drivers/gpu/drm/amd/amdkfd/kfd_topology.c

   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
  32 #include "kfd_priv.h"
  33 #include "kfd_crat.h"
  34 #include "kfd_topology.h"
  35 #include "kfd_device_queue_manager.h"
  36
  37 static struct list_head topology_device_list;
  38 static int topology_crat_parsed;
  39 static struct kfd_system_properties sys_props;
  40
  41 static DECLARE_RWSEM(topology_lock);
  42
  43 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
  44 {
  45         struct kfd_topology_device *top_dev;
  46         struct kfd_dev *device = NULL;
  47
  48         down_read(&topology_lock);
  49
  50         list_for_each_entry(top_dev, &topology_device_list, list)
  51                 if (top_dev->gpu_id == gpu_id) {
  52                         device = top_dev->gpu;
  53                         break;
  54                 }
  55
  56         up_read(&topology_lock);
  57
  58         return device;
  59 }
  60
  61 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
  62 {
  63         struct kfd_topology_device *top_dev;
  64         struct kfd_dev *device = NULL;
  65
  66         down_read(&topology_lock);
  67
  68         list_for_each_entry(top_dev, &topology_device_list, list)
  69                 if (top_dev->gpu->pdev == pdev) {
  70                         device = top_dev->gpu;
  71                         break;
  72                 }
  73
  74         up_read(&topology_lock);
  75
  76         return device;
  77 }
  78
  79 static int kfd_topology_get_crat_acpi(void *crat_image, size_t *size)
  80 {
  81         struct acpi_table_header *crat_table;
  82         acpi_status status;
  83
  84         if (!size)
  85                 return -EINVAL;
  86
  87         /*
  88          * Fetch the CRAT table from ACPI
  89          */
  90         status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table);
  91         if (status == AE_NOT_FOUND) {
  92                 pr_warn("CRAT table not found\n");
  93                 return -ENODATA;
  94         } else if (ACPI_FAILURE(status)) {
  95                 const char *err = acpi_format_exception(status);
  96
  97                 pr_err("CRAT table error: %s\n", err);
  98                 return -EINVAL;
  99         }
 100
 101         if (*size >= crat_table->length && crat_image != NULL)
 102                 memcpy(crat_image, crat_table, crat_table->length);
 103
 104         *size = crat_table->length;
 105
 106         return 0;
 107 }
 108
 109 static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
 110                 struct crat_subtype_computeunit *cu)
 111 {
 112         dev->node_props.cpu_cores_count = cu->num_cpu_cores;
 113         dev->node_props.cpu_core_id_base = cu->processor_id_low;
 114         if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
 115                 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
 116
 117         pr_info("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
 118                         cu->processor_id_low);
 119 }
 120
 121 static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
 122                 struct crat_subtype_computeunit *cu)
 123 {
 124         dev->node_props.simd_id_base = cu->processor_id_low;
 125         dev->node_props.simd_count = cu->num_simd_cores;
 126         dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
 127         dev->node_props.max_waves_per_simd = cu->max_waves_simd;
 128         dev->node_props.wave_front_size = cu->wave_front_size;
 129         dev->node_props.mem_banks_count = cu->num_banks;
 130         dev->node_props.array_count = cu->num_arrays;
 131         dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
 132         dev->node_props.simd_per_cu = cu->num_simd_per_cu;
 133         dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
 134         if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
 135                 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
 136         pr_info("CU GPU: simds=%d id_base=%d\n", cu->num_simd_cores,
 137                                 cu->processor_id_low);
 138 }
 139
 140 /* kfd_parse_subtype_cu is called when the topology mutex is already acquired */
 141 static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu)
 142 {
 143         struct kfd_topology_device *dev;
 144         int i = 0;
 145
 146         pr_info("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
 147                         cu->proximity_domain, cu->hsa_capability);
 148         list_for_each_entry(dev, &topology_device_list, list) {
 149                 if (cu->proximity_domain == i) {
 150                         if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
 151                                 kfd_populated_cu_info_cpu(dev, cu);
 152
 153                         if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
 154                                 kfd_populated_cu_info_gpu(dev, cu);
 155                         break;
 156                 }
 157                 i++;
 158         }
 159
 160         return 0;
 161 }
 162
 163 /*
 164  * kfd_parse_subtype_mem is called when the topology mutex is
 165  * already acquired
 166  */
 167 static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem)
 168 {
 169         struct kfd_mem_properties *props;
 170         struct kfd_topology_device *dev;
 171         int i = 0;
 172
 173         pr_info("Found memory entry in CRAT table with proximity_domain=%d\n",
 174                         mem->promixity_domain);
 175         list_for_each_entry(dev, &topology_device_list, list) {
 176                 if (mem->promixity_domain == i) {
 177                         props = kfd_alloc_struct(props);
 178                         if (props == NULL)
 179                                 return -ENOMEM;
 180
 181                         if (dev->node_props.cpu_cores_count == 0)
 182                                 props->heap_type = HSA_MEM_HEAP_TYPE_FB_PRIVATE;
 183                         else
 184                                 props->heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
 185
 186                         if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
