drm/radeon/kms: Add initial support for async DMA on r6xx/r7xx

Uses the new multi-ring infrastucture.  6xx/7xx has a single
async DMA ring.

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

7 files changed, 630 insertions, 12 deletions

diff --git a/drivers/gpu/drm/radeon/r600.c b/drivers/gpu/drm/radeon/r600.c
index cda280d157da..ee06c8781cd4 100644
--- a/drivers/gpu/drm/radeon/r600.c
+++ b/drivers/gpu/drm/radeon/r600.c
@@ -1370,6 +1370,29 @@ bool r600_gpu_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
 	return radeon_ring_test_lockup(rdev, ring);
 }
 
+/**
+ * r600_dma_is_lockup - Check if the DMA engine is locked up
+ *
+ * @rdev: radeon_device pointer
+ * @ring: radeon_ring structure holding ring information
+ *
+ * Check if the async DMA engine is locked up (r6xx-evergreen).
+ * Returns true if the engine appears to be locked up, false if not.
+ */
+bool r600_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
+{
+	u32 dma_status_reg;
+
+	dma_status_reg = RREG32(DMA_STATUS_REG);
+	if (dma_status_reg & DMA_IDLE) {
+		radeon_ring_lockup_update(ring);
+		return false;
+	}
+	/* force ring activities */
+	radeon_ring_force_activity(rdev, ring);
+	return radeon_ring_test_lockup(rdev, ring);
+}
+
 int r600_asic_reset(struct radeon_device *rdev)
 {
 	return r600_gpu_soft_reset(rdev);
@@ -1594,6 +1617,7 @@ static void r600_gpu_init(struct radeon_device *rdev)
 	WREG32(GB_TILING_CONFIG, tiling_config);
 	WREG32(DCP_TILING_CONFIG, tiling_config & 0xffff);
 	WREG32(HDP_TILING_CONFIG, tiling_config & 0xffff);
+	WREG32(DMA_TILING_CONFIG, tiling_config & 0xffff);
 
 	tmp = R6XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8);
 	WREG32(VGT_OUT_DEALLOC_CNTL, (tmp * 4) & DEALLOC_DIST_MASK);
@@ -1871,6 +1895,7 @@ void r600_cp_stop(struct radeon_device *rdev)
 	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
 	WREG32(R_0086D8_CP_ME_CNTL, S_0086D8_CP_ME_HALT(1));
 	WREG32(SCRATCH_UMSK, 0);
+	rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
 }
 
 int r600_init_microcode(struct radeon_device *rdev)
@@ -2196,6 +2221,128 @@ void r600_cp_fini(struct radeon_device *rdev)
 	radeon_scratch_free(rdev, ring->rptr_save_reg);
 }
 
+/*
+ * DMA
+ * Starting with R600, the GPU has an asynchronous
+ * DMA engine.  The programming model is very similar
+ * to the 3D engine (ring buffer, IBs, etc.), but the
+ * DMA controller has it's own packet format that is
+ * different form the PM4 format used by the 3D engine.
+ * It supports copying data, writing embedded data,
+ * solid fills, and a number of other things.  It also
+ * has support for tiling/detiling of buffers.
+ */
+/**
+ * r600_dma_stop - stop the async dma engine
+ *
+ * @rdev: radeon_device pointer
+ *
+ * Stop the async dma engine (r6xx-evergreen).
+ */
+void r600_dma_stop(struct radeon_device *rdev)
+{
+	u32 rb_cntl = RREG32(DMA_RB_CNTL);
+
+	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
+
+	rb_cntl &= ~DMA_RB_ENABLE;
+	WREG32(DMA_RB_CNTL, rb_cntl);
+
+	rdev->ring[R600_RING_TYPE_DMA_INDEX].ready = false;
+}
+
+/**
+ * r600_dma_resume - setup and start the async dma engine
+ *
+ * @rdev: radeon_device pointer
+ *
+ * Set up the DMA ring buffer and enable it. (r6xx-evergreen).
+ * Returns 0 for success, error for failure.
+ */
+int r600_dma_resume(struct radeon_device *rdev)
+{
+	struct radeon_ring *ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
+	u32 rb_cntl, dma_cntl;
+	u32 rb_bufsz;
+	int r;
+
+	/* Reset dma */
+	if (rdev->family >= CHIP_RV770)
+		WREG32(SRBM_SOFT_RESET, RV770_SOFT_RESET_DMA);
+	else
+		WREG32(SRBM_SOFT_RESET, SOFT_RESET_DMA);
+	RREG32(SRBM_SOFT_RESET);
+	udelay(50);
+	WREG32(SRBM_SOFT_RESET, 0);
+
+	WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL, 0);
+	WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL, 0);
+
+	/* Set ring buffer size in dwords */
+	rb_bufsz = drm_order(ring->ring_size / 4);
+	rb_cntl = rb_bufsz << 1;
+#ifdef __BIG_ENDIAN
+	rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
+#endif
+	WREG32(DMA_RB_CNTL, rb_cntl);
+
+	/* Initialize the ring buffer's read and write pointers */
+	WREG32(DMA_RB_RPTR, 0);
+	WREG32(DMA_RB_WPTR, 0);
+
+	/* set the wb address whether it's enabled or not */
+	WREG32(DMA_RB_RPTR_ADDR_HI,
+	       upper_32_bits(rdev->wb.gpu_addr + R600_WB_DMA_RPTR_OFFSET) & 0xFF);
+	WREG32(DMA_RB_RPTR_ADDR_LO,
+	       ((rdev->wb.gpu_addr + R600_WB_DMA_RPTR_OFFSET) & 0xFFFFFFFC));
+
+	if (rdev->wb.enabled)
+		rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
+
+	WREG32(DMA_RB_BASE, ring->gpu_addr >> 8);
+
+	/* enable DMA IBs */
+	WREG32(DMA_IB_CNTL, DMA_IB_ENABLE);
+
+	dma_cntl = RREG32(DMA_CNTL);
+	dma_cntl &= ~CTXEMPTY_INT_ENABLE;
+	WREG32(DMA_CNTL, dma_cntl);
+
+	if (rdev->family >= CHIP_RV770)
+		WREG32(DMA_MODE, 1);
+
+	ring->wptr = 0;
+	WREG32(DMA_RB_WPTR, ring->wptr << 2);
+
+	ring->rptr = RREG32(DMA_RB_RPTR) >> 2;
+
+	WREG32(DMA_RB_CNTL, rb_cntl | DMA_RB_ENABLE);
+
+	ring->ready = true;
+
+	r = radeon_ring_test(rdev, R600_RING_TYPE_DMA_INDEX, ring);
+	if (r) {
+		ring->ready = false;
+		return r;
+	}
+
+	radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
+
+	return 0;
+}
+
+/**
+ * r600_dma_fini - tear down the async dma engine
+ *
+ * @rdev: radeon_device pointer
+ *
+ * Stop the async dma engine and free the ring (r6xx-evergreen).
+ */
+void r600_dma_fini(struct radeon_device *rdev)
+{
+	r600_dma_stop(rdev);
+	radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX]);
+}
 
