xref: /linux/drivers/gpu/drm/radeon/cik_sdma.c (revision 3932b9ca55b0be314a36d3e84faff3e823c081f5)
1 /*
2  * Copyright 2013 Advanced Micro Devices, Inc.
3  *
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:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
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.
21  *
22  * Authors: Alex Deucher
23  */
24 #include <linux/firmware.h>
25 #include <drm/drmP.h>
26 #include "radeon.h"
27 #include "radeon_ucode.h"
28 #include "radeon_asic.h"
29 #include "radeon_trace.h"
30 #include "cikd.h"
31 
32 /* sdma */
33 #define CIK_SDMA_UCODE_SIZE 1050
34 #define CIK_SDMA_UCODE_VERSION 64
35 
36 u32 cik_gpu_check_soft_reset(struct radeon_device *rdev);
37 
38 /*
39  * sDMA - System DMA
40  * Starting with CIK, the GPU has new asynchronous
41  * DMA engines.  These engines are used for compute
42  * and gfx.  There are two DMA engines (SDMA0, SDMA1)
43  * and each one supports 1 ring buffer used for gfx
44  * and 2 queues used for compute.
45  *
46  * The programming model is very similar to the CP
47  * (ring buffer, IBs, etc.), but sDMA has it's own
48  * packet format that is different from the PM4 format
49  * used by the CP. sDMA supports copying data, writing
50  * embedded data, solid fills, and a number of other
51  * things.  It also has support for tiling/detiling of
52  * buffers.
53  */
54 
55 /**
56  * cik_sdma_get_rptr - get the current read pointer
57  *
58  * @rdev: radeon_device pointer
59  * @ring: radeon ring pointer
60  *
61  * Get the current rptr from the hardware (CIK+).
62  */
63 uint32_t cik_sdma_get_rptr(struct radeon_device *rdev,
64 			   struct radeon_ring *ring)
65 {
66 	u32 rptr, reg;
67 
68 	if (rdev->wb.enabled) {
69 		rptr = rdev->wb.wb[ring->rptr_offs/4];
70 	} else {
71 		if (ring->idx == R600_RING_TYPE_DMA_INDEX)
72 			reg = SDMA0_GFX_RB_RPTR + SDMA0_REGISTER_OFFSET;
73 		else
74 			reg = SDMA0_GFX_RB_RPTR + SDMA1_REGISTER_OFFSET;
75 
76 		rptr = RREG32(reg);
77 	}
78 
79 	return (rptr & 0x3fffc) >> 2;
80 }
81 
82 /**
83  * cik_sdma_get_wptr - get the current write pointer
84  *
85  * @rdev: radeon_device pointer
86  * @ring: radeon ring pointer
87  *
88  * Get the current wptr from the hardware (CIK+).
89  */
90 uint32_t cik_sdma_get_wptr(struct radeon_device *rdev,
91 			   struct radeon_ring *ring)
92 {
93 	u32 reg;
94 
95 	if (ring->idx == R600_RING_TYPE_DMA_INDEX)
96 		reg = SDMA0_GFX_RB_WPTR + SDMA0_REGISTER_OFFSET;
97 	else
98 		reg = SDMA0_GFX_RB_WPTR + SDMA1_REGISTER_OFFSET;
99 
100 	return (RREG32(reg) & 0x3fffc) >> 2;
101 }
102 
103 /**
104  * cik_sdma_set_wptr - commit the write pointer
105  *
106  * @rdev: radeon_device pointer
107  * @ring: radeon ring pointer
108  *
109  * Write the wptr back to the hardware (CIK+).
110  */
111 void cik_sdma_set_wptr(struct radeon_device *rdev,
112 		       struct radeon_ring *ring)
113 {
114 	u32 reg;
115 
116 	if (ring->idx == R600_RING_TYPE_DMA_INDEX)
117 		reg = SDMA0_GFX_RB_WPTR + SDMA0_REGISTER_OFFSET;
118 	else
119 		reg = SDMA0_GFX_RB_WPTR + SDMA1_REGISTER_OFFSET;
120 
121 	WREG32(reg, (ring->wptr << 2) & 0x3fffc);
122 	(void)RREG32(reg);
123 }
124 
125 /**
126  * cik_sdma_ring_ib_execute - Schedule an IB on the DMA engine
127  *
128  * @rdev: radeon_device pointer
129  * @ib: IB object to schedule
130  *
131  * Schedule an IB in the DMA ring (CIK).
