1 /*	$NetBSD: radeon_uvd.c,v 1.7 2021/12/18 23:45:43 riastradh Exp $	*/
2 
3 /*
4  * Copyright 2011 Advanced Micro Devices, Inc.
5  * All Rights Reserved.
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a
8  * copy of this software and associated documentation files (the
9  * "Software"), to deal in the Software without restriction, including
10  * without limitation the rights to use, copy, modify, merge, publish,
11  * distribute, sub license, and/or sell copies of the Software, and to
12  * permit persons to whom the Software is furnished to do so, subject to
13  * the following conditions:
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
22  *
23  * The above copyright notice and this permission notice (including the
24  * next paragraph) shall be included in all copies or substantial portions
25  * of the Software.
26  *
27  */
28 /*
29  * Authors:
30  *    Christian König <deathsimple@vodafone.de>
31  */
32 
33 #include <sys/cdefs.h>
34 __KERNEL_RCSID(0, "$NetBSD: radeon_uvd.c,v 1.7 2021/12/18 23:45:43 riastradh Exp $");
35 
36 #include <linux/firmware.h>
37 #include <linux/module.h>
38 
39 #include <drm/drm.h>
40 
41 #include "radeon.h"
42 #include "radeon_ucode.h"
43 #include "r600d.h"
44 
45 #include <linux/nbsd-namespace.h>
46 
47 /* 1 second timeout */
48 #define UVD_IDLE_TIMEOUT_MS	1000
49 
50 /* Firmware Names */
51 #define FIRMWARE_R600		"radeon/R600_uvd.bin"
52 #define FIRMWARE_RS780		"radeon/RS780_uvd.bin"
53 #define FIRMWARE_RV770		"radeon/RV770_uvd.bin"
54 #define FIRMWARE_RV710		"radeon/RV710_uvd.bin"
55 #define FIRMWARE_CYPRESS	"radeon/CYPRESS_uvd.bin"
56 #define FIRMWARE_SUMO		"radeon/SUMO_uvd.bin"
57 #define FIRMWARE_TAHITI		"radeon/TAHITI_uvd.bin"
58 #define FIRMWARE_BONAIRE_LEGACY	"radeon/BONAIRE_uvd.bin"
59 #define FIRMWARE_BONAIRE	"radeon/bonaire_uvd.bin"
60 
61 MODULE_FIRMWARE(FIRMWARE_R600);
62 MODULE_FIRMWARE(FIRMWARE_RS780);
63 MODULE_FIRMWARE(FIRMWARE_RV770);
64 MODULE_FIRMWARE(FIRMWARE_RV710);
65 MODULE_FIRMWARE(FIRMWARE_CYPRESS);
66 MODULE_FIRMWARE(FIRMWARE_SUMO);
67 MODULE_FIRMWARE(FIRMWARE_TAHITI);
68 MODULE_FIRMWARE(FIRMWARE_BONAIRE_LEGACY);
69 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
70 
71 static void radeon_uvd_idle_work_handler(struct work_struct *work);
72 
radeon_uvd_init(struct radeon_device * rdev)73 int radeon_uvd_init(struct radeon_device *rdev)
74 {
75 	unsigned long bo_size;
76 	const char *fw_name = NULL, *legacy_fw_name = NULL;
77 	int i, r;
78 
79 	INIT_DELAYED_WORK(&rdev->uvd.idle_work, radeon_uvd_idle_work_handler);
80 
81 	switch (rdev->family) {
82 	case CHIP_RV610:
83 	case CHIP_RV630:
84 	case CHIP_RV670:
85 	case CHIP_RV620:
86 	case CHIP_RV635:
87 		legacy_fw_name = FIRMWARE_R600;
88 		break;
89 
90 	case CHIP_RS780:
91 	case CHIP_RS880:
92 		legacy_fw_name = FIRMWARE_RS780;
93 		break;
94 
95 	case CHIP_RV770:
96 		legacy_fw_name = FIRMWARE_RV770;
97 		break;
98 
99 	case CHIP_RV710:
100 	case CHIP_RV730:
101 	case CHIP_RV740:
102 		legacy_fw_name = FIRMWARE_RV710;
103 		break;
104 
105 	case CHIP_CYPRESS:
106 	case CHIP_HEMLOCK:
107 	case CHIP_JUNIPER:
108 	case CHIP_REDWOOD:
109 	case CHIP_CEDAR:
110 		legacy_fw_name = FIRMWARE_CYPRESS;
111 		break;
112 
113 	case CHIP_SUMO:
114 	case CHIP_SUMO2:
115 	case CHIP_PALM:
116 	case CHIP_CAYMAN:
117 	case CHIP_BARTS:
118 	case CHIP_TURKS:
119 	case CHIP_CAICOS:
120 		legacy_fw_name = FIRMWARE_SUMO;
121 		break;
122 
123 	case CHIP_TAHITI:
124 	case CHIP_VERDE:
125 	case CHIP_PITCAIRN:
126 	case CHIP_ARUBA:
127 	case CHIP_OLAND:
128 		legacy_fw_name = FIRMWARE_TAHITI;
129 		break;
130 
131 	case CHIP_BONAIRE:
132 	case CHIP_KABINI:
133 	case CHIP_KAVERI:
134 	case CHIP_HAWAII:
135 	case CHIP_MULLINS:
136 		legacy_fw_name = FIRMWARE_BONAIRE_LEGACY;
137 		fw_name = FIRMWARE_BONAIRE;
138 		break;
139 
140 	default:
141 		return -EINVAL;
142 	}
143 
144 	rdev->uvd.fw_header_present = false;
145 	rdev->uvd.max_handles = RADEON_DEFAULT_UVD_HANDLES;
146 	if (fw_name) {
147 		/* Let's try to load the newer firmware first */