 /*
  * GPU scratch registers helpers function.
@@ -2252,6 +2399,64 @@ int r600_ring_test(struct radeon_device *rdev, struct radeon_ring *ring)
 	return r;
 }
 
+/**
+ * r600_dma_ring_test - simple async dma engine test
+ *
+ * @rdev: radeon_device pointer
+ * @ring: radeon_ring structure holding ring information
+ *
+ * Test the DMA engine by writing using it to write an
+ * value to memory. (r6xx-SI).
+ * Returns 0 for success, error for failure.
+ */
+int r600_dma_ring_test(struct radeon_device *rdev,
+		       struct radeon_ring *ring)
+{
+	unsigned i;
+	int r;
+	void __iomem *ptr = (void *)rdev->vram_scratch.ptr;
+	u32 tmp;
+
+	if (!ptr) {
+		DRM_ERROR("invalid vram scratch pointer\n");
+		return -EINVAL;
+	}
+
+	tmp = 0xCAFEDEAD;
+	writel(tmp, ptr);
+
+	r = radeon_ring_lock(rdev, ring, 4);
+	if (r) {
+		DRM_ERROR("radeon: dma failed to lock ring %d (%d).\n", ring->idx, r);
+		return r;
+	}
+	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1));
+	radeon_ring_write(ring, rdev->vram_scratch.gpu_addr & 0xfffffffc);
+	radeon_ring_write(ring, upper_32_bits(rdev->vram_scratch.gpu_addr) & 0xff);
+	radeon_ring_write(ring, 0xDEADBEEF);
+	radeon_ring_unlock_commit(rdev, ring);
+
+	for (i = 0; i < rdev->usec_timeout; i++) {
+		tmp = readl(ptr);
+		if (tmp == 0xDEADBEEF)
+			break;
+		DRM_UDELAY(1);
+	}
+
+	if (i < rdev->usec_timeout) {
+		DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
+	} else {
+		DRM_ERROR("radeon: ring %d test failed (0x%08X)\n",
+			  ring->idx, tmp);
+		r = -EINVAL;
+	}
+	return r;
+}
+
+/*
+ * CP fences/semaphores
+ */
+
 void r600_fence_ring_emit(struct radeon_device *rdev,
 			  struct radeon_fence *fence)
 {
@@ -2315,6 +2520,58 @@ void r600_semaphore_ring_emit(struct radeon_device *rdev,
 	radeon_ring_write(ring, (upper_32_bits(addr) & 0xff) | sel);
 }
 
+/*
+ * DMA fences/semaphores
+ */
+
+/**
+ * r600_dma_fence_ring_emit - emit a fence on the DMA ring
+ *
+ * @rdev: radeon_device pointer
+ * @fence: radeon fence object
+ *
+ * Add a DMA fence packet to the ring to write
+ * the fence seq number and DMA trap packet to generate
+ * an interrupt if needed (r6xx-r7xx).
+ */
+void r600_dma_fence_ring_emit(struct radeon_device *rdev,
+			      struct radeon_fence *fence)
+{
+	struct radeon_ring *ring = &rdev->ring[fence->ring];
+	u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
+	/* write the fence */
+	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0));
+	radeon_ring_write(ring, addr & 0xfffffffc);
+	radeon_ring_write(ring, (upper_32_bits(addr) & 0xff));
+	radeon_ring_write(ring, fence->seq);
+	/* generate an interrupt */
+	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_TRAP, 0, 0, 0));
+}
+
+/**
+ * r600_dma_semaphore_ring_emit - emit a semaphore on the dma ring
+ *
+ * @rdev: radeon_device pointer
+ * @ring: radeon_ring structure holding ring information
+ * @semaphore: radeon semaphore object
+ * @emit_wait: wait or signal semaphore
+ *
+ * Add a DMA semaphore packet to the ring wait on or signal
+ * other rings (r6xx-SI).
+ */
+void r600_dma_semaphore_ring_emit(struct radeon_device *rdev,
+				  struct radeon_ring *ring,
+				  struct radeon_semaphore *semaphore,
+				  bool emit_wait)
+{
+	u64 addr = semaphore->gpu_addr;
+	u32 s = emit_wait ? 0 : 1;
+
+	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SEMAPHORE, 0, s, 0));
+	radeon_ring_write(ring, addr & 0xfffffffc);
+	radeon_ring_write(ring, upper_32_bits(addr) & 0xff);
+}
+
 int r600_copy_blit(struct radeon_device *rdev,
 		   uint64_t src_offset,
 		   uint64_t dst_offset,
@@ -2334,6 +2591,80 @@ int r600_copy_blit(struct radeon_device *rdev,
 	return 0;
 }
 
