1 /*
2  * Copyright 2011 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Christian König <deathsimple@vodafone.de>
29  */
30 
31 #include <linux/firmware.h>
32 #include <linux/module.h>
33 #include <drm/drmP.h>
34 #include <drm/drm.h>
35 
36 #include "radeon.h"
37 #include "r600d.h"
38 
39 /* 1 second timeout */
40 #define UVD_IDLE_TIMEOUT_MS	1000
41 
42 /* Firmware Names */
43 #define FIRMWARE_RV710		"radeon/RV710_uvd.bin"
44 #define FIRMWARE_CYPRESS	"radeon/CYPRESS_uvd.bin"
45 #define FIRMWARE_SUMO		"radeon/SUMO_uvd.bin"
46 #define FIRMWARE_TAHITI		"radeon/TAHITI_uvd.bin"
47 #define FIRMWARE_BONAIRE	"radeon/BONAIRE_uvd.bin"
48 
49 MODULE_FIRMWARE(FIRMWARE_RV710);
50 MODULE_FIRMWARE(FIRMWARE_CYPRESS);
51 MODULE_FIRMWARE(FIRMWARE_SUMO);
52 MODULE_FIRMWARE(FIRMWARE_TAHITI);
53 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
54 
55 static void radeon_uvd_idle_work_handler(struct work_struct *work);
56 
57 int radeon_uvd_init(struct radeon_device *rdev)
58 {
59 	unsigned long bo_size;
60 	const char *fw_name;
61 	int i, r;
62 
63 	INIT_DELAYED_WORK(&rdev->uvd.idle_work, radeon_uvd_idle_work_handler);
64 
65 	switch (rdev->family) {
66 	case CHIP_RV710:
67 	case CHIP_RV730:
68 	case CHIP_RV740:
69 		fw_name = FIRMWARE_RV710;
70 		break;
71 
72 	case CHIP_CYPRESS:
73 	case CHIP_HEMLOCK:
74 	case CHIP_JUNIPER:
75 	case CHIP_REDWOOD:
76 	case CHIP_CEDAR:
77 		fw_name = FIRMWARE_CYPRESS;
78 		break;
79 
80 	case CHIP_SUMO:
81 	case CHIP_SUMO2:
82 	case CHIP_PALM:
83 	case CHIP_CAYMAN:
84 	case CHIP_BARTS:
85 	case CHIP_TURKS:
86 	case CHIP_CAICOS:
87 		fw_name = FIRMWARE_SUMO;
88 		break;
89 
90 	case CHIP_TAHITI:
91 	case CHIP_VERDE:
92 	case CHIP_PITCAIRN:
93 	case CHIP_ARUBA:
94 		fw_name = FIRMWARE_TAHITI;
95 		break;
96 
97 	case CHIP_BONAIRE:
98 	case CHIP_KABINI:
99 	case CHIP_KAVERI:
100 		fw_name = FIRMWARE_BONAIRE;
101 		break;
102 
103 	default:
104 		return -EINVAL;
105 	}
106 
107 	r = request_firmware(&rdev->uvd_fw, fw_name, rdev->dev);
108 	if (r) {
109 		dev_err(rdev->dev, "radeon_uvd: Can't load firmware \"%s\"\n",
110 			fw_name);
111 		return r;
112 	}
113 
114 	bo_size = RADEON_GPU_PAGE_ALIGN(rdev->uvd_fw->size + 8) +
115 		  RADEON_UVD_STACK_SIZE + RADEON_UVD_HEAP_SIZE;
116 	r = radeon_bo_create(rdev, bo_size, PAGE_SIZE, true,
117 			     RADEON_GEM_DOMAIN_VRAM, NULL, &rdev->uvd.vcpu_bo);
118 	if (r) {
119 		dev_err(rdev->dev, "(%d) failed to allocate UVD bo\n", r);
120 		return r;
121 	}
122 
123 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
124 	if (r) {
125 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
126 		dev_err(rdev->dev, "(%d) failed to reserve UVD bo\n", r);
127 		return r;
128 	}
129 
130 	r = radeon_bo_pin(rdev->uvd.vcpu_bo, RADEON_GEM_DOMAIN_VRAM,
131 			  &rdev->uvd.gpu_addr);
132 	if (r) {
133 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
134 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
135 		dev_err(rdev->dev, "(%d) UVD bo pin failed\n", r);
136 		return r;
137 	}
138 
