sys/dev/drm/radeon/radeon_uvd.c

   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
  35 #include "radeon.h"
  36 #include "r600d.h"
  37
  38 /* 1 second timeout */
  39 #define UVD_IDLE_TIMEOUT_MS     1000
  40
  41 /* Firmware Names */
  42 #define FIRMWARE_RV710          "radeonkmsfw_RV710_uvd"
  43 #define FIRMWARE_CYPRESS        "radeonkmsfw_CYPRESS_uvd"
  44 #define FIRMWARE_SUMO           "radeonkmsfw_SUMO_uvd"
  45 #define FIRMWARE_TAHITI         "radeonkmsfw_TAHITI_uvd"
  46 #define FIRMWARE_BONAIRE        "radeonkmsfw_BONAIRE_uvd"
  47
  48 MODULE_FIRMWARE(FIRMWARE_RV710);
  49 MODULE_FIRMWARE(FIRMWARE_CYPRESS);
  50 MODULE_FIRMWARE(FIRMWARE_SUMO);
  51 MODULE_FIRMWARE(FIRMWARE_TAHITI);
  52 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
  53
  54 static void radeon_uvd_idle_work_handler(struct work_struct *work);
  55
  56 int radeon_uvd_init(struct radeon_device *rdev)
  57 {
  58         unsigned long bo_size;
  59         const char *fw_name;
  60         int i, r;
  61
  62         INIT_DELAYED_WORK(&rdev->uvd.idle_work, radeon_uvd_idle_work_handler);
  63
  64         switch (rdev->family) {
  65         case CHIP_RV710:
  66         case CHIP_RV730:
  67         case CHIP_RV740:
  68                 fw_name = FIRMWARE_RV710;
  69                 break;
  70
  71         case CHIP_CYPRESS:
  72         case CHIP_HEMLOCK:
  73         case CHIP_JUNIPER:
  74         case CHIP_REDWOOD:
  75         case CHIP_CEDAR:
  76                 fw_name = FIRMWARE_CYPRESS;
  77                 break;
  78
  79         case CHIP_SUMO:
  80         case CHIP_SUMO2:
  81         case CHIP_PALM:
  82         case CHIP_CAYMAN:
  83         case CHIP_BARTS:
  84         case CHIP_TURKS:
  85         case CHIP_CAICOS:
  86                 fw_name = FIRMWARE_SUMO;
  87                 break;
  88
  89         case CHIP_TAHITI:
  90         case CHIP_VERDE:
  91         case CHIP_PITCAIRN:
  92         case CHIP_ARUBA:
  93         case CHIP_OLAND:
  94                 fw_name = FIRMWARE_TAHITI;
  95                 break;
  96
  97         case CHIP_BONAIRE:
  98         case CHIP_KABINI:
  99         case CHIP_KAVERI:
 100         case CHIP_HAWAII:
 101         case CHIP_MULLINS:
 102                 fw_name = FIRMWARE_BONAIRE;
 103                 break;
 104
 105         default:
 106                 return -EINVAL;
 107         }
 108
 109         r = request_firmware(&rdev->uvd_fw, fw_name, rdev->dev);
 110         if (r) {
 111                 dev_err(rdev->dev, "radeon_uvd: Can't load firmware \"%s\"\n",
 112                         fw_name);
 113                 return r;
 114         }
 115
 116         bo_size = RADEON_GPU_PAGE_ALIGN(rdev->uvd_fw->datasize + 8) +
 117                   RADEON_UVD_STACK_SIZE + RADEON_UVD_HEAP_SIZE +
 118                   RADEON_GPU_PAGE_SIZE;
 119         r = radeon_bo_create(rdev, bo_size, PAGE_SIZE, true,
 120                              RADEON_GEM_DOMAIN_VRAM, 0, 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
 159 void 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
 180 int radeon_uvd_suspend(struct radeon_device *rdev)
 181 {
 182         unsigned size;
 183         char *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->datasize;
 198
 199         ptr = rdev->uvd.cpu_addr;
 200         ptr += rdev->uvd_fw->datasize;
 201
 202         rdev->uvd.saved_bo = kmalloc(size, M_DRM, M_WAITOK);
 203         memcpy(rdev->uvd.saved_bo, ptr, size);
 204
 205         return 0;
 206 }
 207
 208 int radeon_uvd_resume(struct radeon_device *rdev)
 209 {
 210         unsigned size;
 211         char *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->datasize);
 217
 218         size = radeon_bo_size(rdev->uvd.vcpu_bo);
 219         size -= rdev->uvd_fw->datasize;
 220
 221         ptr = rdev->uvd.cpu_addr;
 222         ptr += rdev->uvd_fw->datasize;
 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
 234 void 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
 240 void 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
 266 static 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
 354 static 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 = (uint32_t*)((uint8_t*)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_cmpset(&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
