xref: /linux/drivers/gpu/drm/radeon/radeon_uvd.c (revision 22511e665eadc2c49021886d07aaef39db90ca82)
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 
34 #include <drm/drm.h>
35 
36 #include "radeon.h"
37 #include "radeon_ucode.h"
38 #include "r600d.h"
39 
40 /* 1 second timeout */
41 #define UVD_IDLE_TIMEOUT_MS	1000
42 
43 /* Firmware Names */
44 #define FIRMWARE_R600		"radeon/R600_uvd.bin"
45 #define FIRMWARE_RS780		"radeon/RS780_uvd.bin"
46 #define FIRMWARE_RV770		"radeon/RV770_uvd.bin"
47 #define FIRMWARE_RV710		"radeon/RV710_uvd.bin"
48 #define FIRMWARE_CYPRESS	"radeon/CYPRESS_uvd.bin"
49 #define FIRMWARE_SUMO		"radeon/SUMO_uvd.bin"
50 #define FIRMWARE_TAHITI		"radeon/TAHITI_uvd.bin"
51 #define FIRMWARE_BONAIRE_LEGACY	"radeon/BONAIRE_uvd.bin"
52 #define FIRMWARE_BONAIRE	"radeon/bonaire_uvd.bin"
53 
54 MODULE_FIRMWARE(FIRMWARE_R600);
55 MODULE_FIRMWARE(FIRMWARE_RS780);
56 MODULE_FIRMWARE(FIRMWARE_RV770);
57 MODULE_FIRMWARE(FIRMWARE_RV710);
58 MODULE_FIRMWARE(FIRMWARE_CYPRESS);
59 MODULE_FIRMWARE(FIRMWARE_SUMO);
60 MODULE_FIRMWARE(FIRMWARE_TAHITI);
61 MODULE_FIRMWARE(FIRMWARE_BONAIRE_LEGACY);
62 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
63 
64 static void radeon_uvd_idle_work_handler(struct work_struct *work);
65 
66 int radeon_uvd_init(struct radeon_device *rdev)
67 {
68 	unsigned long bo_size;
69 	const char *fw_name = NULL, *legacy_fw_name = NULL;
70 	int i, r;
71 
72 	INIT_DELAYED_WORK(&rdev->uvd.idle_work, radeon_uvd_idle_work_handler);
73 
74 	switch (rdev->family) {
75 	case CHIP_RV610:
76 	case CHIP_RV630:
77 	case CHIP_RV670:
78 	case CHIP_RV620:
79 	case CHIP_RV635:
80 		legacy_fw_name = FIRMWARE_R600;
81 		break;
82 
83 	case CHIP_RS780:
84 	case CHIP_RS880:
85 		legacy_fw_name = FIRMWARE_RS780;
86 		break;
87 
88 	case CHIP_RV770:
89 		legacy_fw_name = FIRMWARE_RV770;
90 		break;
91 
92 	case CHIP_RV710:
93 	case CHIP_RV730:
94 	case CHIP_RV740:
95 		legacy_fw_name = FIRMWARE_RV710;
96 		break;
97 
98 	case CHIP_CYPRESS:
99 	case CHIP_HEMLOCK:
100 	case CHIP_JUNIPER:
101 	case CHIP_REDWOOD:
102 	case CHIP_CEDAR:
103 		legacy_fw_name = FIRMWARE_CYPRESS;
104 		break;
105 
106 	case CHIP_SUMO:
107 	case CHIP_SUMO2:
108 	case CHIP_PALM:
109 	case CHIP_CAYMAN:
110 	case CHIP_BARTS:
111 	case CHIP_TURKS:
112 	case CHIP_CAICOS:
113 		legacy_fw_name = FIRMWARE_SUMO;
114 		break;
115 
116 	case CHIP_TAHITI:
117 	case CHIP_VERDE:
118 	case CHIP_PITCAIRN:
119 	case CHIP_ARUBA:
120 	case CHIP_OLAND:
121 		legacy_fw_name = FIRMWARE_TAHITI;
122 		break;
123 
124 	case CHIP_BONAIRE:
125 	case CHIP_KABINI:
126 	case CHIP_KAVERI:
127 	case CHIP_HAWAII:
128 	case CHIP_MULLINS:
129 		legacy_fw_name = FIRMWARE_BONAIRE_LEGACY;
130 		fw_name = FIRMWARE_BONAIRE;
131 		break;
132 
133 	default:
134 		return -EINVAL;
135 	}
136 
137 	rdev->uvd.fw_header_present = false;
138 	rdev->uvd.max_handles = RADEON_DEFAULT_UVD_HANDLES;
139 	if (fw_name) {
140 		/* Let's try to load the newer firmware first */
141 		r = request_firmware(&rdev->uvd_fw, fw_name, rdev->dev);
142 		if (r) {
143 			dev_err(rdev->dev, "radeon_uvd: Can't load firmware \"%s\"\n",
144 				fw_name);
145 		} else {
146 			struct common_firmware_header *hdr = (void *)rdev->uvd_fw->data;
147 			unsigned version_major, version_minor, family_id;
148 
149 			r = radeon_ucode_validate(rdev->uvd_fw);
150 			if (r)
151 				return r;
152 
153 			rdev->uvd.fw_header_present = true;
154 
155 			family_id = (__force u32)(hdr->ucode_version) & 0xff;
156 			version_major = (le32_to_cpu((__force __le32)(hdr->ucode_version))
157 							 >> 24) & 0xff;
158 			version_minor = (le32_to_cpu((__force __le32)(hdr->ucode_version))
159 							 >> 8) & 0xff;
160 			DRM_INFO("Found UVD firmware Version: %u.%u Family ID: %u\n",
161 				 version_major, version_minor, family_id);
162 
163 			/*
164 			 * Limit the number of UVD handles depending on
165 			 * microcode major and minor versions.
