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 <drm/ttm/ttm_tt.h> 32 #include <linux/module.h> 33 #include <linux/dma-buf.h> 34 #include "amdgpu_xgmi.h" 35 #include <uapi/linux/kfd_ioctl.h> 36 #include "amdgpu_ras.h" 37 #include "amdgpu_umc.h" 38 #include "amdgpu_reset.h" 39 #if IS_ENABLED(CONFIG_HSA_AMD) 40 #include "kfd_priv.h" 41 #endif 42 #include "kfd_svm.h" 43 44 /* Total memory size in system memory and all GPU VRAM. Used to 45 * estimate worst case amount of memory to reserve for page tables 46 */ 47 uint64_t amdgpu_amdkfd_total_mem_size; 48 49 static bool kfd_initialized; 50 51 int amdgpu_amdkfd_init(void) 52 { 53 struct sysinfo si; 54 int ret; 55 56 si_meminfo(&si); 57 amdgpu_amdkfd_total_mem_size = si.freeram - si.freehigh; 58 amdgpu_amdkfd_total_mem_size *= si.mem_unit; 59 60 ret = kgd2kfd_init(); 61 kfd_initialized = !ret; 62 63 return ret; 64 } 65 66 void amdgpu_amdkfd_fini(void) 67 { 68 if (kfd_initialized) { 69 kgd2kfd_exit(); 70 kfd_initialized = false; 71 } 72 } 73 74 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev) 75 { 76 bool vf = amdgpu_sriov_vf(adev); 77 78 if (!kfd_initialized) 79 return; 80 81 adev->kfd.dev = kgd2kfd_probe(adev, vf); 82 } 83 84 /** 85 * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to 86 * setup amdkfd 87 * 88 * @adev: amdgpu_device pointer 89 * @aperture_base: output returning doorbell aperture base physical address 90 * @aperture_size: output returning doorbell aperture size in bytes 91 * @start_offset: output returning # of doorbell bytes reserved for amdgpu. 92 * 93 * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up, 94 * takes doorbells required for its own rings and reports the setup to amdkfd. 95 * amdgpu reserved doorbells are at the start of the doorbell aperture. 96 */ 97 static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev, 98 phys_addr_t *aperture_base, 99 size_t *aperture_size, 100 size_t *start_offset) 101 { 102 /* 103 * The first num_kernel_doorbells are used by amdgpu. 104 * amdkfd takes whatever's left in the aperture. 105 */ 106 if (adev->enable_mes) { 107 /* 108 * With MES enabled, we only need to initialize 109 * the base address. The size and offset are 110 * not initialized as AMDGPU manages the whole 111 * doorbell space. 112 */ 113 *aperture_base = adev->doorbell.base; 114 *aperture_size = 0; 115 *start_offset = 0; 116 } else if (adev->doorbell.size > adev->doorbell.num_kernel_doorbells * 117 sizeof(u32)) { 118 *aperture_base = adev->doorbell.base; 119 *aperture_size = adev->doorbell.size; 120 *start_offset = adev->doorbell.num_kernel_doorbells * sizeof(u32); 121 } else { 122 *aperture_base = 0; 123 *aperture_size = 0; 124 *start_offset = 0; 125 } 126 } 127 128 129 static void amdgpu_amdkfd_reset_work(struct work_struct *work) 130 { 131 struct amdgpu_device *adev = container_of(work, struct amdgpu_device, 132 kfd.reset_work); 133 134 struct amdgpu_reset_context reset_context; 135 136 memset(&reset_context, 0, sizeof(reset_context)); 137 138 reset_context.method = AMD_RESET_METHOD_NONE; 139 reset_context.reset_req_dev = adev; 140 reset_context.src = adev->enable_mes ? 