139 	r = radeon_bo_kmap(rdev->uvd.vcpu_bo, &rdev->uvd.cpu_addr);
140 	if (r) {
141 		dev_err(rdev->dev, "(%d) UVD map failed\n", r);
142 		return r;
143 	}
144 
145 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
146 
147 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
148 		atomic_set(&rdev->uvd.handles[i], 0);
149 		rdev->uvd.filp[i] = NULL;
150 	}
151 
152 	return 0;
153 }
154 
155 void radeon_uvd_fini(struct radeon_device *rdev)
156 {
157 	int r;
158 
159 	if (rdev->uvd.vcpu_bo == NULL)
160 		return;
161 
162 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
163 	if (!r) {
164 		radeon_bo_kunmap(rdev->uvd.vcpu_bo);
165 		radeon_bo_unpin(rdev->uvd.vcpu_bo);
166 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
167 	}
168 
169 	radeon_bo_unref(&rdev->uvd.vcpu_bo);
170 
171 	release_firmware(rdev->uvd_fw);
172 }
173 
174 int radeon_uvd_suspend(struct radeon_device *rdev)
175 {
176 	unsigned size;
177 	void *ptr;
178 	int i;
179 
180 	if (rdev->uvd.vcpu_bo == NULL)
181 		return 0;
182 
183 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i)
184 		if (atomic_read(&rdev->uvd.handles[i]))
185 			break;
186 
187 	if (i == RADEON_MAX_UVD_HANDLES)
188 		return 0;
189 
190 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
191 	size -= rdev->uvd_fw->size;
192 
193 	ptr = rdev->uvd.cpu_addr;
194 	ptr += rdev->uvd_fw->size;
195 
196 	rdev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
197 	memcpy(rdev->uvd.saved_bo, ptr, size);
198 
199 	return 0;
200 }
201 
202 int radeon_uvd_resume(struct radeon_device *rdev)
203 {
204 	unsigned size;
205 	void *ptr;
206 
207 	if (rdev->uvd.vcpu_bo == NULL)
208 		return -EINVAL;
209 
210 	memcpy(rdev->uvd.cpu_addr, rdev->uvd_fw->data, rdev->uvd_fw->size);
211 
212 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
213 	size -= rdev->uvd_fw->size;
214 
215 	ptr = rdev->uvd.cpu_addr;
216 	ptr += rdev->uvd_fw->size;
217 
218 	if (rdev->uvd.saved_bo != NULL) {
219 		memcpy(ptr, rdev->uvd.saved_bo, size);
220 		kfree(rdev->uvd.saved_bo);
221 		rdev->uvd.saved_bo = NULL;
222 	} else
223 		memset(ptr, 0, size);
224 
225 	return 0;
226 }
227 
228 void radeon_uvd_force_into_uvd_segment(struct radeon_bo *rbo)
229 {
230 	rbo->placement.fpfn = 0 >> PAGE_SHIFT;
231 	rbo->placement.lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
232 }
233 
234 void radeon_uvd_free_handles(struct radeon_device *rdev, struct drm_file *filp)
235 {
236 	int i, r;
237 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
238 		uint32_t handle = atomic_read(&rdev->uvd.handles[i]);
239 		if (handle != 0 && rdev->uvd.filp[i] == filp) {
240 			struct radeon_fence *fence;
241 
242 			r = radeon_uvd_get_destroy_msg(rdev,
243 				R600_RING_TYPE_UVD_INDEX, handle, &fence);
244 			if (r) {
245 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
246 				continue;
247 			}
248 
249 			radeon_fence_wait(fence, false);
250 			radeon_fence_unref(&fence);
251 
252 			rdev->uvd.filp[i] = NULL;
253 			atomic_set(&rdev->uvd.handles[i], 0);
254 		}
255 	}
256 }
257 
