[PATCH 20/27] drm/radeon: rework recursive gpu reset handling

[j.glisse at gmail.com]

From: Christian König <deathsimple at vodafone.de>

Instead of all this humpy pumpy with recursive
mutex (which also fixes only halve of the problem)
move the actual gpu reset out of the fence code,
return -EDEADLK and then reset the gpu in the
calling ioctl function.

v2: Split removal of radeon_mutex into separate patch.
    Return -EAGAIN if reset is successful.

Signed-off-by: Christian König <deathsimple at vodafone.de>
Reviewed-by: Jerome Glisse <jglisse at redhat.com>
---
 drivers/gpu/drm/radeon/radeon_cs.c     |   13 +++++++++++++
 drivers/gpu/drm/radeon/radeon_device.c |    5 -----
 drivers/gpu/drm/radeon/radeon_fence.c  |   10 +++-------
 drivers/gpu/drm/radeon/radeon_gem.c    |   16 ++++++++++++++++
 4 files changed, 32 insertions(+), 12 deletions(-)

diff --git a/drivers/gpu/drm/radeon/radeon_cs.c b/drivers/gpu/drm/radeon/radeon_cs.c
index 64b86e7..a0826bb 100644
--- a/drivers/gpu/drm/radeon/radeon_cs.c
+++ b/drivers/gpu/drm/radeon/radeon_cs.c
@@ -489,6 +489,16 @@ out:
 	return r;
 }
 
+static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
+{
+	if (r == -EDEADLK) {
+		r = radeon_gpu_reset(rdev);
+		if (!r)
+			r = -EAGAIN;
+	}
+	return r;
+}
+
 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 {
 	struct radeon_device *rdev = dev->dev_private;
@@ -510,6 +520,7 @@ int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 	if (r) {
 		DRM_ERROR("Failed to initialize parser !\n");
 		radeon_cs_parser_fini(&parser, r);
+		r = radeon_cs_handle_lockup(rdev, r);
 		radeon_mutex_unlock(&rdev->cs_mutex);
 		return r;
 	}
@@ -518,6 +529,7 @@ int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 		if (r != -ERESTARTSYS)
 			DRM_ERROR("Failed to parse relocation %d!\n", r);
 		radeon_cs_parser_fini(&parser, r);
+		r = radeon_cs_handle_lockup(rdev, r);
 		radeon_mutex_unlock(&rdev->cs_mutex);
 		return r;
 	}
@@ -531,6 +543,7 @@ int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 	}
 out:
 	radeon_cs_parser_fini(&parser, r);
+	r = radeon_cs_handle_lockup(rdev, r);
 	radeon_mutex_unlock(&rdev->cs_mutex);
 	return r;
 }
diff --git a/drivers/gpu/drm/radeon/radeon_device.c b/drivers/gpu/drm/radeon/radeon_device.c
index 26fb9da..1dac27d 100644
--- a/drivers/gpu/drm/radeon/radeon_device.c
+++ b/drivers/gpu/drm/radeon/radeon_device.c
@@ -984,9 +984,6 @@ int radeon_gpu_reset(struct radeon_device *rdev)
 	int r;
 	int resched;
 
-	/* Prevent CS ioctl from interfering */
-	radeon_mutex_lock(&rdev->cs_mutex);
-
 	radeon_save_bios_scratch_regs(rdev);
 	/* block TTM */
 	resched = ttm_bo_lock_delayed_workqueue(&rdev->mman.bdev);
@@ -1001,8 +998,6 @@ int radeon_gpu_reset(struct radeon_device *rdev)
 		ttm_bo_unlock_delayed_workqueue(&rdev->mman.bdev, resched);
 	}
 
