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 #include "amdgpu_amdkfd.h" 40 #include "amdgpu_ras.h" 41 42 /* topology_device_list - Master list of all topology devices */ 43 static struct list_head topology_device_list; 44 static struct kfd_system_properties sys_props; 45 46 static DECLARE_RWSEM(topology_lock); 47 static atomic_t topology_crat_proximity_domain; 48 49 struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 50 uint32_t proximity_domain) 51 { 52 struct kfd_topology_device *top_dev; 53 struct kfd_topology_device *device = NULL; 54 55 down_read(&topology_lock); 56 57 list_for_each_entry(top_dev, &topology_device_list, list) 58 if (top_dev->proximity_domain == proximity_domain) { 59 device = top_dev; 60 break; 61 } 62 63 up_read(&topology_lock); 64 65 return device; 66 } 67 68 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id) 69 { 70 struct kfd_topology_device *top_dev = NULL; 71 struct kfd_topology_device *ret = NULL; 72 73 down_read(&topology_lock); 74 75 list_for_each_entry(top_dev, &topology_device_list, list) 76 if (top_dev->gpu_id == gpu_id) { 77 ret = top_dev; 78 break; 79 } 80 81 up_read(&topology_lock); 82 83 return ret; 84 } 85 86 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id) 87 { 88 struct kfd_topology_device *top_dev; 89 90 top_dev = kfd_topology_device_by_id(gpu_id); 91 if (!top_dev) 92 return NULL; 93 94 return top_dev->gpu; 95 } 96 97 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev) 98 { 99 struct kfd_topology_device *top_dev; 100 struct kfd_dev *device = NULL; 101 102 down_read(&topology_lock); 103 104 list_for_each_entry(top_dev, &topology_device_list, list) 105 if (top_dev->gpu && top_dev->gpu->pdev == pdev) { 106 device = top_dev->gpu; 107 break; 108 } 109 110 up_read(&topology_lock); 111 112 return device; 113 } 114 115 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd) 116 { 117 struct kfd_topology_device *top_dev; 118 struct kfd_dev *device = NULL; 119 120 down_read(&topology_lock); 121 122 list_for_each_entry(top_dev, &topology_device_list, list) 123 if (top_dev->gpu && top_dev->gpu->kgd == kgd) { 124 device = top_dev->gpu; 125 break; 126 } 127 128 up_read(&topology_lock); 129 130 return device; 131 } 132 133 /* Called with write topology_lock acquired */ 134 static void kfd_release_topology_device(struct kfd_topology_device *dev) 135 { 136 struct kfd_mem_properties *mem; 137 struct kfd_cache_properties *cache; 138 struct kfd_iolink_properties *iolink; 139 struct kfd_perf_properties *perf; 140 141 list_del(&dev->list); 142 143 while (dev->mem_props.next != &dev->mem_props) { 144 mem = container_of(dev->mem_props.next, 145 struct kfd_mem_properties, list); 146 list_del(&mem->list); 147 kfree(mem); 148 } 149 150 while (dev->cache_props.next != &dev->cache_props) { 151 cache = container_of(dev->cache_props.next, 152 struct kfd_cache_properties, list); 153 list_del(&cache->list); 154 kfree(cache); 155 } 156 157 while (dev->io_link_props.next != &dev->io_link_props) { 158 iolink = container_of(dev->io_link_props.next, 159 struct kfd_iolink_properties, list); 160 list_del(&iolink->list); 161 kfree(iolink); 162 } 163 164 while (dev->perf_props.next != &dev->perf_props) { 165 perf = container_of(dev->perf_props.next, 166 struct kfd_perf_properties, list); 167 list_del(&perf->list); 168 kfree(perf); 169 } 170 171 kfree(dev); 172 } 173 174 void kfd_release_topology_device_list(struct list_head *device_list) 175 { 176 struct kfd_topology_device *dev; 177 178 while (!list_empty(device_list)) { 179 dev = list_first_entry(device_list, 180 struct kfd_topology_device, list); 181 kfd_release_topology_device(dev); 182 } 183 } 184 185 static void kfd_release_live_view(void) 186 { 187 kfd_release_topology_device_list(&topology_device_list); 188 memset(&sys_props, 0, sizeof(sys_props)); 189 } 190 191 struct kfd_topology_device *kfd_create_topology_device( 192 struct list_head *device_list) 193 { 194 struct kfd_topology_device *dev; 195 196 dev = kfd_alloc_struct(dev); 197 if (!dev) { 198 pr_err("No memory to allocate a topology device"); 199 return NULL; 200 } 201 202 INIT_LIST_HEAD(&dev->mem_props); 203 INIT_LIST_HEAD(&dev->cache_props); 204 INIT_LIST_HEAD(&dev->io_link_props); 205 INIT_LIST_HEAD(&dev->perf_props); 206 207 list_add_tail(&dev->list, device_list); 208 209 return dev; 210 } 211 212 213 #define sysfs_show_gen_prop(buffer, fmt, ...) \ 214 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__) 215 #define sysfs_show_32bit_prop(buffer, name, value) \ 216 sysfs_show_gen_prop(buffer, "%s %u\n", name, value) 217 #define sysfs_show_64bit_prop(buffer, name, value) \ 218 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value) 219 #define sysfs_show_32bit_val(buffer, value) \ 220 sysfs_show_gen_prop(buffer, "%u\n", value) 221 #define sysfs_show_str_val(buffer, value) \ 222 sysfs_show_gen_prop(buffer, "%s\n", value) 223 224 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr, 225 char *buffer) 226 { 227 ssize_t ret; 228 229 /* Making sure that the buffer is an empty string */ 230 buffer[0] = 0; 231 232 if (attr == &sys_props.attr_genid) { 233 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count); 234 } else if (attr == &sys_props.attr_props) { 235 sysfs_show_64bit_prop(buffer, "platform_oem", 236 sys_props.platform_oem); 237 sysfs_show_64bit_prop(buffer, "platform_id", 238 sys_props.platform_id); 239 ret = sysfs_show_64bit_prop(buffer, "platform_rev", 