 187                                 props->flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
 188                         if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
 189                                 props->flags |= HSA_MEM_FLAGS_NON_VOLATILE;
 190
 191                         props->size_in_bytes =
 192                                 ((uint64_t)mem->length_high << 32) +
 193                                                         mem->length_low;
 194                         props->width = mem->width;
 195
 196                         dev->mem_bank_count++;
 197                         list_add_tail(&props->list, &dev->mem_props);
 198
 199                         break;
 200                 }
 201                 i++;
 202         }
 203
 204         return 0;
 205 }
 206
 207 /*
 208  * kfd_parse_subtype_cache is called when the topology mutex
 209  * is already acquired
 210  */
 211 static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache)
 212 {
 213         struct kfd_cache_properties *props;
 214         struct kfd_topology_device *dev;
 215         uint32_t id;
 216
 217         id = cache->processor_id_low;
 218
 219         pr_info("Found cache entry in CRAT table with processor_id=%d\n", id);
 220         list_for_each_entry(dev, &topology_device_list, list)
 221                 if (id == dev->node_props.cpu_core_id_base ||
 222                     id == dev->node_props.simd_id_base) {
 223                         props = kfd_alloc_struct(props);
 224                         if (props == NULL)
 225                                 return -ENOMEM;
 226
 227                         props->processor_id_low = id;
 228                         props->cache_level = cache->cache_level;
 229                         props->cache_size = cache->cache_size;
 230                         props->cacheline_size = cache->cache_line_size;
 231                         props->cachelines_per_tag = cache->lines_per_tag;
 232                         props->cache_assoc = cache->associativity;
 233                         props->cache_latency = cache->cache_latency;
 234
 235                         if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
 236                                 props->cache_type |= HSA_CACHE_TYPE_DATA;
 237                         if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
 238                                 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
 239                         if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
 240                                 props->cache_type |= HSA_CACHE_TYPE_CPU;
 241                         if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
 242                                 props->cache_type |= HSA_CACHE_TYPE_HSACU;
 243
 244                         dev->cache_count++;
 245                         dev->node_props.caches_count++;
 246                         list_add_tail(&props->list, &dev->cache_props);
 247
 248                         break;
 249                 }
 250
 251         return 0;
 252 }
 253
 254 /*
 255  * kfd_parse_subtype_iolink is called when the topology mutex
 256  * is already acquired
 257  */
 258 static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink)
 259 {
 260         struct kfd_iolink_properties *props;
 261         struct kfd_topology_device *dev;
 262         uint32_t i = 0;
 263         uint32_t id_from;
 264         uint32_t id_to;
 265
 266         id_from = iolink->proximity_domain_from;
 267         id_to = iolink->proximity_domain_to;
 268
 269         pr_info("Found IO link entry in CRAT table with id_from=%d\n", id_from);
 270         list_for_each_entry(dev, &topology_device_list, list) {
 271                 if (id_from == i) {
 272                         props = kfd_alloc_struct(props);
 273                         if (props == NULL)
 274                                 return -ENOMEM;
 275
 276                         props->node_from = id_from;
 277                         props->node_to = id_to;
 278                         props->ver_maj = iolink->version_major;
 279                         props->ver_min = iolink->version_minor;
 280
 281                         /*
 282                          * weight factor (derived from CDIR), currently always 1
 283                          */
 284                         props->weight = 1;
 285
 286                         props->min_latency = iolink->minimum_latency;
 287                         props->max_latency = iolink->maximum_latency;
 288                         props->min_bandwidth = iolink->minimum_bandwidth_mbs;
 289                         props->max_bandwidth = iolink->maximum_bandwidth_mbs;
 290                         props->rec_transfer_size =
 291                                         iolink->recommended_transfer_size;
 292
 293                         dev->io_link_count++;
 294                         dev->node_props.io_links_count++;
 295                         list_add_tail(&props->list, &dev->io_link_props);
 296
 297                         break;
 298                 }
 299                 i++;
 300         }
 301
 302         return 0;
 303 }
 304
 305 static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr)
 306 {
 307         struct crat_subtype_computeunit *cu;
 308         struct crat_subtype_memory *mem;
 309         struct crat_subtype_cache *cache;
 310         struct crat_subtype_iolink *iolink;
 311         int ret = 0;
 312
 313         switch (sub_type_hdr->type) {
 314         case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
 315                 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
 316                 ret = kfd_parse_subtype_cu(cu);
 317                 break;
 318         case CRAT_SUBTYPE_MEMORY_AFFINITY:
 319                 mem = (struct crat_subtype_memory *)sub_type_hdr;
 320                 ret = kfd_parse_subtype_mem(mem);
 321                 break;
 322         case CRAT_SUBTYPE_CACHE_AFFINITY:
 323                 cache = (struct crat_subtype_cache *)sub_type_hdr;
 324                 ret = kfd_parse_subtype_cache(cache);
 325                 break;
 326         case CRAT_SUBTYPE_TLB_AFFINITY:
 327                 /*
