1 /* 2 * Copyright 2014 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 */ 22 23 #include <linux/types.h> 24 #include <linux/kernel.h> 25 #include <linux/pci.h> 26 #include <linux/errno.h> 27 #include <linux/acpi.h> 28 #include <linux/hash.h> 29 #include <linux/cpufreq.h> 30 #include <linux/log2.h> 31 #include <linux/dmi.h> 32 #include <linux/atomic.h> 33 34 #include "kfd_priv.h" 35 #include "kfd_crat.h" 36 #include "kfd_topology.h" 37 #include "kfd_device_queue_manager.h" 38 #include "kfd_iommu.h" 39 40 /* topology_device_list - Master list of all topology devices */ 41 static struct list_head topology_device_list; 42 static struct kfd_system_properties sys_props; 43 44 static DECLARE_RWSEM(topology_lock); 45 static atomic_t topology_crat_proximity_domain; 46 47 struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 48 uint32_t proximity_domain) 49 { 50 struct kfd_topology_device *top_dev; 51 struct kfd_topology_device *device = NULL; 52 53 down_read(&topology_lock); 54 55 list_for_each_entry(top_dev, &topology_device_list, list) 56 if (top_dev->proximity_domain == proximity_domain) { 57 device = top_dev; 58 break; 59 } 60 61 up_read(&topology_lock); 62 63 return device; 64 } 65 66 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id) 67 { 68 struct kfd_topology_device *top_dev = NULL; 69 struct kfd_topology_device *ret = NULL; 70 71 down_read(&topology_lock); 72 73 list_for_each_entry(top_dev, &topology_device_list, list) 74 if (top_dev->gpu_id == gpu_id) { 75 ret = top_dev; 76 break; 77 } 78 79 up_read(&topology_lock); 80 81 return ret; 82 } 83 84 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id) 85 { 86 struct kfd_topology_device *top_dev; 87 88 top_dev = kfd_topology_device_by_id(gpu_id); 89 if (!top_dev) 90 return NULL; 91 92 return top_dev->gpu; 93 } 94 95 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev) 96 { 97 struct kfd_topology_device *top_dev; 98 struct kfd_dev *device = NULL; 99 100 down_read(&topology_lock); 101 102 list_for_each_entry(top_dev, &topology_device_list, list) 103 if (top_dev->gpu->pdev == pdev) { 104 device = top_dev->gpu; 105 break; 106 } 107 108 up_read(&topology_lock); 109 110 return device; 111 } 112 113 /* Called with write topology_lock acquired */ 114 static void kfd_release_topology_device(struct kfd_topology_device *dev) 115 { 116 struct kfd_mem_properties *mem; 117 struct kfd_cache_properties *cache; 118 struct kfd_iolink_properties *iolink; 119 struct kfd_perf_properties *perf; 120 121 list_del(&dev->list); 122 123 while (dev->mem_props.next != &dev->mem_props) { 124 mem = container_of(dev->mem_props.next, 125 struct kfd_mem_properties, list); 126 list_del(&mem->list); 127 kfree(mem); 128 } 129 130 while (dev->cache_props.next != &dev->cache_props) { 131 cache = container_of(dev->cache_props.next, 132 struct kfd_cache_properties, list); 133 list_del(&cache->list); 134 kfree(cache); 135 } 136 137 while (dev->io_link_props.next != &dev->io_link_props) { 138 iolink = container_of(dev->io_link_props.next, 139 struct kfd_iolink_properties, list); 140 list_del(&iolink->list); 141 kfree(iolink); 142 } 143 144 while (dev->perf_props.next != &dev->perf_props) { 145 perf = container_of(dev->perf_props.next, 146 struct kfd_perf_properties, list); 147 list_del(&perf->list); 148 kfree(perf); 149 } 150 151 kfree(dev); 152 } 153 154 void kfd_release_topology_device_list(struct list_head *device_list) 155 { 156 struct kfd_topology_device *dev; 157 158 while (!list_empty(device_list)) { 159 dev = list_first_entry(device_list, 160 struct kfd_topology_device, list); 161 kfd_release_topology_device(dev); 162 } 163 } 164 165 static void kfd_release_live_view(void) 166 { 167 kfd_release_topology_device_list(&topology_device_list); 168 memset(&sys_props, 0, sizeof(sys_props)); 169 } 170 171 struct kfd_topology_device *kfd_create_topology_device( 172 struct list_head *device_list) 173 { 174 struct kfd_topology_device *dev; 175 176 dev = kfd_alloc_struct(dev); 177 if (!dev) { 178 pr_err("No memory to allocate a topology device"); 179 return NULL; 180 } 181 182 INIT_LIST_HEAD(&dev->mem_props); 183 INIT_LIST_HEAD(&dev->cache_props); 