132  */
133 void cik_sdma_ring_ib_execute(struct radeon_device *rdev,
134 			      struct radeon_ib *ib)
135 {
136 	struct radeon_ring *ring = &rdev->ring[ib->ring];
137 	u32 extra_bits = (ib->vm ? ib->vm->id : 0) & 0xf;
138 
139 	if (rdev->wb.enabled) {
140 		u32 next_rptr = ring->wptr + 5;
141 		while ((next_rptr & 7) != 4)
142 			next_rptr++;
143 		next_rptr += 4;
144 		radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
145 		radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
146 		radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr));
147 		radeon_ring_write(ring, 1); /* number of DWs to follow */
148 		radeon_ring_write(ring, next_rptr);
149 	}
150 
151 	/* IB packet must end on a 8 DW boundary */
152 	while ((ring->wptr & 7) != 4)
153 		radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0));
154 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
155 	radeon_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
156 	radeon_ring_write(ring, upper_32_bits(ib->gpu_addr));
157 	radeon_ring_write(ring, ib->length_dw);
158 
159 }
160 
161 /**
162  * cik_sdma_hdp_flush_ring_emit - emit an hdp flush on the DMA ring
163  *
164  * @rdev: radeon_device pointer
165  * @ridx: radeon ring index
166  *
167  * Emit an hdp flush packet on the requested DMA ring.
168  */
169 static void cik_sdma_hdp_flush_ring_emit(struct radeon_device *rdev,
170 					 int ridx)
171 {
172 	struct radeon_ring *ring = &rdev->ring[ridx];
173 	u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
174 			  SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
175 	u32 ref_and_mask;
176 
177 	if (ridx == R600_RING_TYPE_DMA_INDEX)
178 		ref_and_mask = SDMA0;
179 	else
180 		ref_and_mask = SDMA1;
181 
182 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
183 	radeon_ring_write(ring, GPU_HDP_FLUSH_DONE);
184 	radeon_ring_write(ring, GPU_HDP_FLUSH_REQ);
185 	radeon_ring_write(ring, ref_and_mask); /* reference */
186 	radeon_ring_write(ring, ref_and_mask); /* mask */
187 	radeon_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
188 }
189 
190 /**
191  * cik_sdma_fence_ring_emit - emit a fence on the DMA ring
192  *
193  * @rdev: radeon_device pointer
194  * @fence: radeon fence object
195  *
196  * Add a DMA fence packet to the ring to write
197  * the fence seq number and DMA trap packet to generate
198  * an interrupt if needed (CIK).
199  */
200 void cik_sdma_fence_ring_emit(struct radeon_device *rdev,
201 			      struct radeon_fence *fence)
202 {
203 	struct radeon_ring *ring = &rdev->ring[fence->ring];
204 	u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
205 
206 	/* write the fence */
207 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
208 	radeon_ring_write(ring, lower_32_bits(addr));
209 	radeon_ring_write(ring, upper_32_bits(addr));
210 	radeon_ring_write(ring, fence->seq);
211 	/* generate an interrupt */
212 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
213 	/* flush HDP */
214 	cik_sdma_hdp_flush_ring_emit(rdev, fence->ring);
215 }
216 
217 /**
218  * cik_sdma_semaphore_ring_emit - emit a semaphore on the dma ring
219  *
220  * @rdev: radeon_device pointer
221  * @ring: radeon_ring structure holding ring information
222  * @semaphore: radeon semaphore object
223  * @emit_wait: wait or signal semaphore
224  *
225  * Add a DMA semaphore packet to the ring wait on or signal
226  * other rings (CIK).
227  */
228 bool cik_sdma_semaphore_ring_emit(struct radeon_device *rdev,
229 				  struct radeon_ring *ring,
230 				  struct radeon_semaphore *semaphore,
231 				  bool emit_wait)
232 {
233 	u64 addr = semaphore->gpu_addr;
234 	u32 extra_bits = emit_wait ? 0 : SDMA_SEMAPHORE_EXTRA_S;
235 
236 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SEMAPHORE, 0, extra_bits));
237 	radeon_ring_write(ring, addr & 0xfffffff8);
238 	radeon_ring_write(ring, upper_32_bits(addr));
239 
240 	return true;
241 }
242 
243 /**
244  * cik_sdma_gfx_stop - stop the gfx async dma engines
245  *
246  * @rdev: radeon_device pointer
247  *
248  * Stop the gfx async dma ring buffers (CIK).
