2 * Copyright 2013 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
22 * Authors: Alex Deucher
26 #include "radeon_asic.h"
29 u32 r600_gpu_check_soft_reset(struct radeon_device *rdev);
33 * Starting with R600, the GPU has an asynchronous
34 * DMA engine. The programming model is very similar
35 * to the 3D engine (ring buffer, IBs, etc.), but the
36 * DMA controller has it's own packet format that is
37 * different form the PM4 format used by the 3D engine.
38 * It supports copying data, writing embedded data,
39 * solid fills, and a number of other things. It also
40 * has support for tiling/detiling of buffers.
44 * r600_dma_get_rptr - get the current read pointer
46 * @rdev: radeon_device pointer
47 * @ring: radeon ring pointer
49 * Get the current rptr from the hardware (r6xx+).
51 uint32_t r600_dma_get_rptr(struct radeon_device *rdev,
52 struct radeon_ring *ring)
57 rptr = rdev->wb.wb[ring->rptr_offs/4];
59 rptr = RREG32(DMA_RB_RPTR);
61 return (rptr & 0x3fffc) >> 2;
65 * r600_dma_get_wptr - get the current write pointer
67 * @rdev: radeon_device pointer
68 * @ring: radeon ring pointer
70 * Get the current wptr from the hardware (r6xx+).
72 uint32_t r600_dma_get_wptr(struct radeon_device *rdev,
73 struct radeon_ring *ring)
75 return (RREG32(DMA_RB_WPTR) & 0x3fffc) >> 2;
79 * r600_dma_set_wptr - commit the write pointer
81 * @rdev: radeon_device pointer
82 * @ring: radeon ring pointer
84 * Write the wptr back to the hardware (r6xx+).
86 void r600_dma_set_wptr(struct radeon_device *rdev,
87 struct radeon_ring *ring)
89 WREG32(DMA_RB_WPTR, (ring->wptr << 2) & 0x3fffc);
93 * r600_dma_stop - stop the async dma engine
95 * @rdev: radeon_device pointer
97 * Stop the async dma engine (r6xx-evergreen).
99 void r600_dma_stop(struct radeon_device *rdev)
101 u32 rb_cntl = RREG32(DMA_RB_CNTL);
103 if (rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX)
104 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
106 rb_cntl &= ~DMA_RB_ENABLE;
107 WREG32(DMA_RB_CNTL, rb_cntl);
109 rdev->ring[R600_RING_TYPE_DMA_INDEX].ready = false;
113 * r600_dma_resume - setup and start the async dma engine
115 * @rdev: radeon_device pointer
117 * Set up the DMA ring buffer and enable it. (r6xx-evergreen).
118 * Returns 0 for success, error for failure.
120 int r600_dma_resume(struct radeon_device *rdev)
122 struct radeon_ring *ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
123 u32 rb_cntl, dma_cntl, ib_cntl;
128 if (rdev->family >= CHIP_RV770)
129 WREG32(SRBM_SOFT_RESET, RV770_SOFT_RESET_DMA);
131 WREG32(SRBM_SOFT_RESET, SOFT_RESET_DMA);
132 RREG32(SRBM_SOFT_RESET);
134 WREG32(SRBM_SOFT_RESET, 0);
136 WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL, 0);
137 WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL, 0);
139 /* Set ring buffer size in dwords */
140 rb_bufsz = order_base_2(ring->ring_size / 4);
141 rb_cntl = rb_bufsz << 1;
143 rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
145 WREG32(DMA_RB_CNTL, rb_cntl);
147 /* Initialize the ring buffer's read and write pointers */
148 WREG32(DMA_RB_RPTR, 0);
149 WREG32(DMA_RB_WPTR, 0);
151 /* set the wb address whether it's enabled or not */
152 WREG32(DMA_RB_RPTR_ADDR_HI,
153 upper_32_bits(rdev->wb.gpu_addr + R600_WB_DMA_RPTR_OFFSET) & 0xFF);
154 WREG32(DMA_RB_RPTR_ADDR_LO,
155 ((rdev->wb.gpu_addr + R600_WB_DMA_RPTR_OFFSET) & 0xFFFFFFFC));
157 if (rdev->wb.enabled)
158 rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
160 WREG32(DMA_RB_BASE, ring->gpu_addr >> 8);
163 ib_cntl = DMA_IB_ENABLE;
165 ib_cntl |= DMA_IB_SWAP_ENABLE;
167 WREG32(DMA_IB_CNTL, ib_cntl);
169 dma_cntl = RREG32(DMA_CNTL);
170 dma_cntl &= ~CTXEMPTY_INT_ENABLE;
171 WREG32(DMA_CNTL, dma_cntl);
173 if (rdev->family >= CHIP_RV770)
177 WREG32(DMA_RB_WPTR, ring->wptr << 2);
179 WREG32(DMA_RB_CNTL, rb_cntl | DMA_RB_ENABLE);
183 r = radeon_ring_test(rdev, R600_RING_TYPE_DMA_INDEX, ring);
189 if (rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX)
190 radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
196 * r600_dma_fini - tear down the async dma engine
198 * @rdev: radeon_device pointer
200 * Stop the async dma engine and free the ring (r6xx-evergreen).
