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

   1 /*
   2  * Copyright 2015-2017 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/pci.h>
  24 #include <linux/acpi.h>
  25 #include "kfd_crat.h"
  26 #include "kfd_priv.h"
  27 #include "kfd_topology.h"
  28 #include "kfd_iommu.h"
  29
  30 /* GPU Processor ID base for dGPUs for which VCRAT needs to be created.
  31  * GPU processor ID are expressed with Bit[31]=1.
  32  * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs
  33  * used in the CRAT.
  34  */
  35 static uint32_t gpu_processor_id_low = 0x80001000;
  36
  37 /* Return the next available gpu_processor_id and increment it for next GPU
  38  *      @total_cu_count - Total CUs present in the GPU including ones
  39  *                        masked off
  40  */
  41 static inline unsigned int get_and_inc_gpu_processor_id(
  42                                 unsigned int total_cu_count)
  43 {
  44         int current_id = gpu_processor_id_low;
  45
  46         gpu_processor_id_low += total_cu_count;
  47         return current_id;
  48 }
  49
  50 /* Static table to describe GPU Cache information */
  51 struct kfd_gpu_cache_info {
  52         uint32_t        cache_size;
  53         uint32_t        cache_level;
  54         uint32_t        flags;
  55         /* Indicates how many Compute Units share this cache
  56          * Value = 1 indicates the cache is not shared
  57          */
  58         uint32_t        num_cu_shared;
  59 };
  60
  61 static struct kfd_gpu_cache_info kaveri_cache_info[] = {
  62         {
  63                 /* TCP L1 Cache per CU */
  64                 .cache_size = 16,
  65                 .cache_level = 1,
  66                 .flags = (CRAT_CACHE_FLAGS_ENABLED |
  67                                 CRAT_CACHE_FLAGS_DATA_CACHE |
  68                                 CRAT_CACHE_FLAGS_SIMD_CACHE),
  69                 .num_cu_shared = 1,
  70
  71         },
  72         {
  73                 /* Scalar L1 Instruction Cache (in SQC module) per bank */
  74                 .cache_size = 16,
  75                 .cache_level = 1,
  76                 .flags = (CRAT_CACHE_FLAGS_ENABLED |
  77                                 CRAT_CACHE_FLAGS_INST_CACHE |
  78                                 CRAT_CACHE_FLAGS_SIMD_CACHE),
  79                 .num_cu_shared = 2,
  80         },
  81         {
  82                 /* Scalar L1 Data Cache (in SQC module) per bank */
  83                 .cache_size = 8,
  84                 .cache_level = 1,
  85                 .flags = (CRAT_CACHE_FLAGS_ENABLED |
  86                                 CRAT_CACHE_FLAGS_DATA_CACHE |
  87                                 CRAT_CACHE_FLAGS_SIMD_CACHE),
  88                 .num_cu_shared = 2,
  89         },
  90
  91         /* TODO: Add L2 Cache information */
  92 };
  93
  94
  95 static struct kfd_gpu_cache_info carrizo_cache_info[] = {
  96         {
  97                 /* TCP L1 Cache per CU */
  98                 .cache_size = 16,
  99                 .cache_level = 1,
 100                 .flags = (CRAT_CACHE_FLAGS_ENABLED |
 101                                 CRAT_CACHE_FLAGS_DATA_CACHE |
 102                                 CRAT_CACHE_FLAGS_SIMD_CACHE),
 103                 .num_cu_shared = 1,
 104         },
 105         {
 106                 /* Scalar L1 Instruction Cache (in SQC module) per bank */
 107                 .cache_size = 8,
 108                 .cache_level = 1,
 109                 .flags = (CRAT_CACHE_FLAGS_ENABLED |
 110                                 CRAT_CACHE_FLAGS_INST_CACHE |
 111                                 CRAT_CACHE_FLAGS_SIMD_CACHE),
 112                 .num_cu_shared = 4,
 113         },
 114         {
 115                 /* Scalar L1 Data Cache (in SQC module) per bank. */
 116                 .cache_size = 4,
 117                 .cache_level = 1,
 118                 .flags = (CRAT_CACHE_FLAGS_ENABLED |
 119                                 CRAT_CACHE_FLAGS_DATA_CACHE |
 120                                 CRAT_CACHE_FLAGS_SIMD_CACHE),
 121                 .num_cu_shared = 4,
 122         },
 123
 124         /* TODO: Add L2 Cache information */
 125 };
 126
 127 /* NOTE: In future if more information is added to struct kfd_gpu_cache_info
 128  * the following ASICs may need a separate table.
