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_RV710		"radeon/RV710_uvd.bin"
44 #define FIRMWARE_CYPRESS	"radeon/CYPRESS_uvd.bin"
45 #define FIRMWARE_SUMO		"radeon/SUMO_uvd.bin"
46 #define FIRMWARE_TAHITI		"radeon/TAHITI_uvd.bin"
47 #define FIRMWARE_BONAIRE	"radeon/BONAIRE_uvd.bin"
48 
49 MODULE_FIRMWARE(FIRMWARE_RV710);
50 MODULE_FIRMWARE(FIRMWARE_CYPRESS);
51 MODULE_FIRMWARE(FIRMWARE_SUMO);
52 MODULE_FIRMWARE(FIRMWARE_TAHITI);
53 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
54 
55 static void radeon_uvd_idle_work_handler(struct work_struct *work);
56 
57 int radeon_uvd_init(struct radeon_device *rdev)
58 {
59 	unsigned long bo_size;
60 	const char *fw_name;
61 	int i, r;
62 
63 	INIT_DELAYED_WORK(&rdev->uvd.idle_work, radeon_uvd_idle_work_handler);
64 
65 	switch (rdev->family) {
66 	case CHIP_RV710:
67 	case CHIP_RV730:
68 	case CHIP_RV740:
69 		fw_name = FIRMWARE_RV710;
70 		break;
71 
72 	case CHIP_CYPRESS:
73 	case CHIP_HEMLOCK:
74 	case CHIP_JUNIPER:
75 	case CHIP_REDWOOD:
76 	case CHIP_CEDAR:
77 		fw_name = FIRMWARE_CYPRESS;
78 		break;
79 
80 	case CHIP_SUMO:
81 	case CHIP_SUMO2:
82 	case CHIP_PALM:
83 	case CHIP_CAYMAN:
84 	case CHIP_BARTS:
85 	case CHIP_TURKS:
86 	case CHIP_CAICOS:
87 		fw_name = FIRMWARE_SUMO;
88 		break;
89 
90 	case CHIP_TAHITI:
91 	case CHIP_VERDE:
92 	case CHIP_PITCAIRN:
93 	case CHIP_ARUBA:
94 		fw_name = FIRMWARE_TAHITI;
95 		break;
96 
97 	case CHIP_BONAIRE:
98 	case CHIP_KABINI:
99 	case CHIP_KAVERI:
100 		fw_name = FIRMWARE_BONAIRE;
101 		break;
102 
103 	default:
104 		return -EINVAL;
105 	}
106 
107 	r = request_firmware(&rdev->uvd_fw, fw_name, rdev->dev);
108 	if (r) {
109 		dev_err(rdev->dev, "radeon_uvd: Can't load firmware \"%s\"\n",
110 			fw_name);
111 		return r;
112 	}
113 
114 	bo_size = RADEON_GPU_PAGE_ALIGN(rdev->uvd_fw->size + 8) +
115 		  RADEON_UVD_STACK_SIZE + RADEON_UVD_HEAP_SIZE;
116 	r = radeon_bo_create(rdev, bo_size, PAGE_SIZE, true,
117 			     RADEON_GEM_DOMAIN_VRAM, NULL, &rdev->uvd.vcpu_bo);
118 	if (r) {
119 		dev_err(rdev->dev, "(%d) failed to allocate UVD bo\n", r);
120 		return r;
121 	}
122 
123 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
124 	if (r) {
125 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
126 		dev_err(rdev->dev, "(%d) failed to reserve UVD bo\n", r);
127 		return r;
128 	}
129 
130 	r = radeon_bo_pin(rdev->uvd.vcpu_bo, RADEON_GEM_DOMAIN_VRAM,
131 			  &rdev->uvd.gpu_addr);
132 	if (r) {
133 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
134 		radeon_bo_unref(&rdev->uvd.vcpu_bo);
135 		dev_err(rdev->dev, "(%d) UVD bo pin failed\n", r);
136 		return r;
137 	}
138 
139 	r = radeon_bo_kmap(rdev->uvd.vcpu_bo, &rdev->uvd.cpu_addr);
140 	if (r) {
141 		dev_err(rdev->dev, "(%d) UVD map failed\n", r);
142 		return r;
143 	}
144 
145 	radeon_bo_unreserve(rdev->uvd.vcpu_bo);
146 
147 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
148 		atomic_set(&rdev->uvd.handles[i], 0);
149 		rdev->uvd.filp[i] = NULL;
150 		rdev->uvd.img_size[i] = 0;
151 	}
152 
153 	return 0;
154 }
155 
156 void radeon_uvd_fini(struct radeon_device *rdev)
157 {
158 	int r;
159 
