drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c

   1 /*
   2  * Copyright 2013 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  * Authors: Christian König <[email protected]>
  26  */
  27
  28 #include <linux/firmware.h>
  29 #include <linux/module.h>
  30 #include <drm/drmP.h>
  31 #include <drm/drm.h>
  32
  33 #include "amdgpu.h"
  34 #include "amdgpu_pm.h"
  35 #include "amdgpu_vce.h"
  36 #include "cikd.h"
  37
  38 /* 1 second timeout */
  39 #define VCE_IDLE_TIMEOUT_MS     1000
  40
  41 /* Firmware Names */
  42 #ifdef CONFIG_DRM_AMDGPU_CIK
  43 #define FIRMWARE_BONAIRE        "radeon/bonaire_vce.bin"
  44 #define FIRMWARE_KABINI         "radeon/kabini_vce.bin"
  45 #define FIRMWARE_KAVERI         "radeon/kaveri_vce.bin"
  46 #define FIRMWARE_HAWAII         "radeon/hawaii_vce.bin"
  47 #define FIRMWARE_MULLINS        "radeon/mullins_vce.bin"
  48 #endif
  49 #define FIRMWARE_TONGA          "amdgpu/tonga_vce.bin"
  50 #define FIRMWARE_CARRIZO        "amdgpu/carrizo_vce.bin"
  51 #define FIRMWARE_FIJI           "amdgpu/fiji_vce.bin"
  52
  53 #ifdef CONFIG_DRM_AMDGPU_CIK
  54 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
  55 MODULE_FIRMWARE(FIRMWARE_KABINI);
  56 MODULE_FIRMWARE(FIRMWARE_KAVERI);
  57 MODULE_FIRMWARE(FIRMWARE_HAWAII);
  58 MODULE_FIRMWARE(FIRMWARE_MULLINS);
  59 #endif
  60 MODULE_FIRMWARE(FIRMWARE_TONGA);
  61 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
  62 MODULE_FIRMWARE(FIRMWARE_FIJI);
  63
  64 static void amdgpu_vce_idle_work_handler(struct work_struct *work);
  65
  66 /**
  67  * amdgpu_vce_init - allocate memory, load vce firmware
  68  *
  69  * @adev: amdgpu_device pointer
  70  *
  71  * First step to get VCE online, allocate memory and load the firmware
  72  */
  73 int amdgpu_vce_sw_init(struct amdgpu_device *adev, unsigned long size)
  74 {
  75         const char *fw_name;
  76         const struct common_firmware_header *hdr;
  77         unsigned ucode_version, version_major, version_minor, binary_id;
  78         int i, r;
  79
  80         INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
  81
  82         switch (adev->asic_type) {
  83 #ifdef CONFIG_DRM_AMDGPU_CIK
  84         case CHIP_BONAIRE:
  85                 fw_name = FIRMWARE_BONAIRE;
  86                 break;
  87         case CHIP_KAVERI:
  88                 fw_name = FIRMWARE_KAVERI;
  89                 break;
  90         case CHIP_KABINI:
  91                 fw_name = FIRMWARE_KABINI;
  92                 break;
  93         case CHIP_HAWAII:
  94                 fw_name = FIRMWARE_HAWAII;
  95                 break;
  96         case CHIP_MULLINS:
  97                 fw_name = FIRMWARE_MULLINS;
  98                 break;
  99 #endif
 100         case CHIP_TONGA:
 101                 fw_name = FIRMWARE_TONGA;
 102                 break;
 103         case CHIP_CARRIZO:
 104                 fw_name = FIRMWARE_CARRIZO;
 105                 break;
 106         case CHIP_FIJI:
 107                 fw_name = FIRMWARE_FIJI;
 108                 break;
 109
 110         default:
 111                 return -EINVAL;
 112         }
 113
 114         r = request_firmware(&adev->vce.fw, fw_name, adev->dev);
 115         if (r) {
 116                 dev_err(adev->dev, "amdgpu_vce: Can't load firmware \"%s\"\n",
 117                         fw_name);
 118                 return r;
 119         }
 120
 121         r = amdgpu_ucode_validate(adev->vce.fw);
 122         if (r) {
 123                 dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
 124                         fw_name);
 125                 release_firmware(adev->vce.fw);
 126                 adev->vce.fw = NULL;
 127                 return r;
 128         }
 129
 130         hdr = (const struct common_firmware_header *)adev->vce.fw->data;
 131
 132         ucode_version = le32_to_cpu(hdr->ucode_version);
 133         version_major = (ucode_version >> 20) & 0xfff;
 134         version_minor = (ucode_version >> 8) & 0xfff;
 135         binary_id = ucode_version & 0xff;
 136         DRM_INFO("Found VCE firmware Version: %hhd.%hhd Binary ID: %hhd\n",
 137                 version_major, version_minor, binary_id);
 138         adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
 139                                 (binary_id << 8));
 140
 141         /* allocate firmware, stack and heap BO */
