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 sysfs_show_32bit_prop(buffer, "drm_render_minor", 445 dev->node_props.drm_render_minor); 446 sysfs_show_64bit_prop(buffer, "hive_id", 447 dev->node_props.hive_id); 448 449 if (dev->gpu) { 450 log_max_watch_addr = 451 __ilog2_u32(dev->gpu->device_info->num_of_watch_points); 452 453 if (log_max_watch_addr) { 454 dev->node_props.capability |= 455 HSA_CAP_WATCH_POINTS_SUPPORTED; 456 457 dev->node_props.capability |= 458 ((log_max_watch_addr << 459 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) & 460 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK); 461 } 462 463 if (dev->gpu->device_info->asic_family == CHIP_TONGA) 464 dev->node_props.capability |= 465 HSA_CAP_AQL_QUEUE_DOUBLE_MAP; 466 467 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute", 468 dev->node_props.max_engine_clk_fcompute); 469 470 sysfs_show_64bit_prop(buffer, "local_mem_size", 471 (unsigned long long int) 0); 472 473 sysfs_show_32bit_prop(buffer, "fw_version", 474 dev->gpu->kfd2kgd->get_fw_version( 475 dev->gpu->kgd, 476 KGD_ENGINE_MEC1)); 477 sysfs_show_32bit_prop(buffer, "capability", 478 dev->node_props.capability); 479 } 480 481 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute", 482 cpufreq_quick_get_max(0)/1000); 483 } 484 485 static const struct sysfs_ops node_ops = { 486 .show = node_show, 487 }; 488 489 static struct kobj_type node_type = { 490 .release = kfd_topology_kobj_release, 491 .sysfs_ops = &node_ops, 492 }; 493 494 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr) 495 { 496 sysfs_remove_file(kobj, attr); 497 kobject_del(kobj); 498 kobject_put(kobj); 499 } 500 501 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) 502 { 503 struct kfd_iolink_properties *iolink; 504 struct kfd_cache_properties *cache; 505 struct kfd_mem_properties *mem; 506 struct kfd_perf_properties *perf; 507 508 if (dev->kobj_iolink) { 509 list_for_each_entry(iolink, &dev->io_link_props, list) 510 if (iolink->kobj) { 511 kfd_remove_sysfs_file(iolink->kobj, 512 &iolink->attr); 513 iolink->kobj = NULL; 514 } 515 kobject_del(dev->kobj_iolink); 516 kobject_put(dev->kobj_iolink); 517 dev->kobj_iolink = NULL; 518 } 519 520 if (dev->kobj_cache) { 521 list_for_each_entry(cache, &dev->cache_props, list) 522 if (cache->kobj) { 523 kfd_remove_sysfs_file(cache->kobj, 524 &cache->attr); 525 cache->kobj = NULL; 526 } 527 kobject_del(dev->kobj_cache); 528 kobject_put(dev->kobj_cache); 529 dev->kobj_cache = NULL; 530 } 531 532 if (dev->kobj_mem) { 533 list_for_each_entry(mem, &dev->mem_props, list) 534 if (mem->kobj) { 535 kfd_remove_sysfs_file(mem->kobj, &mem->attr); 536 mem->kobj = NULL; 537 } 538 kobject_del(dev->kobj_mem); 539 kobject_put(dev->kobj_mem); 540 dev->kobj_mem = NULL; 541 } 542 543 if (dev->kobj_perf) { 544 list_for_each_entry(perf, &dev->perf_props, list) { 545 kfree(perf->attr_group); 546 perf->attr_group = NULL; 547 } 548 kobject_del(dev->kobj_perf); 549 kobject_put(dev->kobj_perf); 550 dev->kobj_perf = NULL; 551 } 552 553 if (dev->kobj_node) { 554 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid); 555 sysfs_remove_file(dev->kobj_node, &dev->attr_name); 556 sysfs_remove_file(dev->kobj_node, &dev->attr_props); 557 kobject_del(dev->kobj_node); 558 kobject_put(dev->kobj_node); 559 dev->kobj_node = NULL; 560 } 561 } 562 563 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, 564 uint32_t id) 565 { 566 struct kfd_iolink_properties *iolink; 567 struct kfd_cache_properties *cache; 568 struct kfd_mem_properties *mem; 569 struct kfd_perf_properties *perf; 570 int ret; 571 