 328                  * For now, nothing to do here
 329                  */
 330                 pr_info("Found TLB entry in CRAT table (not processing)\n");
 331                 break;
 332         case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
 333                 /*
 334                  * For now, nothing to do here
 335                  */
 336                 pr_info("Found CCOMPUTE entry in CRAT table (not processing)\n");
 337                 break;
 338         case CRAT_SUBTYPE_IOLINK_AFFINITY:
 339                 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
 340                 ret = kfd_parse_subtype_iolink(iolink);
 341                 break;
 342         default:
 343                 pr_warn("Unknown subtype (%d) in CRAT\n",
 344                                 sub_type_hdr->type);
 345         }
 346
 347         return ret;
 348 }
 349
 350 static void kfd_release_topology_device(struct kfd_topology_device *dev)
 351 {
 352         struct kfd_mem_properties *mem;
 353         struct kfd_cache_properties *cache;
 354         struct kfd_iolink_properties *iolink;
 355
 356         list_del(&dev->list);
 357
 358         while (dev->mem_props.next != &dev->mem_props) {
 359                 mem = container_of(dev->mem_props.next,
 360                                 struct kfd_mem_properties, list);
 361                 list_del(&mem->list);
 362                 kfree(mem);
 363         }
 364
 365         while (dev->cache_props.next != &dev->cache_props) {
 366                 cache = container_of(dev->cache_props.next,
 367                                 struct kfd_cache_properties, list);
 368                 list_del(&cache->list);
 369                 kfree(cache);
 370         }
 371
 372         while (dev->io_link_props.next != &dev->io_link_props) {
 373                 iolink = container_of(dev->io_link_props.next,
 374                                 struct kfd_iolink_properties, list);
 375                 list_del(&iolink->list);
 376                 kfree(iolink);
 377         }
 378
 379         kfree(dev);
 380
 381         sys_props.num_devices--;
 382 }
 383
 384 static void kfd_release_live_view(void)
 385 {
 386         struct kfd_topology_device *dev;
 387
 388         while (topology_device_list.next != &topology_device_list) {
 389                 dev = container_of(topology_device_list.next,
 390                                  struct kfd_topology_device, list);
 391                 kfd_release_topology_device(dev);
 392 }
 393
 394         memset(&sys_props, 0, sizeof(sys_props));
 395 }
 396
 397 static struct kfd_topology_device *kfd_create_topology_device(void)
 398 {
 399         struct kfd_topology_device *dev;
 400
 401         dev = kfd_alloc_struct(dev);
 402         if (!dev) {
 403                 pr_err("No memory to allocate a topology device");
 404                 return NULL;
 405         }
 406
 407         INIT_LIST_HEAD(&dev->mem_props);
 408         INIT_LIST_HEAD(&dev->cache_props);
 409         INIT_LIST_HEAD(&dev->io_link_props);
 410
 411         list_add_tail(&dev->list, &topology_device_list);
 412         sys_props.num_devices++;
 413
 414         return dev;
 415 }
 416
 417 static int kfd_parse_crat_table(void *crat_image)
 418 {
 419         struct kfd_topology_device *top_dev;
 420         struct crat_subtype_generic *sub_type_hdr;
 421         uint16_t node_id;
 422         int ret;
 423         struct crat_header *crat_table = (struct crat_header *)crat_image;
 424         uint16_t num_nodes;
 425         uint32_t image_len;
 426
 427         if (!crat_image)
 428                 return -EINVAL;
 429
 430         num_nodes = crat_table->num_domains;
 431         image_len = crat_table->length;
 432
 433         pr_info("Parsing CRAT table with %d nodes\n", num_nodes);
 434
 435         for (node_id = 0; node_id < num_nodes; node_id++) {
 436                 top_dev = kfd_create_topology_device();
 437                 if (!top_dev) {
 438                         kfd_release_live_view();
 439                         return -ENOMEM;
 440                 }
 441         }
 442
 443         sys_props.platform_id =
 444                 (*((uint64_t *)crat_table->oem_id)) & CRAT_OEMID_64BIT_MASK;
 445         sys_props.platform_oem = *((uint64_t *)crat_table->oem_table_id);
 446         sys_props.platform_rev = crat_table->revision;
 447
 448         sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
 449         while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
 450                         ((char *)crat_image) + image_len) {
 451                 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
 452                         ret = kfd_parse_subtype(sub_type_hdr);
 453                         if (ret != 0) {
 454                                 kfd_release_live_view();
 455                                 return ret;
 456                         }
 457                 }
 458
 459                 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
 460                                 sub_type_hdr->length);
 461         }
 462
 463         sys_props.generation_count++;
 464         topology_crat_parsed = 1;
 465
 466         return 0;
 467 }
 468
 469
 470 #define sysfs_show_gen_prop(buffer, fmt, ...) \
 471                 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
 472 #define sysfs_show_32bit_prop(buffer, name, value) \
 473                 sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
 474 #define sysfs_show_64bit_prop(buffer, name, value) \
 475                 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
 476 #define sysfs_show_32bit_val(buffer, value) \
 477                 sysfs_show_gen_prop(buffer, "%u\n", value)
 478 #define sysfs_show_str_val(buffer, value) \
 479                 sysfs_show_gen_prop(buffer, "%s\n", value)
 480
 481 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
 482                 char *buffer)
 483 {
 484         ssize_t ret;
 485
 486         /* Making sure that the buffer is an empty string */
 487         buffer[0] = 0;
 488
 489         if (attr == &sys_props.attr_genid) {
 490                 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