148 		r = request_firmware(&rdev->uvd_fw, fw_name, rdev->dev);
149 		if (r) {
150 			dev_err(rdev->dev, "radeon_uvd: Can't load firmware \"%s\"\n",
151 				fw_name);
152 		} else {
153 			struct common_firmware_header *hdr = (void *)rdev->uvd_fw->data;
154 			unsigned version_major, version_minor, family_id;
155 
156 			r = radeon_ucode_validate(rdev->uvd_fw);
157 			if (r)
158 				return r;
159 
160 			rdev->uvd.fw_header_present = true;
161 
162 			family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
163 			version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
164 			version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
165 			DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
166 				 version_major, version_minor, family_id);
167 
168 			/*
169 			 * Limit the number of UVD handles depending on
170 			 * microcode major and minor versions.
171 			 */
172 			if ((version_major >= 0x01) && (version_minor >= 0x37))
173 				rdev->uvd.max_handles = RADEON_MAX_UVD_HANDLES;
174 		}
175 	}
176 
177 	/*
178 	 * In case there is only legacy firmware, or we encounter an error
179 	 * while loading the new firmware, we fall back to loading the legacy
180 	 * firmware now.
181 	 */
182 	if (!fw_name || r) {
183 		r = request_firmware(&rdev->uvd_fw, legacy_fw_name, rdev->dev);
184 		if (r) {
185 			dev_err(rdev->dev, "radeon_uvd: Can't load firmware \"%s\"\n",
186 				legacy_fw_name);
187 			return r;
188 		}
189 	}
190 
191 	bo_size = RADEON_GPU_PAGE_ALIGN(rdev->uvd_fw->size + 8) +
192 		  RADEON_UVD_STACK_SIZE + RADEON_UVD_HEAP_SIZE +
193 		  RADEON_UVD_SESSION_SIZE * rdev->uvd.max_handles;
194 	r = radeon_bo_create(rdev, bo_size, PAGE_SIZE, true,
195 			     RADEON_GEM_DOMAIN_VRAM, 0, NULL,
196 			     NULL, &rdev->uvd.vcpu_bo);
197 	if (r) {
198 		dev_err(rdev->dev, "(%d) failed to allocate UVD bo\n", r);
199 		return r;
200 	}
201 
202 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
203 	if (r) {
204 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
205 		dev_err(rdev->dev, "(%d) failed to reserve UVD bo\n", r);
206 		return r;
207 	}
208 
209 	r = radeon_bo_pin(rdev->uvd.vcpu_bo, RADEON_GEM_DOMAIN_VRAM,
210 			  &rdev->uvd.gpu_addr);
211 	if (r) {
212 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
213 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
214 		dev_err(rdev->dev, "(%d) UVD bo pin failed\n", r);
215 		return r;
216 	}
217 
218 	r = radeon_bo_kmap(rdev->uvd.vcpu_bo, &rdev->uvd.cpu_addr);
219 	if (r) {
220 		dev_err(rdev->dev, "(%d) UVD map failed\n", r);
221 		return r;
222 	}
223 
224 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
225 
226 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
227 		atomic_set(&rdev->uvd.handles[i], 0);
228 		rdev->uvd.filp[i] = NULL;
229 		rdev->uvd.img_size[i] = 0;
230 	}
231 
232 	return 0;
233 }
234 
radeon_uvd_fini(struct radeon_device * rdev)235 void radeon_uvd_fini(struct radeon_device *rdev)
236 {
237 	int r;
238 
239 	if (rdev->uvd.vcpu_bo == NULL)
240 		return;
241 
242 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
243 	if (!r) {
244 		radeon_bo_kunmap(rdev->uvd.vcpu_bo);
245 		radeon_bo_unpin(rdev->uvd.vcpu_bo);
246 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
247 	}
248 
249 	radeon_bo_unref(&rdev->uvd.vcpu_bo);
250 
251 	radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_UVD_INDEX]);
252 
253 	release_firmware(rdev->uvd_fw);
254 }
255 
radeon_uvd_suspend(struct radeon_device * rdev)256 int radeon_uvd_suspend(struct radeon_device *rdev)
257 {
258 	int i, r;
259 
260 	if (rdev->uvd.vcpu_bo == NULL)
261 		return 0;
262 
263 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
264 		uint32_t handle = atomic_read(&rdev->uvd.handles[i]);
265 		if (handle != 0) {
266 			struct radeon_fence *fence;
267 
268 			radeon_uvd_note_usage(rdev);
269 
270 			r = radeon_uvd_get_destroy_msg(rdev,
271 				R600_RING_TYPE_UVD_INDEX, handle, &fence);
272 			if (r) {
273 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
274 				continue;
275 			}
276 
