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 = le32_to_cpu(hdr->ucode_version) & 0xff;
156 			version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
157 			version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
158 			DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
159 				 version_major, version_minor, family_id);
160 
161 			/*
162 			 * Limit the number of UVD handles depending on
163 			 * microcode major and minor versions.
164 			 */
165 			if ((version_major >= 0x01) && (version_minor >= 0x37))
166 				rdev->uvd.max_handles = RADEON_MAX_UVD_HANDLES;
167 		}
168 	}
169 
170 	/*
171 	 * In case there is only legacy firmware, or we encounter an error
172 	 * while loading the new firmware, we fall back to loading the legacy
173 	 * firmware now.
174 	 */
175 	if (!fw_name || r) {
176 		r = request_firmware(&rdev->uvd_fw, legacy_fw_name, rdev->dev);
177 		if (r) {
178 			dev_err(rdev->dev, "radeon_uvd: Can't load firmware \"%s\"\n",
179 				legacy_fw_name);
180 			return r;
181 		}
182 	}
183 
184 	bo_size = RADEON_GPU_PAGE_ALIGN(rdev->uvd_fw->size + 8) +
185 		  RADEON_UVD_STACK_SIZE + RADEON_UVD_HEAP_SIZE +
186 		  RADEON_UVD_SESSION_SIZE * rdev->uvd.max_handles;
187 	r = radeon_bo_create(rdev, bo_size, PAGE_SIZE, true,
188 			     RADEON_GEM_DOMAIN_VRAM, 0, NULL,
189 			     NULL, &rdev->uvd.vcpu_bo);
190 	if (r) {
191 		dev_err(rdev->dev, "(%d) failed to allocate UVD bo\n", r);
192 		return r;
193 	}
194 
195 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
196 	if (r) {
197 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
198 		dev_err(rdev->dev, "(%d) failed to reserve UVD bo\n", r);
199 		return r;
200 	}
201 
202 	r = radeon_bo_pin(rdev->uvd.vcpu_bo, RADEON_GEM_DOMAIN_VRAM,
203 			  &rdev->uvd.gpu_addr);
204 	if (r) {
205 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
206 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
207 		dev_err(rdev->dev, "(%d) UVD bo pin failed\n", r);
208 		return r;
209 	}
210 
211 	r = radeon_bo_kmap(rdev->uvd.vcpu_bo, &rdev->uvd.cpu_addr);
212 	if (r) {
213 		dev_err(rdev->dev, "(%d) UVD map failed\n", r);
214 		return r;
215 	}
216 
217 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
218 
219 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
220 		atomic_set(&rdev->uvd.handles[i], 0);
221 		rdev->uvd.filp[i] = NULL;
222 		rdev->uvd.img_size[i] = 0;
223 	}
224 
225 	return 0;
226 }
227 
228 void radeon_uvd_fini(struct radeon_device *rdev)
229 {
230 	int r;
231 
232 	if (rdev->uvd.vcpu_bo == NULL)
233 		return;
234 
235 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
236 	if (!r) {
237 		radeon_bo_kunmap(rdev->uvd.vcpu_bo);
238 		radeon_bo_unpin(rdev->uvd.vcpu_bo);
239 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
240 	}
241 
242 	radeon_bo_unref(&rdev->uvd.vcpu_bo);
243 
244 	radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_UVD_INDEX]);
245 
246 	release_firmware(rdev->uvd_fw);
247 }
248 
249 int radeon_uvd_suspend(struct radeon_device *rdev)
250 {
251 	int i, r;
252 
253 	if (rdev->uvd.vcpu_bo == NULL)
254 		return 0;
255 
256 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
257 		uint32_t handle = atomic_read(&rdev->uvd.handles[i]);
258 		if (handle != 0) {
259 			struct radeon_fence *fence;
260 
261 			radeon_uvd_note_usage(rdev);
262 
263 			r = radeon_uvd_get_destroy_msg(rdev,
264 				R600_RING_TYPE_UVD_INDEX, handle, &fence);
265 			if (r) {
