drm/radeon: rework page flip handling v4
authorChristian König <christian.koenig@amd.com>
Tue, 3 Jun 2014 22:13:21 +0000 (18:13 -0400)
committerDave Airlie <airlied@redhat.com>
Wed, 4 Jun 2014 01:59:44 +0000 (11:59 +1000)
Instead of trying to flip inside the vblank period when
the buffer is idle, offload blocking for idle to a kernel
thread and program the flip directly into the hardware.

v2: add error handling, fix EBUSY handling
v3: add proper exclusive_lock handling
v4: update crtc->primary->fb when the flip actually happens

Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
drivers/gpu/drm/radeon/radeon.h
drivers/gpu/drm/radeon/radeon_display.c
drivers/gpu/drm/radeon/radeon_mode.h

index 2eaa3300099e4e1c96b0ac87451dcb38de349db3..dd4da88b3ab18ddc2e50e603b4078fc125725d37 100644 (file)
@@ -676,14 +676,16 @@ void radeon_doorbell_free(struct radeon_device *rdev, u32 doorbell);
  * IRQS.
  */
 
-struct radeon_unpin_work {
-       struct work_struct work;
-       struct radeon_device *rdev;
-       int crtc_id;
-       struct radeon_fence *fence;
+struct radeon_flip_work {
+       struct work_struct              flip_work;
+       struct work_struct              unpin_work;
+       struct radeon_device            *rdev;
+       int                             crtc_id;
+       struct drm_framebuffer          *fb;
        struct drm_pending_vblank_event *event;
-       struct radeon_bo *old_rbo;
-       u64 new_crtc_base;
+       struct radeon_bo                *old_rbo;
+       struct radeon_bo                *new_rbo;
+       struct radeon_fence             *fence;
 };
 
 struct r500_irq_stat_regs {
index 88e3cbe11dad7d0c1d13769de890d62a30b71865..d96d1ca74e7064bc75980452c134a7075b2d5f10 100644 (file)
@@ -249,16 +249,21 @@ static void radeon_crtc_destroy(struct drm_crtc *crtc)
        struct radeon_crtc *radeon_crtc = to_radeon_crtc(crtc);
 
        drm_crtc_cleanup(crtc);
+       destroy_workqueue(radeon_crtc->flip_queue);
        kfree(radeon_crtc);
 }
 
-/*
- * Handle unpin events outside the interrupt handler proper.
+/**
+ * radeon_unpin_work_func - unpin old buffer object
+ *
+ * @__work - kernel work item
+ *
+ * Unpin the old frame buffer object outside of the interrupt handler
  */
 static void radeon_unpin_work_func(struct work_struct *__work)
 {
-       struct radeon_unpin_work *work =
-               container_of(__work, struct radeon_unpin_work, work);
+       struct radeon_flip_work *work =
+               container_of(__work, struct radeon_flip_work, unpin_work);
        int r;
 
        /* unpin of the old buffer */
@@ -279,30 +284,19 @@ static void radeon_unpin_work_func(struct work_struct *__work)
 void radeon_crtc_handle_vblank(struct radeon_device *rdev, int crtc_id)
 {
        struct radeon_crtc *radeon_crtc = rdev->mode_info.crtcs[crtc_id];
-       struct radeon_unpin_work *work;
+       struct radeon_flip_work *work;
        unsigned long flags;
        u32 update_pending;
        int vpos, hpos;
 
        spin_lock_irqsave(&rdev->ddev->event_lock, flags);
-       work = radeon_crtc->unpin_work;
-       if (work == NULL ||
-           (work->fence && !radeon_fence_signaled(work->fence))) {
+       work = radeon_crtc->flip_work;
+       if (work == NULL) {
                spin_unlock_irqrestore(&rdev->ddev->event_lock, flags);
                return;
        }
-       /* New pageflip, or just completion of a previous one? */
-       if (!radeon_crtc->deferred_flip_completion) {
-               /* do the flip (mmio) */
-               radeon_page_flip(rdev, crtc_id, work->new_crtc_base);
-               update_pending = radeon_page_flip_pending(rdev, crtc_id);
-       } else {
-               /* This is just a completion of a flip queued in crtc
-                * at last invocation. Make sure we go directly to
-                * completion routine.
-                */
-               update_pending = 0;
-       }
+
+       update_pending = radeon_page_flip_pending(rdev, crtc_id);
 