184 INIT_LIST_HEAD(&dev->io_link_props); 185 INIT_LIST_HEAD(&dev->perf_props); 186 187 list_add_tail(&dev->list, device_list); 188 189 return dev; 190 } 191 192 193 #define sysfs_show_gen_prop(buffer, fmt, ...) \ 194 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__) 195 #define sysfs_show_32bit_prop(buffer, name, value) \ 196 sysfs_show_gen_prop(buffer, "%s %u\n", name, value) 197 #define sysfs_show_64bit_prop(buffer, name, value) \ 198 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value) 199 #define sysfs_show_32bit_val(buffer, value) \ 200 sysfs_show_gen_prop(buffer, "%u\n", value) 201 #define sysfs_show_str_val(buffer, value) \ 202 sysfs_show_gen_prop(buffer, "%s\n", value) 203 204 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr, 205 char *buffer) 206 { 207 ssize_t ret; 208 209 /* Making sure that the buffer is an empty string */ 210 buffer[0] = 0; 211 212 if (attr == &sys_props.attr_genid) { 213 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count); 214 } else if (attr == &sys_props.attr_props) { 215 sysfs_show_64bit_prop(buffer, "platform_oem", 216 sys_props.platform_oem); 217 sysfs_show_64bit_prop(buffer, "platform_id", 218 sys_props.platform_id); 219 ret = sysfs_show_64bit_prop(buffer, "platform_rev", 220 sys_props.platform_rev); 221 } else { 222 ret = -EINVAL; 223 } 224 225 return ret; 226 } 227 228 static void kfd_topology_kobj_release(struct kobject *kobj) 229 { 230 kfree(kobj); 231 } 232 233 static const struct sysfs_ops sysprops_ops = { 234 .show = sysprops_show, 235 }; 236 237 static struct kobj_type sysprops_type = { 238 .release = kfd_topology_kobj_release, 239 .sysfs_ops = &sysprops_ops, 240 }; 241 242 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr, 243 char *buffer) 244 { 245 ssize_t ret; 246 struct kfd_iolink_properties *iolink; 247 248 /* Making sure that the buffer is an empty string */ 249 buffer[0] = 0; 250 251 iolink = container_of(attr, struct kfd_iolink_properties, attr); 252 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type); 253 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj); 254 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min); 255 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from); 256 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to); 257 sysfs_show_32bit_prop(buffer, "weight", iolink->weight); 258 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency); 259 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency); 260 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth); 261 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth); 262 sysfs_show_32bit_prop(buffer, "recommended_transfer_size", 263 iolink->rec_transfer_size); 264 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags); 265 266 return ret; 267 } 268 269 static const struct sysfs_ops iolink_ops = { 270 .show = iolink_show, 271 }; 272 273 static struct kobj_type iolink_type = { 274 .release = kfd_topology_kobj_release, 275 .sysfs_ops = &iolink_ops, 276 }; 277 278 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr, 279 char *buffer) 280 { 281 ssize_t ret; 282 struct kfd_mem_properties *mem; 283 284 /* Making sure that the buffer is an empty string */ 285 buffer[0] = 0; 286 287 mem = container_of(attr, struct kfd_mem_properties, attr); 288 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type); 289 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes); 290 sysfs_show_32bit_prop(buffer, "flags", mem->flags); 291 sysfs_show_32bit_prop(buffer, "width", mem->width); 292 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max); 293 294 return ret; 295 } 296 297 static const struct sysfs_ops mem_ops = { 298 .show = mem_show, 299 }; 300 301 static struct kobj_type mem_type = { 302 .release = kfd_topology_kobj_release, 303 .sysfs_ops = &mem_ops, 304 }; 305 306 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr, 307 char *buffer) 308 { 309 ssize_t ret; 310 uint32_t i, j; 311 struct kfd_cache_properties *cache; 312 313 /* Making sure that the buffer is an empty string */ 314 buffer[0] = 0; 315 316 cache = container_of(attr, struct kfd_cache_properties, attr); 317 