +	if (obj->funcs && obj->funcs->status) {
> +		s = obj->funcs->status(obj);
> +		drm_data->supported_status |= s;
> +	}
> +
> +	if (drm_gem_object_is_shared_for_memory_stats(obj))
> +		drm_data->status.shared += obj->size;
> +	else
> +		drm_data->status.private += obj->size;
> +
> +	if (s & DRM_GEM_OBJECT_RESIDENT) {
> +		drm_data->status.resident += add_size;
> +	} else {
> +		/* If already purged or not yet backed by pages, don't
> +		 * count it as purgeable:
> +		 */
> +		s &= ~DRM_GEM_OBJECT_PURGEABLE;
> +	}
> +
> +	if (!dma_resv_test_signaled(obj->resv, dma_resv_usage_rw(true))) {
> +		drm_data->status.active += add_size;
> +		drm_data->supported_status |= DRM_GEM_OBJECT_ACTIVE;
> +
> +		/* If still active, don't count as purgeable: */
> +		s &= ~DRM_GEM_OBJECT_PURGEABLE;
> +	}
> +
> +	if (s & DRM_GEM_OBJECT_PURGEABLE)
> +		drm_data->status.purgeable += add_size;
> +
> +	return 0;
> +}
> +
>  /**
>   * drm_show_memory_stats - Helper to collect and show standard fdinfo memory stats
>   * @p: the printer to print output to
> @@ -902,50 +954,13 @@ EXPORT_SYMBOL(drm_print_memory_stats);
>   */
>  void drm_show_memory_stats(struct drm_printer *p, struct drm_file *file)
>  {
> -	struct drm_gem_object *obj;
> -	struct drm_memory_stats status = {};
> -	enum drm_gem_object_status supported_status = 0;
> -	int id;
> +	struct drm_bo_print_data data = {};
>  
>  	spin_lock(&file->table_lock);
> -	idr_for_each_entry (&file->object_idr, obj, id) {
> -		enum drm_gem_object_status s = 0;
> -		size_t add_size = (obj->funcs && obj->funcs->rss) ?
> -			obj->funcs->rss(obj) : obj->size;
> -
> -		if (obj->funcs && obj->funcs->status) {
> -			s = obj->funcs->status(obj);
> -			supported_status |= s;
> -		}
> -
> -		if (drm_gem_object_is_shared_for_memory_stats(obj))
> -			status.shared += obj->size;
> -		else
> -			status.private += obj->size;
> -
> -		if (s & DRM_GEM_OBJECT_RESIDENT) {
> -			status.resident += add_size;
> -		} else {
> -			/* If already purged or not yet backed by pages, don't
> -			 * count it as purgeable:
> -			 */
> -			s &= ~DRM_GEM_OBJECT_PURGEABLE;
> -		}
> -
> -		if (!dma_resv_test_signaled(obj->resv, dma_resv_usage_rw(true))) {
> -			status.active += add_size;
> -			supported_status |= DRM_GEM_OBJECT_ACTIVE;
> -
> -			/* If still active, don't count as purgeable: */
> -			s &= ~DRM_GEM_OBJECT_PURGEABLE;
> -		}
> -
> -		if (s & DRM_GEM_OBJECT_PURGEABLE)
> -			status.purgeable += add_size;
> -	}
> +	idr_for_each(&file->object_idr, &drm_bo_memory_stats, &data);
>  	spin_unlock(&file->table_lock);
>  
> -	drm_print_memory_stats(p, &status, supported_status, "memory");
> +	drm_print_memory_stats(p, &data.status, data.supported_status, "memory");
>  }
>  EXPORT_SYMBOL(drm_show_memory_stats);
>  
> -- 
> 2.49.0
> 

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

Re: [PATCH 1/8] drm/gem: Fix race in drm_gem_handle_create_tail()

[Simona Vetter]

On Wed, May 28, 2025 at 11:12:59AM +0200, Simona Vetter wrote:
> Object creation is a careful dance where we must guarantee that the
> object is fully constructed before it is visible to other threads, and
> GEM buffer objects are no difference.
> 
> Final publishing happens by calling drm_gem_handle_create(). After
> that the only allowed thing to do is call drm_gem_object_put() because
> a concurrent call to the GEM_CLOSE ioctl with a correctly guessed id
> (which is trivial since we have a linear allocator) can already tear
> down the object again.
> 
> Luckily most drivers get this right, the very few exceptions I've
> pinged the relevant maintainers for. Unfortunately we also need
> drm_gem_handle_create() when creating additional handles for an
> already existing object (e.g. GETFB ioctl or the various bo import
> ioctl), and hence we cannot have a drm_gem_handle_create_and_put() as
> the only exported function to stop these issues from happening.
> 
> Now unfortunately the implementation of drm_gem_handle_create() isn't
> living up to standards: It does correctly finishe object
> initialization at the global level, and hence is safe against a
> concurrent tear down. But it also sets up the file-private aspects of
> the handle, and that part goes wrong: We fully register the object in
> the drm_file.object_idr before calling drm_vma_node_allow() or
> obj->funcs->open, which opens up races against concurrent removal of
> that handle in drm_gem_handle_delete().
> 
> Fix this with the usual two-stage approach of first reserving the
> handle id, and then only registering the object after we've completed
> the file-private setup.
> 
> Jacek reported this with a testcase of concurrently calling GEM_CLOSE
> on a freshly-created object (which also destroys the object), but it
> should be possible to hit this with just additional handles created
> through import or GETFB without completed destroying the underlying
> object with the concurrent GEM_CLOSE ioctl calls.
> 
> Note that the close-side of this race was fixed in f6cd7daecff5 ("drm:
> Release driver references to handle before making it available
> again"), which means a cool 9 years have passed until someone noticed
> that we need to make this symmetry or there's still gaps left :-/
> Without the 2-stage close approach we'd still have a race, therefore
> that's an integral part of this bugfix.
> 
> More importantly, this means we can have NULL pointers behind
> allocated id in our drm_file.object_idr. We need to check for that
> now:
> 
> - drm_gem_handle_delete() checks for ERR_OR_NULL already
> 
> - drm_gem.c:object_lookup() also chekcs for NULL
> 
> - drm_gem_release() should never be called if there's another thread
>   still existing that could call into an IOCTL that creates a new
>   handle, so cannot race. For paranoia I added a NULL check to
>   drm_gem_object_release_handle() though.
> 
> - most drivers (etnaviv, i915, msm) are find because they use
>   idr_find, which maps both ENOENT and NULL to NULL.
> 
> - vmgfx is already broken vmw_debugfs_gem_info_show() because NULL
>   pointers might exist due to drm_gem_handle_delete(). This needs a
>   separate patch. This is because idr_for_each_entry terminates on the
>   first NULL entry and so might not iterate over everything.
> 
> - similar for amd in amdgpu_debugfs_gem_info_show() and
>   amdgpu_gem_force_release(). The latter is really questionable though
>   since it's a best effort hack and there's no way to close all the
>   races. Needs separate patches.
> 
> - xe is really broken because it not uses idr_for_each_entry() but
>   also drops the drm_file.table_lock, which can wreak the idr iterator
>   state if you're unlucky enough. Maybe another reason to look into
>   the drm fdinfo memory stats instead of hand-rolling too much.
> 
> - drm_show_memory_stats() is also broken since it uses
>   idr_for_each_entry. But since that's a preexisting bug I'll follow
>   up with a separate patch.