166 			 */
167 			if ((version_major >= 0x01) && (version_minor >= 0x37))
168 				rdev->uvd.max_handles = RADEON_MAX_UVD_HANDLES;
169 		}
170 	}
171 
172 	/*
173 	 * In case there is only legacy firmware, or we encounter an error
174 	 * while loading the new firmware, we fall back to loading the legacy
175 	 * firmware now.
176 	 */
177 	if (!fw_name || r) {
178 		r = request_firmware(&rdev->uvd_fw, legacy_fw_name, rdev->dev);
179 		if (r) {
180 			dev_err(rdev->dev, "radeon_uvd: Can't load firmware \"%s\"\n",
181 				legacy_fw_name);
182 			return r;
183 		}
184 	}
185 
186 	bo_size = RADEON_GPU_PAGE_ALIGN(rdev->uvd_fw->size + 8) +
187 		  RADEON_UVD_STACK_SIZE + RADEON_UVD_HEAP_SIZE +
188 		  RADEON_UVD_SESSION_SIZE * rdev->uvd.max_handles;
189 	r = radeon_bo_create(rdev, bo_size, PAGE_SIZE, true,
190 			     RADEON_GEM_DOMAIN_VRAM, 0, NULL,
191 			     NULL, &rdev->uvd.vcpu_bo);
192 	if (r) {
193 		dev_err(rdev->dev, "(%d) failed to allocate UVD bo\n", r);
194 		return r;
195 	}
196 
197 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
198 	if (r) {
199 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
200 		dev_err(rdev->dev, "(%d) failed to reserve UVD bo\n", r);
201 		return r;
202 	}
203 
204 	r = radeon_bo_pin(rdev->uvd.vcpu_bo, RADEON_GEM_DOMAIN_VRAM,
205 			  &rdev->uvd.gpu_addr);
206 	if (r) {
207 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
208 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
209 		dev_err(rdev->dev, "(%d) UVD bo pin failed\n", r);
210 		return r;
211 	}
212 
213 	r = radeon_bo_kmap(rdev->uvd.vcpu_bo, &rdev->uvd.cpu_addr);
214 	if (r) {
215 		dev_err(rdev->dev, "(%d) UVD map failed\n", r);
216 		return r;
217 	}
218 
219 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
220 
221 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
222 		atomic_set(&rdev->uvd.handles[i], 0);
223 		rdev->uvd.filp[i] = NULL;
224 		rdev->uvd.img_size[i] = 0;
225 	}
226 
227 	return 0;
228 }
229 
230 void radeon_uvd_fini(struct radeon_device *rdev)
231 {
232 	int r;
233 
234 	if (rdev->uvd.vcpu_bo == NULL)
235 		return;
236 
237 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
238 	if (!r) {
239 		radeon_bo_kunmap(rdev->uvd.vcpu_bo);
240 		radeon_bo_unpin(rdev->uvd.vcpu_bo);
241 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
242 	}
243 
244 	radeon_bo_unref(&rdev->uvd.vcpu_bo);
245 
246 	radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_UVD_INDEX]);
247 
248 	release_firmware(rdev->uvd_fw);
249 }
250 
251 int radeon_uvd_suspend(struct radeon_device *rdev)
252 {
253 	int i, r;
254 
255 	if (rdev->uvd.vcpu_bo == NULL)
256 		return 0;
257 
258 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
259 		uint32_t handle = atomic_read(&rdev->uvd.handles[i]);
260 		if (handle != 0) {
261 			struct radeon_fence *fence;
262 
263 			radeon_uvd_note_usage(rdev);
264 
265 			r = radeon_uvd_get_destroy_msg(rdev,
266 				R600_RING_TYPE_UVD_INDEX, handle, &fence);
267 			if (r) {
268 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
269 				continue;
270 			}
271 
272 			radeon_fence_wait(fence, false);
273 			radeon_fence_unref(&fence);
274 
275 			rdev->uvd.filp[i] = NULL;
276 			atomic_set(&rdev->uvd.handles[i], 0);
277 		}
278 	}
279 
280 	return 0;
281 }
282 
283 int radeon_uvd_resume(struct radeon_device *rdev)
284 {
285 	unsigned size;
286 	void *ptr;
287 
288 	if (rdev->uvd.vcpu_bo == NULL)
289 		return -EINVAL;
290 
291 	memcpy_toio((void __iomem *)rdev->uvd.cpu_addr, rdev->uvd_fw->data, rdev->uvd_fw->size);
292 
293 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
294 	size -= rdev->uvd_fw->size;
295 
296 	ptr = rdev->uvd.cpu_addr;
297 	ptr += rdev->uvd_fw->size;
