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 41 /* topology_device_list - Master list of all topology devices */ 42 static struct list_head topology_device_list; 43 static struct kfd_system_properties sys_props; 44 45 static DECLARE_RWSEM(topology_lock); 46 static atomic_t topology_crat_proximity_domain; 47 48 struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 49 uint32_t proximity_domain) 50 { 51 struct kfd_topology_device *top_dev; 52 struct kfd_topology_device *device = NULL; 53 54 down_read(&topology_lock); 55 56 list_for_each_entry(top_dev, &topology_device_list, list) 57 if (top_dev->proximity_domain == proximity_domain) { 58 device = top_dev; 59 break; 60 } 61 62 up_read(&topology_lock); 63 64 return device; 65 } 66 67 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id) 68 { 69 struct kfd_topology_device *top_dev = NULL; 70 struct kfd_topology_device *ret = NULL; 71 72 down_read(&topology_lock); 73 74 list_for_each_entry(top_dev, &topology_device_list, list) 75 if (top_dev->gpu_id == gpu_id) { 76 ret = top_dev; 77 break; 78 } 79 80 up_read(&topology_lock); 81 82 return ret; 83 } 84 85 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id) 86 { 87 struct kfd_topology_device *top_dev; 88 89 top_dev = kfd_topology_device_by_id(gpu_id); 90 if (!top_dev) 91 return NULL; 92 93 return top_dev->gpu; 94 } 95 96 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev) 97 { 98 struct kfd_topology_device *top_dev; 99 struct kfd_dev *device = NULL; 100 101 down_read(&topology_lock); 102 103 list_for_each_entry(top_dev, &topology_device_list, list) 104 if (top_dev->gpu->pdev == pdev) { 105 device = top_dev->gpu; 106 break; 107 } 108 109 up_read(&topology_lock); 110 111 return device; 112 } 113 114 /* Called with write topology_lock acquired */ 115 static void kfd_release_topology_device(struct kfd_topology_device *dev) 116 { 117 struct kfd_mem_properties *mem; 118 struct kfd_cache_properties *cache; 119 struct kfd_iolink_properties *iolink; 120 struct kfd_perf_properties *perf; 121 122 list_del(&dev->list); 123 124 while (dev->mem_props.next != &dev->mem_props) { 125 mem = container_of(dev->mem_props.next, 126 struct kfd_mem_properties, list); 127 list_del(&mem->list); 128 kfree(mem); 129 } 130 131 while (dev->cache_props.next != &dev->cache_props) { 132 cache = container_of(dev->cache_props.next, 133 struct kfd_cache_properties, list); 134 list_del(&cache->list); 135 kfree(cache); 136 } 137 138 while (dev->io_link_props.next != &dev->io_link_props) { 139 iolink = container_of(dev->io_link_props.next, 140 struct kfd_iolink_properties, list); 141 list_del(&iolink->list); 142 kfree(iolink); 143 } 144 145 while (dev->perf_props.next != &dev->perf_props) { 146 perf = container_of(dev->perf_props.next, 147 struct kfd_perf_properties, list); 148 list_del(&perf->list); 149 kfree(perf); 150 } 151 152 kfree(dev); 153 } 154 155 void kfd_release_topology_device_list(struct list_head *device_list) 156 { 157 struct kfd_topology_device *dev; 158 159 while (!list_empty(device_list)) { 160 dev = list_first_entry(device_list, 161 struct kfd_topology_device, list); 162 kfd_release_topology_device(dev); 163 } 164 } 165 166 static void kfd_release_live_view(void) 167 { 168 kfd_release_topology_device_list(&topology_device_list); 169 memset(&sys_props, 0, sizeof(sys_props)); 170 } 171 172 struct kfd_topology_device *kfd_create_topology_device( 173 struct list_head *device_list) 174 { 175 struct kfd_topology_device *dev; 176 177 dev = kfd_alloc_struct(dev); 178 if (!dev) { 179 pr_err("No memory to allocate a topology device"); 180 return NULL; 181 } 182 183 INIT_LIST_HEAD(&dev->mem_props); 184 INIT_LIST_HEAD(&dev->cache_props); 185 INIT_LIST_HEAD(&dev->io_link_props); 186 INIT_LIST_HEAD(&dev->perf_props); 187 188 list_add_tail(&dev->list, device_list); 189 190 return dev; 191 } 192 193 194 #define sysfs_show_gen_prop(buffer, fmt, ...) \ 195 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__) 196 #define sysfs_show_32bit_prop(buffer, name, value) \ 197 sysfs_show_gen_prop(buffer, "%s %u\n", name, value) 198 #define sysfs_show_64bit_prop(buffer, name, value) \ 199 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value) 200 #define sysfs_show_32bit_val(buffer, value) \ 201 sysfs_show_gen_prop(buffer, "%u\n", value) 202 #define sysfs_show_str_val(buffer, value) \ 203 sysfs_show_gen_prop(buffer, "%s\n", value) 204 205 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr, 206 char *buffer) 207 { 208 ssize_t ret; 209 210 /* Making sure that the buffer is an empty string */ 211 buffer[0] = 0; 212 213 if (attr == &sys_props.attr_genid) { 214 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count); 215 } else if (attr == &sys_props.attr_props) { 216 sysfs_show_64bit_prop(buffer, "platform_oem", 217 sys_props.platform_oem); 218 sysfs_show_64bit_prop(buffer, "platform_id", 219 sys_props.platform_id); 220 ret = sysfs_show_64bit_prop(buffer, "platform_rev", 221 sys_props.platform_rev); 222 } else { 223 ret = -EINVAL; 224 } 225 226 return ret; 227 } 228 229 static void kfd_topology_kobj_release(struct kobject *kobj) 230 { 231 kfree(kobj); 232 } 233 234 static const struct sysfs_ops sysprops_ops = { 235 .show = sysprops_show, 236 }; 237 238 static struct kobj_type