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 	case CHIP_OLAND:
95 		fw_name = FIRMWARE_TAHITI;
96 		break;
97 
98 	case CHIP_BONAIRE:
99 	case CHIP_KABINI:
100 	case CHIP_KAVERI:
101 	case CHIP_HAWAII:
102 		fw_name = FIRMWARE_BONAIRE;
103 		break;
104 
105 	default:
106 		return -EINVAL;
107 	}
108 
109 	r = request_firmware(&rdev->uvd_fw, fw_name, rdev->dev);
110 	if (r) {
111 		dev_err(rdev->dev, "radeon_uvd: Can't load firmware \"%s\"\n",
112 			fw_name);
113 		return r;
114 	}
115 
116 	bo_size = RADEON_GPU_PAGE_ALIGN(rdev->uvd_fw->size + 8) +
117 		  RADEON_UVD_STACK_SIZE + RADEON_UVD_HEAP_SIZE;
118 	r = radeon_bo_create(rdev, bo_size, PAGE_SIZE, true,
119 			     RADEON_GEM_DOMAIN_VRAM, NULL, &rdev->uvd.vcpu_bo);
120 	if (r) {
121 		dev_err(rdev->dev, "(%d) failed to allocate UVD bo\n", r);
122 		return r;
123 	}
124 
125 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
126 	if (r) {
127 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
128 		dev_err(rdev->dev, "(%d) failed to reserve UVD bo\n", r);
129 		return r;
130 	}
131 
132 	r = radeon_bo_pin(rdev->uvd.vcpu_bo, RADEON_GEM_DOMAIN_VRAM,
133 			  &rdev->uvd.gpu_addr);
134 	if (r) {
135 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
136 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
137 		dev_err(rdev->dev, "(%d) UVD bo pin failed\n", r);
138 		return r;
139 	}
140 
141 	r = radeon_bo_kmap(rdev->uvd.vcpu_bo, &rdev->uvd.cpu_addr);
142 	if (r) {
143 		dev_err(rdev->dev, "(%d) UVD map failed\n", r);
144 		return r;
145 	}
146 
147 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
148 
149 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
150 		atomic_set(&rdev->uvd.handles[i], 0);
151 		rdev->uvd.filp[i] = NULL;
152 		rdev->uvd.img_size[i] = 0;
153 	}
154 
155 	return 0;
156 }
157 
158 void radeon_uvd_fini(struct radeon_device *rdev)
159 {
160 	int r;
161 
162 	if (rdev->uvd.vcpu_bo == NULL)
163 		return;
164 
165 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
166 	if (!r) {
167 		radeon_bo_kunmap(rdev->uvd.vcpu_bo);
168 		radeon_bo_unpin(rdev->uvd.vcpu_bo);
169 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
170 	}
171 
172 	radeon_bo_unref(&rdev->uvd.vcpu_bo);
173 
174 	radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_UVD_INDEX]);
175 
176 	release_firmware(rdev->uvd_fw);
177 }
178 
179 int radeon_uvd_suspend(struct radeon_device *rdev)
180 {
181 	unsigned size;
182 	void *ptr;
183 	int i;
184 
185 	if (rdev->uvd.vcpu_bo == NULL)
186 		return 0;
187 
188 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i)
189 		if (atomic_read(&rdev->uvd.handles[i]))
190 			break;
191 
192 	if (i == RADEON_MAX_UVD_HANDLES)
193 		return 0;
194 
195 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
196 	size -= rdev->uvd_fw->size;
197 
198 	ptr = rdev->uvd.cpu_addr;
199 	ptr += rdev->uvd_fw->size;
200 
201 	rdev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
202 	memcpy(rdev->uvd.saved_bo, ptr, size);
203 
204 	return 0;
205 }
206 
207 int radeon_uvd_resume(struct radeon_device *rdev)
208 {
209 	unsigned size;
210 	void *ptr;
211 
212 	if (rdev->uvd.vcpu_bo == NULL)
213 		return -EINVAL;
214 
215 	memcpy(rdev->uvd.cpu_addr, rdev->uvd_fw->data, rdev->uvd_fw->size);
216 
217 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
218 	size -= rdev->uvd_fw->size;
219 
220 	ptr = rdev->uvd.cpu_addr;
221 	ptr += rdev->uvd_fw->size;
222 
223 	if (rdev->uvd.saved_bo != NULL) {
224 		memcpy(ptr, rdev->uvd.saved_bo, size);
