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 if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu)) 273 return -EPERM; 274 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type); 275 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj); 276 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min); 277 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from); 278 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to); 279 sysfs_show_32bit_prop(buffer, "weight", iolink->weight); 280 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency); 281 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency); 282 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth); 283 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth); 284 sysfs_show_32bit_prop(buffer, "recommended_transfer_size", 285 iolink->rec_transfer_size); 286 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags); 287 288 return ret; 289 } 290 291 static const struct sysfs_ops iolink_ops = { 292 .show = iolink_show, 293 }; 294 295 static struct kobj_type iolink_type = { 296 .release = kfd_topology_kobj_release, 297 .sysfs_ops = &iolink_ops, 298 }; 299 300 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr, 301 char *buffer) 302 { 303 ssize_t ret; 304 struct kfd_mem_properties *mem; 305 306 /* Making sure that the buffer is an empty string */ 307 buffer[0] = 0; 308 309 mem = container_of(attr, struct kfd_mem_properties, attr); 310 if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu)) 311 return -EPERM; 312 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type); 313 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes); 314 sysfs_show_32bit_prop(buffer, "flags", mem->flags); 315 sysfs_show_32bit_prop(buffer, "width", mem->width); 316 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max); 317 318 return ret; 319 } 320 321 static const struct sysfs_ops mem_ops = { 322 .show = mem_show, 323 }; 324 325 static struct kobj_type mem_type = { 326 .release = kfd_topology_kobj_release, 327 .sysfs_ops = &mem_ops, 328 }; 329 330 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr, 331 char *buffer) 332 { 333 ssize_t ret; 334 uint32_t i, j; 335 struct kfd_cache_properties *cache; 336 337 /* Making sure that the buffer is an empty string */ 338 buffer[0] = 0; 339 340 cache = container_of(attr, struct kfd_cache_properties, attr); 341 if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu)) 342 return -EPERM; 343 sysfs_show_32bit_prop(buffer, "processor_id_low", 344 cache->processor_id_low); 345 sysfs_show_32bit_prop(buffer, "level", cache->cache_level); 346 sysfs_show_32bit_prop(buffer, "size", cache->cache_size); 347 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size); 348 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag", 349 cache->cachelines_per_tag); 350 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc); 351 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency); 352 sysfs_show_32bit_prop(buffer, "type", cache->cache_type); 353 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer); 354 for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++) 355 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) { 356 /* Check each bit */ 357 if (cache->sibling_map[i] & (1 << j)) 358 ret = snprintf(buffer, PAGE_SIZE, 359 "%s%d%s", buffer, 1, ","); 360 else 361 ret = snprintf(buffer, PAGE_SIZE, 362 "%s%d%s", buffer, 0, ","); 363 } 364 /* Replace the last "," with end of line */ 365 *(buffer + strlen(buffer) - 1) = 0xA; 366 return ret; 367 } 368 369 static const struct sysfs_ops cache_ops = { 370 .show = kfd_cache_show, 371 }; 372 373 static struct kobj_type cache_type = { 374 .release = kfd_topology_kobj_release, 375 .sysfs_ops = &cache_ops, 376 }; 377 378 /****** Sysfs of Performance Counters ******/ 379 380 struct kfd_perf_attr { 381 struct kobj_attribute attr; 382 uint32_t data; 383 }; 384 385 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs, 386 char *buf) 387 { 388 struct kfd_perf_attr *attr; 389 390 buf[0] = 0; 391 attr = container_of(attrs, struct kfd_perf_attr, attr); 392 if (!attr->data) /* invalid data for PMC */ 393 return 0; 394 else 395 return sysfs_show_32bit_val(buf, attr->data); 396 } 397 398 #define KFD_PERF_DESC(_name, _data) \ 399 { \ 400 .attr = __ATTR(_name, 0444, perf_show, NULL), \ 401 .data = _data, \ 402 } 403 404 static struct kfd_perf_attr perf_attr_iommu[] = { 405 KFD_PERF_DESC(max_concurrent, 0), 406 KFD_PERF_DESC(num_counters, 0), 407 KFD_PERF_DESC(counter_ids, 0), 408 }; 409 /****************************************/ 410 411 static ssize_t node_show(struct kobject *kobj, struct attribute *attr, 412 char *buffer) 413 { 414 struct kfd_topology_device *dev; 415 uint32_t log_max_watch_addr; 416 417 /* Making sure that the buffer is an empty string */ 418 buffer[0] = 0; 419 420 if (strcmp(attr->name, "gpu_id") == 0) { 421 dev = container_of(attr, struct kfd_topology_device, 