141 AMDGPU_RESET_SRC_MES : 142 AMDGPU_RESET_SRC_HWS; 143 clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); 144 145 amdgpu_device_gpu_recover(adev, NULL, &reset_context); 146 } 147 148 static const struct drm_client_funcs kfd_client_funcs = { 149 .unregister = drm_client_release, 150 }; 151 152 int amdgpu_amdkfd_drm_client_create(struct amdgpu_device *adev) 153 { 154 int ret; 155 156 if (!adev->kfd.init_complete || adev->kfd.client.dev) 157 return 0; 158 159 ret = drm_client_init(&adev->ddev, &adev->kfd.client, "kfd", 160 &kfd_client_funcs); 161 if (ret) { 162 dev_err(adev->dev, "Failed to init DRM client: %d\n", 163 ret); 164 return ret; 165 } 166 167 drm_client_register(&adev->kfd.client); 168 169 return 0; 170 } 171 172 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) 173 { 174 int i; 175 int last_valid_bit; 176 177 amdgpu_amdkfd_gpuvm_init_mem_limits(); 178 179 if (adev->kfd.dev) { 180 struct kgd2kfd_shared_resources gpu_resources = { 181 .compute_vmid_bitmap = adev->vm_manager.vmid_uq_mask_gfxhub, 182 .num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec, 183 .num_queue_per_pipe = adev->gfx.mec.num_queue_per_pipe, 184 .gpuvm_size = min(adev->vm_manager.max_pfn 185 << AMDGPU_GPU_PAGE_SHIFT, 186 AMDGPU_GMC_HOLE_START), 187 .drm_render_minor = adev_to_drm(adev)->render->index, 188 .sdma_doorbell_idx = adev->doorbell_index.sdma_engine, 189 .enable_mes = adev->enable_mes, 190 }; 191 192 /* this is going to have a few of the MSBs set that we need to 193 * clear 194 */ 195 bitmap_complement(gpu_resources.cp_queue_bitmap, 196 adev->gfx.mec_bitmap[0].queue_bitmap, 197 AMDGPU_MAX_QUEUES); 198 199 /* According to linux/bitmap.h we shouldn't use bitmap_clear if 200 * nbits is not compile time constant 201 */ 202 last_valid_bit = 1 /* only first MEC can have compute queues */ 203 * adev->gfx.mec.num_pipe_per_mec 204 * adev->gfx.mec.num_queue_per_pipe; 205 for (i = last_valid_bit; i < AMDGPU_MAX_QUEUES; ++i) 206 clear_bit(i, gpu_resources.cp_queue_bitmap); 207 208 amdgpu_doorbell_get_kfd_info(adev, 209 &gpu_resources.doorbell_physical_address, 210 &gpu_resources.doorbell_aperture_size, 211 &gpu_resources.doorbell_start_offset); 212 213 /* Since SOC15, BIF starts to statically use the 214 * lower 12 bits of doorbell addresses for routing 215 * based on settings in registers like 216 * SDMA0_DOORBELL_RANGE etc.. 217 * In order to route a doorbell to CP engine, the lower 218 * 12 bits of its address has to be outside the range 219 * set for SDMA, VCN, and IH blocks. 220 */ 221 if (adev->asic_type >= CHIP_VEGA10) { 222 gpu_resources.non_cp_doorbells_start = 223 adev->doorbell_index.first_non_cp; 224 gpu_resources.non_cp_doorbells_end = 225 adev->doorbell_index.last_non_cp; 226 } 227 228 adev->kfd.init_complete = kgd2kfd_device_init(adev->kfd.dev, 229 &gpu_resources); 230 231 amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size; 232 233 INIT_WORK(&adev->kfd.reset_work, amdgpu_amdkfd_reset_work); 234 } 235 } 236 237 void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev) 238 { 239 if (adev->kfd.dev) { 240 kgd2kfd_device_exit(adev->kfd.dev); 241 adev->kfd.dev = NULL; 242 amdgpu_amdkfd_total_mem_size -= adev->gmc.real_vram_size; 243 } 244 } 245 246 int amdgpu_amdkfd_prepare_partition_switch(struct amdgpu_device *adev) 247 { 248 struct kfd_dev *kfd = adev->kfd.dev; 249 int r; 250 251 if (!kfd) 252 return 0; 253 r = kgd2kfd_check_and_lock_kfd(kfd); 254 if (r) 255 return r; 256 257 adev->kfd.init_complete = false; 258 