xref: /linux/drivers/gpu/drm/radeon/radeon_uvd.c (revision c4ee0af3fa0dc65f690fc908f02b8355f9576ea0)
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 	case CHIP_HAWAII:
101 		fw_name = FIRMWARE_BONAIRE;
102 		break;
103 
104 	default:
105 		return -EINVAL;
106 	}
107 
108 	r = request_firmware(&rdev->uvd_fw, fw_name, rdev->dev);
109 	if (r) {
110 		dev_err(rdev->dev, "radeon_uvd: Can't load firmware \"%s\"\n",
111 			fw_name);
112 		return r;
113 	}
114 
115 	bo_size = RADEON_GPU_PAGE_ALIGN(rdev->uvd_fw->size + 8) +
116 		  RADEON_UVD_STACK_SIZE + RADEON_UVD_HEAP_SIZE;
117 	r = radeon_bo_create(rdev, bo_size, PAGE_SIZE, true,
118 			     RADEON_GEM_DOMAIN_VRAM, NULL, &rdev->uvd.vcpu_bo);
119 	if (r) {
120 		dev_err(rdev->dev, "(%d) failed to allocate UVD bo\n", r);
121 		return r;
122 	}
123 
124 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
125 	if (r) {
126 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
127 		dev_err(rdev->dev, "(%d) failed to reserve UVD bo\n", r);
128 		return r;
129 	}
130 
131 	r = radeon_bo_pin(rdev->uvd.vcpu_bo, RADEON_GEM_DOMAIN_VRAM,
132 			  &rdev->uvd.gpu_addr);
133 	if (r) {
134 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
135 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
136 		dev_err(rdev->dev, "(%d) UVD bo pin failed\n", r);
137 		return r;
138 	}
139 
140 	r = radeon_bo_kmap(rdev->uvd.vcpu_bo, &rdev->uvd.cpu_addr);
141 	if (r) {
142 		dev_err(rdev->dev, "(%d) UVD map failed\n", r);
143 		return r;
144 	}
145 
146 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
147 
148 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
149 		atomic_set(&rdev->uvd.handles[i], 0);
150 		rdev->uvd.filp[i] = NULL;
151 		rdev->uvd.img_size[i] = 0;
152 	}
153 
154 	return 0;
155 }
156 
157 void radeon_uvd_fini(struct radeon_device *rdev)
158 {
159 	int r;
160 
161 	if (rdev->uvd.vcpu_bo == NULL)
162 		return;
163 
164 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
165 	if (!r) {
166 		radeon_bo_kunmap(rdev->uvd.vcpu_bo);
167 		radeon_bo_unpin(rdev->uvd.vcpu_bo);
168 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
169 	}
170 
171 	radeon_bo_unref(&rdev->uvd.vcpu_bo);
172 
173 	release_firmware(rdev->uvd_fw);
174 }
175 
176 int radeon_uvd_suspend(struct radeon_device *rdev)
177 {
178 	unsigned size;
179 	void *ptr;
180 	int i;
181 
182 	if (rdev->uvd.vcpu_bo == NULL)
183 		return 0;
184 
185 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i)
186 		if (atomic_read(&rdev->uvd.handles[i]))
187 			break;
188 
189 	if (i == RADEON_MAX_UVD_HANDLES)
190 		return 0;
191 
192 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
193 	size -= rdev->uvd_fw->size;
194 
195 	ptr = rdev->uvd.cpu_addr;
196 	ptr += rdev->uvd_fw->size;
197 
198 	rdev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
199 	memcpy(rdev->uvd.saved_bo, ptr, size);
200 
201 	return 0;
202 }
203 
204 int radeon_uvd_resume(struct radeon_device *rdev)
205 {
206 	unsigned size;
207 	void *ptr;
208 
209 	if (rdev->uvd.vcpu_bo == NULL)
210 		return -EINVAL;
211 
212 	memcpy(rdev->uvd.cpu_addr, rdev->uvd_fw->data, rdev->uvd_fw->size);
213 
214 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
215 	size -= rdev->uvd_fw->size;
216 
217 	ptr = rdev->uvd.cpu_addr;
218 	ptr += rdev->uvd_fw->size;
219 
220 	if (rdev->uvd.saved_bo != NULL) {
221 		memcpy(ptr, rdev->uvd.saved_bo, size);
222 		kfree(rdev->uvd.saved_bo);
223 		rdev->uvd.saved_bo = NULL;
224 	} else