sysfs_show_32bit_prop(buffer, "processor_id_low", 318 cache->processor_id_low); 319 sysfs_show_32bit_prop(buffer, "level", cache->cache_level); 320 sysfs_show_32bit_prop(buffer, "size", cache->cache_size); 321 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size); 322 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag", 323 cache->cachelines_per_tag); 324 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc); 325 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency); 326 sysfs_show_32bit_prop(buffer, "type", cache->cache_type); 327 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer); 328 for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++) 329 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) { 330 /* Check each bit */ 331 if (cache->sibling_map[i] & (1 << j)) 332 ret = snprintf(buffer, PAGE_SIZE, 333 "%s%d%s", buffer, 1, ","); 334 else 335 ret = snprintf(buffer, PAGE_SIZE, 336 "%s%d%s", buffer, 0, ","); 337 } 338 /* Replace the last "," with end of line */ 339 *(buffer + strlen(buffer) - 1) = 0xA; 340 return ret; 341 } 342 343 static const struct sysfs_ops cache_ops = { 344 .show = kfd_cache_show, 345 }; 346 347 static struct kobj_type cache_type = { 348 .release = kfd_topology_kobj_release, 349 .sysfs_ops = &cache_ops, 350 }; 351 352 /****** Sysfs of Performance Counters ******/ 353 354 struct kfd_perf_attr { 355 struct kobj_attribute attr; 356 uint32_t data; 357 }; 358 359 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs, 360 char *buf) 361 { 362 struct kfd_perf_attr *attr; 363 364 buf[0] = 0; 365 attr = container_of(attrs, struct kfd_perf_attr, attr); 366 if (!attr->data) /* invalid data for PMC */ 367 return 0; 368 else 369 return sysfs_show_32bit_val(buf, attr->data); 370 } 371 372 #define KFD_PERF_DESC(_name, _data) \ 373 { \ 374 .attr = __ATTR(_name, 0444, perf_show, NULL), \ 375 .data = _data, \ 376 } 377 378 static struct kfd_perf_attr perf_attr_iommu[] = { 379 KFD_PERF_DESC(max_concurrent, 0), 380 KFD_PERF_DESC(num_counters, 0), 381 KFD_PERF_DESC(counter_ids, 0), 382 }; 383 /****************************************/ 384 385 static ssize_t node_show(struct kobject *kobj, struct attribute *attr, 386 char *buffer) 387 { 388 struct kfd_topology_device *dev; 389 char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE]; 390 uint32_t i; 391 uint32_t log_max_watch_addr; 392 393 /* Making sure that the buffer is an empty string */ 394 buffer[0] = 0; 395 396 if (strcmp(attr->name, "gpu_id") == 0) { 397 dev = container_of(attr, struct kfd_topology_device, 398 attr_gpuid); 399 return sysfs_show_32bit_val(buffer, dev->gpu_id); 400 } 401 402 if (strcmp(attr->name, "name") == 0) { 403 dev = container_of(attr, struct kfd_topology_device, 404 attr_name); 405 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) { 406 public_name[i] = 407 (char)dev->node_props.marketing_name[i]; 408 if (dev->node_props.marketing_name[i] == 0) 409 break; 410 } 411 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0; 412 return sysfs_show_str_val(buffer, public_name); 413 } 414 415 dev = container_of(attr, struct kfd_topology_device, 416 attr_props); 417 sysfs_show_32bit_prop(buffer, "cpu_cores_count", 418 dev->node_props.cpu_cores_count); 419 sysfs_show_32bit_prop(buffer, "simd_count", 420 dev->node_props.simd_count); 421 sysfs_show_32bit_prop(buffer, "mem_banks_count", 422 dev->node_props.mem_banks_count); 423 sysfs_show_32bit_prop(buffer, "caches_count", 424 dev->node_props.caches_count); 425 sysfs_show_32bit_prop(buffer, "io_links_count", 426 dev->node_props.io_links_count); 427 sysfs_show_32bit_prop(buffer, "cpu_core_id_base", 428 dev->node_props.cpu_core_id_base); 429 sysfs_show_32bit_prop(buffer, "simd_id_base", 430 dev->node_props.simd_id_base); 431 sysfs_show_32bit_prop(buffer, "max_waves_per_simd", 432 dev->node_props.max_waves_per_simd); 433 sysfs_show_32bit_prop(buffer, "lds_size_in_kb", 434 dev->node_props.lds_size_in_kb); 435 sysfs_show_32bit_prop(buffer, "gds_size_in_kb", 436 dev->node_props.gds_size_in_kb); 437 sysfs_show_32bit_prop(buffer, "wave_front_size", 438 dev->node_props.wave_front_size); 439 sysfs_show_32bit_prop(buffer, "array_count", 440 