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 	struct dma_fence *f;
473 	void *ptr;
474 
475 	int i, r;
476 
477 	if (offset & 0x3F) {
478 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
479 		return -EINVAL;
480 	}
481 
482 	f = dma_resv_excl_fence(bo->tbo.base.resv);
483 	if (f) {
484 		r = radeon_fence_wait((struct radeon_fence *)f, false);
485 		if (r) {
486 			DRM_ERROR("Failed waiting for UVD message (%d)!\n", r);
487 			return r;
488 		}
489 	}
490 
491 	r = radeon_bo_kmap(bo, &ptr);
492 	if (r) {
493 		DRM_ERROR("Failed mapping the UVD message (%d)!\n", r);
494 		return r;
495 	}
496 
497 	msg = ptr + offset;
498 
499 	msg_type = msg[1];
500 	handle = msg[2];
501 
502 	if (handle == 0) {
503 		DRM_ERROR("Invalid UVD handle!\n");
504 		return -EINVAL;
505 	}
506 
507 	switch (msg_type) {
508 	case 0:
509 		/* it's a create msg, calc image size (width * height) */
510 		img_size = msg[7] * msg[8];
511 
512 		r = radeon_uvd_validate_codec(p, msg[4]);
513 		radeon_bo_kunmap(bo);
514 		if (r)
515 			return r;
516 
517 		/* try to alloc a new handle */
518 		for (i = 0; i < p->rdev->uvd.max_handles; ++i) {
519 			if (atomic_read(&p->rdev->uvd.handles[i]) == handle) {
520 				DRM_ERROR("Handle 0x%x already in use!\n", handle);
521 				return -EINVAL;
522 			}
523 
524 			if (!atomic_cmpxchg(&p->rdev->uvd.handles[i], 0, handle)) {
525 				p->rdev->uvd.filp[i] = p->filp;
526 				p->rdev->uvd.img_size[i] = img_size;
527 				return 0;
528 			}
529 		}
530 
531 		DRM_ERROR("No more free UVD handles!\n");
532 		return -EINVAL;
533 
534 	case 1:
535 		/* it's a decode msg, validate codec and calc buffer sizes */
536 		r = radeon_uvd_validate_codec(p, msg[4]);
537 		if (!r)
538 			r = radeon_uvd_cs_msg_decode(msg, buf_sizes);
539 		radeon_bo_kunmap(bo);
540 		if (r)
541 			return r;
542 
543 		/* validate the handle */
544 		for (i = 0; i < p->rdev->uvd.max_handles; ++i) {
545 			if (atomic_read(&p->rdev->uvd.handles[i]) == handle) {
546 				if (p->rdev->uvd.filp[i] != p->filp) {
547 					DRM_ERROR("UVD handle collision detected!\n");
548 					return -EINVAL;
549 				}
550 				return 0;
551 			}
552 		}
553 
554 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
555 		return -ENOENT;
556 
557 	case 2:
558 		/* it's a destroy msg, free the handle */
559 		for (i = 0; i < p->rdev->uvd.max_handles; ++i)
560 			atomic_cmpxchg(&p->rdev->uvd.handles[i], handle, 0);
561 		radeon_bo_kunmap(bo);
562 		return 0;
563 
564 	default:
565 
566 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
567 		return -EINVAL;
568 	}
569 
570 	BUG();
571 	return -EINVAL;
572 }
573 
574 static int radeon_uvd_cs_reloc(struct radeon_cs_parser *p,
575 			       int data0, int data1,
576 			       unsigned buf_sizes[], bool *has_msg_cmd)
577 {
578 	struct radeon_cs_chunk *relocs_chunk;
579 	struct radeon_bo_list *reloc;
580 	unsigned idx, cmd, offset;
581 	uint64_t start, end;
582 	int r;
583 
584 	relocs_chunk = p->chunk_relocs;
585 	offset = radeon_get_ib_value(p, data0);
586 	idx = radeon_get_ib_value(p, data1);
587 	if (idx >= relocs_chunk->length_dw) {
588 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
589 			  idx, relocs_chunk->length_dw);
590 		return -EINVAL;
591 	}
592 
593 	reloc = &p->relocs[(idx / 4)];
594 	start = reloc->gpu_offset;
595 	end = start + radeon_bo_size(reloc->robj);
596 	start += offset;
597 
598 	p->ib.ptr[data0] = start & 0xFFFFFFFF;
599 	p->ib.ptr[data1] = start >> 32;
600 