225 		memset(ptr, 0, size);
226 
227 	return 0;
228 }
229 
230 void radeon_uvd_force_into_uvd_segment(struct radeon_bo *rbo)
231 {
232 	rbo->placement.fpfn = 0 >> PAGE_SHIFT;
233 	rbo->placement.lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
234 }
235 
236 void radeon_uvd_free_handles(struct radeon_device *rdev, struct drm_file *filp)
237 {
238 	int i, r;
239 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
240 		uint32_t handle = atomic_read(&rdev->uvd.handles[i]);
241 		if (handle != 0 && rdev->uvd.filp[i] == filp) {
242 			struct radeon_fence *fence;
243 
244 			radeon_uvd_note_usage(rdev);
245 
246 			r = radeon_uvd_get_destroy_msg(rdev,
247 				R600_RING_TYPE_UVD_INDEX, handle, &fence);
248 			if (r) {
249 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
250 				continue;
251 			}
252 
253 			radeon_fence_wait(fence, false);
254 			radeon_fence_unref(&fence);
255 
256 			rdev->uvd.filp[i] = NULL;
257 			atomic_set(&rdev->uvd.handles[i], 0);
258 		}
259 	}
260 }
261 
262 static int radeon_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
263 {
264 	unsigned stream_type = msg[4];
265 	unsigned width = msg[6];
266 	unsigned height = msg[7];
267 	unsigned dpb_size = msg[9];
268 	unsigned pitch = msg[28];
269 
270 	unsigned width_in_mb = width / 16;
271 	unsigned height_in_mb = ALIGN(height / 16, 2);
272 
273 	unsigned image_size, tmp, min_dpb_size;
274 
275 	image_size = width * height;
276 	image_size += image_size / 2;
277 	image_size = ALIGN(image_size, 1024);
278 
279 	switch (stream_type) {
280 	case 0: /* H264 */
281 
282 		/* reference picture buffer */
283 		min_dpb_size = image_size * 17;
284 
285 		/* macroblock context buffer */
286 		min_dpb_size += width_in_mb * height_in_mb * 17 * 192;
287 
288 		/* IT surface buffer */
289 		min_dpb_size += width_in_mb * height_in_mb * 32;
290 		break;
291 
292 	case 1: /* VC1 */
293 
294 		/* reference picture buffer */
295 		min_dpb_size = image_size * 3;
296 
297 		/* CONTEXT_BUFFER */
298 		min_dpb_size += width_in_mb * height_in_mb * 128;
299 
300 		/* IT surface buffer */
301 		min_dpb_size += width_in_mb * 64;
302 
303 		/* DB surface buffer */
304 		min_dpb_size += width_in_mb * 128;
305 
306 		/* BP */
307 		tmp = max(width_in_mb, height_in_mb);
308 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
309 		break;
310 
311 	case 3: /* MPEG2 */
312 
313 		/* reference picture buffer */
314 		min_dpb_size = image_size * 3;
315 		break;
316 
317 	case 4: /* MPEG4 */
318 
319 		/* reference picture buffer */
320 		min_dpb_size = image_size * 3;
321 
322 		/* CM */
323 		min_dpb_size += width_in_mb * height_in_mb * 64;
324 
325 		/* IT surface buffer */
326 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
327 		break;
328 
329 	default:
330 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
331 		return -EINVAL;
332 	}
333 
334 	if (width > pitch) {
335 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
336 		return -EINVAL;
337 	}
338 
339 	if (dpb_size < min_dpb_size) {
340 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
341 			  dpb_size, min_dpb_size);
342 		return -EINVAL;
343 	}
344 
345 	buf_sizes[0x1] = dpb_size;
346 	buf_sizes[0x2] = image_size;
347 	return 0;
348 }
349 
350 static int radeon_uvd_cs_msg(struct radeon_cs_parser *p, struct radeon_bo *bo,
351 			     unsigned offset, unsigned buf_sizes[])
352 {
353 	int32_t *msg, msg_type, handle;
354 	unsigned img_size = 0;
355 	void *ptr;
356 
357 	int i, r;