 142
 143         r = amdgpu_bo_create(adev, size, PAGE_SIZE, true,
 144                              AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &adev->vce.vcpu_bo);
 145         if (r) {
 146                 dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
 147                 return r;
 148         }
 149
 150         r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
 151         if (r) {
 152                 amdgpu_bo_unref(&adev->vce.vcpu_bo);
 153                 dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
 154                 return r;
 155         }
 156
 157         r = amdgpu_bo_pin(adev->vce.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
 158                           &adev->vce.gpu_addr);
 159         amdgpu_bo_unreserve(adev->vce.vcpu_bo);
 160         if (r) {
 161                 amdgpu_bo_unref(&adev->vce.vcpu_bo);
 162                 dev_err(adev->dev, "(%d) VCE bo pin failed\n", r);
 163                 return r;
 164         }
 165
 166         for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
 167                 atomic_set(&adev->vce.handles[i], 0);
 168                 adev->vce.filp[i] = NULL;
 169         }
 170
 171         return 0;
 172 }
 173
 174 /**
 175  * amdgpu_vce_fini - free memory
 176  *
 177  * @adev: amdgpu_device pointer
 178  *
 179  * Last step on VCE teardown, free firmware memory
 180  */
 181 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
 182 {
 183         if (adev->vce.vcpu_bo == NULL)
 184                 return 0;
 185
 186         amdgpu_bo_unref(&adev->vce.vcpu_bo);
 187
 188         amdgpu_ring_fini(&adev->vce.ring[0]);
 189         amdgpu_ring_fini(&adev->vce.ring[1]);
 190
 191         release_firmware(adev->vce.fw);
 192
 193         return 0;
 194 }
 195
 196 /**
 197  * amdgpu_vce_suspend - unpin VCE fw memory
 198  *
 199  * @adev: amdgpu_device pointer
 200  *
 201  */
 202 int amdgpu_vce_suspend(struct amdgpu_device *adev)
 203 {
 204         int i;
 205
 206         if (adev->vce.vcpu_bo == NULL)
 207                 return 0;
 208
 209         for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
 210                 if (atomic_read(&adev->vce.handles[i]))
 211                         break;
 212
 213         if (i == AMDGPU_MAX_VCE_HANDLES)
 214                 return 0;
 215
 216         /* TODO: suspending running encoding sessions isn't supported */
 217         return -EINVAL;
 218 }
 219
 220 /**
 221  * amdgpu_vce_resume - pin VCE fw memory
 222  *
 223  * @adev: amdgpu_device pointer
 224  *
 225  */
 226 int amdgpu_vce_resume(struct amdgpu_device *adev)
 227 {
 228         void *cpu_addr;
 229         const struct common_firmware_header *hdr;
 230         unsigned offset;
 231         int r;
 232
 233         if (adev->vce.vcpu_bo == NULL)
 234                 return -EINVAL;
 235
 236         r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
 237         if (r) {
 238                 dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
 239                 return r;
 240         }
 241
 242         r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
 243         if (r) {
 244                 amdgpu_bo_unreserve(adev->vce.vcpu_bo);
 245                 dev_err(adev->dev, "(%d) VCE map failed\n", r);
 246                 return r;
 247         }
 248
 249         hdr = (const struct common_firmware_header *)adev->vce.fw->data;
 250         offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
 251         memcpy(cpu_addr, (adev->vce.fw->data) + offset,
 252                 (adev->vce.fw->size) - offset);
 253
 254         amdgpu_bo_kunmap(adev->vce.vcpu_bo);
 255
 256         amdgpu_bo_unreserve(adev->vce.vcpu_bo);
 257
 258         return 0;
 259 }
 260
 261 /**
 262  * amdgpu_vce_idle_work_handler - power off VCE
 263  *
 264  * @work: pointer to work structure
 265  *
 266  * power of VCE when it's not used any more
 267  */
 268 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
 269 {
 270         struct amdgpu_device *adev =
 271                 container_of(work, struct amdgpu_device, vce.idle_work.work);
 272
 273         if ((amdgpu_fence_count_emitted(&adev->vce.ring[0]) == 0) &&
 274             (amdgpu_fence_count_emitted(&adev->vce.ring[1]) == 0)) {
 275                 if (adev->pm.dpm_enabled) {
 276                         amdgpu_dpm_enable_vce(adev, false);