 491         } else if (attr == &sys_props.attr_props) {
 492                 sysfs_show_64bit_prop(buffer, "platform_oem",
 493                                 sys_props.platform_oem);
 494                 sysfs_show_64bit_prop(buffer, "platform_id",
 495                                 sys_props.platform_id);
 496                 ret = sysfs_show_64bit_prop(buffer, "platform_rev",
 497                                 sys_props.platform_rev);
 498         } else {
 499                 ret = -EINVAL;
 500         }
 501
 502         return ret;
 503 }
 504
 505 static const struct sysfs_ops sysprops_ops = {
 506         .show = sysprops_show,
 507 };
 508
 509 static struct kobj_type sysprops_type = {
 510         .sysfs_ops = &sysprops_ops,
 511 };
 512
 513 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
 514                 char *buffer)
 515 {
 516         ssize_t ret;
 517         struct kfd_iolink_properties *iolink;
 518
 519         /* Making sure that the buffer is an empty string */
 520         buffer[0] = 0;
 521
 522         iolink = container_of(attr, struct kfd_iolink_properties, attr);
 523         sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
 524         sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
 525         sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
 526         sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
 527         sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
 528         sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
 529         sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
 530         sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
 531         sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
 532         sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
 533         sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
 534                         iolink->rec_transfer_size);
 535         ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
 536
 537         return ret;
 538 }
 539
 540 static const struct sysfs_ops iolink_ops = {
 541         .show = iolink_show,
 542 };
 543
 544 static struct kobj_type iolink_type = {
 545         .sysfs_ops = &iolink_ops,
 546 };
 547
 548 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
 549                 char *buffer)
 550 {
 551         ssize_t ret;
 552         struct kfd_mem_properties *mem;
 553
 554         /* Making sure that the buffer is an empty string */
 555         buffer[0] = 0;
 556
 557         mem = container_of(attr, struct kfd_mem_properties, attr);
 558         sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
 559         sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
 560         sysfs_show_32bit_prop(buffer, "flags", mem->flags);
 561         sysfs_show_32bit_prop(buffer, "width", mem->width);
 562         ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
 563
 564         return ret;
 565 }
 566
 567 static const struct sysfs_ops mem_ops = {
 568         .show = mem_show,
 569 };
 570
 571 static struct kobj_type mem_type = {
 572         .sysfs_ops = &mem_ops,
 573 };
 574
 575 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
 576                 char *buffer)
 577 {
 578         ssize_t ret;
 579         uint32_t i;
 580         struct kfd_cache_properties *cache;
 581
 582         /* Making sure that the buffer is an empty string */
 583         buffer[0] = 0;
 584
 585         cache = container_of(attr, struct kfd_cache_properties, attr);
 586         sysfs_show_32bit_prop(buffer, "processor_id_low",
 587                         cache->processor_id_low);
 588         sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
 589         sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
 590         sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
 591         sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
 592                         cache->cachelines_per_tag);
 593         sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
 594         sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
 595         sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
 596         snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
 597         for (i = 0; i < KFD_TOPOLOGY_CPU_SIBLINGS; i++)
 598                 ret = snprintf(buffer, PAGE_SIZE, "%s%d%s",
 599                                 buffer, cache->sibling_map[i],
 600                                 (i == KFD_TOPOLOGY_CPU_SIBLINGS-1) ?
 601                                                 "\n" : ",");
 602
 603         return ret;
 604 }
 605
 606 static const struct sysfs_ops cache_ops = {
 607         .show = kfd_cache_show,
 608 };
 609
 610 static struct kobj_type cache_type = {
 611         .sysfs_ops = &cache_ops,
 612 };
 613
 614 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
 615                 char *buffer)
 616 {
 617         struct kfd_topology_device *dev;
 618         char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
 619         uint32_t i;
 620         uint32_t log_max_watch_addr;
 621
 622         /* Making sure that the buffer is an empty string */
 623         buffer[0] = 0;
 624
 625         if (strcmp(attr->name, "gpu_id") == 0) {
 626                 dev = container_of(attr, struct kfd_topology_device,
 627                                 attr_gpuid);
 628                 return sysfs_show_32bit_val(buffer, dev->gpu_id);
 629         }
 630
 631         if (strcmp(attr->name, "name") == 0) {
 632                 dev = container_of(attr, struct kfd_topology_device,
 633                                 attr_name);
 634                 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
 635                         public_name[i] =
 636                                         (char)dev->node_props.marketing_name[i];
 637                         if (dev->node_props.marketing_name[i] == 0)
 638                                 break;
 639                 }
 640                 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
 641                 return sysfs_show_str_val(buffer, public_name);
 642         }
 643
 644         dev = container_of(attr, struct kfd_topology_device,