258 static int radeon_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
259 {
260 	unsigned stream_type = msg[4];
261 	unsigned width = msg[6];
262 	unsigned height = msg[7];
263 	unsigned dpb_size = msg[9];
264 	unsigned pitch = msg[28];
265 
266 	unsigned width_in_mb = width / 16;
267 	unsigned height_in_mb = ALIGN(height / 16, 2);
268 
269 	unsigned image_size, tmp, min_dpb_size;
270 
271 	image_size = width * height;
272 	image_size += image_size / 2;
273 	image_size = ALIGN(image_size, 1024);
274 
275 	switch (stream_type) {
276 	case 0: /* H264 */
277 
278 		/* reference picture buffer */
279 		min_dpb_size = image_size * 17;
280 
281 		/* macroblock context buffer */
282 		min_dpb_size += width_in_mb * height_in_mb * 17 * 192;
283 
284 		/* IT surface buffer */
285 		min_dpb_size += width_in_mb * height_in_mb * 32;
286 		break;
287 
288 	case 1: /* VC1 */
289 
290 		/* reference picture buffer */
291 		min_dpb_size = image_size * 3;
292 
293 		/* CONTEXT_BUFFER */
294 		min_dpb_size += width_in_mb * height_in_mb * 128;
295 
296 		/* IT surface buffer */
297 		min_dpb_size += width_in_mb * 64;
298 
299 		/* DB surface buffer */
300 		min_dpb_size += width_in_mb * 128;
301 
302 		/* BP */
303 		tmp = max(width_in_mb, height_in_mb);
304 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
305 		break;
306 
307 	case 3: /* MPEG2 */
308 
309 		/* reference picture buffer */
310 		min_dpb_size = image_size * 3;
311 		break;
312 
313 	case 4: /* MPEG4 */
314 
315 		/* reference picture buffer */
316 		min_dpb_size = image_size * 3;
317 
318 		/* CM */
319 		min_dpb_size += width_in_mb * height_in_mb * 64;
320 
321 		/* IT surface buffer */
322 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
323 		break;
324 
325 	default:
326 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
327 		return -EINVAL;
328 	}
329 
330 	if (width > pitch) {
331 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
332 		return -EINVAL;
333 	}
334 
335 	if (dpb_size < min_dpb_size) {
336 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
337 			  dpb_size, min_dpb_size);
338 		return -EINVAL;
339 	}
340 
341 	buf_sizes[0x1] = dpb_size;
342 	buf_sizes[0x2] = image_size;
343 	return 0;
344 }
345 
346 static int radeon_uvd_cs_msg(struct radeon_cs_parser *p, struct radeon_bo *bo,
347 			     unsigned offset, unsigned buf_sizes[])
348 {
349 	int32_t *msg, msg_type, handle;
350 	void *ptr;
351 
352 	int i, r;
353 
354 	if (offset & 0x3F) {
355 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
356 		return -EINVAL;
357 	}
358 
359 	if (bo->tbo.sync_obj) {
360 		r = radeon_fence_wait(bo->tbo.sync_obj, false);
361 		if (r) {
362 			DRM_ERROR("Failed waiting for UVD message (%d)!\n", r);
363 			return r;
364 		}
365 	}
366 
367 	r = radeon_bo_kmap(bo, &ptr);
368 	if (r) {
369 		DRM_ERROR("Failed mapping the UVD message (%d)!\n", r);
370 		return r;
371 	}
372 
373 	msg = ptr + offset;
374 
375 	msg_type = msg[1];
376 	handle = msg[2];
377 
378 	if (handle == 0) {
379 		DRM_ERROR("Invalid UVD handle!\n");