uint32_t i, num_attrs; 572 struct attribute **attrs; 573 574 if (WARN_ON(dev->kobj_node)) 575 return -EEXIST; 576 577 /* 578 * Creating the sysfs folders 579 */ 580 dev->kobj_node = kfd_alloc_struct(dev->kobj_node); 581 if (!dev->kobj_node) 582 return -ENOMEM; 583 584 ret = kobject_init_and_add(dev->kobj_node, &node_type, 585 sys_props.kobj_nodes, "%d", id); 586 if (ret < 0) 587 return ret; 588 589 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node); 590 if (!dev->kobj_mem) 591 return -ENOMEM; 592 593 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node); 594 if (!dev->kobj_cache) 595 return -ENOMEM; 596 597 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node); 598 if (!dev->kobj_iolink) 599 return -ENOMEM; 600 601 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node); 602 if (!dev->kobj_perf) 603 return -ENOMEM; 604 605 /* 606 * Creating sysfs files for node properties 607 */ 608 dev->attr_gpuid.name = "gpu_id"; 609 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE; 610 sysfs_attr_init(&dev->attr_gpuid); 611 dev->attr_name.name = "name"; 612 dev->attr_name.mode = KFD_SYSFS_FILE_MODE; 613 sysfs_attr_init(&dev->attr_name); 614 dev->attr_props.name = "properties"; 615 dev->attr_props.mode = KFD_SYSFS_FILE_MODE; 616 sysfs_attr_init(&dev->attr_props); 617 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid); 618 if (ret < 0) 619 return ret; 620 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name); 621 if (ret < 0) 622 return ret; 623 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props); 624 if (ret < 0) 625 return ret; 626 627 i = 0; 628 list_for_each_entry(mem, &dev->mem_props, list) { 629 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 630 if (!mem->kobj) 631 return -ENOMEM; 632 ret = kobject_init_and_add(mem->kobj, &mem_type, 633 dev->kobj_mem, "%d", i); 634 if (ret < 0) 635 return ret; 636 637 mem->attr.name = "properties"; 638 mem->attr.mode = KFD_SYSFS_FILE_MODE; 639 sysfs_attr_init(&mem->attr); 640 ret = sysfs_create_file(mem->kobj, &mem->attr); 641 if (ret < 0) 642 return ret; 643 i++; 644 } 645 646 i = 0; 647 list_for_each_entry(cache, &dev->cache_props, list) { 648 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 649 if (!cache->kobj) 650 return -ENOMEM; 651 ret = kobject_init_and_add(cache->kobj, &cache_type, 652 dev->kobj_cache, "%d", i); 653 if (ret < 0) 654 return ret; 655 656 cache->attr.name = "properties"; 657 cache->attr.mode = KFD_SYSFS_FILE_MODE; 658 sysfs_attr_init(&cache->attr); 659 ret = sysfs_create_file(cache->kobj, &cache->attr); 660 if (ret < 0) 661 return ret; 662 i++; 663 } 664 665 i = 0; 666 list_for_each_entry(iolink, &dev->io_link_props, list) { 667 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 668 if (!iolink->kobj) 669 return -ENOMEM; 670 ret = kobject_init_and_add(iolink->kobj, &iolink_type, 671 dev->kobj_iolink, "%d", i); 672 if (ret < 0) 673 return ret; 674 675 iolink->attr.name = "properties"; 676 iolink->attr.mode = KFD_SYSFS_FILE_MODE; 677 sysfs_attr_init(&iolink->attr); 678 ret = sysfs_create_file(iolink->kobj, &iolink->attr); 679 if (ret < 0) 680 return ret; 681 i++; 682 } 683 684 /* All hardware blocks have the same number of attributes. */ 685 num_attrs = ARRAY_SIZE(perf_attr_iommu); 686 list_for_each_entry(perf, &dev->perf_props, list) { 687 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr) 688 * num_attrs + sizeof(struct attribute_group), 689 GFP_KERNEL); 690 if (!perf->attr_group) 691 return -ENOMEM; 692 693 attrs = (struct attribute **)(perf->attr_group + 1); 694 if (!strcmp(perf->block_name, "iommu")) { 