 645                         attr_props);
 646         sysfs_show_32bit_prop(buffer, "cpu_cores_count",
 647                         dev->node_props.cpu_cores_count);
 648         sysfs_show_32bit_prop(buffer, "simd_count",
 649                         dev->node_props.simd_count);
 650
 651         if (dev->mem_bank_count < dev->node_props.mem_banks_count) {
 652                 pr_info_once("mem_banks_count truncated from %d to %d\n",
 653                                 dev->node_props.mem_banks_count,
 654                                 dev->mem_bank_count);
 655                 sysfs_show_32bit_prop(buffer, "mem_banks_count",
 656                                 dev->mem_bank_count);
 657         } else {
 658                 sysfs_show_32bit_prop(buffer, "mem_banks_count",
 659                                 dev->node_props.mem_banks_count);
 660         }
 661
 662         sysfs_show_32bit_prop(buffer, "caches_count",
 663                         dev->node_props.caches_count);
 664         sysfs_show_32bit_prop(buffer, "io_links_count",
 665                         dev->node_props.io_links_count);
 666         sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
 667                         dev->node_props.cpu_core_id_base);
 668         sysfs_show_32bit_prop(buffer, "simd_id_base",
 669                         dev->node_props.simd_id_base);
 670         sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
 671                         dev->node_props.max_waves_per_simd);
 672         sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
 673                         dev->node_props.lds_size_in_kb);
 674         sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
 675                         dev->node_props.gds_size_in_kb);
 676         sysfs_show_32bit_prop(buffer, "wave_front_size",
 677                         dev->node_props.wave_front_size);
 678         sysfs_show_32bit_prop(buffer, "array_count",
 679                         dev->node_props.array_count);
 680         sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
 681                         dev->node_props.simd_arrays_per_engine);
 682         sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
 683                         dev->node_props.cu_per_simd_array);
 684         sysfs_show_32bit_prop(buffer, "simd_per_cu",
 685                         dev->node_props.simd_per_cu);
 686         sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
 687                         dev->node_props.max_slots_scratch_cu);
 688         sysfs_show_32bit_prop(buffer, "vendor_id",
 689                         dev->node_props.vendor_id);
 690         sysfs_show_32bit_prop(buffer, "device_id",
 691                         dev->node_props.device_id);
 692         sysfs_show_32bit_prop(buffer, "location_id",
 693                         dev->node_props.location_id);
 694
 695         if (dev->gpu) {
 696                 log_max_watch_addr =
 697                         __ilog2_u32(dev->gpu->device_info->num_of_watch_points);
 698
 699                 if (log_max_watch_addr) {
 700                         dev->node_props.capability |=
 701                                         HSA_CAP_WATCH_POINTS_SUPPORTED;
 702
 703                         dev->node_props.capability |=
 704                                 ((log_max_watch_addr <<
 705                                         HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
 706                                 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
 707                 }
 708
 709                 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
 710                         dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(
 711                                         dev->gpu->kgd));
 712
 713                 sysfs_show_64bit_prop(buffer, "local_mem_size",
 714                                 (unsigned long long int) 0);
 715
 716                 sysfs_show_32bit_prop(buffer, "fw_version",
 717                         dev->gpu->kfd2kgd->get_fw_version(
 718                                                 dev->gpu->kgd,
 719                                                 KGD_ENGINE_MEC1));
 720                 sysfs_show_32bit_prop(buffer, "capability",
 721                                 dev->node_props.capability);
 722         }
 723
 724         return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
 725                                         cpufreq_quick_get_max(0)/1000);
 726 }
 727
 728 static const struct sysfs_ops node_ops = {
 729         .show = node_show,
 730 };
 731
 732 static struct kobj_type node_type = {
 733         .sysfs_ops = &node_ops,
 734 };
 735
 736 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
 737 {
 738         sysfs_remove_file(kobj, attr);
 739         kobject_del(kobj);
 740         kobject_put(kobj);
 741 }
 742
 743 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
 744 {
 745         struct kfd_iolink_properties *iolink;
 746         struct kfd_cache_properties *cache;
 747         struct kfd_mem_properties *mem;
 748
 749         if (dev->kobj_iolink) {
 750                 list_for_each_entry(iolink, &dev->io_link_props, list)
 751                         if (iolink->kobj) {
 752                                 kfd_remove_sysfs_file(iolink->kobj,
 753                                                         &iolink->attr);
 754                                 iolink->kobj = NULL;
 755                         }
 756                 kobject_del(dev->kobj_iolink);
 757                 kobject_put(dev->kobj_iolink);
 758                 dev->kobj_iolink = NULL;
 759         }
 760
 761         if (dev->kobj_cache) {
 762                 list_for_each_entry(cache, &dev->cache_props, list)
 763                         if (cache->kobj) {
 764                                 kfd_remove_sysfs_file(cache->kobj,
 765                                                         &cache->attr);
 766                                 cache->kobj = NULL;
 767                         }
 768                 kobject_del(dev->kobj_cache);
 769                 kobject_put(dev->kobj_cache);