422 attr_gpuid); 423 if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) 424 return -EPERM; 425 return sysfs_show_32bit_val(buffer, dev->gpu_id); 426 } 427 428 if (strcmp(attr->name, "name") == 0) { 429 dev = container_of(attr, struct kfd_topology_device, 430 attr_name); 431 432 if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) 433 return -EPERM; 434 return sysfs_show_str_val(buffer, dev->node_props.name); 435 } 436 437 dev = container_of(attr, struct kfd_topology_device, 438 attr_props); 439 if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) 440 return -EPERM; 441 sysfs_show_32bit_prop(buffer, "cpu_cores_count", 442 dev->node_props.cpu_cores_count); 443 sysfs_show_32bit_prop(buffer, "simd_count", 444 dev->node_props.simd_count); 445 sysfs_show_32bit_prop(buffer, "mem_banks_count", 446 dev->node_props.mem_banks_count); 447 sysfs_show_32bit_prop(buffer, "caches_count", 448 dev->node_props.caches_count); 449 sysfs_show_32bit_prop(buffer, "io_links_count", 450 dev->node_props.io_links_count); 451 sysfs_show_32bit_prop(buffer, "cpu_core_id_base", 452 dev->node_props.cpu_core_id_base); 453 sysfs_show_32bit_prop(buffer, "simd_id_base", 454 dev->node_props.simd_id_base); 455 sysfs_show_32bit_prop(buffer, "max_waves_per_simd", 456 dev->node_props.max_waves_per_simd); 457 sysfs_show_32bit_prop(buffer, "lds_size_in_kb", 458 dev->node_props.lds_size_in_kb); 459 sysfs_show_32bit_prop(buffer, "gds_size_in_kb", 460 dev->node_props.gds_size_in_kb); 461 sysfs_show_32bit_prop(buffer, "num_gws", 462 dev->node_props.num_gws); 463 sysfs_show_32bit_prop(buffer, "wave_front_size", 464 dev->node_props.wave_front_size); 465 sysfs_show_32bit_prop(buffer, "array_count", 466 dev->node_props.array_count); 467 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine", 468 dev->node_props.simd_arrays_per_engine); 469 sysfs_show_32bit_prop(buffer, "cu_per_simd_array", 470 dev->node_props.cu_per_simd_array); 471 sysfs_show_32bit_prop(buffer, "simd_per_cu", 472 dev->node_props.simd_per_cu); 473 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu", 474 dev->node_props.max_slots_scratch_cu); 475 sysfs_show_32bit_prop(buffer, "vendor_id", 476 dev->node_props.vendor_id); 477 sysfs_show_32bit_prop(buffer, "device_id", 478 dev->node_props.device_id); 479 sysfs_show_32bit_prop(buffer, "location_id", 480 dev->node_props.location_id); 481 sysfs_show_32bit_prop(buffer, "drm_render_minor", 482 dev->node_props.drm_render_minor); 483 sysfs_show_64bit_prop(buffer, "hive_id", 484 dev->node_props.hive_id); 485 sysfs_show_32bit_prop(buffer, "num_sdma_engines", 486 dev->node_props.num_sdma_engines); 487 sysfs_show_32bit_prop(buffer, "num_sdma_xgmi_engines", 488 dev->node_props.num_sdma_xgmi_engines); 489 sysfs_show_32bit_prop(buffer, "num_sdma_queues_per_engine", 490 dev->node_props.num_sdma_queues_per_engine); 491 sysfs_show_32bit_prop(buffer, "num_cp_queues", 492 dev->node_props.num_cp_queues); 493 494 if (dev->gpu) { 495 log_max_watch_addr = 496 __ilog2_u32(dev->gpu->device_info->num_of_watch_points); 497 498 if (log_max_watch_addr) { 499 dev->node_props.capability |= 500 HSA_CAP_WATCH_POINTS_SUPPORTED; 501 502 dev->node_props.capability |= 503 ((log_max_watch_addr << 504 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) & 505 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK); 506 } 507 508 if (dev->gpu->device_info->asic_family == CHIP_TONGA) 509 dev->node_props.capability |= 510 HSA_CAP_AQL_QUEUE_DOUBLE_MAP; 511 512 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute", 513 dev->node_props.max_engine_clk_fcompute); 514 515 sysfs_show_64bit_prop(buffer, "local_mem_size", 516 (unsigned long long int) 0); 517 518 sysfs_show_32bit_prop(buffer, "fw_version", 519 dev->gpu->mec_fw_version); 520 sysfs_show_32bit_prop(buffer, "capability", 521 dev->node_props.capability); 522 sysfs_show_32bit_prop(buffer, "sdma_fw_version", 523 dev->gpu->sdma_fw_version); 524 } 525 526 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute", 527 cpufreq_quick_get_max(0)/1000); 528 } 529 530 static const struct sysfs_ops node_ops = { 531 .show = node_show, 532 }; 533 534 static struct kobj_type node_type = { 535 .release = kfd_topology_kobj_release, 536 .sysfs_ops = &node_ops, 537 }; 538 539 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr) 540 { 541 sysfs_remove_file(kobj, attr); 542 kobject_del(kobj); 543 kobject_put(kobj); 544 } 545 546 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) 547 { 548 struct kfd_iolink_properties *iolink; 549 struct kfd_cache_properties *cache; 550 struct kfd_mem_properties *mem; 551 struct kfd_perf_properties *perf; 552 553 if (dev->kobj_iolink) { 554 list_for_each_entry(iolink, &dev->io_link_props, list) 555 if (iolink->kobj) { 556 kfd_remove_sysfs_file(iolink->kobj, 557 &iolink->attr); 558 iolink->kobj = NULL; 559 } 560 kobject_del(dev->kobj_iolink); 561 kobject_put(dev->kobj_iolink); 562 