dev->node_props.array_count); 441 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine", 442 dev->node_props.simd_arrays_per_engine); 443 sysfs_show_32bit_prop(buffer, "cu_per_simd_array", 444 dev->node_props.cu_per_simd_array); 445 sysfs_show_32bit_prop(buffer, "simd_per_cu", 446 dev->node_props.simd_per_cu); 447 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu", 448 dev->node_props.max_slots_scratch_cu); 449 sysfs_show_32bit_prop(buffer, "vendor_id", 450 dev->node_props.vendor_id); 451 sysfs_show_32bit_prop(buffer, "device_id", 452 dev->node_props.device_id); 453 sysfs_show_32bit_prop(buffer, "location_id", 454 dev->node_props.location_id); 455 sysfs_show_32bit_prop(buffer, "drm_render_minor", 456 dev->node_props.drm_render_minor); 457 sysfs_show_64bit_prop(buffer, "hive_id", 458 dev->node_props.hive_id); 459 460 if (dev->gpu) { 461 log_max_watch_addr = 462 __ilog2_u32(dev->gpu->device_info->num_of_watch_points); 463 464 if (log_max_watch_addr) { 465 dev->node_props.capability |= 466 HSA_CAP_WATCH_POINTS_SUPPORTED; 467 468 dev->node_props.capability |= 469 ((log_max_watch_addr << 470 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) & 471 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK); 472 } 473 474 if (dev->gpu->device_info->asic_family == CHIP_TONGA) 475 dev->node_props.capability |= 476 HSA_CAP_AQL_QUEUE_DOUBLE_MAP; 477 478 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute", 479 dev->node_props.max_engine_clk_fcompute); 480 481 sysfs_show_64bit_prop(buffer, "local_mem_size", 482 (unsigned long long int) 0); 483 484 sysfs_show_32bit_prop(buffer, "fw_version", 485 dev->gpu->mec_fw_version); 486 sysfs_show_32bit_prop(buffer, "capability", 487 dev->node_props.capability); 488 sysfs_show_32bit_prop(buffer, "sdma_fw_version", 489 dev->gpu->sdma_fw_version); 490 } 491 492 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute", 493 cpufreq_quick_get_max(0)/1000); 494 } 495 496 static const struct sysfs_ops node_ops = { 497 .show = node_show, 498 }; 499 500 static struct kobj_type node_type = { 501 .release = kfd_topology_kobj_release, 502 .sysfs_ops = &node_ops, 503 }; 504 505 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr) 506 { 507 sysfs_remove_file(kobj, attr); 508 kobject_del(kobj); 509 kobject_put(kobj); 510 } 511 512 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) 513 { 514 struct kfd_iolink_properties *iolink; 515 struct kfd_cache_properties *cache; 516 struct kfd_mem_properties *mem; 517 struct kfd_perf_properties *perf; 518 519 if (dev->kobj_iolink) { 520 list_for_each_entry(iolink, &dev->io_link_props, list) 521 if (iolink->kobj) { 522 kfd_remove_sysfs_file(iolink->kobj, 523 &iolink->attr); 524 iolink->kobj = NULL; 525 } 526 kobject_del(dev->kobj_iolink); 527 kobject_put(dev->kobj_iolink); 528 dev->kobj_iolink = NULL; 529 } 530 531 if (dev->kobj_cache) { 532 list_for_each_entry(cache, &dev->cache_props, list) 533 if (cache->kobj) { 534 kfd_remove_sysfs_file(cache->kobj, 535 &cache->attr); 536 cache->kobj = NULL; 537 } 538 kobject_del(dev->kobj_cache); 539 kobject_put(dev->kobj_cache); 540 dev->kobj_cache = NULL; 541 } 542 543 if (dev->kobj_mem) { 544 list_for_each_entry(mem, &dev->mem_props, list) 545 if (mem->kobj) { 546 kfd_remove_sysfs_file(mem->kobj, &mem->attr); 547 mem->kobj = NULL; 548 } 549 kobject_del(dev->kobj_mem); 550 kobject_put(dev->kobj_mem); 551 dev->kobj_mem = NULL; 552 } 553 554 if (dev->kobj_perf) { 555 list_for_each_entry(perf, &dev->perf_props, list) { 556 kfree(perf->attr_group); 557 perf->attr_group = NULL; 558 } 559 kobject_del(dev->kobj_perf); 560 kobject_put(dev->kobj_perf); 561 dev->kobj_perf = NULL; 562 } 563 564 if (dev->kobj_node) { 565 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid); 566 sysfs_remove_file(dev->kobj_node, &dev->attr_name); 567 sysfs_remove_file(dev->kobj_node, &dev->attr_props); 568 kobject_del(dev->kobj_node); 569 kobject_put(dev->kobj_node); 570 dev->kobj_node = NULL; 571 } 572 } 573 574 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, 575 uint32_t id) 576 { 577 struct kfd_iolink_properties *iolink; 578 struct kfd_cache_properties *cache; 579 struct kfd_mem_properties *mem; 580 struct kfd_perf_properties *perf; 581 int ret; 582 uint32_t i, num_attrs; 583 struct attribute **attrs; 584 585 if (WARN_ON(dev->kobj_node)) 586 return -EEXIST; 587 588 /* 589 * Creating the sysfs folders 590 */ 591 dev->kobj_node = kfd_alloc_struct(dev->kobj_node); 592 if (!dev->kobj_node) 593 return -ENOMEM; 594 595 ret = kobject_init_and_add(dev->kobj_node, &node_type, 596 sys_props.kobj_nodes, "%d", id); 597 if (ret < 0) 598 return ret; 599 600 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node); 601 if (!dev->kobj_mem) 602 return -ENOMEM; 603 604 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node); 605 if (!dev->kobj_cache) 606 return -ENOMEM; 607 608 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node); 609 if (!dev->kobj_iolink) 610 return -ENOMEM; 611 612 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node); 613 if (!dev->kobj_perf) 614 return -ENOMEM; 615 616 /* 617 * Creating sysfs files for node properties 618 */ 619 dev->attr_gpuid.name = "gpu_id"; 620 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE; 621 sysfs_attr_init(&dev->attr_gpuid); 622 dev->attr_name.name = "name"; 623 dev->attr_name.mode = KFD_SYSFS_FILE_MODE; 624 sysfs_attr_init(&dev->attr_name); 625 dev->attr_props.name = "properties"; 626 dev->attr_props.mode = KFD_SYSFS_FILE_MODE; 627 sysfs_attr_init(&dev->attr_props); 628 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid); 629 if (ret < 0) 630 return ret; 631 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name); 632 if (ret < 0) 633 return ret; 634 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props); 635 if (ret < 0) 636 return ret; 637 638 i = 0; 639 list_for_each_entry(mem, &dev->mem_props, list) { 640 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 641 if (!mem->kobj) 642 return -ENOMEM; 643 ret = kobject_init_and_add(mem->kobj, &mem_type, 644 dev->kobj_mem, "%d", i); 645 if (ret < 0) 646 return ret; 647 648 mem->attr.name = "properties"; 649 mem->attr.mode = KFD_SYSFS_FILE_MODE; 650 sysfs_attr_init(&mem->attr); 651 ret = sysfs_create_file(mem->kobj, &mem->attr); 652 if (ret < 0) 653 return ret; 654 i++; 655 } 656 657 i = 0; 658 list_for_each_entry(cache, &dev->cache_props, list) { 659 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 660 if (!cache->kobj) 661 return -ENOMEM; 662 ret = kobject_init_and_add(cache->kobj, &cache_type, 663 dev->kobj_cache, "%d", i); 664 if (ret < 0) 665 return ret; 666 667 cache->attr.name = "properties"; 668 cache->attr.mode = KFD_SYSFS_FILE_MODE; 669 sysfs_attr_init(&cache->attr); 670 ret = sysfs_create_file(cache->kobj, &cache->attr); 671 if (ret < 0) 672 return ret; 673 i++; 674 } 675 676 i = 0; 677 list_for_each_entry(iolink, &dev->io_link_props, list) { 678 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); 679 if (!iolink->kobj) 680 return -ENOMEM; 681 ret = kobject_init_and_add(iolink->kobj, &iolink_type, 682 dev->kobj_iolink, "%d", i); 683 if (ret < 0) 684 return ret; 685 686 iolink->attr.name = "properties"; 687 iolink->attr.mode = KFD_SYSFS_FILE_MODE; 688 sysfs_attr_init(&iolink->attr); 689 ret = sysfs_create_file(iolink->kobj, &iolink->attr); 690 if (ret < 0) 691 return ret; 692 i++; 693 } 694 695 /* All hardware blocks have the same number of attributes. */ 696 num_attrs = ARRAY_SIZE(perf_attr_iommu); 697 list_for_each_entry(perf, &dev->perf_props, list) { 698 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr) 699 * num_attrs + sizeof(struct attribute_group), 700 GFP_KERNEL); 701 if (!perf->attr_group) 702 return -ENOMEM; 703 704 attrs = (struct attribute **)(perf->attr_group + 1); 705 if (!strcmp(perf->block_name, "iommu")) { 706 /* Information of IOMMU's num_counters and counter_ids is shown 707 * under /sys/bus/event_source/devices/amd_iommu. We don't 708 * duplicate here. 709 */ 710 perf_attr_iommu[0].data = perf->max_concurrent; 711 for (i = 0; i < num_attrs; i++) 712 attrs[i] = &perf_attr_iommu[i].attr.attr; 713 } 714 perf->attr_group->name = perf->block_name; 715 perf->attr_group->attrs = attrs; 716 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group); 717 if (ret < 0) 718 return ret; 719 } 720 721 