xref: /linux/drivers/gpu/drm/radeon/cik_sdma.c (revision 93d90ad708b8da6efc0e487b66111aa9db7f70c7)
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->ids[ib->ring].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_ctx_switch_enable - enable/disable sdma engine preemption
287  *
288  * @rdev: radeon_device pointer
289  * @enable: enable/disable preemption.
290  *
291  * Halt or unhalt the async dma engines (CIK).
292  */
293 void cik_sdma_ctx_switch_enable(struct radeon_device *rdev, bool enable)
294 {
295 	uint32_t reg_offset, value;
296 	int i;
297 
298 	for (i = 0; i < 2; i++) {
299 		if (i == 0)
300 			reg_offset = SDMA0_REGISTER_OFFSET;
301 		else
302 			reg_offset = SDMA1_REGISTER_OFFSET;
303 		value = RREG32(SDMA0_CNTL + reg_offset);
304 		if (enable)
305 			value |= AUTO_CTXSW_ENABLE;
306 		else
307 			value &= ~AUTO_CTXSW_ENABLE;
308 		WREG32(SDMA0_CNTL + reg_offset, value);
309 	}
310 }
311 
312 /**
313  * cik_sdma_enable - stop the async dma engines
314  *
315  * @rdev: radeon_device pointer
316  * @enable: enable/disable the DMA MEs.
317  *
318  * Halt or unhalt the async dma engines (CIK).
319  */
320 void cik_sdma_enable(struct radeon_device *rdev, bool enable)
321 {
322 	u32 me_cntl, reg_offset;
323 	int i;
324 
325 	if (enable == false) {
326 		cik_sdma_gfx_stop(rdev);
327 		cik_sdma_rlc_stop(rdev);
328 	}
329 
330 	for (i = 0; i < 2; i++) {
331 		if (i == 0)
332 			reg_offset = SDMA0_REGISTER_OFFSET;
333 		else
334 			reg_offset = SDMA1_REGISTER_OFFSET;
335 		me_cntl = RREG32(SDMA0_ME_CNTL + reg_offset);
336 		if (enable)
337 			me_cntl &= ~SDMA_HALT;
338 		else
339 			me_cntl |= SDMA_HALT;
340 		WREG32(SDMA0_ME_CNTL + reg_offset, me_cntl);
341 	}
342 
343 	cik_sdma_ctx_switch_enable(rdev, enable);
344 }
345 
346 /**
347  * cik_sdma_gfx_resume - setup and start the async dma engines
348  *
349  * @rdev: radeon_device pointer
350  *
351  * Set up the gfx DMA ring buffers and enable them (CIK).
352  * Returns 0 for success, error for failure.
353  */
354 static int cik_sdma_gfx_resume(struct radeon_device *rdev)
355 {
356 	struct radeon_ring *ring;
357 	u32 rb_cntl, ib_cntl;
358 	u32 rb_bufsz;
359 	u32 reg_offset, wb_offset;
360 	int i, r;
361 
362 	for (i = 0; i < 2; i++) {
363 		if (i == 0) {
364 			ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
365 			reg_offset = SDMA0_REGISTER_OFFSET;
366 			wb_offset = R600_WB_DMA_RPTR_OFFSET;
367 		} else {
368 			ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
369 			reg_offset = SDMA1_REGISTER_OFFSET;
370 			wb_offset = CAYMAN_WB_DMA1_RPTR_OFFSET;
371 		}
372 
373 		WREG32(SDMA0_SEM_INCOMPLETE_TIMER_CNTL + reg_offset, 0);
374 		WREG32(SDMA0_SEM_WAIT_FAIL_TIMER_CNTL + reg_offset, 0);
375 
376 		/* Set ring buffer size in dwords */
377 		rb_bufsz = order_base_2(ring->ring_size / 4);
378 		rb_cntl = rb_bufsz << 1;
379 #ifdef __BIG_ENDIAN
380 		rb_cntl |= SDMA_RB_SWAP_ENABLE | SDMA_RPTR_WRITEBACK_SWAP_ENABLE;
381 #endif
382 		WREG32(SDMA0_GFX_RB_CNTL + reg_offset, rb_cntl);
383 
384 		/* Initialize the ring buffer's read and write pointers */
385 		WREG32(SDMA0_GFX_RB_RPTR + reg_offset, 0);
386 		WREG32(SDMA0_GFX_RB_WPTR + reg_offset, 0);
387 
388 		/* set the wb address whether it's enabled or not */
389 		WREG32(SDMA0_GFX_RB_RPTR_ADDR_HI + reg_offset,
390 		       upper_32_bits(rdev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
391 		WREG32(SDMA0_GFX_RB_RPTR_ADDR_LO + reg_offset,
392 		       ((rdev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
393 
394 		if (rdev->wb.enabled)
395 			rb_cntl |= SDMA_RPTR_WRITEBACK_ENABLE;
396 
397 		WREG32(SDMA0_GFX_RB_BASE + reg_offset, ring->gpu_addr >> 8);
398 		WREG32(SDMA0_GFX_RB_BASE_HI + reg_offset, ring->gpu_addr >> 40);
399 
400 		ring->wptr = 0;
