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

   1 // SPDX-License-Identifier: MIT
   2 /*
   3  * Copyright 2014 Advanced Micro Devices, Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be included in
  13  * all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  21  * OTHER DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #include "amdgpu_amdkfd.h"
  25 #include "amd_pcie.h"
  26 #include "amd_shared.h"
  27
  28 #include "amdgpu.h"
  29 #include "amdgpu_gfx.h"
  30 #include "amdgpu_dma_buf.h"
  31 #include <linux/module.h>
  32 #include <linux/dma-buf.h>
  33 #include "amdgpu_xgmi.h"
  34 #include <uapi/linux/kfd_ioctl.h>
  35 #include "amdgpu_ras.h"
  36 #include "amdgpu_umc.h"
  37 #include "amdgpu_reset.h"
  38
  39 /* Total memory size in system memory and all GPU VRAM. Used to
  40  * estimate worst case amount of memory to reserve for page tables
  41  */
  42 uint64_t amdgpu_amdkfd_total_mem_size;
  43
  44 static bool kfd_initialized;
  45
  46 int amdgpu_amdkfd_init(void)
  47 {
  48         struct sysinfo si;
  49         int ret;
  50
  51         si_meminfo(&si);
  52         amdgpu_amdkfd_total_mem_size = si.freeram - si.freehigh;
  53         amdgpu_amdkfd_total_mem_size *= si.mem_unit;
  54
  55         ret = kgd2kfd_init();
  56         amdgpu_amdkfd_gpuvm_init_mem_limits();
  57         kfd_initialized = !ret;
  58
  59         return ret;
  60 }
  61
  62 void amdgpu_amdkfd_fini(void)
  63 {
  64         if (kfd_initialized) {
  65                 kgd2kfd_exit();
  66                 kfd_initialized = false;
  67         }
  68 }
  69
  70 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
  71 {
  72         bool vf = amdgpu_sriov_vf(adev);
  73
  74         if (!kfd_initialized)
  75                 return;
  76
  77         adev->kfd.dev = kgd2kfd_probe(adev, vf);
  78
  79         if (adev->kfd.dev)
  80                 amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size;
  81 }
  82
  83 /**
  84  * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
  85  *                                setup amdkfd
  86  *
  87  * @adev: amdgpu_device pointer
  88  * @aperture_base: output returning doorbell aperture base physical address
  89  * @aperture_size: output returning doorbell aperture size in bytes
  90  * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
  91  *
  92  * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
  93  * takes doorbells required for its own rings and reports the setup to amdkfd.
  94  * amdgpu reserved doorbells are at the start of the doorbell aperture.
  95  */
  96 static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
  97                                          phys_addr_t *aperture_base,
  98                                          size_t *aperture_size,
  99                                          size_t *start_offset)
 100 {
 101         /*
 102          * The first num_doorbells are used by amdgpu.
 103          * amdkfd takes whatever's left in the aperture.
 104          */
 105         if (adev->enable_mes) {
 106                 /*
 107                  * With MES enabled, we only need to initialize
 108                  * the base address. The size and offset are
 109                  * not initialized as AMDGPU manages the whole
 110                  * doorbell space.