 277                 } else {
 278                         amdgpu_asic_set_vce_clocks(adev, 0, 0);
 279                 }
 280         } else {
 281                 schedule_delayed_work(&adev->vce.idle_work,
 282                                       msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
 283         }
 284 }
 285
 286 /**
 287  * amdgpu_vce_note_usage - power up VCE
 288  *
 289  * @adev: amdgpu_device pointer
 290  *
 291  * Make sure VCE is powerd up when we want to use it
 292  */
 293 static void amdgpu_vce_note_usage(struct amdgpu_device *adev)
 294 {
 295         bool streams_changed = false;
 296         bool set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
 297         set_clocks &= schedule_delayed_work(&adev->vce.idle_work,
 298                                             msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
 299
 300         if (adev->pm.dpm_enabled) {
 301                 /* XXX figure out if the streams changed */
 302                 streams_changed = false;
 303         }
 304
 305         if (set_clocks || streams_changed) {
 306                 if (adev->pm.dpm_enabled) {
 307                         amdgpu_dpm_enable_vce(adev, true);
 308                 } else {
 309                         amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
 310                 }
 311         }
 312 }
 313
 314 /**
 315  * amdgpu_vce_free_handles - free still open VCE handles
 316  *
 317  * @adev: amdgpu_device pointer
 318  * @filp: drm file pointer
 319  *
 320  * Close all VCE handles still open by this file pointer
 321  */
 322 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
 323 {
 324         struct amdgpu_ring *ring = &adev->vce.ring[0];
 325         int i, r;
 326         for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
 327                 uint32_t handle = atomic_read(&adev->vce.handles[i]);
 328                 if (!handle || adev->vce.filp[i] != filp)
 329                         continue;
 330
 331                 amdgpu_vce_note_usage(adev);
 332
 333                 r = amdgpu_vce_get_destroy_msg(ring, handle, NULL);
 334                 if (r)
 335                         DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
 336
 337                 adev->vce.filp[i] = NULL;
 338                 atomic_set(&adev->vce.handles[i], 0);
 339         }
 340 }
 341
 342 static int amdgpu_vce_free_job(
 343         struct amdgpu_cs_parser *sched_job)
 344 {
 345         amdgpu_ib_free(sched_job->adev, sched_job->ibs);
 346         kfree(sched_job->ibs);
 347         return 0;
 348 }
 349
 350 /**
 351  * amdgpu_vce_get_create_msg - generate a VCE create msg
 352  *
 353  * @adev: amdgpu_device pointer
 354  * @ring: ring we should submit the msg to
 355  * @handle: VCE session handle to use
 356  * @fence: optional fence to return
 357  *
 358  * Open up a stream for HW test
 359  */
 360 int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
 361                               struct fence **fence)
 362 {
 363         const unsigned ib_size_dw = 1024;
 364         struct amdgpu_ib *ib = NULL;
 365         struct fence *f = NULL;
 366         struct amdgpu_device *adev = ring->adev;
 367         uint64_t dummy;
 368         int i, r;
 369
 370         ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
 371         if (!ib)
 372                 return -ENOMEM;
 373         r = amdgpu_ib_get(ring, NULL, ib_size_dw * 4, ib);
 374         if (r) {
 375                 DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
 376                 kfree(ib);
 377                 return r;
 378         }
 379
 380         dummy = ib->gpu_addr + 1024;
 381
 382         /* stitch together an VCE create msg */
 383         ib->length_dw = 0;
 384         ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
 385         ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
 386         ib->ptr[ib->length_dw++] = handle;
 387
 388         ib->ptr[ib->length_dw++] = 0x00000030; /* len */
 389         ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
 390         ib->ptr[ib->length_dw++] = 0x00000000;
 391         ib->ptr[ib->length_dw++] = 0x00000042;
 392         ib->ptr[ib->length_dw++] = 0x0000000a;
 393         ib->ptr[ib->length_dw++] = 0x00000001;
 394         ib->ptr[ib->length_dw++] = 0x00000080;