401 		WREG32(SDMA0_GFX_RB_WPTR + reg_offset, ring->wptr << 2);
402 
403 		/* enable DMA RB */
404 		WREG32(SDMA0_GFX_RB_CNTL + reg_offset, rb_cntl | SDMA_RB_ENABLE);
405 
406 		ib_cntl = SDMA_IB_ENABLE;
407 #ifdef __BIG_ENDIAN
408 		ib_cntl |= SDMA_IB_SWAP_ENABLE;
409 #endif
410 		/* enable DMA IBs */
411 		WREG32(SDMA0_GFX_IB_CNTL + reg_offset, ib_cntl);
412 
413 		ring->ready = true;
414 
415 		r = radeon_ring_test(rdev, ring->idx, ring);
416 		if (r) {
417 			ring->ready = false;
418 			return r;
419 		}
420 	}
421 
422 	if ((rdev->asic->copy.copy_ring_index == R600_RING_TYPE_DMA_INDEX) ||
423 	    (rdev->asic->copy.copy_ring_index == CAYMAN_RING_TYPE_DMA1_INDEX))
424 		radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
425 
426 	return 0;
427 }
428 
429 /**
430  * cik_sdma_rlc_resume - setup and start the async dma engines
431  *
432  * @rdev: radeon_device pointer
433  *
434  * Set up the compute DMA queues and enable them (CIK).
435  * Returns 0 for success, error for failure.
436  */
437 static int cik_sdma_rlc_resume(struct radeon_device *rdev)
438 {
439 	/* XXX todo */
440 	return 0;
441 }
442 
443 /**
444  * cik_sdma_load_microcode - load the sDMA ME ucode
445  *
446  * @rdev: radeon_device pointer
447  *
448  * Loads the sDMA0/1 ucode.
449  * Returns 0 for success, -EINVAL if the ucode is not available.
450  */
451 static int cik_sdma_load_microcode(struct radeon_device *rdev)
452 {
453 	int i;
454 
455 	if (!rdev->sdma_fw)
456 		return -EINVAL;
457 
458 	/* halt the MEs */
459 	cik_sdma_enable(rdev, false);
460 
461 	if (rdev->new_fw) {
462 		const struct sdma_firmware_header_v1_0 *hdr =
463 			(const struct sdma_firmware_header_v1_0 *)rdev->sdma_fw->data;
464 		const __le32 *fw_data;
465 		u32 fw_size;
466 
467 		radeon_ucode_print_sdma_hdr(&hdr->header);
468 
469 		/* sdma0 */
470 		fw_data = (const __le32 *)
471 			(rdev->sdma_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
472 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
473 		WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
474 		for (i = 0; i < fw_size; i++)
475 			WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, le32_to_cpup(fw_data++));
476 		WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
477 
478 		/* sdma1 */
479 		fw_data = (const __le32 *)
480 			(rdev->sdma_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
481 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
482 		WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
483 		for (i = 0; i < fw_size; i++)
484 			WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, le32_to_cpup(fw_data++));
485 		WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
486 	} else {
487 		const __be32 *fw_data;
488 
489 		/* sdma0 */
490 		fw_data = (const __be32 *)rdev->sdma_fw->data;
491 		WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
492 		for (i = 0; i < CIK_SDMA_UCODE_SIZE; i++)
493 			WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, be32_to_cpup(fw_data++));
494 		WREG32(SDMA0_UCODE_DATA + SDMA0_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
495 
496 		/* sdma1 */
497 		fw_data = (const __be32 *)rdev->sdma_fw->data;
498 		WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
499 		for (i = 0; i < CIK_SDMA_UCODE_SIZE; i++)
500 			WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, be32_to_cpup(fw_data++));
501 		WREG32(SDMA0_UCODE_DATA + SDMA1_REGISTER_OFFSET, CIK_SDMA_UCODE_VERSION);
502 	}
503 
504 	WREG32(SDMA0_UCODE_ADDR + SDMA0_REGISTER_OFFSET, 0);
505 	WREG32(SDMA0_UCODE_ADDR + SDMA1_REGISTER_OFFSET, 0);
506 	return 0;
507 }
508 
509 /**
510  * cik_sdma_resume - setup and start the async dma engines
511  *
512  * @rdev: radeon_device pointer
513  *
514  * Set up the DMA engines and enable them (CIK).