kgd2kfd_device_fini(kfd); 259 adev->kfd.dev = NULL; 260 amdgpu_amdkfd_total_mem_size -= adev->gmc.real_vram_size; 261 262 return 0; 263 } 264 265 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev, 266 const void *ih_ring_entry) 267 { 268 if (adev->kfd.dev) 269 kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry); 270 } 271 272 void amdgpu_amdkfd_teardown_processes(struct amdgpu_device *adev) 273 { 274 kgd2kfd_teardown_processes(adev); 275 } 276 277 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool suspend_proc) 278 { 279 if (adev->kfd.dev) { 280 if (adev->in_s0ix) 281 kgd2kfd_stop_sched_all_nodes(adev->kfd.dev); 282 else 283 kgd2kfd_suspend(adev->kfd.dev, suspend_proc); 284 } 285 } 286 287 int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool resume_proc) 288 { 289 int r = 0; 290 291 if (adev->kfd.dev) { 292 if (adev->in_s0ix) 293 r = kgd2kfd_start_sched_all_nodes(adev->kfd.dev); 294 else 295 r = kgd2kfd_resume(adev->kfd.dev, resume_proc); 296 } 297 298 return r; 299 } 300 301 void amdgpu_amdkfd_suspend_process(struct amdgpu_device *adev) 302 { 303 if (adev->kfd.dev) 304 kgd2kfd_suspend_process(adev->kfd.dev); 305 } 306 307 int amdgpu_amdkfd_resume_process(struct amdgpu_device *adev) 308 { 309 int r = 0; 310 311 if (adev->kfd.dev) 312 r = kgd2kfd_resume_process(adev->kfd.dev); 313 314 return r; 315 } 316 317 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev, 318 struct amdgpu_reset_context *reset_context) 319 { 320 int r = 0; 321 322 if (adev->kfd.dev) 323 r = kgd2kfd_pre_reset(adev->kfd.dev, reset_context); 324 325 return r; 326 } 327 328 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev) 329 { 330 int r = 0; 331 332 if (adev->kfd.dev) 333 r = kgd2kfd_post_reset(adev->kfd.dev); 334 335 return r; 336 } 337 338 void amdgpu_amdkfd_gpu_reset(struct amdgpu_device *adev) 339 { 340 if (amdgpu_device_should_recover_gpu(adev)) 341 (void)amdgpu_reset_domain_schedule(adev->reset_domain, &adev->kfd.reset_work); 342 } 343 344 void amdgpu_amdkfd_clear_kfd_mapping(struct amdgpu_device *adev) 345 { 346 #if IS_ENABLED(CONFIG_HSA_AMD) 347 struct kfd_dev *kfd = adev->kfd.dev; 348 unsigned int i; 349 350 if (!kfd || !kfd->init_complete) 351 return; 352 353 for (i = 0; i < kfd->num_nodes; i++) { 354 struct kfd_node *node = kfd->nodes[i]; 355 356 kfd_dev_unmap_mapping_range(KFD_MMAP_TYPE_DOORBELL | 357 KFD_MMAP_GPU_ID(node->id), 358 kfd_doorbell_process_slice(kfd)); 359 kfd_dev_unmap_mapping_range(KFD_MMAP_TYPE_MMIO | 360 KFD_MMAP_GPU_ID(node->id), 361 PAGE_SIZE); 362 } 363 #endif 364 } 365 366 int amdgpu_amdkfd_alloc_kernel_mem(struct amdgpu_device *adev, size_t size, 367 u32 domain, void **mem_obj, uint64_t *gpu_addr, 368 void **cpu_ptr, bool cp_mqd_gfx9) 369 { 370 struct amdgpu_bo *bo = NULL; 371 struct amdgpu_bo_param bp; 372 int r; 373 void *cpu_ptr_tmp = NULL; 374 375 memset(&bp, 0, sizeof(bp)); 376 bp.size = size; 377 bp.byte_align = PAGE_SIZE; 378 bp.domain = domain; 379 bp.flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS | 380 AMDGPU_GEM_CREATE_CPU_GTT_USWC; 381 bp.type = ttm_bo_type_kernel; 382 bp.resv = NULL; 383 bp.bo_ptr_size = sizeof(struct amdgpu_bo); 384 385 if (cp_mqd_gfx9) 386 bp.flags |= AMDGPU_GEM_CREATE_CP_MQD_GFX9; 387 388 r = amdgpu_bo_create(adev, &bp, &bo); 389 if (r) { 390 dev_err(adev->dev, 391 "failed to