return 0; 722 } 723 724 /* Called with write topology lock acquired */ 725 static int kfd_build_sysfs_node_tree(void) 726 { 727 struct kfd_topology_device *dev; 728 int ret; 729 uint32_t i = 0; 730 731 list_for_each_entry(dev, &topology_device_list, list) { 732 ret = kfd_build_sysfs_node_entry(dev, i); 733 if (ret < 0) 734 return ret; 735 i++; 736 } 737 738 return 0; 739 } 740 741 /* Called with write topology lock acquired */ 742 static void kfd_remove_sysfs_node_tree(void) 743 { 744 struct kfd_topology_device *dev; 745 746 list_for_each_entry(dev, &topology_device_list, list) 747 kfd_remove_sysfs_node_entry(dev); 748 } 749 750 static int kfd_topology_update_sysfs(void) 751 { 752 int ret; 753 754 pr_info("Creating topology SYSFS entries\n"); 755 if (!sys_props.kobj_topology) { 756 sys_props.kobj_topology = 757 kfd_alloc_struct(sys_props.kobj_topology); 758 if (!sys_props.kobj_topology) 759 return -ENOMEM; 760 761 ret = kobject_init_and_add(sys_props.kobj_topology, 762 &sysprops_type, &kfd_device->kobj, 763 "topology"); 764 if (ret < 0) 765 return ret; 766 767 sys_props.kobj_nodes = kobject_create_and_add("nodes", 768 sys_props.kobj_topology); 769 if (!sys_props.kobj_nodes) 770 return -ENOMEM; 771 772 sys_props.attr_genid.name = "generation_id"; 773 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE; 774 sysfs_attr_init(&sys_props.attr_genid); 775 ret = sysfs_create_file(sys_props.kobj_topology, 776 &sys_props.attr_genid); 777 if (ret < 0) 778 return ret; 779 780 sys_props.attr_props.name = "system_properties"; 781 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE; 782 sysfs_attr_init(&sys_props.attr_props); 783 ret = sysfs_create_file(sys_props.kobj_topology, 784 &sys_props.attr_props); 785 if (ret < 0) 786 return ret; 787 } 788 789 kfd_remove_sysfs_node_tree(); 790 791 return kfd_build_sysfs_node_tree(); 792 } 793 794 static void kfd_topology_release_sysfs(void) 795 { 796 kfd_remove_sysfs_node_tree(); 797 if (sys_props.kobj_topology) { 798 sysfs_remove_file(sys_props.kobj_topology, 799 &sys_props.attr_genid); 800 sysfs_remove_file(sys_props.kobj_topology, 801 &sys_props.attr_props); 802 if (sys_props.kobj_nodes) { 803 kobject_del(sys_props.kobj_nodes); 804 kobject_put(sys_props.kobj_nodes); 805 sys_props.kobj_nodes = NULL; 806 } 807 kobject_del(sys_props.kobj_topology); 808 kobject_put(sys_props.kobj_topology); 809 sys_props.kobj_topology = NULL; 810 } 811 } 812 813 /* Called with write topology_lock acquired */ 814 static void kfd_topology_update_device_list(struct list_head *temp_list, 815 struct list_head *master_list) 816 { 817 while (!list_empty(temp_list)) { 818 list_move_tail(temp_list->next, master_list); 819 sys_props.num_devices++; 820 } 821 } 822 823 static void kfd_debug_print_topology(void) 824 { 825 struct kfd_topology_device *dev; 826 827 down_read(&topology_lock); 828 829 dev = list_last_entry(&topology_device_list, 830 struct kfd_topology_device, list); 831 if (dev) { 832 if (dev->node_props.cpu_cores_count && 833 dev->node_props.simd_count) { 834 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n", 835 dev->node_props.device_id, 836 dev->node_props.vendor_id); 837 } else if (dev->node_props.cpu_cores_count) 838 pr_info("Topology: Add CPU node\n"); 839 else if (dev->node_props.simd_count) 840 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n", 841 dev->node_props.device_id, 842 dev->node_props.vendor_id); 843 } 844 up_read(&topology_lock); 845 } 846 847 /* Helper function for intializing platform_xx members of 848 * kfd_system_properties. Uses OEM info from the last CPU/APU node. 849 */ 850 static void kfd_update_system_properties(void) 851 { 852 struct kfd_topology_device *dev; 853 854 down_read(&topology_lock); 855 dev = list_last_entry(&topology_device_list, 856 struct kfd_topology_device, list); 857 if (dev) { 858 sys_props.platform_id = 859 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK; 860 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id); 861 sys_props.platform_rev = dev->oem_revision; 862 } 863 up_read(&topology_lock); 864 } 865 866 static void find_system_memory(const struct dmi_header *dm, 867 void *private) 