515  * Returns 0 for success, error for failure.
516  */
517 int cik_sdma_resume(struct radeon_device *rdev)
518 {
519 	int r;
520 
521 	r = cik_sdma_load_microcode(rdev);
522 	if (r)
523 		return r;
524 
525 	/* unhalt the MEs */
526 	cik_sdma_enable(rdev, true);
527 
528 	/* start the gfx rings and rlc compute queues */
529 	r = cik_sdma_gfx_resume(rdev);
530 	if (r)
531 		return r;
532 	r = cik_sdma_rlc_resume(rdev);
533 	if (r)
534 		return r;
535 
536 	return 0;
537 }
538 
539 /**
540  * cik_sdma_fini - tear down the async dma engines
541  *
542  * @rdev: radeon_device pointer
543  *
544  * Stop the async dma engines and free the rings (CIK).
545  */
546 void cik_sdma_fini(struct radeon_device *rdev)
547 {
548 	/* halt the MEs */
549 	cik_sdma_enable(rdev, false);
550 	radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX]);
551 	radeon_ring_fini(rdev, &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX]);
552 	/* XXX - compute dma queue tear down */
553 }
554 
555 /**
556  * cik_copy_dma - copy pages using the DMA engine
557  *
558  * @rdev: radeon_device pointer
559  * @src_offset: src GPU address
560  * @dst_offset: dst GPU address
561  * @num_gpu_pages: number of GPU pages to xfer
562  * @resv: reservation object to sync to
563  *
564  * Copy GPU paging using the DMA engine (CIK).
565  * Used by the radeon ttm implementation to move pages if
566  * registered as the asic copy callback.
567  */
568 struct radeon_fence *cik_copy_dma(struct radeon_device *rdev,
569 				  uint64_t src_offset, uint64_t dst_offset,
570 				  unsigned num_gpu_pages,
571 				  struct reservation_object *resv)
572 {
573 	struct radeon_fence *fence;
574 	struct radeon_sync sync;
575 	int ring_index = rdev->asic->copy.dma_ring_index;
576 	struct radeon_ring *ring = &rdev->ring[ring_index];
577 	u32 size_in_bytes, cur_size_in_bytes;
578 	int i, num_loops;
579 	int r = 0;
580 
581 	radeon_sync_create(&sync);
582 
583 	size_in_bytes = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT);
584 	num_loops = DIV_ROUND_UP(size_in_bytes, 0x1fffff);
585 	r = radeon_ring_lock(rdev, ring, num_loops * 7 + 14);
586 	if (r) {
587 		DRM_ERROR("radeon: moving bo (%d).\n", r);
588 		radeon_sync_free(rdev, &sync, NULL);
589 		return ERR_PTR(r);
590 	}
591 
592 	radeon_sync_resv(rdev, &sync, resv, false);
593 	radeon_sync_rings(rdev, &sync, ring->idx);
594 
595 	for (i = 0; i < num_loops; i++) {
596 		cur_size_in_bytes = size_in_bytes;
597 		if (cur_size_in_bytes > 0x1fffff)
598 			cur_size_in_bytes = 0x1fffff;
599 		size_in_bytes -= cur_size_in_bytes;
600 		radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0));
601 		radeon_ring_write(ring, cur_size_in_bytes);
602 		radeon_ring_write(ring, 0); /* src/dst endian swap */
603 		radeon_ring_write(ring, lower_32_bits(src_offset));
604 		radeon_ring_write(ring, upper_32_bits(src_offset));
605 		radeon_ring_write(ring, lower_32_bits(dst_offset));
606 		radeon_ring_write(ring, upper_32_bits(dst_offset));
607 		src_offset += cur_size_in_bytes;
608 		dst_offset += cur_size_in_bytes;
609 	}
610 
611 	r = radeon_fence_emit(rdev, &fence, ring->idx);
612 	if (r) {
613 		radeon_ring_unlock_undo(rdev, ring);
614 		radeon_sync_free(rdev, &sync, NULL);
615 		return ERR_PTR(r);
616 	}
617 
618 	radeon_ring_unlock_commit(rdev, ring, false);
619 	radeon_sync_free(rdev, &sync, fence);
620 
621 	return fence;
622 }
623 
624 /**
625  * cik_sdma_ring_test - simple async dma engine test
626  *
627  * @rdev: radeon_device pointer
628  * @ring: radeon_ring structure holding ring information
629  *
630  * Test the DMA engine by writing using it to write an
631  * value to memory. (CIK).
