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 /* TODO - check & update Vega10 cache details */
 136 #define vega10_cache_info carrizo_cache_info
 137 #define raven_cache_info carrizo_cache_info
 138
 139 static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
 140                 struct crat_subtype_computeunit *cu)
 141 {
 142         dev->node_props.cpu_cores_count = cu->num_cpu_cores;
 143         dev->node_props.cpu_core_id_base = cu->processor_id_low;
 144         if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
 145                 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
 146
 147         pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
 148                         cu->processor_id_low);
 149 }
 150
 151 static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
 152                 struct crat_subtype_computeunit *cu)
 153 {
 154         dev->node_props.simd_id_base = cu->processor_id_low;
 155         dev->node_props.simd_count = cu->num_simd_cores;
 156         dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
 157         dev->node_props.max_waves_per_simd = cu->max_waves_simd;
 158         dev->node_props.wave_front_size = cu->wave_front_size;
 159         dev->node_props.array_count = cu->array_count;
 160         dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
 161         dev->node_props.simd_per_cu = cu->num_simd_per_cu;
 162         dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
 163         if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
 164                 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
 165         pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low);
 166 }
 167
 168 /* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct
 169  * topology device present in the device_list
 170  */
 171 static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
 172                                 struct list_head *device_list)
 173 {
 174         struct kfd_topology_device *dev;
 175
 176         pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
 177                         cu->proximity_domain, cu->hsa_capability);
 178         list_for_each_entry(dev, device_list, list) {
 179                 if (cu->proximity_domain == dev->proximity_domain) {
 180                         if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
 181                                 kfd_populated_cu_info_cpu(dev, cu);
 182
 183                         if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
 184                                 kfd_populated_cu_info_gpu(dev, cu);
 185                         break;
 186                 }
 187         }
 188
 189         return 0;
 190 }
 191
 192 static struct kfd_mem_properties *
 193 find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width,
 194                 struct kfd_topology_device *dev)
 195 {
 196         struct kfd_mem_properties *props;
 197
 198         list_for_each_entry(props, &dev->mem_props, list) {
 199                 if (props->heap_type == heap_type
 200                                 && props->flags == flags
 201                                 && props->width == width)
 202                         return props;
 203         }
 204
 205         return NULL;
 206 }
 207 /* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
 208  * topology device present in the device_list
 209  */
 210 static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
 211                                 struct list_head *device_list)
 212 {
 213         struct kfd_mem_properties *props;
 214         struct kfd_topology_device *dev;
 215         uint32_t heap_type;
 216         uint64_t size_in_bytes;
 217         uint32_t flags = 0;
 218         uint32_t width;
 219
 220         pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
 221                         mem->proximity_domain);
 222         list_for_each_entry(dev, device_list, list) {
 223                 if (mem->proximity_domain == dev->proximity_domain) {
 224                         /* We're on GPU node */
 225                         if (dev->node_props.cpu_cores_count == 0) {
 226                                 /* APU */
 227                                 if (mem->visibility_type == 0)
 228                                         heap_type =
 229                                                 HSA_MEM_HEAP_TYPE_FB_PRIVATE;
 230                                 /* dGPU */
 231                                 else
 232                                         heap_type = mem->visibility_type;
 233                         } else
 234                                 heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
 235
 236                         if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
 237                                 flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
 238                         if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
 239                                 flags |= HSA_MEM_FLAGS_NON_VOLATILE;
 240
 241                         size_in_bytes =
 242                                 ((uint64_t)mem->length_high << 32) +
 243                                                         mem->length_low;
 244                         width = mem->width;
 245
 246                         /* Multiple banks of the same type are aggregated into
 247                          * one. User mode doesn't care about multiple physical
 248                          * memory segments. It's managed as a single virtual
 249                          * heap for user mode.