sysprops_type = { 239 .release = kfd_topology_kobj_release, 240 .sysfs_ops = &sysprops_ops, 241 }; 242 243 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr, 244 char *buffer) 245 { 246 ssize_t ret; 247 struct kfd_iolink_properties *iolink; 248 249 /* Making sure that the buffer is an empty string */ 250 buffer[0] = 0; 251 252 iolink = container_of(attr, struct kfd_iolink_properties, attr); 253 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type); 254 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj); 255 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min); 256 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from); 257 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to); 258 sysfs_show_32bit_prop(buffer, "weight", iolink->weight); 259 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency); 260 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency); 261 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth); 262 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth); 263 sysfs_show_32bit_prop(buffer, "recommended_transfer_size", 264 iolink->rec_transfer_size); 265 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags); 266 267 return ret; 268 } 269 270 static const struct sysfs_ops iolink_ops = { 271 .show = iolink_show, 272 }; 273 274 static struct kobj_type iolink_type = { 275 .release = kfd_topology_kobj_release, 276 .sysfs_ops = &iolink_ops, 277 }; 278 279 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr, 280 char *buffer) 281 { 282 ssize_t ret; 283 struct kfd_mem_properties *mem; 284 285 /* Making sure that the buffer is an empty string */ 286 buffer[0] = 0; 287 288 mem = container_of(attr, struct kfd_mem_properties, attr); 289 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type); 290 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes); 291 sysfs_show_32bit_prop(buffer, "flags", mem->flags); 292 sysfs_show_32bit_prop(buffer, "width", mem->width); 293 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max); 294 295 return ret; 296 } 297 298 static const struct sysfs_ops mem_ops = { 299 .show = mem_show, 300 }; 301 302 static struct kobj_type mem_type = { 303 .release = kfd_topology_kobj_release, 304 .sysfs_ops = &mem_ops, 305 }; 306 307 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr, 308 char *buffer) 309 { 310 ssize_t ret; 311 uint32_t i, j; 312 struct kfd_cache_properties *cache; 313 314 /* Making sure that the buffer is an empty string */ 315 buffer[0] = 0; 316 317 cache = container_of(attr, struct kfd_cache_properties, attr); 318 sysfs_show_32bit_prop(buffer, "processor_id_low", 319 cache->processor_id_low); 320 sysfs_show_32bit_prop(buffer, "level", cache->cache_level); 321 sysfs_show_32bit_prop(buffer, "size", cache->cache_size); 322 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size); 323 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag", 324 cache->cachelines_per_tag); 325 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc); 326 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency); 327 sysfs_show_32bit_prop(buffer, "type", cache->cache_type); 328 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer); 329 for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++) 330 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) { 331 /* Check each bit */ 332 if (cache->sibling_map[i] & (1 << j)) 333 ret = snprintf(buffer, PAGE_SIZE, 334 "%s%d%s", buffer, 1, ","); 335 else 336 ret = snprintf(buffer, PAGE_SIZE, 337 "%s%d%s", buffer, 0, ","); 338 } 339 /* Replace the last "," with end of line */ 340 *(buffer + strlen(buffer) - 1) = 0xA; 341 return ret; 342 } 343 344 static const struct sysfs_ops cache_ops = { 345 .show = kfd_cache_show, 346 }; 347 348 static struct kobj_type cache_type = { 349 .release = kfd_topology_kobj_release, 350 .sysfs_ops = &cache_ops, 351 }; 352 353 /****** Sysfs of Performance Counters ******/ 354 355 struct kfd_perf_attr { 356 struct kobj_attribute attr; 357 uint32_t data; 358 }; 359 360 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs, 361 char *buf) 362 { 363 struct kfd_perf_attr *attr; 364 365 buf[0] = 0; 366 attr = container_of(attrs, struct kfd_perf_attr, attr); 367 if (!attr->data) /* invalid data for PMC */ 368 return 0; 369 else 370 return sysfs_show_32bit_val(buf, attr->data); 371 } 372 373 #define KFD_PERF_DESC(_name, _data) \ 374 { \ 375 .attr = __ATTR(_name, 0444, perf_show, NULL), \ 376 .data = _data, \ 377 } 378 379 static struct kfd_perf_attr perf_attr_iommu[] = { 380 KFD_PERF_DESC(max_concurrent, 0), 381 KFD_PERF_DESC(num_counters, 0), 382 KFD_PERF_DESC(counter_ids, 0), 383 }; 384 /****************************************/ 385 386 static ssize_t node_show(struct kobject *kobj, struct attribute *attr, 387 char *buffer) 388 { 389 struct kfd_topology_device *dev; 390 char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE]; 391 uint32_t i; 392 uint32_t log_max_watch_addr; 393 394 /* Making sure that the buffer is an empty string */ 395 buffer[0] = 0; 396 397 if (strcmp(attr->name, "gpu_id") == 0) { 398 dev = container_of(attr, struct kfd_topology_device, 399 attr_gpuid); 400 return sysfs_show_32bit_val(buffer, dev->gpu_id); 401 } 402 403 if (strcmp(attr->name, "name") == 0) { 404 dev = container_of(attr, struct