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

   1 /*
   2  * Copyright 2014 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 #include "amdgpu_amdkfd.h"
  24 #include "amd_shared.h"
  25
  26 #include "amdgpu.h"
  27 #include "amdgpu_gfx.h"
  28 #include "amdgpu_dma_buf.h"
  29 #include <linux/module.h>
  30 #include <linux/dma-buf.h>
  31 #include "amdgpu_xgmi.h"
  32
  33 static const unsigned int compute_vmid_bitmap = 0xFF00;
  34
  35 /* Total memory size in system memory and all GPU VRAM. Used to
  36  * estimate worst case amount of memory to reserve for page tables
  37  */
  38 uint64_t amdgpu_amdkfd_total_mem_size;
  39
  40 int amdgpu_amdkfd_init(void)
  41 {
  42         struct sysinfo si;
  43         int ret;
  44
  45         si_meminfo(&si);
  46         amdgpu_amdkfd_total_mem_size = si.totalram - si.totalhigh;
  47         amdgpu_amdkfd_total_mem_size *= si.mem_unit;
  48
  49 #ifdef CONFIG_HSA_AMD
  50         ret = kgd2kfd_init();
  51         amdgpu_amdkfd_gpuvm_init_mem_limits();
  52 #else
  53         ret = -ENOENT;
  54 #endif
  55
  56         return ret;
  57 }
  58
  59 void amdgpu_amdkfd_fini(void)
  60 {
  61         kgd2kfd_exit();
  62 }
  63
  64 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
  65 {
  66         bool vf = amdgpu_sriov_vf(adev);
  67
  68         adev->kfd.dev = kgd2kfd_probe((struct kgd_dev *)adev,
  69                                       adev->pdev, adev->asic_type, vf);
  70
  71         if (adev->kfd.dev)
  72                 amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size;
  73 }
  74
  75 /**
  76  * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
  77  *                                setup amdkfd
  78  *
  79  * @adev: amdgpu_device pointer
  80  * @aperture_base: output returning doorbell aperture base physical address
  81  * @aperture_size: output returning doorbell aperture size in bytes
  82  * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
  83  *
  84  * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
  85  * takes doorbells required for its own rings and reports the setup to amdkfd.
  86  * amdgpu reserved doorbells are at the start of the doorbell aperture.
  87  */
  88 static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
  89                                          phys_addr_t *aperture_base,
  90                                          size_t *aperture_size,
  91                                          size_t *start_offset)
  92 {
  93         /*
  94          * The first num_doorbells are used by amdgpu.
  95          * amdkfd takes whatever's left in the aperture.
  96          */
  97         if (adev->doorbell.size > adev->doorbell.num_doorbells * sizeof(u32)) {
  98                 *aperture_base = adev->doorbell.base;
  99                 *aperture_size = adev->doorbell.size;
 100                 *start_offset = adev->doorbell.num_doorbells * sizeof(u32);
 101         } else {
 102                 *aperture_base = 0;
 103                 *aperture_size = 0;
 104                 *start_offset = 0;
 105         }
 106 }
 107
 108 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
 109 {
 110         int i;
 111         int last_valid_bit;
 112
 113         if (adev->kfd.dev) {
 114                 struct kgd2kfd_shared_resources gpu_resources = {
 115                         .compute_vmid_bitmap = compute_vmid_bitmap,
 116                         .num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec,
 117                         .num_queue_per_pipe = adev->gfx.mec.num_queue_per_pipe,
 118                         .gpuvm_size = min(adev->vm_manager.max_pfn
 119                                           << AMDGPU_GPU_PAGE_SHIFT,
 120                                           AMDGPU_GMC_HOLE_START),
 121                         .drm_render_minor = adev->ddev->render->index,
 122                         .sdma_doorbell_idx = adev->doorbell_index.sdma_engine,
 123
 124                 };
 125
 126                 /* this is going to have a few of the MSBs set that we need to
 127                  * clear
 128                  */
 129                 bitmap_complement(gpu_resources.queue_bitmap,
 130                                   adev->gfx.mec.queue_bitmap,
 131                                   KGD_MAX_QUEUES);
 132
 133                 /* According to linux/bitmap.h we shouldn't use bitmap_clear if
 134                  * nbits is not compile time constant
 135                  */
 136                 last_valid_bit = 1 /* only first MEC can have compute queues */
 137                                 * adev->gfx.mec.num_pipe_per_mec
 138                                 * adev->gfx.mec.num_queue_per_pipe;
 139                 for (i = last_valid_bit; i < KGD_MAX_QUEUES; ++i)
 140                         clear_bit(i, gpu_resources.queue_bitmap);
 141
 142                 amdgpu_doorbell_get_kfd_info(adev,
 143                                 &gpu_resources.doorbell_physical_address,
 144                                 &gpu_resources.doorbell_aperture_size,
 145                                 &gpu_resources.doorbell_start_offset);
 146
 147                 /* Since SOC15, BIF starts to statically use the
 148                  * lower 12 bits of doorbell addresses for routing
 149                  * based on settings in registers like
 150                  * SDMA0_DOORBELL_RANGE etc..
