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 #include <linux/dmi.h> 32 #include <linux/atomic.h> 33 34 #include "kfd_priv.h" 35 #include "kfd_crat.h" 36 #include "kfd_topology.h" 37 #include "kfd_device_queue_manager.h" 38 #include "kfd_iommu.h" 39 40 /* topology_device_list - Master list of all topology devices */ 41 static struct list_head topology_device_list; 42 static struct kfd_system_properties sys_props; 43 44 static DECLARE_RWSEM(topology_lock); 45 static atomic_t topology_crat_proximity_domain; 46 47 struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 48 uint32_t proximity_domain) 49 { 50 struct kfd_topology_device *top_dev; 51 struct kfd_topology_device *device = NULL; 52 53 down_read(&topology_lock); 54 55 list_for_each_entry(top_dev, &topology_device_list, list) 56 if (top_dev->proximity_domain == proximity_domain) { 57 device = top_dev; 58 break; 59 } 60 61 up_read(&topology_lock); 62 63 return device; 64 } 65 66 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id) 67 { 68 struct kfd_topology_device *top_dev; 69 struct kfd_dev *device = NULL; 70 71 down_read(&topology_lock); 72 73 list_for_each_entry(top_dev, &topology_device_list, list) 74 if (top_dev->gpu_id == gpu_id) { 75 device = top_dev->gpu; 76 break; 77 } 78 79 up_read(&topology_lock); 80 81 return device; 82 } 83 84 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev) 85 { 86 struct kfd_topology_device *top_dev; 87 struct kfd_dev *device = NULL; 88 89 down_read(&topology_lock); 90 91 list_for_each_entry(top_dev, &topology_device_list, list) 92 if (top_dev->gpu->pdev == pdev) { 93 device = top_dev->gpu; 94 break; 95 } 96 97 up_read(&topology_lock); 98 99 return device; 100 } 101 102 /* Called with write topology_lock acquired */ 103 static void kfd_release_topology_device(struct kfd_topology_device *dev) 104 { 105 struct kfd_mem_properties *mem; 106 struct kfd_cache_properties *cache; 107 struct kfd_iolink_properties *iolink; 108 struct kfd_perf_properties *perf; 109 110 list_del(&dev->list); 111 112 while (dev->mem_props.next != &dev->mem_props) { 113 mem = container_of(dev->mem_props.next, 114 struct kfd_mem_properties, list); 115 list_del(&mem->list); 116 kfree(mem); 117 } 118 119 while (dev->cache_props.next != &dev->cache_props) { 120 cache = container_of(dev->cache_props.next, 121 struct kfd_cache_properties, list); 122 list_del(&cache->list); 123 kfree(cache); 124 } 125 126 while (dev->io_link_props.next != &dev->io_link_props) { 127 iolink = container_of(dev->io_link_props.next, 128 struct kfd_iolink_properties, list); 129 list_del(&iolink->list); 130 kfree(iolink); 131 } 132 133 while (dev->perf_props.next != &dev->perf_props) { 134 perf = container_of(dev->perf_props.next, 135 struct kfd_perf_properties, list); 136 list_del(&perf->list); 137 kfree(perf); 138 } 139 140 kfree(dev); 141 } 142 143 void kfd_release_topology_device_list(struct list_head *device_list) 144 { 145 struct kfd_topology_device *dev; 146 147 while (!list_empty(device_list)) { 148 dev = list_first_entry(device_list, 149 struct kfd_topology_device, list); 150 kfd_release_topology_device(dev); 151 } 152 } 153 154 static void kfd_release_live_view(void) 155 { 156 kfd_release_topology_device_list(&topology_device_list); 157 memset(&sys_props, 0, sizeof(sys_props)); 158 } 159 160 struct kfd_topology_device *kfd_create_topology_device( 161 struct list_head *device_list) 162 { 163 struct kfd_topology_device *dev; 164 165 dev = kfd_alloc_struct(dev); 166 if (!dev) { 167 pr_err("No memory to allocate a topology device"); 168 return NULL; 169 } 170 171 INIT_LIST_HEAD(&dev->mem_props); 172 INIT_LIST_HEAD(&dev->cache_props); 173 INIT_LIST_HEAD(&dev->io_link_props); 174 INIT_LIST_HEAD(&dev->perf_props); 175 176 list_add_tail(&dev->list, device_list); 177 178 return dev; 179 } 180 181 182 #define sysfs_show_gen_prop(buffer, fmt, ...) \ 183 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__) 184 #define sysfs_show_32bit_prop(buffer, name, value) \ 185 sysfs_show_gen_prop(buffer, "%s %u\n", name, value) 186 #define sysfs_show_64bit_prop(buffer, name, value) \ 187 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value) 188 #define sysfs_show_32bit_val(buffer, value) \ 189 sysfs_show_gen_prop(buffer, "%u\n", value) 190 #define sysfs_show_str_val(buffer, value) \ 191 sysfs_show_gen_prop(buffer, "%s\n", value) 192 193 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr, 194 char *buffer) 195 { 196 ssize_t ret; 197 198 /* Making sure that the buffer is an empty string */ 199 buffer[0] = 0; 200 201 if (attr == &sys_props.attr_genid) { 202 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count); 203 } else if (attr == &sys_props.attr_props) { 204 sysfs_show_64bit_prop(buffer, "platform_oem", 205 sys_props.platform_oem); 206 sysfs_show_64bit_prop(buffer, "platform_id", 207 sys_props.platform_id); 208 ret = sysfs_show_64bit_prop(buffer, "platform_rev", 209 sys_props.platform_rev); 210 } else { 211 ret = -EINVAL; 212 } 213 214 return ret; 215 } 216 217 static void kfd_topology_kobj_release(struct kobject *kobj) 218 { 