380 		return -EINVAL;
381 	}
382 
383 	if (msg_type == 1) {
384 		/* it's a decode msg, calc buffer sizes */
385 		r = radeon_uvd_cs_msg_decode(msg, buf_sizes);
386 		radeon_bo_kunmap(bo);
387 		if (r)
388 			return r;
389 
390 	} else if (msg_type == 2) {
391 		/* it's a destroy msg, free the handle */
392 		for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i)
393 			atomic_cmpxchg(&p->rdev->uvd.handles[i], handle, 0);
394 		radeon_bo_kunmap(bo);
395 		return 0;
396 	} else {
397 		radeon_bo_kunmap(bo);
398 
399 		if (msg_type != 0) {
400 			DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
401 			return -EINVAL;
402 		}
403 
404 		/* it's a create msg, no special handling needed */
405 	}
406 
407 	/* create or decode, validate the handle */
408 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
409 		if (atomic_read(&p->rdev->uvd.handles[i]) == handle)
410 			return 0;
411 	}
412 
413 	/* handle not found try to alloc a new one */
414 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
415 		if (!atomic_cmpxchg(&p->rdev->uvd.handles[i], 0, handle)) {
416 			p->rdev->uvd.filp[i] = p->filp;
417 			return 0;
418 		}
419 	}
420 
421 	DRM_ERROR("No more free UVD handles!\n");
422 	return -EINVAL;
423 }
424 
425 static int radeon_uvd_cs_reloc(struct radeon_cs_parser *p,
426 			       int data0, int data1,
427 			       unsigned buf_sizes[], bool *has_msg_cmd)
428 {
429 	struct radeon_cs_chunk *relocs_chunk;
430 	struct radeon_cs_reloc *reloc;
431 	unsigned idx, cmd, offset;
432 	uint64_t start, end;
433 	int r;
434 
435 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
436 	offset = radeon_get_ib_value(p, data0);
437 	idx = radeon_get_ib_value(p, data1);
438 	if (idx >= relocs_chunk->length_dw) {
439 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
440 			  idx, relocs_chunk->length_dw);
441 		return -EINVAL;
442 	}
443 
444 	reloc = p->relocs_ptr[(idx / 4)];
445 	start = reloc->lobj.gpu_offset;
446 	end = start + radeon_bo_size(reloc->robj);
447 	start += offset;
448 
449 	p->ib.ptr[data0] = start & 0xFFFFFFFF;
450 	p->ib.ptr[data1] = start >> 32;
451 
452 	cmd = radeon_get_ib_value(p, p->idx) >> 1;
453 
454 	if (cmd < 0x4) {
455 		if ((end - start) < buf_sizes[cmd]) {
456 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
457 				  (unsigned)(end - start), buf_sizes[cmd]);
458 			return -EINVAL;
459 		}
460 
461 	} else if (cmd != 0x100) {
462 		DRM_ERROR("invalid UVD command %X!\n", cmd);
463 		return -EINVAL;
464 	}
465 
466 	if ((start >> 28) != (end >> 28)) {
467 		DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
468 			  start, end);
469 		return -EINVAL;
470 	}
471 
472 	/* TODO: is this still necessary on NI+ ? */
473 	if ((cmd == 0 || cmd == 0x3) &&
474 	    (start >> 28) != (p->rdev->uvd.gpu_addr >> 28)) {
475 		DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
476 			  start, end);
477 		return -EINVAL;
478 	}
479 
480 	if (cmd == 0) {
481 		if (*has_msg_cmd) {
482 			DRM_ERROR("More than one message in a UVD-IB!\n");