 129  */
 130 #define hawaii_cache_info kaveri_cache_info
 131 #define tonga_cache_info carrizo_cache_info
 132 #define fiji_cache_info  carrizo_cache_info
 133 #define polaris10_cache_info carrizo_cache_info
 134 #define polaris11_cache_info carrizo_cache_info
 135
 136 static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
 137                 struct crat_subtype_computeunit *cu)
 138 {
 139         dev->node_props.cpu_cores_count = cu->num_cpu_cores;
 140         dev->node_props.cpu_core_id_base = cu->processor_id_low;
 141         if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
 142                 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
 143
 144         pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
 145                         cu->processor_id_low);
 146 }
 147
 148 static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
 149                 struct crat_subtype_computeunit *cu)
 150 {
 151         dev->node_props.simd_id_base = cu->processor_id_low;
 152         dev->node_props.simd_count = cu->num_simd_cores;
 153         dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
 154         dev->node_props.max_waves_per_simd = cu->max_waves_simd;
 155         dev->node_props.wave_front_size = cu->wave_front_size;
 156         dev->node_props.array_count = cu->array_count;
 157         dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
 158         dev->node_props.simd_per_cu = cu->num_simd_per_cu;
 159         dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
 160         if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
 161                 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
 162         pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low);
 163 }
 164
 165 /* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct
 166  * topology device present in the device_list
 167  */
 168 static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
 169                                 struct list_head *device_list)
 170 {
 171         struct kfd_topology_device *dev;
 172
 173         pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
 174                         cu->proximity_domain, cu->hsa_capability);
 175         list_for_each_entry(dev, device_list, list) {
 176                 if (cu->proximity_domain == dev->proximity_domain) {
 177                         if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
 178                                 kfd_populated_cu_info_cpu(dev, cu);
 179
 180                         if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
 181                                 kfd_populated_cu_info_gpu(dev, cu);
 182                         break;
 183                 }
 184         }
 185
 186         return 0;
 187 }
 188
 189 /* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
 190  * topology device present in the device_list
 191  */
 192 static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
 193                                 struct list_head *device_list)
 194 {
 195         struct kfd_mem_properties *props;
 196         struct kfd_topology_device *dev;
 197
 198         pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
 199                         mem->proximity_domain);
 200         list_for_each_entry(dev, device_list, list) {
 201                 if (mem->proximity_domain == dev->proximity_domain) {
 202                         props = kfd_alloc_struct(props);
 203                         if (!props)
 204                                 return -ENOMEM;
 205
 206                         /* We're on GPU node */
 207                         if (dev->node_props.cpu_cores_count == 0) {
 208                                 /* APU */
 209                                 if (mem->visibility_type == 0)
 210                                         props->heap_type =
 211                                                 HSA_MEM_HEAP_TYPE_FB_PRIVATE;
 212                                 /* dGPU */
 213                                 else
 214                                         props->heap_type = mem->visibility_type;
 215                         } else
 216                                 props->heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
 217
 218                         if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
 219                                 props->flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
 220                         if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
 221                                 props->flags |= HSA_MEM_FLAGS_NON_VOLATILE;
 222
 223                         props->size_in_bytes =
 224                                 ((uint64_t)mem->length_high << 32) +
 225                                                         mem->length_low;
 226                         props->width = mem->width;
 227
 228                         dev->node_props.mem_banks_count++;
 229                         list_add_tail(&props->list, &dev->mem_props);
 230
 231                         break;
 232                 }
 233         }
 234
 235         return 0;
 236 }
 237
 238 /* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct
 239  * topology device present in the device_list
 240  */
 241 static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache,
 242                         struct list_head *device_list)
 243 {
 244         struct kfd_cache_properties *props;
 245         struct kfd_topology_device *dev;
 246         uint32_t id;
 247         uint32_t total_num_of_cu;
 248
 249         id = cache->processor_id_low;
 250
 251         pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id);
 252         list_for_each_entry(dev, device_list, list) {
 253                 total_num_of_cu = (dev->node_props.array_count *
 254                                         dev->node_props.cu_per_simd_array);
 255
 256                 /* Cache infomration in CRAT doesn't have proximity_domain
 257                  * information as it is associated with a CPU core or GPU
 258                  * Compute Unit. So map the cache using CPU core Id or SIMD
 259                  * (GPU) ID.
 260                  * TODO: This works because currently we can safely assume that
 261                  *  Compute Units are parsed before caches are parsed. In
 262                  *  future, remove this dependency
 263                  */
 264                 if ((id >= dev->node_props.cpu_core_id_base &&
 265                         id <= dev->node_props.cpu_core_id_base +
 266                                 dev->node_props.cpu_cores_count) ||
 267                         (id >= dev->node_props.simd_id_base &&
 268                         id < dev->node_props.simd_id_base +
 269                                 total_num_of_cu)) {
 270                         props = kfd_alloc_struct(props);
 271                         if (!props)
 272                                 return -ENOMEM;
 273
 274                         props->processor_id_low = id;
 275                         props->cache_level = cache->cache_level;
 276                         props->cache_size = cache->cache_size;
 277                         props->cacheline_size = cache->cache_line_size;
 278                         props->cachelines_per_tag = cache->lines_per_tag;
 279                         props->cache_assoc = cache->associativity;
 280                         props->cache_latency = cache->cache_latency;
 281                         memcpy(props->sibling_map, cache->sibling_map,
 282                                         sizeof(props->sibling_map));
 283
 284                         if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
 285                                 props->cache_type |= HSA_CACHE_TYPE_DATA;
 286                         if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
 287                                 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
 288                         if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
 289                                 props->cache_type |= HSA_CACHE_TYPE_CPU;
 290                         if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
 291                                 props->cache_type |= HSA_CACHE_TYPE_HSACU;
 292
 293                         dev->cache_count++;
 294                         dev->node_props.caches_count++;
 295                         list_add_tail(&props->list, &dev->cache_props);
 296
 297                         break;
 298                 }
 299         }
 300
 301         return 0;
 302 }
 303
 304 /* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct
 305  * topology device present in the device_list
 306  */
 307 static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
 308                                         struct list_head *device_list)
 309 {
 310         struct kfd_iolink_properties *props = NULL, *props2;
 311         struct kfd_topology_device *dev, *cpu_dev;
 312         uint32_t id_from;
 313         uint32_t id_to;
 314
 315         id_from = iolink->proximity_domain_from;
 316         id_to = iolink->proximity_domain_to;
 317
 318         pr_debug("Found IO link entry in CRAT table with id_from=%d\n",
 319                         id_from);
 320         list_for_each_entry(dev, device_list, list) {
 321                 if (id_from == dev->proximity_domain) {
 322                         props = kfd_alloc_struct(props);
 323                         if (!props)
 324                                 return -ENOMEM;
 325
 326                         props->node_from = id_from;
 327                         props->node_to = id_to;
 328                         props->ver_maj = iolink->version_major;
 329                         props->ver_min = iolink->version_minor;
 330                         props->iolink_type = iolink->io_interface_type;
 331
 332                         if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
 333                                 props->weight = 20;
 334                         else
 335                                 props->weight = node_distance(id_from, id_to);
 336
 337                         props->min_latency = iolink->minimum_latency;
 338                         props->max_latency = iolink->maximum_latency;
 339                         props->min_bandwidth = iolink->minimum_bandwidth_mbs;
 340                         props->max_bandwidth = iolink->maximum_bandwidth_mbs;
 341                         props->rec_transfer_size =
 342                                         iolink->recommended_transfer_size;
 343
 344                         dev->io_link_count++;
 345                         dev->node_props.io_links_count++;
 346                         list_add_tail(&props->list, &dev->io_link_props);
 347                         break;
 348                 }
 349         }
 350
 351         /* CPU topology is created before GPUs are detected, so CPU->GPU
 352          * links are not built at that time. If a PCIe type is discovered, it
 353          * means a GPU is detected and we are adding GPU->CPU to the topology.