632  * Returns 0 for success, error for failure.
633  */
634 int cik_sdma_ring_test(struct radeon_device *rdev,
635 		       struct radeon_ring *ring)
636 {
637 	unsigned i;
638 	int r;
639 	unsigned index;
640 	u32 tmp;
641 	u64 gpu_addr;
642 
643 	if (ring->idx == R600_RING_TYPE_DMA_INDEX)
644 		index = R600_WB_DMA_RING_TEST_OFFSET;
645 	else
646 		index = CAYMAN_WB_DMA1_RING_TEST_OFFSET;
647 
648 	gpu_addr = rdev->wb.gpu_addr + index;
649 
650 	tmp = 0xCAFEDEAD;
651 	rdev->wb.wb[index/4] = cpu_to_le32(tmp);
652 
653 	r = radeon_ring_lock(rdev, ring, 5);
654 	if (r) {
655 		DRM_ERROR("radeon: dma failed to lock ring %d (%d).\n", ring->idx, r);
656 		return r;
657 	}
658 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
659 	radeon_ring_write(ring, lower_32_bits(gpu_addr));
660 	radeon_ring_write(ring, upper_32_bits(gpu_addr));
661 	radeon_ring_write(ring, 1); /* number of DWs to follow */
662 	radeon_ring_write(ring, 0xDEADBEEF);
663 	radeon_ring_unlock_commit(rdev, ring, false);
664 
665 	for (i = 0; i < rdev->usec_timeout; i++) {
666 		tmp = le32_to_cpu(rdev->wb.wb[index/4]);
667 		if (tmp == 0xDEADBEEF)
668 			break;
669 		DRM_UDELAY(1);
670 	}
671 
672 	if (i < rdev->usec_timeout) {
673 		DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
674 	} else {
675 		DRM_ERROR("radeon: ring %d test failed (0x%08X)\n",
676 			  ring->idx, tmp);
677 		r = -EINVAL;
678 	}
679 	return r;
680 }
681 
682 /**
683  * cik_sdma_ib_test - test an IB on the DMA engine
684  *
685  * @rdev: radeon_device pointer
686  * @ring: radeon_ring structure holding ring information
687  *
688  * Test a simple IB in the DMA ring (CIK).
689  * Returns 0 on success, error on failure.
690  */
691 int cik_sdma_ib_test(struct radeon_device *rdev, struct radeon_ring *ring)
692 {
693 	struct radeon_ib ib;
694 	unsigned i;
695 	unsigned index;
696 	int r;
697 	u32 tmp = 0;
698 	u64 gpu_addr;
699 
700 	if (ring->idx == R600_RING_TYPE_DMA_INDEX)
701 		index = R600_WB_DMA_RING_TEST_OFFSET;
702 	else
703 		index = CAYMAN_WB_DMA1_RING_TEST_OFFSET;
704 
705 	gpu_addr = rdev->wb.gpu_addr + index;
706 
707 	tmp = 0xCAFEDEAD;
708 	rdev->wb.wb[index/4] = cpu_to_le32(tmp);
709 
710 	r = radeon_ib_get(rdev, ring->idx, &ib, NULL, 256);
711 	if (r) {
712 		DRM_ERROR("radeon: failed to get ib (%d).\n", r);
713 		return r;
714 	}
715 
716 	ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
717 	ib.ptr[1] = lower_32_bits(gpu_addr);
718 	ib.ptr[2] = upper_32_bits(gpu_addr);
719 	ib.ptr[3] = 1;
720 	ib.ptr[4] = 0xDEADBEEF;
721 	ib.length_dw = 5;
722 
723 	r = radeon_ib_schedule(rdev, &ib, NULL, false);
724 	if (r) {
725 		radeon_ib_free(rdev, &ib);
726 		DRM_ERROR("radeon: failed to schedule ib (%d).\n", r);
727 		return r;
728 	}
729 	r = radeon_fence_wait(ib.fence, false);
730 	if (r) {
731 		DRM_ERROR("radeon: fence wait failed (%d).\n", r);
732 		return r;
733 	}
734 	for (i = 0; i < rdev->usec_timeout; i++) {
735 		tmp = le32_to_cpu(rdev->wb.wb[index/4]);
736 		if (tmp == 0xDEADBEEF)
737 			break;
738 		DRM_UDELAY(1);
739 	}
740 	if (i < rdev->usec_timeout) {
741 		DRM_INFO("ib test on ring %d succeeded in %u usecs\n", ib.fence->ring, i);
742 	} else {
743 		DRM_ERROR("radeon: ib test failed (0x%08X)\n", tmp);
744 		r = -EINVAL;
745 	}
746 	radeon_ib_free(rdev, &ib);
747 	return r;
748 }
749 
750 /**
751  * cik_sdma_is_lockup - Check if the DMA engine is locked up
752  *
753  * @rdev: radeon_device pointer
754  * @ring: radeon_ring structure holding ring information
755  *
756  * Check if the async DMA engine is locked up (CIK).