202 void r600_dma_fini(struct radeon_device *rdev)
205 radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX]);
209 * r600_dma_is_lockup - Check if the DMA engine is locked up
211 * @rdev: radeon_device pointer
212 * @ring: radeon_ring structure holding ring information
214 * Check if the async DMA engine is locked up.
215 * Returns true if the engine appears to be locked up, false if not.
217 bool r600_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
219 u32 reset_mask = r600_gpu_check_soft_reset(rdev);
221 if (!(reset_mask & RADEON_RESET_DMA)) {
222 radeon_ring_lockup_update(rdev, ring);
225 return radeon_ring_test_lockup(rdev, ring);
230 * r600_dma_ring_test - simple async dma engine test
232 * @rdev: radeon_device pointer
233 * @ring: radeon_ring structure holding ring information
235 * Test the DMA engine by writing using it to write an
236 * value to memory. (r6xx-SI).
237 * Returns 0 for success, error for failure.
239 int r600_dma_ring_test(struct radeon_device *rdev,
240 struct radeon_ring *ring)
244 void __iomem *ptr = (void *)rdev->vram_scratch.ptr;
248 DRM_ERROR("invalid vram scratch pointer\n");
255 r = radeon_ring_lock(rdev, ring, 4);
257 DRM_ERROR("radeon: dma failed to lock ring %d (%d).\n", ring->idx, r);
260 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1));
261 radeon_ring_write(ring, rdev->vram_scratch.gpu_addr & 0xfffffffc);
262 radeon_ring_write(ring, upper_32_bits(rdev->vram_scratch.gpu_addr) & 0xff);
263 radeon_ring_write(ring, 0xDEADBEEF);
264 radeon_ring_unlock_commit(rdev, ring);
266 for (i = 0; i < rdev->usec_timeout; i++) {
268 if (tmp == 0xDEADBEEF)
273 if (i < rdev->usec_timeout) {
274 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
276 DRM_ERROR("radeon: ring %d test failed (0x%08X)\n",
284 * r600_dma_fence_ring_emit - emit a fence on the DMA ring
286 * @rdev: radeon_device pointer
287 * @fence: radeon fence object
289 * Add a DMA fence packet to the ring to write
290 * the fence seq number and DMA trap packet to generate
291 * an interrupt if needed (r6xx-r7xx).
293 void r600_dma_fence_ring_emit(struct radeon_device *rdev,
294 struct radeon_fence *fence)
296 struct radeon_ring *ring = &rdev->ring[fence->ring];
297 u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
299 /* write the fence */
300 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0));
301 radeon_ring_write(ring, addr & 0xfffffffc);
302 radeon_ring_write(ring, (upper_32_bits(addr) & 0xff));
303 radeon_ring_write(ring, lower_32_bits(fence->seq));
304 /* generate an interrupt */
305 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_TRAP, 0, 0, 0));
309 * r600_dma_semaphore_ring_emit - emit a semaphore on the dma ring
311 * @rdev: radeon_device pointer
312 * @ring: radeon_ring structure holding ring information
313 * @semaphore: radeon semaphore object
314 * @emit_wait: wait or signal semaphore
316 * Add a DMA semaphore packet to the ring wait on or signal
317 * other rings (r6xx-SI).
319 bool r600_dma_semaphore_ring_emit(struct radeon_device *rdev,
320 struct radeon_ring *ring,
321 struct radeon_semaphore *semaphore,
324 u64 addr = semaphore->gpu_addr;
325 u32 s = emit_wait ? 0 : 1;
327 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SEMAPHORE, 0, s, 0));
328 radeon_ring_write(ring, addr & 0xfffffffc);
329 radeon_ring_write(ring, upper_32_bits(addr) & 0xff);
335 * r600_dma_ib_test - test an IB on the DMA engine
337 * @rdev: radeon_device pointer
338 * @ring: radeon_ring structure holding ring information
340 * Test a simple IB in the DMA ring (r6xx-SI).
341 * Returns 0 on success, error on failure.