240 sys_props.platform_rev); 241 } else { 242 ret = -EINVAL; 243 } 244 245 return ret; 246 } 247 248 static void kfd_topology_kobj_release(struct kobject *kobj) 249 { 250 kfree(kobj); 251 } 252 253 static const struct sysfs_ops sysprops_ops = { 254 .show = sysprops_show, 255 }; 256 257 static struct kobj_type sysprops_type = { 258 .release = kfd_topology_kobj_release, 259 .sysfs_ops = &sysprops_ops, 260 }; 261 262 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr, 263 char *buffer) 264 { 265 ssize_t ret; 266 struct kfd_iolink_properties *iolink; 267 268 /* Making sure that the buffer is an empty string */ 269 buffer[0] = 0; 270 271 iolink = container_of(attr, struct kfd_iolink_properties, attr); 272 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type); 273 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj); 274 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min); 275 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from); 276 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to); 277 sysfs_show_32bit_prop(buffer, "weight", iolink->weight); 278 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency); 279 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency); 280 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth); 281 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth); 282 sysfs_show_32bit_prop(buffer, "recommended_transfer_size", 283 iolink->rec_transfer_size); 284 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags); 285 286 return ret; 287 } 288 289 static const struct sysfs_ops iolink_ops = { 290 .show = iolink_show, 291 }; 292 293 static struct kobj_type iolink_type = { 294 .release = kfd_topology_kobj_release, 295 .sysfs_ops = &iolink_ops, 296 }; 297 298 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr, 299 char *buffer) 300 { 301 ssize_t ret; 302 struct kfd_mem_properties *mem; 303 304 /* Making sure that the buffer is an empty string */ 305 buffer[0] = 0; 306 307 mem = container_of(attr, struct kfd_mem_properties, attr); 308 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type); 309 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes); 310 sysfs_show_32bit_prop(buffer, "flags", mem->flags); 311 sysfs_show_32bit_prop(buffer, "width", mem->width); 312 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max); 313 314 return ret; 315 } 316 317 static const struct sysfs_ops mem_ops = { 318 .show = mem_show, 319 }; 320 321 static struct kobj_type mem_type = { 322 .release = kfd_topology_kobj_release, 323 .sysfs_ops = &mem_ops, 324 }; 325 326 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr, 327 char *buffer) 328 { 329 ssize_t ret; 330 uint32_t i, j; 331 struct kfd_cache_properties *cache; 332 333 /* Making sure that the buffer is an empty string */ 334 buffer[0] = 0; 335 336 cache = container_of(attr, struct kfd_cache_properties, attr); 337 sysfs_show_32bit_prop(buffer, "processor_id_low", 338 cache->processor_id_low); 339 sysfs_show_32bit_prop(buffer, "level", cache->cache_level); 340 sysfs_show_32bit_prop(buffer, "size", cache->cache_size); 341 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size); 342 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag", 343 cache->cachelines_per_tag); 344 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc); 345 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency); 346 sysfs_show_32bit_prop(buffer, "type", cache->cache_type); 347 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer); 348 for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++) 349 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) { 350 /* Check each bit */ 351 if (cache->sibling_map[i] & (1 << j)) 352 ret = snprintf(buffer, PAGE_SIZE, 353 "%s%d%s", buffer, 1, ","); 354 else 355 ret = snprintf(buffer, PAGE_SIZE, 356 "%s%d%s", buffer, 0, ","); 357 } 358 /* Replace the last "," with end of line */ 359 *(buffer + strlen(buffer) - 1) = 0xA; 360 return ret; 361 } 362 363 static const struct sysfs_ops cache_ops = { 364 .show = kfd_cache_show, 365 }; 366 367 static struct kobj_type cache_type = { 368 .release = kfd_topology_kobj_release, 369 .sysfs_ops = &cache_ops, 370 }; 371 372 /****** Sysfs of Performance Counters ******/ 373 374 struct kfd_perf_attr { 375 struct kobj_attribute attr; 376 uint32_t data; 377 }; 378 379 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs, 380 char *buf) 381 { 382 struct kfd_perf_attr *attr; 383 384 buf[0] = 0; 385 attr = container_of(attrs, struct kfd_perf_attr, attr); 386 if (!attr->data) /* invalid data for PMC */ 387 return 0; 388 else 389 return sysfs_show_32bit_val(buf, attr->data); 390 } 391 392 #define KFD_PERF_DESC(_name, _data) \ 393 { \ 394 .attr = __ATTR(_name, 0444, perf_show, NULL), \ 395 .data = _data, \ 396 } 397 398 static struct kfd_perf_attr perf_attr_iommu[] = { 399 KFD_PERF_DESC(max_concurrent, 0), 400 KFD_PERF_DESC(num_counters, 0), 401 KFD_PERF_DESC(counter_ids, 0), 402 }; 403 /****************************************/ 404 405 static ssize_t node_show(struct kobject *kobj, struct attribute *attr, 406 char *buffer) 407 { 408 struct kfd_topology_device *dev; 409 char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE]; 410 uint32_t i; 411 uint32_t log_max_watch_addr; 412 413 /* Making sure that the buffer is an empty string */ 414 buffer[0] = 0; 415 416 if (strcmp(attr->name, "gpu_id") == 0) { 417 dev = container_of(attr, struct