 151                  * In order to route a doorbell to CP engine, the lower
 152                  * 12 bits of its address has to be outside the range
 153                  * set for SDMA, VCN, and IH blocks.
 154                  */
 155                 if (adev->asic_type >= CHIP_VEGA10) {
 156                         gpu_resources.non_cp_doorbells_start =
 157                                         adev->doorbell_index.first_non_cp;
 158                         gpu_resources.non_cp_doorbells_end =
 159                                         adev->doorbell_index.last_non_cp;
 160                 }
 161
 162                 kgd2kfd_device_init(adev->kfd.dev, adev->ddev, &gpu_resources);
 163         }
 164 }
 165
 166 void amdgpu_amdkfd_device_fini(struct amdgpu_device *adev)
 167 {
 168         if (adev->kfd.dev) {
 169                 kgd2kfd_device_exit(adev->kfd.dev);
 170                 adev->kfd.dev = NULL;
 171         }
 172 }
 173
 174 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
 175                 const void *ih_ring_entry)
 176 {
 177         if (adev->kfd.dev)
 178                 kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry);
 179 }
 180
 181 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev)
 182 {
 183         if (adev->kfd.dev)
 184                 kgd2kfd_suspend(adev->kfd.dev);
 185 }
 186
 187 int amdgpu_amdkfd_resume(struct amdgpu_device *adev)
 188 {
 189         int r = 0;
 190
 191         if (adev->kfd.dev)
 192                 r = kgd2kfd_resume(adev->kfd.dev);
 193
 194         return r;
 195 }
 196
 197 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev)
 198 {
 199         int r = 0;
 200
 201         if (adev->kfd.dev)
 202                 r = kgd2kfd_pre_reset(adev->kfd.dev);
 203
 204         return r;
 205 }
 206
 207 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev)
 208 {
 209         int r = 0;
 210
 211         if (adev->kfd.dev)
 212                 r = kgd2kfd_post_reset(adev->kfd.dev);
 213
 214         return r;
 215 }
 216
 217 void amdgpu_amdkfd_gpu_reset(struct kgd_dev *kgd)
 218 {
 219         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 220
 221         if (amdgpu_device_should_recover_gpu(adev))
 222                 amdgpu_device_gpu_recover(adev, NULL);
 223 }
 224
 225 int amdgpu_amdkfd_alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
 226                                 void **mem_obj, uint64_t *gpu_addr,
 227                                 void **cpu_ptr, bool mqd_gfx9)
 228 {
 229         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 230         struct amdgpu_bo *bo = NULL;
 231         struct amdgpu_bo_param bp;
 232         int r;
 233         void *cpu_ptr_tmp = NULL;
 234
 235         memset(&bp, 0, sizeof(bp));
 236         bp.size = size;
 237         bp.byte_align = PAGE_SIZE;
 238         bp.domain = AMDGPU_GEM_DOMAIN_GTT;
 239         bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
 240         bp.type = ttm_bo_type_kernel;
 241         bp.resv = NULL;
 242
 243         if (mqd_gfx9)
 244                 bp.flags |= AMDGPU_GEM_CREATE_MQD_GFX9;
 245
 246         r = amdgpu_bo_create(adev, &bp, &bo);
 247         if (r) {
 248                 dev_err(adev->dev,
 249                         "failed to allocate BO for amdkfd (%d)\n", r);
 250                 return r;
 251         }
 252
 253         /* map the buffer */
 254         r = amdgpu_bo_reserve(bo, true);
 255         if (r) {
 256                 dev_err(adev->dev, "(%d) failed to reserve bo for amdkfd\n", r);
 257                 goto allocate_mem_reserve_bo_failed;
 258         }
 259
 260         r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
 261         if (r) {
 262                 dev_err(adev->dev, "(%d) failed to pin bo for amdkfd\n", r);
 263                 goto allocate_mem_pin_bo_failed;
 264         }
 265
 266         r = amdgpu_ttm_alloc_gart(&bo->tbo);
 267         if (r) {
 268                 dev_err(adev->dev, "%p bind failed\n", bo);
 269                 goto allocate_mem_kmap_bo_failed;