 395         ib->ptr[ib->length_dw++] = 0x00000060;
 396         ib->ptr[ib->length_dw++] = 0x00000100;
 397         ib->ptr[ib->length_dw++] = 0x00000100;
 398         ib->ptr[ib->length_dw++] = 0x0000000c;
 399         ib->ptr[ib->length_dw++] = 0x00000000;
 400
 401         ib->ptr[ib->length_dw++] = 0x00000014; /* len */
 402         ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
 403         ib->ptr[ib->length_dw++] = upper_32_bits(dummy);
 404         ib->ptr[ib->length_dw++] = dummy;
 405         ib->ptr[ib->length_dw++] = 0x00000001;
 406
 407         for (i = ib->length_dw; i < ib_size_dw; ++i)
 408                 ib->ptr[i] = 0x0;
 409
 410         r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
 411                                                  &amdgpu_vce_free_job,
 412                                                  AMDGPU_FENCE_OWNER_UNDEFINED,
 413                                                  &f);
 414         if (r)
 415                 goto err;
 416         if (fence)
 417                 *fence = fence_get(f);
 418         if (amdgpu_enable_scheduler)
 419                 return 0;
 420 err:
 421         amdgpu_ib_free(adev, ib);
 422         kfree(ib);
 423         return r;
 424 }
 425
 426 /**
 427  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
 428  *
 429  * @adev: amdgpu_device pointer
 430  * @ring: ring we should submit the msg to
 431  * @handle: VCE session handle to use
 432  * @fence: optional fence to return
 433  *
 434  * Close up a stream for HW test or if userspace failed to do so
 435  */
 436 int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
 437                                struct fence **fence)
 438 {
 439         const unsigned ib_size_dw = 1024;
 440         struct amdgpu_ib *ib = NULL;
 441         struct fence *f = NULL;
 442         struct amdgpu_device *adev = ring->adev;
 443         uint64_t dummy;
 444         int i, r;
 445
 446         ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
 447         if (!ib)
 448                 return -ENOMEM;
 449
 450         r = amdgpu_ib_get(ring, NULL, ib_size_dw * 4, ib);
 451         if (r) {
 452                 kfree(ib);
 453                 DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
 454                 return r;
 455         }
 456
 457         dummy = ib->gpu_addr + 1024;
 458
 459         /* stitch together an VCE destroy msg */
 460         ib->length_dw = 0;
 461         ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
 462         ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
 463         ib->ptr[ib->length_dw++] = handle;
 464
 465         ib->ptr[ib->length_dw++] = 0x00000014; /* len */
 466         ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
 467         ib->ptr[ib->length_dw++] = upper_32_bits(dummy);
 468         ib->ptr[ib->length_dw++] = dummy;
 469         ib->ptr[ib->length_dw++] = 0x00000001;
 470
 471         ib->ptr[ib->length_dw++] = 0x00000008; /* len */
 472         ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
 473
 474         for (i = ib->length_dw; i < ib_size_dw; ++i)
 475                 ib->ptr[i] = 0x0;
 476         r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
 477                                                  &amdgpu_vce_free_job,
 478                                                  AMDGPU_FENCE_OWNER_UNDEFINED,
 479                                                  &f);
 480         if (r)
 481                 goto err;
 482         if (fence)
 483                 *fence = fence_get(f);
 484         if (amdgpu_enable_scheduler)
 485                 return 0;
 486 err:
 487         amdgpu_ib_free(adev, ib);
 488         kfree(ib);
 489         return r;
 490 }
 491
 492 /**
 493  * amdgpu_vce_cs_reloc - command submission relocation
 494  *
 495  * @p: parser context
 496  * @lo: address of lower dword
 497  * @hi: address of higher dword
 498  * @size: minimum size
 499  *
 500  * Patch relocation inside command stream with real buffer address
 501  */
 502 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, uint32_t ib_idx,
 503                                int lo, int hi, unsigned size, uint32_t index)
 504 {
 505         struct amdgpu_bo_va_mapping *mapping;
 506         struct amdgpu_ib *ib = &p->ibs[ib_idx];
 507         struct amdgpu_bo *bo;
 508         uint64_t addr;
 509