 354          * At this time, also add the corresponded CPU->GPU link.
 355          */
 356         if (props && props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) {
 357                 cpu_dev = kfd_topology_device_by_proximity_domain(id_to);
 358                 if (!cpu_dev)
 359                         return -ENODEV;
 360                 /* same everything but the other direction */
 361                 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL);
 362                 props2->node_from = id_to;
 363                 props2->node_to = id_from;
 364                 props2->kobj = NULL;
 365                 cpu_dev->io_link_count++;
 366                 cpu_dev->node_props.io_links_count++;
 367                 list_add_tail(&props2->list, &cpu_dev->io_link_props);
 368         }
 369
 370         return 0;
 371 }
 372
 373 /* kfd_parse_subtype - parse subtypes and attach it to correct topology device
 374  * present in the device_list
 375  *      @sub_type_hdr - subtype section of crat_image
 376  *      @device_list - list of topology devices present in this crat_image
 377  */
 378 static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr,
 379                                 struct list_head *device_list)
 380 {
 381         struct crat_subtype_computeunit *cu;
 382         struct crat_subtype_memory *mem;
 383         struct crat_subtype_cache *cache;
 384         struct crat_subtype_iolink *iolink;
 385         int ret = 0;
 386
 387         switch (sub_type_hdr->type) {
 388         case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
 389                 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
 390                 ret = kfd_parse_subtype_cu(cu, device_list);
 391                 break;
 392         case CRAT_SUBTYPE_MEMORY_AFFINITY:
 393                 mem = (struct crat_subtype_memory *)sub_type_hdr;
 394                 ret = kfd_parse_subtype_mem(mem, device_list);
 395                 break;
 396         case CRAT_SUBTYPE_CACHE_AFFINITY:
 397                 cache = (struct crat_subtype_cache *)sub_type_hdr;
 398                 ret = kfd_parse_subtype_cache(cache, device_list);
 399                 break;
 400         case CRAT_SUBTYPE_TLB_AFFINITY:
 401                 /*
 402                  * For now, nothing to do here
 403                  */
 404                 pr_debug("Found TLB entry in CRAT table (not processing)\n");
 405                 break;
 406         case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
 407                 /*
 408                  * For now, nothing to do here
 409                  */
 410                 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n");
 411                 break;
 412         case CRAT_SUBTYPE_IOLINK_AFFINITY:
 413                 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
 414                 ret = kfd_parse_subtype_iolink(iolink, device_list);
 415                 break;
 416         default:
 417                 pr_warn("Unknown subtype %d in CRAT\n",
 418                                 sub_type_hdr->type);
 419         }
 420
 421         return ret;
 422 }
 423
 424 /* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT
 425  * create a kfd_topology_device and add in to device_list. Also parse
 426  * CRAT subtypes and attach it to appropriate kfd_topology_device
 427  *      @crat_image - input image containing CRAT
 428  *      @device_list - [OUT] list of kfd_topology_device generated after
 429  *                     parsing crat_image
 430  *      @proximity_domain - Proximity domain of the first device in the table
 431  *
 432  *      Return - 0 if successful else -ve value
 433  */
 434 int kfd_parse_crat_table(void *crat_image, struct list_head *device_list,
 435                          uint32_t proximity_domain)
 436 {
 437         struct kfd_topology_device *top_dev = NULL;
 438         struct crat_subtype_generic *sub_type_hdr;
 439         uint16_t node_id;
 440         int ret = 0;
 441         struct crat_header *crat_table = (struct crat_header *)crat_image;
 442         uint16_t num_nodes;
 443         uint32_t image_len;
 444
 445         if (!crat_image)
 446                 return -EINVAL;
 447
 448         if (!list_empty(device_list)) {
 449                 pr_warn("Error device list should be empty\n");
 450                 return -EINVAL;
 451         }
 452
 453         num_nodes = crat_table->num_domains;
 454         image_len = crat_table->length;
 455
 456         pr_info("Parsing CRAT table with %d nodes\n", num_nodes);
 457
 458         for (node_id = 0; node_id < num_nodes; node_id++) {
 459                 top_dev = kfd_create_topology_device(device_list);
 460                 if (!top_dev)
 461                         break;
 462                 top_dev->proximity_domain = proximity_domain++;
 463         }
 464
 465         if (!top_dev) {
 466                 ret = -ENOMEM;
 467                 goto err;
 468         }
 469
 470         memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH);
 471         memcpy(top_dev->oem_table_id, crat_table->oem_table_id,
 472                         CRAT_OEMTABLEID_LENGTH);
 473         top_dev->oem_revision = crat_table->oem_revision;
 474
 475         sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
 476         while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
 477                         ((char *)crat_image) + image_len) {
 478                 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
 479                         ret = kfd_parse_subtype(sub_type_hdr, device_list);
 480                         if (ret)
 481                                 break;
 482                 }
 483
 484                 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
 485                                 sub_type_hdr->length);