695 /* Information of IOMMU's num_counters and counter_ids is shown 696 * under /sys/bus/event_source/devices/amd_iommu. We don't 697 * duplicate here. 698 */ 699 perf_attr_iommu[0].data = perf->max_concurrent; 700 for (i = 0; i < num_attrs; i++) 701 attrs[i] = &perf_attr_iommu[i].attr.attr; 702 } 703 perf->attr_group->name = perf->block_name; 704 perf->attr_group->attrs = attrs; 705 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group); 706 if (ret < 0) 707 return ret; 708 } 709 710 return 0; 711 } 712 713 /* Called with write topology lock acquired */ 714 static int kfd_build_sysfs_node_tree(void) 715 { 716 struct kfd_topology_device *dev; 717 int ret; 718 uint32_t i = 0; 719 720 list_for_each_entry(dev, &topology_device_list, list) { 721 ret = kfd_build_sysfs_node_entry(dev, i); 722 if (ret < 0) 723 return ret; 724 i++; 725 } 726 727 return 0; 728 } 729 730 /* Called with write topology lock acquired */ 731 static void kfd_remove_sysfs_node_tree(void) 732 { 733 struct kfd_topology_device *dev; 734 735 list_for_each_entry(dev, &topology_device_list, list) 736 kfd_remove_sysfs_node_entry(dev); 737 } 738 739 static int kfd_topology_update_sysfs(void) 740 { 741 int ret; 742 743 pr_info("Creating topology SYSFS entries\n"); 744 if (!sys_props.kobj_topology) { 745 sys_props.kobj_topology = 746 kfd_alloc_struct(sys_props.kobj_topology); 747 if (!sys_props.kobj_topology) 748 return -ENOMEM; 749 750 ret = kobject_init_and_add(sys_props.kobj_topology, 751 &sysprops_type, &kfd_device->kobj, 752 "topology"); 753 if (ret < 0) 754 return ret; 755 756 sys_props.kobj_nodes = kobject_create_and_add("nodes", 757 sys_props.kobj_topology); 758 if (!sys_props.kobj_nodes) 759 return -ENOMEM; 760 761 sys_props.attr_genid.name = "generation_id"; 762 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE; 763 sysfs_attr_init(&sys_props.attr_genid); 764 ret = sysfs_create_file(sys_props.kobj_topology, 765 &sys_props.attr_genid); 766 if (ret < 0) 767 return ret; 768 769 sys_props.attr_props.name = "system_properties"; 770 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE; 771 sysfs_attr_init(&sys_props.attr_props); 772 ret = sysfs_create_file(sys_props.kobj_topology, 773 &sys_props.attr_props); 774 if (ret < 0) 775 return ret; 776 } 777 778 kfd_remove_sysfs_node_tree(); 779 780 return kfd_build_sysfs_node_tree(); 781 } 782 783 static void kfd_topology_release_sysfs(void) 784 { 785 kfd_remove_sysfs_node_tree(); 786 if (sys_props.kobj_topology) { 787 sysfs_remove_file(sys_props.kobj_topology, 788 &sys_props.attr_genid); 789 sysfs_remove_file(sys_props.kobj_topology, 790 &sys_props.attr_props); 791 if (sys_props.kobj_nodes) { 792 kobject_del(sys_props.kobj_nodes); 793 kobject_put(sys_props.kobj_nodes); 794 sys_props.kobj_nodes = NULL; 795 } 796 kobject_del(sys_props.kobj_topology); 797 kobject_put(sys_props.kobj_topology); 798 sys_props.kobj_topology = NULL; 799 } 800 } 801 802 /* Called with write topology_lock acquired */ 803 static void kfd_topology_update_device_list(struct list_head *temp_list, 804 struct list_head *master_list) 805 { 806 while (!list_empty(temp_list)) { 807 list_move_tail(temp_list->next, master_list); 808 sys_props.num_devices++; 809 } 810 } 811 812 static void kfd_debug_print_topology(void) 813 { 814 struct kfd_topology_device *dev; 815 816 down_read(&topology_lock); 817 818 dev = list_last_entry(&topology_device_list, 819 struct kfd_topology_device, list); 820 if (dev) { 821 if (dev->node_props.cpu_cores_count && 822 dev->node_props.simd_count) { 823 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n", 