kfd_topology_device, 405 attr_name); 406 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) { 407 public_name[i] = 408 (char)dev->node_props.marketing_name[i]; 409 if (dev->node_props.marketing_name[i] == 0) 410 break; 411 } 412 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0; 413 return sysfs_show_str_val(buffer, public_name); 414 } 415 416 dev = container_of(attr, struct kfd_topology_device, 417 attr_props); 418 sysfs_show_32bit_prop(buffer, "cpu_cores_count", 419 dev->node_props.cpu_cores_count); 420 sysfs_show_32bit_prop(buffer, "simd_count", 421 dev->node_props.simd_count); 422 sysfs_show_32bit_prop(buffer, "mem_banks_count", 423 dev->node_props.mem_banks_count); 424 sysfs_show_32bit_prop(buffer, "caches_count", 425 dev->node_props.caches_count); 426 sysfs_show_32bit_prop(buffer, "io_links_count", 427 dev->node_props.io_links_count); 428 sysfs_show_32bit_prop(buffer, "cpu_core_id_base", 429 dev->node_props.cpu_core_id_base); 430 sysfs_show_32bit_prop(buffer, "simd_id_base", 431 dev->node_props.simd_id_base); 432 sysfs_show_32bit_prop(buffer, "max_waves_per_simd", 433 dev->node_props.max_waves_per_simd); 434 sysfs_show_32bit_prop(buffer, "lds_size_in_kb", 435 dev->node_props.lds_size_in_kb); 436 sysfs_show_32bit_prop(buffer, "gds_size_in_kb", 437 dev->node_props.gds_size_in_kb); 438 sysfs_show_32bit_prop(buffer, "wave_front_size", 439 dev->node_props.wave_front_size); 440 sysfs_show_32bit_prop(buffer, "array_count", 441 dev->node_props.array_count); 442 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine", 443 dev->node_props.simd_arrays_per_engine); 444 sysfs_show_32bit_prop(buffer, "cu_per_simd_array", 445 dev->node_props.cu_per_simd_array); 446 sysfs_show_32bit_prop(buffer, "simd_per_cu", 447 dev->node_props.simd_per_cu); 448 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu", 449 dev->node_props.max_slots_scratch_cu); 450 sysfs_show_32bit_prop(buffer, "vendor_id", 451 dev->node_props.vendor_id); 452 sysfs_show_32bit_prop(buffer, "device_id", 453 dev->node_props.device_id); 454 sysfs_show_32bit_prop(buffer, "location_id", 455 dev->node_props.location_id); 456 sysfs_show_32bit_prop(buffer, "drm_render_minor", 457 dev->node_props.drm_render_minor); 458 sysfs_show_64bit_prop(buffer, "hive_id", 459 dev->node_props.hive_id); 460 461 if (dev->gpu) { 462 log_max_watch_addr = 463 __ilog2_u32(dev->gpu->device_info->num_of_watch_points); 464 465 if (log_max_watch_addr) { 466 dev->node_props.capability |= 467 HSA_CAP_WATCH_POINTS_SUPPORTED; 468 469 dev->node_props.capability |= 470 ((log_max_watch_addr << 471 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) & 472 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK); 473 } 474 475 if (dev->gpu->device_info->asic_family == CHIP_TONGA) 476 dev->node_props.capability |= 477 HSA_CAP_AQL_QUEUE_DOUBLE_MAP; 478 479 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute", 480 dev->node_props.max_engine_clk_fcompute); 481 482 sysfs_show_64bit_prop(buffer, "local_mem_size", 483 (unsigned long long int) 0); 484 485 sysfs_show_32bit_prop(buffer, "fw_version", 486 dev->gpu->mec_fw_version); 487 sysfs_show_32bit_prop(buffer, "capability", 488 dev->node_props.capability); 489 sysfs_show_32bit_prop(buffer, "sdma_fw_version", 490 dev->gpu->sdma_fw_version); 491 } 492 493 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute", 494 cpufreq_quick_get_max(0)/1000); 495 } 496 497 static const struct sysfs_ops node_ops = { 498 .show = node_show, 499 }; 500 501 static struct kobj_type node_type = { 502 .release = kfd_topology_kobj_release, 503 .sysfs_ops = &node_ops, 504 }; 505 506 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr) 507 { 508 sysfs_remove_file(kobj, attr); 509 kobject_del(kobj); 510 kobject_put(kobj); 511 } 512 513 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) 514 { 515 struct kfd_iolink_properties *iolink; 516 struct kfd_cache_properties *cache; 517 struct kfd_mem_properties *mem; 518 struct kfd_perf_properties *perf; 519 520 if (dev->kobj_iolink) { 521 list_for_each_entry(iolink, &dev->io_link_props, list) 522 if (iolink->kobj) { 523 kfd_remove_sysfs_file(iolink->kobj, 524 &iolink->attr); 525 iolink->kobj = NULL; 526 } 527 kobject_del(dev->kobj_iolink); 528 kobject_put(dev->kobj_iolink); 529 dev->kobj_iolink = NULL; 530 } 531 532 if (dev->kobj_cache) { 533 list_for_each_entry(cache, &dev->cache_props, list) 534 if (cache->kobj) { 535 kfd_remove_sysfs_file(cache->kobj, 536 &cache->attr); 537 cache->kobj = NULL; 538 } 539 kobject_del(dev->kobj_cache); 540 kobject_put(dev->kobj_cache); 541 dev->kobj_cache = NULL; 542 } 543 544 if (dev->kobj_mem) { 545 list_for_each_entry(mem, &dev->mem_props, list) 546 if (mem->kobj) { 547 kfd_remove_sysfs_file(mem->kobj, &mem->attr); 548 mem->kobj = NULL; 549 } 550 kobject_del(dev->kobj_mem); 551 kobject_put(dev->kobj_mem); 552 dev->kobj_mem = NULL; 553 } 554 555 if (dev->kobj_perf) { 556 list_for_each_entry(perf, &dev->perf_props, list) { 557 kfree(perf->attr_group); 558 perf->attr_group = NULL; 559 } 560 kobject_del(dev->kobj_perf); 561 kobject_put(dev->kobj_perf); 562 dev->kobj_perf = NULL; 563 } 564 565 if (dev->kobj_node) { 566 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid); 567 sysfs_remove_file(dev->kobj_node, &dev->attr_name); 