1 /* 2 * Copyright 2014 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 */ 22 23 #include <linux/types.h> 24 #include <linux/kernel.h> 25 #include <linux/pci.h> 26 #include <linux/errno.h> 27 #include <linux/acpi.h> 28 #include <linux/hash.h> 29 #include <linux/cpufreq.h> 30 #include <linux/log2.h> 31 32 #include "kfd_priv.h" 33 #include "kfd_crat.h" 34 #include "kfd_topology.h" 35 36 static struct list_head topology_device_list; 37 static int topology_crat_parsed; 38 static struct kfd_system_properties sys_props; 39 40 static DECLARE_RWSEM(topology_lock); 41 42 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id) 43 { 44 struct kfd_topology_device *top_dev; 45 struct kfd_dev *device = NULL; 46 47 down_read(&topology_lock); 48 49 list_for_each_entry(top_dev, &topology_device_list, list) 50 if (top_dev->gpu_id == gpu_id) { 51 device = top_dev->gpu; 52 break; 53 } 54 55 up_read(&topology_lock); 56 57 return device; 58 } 59 60 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev) 61 { 62 struct kfd_topology_device *top_dev; 63 struct kfd_dev *device = NULL; 64 65 down_read(&topology_lock); 66 67 list_for_each_entry(top_dev, &topology_device_list, list) 68 if (top_dev->gpu->pdev == pdev) { 69 device = top_dev->gpu; 70 break; 71 } 72 73 up_read(&topology_lock); 74 75 return device; 76 } 77 78 static int kfd_topology_get_crat_acpi(void *crat_image, size_t *size) 79 { 80 struct acpi_table_header *crat_table; 81 acpi_status status; 82 83 if (!size) 84 return -EINVAL; 85 86 /* 87 * Fetch the CRAT table from ACPI 88 */ 89 status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table); 90 if (status == AE_NOT_FOUND) { 91 pr_warn("CRAT table not found\n"); 92 return -ENODATA; 93 } else if (ACPI_FAILURE(status)) { 94 const char *err = acpi_format_exception(status); 95 96 pr_err("CRAT table error: %s\n", err); 97 return -EINVAL; 98 } 99 100 if (*size >= crat_table->length && crat_image != NULL) 101 memcpy(crat_image, crat_table, crat_table->length); 102 103 *size = crat_table->length; 104 105 return 0; 106 } 107 108 static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev, 109 struct crat_subtype_computeunit *cu) 110 { 111 BUG_ON(!dev); 112 BUG_ON(!cu); 113 114 dev->node_props.cpu_cores_count = cu->num_cpu_cores; 115 dev->node_props.cpu_core_id_base = cu->processor_id_low; 116 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT) 117 dev->node_props.capability |= HSA_CAP_ATS_PRESENT; 118 119 pr_info("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores, 120 cu->processor_id_low); 121 } 122 123 static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev, 124 struct crat_subtype_computeunit *cu) 125 { 126 BUG_ON(!dev); 127 BUG_ON(!cu); 128 129 dev->node_props.simd_id_base = cu->processor_id_low; 130 dev->node_props.simd_count = cu->num_simd_cores; 131 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb; 132 dev->node_props.max_waves_per_simd = cu->max_waves_simd; 133 dev->node_props.wave_front_size = cu->wave_front_size; 134 dev->node_props.mem_banks_count = cu->num_banks; 135 dev->node_props.array_count = cu->num_arrays; 136 dev->node_props.cu_per_simd_array = cu->num_cu_per_array; 137 dev->node_props.simd_per_cu = cu->num_simd_per_cu; 138 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu; 139 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE) 140 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE; 141 pr_info("CU GPU: simds=%d id_base=%d\n", cu->num_simd_cores, 142 cu->processor_id_low); 143 } 144 145 /* kfd_parse_subtype_cu is called when the topology mutex is already acquired */ 146 static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu) 147 { 148 struct kfd_topology_device *dev; 149 int i = 0; 150 151 BUG_ON(!cu); 152 153 pr_info("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n", 154 cu->proximity_domain, cu->hsa_capability); 155 list_for_each_entry(dev, &topology_device_list, list) { 156 if (cu->proximity_domain == i) { 157 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT) 158 kfd_populated_cu_info_cpu(dev, cu); 159 160 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT) 161 kfd_populated_cu_info_gpu(dev, cu); 162 break; 163 } 164 i++; 165 } 166 167 return 0; 168 } 169 170 /* 171 * kfd_parse_subtype_mem is called when the topology mutex is 172 * already acquired 173 */ 174 static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem) 175 { 176 struct