I've already reworded the commit message locally since I now think
idr_for_each_entry is entirely fine.
-Sima

> 
> Reported-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
> Cc: stable@vger.kernel.org
> Cc: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> Cc: Maxime Ripard <mripard@kernel.org>
> Cc: Thomas Zimmermann <tzimmermann@suse.de>
> Cc: David Airlie <airlied@gmail.com>
> Cc: Simona Vetter <simona@ffwll.ch>
> Signed-off-by: Simona Vetter <simona.vetter@intel.com>
> Signed-off-by: Simona Vetter <simona.vetter@ffwll.ch>
> ---
>  drivers/gpu/drm/drm_gem.c | 10 +++++++++-
>  include/drm/drm_file.h    |  3 +++
>  2 files changed, 12 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
> index 1e659d2660f7..e4e20dda47b1 100644
> --- a/drivers/gpu/drm/drm_gem.c
> +++ b/drivers/gpu/drm/drm_gem.c
> @@ -279,6 +279,9 @@ drm_gem_object_release_handle(int id, void *ptr, void *data)
>  	struct drm_file *file_priv = data;
>  	struct drm_gem_object *obj = ptr;
>  
> +	if (WARN_ON(!data))
> +		return 0;
> +
>  	if (obj->funcs->close)
>  		obj->funcs->close(obj, file_priv);
>  
> @@ -399,7 +402,7 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
>  	idr_preload(GFP_KERNEL);
>  	spin_lock(&file_priv->table_lock);
>  
> -	ret = idr_alloc(&file_priv->object_idr, obj, 1, 0, GFP_NOWAIT);
> +	ret = idr_alloc(&file_priv->object_idr, NULL, 1, 0, GFP_NOWAIT);
>  
>  	spin_unlock(&file_priv->table_lock);
>  	idr_preload_end();
> @@ -420,6 +423,11 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
>  			goto err_revoke;
>  	}
>  
> +	/* mirrors drm_gem_handle_delete to avoid races */
> +	spin_lock(&file_priv->table_lock);
> +	obj = idr_replace(&file_priv->object_idr, obj, handle);
> +	WARN_ON(obj != NULL);
> +	spin_unlock(&file_priv->table_lock);
>  	*handlep = handle;
>  	return 0;
>  
> diff --git a/include/drm/drm_file.h b/include/drm/drm_file.h
> index 5c3b2aa3e69d..d344d41e6cfe 100644
> --- a/include/drm/drm_file.h
> +++ b/include/drm/drm_file.h
> @@ -300,6 +300,9 @@ struct drm_file {
>  	 *
>  	 * Mapping of mm object handles to object pointers. Used by the GEM
>  	 * subsystem. Protected by @table_lock.
> +	 *
> +	 * Note that allocated entries might be NULL as a transient state when
> +	 * creating or deleting a handle.
>  	 */
>  	struct idr object_idr;
>  
> -- 
> 2.49.0
> 

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

Re: [PATCH 1/8] drm/gem: Fix race in drm_gem_handle_create_tail()

[Jacek Lawrynowicz]

This fixes the race for me.

