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 <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 "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        msecs_to_jiffies(1000)
  44
  45 /* Firmware versions for VI */
  46 #define FW_1_65_10      ((1 << 24) | (65 << 16) | (10 << 8))
  47 #define FW_1_87_11      ((1 << 24) | (87 << 16) | (11 << 8))
  48 #define FW_1_87_12      ((1 << 24) | (87 << 16) | (12 << 8))
  49 #define FW_1_37_15      ((1 << 24) | (37 << 16) | (15 << 8))
  50
  51 /* Polaris10/11 firmware version */
  52 #define FW_1_66_16      ((1 << 24) | (66 << 16) | (16 << 8))
  53
  54 /* Firmware Names */
  55 #ifdef CONFIG_DRM_AMDGPU_CIK
  56 #define FIRMWARE_BONAIRE        "amdgpu/bonaire_uvd.bin"
  57 #define FIRMWARE_KABINI "amdgpu/kabini_uvd.bin"
  58 #define FIRMWARE_KAVERI "amdgpu/kaveri_uvd.bin"
  59 #define FIRMWARE_HAWAII "amdgpu/hawaii_uvd.bin"
  60 #define FIRMWARE_MULLINS        "amdgpu/mullins_uvd.bin"
  61 #endif
  62 #define FIRMWARE_TONGA          "amdgpu/tonga_uvd.bin"
  63 #define FIRMWARE_CARRIZO        "amdgpu/carrizo_uvd.bin"
  64 #define FIRMWARE_FIJI           "amdgpu/fiji_uvd.bin"
  65 #define FIRMWARE_STONEY         "amdgpu/stoney_uvd.bin"
  66 #define FIRMWARE_POLARIS10      "amdgpu/polaris10_uvd.bin"
  67 #define FIRMWARE_POLARIS11      "amdgpu/polaris11_uvd.bin"
  68 #define FIRMWARE_POLARIS12      "amdgpu/polaris12_uvd.bin"
  69 #define FIRMWARE_VEGAM          "amdgpu/vegam_uvd.bin"
  70
  71 #define FIRMWARE_VEGA10         "amdgpu/vega10_uvd.bin"
  72 #define FIRMWARE_VEGA12         "amdgpu/vega12_uvd.bin"
  73 #define FIRMWARE_VEGA20         "amdgpu/vega20_uvd.bin"
  74
  75 /* These are common relative offsets for all asics, from uvd_7_0_offset.h,  */
  76 #define UVD_GPCOM_VCPU_CMD              0x03c3
  77 #define UVD_GPCOM_VCPU_DATA0    0x03c4
  78 #define UVD_GPCOM_VCPU_DATA1    0x03c5
  79 #define UVD_NO_OP                               0x03ff
  80 #define UVD_BASE_SI                             0x3800
  81
  82 /**
  83  * amdgpu_uvd_cs_ctx - Command submission parser context
  84  *
  85  * Used for emulating virtual memory support on UVD 4.2.
  86  */
  87 struct amdgpu_uvd_cs_ctx {
  88         struct amdgpu_cs_parser *parser;
  89         unsigned reg, count;
  90         unsigned data0, data1;
  91         unsigned idx;
  92         unsigned ib_idx;
  93
  94         /* does the IB has a msg command */
  95         bool has_msg_cmd;
  96
  97         /* minimum buffer sizes */
  98         unsigned *buf_sizes;
  99 };
 100
 101 #ifdef CONFIG_DRM_AMDGPU_CIK
 102 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
 103 MODULE_FIRMWARE(FIRMWARE_KABINI);
 104 MODULE_FIRMWARE(FIRMWARE_KAVERI);
 105 MODULE_FIRMWARE(FIRMWARE_HAWAII);
 106 MODULE_FIRMWARE(FIRMWARE_MULLINS);
 107 #endif
 108 MODULE_FIRMWARE(FIRMWARE_TONGA);
 109 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
 110 MODULE_FIRMWARE(FIRMWARE_FIJI);
 111 MODULE_FIRMWARE(FIRMWARE_STONEY);
 112 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
 113 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
 114 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
 115 MODULE_FIRMWARE(FIRMWARE_VEGAM);
 116
 117 MODULE_FIRMWARE(FIRMWARE_VEGA10);
 118 MODULE_FIRMWARE(FIRMWARE_VEGA12);
 119 MODULE_FIRMWARE(FIRMWARE_VEGA20);
 120
 121 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
 122
 123 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
 124 {
 125         unsigned long bo_size;
 126         const char *fw_name;
 127         const struct common_firmware_header *hdr;
 128         unsigned family_id;
 129         int i, j, r;
 130
 131         INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
 132
 133         switch (adev->asic_type) {
 134 #ifdef CONFIG_DRM_AMDGPU_CIK
 135         case CHIP_BONAIRE:
 136                 fw_name = FIRMWARE_BONAIRE;
 137                 break;
 138         case CHIP_KABINI:
 139                 fw_name = FIRMWARE_KABINI;
 140                 break;
 141         case CHIP_KAVERI:
 142                 fw_name = FIRMWARE_KAVERI;
 143                 break;
 144         case CHIP_HAWAII:
 145                 fw_name = FIRMWARE_HAWAII;
 146                 break;
 147         case CHIP_MULLINS:
 148                 fw_name = FIRMWARE_MULLINS;
 149                 break;
 150 #endif
 151         case CHIP_TONGA:
 152                 fw_name = FIRMWARE_TONGA;
 153                 break;
 154         case CHIP_FIJI:
 155                 fw_name = FIRMWARE_FIJI;
 156                 break;
 157         case CHIP_CARRIZO:
 158                 fw_name = FIRMWARE_CARRIZO;
 159                 break;
 160         case CHIP_STONEY:
 161                 fw_name = FIRMWARE_STONEY;
 162                 break;
 163         case CHIP_POLARIS10:
 164                 fw_name = FIRMWARE_POLARIS10;
 165                 break;
 166         case CHIP_POLARIS11:
 167                 fw_name = FIRMWARE_POLARIS11;
 168                 break;
 169         case CHIP_POLARIS12:
 170                 fw_name = FIRMWARE_POLARIS12;
 171                 break;
 172         case CHIP_VEGA10:
 173                 fw_name = FIRMWARE_VEGA10;
 174                 break;
 175         case CHIP_VEGA12:
 176                 fw_name = FIRMWARE_VEGA12;
 177                 break;
 178         case CHIP_VEGAM:
 179                 fw_name = FIRMWARE_VEGAM;
 180                 break;
 181         case CHIP_VEGA20:
 182                 fw_name = FIRMWARE_VEGA20;
 183                 break;
 184         default:
 185                 return -EINVAL;
 186         }
 187
 188         r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
 189         if (r) {
 190                 dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
 191                         fw_name);
 192                 return r;
 193         }
 194
 195         r = amdgpu_ucode_validate(adev->uvd.fw);
 196         if (r) {
 197                 dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
 198                         fw_name);
 199                 release_firmware(adev->uvd.fw);
 200                 adev->uvd.fw = NULL;
 201                 return r;
 202         }
 203
 204         /* Set the default UVD handles that the firmware can handle */
 205         adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
 206