allocate BO for amdkfd (%d)\n", r); 392 return r; 393 } 394 395 /* map the buffer */ 396 r = amdgpu_bo_reserve(bo, true); 397 if (r) { 398 dev_err(adev->dev, "(%d) failed to reserve bo for amdkfd\n", r); 399 goto allocate_mem_reserve_bo_failed; 400 } 401 402 r = amdgpu_bo_pin(bo, domain); 403 if (r) { 404 dev_err(adev->dev, "(%d) failed to pin bo for amdkfd\n", r); 405 goto allocate_mem_pin_bo_failed; 406 } 407 408 r = amdgpu_ttm_alloc_gart(&bo->tbo); 409 if (r) { 410 dev_err(adev->dev, "%p bind failed\n", bo); 411 goto allocate_mem_kmap_bo_failed; 412 } 413 414 r = amdgpu_bo_kmap(bo, &cpu_ptr_tmp); 415 if (r) { 416 dev_err(adev->dev, 417 "(%d) failed to map bo to kernel for amdkfd\n", r); 418 goto allocate_mem_kmap_bo_failed; 419 } 420 421 *mem_obj = bo; 422 *gpu_addr = amdgpu_bo_gpu_offset(bo); 423 *cpu_ptr = cpu_ptr_tmp; 424 425 amdgpu_bo_unreserve(bo); 426 427 return 0; 428 429 allocate_mem_kmap_bo_failed: 430 amdgpu_bo_unpin(bo); 431 allocate_mem_pin_bo_failed: 432 amdgpu_bo_unreserve(bo); 433 allocate_mem_reserve_bo_failed: 434 amdgpu_bo_unref(&bo); 435 436 return r; 437 } 438 439 void amdgpu_amdkfd_free_kernel_mem(struct amdgpu_device *adev, void **mem_obj) 440 { 441 struct amdgpu_bo **bo = (struct amdgpu_bo **) mem_obj; 442 443 if (!bo || !*bo) 444 return; 445 446 (void)amdgpu_bo_reserve(*bo, true); 447 amdgpu_bo_kunmap(*bo); 448 amdgpu_bo_unpin(*bo); 449 amdgpu_bo_unreserve(*bo); 450 amdgpu_bo_unref(bo); 451 } 452 453 int amdgpu_amdkfd_alloc_gws(struct amdgpu_device *adev, size_t size, 454 void **mem_obj) 455 { 456 struct amdgpu_bo *bo = NULL; 457 struct amdgpu_bo_user *ubo; 458 struct amdgpu_bo_param bp; 459 int r; 460 461 memset(&bp, 0, sizeof(bp)); 462 bp.size = size; 463 bp.byte_align = 1; 464 bp.domain = AMDGPU_GEM_DOMAIN_GWS; 465 bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS; 466 bp.type = ttm_bo_type_device; 467 bp.resv = NULL; 468 bp.bo_ptr_size = sizeof(struct amdgpu_bo); 469 470 r = amdgpu_bo_create_user(adev, &bp, &ubo); 471 if (r) { 472 dev_err(adev->dev, 473 "failed to allocate gws BO for amdkfd (%d)\n", r); 474 return r; 475 } 476 477 bo = &ubo->bo; 478 *mem_obj = bo; 479 return 0; 480 } 481 482 void amdgpu_amdkfd_free_gws(struct amdgpu_device *adev, void *mem_obj) 483 { 484 struct amdgpu_bo *bo = (struct amdgpu_bo *)mem_obj; 485 486 amdgpu_bo_unref(&bo); 487 } 488 489 uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev, 490 enum kgd_engine_type type) 491 { 492 switch (type) { 493 case KGD_ENGINE_PFP: 494 return adev->gfx.pfp_fw_version; 495 496 case KGD_ENGINE_ME: 497 return adev->gfx.me_fw_version; 498 499 case KGD_ENGINE_CE: 500 return adev->gfx.ce_fw_version; 501 502 case KGD_ENGINE_MEC1: 503 return adev->gfx.mec_fw_version; 504 505 case KGD_ENGINE_MEC2: 506 return adev->gfx.mec2_fw_version; 507 508 case KGD_ENGINE_RLC: 509 return adev->gfx.rlc_fw_version; 510 511 case KGD_ENGINE_SDMA1: 512 return adev->sdma.instance[0].fw_version; 513 514 case KGD_ENGINE_SDMA2: 515 return adev->sdma.instance[1].fw_version; 516 517 default: 518 return 0; 519 } 520 521 return 0; 522 } 523 524 void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev, 525 struct kfd_local_mem_info *mem_info, 526 struct amdgpu_xcp *xcp) 527 { 528 memset(mem_info, 0, sizeof(*mem_info)); 529 530 if (xcp) { 531 if (adev->gmc.real_vram_size == adev->gmc.visible_vram_size) 532 mem_info->local_mem_size_public = 533 KFD_XCP_MEMORY_SIZE(adev, xcp->id); 534 else 535 mem_info->local_mem_size_private = 536 KFD_XCP_MEMORY_SIZE(adev, xcp->id); 537 } else if (adev->apu_prefer_gtt) { 538 mem_info->local_mem_size_public = (ttm_tt_pages_limit() << PAGE_SHIFT); 539 mem_info->local_mem_size_private = 0; 540 } else { 541 mem_info->local_mem_size_public = adev->gmc.visible_vram_size; 542 mem_info->local_mem_size_private = adev->gmc.real_vram_size - 543 adev->gmc.visible_vram_size; 544 } 545 mem_info->vram_width = adev->gmc.vram_width; 546 547 pr_debug("Address base: %pap public 0x%llx private 0x%llx\n", 548 &adev->gmc.aper_base, 549 mem_info->local_mem_size_public, 550 mem_info->local_mem_size_private); 551 552 if (adev->pm.dpm_enabled) { 553 if (amdgpu_emu_mode == 1) 554 mem_info->mem_clk_max = 0; 555 else 556 mem_info->mem_clk_max = amdgpu_dpm_get_mclk(adev, false) / 100; 557 } else 558 mem_info->mem_clk_max = 100; 559 } 560 561 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev) 562 { 563 if (adev->gfx.funcs->get_gpu_clock_counter) 564 return adev->gfx.funcs->get_gpu_clock_counter(adev); 565 return 0; 566 } 567 568 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev) 569 { 570 /* the sclk is in quantas of 10kHz */ 571 if (adev->pm.dpm_enabled) 572 return amdgpu_dpm_get_sclk(adev, false) / 100; 573 else 574 return 100; 575 } 576 577 int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd, 578 struct amdgpu_device **dmabuf_adev, 579 uint64_t *bo_size, void **metadata_buffer, 580 size_t buffer_size, uint32_t *metadata_size, 581 uint32_t *flags, int8_t *xcp_id) 582 { 583 struct dma_buf *dma_buf; 584 struct drm_gem_object *obj; 585 struct amdgpu_bo *bo; 586 uint64_t metadata_flags; 587 int r = -EINVAL; 588 589 dma_buf = dma_buf_get(dma_buf_fd); 590 if (IS_ERR(dma_buf)) 591 return PTR_ERR(dma_buf); 592 593 if (dma_buf->ops != &amdgpu_dmabuf_ops) 594 /* Can't handle non-graphics buffers */ 595 goto out_put; 596 597 obj = dma_buf->priv; 598 if (obj->dev->driver != adev_to_drm(adev)->driver) 599 /* Can't handle buffers from different drivers */ 600 goto out_put; 601 602 adev = drm_to_adev(obj->dev); 603 bo = gem_to_amdgpu_bo(obj); 604 if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM | 605 AMDGPU_GEM_DOMAIN_GTT))) 606 /* Only VRAM and GTT BOs are supported */ 607 goto out_put; 608 609 r = 0; 610 if (dmabuf_adev) 611 *dmabuf_adev = adev; 612 if (bo_size) 613 *bo_size = amdgpu_bo_size(bo); 614 if (metadata_buffer) { 615 /* first get metadata_size by buffer = NULL */ 616 r = amdgpu_bo_get_metadata(bo, NULL, 0, 617 metadata_size, NULL); 618 if (r) 619 goto out_put; 620 621 /* user buf_size is bigger than bo metadata_size 622 * allocate a buf at kernel space and copy */ 623 if (*metadata_size <= buffer_size) { 624 *metadata_buffer = kzalloc(*metadata_size, GFP_KERNEL); 625 626 if (!*metadata_buffer) { 627 r = -ENOMEM; 628 goto out_put; 629 } 630 631 r = amdgpu_bo_get_metadata(bo, *metadata_buffer, *metadata_size, 632 NULL, &metadata_flags); 633 } else 634 r = -EINVAL; 635 } 636 if (flags) { 637 *flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ? 