drivers/gpu/drm/amd/amdgpu/amdgpu_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 <[email protected]>
  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 "amdgpu.h"
  37 #include "amdgpu_pm.h"
  38 #include "amdgpu_uvd.h"
  39 #include "cikd.h"
  40 #include "uvd/uvd_4_2_d.h"
  41
  42 /* 1 second timeout */
  43 #define UVD_IDLE_TIMEOUT_MS     1000
  44
  45 /* Firmware Names */
  46 #ifdef CONFIG_DRM_AMDGPU_CIK
  47 #define FIRMWARE_BONAIRE        "radeon/bonaire_uvd.bin"
  48 #define FIRMWARE_KABINI         "radeon/kabini_uvd.bin"
  49 #define FIRMWARE_KAVERI         "radeon/kaveri_uvd.bin"
  50 #define FIRMWARE_HAWAII         "radeon/hawaii_uvd.bin"
  51 #define FIRMWARE_MULLINS        "radeon/mullins_uvd.bin"
  52 #endif
  53 #define FIRMWARE_TONGA          "amdgpu/tonga_uvd.bin"
  54 #define FIRMWARE_CARRIZO        "amdgpu/carrizo_uvd.bin"
  55 #define FIRMWARE_FIJI           "amdgpu/fiji_uvd.bin"
  56
  57 /**
  58  * amdgpu_uvd_cs_ctx - Command submission parser context
  59  *
  60  * Used for emulating virtual memory support on UVD 4.2.
  61  */
  62 struct amdgpu_uvd_cs_ctx {
  63         struct amdgpu_cs_parser *parser;
  64         unsigned reg, count;
  65         unsigned data0, data1;
  66         unsigned idx;
  67         unsigned ib_idx;
  68
  69         /* does the IB has a msg command */
  70         bool has_msg_cmd;
  71
  72         /* minimum buffer sizes */
  73         unsigned *buf_sizes;
  74 };
  75
  76 #ifdef CONFIG_DRM_AMDGPU_CIK
  77 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
  78 MODULE_FIRMWARE(FIRMWARE_KABINI);
  79 MODULE_FIRMWARE(FIRMWARE_KAVERI);
  80 MODULE_FIRMWARE(FIRMWARE_HAWAII);
  81 MODULE_FIRMWARE(FIRMWARE_MULLINS);
  82 #endif
  83 MODULE_FIRMWARE(FIRMWARE_TONGA);
  84 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
  85 MODULE_FIRMWARE(FIRMWARE_FIJI);
  86
  87 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev);
  88 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
  89
  90 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
  91 {
  92         unsigned long bo_size;
  93         const char *fw_name;
  94         const struct common_firmware_header *hdr;
  95         unsigned version_major, version_minor, family_id;
  96         int i, r;
  97
  98         INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
  99
 100         switch (adev->asic_type) {
 101 #ifdef CONFIG_DRM_AMDGPU_CIK
 102         case CHIP_BONAIRE:
 103                 fw_name = FIRMWARE_BONAIRE;
 104                 break;
 105         case CHIP_KABINI:
 106                 fw_name = FIRMWARE_KABINI;
 107                 break;
 108         case CHIP_KAVERI:
 109                 fw_name = FIRMWARE_KAVERI;
 110                 break;
 111         case CHIP_HAWAII:
 112                 fw_name = FIRMWARE_HAWAII;
 113                 break;
 114         case CHIP_MULLINS:
 115                 fw_name = FIRMWARE_MULLINS;
 116                 break;
 117 #endif
 118         case CHIP_TONGA:
 119                 fw_name = FIRMWARE_TONGA;
 120                 break;
 121         case CHIP_FIJI:
 122                 fw_name = FIRMWARE_FIJI;
 123                 break;
 124         case CHIP_CARRIZO:
 125                 fw_name = FIRMWARE_CARRIZO;
 126                 break;
 127         default:
 128                 return -EINVAL;
 129         }
 130
 131         r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
 132         if (r) {
 133                 dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
 134                         fw_name);
 135                 return r;
 136         }
 137
 138         r = amdgpu_ucode_validate(adev->uvd.fw);
 139         if (r) {
 140                 dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
 141                         fw_name);
 142                 release_firmware(adev->uvd.fw);
 143                 adev->uvd.fw = NULL;
 144                 return r;
 145         }
 146
 147         hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 148         family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
 149         version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
 150         version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
 151         DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
 152                 version_major, version_minor, family_id);
 153