dev->kobj_iolink = NULL; 563 } 564 565 if (dev->kobj_cache) { 566 list_for_each_entry(cache, &dev->cache_props, list) 567 if (cache->kobj) { 568 kfd_remove_sysfs_file(cache->kobj, 569 &cache->attr); 570 cache->kobj = NULL; 571 } 572 kobject_del(dev->kobj_cache); 573 kobject_put(dev->kobj_cache); 574 dev->kobj_cache = NULL; 575 } 576 577 if (dev->kobj_mem) { 578 list_for_each_entry(mem, &dev->mem_props, list) 579 if (mem->kobj) { 580 kfd_remove_sysfs_file(mem->kobj, &mem->attr); 581 mem->kobj = NULL; 582 } 583 kobject_del(dev->kobj_mem); 584 kobject_put(dev->kobj_mem); 585 dev->kobj_mem = NULL; 586 } 587 588 if (dev->kobj_perf) { 589 list_for_each_entry(perf, &dev->perf_props, list) { 590 kfree(perf->attr_group); 591 perf->attr_group = NULL; 592 } 593 kobject_del(dev->kobj_perf); 594 kobject_put(dev->kobj_perf); 595 dev->kobj_perf = NULL; 596 } 597 598 if (dev->kobj_node) { 599 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid); 600 sysfs_remove_file(dev->kobj_node, &dev->attr_name); 601 sysfs_remove_file(dev->kobj_node, &dev->attr_props); 602 kobject_del(dev->kobj_node); 603 kobject_put(dev->kobj_node); 604 dev->kobj_node = NULL; 605 } 606 } 607 608 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, 609 uint32_t id) 610 { 611 struct kfd_iolink_properties *iolink; 612 struct kfd_cache_properties *cache; 613 struct kfd_mem_properties *mem; 614 struct kfd_perf_properties *perf; 615 int ret; 616 uint32_t i, num_attrs; 617 struct attribute **attrs; 618 619 if (WARN_ON(dev->kobj_node)) 620 return -EEXIST; 621 622 /* 623 * Creating the sysfs folders 624 */ 625 dev->kobj_node = kfd_alloc_struct(dev->kobj_node); 626 if (!dev->kobj_node) 627 return -ENOMEM; 628 629 ret = kobject_init_and_add(dev->kobj_node, &node_type, 630 sys_props.kobj_nodes, "%d", id); 631 if (ret < 0) 632 return ret; 633 634 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node); 635 if (!dev->kobj_mem) 636 return -ENOMEM; 637 638 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node); 639 if (!dev->kobj_cache) 640 return -ENOMEM; 641 642 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node); 643 if (!dev->kobj_iolink) 644 return -ENOMEM; 645 646 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node); 647 if (!dev->kobj_perf) 648 return -ENOMEM; 649 650 /* 651 * Creating sysfs files for node properties 652 */ 653 dev->attr_gpuid.name = "gpu_id"; 654 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE; 655 sysfs_attr_init(&dev->attr_gpuid); 656 dev->attr_name.name = "name"; 657 dev->attr_name.mode = KFD_SYSFS_FILE_MODE; 658 sysfs_attr_init(&dev->attr_name); 659 dev->attr_props.name = "properties"; 660 dev->attr_props.mode = KFD_SYSFS_FILE_MODE; 661 sysfs_attr_init(&dev->attr_props); 662 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid); 663 if (ret < 0) 664 return ret; 665 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name); 666 if (ret < 0) 667 return ret; 668 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props); 669 if (ret < 0) 670 return ret; 671 672 i = 0; 673 list_for_each_entry(mem, &dev->mem_props, list) { 674 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 675 if (!mem->kobj) 676 return -ENOMEM; 677 ret = kobject_init_and_add(mem->kobj, &mem_type, 678 dev->kobj_mem, "%d", i); 679 if (ret < 0) 680 return ret; 681 682 mem->attr.name = "properties"; 683 mem->attr.mode = KFD_SYSFS_FILE_MODE; 684 sysfs_attr_init(&mem->attr); 685 ret = sysfs_create_file(mem->kobj, &mem->attr); 686 if (ret < 0) 687 return ret; 688 i++; 689 } 690 691 i = 0; 692 list_for_each_entry(cache, &dev->cache_props, list) { 693 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 694 if (!cache->kobj) 695 return -ENOMEM; 696 ret = kobject_init_and_add(cache->kobj, &cache_type, 697 dev->kobj_cache, "%d", i); 698 if (ret < 0) 699 return ret; 700 701 cache->attr.name = "properties"; 702 cache->attr.mode = KFD_SYSFS_FILE_MODE; 703 sysfs_attr_init(&cache->attr); 704 ret = sysfs_create_file(cache->kobj, &cache->attr); 705 if (ret < 0) 706 return ret; 707 i++; 708 } 709 710 i = 0; 711 list_for_each_entry(iolink, &dev->io_link_props, list) { 712 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 713 if (!iolink->kobj) 714 return -ENOMEM; 715 ret = kobject_init_and_add(iolink->kobj, &iolink_type, 716 dev->kobj_iolink, "%d", i); 717 if (ret < 0) 718 return ret; 719 720 iolink->attr.name = "properties"; 721 iolink->attr.mode = KFD_SYSFS_FILE_MODE; 722 sysfs_attr_init(&iolink->attr); 723 ret = sysfs_create_file(iolink->kobj, &iolink->attr); 724 if (ret < 0) 725 return ret; 726 i++; 727 } 728 729 /* All hardware blocks have the same number of attributes. */ 730 num_attrs = ARRAY_SIZE(perf_attr_iommu); 731 list_for_each_entry(perf, &dev->perf_props, list) { 732 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr) 733 * num_attrs + sizeof(struct attribute_group), 734 GFP_KERNEL); 735 if (!perf->attr_group) 736 return -ENOMEM; 737 738 