601 	cmd = radeon_get_ib_value(p, p->idx) >> 1;
602 
603 	if (cmd < 0x4) {
604 		if (end <= start) {
605 			DRM_ERROR("invalid reloc offset %X!\n", offset);
606 			return -EINVAL;
607 		}
608 		if ((end - start) < buf_sizes[cmd]) {
609 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
610 				  (unsigned)(end - start), buf_sizes[cmd]);
611 			return -EINVAL;
612 		}
613 
614 	} else if (cmd != 0x100) {
615 		DRM_ERROR("invalid UVD command %X!\n", cmd);
616 		return -EINVAL;
617 	}
618 
619 	if ((start >> 28) != ((end - 1) >> 28)) {
620 		DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
621 			  start, end);
622 		return -EINVAL;
623 	}
624 
625 	/* TODO: is this still necessary on NI+ ? */
626 	if ((cmd == 0 || cmd == 0x3) &&
627 	    (start >> 28) != (p->rdev->uvd.gpu_addr >> 28)) {
628 		DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
629 			  start, end);
630 		return -EINVAL;
631 	}
632 
633 	if (cmd == 0) {
634 		if (*has_msg_cmd) {
635 			DRM_ERROR("More than one message in a UVD-IB!\n");
636 			return -EINVAL;
637 		}
638 		*has_msg_cmd = true;
639 		r = radeon_uvd_cs_msg(p, reloc->robj, offset, buf_sizes);
640 		if (r)
641 			return r;
642 	} else if (!*has_msg_cmd) {
643 		DRM_ERROR("Message needed before other commands are send!\n");
644 		return -EINVAL;
645 	}
646 
647 	return 0;
648 }
649 
650 static int radeon_uvd_cs_reg(struct radeon_cs_parser *p,
651 			     struct radeon_cs_packet *pkt,
652 			     int *data0, int *data1,
653 			     unsigned buf_sizes[],
654 			     bool *has_msg_cmd)
655 {
656 	int i, r;
657 
658 	p->idx++;
659 	for (i = 0; i <= pkt->count; ++i) {
660 		switch (pkt->reg + i*4) {
661 		case UVD_GPCOM_VCPU_DATA0:
662 			*data0 = p->idx;
663 			break;
664 		case UVD_GPCOM_VCPU_DATA1:
665 			*data1 = p->idx;
666 			break;
667 		case UVD_GPCOM_VCPU_CMD:
668 			r = radeon_uvd_cs_reloc(p, *data0, *data1,
669 						buf_sizes, has_msg_cmd);
670 			if (r)
671 				return r;
672 			break;
673 		case UVD_ENGINE_CNTL:
674 		case UVD_NO_OP:
675 			break;
676 		default:
677 			DRM_ERROR("Invalid reg 0x%X!\n",
678 				  pkt->reg + i*4);
679 			return -EINVAL;
680 		}
681 		p->idx++;
682 	}
683 	return 0;
684 }
685 
686 int radeon_uvd_cs_parse(struct radeon_cs_parser *p)
687 {
688 	struct radeon_cs_packet pkt;
689 	int r, data0 = 0, data1 = 0;
690 
691 	/* does the IB has a msg command */
692 	bool has_msg_cmd = false;
693 
694 	/* minimum buffer sizes */
695 	unsigned buf_sizes[] = {
696 		[0x00000000]	=	2048,
697 		[0x00000001]	=	32 * 1024 * 1024,
698 		[0x00000002]	=	2048 * 1152 * 3,
699 		[0x00000003]	=	2048,
700 	};
701 
702 	if (p->chunk_ib->length_dw % 16) {
703 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
704 			  p->chunk_ib->length_dw);
705 		return -EINVAL;
706 	}
707 
708 	if (p->chunk_relocs == NULL) {
709 		DRM_ERROR("No relocation chunk !\n");
710 		return -EINVAL;
711 	}
712 
713 
714 	do {
715 		r = radeon_cs_packet_parse(p, &pkt, p->idx);
716 		if (r)
717 			return r;
718 		switch (pkt.type) {
719 		case RADEON_PACKET_TYPE0:
720 			r = radeon_uvd_cs_reg(p, &pkt, &data0, &data1,
721 					      buf_sizes, &has_msg_cmd);
722 			if (r)
723 				return r;
724 			break;
725 		case RADEON_PACKET_TYPE2:
726 			p->idx += pkt.count + 2;
727 			break;
728 		default:
729 			DRM_ERROR("Unknown packet type %d !\n", pkt.type);
730 			return -EINVAL;
731 		}
732 	} while (p->idx < p->chunk_ib->length_dw);
733 
734 	if (!has_msg_cmd) {