kfd_topology_device, 418 attr_gpuid); 419 return sysfs_show_32bit_val(buffer, dev->gpu_id); 420 } 421 422 if (strcmp(attr->name, "name") == 0) { 423 dev = container_of(attr, struct kfd_topology_device, 424 attr_name); 425 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) { 426 public_name[i] = 427 (char)dev->node_props.marketing_name[i]; 428 if (dev->node_props.marketing_name[i] == 0) 429 break; 430 } 431 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0; 432 return sysfs_show_str_val(buffer, public_name); 433 } 434 435 dev = container_of(attr, struct kfd_topology_device, 436 attr_props); 437 sysfs_show_32bit_prop(buffer, "cpu_cores_count", 438 dev->node_props.cpu_cores_count); 439 sysfs_show_32bit_prop(buffer, "simd_count", 440 dev->node_props.simd_count); 441 sysfs_show_32bit_prop(buffer, "mem_banks_count", 442 dev->node_props.mem_banks_count); 443 sysfs_show_32bit_prop(buffer, "caches_count", 444 dev->node_props.caches_count); 445 sysfs_show_32bit_prop(buffer, "io_links_count", 446 dev->node_props.io_links_count); 447 sysfs_show_32bit_prop(buffer, "cpu_core_id_base", 448 dev->node_props.cpu_core_id_base); 449 sysfs_show_32bit_prop(buffer, "simd_id_base", 450 dev->node_props.simd_id_base); 451 sysfs_show_32bit_prop(buffer, "max_waves_per_simd", 452 dev->node_props.max_waves_per_simd); 453 sysfs_show_32bit_prop(buffer, "lds_size_in_kb", 454 dev->node_props.lds_size_in_kb); 455 sysfs_show_32bit_prop(buffer, "gds_size_in_kb", 456 dev->node_props.gds_size_in_kb); 457 sysfs_show_32bit_prop(buffer, "wave_front_size", 458 dev->node_props.wave_front_size); 459 sysfs_show_32bit_prop(buffer, "array_count", 460 dev->node_props.array_count); 461 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine", 462 dev->node_props.simd_arrays_per_engine); 463 sysfs_show_32bit_prop(buffer, "cu_per_simd_array", 464 dev->node_props.cu_per_simd_array); 465 sysfs_show_32bit_prop(buffer, "simd_per_cu", 466 dev->node_props.simd_per_cu); 467 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu", 468 dev->node_props.max_slots_scratch_cu); 469 sysfs_show_32bit_prop(buffer, "vendor_id", 470 dev->node_props.vendor_id); 471 sysfs_show_32bit_prop(buffer, "device_id", 472 dev->node_props.device_id); 473 sysfs_show_32bit_prop(buffer, "location_id", 474 dev->node_props.location_id); 475 sysfs_show_32bit_prop(buffer, "drm_render_minor", 476 dev->node_props.drm_render_minor); 477 sysfs_show_64bit_prop(buffer, "hive_id", 478 dev->node_props.hive_id); 479 480 if (dev->gpu) { 481 log_max_watch_addr = 482 __ilog2_u32(dev->gpu->device_info->num_of_watch_points); 483 484 if (log_max_watch_addr) { 485 dev->node_props.capability |= 486 HSA_CAP_WATCH_POINTS_SUPPORTED; 487 488 dev->node_props.capability |= 489 ((log_max_watch_addr << 490 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) & 491 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK); 492 } 493 494 if (dev->gpu->device_info->asic_family == CHIP_TONGA) 495 dev->node_props.capability |= 496 HSA_CAP_AQL_QUEUE_DOUBLE_MAP; 497 498 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute", 499 dev->node_props.max_engine_clk_fcompute); 500 501 sysfs_show_64bit_prop(buffer, "local_mem_size", 502 (unsigned long long int) 0); 503 504 sysfs_show_32bit_prop(buffer, "fw_version", 505 dev->gpu->mec_fw_version); 506 sysfs_show_32bit_prop(buffer, "capability", 507 dev->node_props.capability); 508 sysfs_show_32bit_prop(buffer, "sdma_fw_version", 509 dev->gpu->sdma_fw_version); 510 } 511 512 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute", 513 cpufreq_quick_get_max(0)/1000); 514 } 515 516 static const struct sysfs_ops node_ops = { 517 .show = node_show, 518 }; 519 520 static struct kobj_type node_type = { 521 .release = kfd_topology_kobj_release, 522 .sysfs_ops = &node_ops, 523 }; 524 525 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr) 526 { 527 sysfs_remove_file(kobj, attr); 528 kobject_del(kobj); 529 kobject_put(kobj); 530 } 531 532 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) 533 { 534 struct kfd_iolink_properties *iolink; 535 struct kfd_cache_properties *cache; 536 struct kfd_mem_properties *mem; 537 struct kfd_perf_properties *perf; 538 539 if (dev->kobj_iolink) { 540 list_for_each_entry(iolink, &dev->io_link_props, list) 541 if (iolink->kobj) { 542 kfd_remove_sysfs_file(iolink->kobj, 543 &iolink->attr); 544 iolink->kobj = NULL; 545 } 546 kobject_del(dev->kobj_iolink); 547 kobject_put(dev->kobj_iolink); 548 dev->kobj_iolink = NULL; 549 } 550 551 if (dev->kobj_cache) { 552 list_for_each_entry(cache, &dev->cache_props, list) 553 if (cache->kobj) { 554 kfd_remove_sysfs_file(cache->kobj, 555 &cache->attr); 556 cache->kobj = NULL; 557 } 558 kobject_del(dev->kobj_cache); 559 kobject_put(dev->kobj_cache); 560 dev->kobj_cache = NULL; 561 } 562 563 if (dev->kobj_mem) { 564 list_for_each_entry(mem, &dev->mem_props, list) 565 if (mem->kobj) { 566 kfd_remove_sysfs_file(mem->kobj, &mem->attr); 567 mem->kobj = NULL; 568 } 569 kobject_del(dev->kobj_mem); 570 kobject_put(dev->kobj_mem); 571 dev->kobj_mem = NULL; 572 } 573 574 if (dev->kobj_perf) { 575 list_for_each_entry(perf, &dev->perf_props, list) { 576 kfree(perf->attr_group); 577 perf->attr_group = NULL; 578 } 579 kobject_del(dev->kobj_perf); 580 kobject_put(dev->kobj_perf); 581 dev->kobj_perf = NULL; 582 } 583 584 if (dev->kobj_node) { 585 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid); 586 sysfs_remove_file(dev->kobj_node, &dev->attr_name); 