219 kfree(kobj); 220 } 221 222 static const struct sysfs_ops sysprops_ops = { 223 .show = sysprops_show, 224 }; 225 226 static struct kobj_type sysprops_type = { 227 .release = kfd_topology_kobj_release, 228 .sysfs_ops = &sysprops_ops, 229 }; 230 231 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr, 232 char *buffer) 233 { 234 ssize_t ret; 235 struct kfd_iolink_properties *iolink; 236 237 /* Making sure that the buffer is an empty string */ 238 buffer[0] = 0; 239 240 iolink = container_of(attr, struct kfd_iolink_properties, attr); 241 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type); 242 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj); 243 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min); 244 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from); 245 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to); 246 sysfs_show_32bit_prop(buffer, "weight", iolink->weight); 247 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency); 248 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency); 249 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth); 250 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth); 251 sysfs_show_32bit_prop(buffer, "recommended_transfer_size", 252 iolink->rec_transfer_size); 253 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags); 254 255 return ret; 256 } 257 258 static const struct sysfs_ops iolink_ops = { 259 .show = iolink_show, 260 }; 261 262 static struct kobj_type iolink_type = { 263 .release = kfd_topology_kobj_release, 264 .sysfs_ops = &iolink_ops, 265 }; 266 267 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr, 268 char *buffer) 269 { 270 ssize_t ret; 271 struct kfd_mem_properties *mem; 272 273 /* Making sure that the buffer is an empty string */ 274 buffer[0] = 0; 275 276 mem = container_of(attr, struct kfd_mem_properties, attr); 277 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type); 278 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes); 279 sysfs_show_32bit_prop(buffer, "flags", mem->flags); 280 sysfs_show_32bit_prop(buffer, "width", mem->width); 281 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max); 282 283 return ret; 284 } 285 286 static const struct sysfs_ops mem_ops = { 287 .show = mem_show, 288 }; 289 290 static struct kobj_type mem_type = { 291 .release = kfd_topology_kobj_release, 292 .sysfs_ops = &mem_ops, 293 }; 294 295 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr, 296 char *buffer) 297 { 298 ssize_t ret; 299 uint32_t i, j; 300 struct kfd_cache_properties *cache; 301 302 /* Making sure that the buffer is an empty string */ 303 buffer[0] = 0; 304 305 cache = container_of(attr, struct kfd_cache_properties, attr); 306 sysfs_show_32bit_prop(buffer, "processor_id_low", 307 cache->processor_id_low); 308 sysfs_show_32bit_prop(buffer, "level", cache->cache_level); 309 sysfs_show_32bit_prop(buffer, "size", cache->cache_size); 310 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size); 311 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag", 312 cache->cachelines_per_tag); 313 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc); 314 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency); 315 sysfs_show_32bit_prop(buffer, "type", cache->cache_type); 316 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer); 317 for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++) 318 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) { 319 /* Check each bit */ 320 if (cache->sibling_map[i] & (1 << j)) 321 ret = snprintf(buffer, PAGE_SIZE, 322 "%s%d%s", buffer, 1, ","); 323 else 324 ret = snprintf(buffer, PAGE_SIZE, 325 "%s%d%s", buffer, 0, ","); 326 } 327 /* Replace the last "," with end of line */ 328 *(buffer + strlen(buffer) - 1) = 0xA; 329 return ret; 330 } 331 332 static const struct sysfs_ops cache_ops = { 333 .show = kfd_cache_show, 334 }; 335 336 static struct kobj_type cache_type = { 337 .release = kfd_topology_kobj_release, 338 .sysfs_ops = &cache_ops, 339 }; 340 341 /****** Sysfs of Performance Counters ******/ 342 343 struct kfd_perf_attr { 344 struct kobj_attribute attr; 345 uint32_t data; 346 }; 347 348 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs, 349 char *buf) 350 { 351 struct kfd_perf_attr *attr; 352 353 buf[0] = 0; 354 attr = container_of(attrs, struct kfd_perf_attr, attr); 355 if (!attr->data) /* invalid data for PMC */ 356 return 0; 357 else 358 return sysfs_show_32bit_val(buf, attr->data); 359 } 360 361 #define KFD_PERF_DESC(_name, _data) \ 362 { \ 363 .attr = __ATTR(_name, 0444, perf_show, NULL), \ 364 .data = _data, \ 365 } 366 367 static struct kfd_perf_attr perf_attr_iommu[] = { 368 KFD_PERF_DESC(max_concurrent, 0), 369 KFD_PERF_DESC(num_counters, 0), 370 KFD_PERF_DESC(counter_ids, 0), 371 }; 372 /****************************************/ 373 374 static ssize_t node_show(struct kobject *kobj, struct attribute *attr, 375 char *buffer) 376 { 377 struct kfd_topology_device *dev; 378 char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE]; 379 uint32_t i; 380 uint32_t log_max_watch_addr; 381 382 /* Making sure