568 sysfs_remove_file(dev->kobj_node, &dev->attr_props); 569 kobject_del(dev->kobj_node); 570 kobject_put(dev->kobj_node); 571 dev->kobj_node = NULL; 572 } 573 } 574 575 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, 576 uint32_t id) 577 { 578 struct kfd_iolink_properties *iolink; 579 struct kfd_cache_properties *cache; 580 struct kfd_mem_properties *mem; 581 struct kfd_perf_properties *perf; 582 int ret; 583 uint32_t i, num_attrs; 584 struct attribute **attrs; 585 586 if (WARN_ON(dev->kobj_node)) 587 return -EEXIST; 588 589 /* 590 * Creating the sysfs folders 591 */ 592 dev->kobj_node = kfd_alloc_struct(dev->kobj_node); 593 if (!dev->kobj_node) 594 return -ENOMEM; 595 596 ret = kobject_init_and_add(dev->kobj_node, &node_type, 597 sys_props.kobj_nodes, "%d", id); 598 if (ret < 0) 599 return ret; 600 601 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node); 602 if (!dev->kobj_mem) 603 return -ENOMEM; 604 605 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node); 606 if (!dev->kobj_cache) 607 return -ENOMEM; 608 609 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node); 610 if (!dev->kobj_iolink) 611 return -ENOMEM; 612 613 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node); 614 if (!dev->kobj_perf) 615 return -ENOMEM; 616 617 /* 618 * Creating sysfs files for node properties 619 */ 620 dev->attr_gpuid.name = "gpu_id"; 621 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE; 622 sysfs_attr_init(&dev->attr_gpuid); 623 dev->attr_name.name = "name"; 624 dev->attr_name.mode = KFD_SYSFS_FILE_MODE; 625 sysfs_attr_init(&dev->attr_name); 626 dev->attr_props.name = "properties"; 627 dev->attr_props.mode = KFD_SYSFS_FILE_MODE; 628 sysfs_attr_init(&dev->attr_props); 629 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid); 630 if (ret < 0) 631 return ret; 632 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name); 633 if (ret < 0) 634 return ret; 635 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props); 636 if (ret < 0) 637 return ret; 638 639 i = 0; 640 list_for_each_entry(mem, &dev->mem_props, list) { 641 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 642 if (!mem->kobj) 643 return -ENOMEM; 644 ret = kobject_init_and_add(mem->kobj, &mem_type, 645 dev->kobj_mem, "%d", i); 646 if (ret < 0) 647 return ret; 648 649 mem->attr.name = "properties"; 650 mem->attr.mode = KFD_SYSFS_FILE_MODE; 651 sysfs_attr_init(&mem->attr); 652 ret = sysfs_create_file(mem->kobj, &mem->attr); 653 if (ret < 0) 654 return ret; 655 i++; 656 } 657 658 i = 0; 659 list_for_each_entry(cache, &dev->cache_props, list) { 660 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 661 if (!cache->kobj) 662 return -ENOMEM; 663 ret = kobject_init_and_add(cache->kobj, &cache_type, 664 dev->kobj_cache, "%d", i); 665 if (ret < 0) 666 return ret; 667 668 cache->attr.name = "properties"; 669 cache->attr.mode = KFD_SYSFS_FILE_MODE; 670 sysfs_attr_init(&cache->attr); 671 ret = sysfs_create_file(cache->kobj, &cache->attr); 672 if (ret < 0) 673 return ret; 674 i++; 675 } 676 677 i = 0; 678 list_for_each_entry(iolink, &dev->io_link_props, list) { 679 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 680 if (!iolink->kobj) 681 return -ENOMEM; 682 ret = kobject_init_and_add(iolink->kobj, &iolink_type, 683 dev->kobj_iolink, "%d", i); 684 if (ret < 0) 685 return ret; 686 687 iolink->attr.name = "properties"; 688 iolink->attr.mode = KFD_SYSFS_FILE_MODE; 689 sysfs_attr_init(&iolink->attr); 690 ret = sysfs_create_file(iolink->kobj, &iolink->attr); 691 if (ret < 0) 692 return ret; 693 i++; 694 } 695 696 /* All hardware blocks have the same number of attributes. */ 697 num_attrs = ARRAY_SIZE(perf_attr_iommu); 698 list_for_each_entry(perf, &dev->perf_props, list) { 699 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr) 700 * num_attrs + sizeof(struct attribute_group), 701 GFP_KERNEL); 702 if (!perf->attr_group) 703 return -ENOMEM; 704 705 attrs = (struct attribute **)(perf->attr_group + 1); 706 if (!strcmp(perf->block_name, "iommu")) { 707 /* Information of IOMMU's num_counters and counter_ids is shown 708 * under /sys/bus/event_source/devices/amd_iommu. We don't 709 * duplicate here. 710 */ 711 perf_attr_iommu[0].data = perf->max_concurrent; 712 for (i = 0; i < num_attrs; i++) 713 attrs[i] = &perf_attr_iommu[i].attr.attr; 714 } 715 perf->attr_group->name = perf->block_name; 716 perf->attr_group->attrs = attrs; 717 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group); 718 if (ret < 0) 719 return ret; 720 } 721 722 return 0; 723 } 724 725 /* Called with write topology lock acquired */ 726 static int kfd_build_sysfs_node_tree(void) 727 { 728 struct kfd_topology_device *dev; 729 int ret; 730 uint32_t i = 0; 731 732 list_for_each_entry(dev, &topology_device_list, list) { 733 ret = kfd_build_sysfs_node_entry(dev, i); 734 if (ret < 0) 735 return ret; 736 i++; 737 } 738 739 return 0; 740 } 741 742 /* Called with write topology lock acquired */ 743 static void kfd_remove_sysfs_node_tree(void) 744 { 745 struct kfd_topology_device *dev; 746 747 list_for_each_entry(dev, &topology_device_list, list) 748 kfd_remove_sysfs_node_entry(dev); 749 } 750 751 static int kfd_topology_update_sysfs(void) 752 { 753 int ret; 754 755 pr_info("Creating topology