Tested-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>

On 5/28/2025 11:12 AM, Simona Vetter wrote:
> Object creation is a careful dance where we must guarantee that the
> object is fully constructed before it is visible to other threads, and
> GEM buffer objects are no difference.
> 
> Final publishing happens by calling drm_gem_handle_create(). After
> that the only allowed thing to do is call drm_gem_object_put() because
> a concurrent call to the GEM_CLOSE ioctl with a correctly guessed id
> (which is trivial since we have a linear allocator) can already tear
> down the object again.
> 
> Luckily most drivers get this right, the very few exceptions I've
> pinged the relevant maintainers for. Unfortunately we also need
> drm_gem_handle_create() when creating additional handles for an
> already existing object (e.g. GETFB ioctl or the various bo import
> ioctl), and hence we cannot have a drm_gem_handle_create_and_put() as
> the only exported function to stop these issues from happening.
> 
> Now unfortunately the implementation of drm_gem_handle_create() isn't
> living up to standards: It does correctly finishe object
> initialization at the global level, and hence is safe against a
> concurrent tear down. But it also sets up the file-private aspects of
> the handle, and that part goes wrong: We fully register the object in
> the drm_file.object_idr before calling drm_vma_node_allow() or
> obj->funcs->open, which opens up races against concurrent removal of
> that handle in drm_gem_handle_delete().
> 
> Fix this with the usual two-stage approach of first reserving the
> handle id, and then only registering the object after we've completed
> the file-private setup.
> 
> Jacek reported this with a testcase of concurrently calling GEM_CLOSE
> on a freshly-created object (which also destroys the object), but it
> should be possible to hit this with just additional handles created
> through import or GETFB without completed destroying the underlying
> object with the concurrent GEM_CLOSE ioctl calls.
> 
> Note that the close-side of this race was fixed in f6cd7daecff5 ("drm:
> Release driver references to handle before making it available
> again"), which means a cool 9 years have passed until someone noticed
> that we need to make this symmetry or there's still gaps left :-/
> Without the 2-stage close approach we'd still have a race, therefore
> that's an integral part of this bugfix.
> 
> More importantly, this means we can have NULL pointers behind
> allocated id in our drm_file.object_idr. We need to check for that
> now:
> 
> - drm_gem_handle_delete() checks for ERR_OR_NULL already
> 
> - drm_gem.c:object_lookup() also chekcs for NULL
> 
> - drm_gem_release() should never be called if there's another thread
>   still existing that could call into an IOCTL that creates a new
>   handle, so cannot race. For paranoia I added a NULL check to
>   drm_gem_object_release_handle() though.
> 
> - most drivers (etnaviv, i915, msm) are find because they use
>   idr_find, which maps both ENOENT and NULL to NULL.
> 
> - vmgfx is already broken vmw_debugfs_gem_info_show() because NULL
>   pointers might exist due to drm_gem_handle_delete(). This needs a
>   separate patch. This is because idr_for_each_entry terminates on the
>   first NULL entry and so might not iterate over everything.
> 
> - similar for amd in amdgpu_debugfs_gem_info_show() and
>   amdgpu_gem_force_release(). The latter is really questionable though
>   since it's a best effort hack and there's no way to close all the
>   races. Needs separate patches.
> 
> - xe is really broken because it not uses idr_for_each_entry() but
>   also drops the drm_file.table_lock, which can wreak the idr iterator
>   state if you're unlucky enough. Maybe another reason to look into
>   the drm fdinfo memory stats instead of hand-rolling too much.
> 
> - drm_show_memory_stats() is also broken since it uses
>   idr_for_each_entry. But since that's a preexisting bug I'll follow
>   up with a separate patch.
> 
> Reported-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
> Cc: stable@vger.kernel.org
> Cc: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> Cc: Maxime Ripard <mripard@kernel.org>
> Cc: Thomas Zimmermann <tzimmermann@suse.de>
> Cc: David Airlie <airlied@gmail.com>
> Cc: Simona Vetter <simona@ffwll.ch>
> Signed-off-by: Simona Vetter <simona.vetter@intel.com>
> Signed-off-by: Simona Vetter <simona.vetter@ffwll.ch>
> ---
>  drivers/gpu/drm/drm_gem.c | 10 +++++++++-
>  include/drm/drm_file.h    |  3 +++
>  2 files changed, 12 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
> index 1e659d2660f7..e4e20dda47b1 100644
> --- a/drivers/gpu/drm/drm_gem.c
> +++ b/drivers/gpu/drm/drm_gem.c
> @@ -279,6 +279,9 @@ drm_gem_object_release_handle(int id, void *ptr, void *data)
>  	struct drm_file *file_priv = data;
>  	struct drm_gem_object *obj = ptr;
>  
> +	if (WARN_ON(!data))
> +		return 0;
> +
>  	if (obj->funcs->close)
>  		obj->funcs->close(obj, file_priv);
>  
> @@ -399,7 +402,7 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
>  	idr_preload(GFP_KERNEL);
>  	spin_lock(&file_priv->table_lock);
>  
> -	ret = idr_alloc(&file_priv->object_idr, obj, 1, 0, GFP_NOWAIT);
> +	ret = idr_alloc(&file_priv->object_idr, NULL, 1, 0, GFP_NOWAIT);
>  
>  	spin_unlock(&file_priv->table_lock);
>  	idr_preload_end();
> @@ -420,6 +423,11 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
>  			goto err_revoke;
>  	}
>  
> +	/* mirrors drm_gem_handle_delete to avoid races */
> +	spin_lock(&file_priv->table_lock);
> +	obj = idr_replace(&file_priv->object_idr, obj, handle);
> +	WARN_ON(obj != NULL);
> +	spin_unlock(&file_priv->table_lock);
>  	*handlep = handle;
>  	return 0;
>  
> diff --git a/include/drm/drm_file.h b/include/drm/drm_file.h
> index 5c3b2aa3e69d..d344d41e6cfe 100644
> --- a/include/drm/drm_file.h
> +++ b/include/drm/drm_file.h
> @@ -300,6 +300,9 @@ struct drm_file {
>  	 *
>  	 * Mapping of mm object handles to object pointers. Used by the GEM
>  	 * subsystem. Protected by @table_lock.
> +	 *
> +	 * Note that allocated entries might be NULL as a transient state when
> +	 * creating or deleting a handle.
>  	 */
>  	struct idr object_idr;
>  

Re: [PATCH 2/8] drm/fdinfo: Switch to idr_for_each() in drm_show_memory_stats()

[kernel test robot]

Hi Simona,

kernel test robot noticed the following build warnings:

[auto build test WARNING on next-20250527]
[also build test WARNING on linus/master v6.15]
[cannot apply to v6.15 v6.15-rc7 v6.15-rc6]
[If your patch is applied to the wrong git tree, kindly drop us a note.
And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch#_base_tree_information]

url:    https://github.com/intel-lab-lkp/linux/commits/Simona-Vetter/drm-gem-Fix-race-in-drm_gem_handle_create_tail/20250528-171524
base:   next-20250527
patch link:    https://lore.kernel.org/r/20250528091307.1894940-3-simona.vetter%40ffwll.ch
patch subject: [PATCH 2/8] drm/fdinfo: Switch to idr_for_each() in drm_show_memory_stats()
config: riscv-randconfig-001-20250529 (https://download.01.org/0day-ci/archive/20250529/202505290334.GjoY9qsk-lkp@intel.com/config)
compiler: clang version 21.0.0git (https://github.com/llvm/llvm-project f819f46284f2a79790038e1f6649172789734ae8)
reproduce (this is a W=1 build): (https://download.01.org/0day-ci/archive/20250529/202505290334.GjoY9qsk-lkp@intel.com/reproduce)

If you fix the issue in a separate patch/commit (i.e. not just a new version of
the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <lkp@intel.com>
| Closes: https://lore.kernel.org/oe-kbuild-all/202505290334.GjoY9qsk-lkp@intel.com/

All warnings (new ones prefixed by >>):

>> drivers/gpu/drm/drm_file.c:916:3: warning: variable 'drm_data' is uninitialized when used here [-Wuninitialized]
     916 |                 drm_data->supported_status |= s;
         |                 ^~~~~~~~
   drivers/gpu/drm/drm_file.c:903:36: note: initialize the variable 'drm_data' to silence this warning
     903 |         struct drm_bo_print_data *drm_data;
         |                                           ^
         |                                            = NULL
   1 warning generated.