638 KFD_IOC_ALLOC_MEM_FLAGS_VRAM 639 : KFD_IOC_ALLOC_MEM_FLAGS_GTT; 640 641 if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) 642 *flags |= KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC; 643 } 644 if (xcp_id) 645 *xcp_id = bo->xcp_id; 646 647 out_put: 648 dma_buf_put(dma_buf); 649 return r; 650 } 651 652 int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool is_min) 653 { 654 int num_lanes_shift = (is_min ? ffs(adev->pm.pcie_mlw_mask) : 655 fls(adev->pm.pcie_mlw_mask)) - 1; 656 int gen_speed_shift = (is_min ? ffs(adev->pm.pcie_gen_mask & 657 CAIL_PCIE_LINK_SPEED_SUPPORT_MASK) : 658 fls(adev->pm.pcie_gen_mask & 659 CAIL_PCIE_LINK_SPEED_SUPPORT_MASK)) - 1; 660 uint32_t num_lanes_mask = 1 << num_lanes_shift; 661 uint32_t gen_speed_mask = 1 << gen_speed_shift; 662 int num_lanes_factor = 0, gen_speed_mbits_factor = 0; 663 664 switch (num_lanes_mask) { 665 case CAIL_PCIE_LINK_WIDTH_SUPPORT_X1: 666 num_lanes_factor = 1; 667 break; 668 case CAIL_PCIE_LINK_WIDTH_SUPPORT_X2: 669 num_lanes_factor = 2; 670 break; 671 case CAIL_PCIE_LINK_WIDTH_SUPPORT_X4: 672 num_lanes_factor = 4; 673 break; 674 case CAIL_PCIE_LINK_WIDTH_SUPPORT_X8: 675 num_lanes_factor = 8; 676 break; 677 case CAIL_PCIE_LINK_WIDTH_SUPPORT_X12: 678 num_lanes_factor = 12; 679 break; 680 case CAIL_PCIE_LINK_WIDTH_SUPPORT_X16: 681 num_lanes_factor = 16; 682 break; 683 case CAIL_PCIE_LINK_WIDTH_SUPPORT_X32: 684 num_lanes_factor = 32; 685 break; 686 } 687 688 switch (gen_speed_mask) { 689 case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1: 690 gen_speed_mbits_factor = 2500; 691 break; 692 case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2: 693 gen_speed_mbits_factor = 5000; 694 break; 695 case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3: 696 gen_speed_mbits_factor = 8000; 697 break; 698 case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4: 699 gen_speed_mbits_factor = 16000; 700 break; 701 case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5: 702 gen_speed_mbits_factor = 32000; 703 break; 704 } 705 706 return (num_lanes_factor * gen_speed_mbits_factor)/BITS_PER_BYTE; 707 } 708 709 int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev, 710 enum kgd_engine_type engine, 711 uint32_t vmid, uint64_t gpu_addr, 712 uint32_t *ib_cmd, uint32_t ib_len) 713 { 714 struct amdgpu_job *job; 715 struct amdgpu_ib *ib; 716 struct amdgpu_ring *ring; 717 struct dma_fence *f = NULL; 718 int ret; 719 720 switch (engine) { 721 case KGD_ENGINE_MEC1: 722 ring = &adev->gfx.compute_ring[0]; 723 break; 724 case KGD_ENGINE_SDMA1: 725 ring = &adev->sdma.instance[0].ring; 726 break; 727 case KGD_ENGINE_SDMA2: 728 ring = &adev->sdma.instance[1].ring; 729 break; 730 default: 731 pr_err("Invalid engine in IB submission: %d\n", engine); 732 ret = -EINVAL; 733 goto err; 734 } 735 736 ret = amdgpu_job_alloc(adev, NULL, NULL, NULL, 1, 0, GFP_KERNEL, 737 &job); 738 if (ret) 739 goto err; 740 741 ib = &job->ibs[0]; 742 memset(ib, 0, sizeof(struct amdgpu_ib)); 743 744 ib->gpu_addr = gpu_addr; 745 ib->ptr = ib_cmd; 746 ib->length_dw = ib_len; 747 /* This works for NO_HWS. TODO: need to handle without knowing VMID */ 748 job->vmid = vmid; 749 job->num_ibs = 1; 750 751 ret = amdgpu_ib_schedule(ring, job, &f); 752 753 if (ret) { 754 drm_err(adev_to_drm(adev), "failed to schedule IB.