kfd_mem_properties *props; 177 struct kfd_topology_device *dev; 178 int i = 0; 179 180 BUG_ON(!mem); 181 182 pr_info("Found memory entry in CRAT table with proximity_domain=%d\n", 183 mem->promixity_domain); 184 list_for_each_entry(dev, &topology_device_list, list) { 185 if (mem->promixity_domain == i) { 186 props = kfd_alloc_struct(props); 187 if (props == NULL) 188 return -ENOMEM; 189 190 if (dev->node_props.cpu_cores_count == 0) 191 props->heap_type = HSA_MEM_HEAP_TYPE_FB_PRIVATE; 192 else 193 props->heap_type = HSA_MEM_HEAP_TYPE_SYSTEM; 194 195 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE) 196 props->flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE; 197 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE) 198 props->flags |= HSA_MEM_FLAGS_NON_VOLATILE; 199 200 props->size_in_bytes = 201 ((uint64_t)mem->length_high << 32) + 202 mem->length_low; 203 props->width = mem->width; 204 205 dev->mem_bank_count++; 206 list_add_tail(&props->list, &dev->mem_props); 207 208 break; 209 } 210 i++; 211 } 212 213 return 0; 214 } 215 216 /* 217 * kfd_parse_subtype_cache is called when the topology mutex 218 * is already acquired 219 */ 220 static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache) 221 { 222 struct kfd_cache_properties *props; 223 struct kfd_topology_device *dev; 224 uint32_t id; 225 226 BUG_ON(!cache); 227 228 id = cache->processor_id_low; 229 230 pr_info("Found cache entry in CRAT table with processor_id=%d\n", id); 231 list_for_each_entry(dev, &topology_device_list, list) 232 if (id == dev->node_props.cpu_core_id_base || 233 id == dev->node_props.simd_id_base) { 234 props = kfd_alloc_struct(props); 235 if (props == NULL) 236 return -ENOMEM; 237 238 props->processor_id_low = id; 239 props->cache_level = cache->cache_level; 240 props->cache_size = cache->cache_size; 241 props->cacheline_size = cache->cache_line_size; 242 props->cachelines_per_tag = cache->lines_per_tag; 243 props->cache_assoc = cache->associativity; 244 props->cache_latency = cache->cache_latency; 245 246 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE) 247 props->cache_type |= HSA_CACHE_TYPE_DATA; 248 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE) 249 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION; 250 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE) 251 props->cache_type |= HSA_CACHE_TYPE_CPU; 252 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE) 253 props->cache_type |= HSA_CACHE_TYPE_HSACU; 254 255 dev->cache_count++; 256 dev->node_props.caches_count++; 257 list_add_tail(&props->list, &dev->cache_props); 258 259 break; 260 } 261 262 return 0; 263 } 264 265 /* 266 * kfd_parse_subtype_iolink is called when the topology mutex 267 * is already acquired 268 */ 269 static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink) 270 { 271 struct kfd_iolink_properties *props; 272 struct kfd_topology_device *dev; 273 uint32_t i = 0; 274 uint32_t id_from; 275 uint32_t id_to; 276 277 BUG_ON(!iolink); 278 279 id_from = iolink->proximity_domain_from; 280 id_to = iolink->proximity_domain_to; 281 282 pr_info("Found IO link entry in CRAT table with id_from=%d\n", id_from); 283 list_for_each_entry(dev, &topology_device_list, list) { 284 if (id_from == i) { 285 props = kfd_alloc_struct(props); 286 if (props == NULL) 287 return -ENOMEM; 288 289 props->node_from = id_from; 290 props->node_to = id_to; 291 props->ver_maj = iolink->version_major; 292 props->ver_min = iolink->version_minor; 293 294 /* 295 * weight factor (derived from CDIR), currently always 1 296 */ 297 props->weight = 1; 298 299 props->min_latency = iolink->minimum_latency; 300 props->max_latency = iolink->maximum_latency; 301 props->min_bandwidth = iolink->minimum_bandwidth_mbs; 302 props->max_bandwidth = iolink->maximum_bandwidth_mbs; 303 props->rec_transfer_size = 304 iolink->recommended_transfer_size; 305 306 dev->io_link_count++; 307 dev->node_props.io_links_count++; 308 list_add_tail(&props->list, &dev->io_link_props); 309 310 break; 311 } 312 i++; 313 } 314 315 return 0; 316 } 317 318 static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr) 319 { 320 struct crat_subtype_computeunit *cu; 321 struct crat_subtype_memory *mem; 322 struct crat_subtype_cache *cache; 323 struct crat_subtype_iolink *iolink; 324 int ret = 0; 325 326 BUG_ON(!sub_type_hdr); 327 328 switch (sub_type_hdr->type) { 329 