277 			radeon_fence_wait(fence, false);
278 			radeon_fence_unref(&fence);
279 
280 			rdev->uvd.filp[i] = NULL;
281 			atomic_set(&rdev->uvd.handles[i], 0);
282 		}
283 	}
284 
285 	return 0;
286 }
287 
radeon_uvd_resume(struct radeon_device * rdev)288 int radeon_uvd_resume(struct radeon_device *rdev)
289 {
290 	unsigned size;
291 	void *ptr;
292 
293 	if (rdev->uvd.vcpu_bo == NULL)
294 		return -EINVAL;
295 
296 	memcpy(rdev->uvd.cpu_addr, rdev->uvd_fw->data, rdev->uvd_fw->size);
297 
298 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
299 	size -= rdev->uvd_fw->size;
300 
301 	ptr = rdev->uvd.cpu_addr;
302 	ptr += rdev->uvd_fw->size;
303 
304 	memset(ptr, 0, size);
305 
306 	return 0;
307 }
308 
radeon_uvd_force_into_uvd_segment(struct radeon_bo * rbo,uint32_t allowed_domains)309 void radeon_uvd_force_into_uvd_segment(struct radeon_bo *rbo,
310 				       uint32_t allowed_domains)
311 {
312 	int i;
313 
314 	for (i = 0; i < rbo->placement.num_placement; ++i) {
315 		rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
316 		rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
317 	}
318 
319 	/* If it must be in VRAM it must be in the first segment as well */
320 	if (allowed_domains == RADEON_GEM_DOMAIN_VRAM)
321 		return;
322 
323 	/* abort if we already have more than one placement */
324 	if (rbo->placement.num_placement > 1)
325 		return;
326 
327 	/* add another 256MB segment */
328 	rbo->placements[1] = rbo->placements[0];
329 	rbo->placements[1].fpfn += (256 * 1024 * 1024) >> PAGE_SHIFT;
330 	rbo->placements[1].lpfn += (256 * 1024 * 1024) >> PAGE_SHIFT;
331 	rbo->placement.num_placement++;
332 	rbo->placement.num_busy_placement++;
333 }
334 
radeon_uvd_free_handles(struct radeon_device * rdev,struct drm_file * filp)335 void radeon_uvd_free_handles(struct radeon_device *rdev, struct drm_file *filp)
336 {
337 	int i, r;
338 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
339 		uint32_t handle = atomic_read(&rdev->uvd.handles[i]);
340 		if (handle != 0 && rdev->uvd.filp[i] == filp) {
341 			struct radeon_fence *fence;
342 
343 			radeon_uvd_note_usage(rdev);
344 
345 			r = radeon_uvd_get_destroy_msg(rdev,
346 				R600_RING_TYPE_UVD_INDEX, handle, &fence);
347 			if (r) {
348 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
349 				continue;
350 			}
351 
352 			radeon_fence_wait(fence, false);
353 			radeon_fence_unref(&fence);
354 
355 			rdev->uvd.filp[i] = NULL;
356 			atomic_set(&rdev->uvd.handles[i], 0);
357 		}
358 	}
359 }
360 
radeon_uvd_cs_msg_decode(uint32_t * msg,unsigned buf_sizes[])361 static int radeon_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
362 {
363 	unsigned stream_type = msg[4];
364 	unsigned width = msg[6];
365 	unsigned height = msg[7];
366 	unsigned dpb_size = msg[9];
367 	unsigned pitch = msg[28];
368 
369 	unsigned width_in_mb = width / 16;
370 	unsigned height_in_mb = ALIGN(height / 16, 2);
371 
372 	unsigned image_size, tmp, min_dpb_size;
373 
374 	image_size = width * height;
375 	image_size += image_size / 2;
376 	image_size = ALIGN(image_size, 1024);
377 
378 	switch (stream_type) {
379 	case 0: /* H264 */
380 
381 		/* reference picture buffer */
382 		min_dpb_size = image_size * 17;
383 
384 		/* macroblock context buffer */
385 		min_dpb_size += width_in_mb * height_in_mb * 17 * 192;
386 
387 		/* IT surface buffer */
388 		min_dpb_size += width_in_mb * height_in_mb * 32;
389 		break;
390 
391 	case 1: /* VC1 */
392 
393 		/* reference picture buffer */
394 		min_dpb_size = image_size * 3;
395 
396 		/* CONTEXT_BUFFER */
397 		min_dpb_size += width_in_mb * height_in_mb * 128;
398 
399 		/* IT surface buffer */
400 		min_dpb_size += width_in_mb * 64;
401 
402 		/* DB surface buffer */
403 		min_dpb_size += width_in_mb * 128;
404 
405 		/* BP */
406 		tmp = max(width_in_mb, height_in_mb);
407 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
408 		break;
409 
410 	case 3: /* MPEG2 */
411 
412 		/* reference picture buffer */
413 		min_dpb_size = image_size * 3;
414 		break;
415 
416 	case 4: /* MPEG4 */