 154         bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
 155                  +  AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE;
 156         r = amdgpu_bo_create(adev, bo_size, PAGE_SIZE, true,
 157                              AMDGPU_GEM_DOMAIN_VRAM,
 158                              AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
 159                              NULL, NULL, &adev->uvd.vcpu_bo);
 160         if (r) {
 161                 dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
 162                 return r;
 163         }
 164
 165         r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
 166         if (r) {
 167                 amdgpu_bo_unref(&adev->uvd.vcpu_bo);
 168                 dev_err(adev->dev, "(%d) failed to reserve UVD bo\n", r);
 169                 return r;
 170         }
 171
 172         r = amdgpu_bo_pin(adev->uvd.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
 173                           &adev->uvd.gpu_addr);
 174         if (r) {
 175                 amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
 176                 amdgpu_bo_unref(&adev->uvd.vcpu_bo);
 177                 dev_err(adev->dev, "(%d) UVD bo pin failed\n", r);
 178                 return r;
 179         }
 180
 181         r = amdgpu_bo_kmap(adev->uvd.vcpu_bo, &adev->uvd.cpu_addr);
 182         if (r) {
 183                 dev_err(adev->dev, "(%d) UVD map failed\n", r);
 184                 return r;
 185         }
 186
 187         amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
 188
 189         for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
 190                 atomic_set(&adev->uvd.handles[i], 0);
 191                 adev->uvd.filp[i] = NULL;
 192         }
 193
 194         /* from uvd v5.0 HW addressing capacity increased to 64 bits */
 195         if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
 196                 adev->uvd.address_64_bit = true;
 197
 198         return 0;
 199 }
 200
 201 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
 202 {
 203         int r;
 204
 205         if (adev->uvd.vcpu_bo == NULL)
 206                 return 0;
 207
 208         r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
 209         if (!r) {
 210                 amdgpu_bo_kunmap(adev->uvd.vcpu_bo);
 211                 amdgpu_bo_unpin(adev->uvd.vcpu_bo);
 212                 amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
 213         }
 214
 215         amdgpu_bo_unref(&adev->uvd.vcpu_bo);
 216
 217         amdgpu_ring_fini(&adev->uvd.ring);
 218
 219         release_firmware(adev->uvd.fw);
 220
 221         return 0;
 222 }
 223
 224 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
 225 {
 226         struct amdgpu_ring *ring = &adev->uvd.ring;
 227         int i, r;
 228
 229         if (adev->uvd.vcpu_bo == NULL)
 230                 return 0;
 231
 232         for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
 233                 uint32_t handle = atomic_read(&adev->uvd.handles[i]);
 234                 if (handle != 0) {
 235                         struct fence *fence;
 236
 237                         amdgpu_uvd_note_usage(adev);
 238
 239                         r = amdgpu_uvd_get_destroy_msg(ring, handle, &fence);
 240                         if (r) {
 241                                 DRM_ERROR("Error destroying UVD (%d)!\n", r);
 242                                 continue;
 243                         }
 244
 245                         fence_wait(fence, false);
 246                         fence_put(fence);
 247
 248                         adev->uvd.filp[i] = NULL;
 249                         atomic_set(&adev->uvd.handles[i], 0);
 250                 }
 251         }
 252
 253         return 0;
 254 }
 255
 256 int amdgpu_uvd_resume(struct amdgpu_device *adev)
 257 {
 258         unsigned size;
 259         void *ptr;
 260         const struct common_firmware_header *hdr;
 261         unsigned offset;
 262
 263         if (adev->uvd.vcpu_bo == NULL)
 264                 return -EINVAL;
 265
 266         hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 267         offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
 268         memcpy(adev->uvd.cpu_addr, (adev->uvd.fw->data) + offset,
 269                 (adev->uvd.fw->size) - offset);
 270
 271         size = amdgpu_bo_size(adev->uvd.vcpu_bo);