587 sysfs_remove_file(dev->kobj_node, &dev->attr_props); 588 kobject_del(dev->kobj_node); 589 kobject_put(dev->kobj_node); 590 dev->kobj_node = NULL; 591 } 592 } 593 594 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, 595 uint32_t id) 596 { 597 struct kfd_iolink_properties *iolink; 598 struct kfd_cache_properties *cache; 599 struct kfd_mem_properties *mem; 600 struct kfd_perf_properties *perf; 601 int ret; 602 uint32_t i, num_attrs; 603 struct attribute **attrs; 604 605 if (WARN_ON(dev->kobj_node)) 606 return -EEXIST; 607 608 /* 609 * Creating the sysfs folders 610 */ 611 dev->kobj_node = kfd_alloc_struct(dev->kobj_node); 612 if (!dev->kobj_node) 613 return -ENOMEM; 614 615 ret = kobject_init_and_add(dev->kobj_node, &node_type, 616 sys_props.kobj_nodes, "%d", id); 617 if (ret < 0) 618 return ret; 619 620 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node); 621 if (!dev->kobj_mem) 622 return -ENOMEM; 623 624 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node); 625 if (!dev->kobj_cache) 626 return -ENOMEM; 627 628 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node); 629 if (!dev->kobj_iolink) 630 return -ENOMEM; 631 632 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node); 633 if (!dev->kobj_perf) 634 return -ENOMEM; 635 636 /* 637 * Creating sysfs files for node properties 638 */ 639 dev->attr_gpuid.name = "gpu_id"; 640 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE; 641 sysfs_attr_init(&dev->attr_gpuid); 642 dev->attr_name.name = "name"; 643 dev->attr_name.mode = KFD_SYSFS_FILE_MODE; 644 sysfs_attr_init(&dev->attr_name); 645 dev->attr_props.name = "properties"; 646 dev->attr_props.mode = KFD_SYSFS_FILE_MODE; 647 sysfs_attr_init(&dev->attr_props); 648 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid); 649 if (ret < 0) 650 return ret; 651 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name); 652 if (ret < 0) 653 return ret; 654 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props); 655 if (ret < 0) 656 return ret; 657 658 i = 0; 659 list_for_each_entry(mem, &dev->mem_props, list) { 660 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 661 if (!mem->kobj) 662 return -ENOMEM; 663 ret = kobject_init_and_add(mem->kobj, &mem_type, 664 dev->kobj_mem, "%d", i); 665 if (ret < 0) 666 return ret; 667 668 mem->attr.name = "properties"; 669 mem->attr.mode = KFD_SYSFS_FILE_MODE; 670 sysfs_attr_init(&mem->attr); 671 ret = sysfs_create_file(mem->kobj, &mem->attr); 672 if (ret < 0) 673 return ret; 674 i++; 675 } 676 677 i = 0; 678 list_for_each_entry(cache, &dev->cache_props, list) { 679 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 680 if (!cache->kobj) 681 return -ENOMEM; 682 ret = kobject_init_and_add(cache->kobj, &cache_type, 683 dev->kobj_cache, "%d", i); 684 if (ret < 0) 685 return ret; 686 687 cache->attr.name = "properties"; 688 cache->attr.mode = KFD_SYSFS_FILE_MODE; 689 sysfs_attr_init(&cache->attr); 690 ret = sysfs_create_file(cache->kobj, &cache->attr); 691 if (ret < 0) 692 return ret; 693 i++; 694 } 695 696 i = 0; 697 list_for_each_entry(iolink, &dev->io_link_props, list) { 698 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 699 if (!iolink->kobj) 700 return -ENOMEM; 701 ret = kobject_init_and_add(iolink->kobj, &iolink_type, 702 dev->kobj_iolink, "%d", i); 703 if (ret < 0) 704 return ret; 705 706 iolink->attr.name = "properties"; 707 iolink->attr.mode = KFD_SYSFS_FILE_MODE; 708 sysfs_attr_init(&iolink->attr); 709 ret = sysfs_create_file(iolink->kobj, &iolink->attr); 710 if (ret < 0) 711 return ret; 712 i++; 713 } 714 715 /* All hardware blocks have the same number of attributes. */ 716 num_attrs = ARRAY_SIZE(perf_attr_iommu); 717 list_for_each_entry(perf, &dev->perf_props, list) { 718 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr) 719 * num_attrs + sizeof(struct attribute_group), 720 GFP_KERNEL); 721 if (!perf->attr_group) 722 return -ENOMEM; 723 724 attrs = (struct attribute **)(perf->attr_group + 1); 725 if (!strcmp(perf->block_name, "iommu")) { 726 /* Information of IOMMU's num_counters and counter_ids is shown 727 * under /sys/bus/event_source/devices/amd_iommu. We don't 728 * duplicate here. 729 */ 730 perf_attr_iommu[0].data = perf->max_concurrent; 731 for (i = 0; i < num_attrs; i++) 732 attrs[i] = &perf_attr_iommu[i].attr.attr; 733 } 734 perf->attr_group->name = perf->block_name; 735 perf->attr_group->attrs = attrs; 736 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group); 737 if (ret < 0) 738 return ret; 739 } 740 741 return 0; 742 } 743 744 /* Called with write topology lock acquired */ 745 static int kfd_build_sysfs_node_tree(void) 746 { 747 struct kfd_topology_device *dev; 748 int ret; 749 uint32_t i = 0; 750 751 list_for_each_entry(dev, &topology_device_list, list) { 752 ret = kfd_build_sysfs_node_entry(dev, i); 753 if (ret < 0) 754 return ret; 755 i++; 756 } 757 758 return 0; 759 } 760 761 /* Called with write topology lock acquired */ 762 static void kfd_remove_sysfs_node_tree(void) 763 { 764 struct kfd_topology_device *dev; 765 766 list_for_each_entry(dev, &topology_device_list, list) 767 kfd_remove_sysfs_node_entry(dev); 768 } 769 770 static int kfd_topology_update_sysfs(void) 771 { 772 int ret; 773 774 pr_info("Creating topology SYSFS entries\n"); 775 if (!sys_props.kobj_topology) { 776 sys_props.kobj_topology = 777 kfd_alloc_struct(sys_props.kobj_topology); 778 if (!sys_props.kobj_topology) 