that the buffer is an empty string */ 383 buffer[0] = 0; 384 385 if (strcmp(attr->name, "gpu_id") == 0) { 386 dev = container_of(attr, struct kfd_topology_device, 387 attr_gpuid); 388 return sysfs_show_32bit_val(buffer, dev->gpu_id); 389 } 390 391 if (strcmp(attr->name, "name") == 0) { 392 dev = container_of(attr, struct kfd_topology_device, 393 attr_name); 394 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) { 395 public_name[i] = 396 (char)dev->node_props.marketing_name[i]; 397 if (dev->node_props.marketing_name[i] == 0) 398 break; 399 } 400 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0; 401 return sysfs_show_str_val(buffer, public_name); 402 } 403 404 dev = container_of(attr, struct kfd_topology_device, 405 attr_props); 406 sysfs_show_32bit_prop(buffer, "cpu_cores_count", 407 dev->node_props.cpu_cores_count); 408 sysfs_show_32bit_prop(buffer, "simd_count", 409 dev->node_props.simd_count); 410 sysfs_show_32bit_prop(buffer, "mem_banks_count", 411 dev->node_props.mem_banks_count); 412 sysfs_show_32bit_prop(buffer, "caches_count", 413 dev->node_props.caches_count); 414 sysfs_show_32bit_prop(buffer, "io_links_count", 415 dev->node_props.io_links_count); 416 sysfs_show_32bit_prop(buffer, "cpu_core_id_base", 417 dev->node_props.cpu_core_id_base); 418 sysfs_show_32bit_prop(buffer, "simd_id_base", 419 dev->node_props.simd_id_base); 420 sysfs_show_32bit_prop(buffer, "max_waves_per_simd", 421 dev->node_props.max_waves_per_simd); 422 sysfs_show_32bit_prop(buffer, "lds_size_in_kb", 423 dev->node_props.lds_size_in_kb); 424 sysfs_show_32bit_prop(buffer, "gds_size_in_kb", 425 dev->node_props.gds_size_in_kb); 426 sysfs_show_32bit_prop(buffer, "wave_front_size", 427 dev->node_props.wave_front_size); 428 sysfs_show_32bit_prop(buffer, "array_count", 429 dev->node_props.array_count); 430 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine", 431 dev->node_props.simd_arrays_per_engine); 432 sysfs_show_32bit_prop(buffer, "cu_per_simd_array", 433 dev->node_props.cu_per_simd_array); 434 sysfs_show_32bit_prop(buffer, "simd_per_cu", 435 dev->node_props.simd_per_cu); 436 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu", 437 dev->node_props.max_slots_scratch_cu); 438 sysfs_show_32bit_prop(buffer, "vendor_id", 439 dev->node_props.vendor_id); 440 sysfs_show_32bit_prop(buffer, "device_id", 441 dev->node_props.device_id); 442 sysfs_show_32bit_prop(buffer, "location_id", 443 dev->node_props.location_id); 444 445 if (dev->gpu) { 446 log_max_watch_addr = 447 __ilog2_u32(dev->gpu->device_info->num_of_watch_points); 448 449 if (log_max_watch_addr) { 450 dev->node_props.capability |= 451 HSA_CAP_WATCH_POINTS_SUPPORTED; 452 453 dev->node_props.capability |= 454 ((log_max_watch_addr << 455 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) & 456 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK); 457 } 458 459 if (dev->gpu->device_info->asic_family == CHIP_TONGA) 460 dev->node_props.capability |= 461 HSA_CAP_AQL_QUEUE_DOUBLE_MAP; 462 463 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute", 464 dev->node_props.max_engine_clk_fcompute); 465 466 sysfs_show_64bit_prop(buffer, "local_mem_size", 467 (unsigned long long int) 0); 468 469 sysfs_show_32bit_prop(buffer, "fw_version", 470 dev->gpu->kfd2kgd->get_fw_version( 471 dev->gpu->kgd, 472 KGD_ENGINE_MEC1)); 473 sysfs_show_32bit_prop(buffer, "capability", 474 dev->node_props.capability); 475 } 476 477 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute", 478 cpufreq_quick_get_max(0)/1000); 479 } 480 481 static const struct sysfs_ops node_ops = { 482 .show = node_show, 483 }; 484 485 static struct kobj_type node_type = { 486 .release = kfd_topology_kobj_release, 487 .sysfs_ops = &node_ops, 488 }; 489 490 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr) 491 { 492 sysfs_remove_file(kobj, attr); 493 kobject_del(kobj); 494 kobject_put(kobj); 495 } 496 497 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) 498 { 499 struct kfd_iolink_properties *iolink; 500 struct kfd_cache_properties *cache; 501 struct kfd_mem_properties *mem; 502 struct kfd_perf_properties *perf; 503 504 if (dev->kobj_iolink) { 505 list_for_each_entry(iolink, &dev->io_link_props, list) 506 if (iolink->kobj) { 507 kfd_remove_sysfs_file(iolink->kobj, 508 &iolink->attr); 509 iolink->kobj = NULL; 510 } 511 kobject_del(dev->kobj_iolink); 512 kobject_put(dev->kobj_iolink); 513 dev->kobj_iolink = NULL; 514 } 515 516 if (dev->kobj_cache) { 517 list_for_each_entry(cache, &dev->cache_props, list) 518 if (cache->kobj) { 519 kfd_remove_sysfs_file(cache->kobj, 520 &cache->attr); 521 cache->kobj = NULL; 522 } 523 kobject_del(dev->kobj_cache); 524 kobject_put(dev->kobj_cache); 525 dev->kobj_cache = NULL; 526 } 527 528 if (dev->kobj_mem) { 529 list_for_each_entry(mem, &dev->mem_props, list) 530 if (mem->kobj) { 531 kfd_remove_sysfs_file(mem->kobj, &mem->attr); 532 mem->kobj = NULL; 533 } 534 kobject_del(dev->kobj_mem); 535 kobject_put(dev->kobj_mem); 536 dev->kobj_mem = NULL; 537 } 538 539 if (dev->kobj_perf) { 540 list_for_each_entry(perf, &dev->perf_props, list) { 541 