 272         size -= le32_to_cpu(hdr->ucode_size_bytes);
 273         ptr = adev->uvd.cpu_addr;
 274         ptr += le32_to_cpu(hdr->ucode_size_bytes);
 275
 276         memset(ptr, 0, size);
 277
 278         return 0;
 279 }
 280
 281 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
 282 {
 283         struct amdgpu_ring *ring = &adev->uvd.ring;
 284         int i, r;
 285
 286         for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
 287                 uint32_t handle = atomic_read(&adev->uvd.handles[i]);
 288                 if (handle != 0 && adev->uvd.filp[i] == filp) {
 289                         struct fence *fence;
 290
 291                         amdgpu_uvd_note_usage(adev);
 292
 293                         r = amdgpu_uvd_get_destroy_msg(ring, handle, &fence);
 294                         if (r) {
 295                                 DRM_ERROR("Error destroying UVD (%d)!\n", r);
 296                                 continue;
 297                         }
 298
 299                         fence_wait(fence, false);
 300                         fence_put(fence);
 301
 302                         adev->uvd.filp[i] = NULL;
 303                         atomic_set(&adev->uvd.handles[i], 0);
 304                 }
 305         }
 306 }
 307
 308 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *rbo)
 309 {
 310         int i;
 311         for (i = 0; i < rbo->placement.num_placement; ++i) {
 312                 rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
 313                 rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
 314         }
 315 }
 316
 317 /**
 318  * amdgpu_uvd_cs_pass1 - first parsing round
 319  *
 320  * @ctx: UVD parser context
 321  *
 322  * Make sure UVD message and feedback buffers are in VRAM and
 323  * nobody is violating an 256MB boundary.
 324  */
 325 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
 326 {
 327         struct amdgpu_bo_va_mapping *mapping;
 328         struct amdgpu_bo *bo;
 329         uint32_t cmd, lo, hi;
 330         uint64_t addr;
 331         int r = 0;
 332
 333         lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
 334         hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
 335         addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
 336
 337         mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
 338         if (mapping == NULL) {
 339                 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
 340                 return -EINVAL;
 341         }
 342
 343         if (!ctx->parser->adev->uvd.address_64_bit) {
 344                 /* check if it's a message or feedback command */
 345                 cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
 346                 if (cmd == 0x0 || cmd == 0x3) {
 347                         /* yes, force it into VRAM */
 348                         uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
 349                         amdgpu_ttm_placement_from_domain(bo, domain);
 350                 }
 351                 amdgpu_uvd_force_into_uvd_segment(bo);
 352
 353                 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
 354         }
 355
 356         return r;
 357 }
 358
 359 /**
 360  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
 361  *
 362  * @msg: pointer to message structure
 363  * @buf_sizes: returned buffer sizes
 364  *
 365  * Peek into the decode message and calculate the necessary buffer sizes.
 366  */
 367 static int amdgpu_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
 368 {
 369         unsigned stream_type = msg[4];
 370         unsigned width = msg[6];
 371         unsigned height = msg[7];
 372         unsigned dpb_size = msg[9];
 373         unsigned pitch = msg[28];
 374         unsigned level = msg[57];
 375
 376         unsigned width_in_mb = width / 16;
 377         unsigned height_in_mb = ALIGN(height / 16, 2);
 378         unsigned fs_in_mb = width_in_mb * height_in_mb;
 379
 380         unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
 381         unsigned min_ctx_size = 0;
 382
 383         image_size = width * height;
 384         image_size += image_size / 2;
 385         image_size = ALIGN(image_size, 1024);
 386
 387         switch (stream_type) {
 388         case 0: /* H264 */
 389         case 7: /* H264 Perf */
 390                 switch(level) {