779 return -ENOMEM; 780 781 ret = kobject_init_and_add(sys_props.kobj_topology, 782 &sysprops_type, &kfd_device->kobj, 783 "topology"); 784 if (ret < 0) 785 return ret; 786 787 sys_props.kobj_nodes = kobject_create_and_add("nodes", 788 sys_props.kobj_topology); 789 if (!sys_props.kobj_nodes) 790 return -ENOMEM; 791 792 sys_props.attr_genid.name = "generation_id"; 793 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE; 794 sysfs_attr_init(&sys_props.attr_genid); 795 ret = sysfs_create_file(sys_props.kobj_topology, 796 &sys_props.attr_genid); 797 if (ret < 0) 798 return ret; 799 800 sys_props.attr_props.name = "system_properties"; 801 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE; 802 sysfs_attr_init(&sys_props.attr_props); 803 ret = sysfs_create_file(sys_props.kobj_topology, 804 &sys_props.attr_props); 805 if (ret < 0) 806 return ret; 807 } 808 809 kfd_remove_sysfs_node_tree(); 810 811 return kfd_build_sysfs_node_tree(); 812 } 813 814 static void kfd_topology_release_sysfs(void) 815 { 816 kfd_remove_sysfs_node_tree(); 817 if (sys_props.kobj_topology) { 818 sysfs_remove_file(sys_props.kobj_topology, 819 &sys_props.attr_genid); 820 sysfs_remove_file(sys_props.kobj_topology, 821 &sys_props.attr_props); 822 if (sys_props.kobj_nodes) { 823 kobject_del(sys_props.kobj_nodes); 824 kobject_put(sys_props.kobj_nodes); 825 sys_props.kobj_nodes = NULL; 826 } 827 kobject_del(sys_props.kobj_topology); 828 kobject_put(sys_props.kobj_topology); 829 sys_props.kobj_topology = NULL; 830 } 831 } 832 833 /* Called with write topology_lock acquired */ 834 static void kfd_topology_update_device_list(struct list_head *temp_list, 835 struct list_head *master_list) 836 { 837 while (!list_empty(temp_list)) { 838 list_move_tail(temp_list->next, master_list); 839 sys_props.num_devices++; 840 } 841 } 842 843 static void kfd_debug_print_topology(void) 844 { 845 struct kfd_topology_device *dev; 846 847 down_read(&topology_lock); 848 849 dev = list_last_entry(&topology_device_list, 850 struct kfd_topology_device, list); 851 if (dev) { 852 if (dev->node_props.cpu_cores_count && 853 dev->node_props.simd_count) { 854 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n", 855 dev->node_props.device_id, 856 dev->node_props.vendor_id); 857 } else if (dev->node_props.cpu_cores_count) 858 pr_info("Topology: Add CPU node\n"); 859 else if (dev->node_props.simd_count) 860 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n", 861 dev->node_props.device_id, 862 dev->node_props.vendor_id); 863 } 864 up_read(&topology_lock); 865 } 866 867 /* Helper function for intializing platform_xx members of 868 * kfd_system_properties. Uses OEM info from the last CPU/APU node. 869 */ 870 static void kfd_update_system_properties(void) 871 { 872 struct kfd_topology_device *dev; 873 874 down_read(&topology_lock); 875 dev = list_last_entry(&topology_device_list, 876 struct kfd_topology_device, list); 877 if (dev) { 878 sys_props.platform_id = 879 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK; 880 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id); 881 sys_props.platform_rev = dev->oem_revision; 882 } 883 up_read(&topology_lock); 884 } 885 886 static void find_system_memory(const struct dmi_header *dm, 887 void *private) 888 { 889 struct kfd_mem_properties *mem; 890 u16 mem_width, mem_clock; 891 struct kfd_topology_device *kdev = 892 (struct kfd_topology_device *)private; 893 const u8 *dmi_data = (const u8 *)(dm + 1); 894 895 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) { 896 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6)); 897 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11)); 898 list_for_each_entry(mem, &kdev->mem_props, list) { 899 if (mem_width != 0xFFFF && mem_width != 0) 900 mem->width = mem_width; 901 if (mem_clock != 0) 902 mem->mem_clk_max = mem_clock; 903 } 904 } 905 } 906 907 /* 908 * Performance counters information is not part of CRAT but we would like to 909 * put them in the sysfs under topology directory for Thunk to get the data. 910 * This function is called before updating the sysfs. 911 */ 912 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev) 913 { 914 /* These are the only counters supported so far */ 915 return kfd_iommu_add_perf_counters(kdev); 916 } 917 918 /* kfd_add_non_crat_information - Add information that is not currently 919 * defined in CRAT but is necessary for KFD topology 920 * @dev - topology device to which addition info is added 921 */ 922 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev) 923 { 924 /* Check if CPU only node. */ 925 if (!kdev->gpu) { 926 /* Add system memory information */ 927 dmi_walk(find_system_memory, kdev); 928 } 929 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */ 930 } 931 932 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices. 933 * Ignore CRAT for all other devices. AMD APU is identified if both CPU 934 * and GPU cores are present. 935 * @device_list - topology device list created by parsing ACPI CRAT table. 936 * @return - TRUE if invalid, FALSE is valid. 