824 dev->node_props.device_id, 825 dev->node_props.vendor_id); 826 } else if (dev->node_props.cpu_cores_count) 827 pr_info("Topology: Add CPU node\n"); 828 else if (dev->node_props.simd_count) 829 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n", 830 dev->node_props.device_id, 831 dev->node_props.vendor_id); 832 } 833 up_read(&topology_lock); 834 } 835 836 /* Helper function for intializing platform_xx members of 837 * kfd_system_properties. Uses OEM info from the last CPU/APU node. 838 */ 839 static void kfd_update_system_properties(void) 840 { 841 struct kfd_topology_device *dev; 842 843 down_read(&topology_lock); 844 dev = list_last_entry(&topology_device_list, 845 struct kfd_topology_device, list); 846 if (dev) { 847 sys_props.platform_id = 848 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK; 849 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id); 850 sys_props.platform_rev = dev->oem_revision; 851 } 852 up_read(&topology_lock); 853 } 854 855 static void find_system_memory(const struct dmi_header *dm, 856 void *private) 857 { 858 struct kfd_mem_properties *mem; 859 u16 mem_width, mem_clock; 860 struct kfd_topology_device *kdev = 861 (struct kfd_topology_device *)private; 862 const u8 *dmi_data = (const u8 *)(dm + 1); 863 864 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) { 865 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6)); 866 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11)); 867 list_for_each_entry(mem, &kdev->mem_props, list) { 868 if (mem_width != 0xFFFF && mem_width != 0) 869 mem->width = mem_width; 870 if (mem_clock != 0) 871 mem->mem_clk_max = mem_clock; 872 } 873 } 874 } 875 876 /* 877 * Performance counters information is not part of CRAT but we would like to 878 * put them in the sysfs under topology directory for Thunk to get the data. 879 * This function is called before updating the sysfs. 880 */ 881 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev) 882 { 883 /* These are the only counters supported so far */ 884 return kfd_iommu_add_perf_counters(kdev); 885 } 886 887 /* kfd_add_non_crat_information - Add information that is not currently 888 * defined in CRAT but is necessary for KFD topology 889 * @dev - topology device to which addition info is added 890 */ 891 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev) 892 { 893 /* Check if CPU only node. */ 894 if (!kdev->gpu) { 895 /* Add system memory information */ 896 dmi_walk(find_system_memory, kdev); 897 } 898 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */ 899 } 900 901 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices. 902 * Ignore CRAT for all other devices. AMD APU is identified if both CPU 903 * and GPU cores are present. 904 * @device_list - topology device list created by parsing ACPI CRAT table. 905 * @return - TRUE if invalid, FALSE is valid. 906 */ 907 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list) 908 { 909 struct kfd_topology_device *dev; 910 911 list_for_each_entry(dev, device_list, list) { 912 if (dev->node_props.cpu_cores_count && 913 dev->node_props.simd_count) 914 return false; 915 } 916 pr_info("Ignoring ACPI CRAT on non-APU system\n"); 917 return true; 918 } 919 920 int kfd_topology_init(void) 921 { 922 void *crat_image = NULL; 923 size_t image_size = 0; 924 int ret; 925 struct list_head temp_topology_device_list; 926 int cpu_only_node = 0; 927 struct kfd_topology_device *kdev; 928 int proximity_domain; 929 930 /* topology_device_list - Master list of all topology devices 931 * temp_topology_device_list - temporary list