1 /* 2 * Copyright 2015-2017 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 <linux/pci.h> 24 #include <linux/acpi.h> 25 #include "kfd_crat.h" 26 #include "kfd_priv.h" 27 #include "kfd_topology.h" 28 #include "kfd_iommu.h" 29 #include "amdgpu_amdkfd.h" 30 31 /* GPU Processor ID base for dGPUs for which VCRAT needs to be created. 32 * GPU processor ID are expressed with Bit[31]=1. 33 * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs 34 * used in the CRAT. 35 */ 36 static uint32_t gpu_processor_id_low = 0x80001000; 37 38 /* Return the next available gpu_processor_id and increment it for next GPU 39 * @total_cu_count - Total CUs present in the GPU including ones 40 * masked off 41 */ 42 static inline unsigned int get_and_inc_gpu_processor_id( 43 unsigned int total_cu_count) 44 { 45 int current_id = gpu_processor_id_low; 46 47 gpu_processor_id_low += total_cu_count; 48 return current_id; 49 } 50 51 /* Static table to describe GPU Cache information */ 52 struct kfd_gpu_cache_info { 53 uint32_t cache_size; 54 uint32_t cache_level; 55 uint32_t flags; 56 /* Indicates how many Compute Units share this cache 57 * Value = 1 indicates the cache is not shared 58 */ 59 uint32_t num_cu_shared; 60 }; 61 62 static struct kfd_gpu_cache_info kaveri_cache_info[] = { 63 { 64 /* TCP L1 Cache per CU */ 65 .cache_size = 16, 66 .cache_level = 1, 67 .flags = (CRAT_CACHE_FLAGS_ENABLED | 68 CRAT_CACHE_FLAGS_DATA_CACHE | 69 CRAT_CACHE_FLAGS_SIMD_CACHE), 70 .num_cu_shared = 1, 71 72 }, 73 { 74 /* Scalar L1 Instruction Cache (in SQC module) per bank */ 75 .cache_size = 16, 76 .cache_level = 1, 77 .flags = (CRAT_CACHE_FLAGS_ENABLED | 78 CRAT_CACHE_FLAGS_INST_CACHE | 79 CRAT_CACHE_FLAGS_SIMD_CACHE), 80 .num_cu_shared = 2, 81 }, 82 { 83 /* Scalar L1 Data Cache (in SQC module) per bank */ 84 .cache_size = 8, 85 .cache_level = 1, 86 .flags = (CRAT_CACHE_FLAGS_ENABLED | 87 CRAT_CACHE_FLAGS_DATA_CACHE | 88 CRAT_CACHE_FLAGS_SIMD_CACHE), 89 .num_cu_shared = 2, 90 }, 91 92 /* TODO: Add L2 Cache information */ 93 }; 94 95 96 static struct kfd_gpu_cache_info carrizo_cache_info[] = { 97 { 98 /* TCP L1 Cache per CU */ 99 .cache_size = 16, 100 .cache_level = 1, 101 .flags = (CRAT_CACHE_FLAGS_ENABLED | 102 CRAT_CACHE_FLAGS_DATA_CACHE | 103 CRAT_CACHE_FLAGS_SIMD_CACHE), 104 .num_cu_shared = 1, 105 }, 106 { 107 /* Scalar L1 Instruction Cache (in SQC module) per bank */ 108 .cache_size = 8, 109 .cache_level = 1, 110 .flags = (CRAT_CACHE_FLAGS_ENABLED | 111 CRAT_CACHE_FLAGS_INST_CACHE | 112 CRAT_CACHE_FLAGS_SIMD_CACHE), 113 .num_cu_shared = 4, 114 }, 115 { 116 /* Scalar L1 Data Cache (in SQC module) per bank. */ 117 .cache_size = 4, 118 .cache_level = 1, 119 .flags = (CRAT_CACHE_FLAGS_ENABLED | 120 CRAT_CACHE_FLAGS_DATA_CACHE | 121 CRAT_CACHE_FLAGS_SIMD_CACHE), 122 .num_cu_shared = 4, 123 }, 124 125 /* TODO: Add L2 Cache information */ 126 }; 127 128 /* NOTE: In future if more information is added to struct kfd_gpu_cache_info 129 * the following ASICs may need a separate table. 130 */ 131 #define hawaii_cache_info kaveri_cache_info 132 #define tonga_cache_info carrizo_cache_info 133 #define fiji_cache_info carrizo_cache_info 134 #define polaris10_cache_info carrizo_cache_info 135 #define polaris11_cache_info carrizo_cache_info 136 #define polaris12_cache_info carrizo_cache_info 137 #define vegam_cache_info carrizo_cache_info 138 /* TODO - check & update Vega10 cache details */ 139 #define vega10_cache_info carrizo_cache_info 140 #define raven_cache_info carrizo_cache_info 141 /* TODO - check & update Navi10 cache details */ 142 #define navi10_cache_info carrizo_cache_info 143 144 static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev, 145 struct crat_subtype_computeunit *cu) 146 { 147 dev->node_props.cpu_cores_count = cu->num_cpu_cores; 148 dev->node_props.cpu_core_id_base = cu->processor_id_low; 149 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT) 150 dev->node_props.capability |= HSA_CAP_ATS_PRESENT; 151 152 pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores, 153 cu->processor_id_low); 154 } 155 156 static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev, 157 struct crat_subtype_computeunit *cu) 158 { 159 dev->node_props.simd_id_base = cu->processor_id_low; 160 dev->node_props.simd_count = cu->num_simd_cores; 161 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb; 162 dev->node_props.max_waves_per_simd = cu->max_waves_simd; 163 dev->node_props.wave_front_size = cu->wave_front_size; 164 dev->node_props.array_count = cu->array_count; 165 dev->node_props.cu_per_simd_array = cu->num_cu_per_array; 166 dev->node_props.simd_per_cu = cu->num_simd_per_cu; 167 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu; 168 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE) 169 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE; 170 pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low); 171 } 172 173 /* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct 174 * topology device present in the device_list 175 */ 176 static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu, 177 struct list_head *device_list) 178 { 179 struct kfd_topology_device *dev; 180 181 pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n", 182 cu->proximity_domain, cu->hsa_capability); 183 list_for_each_entry(dev, device_list, list) { 184 if (cu->proximity_domain == dev->proximity_domain) { 185 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT) 186 kfd_populated_cu_info_cpu(dev, cu); 187 188 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT) 189 kfd_populated_cu_info_gpu(dev, cu); 