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY: 330 cu = (struct crat_subtype_computeunit *)sub_type_hdr; 331 ret = kfd_parse_subtype_cu(cu); 332 break; 333 case CRAT_SUBTYPE_MEMORY_AFFINITY: 334 mem = (struct crat_subtype_memory *)sub_type_hdr; 335 ret = kfd_parse_subtype_mem(mem); 336 break; 337 case CRAT_SUBTYPE_CACHE_AFFINITY: 338 cache = (struct crat_subtype_cache *)sub_type_hdr; 339 ret = kfd_parse_subtype_cache(cache); 340 break; 341 case CRAT_SUBTYPE_TLB_AFFINITY: 342 /* 343 * For now, nothing to do here 344 */ 345 pr_info("Found TLB entry in CRAT table (not processing)\n"); 346 break; 347 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY: 348 /* 349 * For now, nothing to do here 350 */ 351 pr_info("Found CCOMPUTE entry in CRAT table (not processing)\n"); 352 break; 353 case CRAT_SUBTYPE_IOLINK_AFFINITY: 354 iolink = (struct crat_subtype_iolink *)sub_type_hdr; 355 ret = kfd_parse_subtype_iolink(iolink); 356 break; 357 default: 358 pr_warn("Unknown subtype (%d) in CRAT\n", 359 sub_type_hdr->type); 360 } 361 362 return ret; 363 } 364 365 static void kfd_release_topology_device(struct kfd_topology_device *dev) 366 { 367 struct kfd_mem_properties *mem; 368 struct kfd_cache_properties *cache; 369 struct kfd_iolink_properties *iolink; 370 371 BUG_ON(!dev); 372 373 list_del(&dev->list); 374 375 while (dev->mem_props.next != &dev->mem_props) { 376 mem = container_of(dev->mem_props.next, 377 struct kfd_mem_properties, list); 378 list_del(&mem->list); 379 kfree(mem); 380 } 381 382 while (dev->cache_props.next != &dev->cache_props) { 383 cache = container_of(dev->cache_props.next, 384 struct kfd_cache_properties, list); 385 list_del(&cache->list); 386 kfree(cache); 387 } 388 389 while (dev->io_link_props.next != &dev->io_link_props) { 390 iolink = container_of(dev->io_link_props.next, 391 struct kfd_iolink_properties, list); 392 list_del(&iolink->list); 393 kfree(iolink); 394 } 395 396 kfree(dev); 397 398 sys_props.num_devices--; 399 } 400 401 static void kfd_release_live_view(void) 402 { 403 struct kfd_topology_device *dev; 404 405 while (topology_device_list.next != &topology_device_list) { 406 dev = container_of(topology_device_list.next, 407 struct kfd_topology_device, list); 408 kfd_release_topology_device(dev); 409 } 410 411 memset(&sys_props, 0, sizeof(sys_props)); 412 } 413 414 static struct kfd_topology_device *kfd_create_topology_device(void) 415 { 416 struct kfd_topology_device *dev; 417 418 dev = kfd_alloc_struct(dev); 419 if (dev == NULL) { 420 pr_err("No memory to allocate a topology device"); 421 return NULL; 422 } 423 424 INIT_LIST_HEAD(&dev->mem_props); 425 INIT_LIST_HEAD(&dev->cache_props); 426 INIT_LIST_HEAD(&dev->io_link_props); 427 428 list_add_tail(&dev->list, &topology_device_list); 429 sys_props.num_devices++; 430 431 return dev; 432 } 433 434 static int kfd_parse_crat_table(void *crat_image) 435 { 436 struct kfd_topology_device *top_dev; 437 struct crat_subtype_generic *sub_type_hdr; 438 uint16_t node_id; 439 int ret; 440 struct crat_header *crat_table = (struct crat_header *)crat_image; 441 uint16_t num_nodes; 442 uint32_t image_len; 443 444 if (!crat_image) 445 return -EINVAL; 446 447 num_nodes = crat_table->num_domains; 448 image_len = crat_table->length; 449 450 pr_info("Parsing CRAT table with %d nodes\n", num_nodes); 451 452 for (node_id = 0; node_id < num_nodes; node_id++) { 453 top_dev = kfd_create_topology_device(); 454 if (!top_dev) { 455 kfd_release_live_view(); 456 return -ENOMEM; 457 } 458 } 459 460 sys_props.platform_id = 461 (*((uint64_t *)crat_table->oem_id)) & CRAT_OEMID_64BIT_MASK; 462 sys_props.platform_oem = *((uint64_t *)crat_table->oem_table_id); 463 sys_props.platform_rev = crat_table->revision; 464 465 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1); 466 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) < 467 ((char *)crat_image) + image_len) { 468 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) { 469 ret = kfd_parse_subtype(sub_type_hdr); 470 if (ret != 0) { 471 kfd_release_live_view(); 472 return ret; 473 } 474 } 475 476 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 477 sub_type_hdr->length); 478 } 479 480 sys_props.generation_count++; 481 topology_crat_parsed = 1; 482 483 return 0; 484 } 485 486 487 #define sysfs_show_gen_prop(buffer, fmt, ...) \ 488 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__) 489 #define sysfs_show_32bit_prop(buffer, name, value) \ 