757  * Returns true if the engine appears to be locked up, false if not.
758  */
759 bool cik_sdma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
760 {
761 	u32 reset_mask = cik_gpu_check_soft_reset(rdev);
762 	u32 mask;
763 
764 	if (ring->idx == R600_RING_TYPE_DMA_INDEX)
765 		mask = RADEON_RESET_DMA;
766 	else
767 		mask = RADEON_RESET_DMA1;
768 
769 	if (!(reset_mask & mask)) {
770 		radeon_ring_lockup_update(rdev, ring);
771 		return false;
772 	}
773 	return radeon_ring_test_lockup(rdev, ring);
774 }
775 
776 /**
777  * cik_sdma_vm_copy_pages - update PTEs by copying them from the GART
778  *
779  * @rdev: radeon_device pointer
780  * @ib: indirect buffer to fill with commands
781  * @pe: addr of the page entry
782  * @src: src addr to copy from
783  * @count: number of page entries to update
784  *
785  * Update PTEs by copying them from the GART using sDMA (CIK).
786  */
787 void cik_sdma_vm_copy_pages(struct radeon_device *rdev,
788 			    struct radeon_ib *ib,
789 			    uint64_t pe, uint64_t src,
790 			    unsigned count)
791 {
792 	while (count) {
793 		unsigned bytes = count * 8;
794 		if (bytes > 0x1FFFF8)
795 			bytes = 0x1FFFF8;
796 
797 		ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
798 			SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
799 		ib->ptr[ib->length_dw++] = bytes;
800 		ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
801 		ib->ptr[ib->length_dw++] = lower_32_bits(src);
802 		ib->ptr[ib->length_dw++] = upper_32_bits(src);
803 		ib->ptr[ib->length_dw++] = lower_32_bits(pe);
804 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
805 
806 		pe += bytes;
807 		src += bytes;
808 		count -= bytes / 8;
809 	}
810 }
811 
812 /**
813  * cik_sdma_vm_write_pages - update PTEs by writing them manually
814  *
815  * @rdev: radeon_device pointer
816  * @ib: indirect buffer to fill with commands
817  * @pe: addr of the page entry
818  * @addr: dst addr to write into pe
819  * @count: number of page entries to update
820  * @incr: increase next addr by incr bytes
821  * @flags: access flags
822  *
823  * Update PTEs by writing them manually using sDMA (CIK).
824  */
825 void cik_sdma_vm_write_pages(struct radeon_device *rdev,
826 			     struct radeon_ib *ib,
827 			     uint64_t pe,
828 			     uint64_t addr, unsigned count,
829 			     uint32_t incr, uint32_t flags)
830 {
831 	uint64_t value;
832 	unsigned ndw;
833 
834 	while (count) {
835 		ndw = count * 2;
836 		if (ndw > 0xFFFFE)
837 			ndw = 0xFFFFE;
838 
839 		/* for non-physically contiguous pages (system) */
840 		ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
841 			SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
842 		ib->ptr[ib->length_dw++] = pe;
843 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
844 		ib->ptr[ib->length_dw++] = ndw;
845 		for (; ndw > 0; ndw -= 2, --count, pe += 8) {
846 			if (flags & R600_PTE_SYSTEM) {
847 				value = radeon_vm_map_gart(rdev, addr);
848 				value &= 0xFFFFFFFFFFFFF000ULL;
849 			} else if (flags & R600_PTE_VALID) {
850 				value = addr;
851 			} else {
852 				value = 0;
853 			}
854 			addr += incr;
855 			value |= flags;
856 			ib->ptr[ib->length_dw++] = value;
857 			ib->ptr[ib->length_dw++] = upper_32_bits(value);
858 		}
859 	}
860 }
861 
862 /**
863  * cik_sdma_vm_set_pages - update the page tables using sDMA
864  *
865  * @rdev: radeon_device pointer
866  * @ib: indirect buffer to fill with commands
867  * @pe: addr of the page entry
868  * @addr: dst addr to write into pe
869  * @count: number of page entries to update
870  * @incr: increase next addr by incr bytes
871  * @flags: access flags
872  *
873  * Update the page tables using sDMA (CIK).