249  */
250 static void cik_sdma_gfx_stop(struct radeon_device *rdev)
251 {
252 	u32 rb_cntl, reg_offset;
253 	int i;
254 
255 	if ((rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX) ||
256 	    (rdev->asic->copy.copy_ring_index == CAYMAN_RING_TYPE_DMA1_INDEX))
257 		radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
258 
259 	for (i = 0; i < 2; i++) {
260 		if (i == 0)
261 			reg_offset = SDMA0_REGISTER_OFFSET;
262 		else
263 			reg_offset = SDMA1_REGISTER_OFFSET;
264 		rb_cntl = RREG32(SDMA0_GFX_RB_CNTL + reg_offset);
265 		rb_cntl &= ~SDMA_RB_ENABLE;
266 		WREG32(SDMA0_GFX_RB_CNTL + reg_offset, rb_cntl);
267 		WREG32(SDMA0_GFX_IB_CNTL + reg_offset, 0);
268 	}
269 	rdev->ring[R600_RING_TYPE_DMA_INDEX].ready = false;
270 	rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX].ready = false;
271 }
272 
273 /**
274  * cik_sdma_rlc_stop - stop the compute async dma engines
275  *
276  * @rdev: radeon_device pointer
277  *
278  * Stop the compute async dma queues (CIK).
279  */
280 static void cik_sdma_rlc_stop(struct radeon_device *rdev)
281 {
282 	/* XXX todo */
283 }
284 
285 /**
286  * cik_sdma_enable - stop the async dma engines
287  *
288  * @rdev: radeon_device pointer
289  * @enable: enable/disable the DMA MEs.
290  *
291  * Halt or unhalt the async dma engines (CIK).
292  */
293 void cik_sdma_enable(struct radeon_device *rdev, bool enable)
294 {
295 	u32 me_cntl, reg_offset;
296 	int i;
297 
298 	if (enable == false) {
299 		cik_sdma_gfx_stop(rdev);
300 		cik_sdma_rlc_stop(rdev);
301 	}
302 
303 	for (i = 0; i < 2; i++) {
304 		if (i == 0)
305 			reg_offset = SDMA0_REGISTER_OFFSET;
306 		else
307 			reg_offset = SDMA1_REGISTER_OFFSET;
308 		me_cntl = RREG32(SDMA0_ME_CNTL + reg_offset);
309 		if (enable)
310 			me_cntl &= ~SDMA_HALT;
311 		else
312 			me_cntl |= SDMA_HALT;
313 		WREG32(SDMA0_ME_CNTL + reg_offset, me_cntl);
314 	}
315 }
316 
317 /**
318  * cik_sdma_gfx_resume - setup and start the async dma engines
319  *
320  * @rdev: radeon_device pointer
321  *
322  * Set up the gfx DMA ring buffers and enable them (CIK).
323  * Returns 0 for success, error for failure.
324  */
325 static int cik_sdma_gfx_resume(struct radeon_device *rdev)
326 {
327 	struct radeon_ring *ring;
328 	u32 rb_cntl, ib_cntl;
329 	u32 rb_bufsz;
330 	u32 reg_offset, wb_offset;
331 	int i, r;
332 
333 	for (i = 0; i < 2; i++) {
334 		if (i == 0) {
335 			ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
336 			reg_offset = SDMA0_REGISTER_OFFSET;
337 			wb_offset = R600_WB_DMA_RPTR_OFFSET;
338 		} else {
339 			ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
340 			reg_offset = SDMA1_REGISTER_OFFSET;
341 			wb_offset = CAYMAN_WB_DMA1_RPTR_OFFSET;
342 		}
343 
344 		WREG32(SDMA0_SEM_INCOMPLETE_TIMER_CNTL + reg_offset, 0);
345 		WREG32(SDMA0_SEM_WAIT_FAIL_TIMER_CNTL + reg_offset, 0);
346 
347 		/* Set ring buffer size in dwords */
348 		rb_bufsz = order_base_2(ring->ring_size / 4);
349 		rb_cntl = rb_bufsz << 1;
350 #ifdef __BIG_ENDIAN
351 		rb_cntl |= SDMA_RB_SWAP_ENABLE | SDMA_RPTR_WRITEBACK_SWAP_ENABLE;
352 #endif
353 		WREG32(SDMA0_GFX_RB_CNTL + reg_offset, rb_cntl);
354 
355 		/* Initialize the ring buffer's read and write pointers */
356 		WREG32(SDMA0_GFX_RB_RPTR + reg_offset, 0);
357 		WREG32(SDMA0_GFX_RB_WPTR + reg_offset, 0);
358 
359 		/* set the wb address whether it's enabled or not */
360 		WREG32(SDMA0_GFX_RB_RPTR_ADDR_HI + reg_offset,
361 		       upper_32_bits(rdev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
362 		WREG32(SDMA0_GFX_RB_RPTR_ADDR_LO + reg_offset,
363 		       ((rdev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
364 
365 		if (rdev->wb.enabled)
366 			rb_cntl |= SDMA_RPTR_WRITEBACK_ENABLE;
367 
368 		WREG32(SDMA0_GFX_RB_BASE + reg_offset, ring->gpu_addr >> 8);