attrs = (struct attribute **)(perf->attr_group + 1); 739 if (!strcmp(perf->block_name, "iommu")) { 740 /* Information of IOMMU's num_counters and counter_ids is shown 741 * under /sys/bus/event_source/devices/amd_iommu. We don't 742 * duplicate here. 743 */ 744 perf_attr_iommu[0].data = perf->max_concurrent; 745 for (i = 0; i < num_attrs; i++) 746 attrs[i] = &perf_attr_iommu[i].attr.attr; 747 } 748 perf->attr_group->name = perf->block_name; 749 perf->attr_group->attrs = attrs; 750 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group); 751 if (ret < 0) 752 return ret; 753 } 754 755 return 0; 756 } 757 758 /* Called with write topology lock acquired */ 759 static int kfd_build_sysfs_node_tree(void) 760 { 761 struct kfd_topology_device *dev; 762 int ret; 763 uint32_t i = 0; 764 765 list_for_each_entry(dev, &topology_device_list, list) { 766 ret = kfd_build_sysfs_node_entry(dev, i); 767 if (ret < 0) 768 return ret; 769 i++; 770 } 771 772 return 0; 773 } 774 775 /* Called with write topology lock acquired */ 776 static void kfd_remove_sysfs_node_tree(void) 777 { 778 struct kfd_topology_device *dev; 779 780 list_for_each_entry(dev, &topology_device_list, list) 781 kfd_remove_sysfs_node_entry(dev); 782 } 783 784 static int kfd_topology_update_sysfs(void) 785 { 786 int ret; 787 788 pr_info("Creating topology SYSFS entries\n"); 789 if (!sys_props.kobj_topology) { 790 sys_props.kobj_topology = 791 kfd_alloc_struct(sys_props.kobj_topology); 792 if (!sys_props.kobj_topology) 793 return -ENOMEM; 794 795 ret = kobject_init_and_add(sys_props.kobj_topology, 796 &sysprops_type, &kfd_device->kobj, 797 "topology"); 798 if (ret < 0) 799 return ret; 800 801 sys_props.kobj_nodes = kobject_create_and_add("nodes", 802 sys_props.kobj_topology); 803 if (!sys_props.kobj_nodes) 804 return -ENOMEM; 805 806 sys_props.attr_genid.name = "generation_id"; 807 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE; 808 sysfs_attr_init(&sys_props.attr_genid); 809 ret = sysfs_create_file(sys_props.kobj_topology, 810 &sys_props.attr_genid); 811 if (ret < 0) 812 return ret; 813 814 sys_props.attr_props.name = "system_properties"; 815 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE; 816 sysfs_attr_init(&sys_props.attr_props); 817 ret = sysfs_create_file(sys_props.kobj_topology, 818 &sys_props.attr_props); 819 if (ret < 0) 820 return ret; 821 } 822 823 kfd_remove_sysfs_node_tree(); 824 825 return kfd_build_sysfs_node_tree(); 826 } 827 828 static void kfd_topology_release_sysfs(void) 829 { 830 kfd_remove_sysfs_node_tree(); 831 if (sys_props.kobj_topology) { 832 sysfs_remove_file(sys_props.kobj_topology, 833 &sys_props.attr_genid); 834 sysfs_remove_file(sys_props.kobj_topology, 835 &sys_props.attr_props); 836 if (sys_props.kobj_nodes) { 837 kobject_del(sys_props.kobj_nodes); 838 kobject_put(sys_props.kobj_nodes); 839 sys_props.kobj_nodes = NULL; 840 } 841 kobject_del(sys_props.kobj_topology); 842 kobject_put(sys_props.kobj_topology); 843 sys_props.kobj_topology = NULL; 844 } 845 } 846 847 /* Called with write topology_lock acquired */ 848 static void kfd_topology_update_device_list(struct list_head *temp_list, 849 struct list_head *master_list) 850 { 851 while (!list_empty(temp_list)) { 852 list_move_tail(temp_list->next, master_list); 853 sys_props.num_devices++; 854 } 855 } 856 857 static void kfd_debug_print_topology(void) 858 { 859 struct kfd_topology_device *dev; 860 861 down_read(&topology_lock); 862 863 dev = list_last_entry(&topology_device_list, 864 struct kfd_topology_device, list); 865 if (dev) { 866 if (dev->node_props.cpu_cores_count && 867 dev->node_props.simd_count) { 868 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n", 869 dev->node_props.device_id, 870 dev->node_props.vendor_id); 871 } else if (dev->node_props.cpu_cores_count) 872 pr_info("Topology: Add CPU node\n"); 873 else if (dev->node_props.simd_count) 874 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n", 875 dev->node_props.device_id, 876 dev->node_props.vendor_id); 877 } 878 up_read(&topology_lock); 879 } 880 881 /* Helper function for intializing platform_xx members of 882 * kfd_system_properties. Uses OEM info from the last CPU/APU node. 883 */ 884 static void kfd_update_system_properties(void) 885 { 886 struct kfd_topology_device *dev; 887 888 down_read(&topology_lock); 889 dev = list_last_entry(&topology_device_list, 890 struct kfd_topology_device, list); 891 if (dev) { 892 sys_props.platform_id = 893 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK; 894 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id); 895 sys_props.platform_rev = dev->oem_revision; 896 } 897 up_read(&topology_lock); 898 } 899 900 static void find_system_memory(const struct dmi_header *dm, 901 void *private) 902 { 903 struct kfd_mem_properties *mem; 904 u16 mem_width, mem_clock; 905 struct kfd_topology_device *kdev = 906 (struct kfd_topology_device *)private; 907 const u8 *dmi_data = (const u8 *)(dm + 1); 908 909 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) { 910 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6)); 911 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11)); 912 list_for_each_entry(mem, &kdev->mem_props, list) { 913 if (mem_width != 0xFFFF && mem_width != 0) 914 mem->width = mem_width; 915 if (mem_clock != 0) 916 mem->mem_clk_max = mem_clock; 917 } 918 } 919 } 920 921 /* 922 * Performance counters information is not part of CRAT but we would like to 923 * put them in the sysfs under topology directory for Thunk to get the data. 924 * This function is called before updating the sysfs. 