 207         hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 208         family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
 209
 210         if (adev->asic_type < CHIP_VEGA20) {
 211                 unsigned version_major, version_minor;
 212
 213                 version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
 214                 version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
 215                 DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
 216                         version_major, version_minor, family_id);
 217
 218                 /*
 219                  * Limit the number of UVD handles depending on microcode major
 220                  * and minor versions. The firmware version which has 40 UVD
 221                  * instances support is 1.80. So all subsequent versions should
 222                  * also have the same support.
 223                  */
 224                 if ((version_major > 0x01) ||
 225                     ((version_major == 0x01) && (version_minor >= 0x50)))
 226                         adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
 227
 228                 adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
 229                                         (family_id << 8));
 230
 231                 if ((adev->asic_type == CHIP_POLARIS10 ||
 232                      adev->asic_type == CHIP_POLARIS11) &&
 233                     (adev->uvd.fw_version < FW_1_66_16))
 234                         DRM_ERROR("POLARIS10/11 UVD firmware version %hu.%hu is too old.\n",
 235                                   version_major, version_minor);
 236         } else {
 237                 unsigned int enc_major, enc_minor, dec_minor;
 238
 239                 dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
 240                 enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
 241                 enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
 242                 DRM_INFO("Found UVD firmware ENC: %hu.%hu DEC: .%hu Family ID: %hu\n",
 243                         enc_major, enc_minor, dec_minor, family_id);
 244
 245                 adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
 246
 247                 adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
 248         }
 249
 250         bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
 251                   +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
 252         if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
 253                 bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
 254
 255         for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
 256                 if (adev->uvd.harvest_config & (1 << j))
 257                         continue;
 258                 r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
 259                                             AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.inst[j].vcpu_bo,
 260                                             &adev->uvd.inst[j].gpu_addr, &adev->uvd.inst[j].cpu_addr);
 261                 if (r) {
 262                         dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
 263                         return r;
 264                 }
 265         }
 266
 267         for (i = 0; i < adev->uvd.max_handles; ++i) {
 268                 atomic_set(&adev->uvd.handles[i], 0);
 269                 adev->uvd.filp[i] = NULL;
 270         }
 271
 272         /* from uvd v5.0 HW addressing capacity increased to 64 bits */
 273         if (!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
 274                 adev->uvd.address_64_bit = true;
 275
 276         switch (adev->asic_type) {
 277         case CHIP_TONGA:
 278                 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
 279                 break;
 280         case CHIP_CARRIZO:
 281                 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
 282                 break;
 283         case CHIP_FIJI:
 284                 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
 285                 break;
 286         case CHIP_STONEY:
 287                 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
 288                 break;
 289         default:
 290                 adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
 291         }
 292
 293         return 0;
 294 }
 295
 296 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
 297 {
 298         int i, j;
 299
 300         drm_sched_entity_destroy(&adev->uvd.entity);
 301
 302         for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
 303                 if (adev->uvd.harvest_config & (1 << j))
 304                         continue;
 305                 kvfree(adev->uvd.inst[j].saved_bo);
 306
 307                 amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
 308                                       &adev->uvd.inst[j].gpu_addr,
 309                                       (void **)&adev->uvd.inst[j].cpu_addr);
 310
 311                 amdgpu_ring_fini(&adev->uvd.inst[j].ring);
 312
 313                 for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
 314                         amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
 315         }
 316         release_firmware(adev->uvd.fw);
 317
 318         return 0;
 319 }
 320
 321 /**
 322  * amdgpu_uvd_entity_init - init entity
 323  *
 324  * @adev: amdgpu_device pointer
 325  *
 326  */
 327 int amdgpu_uvd_entity_init(struct amdgpu_device *adev)
 328 {
 329         struct amdgpu_ring *ring;
 330         struct drm_sched_rq *rq;
 331         int r;
 332
 333         ring = &adev->uvd.inst[0].ring;
 334         rq = &ring->sched.sched_rq[DRM_SCHED_PRIORITY_NORMAL];
 335         r = drm_sched_entity_init(&adev->uvd.entity, &rq, 1, NULL);
 336         if (r) {
 337                 DRM_ERROR("Failed setting up UVD kernel entity.\n");
 338                 return r;
 339         }
 340
 341         return 0;
 342 }
 343
 344 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
 345 {
 346         unsigned size;
 347         void *ptr;
 348         int i, j;
 349
 350         cancel_delayed_work_sync(&adev->uvd.idle_work);
 351
 352         /* only valid for physical mode */
 353         if (adev->asic_type < CHIP_POLARIS10) {
 354                 for (i = 0; i < adev->uvd.max_handles; ++i)