266 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
267 				continue;
268 			}
269 
270 			radeon_fence_wait(fence, false);
271 			radeon_fence_unref(&fence);
272 
273 			rdev->uvd.filp[i] = NULL;
274 			atomic_set(&rdev->uvd.handles[i], 0);
275 		}
276 	}
277 
278 	return 0;
279 }
280 
281 int radeon_uvd_resume(struct radeon_device *rdev)
282 {
283 	unsigned size;
284 	void *ptr;
285 
286 	if (rdev->uvd.vcpu_bo == NULL)
287 		return -EINVAL;
288 
289 	memcpy(rdev->uvd.cpu_addr, rdev->uvd_fw->data, rdev->uvd_fw->size);
290 
291 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
292 	size -= rdev->uvd_fw->size;
293 
294 	ptr = rdev->uvd.cpu_addr;
295 	ptr += rdev->uvd_fw->size;
296 
297 	memset(ptr, 0, size);
298 
299 	return 0;
300 }
301 
302 void radeon_uvd_force_into_uvd_segment(struct radeon_bo *rbo,
303 				       uint32_t allowed_domains)
304 {
305 	int i;
306 
307 	for (i = 0; i < rbo->placement.num_placement; ++i) {
308 		rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
309 		rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
310 	}
311 
312 	/* If it must be in VRAM it must be in the first segment as well */
313 	if (allowed_domains == RADEON_GEM_DOMAIN_VRAM)
314 		return;
315 
316 	/* abort if we already have more than one placement */
317 	if (rbo->placement.num_placement > 1)
318 		return;
319 
320 	/* add another 256MB segment */
321 	rbo->placements[1] = rbo->placements[0];
322 	rbo->placements[1].fpfn += (256 * 1024 * 1024) >> PAGE_SHIFT;
323 	rbo->placements[1].lpfn += (256 * 1024 * 1024) >> PAGE_SHIFT;
324 	rbo->placement.num_placement++;
325 	rbo->placement.num_busy_placement++;
326 }
327 
328 void radeon_uvd_free_handles(struct radeon_device *rdev, struct drm_file *filp)
329 {
330 	int i, r;
331 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
332 		uint32_t handle = atomic_read(&rdev->uvd.handles[i]);
333 		if (handle != 0 && rdev->uvd.filp[i] == filp) {
334 			struct radeon_fence *fence;
335 
336 			radeon_uvd_note_usage(rdev);
337 
338 			r = radeon_uvd_get_destroy_msg(rdev,
339 				R600_RING_TYPE_UVD_INDEX, handle, &fence);
340 			if (r) {
341 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
342 				continue;
343 			}
344 
345 			radeon_fence_wait(fence, false);
346 			radeon_fence_unref(&fence);
347 
348 			rdev->uvd.filp[i] = NULL;
349 			atomic_set(&rdev->uvd.handles[i], 0);
350 		}
351 	}
352 }
353 
354 static int radeon_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
355 {
356 	unsigned stream_type = msg[4];
357 	unsigned width = msg[6];
358 	unsigned height = msg[7];
359 	unsigned dpb_size = msg[9];
360 	unsigned pitch = msg[28];
361 
362 	unsigned width_in_mb = width / 16;
363 	unsigned height_in_mb = ALIGN(height / 16, 2);
364 
365 	unsigned image_size, tmp, min_dpb_size;
366 
367 	image_size = width * height;
368 	image_size += image_size / 2;
369 	image_size = ALIGN(image_size, 1024);
370 
371 	switch (stream_type) {
372 	case 0: /* H264 */
373 
374 		/* reference picture buffer */
375 		min_dpb_size = image_size * 17;
376 
377 		/* macroblock context buffer */
378 		min_dpb_size += width_in_mb * height_in_mb * 17 * 192;
379 
380 		/* IT surface buffer */
381 		min_dpb_size += width_in_mb * height_in_mb * 32;
382 		break;
383 
384 	case 1: /* VC1 */
385 
386 		/* reference picture buffer */
387 		min_dpb_size = image_size * 3;
388 
389 		/* CONTEXT_BUFFER */
390 		min_dpb_size += width_in_mb * height_in_mb * 128;
391 
392 		/* IT surface buffer */