+/**
+ * r600_copy_dma - copy pages using the DMA engine
+ *
+ * @rdev: radeon_device pointer
+ * @src_offset: src GPU address
+ * @dst_offset: dst GPU address
+ * @num_gpu_pages: number of GPU pages to xfer
+ * @fence: radeon fence object
+ *
+ * Copy GPU paging using the DMA engine (r6xx-r7xx).
+ * Used by the radeon ttm implementation to move pages if
+ * registered as the asic copy callback.
+ */
+int r600_copy_dma(struct radeon_device *rdev,
+		  uint64_t src_offset, uint64_t dst_offset,
+		  unsigned num_gpu_pages,
+		  struct radeon_fence **fence)
+{
+	struct radeon_semaphore *sem = NULL;
+	int ring_index = rdev->asic->copy.dma_ring_index;
+	struct radeon_ring *ring = &rdev->ring[ring_index];
+	u32 size_in_dw, cur_size_in_dw;
+	int i, num_loops;
+	int r = 0;
+
+	r = radeon_semaphore_create(rdev, &sem);
+	if (r) {
+		DRM_ERROR("radeon: moving bo (%d).\n", r);
+		return r;
+	}
+
+	size_in_dw = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT) / 4;
+	num_loops = DIV_ROUND_UP(size_in_dw, 0xffff);
+	r = radeon_ring_lock(rdev, ring, num_loops * 5 + 8);
+	if (r) {
+		DRM_ERROR("radeon: moving bo (%d).\n", r);
+		radeon_semaphore_free(rdev, &sem, NULL);
+		return r;
+	}
+
+	if (radeon_fence_need_sync(*fence, ring->idx)) {
+		radeon_semaphore_sync_rings(rdev, sem, (*fence)->ring,
+					    ring->idx);
+		radeon_fence_note_sync(*fence, ring->idx);
+	} else {
+		radeon_semaphore_free(rdev, &sem, NULL);
+	}
+
+	for (i = 0; i < num_loops; i++) {
+		cur_size_in_dw = size_in_dw;
+		if (cur_size_in_dw > 0xFFFF)
+			cur_size_in_dw = 0xFFFF;
+		size_in_dw -= cur_size_in_dw;
+		radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_COPY, 0, 0, cur_size_in_dw));
+		radeon_ring_write(ring, dst_offset & 0xfffffffc);
+		radeon_ring_write(ring, src_offset & 0xfffffffc);
+		radeon_ring_write(ring, upper_32_bits(dst_offset) & 0xff);
+		radeon_ring_write(ring, upper_32_bits(src_offset) & 0xff);
+		src_offset += cur_size_in_dw * 4;
+		dst_offset += cur_size_in_dw * 4;
+	}
+
+	r = radeon_fence_emit(rdev, fence, ring->idx);
+	if (r) {
+		radeon_ring_unlock_undo(rdev, ring);
+		return r;
+	}
+
+	radeon_ring_unlock_commit(rdev, ring);
+	radeon_semaphore_free(rdev, &sem, *fence);
+
+	return r;
+}
+
 int r600_set_surface_reg(struct radeon_device *rdev, int reg,
 			 uint32_t tiling_flags, uint32_t pitch,
 			 uint32_t offset, uint32_t obj_size)
@@ -2349,7 +2680,7 @@ void r600_clear_surface_reg(struct radeon_device *rdev, int reg)
 
 static int r600_startup(struct radeon_device *rdev)
 {
-	struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
+	struct radeon_ring *ring;
 	int r;
 
 	/* enable pcie gen2 link */
@@ -2394,6 +2725,12 @@ static int r600_startup(struct radeon_device *rdev)
 		return r;
 	}
 
+	r = radeon_fence_driver_start_ring(rdev, R600_RING_TYPE_DMA_INDEX);
+	if (r) {
+		dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
+		return r;
+	}
+
 	/* Enable IRQ */
 	r = r600_irq_init(rdev);
 	if (r) {
@@ -2403,12 +2740,20 @@ static int r600_startup(struct radeon_device *rdev)
 	}
 	r600_irq_set(rdev);
 
+	ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
 	r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
 			     R600_CP_RB_RPTR, R600_CP_RB_WPTR,
 			     0, 0xfffff, RADEON_CP_PACKET2);
+	if (r)
+		return r;
 
+	ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
+	r = radeon_ring_init(rdev, ring, ring->ring_size, R600_WB_DMA_RPTR_OFFSET,
+			     DMA_RB_RPTR, DMA_RB_WPTR,
+			     2, 0x3fffc, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0));
 	if (r)
 		return r;
+
 	r = r600_cp_load_microcode(rdev);
 	if (r)
 		return r;
@@ -2416,6 +2761,10 @@ static int r600_startup(struct radeon_device *rdev)
 	if (r)
 		return r;
 
+	r = r600_dma_resume(rdev);
+	if (r)
+		return r;
+
 	r = radeon_ib_pool_init(rdev);
 	if (r) {
 		dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
@@ -2471,7 +2820,7 @@ int r600_suspend(struct radeon_device *rdev)
 {
 	r600_audio_fini(rdev);
 	r600_cp_stop(rdev);
-	rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
+	r600_dma_stop(rdev);
 	r600_irq_suspend(rdev);
 	radeon_wb_disable(rdev);
 	r600_pcie_gart_disable(rdev);
@@ -2544,6 +2893,9 @@ int r600_init(struct radeon_device *rdev)
 	rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ring_obj = NULL;
 	r600_ring_init(rdev, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX], 1024 * 1024);
 