160 	if (rdev->uvd.vcpu_bo == NULL)
161 		return;
162 
163 	r = radeon_bo_reserve(rdev->uvd.vcpu_bo, false);
164 	if (!r) {
165 		radeon_bo_kunmap(rdev->uvd.vcpu_bo);
166 		radeon_bo_unpin(rdev->uvd.vcpu_bo);
167 		radeon_bo_unreserve(rdev->uvd.vcpu_bo);
168 	}
169 
170 	radeon_bo_unref(&rdev->uvd.vcpu_bo);
171 
172 	release_firmware(rdev->uvd_fw);
173 }
174 
175 int radeon_uvd_suspend(struct radeon_device *rdev)
176 {
177 	unsigned size;
178 	void *ptr;
179 	int i;
180 
181 	if (rdev->uvd.vcpu_bo == NULL)
182 		return 0;
183 
184 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i)
185 		if (atomic_read(&rdev->uvd.handles[i]))
186 			break;
187 
188 	if (i == RADEON_MAX_UVD_HANDLES)
189 		return 0;
190 
191 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
192 	size -= rdev->uvd_fw->size;
193 
194 	ptr = rdev->uvd.cpu_addr;
195 	ptr += rdev->uvd_fw->size;
196 
197 	rdev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
198 	memcpy(rdev->uvd.saved_bo, ptr, size);
199 
200 	return 0;
201 }
202 
203 int radeon_uvd_resume(struct radeon_device *rdev)
204 {
205 	unsigned size;
206 	void *ptr;
207 
208 	if (rdev->uvd.vcpu_bo == NULL)
209 		return -EINVAL;
210 
211 	memcpy(rdev->uvd.cpu_addr, rdev->uvd_fw->data, rdev->uvd_fw->size);
212 
213 	size = radeon_bo_size(rdev->uvd.vcpu_bo);
214 	size -= rdev->uvd_fw->size;
215 
216 	ptr = rdev->uvd.cpu_addr;
217 	ptr += rdev->uvd_fw->size;
218 
219 	if (rdev->uvd.saved_bo != NULL) {
220 		memcpy(ptr, rdev->uvd.saved_bo, size);
221 		kfree(rdev->uvd.saved_bo);
222 		rdev->uvd.saved_bo = NULL;
223 	} else
224 		memset(ptr, 0, size);
225 
226 	return 0;
227 }
228 
229 void radeon_uvd_force_into_uvd_segment(struct radeon_bo *rbo)
230 {
231 	rbo->placement.fpfn = 0 >> PAGE_SHIFT;
232 	rbo->placement.lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
233 }
234 
235 void radeon_uvd_free_handles(struct radeon_device *rdev, struct drm_file *filp)
236 {
237 	int i, r;
238 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
239 		uint32_t handle = atomic_read(&rdev->uvd.handles[i]);
240 		if (handle != 0 && rdev->uvd.filp[i] == filp) {
241 			struct radeon_fence *fence;
242 
243 			r = radeon_uvd_get_destroy_msg(rdev,
244 				R600_RING_TYPE_UVD_INDEX, handle, &fence);
245 			if (r) {
246 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
247 				continue;
248 			}
249 
250 			radeon_fence_wait(fence, false);
251 			radeon_fence_unref(&fence);
252 
253 			rdev->uvd.filp[i] = NULL;
254 			atomic_set(&rdev->uvd.handles[i], 0);
255 		}
256 	}
257 }
258 
259 static int radeon_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
260 {
261 	unsigned stream_type = msg[4];
262 	unsigned width = msg[6];
263 	unsigned height = msg[7];
264 	unsigned dpb_size = msg[9];
265 	unsigned pitch = msg[28];
266 
267 	unsigned width_in_mb = width / 16;
268 	unsigned height_in_mb = ALIGN(height / 16, 2);
269 
270 	unsigned image_size, tmp, min_dpb_size;
271 
272 	image_size = width * height;
273 	image_size += image_size / 2;
274 	image_size = ALIGN(image_size, 1024);
275 
276 	switch (stream_type) {
277 	case 0: /* H264 */
278 
279 		/* reference picture buffer */