925 */ 926 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev) 927 { 928 /* These are the only counters supported so far */ 929 return kfd_iommu_add_perf_counters(kdev); 930 } 931 932 /* kfd_add_non_crat_information - Add information that is not currently 933 * defined in CRAT but is necessary for KFD topology 934 * @dev - topology device to which addition info is added 935 */ 936 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev) 937 { 938 /* Check if CPU only node. */ 939 if (!kdev->gpu) { 940 /* Add system memory information */ 941 dmi_walk(find_system_memory, kdev); 942 } 943 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */ 944 } 945 946 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices. 947 * Ignore CRAT for all other devices. AMD APU is identified if both CPU 948 * and GPU cores are present. 949 * @device_list - topology device list created by parsing ACPI CRAT table. 950 * @return - TRUE if invalid, FALSE is valid. 951 */ 952 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list) 953 { 954 struct kfd_topology_device *dev; 955 956 list_for_each_entry(dev, device_list, list) { 957 if (dev->node_props.cpu_cores_count && 958 dev->node_props.simd_count) 959 return false; 960 } 961 pr_info("Ignoring ACPI CRAT on non-APU system\n"); 962 return true; 963 } 964 965 int kfd_topology_init(void) 966 { 967 void *crat_image = NULL; 968 size_t image_size = 0; 969 int ret; 970 struct list_head temp_topology_device_list; 971 int cpu_only_node = 0; 972 struct kfd_topology_device *kdev; 973 int proximity_domain; 974 975 /* topology_device_list - Master list of all topology devices 976 * temp_topology_device_list - temporary list created while parsing CRAT 977 * or VCRAT. Once parsing is complete the contents of list is moved to 978 * topology_device_list 979 */ 980 981 /* Initialize the head for the both the lists */ 982 INIT_LIST_HEAD(&topology_device_list); 983 INIT_LIST_HEAD(&temp_topology_device_list); 984 init_rwsem(&topology_lock); 985 986 memset(&sys_props, 0, sizeof(sys_props)); 987 988 /* Proximity domains in ACPI CRAT tables start counting at 989 * 0. The same should be true for virtual CRAT tables created 990 * at this stage. GPUs added later in kfd_topology_add_device 991 * use a counter. 992 */ 993 proximity_domain = 0; 994 995 /* 996 * Get the CRAT image from the ACPI. If ACPI doesn't have one 997 * or if ACPI CRAT is invalid create a virtual CRAT. 998 * NOTE: The current implementation expects all AMD APUs to have 999 * CRAT. If no CRAT is available, it is assumed to be a CPU 1000 */ 1001 ret = kfd_create_crat_image_acpi(&crat_image, &image_size); 1002 if (!ret) { 1003 ret = kfd_parse_crat_table(crat_image, 1004 &temp_topology_device_list, 1005 proximity_domain); 1006 if (ret || 1007 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) { 1008 kfd_release_topology_device_list( 1009 &temp_topology_device_list); 1010 kfd_destroy_crat_image(crat_image); 1011 crat_image = NULL; 1012 } 1013 } 1014 1015 if (!crat_image) { 1016 ret = kfd_create_crat_image_virtual(&crat_image, &image_size, 1017 COMPUTE_UNIT_CPU, NULL, 1018 proximity_domain); 1019 cpu_only_node = 1; 1020 if (ret) { 1021 pr_err("Error creating VCRAT table for CPU\n"); 1022 return ret; 1023 } 1024 1025 ret = kfd_parse_crat_table(crat_image, 1026 &temp_topology_device_list, 1027 proximity_domain); 1028 if (ret) { 1029 pr_err("Error parsing VCRAT table for CPU\n"); 1030 goto err; 1031 } 1032 } 1033 1034 kdev = list_first_entry(&temp_topology_device_list, 1035 struct kfd_topology_device, list); 1036 kfd_add_perf_to_topology(kdev); 1037 1038 down_write(&topology_lock); 1039 kfd_topology_update_device_list(&temp_topology_device_list, 1040 &topology_device_list); 1041 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1); 1042 ret = kfd_topology_update_sysfs(); 1043 up_write(&topology_lock); 1044 1045 if (!ret) { 1046 sys_props.generation_count++; 1047 kfd_update_system_properties(); 1048 kfd_debug_print_topology(); 1049 pr_info("Finished initializing topology\n"); 1050 } else 1051 pr_err("Failed to update topology in sysfs ret=%d\n", ret); 1052 1053 /* For nodes with GPU, this information gets added 1054 * when GPU is detected (kfd_topology_add_device). 