        /* Has the pageflip already completed in crtc, or is it certain
         * to complete in this vblank?
@@ -320,19 +314,9 @@ void radeon_crtc_handle_vblank(struct radeon_device *rdev, int crtc_id)
                 */
                update_pending = 0;
        }
-       if (update_pending) {
-               /* crtc didn't flip in this target vblank interval,
-                * but flip is pending in crtc. It will complete it
-                * in next vblank interval, so complete the flip at
-                * next vblank irq.
-                */
-               radeon_crtc->deferred_flip_completion = 1;
-               spin_unlock_irqrestore(&rdev->ddev->event_lock, flags);
-               return;
-       } else {
-               spin_unlock_irqrestore(&rdev->ddev->event_lock, flags);
+       spin_unlock_irqrestore(&rdev->ddev->event_lock, flags);
+       if (!update_pending)
                radeon_crtc_handle_flip(rdev, crtc_id);
-       }
 }
 
 /**
@@ -346,7 +330,7 @@ void radeon_crtc_handle_vblank(struct radeon_device *rdev, int crtc_id)
 void radeon_crtc_handle_flip(struct radeon_device *rdev, int crtc_id)
 {
        struct radeon_crtc *radeon_crtc = rdev->mode_info.crtcs[crtc_id];
-       struct radeon_unpin_work *work;
+       struct radeon_flip_work *work;
        unsigned long flags;
 
        /* this can happen at init */
@@ -354,15 +338,14 @@ void radeon_crtc_handle_flip(struct radeon_device *rdev, int crtc_id)
                return;
 
        spin_lock_irqsave(&rdev->ddev->event_lock, flags);
-       work = radeon_crtc->unpin_work;
+       work = radeon_crtc->flip_work;
        if (work == NULL) {
                spin_unlock_irqrestore(&rdev->ddev->event_lock, flags);
                return;
        }
 
-       /* Pageflip (will be) certainly completed in this vblank. Clean up. */
-       radeon_crtc->unpin_work = NULL;
-       radeon_crtc->deferred_flip_completion = 0;
+       /* Pageflip completed. Clean up. */
+       radeon_crtc->flip_work = NULL;
 
        /* wakeup userspace */
        if (work->event)
@@ -372,83 +355,69 @@ void radeon_crtc_handle_flip(struct radeon_device *rdev, int crtc_id)
 
        radeon_fence_unref(&work->fence);
        radeon_irq_kms_pflip_irq_get(rdev, work->crtc_id);
-       schedule_work(&work->work);
+       queue_work(radeon_crtc->flip_queue, &work->unpin_work);
 }
 
-static int radeon_crtc_page_flip(struct drm_crtc *crtc,
-                                struct drm_framebuffer *fb,
-                                struct drm_pending_vblank_event *event,
-                                uint32_t page_flip_flags)
+/**
+ * radeon_flip_work_func - page flip framebuffer
+ *
+ * @work - kernel work item
+ *
+ * Wait for the buffer object to become idle and do the actual page flip
+ */
+static void radeon_flip_work_func(struct work_struct *__work)
 {
-       struct drm_device *dev = crtc->dev;
-       struct radeon_device *rdev = dev->dev_private;
-       struct radeon_crtc *radeon_crtc = to_radeon_crtc(crtc);
-       struct radeon_framebuffer *old_radeon_fb;
-       struct radeon_framebuffer *new_radeon_fb;
-       struct drm_gem_object *obj;
-       struct radeon_bo *rbo;
-       struct radeon_unpin_work *work;
-       unsigned long flags;
-       u32 tiling_flags, pitch_pixels;
-       u64 base;
-       int r;
+       struct radeon_flip_work *work =
+               container_of(__work, struct radeon_flip_work, flip_work);
+       struct radeon_device *rdev = work->rdev;
+       struct radeon_crtc *radeon_crtc = rdev->mode_info.crtcs[work->crtc_id];
 
-       work = kzalloc(sizeof *work, GFP_KERNEL);
-       if (work == NULL)
-               return -ENOMEM;
+       struct drm_crtc *crtc = &radeon_crtc->base;
+       struct drm_framebuffer *fb = work->fb;
 