190 break; 191 } 192 } 193 194 return 0; 195 } 196 197 static struct kfd_mem_properties * 198 find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width, 199 struct kfd_topology_device *dev) 200 { 201 struct kfd_mem_properties *props; 202 203 list_for_each_entry(props, &dev->mem_props, list) { 204 if (props->heap_type == heap_type 205 && props->flags == flags 206 && props->width == width) 207 return props; 208 } 209 210 return NULL; 211 } 212 /* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct 213 * topology device present in the device_list 214 */ 215 static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem, 216 struct list_head *device_list) 217 { 218 struct kfd_mem_properties *props; 219 struct kfd_topology_device *dev; 220 uint32_t heap_type; 221 uint64_t size_in_bytes; 222 uint32_t flags = 0; 223 uint32_t width; 224 225 pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n", 226 mem->proximity_domain); 227 list_for_each_entry(dev, device_list, list) { 228 if (mem->proximity_domain == dev->proximity_domain) { 229 /* We're on GPU node */ 230 if (dev->node_props.cpu_cores_count == 0) { 231 /* APU */ 232 if (mem->visibility_type == 0) 233 heap_type = 234 HSA_MEM_HEAP_TYPE_FB_PRIVATE; 235 /* dGPU */ 236 else 237 heap_type = mem->visibility_type; 238 } else 239 heap_type = HSA_MEM_HEAP_TYPE_SYSTEM; 240 241 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE) 242 flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE; 243 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE) 244 flags |= HSA_MEM_FLAGS_NON_VOLATILE; 245 246 size_in_bytes = 247 ((uint64_t)mem->length_high << 32) + 248 mem->length_low; 249 width = mem->width; 250 251 /* Multiple banks of the same type are aggregated into 252 * one. User mode doesn't care about multiple physical 253 * memory segments. It's managed as a single virtual 254 * heap for user mode. 255 */ 256 props = find_subtype_mem(heap_type, flags, width, dev); 257 if (props) { 258 props->size_in_bytes += size_in_bytes; 259 break; 260 } 261 262 props = kfd_alloc_struct(props); 263 if (!props) 264 return -ENOMEM; 265 266 props->heap_type = heap_type; 267 props->flags = flags; 268 props->size_in_bytes = size_in_bytes; 269 props->width = width; 270 271 dev->node_props.mem_banks_count++; 272 list_add_tail(&props->list, &dev->mem_props); 273 274 break; 275 } 276 } 277 278 return 0; 279 } 280 281 /* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct 282 * topology device present in the device_list 283 */ 284 static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache, 285 struct list_head *device_list) 286 { 287 struct kfd_cache_properties *props; 288 struct kfd_topology_device *dev; 289 uint32_t id; 290 uint32_t total_num_of_cu; 291 292 id = cache->processor_id_low; 293 294 pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id); 295 list_for_each_entry(dev, device_list, list) { 296 total_num_of_cu = (dev->node_props.array_count * 297 dev->node_props.cu_per_simd_array); 298 299 /* Cache infomration in CRAT doesn't have proximity_domain 300 * information as it is associated with a CPU core or GPU 301 * Compute Unit. So map the cache using CPU core Id or SIMD 302 * (GPU) ID. 303 * TODO: This works because currently we can safely assume that 304 * Compute Units are parsed before caches are parsed. In 305 * future, remove this dependency 306 */ 307 if ((id >= dev->node_props.cpu_core_id_base && 308 id <= dev->node_props.cpu_core_id_base + 309 dev->node_props.cpu_cores_count) || 310 (id >= dev->node_props.simd_id_base && 311 id < dev->node_props.simd_id_base + 312 total_num_of_cu)) { 313 props = kfd_alloc_struct(props); 314 if (!props) 315 return -ENOMEM; 316 317 props->processor_id_low = id; 318 props->cache_level = cache->cache_level; 319 props->cache_size = cache->cache_size; 320 props->cacheline_size = cache->cache_line_size; 321 props->cachelines_per_tag = cache->lines_per_tag; 322 props->cache_assoc = cache->associativity; 323 props->cache_latency = cache->cache_latency; 324 memcpy(props->sibling_map, cache->sibling_map, 325 sizeof(props->sibling_map)); 326 327 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE) 328 props->cache_type |= HSA_CACHE_TYPE_DATA; 329 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE) 330 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION; 331 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE) 332 props->cache_type |= HSA_CACHE_TYPE_CPU; 333 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE) 334 props->cache_type |= HSA_CACHE_TYPE_HSACU; 335 336 dev->cache_count++; 337 dev->node_props.caches_count++; 338 list_add_tail(&props->list, &dev->cache_props); 339 340 break; 341 } 342 } 343 344 return 0; 345 } 346 347 /* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct 348 * topology device present in the device_list 349 */ 350 static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink, 351 struct list_head *device_list) 352 { 353 struct kfd_iolink_properties *props = NULL, *props2; 354 struct kfd_topology_device *dev, *to_dev; 355 uint32_t id_from; 356 uint32_t id_to; 357 358 id_from = iolink->proximity_domain_from; 359 id_to = iolink->proximity_domain_to; 360 361 pr_debug("Found IO link entry in CRAT table with id_from=%d, id_to %d\n", 362 id_from, id_to); 363 list_for_each_entry(dev, device_list, list) { 364 if (id_from == dev->proximity_domain) { 365 props = kfd_alloc_struct(props); 366 if (!props) 367 return -ENOMEM; 368 369 props->node_from = id_from; 370 props->node_to = id_to; 371 props->ver_maj = iolink->version_major; 372 props->ver_min = iolink->version_minor; 373 props->iolink_type = iolink->io_interface_type; 374 375 if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) 376 props->weight = 20; 377 else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI) 378 props->weight = 15 * iolink->num_hops_xgmi; 379 else 380 props->weight = node_distance(id_from, id_to); 381 382 props->min_latency = iolink->minimum_latency; 383 props->max_latency = iolink->maximum_latency; 384 props->min_bandwidth = iolink->minimum_bandwidth_mbs; 385 props->max_bandwidth = iolink->maximum_bandwidth_mbs; 386 props->rec_transfer_size = 387 iolink->recommended_transfer_size; 388 389 dev->io_link_count++; 390 dev->node_props.io_links_count++; 391 list_add_tail(&props->list, &dev->io_link_props); 392 break; 393 } 394 } 395 396 /* CPU topology is created before GPUs are detected, so CPU->GPU 397 * links are not built at that time. If a PCIe type is discovered, it 398 * means a GPU is detected and we are adding GPU->CPU to the topology. 399 * At this time, also add the corresponded CPU->GPU link if GPU 400 * is large bar. 401 * For xGMI, we only added the link with one direction in the crat 402 * table, add corresponded reversed direction link now. 403 */ 404 if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) { 405 to_dev = kfd_topology_device_by_proximity_domain(id_to); 406 if (!to_dev) 407 return -ENODEV; 408 /* same everything but the other direction */ 409 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL); 410 props2->node_from = id_to; 411 props2->node_to = id_from; 412 props2->kobj = NULL; 413 to_dev->io_link_count++; 414 to_dev->node_props.io_links_count++; 415 list_add_tail(&props2->list, &to_dev->io_link_props); 416 } 417 418 return 0; 419 } 420 421 /* kfd_parse_subtype - parse subtypes and attach it to correct topology device 422 * present in the device_list 423 * @sub_type_hdr - subtype section of crat_image 424 * @device_list - list of topology devices present in this crat_image 425 */ 426 static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr, 427 struct list_head *device_list) 428 { 429 struct crat_subtype_computeunit *cu; 430 struct crat_subtype_memory *mem; 431 struct crat_subtype_cache *cache; 432 struct crat_subtype_iolink *iolink; 433 int ret = 0; 434 435 switch (sub_type_hdr->type) { 436 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY: 437 cu = (struct crat_subtype_computeunit *)sub_type_hdr; 438 ret = kfd_parse_subtype_cu(cu, device_list); 439 break; 440 case CRAT_SUBTYPE_MEMORY_AFFINITY: 441 mem = (struct crat_subtype_memory *)sub_type_hdr; 442 ret = kfd_parse_subtype_mem(mem, device_list); 443 break; 444 case CRAT_SUBTYPE_CACHE_AFFINITY: 445 cache = (struct crat_subtype_cache *)sub_type_hdr; 446 ret = kfd_parse_subtype_cache(cache, device_list); 447 break; 448 case CRAT_SUBTYPE_TLB_AFFINITY: 449 /* 450 * For now, nothing to do here 451 */ 452 pr_debug("Found TLB entry in CRAT table (not processing)\n"); 453 break; 454 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY: 455 /* 456 * For now, nothing to do here 457 */ 458 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n"); 459 break; 460 case CRAT_SUBTYPE_IOLINK_AFFINITY: 461 iolink = (struct crat_subtype_iolink *)sub_type_hdr; 462 ret = kfd_parse_subtype_iolink(iolink, device_list); 463 break; 464 default: 465 pr_warn("Unknown subtype %d in CRAT\n", 466 sub_type_hdr->type); 467 } 468 469 return ret; 470 } 471 472 /* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT 473 * create a kfd_topology_device and add in to device_list. Also parse 474 * CRAT subtypes and attach it to appropriate kfd_topology_device 475 * @crat_image - input image containing CRAT 476 * @device_list - [OUT] list of kfd_topology_device generated after 477 * parsing crat_image 478 * @proximity_domain - Proximity domain of the first device in the table 479 * 480 * Return - 0 if successful else -ve value 481 */ 482 int kfd_parse_crat_table(void *crat_image, struct list_head *device_list, 483 uint32_t proximity_domain) 484 { 485 struct kfd_topology_device *top_dev = NULL; 486 struct crat_subtype_generic *sub_type_hdr; 487 uint16_t node_id; 488 int ret = 0; 489 struct crat_header *crat_table = (struct crat_header *)crat_image; 490 uint16_t num_nodes; 491 uint32_t image_len; 492 493 if (!crat_image) 494 return -EINVAL; 495 496 if (!list_empty(device_list)) { 497 pr_warn("Error device list should be empty\n"); 498 return -EINVAL; 499 } 500 501 num_nodes = crat_table->num_domains; 502 image_len = crat_table->length; 503 504 pr_info("Parsing CRAT table with %d nodes\n", num_nodes); 505 506 for (node_id = 0; node_id < num_nodes; node_id++) { 507 top_dev = kfd_create_topology_device(device_list); 508 if (!top_dev) 509 break; 510 top_dev->proximity_domain = proximity_domain++; 511 } 512 513 if (!top_dev) { 514 ret = -ENOMEM; 515 goto err; 516 } 517 518 memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH); 519 memcpy(top_dev->oem_table_id, crat_table->oem_table_id, 520 CRAT_OEMTABLEID_LENGTH); 521 top_dev->oem_revision = crat_table->oem_revision; 522 523 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1); 524 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) < 525 ((char *)crat_image) + image_len) { 526 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) { 527 ret = kfd_parse_subtype(sub_type_hdr, device_list); 528 if (ret) 529 break; 530 } 531 532 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 533 sub_type_hdr->length); 534 } 535 536 err: 537 if (ret) 538 kfd_release_topology_device_list(device_list); 539 540 return ret; 541 } 542 543 