298 
299 	memset_io((void __iomem *)ptr, 0, size);
300 
301 	return 0;
302 }
303 
304 void radeon_uvd_force_into_uvd_segment(struct radeon_bo *rbo,
305 				       uint32_t allowed_domains)
306 {
307 	int i;
308 
309 	for (i = 0; i < rbo->placement.num_placement; ++i) {
310 		rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
311 		rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
312 	}
313 
314 	/* If it must be in VRAM it must be in the first segment as well */
315 	if (allowed_domains == RADEON_GEM_DOMAIN_VRAM)
316 		return;
317 
318 	/* abort if we already have more than one placement */
319 	if (rbo->placement.num_placement > 1)
320 		return;
321 
322 	/* add another 256MB segment */
323 	rbo->placements[1] = rbo->placements[0];
324 	rbo->placements[1].fpfn += (256 * 1024 * 1024) >> PAGE_SHIFT;
325 	rbo->placements[1].lpfn += (256 * 1024 * 1024) >> PAGE_SHIFT;
326 	rbo->placement.num_placement++;
327 	rbo->placement.num_busy_placement++;
328 }
329 
330 void radeon_uvd_free_handles(struct radeon_device *rdev, struct drm_file *filp)
331 {
332 	int i, r;
333 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
334 		uint32_t handle = atomic_read(&rdev->uvd.handles[i]);
335 		if (handle != 0 && rdev->uvd.filp[i] == filp) {
336 			struct radeon_fence *fence;
337 
338 			radeon_uvd_note_usage(rdev);
339 
340 			r = radeon_uvd_get_destroy_msg(rdev,
341 				R600_RING_TYPE_UVD_INDEX, handle, &fence);
342 			if (r) {
343 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
344 				continue;
345 			}
346 
347 			radeon_fence_wait(fence, false);
348 			radeon_fence_unref(&fence);
349 
350 			rdev->uvd.filp[i] = NULL;
351 			atomic_set(&rdev->uvd.handles[i], 0);
352 		}
353 	}
354 }
355 
356 static int radeon_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
357 {
358 	unsigned stream_type = msg[4];
359 	unsigned width = msg[6];
360 	unsigned height = msg[7];
361 	unsigned dpb_size = msg[9];
362 	unsigned pitch = msg[28];
363 
364 	unsigned width_in_mb = width / 16;
365 	unsigned height_in_mb = ALIGN(height / 16, 2);
366 
367 	unsigned image_size, tmp, min_dpb_size;
368 
369 	image_size = width * height;
370 	image_size += image_size / 2;
371 	image_size = ALIGN(image_size, 1024);
372 
373 	switch (stream_type) {
374 	case 0: /* H264 */
375 
376 		/* reference picture buffer */
377 		min_dpb_size = image_size * 17;
378 
379 		/* macroblock context buffer */
380 		min_dpb_size += width_in_mb * height_in_mb * 17 * 192;
381 
382 		/* IT surface buffer */
383 		min_dpb_size += width_in_mb * height_in_mb * 32;
384 		break;
385 
386 	case 1: /* VC1 */
387 
388 		/* reference picture buffer */
389 		min_dpb_size = image_size * 3;
390 
391 		/* CONTEXT_BUFFER */
392 		min_dpb_size += width_in_mb * height_in_mb * 128;
393 
394 		/* IT surface buffer */
395 		min_dpb_size += width_in_mb * 64;
396 
397 		/* DB surface buffer */
398 		min_dpb_size += width_in_mb * 128;
399 
400 		/* BP */
401 		tmp = max(width_in_mb, height_in_mb);
402 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
403 		break;
404 
405 	case 3: /* MPEG2 */
406 
407 		/* reference picture buffer */
408 		min_dpb_size = image_size * 3;
409 		break;
410 
411 	case 4: /* MPEG4 */
412 
413 		/* reference picture buffer */
414 		min_dpb_size = image_size * 3;
415 
416 		/* CM */
417 		min_dpb_size += width_in_mb * height_in_mb * 64;
418 
419 		/* IT surface buffer */
420 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
421 		break;
422 
423 	default:
424 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
425 		return -EINVAL;
426 	}
427 
428 	if (width > pitch) {
429 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
430 		return -EINVAL;
431 	}