874  */
875 void cik_sdma_vm_set_pages(struct radeon_device *rdev,
876 			   struct radeon_ib *ib,
877 			   uint64_t pe,
878 			   uint64_t addr, unsigned count,
879 			   uint32_t incr, uint32_t flags)
880 {
881 	uint64_t value;
882 	unsigned ndw;
883 
884 	while (count) {
885 		ndw = count;
886 		if (ndw > 0x7FFFF)
887 			ndw = 0x7FFFF;
888 
889 		if (flags & R600_PTE_VALID)
890 			value = addr;
891 		else
892 			value = 0;
893 
894 		/* for physically contiguous pages (vram) */
895 		ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
896 		ib->ptr[ib->length_dw++] = pe; /* dst addr */
897 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
898 		ib->ptr[ib->length_dw++] = flags; /* mask */
899 		ib->ptr[ib->length_dw++] = 0;
900 		ib->ptr[ib->length_dw++] = value; /* value */
901 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
902 		ib->ptr[ib->length_dw++] = incr; /* increment size */
903 		ib->ptr[ib->length_dw++] = 0;
904 		ib->ptr[ib->length_dw++] = ndw; /* number of entries */
905 
906 		pe += ndw * 8;
907 		addr += ndw * incr;
908 		count -= ndw;
909 	}
910 }
911 
912 /**
913  * cik_sdma_vm_pad_ib - pad the IB to the required number of dw
914  *
915  * @ib: indirect buffer to fill with padding
916  *
917  */
918 void cik_sdma_vm_pad_ib(struct radeon_ib *ib)
919 {
920 	while (ib->length_dw & 0x7)
921 		ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
922 }
923 
924 /**
925  * cik_dma_vm_flush - cik vm flush using sDMA
926  *
927  * @rdev: radeon_device pointer
928  *
929  * Update the page table base and flush the VM TLB
930  * using sDMA (CIK).
931  */
932 void cik_dma_vm_flush(struct radeon_device *rdev, struct radeon_ring *ring,
933 		      unsigned vm_id, uint64_t pd_addr)
934 {
935 	u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
936 			  SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
937 
938 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
939 	if (vm_id < 8) {
940 		radeon_ring_write(ring, (VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm_id << 2)) >> 2);
941 	} else {
942 		radeon_ring_write(ring, (VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm_id - 8) << 2)) >> 2);
943 	}
944 	radeon_ring_write(ring, pd_addr >> 12);
945 
946 	/* update SH_MEM_* regs */
947 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
948 	radeon_ring_write(ring, SRBM_GFX_CNTL >> 2);
949 	radeon_ring_write(ring, VMID(vm_id));
950 
951 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
952 	radeon_ring_write(ring, SH_MEM_BASES >> 2);
953 	radeon_ring_write(ring, 0);
954 
955 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
956 	radeon_ring_write(ring, SH_MEM_CONFIG >> 2);
957 	radeon_ring_write(ring, 0);
958 
959 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
960 	radeon_ring_write(ring, SH_MEM_APE1_BASE >> 2);
961 	radeon_ring_write(ring, 1);
962 
963 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
964 	radeon_ring_write(ring, SH_MEM_APE1_LIMIT >> 2);
965 	radeon_ring_write(ring, 0);
966 
967 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
968 	radeon_ring_write(ring, SRBM_GFX_CNTL >> 2);
969 	radeon_ring_write(ring, VMID(0));
970 
971 	/* flush HDP */
972 	cik_sdma_hdp_flush_ring_emit(rdev, ring->idx);
973 
974 	/* flush TLB */
975 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
976 	radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
977 	radeon_ring_write(ring, 1 << vm_id);
978 
979 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
980 	radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
981 	radeon_ring_write(ring, 0);
982 	radeon_ring_write(ring, 0); /* reference */
983 	radeon_ring_write(ring, 0); /* mask */
984 	radeon_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
985 }
986 
987