 355                         if (atomic_read(&adev->uvd.handles[i]))
 356                                 break;
 357
 358                 if (i == adev->uvd.max_handles)
 359                         return 0;
 360         }
 361
 362         for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
 363                 if (adev->uvd.harvest_config & (1 << j))
 364                         continue;
 365                 if (adev->uvd.inst[j].vcpu_bo == NULL)
 366                         continue;
 367
 368                 size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
 369                 ptr = adev->uvd.inst[j].cpu_addr;
 370
 371                 adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
 372                 if (!adev->uvd.inst[j].saved_bo)
 373                         return -ENOMEM;
 374
 375                 memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
 376         }
 377         return 0;
 378 }
 379
 380 int amdgpu_uvd_resume(struct amdgpu_device *adev)
 381 {
 382         unsigned size;
 383         void *ptr;
 384         int i;
 385
 386         for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
 387                 if (adev->uvd.harvest_config & (1 << i))
 388                         continue;
 389                 if (adev->uvd.inst[i].vcpu_bo == NULL)
 390                         return -EINVAL;
 391
 392                 size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
 393                 ptr = adev->uvd.inst[i].cpu_addr;
 394
 395                 if (adev->uvd.inst[i].saved_bo != NULL) {
 396                         memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
 397                         kvfree(adev->uvd.inst[i].saved_bo);
 398                         adev->uvd.inst[i].saved_bo = NULL;
 399                 } else {
 400                         const struct common_firmware_header *hdr;
 401                         unsigned offset;
 402
 403                         hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 404                         if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
 405                                 offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
 406                                 memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
 407                                             le32_to_cpu(hdr->ucode_size_bytes));
 408                                 size -= le32_to_cpu(hdr->ucode_size_bytes);
 409                                 ptr += le32_to_cpu(hdr->ucode_size_bytes);
 410                         }
 411                         memset_io(ptr, 0, size);
 412                         /* to restore uvd fence seq */
 413                         amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring);
 414                 }
 415         }
 416         return 0;
 417 }
 418
 419 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
 420 {
 421         struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
 422         int i, r;
 423
 424         for (i = 0; i < adev->uvd.max_handles; ++i) {
 425                 uint32_t handle = atomic_read(&adev->uvd.handles[i]);
 426
 427                 if (handle != 0 && adev->uvd.filp[i] == filp) {
 428                         struct dma_fence *fence;
 429
 430                         r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
 431                                                        &fence);
 432                         if (r) {
 433                                 DRM_ERROR("Error destroying UVD %d!\n", r);
 434                                 continue;
 435                         }
 436
 437                         dma_fence_wait(fence, false);
 438                         dma_fence_put(fence);
 439
 440                         adev->uvd.filp[i] = NULL;
 441                         atomic_set(&adev->uvd.handles[i], 0);
 442                 }
 443         }
 444 }
 445
 446 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
 447 {
 448         int i;
 449         for (i = 0; i < abo->placement.num_placement; ++i) {
 450                 abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
 451                 abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
 452         }
 453 }
 454
 455 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
 456 {
 457         uint32_t lo, hi;
 458         uint64_t addr;
 459
 460         lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
 461         hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
 462         addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
 463
 464         return addr;
 465 }
 466
 467 /**
 468  * amdgpu_uvd_cs_pass1 - first parsing round
 469  *
 470  * @ctx: UVD parser context
 471  *
 472  * Make sure UVD message and feedback buffers are in VRAM and
 473  * nobody is violating an 256MB boundary.
 474  */
 475 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
 476 {
 477         struct ttm_operation_ctx tctx = { false, false };
 478         struct amdgpu_bo_va_mapping *mapping;
 479         struct amdgpu_bo *bo;
 480         uint32_t cmd;
 481         uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
 482         int r = 0;
 483
 484         r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
 485         if (r) {
 486                 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
 487                 return r;
 488         }
 489
 490         if (!ctx->parser->adev->uvd.address_64_bit) {
 491                 /* check if it's a message or feedback command */
 492                 cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
 493                 if (cmd == 0x0 || cmd == 0x3) {
 494                         /* yes, force it into VRAM */
 495                         uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
 496                         amdgpu_bo_placement_from_domain(bo, domain);
 497                 }
 498                 amdgpu_uvd_force_into_uvd_segment(bo);
 499
 500                 r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
 501         }
 502
 503         return r;
 504 }
 505
 506 /**
 507  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
 508  *
 509  * @msg: pointer to message structure
 510  * @buf_sizes: returned buffer sizes
 511  *
 512  * Peek into the decode message and calculate the necessary buffer sizes.