 510         if (index == 0xffffffff)
 511                 index = 0;
 512
 513         addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
 514                ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
 515         addr += ((uint64_t)size) * ((uint64_t)index);
 516
 517         mapping = amdgpu_cs_find_mapping(p, addr, &bo);
 518         if (mapping == NULL) {
 519                 DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
 520                           addr, lo, hi, size, index);
 521                 return -EINVAL;
 522         }
 523
 524         if ((addr + (uint64_t)size) >
 525             ((uint64_t)mapping->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
 526                 DRM_ERROR("BO to small for addr 0x%010Lx %d %d\n",
 527                           addr, lo, hi);
 528                 return -EINVAL;
 529         }
 530
 531         addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
 532         addr += amdgpu_bo_gpu_offset(bo);
 533         addr -= ((uint64_t)size) * ((uint64_t)index);
 534
 535         ib->ptr[lo] = addr & 0xFFFFFFFF;
 536         ib->ptr[hi] = addr >> 32;
 537
 538         return 0;
 539 }
 540
 541 /**
 542  * amdgpu_vce_validate_handle - validate stream handle
 543  *
 544  * @p: parser context
 545  * @handle: handle to validate
 546  * @allocated: allocated a new handle?
 547  *
 548  * Validates the handle and return the found session index or -EINVAL
 549  * we we don't have another free session index.
 550  */
 551 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
 552                                       uint32_t handle, bool *allocated)
 553 {
 554         unsigned i;
 555
 556         *allocated = false;
 557
 558         /* validate the handle */
 559         for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
 560                 if (atomic_read(&p->adev->vce.handles[i]) == handle) {
 561                         if (p->adev->vce.filp[i] != p->filp) {
 562                                 DRM_ERROR("VCE handle collision detected!\n");
 563                                 return -EINVAL;
 564                         }
 565                         return i;
 566                 }
 567         }
 568
 569         /* handle not found try to alloc a new one */
 570         for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
 571                 if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
 572                         p->adev->vce.filp[i] = p->filp;
 573                         p->adev->vce.img_size[i] = 0;
 574                         *allocated = true;
 575                         return i;
 576                 }
 577         }
 578
 579         DRM_ERROR("No more free VCE handles!\n");
 580         return -EINVAL;
 581 }
 582
 583 /**
 584  * amdgpu_vce_cs_parse - parse and validate the command stream
 585  *
 586  * @p: parser context
 587  *
 588  */
 589 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
 590 {
 591         struct amdgpu_ib *ib = &p->ibs[ib_idx];
 592         unsigned fb_idx = 0, bs_idx = 0;
 593         int session_idx = -1;
 594         bool destroyed = false;
 595         bool created = false;
 596         bool allocated = false;
 597         uint32_t tmp, handle = 0;
 598         uint32_t *size = &tmp;
 599         int i, r = 0, idx = 0;
 600
 601         amdgpu_vce_note_usage(p->adev);
 602
 603         while (idx < ib->length_dw) {
 604                 uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
 605                 uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
 606
 607                 if ((len < 8) || (len & 3)) {
 608                         DRM_ERROR("invalid VCE command length (%d)!\n", len);
 609                         r = -EINVAL;
 610                         goto out;
 611                 }
 612
 613                 if (destroyed) {
 614                         DRM_ERROR("No other command allowed after destroy!\n");
 615                         r = -EINVAL;
 616                         goto out;
 617                 }
 618
 619                 switch (cmd) {
 620                 case 0x00000001: // session
 621                         handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
 622                         session_idx = amdgpu_vce_validate_handle(p, handle,
 623                                                                  &allocated);