1055 */ 1056 if (cpu_only_node) { 1057 /* Add additional information to CPU only node created above */ 1058 down_write(&topology_lock); 1059 kdev = list_first_entry(&topology_device_list, 1060 struct kfd_topology_device, list); 1061 up_write(&topology_lock); 1062 kfd_add_non_crat_information(kdev); 1063 } 1064 1065 err: 1066 kfd_destroy_crat_image(crat_image); 1067 return ret; 1068 } 1069 1070 void kfd_topology_shutdown(void) 1071 { 1072 down_write(&topology_lock); 1073 kfd_topology_release_sysfs(); 1074 kfd_release_live_view(); 1075 up_write(&topology_lock); 1076 } 1077 1078 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) 1079 { 1080 uint32_t hashout; 1081 uint32_t buf[7]; 1082 uint64_t local_mem_size; 1083 int i; 1084 struct kfd_local_mem_info local_mem_info; 1085 1086 if (!gpu) 1087 return 0; 1088 1089 amdgpu_amdkfd_get_local_mem_info(gpu->kgd, &local_mem_info); 1090 1091 local_mem_size = local_mem_info.local_mem_size_private + 1092 local_mem_info.local_mem_size_public; 1093 1094 buf[0] = gpu->pdev->devfn; 1095 buf[1] = gpu->pdev->subsystem_vendor | 1096 (gpu->pdev->subsystem_device << 16); 1097 buf[2] = pci_domain_nr(gpu->pdev->bus); 1098 buf[3] = gpu->pdev->device; 1099 buf[4] = gpu->pdev->bus->number; 1100 buf[5] = lower_32_bits(local_mem_size); 1101 buf[6] = upper_32_bits(local_mem_size); 1102 1103 for (i = 0, hashout = 0; i < 7; i++) 1104 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); 1105 1106 return hashout; 1107 } 1108 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If 1109 * the GPU device is not already present in the topology device 1110 * list then return NULL. This means a new topology device has to 1111 * be created for this GPU. 1112 */ 1113 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) 1114 { 1115 struct kfd_topology_device *dev; 1116 struct kfd_topology_device *out_dev = NULL; 1117 struct kfd_mem_properties *mem; 1118 struct kfd_cache_properties *cache; 1119 struct kfd_iolink_properties *iolink; 1120 1121 down_write(&topology_lock); 1122 list_for_each_entry(dev, &topology_device_list, list) { 1123 /* Discrete GPUs need their own topology device list 1124 * entries. Don't assign them to CPU/APU nodes. 1125 */ 1126 if (!gpu->device_info->needs_iommu_device && 1127 dev->node_props.cpu_cores_count) 1128 continue; 1129 1130 if (!dev->gpu && (dev->node_props.simd_count > 0)) { 1131 dev->gpu = gpu; 1132 out_dev = dev; 1133 1134 list_for_each_entry(mem, &dev->mem_props, list) 1135 mem->gpu = dev->gpu; 1136 list_for_each_entry(cache, &dev->cache_props, list) 1137 cache->gpu = dev->gpu; 1138 list_for_each_entry(iolink, &dev->io_link_props, list) 1139 iolink->gpu = dev->gpu; 1140 break; 1141 } 1142 } 1143 up_write(&topology_lock); 1144 return out_dev; 1145 } 1146 1147 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival) 1148 { 1149 /* 1150 * TODO: Generate an event for thunk about the arrival/removal 1151 * of the GPU 1152 */ 1153 } 1154 1155 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info, 1156 * patch this after CRAT parsing. 1157 */ 1158 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) 1159 { 1160 struct kfd_mem_properties *mem; 1161 struct kfd_local_mem_info local_mem_info; 1162 1163 if (!dev) 1164 return; 1165 1166 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with 1167 * single bank of VRAM local memory. 1168 * for dGPUs - VCRAT reports only one bank of Local Memory 1169 * for APUs - If CRAT from ACPI reports more than one bank, then 1170 * all the banks will report the same mem_clk_max information 1171 */ 1172 amdgpu_amdkfd_get_local_mem_info(dev->gpu->kgd, &local_mem_info); 1173 1174 list_for_each_entry(mem, &dev->mem_props, list) 1175 mem->mem_clk_max = local_mem_info.mem_clk_max; 1176 } 1177 1178 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev) 1179 { 1180 struct kfd_iolink_properties *link, *cpu_link; 1181 struct kfd_topology_device *cpu_dev; 1182 uint32_t cap; 1183 uint32_t cpu_flag = CRAT_IOLINK_FLAGS_ENABLED; 1184 uint32_t flag = CRAT_IOLINK_FLAGS_ENABLED; 1185 1186 if (!dev || !dev->gpu) 1187 return; 1188 1189 pcie_capability_read_dword(dev->gpu->pdev, 1190 PCI_EXP_DEVCAP2, &cap); 1191 1192 if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | 1193 PCI_EXP_DEVCAP2_ATOMIC_COMP64))) 1194 cpu_flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | 1195 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; 1196 1197 if (!dev->gpu->pci_atomic_requested || 1198 dev->gpu->device_info->asic_family == CHIP_HAWAII) 1199 flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | 1200 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; 1201 1202 /* GPU only creates direct links so apply flags setting to all */ 1203 list_for_each_entry(link, &dev->io_link_props, list) { 1204 link->flags = flag; 1205 cpu_dev = kfd_topology_device_by_proximity_domain( 1206 link->node_to); 1207 if (cpu_dev) { 1208 list_for_each_entry(cpu_link, 1209 &cpu_dev->io_link_props, list) 