vim +/drm_data +916 drivers/gpu/drm/drm_file.c

   899	
   900	static int
   901	drm_bo_memory_stats(int id, void *ptr, void *data)
   902	{
   903		struct drm_bo_print_data *drm_data;
   904		struct drm_gem_object *obj = ptr;
   905		enum drm_gem_object_status s = 0;
   906		size_t add_size;
   907	
   908		if (!obj)
   909			return 0;
   910	
   911		add_size = (obj->funcs && obj->funcs->rss) ?
   912			obj->funcs->rss(obj) : obj->size;
   913	
   914		if (obj->funcs && obj->funcs->status) {
   915			s = obj->funcs->status(obj);
 > 916			drm_data->supported_status |= s;
   917		}
   918	
   919		if (drm_gem_object_is_shared_for_memory_stats(obj))
   920			drm_data->status.shared += obj->size;
   921		else
   922			drm_data->status.private += obj->size;
   923	
   924		if (s & DRM_GEM_OBJECT_RESIDENT) {
   925			drm_data->status.resident += add_size;
   926		} else {
   927			/* If already purged or not yet backed by pages, don't
   928			 * count it as purgeable:
   929			 */
   930			s &= ~DRM_GEM_OBJECT_PURGEABLE;
   931		}
   932	
   933		if (!dma_resv_test_signaled(obj->resv, dma_resv_usage_rw(true))) {
   934			drm_data->status.active += add_size;
   935			drm_data->supported_status |= DRM_GEM_OBJECT_ACTIVE;
   936	
   937			/* If still active, don't count as purgeable: */
   938			s &= ~DRM_GEM_OBJECT_PURGEABLE;
   939		}
   940	
   941		if (s & DRM_GEM_OBJECT_PURGEABLE)
   942			drm_data->status.purgeable += add_size;
   943	
   944		return 0;
   945	}
   946	

-- 
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests/wiki

Re: [PATCH 1/8] drm/gem: Fix race in drm_gem_handle_create_tail()

[Thomas Zimmermann]

Hi

Am 28.05.25 um 11:12 schrieb Simona Vetter:
> Object creation is a careful dance where we must guarantee that the
> object is fully constructed before it is visible to other threads, and
> GEM buffer objects are no difference.
>
> Final publishing happens by calling drm_gem_handle_create(). After
> that the only allowed thing to do is call drm_gem_object_put() because
> a concurrent call to the GEM_CLOSE ioctl with a correctly guessed id
> (which is trivial since we have a linear allocator) can already tear
> down the object again.
>
> Luckily most drivers get this right, the very few exceptions I've
> pinged the relevant maintainers for. Unfortunately we also need
> drm_gem_handle_create() when creating additional handles for an
> already existing object (e.g. GETFB ioctl or the various bo import
> ioctl), and hence we cannot have a drm_gem_handle_create_and_put() as
> the only exported function to stop these issues from happening.
>
> Now unfortunately the implementation of drm_gem_handle_create() isn't
> living up to standards: It does correctly finishe object
> initialization at the global level, and hence is safe against a
> concurrent tear down. But it also sets up the file-private aspects of
> the handle, and that part goes wrong: We fully register the object in
> the drm_file.object_idr before calling drm_vma_node_allow() or
> obj->funcs->open, which opens up races against concurrent removal of
> that handle in drm_gem_handle_delete().
>
> Fix this with the usual two-stage approach of first reserving the
> handle id, and then only registering the object after we've completed
> the file-private setup.
>
> Jacek reported this with a testcase of concurrently calling GEM_CLOSE
> on a freshly-created object (which also destroys the object), but it
> should be possible to hit this with just additional handles created
> through import or GETFB without completed destroying the underlying
> object with the concurrent GEM_CLOSE ioctl calls.
>
> Note that the close-side of this race was fixed in f6cd7daecff5 ("drm:
> Release driver references to handle before making it available
> again"), which means a cool 9 years have passed until someone noticed
> that we need to make this symmetry or there's still gaps left :-/
> Without the 2-stage close approach we'd still have a race, therefore
> that's an integral part of this bugfix.
>
> More importantly, this means we can have NULL pointers behind
> allocated id in our drm_file.object_idr. We need to check for that
> now:
>
> - drm_gem_handle_delete() checks for ERR_OR_NULL already
>
> - drm_gem.c:object_lookup() also chekcs for NULL
>
> - drm_gem_release() should never be called if there's another thread
>    still existing that could call into an IOCTL that creates a new
>    handle, so cannot race. For paranoia I added a NULL check to
>    drm_gem_object_release_handle() though.
>
> - most drivers (etnaviv, i915, msm) are find because they use
>    idr_find, which maps both ENOENT and NULL to NULL.
>
> - vmgfx is already broken vmw_debugfs_gem_info_show() because NULL
>    pointers might exist due to drm_gem_handle_delete(). This needs a
>    separate patch. This is because idr_for_each_entry terminates on the
>    first NULL entry and so might not iterate over everything.
>
> - similar for amd in amdgpu_debugfs_gem_info_show() and
>    amdgpu_gem_force_release(). The latter is really questionable though
>    since it's a best effort hack and there's no way to close all the
>    races. Needs separate patches.
>
> - xe is really broken because it not uses idr_for_each_entry() but
>    also drops the drm_file.table_lock, which can wreak the idr iterator
>    state if you're unlucky enough. Maybe another reason to look into
>    the drm fdinfo memory stats instead of hand-rolling too much.
>
> - drm_show_memory_stats() is also broken since it uses
>    idr_for_each_entry. But since that's a preexisting bug I'll follow
>    up with a separate patch.
>
> Reported-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
> Cc: stable@vger.kernel.org
> Cc: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> Cc: Maxime Ripard <mripard@kernel.org>
> Cc: Thomas Zimmermann <tzimmermann@suse.de>
> Cc: David Airlie <airlied@gmail.com>
> Cc: Simona Vetter <simona@ffwll.ch>
> Signed-off-by: Simona Vetter <simona.vetter@intel.com>
> Signed-off-by: Simona Vetter <simona.vetter@ffwll.ch>
> ---
>   drivers/gpu/drm/drm_gem.c | 10 +++++++++-
>   include/drm/drm_file.h    |  3 +++
>   2 files changed, 12 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
> index 1e659d2660f7..e4e20dda47b1 100644
> --- a/drivers/gpu/drm/drm_gem.c
> +++ b/drivers/gpu/drm/drm_gem.c
> @@ -279,6 +279,9 @@ drm_gem_object_release_handle(int id, void *ptr, void *data)
>   	struct drm_file *file_priv = data;
>   	struct drm_gem_object *obj = ptr;
>   
> +	if (WARN_ON(!data))
> +		return 0;
> +
>   	if (obj->funcs->close)
>   		obj->funcs->close(obj, file_priv);
>   
> @@ -399,7 +402,7 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
>   	idr_preload(GFP_KERNEL);
>   	spin_lock(&file_priv->table_lock);
>   
> -	ret = idr_alloc(&file_priv->object_idr, obj, 1, 0, GFP_NOWAIT);
> +	ret = idr_alloc(&file_priv->object_idr, NULL, 1, 0, GFP_NOWAIT);
>   
>   	spin_unlock(&file_priv->table_lock);
>   	idr_preload_end();
> @@ -420,6 +423,11 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
>   			goto err_revoke;
>   	}
>   
> +	/* mirrors drm_gem_handle_delete to avoid races */
> +	spin_lock(&file_priv->table_lock);
> +	obj = idr_replace(&file_priv->object_idr, obj, handle);
> +	WARN_ON(obj != NULL);

A DRM print function would be preferable. The obj here is an errno 
pointer. Should the errno code be part of the error message?

