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