SYSFS entries\n"); 756 if (!sys_props.kobj_topology) { 757 sys_props.kobj_topology = 758 kfd_alloc_struct(sys_props.kobj_topology); 759 if (!sys_props.kobj_topology) 760 return -ENOMEM; 761 762 ret = kobject_init_and_add(sys_props.kobj_topology, 763 &sysprops_type, &kfd_device->kobj, 764 "topology"); 765 if (ret < 0) 766 return ret; 767 768 sys_props.kobj_nodes = kobject_create_and_add("nodes", 769 sys_props.kobj_topology); 770 if (!sys_props.kobj_nodes) 771 return -ENOMEM; 772 773 sys_props.attr_genid.name = "generation_id"; 774 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE; 775 sysfs_attr_init(&sys_props.attr_genid); 776 ret = sysfs_create_file(sys_props.kobj_topology, 777 &sys_props.attr_genid); 778 if (ret < 0) 779 return ret; 780 781 sys_props.attr_props.name = "system_properties"; 782 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE; 783 sysfs_attr_init(&sys_props.attr_props); 784 ret = sysfs_create_file(sys_props.kobj_topology, 785 &sys_props.attr_props); 786 if (ret < 0) 787 return ret; 788 } 789 790 kfd_remove_sysfs_node_tree(); 791 792 return kfd_build_sysfs_node_tree(); 793 } 794 795 static void kfd_topology_release_sysfs(void) 796 { 797 kfd_remove_sysfs_node_tree(); 798 if (sys_props.kobj_topology) { 799 sysfs_remove_file(sys_props.kobj_topology, 800 &sys_props.attr_genid); 801 sysfs_remove_file(sys_props.kobj_topology, 802 &sys_props.attr_props); 803 if (sys_props.kobj_nodes) { 804 kobject_del(sys_props.kobj_nodes); 805 kobject_put(sys_props.kobj_nodes); 806 sys_props.kobj_nodes = NULL; 807 } 808 kobject_del(sys_props.kobj_topology); 809 kobject_put(sys_props.kobj_topology); 810 sys_props.kobj_topology = NULL; 811 } 812 } 813 814 /* Called with write topology_lock acquired */ 815 static void kfd_topology_update_device_list(struct list_head *temp_list, 816 struct list_head *master_list) 817 { 818 while (!list_empty(temp_list)) { 819 list_move_tail(temp_list->next, master_list); 820 sys_props.num_devices++; 821 } 822 } 823 824 static void kfd_debug_print_topology(void) 825 { 826 struct kfd_topology_device *dev; 827 828 down_read(&topology_lock); 829 830 dev = list_last_entry(&topology_device_list, 831 struct kfd_topology_device, list); 832 if (dev) { 833 if (dev->node_props.cpu_cores_count && 834 dev->node_props.simd_count) { 835 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n", 836 dev->node_props.device_id, 837 dev->node_props.vendor_id); 838 } else if (dev->node_props.cpu_cores_count) 839 pr_info("Topology: Add CPU node\n"); 840 else if (dev->node_props.simd_count) 841 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n", 842 dev->node_props.device_id, 843 dev->node_props.vendor_id); 844 } 845 up_read(&topology_lock); 846 } 847 848 /* Helper function for intializing platform_xx members of 849 * kfd_system_properties. Uses OEM info from the last CPU/APU node. 850 */ 851 static void kfd_update_system_properties(void) 852 { 853 struct kfd_topology_device *dev; 854 855 down_read(&topology_lock); 856 dev = list_last_entry(&topology_device_list, 857 struct kfd_topology_device, list); 858 if (dev) { 859 sys_props.platform_id = 860 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK; 861 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id); 862 sys_props.platform_rev = dev->oem_revision; 863 } 864 up_read(&topology_lock); 865 } 866 867 static void find_system_memory(const struct dmi_header *dm, 868 void *private) 869 { 870 struct kfd_mem_properties *mem; 871 u16 mem_width, mem_clock; 872 struct kfd_topology_device *kdev = 873 (struct kfd_topology_device *)private; 874 const u8 *dmi_data = (const u8 *)(dm + 1); 875 876 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) { 877 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6)); 878 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11)); 879 list_for_each_entry(mem, &kdev->mem_props, list) { 880 if (mem_width != 0xFFFF && mem_width != 0) 881 mem->width = mem_width; 882 if (mem_clock != 0) 883 mem->mem_clk_max = mem_clock; 884 } 885 } 886 } 887 888 /* 889 * Performance counters information is not part of CRAT but we would like to 890 * put them in the sysfs under topology directory for Thunk to get the data. 891 * This function is called before updating the sysfs. 892 */ 893 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev) 894 { 895 /* These are the only counters supported so far */ 896 return kfd_iommu_add_perf_counters(kdev); 897 } 898 899 /* kfd_add_non_crat_information - Add information that is not currently 900 * defined in CRAT but is necessary for KFD topology 901 * @dev - topology device to which addition info is added 902 */ 903 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev) 904 { 905 /* Check if CPU only node. */ 906 if (!kdev->gpu) { 907 /* Add system memory information */ 908 dmi_walk(find_system_memory, kdev); 909 } 910 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */ 911 } 912 913 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices. 914 * Ignore CRAT for all other devices. AMD APU is identified if both CPU 915 * and GPU cores are present. 916 * @device_list - topology device list created by parsing ACPI CRAT table. 917 * @return - TRUE if invalid, FALSE is valid. 