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */ 544 static int fill_in_pcache(struct crat_subtype_cache *pcache, 545 struct kfd_gpu_cache_info *pcache_info, 546 struct kfd_cu_info *cu_info, 547 int mem_available, 548 int cu_bitmask, 549 int cache_type, unsigned int cu_processor_id, 550 int cu_block) 551 { 552 unsigned int cu_sibling_map_mask; 553 int first_active_cu; 554 555 /* First check if enough memory is available */ 556 if (sizeof(struct crat_subtype_cache) > mem_available) 557 return -ENOMEM; 558 559 cu_sibling_map_mask = cu_bitmask; 560 cu_sibling_map_mask >>= cu_block; 561 cu_sibling_map_mask &= 562 ((1 << pcache_info[cache_type].num_cu_shared) - 1); 563 first_active_cu = ffs(cu_sibling_map_mask); 564 565 /* CU could be inactive. In case of shared cache find the first active 566 * CU. and incase of non-shared cache check if the CU is inactive. If 567 * inactive active skip it 568 */ 569 if (first_active_cu) { 570 memset(pcache, 0, sizeof(struct crat_subtype_cache)); 571 pcache->type = CRAT_SUBTYPE_CACHE_AFFINITY; 572 pcache->length = sizeof(struct crat_subtype_cache); 573 pcache->flags = pcache_info[cache_type].flags; 574 pcache->processor_id_low = cu_processor_id 575 + (first_active_cu - 1); 576 pcache->cache_level = pcache_info[cache_type].cache_level; 577 pcache->cache_size = pcache_info[cache_type].cache_size; 578 579 /* Sibling map is w.r.t processor_id_low, so shift out 580 * inactive CU 581 */ 582 cu_sibling_map_mask = 583 cu_sibling_map_mask >> (first_active_cu - 1); 584 585 pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF); 586 pcache->sibling_map[1] = 587 (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF); 588 pcache->sibling_map[2] = 589 (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF); 590 pcache->sibling_map[3] = 591 (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF); 592 return 0; 593 } 594 return 1; 595 } 596 597 /* kfd_fill_gpu_cache_info - Fill GPU cache info using kfd_gpu_cache_info 598 * tables 599 * 600 * @kdev - [IN] GPU device 601 * @gpu_processor_id - [IN] GPU processor ID to which these caches 602 * associate 603 * @available_size - [IN] Amount of memory available in pcache 604 * @cu_info - [IN] Compute Unit info obtained from KGD 605 * @pcache - [OUT] memory into which cache data is to be filled in. 606 * @size_filled - [OUT] amount of data used up in pcache. 607 * @num_of_entries - [OUT] number of caches added 608 */ 609 static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev, 610 int gpu_processor_id, 611 int available_size, 612 struct kfd_cu_info *cu_info, 613 struct crat_subtype_cache *pcache, 614 int *size_filled, 615 int *num_of_entries) 616 { 617 struct kfd_gpu_cache_info *pcache_info; 618 int num_of_cache_types = 0; 619 int i, j, k; 620 int ct = 0; 621 int mem_available = available_size; 622 unsigned int cu_processor_id; 623 int ret; 624 625 switch (kdev->device_info->asic_family) { 626 case CHIP_KAVERI: 627 pcache_info = kaveri_cache_info; 628 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info); 629 break; 630 case CHIP_HAWAII: 631 pcache_info = hawaii_cache_info; 632 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info); 633 break; 634 case CHIP_CARRIZO: 635 pcache_info = carrizo_cache_info; 636 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info); 637 break; 638 case CHIP_TONGA: 639 pcache_info = tonga_cache_info; 640 num_of_cache_types = ARRAY_SIZE(tonga_cache_info); 641 break; 642 case CHIP_FIJI: 643 pcache_info = fiji_cache_info; 644 num_of_cache_types = ARRAY_SIZE(fiji_cache_info); 645 break; 646 case CHIP_POLARIS10: 647 pcache_info = polaris10_cache_info; 648 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info); 649 break; 650 case CHIP_POLARIS11: 651 pcache_info = polaris11_cache_info; 652 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info); 653 break; 654 case CHIP_POLARIS12: 655 pcache_info = polaris12_cache_info; 656 num_of_cache_types = ARRAY_SIZE(polaris12_cache_info); 657 break; 658 case CHIP_VEGAM: 659 pcache_info = vegam_cache_info; 660 num_of_cache_types = ARRAY_SIZE(vegam_cache_info); 661 break; 662 case CHIP_VEGA10: 663 case CHIP_VEGA12: 664 case CHIP_VEGA20: 665 case CHIP_ARCTURUS: 666 pcache_info = vega10_cache_info; 667 num_of_cache_types = ARRAY_SIZE(vega10_cache_info); 668 break; 669 case CHIP_RAVEN: 670 pcache_info = raven_cache_info; 671 num_of_cache_types = ARRAY_SIZE(raven_cache_info); 672 case CHIP_NAVI10: 673 pcache_info = navi10_cache_info; 674 num_of_cache_types = ARRAY_SIZE(navi10_cache_info); 675 break; 676 default: 677 return -EINVAL; 678 } 679 680 *size_filled = 0; 681 *num_of_entries = 0; 682 683 /* For each type of cache listed in the kfd_gpu_cache_info table, 684 * go through all available Compute Units. 685 * The [i,j,k] loop will 686 * if kfd_gpu_cache_info.num_cu_shared = 1 687 * will parse through all available CU 688 * If (kfd_gpu_cache_info.num_cu_shared != 1) 689 * then it will consider only one CU from 690 * the shared unit 691 */ 692 693 for (ct = 0; ct < num_of_cache_types; ct++) { 694 cu_processor_id = gpu_processor_id; 695 for (i = 0; i < cu_info->num_shader_engines; i++) { 696 for (j = 0; j < cu_info->num_shader_arrays_per_engine; 697 j++) { 698 for (k = 0; k < cu_info->num_cu_per_sh; 699 k += pcache_info[ct].num_cu_shared) { 700 701 ret = fill_in_pcache(pcache, 702 pcache_info, 703 cu_info, 704 mem_available, 705 cu_info->cu_bitmap[i][j], 706 ct, 707 cu_processor_id, 708 k); 709 710 if (ret < 0) 711 break; 712 713 if (!ret) { 714 pcache++; 715 (*num_of_entries)++; 716 mem_available -= 717 sizeof(*pcache); 718 (*size_filled) += 719 sizeof(*pcache); 720 } 721 722 /* Move to next CU block */ 723 cu_processor_id += 724 pcache_info[ct].num_cu_shared; 725 } 726 } 727 } 728 } 729 730 pr_debug("Added [%d] GPU cache entries\n", *num_of_entries); 731 732 return 0; 733 } 734 735 /* 736 * kfd_create_crat_image_acpi - Allocates memory for CRAT image and 737 * copies CRAT from ACPI (if available). 