393 		min_dpb_size += width_in_mb * 64;
394 
395 		/* DB surface buffer */
396 		min_dpb_size += width_in_mb * 128;
397 
398 		/* BP */
399 		tmp = max(width_in_mb, height_in_mb);
400 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
401 		break;
402 
403 	case 3: /* MPEG2 */
404 
405 		/* reference picture buffer */
406 		min_dpb_size = image_size * 3;
407 		break;
408 
409 	case 4: /* MPEG4 */
410 
411 		/* reference picture buffer */
412 		min_dpb_size = image_size * 3;
413 
414 		/* CM */
415 		min_dpb_size += width_in_mb * height_in_mb * 64;
416 
417 		/* IT surface buffer */
418 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
419 		break;
420 
421 	default:
422 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
423 		return -EINVAL;
424 	}
425 
426 	if (width > pitch) {
427 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
428 		return -EINVAL;
429 	}
430 
431 	if (dpb_size < min_dpb_size) {
432 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
433 			  dpb_size, min_dpb_size);
434 		return -EINVAL;
435 	}
436 
437 	buf_sizes[0x1] = dpb_size;
438 	buf_sizes[0x2] = image_size;
439 	return 0;
440 }
441 
442 static int radeon_uvd_validate_codec(struct radeon_cs_parser *p,
443 				     unsigned stream_type)
444 {
445 	switch (stream_type) {
446 	case 0: /* H264 */
447 	case 1: /* VC1 */
448 		/* always supported */
449 		return 0;
450 
451 	case 3: /* MPEG2 */
452 	case 4: /* MPEG4 */
453 		/* only since UVD 3 */
454 		if (p->rdev->family >= CHIP_PALM)
455 			return 0;
456 
457 		/* fall through */
458 	default:
459 		DRM_ERROR("UVD codec not supported by hardware %d!\n",
460 			  stream_type);
461 		return -EINVAL;
462 	}
463 }
464 
465 static int radeon_uvd_cs_msg(struct radeon_cs_parser *p, struct radeon_bo *bo,
466 			     unsigned offset, unsigned buf_sizes[])
467 {
468 	int32_t *msg, msg_type, handle;
469 	unsigned img_size = 0;
470 	struct dma_fence *f;
471 	void *ptr;
472 
473 	int i, r;
474 
475 	if (offset & 0x3F) {
476 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
477 		return -EINVAL;
478 	}
479 
480 	f = dma_resv_get_excl(bo->tbo.base.resv);
481 	if (f) {
482 		r = radeon_fence_wait((struct radeon_fence *)f, false);
483 		if (r) {
484 			DRM_ERROR("Failed waiting for UVD message (%d)!\n", r);
485 			return r;
486 		}
487 	}
488 
489 	r = radeon_bo_kmap(bo, &ptr);
490 	if (r) {
491 		DRM_ERROR("Failed mapping the UVD message (%d)!\n", r);
492 		return r;
493 	}
494 
495 	msg = ptr + offset;
496 
497 	msg_type = msg[1];
498 	handle = msg[2];
499 
500 	if (handle == 0) {
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 
564 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
565 		return -EINVAL;
566 	}
567 
568 	BUG();
569 	return -EINVAL;
570 }
571 
572 static int radeon_uvd_cs_reloc(struct radeon_cs_parser *p,
573 			       int data0, int data1,
574 			       unsigned buf_sizes[], bool *has_msg_cmd)
575 {
576 	struct radeon_cs_chunk *relocs_chunk;
577 	struct radeon_bo_list *reloc;
578 	unsigned idx, cmd, offset;
579 	uint64_t start, end;
580 	int r;
581 
582 	relocs_chunk = p->chunk_relocs;
583 	offset = radeon_get_ib_value(p, data0);
584 	idx = radeon_get_ib_value(p, data1);
585 	if (idx >= relocs_chunk->length_dw) {
586 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
587 			  idx, relocs_chunk->length_dw);
588 		return -EINVAL;
589 	}
590 
591 	reloc = &p->relocs[(idx / 4)];
592 	start = reloc->gpu_offset;
593 	end = start + radeon_bo_size(reloc->robj);