 513  */
 514 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
 515         unsigned buf_sizes[])
 516 {
 517         unsigned stream_type = msg[4];
 518         unsigned width = msg[6];
 519         unsigned height = msg[7];
 520         unsigned dpb_size = msg[9];
 521         unsigned pitch = msg[28];
 522         unsigned level = msg[57];
 523
 524         unsigned width_in_mb = width / 16;
 525         unsigned height_in_mb = ALIGN(height / 16, 2);
 526         unsigned fs_in_mb = width_in_mb * height_in_mb;
 527
 528         unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
 529         unsigned min_ctx_size = ~0;
 530
 531         image_size = width * height;
 532         image_size += image_size / 2;
 533         image_size = ALIGN(image_size, 1024);
 534
 535         switch (stream_type) {
 536         case 0: /* H264 */
 537                 switch(level) {
 538                 case 30:
 539                         num_dpb_buffer = 8100 / fs_in_mb;
 540                         break;
 541                 case 31:
 542                         num_dpb_buffer = 18000 / fs_in_mb;
 543                         break;
 544                 case 32:
 545                         num_dpb_buffer = 20480 / fs_in_mb;
 546                         break;
 547                 case 41:
 548                         num_dpb_buffer = 32768 / fs_in_mb;
 549                         break;
 550                 case 42:
 551                         num_dpb_buffer = 34816 / fs_in_mb;
 552                         break;
 553                 case 50:
 554                         num_dpb_buffer = 110400 / fs_in_mb;
 555                         break;
 556                 case 51:
 557                         num_dpb_buffer = 184320 / fs_in_mb;
 558                         break;
 559                 default:
 560                         num_dpb_buffer = 184320 / fs_in_mb;
 561                         break;
 562                 }
 563                 num_dpb_buffer++;
 564                 if (num_dpb_buffer > 17)
 565                         num_dpb_buffer = 17;
 566
 567                 /* reference picture buffer */
 568                 min_dpb_size = image_size * num_dpb_buffer;
 569
 570                 /* macroblock context buffer */
 571                 min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
 572
 573                 /* IT surface buffer */
 574                 min_dpb_size += width_in_mb * height_in_mb * 32;
 575                 break;
 576
 577         case 1: /* VC1 */
 578
 579                 /* reference picture buffer */
 580                 min_dpb_size = image_size * 3;
 581
 582                 /* CONTEXT_BUFFER */
 583                 min_dpb_size += width_in_mb * height_in_mb * 128;
 584
 585                 /* IT surface buffer */
 586                 min_dpb_size += width_in_mb * 64;
 587
 588                 /* DB surface buffer */
 589                 min_dpb_size += width_in_mb * 128;
 590
 591                 /* BP */
 592                 tmp = max(width_in_mb, height_in_mb);
 593                 min_dpb_size += ALIGN(tmp * 7 * 16, 64);
 594                 break;
 595
 596         case 3: /* MPEG2 */
 597
 598                 /* reference picture buffer */
 599                 min_dpb_size = image_size * 3;
 600                 break;
 601
 602         case 4: /* MPEG4 */
 603
 604                 /* reference picture buffer */
 605                 min_dpb_size = image_size * 3;
 606
 607                 /* CM */
 608                 min_dpb_size += width_in_mb * height_in_mb * 64;
 609
 610                 /* IT surface buffer */
 611                 min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
 612                 break;
 613
 614         case 7: /* H264 Perf */
 615                 switch(level) {
 616                 case 30:
 617                         num_dpb_buffer = 8100 / fs_in_mb;
 618                         break;
 619                 case 31:
 620                         num_dpb_buffer = 18000 / fs_in_mb;
 621                         break;
 622                 case 32:
 623                         num_dpb_buffer = 20480 / fs_in_mb;
 624                         break;
 625                 case 41:
 626                         num_dpb_buffer = 32768 / fs_in_mb;
 627                         break;
 628                 case 42:
 629                         num_dpb_buffer = 34816 / fs_in_mb;
 630                         break;
 631                 case 50:
 632                         num_dpb_buffer = 110400 / fs_in_mb;
 633                         break;
 634                 case 51:
 635                         num_dpb_buffer = 184320 / fs_in_mb;
 636                         break;
 637                 default:
 638                         num_dpb_buffer = 184320 / fs_in_mb;
 639                         break;
 640                 }
 641                 num_dpb_buffer++;
 642                 if (num_dpb_buffer > 17)
 643                         num_dpb_buffer = 17;
 644
 645                 /* reference picture buffer */
 646                 min_dpb_size = image_size * num_dpb_buffer;
 647
 648                 if (!adev->uvd.use_ctx_buf){
 649                         /* macroblock context buffer */
 650                         min_dpb_size +=
 651                                 width_in_mb * height_in_mb * num_dpb_buffer * 192;
 652
 653                         /* IT surface buffer */
 654                         min_dpb_size += width_in_mb * height_in_mb * 32;