343 int r600_dma_ib_test(struct radeon_device *rdev, struct radeon_ring *ring)
348 void __iomem *ptr = (void *)rdev->vram_scratch.ptr;
352 DRM_ERROR("invalid vram scratch pointer\n");
359 r = radeon_ib_get(rdev, ring->idx, &ib, NULL, 256);
361 DRM_ERROR("radeon: failed to get ib (%d).\n", r);
365 ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1);
366 ib.ptr[1] = rdev->vram_scratch.gpu_addr & 0xfffffffc;
367 ib.ptr[2] = upper_32_bits(rdev->vram_scratch.gpu_addr) & 0xff;
368 ib.ptr[3] = 0xDEADBEEF;
371 r = radeon_ib_schedule(rdev, &ib, NULL);
373 radeon_ib_free(rdev, &ib);
374 DRM_ERROR("radeon: failed to schedule ib (%d).\n", r);
377 r = radeon_fence_wait(ib.fence, false);
379 DRM_ERROR("radeon: fence wait failed (%d).\n", r);
382 for (i = 0; i < rdev->usec_timeout; i++) {
384 if (tmp == 0xDEADBEEF)
388 if (i < rdev->usec_timeout) {
389 DRM_INFO("ib test on ring %d succeeded in %u usecs\n", ib.fence->ring, i);
391 DRM_ERROR("radeon: ib test failed (0x%08X)\n", tmp);
394 radeon_ib_free(rdev, &ib);
399 * r600_dma_ring_ib_execute - Schedule an IB on the DMA engine
401 * @rdev: radeon_device pointer
402 * @ib: IB object to schedule
404 * Schedule an IB in the DMA ring (r6xx-r7xx).
406 void r600_dma_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
408 struct radeon_ring *ring = &rdev->ring[ib->ring];
410 if (rdev->wb.enabled) {
411 u32 next_rptr = ring->wptr + 4;
412 while ((next_rptr & 7) != 5)
415 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1));
416 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
417 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xff);
418 radeon_ring_write(ring, next_rptr);
421 /* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
422 * Pad as necessary with NOPs.
424 while ((ring->wptr & 7) != 5)
425 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0));
426 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_INDIRECT_BUFFER, 0, 0, 0));
427 radeon_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
428 radeon_ring_write(ring, (ib->length_dw << 16) | (upper_32_bits(ib->gpu_addr) & 0xFF));
433 * r600_copy_dma - copy pages using the DMA engine
435 * @rdev: radeon_device pointer
436 * @src_offset: src GPU address
437 * @dst_offset: dst GPU address
438 * @num_gpu_pages: number of GPU pages to xfer
439 * @fence: radeon fence object
441 * Copy GPU paging using the DMA engine (r6xx).
442 * Used by the radeon ttm implementation to move pages if
443 * registered as the asic copy callback.
445 int r600_copy_dma(struct radeon_device *rdev,
446 uint64_t src_offset, uint64_t dst_offset,
447 unsigned num_gpu_pages,
448 struct radeon_fence **fence)
450 struct radeon_semaphore *sem = NULL;
451 int ring_index = rdev->asic->copy.dma_ring_index;
452 struct radeon_ring *ring = &rdev->ring[ring_index];
453 u32 size_in_dw, cur_size_in_dw;
457 r = radeon_semaphore_create(rdev, &sem);
459 DRM_ERROR("radeon: moving bo (%d).\n", r);
463 size_in_dw = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT) / 4;
464 num_loops = DIV_ROUND_UP(size_in_dw, 0xFFFE);
465 r = radeon_ring_lock(rdev, ring, num_loops * 4 + 8);
467 DRM_ERROR("radeon: moving bo (%d).\n", r);
468 radeon_semaphore_free(rdev, &sem, NULL);
472 radeon_semaphore_sync_to(sem, *fence);
473 radeon_semaphore_sync_rings(rdev, sem, ring->idx);
475 for (i = 0; i < num_loops; i++) {
476 cur_size_in_dw = size_in_dw;
477 if (cur_size_in_dw > 0xFFFE)
478 cur_size_in_dw = 0xFFFE;
479 size_in_dw -= cur_size_in_dw;
480 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_COPY, 0, 0, cur_size_in_dw));
481 radeon_ring_write(ring, dst_offset & 0xfffffffc);
482 radeon_ring_write(ring, src_offset & 0xfffffffc);
483 radeon_ring_write(ring, (((upper_32_bits(dst_offset) & 0xff) << 16) |
484 (upper_32_bits(src_offset) & 0xff)));
485 src_offset += cur_size_in_dw * 4;
486 dst_offset += cur_size_in_dw * 4;
489 r = radeon_fence_emit(rdev, fence, ring->idx);
491 radeon_ring_unlock_undo(rdev, ring);
495 radeon_ring_unlock_commit(rdev, ring);
496 radeon_semaphore_free(rdev, &sem, *fence);