594 	start += offset;
595 
596 	p->ib.ptr[data0] = start & 0xFFFFFFFF;
597 	p->ib.ptr[data1] = start >> 32;
598 
599 	cmd = radeon_get_ib_value(p, p->idx) >> 1;
600 
601 	if (cmd < 0x4) {
602 		if (end <= start) {
603 			DRM_ERROR("invalid reloc offset %X!\n", offset);
604 			return -EINVAL;
605 		}
606 		if ((end - start) < buf_sizes[cmd]) {
607 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
608 				  (unsigned)(end - start), buf_sizes[cmd]);
609 			return -EINVAL;
610 		}
611 
612 	} else if (cmd != 0x100) {
613 		DRM_ERROR("invalid UVD command %X!\n", cmd);
614 		return -EINVAL;
615 	}
616 
617 	if ((start >> 28) != ((end - 1) >> 28)) {
618 		DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
619 			  start, end);
620 		return -EINVAL;
621 	}
622 
623 	/* TODO: is this still necessary on NI+ ? */
624 	if ((cmd == 0 || cmd == 0x3) &&
625 	    (start >> 28) != (p->rdev->uvd.gpu_addr >> 28)) {
626 		DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
627 			  start, end);
628 		return -EINVAL;
629 	}
630 
631 	if (cmd == 0) {
632 		if (*has_msg_cmd) {
633 			DRM_ERROR("More than one message in a UVD-IB!\n");
634 			return -EINVAL;
635 		}
636 		*has_msg_cmd = true;
637 		r = radeon_uvd_cs_msg(p, reloc->robj, offset, buf_sizes);
638 		if (r)
639 			return r;
640 	} else if (!*has_msg_cmd) {
641 		DRM_ERROR("Message needed before other commands are send!\n");
642 		return -EINVAL;
643 	}
644 
645 	return 0;
646 }
647 
648 static int radeon_uvd_cs_reg(struct radeon_cs_parser *p,
649 			     struct radeon_cs_packet *pkt,
650 			     int *data0, int *data1,
651 			     unsigned buf_sizes[],
652 			     bool *has_msg_cmd)
653 {
654 	int i, r;
655 
656 	p->idx++;
657 	for (i = 0; i <= pkt->count; ++i) {
658 		switch (pkt->reg + i*4) {
659 		case UVD_GPCOM_VCPU_DATA0:
660 			*data0 = p->idx;
661 			break;
662 		case UVD_GPCOM_VCPU_DATA1:
663 			*data1 = p->idx;
664 			break;
665 		case UVD_GPCOM_VCPU_CMD:
666 			r = radeon_uvd_cs_reloc(p, *data0, *data1,
667 						buf_sizes, has_msg_cmd);
668 			if (r)
669 				return r;
670 			break;
671 		case UVD_ENGINE_CNTL:
672 		case UVD_NO_OP:
673 			break;
674 		default:
675 			DRM_ERROR("Invalid reg 0x%X!\n",
676 				  pkt->reg + i*4);
677 			return -EINVAL;
678 		}
679 		p->idx++;
680 	}
681 	return 0;
682 }
683 
684 int radeon_uvd_cs_parse(struct radeon_cs_parser *p)
685 {
686 	struct radeon_cs_packet pkt;
687 	int r, data0 = 0, data1 = 0;
688 
689 	/* does the IB has a msg command */
690 	bool has_msg_cmd = false;
691 
692 	/* minimum buffer sizes */
693 	unsigned buf_sizes[] = {
694 		[0x00000000]	=	2048,
695 		[0x00000001]	=	32 * 1024 * 1024,
696 		[0x00000002]	=	2048 * 1152 * 3,
697 		[0x00000003]	=	2048,
698 	};
699 
700 	if (p->chunk_ib->length_dw % 16) {
701 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
702 			  p->chunk_ib->length_dw);
703 		return -EINVAL;
704 	}
705 
706 	if (p->chunk_relocs == NULL) {
707 		DRM_ERROR("No relocation chunk !\n");
708 		return -EINVAL;
709 	}
710 
711 
712 	do {
713 		r = radeon_cs_packet_parse(p, &pkt, p->idx);
714 		if (r)
715 			return r;
716 		switch (pkt.type) {
717 		case RADEON_PACKET_TYPE0:
718 			r = radeon_uvd_cs_reg(p, &pkt, &data0, &data1,
719 					      buf_sizes, &has_msg_cmd);
720 			if (r)
721 				return r;
722 			break;
723 		case RADEON_PACKET_TYPE2:
724 			p->idx += pkt.count + 2;
725 			break;
726 		default:
727 			DRM_ERROR("Unknown packet type %d !\n", pkt.type);