+	rdev->ring[R600_RING_TYPE_DMA_INDEX].ring_obj = NULL;
+	r600_ring_init(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX], 64 * 1024);
+
 	rdev->ih.ring_obj = NULL;
 	r600_ih_ring_init(rdev, 64 * 1024);
 
@@ -2556,6 +2908,7 @@ int r600_init(struct radeon_device *rdev)
 	if (r) {
 		dev_err(rdev->dev, "disabling GPU acceleration\n");
 		r600_cp_fini(rdev);
+		r600_dma_fini(rdev);
 		r600_irq_fini(rdev);
 		radeon_wb_fini(rdev);
 		radeon_ib_pool_fini(rdev);
@@ -2572,6 +2925,7 @@ void r600_fini(struct radeon_device *rdev)
 	r600_audio_fini(rdev);
 	r600_blit_fini(rdev);
 	r600_cp_fini(rdev);
+	r600_dma_fini(rdev);
 	r600_irq_fini(rdev);
 	radeon_wb_fini(rdev);
 	radeon_ib_pool_fini(rdev);
@@ -2674,6 +3028,104 @@ free_scratch:
 	return r;
 }
 
+/**
+ * r600_dma_ib_test - test an IB on the DMA engine
+ *
+ * @rdev: radeon_device pointer
+ * @ring: radeon_ring structure holding ring information
+ *
+ * Test a simple IB in the DMA ring (r6xx-SI).
+ * Returns 0 on success, error on failure.
+ */
+int r600_dma_ib_test(struct radeon_device *rdev, struct radeon_ring *ring)
+{
+	struct radeon_ib ib;
+	unsigned i;
+	int r;
+	void __iomem *ptr = (void *)rdev->vram_scratch.ptr;
+	u32 tmp = 0;
+
+	if (!ptr) {
+		DRM_ERROR("invalid vram scratch pointer\n");
+		return -EINVAL;
+	}
+
+	tmp = 0xCAFEDEAD;
+	writel(tmp, ptr);
+
+	r = radeon_ib_get(rdev, ring->idx, &ib, NULL, 256);
+	if (r) {
+		DRM_ERROR("radeon: failed to get ib (%d).\n", r);
+		return r;
+	}
+
+	ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1);
+	ib.ptr[1] = rdev->vram_scratch.gpu_addr & 0xfffffffc;
+	ib.ptr[2] = upper_32_bits(rdev->vram_scratch.gpu_addr) & 0xff;
+	ib.ptr[3] = 0xDEADBEEF;
+	ib.length_dw = 4;
+
+	r = radeon_ib_schedule(rdev, &ib, NULL);
+	if (r) {
+		radeon_ib_free(rdev, &ib);
+		DRM_ERROR("radeon: failed to schedule ib (%d).\n", r);
+		return r;
+	}
+	r = radeon_fence_wait(ib.fence, false);
+	if (r) {
+		DRM_ERROR("radeon: fence wait failed (%d).\n", r);
+		return r;
+	}
+	for (i = 0; i < rdev->usec_timeout; i++) {
+		tmp = readl(ptr);
+		if (tmp == 0xDEADBEEF)
+			break;
+		DRM_UDELAY(1);
+	}
+	if (i < rdev->usec_timeout) {
+		DRM_INFO("ib test on ring %d succeeded in %u usecs\n", ib.fence->ring, i);
+	} else {
+		DRM_ERROR("radeon: ib test failed (0x%08X)\n", tmp);
+		r = -EINVAL;
+	}
+	radeon_ib_free(rdev, &ib);
+	return r;
+}
+
+/**
+ * r600_dma_ring_ib_execute - Schedule an IB on the DMA engine
+ *
+ * @rdev: radeon_device pointer
+ * @ib: IB object to schedule
+ *
+ * Schedule an IB in the DMA ring (r6xx-r7xx).
+ */
+void r600_dma_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
+{
+	struct radeon_ring *ring = &rdev->ring[ib->ring];
+
+	if (rdev->wb.enabled) {
+		u32 next_rptr = ring->wptr + 4;
+		while ((next_rptr & 7) != 5)
+			next_rptr++;
+		next_rptr += 3;
+		radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1));
+		radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
+		radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xff);
+		radeon_ring_write(ring, next_rptr);
+	}
+
+	/* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
+	 * Pad as necessary with NOPs.
+	 */
+	while ((ring->wptr & 7) != 5)
+		radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0));
+	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_INDIRECT_BUFFER, 0, 0, 0));
+	radeon_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
+	radeon_ring_write(ring, (ib->length_dw << 16) | (upper_32_bits(ib->gpu_addr) & 0xFF));
+
+}
+
 /*
  * Interrupts
  *
@@ -2865,6 +3317,8 @@ static void r600_disable_interrupt_state(struct radeon_device *rdev)
 	u32 tmp;
 