417 
418 		/* reference picture buffer */
419 		min_dpb_size = image_size * 3;
420 
421 		/* CM */
422 		min_dpb_size += width_in_mb * height_in_mb * 64;
423 
424 		/* IT surface buffer */
425 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
426 		break;
427 
428 	default:
429 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
430 		return -EINVAL;
431 	}
432 
433 	if (width > pitch) {
434 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
435 		return -EINVAL;
436 	}
437 
438 	if (dpb_size < min_dpb_size) {
439 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
440 			  dpb_size, min_dpb_size);
441 		return -EINVAL;
442 	}
443 
444 	buf_sizes[0x1] = dpb_size;
445 	buf_sizes[0x2] = image_size;
446 	return 0;
447 }
448 
radeon_uvd_validate_codec(struct radeon_cs_parser * p,unsigned stream_type)449 static int radeon_uvd_validate_codec(struct radeon_cs_parser *p,
450 				     unsigned stream_type)
451 {
452 	switch (stream_type) {
453 	case 0: /* H264 */
454 	case 1: /* VC1 */
455 		/* always supported */
456 		return 0;
457 
458 	case 3: /* MPEG2 */
459 	case 4: /* MPEG4 */
460 		/* only since UVD 3 */
461 		if (p->rdev->family >= CHIP_PALM)
462 			return 0;
463 
464 		/* fall through */
465 	default:
466 		DRM_ERROR("UVD codec not supported by hardware %d!\n",
467 			  stream_type);
468 		return -EINVAL;
469 	}
470 }
471 
radeon_uvd_cs_msg(struct radeon_cs_parser * p,struct radeon_bo * bo,unsigned offset,unsigned buf_sizes[])472 static int radeon_uvd_cs_msg(struct radeon_cs_parser *p, struct radeon_bo *bo,
473 			     unsigned offset, unsigned buf_sizes[])
474 {
475 	int32_t *msg, msg_type, handle;
476 	unsigned img_size = 0;
477 	struct dma_fence *f;
478 	void *ptr;
479 
480 	int i, r;
481 
482 	if (offset & 0x3F) {
483 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
484 		return -EINVAL;
485 	}
486 
487 	f = dma_resv_get_excl(bo->tbo.base.resv);
488 	if (f) {
489 		r = radeon_fence_wait((struct radeon_fence *)f, false);
490 		if (r) {
491 			DRM_ERROR("Failed waiting for UVD message (%d)!\n", r);
492 			return r;
493 		}
494 	}
495 
496 	r = radeon_bo_kmap(bo, &ptr);
497 	if (r) {
498 		DRM_ERROR("Failed mapping the UVD message (%d)!\n", r);
499 		return r;
500 	}
501 
502 	msg = ptr + offset;
503 
504 	msg_type = msg[1];
505 	handle = msg[2];
506 
507 	if (handle == 0) {
508 		DRM_ERROR("Invalid UVD handle!\n");
509 		return -EINVAL;
510 	}
511 
512 	switch (msg_type) {
513 	case 0:
514 		/* it's a create msg, calc image size (width * height) */
515 		img_size = msg[7] * msg[8];
516 
517 		r = radeon_uvd_validate_codec(p, msg[4]);
518 		radeon_bo_kunmap(bo);
519 		if (r)
520 			return r;
521 
522 		/* try to alloc a new handle */
523 		for (i = 0; i < p->rdev->uvd.max_handles; ++i) {
524 			if (atomic_read(&p->rdev->uvd.handles[i]) == handle) {
525 				DRM_ERROR("Handle 0x%x already in use!\n", handle);
526 				return -EINVAL;
527 			}
528 
529 			if (!atomic_cmpxchg(&p->rdev->uvd.handles[i], 0, handle)) {
530 				p->rdev->uvd.filp[i] = p->filp;
531 				p->rdev->uvd.img_size[i] = img_size;
532 				return 0;
533 			}
534 		}
535 
536 		DRM_ERROR("No more free UVD handles!\n");
537 		return -EINVAL;
538 
539 	case 1:
540 		/* it's a decode msg, validate codec and calc buffer sizes */
541 		r = radeon_uvd_validate_codec(p, msg[4]);
542 		if (!r)
543 			r = radeon_uvd_cs_msg_decode(msg, buf_sizes);
544 		radeon_bo_kunmap(bo);
545 		if (r)
546 			return r;
547 
548 		/* validate the handle */
549 		for (i = 0; i < p->rdev->uvd.max_handles; ++i) {
550 			if (atomic_read(&p->rdev->uvd.handles[i]) == handle) {
551 				if (p->rdev->uvd.filp[i] != p->filp) {
552 					DRM_ERROR("UVD handle collision detected!\n");
553 					return -EINVAL;
554 				}
555 				return 0;
556 			}
557 		}
558 
559 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
560 		return -ENOENT;
561 
562 	case 2:
563 		/* it's a destroy msg, free the handle */
564 		for (i = 0; i < p->rdev->uvd.max_handles; ++i)
565 			atomic_cmpxchg(&p->rdev->uvd.handles[i], handle, 0);