868 { 869 struct kfd_mem_properties *mem; 870 u16 mem_width, mem_clock; 871 struct kfd_topology_device *kdev = 872 (struct kfd_topology_device *)private; 873 const u8 *dmi_data = (const u8 *)(dm + 1); 874 875 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) { 876 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6)); 877 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11)); 878 list_for_each_entry(mem, &kdev->mem_props, list) { 879 if (mem_width != 0xFFFF && mem_width != 0) 880 mem->width = mem_width; 881 if (mem_clock != 0) 882 mem->mem_clk_max = mem_clock; 883 } 884 } 885 } 886 887 /* 888 * Performance counters information is not part of CRAT but we would like to 889 * put them in the sysfs under topology directory for Thunk to get the data. 890 * This function is called before updating the sysfs. 891 */ 892 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev) 893 { 894 /* These are the only counters supported so far */ 895 return kfd_iommu_add_perf_counters(kdev); 896 } 897 898 /* kfd_add_non_crat_information - Add information that is not currently 899 * defined in CRAT but is necessary for KFD topology 900 * @dev - topology device to which addition info is added 901 */ 902 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev) 903 { 904 /* Check if CPU only node. */ 905 if (!kdev->gpu) { 906 /* Add system memory information */ 907 dmi_walk(find_system_memory, kdev); 908 } 909 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */ 910 } 911 912 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices. 913 * Ignore CRAT for all other devices. AMD APU is identified if both CPU 914 * and GPU cores are present. 915 * @device_list - topology device list created by parsing ACPI CRAT table. 916 * @return - TRUE if invalid, FALSE is valid. 917 */ 918 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list) 919 { 920 struct kfd_topology_device *dev; 921 922 list_for_each_entry(dev, device_list, list) { 923 if (dev->node_props.cpu_cores_count && 924 dev->node_props.simd_count) 925 return false; 926 } 927 pr_info("Ignoring ACPI CRAT on non-APU system\n"); 928 return true; 929 } 930 931 int kfd_topology_init(void) 932 { 933 void *crat_image = NULL; 934 size_t image_size = 0; 935 int ret; 936 struct list_head temp_topology_device_list; 937 int cpu_only_node = 0; 938 struct kfd_topology_device *kdev; 939 int proximity_domain; 940 941 /* topology_device_list - Master list of all topology devices 942 * temp_topology_device_list - temporary list created while parsing CRAT 943 * or VCRAT. Once parsing is complete the contents of list is moved to 944 * topology_device_list 945 */ 946 947 /* Initialize the head for the both the lists */ 948 INIT_LIST_HEAD(&topology_device_list); 949 INIT_LIST_HEAD(&temp_topology_device_list); 950 init_rwsem(&topology_lock); 951 952 memset(&sys_props, 0, sizeof(sys_props)); 953 954 /* Proximity domains in ACPI CRAT tables start counting at 955 * 0. The same should be true for virtual CRAT tables created 956 * at this stage. GPUs added later in kfd_topology_add_device 957 * use a counter. 958 */ 959 proximity_domain = 0; 960 961 /* 962 * Get the CRAT image from the ACPI. If ACPI doesn't have one 963 * or if ACPI CRAT is invalid create a virtual CRAT. 964 * NOTE: The current implementation expects all AMD APUs to have 965 * CRAT. If no CRAT is available, it is assumed to be a CPU 966 */ 967 ret = kfd_create_crat_image_acpi(&crat_image, &image_size); 968 if (!ret) { 969 ret = kfd_parse_crat_table(crat_image, 970 &temp_topology_device_list, 971 proximity_domain); 972 if (ret || 973 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) { 974 kfd_release_topology_device_list( 975 &temp_topology_device_list); 976 kfd_destroy_crat_image(crat_image); 977 crat_image = NULL; 978 } 979 } 980 981 if (!crat_image) { 982 ret = kfd_create_crat_image_virtual(&crat_image, &image_size, 983 COMPUTE_UNIT_CPU, NULL, 984 proximity_domain); 985 cpu_only_node = 1; 986 if (ret) { 987 pr_err("Error creating VCRAT table for CPU\n"); 988 return ret; 989 } 990 991 ret = kfd_parse_crat_table(crat_image, 992 &temp_topology_device_list, 993 proximity_domain); 994 