280 		min_dpb_size = image_size * 17;
281 
282 		/* macroblock context buffer */
283 		min_dpb_size += width_in_mb * height_in_mb * 17 * 192;
284 
285 		/* IT surface buffer */
286 		min_dpb_size += width_in_mb * height_in_mb * 32;
287 		break;
288 
289 	case 1: /* VC1 */
290 
291 		/* reference picture buffer */
292 		min_dpb_size = image_size * 3;
293 
294 		/* CONTEXT_BUFFER */
295 		min_dpb_size += width_in_mb * height_in_mb * 128;
296 
297 		/* IT surface buffer */
298 		min_dpb_size += width_in_mb * 64;
299 
300 		/* DB surface buffer */
301 		min_dpb_size += width_in_mb * 128;
302 
303 		/* BP */
304 		tmp = max(width_in_mb, height_in_mb);
305 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
306 		break;
307 
308 	case 3: /* MPEG2 */
309 
310 		/* reference picture buffer */
311 		min_dpb_size = image_size * 3;
312 		break;
313 
314 	case 4: /* MPEG4 */
315 
316 		/* reference picture buffer */
317 		min_dpb_size = image_size * 3;
318 
319 		/* CM */
320 		min_dpb_size += width_in_mb * height_in_mb * 64;
321 
322 		/* IT surface buffer */
323 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
324 		break;
325 
326 	default:
327 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
328 		return -EINVAL;
329 	}
330 
331 	if (width > pitch) {
332 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
333 		return -EINVAL;
334 	}
335 
336 	if (dpb_size < min_dpb_size) {
337 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
338 			  dpb_size, min_dpb_size);
339 		return -EINVAL;
340 	}
341 
342 	buf_sizes[0x1] = dpb_size;
343 	buf_sizes[0x2] = image_size;
344 	return 0;
345 }
346 
347 static int radeon_uvd_cs_msg(struct radeon_cs_parser *p, struct radeon_bo *bo,
348 			     unsigned offset, unsigned buf_sizes[])
349 {
350 	int32_t *msg, msg_type, handle;
351 	unsigned img_size = 0;
352 	void *ptr;
353 
354 	int i, r;
355 
356 	if (offset & 0x3F) {
357 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
358 		return -EINVAL;
359 	}
360 
361 	if (bo->tbo.sync_obj) {
362 		r = radeon_fence_wait(bo->tbo.sync_obj, false);
363 		if (r) {
364 			DRM_ERROR("Failed waiting for UVD message (%d)!\n", r);
365 			return r;
366 		}
367 	}
368 
369 	r = radeon_bo_kmap(bo, &ptr);
370 	if (r) {
371 		DRM_ERROR("Failed mapping the UVD message (%d)!\n", r);
372 		return r;
373 	}
374 
375 	msg = ptr + offset;
376 
377 	msg_type = msg[1];
378 	handle = msg[2];
379 
380 	if (handle == 0) {
381 		DRM_ERROR("Invalid UVD handle!\n");
382 		return -EINVAL;
383 	}
384 
385 	if (msg_type == 1) {
386 		/* it's a decode msg, calc buffer sizes */
387 		r = radeon_uvd_cs_msg_decode(msg, buf_sizes);
388 		/* calc image size (width * height) */
389 		img_size = msg[6] * msg[7];
390 		radeon_bo_kunmap(bo);
391 		if (r)
392 			return r;
393 
394 	} else if (msg_type == 2) {
395 		/* it's a destroy msg, free the handle */
396 		for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i)
397 			atomic_cmpxchg(&p->rdev->uvd.handles[i], handle, 0);
398 		radeon_bo_kunmap(bo);
399 		return 0;
400 	} else {
401 		/* it's a create msg, calc image size (width * height) */
402 		img_size = msg[7] * msg[8];
403 		radeon_bo_kunmap(bo);