225 		kfree(rdev->uvd.saved_bo);
226 		rdev->uvd.saved_bo = NULL;
227 	} else
228 		memset(ptr, 0, size);
229 
230 	return 0;
231 }
232 
233 void radeon_uvd_force_into_uvd_segment(struct radeon_bo *rbo)
234 {
235 	rbo->placement.fpfn = 0 >> PAGE_SHIFT;
236 	rbo->placement.lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
237 }
238 
239 void radeon_uvd_free_handles(struct radeon_device *rdev, struct drm_file *filp)
240 {
241 	int i, r;
242 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
243 		uint32_t handle = atomic_read(&rdev->uvd.handles[i]);
244 		if (handle != 0 && rdev->uvd.filp[i] == filp) {
245 			struct radeon_fence *fence;
246 
247 			radeon_uvd_note_usage(rdev);
248 
249 			r = radeon_uvd_get_destroy_msg(rdev,
250 				R600_RING_TYPE_UVD_INDEX, handle, &fence);
251 			if (r) {
252 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
253 				continue;
254 			}
255 
256 			radeon_fence_wait(fence, false);
257 			radeon_fence_unref(&fence);
258 
259 			rdev->uvd.filp[i] = NULL;
260 			atomic_set(&rdev->uvd.handles[i], 0);
261 		}
262 	}
263 }
264 
265 static int radeon_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
266 {
267 	unsigned stream_type = msg[4];
268 	unsigned width = msg[6];
269 	unsigned height = msg[7];
270 	unsigned dpb_size = msg[9];
271 	unsigned pitch = msg[28];
272 
273 	unsigned width_in_mb = width / 16;
274 	unsigned height_in_mb = ALIGN(height / 16, 2);
275 
276 	unsigned image_size, tmp, min_dpb_size;
277 
278 	image_size = width * height;
279 	image_size += image_size / 2;
280 	image_size = ALIGN(image_size, 1024);
281 
282 	switch (stream_type) {
283 	case 0: /* H264 */
284 
285 		/* reference picture buffer */
286 		min_dpb_size = image_size * 17;
287 
288 		/* macroblock context buffer */
289 		min_dpb_size += width_in_mb * height_in_mb * 17 * 192;
290 
291 		/* IT surface buffer */
292 		min_dpb_size += width_in_mb * height_in_mb * 32;
293 		break;
294 
295 	case 1: /* VC1 */
296 
297 		/* reference picture buffer */
298 		min_dpb_size = image_size * 3;
299 
300 		/* CONTEXT_BUFFER */
301 		min_dpb_size += width_in_mb * height_in_mb * 128;
302 
303 		/* IT surface buffer */
304 		min_dpb_size += width_in_mb * 64;
305 
306 		/* DB surface buffer */
307 		min_dpb_size += width_in_mb * 128;
308 
309 		/* BP */
310 		tmp = max(width_in_mb, height_in_mb);
311 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
312 		break;
313 
314 	case 3: /* MPEG2 */
315 
316 		/* reference picture buffer */
317 		min_dpb_size = image_size * 3;
318 		break;
319 
320 	case 4: /* MPEG4 */
321 
322 		/* reference picture buffer */
323 		min_dpb_size = image_size * 3;
324 
325 		/* CM */
326 		min_dpb_size += width_in_mb * height_in_mb * 64;
327 
328 		/* IT surface buffer */
329 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
330 		break;
331 
332 	default:
333 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
334 		return -EINVAL;
335 	}
336 
337 	if (width > pitch) {
338 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
339 		return -EINVAL;
340 	}
341 
342 	if (dpb_size < min_dpb_size) {
343 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
344 			  dpb_size, min_dpb_size);
345 		return -EINVAL;
346 	}
347 
348 	buf_sizes[0x1] = dpb_size;
349 	buf_sizes[0x2] = image_size;
350 	return 0;
351 }
352 
353 static int radeon_uvd_cs_msg(struct radeon_cs_parser *p, struct radeon_bo *bo,
354 			     unsigned offset, unsigned buf_sizes[])
355 {
356 	int32_t *msg, msg_type, handle;