735 		DRM_ERROR("UVD-IBs need a msg command!\n");
736 		return -EINVAL;
737 	}
738 
739 	return 0;
740 }
741 
742 static int radeon_uvd_send_msg(struct radeon_device *rdev,
743 			       int ring, uint64_t addr,
744 			       struct radeon_fence **fence)
745 {
746 	struct radeon_ib ib;
747 	int i, r;
748 
749 	r = radeon_ib_get(rdev, ring, &ib, NULL, 64);
750 	if (r)
751 		return r;
752 
753 	ib.ptr[0] = PACKET0(UVD_GPCOM_VCPU_DATA0, 0);
754 	ib.ptr[1] = addr;
755 	ib.ptr[2] = PACKET0(UVD_GPCOM_VCPU_DATA1, 0);
756 	ib.ptr[3] = addr >> 32;
757 	ib.ptr[4] = PACKET0(UVD_GPCOM_VCPU_CMD, 0);
758 	ib.ptr[5] = 0;
759 	for (i = 6; i < 16; i += 2) {
760 		ib.ptr[i] = PACKET0(UVD_NO_OP, 0);
761 		ib.ptr[i+1] = 0;
762 	}
763 	ib.length_dw = 16;
764 
765 	r = radeon_ib_schedule(rdev, &ib, NULL, false);
766 
767 	if (fence)
768 		*fence = radeon_fence_ref(ib.fence);
769 
770 	radeon_ib_free(rdev, &ib);
771 	return r;
772 }
773 
774 /*
775  * multiple fence commands without any stream commands in between can
776  * crash the vcpu so just try to emmit a dummy create/destroy msg to
777  * avoid this
778  */
779 int radeon_uvd_get_create_msg(struct radeon_device *rdev, int ring,
780 			      uint32_t handle, struct radeon_fence **fence)
781 {
782 	/* we use the last page of the vcpu bo for the UVD message */
783 	uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
784 		RADEON_GPU_PAGE_SIZE;
785 
786 	uint32_t __iomem *msg = (void __iomem *)(rdev->uvd.cpu_addr + offs);
787 	uint64_t addr = rdev->uvd.gpu_addr + offs;
788 
789 	int r, i;
790 
791 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
792 	if (r)
793 		return r;
794 
795 	/* stitch together an UVD create msg */
796 	writel((__force u32)cpu_to_le32(0x00000de4), &msg[0]);
797 	writel(0x0, (void __iomem *)&msg[1]);
798 	writel((__force u32)cpu_to_le32(handle), &msg[2]);
799 	writel(0x0, &msg[3]);
800 	writel(0x0, &msg[4]);
801 	writel(0x0, &msg[5]);
802 	writel(0x0, &msg[6]);
803 	writel((__force u32)cpu_to_le32(0x00000780), &msg[7]);
804 	writel((__force u32)cpu_to_le32(0x00000440), &msg[8]);
805 	writel(0x0, &msg[9]);
806 	writel((__force u32)cpu_to_le32(0x01b37000), &msg[10]);
807 	for (i = 11; i < 1024; ++i)
808 		writel(0x0, &msg[i]);
809 
810 	r = radeon_uvd_send_msg(rdev, ring, addr, fence);
811 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
812 	return r;
813 }
814 
815 int radeon_uvd_get_destroy_msg(struct radeon_device *rdev, int ring,
816 			       uint32_t handle, struct radeon_fence **fence)
817 {
818 	/* we use the last page of the vcpu bo for the UVD message */
819 	uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
820 		RADEON_GPU_PAGE_SIZE;
821 
822 	uint32_t __iomem *msg = (void __iomem *)(rdev->uvd.cpu_addr + offs);
823 	uint64_t addr = rdev->uvd.gpu_addr + offs;
824 
825 	int r, i;
826 
827 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
828 	if (r)
829 		return r;
830 
831 	/* stitch together an UVD destroy msg */
832 	writel((__force u32)cpu_to_le32(0x00000de4), &msg[0]);
833 	writel((__force u32)cpu_to_le32(0x00000002), &msg[1]);
834 	writel((__force u32)cpu_to_le32(handle), &msg[2]);
835 	writel(0x0, &msg[3]);
836 	for (i = 4; i < 1024; ++i)
837 		writel(0x0, &msg[i]);
838 
839 	r = radeon_uvd_send_msg(rdev, ring, addr, fence);
840 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
841 	return r;
842 }
843 
844 /**
845  * radeon_uvd_count_handles - count number of open streams
846  *
847  * @rdev: radeon_device pointer