 655                 } else {
 656                         /* macroblock context buffer */
 657                         min_ctx_size =
 658                                 width_in_mb * height_in_mb * num_dpb_buffer * 192;
 659                 }
 660                 break;
 661
 662         case 8: /* MJPEG */
 663                 min_dpb_size = 0;
 664                 break;
 665
 666         case 16: /* H265 */
 667                 image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
 668                 image_size = ALIGN(image_size, 256);
 669
 670                 num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
 671                 min_dpb_size = image_size * num_dpb_buffer;
 672                 min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
 673                                            * 16 * num_dpb_buffer + 52 * 1024;
 674                 break;
 675
 676         default:
 677                 DRM_ERROR("UVD codec not handled %d!\n", stream_type);
 678                 return -EINVAL;
 679         }
 680
 681         if (width > pitch) {
 682                 DRM_ERROR("Invalid UVD decoding target pitch!\n");
 683                 return -EINVAL;
 684         }
 685
 686         if (dpb_size < min_dpb_size) {
 687                 DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
 688                           dpb_size, min_dpb_size);
 689                 return -EINVAL;
 690         }
 691
 692         buf_sizes[0x1] = dpb_size;
 693         buf_sizes[0x2] = image_size;
 694         buf_sizes[0x4] = min_ctx_size;
 695         return 0;
 696 }
 697
 698 /**
 699  * amdgpu_uvd_cs_msg - handle UVD message
 700  *
 701  * @ctx: UVD parser context
 702  * @bo: buffer object containing the message
 703  * @offset: offset into the buffer object
 704  *
 705  * Peek into the UVD message and extract the session id.
 706  * Make sure that we don't open up to many sessions.
 707  */
 708 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
 709                              struct amdgpu_bo *bo, unsigned offset)
 710 {
 711         struct amdgpu_device *adev = ctx->parser->adev;
 712         int32_t *msg, msg_type, handle;
 713         void *ptr;
 714         long r;
 715         int i;
 716
 717         if (offset & 0x3F) {
 718                 DRM_ERROR("UVD messages must be 64 byte aligned!\n");
 719                 return -EINVAL;
 720         }
 721
 722         r = amdgpu_bo_kmap(bo, &ptr);
 723         if (r) {
 724                 DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
 725                 return r;
 726         }
 727
 728         msg = ptr + offset;
 729
 730         msg_type = msg[1];
 731         handle = msg[2];
 732
 733         if (handle == 0) {
 734                 DRM_ERROR("Invalid UVD handle!\n");
 735                 return -EINVAL;
 736         }
 737
 738         switch (msg_type) {
 739         case 0:
 740                 /* it's a create msg, calc image size (width * height) */
 741                 amdgpu_bo_kunmap(bo);
 742
 743                 /* try to alloc a new handle */
 744                 for (i = 0; i < adev->uvd.max_handles; ++i) {
 745                         if (atomic_read(&adev->uvd.handles[i]) == handle) {
 746                                 DRM_ERROR(")Handle 0x%x already in use!\n",
 747                                           handle);
 748                                 return -EINVAL;
 749                         }
 750
 751                         if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
 752                                 adev->uvd.filp[i] = ctx->parser->filp;
 753                                 return 0;
 754                         }
 755                 }
 756
 757                 DRM_ERROR("No more free UVD handles!\n");
 758                 return -ENOSPC;
 759
 760         case 1:
 761                 /* it's a decode msg, calc buffer sizes */
 762                 r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
 763                 amdgpu_bo_kunmap(bo);
 764                 if (r)
 765                         return r;
 766
 767                 /* validate the handle */
 768                 for (i = 0; i < adev->uvd.max_handles; ++i) {
 769                         if (atomic_read(&adev->uvd.handles[i]) == handle) {
 770                                 if (adev->uvd.filp[i] != ctx->parser->filp) {
 771                                         DRM_ERROR("UVD handle collision detected!\n");
 772                                         return -EINVAL;
 773                                 }
 774                                 return 0;
 775                         }
 776                 }
 777
 778                 DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
 779                 return -ENOENT;
 780
 781         case 2:
 782                 /* it's a destroy msg, free the handle */
 783                 for (i = 0; i < adev->uvd.max_handles; ++i)
 784                         atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
 785                 amdgpu_bo_kunmap(bo);
 786                 return 0;
 787
 788         default:
 789                 DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
 790                 return -EINVAL;
 791         }
 792         BUG();
 793         return -EINVAL;
 794 }
 795
 796 /**
 797  * amdgpu_uvd_cs_pass2 - second parsing round
 798  *
 799  * @ctx: UVD parser context
 800  *
 801  * Patch buffer addresses, make sure buffer sizes are correct.