369 		WREG32(SDMA0_GFX_RB_BASE_HI + reg_offset, ring->gpu_addr >> 40);
370 
371 		ring->wptr = 0;
372 		WREG32(SDMA0_GFX_RB_WPTR + reg_offset, ring->wptr << 2);
373 
374 		/* enable DMA RB */
375 		WREG32(SDMA0_GFX_RB_CNTL + reg_offset, rb_cntl | SDMA_RB_ENABLE);
376 
377 		ib_cntl = SDMA_IB_ENABLE;
378 #ifdef __BIG_ENDIAN
379 		ib_cntl |= SDMA_IB_SWAP_ENABLE;
380 #endif
381 		/* enable DMA IBs */
382 		WREG32(SDMA0_GFX_IB_CNTL + reg_offset, ib_cntl);
383 
384 		ring->ready = true;
385 
386 		r = radeon_ring_test(rdev, ring->idx, ring);
387 		if (r) {
388 			ring->ready = false;
389 			return r;
390 		}
391 	}
392 
393 	if ((rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX) ||
394 	    (rdev->asic->copy.copy_ring_index == CAYMAN_RING_TYPE_DMA1_INDEX))
395 		radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
396 
397 	return 0;
398 }
399 
400 /**
401  * cik_sdma_rlc_resume - setup and start the async dma engines
402  *
403  * @rdev: radeon_device pointer
404  *
405  * Set up the compute DMA queues and enable them (CIK).
406  * Returns 0 for success, error for failure.
407  */
408 static int cik_sdma_rlc_resume(struct radeon_device *rdev)
409 {
410 	/* XXX todo */
411 	return 0;
412 }
413 
414 /**
415  * cik_sdma_load_microcode - load the sDMA ME ucode
416  *
417  * @rdev: radeon_device pointer
418  *
419  * Loads the sDMA0/1 ucode.
420  * Returns 0 for success, -EINVAL if the ucode is not available.
421  */
422 static int cik_sdma_load_microcode(struct radeon_device *rdev)
423 {
424 	int i;
425 
426 	if (!rdev->sdma_fw)
427 		return -EINVAL;
428 
429 	/* halt the MEs */
430 	cik_sdma_enable(rdev, false);
431 
432 	if (rdev->new_fw) {
433 		const struct sdma_firmware_header_v1_0 *hdr =
434 			(const struct sdma_firmware_header_v1_0 *)rdev->sdma_fw->data;
435 		const __le32 *fw_data;
436 		u32 fw_size;
437 
438 		radeon_ucode_print_sdma_hdr(&hdr->header);
439 
440 		/* sdma0 */
441 		fw_data = (const __le32 *)
442 			(rdev->sdma_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
443 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
444 		WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
445 		for (i = 0; i < fw_size; i++)
446 			WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, le32_to_cpup(fw_data++));
447 		WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
448 
449 		/* sdma1 */
450 		fw_data = (const __le32 *)
451 			(rdev->sdma_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
452 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
453 		WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
454 		for (i = 0; i < fw_size; i++)
455 			WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, le32_to_cpup(fw_data++));
456 		WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
457 	} else {
458 		const __be32 *fw_data;
459 
460 		/* sdma0 */
461 		fw_data = (const __be32 *)rdev->sdma_fw->data;
462 		WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
463 		for (i = 0; i < CIK_SDMA_UCODE_SIZE; i++)
464 			WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, be32_to_cpup(fw_data++));
465 		WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
466 
467 		/* sdma1 */
468 		fw_data = (const __be32 *)rdev->sdma_fw->data;
469 		WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
470 		for (i = 0; i < CIK_SDMA_UCODE_SIZE; i++)
471 			WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, be32_to_cpup(fw_data++));
472 		WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
473 	}
474 
475 	WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
476 	WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
477 	return 0;
478 }
479 
480 /**
481  * cik_sdma_resume - setup and start the async dma engines
482  *
483  * @rdev: radeon_device pointer
484  *
485  * Set up the DMA engines and enable them (CIK).