937 */ 938 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list) 939 { 940 struct kfd_topology_device *dev; 941 942 list_for_each_entry(dev, device_list, list) { 943 if (dev->node_props.cpu_cores_count && 944 dev->node_props.simd_count) 945 return false; 946 } 947 pr_info("Ignoring ACPI CRAT on non-APU system\n"); 948 return true; 949 } 950 951 int kfd_topology_init(void) 952 { 953 void *crat_image = NULL; 954 size_t image_size = 0; 955 int ret; 956 struct list_head temp_topology_device_list; 957 int cpu_only_node = 0; 958 struct kfd_topology_device *kdev; 959 int proximity_domain; 960 961 /* topology_device_list - Master list of all topology devices 962 * temp_topology_device_list - temporary list created while parsing CRAT 963 * or VCRAT. Once parsing is complete the contents of list is moved to 964 * topology_device_list 965 */ 966 967 /* Initialize the head for the both the lists */ 968 INIT_LIST_HEAD(&topology_device_list); 969 INIT_LIST_HEAD(&temp_topology_device_list); 970 init_rwsem(&topology_lock); 971 972 memset(&sys_props, 0, sizeof(sys_props)); 973 974 /* Proximity domains in ACPI CRAT tables start counting at 975 * 0. The same should be true for virtual CRAT tables created 976 * at this stage. GPUs added later in kfd_topology_add_device 977 * use a counter. 978 */ 979 proximity_domain = 0; 980 981 /* 982 * Get the CRAT image from the ACPI. If ACPI doesn't have one 983 * or if ACPI CRAT is invalid create a virtual CRAT. 984 * NOTE: The current implementation expects all AMD APUs to have 985 * CRAT. If no CRAT is available, it is assumed to be a CPU 986 */ 987 ret = kfd_create_crat_image_acpi(&crat_image, &image_size); 988 if (!ret) { 989 ret = kfd_parse_crat_table(crat_image, 990 &temp_topology_device_list, 991 proximity_domain); 992 if (ret || 993 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) { 994 kfd_release_topology_device_list( 995 &temp_topology_device_list); 996 kfd_destroy_crat_image(crat_image); 997 crat_image = NULL; 998 } 999 } 1000 1001 if (!crat_image) { 1002 ret = kfd_create_crat_image_virtual(&crat_image, &image_size, 1003 COMPUTE_UNIT_CPU, NULL, 1004 proximity_domain); 1005 cpu_only_node = 1; 1006 if (ret) { 1007 pr_err("Error creating VCRAT table for CPU\n"); 1008 return ret; 1009 } 1010 1011 ret = kfd_parse_crat_table(crat_image, 1012 &temp_topology_device_list, 1013 proximity_domain); 1014 if (ret) { 1015 pr_err("Error parsing VCRAT table for CPU\n"); 1016 goto err; 1017 } 1018 } 1019 1020 kdev = list_first_entry(&temp_topology_device_list, 1021 struct kfd_topology_device, list); 1022 kfd_add_perf_to_topology(kdev); 1023 1024 down_write(&topology_lock); 1025 kfd_topology_update_device_list(&temp_topology_device_list, 1026 &topology_device_list); 1027 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1); 1028 ret = kfd_topology_update_sysfs(); 1029 up_write(&topology_lock); 1030 1031 if (!ret) { 1032 sys_props.generation_count++; 1033 kfd_update_system_properties(); 1034 kfd_debug_print_topology(); 1035 pr_info("Finished initializing topology\n"); 1036 } else 1037 pr_err("Failed to update topology in sysfs ret=%d\n", ret); 1038 1039 /* For nodes with GPU, this information gets added 1040 * when GPU is detected (kfd_topology_add_device). 1041 */ 1042 if (cpu_only_node) { 1043 /* Add additional information to CPU only node created above */ 1044 down_write(&topology_lock); 1045 kdev = list_first_entry(&topology_device_list, 1046 struct kfd_topology_device, list); 1047 up_write(&topology_lock); 1048 kfd_add_non_crat_information(kdev); 1049 } 1050 1051 err: 1052 kfd_destroy_crat_image(crat_image); 1053 return ret; 1054 } 1055 1056 void kfd_topology_shutdown(void) 1057 { 1058 down_write(&topology_lock); 1059 kfd_topology_release_sysfs(); 1060 kfd_release_live_view(); 1061 up_write(&topology_lock); 1062 } 1063 1064 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) 1065 { 1066 uint32_t hashout; 1067 uint32_t buf[7]; 1068 uint64_t local_mem_size; 1069 int i; 1070 struct kfd_local_mem_info local_mem_info; 1071 1072 if (!gpu) 1073 return 0; 1074 1075 amdgpu_amdkfd_get_local_mem_info(gpu->kgd, &local_mem_info); 1076 1077 local_mem_size = local_mem_info.local_mem_size_private + 1078 local_mem_info.local_mem_size_public; 1079 1080 buf[0] = gpu->pdev->devfn; 1081 buf[1] = gpu->pdev->subsystem_vendor; 1082 buf[2] = gpu->pdev->subsystem_device; 1083 buf[3] = gpu->pdev->device; 1084 buf[4] = gpu->pdev->bus->number; 1085 buf[5] = lower_32_bits(local_mem_size); 1086 buf[6] = upper_32_bits(local_mem_size); 1087 1088 for (i = 0, hashout = 0; i < 7; i++) 1089 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); 1090 1091 return hashout; 1092 } 1093 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If 1094 * the GPU device is not already present in the topology device 1095 * list then return NULL. This means a new topology device has to 1096 * be created for this GPU. 1097 */ 1098 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) 1099 { 1100 struct kfd_topology_device *dev; 1101 struct kfd_topology_device *out_dev = NULL; 1102 1103 down_write(&topology_lock); 1104 list_for_each_entry(dev, &topology_device_list, list) { 1105 /* Discrete GPUs need their own topology device list 1106 * entries. Don't assign them to CPU/APU nodes. 1107 */ 1108 if (!gpu->device_info->needs_iommu_device && 1109 dev->node_props.cpu_cores_count) 1110 continue; 1111 1112 if (!dev->gpu && (dev->node_props.simd_count > 0)) { 1113 dev->gpu = gpu; 1114 out_dev = dev; 1115 break; 1116 } 1117 } 1118 up_write(&topology_lock); 1119 return out_dev; 1120 } 1121 1122 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival) 1123 { 1124 /* 1125 * TODO: Generate an event for thunk about the arrival/removal 1126 * of the GPU 1127 */ 1128 } 1129 1130 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info, 1131 * patch this after CRAT parsing. 