kfree(perf->attr_group); 542 perf->attr_group = NULL; 543 } 544 kobject_del(dev->kobj_perf); 545 kobject_put(dev->kobj_perf); 546 dev->kobj_perf = NULL; 547 } 548 549 if (dev->kobj_node) { 550 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid); 551 sysfs_remove_file(dev->kobj_node, &dev->attr_name); 552 sysfs_remove_file(dev->kobj_node, &dev->attr_props); 553 kobject_del(dev->kobj_node); 554 kobject_put(dev->kobj_node); 555 dev->kobj_node = NULL; 556 } 557 } 558 559 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, 560 uint32_t id) 561 { 562 struct kfd_iolink_properties *iolink; 563 struct kfd_cache_properties *cache; 564 struct kfd_mem_properties *mem; 565 struct kfd_perf_properties *perf; 566 int ret; 567 uint32_t i, num_attrs; 568 struct attribute **attrs; 569 570 if (WARN_ON(dev->kobj_node)) 571 return -EEXIST; 572 573 /* 574 * Creating the sysfs folders 575 */ 576 dev->kobj_node = kfd_alloc_struct(dev->kobj_node); 577 if (!dev->kobj_node) 578 return -ENOMEM; 579 580 ret = kobject_init_and_add(dev->kobj_node, &node_type, 581 sys_props.kobj_nodes, "%d", id); 582 if (ret < 0) 583 return ret; 584 585 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node); 586 if (!dev->kobj_mem) 587 return -ENOMEM; 588 589 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node); 590 if (!dev->kobj_cache) 591 return -ENOMEM; 592 593 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node); 594 if (!dev->kobj_iolink) 595 return -ENOMEM; 596 597 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node); 598 if (!dev->kobj_perf) 599 return -ENOMEM; 600 601 /* 602 * Creating sysfs files for node properties 603 */ 604 dev->attr_gpuid.name = "gpu_id"; 605 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE; 606 sysfs_attr_init(&dev->attr_gpuid); 607 dev->attr_name.name = "name"; 608 dev->attr_name.mode = KFD_SYSFS_FILE_MODE; 609 sysfs_attr_init(&dev->attr_name); 610 dev->attr_props.name = "properties"; 611 dev->attr_props.mode = KFD_SYSFS_FILE_MODE; 612 sysfs_attr_init(&dev->attr_props); 613 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid); 614 if (ret < 0) 615 return ret; 616 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name); 617 if (ret < 0) 618 return ret; 619 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props); 620 if (ret < 0) 621 return ret; 622 623 i = 0; 624 list_for_each_entry(mem, &dev->mem_props, list) { 625 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 626 if (!mem->kobj) 627 return -ENOMEM; 628 ret = kobject_init_and_add(mem->kobj, &mem_type, 629 dev->kobj_mem, "%d", i); 630 if (ret < 0) 631 return ret; 632 633 mem->attr.name = "properties"; 634 mem->attr.mode = KFD_SYSFS_FILE_MODE; 635 sysfs_attr_init(&mem->attr); 636 ret = sysfs_create_file(mem->kobj, &mem->attr); 637 if (ret < 0) 638 return ret; 639 i++; 640 } 641 642 i = 0; 643 list_for_each_entry(cache, &dev->cache_props, list) { 644 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 645 if (!cache->kobj) 646 return -ENOMEM; 647 ret = kobject_init_and_add(cache->kobj, &cache_type, 648 dev->kobj_cache, "%d", i); 649 if (ret < 0) 650 return ret; 651 652 cache->attr.name = "properties"; 653 cache->attr.mode = KFD_SYSFS_FILE_MODE; 654 sysfs_attr_init(&cache->attr); 655 ret = sysfs_create_file(cache->kobj, &cache->attr); 656 if (ret < 0) 657 return ret; 658 i++; 659 } 660 661 i = 0; 662 list_for_each_entry(iolink, &dev->io_link_props, list) { 663 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 664 if (!iolink->kobj) 665 return -ENOMEM; 666 ret = kobject_init_and_add(iolink->kobj, &iolink_type, 667 dev->kobj_iolink, "%d", i); 668 if (ret < 0) 669 return ret; 670 671 iolink->attr.name = "properties"; 672 iolink->attr.mode = KFD_SYSFS_FILE_MODE; 673 sysfs_attr_init(&iolink->attr); 674 ret = sysfs_create_file(iolink->kobj, &iolink->attr); 675 if (ret < 0) 676 return ret; 677 i++; 678 } 679 680 /* All hardware blocks have the same number of attributes. */ 681 num_attrs = ARRAY_SIZE(perf_attr_iommu); 682 list_for_each_entry(perf, &dev->perf_props, list) { 683 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr) 684 * num_attrs + sizeof(struct attribute_group), 685 GFP_KERNEL); 686 if (!perf->attr_group) 687 return -ENOMEM; 688 689 attrs = (struct attribute **)(perf->attr_group + 1); 690 if (!strcmp(perf->block_name, "iommu")) { 691 /* Information of IOMMU's num_counters and counter_ids is shown 692 * under /sys/bus/event_source/devices/amd_iommu. We don't 693 * duplicate here. 694 */ 695 perf_attr_iommu[0].data = perf->max_concurrent; 696 for (i = 0; i < num_attrs; i++) 697 attrs[i] = &perf_attr_iommu[i].attr.attr; 698 } 699 perf->attr_group->name = perf->block_name; 700 perf->attr_group->attrs = attrs; 701 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group); 702 if (ret < 0) 703 return ret; 704 } 705 706 return 0; 707 } 708 709 /* Called with write topology lock acquired */ 710 static int kfd_build_sysfs_node_tree(void) 711 { 712 struct kfd_topology_device *dev; 713 int ret; 714 uint32_t i = 0; 715 716 list_for_each_entry(dev, &topology_device_list, list) { 717 ret = kfd_build_sysfs_node_entry(dev, i); 718 if (ret < 0) 719 return ret; 720 i++; 721 } 722 723 return 0; 724 } 725 726 /* Called with write topology lock acquired */ 727 static void kfd_remove_sysfs_node_tree(void) 728 { 729 struct kfd_topology_device *dev; 730 731 list_for_each_entry(dev, &topology_device_list, list) 732 kfd_remove_sysfs_node_entry(dev); 733 } 734 735 static int kfd_topology_update_sysfs(void) 736 { 737 int ret; 738 739 pr_info("Creating topology SYSFS entries\n"); 740 if (!sys_props.kobj_topology) { 741 sys_props.kobj_topology = 742 kfd_alloc_struct(sys_props.kobj_topology); 743 if (!sys_props.kobj_topology) 744 return -ENOMEM; 745 746 ret = kobject_init_and_add(sys_props.kobj_topology, 747 &sysprops_type, &kfd_device->kobj, 748 "topology"); 749 if (ret < 0) 750 return ret; 751 752 sys_props.kobj_nodes = kobject_create_and_add("nodes", 753 sys_props.kobj_topology); 754 if (!sys_props.kobj_nodes) 755 return -ENOMEM; 756 757 sys_props.attr_genid.name = "generation_id"; 758 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE; 759 sysfs_attr_init(&sys_props.attr_genid); 760 ret = sysfs_create_file(sys_props.kobj_topology, 761 &sys_props.attr_genid); 762 if (ret < 0) 763 return ret; 764 765 sys_props.attr_props.name = "system_properties"; 766 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE; 767 sysfs_attr_init(&sys_props.attr_props); 768 ret = sysfs_create_file(sys_props.kobj_topology, 769 &sys_props.attr_props); 770 if (ret < 0) 771 return ret; 772 } 773 774 kfd_remove_sysfs_node_tree(); 775 776 return kfd_build_sysfs_node_tree(); 777 } 778 779 static void kfd_topology_release_sysfs(void) 780 { 781 kfd_remove_sysfs_node_tree(); 782 if (sys_props.kobj_topology) { 783 sysfs_remove_file(sys_props.kobj_topology, 784 &sys_props.attr_genid); 785 sysfs_remove_file(sys_props.kobj_topology, 786 &sys_props.attr_props); 787 if (sys_props.kobj_nodes) { 788 kobject_del(sys_props.kobj_nodes); 789 kobject_put(sys_props.kobj_nodes); 790 sys_props.kobj_nodes = NULL; 791 } 792 kobject_del(sys_props.kobj_topology); 793 kobject_put(sys_props.kobj_topology); 794 sys_props.kobj_topology = NULL; 795 } 796 } 797 798 /* Called with write topology_lock acquired */ 799 static void kfd_topology_update_device_list(struct list_head *temp_list, 800 struct list_head *master_list) 801 { 802 while (!list_empty(temp_list)) { 803 list_move_tail(temp_list->next, master_list); 804 sys_props.num_devices++; 805 } 806 } 807 808 static void kfd_debug_print_topology(void) 809 { 810 struct kfd_topology_device *dev; 811 812 down_read(&topology_lock); 813 814 dev = list_last_entry(&topology_device_list, 815 struct kfd_topology_device, list); 816 if (dev) { 817 if (dev->node_props.cpu_cores_count && 818 dev->node_props.simd_count) { 819 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n", 820 dev->node_props.device_id, 821 dev->node_props.vendor_id); 822 } else if (dev->node_props.cpu_cores_count) 823 pr_info("Topology: Add CPU node\n"); 824 else if (dev->node_props.simd_count) 825 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n", 826 dev->node_props.device_id, 827 dev->node_props.vendor_id); 828 } 829 up_read(&topology_lock); 830 } 831 832 /* Helper function for intializing platform_xx members of 833 * kfd_system_properties. Uses OEM info from the last CPU/APU node. 834 */ 835 static void kfd_update_system_properties(void) 836 { 837 struct kfd_topology_device *dev; 838 839 down_read(&topology_lock); 840 dev = list_last_entry(&topology_device_list, 841 struct kfd_topology_device, list); 842 if (dev) { 843 sys_props.platform_id = 844 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK; 845 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id); 846 sys_props.platform_rev = dev->oem_revision; 847 } 848 up_read(&topology_lock); 849 } 850 851 static void find_system_memory(const struct dmi_header *dm, 852 void *private) 853 { 854 struct kfd_mem_properties *mem; 855 u16 mem_width, mem_clock; 856 struct kfd_topology_device *kdev = 857 (struct kfd_topology_device *)private; 858 const u8 *dmi_data = (const u8 *)(dm + 1); 859 860 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) { 861 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6)); 862 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11)); 863 list_for_each_entry(mem, &kdev->mem_props, list) { 864 if (mem_width != 0xFFFF && mem_width != 0) 865 mem->width = mem_width; 866 if (mem_clock != 0) 867 mem->mem_clk_max = mem_clock; 868 } 869 } 870 } 871 872 /* 873 * Performance counters information is not part of CRAT but we would like to 874 * put them in the sysfs under topology directory for Thunk to get the data. 875 * This function is called before updating the sysfs. 