 391                 case 30:
 392                         num_dpb_buffer = 8100 / fs_in_mb;
 393                         break;
 394                 case 31:
 395                         num_dpb_buffer = 18000 / fs_in_mb;
 396                         break;
 397                 case 32:
 398                         num_dpb_buffer = 20480 / fs_in_mb;
 399                         break;
 400                 case 41:
 401                         num_dpb_buffer = 32768 / fs_in_mb;
 402                         break;
 403                 case 42:
 404                         num_dpb_buffer = 34816 / fs_in_mb;
 405                         break;
 406                 case 50:
 407                         num_dpb_buffer = 110400 / fs_in_mb;
 408                         break;
 409                 case 51:
 410                         num_dpb_buffer = 184320 / fs_in_mb;
 411                         break;
 412                 default:
 413                         num_dpb_buffer = 184320 / fs_in_mb;
 414                         break;
 415                 }
 416                 num_dpb_buffer++;
 417                 if (num_dpb_buffer > 17)
 418                         num_dpb_buffer = 17;
 419
 420                 /* reference picture buffer */
 421                 min_dpb_size = image_size * num_dpb_buffer;
 422
 423                 /* macroblock context buffer */
 424                 min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
 425
 426                 /* IT surface buffer */
 427                 min_dpb_size += width_in_mb * height_in_mb * 32;
 428                 break;
 429
 430         case 1: /* VC1 */
 431
 432                 /* reference picture buffer */
 433                 min_dpb_size = image_size * 3;
 434
 435                 /* CONTEXT_BUFFER */
 436                 min_dpb_size += width_in_mb * height_in_mb * 128;
 437
 438                 /* IT surface buffer */
 439                 min_dpb_size += width_in_mb * 64;
 440
 441                 /* DB surface buffer */
 442                 min_dpb_size += width_in_mb * 128;
 443
 444                 /* BP */
 445                 tmp = max(width_in_mb, height_in_mb);
 446                 min_dpb_size += ALIGN(tmp * 7 * 16, 64);
 447                 break;
 448
 449         case 3: /* MPEG2 */
 450
 451                 /* reference picture buffer */
 452                 min_dpb_size = image_size * 3;
 453                 break;
 454
 455         case 4: /* MPEG4 */
 456
 457                 /* reference picture buffer */
 458                 min_dpb_size = image_size * 3;
 459
 460                 /* CM */
 461                 min_dpb_size += width_in_mb * height_in_mb * 64;
 462
 463                 /* IT surface buffer */
 464                 min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
 465                 break;
 466
 467         case 16: /* H265 */
 468                 image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
 469                 image_size = ALIGN(image_size, 256);
 470
 471                 num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
 472                 min_dpb_size = image_size * num_dpb_buffer;
 473                 min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
 474                                            * 16 * num_dpb_buffer + 52 * 1024;
 475                 break;
 476
 477         default:
 478                 DRM_ERROR("UVD codec not handled %d!\n", stream_type);
 479                 return -EINVAL;
 480         }
 481
 482         if (width > pitch) {
 483                 DRM_ERROR("Invalid UVD decoding target pitch!\n");
 484                 return -EINVAL;
 485         }
 486
 487         if (dpb_size < min_dpb_size) {
 488                 DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
 489                           dpb_size, min_dpb_size);
 490                 return -EINVAL;
 491         }
 492
 493         buf_sizes[0x1] = dpb_size;
 494         buf_sizes[0x2] = image_size;
 495         buf_sizes[0x4] = min_ctx_size;
 496         return 0;
 497 }
 498
 499 /**
 500  * amdgpu_uvd_cs_msg - handle UVD message
 501  *
 502  * @ctx: UVD parser context
 503  * @bo: buffer object containing the message
 504  * @offset: offset into the buffer object
 505  *
 506  * Peek into the UVD message and extract the session id.
 507  * Make sure that we don't open up to many sessions.