918 */ 919 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list) 920 { 921 struct kfd_topology_device *dev; 922 923 list_for_each_entry(dev, device_list, list) { 924 if (dev->node_props.cpu_cores_count && 925 dev->node_props.simd_count) 926 return false; 927 } 928 pr_info("Ignoring ACPI CRAT on non-APU system\n"); 929 return true; 930 } 931 932 int kfd_topology_init(void) 933 { 934 void *crat_image = NULL; 935 size_t image_size = 0; 936 int ret; 937 struct list_head temp_topology_device_list; 938 int cpu_only_node = 0; 939 struct kfd_topology_device *kdev; 940 int proximity_domain; 941 942 /* topology_device_list - Master list of all topology devices 943 * temp_topology_device_list - temporary list created while parsing CRAT 944 * or VCRAT. Once parsing is complete the contents of list is moved to 945 * topology_device_list 946 */ 947 948 /* Initialize the head for the both the lists */ 949 INIT_LIST_HEAD(&topology_device_list); 950 INIT_LIST_HEAD(&temp_topology_device_list); 951 init_rwsem(&topology_lock); 952 953 memset(&sys_props, 0, sizeof(sys_props)); 954 955 /* Proximity domains in ACPI CRAT tables start counting at 956 * 0. The same should be true for virtual CRAT tables created 957 * at this stage. GPUs added later in kfd_topology_add_device 958 * use a counter. 959 */ 960 proximity_domain = 0; 961 962 /* 963 * Get the CRAT image from the ACPI. If ACPI doesn't have one 964 * or if ACPI CRAT is invalid create a virtual CRAT. 965 * NOTE: The current implementation expects all AMD APUs to have 966 * CRAT. If no CRAT is available, it is assumed to be a CPU 967 */ 968 ret = kfd_create_crat_image_acpi(&crat_image, &image_size); 969 if (!ret) { 970 ret = kfd_parse_crat_table(crat_image, 971 &temp_topology_device_list, 972 proximity_domain); 973 if (ret || 974 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) { 975 kfd_release_topology_device_list( 976 &temp_topology_device_list); 977 kfd_destroy_crat_image(crat_image); 978 crat_image = NULL; 979 } 980 } 981 982 if (!crat_image) { 983 ret = kfd_create_crat_image_virtual(&crat_image, &image_size, 984 COMPUTE_UNIT_CPU, NULL, 985 proximity_domain); 986 cpu_only_node = 1; 987 if (ret) { 988 pr_err("Error creating VCRAT table for CPU\n"); 989 return ret; 990 } 991 992 ret = kfd_parse_crat_table(crat_image, 993 &temp_topology_device_list, 994 proximity_domain); 995 if (ret) { 996 pr_err("Error parsing VCRAT table for CPU\n"); 997 goto err; 998 } 999 } 1000 1001 kdev = list_first_entry(&temp_topology_device_list, 1002 struct kfd_topology_device, list); 1003 kfd_add_perf_to_topology(kdev); 1004 1005 down_write(&topology_lock); 1006 kfd_topology_update_device_list(&temp_topology_device_list, 1007 &topology_device_list); 1008 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1); 1009 ret = kfd_topology_update_sysfs(); 1010 up_write(&topology_lock); 1011 1012 if (!ret) { 1013 sys_props.generation_count++; 1014 kfd_update_system_properties(); 1015 kfd_debug_print_topology(); 1016 pr_info("Finished initializing topology\n"); 1017 } else 1018 pr_err("Failed to update topology in sysfs ret=%d\n", ret); 1019 1020 /* For nodes with GPU, this information gets added 1021 * when GPU is detected (kfd_topology_add_device). 1022 */ 1023 if (cpu_only_node) { 1024 /* Add additional information to CPU only node created above */ 1025 down_write(&topology_lock); 1026 kdev = list_first_entry(&topology_device_list, 1027 struct kfd_topology_device, list); 1028 up_write(&topology_lock); 1029 kfd_add_non_crat_information(kdev); 1030 } 1031 1032 err: 1033 kfd_destroy_crat_image(crat_image); 1034 return ret; 1035 } 1036 1037 void kfd_topology_shutdown(void) 1038 { 1039 down_write(&topology_lock); 1040 kfd_topology_release_sysfs(); 1041 kfd_release_live_view(); 1042 up_write(&topology_lock); 1043 } 1044 1045 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) 1046 { 1047 uint32_t hashout; 1048 uint32_t buf[7]; 1049 uint64_t local_mem_size; 1050 int i; 1051 struct kfd_local_mem_info local_mem_info; 1052 1053 if (!gpu) 1054 return 0; 1055 1056 amdgpu_amdkfd_get_local_mem_info(gpu->kgd, &local_mem_info); 1057 1058 local_mem_size = local_mem_info.local_mem_size_private + 1059 local_mem_info.local_mem_size_public; 1060 1061 buf[0] = gpu->pdev->devfn; 1062 buf[1] = gpu->pdev->subsystem_vendor; 1063 buf[2] = gpu->pdev->subsystem_device; 1064 buf[3] = gpu->pdev->device; 1065 buf[4] = gpu->pdev->bus->number; 1066 buf[5] = lower_32_bits(local_mem_size); 1067 buf[6] = upper_32_bits(local_mem_size); 1068 1069 for (i = 0, hashout = 0; i < 7; i++) 1070 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); 1071 1072 return hashout; 1073 } 1074 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If 1075 * the GPU device is not already present in the topology device 1076 * list then return NULL. This means a new topology device has to 1077 * be created for this GPU. 1078 * TODO: Rather than assiging @gpu to first topology device withtout 1079 * gpu attached, it will better to have more stringent check. 