 	WREG32(CP_INT_CNTL, CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
+	tmp = RREG32(DMA_CNTL) & ~TRAP_ENABLE;
+	WREG32(DMA_CNTL, tmp);
 	WREG32(GRBM_INT_CNTL, 0);
 	WREG32(DxMODE_INT_MASK, 0);
 	WREG32(D1GRPH_INTERRUPT_CONTROL, 0);
@@ -3006,6 +3460,7 @@ int r600_irq_set(struct radeon_device *rdev)
 	u32 grbm_int_cntl = 0;
 	u32 hdmi0, hdmi1;
 	u32 d1grph = 0, d2grph = 0;
+	u32 dma_cntl;
 
 	if (!rdev->irq.installed) {
 		WARN(1, "Can't enable IRQ/MSI because no handler is installed\n");
@@ -3040,12 +3495,19 @@ int r600_irq_set(struct radeon_device *rdev)
 		hdmi0 = RREG32(HDMI0_AUDIO_PACKET_CONTROL) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
 		hdmi1 = RREG32(HDMI1_AUDIO_PACKET_CONTROL) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
 	}
+	dma_cntl = RREG32(DMA_CNTL) & ~TRAP_ENABLE;
 
 	if (atomic_read(&rdev->irq.ring_int[RADEON_RING_TYPE_GFX_INDEX])) {
 		DRM_DEBUG("r600_irq_set: sw int\n");
 		cp_int_cntl |= RB_INT_ENABLE;
 		cp_int_cntl |= TIME_STAMP_INT_ENABLE;
 	}
+
+	if (atomic_read(&rdev->irq.ring_int[R600_RING_TYPE_DMA_INDEX])) {
+		DRM_DEBUG("r600_irq_set: sw int dma\n");
+		dma_cntl |= TRAP_ENABLE;
+	}
+
 	if (rdev->irq.crtc_vblank_int[0] ||
 	    atomic_read(&rdev->irq.pflip[0])) {
 		DRM_DEBUG("r600_irq_set: vblank 0\n");
@@ -3090,6 +3552,7 @@ int r600_irq_set(struct radeon_device *rdev)
 	}
 
 	WREG32(CP_INT_CNTL, cp_int_cntl);
+	WREG32(DMA_CNTL, dma_cntl);
 	WREG32(DxMODE_INT_MASK, mode_int);
 	WREG32(D1GRPH_INTERRUPT_CONTROL, d1grph);
 	WREG32(D2GRPH_INTERRUPT_CONTROL, d2grph);
@@ -3469,6 +3932,10 @@ restart_ih:
 			DRM_DEBUG("IH: CP EOP\n");
 			radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
 			break;
+		case 224: /* DMA trap event */
+			DRM_DEBUG("IH: DMA trap\n");
+			radeon_fence_process(rdev, R600_RING_TYPE_DMA_INDEX);
+			break;
 		case 233: /* GUI IDLE */
 			DRM_DEBUG("IH: GUI idle\n");
 			break;
diff --git a/drivers/gpu/drm/radeon/r600d.h b/drivers/gpu/drm/radeon/r600d.h
index fa6f37099ba9..a596c554a3a0 100644
--- a/drivers/gpu/drm/radeon/r600d.h
+++ b/drivers/gpu/drm/radeon/r600d.h
@@ -590,9 +590,59 @@
 #define         WAIT_2D_IDLECLEAN_bit                           (1 << 16)
 #define         WAIT_3D_IDLECLEAN_bit                           (1 << 17)
 
+/* async DMA */
+#define DMA_TILING_CONFIG                                 0x3ec4
+#define DMA_CONFIG                                        0x3e4c
+
+#define DMA_RB_CNTL                                       0xd000
+#       define DMA_RB_ENABLE                              (1 << 0)
+#       define DMA_RB_SIZE(x)                             ((x) << 1) /* log2 */
+#       define DMA_RB_SWAP_ENABLE                         (1 << 9) /* 8IN32 */
+#       define DMA_RPTR_WRITEBACK_ENABLE                  (1 << 12)
+#       define DMA_RPTR_WRITEBACK_SWAP_ENABLE             (1 << 13)  /* 8IN32 */
+#       define DMA_RPTR_WRITEBACK_TIMER(x)                ((x) << 16) /* log2 */
+#define DMA_RB_BASE                                       0xd004
+#define DMA_RB_RPTR                                       0xd008
+#define DMA_RB_WPTR                                       0xd00c
+
+#define DMA_RB_RPTR_ADDR_HI                               0xd01c
+#define DMA_RB_RPTR_ADDR_LO                               0xd020
+
+#define DMA_IB_CNTL                                       0xd024
+#       define DMA_IB_ENABLE                              (1 << 0)
+#       define DMA_IB_SWAP_ENABLE                         (1 << 4)
+#define DMA_IB_RPTR                                       0xd028
+#define DMA_CNTL                                          0xd02c
+#       define TRAP_ENABLE                                (1 << 0)
+#       define SEM_INCOMPLETE_INT_ENABLE                  (1 << 1)
+#       define SEM_WAIT_INT_ENABLE                        (1 << 2)
+#       define DATA_SWAP_ENABLE                           (1 << 3)
+#       define FENCE_SWAP_ENABLE                          (1 << 4)
+#       define CTXEMPTY_INT_ENABLE                        (1 << 28)
+#define DMA_STATUS_REG                                    0xd034
+#       define DMA_IDLE                                   (1 << 0)
+#define DMA_SEM_INCOMPLETE_TIMER_CNTL                     0xd044
+#define DMA_SEM_WAIT_FAIL_TIMER_CNTL                      0xd048
+#define DMA_MODE                                          0xd0bc
+
+/* async DMA packets */
+#define DMA_PACKET(cmd, t, s, n)	((((cmd) & 0xF) << 28) |	\
+					 (((t) & 0x1) << 23) |		\
+					 (((s) & 0x1) << 22) |		\
+					 (((n) & 0xFFFF) << 0))
+/* async DMA Packet types */
+#define	DMA_PACKET_WRITE				  0x2
+#define	DMA_PACKET_COPY					  0x3
+#define	DMA_PACKET_INDIRECT_BUFFER			  0x4
+#define	DMA_PACKET_SEMAPHORE				  0x5
+#define	DMA_PACKET_FENCE				  0x6
+#define	DMA_PACKET_TRAP					  0x7
+#define	DMA_PACKET_CONSTANT_FILL			  0xd /* 7xx only */
+#define	DMA_PACKET_NOP					  0xf
+
 #define IH_RB_CNTL                                        0x3e00
 #       define IH_RB_ENABLE                               (1 << 0)
-#       define IH_IB_SIZE(x)                              ((x) << 1) /* log2 */
+#       define IH_RB_SIZE(x)                              ((x) << 1) /* log2 */
 #       define IH_RB_FULL_DRAIN_ENABLE                    (1 << 6)
 #       define IH_WPTR_WRITEBACK_ENABLE                   (1 << 8)
 #       define IH_WPTR_WRITEBACK_TIMER(x)                 ((x) << 9) /* log2 */
@@ -637,7 +687,9 @@
 #define TN_RLC_CLEAR_STATE_RESTORE_BASE                   0x3f20
 