 270         }
 271
 272         r = amdgpu_bo_kmap(bo, &cpu_ptr_tmp);
 273         if (r) {
 274                 dev_err(adev->dev,
 275                         "(%d) failed to map bo to kernel for amdkfd\n", r);
 276                 goto allocate_mem_kmap_bo_failed;
 277         }
 278
 279         *mem_obj = bo;
 280         *gpu_addr = amdgpu_bo_gpu_offset(bo);
 281         *cpu_ptr = cpu_ptr_tmp;
 282
 283         amdgpu_bo_unreserve(bo);
 284
 285         return 0;
 286
 287 allocate_mem_kmap_bo_failed:
 288         amdgpu_bo_unpin(bo);
 289 allocate_mem_pin_bo_failed:
 290         amdgpu_bo_unreserve(bo);
 291 allocate_mem_reserve_bo_failed:
 292         amdgpu_bo_unref(&bo);
 293
 294         return r;
 295 }
 296
 297 void amdgpu_amdkfd_free_gtt_mem(struct kgd_dev *kgd, void *mem_obj)
 298 {
 299         struct amdgpu_bo *bo = (struct amdgpu_bo *) mem_obj;
 300
 301         amdgpu_bo_reserve(bo, true);
 302         amdgpu_bo_kunmap(bo);
 303         amdgpu_bo_unpin(bo);
 304         amdgpu_bo_unreserve(bo);
 305         amdgpu_bo_unref(&(bo));
 306 }
 307
 308 int amdgpu_amdkfd_alloc_gws(struct kgd_dev *kgd, size_t size,
 309                                 void **mem_obj)
 310 {
 311         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 312         struct amdgpu_bo *bo = NULL;
 313         struct amdgpu_bo_param bp;
 314         int r;
 315
 316         memset(&bp, 0, sizeof(bp));
 317         bp.size = size;
 318         bp.byte_align = 1;
 319         bp.domain = AMDGPU_GEM_DOMAIN_GWS;
 320         bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
 321         bp.type = ttm_bo_type_device;
 322         bp.resv = NULL;
 323
 324         r = amdgpu_bo_create(adev, &bp, &bo);
 325         if (r) {
 326                 dev_err(adev->dev,
 327                         "failed to allocate gws BO for amdkfd (%d)\n", r);
 328                 return r;
 329         }
 330
 331         *mem_obj = bo;
 332         return 0;
 333 }
 334
 335 void amdgpu_amdkfd_free_gws(struct kgd_dev *kgd, void *mem_obj)
 336 {
 337         struct amdgpu_bo *bo = (struct amdgpu_bo *)mem_obj;
 338
 339         amdgpu_bo_unref(&bo);
 340 }
 341
 342 uint32_t amdgpu_amdkfd_get_fw_version(struct kgd_dev *kgd,
 343                                       enum kgd_engine_type type)
 344 {
 345         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 346
 347         switch (type) {
 348         case KGD_ENGINE_PFP:
 349                 return adev->gfx.pfp_fw_version;
 350
 351         case KGD_ENGINE_ME:
 352                 return adev->gfx.me_fw_version;
 353
 354         case KGD_ENGINE_CE:
 355                 return adev->gfx.ce_fw_version;
 356
 357         case KGD_ENGINE_MEC1:
 358                 return adev->gfx.mec_fw_version;
 359
 360         case KGD_ENGINE_MEC2:
 361                 return adev->gfx.mec2_fw_version;
 362
 363         case KGD_ENGINE_RLC:
 364                 return adev->gfx.rlc_fw_version;
 365
 366         case KGD_ENGINE_SDMA1:
 367                 return adev->sdma.instance[0].fw_version;
 368
 369         case KGD_ENGINE_SDMA2:
 370                 return adev->sdma.instance[1].fw_version;
 371
 372         default:
 373                 return 0;
 374         }
 375
 376         return 0;
 377 }
 378
 379 void amdgpu_amdkfd_get_local_mem_info(struct kgd_dev *kgd,
 380                                       struct kfd_local_mem_info *mem_info)
 381 {
 382         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 383         uint64_t address_mask = adev->dev->dma_mask ? ~*adev->dev->dma_mask :
 384                                              ~((1ULL << 32) - 1);
 385         resource_size_t aper_limit = adev->gmc.aper_base + adev->gmc.aper_size;
 386
 387         memset(mem_info, 0, sizeof(*mem_info));
 388         if (!(adev->gmc.aper_base & address_mask || aper_limit & address_mask)) {
 389                 mem_info->local_mem_size_public = adev->gmc.visible_vram_size;
 390                 mem_info->local_mem_size_private = adev->gmc.real_vram_size -
 391                                 adev->gmc.visible_vram_size;