357 	unsigned img_size = 0;
358 	void *ptr;
359 
360 	int i, r;
361 
362 	if (offset & 0x3F) {
363 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
364 		return -EINVAL;
365 	}
366 
367 	if (bo->tbo.sync_obj) {
368 		r = radeon_fence_wait(bo->tbo.sync_obj, false);
369 		if (r) {
370 			DRM_ERROR("Failed waiting for UVD message (%d)!\n", r);
371 			return r;
372 		}
373 	}
374 
375 	r = radeon_bo_kmap(bo, &ptr);
376 	if (r) {
377 		DRM_ERROR("Failed mapping the UVD message (%d)!\n", r);
378 		return r;
379 	}
380 
381 	msg = ptr + offset;
382 
383 	msg_type = msg[1];
384 	handle = msg[2];
385 
386 	if (handle == 0) {
387 		DRM_ERROR("Invalid UVD handle!\n");
388 		return -EINVAL;
389 	}
390 
391 	if (msg_type == 1) {
392 		/* it's a decode msg, calc buffer sizes */
393 		r = radeon_uvd_cs_msg_decode(msg, buf_sizes);
394 		/* calc image size (width * height) */
395 		img_size = msg[6] * msg[7];
396 		radeon_bo_kunmap(bo);
397 		if (r)
398 			return r;
399 
400 	} else if (msg_type == 2) {
401 		/* it's a destroy msg, free the handle */
402 		for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i)
403 			atomic_cmpxchg(&p->rdev->uvd.handles[i], handle, 0);
404 		radeon_bo_kunmap(bo);
405 		return 0;
406 	} else {
407 		/* it's a create msg, calc image size (width * height) */
408 		img_size = msg[7] * msg[8];
409 		radeon_bo_kunmap(bo);
410 
411 		if (msg_type != 0) {
412 			DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
413 			return -EINVAL;
414 		}
415 
416 		/* it's a create msg, no special handling needed */
417 	}
418 
419 	/* create or decode, validate the handle */
420 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
421 		if (atomic_read(&p->rdev->uvd.handles[i]) == handle)
422 			return 0;
423 	}
424 
425 	/* handle not found try to alloc a new one */
426 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
427 		if (!atomic_cmpxchg(&p->rdev->uvd.handles[i], 0, handle)) {
428 			p->rdev->uvd.filp[i] = p->filp;
429 			p->rdev->uvd.img_size[i] = img_size;
430 			return 0;
431 		}
432 	}
433 
434 	DRM_ERROR("No more free UVD handles!\n");
435 	return -EINVAL;
436 }
437 
438 static int radeon_uvd_cs_reloc(struct radeon_cs_parser *p,
439 			       int data0, int data1,
440 			       unsigned buf_sizes[], bool *has_msg_cmd)
441 {
442 	struct radeon_cs_chunk *relocs_chunk;
443 	struct radeon_cs_reloc *reloc;
444 	unsigned idx, cmd, offset;
445 	uint64_t start, end;
446 	int r;
447 
448 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
449 	offset = radeon_get_ib_value(p, data0);
450 	idx = radeon_get_ib_value(p, data1);
451 	if (idx >= relocs_chunk->length_dw) {
452 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
453 			  idx, relocs_chunk->length_dw);
454 		return -EINVAL;
455 	}
456 
457 	reloc = p->relocs_ptr[(idx / 4)];
458 	start = reloc->gpu_offset;
459 	end = start + radeon_bo_size(reloc->robj);
460 	start += offset;
461 
462 	p->ib.ptr[data0] = start & 0xFFFFFFFF;
463 	p->ib.ptr[data1] = start >> 32;
464 
465 	cmd = radeon_get_ib_value(p, p->idx) >> 1;
466 
467 	if (cmd < 0x4) {
468 		if ((end - start) < buf_sizes[cmd]) {
469 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
470 				  (unsigned)(end - start), buf_sizes[cmd]);
471 			return -EINVAL;
472 		}
473 
474 	} else if (cmd != 0x100) {
475 		DRM_ERROR("invalid UVD command %X!\n", cmd);
476 		return -EINVAL;
477 	}
478 
479 	if ((start >> 28) != ((end - 1) >> 28)) {
480 		DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