486  * Returns 0 for success, error for failure.
487  */
488 int cik_sdma_resume(struct radeon_device *rdev)
489 {
490 	int r;
491 
492 	r = cik_sdma_load_microcode(rdev);
493 	if (r)
494 		return r;
495 
496 	/* unhalt the MEs */
497 	cik_sdma_enable(rdev, true);
498 
499 	/* start the gfx rings and rlc compute queues */
500 	r = cik_sdma_gfx_resume(rdev);
501 	if (r)
502 		return r;
503 	r = cik_sdma_rlc_resume(rdev);
504 	if (r)
505 		return r;
506 
507 	return 0;
508 }
509 
510 /**
511  * cik_sdma_fini - tear down the async dma engines
512  *
513  * @rdev: radeon_device pointer
514  *
515  * Stop the async dma engines and free the rings (CIK).
516  */
517 void cik_sdma_fini(struct radeon_device *rdev)
518 {
519 	/* halt the MEs */
520 	cik_sdma_enable(rdev, false);
521 	radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX]);
522 	radeon_ring_fini(rdev, &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX]);
523 	/* XXX - compute dma queue tear down */
524 }
525 
526 /**
527  * cik_copy_dma - copy pages using the DMA engine
528  *
529  * @rdev: radeon_device pointer
530  * @src_offset: src GPU address
531  * @dst_offset: dst GPU address
532  * @num_gpu_pages: number of GPU pages to xfer
533  * @fence: radeon fence object
534  *
535  * Copy GPU paging using the DMA engine (CIK).
536  * Used by the radeon ttm implementation to move pages if
537  * registered as the asic copy callback.
538  */
539 int cik_copy_dma(struct radeon_device *rdev,
540 		 uint64_t src_offset, uint64_t dst_offset,
541 		 unsigned num_gpu_pages,
542 		 struct radeon_fence **fence)
543 {
544 	struct radeon_semaphore *sem = NULL;
545 	int ring_index = rdev->asic->copy.dma_ring_index;
546 	struct radeon_ring *ring = &rdev->ring[ring_index];
547 	u32 size_in_bytes, cur_size_in_bytes;
548 	int i, num_loops;
549 	int r = 0;
550 
551 	r = radeon_semaphore_create(rdev, &sem);
552 	if (r) {
553 		DRM_ERROR("radeon: moving bo (%d).\n", r);
554 		return r;
555 	}
556 
557 	size_in_bytes = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT);
558 	num_loops = DIV_ROUND_UP(size_in_bytes, 0x1fffff);
559 	r = radeon_ring_lock(rdev, ring, num_loops * 7 + 14);
560 	if (r) {
561 		DRM_ERROR("radeon: moving bo (%d).\n", r);
562 		radeon_semaphore_free(rdev, &sem, NULL);
563 		return r;
564 	}
565 
566 	radeon_semaphore_sync_to(sem, *fence);
567 	radeon_semaphore_sync_rings(rdev, sem, ring->idx);
568 
569 	for (i = 0; i < num_loops; i++) {
570 		cur_size_in_bytes = size_in_bytes;
571 		if (cur_size_in_bytes > 0x1fffff)
572 			cur_size_in_bytes = 0x1fffff;
573 		size_in_bytes -= cur_size_in_bytes;
574 		radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0));
575 		radeon_ring_write(ring, cur_size_in_bytes);
576 		radeon_ring_write(ring, 0); /* src/dst endian swap */
577 		radeon_ring_write(ring, lower_32_bits(src_offset));
578 		radeon_ring_write(ring, upper_32_bits(src_offset));
579 		radeon_ring_write(ring, lower_32_bits(dst_offset));
580 		radeon_ring_write(ring, upper_32_bits(dst_offset));
581 		src_offset += cur_size_in_bytes;
582 		dst_offset += cur_size_in_bytes;
583 	}
584 
585 	r = radeon_fence_emit(rdev, fence, ring->idx);
586 	if (r) {
587 		radeon_ring_unlock_undo(rdev, ring);
588 		radeon_semaphore_free(rdev, &sem, NULL);
589 		return r;
590 	}
591 
592 	radeon_ring_unlock_commit(rdev, ring, false);
593 	radeon_semaphore_free(rdev, &sem, *fence);
594 
595 	return r;
596 }
597 
598 /**
599  * cik_sdma_ring_test - simple async dma engine test
600  *
601  * @rdev: radeon_device pointer
602  * @ring: radeon_ring structure holding ring information
603  *
604  * Test the DMA engine by writing using it to write an
605  * value to memory. (CIK).