created while parsing CRAT 932 * or VCRAT. Once parsing is complete the contents of list is moved to 933 * topology_device_list 934 */ 935 936 /* Initialize the head for the both the lists */ 937 INIT_LIST_HEAD(&topology_device_list); 938 INIT_LIST_HEAD(&temp_topology_device_list); 939 init_rwsem(&topology_lock); 940 941 memset(&sys_props, 0, sizeof(sys_props)); 942 943 /* Proximity domains in ACPI CRAT tables start counting at 944 * 0. The same should be true for virtual CRAT tables created 945 * at this stage. GPUs added later in kfd_topology_add_device 946 * use a counter. 947 */ 948 proximity_domain = 0; 949 950 /* 951 * Get the CRAT image from the ACPI. If ACPI doesn't have one 952 * or if ACPI CRAT is invalid create a virtual CRAT. 953 * NOTE: The current implementation expects all AMD APUs to have 954 * CRAT. If no CRAT is available, it is assumed to be a CPU 955 */ 956 ret = kfd_create_crat_image_acpi(&crat_image, &image_size); 957 if (!ret) { 958 ret = kfd_parse_crat_table(crat_image, 959 &temp_topology_device_list, 960 proximity_domain); 961 if (ret || 962 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) { 963 kfd_release_topology_device_list( 964 &temp_topology_device_list); 965 kfd_destroy_crat_image(crat_image); 966 crat_image = NULL; 967 } 968 } 969 970 if (!crat_image) { 971 ret = kfd_create_crat_image_virtual(&crat_image, &image_size, 972 COMPUTE_UNIT_CPU, NULL, 973 proximity_domain); 974 cpu_only_node = 1; 975 if (ret) { 976 pr_err("Error creating VCRAT table for CPU\n"); 977 return ret; 978 } 979 980 ret = kfd_parse_crat_table(crat_image, 981 &temp_topology_device_list, 982 proximity_domain); 983 if (ret) { 984 pr_err("Error parsing VCRAT table for CPU\n"); 985 goto err; 986 } 987 } 988 989 kdev = list_first_entry(&temp_topology_device_list, 990 struct kfd_topology_device, list); 991 kfd_add_perf_to_topology(kdev); 992 993 down_write(&topology_lock); 994 kfd_topology_update_device_list(&temp_topology_device_list, 995 &topology_device_list); 996 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1); 997 ret = kfd_topology_update_sysfs(); 998 up_write(&topology_lock); 999 1000 if (!ret) { 1001 sys_props.generation_count++; 1002 kfd_update_system_properties(); 1003 kfd_debug_print_topology(); 1004 pr_info("Finished initializing topology\n"); 1005 } else 1006 pr_err("Failed to update topology in sysfs ret=%d\n", ret); 1007 1008 /* For nodes with GPU, this information gets added 1009 * when GPU is detected (kfd_topology_add_device). 1010 */ 1011 if (cpu_only_node) { 1012 /* Add additional information to CPU only node created above */ 1013 down_write(&topology_lock); 1014 kdev = list_first_entry(&topology_device_list, 1015 struct kfd_topology_device, list); 1016 up_write(&topology_lock); 1017 kfd_add_non_crat_information(kdev); 1018 } 1019 1020 err: 1021 kfd_destroy_crat_image(crat_image); 1022 return ret; 1023 } 1024 1025 void kfd_topology_shutdown(void) 1026 { 1027 down_write(&topology_lock); 1028 kfd_topology_release_sysfs(); 1029 kfd_release_live_view(); 1030 up_write(&topology_lock); 1031 } 1032 1033 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) 1034 { 1035 uint32_t hashout; 1036 uint32_t buf[7]; 1037 uint64_t local_mem_size; 1038 int i; 1039 struct kfd_local_mem_info local_mem_info; 1040 1041 if (!gpu) 1042 return 0; 1043 1044 gpu->kfd2kgd->get_local_mem_info(gpu->kgd, &local_mem_info); 1045 1046 local_mem_size = local_mem_info.local_mem_size_private + 1047 local_mem_info.local_mem_size_public; 1048 1049 buf[0] = gpu->pdev->devfn; 1050 buf[1] = gpu->pdev->subsystem_vendor; 1051 