481 			  start, end);
482 		return -EINVAL;
483 	}
484 
485 	/* TODO: is this still necessary on NI+ ? */
486 	if ((cmd == 0 || cmd == 0x3) &&
487 	    (start >> 28) != (p->rdev->uvd.gpu_addr >> 28)) {
488 		DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
489 			  start, end);
490 		return -EINVAL;
491 	}
492 
493 	if (cmd == 0) {
494 		if (*has_msg_cmd) {
495 			DRM_ERROR("More than one message in a UVD-IB!\n");
496 			return -EINVAL;
497 		}
498 		*has_msg_cmd = true;
499 		r = radeon_uvd_cs_msg(p, reloc->robj, offset, buf_sizes);
500 		if (r)
501 			return r;
502 	} else if (!*has_msg_cmd) {
503 		DRM_ERROR("Message needed before other commands are send!\n");
504 		return -EINVAL;
505 	}
506 
507 	return 0;
508 }
509 
510 static int radeon_uvd_cs_reg(struct radeon_cs_parser *p,
511 			     struct radeon_cs_packet *pkt,
512 			     int *data0, int *data1,
513 			     unsigned buf_sizes[],
514 			     bool *has_msg_cmd)
515 {
516 	int i, r;
517 
518 	p->idx++;
519 	for (i = 0; i <= pkt->count; ++i) {
520 		switch (pkt->reg + i*4) {
521 		case UVD_GPCOM_VCPU_DATA0:
522 			*data0 = p->idx;
523 			break;
524 		case UVD_GPCOM_VCPU_DATA1:
525 			*data1 = p->idx;
526 			break;
527 		case UVD_GPCOM_VCPU_CMD:
528 			r = radeon_uvd_cs_reloc(p, *data0, *data1,
529 						buf_sizes, has_msg_cmd);
530 			if (r)
531 				return r;
532 			break;
533 		case UVD_ENGINE_CNTL:
534 			break;
535 		default:
536 			DRM_ERROR("Invalid reg 0x%X!\n",
537 				  pkt->reg + i*4);
538 			return -EINVAL;
539 		}
540 		p->idx++;
541 	}
542 	return 0;
543 }
544 
545 int radeon_uvd_cs_parse(struct radeon_cs_parser *p)
546 {
547 	struct radeon_cs_packet pkt;
548 	int r, data0 = 0, data1 = 0;
549 
550 	/* does the IB has a msg command */
551 	bool has_msg_cmd = false;
552 
553 	/* minimum buffer sizes */
554 	unsigned buf_sizes[] = {
555 		[0x00000000]	=	2048,
556 		[0x00000001]	=	32 * 1024 * 1024,
557 		[0x00000002]	=	2048 * 1152 * 3,
558 		[0x00000003]	=	2048,
559 	};
560 
561 	if (p->chunks[p->chunk_ib_idx].length_dw % 16) {
562 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
563 			  p->chunks[p->chunk_ib_idx].length_dw);
564 		return -EINVAL;
565 	}
566 
567 	if (p->chunk_relocs_idx == -1) {
568 		DRM_ERROR("No relocation chunk !\n");
569 		return -EINVAL;
570 	}
571 
572 
573 	do {
574 		r = radeon_cs_packet_parse(p, &pkt, p->idx);
575 		if (r)
576 			return r;
577 		switch (pkt.type) {
578 		case RADEON_PACKET_TYPE0:
579 			r = radeon_uvd_cs_reg(p, &pkt, &data0, &data1,
580 					      buf_sizes, &has_msg_cmd);
581 			if (r)
582 				return r;
583 			break;
584 		case RADEON_PACKET_TYPE2:
585 			p->idx += pkt.count + 2;
586 			break;
587 		default:
588 			DRM_ERROR("Unknown packet type %d !\n", pkt.type);
589 			return -EINVAL;
590 		}
591 	} while (p->idx < p->chunks[p->chunk_ib_idx].length_dw);
592 
593 	if (!has_msg_cmd) {
594 		DRM_ERROR("UVD-IBs need a msg command!\n");
595 		return -EINVAL;
596 	}
597 
598 	return 0;
599 }
600 
601 static int radeon_uvd_send_msg(struct radeon_device *rdev,
602 			       int ring, struct radeon_bo *bo,
603 			       struct radeon_fence **fence)
604 {
605 	struct ttm_validate_buffer tv;
606 	struct ww_acquire_ctx ticket;
607 	struct list_head head;
608 	struct radeon_ib ib;
609 	uint64_t addr;
610 	int i, r;
611 
612 	memset(&tv, 0, sizeof(tv));
613 	tv.bo = &bo->tbo;
614 
615 	INIT_LIST_HEAD(&head);
616 	list_add(&tv.head, &head);
617 
618 	r = ttm_eu_reserve_buffers(&ticket, &head);
619 	if (r)
620 		return r;
