• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2011 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Christian König <deathsimple@vodafone.de>
29  */
30 
31 #include <linux/firmware.h>
32 #include <linux/module.h>
33 #include <drm/drmP.h>
34 #include <drm/drm.h>
35 
36 #include "radeon.h"
37 #include "r600d.h"
38 
39 /* 1 second timeout */
40 #define UVD_IDLE_TIMEOUT_MS	1000
41 
42 /* Firmware Names */
43 #define FIRMWARE_R600		"radeon/R600_uvd.bin"
44 #define FIRMWARE_RS780		"radeon/RS780_uvd.bin"
45 #define FIRMWARE_RV770		"radeon/RV770_uvd.bin"
46 #define FIRMWARE_RV710		"radeon/RV710_uvd.bin"
47 #define FIRMWARE_CYPRESS	"radeon/CYPRESS_uvd.bin"
48 #define FIRMWARE_SUMO		"radeon/SUMO_uvd.bin"
49 #define FIRMWARE_TAHITI		"radeon/TAHITI_uvd.bin"
50 #define FIRMWARE_BONAIRE	"radeon/BONAIRE_uvd.bin"
51 
52 MODULE_FIRMWARE(FIRMWARE_R600);
53 MODULE_FIRMWARE(FIRMWARE_RS780);
54 MODULE_FIRMWARE(FIRMWARE_RV770);
55 MODULE_FIRMWARE(FIRMWARE_RV710);
56 MODULE_FIRMWARE(FIRMWARE_CYPRESS);
57 MODULE_FIRMWARE(FIRMWARE_SUMO);
58 MODULE_FIRMWARE(FIRMWARE_TAHITI);
59 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
60 
61 static void radeon_uvd_idle_work_handler(struct work_struct *work);
62 
radeon_uvd_init(struct radeon_device * rdev)63 int radeon_uvd_init(struct radeon_device *rdev)
64 {
65 	unsigned long bo_size;
66 	const char *fw_name;
67 	int i, r;
68 
69 	INIT_DELAYED_WORK(&rdev->uvd.idle_work, radeon_uvd_idle_work_handler);
70 
71 	switch (rdev->family) {
72 	case CHIP_RV610:
73 	case CHIP_RV630:
74 	case CHIP_RV670:
75 	case CHIP_RV620:
76 	case CHIP_RV635:
77 		fw_name = FIRMWARE_R600;
78 		break;
79 
80 	case CHIP_RS780:
81 	case CHIP_RS880:
82 		fw_name = FIRMWARE_RS780;
83 		break;
84 
85 	case CHIP_RV770:
86 		fw_name = FIRMWARE_RV770;
87 		break;
88 
89 	case CHIP_RV710:
90 	case CHIP_RV730:
91 	case CHIP_RV740:
92 		fw_name = FIRMWARE_RV710;
93 		break;
94 
95 	case CHIP_CYPRESS:
96 	case CHIP_HEMLOCK:
97 	case CHIP_JUNIPER:
98 	case CHIP_REDWOOD:
99 	case CHIP_CEDAR:
100 		fw_name = FIRMWARE_CYPRESS;
101 		break;
102 
103 	case CHIP_SUMO:
104 	case CHIP_SUMO2:
105 	case CHIP_PALM:
106 	case CHIP_CAYMAN:
107 	case CHIP_BARTS:
108 	case CHIP_TURKS:
109 	case CHIP_CAICOS:
110 		fw_name = FIRMWARE_SUMO;
111 		break;
112 
113 	case CHIP_TAHITI:
114 	case CHIP_VERDE:
115 	case CHIP_PITCAIRN:
116 	case CHIP_ARUBA:
117 	case CHIP_OLAND:
118 		fw_name = FIRMWARE_TAHITI;
119 		break;
120 
121 	case CHIP_BONAIRE:
122 	case CHIP_KABINI:
123 	case CHIP_KAVERI:
124 	case CHIP_HAWAII:
125 	case CHIP_MULLINS:
126 		fw_name = FIRMWARE_BONAIRE;
127 		break;
128 
129 	default:
130 		return -EINVAL;
131 	}
132 
133 	r = request_firmware(&rdev->uvd_fw, fw_name, rdev->dev);
134 	if (r) {
135 		dev_err(rdev->dev, "radeon_uvd: Can't load firmware \"%s\"\n",
136 			fw_name);
137 		return r;
138 	}
139 
140 	bo_size = RADEON_GPU_PAGE_ALIGN(rdev->uvd_fw->size + 8) +
141 		  RADEON_UVD_STACK_SIZE + RADEON_UVD_HEAP_SIZE +
142 		  RADEON_GPU_PAGE_SIZE;
143 	r = radeon_bo_create(rdev, bo_size, PAGE_SIZE, true,
144 			     RADEON_GEM_DOMAIN_VRAM, 0, NULL,
145 			     NULL, &rdev->uvd.vcpu_bo);
146 	if (r) {
147 		dev_err(rdev->dev, "(%d) failed to allocate UVD bo\n", r);
148 		return r;
149 	}
150 
151 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
152 	if (r) {
153 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
154 		dev_err(rdev->dev, "(%d) failed to reserve UVD bo\n", r);
155 		return r;
156 	}
157 
158 	r = radeon_bo_pin(rdev->uvd.vcpu_bo, RADEON_GEM_DOMAIN_VRAM,
159 			  &rdev->uvd.gpu_addr);
160 	if (r) {
161 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
162 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