-       work->event = event;
-       work->rdev = rdev;
-       work->crtc_id = radeon_crtc->crtc_id;
-       old_radeon_fb = to_radeon_framebuffer(crtc->primary->fb);
-       new_radeon_fb = to_radeon_framebuffer(fb);
-       /* schedule unpin of the old buffer */
-       obj = old_radeon_fb->obj;
-       /* take a reference to the old object */
-       drm_gem_object_reference(obj);
-       rbo = gem_to_radeon_bo(obj);
-       work->old_rbo = rbo;
-       obj = new_radeon_fb->obj;
-       rbo = gem_to_radeon_bo(obj);
+       uint32_t tiling_flags, pitch_pixels;
+       uint64_t base;
 
-       spin_lock(&rbo->tbo.bdev->fence_lock);
-       if (rbo->tbo.sync_obj)
-               work->fence = radeon_fence_ref(rbo->tbo.sync_obj);
-       spin_unlock(&rbo->tbo.bdev->fence_lock);
+       unsigned long flags;
+       int r;
 
-       INIT_WORK(&work->work, radeon_unpin_work_func);
+        down_read(&rdev->exclusive_lock);
+       while (work->fence) {
+               r = radeon_fence_wait(work->fence, false);
+               if (r == -EDEADLK) {
+                       up_read(&rdev->exclusive_lock);
+                       r = radeon_gpu_reset(rdev);
+                       down_read(&rdev->exclusive_lock);
+               }
 
-       /* We borrow the event spin lock for protecting unpin_work */
-       spin_lock_irqsave(&dev->event_lock, flags);
-       if (radeon_crtc->unpin_work) {
-               DRM_DEBUG_DRIVER("flip queue: crtc already busy\n");
-               r = -EBUSY;
-               goto unlock_free;
+               if (r) {
+                       DRM_ERROR("failed to wait on page flip fence (%d)!\n",
+                                 r);
+                       goto cleanup;
+               } else
+                       radeon_fence_unref(&work->fence);
        }
-       radeon_crtc->unpin_work = work;
-       radeon_crtc->deferred_flip_completion = 0;
-       spin_unlock_irqrestore(&dev->event_lock, flags);
 
        /* pin the new buffer */
        DRM_DEBUG_DRIVER("flip-ioctl() cur_fbo = %p, cur_bbo = %p\n",
-                        work->old_rbo, rbo);
+                        work->old_rbo, work->new_rbo);
 
-       r = radeon_bo_reserve(rbo, false);
+       r = radeon_bo_reserve(work->new_rbo, false);
        if (unlikely(r != 0)) {
                DRM_ERROR("failed to reserve new rbo buffer before flip\n");
-               goto pflip_cleanup;
+               goto cleanup;
        }
        /* Only 27 bit offset for legacy CRTC */
-       r = radeon_bo_pin_restricted(rbo, RADEON_GEM_DOMAIN_VRAM,
+       r = radeon_bo_pin_restricted(work->new_rbo, RADEON_GEM_DOMAIN_VRAM,
                                     ASIC_IS_AVIVO(rdev) ? 0 : 1 << 27, &base);
        if (unlikely(r != 0)) {
-               radeon_bo_unreserve(rbo);
+               radeon_bo_unreserve(work->new_rbo);
                r = -EINVAL;
                DRM_ERROR("failed to pin new rbo buffer before flip\n");
-               goto pflip_cleanup;
+               goto cleanup;
        }
-       radeon_bo_get_tiling_flags(rbo, &tiling_flags, NULL);
-       radeon_bo_unreserve(rbo);
+       radeon_bo_get_tiling_flags(work->new_rbo, &tiling_flags, NULL);
+       radeon_bo_unreserve(work->new_rbo);
 
        if (!ASIC_IS_AVIVO(rdev)) {
                /* crtc offset is from display base addr not FB location */
@@ -486,9 +455,8 @@ static int radeon_crtc_page_flip(struct drm_crtc *crtc,
                base &= ~7;
        }
 
-       spin_lock_irqsave(&dev->event_lock, flags);
-       work->new_crtc_base = base;
-       spin_unlock_irqrestore(&dev->event_lock, flags);
+       /* We borrow the event spin lock for protecting flip_work */
+       spin_lock_irqsave(&crtc->dev->event_lock, flags);
 
        /* update crtc fb */
        crtc->primary->fb = fb;
@@ -496,18 +464,82 @@ static int radeon_crtc_page_flip(struct drm_crtc *crtc,
        /* set the proper interrupt */
        radeon_irq_kms_pflip_irq_get(rdev, radeon_crtc->crtc_id);
 
-       return 0;
+       /* do the flip (mmio) */
+       radeon_page_flip(rdev, radeon_crtc->crtc_id, base);
+
+       spin_unlock_irqrestore(&crtc->dev->event_lock, flags);
+       up_read(&rdev->exclusive_lock);
+
+       return;
 