 770                 dev->kobj_cache = NULL;
 771         }
 772
 773         if (dev->kobj_mem) {
 774                 list_for_each_entry(mem, &dev->mem_props, list)
 775                         if (mem->kobj) {
 776                                 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
 777                                 mem->kobj = NULL;
 778                         }
 779                 kobject_del(dev->kobj_mem);
 780                 kobject_put(dev->kobj_mem);
 781                 dev->kobj_mem = NULL;
 782         }
 783
 784         if (dev->kobj_node) {
 785                 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
 786                 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
 787                 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
 788                 kobject_del(dev->kobj_node);
 789                 kobject_put(dev->kobj_node);
 790                 dev->kobj_node = NULL;
 791         }
 792 }
 793
 794 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
 795                 uint32_t id)
 796 {
 797         struct kfd_iolink_properties *iolink;
 798         struct kfd_cache_properties *cache;
 799         struct kfd_mem_properties *mem;
 800         int ret;
 801         uint32_t i;
 802
 803         if (WARN_ON(dev->kobj_node))
 804                 return -EEXIST;
 805
 806         /*
 807          * Creating the sysfs folders
 808          */
 809         dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
 810         if (!dev->kobj_node)
 811                 return -ENOMEM;
 812
 813         ret = kobject_init_and_add(dev->kobj_node, &node_type,
 814                         sys_props.kobj_nodes, "%d", id);
 815         if (ret < 0)
 816                 return ret;
 817
 818         dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
 819         if (!dev->kobj_mem)
 820                 return -ENOMEM;
 821
 822         dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
 823         if (!dev->kobj_cache)
 824                 return -ENOMEM;
 825
 826         dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
 827         if (!dev->kobj_iolink)
 828                 return -ENOMEM;
 829
 830         /*
 831          * Creating sysfs files for node properties
 832          */
 833         dev->attr_gpuid.name = "gpu_id";
 834         dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
 835         sysfs_attr_init(&dev->attr_gpuid);
 836         dev->attr_name.name = "name";
 837         dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
 838         sysfs_attr_init(&dev->attr_name);
 839         dev->attr_props.name = "properties";
 840         dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
 841         sysfs_attr_init(&dev->attr_props);
 842         ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
 843         if (ret < 0)
 844                 return ret;
 845         ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
 846         if (ret < 0)
 847                 return ret;
 848         ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
 849         if (ret < 0)
 850                 return ret;
 851
 852         i = 0;
 853         list_for_each_entry(mem, &dev->mem_props, list) {
 854                 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 855                 if (!mem->kobj)
 856                         return -ENOMEM;
 857                 ret = kobject_init_and_add(mem->kobj, &mem_type,
 858                                 dev->kobj_mem, "%d", i);
 859                 if (ret < 0)
 860                         return ret;
 861
 862                 mem->attr.name = "properties";
 863                 mem->attr.mode = KFD_SYSFS_FILE_MODE;
 864                 sysfs_attr_init(&mem->attr);
 865                 ret = sysfs_create_file(mem->kobj, &mem->attr);
 866                 if (ret < 0)
 867                         return ret;
 868                 i++;
 869         }
 870
 871         i = 0;
 872         list_for_each_entry(cache, &dev->cache_props, list) {
 873                 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 874                 if (!cache->kobj)
 875                         return -ENOMEM;
 876                 ret = kobject_init_and_add(cache->kobj, &cache_type,
 877                                 dev->kobj_cache, "%d", i);
 878                 if (ret < 0)
 879                         return ret;
 880
 881                 cache->attr.name = "properties";
 882                 cache->attr.mode = KFD_SYSFS_FILE_MODE;
 883                 sysfs_attr_init(&cache->attr);
 884                 ret = sysfs_create_file(cache->kobj, &cache->attr);
 885                 if (ret < 0)
 886                         return ret;
 887                 i++;
 888         }
 889
 890         i = 0;
 891         list_for_each_entry(iolink, &dev->io_link_props, list) {
 892                 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 893                 if (!iolink->kobj)
 894                         return -ENOMEM;
 895                 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
 896                                 dev->kobj_iolink, "%d", i);
 897                 if (ret < 0)
 898                         return ret;
 899
 900                 iolink->attr.name = "properties";
 901                 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
 902                 sysfs_attr_init(&iolink->attr);
 903                 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
 904                 if (ret < 0)
 905                         return ret;
 906                 i++;
 907 }
 908
 909         return 0;
 910 }
 911
 912 static int kfd_build_sysfs_node_tree(void)
 913 {
 914         struct kfd_topology_device *dev;
 915         int ret;
 916         uint32_t i = 0;
 917
 918         list_for_each_entry(dev, &topology_device_list, list) {
 919                 ret = kfd_build_sysfs_node_entry(dev, i);
 920                 if (ret < 0)
 921                         return ret;
 922                 i++;
 923         }
 924
 925         return 0;
 926 }
 927
 928 static void kfd_remove_sysfs_node_tree(void)
 929 {
 930         struct kfd_topology_device *dev;
 931