728 			return -EINVAL;
729 		}
730 	} while (p->idx < p->chunk_ib->length_dw);
731 
732 	if (!has_msg_cmd) {
733 		DRM_ERROR("UVD-IBs need a msg command!\n");
734 		return -EINVAL;
735 	}
736 
737 	return 0;
738 }
739 
740 static int radeon_uvd_send_msg(struct radeon_device *rdev,
741 			       int ring, uint64_t addr,
742 			       struct radeon_fence **fence)
743 {
744 	struct radeon_ib ib;
745 	int i, r;
746 
747 	r = radeon_ib_get(rdev, ring, &ib, NULL, 64);
748 	if (r)
749 		return r;
750 
751 	ib.ptr[0] = PACKET0(UVD_GPCOM_VCPU_DATA0, 0);
752 	ib.ptr[1] = addr;
753 	ib.ptr[2] = PACKET0(UVD_GPCOM_VCPU_DATA1, 0);
754 	ib.ptr[3] = addr >> 32;
755 	ib.ptr[4] = PACKET0(UVD_GPCOM_VCPU_CMD, 0);
756 	ib.ptr[5] = 0;
757 	for (i = 6; i < 16; i += 2) {
758 		ib.ptr[i] = PACKET0(UVD_NO_OP, 0);
759 		ib.ptr[i+1] = 0;
760 	}
761 	ib.length_dw = 16;
762 
763 	r = radeon_ib_schedule(rdev, &ib, NULL, false);
764 
765 	if (fence)
766 		*fence = radeon_fence_ref(ib.fence);
767 
768 	radeon_ib_free(rdev, &ib);
769 	return r;
770 }
771 
772 /*
773  * multiple fence commands without any stream commands in between can
774  * crash the vcpu so just try to emmit a dummy create/destroy msg to
775  * avoid this
776  */
777 int radeon_uvd_get_create_msg(struct radeon_device *rdev, int ring,
778 			      uint32_t handle, struct radeon_fence **fence)
779 {
780 	/* we use the last page of the vcpu bo for the UVD message */
781 	uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
782 		RADEON_GPU_PAGE_SIZE;
783 
784 	uint32_t *msg = rdev->uvd.cpu_addr + offs;
785 	uint64_t addr = rdev->uvd.gpu_addr + offs;
786 
787 	int r, i;
788 
789 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
790 	if (r)
791 		return r;
792 
793 	/* stitch together an UVD create msg */
794 	msg[0] = cpu_to_le32(0x00000de4);
795 	msg[1] = cpu_to_le32(0x00000000);
796 	msg[2] = cpu_to_le32(handle);
797 	msg[3] = cpu_to_le32(0x00000000);
798 	msg[4] = cpu_to_le32(0x00000000);
799 	msg[5] = cpu_to_le32(0x00000000);
800 	msg[6] = cpu_to_le32(0x00000000);
801 	msg[7] = cpu_to_le32(0x00000780);
802 	msg[8] = cpu_to_le32(0x00000440);
803 	msg[9] = cpu_to_le32(0x00000000);
804 	msg[10] = cpu_to_le32(0x01b37000);
805 	for (i = 11; i < 1024; ++i)
806 		msg[i] = cpu_to_le32(0x0);
807 
808 	r = radeon_uvd_send_msg(rdev, ring, addr, fence);
809 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
810 	return r;
811 }
812 
813 int radeon_uvd_get_destroy_msg(struct radeon_device *rdev, int ring,
814 			       uint32_t handle, struct radeon_fence **fence)
815 {
816 	/* we use the last page of the vcpu bo for the UVD message */
817 	uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
818 		RADEON_GPU_PAGE_SIZE;
819 
820 	uint32_t *msg = rdev->uvd.cpu_addr + offs;
821 	uint64_t addr = rdev->uvd.gpu_addr + offs;
822 
823 	int r, i;
824 
825 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
826 	if (r)
827 		return r;
828 
829 	/* stitch together an UVD destroy msg */
830 	msg[0] = cpu_to_le32(0x00000de4);
831 	msg[1] = cpu_to_le32(0x00000002);
832 	msg[2] = cpu_to_le32(handle);
833 	msg[3] = cpu_to_le32(0x00000000);
834 	for (i = 4; i < 1024; ++i)
835 		msg[i] = cpu_to_le32(0x0);
836 
837 	r = radeon_uvd_send_msg(rdev, ring, addr, fence);
838 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
839 	return r;
840 }
841 
842 /**
843  * radeon_uvd_count_handles - count number of open streams
844  *
845  * @rdev: radeon_device pointer