606  * Returns 0 for success, error for failure.
607  */
608 int cik_sdma_ring_test(struct radeon_device *rdev,
609 		       struct radeon_ring *ring)
610 {
611 	unsigned i;
612 	int r;
613 	void __iomem *ptr = (void *)rdev->vram_scratch.ptr;
614 	u32 tmp;
615 
616 	if (!ptr) {
617 		DRM_ERROR("invalid vram scratch pointer\n");
618 		return -EINVAL;
619 	}
620 
621 	tmp = 0xCAFEDEAD;
622 	writel(tmp, ptr);
623 
624 	r = radeon_ring_lock(rdev, ring, 5);
625 	if (r) {
626 		DRM_ERROR("radeon: dma failed to lock ring %d (%d).\n", ring->idx, r);
627 		return r;
628 	}
629 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
630 	radeon_ring_write(ring, rdev->vram_scratch.gpu_addr & 0xfffffffc);
631 	radeon_ring_write(ring, upper_32_bits(rdev->vram_scratch.gpu_addr));
632 	radeon_ring_write(ring, 1); /* number of DWs to follow */
633 	radeon_ring_write(ring, 0xDEADBEEF);
634 	radeon_ring_unlock_commit(rdev, ring, false);
635 
636 	for (i = 0; i < rdev->usec_timeout; i++) {
637 		tmp = readl(ptr);
638 		if (tmp == 0xDEADBEEF)
639 			break;
640 		DRM_UDELAY(1);
641 	}
642 
643 	if (i < rdev->usec_timeout) {
644 		DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
645 	} else {
646 		DRM_ERROR("radeon: ring %d test failed (0x%08X)\n",
647 			  ring->idx, tmp);
648 		r = -EINVAL;
649 	}
650 	return r;
651 }
652 
653 /**
654  * cik_sdma_ib_test - test an IB on the DMA engine
655  *
656  * @rdev: radeon_device pointer
657  * @ring: radeon_ring structure holding ring information
658  *
659  * Test a simple IB in the DMA ring (CIK).
660  * Returns 0 on success, error on failure.
661  */
662 int cik_sdma_ib_test(struct radeon_device *rdev, struct radeon_ring *ring)
663 {
664 	struct radeon_ib ib;
665 	unsigned i;
666 	int r;
667 	void __iomem *ptr = (void *)rdev->vram_scratch.ptr;
668 	u32 tmp = 0;
669 
670 	if (!ptr) {
671 		DRM_ERROR("invalid vram scratch pointer\n");
672 		return -EINVAL;
673 	}
674 
675 	tmp = 0xCAFEDEAD;
676 	writel(tmp, ptr);
677 
678 	r = radeon_ib_get(rdev, ring->idx, &ib, NULL, 256);
679 	if (r) {
680 		DRM_ERROR("radeon: failed to get ib (%d).\n", r);
681 		return r;
682 	}
683 
684 	ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
685 	ib.ptr[1] = rdev->vram_scratch.gpu_addr & 0xfffffffc;
686 	ib.ptr[2] = upper_32_bits(rdev->vram_scratch.gpu_addr);
687 	ib.ptr[3] = 1;
688 	ib.ptr[4] = 0xDEADBEEF;
689 	ib.length_dw = 5;
690 
691 	r = radeon_ib_schedule(rdev, &ib, NULL, false);
692 	if (r) {
693 		radeon_ib_free(rdev, &ib);
694 		DRM_ERROR("radeon: failed to schedule ib (%d).\n", r);
695 		return r;
696 	}
697 	r = radeon_fence_wait(ib.fence, false);
698 	if (r) {
699 		DRM_ERROR("radeon: fence wait failed (%d).\n", r);
700 		return r;
701 	}
702 	for (i = 0; i < rdev->usec_timeout; i++) {
703 		tmp = readl(ptr);
704 		if (tmp == 0xDEADBEEF)
705 			break;
706 		DRM_UDELAY(1);
707 	}
708 	if (i < rdev->usec_timeout) {
709 		DRM_INFO("ib test on ring %d succeeded in %u usecs\n", ib.fence->ring, i);
710 	} else {
711 		DRM_ERROR("radeon: ib test failed (0x%08X)\n", tmp);
712 		r = -EINVAL;
713 	}
714 	radeon_ib_free(rdev, &ib);
715 	return r;
716 }
717 
718 /**
719  * cik_sdma_is_lockup - Check if the DMA engine is locked up
720  *
721  * @rdev: radeon_device pointer
722  * @ring: radeon_ring structure holding ring information
723  *
724  * Check if the async DMA engine is locked up (CIK).