-	radeon_mutex_unlock(&rdev->cs_mutex);
-
 	if (r) {
 		/* bad news, how to tell it to userspace ? */
 		dev_info(rdev->dev, "GPU reset failed\n");
diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c
index 326bdb5..24a4d4b 100644
--- a/drivers/gpu/drm/radeon/radeon_fence.c
+++ b/drivers/gpu/drm/radeon/radeon_fence.c
@@ -264,6 +264,8 @@ int radeon_fence_wait(struct radeon_fence *fence, bool intr)
 			/* change sequence value on all rings, so nobody else things there is a lockup */
 			for (i = 0; i < RADEON_NUM_RINGS; ++i)
 				rdev->fence_drv[i].last_seq -= 0x10000;
+
+			rdev->fence_drv[fence->ring].last_activity = jiffies;
 			write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
 
 			if (radeon_ring_is_lockup(rdev, fence->ring, &rdev->ring[fence->ring])) {
@@ -274,13 +276,7 @@ int radeon_fence_wait(struct radeon_fence *fence, bool intr)
 
 				/* mark the ring as not ready any more */
 				rdev->ring[fence->ring].ready = false;
-				r = radeon_gpu_reset(rdev);
-				if (r)
-					return r;
-
-				write_lock_irqsave(&rdev->fence_lock, irq_flags);
-				rdev->fence_drv[fence->ring].last_activity = jiffies;
-				write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
+				return -EDEADLK;
 			}
 		}
 	}
diff --git a/drivers/gpu/drm/radeon/radeon_gem.c b/drivers/gpu/drm/radeon/radeon_gem.c
index c7008b5..e15cb1f 100644
--- a/drivers/gpu/drm/radeon/radeon_gem.c
+++ b/drivers/gpu/drm/radeon/radeon_gem.c
@@ -154,6 +154,17 @@ void radeon_gem_object_close(struct drm_gem_object *obj,
 	radeon_bo_unreserve(rbo);
 }
 
+static int radeon_gem_handle_lockup(struct radeon_device *rdev, int r)
+{
+	if (r == -EDEADLK) {
+		radeon_mutex_lock(&rdev->cs_mutex);
+		r = radeon_gpu_reset(rdev);
+		if (!r)
+			r = -EAGAIN;
+		radeon_mutex_unlock(&rdev->cs_mutex);
+	}
+	return r;
+}
 
 /*
  * GEM ioctls.
@@ -210,12 +221,14 @@ int radeon_gem_create_ioctl(struct drm_device *dev, void *data,
 					args->initial_domain, false,
 					false, &gobj);
 	if (r) {
+		r = radeon_gem_handle_lockup(rdev, r);
 		return r;
 	}
 	r = drm_gem_handle_create(filp, gobj, &handle);
 	/* drop reference from allocate - handle holds it now */
 	drm_gem_object_unreference_unlocked(gobj);
 	if (r) {
+		r = radeon_gem_handle_lockup(rdev, r);
 		return r;
 	}
 	args->handle = handle;
@@ -245,6 +258,7 @@ int radeon_gem_set_domain_ioctl(struct drm_device *dev, void *data,
 	r = radeon_gem_set_domain(gobj, args->read_domains, args->write_domain);
 
 	drm_gem_object_unreference_unlocked(gobj);
+	r = radeon_gem_handle_lockup(robj->rdev, r);
 	return r;
 }
 
@@ -301,6 +315,7 @@ int radeon_gem_busy_ioctl(struct drm_device *dev, void *data,
 		break;
 	}
 	drm_gem_object_unreference_unlocked(gobj);
+	r = radeon_gem_handle_lockup(robj->rdev, r);
 	return r;
 }
 
@@ -322,6 +337,7 @@ int radeon_gem_wait_idle_ioctl(struct drm_device *dev, void *data,
 	if (robj->rdev->asic->ioctl_wait_idle)
 		robj->rdev->asic->ioctl_wait_idle(robj->rdev, robj);
 	drm_gem_object_unreference_unlocked(gobj);
+	r = radeon_gem_handle_lockup(robj->rdev, r);
 	return r;
 }
 
-- 
1.7.7.6