 392         } else {
 393                 mem_info->local_mem_size_public = 0;
 394                 mem_info->local_mem_size_private = adev->gmc.real_vram_size;
 395         }
 396         mem_info->vram_width = adev->gmc.vram_width;
 397
 398         pr_debug("Address base: %pap limit %pap public 0x%llx private 0x%llx\n",
 399                         &adev->gmc.aper_base, &aper_limit,
 400                         mem_info->local_mem_size_public,
 401                         mem_info->local_mem_size_private);
 402
 403         if (amdgpu_sriov_vf(adev))
 404                 mem_info->mem_clk_max = adev->clock.default_mclk / 100;
 405         else if (adev->powerplay.pp_funcs) {
 406                 if (amdgpu_emu_mode == 1)
 407                         mem_info->mem_clk_max = 0;
 408                 else
 409                         mem_info->mem_clk_max = amdgpu_dpm_get_mclk(adev, false) / 100;
 410         } else
 411                 mem_info->mem_clk_max = 100;
 412 }
 413
 414 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct kgd_dev *kgd)
 415 {
 416         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 417
 418         if (adev->gfx.funcs->get_gpu_clock_counter)
 419                 return adev->gfx.funcs->get_gpu_clock_counter(adev);
 420         return 0;
 421 }
 422
 423 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct kgd_dev *kgd)
 424 {
 425         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 426
 427         /* the sclk is in quantas of 10kHz */
 428         if (amdgpu_sriov_vf(adev))
 429                 return adev->clock.default_sclk / 100;
 430         else if (adev->powerplay.pp_funcs)
 431                 return amdgpu_dpm_get_sclk(adev, false) / 100;
 432         else
 433                 return 100;
 434 }
 435
 436 void amdgpu_amdkfd_get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info)
 437 {
 438         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 439         struct amdgpu_cu_info acu_info = adev->gfx.cu_info;
 440
 441         memset(cu_info, 0, sizeof(*cu_info));
 442         if (sizeof(cu_info->cu_bitmap) != sizeof(acu_info.bitmap))
 443                 return;
 444
 445         cu_info->cu_active_number = acu_info.number;
 446         cu_info->cu_ao_mask = acu_info.ao_cu_mask;
 447         memcpy(&cu_info->cu_bitmap[0], &acu_info.bitmap[0],
 448                sizeof(acu_info.bitmap));
 449         cu_info->num_shader_engines = adev->gfx.config.max_shader_engines;
 450         cu_info->num_shader_arrays_per_engine = adev->gfx.config.max_sh_per_se;
 451         cu_info->num_cu_per_sh = adev->gfx.config.max_cu_per_sh;
 452         cu_info->simd_per_cu = acu_info.simd_per_cu;
 453         cu_info->max_waves_per_simd = acu_info.max_waves_per_simd;
 454         cu_info->wave_front_size = acu_info.wave_front_size;
 455         cu_info->max_scratch_slots_per_cu = acu_info.max_scratch_slots_per_cu;
 456         cu_info->lds_size = acu_info.lds_size;
 457 }
 458
 459 int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd,
 460                                   struct kgd_dev **dma_buf_kgd,
 461                                   uint64_t *bo_size, void *metadata_buffer,
 462                                   size_t buffer_size, uint32_t *metadata_size,
 463                                   uint32_t *flags)
 464 {
 465         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 466         struct dma_buf *dma_buf;
 467         struct drm_gem_object *obj;
 468         struct amdgpu_bo *bo;
 469         uint64_t metadata_flags;
 470         int r = -EINVAL;
 471
 472         dma_buf = dma_buf_get(dma_buf_fd);
 473         if (IS_ERR(dma_buf))
 474                 return PTR_ERR(dma_buf);
 475
 476         if (dma_buf->ops != &amdgpu_dmabuf_ops)
 477                 /* Can't handle non-graphics buffers */
 478                 goto out_put;
 479
 480         obj = dma_buf->priv;
 481         if (obj->dev->driver != adev->ddev->driver)
 482                 /* Can't handle buffers from different drivers */