 111                  */
 112                 *aperture_base = adev->doorbell.base;
 113                 *aperture_size = 0;
 114                 *start_offset = 0;
 115         } else if (adev->doorbell.size > adev->doorbell.num_doorbells *
 116                                                 sizeof(u32)) {
 117                 *aperture_base = adev->doorbell.base;
 118                 *aperture_size = adev->doorbell.size;
 119                 *start_offset = adev->doorbell.num_doorbells * sizeof(u32);
 120         } else {
 121                 *aperture_base = 0;
 122                 *aperture_size = 0;
 123                 *start_offset = 0;
 124         }
 125 }
 126
 127
 128 static void amdgpu_amdkfd_reset_work(struct work_struct *work)
 129 {
 130         struct amdgpu_device *adev = container_of(work, struct amdgpu_device,
 131                                                   kfd.reset_work);
 132
 133         struct amdgpu_reset_context reset_context;
 134
 135         memset(&reset_context, 0, sizeof(reset_context));
 136
 137         reset_context.method = AMD_RESET_METHOD_NONE;
 138         reset_context.reset_req_dev = adev;
 139         clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
 140         clear_bit(AMDGPU_SKIP_MODE2_RESET, &reset_context.flags);
 141
 142         amdgpu_device_gpu_recover(adev, NULL, &reset_context);
 143 }
 144
 145 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
 146 {
 147         int i;
 148         int last_valid_bit;
 149
 150         if (adev->kfd.dev) {
 151                 struct kgd2kfd_shared_resources gpu_resources = {
 152                         .compute_vmid_bitmap =
 153                                 ((1 << AMDGPU_NUM_VMID) - 1) -
 154                                 ((1 << adev->vm_manager.first_kfd_vmid) - 1),
 155                         .num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec,
 156                         .num_queue_per_pipe = adev->gfx.mec.num_queue_per_pipe,
 157                         .gpuvm_size = min(adev->vm_manager.max_pfn
 158                                           << AMDGPU_GPU_PAGE_SHIFT,
 159                                           AMDGPU_GMC_HOLE_START),
 160                         .drm_render_minor = adev_to_drm(adev)->render->index,
 161                         .sdma_doorbell_idx = adev->doorbell_index.sdma_engine,
 162                         .enable_mes = adev->enable_mes,
 163                 };
 164
 165                 /* this is going to have a few of the MSBs set that we need to
 166                  * clear
 167                  */
 168                 bitmap_complement(gpu_resources.cp_queue_bitmap,
 169                                   adev->gfx.mec.queue_bitmap,
 170                                   KGD_MAX_QUEUES);
 171
 172                 /* According to linux/bitmap.h we shouldn't use bitmap_clear if
 173                  * nbits is not compile time constant
 174                  */
 175                 last_valid_bit = 1 /* only first MEC can have compute queues */
 176                                 * adev->gfx.mec.num_pipe_per_mec
 177                                 * adev->gfx.mec.num_queue_per_pipe;
 178                 for (i = last_valid_bit; i < KGD_MAX_QUEUES; ++i)
 179                         clear_bit(i, gpu_resources.cp_queue_bitmap);
 180
 181                 amdgpu_doorbell_get_kfd_info(adev,
 182                                 &gpu_resources.doorbell_physical_address,
 183                                 &gpu_resources.doorbell_aperture_size,
 184                                 &gpu_resources.doorbell_start_offset);
 185
 186                 /* Since SOC15, BIF starts to statically use the
 187                  * lower 12 bits of doorbell addresses for routing
 188                  * based on settings in registers like
 189                  * SDMA0_DOORBELL_RANGE etc..
 190                  * In order to route a doorbell to CP engine, the lower
 191                  * 12 bits of its address has to be outside the range
 192                  * set for SDMA, VCN, and IH blocks.
 193                  */
 194                 if (adev->asic_type >= CHIP_VEGA10) {
 195                         gpu_resources.non_cp_doorbells_start =