 486         }
 487
 488 err:
 489         if (ret)
 490                 kfd_release_topology_device_list(device_list);
 491
 492         return ret;
 493 }
 494
 495 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
 496 static int fill_in_pcache(struct crat_subtype_cache *pcache,
 497                                 struct kfd_gpu_cache_info *pcache_info,
 498                                 struct kfd_cu_info *cu_info,
 499                                 int mem_available,
 500                                 int cu_bitmask,
 501                                 int cache_type, unsigned int cu_processor_id,
 502                                 int cu_block)
 503 {
 504         unsigned int cu_sibling_map_mask;
 505         int first_active_cu;
 506
 507         /* First check if enough memory is available */
 508         if (sizeof(struct crat_subtype_cache) > mem_available)
 509                 return -ENOMEM;
 510
 511         cu_sibling_map_mask = cu_bitmask;
 512         cu_sibling_map_mask >>= cu_block;
 513         cu_sibling_map_mask &=
 514                 ((1 << pcache_info[cache_type].num_cu_shared) - 1);
 515         first_active_cu = ffs(cu_sibling_map_mask);
 516
 517         /* CU could be inactive. In case of shared cache find the first active
 518          * CU. and incase of non-shared cache check if the CU is inactive. If
 519          * inactive active skip it
 520          */
 521         if (first_active_cu) {
 522                 memset(pcache, 0, sizeof(struct crat_subtype_cache));
 523                 pcache->type = CRAT_SUBTYPE_CACHE_AFFINITY;
 524                 pcache->length = sizeof(struct crat_subtype_cache);
 525                 pcache->flags = pcache_info[cache_type].flags;
 526                 pcache->processor_id_low = cu_processor_id
 527                                          + (first_active_cu - 1);
 528                 pcache->cache_level = pcache_info[cache_type].cache_level;
 529                 pcache->cache_size = pcache_info[cache_type].cache_size;
 530
 531                 /* Sibling map is w.r.t processor_id_low, so shift out
 532                  * inactive CU
 533                  */
 534                 cu_sibling_map_mask =
 535                         cu_sibling_map_mask >> (first_active_cu - 1);
 536
 537                 pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF);
 538                 pcache->sibling_map[1] =
 539                                 (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
 540                 pcache->sibling_map[2] =
 541                                 (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
 542                 pcache->sibling_map[3] =
 543                                 (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
 544                 return 0;
 545         }
 546         return 1;
 547 }
 548
 549 /* kfd_fill_gpu_cache_info - Fill GPU cache info using kfd_gpu_cache_info
 550  * tables
 551  *
 552  *      @kdev - [IN] GPU device
 553  *      @gpu_processor_id - [IN] GPU processor ID to which these caches
 554  *                          associate
 555  *      @available_size - [IN] Amount of memory available in pcache
 556  *      @cu_info - [IN] Compute Unit info obtained from KGD
 557  *      @pcache - [OUT] memory into which cache data is to be filled in.
 558  *      @size_filled - [OUT] amount of data used up in pcache.
 559  *      @num_of_entries - [OUT] number of caches added
 560  */
 561 static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,
 562                         int gpu_processor_id,
 563                         int available_size,
 564                         struct kfd_cu_info *cu_info,
 565                         struct crat_subtype_cache *pcache,
 566                         int *size_filled,
 567                         int *num_of_entries)
 568 {
 569         struct kfd_gpu_cache_info *pcache_info;
 570         int num_of_cache_types = 0;
 571         int i, j, k;
 572         int ct = 0;
 573         int mem_available = available_size;
 574         unsigned int cu_processor_id;
 575         int ret;
 576
 577         switch (kdev->device_info->asic_family) {
 578         case CHIP_KAVERI:
 579                 pcache_info = kaveri_cache_info;
 580                 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info);
 581                 break;
 582         case CHIP_HAWAII:
 583                 pcache_info = hawaii_cache_info;
 584                 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info);
 585                 break;
 586         case CHIP_CARRIZO:
 587                 pcache_info = carrizo_cache_info;
 588                 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info);
 589                 break;
 590         case CHIP_TONGA:
 591                 pcache_info = tonga_cache_info;
 592                 num_of_cache_types = ARRAY_SIZE(tonga_cache_info);
 593                 break;
 594         case CHIP_FIJI:
 595                 pcache_info = fiji_cache_info;
 596                 num_of_cache_types = ARRAY_SIZE(fiji_cache_info);
 597                 break;
 598         case CHIP_POLARIS10:
 599                 pcache_info = polaris10_cache_info;
 600                 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info);
 601                 break;
 602         case CHIP_POLARIS11:
 603                 pcache_info = polaris11_cache_info;
 604                 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info);
 605                 break;
 606         default:
 607                 return -EINVAL;
 608         }
 609
 610         *size_filled = 0;
 611         *num_of_entries = 0;
 612
 613         /* For each type of cache listed in the kfd_gpu_cache_info table,
 614          * go through all available Compute Units.