846  * @sd: number of SD streams
847  * @hd: number of HD streams
848  *
849  * Count the number of open SD/HD streams as a hint for power mangement
850  */
851 static void radeon_uvd_count_handles(struct radeon_device *rdev,
852 				     unsigned *sd, unsigned *hd)
853 {
854 	unsigned i;
855 
856 	*sd = 0;
857 	*hd = 0;
858 
859 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
860 		if (!atomic_read(&rdev->uvd.handles[i]))
861 			continue;
862 
863 		if (rdev->uvd.img_size[i] >= 720*576)
864 			++(*hd);
865 		else
866 			++(*sd);
867 	}
868 }
869 
870 static void radeon_uvd_idle_work_handler(struct work_struct *work)
871 {
872 	struct radeon_device *rdev =
873 		container_of(work, struct radeon_device, uvd.idle_work.work);
874 
875 	if (radeon_fence_count_emitted(rdev, R600_RING_TYPE_UVD_INDEX) == 0) {
876 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
877 			radeon_uvd_count_handles(rdev, &rdev->pm.dpm.sd,
878 						 &rdev->pm.dpm.hd);
879 			radeon_dpm_enable_uvd(rdev, false);
880 		} else {
881 			radeon_set_uvd_clocks(rdev, 0, 0);
882 		}
883 	} else {
884 		schedule_delayed_work(&rdev->uvd.idle_work,
885 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
886 	}
887 }
888 
889 void radeon_uvd_note_usage(struct radeon_device *rdev)
890 {
891 	bool streams_changed = false;
892 	bool set_clocks = !cancel_delayed_work_sync(&rdev->uvd.idle_work);
893 	set_clocks &= schedule_delayed_work(&rdev->uvd.idle_work,
894 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
895 
896 	if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
897 		unsigned hd = 0, sd = 0;
898 		radeon_uvd_count_handles(rdev, &sd, &hd);
899 		if ((rdev->pm.dpm.sd != sd) ||
900 		    (rdev->pm.dpm.hd != hd)) {
901 			rdev->pm.dpm.sd = sd;
902 			rdev->pm.dpm.hd = hd;
903 			/* disable this for now */
904 			/*streams_changed = true;*/
905 		}
906 	}
907 
908 	if (set_clocks || streams_changed) {
909 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
910 			radeon_dpm_enable_uvd(rdev, true);
911 		} else {
912 			radeon_set_uvd_clocks(rdev, 53300, 40000);
913 		}
914 	}
915 }
916 
917 static unsigned radeon_uvd_calc_upll_post_div(unsigned vco_freq,
918 					      unsigned target_freq,
919 					      unsigned pd_min,
920 					      unsigned pd_even)
921 {
922 	unsigned post_div = vco_freq / target_freq;
923 
924 	/* adjust to post divider minimum value */
925 	if (post_div < pd_min)
926 		post_div = pd_min;
927 
928 	/* we alway need a frequency less than or equal the target */
929 	if ((vco_freq / post_div) > target_freq)
930 		post_div += 1;
931 
932 	/* post dividers above a certain value must be even */
933 	if (post_div > pd_even && post_div % 2)
934 		post_div += 1;
935 
936 	return post_div;
937 }
938 
939 /**
940  * radeon_uvd_calc_upll_dividers - calc UPLL clock dividers
941  *
942  * @rdev: radeon_device pointer
943  * @vclk: wanted VCLK
944  * @dclk: wanted DCLK
945  * @vco_min: minimum VCO frequency
946  * @vco_max: maximum VCO frequency
947  * @fb_factor: factor to multiply vco freq with
948  * @fb_mask: limit and bitmask for feedback divider
949  * @pd_min: post divider minimum
950  * @pd_max: post divider maximum
951  * @pd_even: post divider must be even above this value
952  * @optimal_fb_div: resulting feedback divider
953  * @optimal_vclk_div: resulting vclk post divider
954  * @optimal_dclk_div: resulting dclk post divider
955  *
956  * Calculate dividers for UVDs UPLL (R6xx-SI, except APUs).