432 
433 	if (dpb_size < min_dpb_size) {
434 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
435 			  dpb_size, min_dpb_size);
436 		return -EINVAL;
437 	}
438 
439 	buf_sizes[0x1] = dpb_size;
440 	buf_sizes[0x2] = image_size;
441 	return 0;
442 }
443 
444 static int radeon_uvd_validate_codec(struct radeon_cs_parser *p,
445 				     unsigned stream_type)
446 {
447 	switch (stream_type) {
448 	case 0: /* H264 */
449 	case 1: /* VC1 */
450 		/* always supported */
451 		return 0;
452 
453 	case 3: /* MPEG2 */
454 	case 4: /* MPEG4 */
455 		/* only since UVD 3 */
456 		if (p->rdev->family >= CHIP_PALM)
457 			return 0;
458 
459 		fallthrough;
460 	default:
461 		DRM_ERROR("UVD codec not supported by hardware %d!\n",
462 			  stream_type);
463 		return -EINVAL;
464 	}
465 }
466 
467 static int radeon_uvd_cs_msg(struct radeon_cs_parser *p, struct radeon_bo *bo,
468 			     unsigned offset, unsigned buf_sizes[])
469 {
470 	int32_t *msg, msg_type, handle;
471 	unsigned img_size = 0;
472 	void *ptr;
473 	long r;
474 	int i;
475 
476 	if (offset & 0x3F) {
477 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
478 		return -EINVAL;
479 	}
480 
481 	r = dma_resv_wait_timeout(bo->tbo.base.resv, false, false,
482 				  MAX_SCHEDULE_TIMEOUT);
483 	if (r <= 0) {
484 		DRM_ERROR("Failed waiting for UVD message (%ld)!\n", r);
485 		return r ? r : -ETIME;
486 	}
487 
488 	r = radeon_bo_kmap(bo, &ptr);
489 	if (r) {
490 		DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r);
491 		return r;
492 	}
493 
494 	msg = ptr + offset;
495 
496 	msg_type = msg[1];
497 	handle = msg[2];
498 
499 	if (handle == 0) {
500 		radeon_bo_kunmap(bo);
501 		DRM_ERROR("Invalid UVD handle!\n");
502 		return -EINVAL;
503 	}
504 
505 	switch (msg_type) {
506 	case 0:
507 		/* it's a create msg, calc image size (width * height) */
508 		img_size = msg[7] * msg[8];
509 
510 		r = radeon_uvd_validate_codec(p, msg[4]);
511 		radeon_bo_kunmap(bo);
512 		if (r)
513 			return r;
514 
515 		/* try to alloc a new handle */
516 		for (i = 0; i < p->rdev->uvd.max_handles; ++i) {
517 			if (atomic_read(&p->rdev->uvd.handles[i]) == handle) {
518 				DRM_ERROR("Handle 0x%x already in use!\n", handle);
519 				return -EINVAL;
520 			}
521 
522 			if (!atomic_cmpxchg(&p->rdev->uvd.handles[i], 0, handle)) {
523 				p->rdev->uvd.filp[i] = p->filp;
524 				p->rdev->uvd.img_size[i] = img_size;
525 				return 0;
526 			}
527 		}
528 
529 		DRM_ERROR("No more free UVD handles!\n");
530 		return -EINVAL;
531 
532 	case 1:
533 		/* it's a decode msg, validate codec and calc buffer sizes */
534 		r = radeon_uvd_validate_codec(p, msg[4]);
535 		if (!r)
536 			r = radeon_uvd_cs_msg_decode(msg, buf_sizes);
537 		radeon_bo_kunmap(bo);
538 		if (r)
539 			return r;
540 
541 		/* validate the handle */
542 		for (i = 0; i < p->rdev->uvd.max_handles; ++i) {
543 			if (atomic_read(&p->rdev->uvd.handles[i]) == handle) {
544 				if (p->rdev->uvd.filp[i] != p->filp) {
545 					DRM_ERROR("UVD handle collision detected!\n");
546 					return -EINVAL;
547 				}
548 				return 0;
549 			}
550 		}
551 
552 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
553 		return -ENOENT;
554 
555 	case 2:
556 		/* it's a destroy msg, free the handle */
557 		for (i = 0; i < p->rdev->uvd.max_handles; ++i)
558 			atomic_cmpxchg(&p->rdev->uvd.handles[i], handle, 0);
559 		radeon_bo_kunmap(bo);
560 		return 0;
561 
562 	default:
563 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
564 	}
565 
566 	radeon_bo_kunmap(bo);
567 	return -EINVAL;
568 }
569 
570 static int radeon_uvd_cs_reloc(struct radeon_cs_parser *p,
571 			       int data0, int data1,
572 			       unsigned buf_sizes[], bool *has_msg_cmd)