163 		dev_err(rdev->dev, "(%d) UVD bo pin failed\n", r);
164 		return r;
165 	}
166 
167 	r = radeon_bo_kmap(rdev->uvd.vcpu_bo, &rdev->uvd.cpu_addr);
168 	if (r) {
169 		dev_err(rdev->dev, "(%d) UVD map failed\n", r);
170 		return r;
171 	}
172 
173 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
174 
175 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
176 		atomic_set(&rdev->uvd.handles[i], 0);
177 		rdev->uvd.filp[i] = NULL;
178 		rdev->uvd.img_size[i] = 0;
179 	}
180 
181 	return 0;
182 }
183 
radeon_uvd_fini(struct radeon_device * rdev)184 void radeon_uvd_fini(struct radeon_device *rdev)
185 {
186 	int r;
187 
188 	if (rdev->uvd.vcpu_bo == NULL)
189 		return;
190 
191 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
192 	if (!r) {
193 		radeon_bo_kunmap(rdev->uvd.vcpu_bo);
194 		radeon_bo_unpin(rdev->uvd.vcpu_bo);
195 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
196 	}
197 
198 	radeon_bo_unref(&rdev->uvd.vcpu_bo);
199 
200 	radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_UVD_INDEX]);
201 
202 	release_firmware(rdev->uvd_fw);
203 }
204 
radeon_uvd_suspend(struct radeon_device * rdev)205 int radeon_uvd_suspend(struct radeon_device *rdev)
206 {
207 	unsigned size;
208 	void *ptr;
209 	int i;
210 
211 	if (rdev->uvd.vcpu_bo == NULL)
212 		return 0;
213 
214 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i)
215 		if (atomic_read(&rdev->uvd.handles[i]))
216 			break;
217 
218 	if (i == RADEON_MAX_UVD_HANDLES)
219 		return 0;
220 
221 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
222 	size -= rdev->uvd_fw->size;
223 
224 	ptr = rdev->uvd.cpu_addr;
225 	ptr += rdev->uvd_fw->size;
226 
227 	rdev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
228 	memcpy(rdev->uvd.saved_bo, ptr, size);
229 
230 	return 0;
231 }
232 
radeon_uvd_resume(struct radeon_device * rdev)233 int radeon_uvd_resume(struct radeon_device *rdev)
234 {
235 	unsigned size;
236 	void *ptr;
237 
238 	if (rdev->uvd.vcpu_bo == NULL)
239 		return -EINVAL;
240 
241 	memcpy(rdev->uvd.cpu_addr, rdev->uvd_fw->data, rdev->uvd_fw->size);
242 
243 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
244 	size -= rdev->uvd_fw->size;
245 
246 	ptr = rdev->uvd.cpu_addr;
247 	ptr += rdev->uvd_fw->size;
248 
249 	if (rdev->uvd.saved_bo != NULL) {
250 		memcpy(ptr, rdev->uvd.saved_bo, size);
251 		kfree(rdev->uvd.saved_bo);
252 		rdev->uvd.saved_bo = NULL;
253 	} else
254 		memset(ptr, 0, size);
255 
256 	return 0;
257 }
258 
radeon_uvd_force_into_uvd_segment(struct radeon_bo * rbo,uint32_t allowed_domains)259 void radeon_uvd_force_into_uvd_segment(struct radeon_bo *rbo,
260 				       uint32_t allowed_domains)
261 {
262 	int i;
263 
264 	for (i = 0; i < rbo->placement.num_placement; ++i) {
265 		rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
266 		rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
267 	}
268 
269 	/* If it must be in VRAM it must be in the first segment as well */
270 	if (allowed_domains == RADEON_GEM_DOMAIN_VRAM)
271 		return;
272 
273 	/* abort if we already have more than one placement */
274 	if (rbo->placement.num_placement > 1)
275 		return;
276 
277 	/* add another 256MB segment */
278 	rbo->placements[1] = rbo->placements[0];
279 	rbo->placements[1].fpfn += (256 * 1024 * 1024) >> PAGE_SHIFT;
280 	rbo->placements[1].lpfn += (256 * 1024 * 1024) >> PAGE_SHIFT;
281 	rbo->placement.num_placement++;
282 	rbo->placement.num_busy_placement++;
283 }
284 
radeon_uvd_free_handles(struct radeon_device * rdev,struct drm_file * filp)285 void radeon_uvd_free_handles(struct radeon_device *rdev, struct drm_file *filp)
286 {
287 	int i, r;
288 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