 932         list_for_each_entry(dev, &topology_device_list, list)
 933                 kfd_remove_sysfs_node_entry(dev);
 934 }
 935
 936 static int kfd_topology_update_sysfs(void)
 937 {
 938         int ret;
 939
 940         pr_info("Creating topology SYSFS entries\n");
 941         if (!sys_props.kobj_topology) {
 942                 sys_props.kobj_topology =
 943                                 kfd_alloc_struct(sys_props.kobj_topology);
 944                 if (!sys_props.kobj_topology)
 945                         return -ENOMEM;
 946
 947                 ret = kobject_init_and_add(sys_props.kobj_topology,
 948                                 &sysprops_type,  &kfd_device->kobj,
 949                                 "topology");
 950                 if (ret < 0)
 951                         return ret;
 952
 953                 sys_props.kobj_nodes = kobject_create_and_add("nodes",
 954                                 sys_props.kobj_topology);
 955                 if (!sys_props.kobj_nodes)
 956                         return -ENOMEM;
 957
 958                 sys_props.attr_genid.name = "generation_id";
 959                 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
 960                 sysfs_attr_init(&sys_props.attr_genid);
 961                 ret = sysfs_create_file(sys_props.kobj_topology,
 962                                 &sys_props.attr_genid);
 963                 if (ret < 0)
 964                         return ret;
 965
 966                 sys_props.attr_props.name = "system_properties";
 967                 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
 968                 sysfs_attr_init(&sys_props.attr_props);
 969                 ret = sysfs_create_file(sys_props.kobj_topology,
 970                                 &sys_props.attr_props);
 971                 if (ret < 0)
 972                         return ret;
 973         }
 974
 975         kfd_remove_sysfs_node_tree();
 976
 977         return kfd_build_sysfs_node_tree();
 978 }
 979
 980 static void kfd_topology_release_sysfs(void)
 981 {
 982         kfd_remove_sysfs_node_tree();
 983         if (sys_props.kobj_topology) {
 984                 sysfs_remove_file(sys_props.kobj_topology,
 985                                 &sys_props.attr_genid);
 986                 sysfs_remove_file(sys_props.kobj_topology,
 987                                 &sys_props.attr_props);
 988                 if (sys_props.kobj_nodes) {
 989                         kobject_del(sys_props.kobj_nodes);
 990                         kobject_put(sys_props.kobj_nodes);
 991                         sys_props.kobj_nodes = NULL;
 992                 }
 993                 kobject_del(sys_props.kobj_topology);
 994                 kobject_put(sys_props.kobj_topology);
 995                 sys_props.kobj_topology = NULL;
 996         }
 997 }
 998
 999 int kfd_topology_init(void)
1000 {
1001         void *crat_image = NULL;
1002         size_t image_size = 0;
1003         int ret;
1004
1005         /*
1006          * Initialize the head for the topology device list
1007          */
1008         INIT_LIST_HEAD(&topology_device_list);
1009         init_rwsem(&topology_lock);
1010         topology_crat_parsed = 0;
1011
1012         memset(&sys_props, 0, sizeof(sys_props));
1013
1014         /*
1015          * Get the CRAT image from the ACPI
1016          */
1017         ret = kfd_topology_get_crat_acpi(crat_image, &image_size);
1018         if (ret == 0 && image_size > 0) {
1019                 pr_info("Found CRAT image with size=%zd\n", image_size);
1020                 crat_image = kmalloc(image_size, GFP_KERNEL);
1021                 if (!crat_image) {
1022                         ret = -ENOMEM;
1023                         pr_err("No memory for allocating CRAT image\n");
1024                         goto err;
1025                 }
1026                 ret = kfd_topology_get_crat_acpi(crat_image, &image_size);
1027
1028                 if (ret == 0) {
1029                         down_write(&topology_lock);
1030                         ret = kfd_parse_crat_table(crat_image);
1031                         if (ret == 0)
1032                                 ret = kfd_topology_update_sysfs();
1033                         up_write(&topology_lock);
1034                 } else {
1035                         pr_err("Couldn't get CRAT table size from ACPI\n");
1036                 }
1037                 kfree(crat_image);
1038         } else if (ret == -ENODATA) {
1039                 ret = 0;
1040         } else {
1041                 pr_err("Couldn't get CRAT table size from ACPI\n");
1042         }
1043
1044 err:
1045         pr_info("Finished initializing topology ret=%d\n", ret);
1046         return ret;
1047 }
1048
1049 void kfd_topology_shutdown(void)
1050 {
1051         kfd_topology_release_sysfs();
1052         kfd_release_live_view();
1053 }
1054
1055 static void kfd_debug_print_topology(void)
1056 {
1057         struct kfd_topology_device *dev;
1058         uint32_t i = 0;
1059
1060         pr_info("DEBUG PRINT OF TOPOLOGY:");
1061         list_for_each_entry(dev, &topology_device_list, list) {
1062                 pr_info("Node: %d\n", i);
1063                 pr_info("\tGPU assigned: %s\n", (dev->gpu ? "yes" : "no"));
1064                 pr_info("\tCPU count: %d\n", dev->node_props.cpu_cores_count);
1065                 pr_info("\tSIMD count: %d", dev->node_props.simd_count);
1066                 i++;
1067         }
1068 }
1069
1070 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1071 {
1072         uint32_t hashout;
1073         uint32_t buf[7];
1074         uint64_t local_mem_size;
1075         int i;
1076
1077         if (!gpu)
1078                 return 0;
1079
1080         local_mem_size = gpu->kfd2kgd->get_vmem_size(gpu->kgd);
1081
1082         buf[0] = gpu->pdev->devfn;
1083         buf[1] = gpu->pdev->subsystem_vendor;
1084         buf[2] = gpu->pdev->subsystem_device;
1085         buf[3] = gpu->pdev->device;
1086         buf[4] = gpu->pdev->bus->number;
1087         buf[5] = lower_32_bits(local_mem_size);
1088         buf[6] = upper_32_bits(local_mem_size);
1089
1090         for (i = 0, hashout = 0; i < 7; i++)