 #define SRBM_SOFT_RESET                                   0xe60
+#       define SOFT_RESET_DMA                             (1 << 12)
 #       define SOFT_RESET_RLC                             (1 << 13)
+#       define RV770_SOFT_RESET_DMA                       (1 << 20)
 
 #define CP_INT_CNTL                                       0xc124
 #       define CNTX_BUSY_INT_ENABLE                       (1 << 19)
diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
index 8c42d54c2e26..461bf53709f5 100644
--- a/drivers/gpu/drm/radeon/radeon.h
+++ b/drivers/gpu/drm/radeon/radeon.h
@@ -109,7 +109,7 @@ extern int radeon_lockup_timeout;
 #define RADEON_BIOS_NUM_SCRATCH			8
 
 /* max number of rings */
-#define RADEON_NUM_RINGS			3
+#define RADEON_NUM_RINGS			4
 
 /* fence seq are set to this number when signaled */
 #define RADEON_FENCE_SIGNALED_SEQ		0LL
@@ -122,6 +122,9 @@ extern int radeon_lockup_timeout;
 #define CAYMAN_RING_TYPE_CP1_INDEX		1
 #define CAYMAN_RING_TYPE_CP2_INDEX		2
 
+/* R600+ has an async dma ring */
+#define R600_RING_TYPE_DMA_INDEX		3
+
 /* hardcode those limit for now */
 #define RADEON_VA_IB_OFFSET			(1 << 20)
 #define RADEON_VA_RESERVED_SIZE			(8 << 20)
@@ -787,6 +790,11 @@ int radeon_ring_init(struct radeon_device *rdev, struct radeon_ring *cp, unsigne
 void radeon_ring_fini(struct radeon_device *rdev, struct radeon_ring *cp);
 
 
+/* r600 async dma */
+void r600_dma_stop(struct radeon_device *rdev);
+int r600_dma_resume(struct radeon_device *rdev);
+void r600_dma_fini(struct radeon_device *rdev);
+
 /*
  * CS.
  */
@@ -883,6 +891,7 @@ struct radeon_wb {
 #define RADEON_WB_CP_RPTR_OFFSET 1024
 #define RADEON_WB_CP1_RPTR_OFFSET 1280
 #define RADEON_WB_CP2_RPTR_OFFSET 1536
+#define R600_WB_DMA_RPTR_OFFSET   1792
 #define R600_WB_IH_WPTR_OFFSET   2048
 #define R600_WB_EVENT_OFFSET     3072
 