566 		radeon_bo_kunmap(bo);
567 		return 0;
568 
569 	default:
570 
571 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
572 		return -EINVAL;
573 	}
574 
575 	BUG();
576 	return -EINVAL;
577 }
578 
radeon_uvd_cs_reloc(struct radeon_cs_parser * p,int data0,int data1,unsigned buf_sizes[],bool * has_msg_cmd)579 static int radeon_uvd_cs_reloc(struct radeon_cs_parser *p,
580 			       int data0, int data1,
581 			       unsigned buf_sizes[], bool *has_msg_cmd)
582 {
583 	struct radeon_cs_chunk *relocs_chunk;
584 	struct radeon_bo_list *reloc;
585 	unsigned idx, cmd, offset;
586 	uint64_t start, end;
587 	int r;
588 
589 	relocs_chunk = p->chunk_relocs;
590 	offset = radeon_get_ib_value(p, data0);
591 	idx = radeon_get_ib_value(p, data1);
592 	if (idx >= relocs_chunk->length_dw) {
593 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
594 			  idx, relocs_chunk->length_dw);
595 		return -EINVAL;
596 	}
597 
598 	reloc = &p->relocs[(idx / 4)];
599 	start = reloc->gpu_offset;
600 	end = start + radeon_bo_size(reloc->robj);
601 	start += offset;
602 
603 	p->ib.ptr[data0] = start & 0xFFFFFFFF;
604 	p->ib.ptr[data1] = start >> 32;
605 
606 	cmd = radeon_get_ib_value(p, p->idx) >> 1;
607 
608 	if (cmd < 0x4) {
609 		if (end <= start) {
610 			DRM_ERROR("invalid reloc offset %X!\n", offset);
611 			return -EINVAL;
612 		}
613 		if ((end - start) < buf_sizes[cmd]) {
614 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
615 				  (unsigned)(end - start), buf_sizes[cmd]);
616 			return -EINVAL;
617 		}
618 
619 	} else if (cmd != 0x100) {
620 		DRM_ERROR("invalid UVD command %X!\n", cmd);
621 		return -EINVAL;
622 	}
623 
624 	if ((start >> 28) != ((end - 1) >> 28)) {
625 		DRM_ERROR("reloc %"PRIX64"-%"PRIX64" crossing 256MB boundary!\n",
626 			  start, end);
627 		return -EINVAL;
628 	}
629 
630 	/* TODO: is this still necessary on NI+ ? */
631 	if ((cmd == 0 || cmd == 0x3) &&
632 	    (start >> 28) != (p->rdev->uvd.gpu_addr >> 28)) {
633 		DRM_ERROR("msg/fb buffer %"PRIX64"-%"PRIX64" out of 256MB segment!\n",
634 			  start, end);
635 		return -EINVAL;
636 	}
637 
638 	if (cmd == 0) {
639 		if (*has_msg_cmd) {
640 			DRM_ERROR("More than one message in a UVD-IB!\n");
641 			return -EINVAL;
642 		}
643 		*has_msg_cmd = true;
644 		r = radeon_uvd_cs_msg(p, reloc->robj, offset, buf_sizes);
645 		if (r)
646 			return r;
647 	} else if (!*has_msg_cmd) {
648 		DRM_ERROR("Message needed before other commands are send!\n");
649 		return -EINVAL;
650 	}
651 
652 	return 0;
653 }
654 
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)655 static int radeon_uvd_cs_reg(struct radeon_cs_parser *p,
656 			     struct radeon_cs_packet *pkt,
657 			     int *data0, int *data1,
658 			     unsigned buf_sizes[],
659 			     bool *has_msg_cmd)
660 {
661 	int i, r;
662 
663 	p->idx++;
664 	for (i = 0; i <= pkt->count; ++i) {
665 		switch (pkt->reg + i*4) {
666 		case UVD_GPCOM_VCPU_DATA0:
667 			*data0 = p->idx;
668 			break;
669 		case UVD_GPCOM_VCPU_DATA1:
670 			*data1 = p->idx;
671 			break;
672 		case UVD_GPCOM_VCPU_CMD:
673 			r = radeon_uvd_cs_reloc(p, *data0, *data1,
674 						buf_sizes, has_msg_cmd);
675 			if (r)
676 				return r;
677 			break;
678 		case UVD_ENGINE_CNTL:
679 		case UVD_NO_OP:
680 			break;
681 		default:
682 			DRM_ERROR("Invalid reg 0x%X!\n",
683 				  pkt->reg + i*4);
684 			return -EINVAL;
685 		}
686 		p->idx++;
687 	}
688 	return 0;
689 }
690 
radeon_uvd_cs_parse(struct radeon_cs_parser * p)691 int radeon_uvd_cs_parse(struct radeon_cs_parser *p)
692 {
693 	struct radeon_cs_packet pkt;
694 	int r, data0 = 0, data1 = 0;
695 
696 	/* does the IB has a msg command */
697 	bool has_msg_cmd = false;
698 
699 	/* minimum buffer sizes */
700 	unsigned buf_sizes[] = {
701 		[0x00000000]	=	2048,
702 		[0x00000001]	=	32 * 1024 * 1024,
703 		[0x00000002]	=	2048 * 1152 * 3,
704 		[0x00000003]	=	2048,
705 	};
706 
707 	if (p->chunk_ib->length_dw % 16) {
708 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
709 			  p->chunk_ib->length_dw);
710 		return -EINVAL;