buf[2] = gpu->pdev->subsystem_device; 1052 buf[3] = gpu->pdev->device; 1053 buf[4] = gpu->pdev->bus->number; 1054 buf[5] = lower_32_bits(local_mem_size); 1055 buf[6] = upper_32_bits(local_mem_size); 1056 1057 for (i = 0, hashout = 0; i < 7; i++) 1058 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); 1059 1060 return hashout; 1061 } 1062 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If 1063 * the GPU device is not already present in the topology device 1064 * list then return NULL. This means a new topology device has to 1065 * be created for this GPU. 1066 * TODO: Rather than assiging @gpu to first topology device withtout 1067 * gpu attached, it will better to have more stringent check. 1068 */ 1069 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) 1070 { 1071 struct kfd_topology_device *dev; 1072 struct kfd_topology_device *out_dev = NULL; 1073 1074 down_write(&topology_lock); 1075 list_for_each_entry(dev, &topology_device_list, list) 1076 if (!dev->gpu && (dev->node_props.simd_count > 0)) { 1077 dev->gpu = gpu; 1078 out_dev = dev; 1079 break; 1080 } 1081 up_write(&topology_lock); 1082 return out_dev; 1083 } 1084 1085 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival) 1086 { 1087 /* 1088 * TODO: Generate an event for thunk about the arrival/removal 1089 * of the GPU 1090 */ 1091 } 1092 1093 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info, 1094 * patch this after CRAT parsing. 1095 */ 1096 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) 1097 { 1098 struct kfd_mem_properties *mem; 1099 struct kfd_local_mem_info local_mem_info; 1100 1101 if (!dev) 1102 return; 1103 1104 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with 1105 * single bank of VRAM local memory. 1106 * for dGPUs - VCRAT reports only one bank of Local Memory 1107 * for APUs - If CRAT from ACPI reports more than one bank, then 1108 * all the banks will report the same mem_clk_max information 1109 */ 1110 dev->gpu->kfd2kgd->get_local_mem_info(dev->gpu->kgd, 1111 &local_mem_info); 1112 1113 list_for_each_entry(mem, &dev->mem_props, list) 1114 mem->mem_clk_max = local_mem_info.mem_clk_max; 1115 } 1116 1117 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev) 1118 { 1119 struct kfd_iolink_properties *link; 1120 1121 if (!dev || !dev->gpu) 1122 return; 1123 1124 /* GPU only creates direck links so apply flags setting to all */ 1125 if (dev->gpu->device_info->asic_family == CHIP_HAWAII) 1126 list_for_each_entry(link, &dev->io_link_props, list) 1127 link->flags = CRAT_IOLINK_FLAGS_ENABLED | 1128 CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | 1129 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; 1130 } 1131 1132 int kfd_topology_add_device(struct kfd_dev *gpu) 1133 { 1134 uint32_t gpu_id; 1135 struct kfd_topology_device *dev; 1136 struct kfd_cu_info cu_info; 1137 int res = 0; 1138 struct list_head temp_topology_device_list; 1139 void *crat_image = NULL; 1140 size_t image_size = 0; 1141 int proximity_domain; 1142 1143 INIT_LIST_HEAD(&temp_topology_device_list); 1144 1145 gpu_id = kfd_generate_gpu_id(gpu); 1146 1147 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id); 1148 1149 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain); 1150 1151 /* Check to see if this gpu device exists in the topology_device_list. 1152 * If so, assign the gpu to that device, 1153 * else create a Virtual CRAT for this gpu device and then parse that 1154 * CRAT to create a new topology device. Once created assign the gpu to 1155 * that topology device 1156 */ 1157 dev = kfd_assign_gpu(gpu); 1158 if (!dev) { 