490 sysfs_show_gen_prop(buffer, "%s %u\n", name, value) 491 #define sysfs_show_64bit_prop(buffer, name, value) \ 492 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value) 493 #define sysfs_show_32bit_val(buffer, value) \ 494 sysfs_show_gen_prop(buffer, "%u\n", value) 495 #define sysfs_show_str_val(buffer, value) \ 496 sysfs_show_gen_prop(buffer, "%s\n", value) 497 498 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr, 499 char *buffer) 500 { 501 ssize_t ret; 502 503 /* Making sure that the buffer is an empty string */ 504 buffer[0] = 0; 505 506 if (attr == &sys_props.attr_genid) { 507 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count); 508 } else if (attr == &sys_props.attr_props) { 509 sysfs_show_64bit_prop(buffer, "platform_oem", 510 sys_props.platform_oem); 511 sysfs_show_64bit_prop(buffer, "platform_id", 512 sys_props.platform_id); 513 ret = sysfs_show_64bit_prop(buffer, "platform_rev", 514 sys_props.platform_rev); 515 } else { 516 ret = -EINVAL; 517 } 518 519 return ret; 520 } 521 522 static const struct sysfs_ops sysprops_ops = { 523 .show = sysprops_show, 524 }; 525 526 static struct kobj_type sysprops_type = { 527 .sysfs_ops = &sysprops_ops, 528 }; 529 530 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr, 531 char *buffer) 532 { 533 ssize_t ret; 534 struct kfd_iolink_properties *iolink; 535 536 /* Making sure that the buffer is an empty string */ 537 buffer[0] = 0; 538 539 iolink = container_of(attr, struct kfd_iolink_properties, attr); 540 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type); 541 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj); 542 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min); 543 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from); 544 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to); 545 sysfs_show_32bit_prop(buffer, "weight", iolink->weight); 546 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency); 547 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency); 548 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth); 549 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth); 550 sysfs_show_32bit_prop(buffer, "recommended_transfer_size", 551 iolink->rec_transfer_size); 552 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags); 553 554 return ret; 555 } 556 557 static const struct sysfs_ops iolink_ops = { 558 .show = iolink_show, 559 }; 560 561 static struct kobj_type iolink_type = { 562 .sysfs_ops = &iolink_ops, 563 }; 564 565 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr, 566 char *buffer) 567 { 568 ssize_t ret; 569 struct kfd_mem_properties *mem; 570 571 /* Making sure that the buffer is an empty string */ 572 buffer[0] = 0; 573 574 mem = container_of(attr, struct kfd_mem_properties, attr); 575 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type); 576 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes); 577 sysfs_show_32bit_prop(buffer, "flags", mem->flags); 578 sysfs_show_32bit_prop(buffer, "width", mem->width); 579 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max); 580 581 return ret; 582 } 583 584 static const struct sysfs_ops mem_ops = { 585 .show = mem_show, 586 }; 587 588 static struct kobj_type mem_type = { 589 .sysfs_ops = &mem_ops, 590 }; 591 592 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr, 593 char *buffer) 594 { 595 ssize_t ret; 596 uint32_t i; 597 struct kfd_cache_properties *cache; 598 599 /* Making sure that the buffer is an empty string */ 600 buffer[0] = 0; 601 602 cache = container_of(attr, struct kfd_cache_properties, attr); 603 sysfs_show_32bit_prop(buffer, "processor_id_low", 604 cache->processor_id_low); 605 sysfs_show_32bit_prop(buffer, "level", cache->cache_level); 606 sysfs_show_32bit_prop(buffer, "size", cache->cache_size); 607 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size); 608 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag", 609 cache->cachelines_per_tag); 610 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc); 611 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency); 612 sysfs_show_32bit_prop(buffer, "type", cache->cache_type); 613 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer); 614 for (i = 0; i < KFD_TOPOLOGY_CPU_SIBLINGS; i++) 615 ret = snprintf(buffer, PAGE_SIZE, "%s%d%s", 616 buffer, cache->sibling_map[i], 617 (i == KFD_TOPOLOGY_CPU_SIBLINGS-1) ? 