 250                          */
 251                         props = find_subtype_mem(heap_type, flags, width, dev);
 252                         if (props) {
 253                                 props->size_in_bytes += size_in_bytes;
 254                                 break;
 255                         }
 256
 257                         props = kfd_alloc_struct(props);
 258                         if (!props)
 259                                 return -ENOMEM;
 260
 261                         props->heap_type = heap_type;
 262                         props->flags = flags;
 263                         props->size_in_bytes = size_in_bytes;
 264                         props->width = width;
 265
 266                         dev->node_props.mem_banks_count++;
 267                         list_add_tail(&props->list, &dev->mem_props);
 268
 269                         break;
 270                 }
 271         }
 272
 273         return 0;
 274 }
 275
 276 /* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct
 277  * topology device present in the device_list
 278  */
 279 static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache,
 280                         struct list_head *device_list)
 281 {
 282         struct kfd_cache_properties *props;
 283         struct kfd_topology_device *dev;
 284         uint32_t id;
 285         uint32_t total_num_of_cu;
 286
 287         id = cache->processor_id_low;
 288
 289         pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id);
 290         list_for_each_entry(dev, device_list, list) {
 291                 total_num_of_cu = (dev->node_props.array_count *
 292                                         dev->node_props.cu_per_simd_array);
 293
 294                 /* Cache infomration in CRAT doesn't have proximity_domain
 295                  * information as it is associated with a CPU core or GPU
 296                  * Compute Unit. So map the cache using CPU core Id or SIMD
 297                  * (GPU) ID.
 298                  * TODO: This works because currently we can safely assume that
 299                  *  Compute Units are parsed before caches are parsed. In
 300                  *  future, remove this dependency
 301                  */
 302                 if ((id >= dev->node_props.cpu_core_id_base &&
 303                         id <= dev->node_props.cpu_core_id_base +
 304                                 dev->node_props.cpu_cores_count) ||
 305                         (id >= dev->node_props.simd_id_base &&
 306                         id < dev->node_props.simd_id_base +
 307                                 total_num_of_cu)) {
 308                         props = kfd_alloc_struct(props);
 309                         if (!props)
 310                                 return -ENOMEM;
 311
 312                         props->processor_id_low = id;
 313                         props->cache_level = cache->cache_level;
 314                         props->cache_size = cache->cache_size;
 315                         props->cacheline_size = cache->cache_line_size;
 316                         props->cachelines_per_tag = cache->lines_per_tag;
 317                         props->cache_assoc = cache->associativity;
 318                         props->cache_latency = cache->cache_latency;
 319                         memcpy(props->sibling_map, cache->sibling_map,
 320                                         sizeof(props->sibling_map));
 321
 322                         if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
 323                                 props->cache_type |= HSA_CACHE_TYPE_DATA;
 324                         if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
 325                                 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
 326                         if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
 327                                 props->cache_type |= HSA_CACHE_TYPE_CPU;
 328                         if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
 329                                 props->cache_type |= HSA_CACHE_TYPE_HSACU;
 330
 331                         dev->cache_count++;
 332                         dev->node_props.caches_count++;
 333                         list_add_tail(&props->list, &dev->cache_props);
 334
 335                         break;
 336                 }
 337         }
 338
 339         return 0;
 340 }
 341
 342 /* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct
 343  * topology device present in the device_list
 344  */
 345 static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
 346                                         struct list_head *device_list)
 347 {
 348         struct kfd_iolink_properties *props = NULL, *props2;
 349         struct kfd_topology_device *dev, *cpu_dev;
 350         uint32_t id_from;