 439 static 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
 515 static 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
 550 int 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
 606 static int radeon_uvd_send_msg(struct radeon_device *rdev,
 607                                int ring, uint64_t addr,
 608                                struct radeon_fence **fence)
 609 {
 610         struct radeon_ib ib;
 611         int i, r;
 612
 613         r = radeon_ib_get(rdev, ring, &ib, NULL, 64);
 614         if (r)
 615                 return r;
 616
 617         ib.ptr[0] = PACKET0(UVD_GPCOM_VCPU_DATA0, 0);
 618         ib.ptr[1] = addr;
 619         ib.ptr[2] = PACKET0(UVD_GPCOM_VCPU_DATA1, 0);
 620         ib.ptr[3] = addr >> 32;
 621         ib.ptr[4] = PACKET0(UVD_GPCOM_VCPU_CMD, 0);
 622         ib.ptr[5] = 0;
 623         for (i = 6; i < 16; ++i)
 624                 ib.ptr[i] = PACKET2(0);
 625         ib.length_dw = 16;
 626
 627         r = radeon_ib_schedule(rdev, &ib, NULL, false);
 628
 629         if (fence)
 630                 *fence = radeon_fence_ref(ib.fence);
 631
 632         radeon_ib_free(rdev, &ib);
 633         return r;
 634 }
 635
 636 /* multiple fence commands without any stream commands in between can
 637    crash the vcpu so just try to emmit a dummy create/destroy msg to
 638    avoid this */
 639 int radeon_uvd_get_create_msg(struct radeon_device *rdev, int ring,
 640                               uint32_t handle, struct radeon_fence **fence)
 641 {
 642         /* we use the last page of the vcpu bo for the UVD message */
 643         uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
 644                 RADEON_GPU_PAGE_SIZE;
 645
 646         uint32_t *msg = (uint32_t*)((uint8_t*)rdev->uvd.cpu_addr + offs);
 647         uint64_t addr = rdev->uvd.gpu_addr + offs;
 648
 649         int r, i;
 650
 651         r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
 652         if (r)
 653                 return r;
 654
 655         /* stitch together an UVD create msg */
 656         msg[0] = cpu_to_le32(0x00000de4);
 657         msg[1] = cpu_to_le32(0x00000000);
 658         msg[2] = cpu_to_le32(handle);
 659         msg[3] = cpu_to_le32(0x00000000);
 660         msg[4] = cpu_to_le32(0x00000000);
 661         msg[5] = cpu_to_le32(0x00000000);
 662         msg[6] = cpu_to_le32(0x00000000);
 663         msg[7] = cpu_to_le32(0x00000780);
 664         msg[8] = cpu_to_le32(0x00000440);
 665         msg[9] = cpu_to_le32(0x00000000);
 666         msg[10] = cpu_to_le32(0x01b37000);
 667         for (i = 11; i < 1024; ++i)
 668                 msg[i] = cpu_to_le32(0x0);
 669
 670         r = radeon_uvd_send_msg(rdev, ring, addr, fence);
 671         radeon_bo_unreserve(rdev->uvd.vcpu_bo);
 672         return r;
 673 }
 674
 675 int radeon_uvd_get_destroy_msg(struct radeon_device *rdev, int ring,
 676                                uint32_t handle, struct radeon_fence **fence)
 677 {
 678         /* we use the last page of the vcpu bo for the UVD message */
 679         uint64_t offs = radeon_bo_size(rdev->uvd.vcpu_bo) -
 680                 RADEON_GPU_PAGE_SIZE;
 681
 682         uint32_t *msg = (uint32_t*)((uint8_t*)rdev->uvd.cpu_addr + offs);
 683         uint64_t addr = rdev->uvd.gpu_addr + offs;
 684
 685         int r, i;
 686
 687         r = radeon_bo_reserve(rdev->uvd.vcpu_bo, true);
 688         if (r)
 689                 return r;
 690
 691         /* stitch together an UVD destroy msg */
 692         msg[0] = cpu_to_le32(0x00000de4);
 693         msg[1] = cpu_to_le32(0x00000002);
 694         msg[2] = cpu_to_le32(handle);
 695         msg[3] = cpu_to_le32(0x00000000);
 696         for (i = 4; i < 1024; ++i)
 697                 msg[i] = cpu_to_le32(0x0);
 698
 699         r = radeon_uvd_send_msg(rdev, ring, addr, fence);
 700         radeon_bo_unreserve(rdev->uvd.vcpu_bo);
 701         return r;
 702 }
 703
 704 /**
 705  * radeon_uvd_count_handles - count number of open streams
 706  *
 707  * @rdev: radeon_device pointer
 708  * @sd: number of SD streams
 709  * @hd: number of HD streams
 710  *
 711  * Count the number of open SD/HD streams as a hint for power mangement
 712  */
 713 static void radeon_uvd_count_handles(struct radeon_device *rdev,