1080 */ 1081 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) 1082 { 1083 struct kfd_topology_device *dev; 1084 struct kfd_topology_device *out_dev = NULL; 1085 1086 down_write(&topology_lock); 1087 list_for_each_entry(dev, &topology_device_list, list) 1088 if (!dev->gpu && (dev->node_props.simd_count > 0)) { 1089 dev->gpu = gpu; 1090 out_dev = dev; 1091 break; 1092 } 1093 up_write(&topology_lock); 1094 return out_dev; 1095 } 1096 1097 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival) 1098 { 1099 /* 1100 * TODO: Generate an event for thunk about the arrival/removal 1101 * of the GPU 1102 */ 1103 } 1104 1105 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info, 1106 * patch this after CRAT parsing. 1107 */ 1108 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) 1109 { 1110 struct kfd_mem_properties *mem; 1111 struct kfd_local_mem_info local_mem_info; 1112 1113 if (!dev) 1114 return; 1115 1116 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with 1117 * single bank of VRAM local memory. 1118 * for dGPUs - VCRAT reports only one bank of Local Memory 1119 * for APUs - If CRAT from ACPI reports more than one bank, then 1120 * all the banks will report the same mem_clk_max information 1121 */ 1122 amdgpu_amdkfd_get_local_mem_info(dev->gpu->kgd, &local_mem_info); 1123 1124 list_for_each_entry(mem, &dev->mem_props, list) 1125 mem->mem_clk_max = local_mem_info.mem_clk_max; 1126 } 1127 1128 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev) 1129 { 1130 struct kfd_iolink_properties *link, *cpu_link; 1131 struct kfd_topology_device *cpu_dev; 1132 uint32_t cap; 1133 uint32_t cpu_flag = CRAT_IOLINK_FLAGS_ENABLED; 1134 uint32_t flag = CRAT_IOLINK_FLAGS_ENABLED; 1135 1136 if (!dev || !dev->gpu) 1137 return; 1138 1139 pcie_capability_read_dword(dev->gpu->pdev, 1140 PCI_EXP_DEVCAP2, &cap); 1141 1142 if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | 1143 PCI_EXP_DEVCAP2_ATOMIC_COMP64))) 1144 cpu_flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | 1145 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; 1146 1147 if (!dev->gpu->pci_atomic_requested || 1148 dev->gpu->device_info->asic_family == CHIP_HAWAII) 1149 flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | 1150 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; 1151 1152 /* GPU only creates direct links so apply flags setting to all */ 1153 list_for_each_entry(link, &dev->io_link_props, list) { 1154 link->flags = flag; 1155 cpu_dev = kfd_topology_device_by_proximity_domain( 1156 link->node_to); 1157 if (cpu_dev) { 1158 list_for_each_entry(cpu_link, 1159 &cpu_dev->io_link_props, list) 1160 if (cpu_link->node_to == link->node_from) 1161 cpu_link->flags = cpu_flag; 1162 } 1163 } 1164 } 1165 1166 int kfd_topology_add_device(struct kfd_dev *gpu) 1167 { 1168 uint32_t gpu_id; 1169 struct kfd_topology_device *dev; 1170 struct kfd_cu_info cu_info; 1171 int res = 0; 1172 struct list_head temp_topology_device_list; 1173 void *crat_image = NULL; 1174 size_t image_size = 0; 1175 int proximity_domain; 1176 1177 INIT_LIST_HEAD(&temp_topology_device_list); 1178 1179 gpu_id = kfd_generate_gpu_id(gpu); 1180 1181 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id); 1182 1183 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain); 1184 1185 /* Check to see if this gpu device exists in the topology_device_list. 1186 * If so, assign the gpu to that device, 1187 * else create a Virtual CRAT for this gpu device and then parse that 1188 * CRAT to create a new topology device. Once created assign the gpu to 1189 * that topology device 1190 */ 1191 dev = kfd_assign_gpu(gpu); 1192 if (!dev) { 1193 res = kfd_create_crat_image_virtual(&crat_image, &image_size, 1194 COMPUTE_UNIT_GPU, gpu, 1195 proximity_domain); 1196 if (res) { 1197 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n", 1198 gpu_id); 1199 return res; 1200 } 1201 res = kfd_parse_crat_table(crat_image, 1202 &temp_topology_device_list, 1203 proximity_domain); 1204 if (res) { 1205 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n", 1206 gpu_id); 1207 goto err; 1208 } 1209 1210 down_write(&topology_lock); 1211 kfd_topology_update_device_list(&temp_topology_device_list, 1212 &topology_device_list); 1213 1214 /* Update the SYSFS tree, since we added another topology 1215 * device 1216 */ 1217 res = kfd_topology_update_sysfs(); 1218 up_write(&topology_lock); 1219 1220 if (!res) 1221 sys_props.generation_count++; 1222 else 1223 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n", 1224 gpu_id, res); 1225 dev = kfd_assign_gpu(gpu); 1226 if (WARN_ON(!dev)) { 1227 res = -ENODEV; 1228 goto err; 1229 } 1230 } 1231 1232 dev->gpu_id = gpu_id; 1233 gpu->id = gpu_id; 1234 1235 /* TODO: Move the following lines to function 1236 * kfd_add_non_crat_information 1237 */ 1238 1239 /* Fill-in additional information that is not available in CRAT but 1240 * needed for the topology 1241 */ 1242 1243 amdgpu_amdkfd_get_cu_info(dev->gpu->kgd, &cu_info); 1244 dev->node_props.simd_arrays_per_engine = 1245 cu_info.num_shader_arrays_per_engine; 1246 1247 dev->node_props.vendor_id = gpu->pdev->vendor; 1248 dev->node_props.device_id = gpu->pdev->device; 1249 dev->node_props.location_id = PCI_DEVID(gpu->pdev->bus->number, 1250 gpu->pdev->devfn); 1251 dev->node_props.max_engine_clk_fcompute = 1252 amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->kgd); 1253 dev->node_props.max_engine_clk_ccompute = 1254 cpufreq_quick_get_max(0) / 1000; 1255 dev->node_props.drm_render_minor = 1256 gpu->shared_resources.drm_render_minor; 1257 