621 
622 	radeon_ttm_placement_from_domain(bo, RADEON_GEM_DOMAIN_VRAM);
623 	radeon_uvd_force_into_uvd_segment(bo);
624 
625 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
626 	if (r)
627 		goto err;
628 
629 	r = radeon_ib_get(rdev, ring, &ib, NULL, 64);
630 	if (r)
631 		goto err;
632 
633 	addr = radeon_bo_gpu_offset(bo);
634 	ib.ptr[0] = PACKET0(UVD_GPCOM_VCPU_DATA0, 0);
635 	ib.ptr[1] = addr;
636 	ib.ptr[2] = PACKET0(UVD_GPCOM_VCPU_DATA1, 0);
637 	ib.ptr[3] = addr >> 32;
638 	ib.ptr[4] = PACKET0(UVD_GPCOM_VCPU_CMD, 0);
639 	ib.ptr[5] = 0;
640 	for (i = 6; i < 16; ++i)
641 		ib.ptr[i] = PACKET2(0);
642 	ib.length_dw = 16;
643 
644 	r = radeon_ib_schedule(rdev, &ib, NULL);
645 	if (r)
646 		goto err;
647 	ttm_eu_fence_buffer_objects(&ticket, &head, ib.fence);
648 
649 	if (fence)
650 		*fence = radeon_fence_ref(ib.fence);
651 
652 	radeon_ib_free(rdev, &ib);
653 	radeon_bo_unref(&bo);
654 	return 0;
655 
656 err:
657 	ttm_eu_backoff_reservation(&ticket, &head);
658 	return r;
659 }
660 
661 /* multiple fence commands without any stream commands in between can
662    crash the vcpu so just try to emmit a dummy create/destroy msg to
663    avoid this */
664 int radeon_uvd_get_create_msg(struct radeon_device *rdev, int ring,
665 			      uint32_t handle, struct radeon_fence **fence)
666 {
667 	struct radeon_bo *bo;
668 	uint32_t *msg;
669 	int r, i;
670 
671 	r = radeon_bo_create(rdev, 1024, PAGE_SIZE, true,
672 			     RADEON_GEM_DOMAIN_VRAM, NULL, &bo);
673 	if (r)
674 		return r;
675 
676 	r = radeon_bo_reserve(bo, false);
677 	if (r) {
678 		radeon_bo_unref(&bo);
679 		return r;
680 	}
681 
682 	r = radeon_bo_kmap(bo, (void **)&msg);
683 	if (r) {
684 		radeon_bo_unreserve(bo);
685 		radeon_bo_unref(&bo);
686 		return r;
687 	}
688 
689 	/* stitch together an UVD create msg */
690 	msg[0] = cpu_to_le32(0x00000de4);
691 	msg[1] = cpu_to_le32(0x00000000);
692 	msg[2] = cpu_to_le32(handle);
693 	msg[3] = cpu_to_le32(0x00000000);
694 	msg[4] = cpu_to_le32(0x00000000);
695 	msg[5] = cpu_to_le32(0x00000000);
696 	msg[6] = cpu_to_le32(0x00000000);
697 	msg[7] = cpu_to_le32(0x00000780);
698 	msg[8] = cpu_to_le32(0x00000440);
699 	msg[9] = cpu_to_le32(0x00000000);
700 	msg[10] = cpu_to_le32(0x01b37000);
701 	for (i = 11; i < 1024; ++i)
702 		msg[i] = cpu_to_le32(0x0);
703 
704 	radeon_bo_kunmap(bo);
705 	radeon_bo_unreserve(bo);
706 
707 	return radeon_uvd_send_msg(rdev, ring, bo, fence);
708 }
709 
710 int radeon_uvd_get_destroy_msg(struct radeon_device *rdev, int ring,
711 			       uint32_t handle, struct radeon_fence **fence)
712 {
713 	struct radeon_bo *bo;
714 	uint32_t *msg;
715 	int r, i;
716 
717 	r = radeon_bo_create(rdev, 1024, PAGE_SIZE, true,
718 			     RADEON_GEM_DOMAIN_VRAM, NULL, &bo);
719 	if (r)
720 		return r;
721 
722 	r = radeon_bo_reserve(bo, false);
723 	if (r) {
724 		radeon_bo_unref(&bo);
725 		return r;
726 	}
727 
728 	r = radeon_bo_kmap(bo, (void **)&msg);
729 	if (r) {
730 		radeon_bo_unreserve(bo);
731 		radeon_bo_unref(&bo);
732 		return r;
733 	}
734 
735 	/* stitch together an UVD destroy msg */
736 	msg[0] = cpu_to_le32(0x00000de4);
737 	msg[1] = cpu_to_le32(0x00000002);
738 	msg[2] = cpu_to_le32(handle);
739 	msg[3] = cpu_to_le32(0x00000000);
740 	for (i = 4; i < 1024; ++i)
741 		msg[i] = cpu_to_le32(0x0);
742 
743 	radeon_bo_kunmap(bo);
744 	radeon_bo_unreserve(bo);
745 
746 	return radeon_uvd_send_msg(rdev, ring, bo, fence);
747 }
748 
749 /**
750  * radeon_uvd_count_handles - count number of open streams