573 {
574 	struct radeon_cs_chunk *relocs_chunk;
575 	struct radeon_bo_list *reloc;
576 	unsigned idx, cmd, offset;
577 	uint64_t start, end;
578 	int r;
579 
580 	relocs_chunk = p->chunk_relocs;
581 	offset = radeon_get_ib_value(p, data0);
582 	idx = radeon_get_ib_value(p, data1);
583 	if (idx >= relocs_chunk->length_dw) {
584 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
585 			  idx, relocs_chunk->length_dw);
586 		return -EINVAL;
587 	}
588 
589 	reloc = &p->relocs[(idx / 4)];
590 	start = reloc->gpu_offset;
591 	end = start + radeon_bo_size(reloc->robj);
592 	start += offset;
593 
594 	p->ib.ptr[data0] = start & 0xFFFFFFFF;
595 	p->ib.ptr[data1] = start >> 32;
596 
597 	cmd = radeon_get_ib_value(p, p->idx) >> 1;
598 
599 	if (cmd < 0x4) {
600 		if (end <= start) {
601 			DRM_ERROR("invalid reloc offset %X!\n", offset);
602 			return -EINVAL;
603 		}
604 		if ((end - start) < buf_sizes[cmd]) {
605 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
606 				  (unsigned)(end - start), buf_sizes[cmd]);
607 			return -EINVAL;
608 		}
609 
610 	} else if (cmd != 0x100) {
611 		DRM_ERROR("invalid UVD command %X!\n", cmd);
612 		return -EINVAL;
613 	}
614 
615 	if ((start >> 28) != ((end - 1) >> 28)) {
616 		DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
617 			  start, end);
618 		return -EINVAL;
619 	}
620 
621 	/* TODO: is this still necessary on NI+ ? */
622 	if ((cmd == 0 || cmd == 0x3) &&
623 	    (start >> 28) != (p->rdev->uvd.gpu_addr >> 28)) {
624 		DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
625 			  start, end);
626 		return -EINVAL;
627 	}
628 
629 	if (cmd == 0) {
630 		if (*has_msg_cmd) {
631 			DRM_ERROR("More than one message in a UVD-IB!\n");
632 			return -EINVAL;
633 		}
634 		*has_msg_cmd = true;
635 		r = radeon_uvd_cs_msg(p, reloc->robj, offset, buf_sizes);
636 		if (r)
637 			return r;
638 	} else if (!*has_msg_cmd) {
639 		DRM_ERROR("Message needed before other commands are send!\n");
640 		return -EINVAL;
641 	}
642 
643 	return 0;
644 }
645 
646 static int radeon_uvd_cs_reg(struct radeon_cs_parser *p,
647 			     struct radeon_cs_packet *pkt,
648 			     int *data0, int *data1,
649 			     unsigned buf_sizes[],
650 			     bool *has_msg_cmd)
651 {
652 	int i, r;
653 
654 	p->idx++;
655 	for (i = 0; i <= pkt->count; ++i) {
656 		switch (pkt->reg + i*4) {
657 		case UVD_GPCOM_VCPU_DATA0:
658 			*data0 = p->idx;
659 			break;
660 		case UVD_GPCOM_VCPU_DATA1:
661 			*data1 = p->idx;
662 			break;
663 		case UVD_GPCOM_VCPU_CMD:
664 			r = radeon_uvd_cs_reloc(p, *data0, *data1,
665 						buf_sizes, has_msg_cmd);
666 			if (r)
667 				return r;
668 			break;
669 		case UVD_ENGINE_CNTL:
670 		case UVD_NO_OP:
671 			break;
672 		default:
673 			DRM_ERROR("Invalid reg 0x%X!\n",
674 				  pkt->reg + i*4);
675 			return -EINVAL;
676 		}
677 		p->idx++;
678 	}
679 	return 0;
680 }
681 
682 int radeon_uvd_cs_parse(struct radeon_cs_parser *p)
683 {
684 	struct radeon_cs_packet pkt;
685 	int r, data0 = 0, data1 = 0;
686 
687 	/* does the IB has a msg command */
688 	bool has_msg_cmd = false;
689 
690 	/* minimum buffer sizes */
691 	unsigned buf_sizes[] = {
692 		[0x00000000]	=	2048,
693 		[0x00000001]	=	32 * 1024 * 1024,
694 		[0x00000002]	=	2048 * 1152 * 3,
695 		[0x00000003]	=	2048,
696 	};
697 
698 	if (p->chunk_ib->length_dw % 16) {
699 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
700 			  p->chunk_ib->length_dw);
701 		return -EINVAL;
702 	}
703 
704 	if (p->chunk_relocs == NULL) {
705 		DRM_ERROR("No relocation chunk !\n");
706 		return -EINVAL;
707 	}
708 
709 
710 	do {
711 		r = radeon_cs_packet_parse(p, &pkt, p->idx);
712 		if (r)
713 			return r;
714 		switch (pkt.type) {