 615          * The [i,j,k] loop will
 616          *              if kfd_gpu_cache_info.num_cu_shared = 1
 617          *                      will parse through all available CU
 618          *              If (kfd_gpu_cache_info.num_cu_shared != 1)
 619          *                      then it will consider only one CU from
 620          *                      the shared unit
 621          */
 622
 623         for (ct = 0; ct < num_of_cache_types; ct++) {
 624                 cu_processor_id = gpu_processor_id;
 625                 for (i = 0; i < cu_info->num_shader_engines; i++) {
 626                         for (j = 0; j < cu_info->num_shader_arrays_per_engine;
 627                                 j++) {
 628                                 for (k = 0; k < cu_info->num_cu_per_sh;
 629                                         k += pcache_info[ct].num_cu_shared) {
 630
 631                                         ret = fill_in_pcache(pcache,
 632                                                 pcache_info,
 633                                                 cu_info,
 634                                                 mem_available,
 635                                                 cu_info->cu_bitmap[i][j],
 636                                                 ct,
 637                                                 cu_processor_id,
 638                                                 k);
 639
 640                                         if (ret < 0)
 641                                                 break;
 642
 643                                         if (!ret) {
 644                                                 pcache++;
 645                                                 (*num_of_entries)++;
 646                                                 mem_available -=
 647                                                         sizeof(*pcache);
 648                                                 (*size_filled) +=
 649                                                         sizeof(*pcache);
 650                                         }
 651
 652                                         /* Move to next CU block */
 653                                         cu_processor_id +=
 654                                                 pcache_info[ct].num_cu_shared;
 655                                 }
 656                         }
 657                 }
 658         }
 659
 660         pr_debug("Added [%d] GPU cache entries\n", *num_of_entries);
 661
 662         return 0;
 663 }
 664
 665 /*
 666  * kfd_create_crat_image_acpi - Allocates memory for CRAT image and
 667  * copies CRAT from ACPI (if available).
 668  * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
 669  *
 670  *      @crat_image: CRAT read from ACPI. If no CRAT in ACPI then
 671  *                   crat_image will be NULL
 672  *      @size: [OUT] size of crat_image
 673  *
 674  *      Return 0 if successful else return error code
 675  */
 676 int kfd_create_crat_image_acpi(void **crat_image, size_t *size)
 677 {
 678         struct acpi_table_header *crat_table;
 679         acpi_status status;
 680         void *pcrat_image;
 681
 682         if (!crat_image)
 683                 return -EINVAL;
 684
 685         *crat_image = NULL;
 686
 687         /* Fetch the CRAT table from ACPI */
 688         status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table);
 689         if (status == AE_NOT_FOUND) {
 690                 pr_warn("CRAT table not found\n");
 691                 return -ENODATA;
 692         } else if (ACPI_FAILURE(status)) {
 693                 const char *err = acpi_format_exception(status);
 694
 695                 pr_err("CRAT table error: %s\n", err);
 696                 return -EINVAL;
 697         }
 698
 699         if (ignore_crat) {
 700                 pr_info("CRAT table disabled by module option\n");
 701                 return -ENODATA;
 702         }
 703
 704         pcrat_image = kmalloc(crat_table->length, GFP_KERNEL);
 705         if (!pcrat_image)
 706                 return -ENOMEM;
 707
 708         memcpy(pcrat_image, crat_table, crat_table->length);
 709
 710         *crat_image = pcrat_image;
 711         *size = crat_table->length;
 712
 713         return 0;
 714 }
 715
 716 /* Memory required to create Virtual CRAT.
 717  * Since there is no easy way to predict the amount of memory required, the
 718  * following amount are allocated for CPU and GPU Virtual CRAT. This is
 719  * expected to cover all known conditions. But to be safe additional check
 720  * is put in the code to ensure we don't overwrite.
 721  */
 722 #define VCRAT_SIZE_FOR_CPU      (2 * PAGE_SIZE)
 723 #define VCRAT_SIZE_FOR_GPU      (3 * PAGE_SIZE)
 724
 725 /* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node
 726  *
 727  *      @numa_node_id: CPU NUMA node id
 728  *      @avail_size: Available size in the memory
 729  *      @sub_type_hdr: Memory into which compute info will be filled in
 730  *
 731  *      Return 0 if successful else return -ve value
 732  */
 733 static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size,
 734                                 int proximity_domain,
 735                                 struct crat_subtype_computeunit *sub_type_hdr)
 736 {
 737         const struct cpumask *cpumask;
 738
 739         *avail_size -= sizeof(struct crat_subtype_computeunit);
 740         if (*avail_size < 0)
 741                 return -ENOMEM;
 742
 743         memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
 744
 745         /* Fill in subtype header data */
 746         sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
 747         sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
 748         sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
 749
 750         cpumask = cpumask_of_node(numa_node_id);
 751
 752         /* Fill in CU data */
 753         sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT;
 754         sub_type_hdr->proximity_domain = proximity_domain;
 755         sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id);
 756         if (sub_type_hdr->processor_id_low == -1)
 757                 return -EINVAL;
 758
 759         sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask);
 760
 761         return 0;
 762 }
 763
 764 /* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node
 765  *
 766  *      @numa_node_id: CPU NUMA node id
 767  *      @avail_size: Available size in the memory
 768  *      @sub_type_hdr: Memory into which compute info will be filled in
 769  *
 770  *      Return 0 if successful else return -ve value
 771  */
 772 static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size,
 773                         int proximity_domain,
 774                         struct crat_subtype_memory *sub_type_hdr)
 775 {
 776         uint64_t mem_in_bytes = 0;
 777         pg_data_t *pgdat;
 778         int zone_type;
 779
 780         *avail_size -= sizeof(struct crat_subtype_memory);
 781         if (*avail_size < 0)
 782                 return -ENOMEM;
 783
 784         memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