diff --git a/drivers/gpu/drm/radeon/radeon_asic.c b/drivers/gpu/drm/radeon/radeon_asic.c
index 654520b95ab7..3cf9b29fb53f 100644
--- a/drivers/gpu/drm/radeon/radeon_asic.c
+++ b/drivers/gpu/drm/radeon/radeon_asic.c
@@ -947,6 +947,15 @@ static struct radeon_asic r600_asic = {
 			.ring_test = &r600_ring_test,
 			.ib_test = &r600_ib_test,
 			.is_lockup = &r600_gpu_is_lockup,
+		},
+		[R600_RING_TYPE_DMA_INDEX] = {
+			.ib_execute = &r600_dma_ring_ib_execute,
+			.emit_fence = &r600_dma_fence_ring_emit,
+			.emit_semaphore = &r600_dma_semaphore_ring_emit,
+			.cs_parse = NULL,
+			.ring_test = &r600_dma_ring_test,
+			.ib_test = &r600_dma_ib_test,
+			.is_lockup = &r600_dma_is_lockup,
 		}
 	},
 	.irq = {
@@ -963,8 +972,8 @@ static struct radeon_asic r600_asic = {
 	.copy = {
 		.blit = &r600_copy_blit,
 		.blit_ring_index = RADEON_RING_TYPE_GFX_INDEX,
-		.dma = NULL,
-		.dma_ring_index = RADEON_RING_TYPE_GFX_INDEX,
+		.dma = &r600_copy_dma,
+		.dma_ring_index = R600_RING_TYPE_DMA_INDEX,
 		.copy = &r600_copy_blit,
 		.copy_ring_index = RADEON_RING_TYPE_GFX_INDEX,
 	},
@@ -1022,6 +1031,15 @@ static struct radeon_asic rs780_asic = {
 			.ring_test = &r600_ring_test,
 			.ib_test = &r600_ib_test,
 			.is_lockup = &r600_gpu_is_lockup,
+		},
+		[R600_RING_TYPE_DMA_INDEX] = {
+			.ib_execute = &r600_dma_ring_ib_execute,
+			.emit_fence = &r600_dma_fence_ring_emit,
+			.emit_semaphore = &r600_dma_semaphore_ring_emit,
+			.cs_parse = NULL,
+			.ring_test = &r600_dma_ring_test,
+			.ib_test = &r600_dma_ib_test,
+			.is_lockup = &r600_dma_is_lockup,
 		}
 	},
 	.irq = {
@@ -1038,8 +1056,8 @@ static struct radeon_asic rs780_asic = {
 	.copy = {
 		.blit = &r600_copy_blit,
 		.blit_ring_index = RADEON_RING_TYPE_GFX_INDEX,
-		.dma = NULL,
-		.dma_ring_index = RADEON_RING_TYPE_GFX_INDEX,
+		.dma = &r600_copy_dma,
+		.dma_ring_index = R600_RING_TYPE_DMA_INDEX,
 		.copy = &r600_copy_blit,
 		.copy_ring_index = RADEON_RING_TYPE_GFX_INDEX,
 	},
@@ -1097,6 +1115,15 @@ static struct radeon_asic rv770_asic = {
 			.ring_test = &r600_ring_test,
 			.ib_test = &r600_ib_test,
 			.is_lockup = &r600_gpu_is_lockup,
+		},
+		[R600_RING_TYPE_DMA_INDEX] = {
+			.ib_execute = &r600_dma_ring_ib_execute,
+			.emit_fence = &r600_dma_fence_ring_emit,
+			.emit_semaphore = &r600_dma_semaphore_ring_emit,
+			.cs_parse = NULL,
+			.ring_test = &r600_dma_ring_test,
+			.ib_test = &r600_dma_ib_test,
+			.is_lockup = &r600_dma_is_lockup,
 		}
 	},
 	.irq = {
@@ -1113,8 +1140,8 @@ static struct radeon_asic rv770_asic = {
 	.copy = {
 		.blit = &r600_copy_blit,
 		.blit_ring_index = RADEON_RING_TYPE_GFX_INDEX,
-		.dma = NULL,
-		.dma_ring_index = RADEON_RING_TYPE_GFX_INDEX,
+		.dma = &r600_copy_dma,
+		.dma_ring_index = R600_RING_TYPE_DMA_INDEX,
 		.copy = &r600_copy_blit,
 		.copy_ring_index = RADEON_RING_TYPE_GFX_INDEX,
 	},
diff --git a/drivers/gpu/drm/radeon/radeon_asic.h b/drivers/gpu/drm/radeon/radeon_asic.h
index 5e3a0e5c6be1..70a5b1f0e43e 100644
--- a/drivers/gpu/drm/radeon/radeon_asic.h
+++ b/drivers/gpu/drm/radeon/radeon_asic.h
@@ -309,6 +309,14 @@ void r600_semaphore_ring_emit(struct radeon_device *rdev,
 			      struct radeon_ring *cp,
 			      struct radeon_semaphore *semaphore,
 			      bool emit_wait);
+void r600_dma_fence_ring_emit(struct radeon_device *rdev,
+			      struct radeon_fence *fence);
+void r600_dma_semaphore_ring_emit(struct radeon_device *rdev,
+				  struct radeon_ring *ring,
+				  struct radeon_semaphore *semaphore,
+				  bool emit_wait);
+void r600_dma_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib);
+bool r600_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring);
 bool r600_gpu_is_lockup(struct radeon_device *rdev, struct radeon_ring *cp);
 int r600_asic_reset(struct radeon_device *rdev);
 int r600_set_surface_reg(struct radeon_device *rdev, int reg,
@@ -316,11 +324,16 @@ int r600_set_surface_reg(struct radeon_device *rdev, int reg,
 			 uint32_t offset, uint32_t obj_size);
 void r600_clear_surface_reg(struct radeon_device *rdev, int reg);
 int r600_ib_test(struct radeon_device *rdev, struct radeon_ring *ring);
+int r600_dma_ib_test(struct radeon_device *rdev, struct radeon_ring *ring);
 void r600_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib);
 int r600_ring_test(struct radeon_device *rdev, struct radeon_ring *cp);
+int r600_dma_ring_test(struct radeon_device *rdev, struct radeon_ring *cp);
 int r600_copy_blit(struct radeon_device *rdev,
 		   uint64_t src_offset, uint64_t dst_offset,
 		   unsigned num_gpu_pages, struct radeon_fence **fence);
+int r600_copy_dma(struct radeon_device *rdev,
+		  uint64_t src_offset, uint64_t dst_offset,
+		  unsigned num_gpu_pages, struct radeon_fence **fence);
 void r600_hpd_init(struct radeon_device *rdev);
 void r600_hpd_fini(struct radeon_device *rdev);
 bool r600_hpd_sense(struct radeon_device *rdev, enum radeon_hpd_id hpd);
diff --git a/drivers/gpu/drm/radeon/rv770.c b/drivers/gpu/drm/radeon/rv770.c
index 79814a08c8e5..87c979c4f721 100644
--- a/drivers/gpu/drm/radeon/rv770.c
+++ b/drivers/gpu/drm/radeon/rv770.c
@@ -316,6 +316,7 @@ void r700_cp_stop(struct radeon_device *rdev)
 	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
 	WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT));
 	WREG32(SCRATCH_UMSK, 0);
+	rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
 }
 
 static int rv770_cp_load_microcode(struct radeon_device *rdev)
@@ -583,6 +584,8 @@ static void rv770_gpu_init(struct radeon_device *rdev)
 	WREG32(GB_TILING_CONFIG, gb_tiling_config);
 	WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff));
 	WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff));
+	WREG32(DMA_TILING_CONFIG, (gb_tiling_config & 0xffff));
+	WREG32(DMA_TILING_CONFIG2, (gb_tiling_config & 0xffff));
 