358 
359 	if (offset & 0x3F) {
360 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
361 		return -EINVAL;
362 	}
363 
364 	if (bo->tbo.sync_obj) {
365 		r = radeon_fence_wait(bo->tbo.sync_obj, false);
366 		if (r) {
367 			DRM_ERROR("Failed waiting for UVD message (%d)!\n", r);
368 			return r;
369 		}
370 	}
371 
372 	r = radeon_bo_kmap(bo, &ptr);
373 	if (r) {
374 		DRM_ERROR("Failed mapping the UVD message (%d)!\n", r);
375 		return r;
376 	}
377 
378 	msg = ptr + offset;
379 
380 	msg_type = msg[1];
381 	handle = msg[2];
382 
383 	if (handle == 0) {
384 		DRM_ERROR("Invalid UVD handle!\n");
385 		return -EINVAL;
386 	}
387 
388 	if (msg_type == 1) {
389 		/* it's a decode msg, calc buffer sizes */
390 		r = radeon_uvd_cs_msg_decode(msg, buf_sizes);
391 		/* calc image size (width * height) */
392 		img_size = msg[6] * msg[7];
393 		radeon_bo_kunmap(bo);
394 		if (r)
395 			return r;
396 
397 	} else if (msg_type == 2) {
398 		/* it's a destroy msg, free the handle */
399 		for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i)
400 			atomic_cmpxchg(&p->rdev->uvd.handles[i], handle, 0);
401 		radeon_bo_kunmap(bo);
402 		return 0;
403 	} else {
404 		/* it's a create msg, calc image size (width * height) */
405 		img_size = msg[7] * msg[8];
406 		radeon_bo_kunmap(bo);
407 
408 		if (msg_type != 0) {
409 			DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
410 			return -EINVAL;
411 		}
412 
413 		/* it's a create msg, no special handling needed */
414 	}
415 
416 	/* create or decode, validate the handle */
417 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
418 		if (atomic_read(&p->rdev->uvd.handles[i]) == handle)
419 			return 0;
420 	}
421 
422 	/* handle not found try to alloc a new one */
423 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
424 		if (!atomic_cmpxchg(&p->rdev->uvd.handles[i], 0, handle)) {
425 			p->rdev->uvd.filp[i] = p->filp;
426 			p->rdev->uvd.img_size[i] = img_size;
427 			return 0;
428 		}
429 	}
430 
431 	DRM_ERROR("No more free UVD handles!\n");
432 	return -EINVAL;
433 }
434 
435 static int radeon_uvd_cs_reloc(struct radeon_cs_parser *p,
436 			       int data0, int data1,
437 			       unsigned buf_sizes[], bool *has_msg_cmd)
438 {
439 	struct radeon_cs_chunk *relocs_chunk;
440 	struct radeon_cs_reloc *reloc;
441 	unsigned idx, cmd, offset;
442 	uint64_t start, end;
443 	int r;
444 
445 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
446 	offset = radeon_get_ib_value(p, data0);
447 	idx = radeon_get_ib_value(p, data1);
448 	if (idx >= relocs_chunk->length_dw) {
449 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
450 			  idx, relocs_chunk->length_dw);
451 		return -EINVAL;
452 	}
453 
454 	reloc = p->relocs_ptr[(idx / 4)];
455 	start = reloc->lobj.gpu_offset;
456 	end = start + radeon_bo_size(reloc->robj);
457 	start += offset;
458 
459 	p->ib.ptr[data0] = start & 0xFFFFFFFF;
460 	p->ib.ptr[data1] = start >> 32;
461 
462 	cmd = radeon_get_ib_value(p, p->idx) >> 1;
463 
464 	if (cmd < 0x4) {
465 		if ((end - start) < buf_sizes[cmd]) {
466 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
467 				  (unsigned)(end - start), buf_sizes[cmd]);
468 			return -EINVAL;
469 		}
470 
471 	} else if (cmd != 0x100) {
472 		DRM_ERROR("invalid UVD command %X!\n", cmd);
473 		return -EINVAL;
474 	}
475 
476 	if ((start >> 28) != (end >> 28)) {
477 		DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
478 			  start, end);
479 		return -EINVAL;