if (ret) { 995 pr_err("Error parsing VCRAT table for CPU\n"); 996 goto err; 997 } 998 } 999 1000 kdev = list_first_entry(&temp_topology_device_list, 1001 struct kfd_topology_device, list); 1002 kfd_add_perf_to_topology(kdev); 1003 1004 down_write(&topology_lock); 1005 kfd_topology_update_device_list(&temp_topology_device_list, 1006 &topology_device_list); 1007 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1); 1008 ret = kfd_topology_update_sysfs(); 1009 up_write(&topology_lock); 1010 1011 if (!ret) { 1012 sys_props.generation_count++; 1013 kfd_update_system_properties(); 1014 kfd_debug_print_topology(); 1015 pr_info("Finished initializing topology\n"); 1016 } else 1017 pr_err("Failed to update topology in sysfs ret=%d\n", ret); 1018 1019 /* For nodes with GPU, this information gets added 1020 * when GPU is detected (kfd_topology_add_device). 1021 */ 1022 if (cpu_only_node) { 1023 /* Add additional information to CPU only node created above */ 1024 down_write(&topology_lock); 1025 kdev = list_first_entry(&topology_device_list, 1026 struct kfd_topology_device, list); 1027 up_write(&topology_lock); 1028 kfd_add_non_crat_information(kdev); 1029 } 1030 1031 err: 1032 kfd_destroy_crat_image(crat_image); 1033 return ret; 1034 } 1035 1036 void kfd_topology_shutdown(void) 1037 { 1038 down_write(&topology_lock); 1039 kfd_topology_release_sysfs(); 1040 kfd_release_live_view(); 1041 up_write(&topology_lock); 1042 } 1043 1044 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) 1045 { 1046 uint32_t hashout; 1047 uint32_t buf[7]; 1048 uint64_t local_mem_size; 1049 int i; 1050 struct kfd_local_mem_info local_mem_info; 1051 1052 if (!gpu) 1053 return 0; 1054 1055 gpu->kfd2kgd->get_local_mem_info(gpu->kgd, &local_mem_info); 1056 1057 local_mem_size = local_mem_info.local_mem_size_private + 1058 local_mem_info.local_mem_size_public; 1059 1060 buf[0] = gpu->pdev->devfn; 1061 buf[1] = gpu->pdev->subsystem_vendor; 1062 buf[2] = gpu->pdev->subsystem_device; 1063 buf[3] = gpu->pdev->device; 1064 buf[4] = gpu->pdev->bus->number; 1065 buf[5] = lower_32_bits(local_mem_size); 1066 buf[6] = upper_32_bits(local_mem_size); 1067 1068 for (i = 0, hashout = 0; i < 7; i++) 1069 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); 1070 1071 return hashout; 1072 } 1073 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If 1074 * the GPU device is not already present in the topology device 1075 * list then return NULL. This means a new topology device has to 1076 * be created for this GPU. 1077 * TODO: Rather than assiging @gpu to first topology device withtout 1078 * gpu attached, it will better to have more stringent check. 1079 */ 1080 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) 1081 { 1082 struct kfd_topology_device *dev; 1083 struct kfd_topology_device *out_dev = NULL; 1084 1085 down_write(&topology_lock); 1086 list_for_each_entry(dev, &topology_device_list, list) 1087 if (!dev->gpu && (dev->node_props.simd_count > 0)) { 1088 dev->gpu = gpu; 1089 out_dev = dev; 1090 break; 1091 } 1092 up_write(&topology_lock); 1093 return out_dev; 1094 } 1095 1096 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival) 1097 { 1098 /* 1099 * TODO: Generate an event for thunk about the arrival/removal 1100 * of the GPU 1101 */ 1102 } 1103 1104 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info, 1105 * patch this after CRAT parsing. 1106 */ 1107 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) 1108 { 1109 struct kfd_mem_properties *mem; 1110 struct kfd_local_mem_info local_mem_info; 1111 1112 if (!dev) 1113 return; 1114 1115 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with 1116 * single bank of VRAM local memory. 1117 * for dGPUs - VCRAT reports only one bank of Local Memory 1118 * for APUs - If CRAT from ACPI reports more than one bank, then 1119 * all the banks will report the same mem_clk_max information 1120 */ 1121 dev->gpu->kfd2kgd->get_local_mem_info(dev->gpu->kgd, 1122 &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 dev->gpu->kfd2kgd->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 dev->gpu->kfd2kgd->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