 624                         if (session_idx < 0)
 625                                 return session_idx;
 626                         size = &p->adev->vce.img_size[session_idx];
 627                         break;
 628
 629                 case 0x00000002: // task info
 630                         fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
 631                         bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
 632                         break;
 633
 634                 case 0x01000001: // create
 635                         created = true;
 636                         if (!allocated) {
 637                                 DRM_ERROR("Handle already in use!\n");
 638                                 r = -EINVAL;
 639                                 goto out;
 640                         }
 641
 642                         *size = amdgpu_get_ib_value(p, ib_idx, idx + 8) *
 643                                 amdgpu_get_ib_value(p, ib_idx, idx + 10) *
 644                                 8 * 3 / 2;
 645                         break;
 646
 647                 case 0x04000001: // config extension
 648                 case 0x04000002: // pic control
 649                 case 0x04000005: // rate control
 650                 case 0x04000007: // motion estimation
 651                 case 0x04000008: // rdo
 652                 case 0x04000009: // vui
 653                 case 0x05000002: // auxiliary buffer
 654                         break;
 655
 656                 case 0x03000001: // encode
 657                         r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 10, idx + 9,
 658                                                 *size, 0);
 659                         if (r)
 660                                 goto out;
 661
 662                         r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 12, idx + 11,
 663                                                 *size / 3, 0);
 664                         if (r)
 665                                 goto out;
 666                         break;
 667
 668                 case 0x02000001: // destroy
 669                         destroyed = true;
 670                         break;
 671
 672                 case 0x05000001: // context buffer
 673                         r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
 674                                                 *size * 2, 0);
 675                         if (r)
 676                                 goto out;
 677                         break;
 678
 679                 case 0x05000004: // video bitstream buffer
 680                         tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
 681                         r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
 682                                                 tmp, bs_idx);
 683                         if (r)
 684                                 goto out;
 685                         break;
 686
 687                 case 0x05000005: // feedback buffer
 688                         r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
 689                                                 4096, fb_idx);
 690                         if (r)
 691                                 goto out;
 692                         break;
 693
 694                 default:
 695                         DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
 696                         r = -EINVAL;
 697                         goto out;
 698                 }
 699
 700                 if (session_idx == -1) {
 701                         DRM_ERROR("no session command at start of IB\n");
 702                         r = -EINVAL;
 703                         goto out;
 704                 }
 705
 706                 idx += len / 4;
 707         }
 708
 709         if (allocated && !created) {
 710                 DRM_ERROR("New session without create command!\n");
 711                 r = -ENOENT;
 712         }
 713
 714 out:
 715         if ((!r && destroyed) || (r && allocated)) {
 716                 /*
 717                  * IB contains a destroy msg or we have allocated an
 718                  * handle and got an error, anyway free the handle
 719                  */
 720                 for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
 721                         atomic_cmpxchg(&p->adev->vce.handles[i], handle, 0);