480 	}
481 
482 	/* TODO: is this still necessary on NI+ ? */
483 	if ((cmd == 0 || cmd == 0x3) &&
484 	    (start >> 28) != (p->rdev->uvd.gpu_addr >> 28)) {
485 		DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
486 			  start, end);
487 		return -EINVAL;
488 	}
489 
490 	if (cmd == 0) {
491 		if (*has_msg_cmd) {
492 			DRM_ERROR("More than one message in a UVD-IB!\n");
493 			return -EINVAL;
494 		}
495 		*has_msg_cmd = true;
496 		r = radeon_uvd_cs_msg(p, reloc->robj, offset, buf_sizes);
497 		if (r)
498 			return r;
499 	} else if (!*has_msg_cmd) {
500 		DRM_ERROR("Message needed before other commands are send!\n");
501 		return -EINVAL;
502 	}
503 
504 	return 0;
505 }
506 
507 static int radeon_uvd_cs_reg(struct radeon_cs_parser *p,
508 			     struct radeon_cs_packet *pkt,
509 			     int *data0, int *data1,
510 			     unsigned buf_sizes[],
511 			     bool *has_msg_cmd)
512 {
513 	int i, r;
514 
515 	p->idx++;
516 	for (i = 0; i <= pkt->count; ++i) {
517 		switch (pkt->reg + i*4) {
518 		case UVD_GPCOM_VCPU_DATA0:
519 			*data0 = p->idx;
520 			break;
521 		case UVD_GPCOM_VCPU_DATA1:
522 			*data1 = p->idx;
523 			break;
524 		case UVD_GPCOM_VCPU_CMD:
525 			r = radeon_uvd_cs_reloc(p, *data0, *data1,
526 						buf_sizes, has_msg_cmd);
527 			if (r)
528 				return r;
529 			break;
530 		case UVD_ENGINE_CNTL:
531 			break;
532 		default:
533 			DRM_ERROR("Invalid reg 0x%X!\n",
534 				  pkt->reg + i*4);
535 			return -EINVAL;
536 		}
537 		p->idx++;
538 	}
539 	return 0;
540 }
541 
542 int radeon_uvd_cs_parse(struct radeon_cs_parser *p)
543 {
544 	struct radeon_cs_packet pkt;
545 	int r, data0 = 0, data1 = 0;
546 
547 	/* does the IB has a msg command */
548 	bool has_msg_cmd = false;
549 
550 	/* minimum buffer sizes */
551 	unsigned buf_sizes[] = {
552 		[0x00000000]	=	2048,
553 		[0x00000001]	=	32 * 1024 * 1024,
554 		[0x00000002]	=	2048 * 1152 * 3,
555 		[0x00000003]	=	2048,
556 	};
557 
558 	if (p->chunks[p->chunk_ib_idx].length_dw % 16) {
559 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
560 			  p->chunks[p->chunk_ib_idx].length_dw);
561 		return -EINVAL;
562 	}
563 
564 	if (p->chunk_relocs_idx == -1) {
565 		DRM_ERROR("No relocation chunk !\n");
566 		return -EINVAL;
567 	}
568 
569 
570 	do {
571 		r = radeon_cs_packet_parse(p, &pkt, p->idx);
572 		if (r)
573 			return r;
574 		switch (pkt.type) {
575 		case RADEON_PACKET_TYPE0:
576 			r = radeon_uvd_cs_reg(p, &pkt, &data0, &data1,
577 					      buf_sizes, &has_msg_cmd);
578 			if (r)
579 				return r;
580 			break;
581 		case RADEON_PACKET_TYPE2:
582 			p->idx += pkt.count + 2;
583 			break;
584 		default:
585 			DRM_ERROR("Unknown packet type %d !\n", pkt.type);
586 			return -EINVAL;
587 		}
588 	} while (p->idx < p->chunks[p->chunk_ib_idx].length_dw);
589 
590 	if (!has_msg_cmd) {
591 		DRM_ERROR("UVD-IBs need a msg command!\n");
592 		return -EINVAL;
593 	}
594 
595 	return 0;
596 }
597 
598 static int radeon_uvd_send_msg(struct radeon_device *rdev,
599 			       int ring, struct radeon_bo *bo,
600 			       struct radeon_fence **fence)
601 {
602 	struct ttm_validate_buffer tv;
603 	struct ww_acquire_ctx ticket;
604 	struct list_head head;
605 	struct radeon_ib ib;
606 	uint64_t addr;
607 	int i, r;
608 
609 	memset(&tv, 0, sizeof(tv));
610 	tv.bo = &bo->tbo;
611 
612 	INIT_LIST_HEAD(&head);
613 	list_add(&tv.head, &head);
614 
615 	r = ttm_eu_reserve_buffers(&ticket, &head);
616 	if (r)
617 		return r;
618 