 785
 786         /* Fill in subtype header data */
 787         sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
 788         sub_type_hdr->length = sizeof(struct crat_subtype_memory);
 789         sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
 790
 791         /* Fill in Memory Subunit data */
 792
 793         /* Unlike si_meminfo, si_meminfo_node is not exported. So
 794          * the following lines are duplicated from si_meminfo_node
 795          * function
 796          */
 797         pgdat = NODE_DATA(numa_node_id);
 798         for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
 799                 mem_in_bytes += pgdat->node_zones[zone_type].managed_pages;
 800         mem_in_bytes <<= PAGE_SHIFT;
 801
 802         sub_type_hdr->length_low = lower_32_bits(mem_in_bytes);
 803         sub_type_hdr->length_high = upper_32_bits(mem_in_bytes);
 804         sub_type_hdr->proximity_domain = proximity_domain;
 805
 806         return 0;
 807 }
 808
 809 static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size,
 810                                 uint32_t *num_entries,
 811                                 struct crat_subtype_iolink *sub_type_hdr)
 812 {
 813         int nid;
 814         struct cpuinfo_x86 *c = &cpu_data(0);
 815         uint8_t link_type;
 816
 817         if (c->x86_vendor == X86_VENDOR_AMD)
 818                 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT;
 819         else
 820                 link_type = CRAT_IOLINK_TYPE_QPI_1_1;
 821
 822         *num_entries = 0;
 823
 824         /* Create IO links from this node to other CPU nodes */
 825         for_each_online_node(nid) {
 826                 if (nid == numa_node_id) /* node itself */
 827                         continue;
 828
 829                 *avail_size -= sizeof(struct crat_subtype_iolink);
 830                 if (*avail_size < 0)
 831                         return -ENOMEM;
 832
 833                 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
 834
 835                 /* Fill in subtype header data */
 836                 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
 837                 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
 838                 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
 839
 840                 /* Fill in IO link data */
 841                 sub_type_hdr->proximity_domain_from = numa_node_id;
 842                 sub_type_hdr->proximity_domain_to = nid;
 843                 sub_type_hdr->io_interface_type = link_type;
 844
 845                 (*num_entries)++;
 846                 sub_type_hdr++;
 847         }
 848
 849         return 0;
 850 }
 851
 852 /* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU
 853  *
 854  *      @pcrat_image: Fill in VCRAT for CPU
 855  *      @size:  [IN] allocated size of crat_image.
 856  *              [OUT] actual size of data filled in crat_image
 857  */
 858 static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size)
 859 {
 860         struct crat_header *crat_table = (struct crat_header *)pcrat_image;
 861         struct acpi_table_header *acpi_table;
 862         acpi_status status;
 863         struct crat_subtype_generic *sub_type_hdr;
 864         int avail_size = *size;
 865         int numa_node_id;
 866         uint32_t entries = 0;
 867         int ret = 0;
 868
 869         if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_CPU)
 870                 return -EINVAL;
 871
 872         /* Fill in CRAT Header.
 873          * Modify length and total_entries as subunits are added.
 874          */
 875         avail_size -= sizeof(struct crat_header);
 876         if (avail_size < 0)
 877                 return -ENOMEM;
 878
 879         memset(crat_table, 0, sizeof(struct crat_header));
 880         memcpy(&crat_table->signature, CRAT_SIGNATURE,
 881                         sizeof(crat_table->signature));
 882         crat_table->length = sizeof(struct crat_header);
 883
 884         status = acpi_get_table("DSDT", 0, &acpi_table);
 885         if (status != AE_OK)
 886                 pr_warn("DSDT table not found for OEM information\n");
 887         else {
 888                 crat_table->oem_revision = acpi_table->revision;
 889                 memcpy(crat_table->oem_id, acpi_table->oem_id,
 890                                 CRAT_OEMID_LENGTH);
 891                 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id,
 892                                 CRAT_OEMTABLEID_LENGTH);
 893         }
 894         crat_table->total_entries = 0;
 895         crat_table->num_domains = 0;
 896
 897         sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
 898
 899         for_each_online_node(numa_node_id) {
 900                 if (kfd_numa_node_to_apic_id(numa_node_id) == -1)
 901                         continue;
 902
 903                 /* Fill in Subtype: Compute Unit */
 904                 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size,
 905                         crat_table->num_domains,
 906                         (struct crat_subtype_computeunit *)sub_type_hdr);
 907                 if (ret < 0)
 908                         return ret;
 909                 crat_table->length += sub_type_hdr->length;
 910                 crat_table->total_entries++;
 911
 912                 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
 913                         sub_type_hdr->length);
 914
 915                 /* Fill in Subtype: Memory */
 916                 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size,
 917                         crat_table->num_domains,
 918                         (struct crat_subtype_memory *)sub_type_hdr);
 919                 if (ret < 0)
 920                         return ret;
 921                 crat_table->length += sub_type_hdr->length;
 922                 crat_table->total_entries++;
 923
 924                 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
 925                         sub_type_hdr->length);
 926
 927                 /* Fill in Subtype: IO Link */
 928                 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size,
 929                                 &entries,
 930                                 (struct crat_subtype_iolink *)sub_type_hdr);
 931                 if (ret < 0)
 932                         return ret;
 933                 crat_table->length += (sub_type_hdr->length * entries);
 934                 crat_table->total_entries += entries;
 935
 936                 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
 937                                 sub_type_hdr->length * entries);
 938
 939                 crat_table->num_domains++;
 940         }
 941
 942         /* TODO: Add cache Subtype for CPU.
 943          * Currently, CPU cache information is available in function
 944          * detect_cache_attributes(cpu) defined in the file
 945          * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not
 946          * exported and to get the same information the code needs to be
 947          * duplicated.