If it fails, why does the function still succeed?

Best regards
Thomas

> +	spin_unlock(&file_priv->table_lock);
>   	*handlep = handle;
>   	return 0;
>   
> diff --git a/include/drm/drm_file.h b/include/drm/drm_file.h
> index 5c3b2aa3e69d..d344d41e6cfe 100644
> --- a/include/drm/drm_file.h
> +++ b/include/drm/drm_file.h
> @@ -300,6 +300,9 @@ struct drm_file {
>   	 *
>   	 * Mapping of mm object handles to object pointers. Used by the GEM
>   	 * subsystem. Protected by @table_lock.
> +	 *
> +	 * Note that allocated entries might be NULL as a transient state when
> +	 * creating or deleting a handle.
>   	 */
>   	struct idr object_idr;
>   

-- 
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)

Re: [PATCH 1/8] drm/gem: Fix race in drm_gem_handle_create_tail()

[Simona Vetter]

On Mon, Jun 02, 2025 at 05:15:58PM +0200, Thomas Zimmermann wrote:
> Hi
> 
> Am 28.05.25 um 11:12 schrieb Simona Vetter:
> > Object creation is a careful dance where we must guarantee that the
> > object is fully constructed before it is visible to other threads, and
> > GEM buffer objects are no difference.
> > 
> > Final publishing happens by calling drm_gem_handle_create(). After
> > that the only allowed thing to do is call drm_gem_object_put() because
> > a concurrent call to the GEM_CLOSE ioctl with a correctly guessed id
> > (which is trivial since we have a linear allocator) can already tear
> > down the object again.
> > 
> > Luckily most drivers get this right, the very few exceptions I've
> > pinged the relevant maintainers for. Unfortunately we also need
> > drm_gem_handle_create() when creating additional handles for an
> > already existing object (e.g. GETFB ioctl or the various bo import
> > ioctl), and hence we cannot have a drm_gem_handle_create_and_put() as
> > the only exported function to stop these issues from happening.
> > 
> > Now unfortunately the implementation of drm_gem_handle_create() isn't
> > living up to standards: It does correctly finishe object
> > initialization at the global level, and hence is safe against a
> > concurrent tear down. But it also sets up the file-private aspects of
> > the handle, and that part goes wrong: We fully register the object in
> > the drm_file.object_idr before calling drm_vma_node_allow() or
> > obj->funcs->open, which opens up races against concurrent removal of
> > that handle in drm_gem_handle_delete().
> > 
> > Fix this with the usual two-stage approach of first reserving the
> > handle id, and then only registering the object after we've completed
> > the file-private setup.
> > 
> > Jacek reported this with a testcase of concurrently calling GEM_CLOSE
> > on a freshly-created object (which also destroys the object), but it
> > should be possible to hit this with just additional handles created
> > through import or GETFB without completed destroying the underlying
> > object with the concurrent GEM_CLOSE ioctl calls.
> > 
> > Note that the close-side of this race was fixed in f6cd7daecff5 ("drm:
> > Release driver references to handle before making it available
> > again"), which means a cool 9 years have passed until someone noticed
> > that we need to make this symmetry or there's still gaps left :-/
> > Without the 2-stage close approach we'd still have a race, therefore
> > that's an integral part of this bugfix.
> > 
> > More importantly, this means we can have NULL pointers behind
> > allocated id in our drm_file.object_idr. We need to check for that
> > now:
> > 
> > - drm_gem_handle_delete() checks for ERR_OR_NULL already
> > 
> > - drm_gem.c:object_lookup() also chekcs for NULL
> > 
> > - drm_gem_release() should never be called if there's another thread
> >    still existing that could call into an IOCTL that creates a new
> >    handle, so cannot race. For paranoia I added a NULL check to
> >    drm_gem_object_release_handle() though.
> > 
> > - most drivers (etnaviv, i915, msm) are find because they use
> >    idr_find, which maps both ENOENT and NULL to NULL.
> > 
> > - vmgfx is already broken vmw_debugfs_gem_info_show() because NULL
> >    pointers might exist due to drm_gem_handle_delete(). This needs a
> >    separate patch. This is because idr_for_each_entry terminates on the
> >    first NULL entry and so might not iterate over everything.
> > 
> > - similar for amd in amdgpu_debugfs_gem_info_show() and
> >    amdgpu_gem_force_release(). The latter is really questionable though
> >    since it's a best effort hack and there's no way to close all the
> >    races. Needs separate patches.
> > 
> > - xe is really broken because it not uses idr_for_each_entry() but
> >    also drops the drm_file.table_lock, which can wreak the idr iterator
> >    state if you're unlucky enough. Maybe another reason to look into
> >    the drm fdinfo memory stats instead of hand-rolling too much.
> > 
> > - drm_show_memory_stats() is also broken since it uses
> >    idr_for_each_entry. But since that's a preexisting bug I'll follow
> >    up with a separate patch.
> > 
> > Reported-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
> > Cc: stable@vger.kernel.org
> > Cc: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
> > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> > Cc: Maxime Ripard <mripard@kernel.org>
> > Cc: Thomas Zimmermann <tzimmermann@suse.de>
> > Cc: David Airlie <airlied@gmail.com>
> > Cc: Simona Vetter <simona@ffwll.ch>
> > Signed-off-by: Simona Vetter <simona.vetter@intel.com>
> > Signed-off-by: Simona Vetter <simona.vetter@ffwll.ch>
> > ---
> >   drivers/gpu/drm/drm_gem.c | 10 +++++++++-
> >   include/drm/drm_file.h    |  3 +++
> >   2 files changed, 12 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
> > index 1e659d2660f7..e4e20dda47b1 100644
> > --- a/drivers/gpu/drm/drm_gem.c
> > +++ b/drivers/gpu/drm/drm_gem.c
> > @@ -279,6 +279,9 @@ drm_gem_object_release_handle(int id, void *ptr, void *data)
> >   	struct drm_file *file_priv = data;
> >   	struct drm_gem_object *obj = ptr;
> > +	if (WARN_ON(!data))
> > +		return 0;
> > +
> >   	if (obj->funcs->close)
> >   		obj->funcs->close(obj, file_priv);
> > @@ -399,7 +402,7 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
> >   	idr_preload(GFP_KERNEL);
> >   	spin_lock(&file_priv->table_lock);
> > -	ret = idr_alloc(&file_priv->object_idr, obj, 1, 0, GFP_NOWAIT);
> > +	ret = idr_alloc(&file_priv->object_idr, NULL, 1, 0, GFP_NOWAIT);
> >   	spin_unlock(&file_priv->table_lock);
> >   	idr_preload_end();
> > @@ -420,6 +423,11 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
> >   			goto err_revoke;
> >   	}
> > +	/* mirrors drm_gem_handle_delete to avoid races */
> > +	spin_lock(&file_priv->table_lock);
> > +	obj = idr_replace(&file_priv->object_idr, obj, handle);
> > +	WARN_ON(obj != NULL);
> 
> A DRM print function would be preferable. The obj here is an errno pointer.
> Should the errno code be part of the error message?
> 
> If it fails, why does the function still succeed?