404 
405 		if (msg_type != 0) {
406 			DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
407 			return -EINVAL;
408 		}
409 
410 		/* it's a create msg, no special handling needed */
411 	}
412 
413 	/* create or decode, validate the handle */
414 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
415 		if (atomic_read(&p->rdev->uvd.handles[i]) == handle)
416 			return 0;
417 	}
418 
419 	/* handle not found try to alloc a new one */
420 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
421 		if (!atomic_cmpxchg(&p->rdev->uvd.handles[i], 0, handle)) {
422 			p->rdev->uvd.filp[i] = p->filp;
423 			p->rdev->uvd.img_size[i] = img_size;
424 			return 0;
425 		}
426 	}
427 
428 	DRM_ERROR("No more free UVD handles!\n");
429 	return -EINVAL;
430 }
431 
432 static int radeon_uvd_cs_reloc(struct radeon_cs_parser *p,
433 			       int data0, int data1,
434 			       unsigned buf_sizes[], bool *has_msg_cmd)
435 {
436 	struct radeon_cs_chunk *relocs_chunk;
437 	struct radeon_cs_reloc *reloc;
438 	unsigned idx, cmd, offset;
439 	uint64_t start, end;
440 	int r;
441 
442 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
443 	offset = radeon_get_ib_value(p, data0);
444 	idx = radeon_get_ib_value(p, data1);
445 	if (idx >= relocs_chunk->length_dw) {
446 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
447 			  idx, relocs_chunk->length_dw);
448 		return -EINVAL;
449 	}
450 
451 	reloc = p->relocs_ptr[(idx / 4)];
452 	start = reloc->lobj.gpu_offset;
453 	end = start + radeon_bo_size(reloc->robj);
454 	start += offset;
455 
456 	p->ib.ptr[data0] = start & 0xFFFFFFFF;
457 	p->ib.ptr[data1] = start >> 32;
458 
459 	cmd = radeon_get_ib_value(p, p->idx) >> 1;
460 
461 	if (cmd < 0x4) {
462 		if ((end - start) < buf_sizes[cmd]) {
463 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
464 				  (unsigned)(end - start), buf_sizes[cmd]);
465 			return -EINVAL;
466 		}
467 
468 	} else if (cmd != 0x100) {
469 		DRM_ERROR("invalid UVD command %X!\n", cmd);
470 		return -EINVAL;
471 	}
472 
473 	if ((start >> 28) != (end >> 28)) {
474 		DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
475 			  start, end);
476 		return -EINVAL;
477 	}
478 
479 	/* TODO: is this still necessary on NI+ ? */
480 	if ((cmd == 0 || cmd == 0x3) &&
481 	    (start >> 28) != (p->rdev->uvd.gpu_addr >> 28)) {
482 		DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
483 			  start, end);
484 		return -EINVAL;
485 	}
486 
487 	if (cmd == 0) {
488 		if (*has_msg_cmd) {
489 			DRM_ERROR("More than one message in a UVD-IB!\n");
490 			return -EINVAL;
491 		}
492 		*has_msg_cmd = true;
493 		r = radeon_uvd_cs_msg(p, reloc->robj, offset, buf_sizes);
494 		if (r)
495 			return r;
496 	} else if (!*has_msg_cmd) {
497 		DRM_ERROR("Message needed before other commands are send!\n");
498 		return -EINVAL;
499 	}
500 
501 	return 0;
502 }
503 
504 static int radeon_uvd_cs_reg(struct radeon_cs_parser *p,
505 			     struct radeon_cs_packet *pkt,
506 			     int *data0, int *data1,
507 			     unsigned buf_sizes[],
508 			     bool *has_msg_cmd)
509 {
510 	int i, r;
511 
512 	p->idx++;
513 	for (i = 0; i <= pkt->count; ++i) {