 196                                         adev->doorbell_index.first_non_cp;
 197                         gpu_resources.non_cp_doorbells_end =
 198                                         adev->doorbell_index.last_non_cp;
 199                 }
 200
 201                 adev->kfd.init_complete = kgd2kfd_device_init(adev->kfd.dev,
 202                                                 adev_to_drm(adev), &gpu_resources);
 203
 204                 INIT_WORK(&adev->kfd.reset_work, amdgpu_amdkfd_reset_work);
 205         }
 206 }
 207
 208 void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev)
 209 {
 210         if (adev->kfd.dev) {
 211                 kgd2kfd_device_exit(adev->kfd.dev);
 212                 adev->kfd.dev = NULL;
 213         }
 214 }
 215
 216 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
 217                 const void *ih_ring_entry)
 218 {
 219         if (adev->kfd.dev)
 220                 kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry);
 221 }
 222
 223 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm)
 224 {
 225         if (adev->kfd.dev)
 226                 kgd2kfd_suspend(adev->kfd.dev, run_pm);
 227 }
 228
 229 int amdgpu_amdkfd_resume_iommu(struct amdgpu_device *adev)
 230 {
 231         int r = 0;
 232
 233         if (adev->kfd.dev)
 234                 r = kgd2kfd_resume_iommu(adev->kfd.dev);
 235
 236         return r;
 237 }
 238
 239 int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm)
 240 {
 241         int r = 0;
 242
 243         if (adev->kfd.dev)
 244                 r = kgd2kfd_resume(adev->kfd.dev, run_pm);
 245
 246         return r;
 247 }
 248
 249 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev)
 250 {
 251         int r = 0;
 252
 253         if (adev->kfd.dev)
 254                 r = kgd2kfd_pre_reset(adev->kfd.dev);
 255
 256         return r;
 257 }
 258
 259 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev)
 260 {
 261         int r = 0;
 262
 263         if (adev->kfd.dev)
 264                 r = kgd2kfd_post_reset(adev->kfd.dev);
 265
 266         return r;
 267 }
 268
 269 void amdgpu_amdkfd_gpu_reset(struct amdgpu_device *adev)
 270 {
 271         if (amdgpu_device_should_recover_gpu(adev))
 272                 amdgpu_reset_domain_schedule(adev->reset_domain,
 273                                              &adev->kfd.reset_work);
 274 }
 275
 276 int amdgpu_amdkfd_alloc_gtt_mem(struct amdgpu_device *adev, size_t size,
 277                                 void **mem_obj, uint64_t *gpu_addr,
 278                                 void **cpu_ptr, bool cp_mqd_gfx9)
 279 {
 280         struct amdgpu_bo *bo = NULL;
 281         struct amdgpu_bo_param bp;
 282         int r;
 283         void *cpu_ptr_tmp = NULL;
 284
 285         memset(&bp, 0, sizeof(bp));
 286         bp.size = size;
 287         bp.byte_align = PAGE_SIZE;
 288         bp.domain = AMDGPU_GEM_DOMAIN_GTT;
 289         bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
 290         bp.type = ttm_bo_type_kernel;
 291         bp.resv = NULL;
 292         bp.bo_ptr_size = sizeof(struct amdgpu_bo);
 293
 294         if (cp_mqd_gfx9)
 295                 bp.flags |= AMDGPU_GEM_CREATE_CP_MQD_GFX9;
 296
 297         r = amdgpu_bo_create(adev, &bp, &bo);
 298         if (r) {
 299                 dev_err(adev->dev,
 300                         "failed to allocate BO for amdkfd (%d)\n", r);
 301                 return r;
 302         }
 303
 304         /* map the buffer */
 305         r = amdgpu_bo_reserve(bo, true);
 306         if (r) {
 307                 dev_err(adev->dev, "(%d) failed to reserve bo for amdkfd\n", r);
 308                 goto allocate_mem_reserve_bo_failed;
 309         }
 310
 311         r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
 312         if (r) {
 313                 dev_err(adev->dev, "(%d) failed to pin bo for amdkfd\n", r);
 314                 goto allocate_mem_pin_bo_failed;
 315         }
 316
 317         r = amdgpu_ttm_alloc_gart(&bo->tbo);
 318         if (r) {
 319                 dev_err(adev->dev, "%p bind failed\n", bo);
 320                 goto allocate_mem_kmap_bo_failed;
 321         }
 322
 323         r = amdgpu_bo_kmap(bo, &cpu_ptr_tmp);
 324         if (r) {