This is an internal error that should never happen, at that point just
bailing out is the way to go.

Also note that the error code here is just to satisfy the function
signature that id_for_each expects, we don't look at it ever (since if
there's no bugs, it should never fail). I learned this because I actually
removed the int return value and stuff didn't compile :-)

I can use drm_WARN_ON if you want me to though?

I'll also explain this in the commit message for the next round.
-Sima

> 
> Best regards
> Thomas
> 
> > +	spin_unlock(&file_priv->table_lock);
> >   	*handlep = handle;
> >   	return 0;
> > diff --git a/include/drm/drm_file.h b/include/drm/drm_file.h
> > index 5c3b2aa3e69d..d344d41e6cfe 100644
> > --- a/include/drm/drm_file.h
> > +++ b/include/drm/drm_file.h
> > @@ -300,6 +300,9 @@ struct drm_file {
> >   	 *
> >   	 * Mapping of mm object handles to object pointers. Used by the GEM
> >   	 * subsystem. Protected by @table_lock.
> > +	 *
> > +	 * Note that allocated entries might be NULL as a transient state when
> > +	 * creating or deleting a handle.
> >   	 */
> >   	struct idr object_idr;
> 
> -- 
> --
> Thomas Zimmermann
> Graphics Driver Developer
> SUSE Software Solutions Germany GmbH
> Frankenstrasse 146, 90461 Nuernberg, Germany
> GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
> HRB 36809 (AG Nuernberg)
> 

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

Re: [PATCH 1/8] drm/gem: Fix race in drm_gem_handle_create_tail()

[Thomas Zimmermann]

Hi

Am 03.06.25 um 13:45 schrieb Simona Vetter:
> On Mon, Jun 02, 2025 at 05:15:58PM +0200, Thomas Zimmermann wrote:
>> Hi
>>
>> Am 28.05.25 um 11:12 schrieb Simona Vetter:
>>> Object creation is a careful dance where we must guarantee that the
>>> object is fully constructed before it is visible to other threads, and
>>> GEM buffer objects are no difference.
>>>
>>> Final publishing happens by calling drm_gem_handle_create(). After
>>> that the only allowed thing to do is call drm_gem_object_put() because
>>> a concurrent call to the GEM_CLOSE ioctl with a correctly guessed id
>>> (which is trivial since we have a linear allocator) can already tear
>>> down the object again.
>>>
>>> Luckily most drivers get this right, the very few exceptions I've
>>> pinged the relevant maintainers for. Unfortunately we also need
>>> drm_gem_handle_create() when creating additional handles for an
>>> already existing object (e.g. GETFB ioctl or the various bo import
>>> ioctl), and hence we cannot have a drm_gem_handle_create_and_put() as
>>> the only exported function to stop these issues from happening.
>>>
>>> Now unfortunately the implementation of drm_gem_handle_create() isn't
>>> living up to standards: It does correctly finishe object
>>> initialization at the global level, and hence is safe against a
>>> concurrent tear down. But it also sets up the file-private aspects of
>>> the handle, and that part goes wrong: We fully register the object in
>>> the drm_file.object_idr before calling drm_vma_node_allow() or
>>> obj->funcs->open, which opens up races against concurrent removal of
>>> that handle in drm_gem_handle_delete().
>>>
>>> Fix this with the usual two-stage approach of first reserving the
>>> handle id, and then only registering the object after we've completed
>>> the file-private setup.
>>>
>>> Jacek reported this with a testcase of concurrently calling GEM_CLOSE
>>> on a freshly-created object (which also destroys the object), but it
>>> should be possible to hit this with just additional handles created
>>> through import or GETFB without completed destroying the underlying
>>> object with the concurrent GEM_CLOSE ioctl calls.
>>>
>>> Note that the close-side of this race was fixed in f6cd7daecff5 ("drm:
>>> Release driver references to handle before making it available
>>> again"), which means a cool 9 years have passed until someone noticed
>>> that we need to make this symmetry or there's still gaps left :-/
>>> Without the 2-stage close approach we'd still have a race, therefore
>>> that's an integral part of this bugfix.
>>>
>>> More importantly, this means we can have NULL pointers behind
>>> allocated id in our drm_file.object_idr. We need to check for that
>>> now:
>>>
>>> - drm_gem_handle_delete() checks for ERR_OR_NULL already
>>>
>>> - drm_gem.c:object_lookup() also chekcs for NULL
>>>
>>> - drm_gem_release() should never be called if there's another thread
>>>     still existing that could call into an IOCTL that creates a new
>>>     handle, so cannot race. For paranoia I added a NULL check to
>>>     drm_gem_object_release_handle() though.
>>>
>>> - most drivers (etnaviv, i915, msm) are find because they use
>>>     idr_find, which maps both ENOENT and NULL to NULL.
>>>
>>> - vmgfx is already broken vmw_debugfs_gem_info_show() because NULL
>>>     pointers might exist due to drm_gem_handle_delete(). This needs a
>>>     separate patch. This is because idr_for_each_entry terminates on the
>>>     first NULL entry and so might not iterate over everything.
>>>
>>> - similar for amd in amdgpu_debugfs_gem_info_show() and
>>>     amdgpu_gem_force_release(). The latter is really questionable though
>>>     since it's a best effort hack and there's no way to close all the
>>>     races. Needs separate patches.
>>>
>>> - xe is really broken because it not uses idr_for_each_entry() but
>>>     also drops the drm_file.table_lock, which can wreak the idr iterator
>>>     state if you're unlucky enough. Maybe another reason to look into
>>>     the drm fdinfo memory stats instead of hand-rolling too much.
>>>
>>> - drm_show_memory_stats() is also broken since it uses
>>>     idr_for_each_entry. But since that's a preexisting bug I'll follow
>>>     up with a separate patch.
>>>
>>> Reported-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
>>> Cc: stable@vger.kernel.org
>>> Cc: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
>>> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
>>> Cc: Maxime Ripard <mripard@kernel.org>
>>> Cc: Thomas Zimmermann <tzimmermann@suse.de>
>>> Cc: David Airlie <airlied@gmail.com>
>>> Cc: Simona Vetter <simona@ffwll.ch>
>>> Signed-off-by: Simona Vetter <simona.vetter@intel.com>
>>> Signed-off-by: Simona Vetter <simona.vetter@ffwll.ch>
>>> ---
>>>    drivers/gpu/drm/drm_gem.c | 10 +++++++++-
>>>    include/drm/drm_file.h    |  3 +++
>>>    2 files changed, 12 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
>>> index 1e659d2660f7..e4e20dda47b1 100644
>>> --- a/drivers/gpu/drm/drm_gem.c
>>> +++ b/drivers/gpu/drm/drm_gem.c
>>> @@ -279,6 +279,9 @@ drm_gem_object_release_handle(int id, void *ptr, void *data)
>>>    	struct drm_file *file_priv = data;
>>>    	struct drm_gem_object *obj = ptr;
>>> +	if (WARN_ON(!data))
>>> +		return 0;
>>> +
>>>    	if (obj->funcs->close)
>>>    		obj->funcs->close(obj, file_priv);
>>> @@ -399,7 +402,7 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
>>>    	idr_preload(GFP_KERNEL);
>>>    	spin_lock(&file_priv->table_lock);
>>> -	ret = idr_alloc(&file_priv->object_idr, obj, 1, 0, GFP_NOWAIT);
>>> +	ret = idr_alloc(&file_priv->object_idr, NULL, 1, 0, GFP_NOWAIT);
>>>    	spin_unlock(&file_priv->table_lock);
>>>    	idr_preload_end();
>>> @@ -420,6 +423,11 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
>>>    			goto err_revoke;
>>>    	}
>>> +	/* mirrors drm_gem_handle_delete to avoid races */
>>> +	spin_lock(&file_priv->table_lock);
>>> +	obj = idr_replace(&file_priv->object_idr, obj, handle);
>>> +	WARN_ON(obj != NULL);
>> A DRM print function would be preferable. The obj here is an errno pointer.
>> Should the errno code be part of the error message?
>>
>> If it fails, why does the function still succeed?
> This is an internal error that should never happen, at that point just
> bailing out is the way to go.
>
> Also note that the error code here is just to satisfy the function
> signature that id_for_each expects, we don't look at it ever (since if
> there's no bugs, it should never fail). I learned this because I actually
> removed the int return value and stuff didn't compile :-)