 802  */
 803 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
 804 {
 805         struct amdgpu_bo_va_mapping *mapping;
 806         struct amdgpu_bo *bo;
 807         uint32_t cmd;
 808         uint64_t start, end;
 809         uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
 810         int r;
 811
 812         r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
 813         if (r) {
 814                 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
 815                 return r;
 816         }
 817
 818         start = amdgpu_bo_gpu_offset(bo);
 819
 820         end = (mapping->last + 1 - mapping->start);
 821         end = end * AMDGPU_GPU_PAGE_SIZE + start;
 822
 823         addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
 824         start += addr;
 825
 826         amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
 827                             lower_32_bits(start));
 828         amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
 829                             upper_32_bits(start));
 830
 831         cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
 832         if (cmd < 0x4) {
 833                 if ((end - start) < ctx->buf_sizes[cmd]) {
 834                         DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 835                                   (unsigned)(end - start),
 836                                   ctx->buf_sizes[cmd]);
 837                         return -EINVAL;
 838                 }
 839
 840         } else if (cmd == 0x206) {
 841                 if ((end - start) < ctx->buf_sizes[4]) {
 842                         DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 843                                           (unsigned)(end - start),
 844                                           ctx->buf_sizes[4]);
 845                         return -EINVAL;
 846                 }
 847         } else if ((cmd != 0x100) && (cmd != 0x204)) {
 848                 DRM_ERROR("invalid UVD command %X!\n", cmd);
 849                 return -EINVAL;
 850         }
 851
 852         if (!ctx->parser->adev->uvd.address_64_bit) {
 853                 if ((start >> 28) != ((end - 1) >> 28)) {
 854                         DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
 855                                   start, end);
 856                         return -EINVAL;
 857                 }
 858
 859                 if ((cmd == 0 || cmd == 0x3) &&
 860                     (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
 861                         DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
 862                                   start, end);
 863                         return -EINVAL;
 864                 }
 865         }
 866
 867         if (cmd == 0) {
 868                 ctx->has_msg_cmd = true;
 869                 r = amdgpu_uvd_cs_msg(ctx, bo, addr);
 870                 if (r)
 871                         return r;
 872         } else if (!ctx->has_msg_cmd) {
 873                 DRM_ERROR("Message needed before other commands are send!\n");
 874                 return -EINVAL;
 875         }
 876
 877         return 0;
 878 }
 879
 880 /**
 881  * amdgpu_uvd_cs_reg - parse register writes
 882  *
 883  * @ctx: UVD parser context
 884  * @cb: callback function
 885  *
 886  * Parse the register writes, call cb on each complete command.
 887  */
 888 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
 889                              int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
 890 {
 891         struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
 892         int i, r;
 893
 894         ctx->idx++;
 895         for (i = 0; i <= ctx->count; ++i) {
 896                 unsigned reg = ctx->reg + i;
 897
 898                 if (ctx->idx >= ib->length_dw) {
 899                         DRM_ERROR("Register command after end of CS!\n");
 900                         return -EINVAL;
 901                 }
 902
 903                 switch (reg) {
 904                 case mmUVD_GPCOM_VCPU_DATA0:
 905                         ctx->data0 = ctx->idx;
 906                         break;
 907                 case mmUVD_GPCOM_VCPU_DATA1:
 908                         ctx->data1 = ctx->idx;
 909                         break;
 910                 case mmUVD_GPCOM_VCPU_CMD:
 911                         r = cb(ctx);
 912                         if (r)
 913                                 return r;
 914                         break;
 915                 case mmUVD_ENGINE_CNTL:
 916                 case mmUVD_NO_OP:
 917                         break;
 918                 default:
 919                         DRM_ERROR("Invalid reg 0x%X!\n", reg);
 920                         return -EINVAL;
 921                 }
 922                 ctx->idx++;
 923         }
 924         return 0;
 925 }
 926
 927 /**
 928  * amdgpu_uvd_cs_packets - parse UVD packets
 929  *
 930  * @ctx: UVD parser context
 931  * @cb: callback function
 932  *
 933  * Parse the command stream packets.
 934  */
 935 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
 936                                  int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
 937 {
 938         struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
 939         int r;
 940
 941         for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
 942                 uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
 943                 unsigned type = CP_PACKET_GET_TYPE(cmd);
 944                 switch (type) {
 945                 case PACKET_TYPE0:
 946                         ctx->reg = CP_PACKET0_GET_REG(cmd);
 947                         ctx->count = CP_PACKET_GET_COUNT(cmd);
 948                         r = amdgpu_uvd_cs_reg(ctx, cb);
 949                         if (r)
 950                                 return r;
 951                         break;
 952                 case PACKET_TYPE2:
 953                         ++ctx->idx;
 954                         break;
 955                 default:
 956                         DRM_ERROR("Unknown packet type %d !\n", type);
 957                         return -EINVAL;
 958                 }
 959         }
 960         return 0;
 961 }
 962
 963 /**
 964  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
 965  *
 966  * @parser: Command submission parser context
 967  *
 968  * Parse the command stream, patch in addresses as necessary.