483 			return -EINVAL;
484 		}
485 		*has_msg_cmd = true;
486 		r = radeon_uvd_cs_msg(p, reloc->robj, offset, buf_sizes);
487 		if (r)
488 			return r;
489 	} else if (!*has_msg_cmd) {
490 		DRM_ERROR("Message needed before other commands are send!\n");
491 		return -EINVAL;
492 	}
493 
494 	return 0;
495 }
496 
497 static int radeon_uvd_cs_reg(struct radeon_cs_parser *p,
498 			     struct radeon_cs_packet *pkt,
499 			     int *data0, int *data1,
500 			     unsigned buf_sizes[],
501 			     bool *has_msg_cmd)
502 {
503 	int i, r;
504 
505 	p->idx++;
506 	for (i = 0; i <= pkt->count; ++i) {
507 		switch (pkt->reg + i*4) {
508 		case UVD_GPCOM_VCPU_DATA0:
509 			*data0 = p->idx;
510 			break;
511 		case UVD_GPCOM_VCPU_DATA1:
512 			*data1 = p->idx;
513 			break;
514 		case UVD_GPCOM_VCPU_CMD:
515 			r = radeon_uvd_cs_reloc(p, *data0, *data1,
516 						buf_sizes, has_msg_cmd);
517 			if (r)
518 				return r;
519 			break;
520 		case UVD_ENGINE_CNTL:
521 			break;
522 		default:
523 			DRM_ERROR("Invalid reg 0x%X!\n",
524 				  pkt->reg + i*4);
525 			return -EINVAL;
526 		}
527 		p->idx++;
528 	}
529 	return 0;
530 }
531 
532 int radeon_uvd_cs_parse(struct radeon_cs_parser *p)
533 {
534 	struct radeon_cs_packet pkt;
535 	int r, data0 = 0, data1 = 0;
536 
537 	/* does the IB has a msg command */
538 	bool has_msg_cmd = false;
539 
540 	/* minimum buffer sizes */
541 	unsigned buf_sizes[] = {
542 		[0x00000000]	=	2048,
543 		[0x00000001]	=	32 * 1024 * 1024,
544 		[0x00000002]	=	2048 * 1152 * 3,
545 		[0x00000003]	=	2048,
546 	};
547 
548 	if (p->chunks[p->chunk_ib_idx].length_dw % 16) {
549 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
550 			  p->chunks[p->chunk_ib_idx].length_dw);
551 		return -EINVAL;
552 	}
553 
554 	if (p->chunk_relocs_idx == -1) {
555 		DRM_ERROR("No relocation chunk !\n");
556 		return -EINVAL;
557 	}
558 
559 
560 	do {
561 		r = radeon_cs_packet_parse(p, &pkt, p->idx);
562 		if (r)
563 			return r;
564 		switch (pkt.type) {
565 		case RADEON_PACKET_TYPE0:
566 			r = radeon_uvd_cs_reg(p, &pkt, &data0, &data1,
567 					      buf_sizes, &has_msg_cmd);
568 			if (r)
569 				return r;
570 			break;
571 		case RADEON_PACKET_TYPE2:
572 			p->idx += pkt.count + 2;
573 			break;
574 		default:
575 			DRM_ERROR("Unknown packet type %d !\n", pkt.type);
576 			return -EINVAL;
577 		}
578 	} while (p->idx < p->chunks[p->chunk_ib_idx].length_dw);
579 
580 	if (!has_msg_cmd) {
581 		DRM_ERROR("UVD-IBs need a msg command!\n");
582 		return -EINVAL;
583 	}
584 
585 	return 0;
586 }
587 
588 static int radeon_uvd_send_msg(struct radeon_device *rdev,
589 			       int ring, struct radeon_bo *bo,
590 			       struct radeon_fence **fence)
591 {
592 	struct ttm_validate_buffer tv;
593 	struct ww_acquire_ctx ticket;
594 	struct list_head head;
595 	struct radeon_ib ib;
596 	uint64_t addr;
597 	int i, r;
598 
599 	memset(&tv, 0, sizeof(tv));
600 	tv.bo = &bo->tbo;
601 
602 	INIT_LIST_HEAD(&head);
603 	list_add(&tv.head, &head);
604 
605 	r = ttm_eu_reserve_buffers(&ticket, &head);