711 	}
712 
713 	if (p->chunk_relocs == NULL) {
714 		DRM_ERROR("No relocation chunk !\n");
715 		return -EINVAL;
716 	}
717 
718 
719 	do {
720 		r = radeon_cs_packet_parse(p, &pkt, p->idx);
721 		if (r)
722 			return r;
723 		switch (pkt.type) {
724 		case RADEON_PACKET_TYPE0:
725 			r = radeon_uvd_cs_reg(p, &pkt, &data0, &data1,
726 					      buf_sizes, &has_msg_cmd);
727 			if (r)
728 				return r;
729 			break;
730 		case RADEON_PACKET_TYPE2:
731 			p->idx += pkt.count + 2;
732 			break;
733 		default:
734 			DRM_ERROR("Unknown packet type %d !\n", pkt.type);
735 			return -EINVAL;
736 		}
737 	} while (p->idx < p->chunk_ib->length_dw);
738 
739 	if (!has_msg_cmd) {
740 		DRM_ERROR("UVD-IBs need a msg command!\n");
741 		return -EINVAL;
742 	}
743 
744 	return 0;
745 }
746 
radeon_uvd_send_msg(struct radeon_device * rdev,int ring,uint64_t addr,struct radeon_fence ** fence)747 static int radeon_uvd_send_msg(struct radeon_device *rdev,
748 			       int ring, uint64_t addr,
749 			       struct radeon_fence **fence)
750 {
751 	struct radeon_ib ib;
752 	int i, r;
753 
754 	r = radeon_ib_get(rdev, ring, &ib, NULL, 64);
755 	if (r)
756 		return r;
757 
758 	ib.ptr[0] = PACKET0(UVD_GPCOM_VCPU_DATA0, 0);
759 	ib.ptr[1] = addr;
760 	ib.ptr[2] = PACKET0(UVD_GPCOM_VCPU_DATA1, 0);
761 	ib.ptr[3] = addr >> 32;
762 	ib.ptr[4] = PACKET0(UVD_GPCOM_VCPU_CMD, 0);
763 	ib.ptr[5] = 0;
764 	for (i = 6; i < 16; i += 2) {
765 		ib.ptr[i] = PACKET0(UVD_NO_OP, 0);
766 		ib.ptr[i+1] = 0;
767 	}
768 	ib.length_dw = 16;
769 
770 	r = radeon_ib_schedule(rdev, &ib, NULL, false);
771 
772 	if (fence)
773 		*fence = radeon_fence_ref(ib.fence);
774 
775 	radeon_ib_free(rdev, &ib);
776 	return r;
777 }
778 
779 /*
780  * multiple fence commands without any stream commands in between can
781  * crash the vcpu so just try to emmit a dummy create/destroy msg to
782  * avoid this
783  */
radeon_uvd_get_create_msg(struct radeon_device * rdev,int ring,uint32_t handle,struct radeon_fence ** fence)784 int radeon_uvd_get_create_msg(struct radeon_device *rdev, int ring,
785 			      uint32_t handle, struct radeon_fence **fence)
786 {
787 	/* we use the last page of the vcpu bo for the UVD message */
788 	uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
789 		RADEON_GPU_PAGE_SIZE;
790 
791 	uint32_t *msg = rdev->uvd.cpu_addr + offs;
792 	uint64_t addr = rdev->uvd.gpu_addr + offs;
793 
794 	int r, i;
795 
796 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
797 	if (r)
798 		return r;
799 
800 	/* stitch together an UVD create msg */
801 	msg[0] = cpu_to_le32(0x00000de4);
802 	msg[1] = cpu_to_le32(0x00000000);
803 	msg[2] = cpu_to_le32(handle);
804 	msg[3] = cpu_to_le32(0x00000000);
805 	msg[4] = cpu_to_le32(0x00000000);
806 	msg[5] = cpu_to_le32(0x00000000);
807 	msg[6] = cpu_to_le32(0x00000000);
808 	msg[7] = cpu_to_le32(0x00000780);
809 	msg[8] = cpu_to_le32(0x00000440);
810 	msg[9] = cpu_to_le32(0x00000000);
811 	msg[10] = cpu_to_le32(0x01b37000);
812 	for (i = 11; i < 1024; ++i)
813 		msg[i] = cpu_to_le32(0x0);
814 
815 	r = radeon_uvd_send_msg(rdev, ring, addr, fence);
816 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
817 	return r;
818 }
819 
radeon_uvd_get_destroy_msg(struct radeon_device * rdev,int ring,uint32_t handle,struct radeon_fence ** fence)820 int radeon_uvd_get_destroy_msg(struct radeon_device *rdev, int ring,
821 			       uint32_t handle, struct radeon_fence **fence)
822 {
823 	/* we use the last page of the vcpu bo for the UVD message */
824 	uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
825 		RADEON_GPU_PAGE_SIZE;
826 
827 	uint32_t *msg = rdev->uvd.cpu_addr + offs;
828 	uint64_t addr = rdev->uvd.gpu_addr + offs;
829 
830 	int r, i;
831 
832 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
833 	if (r)
834 		return r;
835 
836 	/* stitch together an UVD destroy msg */
837 	msg[0] = cpu_to_le32(0x00000de4);
838 	msg[1] = cpu_to_le32(0x00000002);
839 	msg[2] = cpu_to_le32(handle);
840 	msg[3] = cpu_to_le32(0x00000000);
841 	for (i = 4; i < 1024; ++i)