514 		switch (pkt->reg + i*4) {
515 		case UVD_GPCOM_VCPU_DATA0:
516 			*data0 = p->idx;
517 			break;
518 		case UVD_GPCOM_VCPU_DATA1:
519 			*data1 = p->idx;
520 			break;
521 		case UVD_GPCOM_VCPU_CMD:
522 			r = radeon_uvd_cs_reloc(p, *data0, *data1,
523 						buf_sizes, has_msg_cmd);
524 			if (r)
525 				return r;
526 			break;
527 		case UVD_ENGINE_CNTL:
528 			break;
529 		default:
530 			DRM_ERROR("Invalid reg 0x%X!\n",
531 				  pkt->reg + i*4);
532 			return -EINVAL;
533 		}
534 		p->idx++;
535 	}
536 	return 0;
537 }
538 
539 int radeon_uvd_cs_parse(struct radeon_cs_parser *p)
540 {
541 	struct radeon_cs_packet pkt;
542 	int r, data0 = 0, data1 = 0;
543 
544 	/* does the IB has a msg command */
545 	bool has_msg_cmd = false;
546 
547 	/* minimum buffer sizes */
548 	unsigned buf_sizes[] = {
549 		[0x00000000]	=	2048,
550 		[0x00000001]	=	32 * 1024 * 1024,
551 		[0x00000002]	=	2048 * 1152 * 3,
552 		[0x00000003]	=	2048,
553 	};
554 
555 	if (p->chunks[p->chunk_ib_idx].length_dw % 16) {
556 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
557 			  p->chunks[p->chunk_ib_idx].length_dw);
558 		return -EINVAL;
559 	}
560 
561 	if (p->chunk_relocs_idx == -1) {
562 		DRM_ERROR("No relocation chunk !\n");
563 		return -EINVAL;
564 	}
565 
566 
567 	do {
568 		r = radeon_cs_packet_parse(p, &pkt, p->idx);
569 		if (r)
570 			return r;
571 		switch (pkt.type) {
572 		case RADEON_PACKET_TYPE0:
573 			r = radeon_uvd_cs_reg(p, &pkt, &data0, &data1,
574 					      buf_sizes, &has_msg_cmd);
575 			if (r)
576 				return r;
577 			break;
578 		case RADEON_PACKET_TYPE2:
579 			p->idx += pkt.count + 2;
580 			break;
581 		default:
582 			DRM_ERROR("Unknown packet type %d !\n", pkt.type);
583 			return -EINVAL;
584 		}
585 	} while (p->idx < p->chunks[p->chunk_ib_idx].length_dw);
586 
587 	if (!has_msg_cmd) {
588 		DRM_ERROR("UVD-IBs need a msg command!\n");
589 		return -EINVAL;
590 	}
591 
592 	return 0;
593 }
594 
595 static int radeon_uvd_send_msg(struct radeon_device *rdev,
596 			       int ring, struct radeon_bo *bo,
597 			       struct radeon_fence **fence)
598 {
599 	struct ttm_validate_buffer tv;
600 	struct ww_acquire_ctx ticket;
601 	struct list_head head;
602 	struct radeon_ib ib;
603 	uint64_t addr;
604 	int i, r;
605 
606 	memset(&tv, 0, sizeof(tv));
607 	tv.bo = &bo->tbo;
608 
609 	INIT_LIST_HEAD(&head);
610 	list_add(&tv.head, &head);
611 
612 	r = ttm_eu_reserve_buffers(&ticket, &head);
613 	if (r)
614 		return r;
615 
616 	radeon_ttm_placement_from_domain(bo, RADEON_GEM_DOMAIN_VRAM);
617 	radeon_uvd_force_into_uvd_segment(bo);
618 
619 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
620 	if (r)
621 		goto err;
622 
623 	r = radeon_ib_get(rdev, ring, &ib, NULL, 16);
624 	if (r)
625 		goto err;
626 
627 	addr = radeon_bo_gpu_offset(bo);
628 	ib.ptr[0] = PACKET0(UVD_GPCOM_VCPU_DATA0, 0);
629 	ib.ptr[1] = addr;
630 	ib.ptr[2] = PACKET0(UVD_GPCOM_VCPU_DATA1, 0);
631 	ib.ptr[3] = addr >> 32;
632 	ib.ptr[4] = PACKET0(UVD_GPCOM_VCPU_CMD, 0);
633 	ib.ptr[5] = 0;
634 	for (i = 6; i < 16; ++i)
635 		ib.ptr[i] = PACKET2(0);
636 	ib.length_dw = 16;
637 