289 		uint32_t handle = atomic_read(&rdev->uvd.handles[i]);
290 		if (handle != 0 && rdev->uvd.filp[i] == filp) {
291 			struct radeon_fence *fence;
292 
293 			radeon_uvd_note_usage(rdev);
294 
295 			r = radeon_uvd_get_destroy_msg(rdev,
296 				R600_RING_TYPE_UVD_INDEX, handle, &fence);
297 			if (r) {
298 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
299 				continue;
300 			}
301 
302 			radeon_fence_wait(fence, false);
303 			radeon_fence_unref(&fence);
304 
305 			rdev->uvd.filp[i] = NULL;
306 			atomic_set(&rdev->uvd.handles[i], 0);
307 		}
308 	}
309 }
310 
radeon_uvd_cs_msg_decode(uint32_t * msg,unsigned buf_sizes[])311 static int radeon_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
312 {
313 	unsigned stream_type = msg[4];
314 	unsigned width = msg[6];
315 	unsigned height = msg[7];
316 	unsigned dpb_size = msg[9];
317 	unsigned pitch = msg[28];
318 
319 	unsigned width_in_mb = width / 16;
320 	unsigned height_in_mb = ALIGN(height / 16, 2);
321 
322 	unsigned image_size, tmp, min_dpb_size;
323 
324 	image_size = width * height;
325 	image_size += image_size / 2;
326 	image_size = ALIGN(image_size, 1024);
327 
328 	switch (stream_type) {
329 	case 0: /* H264 */
330 
331 		/* reference picture buffer */
332 		min_dpb_size = image_size * 17;
333 
334 		/* macroblock context buffer */
335 		min_dpb_size += width_in_mb * height_in_mb * 17 * 192;
336 
337 		/* IT surface buffer */
338 		min_dpb_size += width_in_mb * height_in_mb * 32;
339 		break;
340 
341 	case 1: /* VC1 */
342 
343 		/* reference picture buffer */
344 		min_dpb_size = image_size * 3;
345 
346 		/* CONTEXT_BUFFER */
347 		min_dpb_size += width_in_mb * height_in_mb * 128;
348 
349 		/* IT surface buffer */
350 		min_dpb_size += width_in_mb * 64;
351 
352 		/* DB surface buffer */
353 		min_dpb_size += width_in_mb * 128;
354 
355 		/* BP */
356 		tmp = max(width_in_mb, height_in_mb);
357 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
358 		break;
359 
360 	case 3: /* MPEG2 */
361 
362 		/* reference picture buffer */
363 		min_dpb_size = image_size * 3;
364 		break;
365 
366 	case 4: /* MPEG4 */
367 
368 		/* reference picture buffer */
369 		min_dpb_size = image_size * 3;
370 
371 		/* CM */
372 		min_dpb_size += width_in_mb * height_in_mb * 64;
373 
374 		/* IT surface buffer */
375 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
376 		break;
377 
378 	default:
379 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
380 		return -EINVAL;
381 	}
382 
383 	if (width > pitch) {
384 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
385 		return -EINVAL;
386 	}
387 
388 	if (dpb_size < min_dpb_size) {
389 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
390 			  dpb_size, min_dpb_size);
391 		return -EINVAL;
392 	}
393 
394 	buf_sizes[0x1] = dpb_size;
395 	buf_sizes[0x2] = image_size;
396 	return 0;
397 }
398 
radeon_uvd_validate_codec(struct radeon_cs_parser * p,unsigned stream_type)399 static int radeon_uvd_validate_codec(struct radeon_cs_parser *p,
400 				     unsigned stream_type)
401 {
402 	switch (stream_type) {
403 	case 0: /* H264 */
404 	case 1: /* VC1 */
405 		/* always supported */
406 		return 0;
407 
408 	case 3: /* MPEG2 */
409 	case 4: /* MPEG4 */
410 		/* only since UVD 3 */
411 		if (p->rdev->family >= CHIP_PALM)
412 			return 0;
413 
414 		/* fall through */
415 	default:
416 		DRM_ERROR("UVD codec not supported by hardware %d!\n",
417 			  stream_type);
418 		return -EINVAL;
419 	}
420 }
421 
radeon_uvd_cs_msg(struct radeon_cs_parser * p,struct radeon_bo * bo,unsigned offset,unsigned buf_sizes[])422 static int radeon_uvd_cs_msg(struct radeon_cs_parser *p, struct radeon_bo *bo,