618 "\n" : ","); 619 620 return ret; 621 } 622 623 static const struct sysfs_ops cache_ops = { 624 .show = kfd_cache_show, 625 }; 626 627 static struct kobj_type cache_type = { 628 .sysfs_ops = &cache_ops, 629 }; 630 631 static ssize_t node_show(struct kobject *kobj, struct attribute *attr, 632 char *buffer) 633 { 634 struct kfd_topology_device *dev; 635 char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE]; 636 uint32_t i; 637 uint32_t log_max_watch_addr; 638 639 /* Making sure that the buffer is an empty string */ 640 buffer[0] = 0; 641 642 if (strcmp(attr->name, "gpu_id") == 0) { 643 dev = container_of(attr, struct kfd_topology_device, 644 attr_gpuid); 645 return sysfs_show_32bit_val(buffer, dev->gpu_id); 646 } 647 648 if (strcmp(attr->name, "name") == 0) { 649 dev = container_of(attr, struct kfd_topology_device, 650 attr_name); 651 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) { 652 public_name[i] = 653 (char)dev->node_props.marketing_name[i]; 654 if (dev->node_props.marketing_name[i] == 0) 655 break; 656 } 657 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0; 658 return sysfs_show_str_val(buffer, public_name); 659 } 660 661 dev = container_of(attr, struct kfd_topology_device, 662 attr_props); 663 sysfs_show_32bit_prop(buffer, "cpu_cores_count", 664 dev->node_props.cpu_cores_count); 665 sysfs_show_32bit_prop(buffer, "simd_count", 666 dev->node_props.simd_count); 667 668 if (dev->mem_bank_count < dev->node_props.mem_banks_count) { 669 pr_info_once("kfd: mem_banks_count truncated from %d to %d\n", 670 dev->node_props.mem_banks_count, 671 dev->mem_bank_count); 672 sysfs_show_32bit_prop(buffer, "mem_banks_count", 673 dev->mem_bank_count); 674 } else { 675 sysfs_show_32bit_prop(buffer, "mem_banks_count", 676 dev->node_props.mem_banks_count); 677 } 678 679 sysfs_show_32bit_prop(buffer, "caches_count", 680 dev->node_props.caches_count); 681 sysfs_show_32bit_prop(buffer, "io_links_count", 682 dev->node_props.io_links_count); 683 sysfs_show_32bit_prop(buffer, "cpu_core_id_base", 684 dev->node_props.cpu_core_id_base); 685 sysfs_show_32bit_prop(buffer, "simd_id_base", 686 dev->node_props.simd_id_base); 687 sysfs_show_32bit_prop(buffer, "max_waves_per_simd", 688 dev->node_props.max_waves_per_simd); 689 sysfs_show_32bit_prop(buffer, "lds_size_in_kb", 690 dev->node_props.lds_size_in_kb); 691 sysfs_show_32bit_prop(buffer, "gds_size_in_kb", 692 dev->node_props.gds_size_in_kb); 693 sysfs_show_32bit_prop(buffer, "wave_front_size", 694 dev->node_props.wave_front_size); 695 sysfs_show_32bit_prop(buffer, "array_count", 696 dev->node_props.array_count); 697 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine", 698 dev->node_props.simd_arrays_per_engine); 699 sysfs_show_32bit_prop(buffer, "cu_per_simd_array", 700 dev->node_props.cu_per_simd_array); 701 sysfs_show_32bit_prop(buffer, "simd_per_cu", 702 dev->node_props.simd_per_cu); 703 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu", 704 dev->node_props.max_slots_scratch_cu); 705 sysfs_show_32bit_prop(buffer, "vendor_id", 706 dev->node_props.vendor_id); 707 sysfs_show_32bit_prop(buffer, "device_id", 708 dev->node_props.device_id); 709 sysfs_show_32bit_prop(buffer, "location_id", 710 dev->node_props.location_id); 711 712 if (dev->gpu) { 713 log_max_watch_addr = 714 __ilog2_u32(dev->gpu->device_info->num_of_watch_points); 715 716 if (log_max_watch_addr) { 717 dev->node_props.capability |= 718 HSA_CAP_WATCH_POINTS_SUPPORTED; 719 720 dev->node_props.capability |= 721 ((log_max_watch_addr << 722 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) & 723 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK); 724 } 725 726 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute", 727 dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz( 728 dev->gpu->kgd)); 729 730 sysfs_show_64bit_prop(buffer, "local_mem_size", 731 (unsigned long long int) 0); 732 733 sysfs_show_32bit_prop(buffer, "fw_version", 734 dev->gpu->kfd2kgd->get_fw_version( 735 dev->gpu->kgd, 736 KGD_ENGINE_MEC1)); 737 sysfs_show_32bit_prop(buffer, "capability", 738 dev->node_props.capability); 739 } 740 741 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute", 742 cpufreq_quick_get_max(0)/1000); 743 } 744 745 static const struct sysfs_ops node_ops = { 746 .show = node_show, 747 }; 748 749 static struct kobj_type node_type = { 750 .sysfs_ops = &node_ops, 751 }; 752 753 