738 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory 739 * 740 * @crat_image: CRAT read from ACPI. If no CRAT in ACPI then 741 * crat_image will be NULL 742 * @size: [OUT] size of crat_image 743 * 744 * Return 0 if successful else return error code 745 */ 746 int kfd_create_crat_image_acpi(void **crat_image, size_t *size) 747 { 748 struct acpi_table_header *crat_table; 749 acpi_status status; 750 void *pcrat_image; 751 752 if (!crat_image) 753 return -EINVAL; 754 755 *crat_image = NULL; 756 757 /* Fetch the CRAT table from ACPI */ 758 status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table); 759 if (status == AE_NOT_FOUND) { 760 pr_warn("CRAT table not found\n"); 761 return -ENODATA; 762 } else if (ACPI_FAILURE(status)) { 763 const char *err = acpi_format_exception(status); 764 765 pr_err("CRAT table error: %s\n", err); 766 return -EINVAL; 767 } 768 769 if (ignore_crat) { 770 pr_info("CRAT table disabled by module option\n"); 771 return -ENODATA; 772 } 773 774 pcrat_image = kmemdup(crat_table, crat_table->length, GFP_KERNEL); 775 if (!pcrat_image) 776 return -ENOMEM; 777 778 *crat_image = pcrat_image; 779 *size = crat_table->length; 780 781 return 0; 782 } 783 784 /* Memory required to create Virtual CRAT. 785 * Since there is no easy way to predict the amount of memory required, the 786 * following amount are allocated for CPU and GPU Virtual CRAT. This is 787 * expected to cover all known conditions. But to be safe additional check 788 * is put in the code to ensure we don't overwrite. 789 */ 790 #define VCRAT_SIZE_FOR_CPU (2 * PAGE_SIZE) 791 #define VCRAT_SIZE_FOR_GPU (4 * PAGE_SIZE) 792 793 /* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node 794 * 795 * @numa_node_id: CPU NUMA node id 796 * @avail_size: Available size in the memory 797 * @sub_type_hdr: Memory into which compute info will be filled in 798 * 799 * Return 0 if successful else return -ve value 800 */ 801 static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size, 802 int proximity_domain, 803 struct crat_subtype_computeunit *sub_type_hdr) 804 { 805 const struct cpumask *cpumask; 806 807 *avail_size -= sizeof(struct crat_subtype_computeunit); 808 if (*avail_size < 0) 809 return -ENOMEM; 810 811 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit)); 812 813 /* Fill in subtype header data */ 814 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY; 815 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit); 816 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 817 818 cpumask = cpumask_of_node(numa_node_id); 819 820 /* Fill in CU data */ 821 sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT; 822 sub_type_hdr->proximity_domain = proximity_domain; 823 sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id); 824 if (sub_type_hdr->processor_id_low == -1) 825 return -EINVAL; 826 827 sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask); 828 829 return 0; 830 } 831 832 /* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node 833 * 834 * @numa_node_id: CPU NUMA node id 835 * @avail_size: Available size in the memory 836 * @sub_type_hdr: Memory into which compute info will be filled in 837 * 838 * Return 0 if successful else return -ve value 839 */ 840 static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size, 841 int proximity_domain, 842 struct crat_subtype_memory *sub_type_hdr) 843 { 844 uint64_t mem_in_bytes = 0; 845 pg_data_t *pgdat; 846 int zone_type; 847 848 *avail_size -= sizeof(struct crat_subtype_memory); 849 if (*avail_size < 0) 850 return -ENOMEM; 851 852 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory)); 853 854 /* Fill in subtype header data */ 855 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY; 856 sub_type_hdr->length = sizeof(struct crat_subtype_memory); 857 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 858 859 /* Fill in Memory Subunit data */ 860 861 /* Unlike si_meminfo, si_meminfo_node is not exported. So 862 * the following lines are duplicated from si_meminfo_node 863 * function 864 */ 865 pgdat = NODE_DATA(numa_node_id); 866 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) 867 mem_in_bytes += zone_managed_pages(&pgdat->node_zones[zone_type]); 868 mem_in_bytes <<= PAGE_SHIFT; 869 870 sub_type_hdr->length_low = lower_32_bits(mem_in_bytes); 871 sub_type_hdr->length_high = upper_32_bits(mem_in_bytes); 872 sub_type_hdr->proximity_domain = proximity_domain; 873 874 return 0; 875 } 876 877 #ifdef CONFIG_X86_64 878 static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size, 879 uint32_t *num_entries, 880 struct crat_subtype_iolink *sub_type_hdr) 881 { 882 int nid; 883 struct cpuinfo_x86 *c = &cpu_data(0); 884 uint8_t link_type; 885 886 if (c->x86_vendor == X86_VENDOR_AMD) 887 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT; 888 else 889 link_type = CRAT_IOLINK_TYPE_QPI_1_1; 890 891 *num_entries = 0; 892 893 /* Create IO links from this node to other CPU nodes */ 894 for_each_online_node(nid) { 895 if (nid == numa_node_id) /* node itself */ 896 continue; 897 898 *avail_size -= sizeof(struct crat_subtype_iolink); 899 if (*avail_size < 0) 900 return -ENOMEM; 901 902 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink)); 903 904 /* Fill in subtype header data */ 905 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY; 906 sub_type_hdr->length = sizeof(struct crat_subtype_iolink); 907 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 908 909 /* Fill in IO link data */ 910 sub_type_hdr->proximity_domain_from = numa_node_id; 911 sub_type_hdr->proximity_domain_to = nid; 912 sub_type_hdr->io_interface_type = link_type; 913 914 (*num_entries)++; 915 sub_type_hdr++; 916 } 917 918 return 0; 919 } 920 #endif 921 922 /* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU 923 * 924 * @pcrat_image: Fill in VCRAT for CPU 925 * @size: [IN] allocated size of crat_image. 