619 	radeon_ttm_placement_from_domain(bo, RADEON_GEM_DOMAIN_VRAM);
620 	radeon_uvd_force_into_uvd_segment(bo);
621 
622 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
623 	if (r)
624 		goto err;
625 
626 	r = radeon_ib_get(rdev, ring, &ib, NULL, 64);
627 	if (r)
628 		goto err;
629 
630 	addr = radeon_bo_gpu_offset(bo);
631 	ib.ptr[0] = PACKET0(UVD_GPCOM_VCPU_DATA0, 0);
632 	ib.ptr[1] = addr;
633 	ib.ptr[2] = PACKET0(UVD_GPCOM_VCPU_DATA1, 0);
634 	ib.ptr[3] = addr >> 32;
635 	ib.ptr[4] = PACKET0(UVD_GPCOM_VCPU_CMD, 0);
636 	ib.ptr[5] = 0;
637 	for (i = 6; i < 16; ++i)
638 		ib.ptr[i] = PACKET2(0);
639 	ib.length_dw = 16;
640 
641 	r = radeon_ib_schedule(rdev, &ib, NULL);
642 	if (r)
643 		goto err;
644 	ttm_eu_fence_buffer_objects(&ticket, &head, ib.fence);
645 
646 	if (fence)
647 		*fence = radeon_fence_ref(ib.fence);
648 
649 	radeon_ib_free(rdev, &ib);
650 	radeon_bo_unref(&bo);
651 	return 0;
652 
653 err:
654 	ttm_eu_backoff_reservation(&ticket, &head);
655 	return r;
656 }
657 
658 /* multiple fence commands without any stream commands in between can
659    crash the vcpu so just try to emmit a dummy create/destroy msg to
660    avoid this */
661 int radeon_uvd_get_create_msg(struct radeon_device *rdev, int ring,
662 			      uint32_t handle, struct radeon_fence **fence)
663 {
664 	struct radeon_bo *bo;
665 	uint32_t *msg;
666 	int r, i;
667 
668 	r = radeon_bo_create(rdev, 1024, PAGE_SIZE, true,
669 			     RADEON_GEM_DOMAIN_VRAM, NULL, &bo);
670 	if (r)
671 		return r;
672 
673 	r = radeon_bo_reserve(bo, false);
674 	if (r) {
675 		radeon_bo_unref(&bo);
676 		return r;
677 	}
678 
679 	r = radeon_bo_kmap(bo, (void **)&msg);
680 	if (r) {
681 		radeon_bo_unreserve(bo);
682 		radeon_bo_unref(&bo);
683 		return r;
684 	}
685 
686 	/* stitch together an UVD create msg */
687 	msg[0] = cpu_to_le32(0x00000de4);
688 	msg[1] = cpu_to_le32(0x00000000);
689 	msg[2] = cpu_to_le32(handle);
690 	msg[3] = cpu_to_le32(0x00000000);
691 	msg[4] = cpu_to_le32(0x00000000);
692 	msg[5] = cpu_to_le32(0x00000000);
693 	msg[6] = cpu_to_le32(0x00000000);
694 	msg[7] = cpu_to_le32(0x00000780);
695 	msg[8] = cpu_to_le32(0x00000440);
696 	msg[9] = cpu_to_le32(0x00000000);
697 	msg[10] = cpu_to_le32(0x01b37000);
698 	for (i = 11; i < 1024; ++i)
699 		msg[i] = cpu_to_le32(0x0);
700 
701 	radeon_bo_kunmap(bo);
702 	radeon_bo_unreserve(bo);
703 
704 	return radeon_uvd_send_msg(rdev, ring, bo, fence);
705 }
706 
707 int radeon_uvd_get_destroy_msg(struct radeon_device *rdev, int ring,
708 			       uint32_t handle, struct radeon_fence **fence)
709 {
710 	struct radeon_bo *bo;
711 	uint32_t *msg;
712 	int r, i;
713 
714 	r = radeon_bo_create(rdev, 1024, PAGE_SIZE, true,
715 			     RADEON_GEM_DOMAIN_VRAM, NULL, &bo);
716 	if (r)
717 		return r;
718 
719 	r = radeon_bo_reserve(bo, false);
720 	if (r) {
721 		radeon_bo_unref(&bo);
722 		return r;
723 	}
724 
725 	r = radeon_bo_kmap(bo, (void **)&msg);
726 	if (r) {
727 		radeon_bo_unreserve(bo);
728 		radeon_bo_unref(&bo);
729 		return r;
730 	}
731 
732 	/* stitch together an UVD destroy msg */
733 	msg[0] = cpu_to_le32(0x00000de4);
734 	msg[1] = cpu_to_le32(0x00000002);
735 	msg[2] = cpu_to_le32(handle);
736 	msg[3] = cpu_to_le32(0x00000000);
737 	for (i = 4; i < 1024; ++i)
738 		msg[i] = cpu_to_le32(0x0);
739 
740 	radeon_bo_kunmap(bo);
741 	radeon_bo_unreserve(bo);
742 
743 	return radeon_uvd_send_msg(rdev, ring, bo, fence);
744 }
745 
746 /**