 351         uint32_t id_to;
 352
 353         id_from = iolink->proximity_domain_from;
 354         id_to = iolink->proximity_domain_to;
 355
 356         pr_debug("Found IO link entry in CRAT table with id_from=%d\n",
 357                         id_from);
 358         list_for_each_entry(dev, device_list, list) {
 359                 if (id_from == dev->proximity_domain) {
 360                         props = kfd_alloc_struct(props);
 361                         if (!props)
 362                                 return -ENOMEM;
 363
 364                         props->node_from = id_from;
 365                         props->node_to = id_to;
 366                         props->ver_maj = iolink->version_major;
 367                         props->ver_min = iolink->version_minor;
 368                         props->iolink_type = iolink->io_interface_type;
 369
 370                         if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
 371                                 props->weight = 20;
 372                         else
 373                                 props->weight = node_distance(id_from, id_to);
 374
 375                         props->min_latency = iolink->minimum_latency;
 376                         props->max_latency = iolink->maximum_latency;
 377                         props->min_bandwidth = iolink->minimum_bandwidth_mbs;
 378                         props->max_bandwidth = iolink->maximum_bandwidth_mbs;
 379                         props->rec_transfer_size =
 380                                         iolink->recommended_transfer_size;
 381
 382                         dev->io_link_count++;
 383                         dev->node_props.io_links_count++;
 384                         list_add_tail(&props->list, &dev->io_link_props);
 385                         break;
 386                 }
 387         }
 388
 389         /* CPU topology is created before GPUs are detected, so CPU->GPU
 390          * links are not built at that time. If a PCIe type is discovered, it
 391          * means a GPU is detected and we are adding GPU->CPU to the topology.
 392          * At this time, also add the corresponded CPU->GPU link.
 393          */
 394         if (props && props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) {
 395                 cpu_dev = kfd_topology_device_by_proximity_domain(id_to);
 396                 if (!cpu_dev)
 397                         return -ENODEV;
 398                 /* same everything but the other direction */
 399                 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL);
 400                 props2->node_from = id_to;
 401                 props2->node_to = id_from;
 402                 props2->kobj = NULL;
 403                 cpu_dev->io_link_count++;
 404                 cpu_dev->node_props.io_links_count++;
 405                 list_add_tail(&props2->list, &cpu_dev->io_link_props);
 406         }
 407
 408         return 0;
 409 }
 410
 411 /* kfd_parse_subtype - parse subtypes and attach it to correct topology device
 412  * present in the device_list
 413  *      @sub_type_hdr - subtype section of crat_image
 414  *      @device_list - list of topology devices present in this crat_image
 415  */
 416 static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr,
 417                                 struct list_head *device_list)
 418 {
 419         struct crat_subtype_computeunit *cu;
 420         struct crat_subtype_memory *mem;
 421         struct crat_subtype_cache *cache;
 422         struct crat_subtype_iolink *iolink;
 423         int ret = 0;
 424
 425         switch (sub_type_hdr->type) {
 426         case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
 427                 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
 428                 ret = kfd_parse_subtype_cu(cu, device_list);
 429                 break;
 430         case CRAT_SUBTYPE_MEMORY_AFFINITY:
 431                 mem = (struct crat_subtype_memory *)sub_type_hdr;
 432                 ret = kfd_parse_subtype_mem(mem, device_list);
 433                 break;
 434         case CRAT_SUBTYPE_CACHE_AFFINITY:
 435                 cache = (struct crat_subtype_cache *)sub_type_hdr;
 436                 ret = kfd_parse_subtype_cache(cache, device_list);
 437                 break;
 438         case CRAT_SUBTYPE_TLB_AFFINITY:
 439                 /*
 440                  * For now, nothing to do here
 441                  */
 442                 pr_debug("Found TLB entry in CRAT table (not processing)\n");
 443                 break;
 444         case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
 445                 /*
 446                  * For now, nothing to do here
 447                  */
 448                 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n");
 449                 break;