 325                 dev_err(adev->dev,
 326                         "(%d) failed to map bo to kernel for amdkfd\n", r);
 327                 goto allocate_mem_kmap_bo_failed;
 328         }
 329
 330         *mem_obj = bo;
 331         *gpu_addr = amdgpu_bo_gpu_offset(bo);
 332         *cpu_ptr = cpu_ptr_tmp;
 333
 334         amdgpu_bo_unreserve(bo);
 335
 336         return 0;
 337
 338 allocate_mem_kmap_bo_failed:
 339         amdgpu_bo_unpin(bo);
 340 allocate_mem_pin_bo_failed:
 341         amdgpu_bo_unreserve(bo);
 342 allocate_mem_reserve_bo_failed:
 343         amdgpu_bo_unref(&bo);
 344
 345         return r;
 346 }
 347
 348 void amdgpu_amdkfd_free_gtt_mem(struct amdgpu_device *adev, void *mem_obj)
 349 {
 350         struct amdgpu_bo *bo = (struct amdgpu_bo *) mem_obj;
 351
 352         amdgpu_bo_reserve(bo, true);
 353         amdgpu_bo_kunmap(bo);
 354         amdgpu_bo_unpin(bo);
 355         amdgpu_bo_unreserve(bo);
 356         amdgpu_bo_unref(&(bo));
 357 }
 358
 359 int amdgpu_amdkfd_alloc_gws(struct amdgpu_device *adev, size_t size,
 360                                 void **mem_obj)
 361 {
 362         struct amdgpu_bo *bo = NULL;
 363         struct amdgpu_bo_user *ubo;
 364         struct amdgpu_bo_param bp;
 365         int r;
 366
 367         memset(&bp, 0, sizeof(bp));
 368         bp.size = size;
 369         bp.byte_align = 1;
 370         bp.domain = AMDGPU_GEM_DOMAIN_GWS;
 371         bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
 372         bp.type = ttm_bo_type_device;
 373         bp.resv = NULL;
 374         bp.bo_ptr_size = sizeof(struct amdgpu_bo);
 375
 376         r = amdgpu_bo_create_user(adev, &bp, &ubo);
 377         if (r) {
 378                 dev_err(adev->dev,
 379                         "failed to allocate gws BO for amdkfd (%d)\n", r);
 380                 return r;
 381         }
 382
 383         bo = &ubo->bo;
 384         *mem_obj = bo;
 385         return 0;
 386 }
 387
 388 void amdgpu_amdkfd_free_gws(struct amdgpu_device *adev, void *mem_obj)
 389 {
 390         struct amdgpu_bo *bo = (struct amdgpu_bo *)mem_obj;
 391
 392         amdgpu_bo_unref(&bo);
 393 }
 394
 395 uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev,
 396                                       enum kgd_engine_type type)
 397 {
 398         switch (type) {
 399         case KGD_ENGINE_PFP:
 400                 return adev->gfx.pfp_fw_version;
 401
 402         case KGD_ENGINE_ME:
 403                 return adev->gfx.me_fw_version;
 404
 405         case KGD_ENGINE_CE:
 406                 return adev->gfx.ce_fw_version;
 407
 408         case KGD_ENGINE_MEC1:
 409                 return adev->gfx.mec_fw_version;
 410
 411         case KGD_ENGINE_MEC2:
 412                 return adev->gfx.mec2_fw_version;
 413
 414         case KGD_ENGINE_RLC:
 415                 return adev->gfx.rlc_fw_version;
 416
 417         case KGD_ENGINE_SDMA1:
 418                 return adev->sdma.instance[0].fw_version;
 419
 420         case KGD_ENGINE_SDMA2:
 421                 return adev->sdma.instance[1].fw_version;
 422
 423         default:
 424                 return 0;
 425         }
 426
 427         return 0;
 428 }
 429
 430 void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev,
 431                                       struct kfd_local_mem_info *mem_info)
 432 {
 433         memset(mem_info, 0, sizeof(*mem_info));
 434
 435         mem_info->local_mem_size_public = adev->gmc.visible_vram_size;
 436         mem_info->local_mem_size_private = adev->gmc.real_vram_size -
 437                                                 adev->gmc.visible_vram_size;
 438
 439         mem_info->vram_width = adev->gmc.vram_width;
 440
 441         pr_debug("Address base: %pap public 0x%llx private 0x%llx\n",
 442                         &adev->gmc.aper_base,
 443                         mem_info->local_mem_size_public,
 444                         mem_info->local_mem_size_private);
 445
 446         if (amdgpu_sriov_vf(adev))
 447                 mem_info->mem_clk_max = adev->clock.default_mclk / 100;
 448         else if (adev->pm.dpm_enabled) {
 449                 if (amdgpu_emu_mode == 1)
 450                         mem_info->mem_clk_max = 0;