 969  */
 970 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
 971 {
 972         struct amdgpu_uvd_cs_ctx ctx = {};
 973         unsigned buf_sizes[] = {
 974                 [0x00000000]    =       2048,
 975                 [0x00000001]    =       0xFFFFFFFF,
 976                 [0x00000002]    =       0xFFFFFFFF,
 977                 [0x00000003]    =       2048,
 978                 [0x00000004]    =       0xFFFFFFFF,
 979         };
 980         struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
 981         int r;
 982
 983         parser->job->vm = NULL;
 984         ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
 985
 986         if (ib->length_dw % 16) {
 987                 DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
 988                           ib->length_dw);
 989                 return -EINVAL;
 990         }
 991
 992         ctx.parser = parser;
 993         ctx.buf_sizes = buf_sizes;
 994         ctx.ib_idx = ib_idx;
 995
 996         /* first round only required on chips without UVD 64 bit address support */
 997         if (!parser->adev->uvd.address_64_bit) {
 998                 /* first round, make sure the buffers are actually in the UVD segment */
 999                 r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1000                 if (r)
1001                         return r;
1002         }
1003
1004         /* second round, patch buffer addresses into the command stream */
1005         r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1006         if (r)
1007                 return r;
1008
1009         if (!ctx.has_msg_cmd) {
1010                 DRM_ERROR("UVD-IBs need a msg command!\n");
1011                 return -EINVAL;
1012         }
1013
1014         return 0;
1015 }
1016
1017 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1018                                bool direct, struct dma_fence **fence)
1019 {
1020         struct amdgpu_device *adev = ring->adev;
1021         struct dma_fence *f = NULL;
1022         struct amdgpu_job *job;
1023         struct amdgpu_ib *ib;
1024         uint32_t data[4];
1025         uint64_t addr;
1026         long r;
1027         int i;
1028         unsigned offset_idx = 0;
1029         unsigned offset[3] = { UVD_BASE_SI, 0, 0 };
1030
1031         amdgpu_bo_kunmap(bo);
1032         amdgpu_bo_unpin(bo);
1033
1034         if (!ring->adev->uvd.address_64_bit) {
1035                 struct ttm_operation_ctx ctx = { true, false };
1036
1037                 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
1038                 amdgpu_uvd_force_into_uvd_segment(bo);
1039                 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1040                 if (r)
1041                         goto err;
1042         }
1043
1044         r = amdgpu_job_alloc_with_ib(adev, 64, &job);
1045         if (r)
1046                 goto err;
1047
1048         if (adev->asic_type >= CHIP_VEGA10) {
1049                 offset_idx = 1 + ring->me;
1050                 offset[1] = adev->reg_offset[UVD_HWIP][0][1];
1051                 offset[2] = adev->reg_offset[UVD_HWIP][1][1];
1052         }
1053
1054         data[0] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA0, 0);
1055         data[1] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA1, 0);
1056         data[2] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_CMD, 0);
1057         data[3] = PACKET0(offset[offset_idx] + UVD_NO_OP, 0);
1058
1059         ib = &job->ibs[0];
1060         addr = amdgpu_bo_gpu_offset(bo);
1061         ib->ptr[0] = data[0];
1062         ib->ptr[1] = addr;
1063         ib->ptr[2] = data[1];
1064         ib->ptr[3] = addr >> 32;
1065         ib->ptr[4] = data[2];
1066         ib->ptr[5] = 0;
1067         for (i = 6; i < 16; i += 2) {
1068                 ib->ptr[i] = data[3];
1069                 ib->ptr[i+1] = 0;
1070         }
1071         ib->length_dw = 16;
1072
1073         if (direct) {
1074                 r = reservation_object_wait_timeout_rcu(bo->tbo.resv,
1075                                                         true, false,
1076                                                         msecs_to_jiffies(10));
1077                 if (r == 0)
1078                         r = -ETIMEDOUT;
1079                 if (r < 0)
1080                         goto err_free;
1081
1082                 r = amdgpu_job_submit_direct(job, ring, &f);
1083                 if (r)
1084                         goto err_free;
1085         } else {
1086                 r = amdgpu_sync_resv(adev, &job->sync, bo->tbo.resv,
1087                                      AMDGPU_FENCE_OWNER_UNDEFINED, false);
1088                 if (r)
1089                         goto err_free;
1090
1091                 r = amdgpu_job_submit(job, &adev->uvd.entity,
1092                                       AMDGPU_FENCE_OWNER_UNDEFINED, &f);
1093                 if (r)
1094                         goto err_free;
1095         }
1096
1097         amdgpu_bo_fence(bo, f, false);
1098         amdgpu_bo_unreserve(bo);
1099         amdgpu_bo_unref(&bo);
1100
1101         if (fence)
1102                 *fence = dma_fence_get(f);
1103         dma_fence_put(f);
1104
1105         return 0;
1106
1107 err_free:
1108         amdgpu_job_free(job);
1109
1110 err:
1111         amdgpu_bo_unreserve(bo);
1112         amdgpu_bo_unref(&bo);
1113         return r;
1114 }
1115
1116 /* multiple fence commands without any stream commands in between can
1117    crash the vcpu so just try to emmit a dummy create/destroy msg to
1118    avoid this */
1119 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1120                               struct dma_fence **fence)
1121 {
1122         struct amdgpu_device *adev = ring->adev;
1123         struct amdgpu_bo *bo = NULL;
1124         uint32_t *msg;
1125         int r, i;
1126
1127         r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1128                                       AMDGPU_GEM_DOMAIN_VRAM,
1129                                       &bo, NULL, (void **)&msg);
1130         if (r)
1131                 return r;
1132
1133         /* stitch together an UVD create msg */
1134         msg[0] = cpu_to_le32(0x00000de4);
1135         msg[1] = cpu_to_le32(0x00000000);
1136         msg[2] = cpu_to_le32(handle);