I see.

>
> I can use drm_WARN_ON if you want me to though?

If you use drm_WARN_ON, you can add

Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>

Best regards
Thomas

>
> I'll also explain this in the commit message for the next round.
> -Sima
>
>> Best regards
>> Thomas
>>
>>> +	spin_unlock(&file_priv->table_lock);
>>>    	*handlep = handle;
>>>    	return 0;
>>> diff --git a/include/drm/drm_file.h b/include/drm/drm_file.h
>>> index 5c3b2aa3e69d..d344d41e6cfe 100644
>>> --- a/include/drm/drm_file.h
>>> +++ b/include/drm/drm_file.h
>>> @@ -300,6 +300,9 @@ struct drm_file {
>>>    	 *
>>>    	 * Mapping of mm object handles to object pointers. Used by the GEM
>>>    	 * subsystem. Protected by @table_lock.
>>> +	 *
>>> +	 * Note that allocated entries might be NULL as a transient state when
>>> +	 * creating or deleting a handle.
>>>    	 */
>>>    	struct idr object_idr;
>> -- 
>> --
>> Thomas Zimmermann
>> Graphics Driver Developer
>> SUSE Software Solutions Germany GmbH
>> Frankenstrasse 146, 90461 Nuernberg, Germany
>> GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
>> HRB 36809 (AG Nuernberg)
>>

-- 
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)

Re: [PATCH 1/8] drm/gem: Fix race in drm_gem_handle_create_tail()

[Simona Vetter]