 948          */
 949
 950         *size = crat_table->length;
 951         pr_info("Virtual CRAT table created for CPU\n");
 952
 953         return 0;
 954 }
 955
 956 static int kfd_fill_gpu_memory_affinity(int *avail_size,
 957                 struct kfd_dev *kdev, uint8_t type, uint64_t size,
 958                 struct crat_subtype_memory *sub_type_hdr,
 959                 uint32_t proximity_domain,
 960                 const struct kfd_local_mem_info *local_mem_info)
 961 {
 962         *avail_size -= sizeof(struct crat_subtype_memory);
 963         if (*avail_size < 0)
 964                 return -ENOMEM;
 965
 966         memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
 967         sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
 968         sub_type_hdr->length = sizeof(struct crat_subtype_memory);
 969         sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
 970
 971         sub_type_hdr->proximity_domain = proximity_domain;
 972
 973         pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n",
 974                         type, size);
 975
 976         sub_type_hdr->length_low = lower_32_bits(size);
 977         sub_type_hdr->length_high = upper_32_bits(size);
 978
 979         sub_type_hdr->width = local_mem_info->vram_width;
 980         sub_type_hdr->visibility_type = type;
 981
 982         return 0;
 983 }
 984
 985 /* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU
 986  * to its NUMA node
 987  *      @avail_size: Available size in the memory
 988  *      @kdev - [IN] GPU device
 989  *      @sub_type_hdr: Memory into which io link info will be filled in
 990  *      @proximity_domain - proximity domain of the GPU node
 991  *
 992  *      Return 0 if successful else return -ve value
 993  */
 994 static int kfd_fill_gpu_direct_io_link(int *avail_size,
 995                         struct kfd_dev *kdev,
 996                         struct crat_subtype_iolink *sub_type_hdr,
 997                         uint32_t proximity_domain)
 998 {
 999         *avail_size -= sizeof(struct crat_subtype_iolink);
1000         if (*avail_size < 0)
1001                 return -ENOMEM;
1002
1003         memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1004
1005         /* Fill in subtype header data */
1006         sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1007         sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1008         sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1009
1010         /* Fill in IOLINK subtype.
1011          * TODO: Fill-in other fields of iolink subtype
1012          */
1013         sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS;
1014         sub_type_hdr->proximity_domain_from = proximity_domain;
1015 #ifdef CONFIG_NUMA
1016         if (kdev->pdev->dev.numa_node == NUMA_NO_NODE)
1017                 sub_type_hdr->proximity_domain_to = 0;
1018         else
1019                 sub_type_hdr->proximity_domain_to = kdev->pdev->dev.numa_node;
1020 #else
1021         sub_type_hdr->proximity_domain_to = 0;
1022 #endif
1023         return 0;
1024 }
1025
1026 /* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU
1027  *
1028  *      @pcrat_image: Fill in VCRAT for GPU
1029  *      @size:  [IN] allocated size of crat_image.
1030  *              [OUT] actual size of data filled in crat_image
1031  */
1032 static int kfd_create_vcrat_image_gpu(void *pcrat_image,
1033                                       size_t *size, struct kfd_dev *kdev,
1034                                       uint32_t proximity_domain)
1035 {
1036         struct crat_header *crat_table = (struct crat_header *)pcrat_image;
1037         struct crat_subtype_generic *sub_type_hdr;
1038         struct crat_subtype_computeunit *cu;
1039         struct kfd_cu_info cu_info;
1040         int avail_size = *size;
1041         uint32_t total_num_of_cu;
1042         int num_of_cache_entries = 0;
1043         int cache_mem_filled = 0;
1044         int ret = 0;
1045         struct kfd_local_mem_info local_mem_info;
1046
1047         if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU)
1048                 return -EINVAL;
1049
1050         /* Fill the CRAT Header.
1051          * Modify length and total_entries as subunits are added.
1052          */
1053         avail_size -= sizeof(struct crat_header);
1054         if (avail_size < 0)
1055                 return -ENOMEM;
1056
1057         memset(crat_table, 0, sizeof(struct crat_header));
1058
1059         memcpy(&crat_table->signature, CRAT_SIGNATURE,
1060                         sizeof(crat_table->signature));
1061         /* Change length as we add more subtypes*/
1062         crat_table->length = sizeof(struct crat_header);
1063         crat_table->num_domains = 1;
1064         crat_table->total_entries = 0;
1065
1066         /* Fill in Subtype: Compute Unit
1067          * First fill in the sub type header and then sub type data
1068          */
1069         avail_size -= sizeof(struct crat_subtype_computeunit);
1070         if (avail_size < 0)
1071                 return -ENOMEM;
1072
1073         sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1);
1074         memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
1075
1076         sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
1077         sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
1078         sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1079
1080         /* Fill CU subtype data */
1081         cu = (struct crat_subtype_computeunit *)sub_type_hdr;
1082         cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT;
1083         cu->proximity_domain = proximity_domain;
1084
1085         kdev->kfd2kgd->get_cu_info(kdev->kgd, &cu_info);
1086         cu->num_simd_per_cu = cu_info.simd_per_cu;
1087         cu->num_simd_cores = cu_info.simd_per_cu * cu_info.cu_active_number;
1088         cu->max_waves_simd = cu_info.max_waves_per_simd;
1089
1090         cu->wave_front_size = cu_info.wave_front_size;
1091         cu->array_count = cu_info.num_shader_arrays_per_engine *
1092                 cu_info.num_shader_engines;
1093         total_num_of_cu = (cu->array_count * cu_info.num_cu_per_sh);
1094         cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu);
1095         cu->num_cu_per_array = cu_info.num_cu_per_sh;
1096         cu->max_slots_scatch_cu = cu_info.max_scratch_slots_per_cu;
1097         cu->num_banks = cu_info.num_shader_engines;
1098         cu->lds_size_in_kb = cu_info.lds_size;
1099
1100         cu->hsa_capability = 0;
1101
1102         /* Check if this node supports IOMMU. During parsing this flag will
1103          * translate to HSA_CAP_ATS_PRESENT
1104          */
1105         if (!kfd_iommu_check_device(kdev))
1106                 cu->hsa_capability |= CRAT_CU_FLAGS_IOMMU_PRESENT;
1107
1108         crat_table->length += sub_type_hdr->length;
1109         crat_table->total_entries++;
1110
1111         /* Fill in Subtype: Memory. Only on systems with large BAR (no
1112          * private FB), report memory as public. On other systems
1113          * report the total FB size (public+private) as a single
1114          * private heap.