 483                 goto out_put;
 484
 485         adev = obj->dev->dev_private;
 486         bo = gem_to_amdgpu_bo(obj);
 487         if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
 488                                     AMDGPU_GEM_DOMAIN_GTT)))
 489                 /* Only VRAM and GTT BOs are supported */
 490                 goto out_put;
 491
 492         r = 0;
 493         if (dma_buf_kgd)
 494                 *dma_buf_kgd = (struct kgd_dev *)adev;
 495         if (bo_size)
 496                 *bo_size = amdgpu_bo_size(bo);
 497         if (metadata_size)
 498                 *metadata_size = bo->metadata_size;
 499         if (metadata_buffer)
 500                 r = amdgpu_bo_get_metadata(bo, metadata_buffer, buffer_size,
 501                                            metadata_size, &metadata_flags);
 502         if (flags) {
 503                 *flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
 504                         ALLOC_MEM_FLAGS_VRAM : ALLOC_MEM_FLAGS_GTT;
 505
 506                 if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
 507                         *flags |= ALLOC_MEM_FLAGS_PUBLIC;
 508         }
 509
 510 out_put:
 511         dma_buf_put(dma_buf);
 512         return r;
 513 }
 514
 515 uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd)
 516 {
 517         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 518
 519         return amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
 520 }
 521
 522 uint64_t amdgpu_amdkfd_get_hive_id(struct kgd_dev *kgd)
 523 {
 524         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 525
 526         return adev->gmc.xgmi.hive_id;
 527 }
 528 uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct kgd_dev *dst, struct kgd_dev *src)
 529 {
 530         struct amdgpu_device *peer_adev = (struct amdgpu_device *)src;
 531         struct amdgpu_device *adev = (struct amdgpu_device *)dst;
 532         int ret = amdgpu_xgmi_get_hops_count(adev, peer_adev);
 533
 534         if (ret < 0) {
 535                 DRM_ERROR("amdgpu: failed to get  xgmi hops count between node %d and %d. ret = %d\n",
 536                         adev->gmc.xgmi.physical_node_id,
 537                         peer_adev->gmc.xgmi.physical_node_id, ret);
 538                 ret = 0;
 539         }
 540         return  (uint8_t)ret;
 541 }
 542
 543 uint64_t amdgpu_amdkfd_get_mmio_remap_phys_addr(struct kgd_dev *kgd)
 544 {
 545         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 546
 547         return adev->rmmio_remap.bus_addr;
 548 }
 549
 550 uint32_t amdgpu_amdkfd_get_num_gws(struct kgd_dev *kgd)
 551 {
 552         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 553
 554         return adev->gds.gws_size;
 555 }
 556
 557 int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine,
 558                                 uint32_t vmid, uint64_t gpu_addr,
 559                                 uint32_t *ib_cmd, uint32_t ib_len)
 560 {
 561         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 562         struct amdgpu_job *job;
 563         struct amdgpu_ib *ib;
 564         struct amdgpu_ring *ring;
 565         struct dma_fence *f = NULL;
 566         int ret;
 567
 568         switch (engine) {
 569         case KGD_ENGINE_MEC1:
 570                 ring = &adev->gfx.compute_ring[0];
 571                 break;
 572         case KGD_ENGINE_SDMA1:
 573                 ring = &adev->sdma.instance[0].ring;
 574                 break;
 575         case KGD_ENGINE_SDMA2:
 576                 ring = &adev->sdma.instance[1].ring;
 577                 break;
 578         default:
 579                 pr_err("Invalid engine in IB submission: %d\n", engine);
 580                 ret = -EINVAL;
 581                 goto err;
 582         }
 583
 584         ret = amdgpu_job_alloc(adev, 1, &job, NULL);
 585         if (ret)
 586                 goto err;
 587
 588         ib = &job->ibs[0];
 589         memset(ib, 0, sizeof(struct amdgpu_ib));
 590
 591         ib->gpu_addr = gpu_addr;
 592         ib->ptr = ib_cmd;