848  * @sd: number of SD streams
849  * @hd: number of HD streams
850  *
851  * Count the number of open SD/HD streams as a hint for power mangement
852  */
853 static void radeon_uvd_count_handles(struct radeon_device *rdev,
854 				     unsigned *sd, unsigned *hd)
855 {
856 	unsigned i;
857 
858 	*sd = 0;
859 	*hd = 0;
860 
861 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
862 		if (!atomic_read(&rdev->uvd.handles[i]))
863 			continue;
864 
865 		if (rdev->uvd.img_size[i] >= 720*576)
866 			++(*hd);
867 		else
868 			++(*sd);
869 	}
870 }
871 
872 static void radeon_uvd_idle_work_handler(struct work_struct *work)
873 {
874 	struct radeon_device *rdev =
875 		container_of(work, struct radeon_device, uvd.idle_work.work);
876 
877 	if (radeon_fence_count_emitted(rdev, R600_RING_TYPE_UVD_INDEX) == 0) {
878 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
879 			radeon_uvd_count_handles(rdev, &rdev->pm.dpm.sd,
880 						 &rdev->pm.dpm.hd);
881 			radeon_dpm_enable_uvd(rdev, false);
882 		} else {
883 			radeon_set_uvd_clocks(rdev, 0, 0);
884 		}
885 	} else {
886 		schedule_delayed_work(&rdev->uvd.idle_work,
887 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
888 	}
889 }
890 
891 void radeon_uvd_note_usage(struct radeon_device *rdev)
892 {
893 	bool streams_changed = false;
894 	bool set_clocks = !cancel_delayed_work_sync(&rdev->uvd.idle_work);
895 	set_clocks &= schedule_delayed_work(&rdev->uvd.idle_work,
896 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
897 
898 	if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
899 		unsigned hd = 0, sd = 0;
900 		radeon_uvd_count_handles(rdev, &sd, &hd);
901 		if ((rdev->pm.dpm.sd != sd) ||
902 		    (rdev->pm.dpm.hd != hd)) {
903 			rdev->pm.dpm.sd = sd;
904 			rdev->pm.dpm.hd = hd;
905 			/* disable this for now */
906 			/*streams_changed = true;*/
907 		}
908 	}
909 
910 	if (set_clocks || streams_changed) {
911 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
912 			radeon_dpm_enable_uvd(rdev, true);
913 		} else {
914 			radeon_set_uvd_clocks(rdev, 53300, 40000);
915 		}
916 	}
917 }
918 
919 static unsigned radeon_uvd_calc_upll_post_div(unsigned vco_freq,
920 					      unsigned target_freq,
921 					      unsigned pd_min,
922 					      unsigned pd_even)
923 {
924 	unsigned post_div = vco_freq / target_freq;
925 
926 	/* adjust to post divider minimum value */
927 	if (post_div < pd_min)
928 		post_div = pd_min;
929 
930 	/* we alway need a frequency less than or equal the target */
931 	if ((vco_freq / post_div) > target_freq)
932 		post_div += 1;
933 
934 	/* post dividers above a certain value must be even */
935 	if (post_div > pd_even && post_div % 2)
936 		post_div += 1;
937 
938 	return post_div;
939 }
940 
941 /**
942  * radeon_uvd_calc_upll_dividers - calc UPLL clock dividers
943  *
944  * @rdev: radeon_device pointer
945  * @vclk: wanted VCLK
946  * @dclk: wanted DCLK
947  * @vco_min: minimum VCO frequency
948  * @vco_max: maximum VCO frequency
949  * @fb_factor: factor to multiply vco freq with
950  * @fb_mask: limit and bitmask for feedback divider
951  * @pd_min: post divider minimum
952  * @pd_max: post divider maximum
953  * @pd_even: post divider must be even above this value
954  * @optimal_fb_div: resulting feedback divider
955  * @optimal_vclk_div: resulting vclk post divider
956  * @optimal_dclk_div: resulting dclk post divider
957  *
958  * Calculate dividers for UVDs UPLL (R6xx-SI, except APUs).