-pflip_cleanup:
-       spin_lock_irqsave(&dev->event_lock, flags);
-       radeon_crtc->unpin_work = NULL;
-unlock_free:
-       spin_unlock_irqrestore(&dev->event_lock, flags);
-       drm_gem_object_unreference_unlocked(old_radeon_fb->obj);
+cleanup:
+       drm_gem_object_unreference_unlocked(&work->old_rbo->gem_base);
        radeon_fence_unref(&work->fence);
        kfree(work);
+       up_read(&rdev->exclusive_lock);
+}
+
+static int radeon_crtc_page_flip(struct drm_crtc *crtc,
+                                struct drm_framebuffer *fb,
+                                struct drm_pending_vblank_event *event,
+                                uint32_t page_flip_flags)
+{
+       struct drm_device *dev = crtc->dev;
+       struct radeon_device *rdev = dev->dev_private;
+       struct radeon_crtc *radeon_crtc = to_radeon_crtc(crtc);
+       struct radeon_framebuffer *old_radeon_fb;
+       struct radeon_framebuffer *new_radeon_fb;
+       struct drm_gem_object *obj;
+       struct radeon_flip_work *work;
+       unsigned long flags;
+
+       work = kzalloc(sizeof *work, GFP_KERNEL);
+       if (work == NULL)
+               return -ENOMEM;
+
+       INIT_WORK(&work->flip_work, radeon_flip_work_func);
+       INIT_WORK(&work->unpin_work, radeon_unpin_work_func);
+
+       work->rdev = rdev;
+       work->crtc_id = radeon_crtc->crtc_id;
+       work->fb = fb;
+       work->event = event;
+
+       /* schedule unpin of the old buffer */
+       old_radeon_fb = to_radeon_framebuffer(crtc->primary->fb);
+       obj = old_radeon_fb->obj;
+
+       /* take a reference to the old object */
+       drm_gem_object_reference(obj);
+       work->old_rbo = gem_to_radeon_bo(obj);
+
+       new_radeon_fb = to_radeon_framebuffer(fb);
+       obj = new_radeon_fb->obj;
+       work->new_rbo = gem_to_radeon_bo(obj);
+
+       spin_lock(&work->new_rbo->tbo.bdev->fence_lock);
+       if (work->new_rbo->tbo.sync_obj)
+               work->fence = radeon_fence_ref(work->new_rbo->tbo.sync_obj);
+       spin_unlock(&work->new_rbo->tbo.bdev->fence_lock);
+
+       /* We borrow the event spin lock for protecting flip_work */
+       spin_lock_irqsave(&crtc->dev->event_lock, flags);
 
-       return r;
+       if (radeon_crtc->flip_work) {
+               DRM_DEBUG_DRIVER("flip queue: crtc already busy\n");
+               spin_unlock_irqrestore(&crtc->dev->event_lock, flags);
+               drm_gem_object_unreference_unlocked(&work->old_rbo->gem_base);
+               radeon_fence_unref(&work->fence);
+               kfree(work);
+               return -EBUSY;
+       }
+       radeon_crtc->flip_work = work;
+
+       spin_unlock_irqrestore(&crtc->dev->event_lock, flags);
+
+       queue_work(radeon_crtc->flip_queue, &work->flip_work);
+
+       return 0;
 }
 
 static int
@@ -577,6 +609,7 @@ static void radeon_crtc_init(struct drm_device *dev, int index)
 
        drm_mode_crtc_set_gamma_size(&radeon_crtc->base, 256);
        radeon_crtc->crtc_id = index;
+       radeon_crtc->flip_queue = create_singlethread_workqueue("radeon-crtc");
        rdev->mode_info.crtcs[index] = radeon_crtc;
 
        if (rdev->family >= CHIP_BONAIRE) {
index 718be1a932ac6b757cce6cd6027b050161af8006..ea72ad889a110c5a678b4028a7721da066b02978 100644 (file)
@@ -325,8 +325,8 @@ struct radeon_crtc {
        struct drm_display_mode native_mode;
        int pll_id;
        /* page flipping */
-       struct radeon_unpin_work *unpin_work;
-       int deferred_flip_completion;
+       struct workqueue_struct *flip_queue;
+       struct radeon_flip_work *flip_work;
        /* pll sharing */
        struct radeon_atom_ss ss;
        bool ss_enabled;