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