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr) 754 { 755 sysfs_remove_file(kobj, attr); 756 kobject_del(kobj); 757 kobject_put(kobj); 758 } 759 760 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) 761 { 762 struct kfd_iolink_properties *iolink; 763 struct kfd_cache_properties *cache; 764 struct kfd_mem_properties *mem; 765 766 BUG_ON(!dev); 767 768 if (dev->kobj_iolink) { 769 list_for_each_entry(iolink, &dev->io_link_props, list) 770 if (iolink->kobj) { 771 kfd_remove_sysfs_file(iolink->kobj, 772 &iolink->attr); 773 iolink->kobj = NULL; 774 } 775 kobject_del(dev->kobj_iolink); 776 kobject_put(dev->kobj_iolink); 777 dev->kobj_iolink = NULL; 778 } 779 780 if (dev->kobj_cache) { 781 list_for_each_entry(cache, &dev->cache_props, list) 782 if (cache->kobj) { 783 kfd_remove_sysfs_file(cache->kobj, 784 &cache->attr); 785 cache->kobj = NULL; 786 } 787 kobject_del(dev->kobj_cache); 788 kobject_put(dev->kobj_cache); 789 dev->kobj_cache = NULL; 790 } 791 792 if (dev->kobj_mem) { 793 list_for_each_entry(mem, &dev->mem_props, list) 794 if (mem->kobj) { 795 kfd_remove_sysfs_file(mem->kobj, &mem->attr); 796 mem->kobj = NULL; 797 } 798 kobject_del(dev->kobj_mem); 799 kobject_put(dev->kobj_mem); 800 dev->kobj_mem = NULL; 801 } 802 803 if (dev->kobj_node) { 804 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid); 805 sysfs_remove_file(dev->kobj_node, &dev->attr_name); 806 sysfs_remove_file(dev->kobj_node, &dev->attr_props); 807 kobject_del(dev->kobj_node); 808 kobject_put(dev->kobj_node); 809 dev->kobj_node = NULL; 810 } 811 } 812 813 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, 814 uint32_t id) 815 { 816 struct kfd_iolink_properties *iolink; 817 struct kfd_cache_properties *cache; 818 struct kfd_mem_properties *mem; 819 int ret; 820 uint32_t i; 821 822 BUG_ON(!dev); 823 824 /* 825 * Creating the sysfs folders 826 */ 827 BUG_ON(dev->kobj_node); 828 dev->kobj_node = kfd_alloc_struct(dev->kobj_node); 829 if (!dev->kobj_node) 830 return -ENOMEM; 831 832 ret = kobject_init_and_add(dev->kobj_node, &node_type, 833 sys_props.kobj_nodes, "%d", id); 834 if (ret < 0) 835 return ret; 836 837 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node); 838 if (!dev->kobj_mem) 839 return -ENOMEM; 840 841 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node); 842 if (!dev->kobj_cache) 843 return -ENOMEM; 844 845 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node); 846 if (!dev->kobj_iolink) 847 return -ENOMEM; 848 849 /* 850 * Creating sysfs files for node properties 851 */ 852 dev->attr_gpuid.name = "gpu_id"; 853 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE; 854 sysfs_attr_init(&dev->attr_gpuid); 855 dev->attr_name.name = "name"; 856 dev->attr_name.mode = KFD_SYSFS_FILE_MODE; 857 sysfs_attr_init(&dev->attr_name); 858 dev->attr_props.name = "properties"; 859 dev->attr_props.mode = KFD_SYSFS_FILE_MODE; 860 sysfs_attr_init(&dev->attr_props); 861 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid); 862 if (ret < 0) 863 return ret; 864 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name); 865 if (ret < 0) 866 return ret; 867 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props); 868 if (ret < 0) 869 return ret; 870 871 i = 0; 872 list_for_each_entry(mem, &dev->mem_props, list) { 873 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 874 if (!mem->kobj) 875 return -ENOMEM; 876 ret = kobject_init_and_add(mem->kobj, &mem_type, 877 dev->kobj_mem, "%d", i); 878 if (ret < 0) 879 return ret; 880 881 mem->attr.name = "properties"; 882 mem->attr.mode = KFD_SYSFS_FILE_MODE; 883 sysfs_attr_init(&mem->attr); 884 ret = sysfs_create_file(mem->kobj, &mem->attr); 885 if (ret < 0) 886 return ret; 887 i++; 888 } 889 890 i = 0; 891 list_for_each_entry(cache, &dev->cache_props, list) { 892 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 893 if (!cache->kobj) 894 return -ENOMEM; 895 ret = kobject_init_and_add(cache->kobj, &cache_type, 896 dev->kobj_cache, "%d", i); 897 if (ret < 0) 898 return ret; 899 900 cache->attr.name = "properties"; 901 cache->attr.mode = KFD_SYSFS_FILE_MODE; 902 sysfs_attr_init(&cache->attr); 903 ret = sysfs_create_file(cache->kobj, &cache->attr); 904 if (ret < 0) 905 return ret; 906 i++; 907 } 908 909 i = 0; 910 list_for_each_entry(iolink, &dev->io_link_props, list) { 911 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 912 if (!iolink->kobj) 913 return -ENOMEM; 914 ret = kobject_init_and_add(iolink->kobj, &iolink_type, 915 dev->kobj_iolink, "%d", i); 916 if (ret < 0) 917 return ret; 918 919 iolink->attr.name = "properties"; 920 iolink->attr.mode = KFD_SYSFS_FILE_MODE; 921 sysfs_attr_init(&iolink->attr); 922 ret = sysfs_create_file(iolink->kobj, &iolink->attr); 923 if (ret < 0) 924 return ret; 925 i++; 926 } 927 928 return 0; 929 } 930 931 static int kfd_build_sysfs_node_tree(void) 932 { 933 struct kfd_topology_device *dev; 934 int ret; 935 uint32_t i = 0; 936 937 list_for_each_entry(dev, &topology_device_list, list) { 938 ret = kfd_build_sysfs_node_entry(dev, i); 939 if (ret < 0) 940 return ret; 941 i++; 942 } 943 944 return 0; 945 } 946 947 static void kfd_remove_sysfs_node_tree(void) 948 { 949 struct kfd_topology_device *dev; 950 951 list_for_each_entry(dev, &topology_device_list, list) 952 kfd_remove_sysfs_node_entry(dev); 953 } 954 955 static int kfd_topology_update_sysfs(void) 956 { 957 int ret; 958 959 pr_info("Creating topology SYSFS entries\n"); 960 if (sys_props.kobj_topology == NULL) { 961 sys_props.kobj_topology = 962 kfd_alloc_struct(sys_props.kobj_topology); 963 if (!sys_props.kobj_topology) 964 return -ENOMEM; 965 966 ret = kobject_init_and_add(sys_props.kobj_topology, 967 &sysprops_type, &kfd_device->kobj, 968 "topology"); 969 if (ret < 0) 970 return ret; 971 972 sys_props.kobj_nodes = kobject_create_and_add("nodes", 973 sys_props.kobj_topology); 974 if (!sys_props.kobj_nodes) 975 return -ENOMEM; 976 977 sys_props.attr_genid.name = "generation_id"; 978 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE; 979 sysfs_attr_init(&sys_props.attr_genid); 980 ret = sysfs_create_file(sys_props.kobj_topology, 981 &sys_props.attr_genid); 982 if (ret < 0) 983 return ret; 984 985 sys_props.attr_props.name = "system_properties"; 986 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE; 987 sysfs_attr_init(&sys_props.attr_props); 988 ret = sysfs_create_file(sys_props.kobj_topology, 989 &sys_props.attr_props); 990 if (ret < 0) 991 return ret; 992 } 993 994 kfd_remove_sysfs_node_tree(); 995 996 return kfd_build_sysfs_node_tree(); 997 } 998 999 static void kfd_topology_release_sysfs(void) 1000 { 1001 kfd_remove_sysfs_node_tree(); 1002 if (sys_props.kobj_topology) { 1003 sysfs_remove_file(sys_props.kobj_topology, 1004 &sys_props.attr_genid); 1005 sysfs_remove_file(sys_props.kobj_topology, 1006 &sys_props.attr_props); 1007 if (sys_props.kobj_nodes) { 1008 kobject_del(sys_props.kobj_nodes); 1009 kobject_put(sys_props.kobj_nodes); 1010 sys_props.kobj_nodes = NULL; 1011 } 1012 kobject_del(sys_props.kobj_topology); 1013 kobject_put(sys_props.kobj_topology); 1014 sys_props.kobj_topology = NULL; 1015 } 1016 } 1017 1018 int kfd_topology_init(void) 1019 { 1020 void *crat_image = NULL; 1021 size_t image_size = 0; 1022 int ret; 1023 1024 /* 1025 * Initialize the head for the topology device list 1026 */ 1027 INIT_LIST_HEAD(&topology_device_list); 1028 init_rwsem(&topology_lock); 1029 topology_crat_parsed = 0; 1030 1031 memset(&sys_props, 0, sizeof(sys_props)); 1032 1033 /* 1034 * Get the CRAT image from the ACPI 1035 */ 1036 ret = kfd_topology_get_crat_acpi(crat_image, &image_size); 1037 if (ret == 0 && image_size > 0) { 1038 pr_info("Found CRAT image with size=%zd\n", image_size); 1039 crat_image = kmalloc(image_size, GFP_KERNEL); 1040 if (!crat_image) { 1041 ret = -ENOMEM; 1042 pr_err("No memory for allocating CRAT image\n"); 1043 goto err; 1044 } 1045 ret = kfd_topology_get_crat_acpi(crat_image, &image_size); 1046 1047 if (ret == 0) { 1048 down_write(&topology_lock); 1049 ret = kfd_parse_crat_table(crat_image); 1050 if (ret == 0) 1051 ret = kfd_topology_update_sysfs(); 1052 up_write(&topology_lock); 1053 } else { 1054 pr_err("Couldn't get CRAT table size from ACPI\n"); 1055 } 1056 kfree(crat_image); 1057 } else if (ret == -ENODATA) { 1058 ret = 0; 1059 } else { 1060 pr_err("Couldn't get CRAT table size from ACPI\n"); 1061 } 1062 1063 err: 1064 pr_info("Finished initializing topology ret=%d\n", ret); 1065 return ret; 1066 } 1067 1068 void kfd_topology_shutdown(void) 1069 { 1070 kfd_topology_release_sysfs(); 1071 kfd_release_live_view(); 1072 } 1073 1074 static void kfd_debug_print_topology(void) 1075 { 1076 struct kfd_topology_device *dev; 1077 uint32_t i = 0; 1078 1079 pr_info("DEBUG PRINT OF TOPOLOGY:"); 1080 list_for_each_entry(dev, &topology_device_list, list) { 1081 pr_info("Node: %d\n", i); 