725  * Returns true if the engine appears to be locked up, false if not.
726  */
727 bool cik_sdma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
728 {
729 	u32 reset_mask = cik_gpu_check_soft_reset(rdev);
730 	u32 mask;
731 
732 	if (ring->idx == R600_RING_TYPE_DMA_INDEX)
733 		mask = RADEON_RESET_DMA;
734 	else
735 		mask = RADEON_RESET_DMA1;
736 
737 	if (!(reset_mask & mask)) {
738 		radeon_ring_lockup_update(rdev, ring);
739 		return false;
740 	}
741 	return radeon_ring_test_lockup(rdev, ring);
742 }
743 
744 /**
745  * cik_sdma_vm_copy_pages - update PTEs by copying them from the GART
746  *
747  * @rdev: radeon_device pointer
748  * @ib: indirect buffer to fill with commands
749  * @pe: addr of the page entry
750  * @src: src addr to copy from
751  * @count: number of page entries to update
752  *
753  * Update PTEs by copying them from the GART using sDMA (CIK).
754  */
755 void cik_sdma_vm_copy_pages(struct radeon_device *rdev,
756 			    struct radeon_ib *ib,
757 			    uint64_t pe, uint64_t src,
758 			    unsigned count)
759 {
760 	while (count) {
761 		unsigned bytes = count * 8;
762 		if (bytes > 0x1FFFF8)
763 			bytes = 0x1FFFF8;
764 
765 		ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
766 			SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
767 		ib->ptr[ib->length_dw++] = bytes;
768 		ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
769 		ib->ptr[ib->length_dw++] = lower_32_bits(src);
770 		ib->ptr[ib->length_dw++] = upper_32_bits(src);
771 		ib->ptr[ib->length_dw++] = lower_32_bits(pe);
772 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
773 
774 		pe += bytes;
775 		src += bytes;
776 		count -= bytes / 8;
777 	}
778 }
779 
780 /**
781  * cik_sdma_vm_write_pages - update PTEs by writing them manually
782  *
783  * @rdev: radeon_device pointer
784  * @ib: indirect buffer to fill with commands
785  * @pe: addr of the page entry
786  * @addr: dst addr to write into pe
787  * @count: number of page entries to update
788  * @incr: increase next addr by incr bytes
789  * @flags: access flags
790  *
791  * Update PTEs by writing them manually using sDMA (CIK).
792  */
793 void cik_sdma_vm_write_pages(struct radeon_device *rdev,
794 			     struct radeon_ib *ib,
795 			     uint64_t pe,
796 			     uint64_t addr, unsigned count,
797 			     uint32_t incr, uint32_t flags)
798 {
799 	uint64_t value;
800 	unsigned ndw;
801 
802 	while (count) {
803 		ndw = count * 2;
804 		if (ndw > 0xFFFFE)
805 			ndw = 0xFFFFE;
806 
807 		/* for non-physically contiguous pages (system) */
808 		ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
809 			SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
810 		ib->ptr[ib->length_dw++] = pe;
811 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
812 		ib->ptr[ib->length_dw++] = ndw;
813 		for (; ndw > 0; ndw -= 2, --count, pe += 8) {
814 			if (flags & R600_PTE_SYSTEM) {
815 				value = radeon_vm_map_gart(rdev, addr);
816 				value &= 0xFFFFFFFFFFFFF000ULL;
817 			} else if (flags & R600_PTE_VALID) {
818 				value = addr;
819 			} else {
820 				value = 0;
821 			}
822 			addr += incr;
823 			value |= flags;
824 			ib->ptr[ib->length_dw++] = value;
825 			ib->ptr[ib->length_dw++] = upper_32_bits(value);
826 		}
827 	}
828 }
829 
830 /**
831  * cik_sdma_vm_set_pages - update the page tables using sDMA
832  *
833  * @rdev: radeon_device pointer
834  * @ib: indirect buffer to fill with commands
835  * @pe: addr of the page entry
836  * @addr: dst addr to write into pe
837  * @count: number of page entries to update
838  * @incr: increase next addr by incr bytes
839  * @flags: access flags
840  *
841  * Update the page tables using sDMA (CIK).