747  * radeon_uvd_count_handles - count number of open streams
748  *
749  * @rdev: radeon_device pointer
750  * @sd: number of SD streams
751  * @hd: number of HD streams
752  *
753  * Count the number of open SD/HD streams as a hint for power mangement
754  */
755 static void radeon_uvd_count_handles(struct radeon_device *rdev,
756 				     unsigned *sd, unsigned *hd)
757 {
758 	unsigned i;
759 
760 	*sd = 0;
761 	*hd = 0;
762 
763 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
764 		if (!atomic_read(&rdev->uvd.handles[i]))
765 			continue;
766 
767 		if (rdev->uvd.img_size[i] >= 720*576)
768 			++(*hd);
769 		else
770 			++(*sd);
771 	}
772 }
773 
774 static void radeon_uvd_idle_work_handler(struct work_struct *work)
775 {
776 	struct radeon_device *rdev =
777 		container_of(work, struct radeon_device, uvd.idle_work.work);
778 
779 	if (radeon_fence_count_emitted(rdev, R600_RING_TYPE_UVD_INDEX) == 0) {
780 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
781 			radeon_dpm_enable_uvd(rdev, false);
782 		} else {
783 			radeon_set_uvd_clocks(rdev, 0, 0);
784 		}
785 	} else {
786 		schedule_delayed_work(&rdev->uvd.idle_work,
787 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
788 	}
789 }
790 
791 void radeon_uvd_note_usage(struct radeon_device *rdev)
792 {
793 	bool streams_changed = false;
794 	bool set_clocks = !cancel_delayed_work_sync(&rdev->uvd.idle_work);
795 	set_clocks &= schedule_delayed_work(&rdev->uvd.idle_work,
796 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
797 
798 	if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
799 		unsigned hd = 0, sd = 0;
800 		radeon_uvd_count_handles(rdev, &sd, &hd);
801 		if ((rdev->pm.dpm.sd != sd) ||
802 		    (rdev->pm.dpm.hd != hd)) {
803 			rdev->pm.dpm.sd = sd;
804 			rdev->pm.dpm.hd = hd;
805 			/* disable this for now */
806 			/*streams_changed = true;*/
807 		}
808 	}
809 
810 	if (set_clocks || streams_changed) {
811 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
812 			radeon_dpm_enable_uvd(rdev, true);
813 		} else {
814 			radeon_set_uvd_clocks(rdev, 53300, 40000);
815 		}
816 	}
817 }
818 
819 static unsigned radeon_uvd_calc_upll_post_div(unsigned vco_freq,
820 					      unsigned target_freq,
821 					      unsigned pd_min,
822 					      unsigned pd_even)
823 {
824 	unsigned post_div = vco_freq / target_freq;
825 
826 	/* adjust to post divider minimum value */
827 	if (post_div < pd_min)
828 		post_div = pd_min;
829 
830 	/* we alway need a frequency less than or equal the target */
831 	if ((vco_freq / post_div) > target_freq)
832 		post_div += 1;
833 
834 	/* post dividers above a certain value must be even */
835 	if (post_div > pd_even && post_div % 2)
836 		post_div += 1;
837 
838 	return post_div;
839 }
840 
841 /**
842  * radeon_uvd_calc_upll_dividers - calc UPLL clock dividers
843  *
844  * @rdev: radeon_device pointer
845  * @vclk: wanted VCLK
846  * @dclk: wanted DCLK
847  * @vco_min: minimum VCO frequency
848  * @vco_max: maximum VCO frequency
849  * @fb_factor: factor to multiply vco freq with
850  * @fb_mask: limit and bitmask for feedback divider
851  * @pd_min: post divider minimum
852  * @pd_max: post divider maximum
853  * @pd_even: post divider must be even above this value
854  * @optimal_fb_div: resulting feedback divider
855  * @optimal_vclk_div: resulting vclk post divider
856  * @optimal_dclk_div: resulting dclk post divider
857  *
858  * Calculate dividers for UVDs UPLL (R6xx-SI, except APUs).