1091                 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1092
1093         return hashout;
1094 }
1095
1096 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1097 {
1098         struct kfd_topology_device *dev;
1099         struct kfd_topology_device *out_dev = NULL;
1100
1101         list_for_each_entry(dev, &topology_device_list, list)
1102                 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1103                         dev->gpu = gpu;
1104                         out_dev = dev;
1105                         break;
1106                 }
1107
1108         return out_dev;
1109 }
1110
1111 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1112 {
1113         /*
1114          * TODO: Generate an event for thunk about the arrival/removal
1115          * of the GPU
1116          */
1117 }
1118
1119 int kfd_topology_add_device(struct kfd_dev *gpu)
1120 {
1121         uint32_t gpu_id;
1122         struct kfd_topology_device *dev;
1123         int res;
1124
1125         gpu_id = kfd_generate_gpu_id(gpu);
1126
1127         pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1128
1129         down_write(&topology_lock);
1130         /*
1131          * Try to assign the GPU to existing topology device (generated from
1132          * CRAT table
1133          */
1134         dev = kfd_assign_gpu(gpu);
1135         if (!dev) {
1136                 pr_info("GPU was not found in the current topology. Extending.\n");
1137                 kfd_debug_print_topology();
1138                 dev = kfd_create_topology_device();
1139                 if (!dev) {
1140                         res = -ENOMEM;
1141                         goto err;
1142                 }
1143                 dev->gpu = gpu;
1144
1145                 /*
1146                  * TODO: Make a call to retrieve topology information from the
1147                  * GPU vBIOS
1148                  */
1149
1150                 /* Update the SYSFS tree, since we added another topology
1151                  * device
1152                  */
1153                 if (kfd_topology_update_sysfs() < 0)
1154                         kfd_topology_release_sysfs();
1155
1156         }
1157
1158         dev->gpu_id = gpu_id;
1159         gpu->id = gpu_id;
1160         dev->node_props.vendor_id = gpu->pdev->vendor;
1161         dev->node_props.device_id = gpu->pdev->device;
1162         dev->node_props.location_id = (gpu->pdev->bus->number << 24) +
1163                         (gpu->pdev->devfn & 0xffffff);
1164         /*
1165          * TODO: Retrieve max engine clock values from KGD
1166          */
1167
1168         if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1169                 dev->node_props.capability |= HSA_CAP_DOORBELL_PACKET_TYPE;
1170                 pr_info("Adding doorbell packet type capability\n");
1171         }
1172
1173         res = 0;
1174
1175 err:
1176         up_write(&topology_lock);
1177
1178         if (res == 0)
1179                 kfd_notify_gpu_change(gpu_id, 1);
1180
1181         return res;
1182 }
1183
1184 int kfd_topology_remove_device(struct kfd_dev *gpu)
1185 {
1186         struct kfd_topology_device *dev;
1187         uint32_t gpu_id;
1188         int res = -ENODEV;
1189
1190         down_write(&topology_lock);
1191
1192         list_for_each_entry(dev, &topology_device_list, list)
1193                 if (dev->gpu == gpu) {
1194                         gpu_id = dev->gpu_id;
1195                         kfd_remove_sysfs_node_entry(dev);
1196                         kfd_release_topology_device(dev);
1197                         res = 0;
1198                         if (kfd_topology_update_sysfs() < 0)
1199                                 kfd_topology_release_sysfs();
1200                         break;
1201                 }
1202
1203         up_write(&topology_lock);
1204
1205         if (res == 0)
1206                 kfd_notify_gpu_change(gpu_id, 0);
1207
1208         return res;
1209 }
1210
1211 /*
1212  * When idx is out of bounds, the function will return NULL
1213  */
1214 struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx)
1215 {
1216
1217         struct kfd_topology_device *top_dev;
1218         struct kfd_dev *device = NULL;
1219         uint8_t device_idx = 0;
1220
1221         down_read(&topology_lock);
1222
1223         list_for_each_entry(top_dev, &topology_device_list, list) {
1224                 if (device_idx == idx) {
1225                         device = top_dev->gpu;
1226                         break;
1227                 }
1228
1229                 device_idx++;
1230         }
1231
1232         up_read(&topology_lock);
1233
1234         return device;
1235
1236 }
1237
1238 #if defined(CONFIG_DEBUG_FS)
1239
1240 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1241 {
1242         struct kfd_topology_device *dev;
1243         unsigned int i = 0;
1244         int r = 0;
1245
1246         down_read(&topology_lock);
1247
1248         list_for_each_entry(dev, &topology_device_list, list) {
1249                 if (!dev->gpu) {
1250                         i++;
1251                         continue;
1252                 }
1253
1254                 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1255                 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1256                 if (r)
1257                         break;
1258         }
1259
1260         up_read(&topology_lock);
1261
1262         return r;
1263 }
1264
1265 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1266 {
1267         struct kfd_topology_device *dev;
1268         unsigned int i = 0;
1269         int r = 0;
1270
1271         down_read(&topology_lock);
1272
1273         list_for_each_entry(dev, &topology_device_list, list) {
1274                 if (!dev->gpu) {
1275                         i++;
1276                         continue;
1277                 }
1278
1279                 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1280                 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets);
1281                 if (r)
1282                         break;
1283         }
1284
1285         up_read(&topology_lock);
1286
1287         return r;
1288 }
1289
1290 #endif