423 			     unsigned offset, unsigned buf_sizes[])
424 {
425 	int32_t *msg, msg_type, handle;
426 	unsigned img_size = 0;
427 	struct fence *f;
428 	void *ptr;
429 
430 	int i, r;
431 
432 	if (offset & 0x3F) {
433 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
434 		return -EINVAL;
435 	}
436 
437 	f = reservation_object_get_excl(bo->tbo.resv);
438 	if (f) {
439 		r = radeon_fence_wait((struct radeon_fence *)f, false);
440 		if (r) {
441 			DRM_ERROR("Failed waiting for UVD message (%d)!\n", r);
442 			return r;
443 		}
444 	}
445 
446 	r = radeon_bo_kmap(bo, &ptr);
447 	if (r) {
448 		DRM_ERROR("Failed mapping the UVD message (%d)!\n", r);
449 		return r;
450 	}
451 
452 	msg = ptr + offset;
453 
454 	msg_type = msg[1];
455 	handle = msg[2];
456 
457 	if (handle == 0) {
458 		DRM_ERROR("Invalid UVD handle!\n");
459 		return -EINVAL;
460 	}
461 
462 	switch (msg_type) {
463 	case 0:
464 		/* it's a create msg, calc image size (width * height) */
465 		img_size = msg[7] * msg[8];
466 
467 		r = radeon_uvd_validate_codec(p, msg[4]);
468 		radeon_bo_kunmap(bo);
469 		if (r)
470 			return r;
471 
472 		/* try to alloc a new handle */
473 		for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
474 			if (atomic_read(&p->rdev->uvd.handles[i]) == handle) {
475 				DRM_ERROR("Handle 0x%x already in use!\n", handle);
476 				return -EINVAL;
477 			}
478 
479 			if (!atomic_cmpxchg(&p->rdev->uvd.handles[i], 0, handle)) {
480 				p->rdev->uvd.filp[i] = p->filp;
481 				p->rdev->uvd.img_size[i] = img_size;
482 				return 0;
483 			}
484 		}
485 
486 		DRM_ERROR("No more free UVD handles!\n");
487 		return -EINVAL;
488 
489 	case 1:
490 		/* it's a decode msg, validate codec and calc buffer sizes */
491 		r = radeon_uvd_validate_codec(p, msg[4]);
492 		if (!r)
493 			r = radeon_uvd_cs_msg_decode(msg, buf_sizes);
494 		radeon_bo_kunmap(bo);
495 		if (r)
496 			return r;
497 
498 		/* validate the handle */
499 		for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
500 			if (atomic_read(&p->rdev->uvd.handles[i]) == handle) {
501 				if (p->rdev->uvd.filp[i] != p->filp) {
502 					DRM_ERROR("UVD handle collision detected!\n");
503 					return -EINVAL;
504 				}
505 				return 0;
506 			}
507 		}
508 
509 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
510 		return -ENOENT;
511 
512 	case 2:
513 		/* it's a destroy msg, free the handle */
514 		for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i)
515 			atomic_cmpxchg(&p->rdev->uvd.handles[i], handle, 0);
516 		radeon_bo_kunmap(bo);
517 		return 0;
518 
519 	default:
520 
521 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
522 		return -EINVAL;
523 	}
524 
525 	BUG();
526 	return -EINVAL;
527 }
528 
radeon_uvd_cs_reloc(struct radeon_cs_parser * p,int data0,int data1,unsigned buf_sizes[],bool * has_msg_cmd)529 static int radeon_uvd_cs_reloc(struct radeon_cs_parser *p,
530 			       int data0, int data1,
531 			       unsigned buf_sizes[], bool *has_msg_cmd)
532 {
533 	struct radeon_cs_chunk *relocs_chunk;
534 	struct radeon_cs_reloc *reloc;
535 	unsigned idx, cmd, offset;
536 	uint64_t start, end;
537 	int r;
538 
539 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
540 	offset = radeon_get_ib_value(p, data0);
541 	idx = radeon_get_ib_value(p, data1);
542 	if (idx >= relocs_chunk->length_dw) {
543 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
544 			  idx, relocs_chunk->length_dw);
545 		return -EINVAL;
546 	}
547 
548 	reloc = p->relocs_ptr[(idx / 4)];
549 	start = reloc->gpu_offset;
550 	end = start + radeon_bo_size(reloc->robj);
551 	start += offset;
552 
553 	p->ib.ptr[data0] = start & 0xFFFFFFFF;
