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