 508  */
 509 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
 510                              struct amdgpu_bo *bo, unsigned offset)
 511 {
 512         struct amdgpu_device *adev = ctx->parser->adev;
 513         int32_t *msg, msg_type, handle;
 514         void *ptr;
 515         long r;
 516         int i;
 517
 518         if (offset & 0x3F) {
 519                 DRM_ERROR("UVD messages must be 64 byte aligned!\n");
 520                 return -EINVAL;
 521         }
 522
 523         r = reservation_object_wait_timeout_rcu(bo->tbo.resv, true, false,
 524                                                 MAX_SCHEDULE_TIMEOUT);
 525         if (r < 0) {
 526                 DRM_ERROR("Failed waiting for UVD message (%ld)!\n", r);
 527                 return r;
 528         }
 529
 530         r = amdgpu_bo_kmap(bo, &ptr);
 531         if (r) {
 532                 DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r);
 533                 return r;
 534         }
 535
 536         msg = ptr + offset;
 537
 538         msg_type = msg[1];
 539         handle = msg[2];
 540
 541         if (handle == 0) {
 542                 DRM_ERROR("Invalid UVD handle!\n");
 543                 return -EINVAL;
 544         }
 545
 546         switch (msg_type) {
 547         case 0:
 548                 /* it's a create msg, calc image size (width * height) */
 549                 amdgpu_bo_kunmap(bo);
 550
 551                 /* try to alloc a new handle */
 552                 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
 553                         if (atomic_read(&adev->uvd.handles[i]) == handle) {
 554                                 DRM_ERROR("Handle 0x%x already in use!\n", handle);
 555                                 return -EINVAL;
 556                         }
 557
 558                         if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
 559                                 adev->uvd.filp[i] = ctx->parser->filp;
 560                                 return 0;
 561                         }
 562                 }
 563
 564                 DRM_ERROR("No more free UVD handles!\n");
 565                 return -EINVAL;
 566
 567         case 1:
 568                 /* it's a decode msg, calc buffer sizes */
 569                 r = amdgpu_uvd_cs_msg_decode(msg, ctx->buf_sizes);
 570                 amdgpu_bo_kunmap(bo);
 571                 if (r)
 572                         return r;
 573
 574                 /* validate the handle */
 575                 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
 576                         if (atomic_read(&adev->uvd.handles[i]) == handle) {
 577                                 if (adev->uvd.filp[i] != ctx->parser->filp) {
 578                                         DRM_ERROR("UVD handle collision detected!\n");
 579                                         return -EINVAL;
 580                                 }
 581                                 return 0;
 582                         }
 583                 }
 584
 585                 DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
 586                 return -ENOENT;
 587
 588         case 2:
 589                 /* it's a destroy msg, free the handle */
 590                 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
 591                         atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
 592                 amdgpu_bo_kunmap(bo);
 593                 return 0;
 594
 595         default:
 596                 DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
 597                 return -EINVAL;
 598         }
 599         BUG();
 600         return -EINVAL;
 601 }
 602
 603 /**
 604  * amdgpu_uvd_cs_pass2 - second parsing round
 605  *
 606  * @ctx: UVD parser context
 607  *
 608  * Patch buffer addresses, make sure buffer sizes are correct.
 609  */
 610 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
 611 {
 612         struct amdgpu_bo_va_mapping *mapping;
 613         struct amdgpu_bo *bo;
 614         struct amdgpu_ib *ib;
 615         uint32_t cmd, lo, hi;
 616         uint64_t start, end;
 617         uint64_t addr;
 618         int r;
 619
 620         lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
 621         hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
 622         addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
 623
 624         mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
 625         if (mapping == NULL)
 626                 return -EINVAL;
 627
 628         start = amdgpu_bo_gpu_offset(bo);
 629
 630         end = (mapping->it.last + 1 - mapping->it.start);
 631         end = end * AMDGPU_GPU_PAGE_SIZE + start;
 632
 633         addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
 634         start += addr;
 635
 636         ib = &ctx->parser->ibs[ctx->ib_idx];
 637         ib->ptr[ctx->data0] = start & 0xFFFFFFFF;
 638         ib->ptr[ctx->data1] = start >> 32;
 639
 640         cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
 641         if (cmd < 0x4) {
 642                 if ((end - start) < ctx->buf_sizes[cmd]) {
 643                         DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 644                                   (unsigned)(end - start),
 645                                   ctx->buf_sizes[cmd]);
 646                         return -EINVAL;
 647                 }
 648
 649         } else if (cmd == 0x206) {
 650                 if ((end - start) < ctx->buf_sizes[4]) {
 651                         DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 652                                           (unsigned)(end - start),
 653                                           ctx->buf_sizes[4]);
 654                         return -EINVAL;
 655                 }
 656         } else if ((cmd != 0x100) && (cmd != 0x204)) {
 657                 DRM_ERROR("invalid UVD command %X!\n", cmd);
 658                 return -EINVAL;
 659         }
 660
 661         if (!ctx->parser->adev->uvd.address_64_bit) {
 662                 if ((start >> 28) != ((end - 1) >> 28)) {
 663                         DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
 664                                   start, end);
 665                         return -EINVAL;
 666                 }
 667
 668                 if ((cmd == 0 || cmd == 0x3) &&
 669                     (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
 670                         DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
 671                                   start, end);
 672                         return -EINVAL;
 673                 }
 674         }
 675
 676         if (cmd == 0) {
 677                 ctx->has_msg_cmd = true;
 678                 r = amdgpu_uvd_cs_msg(ctx, bo, addr);
 679                 if (r)
 680                         return r;
 681         } else if (!ctx->has_msg_cmd) {
 682                 DRM_ERROR("Message needed before other commands are send!\n");
 683                 return -EINVAL;
 684         }
 685
 686         return 0;
 687 }
 688
 689 /**
 690  * amdgpu_uvd_cs_reg - parse register writes
 691  *
 692  * @ctx: UVD parser context
 693  * @cb: callback function
 694  *
 695  * Parse the register writes, call cb on each complete command.