554 	p->ib.ptr[data1] = start >> 32;
555 
556 	cmd = radeon_get_ib_value(p, p->idx) >> 1;
557 
558 	if (cmd < 0x4) {
559 		if (end <= start) {
560 			DRM_ERROR("invalid reloc offset %X!\n", offset);
561 			return -EINVAL;
562 		}
563 		if ((end - start) < buf_sizes[cmd]) {
564 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
565 				  (unsigned)(end - start), buf_sizes[cmd]);
566 			return -EINVAL;
567 		}
568 
569 	} else if (cmd != 0x100) {
570 		DRM_ERROR("invalid UVD command %X!\n", cmd);
571 		return -EINVAL;
572 	}
573 
574 	if ((start >> 28) != ((end - 1) >> 28)) {
575 		DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
576 			  start, end);
577 		return -EINVAL;
578 	}
579 
580 	/* TODO: is this still necessary on NI+ ? */
581 	if ((cmd == 0 || cmd == 0x3) &&
582 	    (start >> 28) != (p->rdev->uvd.gpu_addr >> 28)) {
583 		DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
584 			  start, end);
585 		return -EINVAL;
586 	}
587 
588 	if (cmd == 0) {
589 		if (*has_msg_cmd) {
590 			DRM_ERROR("More than one message in a UVD-IB!\n");
591 			return -EINVAL;
592 		}
593 		*has_msg_cmd = true;
594 		r = radeon_uvd_cs_msg(p, reloc->robj, offset, buf_sizes);
595 		if (r)
596 			return r;
597 	} else if (!*has_msg_cmd) {
598 		DRM_ERROR("Message needed before other commands are send!\n");
599 		return -EINVAL;
600 	}
601 
602 	return 0;
603 }
604 
radeon_uvd_cs_reg(struct radeon_cs_parser * p,struct radeon_cs_packet * pkt,int * data0,int * data1,unsigned buf_sizes[],bool * has_msg_cmd)605 static int radeon_uvd_cs_reg(struct radeon_cs_parser *p,
606 			     struct radeon_cs_packet *pkt,
607 			     int *data0, int *data1,
608 			     unsigned buf_sizes[],
609 			     bool *has_msg_cmd)
610 {
611 	int i, r;
612 
613 	p->idx++;
614 	for (i = 0; i <= pkt->count; ++i) {
615 		switch (pkt->reg + i*4) {
616 		case UVD_GPCOM_VCPU_DATA0:
617 			*data0 = p->idx;
618 			break;
619 		case UVD_GPCOM_VCPU_DATA1:
620 			*data1 = p->idx;
621 			break;
622 		case UVD_GPCOM_VCPU_CMD:
623 			r = radeon_uvd_cs_reloc(p, *data0, *data1,
624 						buf_sizes, has_msg_cmd);
625 			if (r)
626 				return r;
627 			break;
628 		case UVD_ENGINE_CNTL:
629 			break;
630 		default:
631 			DRM_ERROR("Invalid reg 0x%X!\n",
632 				  pkt->reg + i*4);
633 			return -EINVAL;
634 		}
635 		p->idx++;
636 	}
637 	return 0;
638 }
639 
radeon_uvd_cs_parse(struct radeon_cs_parser * p)640 int radeon_uvd_cs_parse(struct radeon_cs_parser *p)
641 {
642 	struct radeon_cs_packet pkt;
643 	int r, data0 = 0, data1 = 0;
644 
645 	/* does the IB has a msg command */
646 	bool has_msg_cmd = false;
647 
648 	/* minimum buffer sizes */
649 	unsigned buf_sizes[] = {
650 		[0x00000000]	=	2048,
651 		[0x00000001]	=	32 * 1024 * 1024,
652 		[0x00000002]	=	2048 * 1152 * 3,
653 		[0x00000003]	=	2048,
654 	};
655 
656 	if (p->chunks[p->chunk_ib_idx].length_dw % 16) {
657 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
658 			  p->chunks[p->chunk_ib_idx].length_dw);
659 		return -EINVAL;
660 	}
661 
662 	if (p->chunk_relocs_idx == -1) {
663 		DRM_ERROR("No relocation chunk !\n");
664 		return -EINVAL;
665 	}
666 
667 
668 	do {
669 		r = radeon_cs_packet_parse(p, &pkt, p->idx);
670 		if (r)
671 			return r;
672 		switch (pkt.type) {
673 		case RADEON_PACKET_TYPE0:
674 			r = radeon_uvd_cs_reg(p, &pkt, &data0, &data1,
675 					      buf_sizes, &has_msg_cmd);
676 			if (r)
677 				return r;
678 			break;
679 		case RADEON_PACKET_TYPE2:
680 			p->idx += pkt.count + 2;
681 			break;
682 		default:
683 			DRM_ERROR("Unknown packet type %d !\n", pkt.type);
684 			return -EINVAL;
685 		}
686 	} while (p->idx < p->chunks[p->chunk_ib_idx].length_dw);
687 
688 	if (!has_msg_cmd) {