 714                                      unsigned *sd, unsigned *hd)
 715 {
 716         unsigned i;
 717
 718         *sd = 0;
 719         *hd = 0;
 720
 721         for (i = 0; i < RADEON_MAX_UVD_HANDLES; ++i) {
 722                 if (!atomic_read(&rdev->uvd.handles[i]))
 723                         continue;
 724
 725                 if (rdev->uvd.img_size[i] >= 720*576)
 726                         ++(*hd);
 727                 else
 728                         ++(*sd);
 729         }
 730 }
 731
 732 static void radeon_uvd_idle_work_handler(struct work_struct *work)
 733 {
 734         struct radeon_device *rdev =
 735                 container_of(work, struct radeon_device, uvd.idle_work.work);
 736
 737         if (radeon_fence_count_emitted(rdev, R600_RING_TYPE_UVD_INDEX) == 0) {
 738                 if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
 739                         radeon_uvd_count_handles(rdev, &rdev->pm.dpm.sd,
 740                                                  &rdev->pm.dpm.hd);
 741                         radeon_dpm_enable_uvd(rdev, false);
 742                 } else {
 743                         radeon_set_uvd_clocks(rdev, 0, 0);
 744                 }
 745         } else {
 746                 schedule_delayed_work(&rdev->uvd.idle_work,
 747                                       msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
 748         }
 749 }
 750
 751 void radeon_uvd_note_usage(struct radeon_device *rdev)
 752 {
 753         bool streams_changed = false;
 754         bool set_clocks = !cancel_delayed_work_sync(&rdev->uvd.idle_work);
 755         set_clocks &= schedule_delayed_work(&rdev->uvd.idle_work,
 756                                             msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
 757
 758         if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
 759                 unsigned hd = 0, sd = 0;
 760                 radeon_uvd_count_handles(rdev, &sd, &hd);
 761                 if ((rdev->pm.dpm.sd != sd) ||
 762                     (rdev->pm.dpm.hd != hd)) {
 763                         rdev->pm.dpm.sd = sd;
 764                         rdev->pm.dpm.hd = hd;
 765                         /* disable this for now */
 766                         /*streams_changed = true;*/
 767                 }
 768         }
 769
 770         if (set_clocks || streams_changed) {
 771                 if ((rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled) {
 772                         radeon_dpm_enable_uvd(rdev, true);
 773                 } else {
 774                         radeon_set_uvd_clocks(rdev, 53300, 40000);
 775                 }
 776         }
 777 }
 778
 779 static unsigned radeon_uvd_calc_upll_post_div(unsigned vco_freq,
 780                                               unsigned target_freq,
 781                                               unsigned pd_min,
 782                                               unsigned pd_even)
 783 {
 784         unsigned post_div = vco_freq / target_freq;
 785
 786         /* adjust to post divider minimum value */
 787         if (post_div < pd_min)
 788                 post_div = pd_min;
 789
 790         /* we alway need a frequency less than or equal the target */
 791         if ((vco_freq / post_div) > target_freq)
 792                 post_div += 1;
 793
 794         /* post dividers above a certain value must be even */
 795         if (post_div > pd_even && post_div % 2)
 796                 post_div += 1;
 797
 798         return post_div;
 799 }
 800
 801 /**
 802  * radeon_uvd_calc_upll_dividers - calc UPLL clock dividers
 803  *
 804  * @rdev: radeon_device pointer
 805  * @vclk: wanted VCLK
 806  * @dclk: wanted DCLK
 807  * @vco_min: minimum VCO frequency
 808  * @vco_max: maximum VCO frequency
 809  * @fb_factor: factor to multiply vco freq with
 810  * @fb_mask: limit and bitmask for feedback divider
 811  * @pd_min: post divider minimum
 812  * @pd_max: post divider maximum
 813  * @pd_even: post divider must be even above this value
 814  * @optimal_fb_div: resulting feedback divider
 815  * @optimal_vclk_div: resulting vclk post divider
 816  * @optimal_dclk_div: resulting dclk post divider
 817  *
 818  * Calculate dividers for UVDs UPLL (R6xx-SI, except APUs).