638 	r = radeon_ib_schedule(rdev, &ib, NULL);
639 	if (r)
640 		goto err;
641 	ttm_eu_fence_buffer_objects(&ticket, &head, ib.fence);
642 
643 	if (fence)
644 		*fence = radeon_fence_ref(ib.fence);
645 
646 	radeon_ib_free(rdev, &ib);
647 	radeon_bo_unref(&bo);
648 	return 0;
649 
650 err:
651 	ttm_eu_backoff_reservation(&ticket, &head);
652 	return r;
653 }
654 
655 /* multiple fence commands without any stream commands in between can
656    crash the vcpu so just try to emmit a dummy create/destroy msg to
657    avoid this */
658 int radeon_uvd_get_create_msg(struct radeon_device *rdev, int ring,
659 			      uint32_t handle, struct radeon_fence **fence)
660 {
661 	struct radeon_bo *bo;
662 	uint32_t *msg;
663 	int r, i;
664 
665 	r = radeon_bo_create(rdev, 1024, PAGE_SIZE, true,
666 			     RADEON_GEM_DOMAIN_VRAM, NULL, &bo);
667 	if (r)
668 		return r;
669 
670 	r = radeon_bo_reserve(bo, false);
671 	if (r) {
672 		radeon_bo_unref(&bo);
673 		return r;
674 	}
675 
676 	r = radeon_bo_kmap(bo, (void **)&msg);
677 	if (r) {
678 		radeon_bo_unreserve(bo);
679 		radeon_bo_unref(&bo);
680 		return r;
681 	}
682 
683 	/* stitch together an UVD create msg */
684 	msg[0] = cpu_to_le32(0x00000de4);
685 	msg[1] = cpu_to_le32(0x00000000);
686 	msg[2] = cpu_to_le32(handle);
687 	msg[3] = cpu_to_le32(0x00000000);
688 	msg[4] = cpu_to_le32(0x00000000);
689 	msg[5] = cpu_to_le32(0x00000000);
690 	msg[6] = cpu_to_le32(0x00000000);
691 	msg[7] = cpu_to_le32(0x00000780);
692 	msg[8] = cpu_to_le32(0x00000440);
693 	msg[9] = cpu_to_le32(0x00000000);
694 	msg[10] = cpu_to_le32(0x01b37000);
695 	for (i = 11; i < 1024; ++i)
696 		msg[i] = cpu_to_le32(0x0);
697 
698 	radeon_bo_kunmap(bo);
699 	radeon_bo_unreserve(bo);
700 
701 	return radeon_uvd_send_msg(rdev, ring, bo, fence);
702 }
703 
704 int radeon_uvd_get_destroy_msg(struct radeon_device *rdev, int ring,
705 			       uint32_t handle, struct radeon_fence **fence)
706 {
707 	struct radeon_bo *bo;
708 	uint32_t *msg;
709 	int r, i;
710 
711 	r = radeon_bo_create(rdev, 1024, PAGE_SIZE, true,
712 			     RADEON_GEM_DOMAIN_VRAM, NULL, &bo);
713 	if (r)
714 		return r;
715 
716 	r = radeon_bo_reserve(bo, false);
717 	if (r) {
718 		radeon_bo_unref(&bo);
719 		return r;
720 	}
721 
722 	r = radeon_bo_kmap(bo, (void **)&msg);
723 	if (r) {
724 		radeon_bo_unreserve(bo);
725 		radeon_bo_unref(&bo);
726 		return r;
727 	}
728 
729 	/* stitch together an UVD destroy msg */
730 	msg[0] = cpu_to_le32(0x00000de4);
731 	msg[1] = cpu_to_le32(0x00000002);
732 	msg[2] = cpu_to_le32(handle);
733 	msg[3] = cpu_to_le32(0x00000000);
734 	for (i = 4; i < 1024; ++i)
735 		msg[i] = cpu_to_le32(0x0);
736 
737 	radeon_bo_kunmap(bo);
738 	radeon_bo_unreserve(bo);
739 
740 	return radeon_uvd_send_msg(rdev, ring, bo, fence);
741 }
742 
743 /**
744  * radeon_uvd_count_handles - count number of open streams
745  *
746  * @rdev: radeon_device pointer
747  * @sd: number of SD streams
748  * @hd: number of HD streams
749  *
750  * Count the number of open SD/HD streams as a hint for power mangement
751  */
752 static void radeon_uvd_count_handles(struct radeon_device *rdev,