689 		DRM_ERROR("UVD-IBs need a msg command!\n");
690 		return -EINVAL;
691 	}
692 
693 	return 0;
694 }
695 
radeon_uvd_send_msg(struct radeon_device * rdev,int ring,uint64_t addr,struct radeon_fence ** fence)696 static int radeon_uvd_send_msg(struct radeon_device *rdev,
697 			       int ring, uint64_t addr,
698 			       struct radeon_fence **fence)
699 {
700 	struct radeon_ib ib;
701 	int i, r;
702 
703 	r = radeon_ib_get(rdev, ring, &ib, NULL, 64);
704 	if (r)
705 		return r;
706 
707 	ib.ptr[0] = PACKET0(UVD_GPCOM_VCPU_DATA0, 0);
708 	ib.ptr[1] = addr;
709 	ib.ptr[2] = PACKET0(UVD_GPCOM_VCPU_DATA1, 0);
710 	ib.ptr[3] = addr >> 32;
711 	ib.ptr[4] = PACKET0(UVD_GPCOM_VCPU_CMD, 0);
712 	ib.ptr[5] = 0;
713 	for (i = 6; i < 16; ++i)
714 		ib.ptr[i] = PACKET2(0);
715 	ib.length_dw = 16;
716 
717 	r = radeon_ib_schedule(rdev, &ib, NULL, false);
718 
719 	if (fence)
720 		*fence = radeon_fence_ref(ib.fence);
721 
722 	radeon_ib_free(rdev, &ib);
723 	return r;
724 }
725 
726 /* multiple fence commands without any stream commands in between can
727    crash the vcpu so just try to emmit a dummy create/destroy msg to
728    avoid this */
radeon_uvd_get_create_msg(struct radeon_device * rdev,int ring,uint32_t handle,struct radeon_fence ** fence)729 int radeon_uvd_get_create_msg(struct radeon_device *rdev, int ring,
730 			      uint32_t handle, struct radeon_fence **fence)
731 {
732 	/* we use the last page of the vcpu bo for the UVD message */
733 	uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
734 		RADEON_GPU_PAGE_SIZE;
735 
736 	uint32_t *msg = rdev->uvd.cpu_addr + offs;
737 	uint64_t addr = rdev->uvd.gpu_addr + offs;
738 
739 	int r, i;
740 
741 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
742 	if (r)
743 		return r;
744 
745 	/* stitch together an UVD create msg */
746 	msg[0] = cpu_to_le32(0x00000de4);
747 	msg[1] = cpu_to_le32(0x00000000);
748 	msg[2] = cpu_to_le32(handle);
749 	msg[3] = cpu_to_le32(0x00000000);
750 	msg[4] = cpu_to_le32(0x00000000);
751 	msg[5] = cpu_to_le32(0x00000000);
752 	msg[6] = cpu_to_le32(0x00000000);
753 	msg[7] = cpu_to_le32(0x00000780);
754 	msg[8] = cpu_to_le32(0x00000440);
755 	msg[9] = cpu_to_le32(0x00000000);
756 	msg[10] = cpu_to_le32(0x01b37000);
757 	for (i = 11; i < 1024; ++i)
758 		msg[i] = cpu_to_le32(0x0);
759 
760 	r = radeon_uvd_send_msg(rdev, ring, addr, fence);
761 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
762 	return r;
763 }
764 
radeon_uvd_get_destroy_msg(struct radeon_device * rdev,int ring,uint32_t handle,struct radeon_fence ** fence)765 int radeon_uvd_get_destroy_msg(struct radeon_device *rdev, int ring,
766 			       uint32_t handle, struct radeon_fence **fence)
767 {
768 	/* we use the last page of the vcpu bo for the UVD message */
769 	uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
770 		RADEON_GPU_PAGE_SIZE;
771 
772 	uint32_t *msg = rdev->uvd.cpu_addr + offs;
773 	uint64_t addr = rdev->uvd.gpu_addr + offs;
774 
775 	int r, i;
776 
777 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
778 	if (r)
779 		return r;
780 
781 	/* stitch together an UVD destroy msg */
782 	msg[0] = cpu_to_le32(0x00000de4);
783 	msg[1] = cpu_to_le32(0x00000002);
784 	msg[2] = cpu_to_le32(handle);
785 	msg[3] = cpu_to_le32(0x00000000);
786 	for (i = 4; i < 1024; ++i)
787 		msg[i] = cpu_to_le32(0x0);
788 
789 	r = radeon_uvd_send_msg(rdev, ring, addr, fence);
790 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
791 	return r;
792 }
793 
794 /**
795  * radeon_uvd_count_handles - count number of open streams
796  *
797  * @rdev: radeon_device pointer
798  * @sd: number of SD streams
799  * @hd: number of HD streams
800  *