 696  */
 697 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
 698                              int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
 699 {
 700         struct amdgpu_ib *ib = &ctx->parser->ibs[ctx->ib_idx];
 701         int i, r;
 702
 703         ctx->idx++;
 704         for (i = 0; i <= ctx->count; ++i) {
 705                 unsigned reg = ctx->reg + i;
 706
 707                 if (ctx->idx >= ib->length_dw) {
 708                         DRM_ERROR("Register command after end of CS!\n");
 709                         return -EINVAL;
 710                 }
 711
 712                 switch (reg) {
 713                 case mmUVD_GPCOM_VCPU_DATA0:
 714                         ctx->data0 = ctx->idx;
 715                         break;
 716                 case mmUVD_GPCOM_VCPU_DATA1:
 717                         ctx->data1 = ctx->idx;
 718                         break;
 719                 case mmUVD_GPCOM_VCPU_CMD:
 720                         r = cb(ctx);
 721                         if (r)
 722                                 return r;
 723                         break;
 724                 case mmUVD_ENGINE_CNTL:
 725                         break;
 726                 default:
 727                         DRM_ERROR("Invalid reg 0x%X!\n", reg);
 728                         return -EINVAL;
 729                 }
 730                 ctx->idx++;
 731         }
 732         return 0;
 733 }
 734
 735 /**
 736  * amdgpu_uvd_cs_packets - parse UVD packets
 737  *
 738  * @ctx: UVD parser context
 739  * @cb: callback function
 740  *
 741  * Parse the command stream packets.
 742  */
 743 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
 744                                  int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
 745 {
 746         struct amdgpu_ib *ib = &ctx->parser->ibs[ctx->ib_idx];
 747         int r;
 748
 749         for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
 750                 uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
 751                 unsigned type = CP_PACKET_GET_TYPE(cmd);
 752                 switch (type) {
 753                 case PACKET_TYPE0:
 754                         ctx->reg = CP_PACKET0_GET_REG(cmd);
 755                         ctx->count = CP_PACKET_GET_COUNT(cmd);
 756                         r = amdgpu_uvd_cs_reg(ctx, cb);
 757                         if (r)
 758                                 return r;
 759                         break;
 760                 case PACKET_TYPE2:
 761                         ++ctx->idx;
 762                         break;
 763                 default:
 764                         DRM_ERROR("Unknown packet type %d !\n", type);
 765                         return -EINVAL;
 766                 }
 767         }
 768         return 0;
 769 }
 770
 771 /**
 772  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
 773  *
 774  * @parser: Command submission parser context
 775  *
 776  * Parse the command stream, patch in addresses as necessary.