 451                 else
 452                         mem_info->mem_clk_max = amdgpu_dpm_get_mclk(adev, false) / 100;
 453         } else
 454                 mem_info->mem_clk_max = 100;
 455 }
 456
 457 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev)
 458 {
 459         if (adev->gfx.funcs->get_gpu_clock_counter)
 460                 return adev->gfx.funcs->get_gpu_clock_counter(adev);
 461         return 0;
 462 }
 463
 464 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev)
 465 {
 466         /* the sclk is in quantas of 10kHz */
 467         if (amdgpu_sriov_vf(adev))
 468                 return adev->clock.default_sclk / 100;
 469         else if (adev->pm.dpm_enabled)
 470                 return amdgpu_dpm_get_sclk(adev, false) / 100;
 471         else
 472                 return 100;
 473 }
 474
 475 void amdgpu_amdkfd_get_cu_info(struct amdgpu_device *adev, struct kfd_cu_info *cu_info)
 476 {
 477         struct amdgpu_cu_info acu_info = adev->gfx.cu_info;
 478
 479         memset(cu_info, 0, sizeof(*cu_info));
 480         if (sizeof(cu_info->cu_bitmap) != sizeof(acu_info.bitmap))
 481                 return;
 482
 483         cu_info->cu_active_number = acu_info.number;
 484         cu_info->cu_ao_mask = acu_info.ao_cu_mask;
 485         memcpy(&cu_info->cu_bitmap[0], &acu_info.bitmap[0],
 486                sizeof(acu_info.bitmap));
 487         cu_info->num_shader_engines = adev->gfx.config.max_shader_engines;
 488         cu_info->num_shader_arrays_per_engine = adev->gfx.config.max_sh_per_se;
 489         cu_info->num_cu_per_sh = adev->gfx.config.max_cu_per_sh;
 490         cu_info->simd_per_cu = acu_info.simd_per_cu;
 491         cu_info->max_waves_per_simd = acu_info.max_waves_per_simd;
 492         cu_info->wave_front_size = acu_info.wave_front_size;
 493         cu_info->max_scratch_slots_per_cu = acu_info.max_scratch_slots_per_cu;
 494         cu_info->lds_size = acu_info.lds_size;
 495 }
 496
 497 int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd,
 498                                   struct amdgpu_device **dmabuf_adev,
 499                                   uint64_t *bo_size, void *metadata_buffer,
 500                                   size_t buffer_size, uint32_t *metadata_size,
 501                                   uint32_t *flags)
 502 {
 503         struct dma_buf *dma_buf;
 504         struct drm_gem_object *obj;
 505         struct amdgpu_bo *bo;
 506         uint64_t metadata_flags;
 507         int r = -EINVAL;
 508
 509         dma_buf = dma_buf_get(dma_buf_fd);
 510         if (IS_ERR(dma_buf))
 511                 return PTR_ERR(dma_buf);
 512
 513         if (dma_buf->ops != &amdgpu_dmabuf_ops)
 514                 /* Can't handle non-graphics buffers */
 515                 goto out_put;
 516
 517         obj = dma_buf->priv;
 518         if (obj->dev->driver != adev_to_drm(adev)->driver)
 519                 /* Can't handle buffers from different drivers */
 520                 goto out_put;
 521
 522         adev = drm_to_adev(obj->dev);
 523         bo = gem_to_amdgpu_bo(obj);
 524         if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
 525                                     AMDGPU_GEM_DOMAIN_GTT)))
 526                 /* Only VRAM and GTT BOs are supported */
 527                 goto out_put;
 528
 529         r = 0;
 530         if (dmabuf_adev)
 531                 *dmabuf_adev = adev;
 532         if (bo_size)
 533                 *bo_size = amdgpu_bo_size(bo);
 534         if (metadata_buffer)
 535                 r = amdgpu_bo_get_metadata(bo, metadata_buffer, buffer_size,
 536                                            metadata_size, &metadata_flags);
 537         if (flags) {
 538                 *flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
 539                                 KFD_IOC_ALLOC_MEM_FLAGS_VRAM
 540                                 : KFD_IOC_ALLOC_MEM_FLAGS_GTT;
 541
 542                 if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
 543                         *flags |= KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC;
 544         }
 545
 546 out_put:
 547         dma_buf_put(dma_buf);
 548         return r;
 549 }
 550