801  * Count the number of open SD/HD streams as a hint for power mangement
802  */
radeon_uvd_count_handles(struct radeon_device * rdev,unsigned * sd,unsigned * hd)803 static void radeon_uvd_count_handles(struct radeon_device *rdev,
804 				     unsigned *sd, unsigned *hd)
805 {
806 	unsigned i;
807 
808 	*sd = 0;
809 	*hd = 0;
810 
811 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
812 		if (!atomic_read(&rdev->uvd.handles[i]))
813 			continue;
814 
815 		if (rdev->uvd.img_size[i] >= 720*576)
816 			++(*hd);
817 		else
818 			++(*sd);
819 	}
820 }
821 
radeon_uvd_idle_work_handler(struct work_struct * work)822 static void radeon_uvd_idle_work_handler(struct work_struct *work)
823 {
824 	struct radeon_device *rdev =
825 		container_of(work, struct radeon_device, uvd.idle_work.work);
826 
827 	if (radeon_fence_count_emitted(rdev, R600_RING_TYPE_UVD_INDEX) == 0) {
828 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
829 			radeon_uvd_count_handles(rdev, &rdev->pm.dpm.sd,
830 						 &rdev->pm.dpm.hd);
831 			radeon_dpm_enable_uvd(rdev, false);
832 		} else {
833 			radeon_set_uvd_clocks(rdev, 0, 0);
834 		}
835 	} else {
836 		schedule_delayed_work(&rdev->uvd.idle_work,
837 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
838 	}
839 }
840 
radeon_uvd_note_usage(struct radeon_device * rdev)841 void radeon_uvd_note_usage(struct radeon_device *rdev)
842 {
843 	bool streams_changed = false;
844 	bool set_clocks = !cancel_delayed_work_sync(&rdev->uvd.idle_work);
845 	set_clocks &= schedule_delayed_work(&rdev->uvd.idle_work,
846 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
847 
848 	if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
849 		unsigned hd = 0, sd = 0;
850 		radeon_uvd_count_handles(rdev, &sd, &hd);
851 		if ((rdev->pm.dpm.sd != sd) ||
852 		    (rdev->pm.dpm.hd != hd)) {
853 			rdev->pm.dpm.sd = sd;
854 			rdev->pm.dpm.hd = hd;
855 			/* disable this for now */
856 			/*streams_changed = true;*/
857 		}
858 	}
859 
860 	if (set_clocks || streams_changed) {
861 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
862 			radeon_dpm_enable_uvd(rdev, true);
863 		} else {
864 			radeon_set_uvd_clocks(rdev, 53300, 40000);
865 		}
866 	}
867 }
868 
radeon_uvd_calc_upll_post_div(unsigned vco_freq,unsigned target_freq,unsigned pd_min,unsigned pd_even)869 static unsigned radeon_uvd_calc_upll_post_div(unsigned vco_freq,
870 					      unsigned target_freq,
871 					      unsigned pd_min,
872 					      unsigned pd_even)
873 {
874 	unsigned post_div = vco_freq / target_freq;
875 
876 	/* adjust to post divider minimum value */
877 	if (post_div < pd_min)
878 		post_div = pd_min;
879 
880 	/* we alway need a frequency less than or equal the target */
881 	if ((vco_freq / post_div) > target_freq)
882 		post_div += 1;
883 
884 	/* post dividers above a certain value must be even */
885 	if (post_div > pd_even && post_div % 2)
886 		post_div += 1;
887 
888 	return post_div;
889 }
890 
891 /**
892  * radeon_uvd_calc_upll_dividers - calc UPLL clock dividers
893  *
894  * @rdev: radeon_device pointer
895  * @vclk: wanted VCLK
896  * @dclk: wanted DCLK
897  * @vco_min: minimum VCO frequency
898  * @vco_max: maximum VCO frequency
899  * @fb_factor: factor to multiply vco freq with
900  * @fb_mask: limit and bitmask for feedback divider
901  * @pd_min: post divider minimum
902  * @pd_max: post divider maximum
903  * @pd_even: post divider must be even above this value
904  * @optimal_fb_div: resulting feedback divider
905  * @optimal_vclk_div: resulting vclk post divider
906  * @optimal_dclk_div: resulting dclk post divider
907  *
908  * Calculate dividers for UVDs UPLL (R6xx-SI, except APUs).