926 * [OUT] actual size of data filled in crat_image 927 */ 928 static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size) 929 { 930 struct crat_header *crat_table = (struct crat_header *)pcrat_image; 931 struct acpi_table_header *acpi_table; 932 acpi_status status; 933 struct crat_subtype_generic *sub_type_hdr; 934 int avail_size = *size; 935 int numa_node_id; 936 #ifdef CONFIG_X86_64 937 uint32_t entries = 0; 938 #endif 939 int ret = 0; 940 941 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_CPU) 942 return -EINVAL; 943 944 /* Fill in CRAT Header. 945 * Modify length and total_entries as subunits are added. 946 */ 947 avail_size -= sizeof(struct crat_header); 948 if (avail_size < 0) 949 return -ENOMEM; 950 951 memset(crat_table, 0, sizeof(struct crat_header)); 952 memcpy(&crat_table->signature, CRAT_SIGNATURE, 953 sizeof(crat_table->signature)); 954 crat_table->length = sizeof(struct crat_header); 955 956 status = acpi_get_table("DSDT", 0, &acpi_table); 957 if (status != AE_OK) 958 pr_warn("DSDT table not found for OEM information\n"); 959 else { 960 crat_table->oem_revision = acpi_table->revision; 961 memcpy(crat_table->oem_id, acpi_table->oem_id, 962 CRAT_OEMID_LENGTH); 963 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id, 964 CRAT_OEMTABLEID_LENGTH); 965 } 966 crat_table->total_entries = 0; 967 crat_table->num_domains = 0; 968 969 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1); 970 971 for_each_online_node(numa_node_id) { 972 if (kfd_numa_node_to_apic_id(numa_node_id) == -1) 973 continue; 974 975 /* Fill in Subtype: Compute Unit */ 976 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size, 977 crat_table->num_domains, 978 (struct crat_subtype_computeunit *)sub_type_hdr); 979 if (ret < 0) 980 return ret; 981 crat_table->length += sub_type_hdr->length; 982 crat_table->total_entries++; 983 984 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 985 sub_type_hdr->length); 986 987 /* Fill in Subtype: Memory */ 988 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size, 989 crat_table->num_domains, 990 (struct crat_subtype_memory *)sub_type_hdr); 991 if (ret < 0) 992 return ret; 993 crat_table->length += sub_type_hdr->length; 994 crat_table->total_entries++; 995 996 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 997 sub_type_hdr->length); 998 999 /* Fill in Subtype: IO Link */ 1000 #ifdef CONFIG_X86_64 1001 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size, 1002 &entries, 1003 (struct crat_subtype_iolink *)sub_type_hdr); 1004 if (ret < 0) 1005 return ret; 1006 crat_table->length += (sub_type_hdr->length * entries); 1007 crat_table->total_entries += entries; 1008 1009 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1010 sub_type_hdr->length * entries); 1011 #else 1012 pr_info("IO link not available for non x86 platforms\n"); 1013 #endif 1014 1015 crat_table->num_domains++; 1016 } 1017 1018 /* TODO: Add cache Subtype for CPU. 1019 * Currently, CPU cache information is available in function 1020 * detect_cache_attributes(cpu) defined in the file 1021 * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not 1022 * exported and to get the same information the code needs to be 1023 * duplicated. 1024 */ 1025 1026 *size = crat_table->length; 1027 pr_info("Virtual CRAT table created for CPU\n"); 1028 1029 return 0; 1030 } 1031 1032 static int kfd_fill_gpu_memory_affinity(int *avail_size, 1033 struct kfd_dev *kdev, uint8_t type, uint64_t size, 1034 struct crat_subtype_memory *sub_type_hdr, 1035 uint32_t proximity_domain, 1036 const struct kfd_local_mem_info *local_mem_info) 1037 { 1038 *avail_size -= sizeof(struct crat_subtype_memory); 1039 if (*avail_size < 0) 1040 return -ENOMEM; 1041 1042 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory)); 1043 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY; 1044 sub_type_hdr->length = sizeof(struct crat_subtype_memory); 1045 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED; 1046 1047 sub_type_hdr->proximity_domain = proximity_domain; 1048 1049 pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n", 1050 type, size); 1051 1052 sub_type_hdr->length_low = lower_32_bits(size); 1053 sub_type_hdr->length_high = upper_32_bits(size); 1054 1055 sub_type_hdr->width = local_mem_info->vram_width; 1056 sub_type_hdr->visibility_type = type; 1057 1058 return 0; 1059 } 1060 1061 /* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU 1062 * to its NUMA node 1063 * @avail_size: Available size in the memory 1064 * @kdev - [IN] GPU device 1065 * @sub_type_hdr: Memory into which io link info will be filled in 1066 * @proximity_domain - proximity domain of the GPU node 1067 * 1068 * Return 0 if successful else return -ve value 1069 */ 1070 static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size, 1071 struct kfd_dev *kdev, 1072 struct crat_subtype_iolink *sub_type_hdr, 1073 uint32_t proximity_domain) 1074 { 1075 *avail_size -= sizeof(struct crat_subtype_iolink); 1076 if (*avail_size < 0) 1077 return -ENOMEM; 1078 1079 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink)); 1080 1081 /* Fill in subtype header data */ 1082 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY; 1083 sub_type_hdr->length = sizeof(struct crat_subtype_iolink); 1084 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED; 1085 if (kfd_dev_is_large_bar(kdev)) 1086 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL; 1087 1088 /* Fill in IOLINK subtype. 