1082 pr_info("\tGPU assigned: %s\n", (dev->gpu ? "yes" : "no")); 1083 pr_info("\tCPU count: %d\n", dev->node_props.cpu_cores_count); 1084 pr_info("\tSIMD count: %d", dev->node_props.simd_count); 1085 i++; 1086 } 1087 } 1088 1089 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) 1090 { 1091 uint32_t hashout; 1092 uint32_t buf[7]; 1093 uint64_t local_mem_size; 1094 int i; 1095 1096 if (!gpu) 1097 return 0; 1098 1099 local_mem_size = gpu->kfd2kgd->get_vmem_size(gpu->kgd); 1100 1101 buf[0] = gpu->pdev->devfn; 1102 buf[1] = gpu->pdev->subsystem_vendor; 1103 buf[2] = gpu->pdev->subsystem_device; 1104 buf[3] = gpu->pdev->device; 1105 buf[4] = gpu->pdev->bus->number; 1106 buf[5] = lower_32_bits(local_mem_size); 1107 buf[6] = upper_32_bits(local_mem_size); 1108 1109 for (i = 0, hashout = 0; i < 7; i++) 1110 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); 1111 1112 return hashout; 1113 } 1114 1115 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) 1116 { 1117 struct kfd_topology_device *dev; 1118 struct kfd_topology_device *out_dev = NULL; 1119 1120 BUG_ON(!gpu); 1121 1122 list_for_each_entry(dev, &topology_device_list, list) 1123 if (dev->gpu == NULL && dev->node_props.simd_count > 0) { 1124 dev->gpu = gpu; 1125 out_dev = dev; 1126 break; 1127 } 1128 1129 return out_dev; 1130 } 1131 1132 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival) 1133 { 1134 /* 1135 * TODO: Generate an event for thunk about the arrival/removal 1136 * of the GPU 1137 */ 1138 } 1139 1140 int kfd_topology_add_device(struct kfd_dev *gpu) 1141 { 1142 uint32_t gpu_id; 1143 struct kfd_topology_device *dev; 1144 int res; 1145 1146 BUG_ON(!gpu); 1147 1148 gpu_id = kfd_generate_gpu_id(gpu); 1149 1150 pr_debug("kfd: Adding new GPU (ID: 0x%x) to topology\n", gpu_id); 1151 1152 down_write(&topology_lock); 1153 /* 1154 * Try to assign the GPU to existing topology device (generated from 1155 * CRAT table 1156 */ 1157 dev = kfd_assign_gpu(gpu); 1158 if (!dev) { 1159 pr_info("GPU was not found in the current topology. Extending.\n"); 1160 kfd_debug_print_topology(); 1161 dev = kfd_create_topology_device(); 1162 if (!dev) { 1163 res = -ENOMEM; 1164 goto err; 1165 } 1166 dev->gpu = gpu; 1167 1168 /* 1169 * TODO: Make a call to retrieve topology information from the 1170 * GPU vBIOS 1171 */ 1172 1173 /* 1174 * Update the SYSFS tree, since we added another topology device 1175 */ 1176 if (kfd_topology_update_sysfs() < 0) 1177 kfd_topology_release_sysfs(); 1178 1179 } 1180 1181 dev->gpu_id = gpu_id; 1182 gpu->id = gpu_id; 1183 dev->node_props.vendor_id = gpu->pdev->vendor; 1184 dev->node_props.device_id = gpu->pdev->device; 1185 dev->node_props.location_id = (gpu->pdev->bus->number << 24) + 1186 (gpu->pdev->devfn & 0xffffff); 1187 /* 1188 * TODO: Retrieve max engine clock values from KGD 1189 */ 1190 1191 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) { 1192 dev->node_props.capability |= HSA_CAP_DOORBELL_PACKET_TYPE; 1193 pr_info("amdkfd: adding doorbell packet type capability\n"); 1194 } 1195 1196 res = 0; 1197 1198 err: 1199 up_write(&topology_lock); 1200 1201 if (res == 0) 1202 kfd_notify_gpu_change(gpu_id, 1); 1203 1204 return res; 1205 } 1206 1207 int kfd_topology_remove_device(struct kfd_dev *gpu) 1208 { 1209 struct kfd_topology_device *dev; 1210 uint32_t gpu_id; 1211 int res = -ENODEV; 1212 1213 BUG_ON(!gpu); 1214 1215 down_write(&topology_lock); 1216 1217 list_for_each_entry(dev, &topology_device_list, list) 1218 if (dev->gpu == gpu) { 1219 gpu_id = dev->gpu_id; 1220 kfd_remove_sysfs_node_entry(dev); 1221 kfd_release_topology_device(dev); 1222 res = 0; 1223 if (kfd_topology_update_sysfs() < 0) 1224 kfd_topology_release_sysfs(); 1225 break; 1226 } 1227 1228 up_write(&topology_lock); 1229 1230 if (res == 0) 1231 kfd_notify_gpu_change(gpu_id, 0); 1232 1233 return res; 1234 } 1235 1236 /* 1237 * When idx is out of bounds, the function will return NULL 1238 */ 1239 struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx) 1240 { 1241 1242 struct kfd_topology_device *top_dev; 1243 struct kfd_dev *device = NULL; 1244 uint8_t device_idx = 0; 1245 1246 down_read(&topology_lock); 1247 1248 list_for_each_entry(top_dev, &topology_device_list, list) { 1249 if (device_idx == idx) { 1250 device = top_dev->gpu; 1251 break; 1252 } 1253 1254 device_idx++; 1255 } 1256 1257 up_read(&topology_lock); 1258 1259 return device; 1260 1261 } 1262