606 	if (r)
607 		return r;
608 
609 	radeon_ttm_placement_from_domain(bo, RADEON_GEM_DOMAIN_VRAM);
610 	radeon_uvd_force_into_uvd_segment(bo);
611 
612 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
613 	if (r)
614 		goto err;
615 
616 	r = radeon_ib_get(rdev, ring, &ib, NULL, 16);
617 	if (r)
618 		goto err;
619 
620 	addr = radeon_bo_gpu_offset(bo);
621 	ib.ptr[0] = PACKET0(UVD_GPCOM_VCPU_DATA0, 0);
622 	ib.ptr[1] = addr;
623 	ib.ptr[2] = PACKET0(UVD_GPCOM_VCPU_DATA1, 0);
624 	ib.ptr[3] = addr >> 32;
625 	ib.ptr[4] = PACKET0(UVD_GPCOM_VCPU_CMD, 0);
626 	ib.ptr[5] = 0;
627 	for (i = 6; i < 16; ++i)
628 		ib.ptr[i] = PACKET2(0);
629 	ib.length_dw = 16;
630 
631 	r = radeon_ib_schedule(rdev, &ib, NULL);
632 	if (r)
633 		goto err;
634 	ttm_eu_fence_buffer_objects(&ticket, &head, ib.fence);
635 
636 	if (fence)
637 		*fence = radeon_fence_ref(ib.fence);
638 
639 	radeon_ib_free(rdev, &ib);
640 	radeon_bo_unref(&bo);
641 	return 0;
642 
643 err:
644 	ttm_eu_backoff_reservation(&ticket, &head);
645 	return r;
646 }
647 
648 /* multiple fence commands without any stream commands in between can
649    crash the vcpu so just try to emmit a dummy create/destroy msg to
650    avoid this */
651 int radeon_uvd_get_create_msg(struct radeon_device *rdev, int ring,
652 			      uint32_t handle, struct radeon_fence **fence)
653 {
654 	struct radeon_bo *bo;
655 	uint32_t *msg;
656 	int r, i;
657 
658 	r = radeon_bo_create(rdev, 1024, PAGE_SIZE, true,
659 			     RADEON_GEM_DOMAIN_VRAM, NULL, &bo);
660 	if (r)
661 		return r;
662 
663 	r = radeon_bo_reserve(bo, false);
664 	if (r) {
665 		radeon_bo_unref(&bo);
666 		return r;
667 	}
668 
669 	r = radeon_bo_kmap(bo, (void **)&msg);
670 	if (r) {
671 		radeon_bo_unreserve(bo);
672 		radeon_bo_unref(&bo);
673 		return r;
674 	}
675 
676 	/* stitch together an UVD create msg */
677 	msg[0] = cpu_to_le32(0x00000de4);
678 	msg[1] = cpu_to_le32(0x00000000);
679 	msg[2] = cpu_to_le32(handle);
680 	msg[3] = cpu_to_le32(0x00000000);
681 	msg[4] = cpu_to_le32(0x00000000);
682 	msg[5] = cpu_to_le32(0x00000000);
683 	msg[6] = cpu_to_le32(0x00000000);
684 	msg[7] = cpu_to_le32(0x00000780);
685 	msg[8] = cpu_to_le32(0x00000440);
686 	msg[9] = cpu_to_le32(0x00000000);
687 	msg[10] = cpu_to_le32(0x01b37000);
688 	for (i = 11; i < 1024; ++i)
689 		msg[i] = cpu_to_le32(0x0);
690 
691 	radeon_bo_kunmap(bo);
692 	radeon_bo_unreserve(bo);
693 
694 	return radeon_uvd_send_msg(rdev, ring, bo, fence);
695 }
696 
697 int radeon_uvd_get_destroy_msg(struct radeon_device *rdev, int ring,
698 			       uint32_t handle, struct radeon_fence **fence)
699 {
700 	struct radeon_bo *bo;
701 	uint32_t *msg;
702 	int r, i;
703 
704 	r = radeon_bo_create(rdev, 1024, PAGE_SIZE, true,
705 			     RADEON_GEM_DOMAIN_VRAM, NULL, &bo);
706 	if (r)
707 		return r;
708 
709 	r = radeon_bo_reserve(bo, false);
710 	if (r) {
711 		radeon_bo_unref(&bo);
712 		return r;
713 	}
714 
715 	r = radeon_bo_kmap(bo, (void **)&msg);
716 	if (r) {
717 		radeon_bo_unreserve(bo);