909  * Returns zero on success -EINVAL on error.
910  */
radeon_uvd_calc_upll_dividers(struct radeon_device * rdev,unsigned vclk,unsigned dclk,unsigned vco_min,unsigned vco_max,unsigned fb_factor,unsigned fb_mask,unsigned pd_min,unsigned pd_max,unsigned pd_even,unsigned * optimal_fb_div,unsigned * optimal_vclk_div,unsigned * optimal_dclk_div)911 int radeon_uvd_calc_upll_dividers(struct radeon_device *rdev,
912 				  unsigned vclk, unsigned dclk,
913 				  unsigned vco_min, unsigned vco_max,
914 				  unsigned fb_factor, unsigned fb_mask,
915 				  unsigned pd_min, unsigned pd_max,
916 				  unsigned pd_even,
917 				  unsigned *optimal_fb_div,
918 				  unsigned *optimal_vclk_div,
919 				  unsigned *optimal_dclk_div)
920 {
921 	unsigned vco_freq, ref_freq = rdev->clock.spll.reference_freq;
922 
923 	/* start off with something large */
924 	unsigned optimal_score = ~0;
925 
926 	/* loop through vco from low to high */
927 	vco_min = max(max(vco_min, vclk), dclk);
928 	for (vco_freq = vco_min; vco_freq <= vco_max; vco_freq += 100) {
929 
930 		uint64_t fb_div = (uint64_t)vco_freq * fb_factor;
931 		unsigned vclk_div, dclk_div, score;
932 
933 		do_div(fb_div, ref_freq);
934 
935 		/* fb div out of range ? */
936 		if (fb_div > fb_mask)
937 			break; /* it can oly get worse */
938 
939 		fb_div &= fb_mask;
940 
941 		/* calc vclk divider with current vco freq */
942 		vclk_div = radeon_uvd_calc_upll_post_div(vco_freq, vclk,
943 							 pd_min, pd_even);
944 		if (vclk_div > pd_max)
945 			break; /* vco is too big, it has to stop */
946 
947 		/* calc dclk divider with current vco freq */
948 		dclk_div = radeon_uvd_calc_upll_post_div(vco_freq, dclk,
949 							 pd_min, pd_even);
950 		if (vclk_div > pd_max)
951 			break; /* vco is too big, it has to stop */
952 
953 		/* calc score with current vco freq */
954 		score = vclk - (vco_freq / vclk_div) + dclk - (vco_freq / dclk_div);
955 
956 		/* determine if this vco setting is better than current optimal settings */
957 		if (score < optimal_score) {
958 			*optimal_fb_div = fb_div;
959 			*optimal_vclk_div = vclk_div;
960 			*optimal_dclk_div = dclk_div;
961 			optimal_score = score;
962 			if (optimal_score == 0)
963 				break; /* it can't get better than this */
964 		}
965 	}
966 
967 	/* did we found a valid setup ? */
968 	if (optimal_score == ~0)
969 		return -EINVAL;
970 
971 	return 0;
972 }
973 
radeon_uvd_send_upll_ctlreq(struct radeon_device * rdev,unsigned cg_upll_func_cntl)974 int radeon_uvd_send_upll_ctlreq(struct radeon_device *rdev,
975 				unsigned cg_upll_func_cntl)
976 {
977 	unsigned i;
978 
979 	/* make sure UPLL_CTLREQ is deasserted */
980 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
981 
982 	mdelay(10);
983 
984 	/* assert UPLL_CTLREQ */
985 	WREG32_P(cg_upll_func_cntl, UPLL_CTLREQ_MASK, ~UPLL_CTLREQ_MASK);
986 
987 	/* wait for CTLACK and CTLACK2 to get asserted */
988 	for (i = 0; i < 100; ++i) {
989 		uint32_t mask = UPLL_CTLACK_MASK | UPLL_CTLACK2_MASK;
990 		if ((RREG32(cg_upll_func_cntl) & mask) == mask)
991 			break;
992 		mdelay(10);
993 	}
994 
995 	/* deassert UPLL_CTLREQ */
996 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
997 
998 	if (i == 100) {
999 		DRM_ERROR("Timeout setting UVD clocks!\n");
1000 		return -ETIMEDOUT;
1001 	}
1002 
1003 	return 0;
1004 }
1005