1210 if (cpu_link->node_to == link->node_from) 1211 cpu_link->flags = cpu_flag; 1212 } 1213 } 1214 } 1215 1216 int kfd_topology_add_device(struct kfd_dev *gpu) 1217 { 1218 uint32_t gpu_id; 1219 struct kfd_topology_device *dev; 1220 struct kfd_cu_info cu_info; 1221 int res = 0; 1222 struct list_head temp_topology_device_list; 1223 void *crat_image = NULL; 1224 size_t image_size = 0; 1225 int proximity_domain; 1226 struct amdgpu_ras *ctx; 1227 1228 INIT_LIST_HEAD(&temp_topology_device_list); 1229 1230 gpu_id = kfd_generate_gpu_id(gpu); 1231 1232 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id); 1233 1234 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain); 1235 1236 /* Check to see if this gpu device exists in the topology_device_list. 1237 * If so, assign the gpu to that device, 1238 * else create a Virtual CRAT for this gpu device and then parse that 1239 * CRAT to create a new topology device. Once created assign the gpu to 1240 * that topology device 1241 */ 1242 dev = kfd_assign_gpu(gpu); 1243 if (!dev) { 1244 res = kfd_create_crat_image_virtual(&crat_image, &image_size, 1245 COMPUTE_UNIT_GPU, gpu, 1246 proximity_domain); 1247 if (res) { 1248 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n", 1249 gpu_id); 1250 return res; 1251 } 1252 res = kfd_parse_crat_table(crat_image, 1253 &temp_topology_device_list, 1254 proximity_domain); 1255 if (res) { 1256 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n", 1257 gpu_id); 1258 goto err; 1259 } 1260 1261 down_write(&topology_lock); 1262 kfd_topology_update_device_list(&temp_topology_device_list, 1263 &topology_device_list); 1264 1265 /* Update the SYSFS tree, since we added another topology 1266 * device 1267 */ 1268 res = kfd_topology_update_sysfs(); 1269 up_write(&topology_lock); 1270 1271 if (!res) 1272 sys_props.generation_count++; 1273 else 1274 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n", 1275 gpu_id, res); 1276 dev = kfd_assign_gpu(gpu); 1277 if (WARN_ON(!dev)) { 1278 res = -ENODEV; 1279 goto err; 1280 } 1281 } 1282 1283 dev->gpu_id = gpu_id; 1284 gpu->id = gpu_id; 1285 1286 /* TODO: Move the following lines to function 1287 * kfd_add_non_crat_information 1288 */ 1289 1290 /* Fill-in additional information that is not available in CRAT but 1291 * needed for the topology 1292 */ 1293 1294 amdgpu_amdkfd_get_cu_info(dev->gpu->kgd, &cu_info); 1295 1296 strncpy(dev->node_props.name, gpu->device_info->asic_name, 1297 KFD_TOPOLOGY_PUBLIC_NAME_SIZE); 1298 1299 dev->node_props.simd_arrays_per_engine = 1300 cu_info.num_shader_arrays_per_engine; 1301 1302 dev->node_props.vendor_id = gpu->pdev->vendor; 1303 dev->node_props.device_id = gpu->pdev->device; 1304 dev->node_props.location_id = pci_dev_id(gpu->pdev); 1305 dev->node_props.max_engine_clk_fcompute = 1306 amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->kgd); 1307 dev->node_props.max_engine_clk_ccompute = 1308 cpufreq_quick_get_max(0) / 1000; 1309 dev->node_props.drm_render_minor = 1310 gpu->shared_resources.drm_render_minor; 1311 1312 dev->node_props.hive_id = gpu->hive_id; 1313 dev->node_props.num_sdma_engines = gpu->device_info->num_sdma_engines; 1314 dev->node_props.num_sdma_xgmi_engines = 1315 gpu->device_info->num_xgmi_sdma_engines; 1316 dev->node_props.num_sdma_queues_per_engine = 1317 gpu->device_info->num_sdma_queues_per_engine; 1318 dev->node_props.num_gws = (hws_gws_support && 1319 dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ? 1320 amdgpu_amdkfd_get_num_gws(dev->gpu->kgd) : 0; 1321 dev->node_props.num_cp_queues = get_queues_num(dev->gpu->dqm); 1322 1323 kfd_fill_mem_clk_max_info(dev); 1324 kfd_fill_iolink_non_crat_info(dev); 1325 1326 switch (dev->gpu->device_info->asic_family) { 1327 case CHIP_KAVERI: 1328 case CHIP_HAWAII: 1329 case CHIP_TONGA: 1330 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 << 1331 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1332 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1333 break; 1334 case CHIP_CARRIZO: 1335 case CHIP_FIJI: 1336 case CHIP_POLARIS10: 1337 case CHIP_POLARIS11: 1338 case CHIP_POLARIS12: 1339 case CHIP_VEGAM: 1340 pr_debug("Adding doorbell packet type capability\n"); 1341 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 << 1342 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1343 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1344 break; 1345 case CHIP_VEGA10: 1346 case CHIP_VEGA12: 1347 case CHIP_VEGA20: 1348 case CHIP_RAVEN: 1349 case CHIP_RENOIR: 1350 case CHIP_ARCTURUS: 1351 case CHIP_NAVI10: 1352 case CHIP_NAVI12: 1353 case CHIP_NAVI14: 1354 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 << 1355 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1356 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1357 break; 1358 default: 1359 WARN(1, "Unexpected ASIC family %u", 1360 dev->gpu->device_info->asic_family); 1361 } 1362 1363 /* 1364 * Overwrite ATS capability