753 				     unsigned *sd, unsigned *hd)
754 {
755 	unsigned i;
756 
757 	*sd = 0;
758 	*hd = 0;
759 
760 	for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
761 		if (!atomic_read(&rdev->uvd.handles[i]))
762 			continue;
763 
764 		if (rdev->uvd.img_size[i] >= 720*576)
765 			++(*hd);
766 		else
767 			++(*sd);
768 	}
769 }
770 
771 static void radeon_uvd_idle_work_handler(struct work_struct *work)
772 {
773 	struct radeon_device *rdev =
774 		container_of(work, struct radeon_device, uvd.idle_work.work);
775 
776 	if (radeon_fence_count_emitted(rdev, R600_RING_TYPE_UVD_INDEX) == 0) {
777 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
778 			radeon_dpm_enable_uvd(rdev, false);
779 		} else {
780 			radeon_set_uvd_clocks(rdev, 0, 0);
781 		}
782 	} else {
783 		schedule_delayed_work(&rdev->uvd.idle_work,
784 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
785 	}
786 }
787 
788 void radeon_uvd_note_usage(struct radeon_device *rdev)
789 {
790 	bool streams_changed = false;
791 	bool set_clocks = !cancel_delayed_work_sync(&rdev->uvd.idle_work);
792 	set_clocks &= schedule_delayed_work(&rdev->uvd.idle_work,
793 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
794 
795 	if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
796 		unsigned hd = 0, sd = 0;
797 		radeon_uvd_count_handles(rdev, &sd, &hd);
798 		if ((rdev->pm.dpm.sd != sd) ||
799 		    (rdev->pm.dpm.hd != hd)) {
800 			rdev->pm.dpm.sd = sd;
801 			rdev->pm.dpm.hd = hd;
802 			/* disable this for now */
803 			/*streams_changed = true;*/
804 		}
805 	}
806 
807 	if (set_clocks || streams_changed) {
808 		if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
809 			radeon_dpm_enable_uvd(rdev, true);
810 		} else {
811 			radeon_set_uvd_clocks(rdev, 53300, 40000);
812 		}
813 	}
814 }
815 
816 static unsigned radeon_uvd_calc_upll_post_div(unsigned vco_freq,
817 					      unsigned target_freq,
818 					      unsigned pd_min,
819 					      unsigned pd_even)
820 {
821 	unsigned post_div = vco_freq / target_freq;
822 
823 	/* adjust to post divider minimum value */
824 	if (post_div < pd_min)
825 		post_div = pd_min;
826 
827 	/* we alway need a frequency less than or equal the target */
828 	if ((vco_freq / post_div) > target_freq)
829 		post_div += 1;
830 
831 	/* post dividers above a certain value must be even */
832 	if (post_div > pd_even && post_div % 2)
833 		post_div += 1;
834 
835 	return post_div;
836 }
837 
838 /**
839  * radeon_uvd_calc_upll_dividers - calc UPLL clock dividers
840  *
841  * @rdev: radeon_device pointer
842  * @vclk: wanted VCLK
843  * @dclk: wanted DCLK
844  * @vco_min: minimum VCO frequency
845  * @vco_max: maximum VCO frequency
846  * @fb_factor: factor to multiply vco freq with
847  * @fb_mask: limit and bitmask for feedback divider
848  * @pd_min: post divider minimum
849  * @pd_max: post divider maximum
850  * @pd_even: post divider must be even above this value
851  * @optimal_fb_div: resulting feedback divider
852  * @optimal_vclk_div: resulting vclk post divider
853  * @optimal_dclk_div: resulting dclk post divider
854  *
855  * Calculate dividers for UVDs UPLL (R6xx-SI, except APUs).