1132 */ 1133 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) 1134 { 1135 struct kfd_mem_properties *mem; 1136 struct kfd_local_mem_info local_mem_info; 1137 1138 if (!dev) 1139 return; 1140 1141 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with 1142 * single bank of VRAM local memory. 1143 * for dGPUs - VCRAT reports only one bank of Local Memory 1144 * for APUs - If CRAT from ACPI reports more than one bank, then 1145 * all the banks will report the same mem_clk_max information 1146 */ 1147 amdgpu_amdkfd_get_local_mem_info(dev->gpu->kgd, &local_mem_info); 1148 1149 list_for_each_entry(mem, &dev->mem_props, list) 1150 mem->mem_clk_max = local_mem_info.mem_clk_max; 1151 } 1152 1153 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev) 1154 { 1155 struct kfd_iolink_properties *link, *cpu_link; 1156 struct kfd_topology_device *cpu_dev; 1157 uint32_t cap; 1158 uint32_t cpu_flag = CRAT_IOLINK_FLAGS_ENABLED; 1159 uint32_t flag = CRAT_IOLINK_FLAGS_ENABLED; 1160 1161 if (!dev || !dev->gpu) 1162 return; 1163 1164 pcie_capability_read_dword(dev->gpu->pdev, 1165 PCI_EXP_DEVCAP2, &cap); 1166 1167 if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | 1168 PCI_EXP_DEVCAP2_ATOMIC_COMP64))) 1169 cpu_flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | 1170 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; 1171 1172 if (!dev->gpu->pci_atomic_requested || 1173 dev->gpu->device_info->asic_family == CHIP_HAWAII) 1174 flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | 1175 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; 1176 1177 /* GPU only creates direct links so apply flags setting to all */ 1178 list_for_each_entry(link, &dev->io_link_props, list) { 1179 link->flags = flag; 1180 cpu_dev = kfd_topology_device_by_proximity_domain( 1181 link->node_to); 1182 if (cpu_dev) { 1183 list_for_each_entry(cpu_link, 1184 &cpu_dev->io_link_props, list) 1185 if (cpu_link->node_to == link->node_from) 1186 cpu_link->flags = cpu_flag; 1187 } 1188 } 1189 } 1190 1191 int kfd_topology_add_device(struct kfd_dev *gpu) 1192 { 1193 uint32_t gpu_id; 1194 struct kfd_topology_device *dev; 1195 struct kfd_cu_info cu_info; 1196 int res = 0; 1197 struct list_head temp_topology_device_list; 1198 void *crat_image = NULL; 1199 size_t image_size = 0; 1200 int proximity_domain; 1201 struct amdgpu_ras *ctx; 1202 1203 INIT_LIST_HEAD(&temp_topology_device_list); 1204 1205 gpu_id = kfd_generate_gpu_id(gpu); 1206 1207 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id); 1208 1209 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain); 1210 1211 /* Check to see if this gpu device exists in the topology_device_list. 1212 * If so, assign the gpu to that device, 1213 * else create a Virtual CRAT for this gpu device and then parse that 1214 * CRAT to create a new topology device. Once created assign the gpu to 1215 * that topology device 1216 */ 1217 dev = kfd_assign_gpu(gpu); 1218 if (!dev) { 1219 res = kfd_create_crat_image_virtual(&crat_image, &image_size, 1220 COMPUTE_UNIT_GPU, gpu, 1221 proximity_domain); 1222 if (res) { 1223 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n", 1224 gpu_id); 1225 return res; 1226 } 1227 res = kfd_parse_crat_table(crat_image, 1228 &temp_topology_device_list, 1229 proximity_domain); 1230 if (res) { 1231 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n", 1232 gpu_id); 1233 goto err; 1234 } 1235 1236 down_write(&topology_lock); 1237 kfd_topology_update_device_list(&temp_topology_device_list, 1238 &topology_device_list); 1239 1240 /* Update the SYSFS tree, since we added another topology 1241 * device 1242 */ 1243 res = kfd_topology_update_sysfs(); 1244 up_write(&topology_lock); 1245 1246 if (!res) 1247 sys_props.generation_count++; 1248 else 1249 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n", 1250 gpu_id, res); 1251 dev = kfd_assign_gpu(gpu); 1252 if (WARN_ON(!dev)) { 1253 res = -ENODEV; 1254 goto err; 1255 } 1256 } 1257 1258 dev->gpu_id = gpu_id; 1259 gpu->id = gpu_id; 1260 1261 /* TODO: Move the following lines to function 1262 * kfd_add_non_crat_information 1263 */ 1264 1265 /* Fill-in additional information that is not available in CRAT but 1266 * needed for the topology 1267 */ 1268 1269 amdgpu_amdkfd_get_cu_info(dev->gpu->kgd, &cu_info); 1270 dev->node_props.simd_arrays_per_engine = 1271 cu_info.num_shader_arrays_per_engine; 1272 1273 dev->node_props.vendor_id = gpu->pdev->vendor; 1274 dev->node_props.device_id = gpu->pdev->device; 1275 dev->node_props.location_id = pci_dev_id(gpu->pdev); 1276 dev->node_props.max_engine_clk_fcompute = 1277 amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->kgd); 1278 dev->node_props.max_engine_clk_ccompute = 1279 cpufreq_quick_get_max(0) / 1000; 1280 dev->node_props.drm_render_minor = 1281 gpu->shared_resources.drm_render_minor; 1282 1283 dev->node_props.hive_id = gpu->hive_id; 1284 1285 kfd_fill_mem_clk_max_info(dev); 1286 kfd_fill_iolink_non_crat_info(dev); 1287 1288 switch (dev->gpu->device_info->asic_family) { 1289 case CHIP_KAVERI: 1290 case CHIP_HAWAII: 1291 case CHIP_TONGA: 1292 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 << 1293 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1294 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1295 break; 1296 case CHIP_CARRIZO: 1297 case CHIP_FIJI: 1298 case CHIP_POLARIS10: 1299 case