 593         ib->length_dw = ib_len;
 594         /* This works for NO_HWS. TODO: need to handle without knowing VMID */
 595         job->vmid = vmid;
 596
 597         ret = amdgpu_ib_schedule(ring, 1, ib, job, &f);
 598         if (ret) {
 599                 DRM_ERROR("amdgpu: failed to schedule IB.\n");
 600                 goto err_ib_sched;
 601         }
 602
 603         ret = dma_fence_wait(f, false);
 604
 605 err_ib_sched:
 606         dma_fence_put(f);
 607         amdgpu_job_free(job);
 608 err:
 609         return ret;
 610 }
 611
 612 void amdgpu_amdkfd_set_compute_idle(struct kgd_dev *kgd, bool idle)
 613 {
 614         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 615
 616         amdgpu_dpm_switch_power_profile(adev,
 617                                         PP_SMC_POWER_PROFILE_COMPUTE,
 618                                         !idle);
 619 }
 620
 621 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid)
 622 {
 623         if (adev->kfd.dev) {
 624                 if ((1 << vmid) & compute_vmid_bitmap)
 625                         return true;
 626         }
 627
 628         return false;
 629 }
 630
 631 int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct kgd_dev *kgd, uint16_t vmid)
 632 {
 633         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 634
 635         if (adev->family == AMDGPU_FAMILY_AI) {
 636                 int i;
 637
 638                 for (i = 0; i < adev->num_vmhubs; i++)
 639                         amdgpu_gmc_flush_gpu_tlb(adev, vmid, i, 0);
 640         } else {
 641                 amdgpu_gmc_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB_0, 0);
 642         }
 643
 644         return 0;
 645 }
 646
 647 int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct kgd_dev *kgd, uint16_t pasid)
 648 {
 649         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 650         uint32_t flush_type = 0;
 651         bool all_hub = false;
 652
 653         if (adev->gmc.xgmi.num_physical_nodes &&
 654                 adev->asic_type == CHIP_VEGA20)
 655                 flush_type = 2;
 656
 657         if (adev->family == AMDGPU_FAMILY_AI)
 658                 all_hub = true;
 659
 660         return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub);
 661 }
 662
 663 bool amdgpu_amdkfd_have_atomics_support(struct kgd_dev *kgd)
 664 {
 665         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 666
 667         return adev->have_atomics_support;
 668 }
 669
 670 #ifndef CONFIG_HSA_AMD
 671 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
 672 {
 673         return false;
 674 }
 675
 676 void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo)
 677 {
 678 }
 679
 680 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
 681                                         struct amdgpu_vm *vm)
 682 {
 683 }
 684
 685 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f)
 686 {
 687         return NULL;
 688 }
 689
 690 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm)
 691 {
 692         return 0;
 693 }
 694
 695 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev,
 696                               unsigned int asic_type, bool vf)
 697 {
 698         return NULL;
 699 }
 700
 701 bool kgd2kfd_device_init(struct kfd_dev *kfd,
 702                          struct drm_device *ddev,
 703                          const struct kgd2kfd_shared_resources *gpu_resources)
 704 {
 705         return false;
 706 }
 707
 708 void kgd2kfd_device_exit(struct kfd_dev *kfd)
 709 {
 710 }
 711
 712 void kgd2kfd_exit(void)
 713 {
 714 }
 715
 716 void kgd2kfd_suspend(struct kfd_dev *kfd)
 717 {
 718 }
 719
 720 int kgd2kfd_resume(struct kfd_dev *kfd)
 721 {
 722         return 0;
 723 }
 724
 725 int kgd2kfd_pre_reset(struct kfd_dev *kfd)
 726 {
 727         return 0;
 728 }
 729
 730 int kgd2kfd_post_reset(struct kfd_dev *kfd)
 731 {
 732         return 0;
 733 }
 734
 735 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
 736 {
 737 }
 738
 739 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd)
 740 {
 741 }
 742 #endif