1258 dev->node_props.hive_id = gpu->hive_id; 1259 1260 kfd_fill_mem_clk_max_info(dev); 1261 kfd_fill_iolink_non_crat_info(dev); 1262 1263 switch (dev->gpu->device_info->asic_family) { 1264 case CHIP_KAVERI: 1265 case CHIP_HAWAII: 1266 case CHIP_TONGA: 1267 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 << 1268 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1269 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1270 break; 1271 case CHIP_CARRIZO: 1272 case CHIP_FIJI: 1273 case CHIP_POLARIS10: 1274 case CHIP_POLARIS11: 1275 pr_debug("Adding doorbell packet type capability\n"); 1276 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 << 1277 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1278 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1279 break; 1280 case CHIP_VEGA10: 1281 case CHIP_VEGA20: 1282 case CHIP_RAVEN: 1283 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 << 1284 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & 1285 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); 1286 break; 1287 default: 1288 WARN(1, "Unexpected ASIC family %u", 1289 dev->gpu->device_info->asic_family); 1290 } 1291 1292 /* Fix errors in CZ CRAT. 1293 * simd_count: Carrizo CRAT reports wrong simd_count, probably 1294 * because it doesn't consider masked out CUs 1295 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd 1296 * capability flag: Carrizo CRAT doesn't report IOMMU flags 1297 */ 1298 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) { 1299 dev->node_props.simd_count = 1300 cu_info.simd_per_cu * cu_info.cu_active_number; 1301 dev->node_props.max_waves_per_simd = 10; 1302 dev->node_props.capability |= HSA_CAP_ATS_PRESENT; 1303 } 1304 1305 kfd_debug_print_topology(); 1306 1307 if (!res) 1308 kfd_notify_gpu_change(gpu_id, 1); 1309 err: 1310 kfd_destroy_crat_image(crat_image); 1311 return res; 1312 } 1313 1314 int kfd_topology_remove_device(struct kfd_dev *gpu) 1315 { 1316 struct kfd_topology_device *dev, *tmp; 1317 uint32_t gpu_id; 1318 int res = -ENODEV; 1319 1320 down_write(&topology_lock); 1321 1322 list_for_each_entry_safe(dev, tmp, &topology_device_list, list) 1323 if (dev->gpu == gpu) { 1324 gpu_id = dev->gpu_id; 1325 kfd_remove_sysfs_node_entry(dev); 1326 kfd_release_topology_device(dev); 1327 sys_props.num_devices--; 1328 res = 0; 1329 if (kfd_topology_update_sysfs() < 0) 1330 kfd_topology_release_sysfs(); 1331 break; 1332 } 1333 1334 up_write(&topology_lock); 1335 1336 if (!res) 1337 kfd_notify_gpu_change(gpu_id, 0); 1338 1339 return res; 1340 } 1341 1342 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD 1343 * topology. If GPU device is found @idx, then valid kfd_dev pointer is 1344 * returned through @kdev 1345 * Return - 0: On success (@kdev will be NULL for non GPU nodes) 1346 * -1: If end of list 1347 */ 1348 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev) 1349 { 1350 1351 struct kfd_topology_device *top_dev; 1352 uint8_t device_idx = 0; 1353 1354 *kdev = NULL; 1355 down_read(&topology_lock); 1356 1357 list_for_each_entry(top_dev, &topology_device_list, list) { 1358 if (device_idx == idx) { 1359 *kdev = top_dev->gpu; 1360 up_read(&topology_lock); 1361 return 0; 1362 } 1363 1364 device_idx++; 1365 } 1366 1367 up_read(&topology_lock); 1368 1369 return -1; 1370 1371 } 1372 1373 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask) 1374 { 1375 const struct cpuinfo_x86 *cpuinfo; 1376 int first_cpu_of_numa_node; 1377 1378 if (!cpumask || cpumask == cpu_none_mask) 1379 return -1; 1380 first_cpu_of_numa_node = cpumask_first(cpumask); 1381 if (first_cpu_of_numa_node >= nr_cpu_ids) 1382 return -1; 1383 cpuinfo = &cpu_data(first_cpu_of_numa_node); 1384 1385 return cpuinfo->apicid; 1386 } 1387 1388 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor 1389 * of the given NUMA node (numa_node_id) 1390 * Return -1 on failure 1391 */ 1392 int kfd_numa_node_to_apic_id(int numa_node_id) 1393 { 1394 if (numa_node_id == -1) { 1395 pr_warn("Invalid NUMA Node. Use online CPU mask\n"); 1396 return kfd_cpumask_to_apic_id(cpu_online_mask); 1397 } 1398 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id)); 1399 } 1400 1401 #if defined(CONFIG_DEBUG_FS) 1402 1403 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data) 1404 { 1405 struct kfd_topology_device *dev; 1406 unsigned int i = 0; 1407 int r = 0; 1408 1409 down_read(&topology_lock); 1410 1411 list_for_each_entry(dev, &topology_device_list, list) { 1412 if (!dev->gpu) { 1413 i++; 1414 continue; 1415 } 1416 1417 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); 1418 r = dqm_debugfs_hqds(m, dev->gpu->dqm); 1419 if (r) 1420 break; 1421 } 1422 1423 up_read(&topology_lock); 1424 1425 return r; 1426 } 1427 1428 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data) 1429 { 1430 struct kfd_topology_device *dev; 1431 unsigned int i = 0; 1432 int r = 0; 1433 1434 down_read(&topology_lock); 1435 1436 list_for_each_entry(dev, &topology_device_list, list) { 1437 if (!dev->gpu) { 1438 i++; 1439 continue; 1440 } 1441 1442 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); 1443 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets); 1444 if (r) 1445 break; 1446 } 1447 1448 up_read(&topology_lock); 1449 1450 return r; 1451 } 1452 1453 #endif 1454