 777  */
 778 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
 779 {
 780         struct amdgpu_uvd_cs_ctx ctx = {};
 781         unsigned buf_sizes[] = {
 782                 [0x00000000]    =       2048,
 783                 [0x00000001]    =       0xFFFFFFFF,
 784                 [0x00000002]    =       0xFFFFFFFF,
 785                 [0x00000003]    =       2048,
 786                 [0x00000004]    =       0xFFFFFFFF,
 787         };
 788         struct amdgpu_ib *ib = &parser->ibs[ib_idx];
 789         int r;
 790
 791         if (ib->length_dw % 16) {
 792                 DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
 793                           ib->length_dw);
 794                 return -EINVAL;
 795         }
 796
 797         ctx.parser = parser;
 798         ctx.buf_sizes = buf_sizes;
 799         ctx.ib_idx = ib_idx;
 800
 801         /* first round, make sure the buffers are actually in the UVD segment */
 802         r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
 803         if (r)
 804                 return r;
 805
 806         /* second round, patch buffer addresses into the command stream */
 807         r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
 808         if (r)
 809                 return r;
 810
 811         if (!ctx.has_msg_cmd) {
 812                 DRM_ERROR("UVD-IBs need a msg command!\n");
 813                 return -EINVAL;
 814         }
 815
 816         amdgpu_uvd_note_usage(ctx.parser->adev);
 817
 818         return 0;
 819 }
 820
 821 static int amdgpu_uvd_free_job(
 822         struct amdgpu_job *job)
 823 {
 824         amdgpu_ib_free(job->adev, job->ibs);
 825         kfree(job->ibs);
 826         return 0;
 827 }
 828
 829 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring,
 830                                struct amdgpu_bo *bo,
 831                                struct fence **fence)
 832 {
 833         struct ttm_validate_buffer tv;
 834         struct ww_acquire_ctx ticket;
 835         struct list_head head;
 836         struct amdgpu_ib *ib = NULL;
 837         struct fence *f = NULL;
 838         struct amdgpu_device *adev = ring->adev;
 839         uint64_t addr;
 840         int i, r;
 841
 842         memset(&tv, 0, sizeof(tv));
 843         tv.bo = &bo->tbo;
 844
 845         INIT_LIST_HEAD(&head);
 846         list_add(&tv.head, &head);
 847
 848         r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
 849         if (r)
 850                 return r;
 851
 852         if (!bo->adev->uvd.address_64_bit) {
 853                 amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
 854                 amdgpu_uvd_force_into_uvd_segment(bo);
 855         }
 856
 857         r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
 858         if (r)
 859                 goto err;
 860         ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
 861         if (!ib) {
 862                 r = -ENOMEM;
 863                 goto err;
 864         }
 865         r = amdgpu_ib_get(ring, NULL, 64, ib);
 866         if (r)
 867                 goto err1;
 868
 869         addr = amdgpu_bo_gpu_offset(bo);
 870         ib->ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
 871         ib->ptr[1] = addr;
 872         ib->ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
 873         ib->ptr[3] = addr >> 32;
 874         ib->ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
 875         ib->ptr[5] = 0;
 876         for (i = 6; i < 16; ++i)
 877                 ib->ptr[i] = PACKET2(0);
 878         ib->length_dw = 16;
 879
 880         r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
 881                                                  &amdgpu_uvd_free_job,
 882                                                  AMDGPU_FENCE_OWNER_UNDEFINED,
 883                                                  &f);
 884         if (r)
 885                 goto err2;
 886
 887         ttm_eu_fence_buffer_objects(&ticket, &head, f);
 888
 889         if (fence)
 890                 *fence = fence_get(f);
 891         amdgpu_bo_unref(&bo);
 892         fence_put(f);
 893         if (amdgpu_enable_scheduler)
 894                 return 0;
 895
 896         amdgpu_ib_free(ring->adev, ib);
 897         kfree(ib);
 898         return 0;
 899 err2:
 900         amdgpu_ib_free(ring->adev, ib);
 901 err1:
 902         kfree(ib);
 903 err:
 904         ttm_eu_backoff_reservation(&ticket, &head);
 905         return r;
 906 }
 907
 908 /* multiple fence commands without any stream commands in between can
 909    crash the vcpu so just try to emmit a dummy create/destroy msg to
 910    avoid this */
 911 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