876 */ 877 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev) 878 { 879 /* These are the only counters supported so far */ 880 return kfd_iommu_add_perf_counters(kdev); 881 } 882 883 /* kfd_add_non_crat_information - Add information that is not currently 884 * defined in CRAT but is necessary for KFD topology 885 * @dev - topology device to which addition info is added 886 */ 887 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev) 888 { 889 /* Check if CPU only node. */ 890 if (!kdev->gpu) { 891 /* Add system memory information */ 892 dmi_walk(find_system_memory, kdev); 893 } 894 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */ 895 } 896 897 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices. 898 * Ignore CRAT for all other devices. AMD APU is identified if both CPU 899 * and GPU cores are present. 900 * @device_list - topology device list created by parsing ACPI CRAT table. 901 * @return - TRUE if invalid, FALSE is valid. 902 */ 903 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list) 904 { 905 struct kfd_topology_device *dev; 906 907 list_for_each_entry(dev, device_list, list) { 908 if (dev->node_props.cpu_cores_count && 909 dev->node_props.simd_count) 910 return false; 911 } 912 pr_info("Ignoring ACPI CRAT on non-APU system\n"); 913 return true; 914 } 915 916 int kfd_topology_init(void) 917 { 918 void *crat_image = NULL; 919 size_t image_size = 0; 920 int ret; 921 struct list_head temp_topology_device_list; 922 int cpu_only_node = 0; 923 struct kfd_topology_device *kdev; 924 int proximity_domain; 925 926 /* topology_device_list - Master list of all topology devices 927 * temp_topology_device_list - temporary list created while parsing CRAT 928 * or VCRAT. Once parsing is complete the contents of list is moved to 929 * topology_device_list 930 */ 931 932 /* Initialize the head for the both the lists */ 933 INIT_LIST_HEAD(&topology_device_list); 934 INIT_LIST_HEAD(&temp_topology_device_list); 935 init_rwsem(&topology_lock); 936 937 memset(&sys_props, 0, sizeof(sys_props)); 938 939 /* Proximity domains in ACPI CRAT tables start counting at 940 * 0. The same should be true for virtual CRAT tables created 941 * at this stage. GPUs added later in kfd_topology_add_device 942 * use a counter. 943 */ 944 proximity_domain = 0; 945 946 /* 947 * Get the CRAT image from the ACPI. If ACPI doesn't have one 948 * or if ACPI CRAT is invalid create a virtual CRAT. 949 * NOTE: The current implementation expects all AMD APUs to have 950 * CRAT. If no CRAT is available, it is assumed to be a CPU 951 */ 952 ret = kfd_create_crat_image_acpi(&crat_image, &image_size); 953 if (!ret) { 954 ret = kfd_parse_crat_table(crat_image, 955 &temp_topology_device_list, 956 proximity_domain); 957 if (ret || 958 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) { 959 kfd_release_topology_device_list( 960 &temp_topology_device_list); 961 kfd_destroy_crat_image(crat_image); 962 crat_image = NULL; 963 } 964 } 965 966 if (!crat_image) { 967 ret = kfd_create_crat_image_virtual(&crat_image, &image_size, 968 COMPUTE_UNIT_CPU, NULL, 969 proximity_domain); 970 cpu_only_node = 1; 971 if (ret) { 972 pr_err("Error creating VCRAT table for CPU\n"); 973 return ret; 974 } 975 976 ret = kfd_parse_crat_table(crat_image, 977 &temp_topology_device_list, 978 proximity_domain); 979 if (ret) { 980 pr_err("Error parsing VCRAT table for CPU\n"); 981 goto err; 982 } 983 } 984 985 kdev = list_first_entry(&temp_topology_device_list, 986 struct kfd_topology_device, list); 987 kfd_add_perf_to_topology(kdev); 988 989 down_write(&topology_lock); 990 kfd_topology_update_device_list(&temp_topology_device_list, 991 &topology_device_list); 992 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1); 993 ret = kfd_topology_update_sysfs(); 994 up_write(&topology_lock); 995 996 if (!ret) { 997 sys_props.generation_count++; 998 kfd_update_system_properties(); 999 kfd_debug_print_topology(); 1000 pr_info("Finished initializing topology\n"); 1001 } else 1002 pr_err("Failed to update topology in sysfs ret=%d\n", ret); 1003 1004 /* For nodes with GPU, this information gets added 1005 * when GPU is detected (kfd_topology_add_device). 1006 */ 1007 if (cpu_only_node) { 1008 /* Add additional information to CPU only node created above */ 1009 down_write(&topology_lock); 1010 kdev = list_first_entry(&topology_device_list, 1011 struct kfd_topology_device, list); 1012 up_write(&topology_lock); 1013 kfd_add_non_crat_information(kdev); 1014 } 1015 1016 err: 1017 kfd_destroy_crat_image(crat_image); 1018 return ret; 1019 } 1020 1021 void kfd_topology_shutdown(void) 1022 { 1023 down_write(&topology_lock); 1024 kfd_topology_release_sysfs(); 1025 kfd_release_live_view(); 1026 up_write(&topology_lock); 1027 } 1028 1029 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) 1030 { 1031 uint32_t hashout; 1032 uint32_t buf[7]; 1033 uint64_t local_mem_size; 1034 int i; 1035 struct kfd_local_mem_info local_mem_info; 1036 1037 if (!gpu) 1038 return 0; 1039 1040 gpu->kfd2kgd->get_local_mem_info(gpu->kgd, &local_mem_info); 1041 1042 local_mem_size = local_mem_info.local_mem_size_private + 1043 local_mem_info.local_mem_size_public; 1044 1045 buf[0] = gpu->pdev->devfn; 1046 buf[1] = gpu->pdev->subsystem_vendor; 1047 buf[2] = gpu->pdev->subsystem_device; 1048 buf[3] = gpu->pdev->device; 1049 buf[4] = gpu->pdev->bus->number; 1050 buf[5] = lower_32_bits(local_mem_size); 1051 buf[6] = upper_32_bits(local_mem_size); 1052 1053 for (i = 0, hashout = 0; i < 7; i++) 1054 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); 1055 1056 return hashout; 1057 } 1058 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If 1059 * the GPU device is not already present in the topology device 1060 * list then return NULL. This means a new topology device has to 1061 * be created for this GPU. 1062 * TODO: Rather than assiging @gpu to first topology device withtout 1063 * gpu attached, it will better to have more stringent check. 