957  * Returns zero on success -EINVAL on error.
958  */
959 int radeon_uvd_calc_upll_dividers(struct radeon_device *rdev,
960 				  unsigned vclk, unsigned dclk,
961 				  unsigned vco_min, unsigned vco_max,
962 				  unsigned fb_factor, unsigned fb_mask,
963 				  unsigned pd_min, unsigned pd_max,
964 				  unsigned pd_even,
965 				  unsigned *optimal_fb_div,
966 				  unsigned *optimal_vclk_div,
967 				  unsigned *optimal_dclk_div)
968 {
969 	unsigned vco_freq, ref_freq = rdev->clock.spll.reference_freq;
970 
971 	/* start off with something large */
972 	unsigned optimal_score = ~0;
973 
974 	/* loop through vco from low to high */
975 	vco_min = max(max(vco_min, vclk), dclk);
976 	for (vco_freq = vco_min; vco_freq <= vco_max; vco_freq += 100) {
977 
978 		uint64_t fb_div = (uint64_t)vco_freq * fb_factor;
979 		unsigned vclk_div, dclk_div, score;
980 
981 		do_div(fb_div, ref_freq);
982 
983 		/* fb div out of range ? */
984 		if (fb_div > fb_mask)
985 			break; /* it can oly get worse */
986 
987 		fb_div &= fb_mask;
988 
989 		/* calc vclk divider with current vco freq */
990 		vclk_div = radeon_uvd_calc_upll_post_div(vco_freq, vclk,
991 							 pd_min, pd_even);
992 		if (vclk_div > pd_max)
993 			break; /* vco is too big, it has to stop */
994 
995 		/* calc dclk divider with current vco freq */
996 		dclk_div = radeon_uvd_calc_upll_post_div(vco_freq, dclk,
997 							 pd_min, pd_even);
998 		if (dclk_div > pd_max)
999 			break; /* vco is too big, it has to stop */
1000 
1001 		/* calc score with current vco freq */
1002 		score = vclk - (vco_freq / vclk_div) + dclk - (vco_freq / dclk_div);
1003 
1004 		/* determine if this vco setting is better than current optimal settings */
1005 		if (score < optimal_score) {
1006 			*optimal_fb_div = fb_div;
1007 			*optimal_vclk_div = vclk_div;
1008 			*optimal_dclk_div = dclk_div;
1009 			optimal_score = score;
1010 			if (optimal_score == 0)
1011 				break; /* it can't get better than this */
1012 		}
1013 	}
1014 
1015 	/* did we found a valid setup ? */
1016 	if (optimal_score == ~0)
1017 		return -EINVAL;
1018 
1019 	return 0;
1020 }
1021 
1022 int radeon_uvd_send_upll_ctlreq(struct radeon_device *rdev,
1023 				unsigned cg_upll_func_cntl)
1024 {
1025 	unsigned i;
1026 
1027 	/* make sure UPLL_CTLREQ is deasserted */
1028 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
1029 
1030 	mdelay(10);
1031 
1032 	/* assert UPLL_CTLREQ */
1033 	WREG32_P(cg_upll_func_cntl, UPLL_CTLREQ_MASK, ~UPLL_CTLREQ_MASK);
1034 
1035 	/* wait for CTLACK and CTLACK2 to get asserted */
1036 	for (i = 0; i < 100; ++i) {
1037 		uint32_t mask = UPLL_CTLACK_MASK | UPLL_CTLACK2_MASK;
1038 		if ((RREG32(cg_upll_func_cntl) & mask) == mask)
1039 			break;
1040 		mdelay(10);
1041 	}
1042 
1043 	/* deassert UPLL_CTLREQ */
1044 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
1045 
1046 	if (i == 100) {
1047 		DRM_ERROR("Timeout setting UVD clocks!\n");
1048 		return -ETIMEDOUT;
1049 	}
1050 
1051 	return 0;
1052 }
1053