 551 uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct amdgpu_device *dst,
 552                                           struct amdgpu_device *src)
 553 {
 554         struct amdgpu_device *peer_adev = src;
 555         struct amdgpu_device *adev = dst;
 556         int ret = amdgpu_xgmi_get_hops_count(adev, peer_adev);
 557
 558         if (ret < 0) {
 559                 DRM_ERROR("amdgpu: failed to get  xgmi hops count between node %d and %d. ret = %d\n",
 560                         adev->gmc.xgmi.physical_node_id,
 561                         peer_adev->gmc.xgmi.physical_node_id, ret);
 562                 ret = 0;
 563         }
 564         return  (uint8_t)ret;
 565 }
 566
 567 int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct amdgpu_device *dst,
 568                                             struct amdgpu_device *src,
 569                                             bool is_min)
 570 {
 571         struct amdgpu_device *adev = dst, *peer_adev;
 572         int num_links;
 573
 574         if (adev->asic_type != CHIP_ALDEBARAN)
 575                 return 0;
 576
 577         if (src)
 578                 peer_adev = src;
 579
 580         /* num links returns 0 for indirect peers since indirect route is unknown. */
 581         num_links = is_min ? 1 : amdgpu_xgmi_get_num_links(adev, peer_adev);
 582         if (num_links < 0) {
 583                 DRM_ERROR("amdgpu: failed to get xgmi num links between node %d and %d. ret = %d\n",
 584                         adev->gmc.xgmi.physical_node_id,
 585                         peer_adev->gmc.xgmi.physical_node_id, num_links);
 586                 num_links = 0;
 587         }
 588
 589         /* Aldebaran xGMI DPM is defeatured so assume x16 x 25Gbps for bandwidth. */
 590         return (num_links * 16 * 25000)/BITS_PER_BYTE;
 591 }
 592
 593 int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool is_min)
 594 {
 595         int num_lanes_shift = (is_min ? ffs(adev->pm.pcie_mlw_mask) :
 596                                                         fls(adev->pm.pcie_mlw_mask)) - 1;
 597         int gen_speed_shift = (is_min ? ffs(adev->pm.pcie_gen_mask &
 598                                                 CAIL_PCIE_LINK_SPEED_SUPPORT_MASK) :
 599                                         fls(adev->pm.pcie_gen_mask &
 600                                                 CAIL_PCIE_LINK_SPEED_SUPPORT_MASK)) - 1;
 601         uint32_t num_lanes_mask = 1 << num_lanes_shift;
 602         uint32_t gen_speed_mask = 1 << gen_speed_shift;
 603         int num_lanes_factor = 0, gen_speed_mbits_factor = 0;
 604
 605         switch (num_lanes_mask) {
 606         case CAIL_PCIE_LINK_WIDTH_SUPPORT_X1:
 607                 num_lanes_factor = 1;
 608                 break;
 609         case CAIL_PCIE_LINK_WIDTH_SUPPORT_X2:
 610                 num_lanes_factor = 2;
 611                 break;
 612         case CAIL_PCIE_LINK_WIDTH_SUPPORT_X4:
 613                 num_lanes_factor = 4;
 614                 break;
 615         case CAIL_PCIE_LINK_WIDTH_SUPPORT_X8:
 616                 num_lanes_factor = 8;
 617                 break;
 618         case CAIL_PCIE_LINK_WIDTH_SUPPORT_X12:
 619                 num_lanes_factor = 12;
 620                 break;
 621         case CAIL_PCIE_LINK_WIDTH_SUPPORT_X16:
 622                 num_lanes_factor = 16;
 623                 break;
 624         case CAIL_PCIE_LINK_WIDTH_SUPPORT_X32:
 625                 num_lanes_factor = 32;
 626                 break;
 627         }
 628
 629         switch (gen_speed_mask) {
 630         case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1:
 631                 gen_speed_mbits_factor = 2500;
 632                 break;
 633         case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2:
 634                 gen_speed_mbits_factor = 5000;
 635                 break;
 636         case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3:
 637                 gen_speed_mbits_factor = 8000;
 638                 break;
 639         case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4:
 640                 gen_speed_mbits_factor = 16000;
 641                 break;
 642         case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5:
 643                 gen_speed_mbits_factor = 32000;
