[PATCH] drm: Avoid drm_global_mutex for simple inc/dec of dev->open_count

[Chris Wilson]

Since drm_global_mutex is a true global mutex across devices, we don't
want to acquire it unless absolutely necessary. For maintaining the
device local open_count, we can use atomic operations on the counter
itself, except when making the transition to/from 0. Here, we tackle the
easy portion of delaying acquiring the drm_global_mutex for the final
release by using atomic_dec_and_mutex_lock(), leaving the global
serialisation across the device opens.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Thomas Hellström (VMware) <thomas_os at shipmail.org>
---
atomic_dec_and_mutex_lock needs pairing with mutex_unlock (you fool)
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |  2 +-
 drivers/gpu/drm/drm_file.c                 | 16 ++++++++--------
 drivers/gpu/drm/i915/i915_switcheroo.c     |  2 +-
 drivers/gpu/drm/nouveau/nouveau_vga.c      |  2 +-
 drivers/gpu/drm/radeon/radeon_device.c     |  2 +-
 include/drm/drm_device.h                   |  2 +-
 6 files changed, 13 insertions(+), 13 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 53d882000101..c3c0356dfa61 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -1136,7 +1136,7 @@ static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
 	* locking inversion with the driver load path. And the access here is
 	* completely racy anyway. So don't bother with locking for now.
 	*/
-	return dev->open_count == 0;
+	return atomic_read(&dev->open_count) == 0;
 }
 
 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c
index e25306c49cc6..1075b3a8b5b1 100644
--- a/drivers/gpu/drm/drm_file.c
+++ b/drivers/gpu/drm/drm_file.c
@@ -220,7 +220,7 @@ void drm_file_free(struct drm_file *file)
 	DRM_DEBUG("pid = %d, device = 0x%lx, open_count = %d\n",
 		  task_pid_nr(current),
 		  (long)old_encode_dev(file->minor->kdev->devt),
-		  READ_ONCE(dev->open_count));
+		  atomic_read(&dev->open_count));
 
 	if (drm_core_check_feature(dev, DRIVER_LEGACY) &&
 	    dev->driver->preclose)
@@ -379,7 +379,7 @@ int drm_open(struct inode *inode, struct file *filp)
 		return PTR_ERR(minor);
 
 	dev = minor->dev;
-	if (!dev->open_count++)
+	if (!atomic_fetch_inc(&dev->open_count))
 		need_setup = 1;
 
 	/* share address_space across all char-devs of a single device */
@@ -398,7 +398,7 @@ int drm_open(struct inode *inode, struct file *filp)
 	return 0;
 
 err_undo:
-	dev->open_count--;
+	atomic_dec(&dev->open_count);
 	drm_minor_release(minor);
 	return retcode;
 }
@@ -440,11 +440,11 @@ int drm_release(struct inode *inode, struct file *filp)
 
 	mutex_lock(&drm_global_mutex);
 
-	DRM_DEBUG("open_count = %d\n", dev->open_count);
+	DRM_DEBUG("open_count = %d\n", atomic_read(&dev->open_count));
 
 	drm_close_helper(filp);
 
-	if (!--dev->open_count)
+	if (atomic_dec_and_test(&dev->open_count))
 		drm_lastclose(dev);
 
 	mutex_unlock(&drm_global_mutex);
@@ -478,10 +478,10 @@ int drm_release_noglobal(struct inode *inode, struct file *filp)
 
 	drm_close_helper(filp);
 
-	mutex_lock(&drm_global_mutex);
-	if (!--dev->open_count)
+	if (atomic_dec_and_mutex_lock(&dev->open_count, &drm_global_mutex)) {
 		drm_lastclose(dev);
-	mutex_unlock(&drm_global_mutex);
+		mutex_unlock(&drm_global_mutex);
+	}
 
 	drm_minor_release(minor);
 
diff --git a/drivers/gpu/drm/i915/i915_switcheroo.c b/drivers/gpu/drm/i915/i915_switcheroo.c
index 39c79e1c5b52..ed69b5d4a375 100644
--- a/drivers/gpu/drm/i915/i915_switcheroo.c
+++ b/drivers/gpu/drm/i915/i915_switcheroo.c
@@ -43,7 +43,7 @@ static bool i915_switcheroo_can_switch(struct pci_dev *pdev)
 	 * locking inversion with the driver load path. And the access here is
 	 * completely racy anyway. So don't bother with locking for now.
 	 */
-	return i915 && i915->drm.open_count == 0;
+	return i915 && atomic_read(&i915->drm.open_count) == 0;
 }
 
 static const struct vga_switcheroo_client_ops i915_switcheroo_ops = {
diff --git a/drivers/gpu/drm/nouveau/nouveau_vga.c b/drivers/gpu/drm/nouveau/nouveau_vga.c
index d865d8aeac3c..c85dd8afa3c3 100644
--- a/drivers/gpu/drm/nouveau/nouveau_vga.c
+++ b/drivers/gpu/drm/nouveau/nouveau_vga.c
@@ -72,7 +72,7 @@ nouveau_switcheroo_can_switch(struct pci_dev *pdev)
 	 * locking inversion with the driver load path. And the access here is
 	 * completely racy anyway. So don't bother with locking for now.
 	 */
-	return dev->open_count == 0;
+	return atomic_read(&dev->open_count) == 0;
 }
 
 static const struct vga_switcheroo_client_ops
diff --git a/drivers/gpu/drm/radeon/radeon_device.c b/drivers/gpu/drm/radeon/radeon_device.c
index a522e092038b..266e3cbbd09b 100644
--- a/drivers/gpu/drm/radeon/radeon_device.c
+++ b/drivers/gpu/drm/radeon/radeon_device.c
@@ -1263,7 +1263,7 @@ static bool radeon_switcheroo_can_switch(struct pci_dev *pdev)
 	 * locking inversion with the driver load path. And the access here is
 	 * completely racy anyway. So don't bother with locking for now.
 	 */
-	return dev->open_count == 0;
+	return atomic_read(&dev->open_count) == 0;
 }
 
 static const struct vga_switcheroo_client_ops radeon_switcheroo_ops = {
diff --git a/include/drm/drm_device.h b/include/drm/drm_device.h
index 1acfc3bbd3fb..bb60a949f416 100644
--- a/include/drm/drm_device.h
+++ b/include/drm/drm_device.h
@@ -144,7 +144,7 @@ struct drm_device {
 	 * Usage counter for outstanding files open,
 	 * protected by drm_global_mutex
 	 */
-	int open_count;
+	atomic_t open_count;
 
 	/** @filelist_mutex: Protects @filelist. */
 	struct mutex filelist_mutex;
-- 
2.25.0