1159 res = kfd_create_crat_image_virtual(&crat_image, &image_size, 1160 COMPUTE_UNIT_GPU, gpu, 1161 proximity_domain); 1162 if (res) { 1163 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n", 1164 gpu_id); 1165 return res; 1166 } 1167 res = kfd_parse_crat_table(crat_image, 1168 &temp_topology_device_list, 1169 proximity_domain); 1170 if (res) { 1171 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n", 1172 gpu_id); 1173 goto err; 1174 } 1175 1176 down_write(&topology_lock); 1177 kfd_topology_update_device_list(&temp_topology_device_list, 1178 &topology_device_list); 1179 1180 /* Update the SYSFS tree, since we added another topology 1181 * device 1182 */ 1183 res = kfd_topology_update_sysfs(); 1184 up_write(&topology_lock); 1185 1186 if (!res) 1187 sys_props.generation_count++; 1188 else 1189 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n", 1190 gpu_id, res); 1191 dev = kfd_assign_gpu(gpu); 1192 if (WARN_ON(!dev)) { 1193 res = -ENODEV; 1194 goto err; 1195 } 1196 } 1197 1198 dev->gpu_id = gpu_id; 1199 gpu->id = gpu_id; 1200 1201 /* TODO: Move the following lines to function 1202 * kfd_add_non_crat_information 1203 */ 1204 1205 /* Fill-in additional information that is not available in CRAT but 1206 * needed for the topology 1207 */ 1208 1209 dev->gpu->kfd2kgd->get_cu_info(dev->gpu->kgd, &cu_info); 1210 dev->node_props.simd_arrays_per_engine = 1211 cu_info.num_shader_arrays_per_engine; 1212 1213 dev->node_props.vendor_id = gpu->pdev->vendor; 1214 dev->node_props.device_id = gpu->pdev->device; 1215 dev->node_props.location_id = PCI_DEVID(gpu->pdev->bus->number, 1216 gpu->pdev->devfn); 1217 dev->node_props.max_engine_clk_fcompute = 1218 dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(dev->gpu->kgd); 1219 dev->node_props.max_engine_clk_ccompute = 1220 cpufreq_quick_get_max(0) / 1000; 1221 dev->node_props.drm_render_minor = 1222 gpu->shared_resources.drm_render_minor; 1223 1224 dev->node_props.hive_id = gpu->hive_id; 1225 1226 kfd_fill_mem_clk_max_info(dev); 1227 kfd_fill_iolink_non_crat_info(dev); 1228 1229 switch (dev->gpu->device_info->asic_family) { 1230 case CHIP_KAVERI: 1231 case CHIP_HAWAII: 1232 case CHIP_TONGA: 1233 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 << 1234 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1235 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1236 break; 1237 case CHIP_CARRIZO: 1238 case CHIP_FIJI: 1239 case CHIP_POLARIS10: 1240 case CHIP_POLARIS11: 1241 pr_debug("Adding doorbell packet type capability\n"); 1242 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 << 1243 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1244 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1245 break; 1246 case CHIP_VEGA10: 1247 case CHIP_RAVEN: 1248 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 << 1249 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1250 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1251 break; 1252 default: 1253 WARN(1, "Unexpected ASIC family %u", 1254 dev->gpu->device_info->asic_family); 1255 } 1256 1257 /* Fix errors in CZ CRAT. 1258 * simd_count: Carrizo CRAT reports wrong simd_count, probably 1259 * because it doesn't consider masked out CUs 1260 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd 1261 * capability flag: Carrizo CRAT doesn't report IOMMU flags 1262 */ 1263 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) { 1264 dev->node_props.simd_count = 1265 cu_info.simd_per_cu * cu_info.cu_active_number; 1266 dev->node_props.max_waves_per_simd = 10; 1267 dev->node_props.capability |= HSA_CAP_ATS_PRESENT; 1268 } 1269 1270 kfd_debug_print_topology(); 1271 1272 if (!res) 1273 kfd_notify_gpu_change(gpu_id, 1); 1274 err: 1275 kfd_destroy_crat_image(crat_image); 1276 return res; 1277 } 1278 1279 int kfd_topology_remove_device(struct kfd_dev *gpu) 1280 { 1281 struct kfd_topology_device *dev, *tmp; 1282 uint32_t gpu_id; 1283 int res = -ENODEV; 1284 1285 down_write(&topology_lock); 1286 1287 list_for_each_entry_safe(dev, tmp, &topology_device_list, list) 1288 if (dev->gpu == gpu) { 1289 gpu_id = dev->gpu_id; 1290 kfd_remove_sysfs_node_entry(dev); 1291 kfd_release_topology_device(dev); 1292 sys_props.num_devices--; 1293 res = 0; 1294 if (kfd_topology_update_sysfs() < 0) 1295 kfd_topology_release_sysfs(); 1296 break; 1297 } 1298 1299 up_write(&topology_lock); 1300 1301 if (!res) 1302 kfd_notify_gpu_change(gpu_id, 0); 1303 1304 return res; 1305 } 1306 1307 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD 1308 * topology. If GPU device is found @idx, then valid kfd_dev pointer is 1309 * returned through @kdev 1310 * Return - 0: On success (@kdev will be NULL for non GPU nodes) 1311 * -1: If end of list 1312 */ 1313 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev) 1314 { 1315 1316 struct kfd_topology_device *top_dev; 1317 uint8_t device_idx = 0; 1318 1319 *kdev = NULL; 1320 down_read(&topology_lock); 1321 1322 list_for_each_entry(top_dev, &topology_device_list, list) { 1323 if (device_idx == idx) { 1324 *kdev = top_dev->gpu; 1325 up_read(&topology_lock); 1326 return 0; 1327 } 1328 1329 device_idx++; 1330 } 1331 1332 up_read(&topology_lock); 1333 1334 return -1; 1335 1336 } 1337 1338 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask) 1339 { 1340 const struct cpuinfo_x86 *cpuinfo; 1341 int first_cpu_of_numa_node; 1342 1343 if (!cpumask || cpumask == cpu_none_mask) 1344 return -1; 1345 first_cpu_of_numa_node = cpumask_first(cpumask); 1346 if (first_cpu_of_numa_node >= nr_cpu_ids) 1347 return -1; 1348 cpuinfo = &cpu_data(first_cpu_of_numa_node); 1349 1350 return cpuinfo->apicid; 1351 } 1352 1353 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor 1354 * of the given NUMA node (numa_node_id) 1355 * Return -1 on failure 1356 */ 1357 int kfd_numa_node_to_apic_id(int numa_node_id) 1358 { 1359 if (numa_node_id == -1) { 1360 pr_warn("Invalid NUMA Node. Use online CPU mask\n"); 1361 return kfd_cpumask_to_apic_id(cpu_online_mask); 1362 } 1363 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id)); 1364 } 1365 1366 #if defined(CONFIG_DEBUG_FS) 1367 1368 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data) 1369 { 1370 struct kfd_topology_device *dev; 1371 unsigned int i = 0; 1372 int r = 0; 1373 1374 down_read(&topology_lock); 1375 1376 list_for_each_entry(dev, &topology_device_list, list) { 1377 if (!dev->gpu) { 1378 i++; 1379 continue; 1380 } 1381 1382 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); 1383 r = dqm_debugfs_hqds(m, dev->gpu->dqm); 1384 if (r) 1385 break; 1386 } 1387 1388 up_read(&topology_lock); 1389 1390 return r; 1391 } 1392 1393 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data) 1394 { 1395 struct kfd_topology_device *dev; 1396 unsigned int i = 0; 1397 int r = 0; 1398 1399 down_read(&topology_lock); 1400 1401 list_for_each_entry(dev, &topology_device_list, list) { 1402 if (!dev->gpu) { 1403 i++; 1404 continue; 1405 } 1406 1407 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); 1408 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets); 1409 if (r) 1410 break; 1411 } 1412 1413 up_read(&topology_lock); 1414 1415 return r; 1416 } 1417 1418 #endif 1419