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