842  */
843 void cik_sdma_vm_set_pages(struct radeon_device *rdev,
844 			   struct radeon_ib *ib,
845 			   uint64_t pe,
846 			   uint64_t addr, unsigned count,
847 			   uint32_t incr, uint32_t flags)
848 {
849 	uint64_t value;
850 	unsigned ndw;
851 
852 	while (count) {
853 		ndw = count;
854 		if (ndw > 0x7FFFF)
855 			ndw = 0x7FFFF;
856 
857 		if (flags & R600_PTE_VALID)
858 			value = addr;
859 		else
860 			value = 0;
861 
862 		/* for physically contiguous pages (vram) */
863 		ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
864 		ib->ptr[ib->length_dw++] = pe; /* dst addr */
865 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
866 		ib->ptr[ib->length_dw++] = flags; /* mask */
867 		ib->ptr[ib->length_dw++] = 0;
868 		ib->ptr[ib->length_dw++] = value; /* value */
869 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
870 		ib->ptr[ib->length_dw++] = incr; /* increment size */
871 		ib->ptr[ib->length_dw++] = 0;
872 		ib->ptr[ib->length_dw++] = ndw; /* number of entries */
873 
874 		pe += ndw * 8;
875 		addr += ndw * incr;
876 		count -= ndw;
877 	}
878 }
879 
880 /**
881  * cik_sdma_vm_pad_ib - pad the IB to the required number of dw
882  *
883  * @ib: indirect buffer to fill with padding
884  *
885  */
886 void cik_sdma_vm_pad_ib(struct radeon_ib *ib)
887 {
888 	while (ib->length_dw & 0x7)
889 		ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
890 }
891 
892 /**
893  * cik_dma_vm_flush - cik vm flush using sDMA
894  *
895  * @rdev: radeon_device pointer
896  *
897  * Update the page table base and flush the VM TLB
898  * using sDMA (CIK).
899  */
900 void cik_dma_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
901 {
902 	struct radeon_ring *ring = &rdev->ring[ridx];
903 
904 	if (vm == NULL)
905 		return;
906 
907 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
908 	if (vm->id < 8) {
909 		radeon_ring_write(ring, (VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2)) >> 2);
910 	} else {
911 		radeon_ring_write(ring, (VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm->id - 8) << 2)) >> 2);
912 	}
913 	radeon_ring_write(ring, vm->pd_gpu_addr >> 12);
914 
915 	/* update SH_MEM_* regs */
916 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
917 	radeon_ring_write(ring, SRBM_GFX_CNTL >> 2);
918 	radeon_ring_write(ring, VMID(vm->id));
919 
920 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
921 	radeon_ring_write(ring, SH_MEM_BASES >> 2);
922 	radeon_ring_write(ring, 0);
923 
924 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
925 	radeon_ring_write(ring, SH_MEM_CONFIG >> 2);
926 	radeon_ring_write(ring, 0);
927 
928 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
929 	radeon_ring_write(ring, SH_MEM_APE1_BASE >> 2);
930 	radeon_ring_write(ring, 1);
931 
932 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
933 	radeon_ring_write(ring, SH_MEM_APE1_LIMIT >> 2);
934 	radeon_ring_write(ring, 0);
935 
936 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
937 	radeon_ring_write(ring, SRBM_GFX_CNTL >> 2);
938 	radeon_ring_write(ring, VMID(0));
939 
940 	/* flush HDP */
941 	cik_sdma_hdp_flush_ring_emit(rdev, ridx);
942 
943 	/* flush TLB */
944 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
945 	radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
946 	radeon_ring_write(ring, 1 << vm->id);
947 }
948 
949