 819  * Returns zero on success -EINVAL on error.
 820  */
 821 int radeon_uvd_calc_upll_dividers(struct radeon_device *rdev,
 822                                   unsigned vclk, unsigned dclk,
 823                                   unsigned vco_min, unsigned vco_max,
 824                                   unsigned fb_factor, unsigned fb_mask,
 825                                   unsigned pd_min, unsigned pd_max,
 826                                   unsigned pd_even,
 827                                   unsigned *optimal_fb_div,
 828                                   unsigned *optimal_vclk_div,
 829                                   unsigned *optimal_dclk_div)
 830 {
 831         unsigned vco_freq, ref_freq = rdev->clock.spll.reference_freq;
 832
 833         /* start off with something large */
 834         unsigned optimal_score = ~0;
 835
 836         /* loop through vco from low to high */
 837         vco_min = max(max(vco_min, vclk), dclk);
 838         for (vco_freq = vco_min; vco_freq <= vco_max; vco_freq += 100) {
 839
 840                 uint64_t fb_div = (uint64_t)vco_freq * fb_factor;
 841                 unsigned vclk_div, dclk_div, score;
 842
 843                 do_div(fb_div, ref_freq);
 844
 845                 /* fb div out of range ? */
 846                 if (fb_div > fb_mask)
 847                         break; /* it can oly get worse */
 848
 849                 fb_div &= fb_mask;
 850
 851                 /* calc vclk divider with current vco freq */
 852                 vclk_div = radeon_uvd_calc_upll_post_div(vco_freq, vclk,
 853                                                          pd_min, pd_even);
 854                 if (vclk_div > pd_max)
 855                         break; /* vco is too big, it has to stop */
 856
 857                 /* calc dclk divider with current vco freq */
 858                 dclk_div = radeon_uvd_calc_upll_post_div(vco_freq, dclk,
 859                                                          pd_min, pd_even);
 860                 if (vclk_div > pd_max)
 861                         break; /* vco is too big, it has to stop */
 862
 863                 /* calc score with current vco freq */
 864                 score = vclk - (vco_freq / vclk_div) + dclk - (vco_freq / dclk_div);
 865
 866                 /* determine if this vco setting is better than current optimal settings */
 867                 if (score < optimal_score) {
 868                         *optimal_fb_div = fb_div;
 869                         *optimal_vclk_div = vclk_div;
 870                         *optimal_dclk_div = dclk_div;
 871                         optimal_score = score;
 872                         if (optimal_score == 0)
 873                                 break; /* it can't get better than this */
 874                 }
 875         }
 876
 877         /* did we found a valid setup ? */
 878         if (optimal_score == ~0)
 879                 return -EINVAL;
 880
 881         return 0;
 882 }
 883
 884 int radeon_uvd_send_upll_ctlreq(struct radeon_device *rdev,
 885                                 unsigned cg_upll_func_cntl)
 886 {
 887         unsigned i;
 888
 889         /* make sure UPLL_CTLREQ is deasserted */
 890         WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
 891
 892         mdelay(10);
 893
 894         /* assert UPLL_CTLREQ */
 895         WREG32_P(cg_upll_func_cntl, UPLL_CTLREQ_MASK, ~UPLL_CTLREQ_MASK);
 896
 897         /* wait for CTLACK and CTLACK2 to get asserted */
 898         for (i = 0; i < 100; ++i) {
 899                 uint32_t mask = UPLL_CTLACK_MASK | UPLL_CTLACK2_MASK;
 900                 if ((RREG32(cg_upll_func_cntl) & mask) == mask)
 901                         break;
 902                 mdelay(10);
 903         }
 904
 905         /* deassert UPLL_CTLREQ */
 906         WREG32_P(cg_upll_func_cntl, 0, ~UPLL_CTLREQ_MASK);
 907
 908         if (i == 100) {
 909                 DRM_ERROR("Timeout setting UVD clocks!\n");
 910                 return -ETIMEDOUT;
 911         }
 912
 913         return 0;
 914 }