\n"); 755 goto err_ib_sched; 756 } 757 758 ret = dma_fence_wait(f, false); 759 /* Drop the returned fence reference after the wait completes */ 760 dma_fence_put(f); 761 762 err_ib_sched: 763 amdgpu_job_free(job); 764 err: 765 return ret; 766 } 767 768 void amdgpu_amdkfd_set_compute_idle(struct amdgpu_device *adev, bool idle) 769 { 770 enum amd_powergating_state state = idle ? AMD_PG_STATE_GATE : AMD_PG_STATE_UNGATE; 771 if ((IP_VERSION_MAJ(amdgpu_ip_version(adev, GC_HWIP, 0)) == 11 && 772 ((adev->mes.kiq_version & AMDGPU_MES_VERSION_MASK) <= 64)) || 773 (IP_VERSION_MAJ(amdgpu_ip_version(adev, GC_HWIP, 0)) == 12)) { 774 pr_debug("GFXOFF is %s\n", idle ? "enabled" : "disabled"); 775 amdgpu_gfx_off_ctrl(adev, idle); 776 } else if ((IP_VERSION_MAJ(amdgpu_ip_version(adev, GC_HWIP, 0)) == 9) && 777 (adev->flags & AMD_IS_APU)) { 778 /* Disable GFXOFF and PG. Temporary workaround 779 * to fix some compute applications issue on GFX9. 780 */ 781 struct amdgpu_ip_block *gfx_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX); 782 if (gfx_block != NULL) 783 gfx_block->version->funcs->set_powergating_state((void *)gfx_block, state); 784 } 785 (void)amdgpu_dpm_switch_power_profile(adev, PP_SMC_POWER_PROFILE_COMPUTE, !idle); 786 787 } 788 789 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid) 790 { 791 if (adev->kfd.dev) 792 return vmid >= adev->vm_manager.first_kfd_vmid; 793 794 return false; 795 } 796 797 bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev) 798 { 799 return adev->have_atomics_support; 800 } 801 802 void amdgpu_amdkfd_debug_mem_fence(struct amdgpu_device *adev) 803 { 804 amdgpu_device_flush_hdp(adev, NULL); 805 } 806 807 bool amdgpu_amdkfd_is_fed(struct amdgpu_device *adev) 808 { 809 return amdgpu_ras_get_fed_status(adev); 810 } 811 812 void amdgpu_amdkfd_ras_pasid_poison_consumption_handler(struct amdgpu_device *adev, 813 enum amdgpu_ras_block block, uint16_t pasid, 814 pasid_notify pasid_fn, void *data, uint32_t reset) 815 { 816 amdgpu_umc_pasid_poison_handler(adev, block, pasid, pasid_fn, data, reset); 817 } 818 819 void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, 820 enum amdgpu_ras_block block, uint32_t reset) 821 { 822 amdgpu_umc_pasid_poison_handler(adev, block, 0, NULL, NULL, reset); 823 } 824 825 int amdgpu_amdkfd_send_close_event_drain_irq(struct amdgpu_device *adev, 826 uint32_t *payload) 827 { 828 int ret; 829 830 /* Device or IH ring is not ready so bail. */ 831 ret = amdgpu_ih_wait_on_checkpoint_process_ts(adev, &adev->irq.ih); 832 if (ret) 833 return ret; 834 835 /* Send payload to fence KFD interrupts */ 836 amdgpu_amdkfd_interrupt(adev, payload); 837 838 return 0; 839 } 840 841 842 u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id) 843 { 844 s8 mem_id = KFD_XCP_MEM_ID(adev, xcp_id); 845 u64 tmp; 846 847 if (adev->gmc.num_mem_partitions && xcp_id >= 0 && mem_id >= 0) { 848 if (adev->gmc.is_app_apu && adev->gmc.num_mem_partitions == 1) { 849 /* In NPS1 mode, we should restrict the vram reporting 850 * tied to the ttm_pages_limit which is 1/2 of the system 851 * memory. For other partition modes, the HBM is uniformly 852 * divided already per numa node reported. If user wants to 853 * go beyond the default ttm limit and maximize the ROCm 854 * allocations, they can go up to max ttm and sysmem limits. 