859  * Returns zero on success -EINVAL on error.
860  */
861 int radeon_uvd_calc_upll_dividers(struct radeon_device *rdev,
862 				  unsigned vclk, unsigned dclk,
863 				  unsigned vco_min, unsigned vco_max,
864 				  unsigned fb_factor, unsigned fb_mask,
865 				  unsigned pd_min, unsigned pd_max,
866 				  unsigned pd_even,
867 				  unsigned *optimal_fb_div,
868 				  unsigned *optimal_vclk_div,
869 				  unsigned *optimal_dclk_div)
870 {
871 	unsigned vco_freq, ref_freq = rdev->clock.spll.reference_freq;
872 
873 	/* start off with something large */
874 	unsigned optimal_score = ~0;
875 
876 	/* loop through vco from low to high */
877 	vco_min = max(max(vco_min, vclk), dclk);
878 	for (vco_freq = vco_min; vco_freq <= vco_max; vco_freq += 100) {
879 
880 		uint64_t fb_div = (uint64_t)vco_freq * fb_factor;
881 		unsigned vclk_div, dclk_div, score;
882 
883 		do_div(fb_div, ref_freq);
884 
885 		/* fb div out of range ? */
886 		if (fb_div > fb_mask)
887 			break; /* it can oly get worse */
888 
889 		fb_div &= fb_mask;
890 
891 		/* calc vclk divider with current vco freq */
892 		vclk_div = radeon_uvd_calc_upll_post_div(vco_freq, vclk,
893 							 pd_min, pd_even);
894 		if (vclk_div > pd_max)
895 			break; /* vco is too big, it has to stop */
896 
897 		/* calc dclk divider with current vco freq */
898 		dclk_div = radeon_uvd_calc_upll_post_div(vco_freq, dclk,
899 							 pd_min, pd_even);
900 		if (vclk_div > pd_max)
901 			break; /* vco is too big, it has to stop */
902 
903 		/* calc score with current vco freq */
904 		score = vclk - (vco_freq / vclk_div) + dclk - (vco_freq / dclk_div);
905 
906 		/* determine if this vco setting is better than current optimal settings */
907 		if (score < optimal_score) {
908 			*optimal_fb_div = fb_div;
909 			*optimal_vclk_div = vclk_div;
910 			*optimal_dclk_div = dclk_div;
911 			optimal_score = score;
912 			if (optimal_score == 0)
913 				break; /* it can't get better than this */
914 		}
915 	}
916 
917 	/* did we found a valid setup ? */
918 	if (optimal_score == ~0)
919 		return -EINVAL;
920 
921 	return 0;
922 }
923 
924 int radeon_uvd_send_upll_ctlreq(struct radeon_device *rdev,
925 				unsigned cg_upll_func_cntl)
926 {
927 	unsigned i;
928 
929 	/* make sure UPLL_CTLREQ is deasserted */
930 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
931 
932 	mdelay(10);
933 
934 	/* assert UPLL_CTLREQ */
935 	WREG32_P(cg_upll_func_cntl, UPLL_CTLREQ_MASK, ~UPLL_CTLREQ_MASK);
936 
937 	/* wait for CTLACK and CTLACK2 to get asserted */
938 	for (i = 0; i < 100; ++i) {
939 		uint32_t mask = UPLL_CTLACK_MASK | UPLL_CTLACK2_MASK;
940 		if ((RREG32(cg_upll_func_cntl) & mask) == mask)
941 			break;
942 		mdelay(10);
943 	}
944 
945 	/* deassert UPLL_CTLREQ */
946 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
947 
948 	if (i == 100) {
949 		DRM_ERROR("Timeout setting UVD clocks!\n");
950 		return -ETIMEDOUT;
951 	}
952 
953 	return 0;
954 }
955