 	WREG32(CGTS_SYS_TCC_DISABLE, 0);
 	WREG32(CGTS_TCC_DISABLE, 0);
@@ -886,7 +889,7 @@ static int rv770_mc_init(struct radeon_device *rdev)
 
 static int rv770_startup(struct radeon_device *rdev)
 {
-	struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
+	struct radeon_ring *ring;
 	int r;
 
 	/* enable pcie gen2 link */
@@ -932,6 +935,12 @@ static int rv770_startup(struct radeon_device *rdev)
 		return r;
 	}
 
+	r = radeon_fence_driver_start_ring(rdev, R600_RING_TYPE_DMA_INDEX);
+	if (r) {
+		dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
+		return r;
+	}
+
 	/* Enable IRQ */
 	r = r600_irq_init(rdev);
 	if (r) {
@@ -941,11 +950,20 @@ static int rv770_startup(struct radeon_device *rdev)
 	}
 	r600_irq_set(rdev);
 
+	ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
 	r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
 			     R600_CP_RB_RPTR, R600_CP_RB_WPTR,
 			     0, 0xfffff, RADEON_CP_PACKET2);
 	if (r)
 		return r;
+
+	ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
+	r = radeon_ring_init(rdev, ring, ring->ring_size, R600_WB_DMA_RPTR_OFFSET,
+			     DMA_RB_RPTR, DMA_RB_WPTR,
+			     2, 0x3fffc, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0));
+	if (r)
+		return r;
+
 	r = rv770_cp_load_microcode(rdev);
 	if (r)
 		return r;
@@ -953,6 +971,10 @@ static int rv770_startup(struct radeon_device *rdev)
 	if (r)
 		return r;
 
+	r = r600_dma_resume(rdev);
+	if (r)
+		return r;
+
 	r = radeon_ib_pool_init(rdev);
 	if (r) {
 		dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
@@ -995,7 +1017,7 @@ int rv770_suspend(struct radeon_device *rdev)
 {
 	r600_audio_fini(rdev);
 	r700_cp_stop(rdev);
-	rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
+	r600_dma_stop(rdev);
 	r600_irq_suspend(rdev);
 	radeon_wb_disable(rdev);
 	rv770_pcie_gart_disable(rdev);
@@ -1066,6 +1088,9 @@ int rv770_init(struct radeon_device *rdev)
 	rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ring_obj = NULL;
 	r600_ring_init(rdev, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX], 1024 * 1024);
 
+	rdev->ring[R600_RING_TYPE_DMA_INDEX].ring_obj = NULL;
+	r600_ring_init(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX], 64 * 1024);
+
 	rdev->ih.ring_obj = NULL;
 	r600_ih_ring_init(rdev, 64 * 1024);
 
@@ -1078,6 +1103,7 @@ int rv770_init(struct radeon_device *rdev)
 	if (r) {
 		dev_err(rdev->dev, "disabling GPU acceleration\n");
 		r700_cp_fini(rdev);
+		r600_dma_fini(rdev);
 		r600_irq_fini(rdev);
 		radeon_wb_fini(rdev);
 		radeon_ib_pool_fini(rdev);
@@ -1093,6 +1119,7 @@ void rv770_fini(struct radeon_device *rdev)
 {
 	r600_blit_fini(rdev);
 	r700_cp_fini(rdev);
+	r600_dma_fini(rdev);
 	r600_irq_fini(rdev);
 	radeon_wb_fini(rdev);
 	radeon_ib_pool_fini(rdev);
diff --git a/drivers/gpu/drm/radeon/rv770d.h b/drivers/gpu/drm/radeon/rv770d.h
index e2d9dc8e751e..20e29d23d348 100644
--- a/drivers/gpu/drm/radeon/rv770d.h
+++ b/drivers/gpu/drm/radeon/rv770d.h
@@ -109,6 +109,9 @@
 #define     PIPE_TILING__SHIFT              1
 #define     PIPE_TILING__MASK               0x0000000e
 
+#define DMA_TILING_CONFIG                               0x3ec8
+#define DMA_TILING_CONFIG2                              0xd0b8
+
 #define	GC_USER_SHADER_PIPE_CONFIG			0x8954
 #define		INACTIVE_QD_PIPES(x)				((x) << 8)
 #define		INACTIVE_QD_PIPES_MASK				0x0000FF00
@@ -358,6 +361,26 @@
 
 #define	WAIT_UNTIL					0x8040
 
+/* async DMA */
+#define DMA_RB_RPTR                                       0xd008
+#define DMA_RB_WPTR                                       0xd00c
+
+/* async DMA packets */
+#define DMA_PACKET(cmd, t, s, n)	((((cmd) & 0xF) << 28) |	\
+					 (((t) & 0x1) << 23) |		\
+					 (((s) & 0x1) << 22) |		\
+					 (((n) & 0xFFFF) << 0))
+/* async DMA Packet types */
+#define	DMA_PACKET_WRITE				  0x2
+#define	DMA_PACKET_COPY					  0x3
+#define	DMA_PACKET_INDIRECT_BUFFER			  0x4
+#define	DMA_PACKET_SEMAPHORE				  0x5
+#define	DMA_PACKET_FENCE				  0x6
+#define	DMA_PACKET_TRAP					  0x7
+#define	DMA_PACKET_CONSTANT_FILL			  0xd
+#define	DMA_PACKET_NOP					  0xf
+
+
 #define	SRBM_STATUS				        0x0E50
 
 /* DCE 3.2 HDMI */