751  *
752  * @rdev: radeon_device pointer
753  * @sd: number of SD streams
754  * @hd: number of HD streams
755  *
756  * Count the number of open SD/HD streams as a hint for power mangement
757  */
758 static void radeon_uvd_count_handles(struct radeon_device *rdev,
759 				     unsigned *sd, unsigned *hd)
760 {
761 	unsigned i;
762 
763 	*sd = 0;
764 	*hd = 0;
765 
766 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
767 		if (!atomic_read(&rdev->uvd.handles[i]))
768 			continue;
769 
770 		if (rdev->uvd.img_size[i] >= 720*576)
771 			++(*hd);
772 		else
773 			++(*sd);
774 	}
775 }
776 
777 static void radeon_uvd_idle_work_handler(struct work_struct *work)
778 {
779 	struct radeon_device *rdev =
780 		container_of(work, struct radeon_device, uvd.idle_work.work);
781 
782 	if (radeon_fence_count_emitted(rdev, R600_RING_TYPE_UVD_INDEX) == 0) {
783 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
784 			radeon_uvd_count_handles(rdev, &rdev->pm.dpm.sd,
785 						 &rdev->pm.dpm.hd);
786 			radeon_dpm_enable_uvd(rdev, false);
787 		} else {
788 			radeon_set_uvd_clocks(rdev, 0, 0);
789 		}
790 	} else {
791 		schedule_delayed_work(&rdev->uvd.idle_work,
792 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
793 	}
794 }
795 
796 void radeon_uvd_note_usage(struct radeon_device *rdev)
797 {
798 	bool streams_changed = false;
799 	bool set_clocks = !cancel_delayed_work_sync(&rdev->uvd.idle_work);
800 	set_clocks &= schedule_delayed_work(&rdev->uvd.idle_work,
801 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
802 
803 	if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
804 		unsigned hd = 0, sd = 0;
805 		radeon_uvd_count_handles(rdev, &sd, &hd);
806 		if ((rdev->pm.dpm.sd != sd) ||
807 		    (rdev->pm.dpm.hd != hd)) {
808 			rdev->pm.dpm.sd = sd;
809 			rdev->pm.dpm.hd = hd;
810 			streams_changed = true;
811 		}
812 	}
813 
814 	if (set_clocks || streams_changed) {
815 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
816 			radeon_dpm_enable_uvd(rdev, true);
817 		} else {
818 			radeon_set_uvd_clocks(rdev, 53300, 40000);
819 		}
820 	}
821 }
822 
823 static unsigned radeon_uvd_calc_upll_post_div(unsigned vco_freq,
824 					      unsigned target_freq,
825 					      unsigned pd_min,
826 					      unsigned pd_even)
827 {
828 	unsigned post_div = vco_freq / target_freq;
829 
830 	/* adjust to post divider minimum value */
831 	if (post_div < pd_min)
832 		post_div = pd_min;
833 
834 	/* we alway need a frequency less than or equal the target */
835 	if ((vco_freq / post_div) > target_freq)
836 		post_div += 1;
837 
838 	/* post dividers above a certain value must be even */
839 	if (post_div > pd_even && post_div % 2)
840 		post_div += 1;
841 
842 	return post_div;
843 }
844 
845 /**
846  * radeon_uvd_calc_upll_dividers - calc UPLL clock dividers
847  *
848  * @rdev: radeon_device pointer
849  * @vclk: wanted VCLK
850  * @dclk: wanted DCLK
851  * @vco_min: minimum VCO frequency
852  * @vco_max: maximum VCO frequency
853  * @fb_factor: factor to multiply vco freq with
854  * @fb_mask: limit and bitmask for feedback divider
855  * @pd_min: post divider minimum
856  * @pd_max: post divider maximum
857  * @pd_even: post divider must be even above this value
858  * @optimal_fb_div: resulting feedback divider
859  * @optimal_vclk_div: resulting vclk post divider
860  * @optimal_dclk_div: resulting dclk post divider
861  *
862  * Calculate dividers for UVDs UPLL (R6xx-SI, except APUs).