718 		radeon_bo_unref(&bo);
719 		return r;
720 	}
721 
722 	/* stitch together an UVD destroy msg */
723 	msg[0] = cpu_to_le32(0x00000de4);
724 	msg[1] = cpu_to_le32(0x00000002);
725 	msg[2] = cpu_to_le32(handle);
726 	msg[3] = cpu_to_le32(0x00000000);
727 	for (i = 4; i < 1024; ++i)
728 		msg[i] = cpu_to_le32(0x0);
729 
730 	radeon_bo_kunmap(bo);
731 	radeon_bo_unreserve(bo);
732 
733 	return radeon_uvd_send_msg(rdev, ring, bo, fence);
734 }
735 
736 static void radeon_uvd_idle_work_handler(struct work_struct *work)
737 {
738 	struct radeon_device *rdev =
739 		container_of(work, struct radeon_device, uvd.idle_work.work);
740 
741 	if (radeon_fence_count_emitted(rdev, R600_RING_TYPE_UVD_INDEX) == 0) {
742 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
743 			mutex_lock(&rdev->pm.mutex);
744 			rdev->pm.dpm.uvd_active = false;
745 			mutex_unlock(&rdev->pm.mutex);
746 			radeon_pm_compute_clocks(rdev);
747 		} else {
748 			radeon_set_uvd_clocks(rdev, 0, 0);
749 		}
750 	} else {
751 		schedule_delayed_work(&rdev->uvd.idle_work,
752 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
753 	}
754 }
755 
756 void radeon_uvd_note_usage(struct radeon_device *rdev)
757 {
758 	bool set_clocks = !cancel_delayed_work_sync(&rdev->uvd.idle_work);
759 	set_clocks &= schedule_delayed_work(&rdev->uvd.idle_work,
760 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
761 	if (set_clocks) {
762 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
763 			/* XXX pick SD/HD/MVC */
764 			radeon_dpm_enable_power_state(rdev, POWER_STATE_TYPE_INTERNAL_UVD);
765 		} else {
766 			radeon_set_uvd_clocks(rdev, 53300, 40000);
767 		}
768 	}
769 }
770 
771 static unsigned radeon_uvd_calc_upll_post_div(unsigned vco_freq,
772 					      unsigned target_freq,
773 					      unsigned pd_min,
774 					      unsigned pd_even)
775 {
776 	unsigned post_div = vco_freq / target_freq;
777 
778 	/* adjust to post divider minimum value */
779 	if (post_div < pd_min)
780 		post_div = pd_min;
781 
782 	/* we alway need a frequency less than or equal the target */
783 	if ((vco_freq / post_div) > target_freq)
784 		post_div += 1;
785 
786 	/* post dividers above a certain value must be even */
787 	if (post_div > pd_even && post_div % 2)
788 		post_div += 1;
789 
790 	return post_div;
791 }
792 
793 /**
794  * radeon_uvd_calc_upll_dividers - calc UPLL clock dividers
795  *
796  * @rdev: radeon_device pointer
797  * @vclk: wanted VCLK
798  * @dclk: wanted DCLK
799  * @vco_min: minimum VCO frequency
800  * @vco_max: maximum VCO frequency
801  * @fb_factor: factor to multiply vco freq with
802  * @fb_mask: limit and bitmask for feedback divider
803  * @pd_min: post divider minimum
804  * @pd_max: post divider maximum
805  * @pd_even: post divider must be even above this value
806  * @optimal_fb_div: resulting feedback divider
807  * @optimal_vclk_div: resulting vclk post divider
808  * @optimal_dclk_div: resulting dclk post divider
809  *
810  * Calculate dividers for UVDs UPLL (R6xx-SI, except APUs).