1064 */ 1065 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) 1066 { 1067 struct kfd_topology_device *dev; 1068 struct kfd_topology_device *out_dev = NULL; 1069 1070 down_write(&topology_lock); 1071 list_for_each_entry(dev, &topology_device_list, list) 1072 if (!dev->gpu && (dev->node_props.simd_count > 0)) { 1073 dev->gpu = gpu; 1074 out_dev = dev; 1075 break; 1076 } 1077 up_write(&topology_lock); 1078 return out_dev; 1079 } 1080 1081 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival) 1082 { 1083 /* 1084 * TODO: Generate an event for thunk about the arrival/removal 1085 * of the GPU 1086 */ 1087 } 1088 1089 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info, 1090 * patch this after CRAT parsing. 1091 */ 1092 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) 1093 { 1094 struct kfd_mem_properties *mem; 1095 struct kfd_local_mem_info local_mem_info; 1096 1097 if (!dev) 1098 return; 1099 1100 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with 1101 * single bank of VRAM local memory. 1102 * for dGPUs - VCRAT reports only one bank of Local Memory 1103 * for APUs - If CRAT from ACPI reports more than one bank, then 1104 * all the banks will report the same mem_clk_max information 1105 */ 1106 dev->gpu->kfd2kgd->get_local_mem_info(dev->gpu->kgd, 1107 &local_mem_info); 1108 1109 list_for_each_entry(mem, &dev->mem_props, list) 1110 mem->mem_clk_max = local_mem_info.mem_clk_max; 1111 } 1112 1113 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev) 1114 { 1115 struct kfd_iolink_properties *link; 1116 1117 if (!dev || !dev->gpu) 1118 return; 1119 1120 /* GPU only creates direck links so apply flags setting to all */ 1121 if (dev->gpu->device_info->asic_family == CHIP_HAWAII) 1122 list_for_each_entry(link, &dev->io_link_props, list) 1123 link->flags = CRAT_IOLINK_FLAGS_ENABLED | 1124 CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | 1125 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; 1126 } 1127 1128 int kfd_topology_add_device(struct kfd_dev *gpu) 1129 { 1130 uint32_t gpu_id; 1131 struct kfd_topology_device *dev; 1132 struct kfd_cu_info cu_info; 1133 int res = 0; 1134 struct list_head temp_topology_device_list; 1135 void *crat_image = NULL; 1136 size_t image_size = 0; 1137 int proximity_domain; 1138 1139 INIT_LIST_HEAD(&temp_topology_device_list); 1140 1141 gpu_id = kfd_generate_gpu_id(gpu); 1142 1143 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id); 1144 1145 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain); 1146 1147 /* Check to see if this gpu device exists in the topology_device_list. 1148 * If so, assign the gpu to that device, 1149 * else create a Virtual CRAT for this gpu device and then parse that 1150 * CRAT to create a new topology device. Once created assign the gpu to 1151 * that topology device 1152 */ 1153 dev = kfd_assign_gpu(gpu); 1154 if (!dev) { 1155 res = kfd_create_crat_image_virtual(&crat_image, &image_size, 1156 COMPUTE_UNIT_GPU, gpu, 1157 proximity_domain); 1158 if (res) { 1159 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n", 1160 gpu_id); 1161 return res; 1162 } 1163 res = kfd_parse_crat_table(crat_image, 1164 &temp_topology_device_list, 1165 proximity_domain); 1166 if (res) { 1167 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n", 1168 gpu_id); 1169 goto err; 1170 } 1171 1172 down_write(&topology_lock); 1173 kfd_topology_update_device_list(&temp_topology_device_list, 1174 &topology_device_list); 1175 1176 /* Update the SYSFS tree, since we added another topology 1177 * device 1178 */ 1179 res = kfd_topology_update_sysfs(); 1180 up_write(&topology_lock); 1181 1182 if (!res) 1183 sys_props.generation_count++; 1184 else 1185 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n", 1186 gpu_id, res); 1187 dev = kfd_assign_gpu(gpu); 1188 if (WARN_ON(!dev)) { 1189 res = -ENODEV; 1190 goto err; 1191 } 1192 } 1193 1194 dev->gpu_id = gpu_id; 1195 gpu->id = gpu_id; 1196 1197 /* TODO: Move the following lines to function 1198 * kfd_add_non_crat_information 1199 */ 1200 1201 /* Fill-in additional information that is not available in CRAT but 1202 * needed for the topology 1203 */ 1204 1205 dev->gpu->kfd2kgd->get_cu_info(dev->gpu->kgd, &cu_info); 1206 dev->node_props.simd_arrays_per_engine = 1207 cu_info.num_shader_arrays_per_engine; 1208 1209 dev->node_props.vendor_id = gpu->pdev->vendor; 1210 dev->node_props.device_id = gpu->pdev->device; 1211 dev->node_props.location_id = PCI_DEVID(gpu->pdev->bus->number, 1212 gpu->pdev->devfn); 1213 dev->node_props.max_engine_clk_fcompute = 1214 dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(dev->gpu->kgd); 1215 dev->node_props.max_engine_clk_ccompute = 1216 cpufreq_quick_get_max(0) / 1000; 1217 1218 kfd_fill_mem_clk_max_info(dev); 1219 kfd_fill_iolink_non_crat_info(dev); 1220 1221 switch (dev->gpu->device_info->asic_family) { 1222 case CHIP_KAVERI: 1223 case CHIP_HAWAII: 1224 case CHIP_TONGA: 1225 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 << 1226 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1227 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1228 break; 1229 case CHIP_CARRIZO: 1230 case CHIP_FIJI: 1231 case CHIP_POLARIS10: 1232 case CHIP_POLARIS11: 1233 pr_debug("Adding doorbell packet type capability\n"); 1234 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 << 1235 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1236 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1237 break; 1238 default: 1239 WARN(1, "Unexpected ASIC family %u", 1240 dev->gpu->device_info->asic_family); 1241 } 1242 1243 /* Fix errors in CZ CRAT. 1244 * simd_count: Carrizo CRAT reports wrong simd_count, probably 1245 * because it doesn't consider masked out CUs 1246 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd 1247 * capability flag: Carrizo CRAT doesn't report IOMMU flags 1248 */ 1249 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) { 1250 dev->node_props.simd_count = 1251 cu_info.simd_per_cu * cu_info.cu_active_number; 1252 dev->node_props.max_waves_per_simd = 10; 1253 dev->node_props.capability |= HSA_CAP_ATS_PRESENT; 1254 } 1255 1256 kfd_debug_print_topology(); 1257 1258 if (!res) 1259 kfd_notify_gpu_change(gpu_id, 1); 1260 err: 1261 kfd_destroy_crat_image(crat_image); 1262 return res; 1263 } 1264 1265 int kfd_topology_remove_device(struct kfd_dev *gpu) 1266 { 1267 struct kfd_topology_device *dev, *tmp; 1268 uint32_t gpu_id; 1269 int res = -ENODEV; 1270 1271 down_write(&topology_lock); 1272 1273 list_for_each_entry_safe(dev, tmp, &topology_device_list, list) 1274 if (dev->gpu == gpu) { 1275 gpu_id = dev->gpu_id; 1276 kfd_remove_sysfs_node_entry(dev); 1277 kfd_release_topology_device(dev); 1278 sys_props.num_devices--; 1279 res = 0; 1280 if (kfd_topology_update_sysfs() < 0) 1281 kfd_topology_release_sysfs(); 1282 break; 1283 } 1284 1285 up_write(&topology_lock); 1286 1287 if (!res) 1288 kfd_notify_gpu_change(gpu_id, 0); 1289 1290 return res; 1291 } 1292 1293 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD 1294 * topology. If GPU device is found @idx, then valid kfd_dev pointer is 1295 * returned through @kdev 1296 * Return - 0: On success (@kdev will be NULL for non GPU nodes) 1297 * -1: If end of list 1298 */ 1299 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev) 1300 { 1301 1302 struct kfd_topology_device *top_dev; 1303 uint8_t device_idx = 0; 1304 1305 *kdev = NULL; 1306 down_read(&topology_lock); 1307 1308 list_for_each_entry(top_dev, &topology_device_list, list) { 1309 if (device_idx == idx) { 1310 *kdev = top_dev->gpu; 1311 up_read(&topology_lock); 1312 return 0; 1313 } 1314 1315 device_idx++; 1316 } 1317 1318 up_read(&topology_lock); 1319 1320 return -1; 1321 1322 } 1323 1324 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask) 1325 { 1326 const struct cpuinfo_x86 *cpuinfo; 1327 int first_cpu_of_numa_node; 1328 1329 if (!cpumask || cpumask == cpu_none_mask) 1330 return -1; 1331 first_cpu_of_numa_node = cpumask_first(cpumask); 1332 if (first_cpu_of_numa_node >= nr_cpu_ids) 1333 return -1; 1334 cpuinfo = &cpu_data(first_cpu_of_numa_node); 1335 1336 return cpuinfo->apicid; 1337 } 1338 1339 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor 1340 * of the given NUMA node (numa_node_id) 1341 * Return -1 on failure 1342 */ 1343 int kfd_numa_node_to_apic_id(int numa_node_id) 1344 { 1345 if (numa_node_id == -1) { 1346 pr_warn("Invalid NUMA Node. Use online CPU mask\n"); 1347 return kfd_cpumask_to_apic_id(cpu_online_mask); 1348 } 1349 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id)); 1350 } 1351 1352 #if defined(CONFIG_DEBUG_FS) 1353 1354 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data) 1355 { 1356 struct kfd_topology_device *dev; 1357 unsigned int i = 0; 1358 int r = 0; 1359 1360 down_read(&topology_lock); 1361 1362 list_for_each_entry(dev, &topology_device_list, list) { 1363 if (!dev->gpu) { 1364 i++; 1365 continue; 1366 } 1367 1368 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); 1369 r = dqm_debugfs_hqds(m, dev->gpu->dqm); 1370 if (r) 1371 break; 1372 } 1373 1374 up_read(&topology_lock); 1375 1376 return r; 1377 } 1378 1379 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data) 1380 { 1381 struct kfd_topology_device *dev; 1382 unsigned int i = 0; 1383 int r = 0; 1384 1385 down_read(&topology_lock); 1386 1387 list_for_each_entry(dev, &topology_device_list, list) { 1388 if (!dev->gpu) { 1389 i++; 1390 continue; 1391 } 1392 1393 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); 1394 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets); 1395 if (r) 1396 break; 1397 } 1398 1399 up_read(&topology_lock); 1400 1401 return r; 1402 } 1403 1404 #endif 1405