 912                               struct fence **fence)
 913 {
 914         struct amdgpu_device *adev = ring->adev;
 915         struct amdgpu_bo *bo;
 916         uint32_t *msg;
 917         int r, i;
 918
 919         r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
 920                              AMDGPU_GEM_DOMAIN_VRAM,
 921                              AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
 922                              NULL, NULL, &bo);
 923         if (r)
 924                 return r;
 925
 926         r = amdgpu_bo_reserve(bo, false);
 927         if (r) {
 928                 amdgpu_bo_unref(&bo);
 929                 return r;
 930         }
 931
 932         r = amdgpu_bo_kmap(bo, (void **)&msg);
 933         if (r) {
 934                 amdgpu_bo_unreserve(bo);
 935                 amdgpu_bo_unref(&bo);
 936                 return r;
 937         }
 938
 939         /* stitch together an UVD create msg */
 940         msg[0] = cpu_to_le32(0x00000de4);
 941         msg[1] = cpu_to_le32(0x00000000);
 942         msg[2] = cpu_to_le32(handle);
 943         msg[3] = cpu_to_le32(0x00000000);
 944         msg[4] = cpu_to_le32(0x00000000);
 945         msg[5] = cpu_to_le32(0x00000000);
 946         msg[6] = cpu_to_le32(0x00000000);
 947         msg[7] = cpu_to_le32(0x00000780);
 948         msg[8] = cpu_to_le32(0x00000440);
 949         msg[9] = cpu_to_le32(0x00000000);
 950         msg[10] = cpu_to_le32(0x01b37000);
 951         for (i = 11; i < 1024; ++i)
 952                 msg[i] = cpu_to_le32(0x0);
 953
 954         amdgpu_bo_kunmap(bo);
 955         amdgpu_bo_unreserve(bo);
 956
 957         return amdgpu_uvd_send_msg(ring, bo, fence);
 958 }
 959
 960 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
 961                                struct fence **fence)
 962 {
 963         struct amdgpu_device *adev = ring->adev;
 964         struct amdgpu_bo *bo;
 965         uint32_t *msg;
 966         int r, i;
 967
 968         r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
 969                              AMDGPU_GEM_DOMAIN_VRAM,
 970                              AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
 971                              NULL, NULL, &bo);
 972         if (r)
 973                 return r;
 974
 975         r = amdgpu_bo_reserve(bo, false);
 976         if (r) {
 977                 amdgpu_bo_unref(&bo);
 978                 return r;
 979         }
 980
 981         r = amdgpu_bo_kmap(bo, (void **)&msg);
 982         if (r) {
 983                 amdgpu_bo_unreserve(bo);
 984                 amdgpu_bo_unref(&bo);
 985                 return r;
 986         }
 987
 988         /* stitch together an UVD destroy msg */
 989         msg[0] = cpu_to_le32(0x00000de4);
 990         msg[1] = cpu_to_le32(0x00000002);
 991         msg[2] = cpu_to_le32(handle);
 992         msg[3] = cpu_to_le32(0x00000000);
 993         for (i = 4; i < 1024; ++i)
 994                 msg[i] = cpu_to_le32(0x0);
 995
 996         amdgpu_bo_kunmap(bo);
 997         amdgpu_bo_unreserve(bo);
 998
 999         return amdgpu_uvd_send_msg(ring, bo, fence);
1000 }
1001
1002 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1003 {
1004         struct amdgpu_device *adev =
1005                 container_of(work, struct amdgpu_device, uvd.idle_work.work);
1006         unsigned i, fences, handles = 0;
1007
1008         fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
1009
1010         for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
1011                 if (atomic_read(&adev->uvd.handles[i]))
1012                         ++handles;
1013
1014         if (fences == 0 && handles == 0) {
1015                 if (adev->pm.dpm_enabled) {
1016                         amdgpu_dpm_enable_uvd(adev, false);
1017                 } else {
1018                         amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1019                 }
1020         } else {
1021                 schedule_delayed_work(&adev->uvd.idle_work,
1022                                       msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1023         }
1024 }
1025
1026 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev)
1027 {
1028         bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1029         set_clocks &= schedule_delayed_work(&adev->uvd.idle_work,
1030                                             msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1031
1032         if (set_clocks) {
1033                 if (adev->pm.dpm_enabled) {
1034                         amdgpu_dpm_enable_uvd(adev, true);
1035                 } else {
1036                         amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1037                 }
1038         }
1039 }