856  * Returns zero on success -EINVAL on error.
857  */
858 int radeon_uvd_calc_upll_dividers(struct radeon_device *rdev,
859 				  unsigned vclk, unsigned dclk,
860 				  unsigned vco_min, unsigned vco_max,
861 				  unsigned fb_factor, unsigned fb_mask,
862 				  unsigned pd_min, unsigned pd_max,
863 				  unsigned pd_even,
864 				  unsigned *optimal_fb_div,
865 				  unsigned *optimal_vclk_div,
866 				  unsigned *optimal_dclk_div)
867 {
868 	unsigned vco_freq, ref_freq = rdev->clock.spll.reference_freq;
869 
870 	/* start off with something large */
871 	unsigned optimal_score = ~0;
872 
873 	/* loop through vco from low to high */
874 	vco_min = max(max(vco_min, vclk), dclk);
875 	for (vco_freq = vco_min; vco_freq <= vco_max; vco_freq += 100) {
876 
877 		uint64_t fb_div = (uint64_t)vco_freq * fb_factor;
878 		unsigned vclk_div, dclk_div, score;
879 
880 		do_div(fb_div, ref_freq);
881 
882 		/* fb div out of range ? */
883 		if (fb_div > fb_mask)
884 			break; /* it can oly get worse */
885 
886 		fb_div &= fb_mask;
887 
888 		/* calc vclk divider with current vco freq */
889 		vclk_div = radeon_uvd_calc_upll_post_div(vco_freq, vclk,
890 							 pd_min, pd_even);
891 		if (vclk_div > pd_max)
892 			break; /* vco is too big, it has to stop */
893 
894 		/* calc dclk divider with current vco freq */
895 		dclk_div = radeon_uvd_calc_upll_post_div(vco_freq, dclk,
896 							 pd_min, pd_even);
897 		if (vclk_div > pd_max)
898 			break; /* vco is too big, it has to stop */
899 
900 		/* calc score with current vco freq */
901 		score = vclk - (vco_freq / vclk_div) + dclk - (vco_freq / dclk_div);
902 
903 		/* determine if this vco setting is better than current optimal settings */
904 		if (score < optimal_score) {
905 			*optimal_fb_div = fb_div;
906 			*optimal_vclk_div = vclk_div;
907 			*optimal_dclk_div = dclk_div;
908 			optimal_score = score;
909 			if (optimal_score == 0)
910 				break; /* it can't get better than this */
911 		}
912 	}
913 
914 	/* did we found a valid setup ? */
915 	if (optimal_score == ~0)
916 		return -EINVAL;
917 
918 	return 0;
919 }
920 
921 int radeon_uvd_send_upll_ctlreq(struct radeon_device *rdev,
922 				unsigned cg_upll_func_cntl)
923 {
924 	unsigned i;
925 
926 	/* make sure UPLL_CTLREQ is deasserted */
927 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
928 
929 	mdelay(10);
930 
931 	/* assert UPLL_CTLREQ */
932 	WREG32_P(cg_upll_func_cntl, UPLL_CTLREQ_MASK, ~UPLL_CTLREQ_MASK);
933 
934 	/* wait for CTLACK and CTLACK2 to get asserted */
935 	for (i = 0; i < 100; ++i) {
936 		uint32_t mask = UPLL_CTLACK_MASK | UPLL_CTLACK2_MASK;
937 		if ((RREG32(cg_upll_func_cntl) & mask) == mask)
938 			break;
939 		mdelay(10);
940 	}
941 
942 	/* deassert UPLL_CTLREQ */
943 	WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
944 
945 	if (i == 100) {
946 		DRM_ERROR("Timeout setting UVD clocks!\n");
947 		return -ETIMEDOUT;
948 	}
949 
950 	return 0;
951 }
952