863  * Returns zero on success -EINVAL on error.
864  */
865 int radeon_uvd_calc_upll_dividers(struct radeon_device *rdev,
866 				  unsigned vclk, unsigned dclk,
867 				  unsigned vco_min, unsigned vco_max,
868 				  unsigned fb_factor, unsigned fb_mask,
869 				  unsigned pd_min, unsigned pd_max,
870 				  unsigned pd_even,
871 				  unsigned *optimal_fb_div,
872 				  unsigned *optimal_vclk_div,
873 				  unsigned *optimal_dclk_div)
874 {
875 	unsigned vco_freq, ref_freq = rdev->clock.spll.reference_freq;
876 
877 	/* start off with something large */
878 	unsigned optimal_score = ~0;
879 
880 	/* loop through vco from low to high */
881 	vco_min = max(max(vco_min, vclk), dclk);
882 	for (vco_freq = vco_min; vco_freq <= vco_max; vco_freq += 100) {
883 
884 		uint64_t fb_div = (uint64_t)vco_freq * fb_factor;
885 		unsigned vclk_div, dclk_div, score;
886 
887 		do_div(fb_div, ref_freq);
888 
889 		/* fb div out of range ? */
890 		if (fb_div > fb_mask)
891 			break; /* it can oly get worse */
892 
893 		fb_div &= fb_mask;
894 
895 		/* calc vclk divider with current vco freq */
896 		vclk_div = radeon_uvd_calc_upll_post_div(vco_freq, vclk,
897 							 pd_min, pd_even);
898 		if (vclk_div > pd_max)
899 			break; /* vco is too big, it has to stop */
900 
901 		/* calc dclk divider with current vco freq */
902 		dclk_div = radeon_uvd_calc_upll_post_div(vco_freq, dclk,
903 							 pd_min, pd_even);
904 		if (vclk_div > pd_max)
905 			break; /* vco is too big, it has to stop */
906 
907 		/* calc score with current vco freq */
908 		score = vclk - (vco_freq / vclk_div) + dclk - (vco_freq / dclk_div);
909 
910 		/* determine if this vco setting is better than current optimal settings */
911 		if (score < optimal_score) {
912 			*optimal_fb_div = fb_div;
913 			*optimal_vclk_div = vclk_div;
914 			*optimal_dclk_div = dclk_div;
915 			optimal_score = score;
916 			if (optimal_score == 0)
917 				break; /* it can't get better than this */
918 		}
919 	}
920 
921 	/* did we found a valid setup ? */
922 	if (optimal_score == ~0)
923 		return -EINVAL;
924 
925 	return 0;
926 }
927 
928 int radeon_uvd_send_upll_ctlreq(struct radeon_device *rdev,
929 				unsigned cg_upll_func_cntl)
930 {
931 	unsigned i;
932 
933 	/* make sure UPLL_CTLREQ is deasserted */
934 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
935 
936 	mdelay(10);
937 
938 	/* assert UPLL_CTLREQ */
939 	WREG32_P(cg_upll_func_cntl, UPLL_CTLREQ_MASK, ~UPLL_CTLREQ_MASK);
940 
941 	/* wait for CTLACK and CTLACK2 to get asserted */
942 	for (i = 0; i < 100; ++i) {
943 		uint32_t mask = UPLL_CTLACK_MASK | UPLL_CTLACK2_MASK;
944 		if ((RREG32(cg_upll_func_cntl) & mask) == mask)
945 			break;
946 		mdelay(10);
947 	}
948 
949 	/* deassert UPLL_CTLREQ */
950 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
951 
952 	if (i == 100) {
953 		DRM_ERROR("Timeout setting UVD clocks!\n");
954 		return -ETIMEDOUT;
955 	}
956 
957 	return 0;
958 }
959