715 		case RADEON_PACKET_TYPE0:
716 			r = radeon_uvd_cs_reg(p, &pkt, &data0, &data1,
717 					      buf_sizes, &has_msg_cmd);
718 			if (r)
719 				return r;
720 			break;
721 		case RADEON_PACKET_TYPE2:
722 			p->idx += pkt.count + 2;
723 			break;
724 		default:
725 			DRM_ERROR("Unknown packet type %d !\n", pkt.type);
726 			return -EINVAL;
727 		}
728 	} while (p->idx < p->chunk_ib->length_dw);
729 
730 	if (!has_msg_cmd) {
731 		DRM_ERROR("UVD-IBs need a msg command!\n");
732 		return -EINVAL;
733 	}
734 
735 	return 0;
736 }
737 
738 static int radeon_uvd_send_msg(struct radeon_device *rdev,
739 			       int ring, uint64_t addr,
740 			       struct radeon_fence **fence)
741 {
742 	struct radeon_ib ib;
743 	int i, r;
744 
745 	r = radeon_ib_get(rdev, ring, &ib, NULL, 64);
746 	if (r)
747 		return r;
748 
749 	ib.ptr[0] = PACKET0(UVD_GPCOM_VCPU_DATA0, 0);
750 	ib.ptr[1] = addr;
751 	ib.ptr[2] = PACKET0(UVD_GPCOM_VCPU_DATA1, 0);
752 	ib.ptr[3] = addr >> 32;
753 	ib.ptr[4] = PACKET0(UVD_GPCOM_VCPU_CMD, 0);
754 	ib.ptr[5] = 0;
755 	for (i = 6; i < 16; i += 2) {
756 		ib.ptr[i] = PACKET0(UVD_NO_OP, 0);
757 		ib.ptr[i+1] = 0;
758 	}
759 	ib.length_dw = 16;
760 
761 	r = radeon_ib_schedule(rdev, &ib, NULL, false);
762 
763 	if (fence)
764 		*fence = radeon_fence_ref(ib.fence);
765 
766 	radeon_ib_free(rdev, &ib);
767 	return r;
768 }
769 
770 /*
771  * multiple fence commands without any stream commands in between can
772  * crash the vcpu so just try to emmit a dummy create/destroy msg to
773  * avoid this
774  */
775 int radeon_uvd_get_create_msg(struct radeon_device *rdev, int ring,
776 			      uint32_t handle, struct radeon_fence **fence)
777 {
778 	/* we use the last page of the vcpu bo for the UVD message */
779 	uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
780 		RADEON_GPU_PAGE_SIZE;
781 
782 	uint32_t __iomem *msg = (void __iomem *)(rdev->uvd.cpu_addr + offs);
783 	uint64_t addr = rdev->uvd.gpu_addr + offs;
784 
785 	int r, i;
786 
787 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
788 	if (r)
789 		return r;
790 
791 	/* stitch together an UVD create msg */
792 	writel((__force u32)cpu_to_le32(0x00000de4), &msg[0]);
793 	writel(0x0, (void __iomem *)&msg[1]);
794 	writel((__force u32)cpu_to_le32(handle), &msg[2]);
795 	writel(0x0, &msg[3]);
796 	writel(0x0, &msg[4]);
797 	writel(0x0, &msg[5]);
798 	writel(0x0, &msg[6]);
799 	writel((__force u32)cpu_to_le32(0x00000780), &msg[7]);
800 	writel((__force u32)cpu_to_le32(0x00000440), &msg[8]);
801 	writel(0x0, &msg[9]);
802 	writel((__force u32)cpu_to_le32(0x01b37000), &msg[10]);
803 	for (i = 11; i < 1024; ++i)
804 		writel(0x0, &msg[i]);
805 
806 	r = radeon_uvd_send_msg(rdev, ring, addr, fence);
807 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
808 	return r;
809 }
810 
811 int radeon_uvd_get_destroy_msg(struct radeon_device *rdev, int ring,
812 			       uint32_t handle, struct radeon_fence **fence)
813 {
814 	/* we use the last page of the vcpu bo for the UVD message */
815 	uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
816 		RADEON_GPU_PAGE_SIZE;
817 
818 	uint32_t __iomem *msg = (void __iomem *)(rdev->uvd.cpu_addr + offs);
819 	uint64_t addr = rdev->uvd.gpu_addr + offs;
820 
821 	int r, i;
822 
823 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
824 	if (r)
825 		return r;
826 
827 	/* stitch together an UVD destroy msg */
828 	writel((__force u32)cpu_to_le32(0x00000de4), &msg[0]);
829 	writel((__force u32)cpu_to_le32(0x00000002), &msg[1]);
830 	writel((__force u32)cpu_to_le32(handle), &msg[2]);
831 	writel(0x0, &msg[3]);
832 	for (i = 4; i < 1024; ++i)
833 		writel(0x0, &msg[i]);
834 