 644                 break;
 645         }
 646
 647         return (num_lanes_factor * gen_speed_mbits_factor)/BITS_PER_BYTE;
 648 }
 649
 650 int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev,
 651                                 enum kgd_engine_type engine,
 652                                 uint32_t vmid, uint64_t gpu_addr,
 653                                 uint32_t *ib_cmd, uint32_t ib_len)
 654 {
 655         struct amdgpu_job *job;
 656         struct amdgpu_ib *ib;
 657         struct amdgpu_ring *ring;
 658         struct dma_fence *f = NULL;
 659         int ret;
 660
 661         switch (engine) {
 662         case KGD_ENGINE_MEC1:
 663                 ring = &adev->gfx.compute_ring[0];
 664                 break;
 665         case KGD_ENGINE_SDMA1:
 666                 ring = &adev->sdma.instance[0].ring;
 667                 break;
 668         case KGD_ENGINE_SDMA2:
 669                 ring = &adev->sdma.instance[1].ring;
 670                 break;
 671         default:
 672                 pr_err("Invalid engine in IB submission: %d\n", engine);
 673                 ret = -EINVAL;
 674                 goto err;
 675         }
 676
 677         ret = amdgpu_job_alloc(adev, 1, &job, NULL);
 678         if (ret)
 679                 goto err;
 680
 681         ib = &job->ibs[0];
 682         memset(ib, 0, sizeof(struct amdgpu_ib));
 683
 684         ib->gpu_addr = gpu_addr;
 685         ib->ptr = ib_cmd;
 686         ib->length_dw = ib_len;
 687         /* This works for NO_HWS. TODO: need to handle without knowing VMID */
 688         job->vmid = vmid;
 689         job->num_ibs = 1;
 690
 691         ret = amdgpu_ib_schedule(ring, 1, ib, job, &f);
 692
 693         if (ret) {
 694                 DRM_ERROR("amdgpu: failed to schedule IB.\n");
 695                 goto err_ib_sched;
 696         }
 697
 698         /* Drop the initial kref_init count (see drm_sched_main as example) */
 699         dma_fence_put(f);
 700         ret = dma_fence_wait(f, false);
 701
 702 err_ib_sched:
 703         amdgpu_job_free(job);
 704 err:
 705         return ret;
 706 }
 707
 708 void amdgpu_amdkfd_set_compute_idle(struct amdgpu_device *adev, bool idle)
 709 {
 710         amdgpu_dpm_switch_power_profile(adev,
 711                                         PP_SMC_POWER_PROFILE_COMPUTE,
 712                                         !idle);
 713 }
 714
 715 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid)
 716 {
 717         if (adev->kfd.dev)
 718                 return vmid >= adev->vm_manager.first_kfd_vmid;
 719
 720         return false;
 721 }
 722
 723 int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct amdgpu_device *adev,
 724                                      uint16_t vmid)
 725 {
 726         if (adev->family == AMDGPU_FAMILY_AI) {
 727                 int i;
 728
 729                 for (i = 0; i < adev->num_vmhubs; i++)
 730                         amdgpu_gmc_flush_gpu_tlb(adev, vmid, i, 0);
 731         } else {
 732                 amdgpu_gmc_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB_0, 0);
 733         }
 734
 735         return 0;
 736 }
 737
 738 int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
 739                                       uint16_t pasid, enum TLB_FLUSH_TYPE flush_type)
 740 {
 741         bool all_hub = false;
 742
 743         if (adev->family == AMDGPU_FAMILY_AI ||
 744             adev->family == AMDGPU_FAMILY_RV)
 745                 all_hub = true;
 746
 747         return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub);
 748 }
 749
 750 bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev)
 751 {
 752         return adev->have_atomics_support;
 753 }
 754
 755 void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, bool reset)
 756 {
 757         struct ras_err_data err_data = {0, 0, 0, NULL};
 758
 759         amdgpu_umc_poison_handler(adev, &err_data, reset);
 760 }
 761
 762 bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev)
 763 {
 764         if (adev->gfx.ras && adev->gfx.ras->query_utcl2_poison_status)
 765                 return adev->gfx.ras->query_utcl2_poison_status(adev);
 766         else
 767                 return false;
 768 }