959  * Returns zero on success -EINVAL on error.
960  */
961 int radeon_uvd_calc_upll_dividers(struct radeon_device *rdev,
962 				  unsigned vclk, unsigned dclk,
963 				  unsigned vco_min, unsigned vco_max,
964 				  unsigned fb_factor, unsigned fb_mask,
965 				  unsigned pd_min, unsigned pd_max,
966 				  unsigned pd_even,
967 				  unsigned *optimal_fb_div,
968 				  unsigned *optimal_vclk_div,
969 				  unsigned *optimal_dclk_div)
970 {
971 	unsigned vco_freq, ref_freq = rdev->clock.spll.reference_freq;
972 
973 	/* start off with something large */
974 	unsigned optimal_score = ~0;
975 
976 	/* loop through vco from low to high */
977 	vco_min = max(max(vco_min, vclk), dclk);
978 	for (vco_freq = vco_min; vco_freq <= vco_max; vco_freq += 100) {
979 
980 		uint64_t fb_div = (uint64_t)vco_freq * fb_factor;
981 		unsigned vclk_div, dclk_div, score;
982 
983 		do_div(fb_div, ref_freq);
984 
985 		/* fb div out of range ? */
986 		if (fb_div > fb_mask)
987 			break; /* it can oly get worse */
988 
989 		fb_div &= fb_mask;
990 
991 		/* calc vclk divider with current vco freq */
992 		vclk_div = radeon_uvd_calc_upll_post_div(vco_freq, vclk,
993 							 pd_min, pd_even);
994 		if (vclk_div > pd_max)
995 			break; /* vco is too big, it has to stop */
996 
997 		/* calc dclk divider with current vco freq */
998 		dclk_div = radeon_uvd_calc_upll_post_div(vco_freq, dclk,
999 							 pd_min, pd_even);
1000 		if (dclk_div > pd_max)
1001 			break; /* vco is too big, it has to stop */
1002 
1003 		/* calc score with current vco freq */
1004 		score = vclk - (vco_freq / vclk_div) + dclk - (vco_freq / dclk_div);
1005 
1006 		/* determine if this vco setting is better than current optimal settings */
1007 		if (score < optimal_score) {
1008 			*optimal_fb_div = fb_div;
1009 			*optimal_vclk_div = vclk_div;
1010 			*optimal_dclk_div = dclk_div;
1011 			optimal_score = score;
1012 			if (optimal_score == 0)
1013 				break; /* it can't get better than this */
1014 		}
1015 	}
1016 
1017 	/* did we found a valid setup ? */
1018 	if (optimal_score == ~0)
1019 		return -EINVAL;
1020 
1021 	return 0;
1022 }
1023 
1024 int radeon_uvd_send_upll_ctlreq(struct radeon_device *rdev,
1025 				unsigned cg_upll_func_cntl)
1026 {
1027 	unsigned i;
1028 
1029 	/* make sure UPLL_CTLREQ is deasserted */
1030 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
1031 
1032 	mdelay(10);
1033 
1034 	/* assert UPLL_CTLREQ */
1035 	WREG32_P(cg_upll_func_cntl, UPLL_CTLREQ_MASK, ~UPLL_CTLREQ_MASK);
1036 
1037 	/* wait for CTLACK and CTLACK2 to get asserted */
1038 	for (i = 0; i < 100; ++i) {
1039 		uint32_t mask = UPLL_CTLACK_MASK | UPLL_CTLACK2_MASK;
1040 		if ((RREG32(cg_upll_func_cntl) & mask) == mask)
1041 			break;
1042 		mdelay(10);
1043 	}
1044 
1045 	/* deassert UPLL_CTLREQ */
1046 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
1047 
1048 	if (i == 100) {
1049 		DRM_ERROR("Timeout setting UVD clocks!\n");
1050 		return -ETIMEDOUT;
1051 	}
1052 
1053 	return 0;
1054 }
1055