855 */ 856 857 tmp = (ttm_tt_pages_limit() << PAGE_SHIFT) / num_online_nodes(); 858 } else { 859 tmp = adev->gmc.mem_partitions[mem_id].size; 860 } 861 862 if (adev->xcp_mgr->mem_alloc_mode == AMDGPU_PARTITION_MEM_CAPPING_EVEN) 863 do_div(tmp, adev->xcp_mgr->num_xcp_per_mem_partition); 864 865 return ALIGN_DOWN(tmp, PAGE_SIZE); 866 } else if (adev->apu_prefer_gtt) { 867 return (ttm_tt_pages_limit() << PAGE_SHIFT); 868 } else { 869 return adev->gmc.real_vram_size; 870 } 871 } 872 873 int amdgpu_amdkfd_unmap_hiq(struct amdgpu_device *adev, u32 doorbell_off, 874 u32 inst) 875 { 876 struct amdgpu_kiq *kiq = &adev->gfx.kiq[inst]; 877 struct amdgpu_ring *kiq_ring = &kiq->ring; 878 struct amdgpu_ring_funcs *ring_funcs; 879 struct amdgpu_ring *ring; 880 int r = 0; 881 882 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues) 883 return -EINVAL; 884 885 if (!kiq_ring->sched.ready || amdgpu_in_reset(adev)) 886 return 0; 887 888 ring_funcs = kzalloc_obj(*ring_funcs); 889 if (!ring_funcs) 890 return -ENOMEM; 891 892 ring = kzalloc_obj(*ring); 893 if (!ring) { 894 r = -ENOMEM; 895 goto free_ring_funcs; 896 } 897 898 ring_funcs->type = AMDGPU_RING_TYPE_COMPUTE; 899 ring->doorbell_index = doorbell_off; 900 ring->funcs = ring_funcs; 901 902 spin_lock(&kiq->ring_lock); 903 904 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size)) { 905 spin_unlock(&kiq->ring_lock); 906 r = -ENOMEM; 907 goto free_ring; 908 } 909 910 kiq->pmf->kiq_unmap_queues(kiq_ring, ring, RESET_QUEUES, 0, 0); 911 912 /* Submit unmap queue packet */ 913 amdgpu_ring_commit(kiq_ring); 914 /* 915 * Ring test will do a basic scratch register change check. Just run 916 * this to ensure that unmap queues that is submitted before got 917 * processed successfully before returning. 918 */ 919 r = amdgpu_ring_test_helper(kiq_ring); 920 921 spin_unlock(&kiq->ring_lock); 922 923 free_ring: 924 kfree(ring); 925 926 free_ring_funcs: 927 kfree(ring_funcs); 928 929 return r; 930 } 931 932 /* Stop scheduling on KFD */ 933 int amdgpu_amdkfd_stop_sched(struct amdgpu_device *adev, uint32_t node_id) 934 { 935 if (!adev->kfd.init_complete) 936 return 0; 937 938 return kgd2kfd_stop_sched(adev->kfd.dev, node_id); 939 } 940 941 /* Start scheduling on KFD */ 942 int amdgpu_amdkfd_start_sched(struct amdgpu_device *adev, uint32_t node_id) 943 { 944 if (!adev->kfd.init_complete) 945 return 0; 946 947 return kgd2kfd_start_sched(adev->kfd.dev, node_id); 948 } 949 950 /* check if there are KFD queues active */ 951 bool amdgpu_amdkfd_compute_active(struct amdgpu_device *adev, uint32_t node_id) 952 { 953 if (!adev->kfd.init_complete) 954 return false; 955 956 return kgd2kfd_compute_active(adev->kfd.dev, node_id); 957 } 958 959 /* Config CGTT_SQ_CLK_CTRL */ 960 int amdgpu_amdkfd_config_sq_perfmon(struct amdgpu_device *adev, uint32_t xcp_id, 961 bool core_override_enable, bool reg_override_enable, bool perfmon_override_enable) 962 { 963 int r; 964 965 if (!adev->kfd.init_complete) 966 return 0; 967 968 r = psp_config_sq_perfmon(&adev->psp, xcp_id, core_override_enable, 969 reg_override_enable, perfmon_override_enable); 970 971 return r; 972 } 973 974 /* Reset an MES queue */ 975 int amdgpu_amdkfd_reset_mes_queue(struct amdgpu_device *adev, 976 uint32_t node_id, 977 int queue_type, 978 int pipe, int queue, 979 unsigned int db) 980 { 981 if (!adev->kfd.init_complete) 982 return 0; 983 984 return kgd2kfd_reset_mes_queue(adev->kfd.dev, node_id, queue_type, 985 pipe, queue, db); 986 } 987