1089 * TODO: Fill-in other fields of iolink subtype 1090 */ 1091 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS; 1092 sub_type_hdr->proximity_domain_from = proximity_domain; 1093 #ifdef CONFIG_NUMA 1094 if (kdev->pdev->dev.numa_node == NUMA_NO_NODE) 1095 sub_type_hdr->proximity_domain_to = 0; 1096 else 1097 sub_type_hdr->proximity_domain_to = kdev->pdev->dev.numa_node; 1098 #else 1099 sub_type_hdr->proximity_domain_to = 0; 1100 #endif 1101 return 0; 1102 } 1103 1104 static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size, 1105 struct kfd_dev *kdev, 1106 struct kfd_dev *peer_kdev, 1107 struct crat_subtype_iolink *sub_type_hdr, 1108 uint32_t proximity_domain_from, 1109 uint32_t proximity_domain_to) 1110 { 1111 *avail_size -= sizeof(struct crat_subtype_iolink); 1112 if (*avail_size < 0) 1113 return -ENOMEM; 1114 1115 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink)); 1116 1117 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY; 1118 sub_type_hdr->length = sizeof(struct crat_subtype_iolink); 1119 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED | 1120 CRAT_IOLINK_FLAGS_BI_DIRECTIONAL; 1121 1122 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI; 1123 sub_type_hdr->proximity_domain_from = proximity_domain_from; 1124 sub_type_hdr->proximity_domain_to = proximity_domain_to; 1125 sub_type_hdr->num_hops_xgmi = 1126 amdgpu_amdkfd_get_xgmi_hops_count(kdev->kgd, peer_kdev->kgd); 1127 return 0; 1128 } 1129 1130 /* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU 1131 * 1132 * @pcrat_image: Fill in VCRAT for GPU 1133 * @size: [IN] allocated size of crat_image. 1134 * [OUT] actual size of data filled in crat_image 1135 */ 1136 static int kfd_create_vcrat_image_gpu(void *pcrat_image, 1137 size_t *size, struct kfd_dev *kdev, 1138 uint32_t proximity_domain) 1139 { 1140 struct crat_header *crat_table = (struct crat_header *)pcrat_image; 1141 struct crat_subtype_generic *sub_type_hdr; 1142 struct kfd_local_mem_info local_mem_info; 1143 struct kfd_topology_device *peer_dev; 1144 struct crat_subtype_computeunit *cu; 1145 struct kfd_cu_info cu_info; 1146 int avail_size = *size; 1147 uint32_t total_num_of_cu; 1148 int num_of_cache_entries = 0; 1149 int cache_mem_filled = 0; 1150 uint32_t nid = 0; 1151 int ret = 0; 1152 1153 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU) 1154 return -EINVAL; 1155 1156 /* Fill the CRAT Header. 1157 * Modify length and total_entries as subunits are added. 1158 */ 1159 avail_size -= sizeof(struct crat_header); 1160 if (avail_size < 0) 1161 return -ENOMEM; 1162 1163 memset(crat_table, 0, sizeof(struct crat_header)); 1164 1165 memcpy(&crat_table->signature, CRAT_SIGNATURE, 1166 sizeof(crat_table->signature)); 1167 /* Change length as we add more subtypes*/ 1168 crat_table->length = sizeof(struct crat_header); 1169 crat_table->num_domains = 1; 1170 crat_table->total_entries = 0; 1171 1172 /* Fill in Subtype: Compute Unit 1173 * First fill in the sub type header and then sub type data 1174 */ 1175 avail_size -= sizeof(struct crat_subtype_computeunit); 1176 if (avail_size < 0) 1177 return -ENOMEM; 1178 1179 sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1); 1180 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit)); 1181 1182 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY; 1183 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit); 1184 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 1185 1186 /* Fill CU subtype data */ 1187 cu = (struct crat_subtype_computeunit *)sub_type_hdr; 1188 cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT; 1189 cu->proximity_domain = proximity_domain; 1190 1191 amdgpu_amdkfd_get_cu_info(kdev->kgd, &cu_info); 1192 cu->num_simd_per_cu = cu_info.simd_per_cu; 1193 cu->num_simd_cores = cu_info.simd_per_cu * cu_info.cu_active_number; 1194 cu->max_waves_simd = cu_info.max_waves_per_simd; 1195 1196 cu->wave_front_size = cu_info.wave_front_size; 1197 cu->array_count = cu_info.num_shader_arrays_per_engine * 1198 cu_info.num_shader_engines; 1199 total_num_of_cu = (cu->array_count * cu_info.num_cu_per_sh); 1200 cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu); 1201 cu->num_cu_per_array = cu_info.num_cu_per_sh; 1202 cu->max_slots_scatch_cu = cu_info.max_scratch_slots_per_cu; 1203 cu->num_banks = cu_info.num_shader_engines; 1204 cu->lds_size_in_kb = cu_info.lds_size; 1205 1206 cu->hsa_capability = 0; 1207 1208 /* Check if this node supports IOMMU. During parsing this flag will 1209 * translate to HSA_CAP_ATS_PRESENT 1210 */ 1211 if (!kfd_iommu_check_device(kdev)) 1212 cu->hsa_capability |= CRAT_CU_FLAGS_IOMMU_PRESENT; 1213 1214 crat_table->length += sub_type_hdr->length; 1215 crat_table->total_entries++; 1216 1217 /* Fill in Subtype: Memory. Only on systems with large BAR (no 1218 * private FB), report memory as public. On other systems 1219 * report the total FB size (public+private) as a single 1220 * private heap. 1221 */ 1222 amdgpu_amdkfd_get_local_mem_info(kdev->kgd, &local_mem_info); 1223 