CHIP_POLARIS11: 1300 case CHIP_POLARIS12: 1301 pr_debug("Adding doorbell packet type capability\n"); 1302 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 << 1303 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1304 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1305 break; 1306 case CHIP_VEGA10: 1307 case CHIP_VEGA12: 1308 case CHIP_VEGA20: 1309 case CHIP_RAVEN: 1310 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 << 1311 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1312 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1313 break; 1314 default: 1315 WARN(1, "Unexpected ASIC family %u", 1316 dev->gpu->device_info->asic_family); 1317 } 1318 1319 /* Fix errors in CZ CRAT. 1320 * simd_count: Carrizo CRAT reports wrong simd_count, probably 1321 * because it doesn't consider masked out CUs 1322 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd 1323 * capability flag: Carrizo CRAT doesn't report IOMMU flags 1324 */ 1325 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) { 1326 dev->node_props.simd_count = 1327 cu_info.simd_per_cu * cu_info.cu_active_number; 1328 dev->node_props.max_waves_per_simd = 10; 1329 dev->node_props.capability |= HSA_CAP_ATS_PRESENT; 1330 } 1331 1332 ctx = amdgpu_ras_get_context((struct amdgpu_device *)(dev->gpu->kgd)); 1333 if (ctx) { 1334 /* kfd only concerns sram ecc on GFX/SDMA and HBM ecc on UMC */ 1335 dev->node_props.capability |= 1336 (((ctx->features & BIT(AMDGPU_RAS_BLOCK__SDMA)) != 0) || 1337 ((ctx->features & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0)) ? 1338 HSA_CAP_SRAM_EDCSUPPORTED : 0; 1339 dev->node_props.capability |= ((ctx->features & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ? 1340 HSA_CAP_MEM_EDCSUPPORTED : 0; 1341 1342 dev->node_props.capability |= (ctx->features != 0) ? 1343 HSA_CAP_RASEVENTNOTIFY : 0; 1344 } 1345 1346 kfd_debug_print_topology(); 1347 1348 if (!res) 1349 kfd_notify_gpu_change(gpu_id, 1); 1350 err: 1351 kfd_destroy_crat_image(crat_image); 1352 return res; 1353 } 1354 1355 int kfd_topology_remove_device(struct kfd_dev *gpu) 1356 { 1357 struct kfd_topology_device *dev, *tmp; 1358 uint32_t gpu_id; 1359 int res = -ENODEV; 1360 1361 down_write(&topology_lock); 1362 1363 list_for_each_entry_safe(dev, tmp, &topology_device_list, list) 1364 if (dev->gpu == gpu) { 1365 gpu_id = dev->gpu_id; 1366 kfd_remove_sysfs_node_entry(dev); 1367 kfd_release_topology_device(dev); 1368 sys_props.num_devices--; 1369 res = 0; 1370 if (kfd_topology_update_sysfs() < 0) 1371 kfd_topology_release_sysfs(); 1372 break; 1373 } 1374 1375 up_write(&topology_lock); 1376 1377 if (!res) 1378 kfd_notify_gpu_change(gpu_id, 0); 1379 1380 return res; 1381 } 1382 1383 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD 1384 * topology. If GPU device is found @idx, then valid kfd_dev pointer is 1385 * returned through @kdev 1386 * Return - 0: On success (@kdev will be NULL for non GPU nodes) 1387 * -1: If end of list 1388 */ 1389 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev) 1390 { 1391 1392 struct kfd_topology_device *top_dev; 1393 uint8_t device_idx = 0; 1394 1395 *kdev = NULL; 1396 down_read(&topology_lock); 1397 1398 list_for_each_entry(top_dev, &topology_device_list, list) { 1399 if (device_idx == idx) { 1400 *kdev = top_dev->gpu; 1401 up_read(&topology_lock); 1402 return 0; 1403 } 1404 1405 device_idx++; 1406 } 1407 1408 up_read(&topology_lock); 1409 1410 return -1; 1411 1412 } 1413 1414 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask) 1415 { 1416 int first_cpu_of_numa_node; 1417 1418 if (!cpumask || cpumask == cpu_none_mask) 1419 return -1; 1420 first_cpu_of_numa_node = cpumask_first(cpumask); 1421 if (first_cpu_of_numa_node >= nr_cpu_ids) 1422 return -1; 1423 #ifdef CONFIG_X86_64 1424 return cpu_data(first_cpu_of_numa_node).apicid; 1425 #else 1426 return first_cpu_of_numa_node; 1427 #endif 1428 } 1429 1430 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor 1431 * of the given NUMA node (numa_node_id) 1432 * Return -1 on failure 1433 */ 1434 int kfd_numa_node_to_apic_id(int numa_node_id) 1435 { 1436 if (numa_node_id == -1) { 1437 pr_warn("Invalid NUMA Node. Use online CPU mask\n"); 1438 return kfd_cpumask_to_apic_id(cpu_online_mask); 1439 } 1440 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id)); 1441 } 1442 1443 #if defined(CONFIG_DEBUG_FS) 1444 1445 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data) 1446 { 1447 struct kfd_topology_device *dev; 1448 unsigned int i = 0; 1449 int r = 0; 1450 1451 down_read(&topology_lock); 1452 1453 list_for_each_entry(dev, &topology_device_list, list) { 1454 if (!dev->gpu) { 1455 i++; 1456 continue; 1457 } 1458 1459 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); 1460 r = dqm_debugfs_hqds(m, dev->gpu->dqm); 1461 if (r) 1462 break; 1463 } 1464 1465 up_read(&topology_lock); 1466 1467 return r; 1468 } 1469 1470 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data) 1471 { 1472 struct kfd_topology_device *dev; 1473 unsigned int i = 0; 1474 int r = 0; 1475 1476 down_read(&topology_lock); 1477 1478 list_for_each_entry(dev, &topology_device_list, list) { 1479 if (!dev->gpu) { 1480 i++; 1481 continue; 1482 } 1483 1484 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); 1485 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets); 1486 if (r) 1487 break; 1488 } 1489 1490 up_read(&topology_lock); 1491 1492 return r; 1493 } 1494 1495 #endif 1496