 450         case CRAT_SUBTYPE_IOLINK_AFFINITY:
 451                 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
 452                 ret = kfd_parse_subtype_iolink(iolink, device_list);
 453                 break;
 454         default:
 455                 pr_warn("Unknown subtype %d in CRAT\n",
 456                                 sub_type_hdr->type);
 457         }
 458
 459         return ret;
 460 }
 461
 462 /* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT
 463  * create a kfd_topology_device and add in to device_list. Also parse
 464  * CRAT subtypes and attach it to appropriate kfd_topology_device
 465  *      @crat_image - input image containing CRAT
 466  *      @device_list - [OUT] list of kfd_topology_device generated after
 467  *                     parsing crat_image
 468  *      @proximity_domain - Proximity domain of the first device in the table
 469  *
 470  *      Return - 0 if successful else -ve value
 471  */
 472 int kfd_parse_crat_table(void *crat_image, struct list_head *device_list,
 473                          uint32_t proximity_domain)
 474 {
 475         struct kfd_topology_device *top_dev = NULL;
 476         struct crat_subtype_generic *sub_type_hdr;
 477         uint16_t node_id;
 478         int ret = 0;
 479         struct crat_header *crat_table = (struct crat_header *)crat_image;
 480         uint16_t num_nodes;
 481         uint32_t image_len;
 482
 483         if (!crat_image)
 484                 return -EINVAL;
 485
 486         if (!list_empty(device_list)) {
 487                 pr_warn("Error device list should be empty\n");
 488                 return -EINVAL;
 489         }
 490
 491         num_nodes = crat_table->num_domains;
 492         image_len = crat_table->length;
 493
 494         pr_info("Parsing CRAT table with %d nodes\n", num_nodes);
 495
 496         for (node_id = 0; node_id < num_nodes; node_id++) {
 497                 top_dev = kfd_create_topology_device(device_list);
 498                 if (!top_dev)
 499                         break;
 500                 top_dev->proximity_domain = proximity_domain++;
 501         }
 502
 503         if (!top_dev) {
 504                 ret = -ENOMEM;
 505                 goto err;
 506         }
 507
 508         memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH);
 509         memcpy(top_dev->oem_table_id, crat_table->oem_table_id,
 510                         CRAT_OEMTABLEID_LENGTH);
 511         top_dev->oem_revision = crat_table->oem_revision;
 512
 513         sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
 514         while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
 515                         ((char *)crat_image) + image_len) {
 516                 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
 517                         ret = kfd_parse_subtype(sub_type_hdr, device_list);
 518                         if (ret)
 519                                 break;
 520                 }
 521
 522                 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
 523                                 sub_type_hdr->length);
 524         }
 525
 526 err:
 527         if (ret)
 528                 kfd_release_topology_device_list(device_list);
 529
 530         return ret;
 531 }
 532
 533 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
 534 static int fill_in_pcache(struct crat_subtype_cache *pcache,
 535                                 struct kfd_gpu_cache_info *pcache_info,
 536                                 struct kfd_cu_info *cu_info,
 537                                 int mem_available,
 538                                 int cu_bitmask,
 539                                 int cache_type, unsigned int cu_processor_id,
 540                                 int cu_block)
 541 {
 542         unsigned int cu_sibling_map_mask;
 543         int first_active_cu;
 544
 545         /* First check if enough memory is available */
 546         if (sizeof(struct crat_subtype_cache) > mem_available)
 547                 return -ENOMEM;
 548
 549         cu_sibling_map_mask = cu_bitmask;
 550         cu_sibling_map_mask >>= cu_block;
 551         cu_sibling_map_mask &=
 552                 ((1 << pcache_info[cache_type].num_cu_shared) - 1);
 553         first_active_cu = ffs(cu_sibling_map_mask);
 554
 555         /* CU could be inactive. In case of shared cache find the first active
 556          * CU. and incase of non-shared cache check if the CU is inactive. If
 557          * inactive active skip it
 558          */
 559         if (first_active_cu) {
 560                 memset(pcache, 0, sizeof(struct crat_subtype_cache));
 561                 pcache->type = CRAT_SUBTYPE_CACHE_AFFINITY;
 562                 pcache->length = sizeof(struct crat_subtype_cache);
 563                 pcache->flags = pcache_info[cache_type].flags;