835 	r = radeon_uvd_send_msg(rdev, ring, addr, fence);
836 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
837 	return r;
838 }
839 
840 /**
841  * radeon_uvd_count_handles - count number of open streams
842  *
843  * @rdev: radeon_device pointer
844  * @sd: number of SD streams
845  * @hd: number of HD streams
846  *
847  * Count the number of open SD/HD streams as a hint for power mangement
848  */
849 static void radeon_uvd_count_handles(struct radeon_device *rdev,
850 				     unsigned *sd, unsigned *hd)
851 {
852 	unsigned i;
853 
854 	*sd = 0;
855 	*hd = 0;
856 
857 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
858 		if (!atomic_read(&rdev->uvd.handles[i]))
859 			continue;
860 
861 		if (rdev->uvd.img_size[i] >= 720*576)
862 			++(*hd);
863 		else
864 			++(*sd);
865 	}
866 }
867 
868 static void radeon_uvd_idle_work_handler(struct work_struct *work)
869 {
870 	struct radeon_device *rdev =
871 		container_of(work, struct radeon_device, uvd.idle_work.work);
872 
873 	if (radeon_fence_count_emitted(rdev, R600_RING_TYPE_UVD_INDEX) == 0) {
874 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
875 			radeon_uvd_count_handles(rdev, &rdev->pm.dpm.sd,
876 						 &rdev->pm.dpm.hd);
877 			radeon_dpm_enable_uvd(rdev, false);
878 		} else {
879 			radeon_set_uvd_clocks(rdev, 0, 0);
880 		}
881 	} else {
882 		schedule_delayed_work(&rdev->uvd.idle_work,
883 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
884 	}
885 }
886 
887 void radeon_uvd_note_usage(struct radeon_device *rdev)
888 {
889 	bool streams_changed = false;
890 	bool set_clocks = !cancel_delayed_work_sync(&rdev->uvd.idle_work);
891 	set_clocks &= schedule_delayed_work(&rdev->uvd.idle_work,
892 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
893 
894 	if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
895 		unsigned hd = 0, sd = 0;
896 		radeon_uvd_count_handles(rdev, &sd, &hd);
897 		if ((rdev->pm.dpm.sd != sd) ||
898 		    (rdev->pm.dpm.hd != hd)) {
899 			rdev->pm.dpm.sd = sd;
900 			rdev->pm.dpm.hd = hd;
901 			/* disable this for now */
902 			/*streams_changed = true;*/
903 		}
904 	}
905 
906 	if (set_clocks || streams_changed) {
907 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
908 			radeon_dpm_enable_uvd(rdev, true);
909 		} else {
910 			radeon_set_uvd_clocks(rdev, 53300, 40000);
911 		}
912 	}
913 }
914 
915 static unsigned radeon_uvd_calc_upll_post_div(unsigned vco_freq,
916 					      unsigned target_freq,
917 					      unsigned pd_min,
918 					      unsigned pd_even)
919 {
920 	unsigned post_div = vco_freq / target_freq;
921 
922 	/* adjust to post divider minimum value */
923 	if (post_div < pd_min)
924 		post_div = pd_min;
925 
926 	/* we alway need a frequency less than or equal the target */
927 	if ((vco_freq / post_div) > target_freq)
928 		post_div += 1;
929 
930 	/* post dividers above a certain value must be even */
931 	if (post_div > pd_even && post_div % 2)
932 		post_div += 1;
933 
934 	return post_div;
935 }
936 
937 /**
938  * radeon_uvd_calc_upll_dividers - calc UPLL clock dividers
939  *
940  * @rdev: radeon_device pointer
941  * @vclk: wanted VCLK
942  * @dclk: wanted DCLK
943  * @vco_min: minimum VCO frequency
944  * @vco_max: maximum VCO frequency
945  * @fb_factor: factor to multiply vco freq with
946  * @fb_mask: limit and bitmask for feedback divider
947  * @pd_min: post divider minimum
948  * @pd_max: post divider maximum
949  * @pd_even: post divider must be even above this value
950  * @optimal_fb_div: resulting feedback divider
951  * @optimal_vclk_div: resulting vclk post divider
952  * @optimal_dclk_div: resulting dclk post divider
953  *
954  * Calculate dividers for UVDs UPLL (R6xx-SI, except APUs).