1137         msg[3] = cpu_to_le32(0x00000000);
1138         msg[4] = cpu_to_le32(0x00000000);
1139         msg[5] = cpu_to_le32(0x00000000);
1140         msg[6] = cpu_to_le32(0x00000000);
1141         msg[7] = cpu_to_le32(0x00000780);
1142         msg[8] = cpu_to_le32(0x00000440);
1143         msg[9] = cpu_to_le32(0x00000000);
1144         msg[10] = cpu_to_le32(0x01b37000);
1145         for (i = 11; i < 1024; ++i)
1146                 msg[i] = cpu_to_le32(0x0);
1147
1148         return amdgpu_uvd_send_msg(ring, bo, true, fence);
1149 }
1150
1151 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1152                                bool direct, struct dma_fence **fence)
1153 {
1154         struct amdgpu_device *adev = ring->adev;
1155         struct amdgpu_bo *bo = NULL;
1156         uint32_t *msg;
1157         int r, i;
1158
1159         r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1160                                       AMDGPU_GEM_DOMAIN_VRAM,
1161                                       &bo, NULL, (void **)&msg);
1162         if (r)
1163                 return r;
1164
1165         /* stitch together an UVD destroy msg */
1166         msg[0] = cpu_to_le32(0x00000de4);
1167         msg[1] = cpu_to_le32(0x00000002);
1168         msg[2] = cpu_to_le32(handle);
1169         msg[3] = cpu_to_le32(0x00000000);
1170         for (i = 4; i < 1024; ++i)
1171                 msg[i] = cpu_to_le32(0x0);
1172
1173         return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1174 }
1175
1176 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1177 {
1178         struct amdgpu_device *adev =
1179                 container_of(work, struct amdgpu_device, uvd.idle_work.work);
1180         unsigned fences = 0, i, j;
1181
1182         for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1183                 if (adev->uvd.harvest_config & (1 << i))
1184                         continue;
1185                 fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1186                 for (j = 0; j < adev->uvd.num_enc_rings; ++j) {
1187                         fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1188                 }
1189         }
1190
1191         if (fences == 0) {
1192                 if (adev->pm.dpm_enabled) {
1193                         amdgpu_dpm_enable_uvd(adev, false);
1194                 } else {
1195                         amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1196                         /* shutdown the UVD block */
1197                         amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1198                                                                AMD_PG_STATE_GATE);
1199                         amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1200                                                                AMD_CG_STATE_GATE);
1201                 }
1202         } else {
1203                 schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1204         }
1205 }
1206
1207 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1208 {
1209         struct amdgpu_device *adev = ring->adev;
1210         bool set_clocks;
1211
1212         if (amdgpu_sriov_vf(adev))
1213                 return;
1214
1215         set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1216         if (set_clocks) {
1217                 if (adev->pm.dpm_enabled) {
1218                         amdgpu_dpm_enable_uvd(adev, true);
1219                 } else {
1220                         amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1221                         amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1222                                                                AMD_CG_STATE_UNGATE);
1223                         amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1224                                                                AMD_PG_STATE_UNGATE);
1225                 }
1226         }
1227 }
1228
1229 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1230 {
1231         if (!amdgpu_sriov_vf(ring->adev))
1232                 schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1233 }
1234
1235 /**
1236  * amdgpu_uvd_ring_test_ib - test ib execution
1237  *
1238  * @ring: amdgpu_ring pointer
1239  *
1240  * Test if we can successfully execute an IB
1241  */
1242 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1243 {
1244         struct dma_fence *fence;
1245         long r;
1246         uint32_t ip_instance = ring->me;
1247
1248         r = amdgpu_uvd_get_create_msg(ring, 1, NULL);
1249         if (r) {
1250                 DRM_ERROR("amdgpu: (%d)failed to get create msg (%ld).\n", ip_instance, r);
1251                 goto error;
1252         }
1253
1254         r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1255         if (r) {
1256                 DRM_ERROR("amdgpu: (%d)failed to get destroy ib (%ld).\n", ip_instance, r);
1257                 goto error;
1258         }
1259
1260         r = dma_fence_wait_timeout(fence, false, timeout);
1261         if (r == 0) {
1262                 DRM_ERROR("amdgpu: (%d)IB test timed out.\n", ip_instance);
1263                 r = -ETIMEDOUT;
1264         } else if (r < 0) {
1265                 DRM_ERROR("amdgpu: (%d)fence wait failed (%ld).\n", ip_instance, r);
1266         } else {
1267                 DRM_DEBUG("ib test on (%d)ring %d succeeded\n", ip_instance, ring->idx);
1268                 r = 0;
1269         }
1270
1271         dma_fence_put(fence);
1272
1273 error:
1274         return r;
1275 }
1276
1277 /**
1278  * amdgpu_uvd_used_handles - returns used UVD handles
1279  *
1280  * @adev: amdgpu_device pointer
1281  *
1282  * Returns the number of UVD handles in use
1283  */
1284 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1285 {
1286         unsigned i;
1287         uint32_t used_handles = 0;
1288
1289         for (i = 0; i < adev->uvd.max_handles; ++i) {
1290                 /*
1291                  * Handles can be freed in any order, and not
1292                  * necessarily linear. So we need to count
1293                  * all non-zero handles.
1294                  */
1295                 if (atomic_read(&adev->uvd.handles[i]))
1296                         used_handles++;
1297         }
1298
1299         return used_handles;
1300 }