842 		msg[i] = cpu_to_le32(0x0);
843 
844 	r = radeon_uvd_send_msg(rdev, ring, addr, fence);
845 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
846 	return r;
847 }
848 
849 /**
850  * radeon_uvd_count_handles - count number of open streams
851  *
852  * @rdev: radeon_device pointer
853  * @sd: number of SD streams
854  * @hd: number of HD streams
855  *
856  * Count the number of open SD/HD streams as a hint for power mangement
857  */
radeon_uvd_count_handles(struct radeon_device * rdev,unsigned * sd,unsigned * hd)858 static void radeon_uvd_count_handles(struct radeon_device *rdev,
859 				     unsigned *sd, unsigned *hd)
860 {
861 	unsigned i;
862 
863 	*sd = 0;
864 	*hd = 0;
865 
866 	for (i = 0; i < rdev->uvd.max_handles; ++i) {
867 		if (!atomic_read(&rdev->uvd.handles[i]))
868 			continue;
869 
870 		if (rdev->uvd.img_size[i] >= 720*576)
871 			++(*hd);
872 		else
873 			++(*sd);
874 	}
875 }
876 
radeon_uvd_idle_work_handler(struct work_struct * work)877 static void radeon_uvd_idle_work_handler(struct work_struct *work)
878 {
879 	struct radeon_device *rdev =
880 		container_of(work, struct radeon_device, uvd.idle_work.work);
881 
882 	if (radeon_fence_count_emitted(rdev, R600_RING_TYPE_UVD_INDEX) == 0) {
883 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
884 			radeon_uvd_count_handles(rdev, &rdev->pm.dpm.sd,
885 						 &rdev->pm.dpm.hd);
886 			radeon_dpm_enable_uvd(rdev, false);
887 		} else {
888 			radeon_set_uvd_clocks(rdev, 0, 0);
889 		}
890 	} else {
891 		schedule_delayed_work(&rdev->uvd.idle_work,
892 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
893 	}
894 }
895 
radeon_uvd_note_usage(struct radeon_device * rdev)896 void radeon_uvd_note_usage(struct radeon_device *rdev)
897 {
898 	bool streams_changed = false;
899 	bool set_clocks = !cancel_delayed_work_sync(&rdev->uvd.idle_work);
900 	set_clocks &= schedule_delayed_work(&rdev->uvd.idle_work,
901 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
902 
903 	if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
904 		unsigned hd = 0, sd = 0;
905 		radeon_uvd_count_handles(rdev, &sd, &hd);
906 		if ((rdev->pm.dpm.sd != sd) ||
907 		    (rdev->pm.dpm.hd != hd)) {
908 			rdev->pm.dpm.sd = sd;
909 			rdev->pm.dpm.hd = hd;
910 			/* disable this for now */
911 			/*streams_changed = true;*/
912 		}
913 	}
914 
915 	if (set_clocks || streams_changed) {
916 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
917 			radeon_dpm_enable_uvd(rdev, true);
918 		} else {
919 			radeon_set_uvd_clocks(rdev, 53300, 40000);
920 		}
921 	}
922 }
923 
radeon_uvd_calc_upll_post_div(unsigned vco_freq,unsigned target_freq,unsigned pd_min,unsigned pd_even)924 static unsigned radeon_uvd_calc_upll_post_div(unsigned vco_freq,
925 					      unsigned target_freq,
926 					      unsigned pd_min,
927 					      unsigned pd_even)
928 {
929 	unsigned post_div = vco_freq / target_freq;
930 
931 	/* adjust to post divider minimum value */
932 	if (post_div < pd_min)
933 		post_div = pd_min;
934 
935 	/* we alway need a frequency less than or equal the target */
936 	if ((vco_freq / post_div) > target_freq)
937 		post_div += 1;
938 
939 	/* post dividers above a certain value must be even */
940 	if (post_div > pd_even && post_div % 2)
941 		post_div += 1;
942 
943 	return post_div;
944 }
945 
946 /**
947  * radeon_uvd_calc_upll_dividers - calc UPLL clock dividers
948  *
949  * @rdev: radeon_device pointer
950  * @vclk: wanted VCLK
951  * @dclk: wanted DCLK
952  * @vco_min: minimum VCO frequency
953  * @vco_max: maximum VCO frequency
954  * @fb_factor: factor to multiply vco freq with
955  * @fb_mask: limit and bitmask for feedback divider
956  * @pd_min: post divider minimum
957  * @pd_max: post divider maximum
958  * @pd_even: post divider must be even above this value
959  * @optimal_fb_div: resulting feedback divider
960  * @optimal_vclk_div: resulting vclk post divider
961  * @optimal_dclk_div: resulting dclk post divider
962  *
963  * Calculate dividers for UVDs UPLL (R6xx-SI, except APUs).