1115          */
1116         kdev->kfd2kgd->get_local_mem_info(kdev->kgd, &local_mem_info);
1117         sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1118                         sub_type_hdr->length);
1119
1120         if (local_mem_info.local_mem_size_private == 0)
1121                 ret = kfd_fill_gpu_memory_affinity(&avail_size,
1122                                 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC,
1123                                 local_mem_info.local_mem_size_public,
1124                                 (struct crat_subtype_memory *)sub_type_hdr,
1125                                 proximity_domain,
1126                                 &local_mem_info);
1127         else
1128                 ret = kfd_fill_gpu_memory_affinity(&avail_size,
1129                                 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE,
1130                                 local_mem_info.local_mem_size_public +
1131                                 local_mem_info.local_mem_size_private,
1132                                 (struct crat_subtype_memory *)sub_type_hdr,
1133                                 proximity_domain,
1134                                 &local_mem_info);
1135         if (ret < 0)
1136                 return ret;
1137
1138         crat_table->length += sizeof(struct crat_subtype_memory);
1139         crat_table->total_entries++;
1140
1141         /* TODO: Fill in cache information. This information is NOT readily
1142          * available in KGD
1143          */
1144         sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1145                 sub_type_hdr->length);
1146         ret = kfd_fill_gpu_cache_info(kdev, cu->processor_id_low,
1147                                 avail_size,
1148                                 &cu_info,
1149                                 (struct crat_subtype_cache *)sub_type_hdr,
1150                                 &cache_mem_filled,
1151                                 &num_of_cache_entries);
1152
1153         if (ret < 0)
1154                 return ret;
1155
1156         crat_table->length += cache_mem_filled;
1157         crat_table->total_entries += num_of_cache_entries;
1158         avail_size -= cache_mem_filled;
1159
1160         /* Fill in Subtype: IO_LINKS
1161          *  Only direct links are added here which is Link from GPU to
1162          *  to its NUMA node. Indirect links are added by userspace.
1163          */
1164         sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1165                 cache_mem_filled);
1166         ret = kfd_fill_gpu_direct_io_link(&avail_size, kdev,
1167                 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain);
1168
1169         if (ret < 0)
1170                 return ret;
1171
1172         crat_table->length += sub_type_hdr->length;
1173         crat_table->total_entries++;
1174
1175         *size = crat_table->length;
1176         pr_info("Virtual CRAT table created for GPU\n");
1177
1178         return ret;
1179 }
1180
1181 /* kfd_create_crat_image_virtual - Allocates memory for CRAT image and
1182  *              creates a Virtual CRAT (VCRAT) image
1183  *
1184  * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
1185  *
1186  *      @crat_image: VCRAT image created because ACPI does not have a
1187  *                   CRAT for this device
1188  *      @size: [OUT] size of virtual crat_image
1189  *      @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device
1190  *              COMPUTE_UNIT_GPU - Create VCRAT for GPU
1191  *              (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU
1192  *                      -- this option is not currently implemented.
1193  *                      The assumption is that all AMD APUs will have CRAT
1194  *      @kdev: Valid kfd_device required if flags contain COMPUTE_UNIT_GPU
1195  *
1196  *      Return 0 if successful else return -ve value
1197  */
1198 int kfd_create_crat_image_virtual(void **crat_image, size_t *size,
1199                                   int flags, struct kfd_dev *kdev,
1200                                   uint32_t proximity_domain)
1201 {
1202         void *pcrat_image = NULL;
1203         int ret = 0;
1204
1205         if (!crat_image)
1206                 return -EINVAL;
1207
1208         *crat_image = NULL;
1209
1210         /* Allocate one VCRAT_SIZE_FOR_CPU for CPU virtual CRAT image and
1211          * VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image. This should cover
1212          * all the current conditions. A check is put not to overwrite beyond
1213          * allocated size
1214          */
1215         switch (flags) {
1216         case COMPUTE_UNIT_CPU:
1217                 pcrat_image = kmalloc(VCRAT_SIZE_FOR_CPU, GFP_KERNEL);
1218                 if (!pcrat_image)
1219                         return -ENOMEM;
1220                 *size = VCRAT_SIZE_FOR_CPU;
1221                 ret = kfd_create_vcrat_image_cpu(pcrat_image, size);
1222                 break;
1223         case COMPUTE_UNIT_GPU:
1224                 if (!kdev)
1225                         return -EINVAL;
1226                 pcrat_image = kmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL);
1227                 if (!pcrat_image)
1228                         return -ENOMEM;
1229                 *size = VCRAT_SIZE_FOR_GPU;
1230                 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev,
1231                                                  proximity_domain);
1232                 break;
1233         case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU):
1234                 /* TODO: */
1235                 ret = -EINVAL;
1236                 pr_err("VCRAT not implemented for APU\n");
1237                 break;
1238         default:
1239                 ret = -EINVAL;
1240         }
1241
1242         if (!ret)
1243                 *crat_image = pcrat_image;
1244         else
1245                 kfree(pcrat_image);
1246
1247         return ret;
1248 }
1249
1250
1251 /* kfd_destroy_crat_image
1252  *
1253  *      @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..)
1254  *
1255  */
1256 void kfd_destroy_crat_image(void *crat_image)
1257 {
1258         kfree(crat_image);
1259 }