955  * Returns zero on success -EINVAL on error.
956  */
957 int radeon_uvd_calc_upll_dividers(struct radeon_device *rdev,
958 				  unsigned vclk, unsigned dclk,
959 				  unsigned vco_min, unsigned vco_max,
960 				  unsigned fb_factor, unsigned fb_mask,
961 				  unsigned pd_min, unsigned pd_max,
962 				  unsigned pd_even,
963 				  unsigned *optimal_fb_div,
964 				  unsigned *optimal_vclk_div,
965 				  unsigned *optimal_dclk_div)
966 {
967 	unsigned vco_freq, ref_freq = rdev->clock.spll.reference_freq;
968 
969 	/* start off with something large */
970 	unsigned optimal_score = ~0;
971 
972 	/* loop through vco from low to high */
973 	vco_min = max(max(vco_min, vclk), dclk);
974 	for (vco_freq = vco_min; vco_freq <= vco_max; vco_freq += 100) {
975 
976 		uint64_t fb_div = (uint64_t)vco_freq * fb_factor;
977 		unsigned vclk_div, dclk_div, score;
978 
979 		do_div(fb_div, ref_freq);
980 
981 		/* fb div out of range ? */
982 		if (fb_div > fb_mask)
983 			break; /* it can oly get worse */
984 
985 		fb_div &= fb_mask;
986 
987 		/* calc vclk divider with current vco freq */
988 		vclk_div = radeon_uvd_calc_upll_post_div(vco_freq, vclk,
989 							 pd_min, pd_even);
990 		if (vclk_div > pd_max)
991 			break; /* vco is too big, it has to stop */
992 
993 		/* calc dclk divider with current vco freq */
994 		dclk_div = radeon_uvd_calc_upll_post_div(vco_freq, dclk,
995 							 pd_min, pd_even);
996 		if (dclk_div > pd_max)
997 			break; /* vco is too big, it has to stop */
998 
999 		/* calc score with current vco freq */
1000 		score = vclk - (vco_freq / vclk_div) + dclk - (vco_freq / dclk_div);
1001 
1002 		/* determine if this vco setting is better than current optimal settings */
1003 		if (score < optimal_score) {
1004 			*optimal_fb_div = fb_div;
1005 			*optimal_vclk_div = vclk_div;
1006 			*optimal_dclk_div = dclk_div;
1007 			optimal_score = score;
1008 			if (optimal_score == 0)
1009 				break; /* it can't get better than this */
1010 		}
1011 	}
1012 
1013 	/* did we found a valid setup ? */
1014 	if (optimal_score == ~0)
1015 		return -EINVAL;
1016 
1017 	return 0;
1018 }
1019 
1020 int radeon_uvd_send_upll_ctlreq(struct radeon_device *rdev,
1021 				unsigned cg_upll_func_cntl)
1022 {
1023 	unsigned i;
1024 
1025 	/* make sure UPLL_CTLREQ is deasserted */
1026 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
1027 
1028 	mdelay(10);
1029 
1030 	/* assert UPLL_CTLREQ */
1031 	WREG32_P(cg_upll_func_cntl, UPLL_CTLREQ_MASK, ~UPLL_CTLREQ_MASK);
1032 
1033 	/* wait for CTLACK and CTLACK2 to get asserted */
1034 	for (i = 0; i < 100; ++i) {
1035 		uint32_t mask = UPLL_CTLACK_MASK | UPLL_CTLACK2_MASK;
1036 		if ((RREG32(cg_upll_func_cntl) & mask) == mask)
1037 			break;
1038 		mdelay(10);
1039 	}
1040 
1041 	/* deassert UPLL_CTLREQ */
1042 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
1043 
1044 	if (i == 100) {
1045 		DRM_ERROR("Timeout setting UVD clocks!\n");
1046 		return -ETIMEDOUT;
1047 	}
1048 
1049 	return 0;
1050 }
1051