964  * Returns zero on success -EINVAL on error.
965  */
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)966 int radeon_uvd_calc_upll_dividers(struct radeon_device *rdev,
967 				  unsigned vclk, unsigned dclk,
968 				  unsigned vco_min, unsigned vco_max,
969 				  unsigned fb_factor, unsigned fb_mask,
970 				  unsigned pd_min, unsigned pd_max,
971 				  unsigned pd_even,
972 				  unsigned *optimal_fb_div,
973 				  unsigned *optimal_vclk_div,
974 				  unsigned *optimal_dclk_div)
975 {
976 	unsigned vco_freq, ref_freq = rdev->clock.spll.reference_freq;
977 
978 	/* start off with something large */
979 	unsigned optimal_score = ~0;
980 
981 	/* loop through vco from low to high */
982 	vco_min = max(max(vco_min, vclk), dclk);
983 	for (vco_freq = vco_min; vco_freq <= vco_max; vco_freq += 100) {
984 
985 		uint64_t fb_div = (uint64_t)vco_freq * fb_factor;
986 		unsigned vclk_div, dclk_div, score;
987 
988 		do_div(fb_div, ref_freq);
989 
990 		/* fb div out of range ? */
991 		if (fb_div > fb_mask)
992 			break; /* it can oly get worse */
993 
994 		fb_div &= fb_mask;
995 
996 		/* calc vclk divider with current vco freq */
997 		vclk_div = radeon_uvd_calc_upll_post_div(vco_freq, vclk,
998 							 pd_min, pd_even);
999 		if (vclk_div > pd_max)
1000 			break; /* vco is too big, it has to stop */
1001 
1002 		/* calc dclk divider with current vco freq */
1003 		dclk_div = radeon_uvd_calc_upll_post_div(vco_freq, dclk,
1004 							 pd_min, pd_even);
1005 		if (dclk_div > pd_max)
1006 			break; /* vco is too big, it has to stop */
1007 
1008 		/* calc score with current vco freq */
1009 		score = vclk - (vco_freq / vclk_div) + dclk - (vco_freq / dclk_div);
1010 
1011 		/* determine if this vco setting is better than current optimal settings */
1012 		if (score < optimal_score) {
1013 			*optimal_fb_div = fb_div;
1014 			*optimal_vclk_div = vclk_div;
1015 			*optimal_dclk_div = dclk_div;
1016 			optimal_score = score;
1017 			if (optimal_score == 0)
1018 				break; /* it can't get better than this */
1019 		}
1020 	}
1021 
1022 	/* did we found a valid setup ? */
1023 	if (optimal_score == ~0)
1024 		return -EINVAL;
1025 
1026 	return 0;
1027 }
1028 
radeon_uvd_send_upll_ctlreq(struct radeon_device * rdev,unsigned cg_upll_func_cntl)1029 int radeon_uvd_send_upll_ctlreq(struct radeon_device *rdev,
1030 				unsigned cg_upll_func_cntl)
1031 {
1032 	unsigned i;
1033 
1034 	/* make sure UPLL_CTLREQ is deasserted */
1035 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
1036 
1037 	mdelay(10);
1038 
1039 	/* assert UPLL_CTLREQ */
1040 	WREG32_P(cg_upll_func_cntl, UPLL_CTLREQ_MASK, ~UPLL_CTLREQ_MASK);
1041 
1042 	/* wait for CTLACK and CTLACK2 to get asserted */
1043 	for (i = 0; i < 100; ++i) {
1044 		uint32_t mask = UPLL_CTLACK_MASK | UPLL_CTLACK2_MASK;
1045 		if ((RREG32(cg_upll_func_cntl) & mask) == mask)
1046 			break;
1047 		mdelay(10);
1048 	}
1049 
1050 	/* deassert UPLL_CTLREQ */
1051 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
1052 
1053 	if (i == 100) {
1054 		DRM_ERROR("Timeout setting UVD clocks!\n");
1055 		return -ETIMEDOUT;
1056 	}
1057 
1058 	return 0;
1059 }
1060