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1224 sub_type_hdr->length); 1225 1226 if (debug_largebar) 1227 local_mem_info.local_mem_size_private = 0; 1228 1229 if (local_mem_info.local_mem_size_private == 0) 1230 ret = kfd_fill_gpu_memory_affinity(&avail_size, 1231 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC, 1232 local_mem_info.local_mem_size_public, 1233 (struct crat_subtype_memory *)sub_type_hdr, 1234 proximity_domain, 1235 &local_mem_info); 1236 else 1237 ret = kfd_fill_gpu_memory_affinity(&avail_size, 1238 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE, 1239 local_mem_info.local_mem_size_public + 1240 local_mem_info.local_mem_size_private, 1241 (struct crat_subtype_memory *)sub_type_hdr, 1242 proximity_domain, 1243 &local_mem_info); 1244 if (ret < 0) 1245 return ret; 1246 1247 crat_table->length += sizeof(struct crat_subtype_memory); 1248 crat_table->total_entries++; 1249 1250 /* TODO: Fill in cache information. This information is NOT readily 1251 * available in KGD 1252 */ 1253 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1254 sub_type_hdr->length); 1255 ret = kfd_fill_gpu_cache_info(kdev, cu->processor_id_low, 1256 avail_size, 1257 &cu_info, 1258 (struct crat_subtype_cache *)sub_type_hdr, 1259 &cache_mem_filled, 1260 &num_of_cache_entries); 1261 1262 if (ret < 0) 1263 return ret; 1264 1265 crat_table->length += cache_mem_filled; 1266 crat_table->total_entries += num_of_cache_entries; 1267 avail_size -= cache_mem_filled; 1268 1269 /* Fill in Subtype: IO_LINKS 1270 * Only direct links are added here which is Link from GPU to 1271 * to its NUMA node. Indirect links are added by userspace. 1272 */ 1273 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1274 cache_mem_filled); 1275 ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev, 1276 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain); 1277 1278 if (ret < 0) 1279 return ret; 1280 1281 crat_table->length += sub_type_hdr->length; 1282 crat_table->total_entries++; 1283 1284 1285 /* Fill in Subtype: IO_LINKS 1286 * Direct links from GPU to other GPUs through xGMI. 1287 * We will loop GPUs that already be processed (with lower value 1288 * of proximity_domain), add the link for the GPUs with same 1289 * hive id (from this GPU to other GPU) . The reversed iolink 1290 * (from other GPU to this GPU) will be added 1291 * in kfd_parse_subtype_iolink. 1292 */ 1293 if (kdev->hive_id) { 1294 for (nid = 0; nid < proximity_domain; ++nid) { 1295 peer_dev = kfd_topology_device_by_proximity_domain(nid); 1296 if (!peer_dev->gpu) 1297 continue; 1298 if (peer_dev->gpu->hive_id != kdev->hive_id) 1299 continue; 1300 sub_type_hdr = (typeof(sub_type_hdr))( 1301 (char *)sub_type_hdr + 1302 sizeof(struct crat_subtype_iolink)); 1303 ret = kfd_fill_gpu_xgmi_link_to_gpu( 1304 &avail_size, kdev, peer_dev->gpu, 1305 (struct crat_subtype_iolink *)sub_type_hdr, 1306 proximity_domain, nid); 1307 if (ret < 0) 1308 return ret; 1309 crat_table->length += sub_type_hdr->length; 1310 crat_table->total_entries++; 1311 } 1312 } 1313 *size = crat_table->length; 1314 pr_info("Virtual CRAT table created for GPU\n"); 1315 1316 return ret; 1317 } 1318 1319 /* kfd_create_crat_image_virtual - Allocates memory for CRAT image and 1320 * creates a Virtual CRAT (VCRAT) image 1321 * 1322 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory 1323 * 1324 * @crat_image: VCRAT image created because ACPI does not have a 1325 * CRAT for this device 1326 * @size: [OUT] size of virtual crat_image 1327 * @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device 1328 * COMPUTE_UNIT_GPU - Create VCRAT for GPU 1329 * (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU 1330 * -- this option is not currently implemented. 1331 * The assumption is that all AMD APUs will have CRAT 1332 * @kdev: Valid kfd_device required if flags contain COMPUTE_UNIT_GPU 1333 * 1334 * Return 0 if successful else return -ve value 1335 */ 1336 int kfd_create_crat_image_virtual(void **crat_image, size_t *size, 1337 int flags, struct kfd_dev *kdev, 1338 uint32_t proximity_domain) 1339 { 1340 void *pcrat_image = NULL; 1341 int ret = 0; 1342 1343 if (!crat_image) 1344 return -EINVAL; 1345 1346 *crat_image = NULL; 1347 1348 /* Allocate one VCRAT_SIZE_FOR_CPU for CPU virtual CRAT image and 1349 * VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image. This should cover 1350 * all the current conditions. A check is put not to overwrite beyond 1351 * allocated size 1352 */ 1353 switch (flags) { 1354 case COMPUTE_UNIT_CPU: 1355 pcrat_image = kmalloc(VCRAT_SIZE_FOR_CPU, GFP_KERNEL); 1356 if (!pcrat_image) 1357 return -ENOMEM; 1358 *size = VCRAT_SIZE_FOR_CPU; 1359 ret = kfd_create_vcrat_image_cpu(pcrat_image, size); 1360 break; 1361 case COMPUTE_UNIT_GPU: 1362 if (!kdev) 1363 return -EINVAL; 1364 pcrat_image = kmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL); 1365 if (!pcrat_image) 1366 return -ENOMEM; 1367 *size = VCRAT_SIZE_FOR_GPU; 1368 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev, 1369 proximity_domain); 1370 break; 1371 case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU): 1372 /* TODO: */ 1373 ret = -EINVAL; 1374 pr_err("VCRAT not implemented for APU\n"); 1375 break; 1376 default: 1377 ret = -EINVAL; 1378 } 1379 1380 if (!ret) 1381 *crat_image = pcrat_image; 1382 else 1383 kfree(pcrat_image); 1384 1385 return ret; 1386 } 1387 1388 1389 /* kfd_destroy_crat_image 1390 * 1391 * @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..) 1392 * 1393 */ 1394 void kfd_destroy_crat_image(void *crat_image) 1395 { 1396 kfree(crat_image); 1397 } 1398