On Tue, Jun 03, 2025 at 01:45:54PM +0200, Simona Vetter wrote:
> On Mon, Jun 02, 2025 at 05:15:58PM +0200, Thomas Zimmermann wrote:
> > Hi
> > 
> > Am 28.05.25 um 11:12 schrieb Simona Vetter:
> > > Object creation is a careful dance where we must guarantee that the
> > > object is fully constructed before it is visible to other threads, and
> > > GEM buffer objects are no difference.
> > > 
> > > Final publishing happens by calling drm_gem_handle_create(). After
> > > that the only allowed thing to do is call drm_gem_object_put() because
> > > a concurrent call to the GEM_CLOSE ioctl with a correctly guessed id
> > > (which is trivial since we have a linear allocator) can already tear
> > > down the object again.
> > > 
> > > Luckily most drivers get this right, the very few exceptions I've
> > > pinged the relevant maintainers for. Unfortunately we also need
> > > drm_gem_handle_create() when creating additional handles for an
> > > already existing object (e.g. GETFB ioctl or the various bo import
> > > ioctl), and hence we cannot have a drm_gem_handle_create_and_put() as
> > > the only exported function to stop these issues from happening.
> > > 
> > > Now unfortunately the implementation of drm_gem_handle_create() isn't
> > > living up to standards: It does correctly finishe object
> > > initialization at the global level, and hence is safe against a
> > > concurrent tear down. But it also sets up the file-private aspects of
> > > the handle, and that part goes wrong: We fully register the object in
> > > the drm_file.object_idr before calling drm_vma_node_allow() or
> > > obj->funcs->open, which opens up races against concurrent removal of
> > > that handle in drm_gem_handle_delete().
> > > 
> > > Fix this with the usual two-stage approach of first reserving the
> > > handle id, and then only registering the object after we've completed
> > > the file-private setup.
> > > 
> > > Jacek reported this with a testcase of concurrently calling GEM_CLOSE
> > > on a freshly-created object (which also destroys the object), but it
> > > should be possible to hit this with just additional handles created
> > > through import or GETFB without completed destroying the underlying
> > > object with the concurrent GEM_CLOSE ioctl calls.
> > > 
> > > Note that the close-side of this race was fixed in f6cd7daecff5 ("drm:
> > > Release driver references to handle before making it available
> > > again"), which means a cool 9 years have passed until someone noticed
> > > that we need to make this symmetry or there's still gaps left :-/
> > > Without the 2-stage close approach we'd still have a race, therefore
> > > that's an integral part of this bugfix.
> > > 
> > > More importantly, this means we can have NULL pointers behind
> > > allocated id in our drm_file.object_idr. We need to check for that
> > > now:
> > > 
> > > - drm_gem_handle_delete() checks for ERR_OR_NULL already
> > > 
> > > - drm_gem.c:object_lookup() also chekcs for NULL
> > > 
> > > - drm_gem_release() should never be called if there's another thread
> > >    still existing that could call into an IOCTL that creates a new
> > >    handle, so cannot race. For paranoia I added a NULL check to
> > >    drm_gem_object_release_handle() though.
> > > 
> > > - most drivers (etnaviv, i915, msm) are find because they use
> > >    idr_find, which maps both ENOENT and NULL to NULL.
> > > 
> > > - vmgfx is already broken vmw_debugfs_gem_info_show() because NULL
> > >    pointers might exist due to drm_gem_handle_delete(). This needs a
> > >    separate patch. This is because idr_for_each_entry terminates on the
> > >    first NULL entry and so might not iterate over everything.
> > > 
> > > - similar for amd in amdgpu_debugfs_gem_info_show() and
> > >    amdgpu_gem_force_release(). The latter is really questionable though
> > >    since it's a best effort hack and there's no way to close all the
> > >    races. Needs separate patches.
> > > 
> > > - xe is really broken because it not uses idr_for_each_entry() but
> > >    also drops the drm_file.table_lock, which can wreak the idr iterator
> > >    state if you're unlucky enough. Maybe another reason to look into
> > >    the drm fdinfo memory stats instead of hand-rolling too much.
> > > 
> > > - drm_show_memory_stats() is also broken since it uses
> > >    idr_for_each_entry. But since that's a preexisting bug I'll follow
> > >    up with a separate patch.
> > > 
> > > Reported-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
> > > Cc: stable@vger.kernel.org
> > > Cc: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
> > > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> > > Cc: Maxime Ripard <mripard@kernel.org>
> > > Cc: Thomas Zimmermann <tzimmermann@suse.de>
> > > Cc: David Airlie <airlied@gmail.com>
> > > Cc: Simona Vetter <simona@ffwll.ch>
> > > Signed-off-by: Simona Vetter <simona.vetter@intel.com>
> > > Signed-off-by: Simona Vetter <simona.vetter@ffwll.ch>
> > > ---
> > >   drivers/gpu/drm/drm_gem.c | 10 +++++++++-
> > >   include/drm/drm_file.h    |  3 +++
> > >   2 files changed, 12 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
> > > index 1e659d2660f7..e4e20dda47b1 100644
> > > --- a/drivers/gpu/drm/drm_gem.c
> > > +++ b/drivers/gpu/drm/drm_gem.c
> > > @@ -279,6 +279,9 @@ drm_gem_object_release_handle(int id, void *ptr, void *data)
> > >   	struct drm_file *file_priv = data;
> > >   	struct drm_gem_object *obj = ptr;
> > > +	if (WARN_ON(!data))
> > > +		return 0;
> > > +
> > >   	if (obj->funcs->close)
> > >   		obj->funcs->close(obj, file_priv);
> > > @@ -399,7 +402,7 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
> > >   	idr_preload(GFP_KERNEL);
> > >   	spin_lock(&file_priv->table_lock);
> > > -	ret = idr_alloc(&file_priv->object_idr, obj, 1, 0, GFP_NOWAIT);
> > > +	ret = idr_alloc(&file_priv->object_idr, NULL, 1, 0, GFP_NOWAIT);
> > >   	spin_unlock(&file_priv->table_lock);
> > >   	idr_preload_end();
> > > @@ -420,6 +423,11 @@ drm_gem_handle_create_tail(struct drm_file *file_priv,
> > >   			goto err_revoke;
> > >   	}
> > > +	/* mirrors drm_gem_handle_delete to avoid races */
> > > +	spin_lock(&file_priv->table_lock);
> > > +	obj = idr_replace(&file_priv->object_idr, obj, handle);
> > > +	WARN_ON(obj != NULL);
> > 
> > A DRM print function would be preferable. The obj here is an errno pointer.
> > Should the errno code be part of the error message?
> > 
> > If it fails, why does the function still succeed?
> 
> This is an internal error that should never happen, at that point just
> bailing out is the way to go.
> 
> Also note that the error code here is just to satisfy the function
> signature that id_for_each expects, we don't look at it ever (since if
> there's no bugs, it should never fail). I learned this because I actually
> removed the int return value and stuff didn't compile :-)

Ok this part was nonsense, I mixed it up with handle_delete(). I still
don't think we should return an error code here, because we've
successfully installed the handle. It's just that something happened with
the idr that should be impossible, so all bets are off.
-Sima

> I can use drm_WARN_ON if you want me to though?
> 
> I'll also explain this in the commit message for the next round.
> -Sima
> 
> > 
> > Best regards
> > Thomas
> > 
> > > +	spin_unlock(&file_priv->table_lock);
> > >   	*handlep = handle;
> > >   	return 0;
> > > diff --git a/include/drm/drm_file.h b/include/drm/drm_file.h
> > > index 5c3b2aa3e69d..d344d41e6cfe 100644
> > > --- a/include/drm/drm_file.h
> > > +++ b/include/drm/drm_file.h
> > > @@ -300,6 +300,9 @@ struct drm_file {
> > >   	 *
> > >   	 * Mapping of mm object handles to object pointers. Used by the GEM
> > >   	 * subsystem. Protected by @table_lock.
> > > +	 *
> > > +	 * Note that allocated entries might be NULL as a transient state when
> > > +	 * creating or deleting a handle.
> > >   	 */
> > >   	struct idr object_idr;
> > 
> > -- 
> > --
> > Thomas Zimmermann
> > Graphics Driver Developer
> > SUSE Software Solutions Germany GmbH
> > Frankenstrasse 146, 90461 Nuernberg, Germany
> > GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
> > HRB 36809 (AG Nuernberg)
> > 
> 
> -- 
> Simona Vetter
> Software Engineer, Intel Corporation
> http://blog.ffwll.ch

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