according to needs_iommu_device to fix 1365 * potential missing corresponding bit in CRAT of BIOS. 1366 */ 1367 if (dev->gpu->device_info->needs_iommu_device) 1368 dev->node_props.capability |= HSA_CAP_ATS_PRESENT; 1369 else 1370 dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT; 1371 1372 /* Fix errors in CZ CRAT. 1373 * simd_count: Carrizo CRAT reports wrong simd_count, probably 1374 * because it doesn't consider masked out CUs 1375 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd 1376 */ 1377 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) { 1378 dev->node_props.simd_count = 1379 cu_info.simd_per_cu * cu_info.cu_active_number; 1380 dev->node_props.max_waves_per_simd = 10; 1381 } 1382 1383 ctx = amdgpu_ras_get_context((struct amdgpu_device *)(dev->gpu->kgd)); 1384 if (ctx) { 1385 /* kfd only concerns sram ecc on GFX/SDMA and HBM ecc on UMC */ 1386 dev->node_props.capability |= 1387 (((ctx->features & BIT(AMDGPU_RAS_BLOCK__SDMA)) != 0) || 1388 ((ctx->features & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0)) ? 1389 HSA_CAP_SRAM_EDCSUPPORTED : 0; 1390 dev->node_props.capability |= ((ctx->features & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ? 1391 HSA_CAP_MEM_EDCSUPPORTED : 0; 1392 1393 dev->node_props.capability |= (ctx->features != 0) ? 1394 HSA_CAP_RASEVENTNOTIFY : 0; 1395 } 1396 1397 kfd_debug_print_topology(); 1398 1399 if (!res) 1400 kfd_notify_gpu_change(gpu_id, 1); 1401 err: 1402 kfd_destroy_crat_image(crat_image); 1403 return res; 1404 } 1405 1406 int kfd_topology_remove_device(struct kfd_dev *gpu) 1407 { 1408 struct kfd_topology_device *dev, *tmp; 1409 uint32_t gpu_id; 1410 int res = -ENODEV; 1411 1412 down_write(&topology_lock); 1413 1414 list_for_each_entry_safe(dev, tmp, &topology_device_list, list) 1415 if (dev->gpu == gpu) { 1416 gpu_id = dev->gpu_id; 1417 kfd_remove_sysfs_node_entry(dev); 1418 kfd_release_topology_device(dev); 1419 sys_props.num_devices--; 1420 res = 0; 1421 if (kfd_topology_update_sysfs() < 0) 1422 kfd_topology_release_sysfs(); 1423 break; 1424 } 1425 1426 up_write(&topology_lock); 1427 1428 if (!res) 1429 kfd_notify_gpu_change(gpu_id, 0); 1430 1431 return res; 1432 } 1433 1434 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD 1435 * topology. If GPU device is found @idx, then valid kfd_dev pointer is 1436 * returned through @kdev 1437 * Return - 0: On success (@kdev will be NULL for non GPU nodes) 1438 * -1: If end of list 1439 */ 1440 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev) 1441 { 1442 1443 struct kfd_topology_device *top_dev; 1444 uint8_t device_idx = 0; 1445 1446 *kdev = NULL; 1447 down_read(&topology_lock); 1448 1449 list_for_each_entry(top_dev, &topology_device_list, list) { 1450 if (device_idx == idx) { 1451 *kdev = top_dev->gpu; 1452 up_read(&topology_lock); 1453 return 0; 1454 } 1455 1456 device_idx++; 1457 } 1458 1459 up_read(&topology_lock); 1460 1461 return -1; 1462 1463 } 1464 1465 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask) 1466 { 1467 int first_cpu_of_numa_node; 1468 1469 if (!cpumask || cpumask == cpu_none_mask) 1470 return -1; 1471 first_cpu_of_numa_node = cpumask_first(cpumask); 1472 if (first_cpu_of_numa_node >= nr_cpu_ids) 1473 return -1; 1474 #ifdef CONFIG_X86_64 1475 return cpu_data(first_cpu_of_numa_node).apicid; 1476 #else 1477 return first_cpu_of_numa_node; 1478 #endif 1479 } 1480 1481 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor 1482 * of the given NUMA node (numa_node_id) 1483 * Return -1 on failure 1484 */ 1485 int kfd_numa_node_to_apic_id(int numa_node_id) 1486 { 1487 if (numa_node_id == -1) { 1488 pr_warn("Invalid NUMA Node. Use online CPU mask\n"); 1489 return kfd_cpumask_to_apic_id(cpu_online_mask); 1490 } 1491 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id)); 1492 } 1493 1494 #if defined(CONFIG_DEBUG_FS) 1495 1496 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data) 1497 { 1498 struct kfd_topology_device *dev; 1499 unsigned int i = 0; 1500 int r = 0; 1501 1502 down_read(&topology_lock); 1503 1504 list_for_each_entry(dev, &topology_device_list, list) { 1505 if (!dev->gpu) { 1506 i++; 1507 continue; 1508 } 1509 1510 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); 1511 r = dqm_debugfs_hqds(m, dev->gpu->dqm); 1512 if (r) 1513 break; 1514 } 1515 1516 up_read(&topology_lock); 1517 1518 return r; 1519 } 1520 1521 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data) 1522 { 1523 struct kfd_topology_device *dev; 1524 unsigned int i = 0; 1525 int r = 0; 1526 1527 down_read(&topology_lock); 1528 1529 list_for_each_entry(dev, &topology_device_list, list) { 1530 if (!dev->gpu) { 1531 i++; 1532 continue; 1533 } 1534 1535 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); 1536 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets); 1537 if (r) 1538 break; 1539 } 1540 1541 up_read(&topology_lock); 1542 1543 return r; 1544 } 1545 1546 #endif 1547