811  * Returns zero on success -EINVAL on error.
812  */
813 int radeon_uvd_calc_upll_dividers(struct radeon_device *rdev,
814 				  unsigned vclk, unsigned dclk,
815 				  unsigned vco_min, unsigned vco_max,
816 				  unsigned fb_factor, unsigned fb_mask,
817 				  unsigned pd_min, unsigned pd_max,
818 				  unsigned pd_even,
819 				  unsigned *optimal_fb_div,
820 				  unsigned *optimal_vclk_div,
821 				  unsigned *optimal_dclk_div)
822 {
823 	unsigned vco_freq, ref_freq = rdev->clock.spll.reference_freq;
824 
825 	/* start off with something large */
826 	unsigned optimal_score = ~0;
827 
828 	/* loop through vco from low to high */
829 	vco_min = max(max(vco_min, vclk), dclk);
830 	for (vco_freq = vco_min; vco_freq <= vco_max; vco_freq += 100) {
831 
832 		uint64_t fb_div = (uint64_t)vco_freq * fb_factor;
833 		unsigned vclk_div, dclk_div, score;
834 
835 		do_div(fb_div, ref_freq);
836 
837 		/* fb div out of range ? */
838 		if (fb_div > fb_mask)
839 			break; /* it can oly get worse */
840 
841 		fb_div &= fb_mask;
842 
843 		/* calc vclk divider with current vco freq */
844 		vclk_div = radeon_uvd_calc_upll_post_div(vco_freq, vclk,
845 							 pd_min, pd_even);
846 		if (vclk_div > pd_max)
847 			break; /* vco is too big, it has to stop */
848 
849 		/* calc dclk divider with current vco freq */
850 		dclk_div = radeon_uvd_calc_upll_post_div(vco_freq, dclk,
851 							 pd_min, pd_even);
852 		if (vclk_div > pd_max)
853 			break; /* vco is too big, it has to stop */
854 
855 		/* calc score with current vco freq */
856 		score = vclk - (vco_freq / vclk_div) + dclk - (vco_freq / dclk_div);
857 
858 		/* determine if this vco setting is better than current optimal settings */
859 		if (score < optimal_score) {
860 			*optimal_fb_div = fb_div;
861 			*optimal_vclk_div = vclk_div;
862 			*optimal_dclk_div = dclk_div;
863 			optimal_score = score;
864 			if (optimal_score == 0)
865 				break; /* it can't get better than this */
866 		}
867 	}
868 
869 	/* did we found a valid setup ? */
870 	if (optimal_score == ~0)
871 		return -EINVAL;
872 
873 	return 0;
874 }
875 
876 int radeon_uvd_send_upll_ctlreq(struct radeon_device *rdev,
877 				unsigned cg_upll_func_cntl)
878 {
879 	unsigned i;
880 
881 	/* make sure UPLL_CTLREQ is deasserted */
882 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
883 
884 	mdelay(10);
885 
886 	/* assert UPLL_CTLREQ */
887 	WREG32_P(cg_upll_func_cntl, UPLL_CTLREQ_MASK, ~UPLL_CTLREQ_MASK);
888 
889 	/* wait for CTLACK and CTLACK2 to get asserted */
890 	for (i = 0; i < 100; ++i) {
891 		uint32_t mask = UPLL_CTLACK_MASK | UPLL_CTLACK2_MASK;
892 		if ((RREG32(cg_upll_func_cntl) & mask) == mask)
893 			break;
894 		mdelay(10);
895 	}
896 
897 	/* deassert UPLL_CTLREQ */
898 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
899 
900 	if (i == 100) {
901 		DRM_ERROR("Timeout setting UVD clocks!\n");
902 		return -ETIMEDOUT;
903 	}
904 
905 	return 0;
906 }
907