 722         }
 723
 724         return r;
 725 }
 726
 727 /**
 728  * amdgpu_vce_ring_emit_semaphore - emit a semaphore command
 729  *
 730  * @ring: engine to use
 731  * @semaphore: address of semaphore
 732  * @emit_wait: true=emit wait, false=emit signal
 733  *
 734  */
 735 bool amdgpu_vce_ring_emit_semaphore(struct amdgpu_ring *ring,
 736                                     struct amdgpu_semaphore *semaphore,
 737                                     bool emit_wait)
 738 {
 739         uint64_t addr = semaphore->gpu_addr;
 740
 741         amdgpu_ring_write(ring, VCE_CMD_SEMAPHORE);
 742         amdgpu_ring_write(ring, (addr >> 3) & 0x000FFFFF);
 743         amdgpu_ring_write(ring, (addr >> 23) & 0x000FFFFF);
 744         amdgpu_ring_write(ring, 0x01003000 | (emit_wait ? 1 : 0));
 745         if (!emit_wait)
 746                 amdgpu_ring_write(ring, VCE_CMD_END);
 747
 748         return true;
 749 }
 750
 751 /**
 752  * amdgpu_vce_ring_emit_ib - execute indirect buffer
 753  *
 754  * @ring: engine to use
 755  * @ib: the IB to execute
 756  *
 757  */
 758 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
 759 {
 760         amdgpu_ring_write(ring, VCE_CMD_IB);
 761         amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
 762         amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
 763         amdgpu_ring_write(ring, ib->length_dw);
 764 }
 765
 766 /**
 767  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
 768  *
 769  * @ring: engine to use
 770  * @fence: the fence
 771  *
 772  */
 773 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
 774                                 unsigned flags)
 775 {
 776         WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
 777
 778         amdgpu_ring_write(ring, VCE_CMD_FENCE);
 779         amdgpu_ring_write(ring, addr);
 780         amdgpu_ring_write(ring, upper_32_bits(addr));
 781         amdgpu_ring_write(ring, seq);
 782         amdgpu_ring_write(ring, VCE_CMD_TRAP);
 783         amdgpu_ring_write(ring, VCE_CMD_END);
 784 }
 785
 786 /**
 787  * amdgpu_vce_ring_test_ring - test if VCE ring is working
 788  *
 789  * @ring: the engine to test on
 790  *
 791  */
 792 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
 793 {
 794         struct amdgpu_device *adev = ring->adev;
 795         uint32_t rptr = amdgpu_ring_get_rptr(ring);
 796         unsigned i;
 797         int r;
 798
 799         r = amdgpu_ring_lock(ring, 16);
 800         if (r) {
 801                 DRM_ERROR("amdgpu: vce failed to lock ring %d (%d).\n",
 802                           ring->idx, r);
 803                 return r;
 804         }
 805         amdgpu_ring_write(ring, VCE_CMD_END);
 806         amdgpu_ring_unlock_commit(ring);
 807
 808         for (i = 0; i < adev->usec_timeout; i++) {
 809                 if (amdgpu_ring_get_rptr(ring) != rptr)
 810                         break;
 811                 DRM_UDELAY(1);
 812         }
 813
 814         if (i < adev->usec_timeout) {
 815                 DRM_INFO("ring test on %d succeeded in %d usecs\n",
 816                          ring->idx, i);
 817         } else {
 818                 DRM_ERROR("amdgpu: ring %d test failed\n",
 819                           ring->idx);
 820                 r = -ETIMEDOUT;
 821         }
 822
 823         return r;
 824 }
 825
 826 /**
 827  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
 828  *
 829  * @ring: the engine to test on
 830  *
 831  */
 832 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring)
 833 {
 834         struct fence *fence = NULL;
 835         int r;
 836
 837         r = amdgpu_vce_get_create_msg(ring, 1, NULL);
 838         if (r) {
 839                 DRM_ERROR("amdgpu: failed to get create msg (%d).\n", r);
 840                 goto error;
 841         }
 842
 843         r = amdgpu_vce_get_destroy_msg(ring, 1, &fence);
 844         if (r) {
 845                 DRM_ERROR("amdgpu: failed to get destroy ib (%d).\n", r);
 846                 goto error;
 847         }
 848
 849         r = fence_wait(fence, false);
 850         if (r) {
 851                 DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
 852         } else {
 853                 DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
 854         }
 855 error:
 856         fence_put(fence);
 857         return r;
 858 }