 564                 pcache->processor_id_low = cu_processor_id
 565                                          + (first_active_cu - 1);
 566                 pcache->cache_level = pcache_info[cache_type].cache_level;
 567                 pcache->cache_size = pcache_info[cache_type].cache_size;
 568
 569                 /* Sibling map is w.r.t processor_id_low, so shift out
 570                  * inactive CU
 571                  */
 572                 cu_sibling_map_mask =
 573                         cu_sibling_map_mask >> (first_active_cu - 1);
 574
 575                 pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF);
 576                 pcache->sibling_map[1] =
 577                                 (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
 578                 pcache->sibling_map[2] =
 579                                 (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
 580                 pcache->sibling_map[3] =
 581                                 (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
 582                 return 0;
 583         }
 584         return 1;
 585 }
 586
 587 /* kfd_fill_gpu_cache_info - Fill GPU cache info using kfd_gpu_cache_info
 588  * tables
 589  *
 590  *      @kdev - [IN] GPU device
 591  *      @gpu_processor_id - [IN] GPU processor ID to which these caches
 592  *                          associate
 593  *      @available_size - [IN] Amount of memory available in pcache
 594  *      @cu_info - [IN] Compute Unit info obtained from KGD
 595  *      @pcache - [OUT] memory into which cache data is to be filled in.
 596  *      @size_filled - [OUT] amount of data used up in pcache.
 597  *      @num_of_entries - [OUT] number of caches added
 598  */
 599 static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,
 600                         int gpu_processor_id,
 601                         int available_size,
 602                         struct kfd_cu_info *cu_info,
 603                         struct crat_subtype_cache *pcache,
 604                         int *size_filled,
 605                         int *num_of_entries)
 606 {
 607         struct kfd_gpu_cache_info *pcache_info;
 608         int num_of_cache_types = 0;
 609         int i, j, k;
 610         int ct = 0;
 611         int mem_available = available_size;
 612         unsigned int cu_processor_id;
 613         int ret;
 614
 615         switch (kdev->device_info->asic_family) {
 616         case CHIP_KAVERI:
 617                 pcache_info = kaveri_cache_info;
 618                 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info);
 619                 break;
 620         case CHIP_HAWAII:
 621                 pcache_info = hawaii_cache_info;
 622                 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info);
 623                 break;
 624         case CHIP_CARRIZO:
 625                 pcache_info = carrizo_cache_info;
 626                 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info);
 627                 break;
 628         case CHIP_TONGA:
 629                 pcache_info = tonga_cache_info;
 630                 num_of_cache_types = ARRAY_SIZE(tonga_cache_info);
 631                 break;
 632         case CHIP_FIJI:
 633                 pcache_info = fiji_cache_info;
 634                 num_of_cache_types = ARRAY_SIZE(fiji_cache_info);
 635                 break;
 636         case CHIP_POLARIS10:
 637                 pcache_info = polaris10_cache_info;
 638                 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info);
 639                 break;
 640         case CHIP_POLARIS11:
 641                 pcache_info = polaris11_cache_info;
 642                 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info);
 643                 break;
 644         case CHIP_VEGA10:
 645                 pcache_info = vega10_cache_info;
 646                 num_of_cache_types = ARRAY_SIZE(vega10_cache_info);
 647                 break;
 648         case CHIP_RAVEN:
 649                 pcache_info = raven_cache_info;
 650                 num_of_cache_types = ARRAY_SIZE(raven_cache_info);
 651                 break;
 652         default:
 653                 return -EINVAL;
 654         }
 655
 656         *size_filled = 0;
 657         *num_of_entries = 0;
 658
 659         /* For each type of cache listed in the kfd_gpu_cache_info table,
 660          * go through all available Compute Units.
 661          * The [i,j,k] loop will
 662          *              if kfd_gpu_cache_info.num_cu_shared = 1
 663          *                      will parse through all available CU
 664          *              If (kfd_gpu_cache_info.num_cu_shared != 1)
 665          *                      then it will consider only one CU from
 666          *                      the shared unit
 667          */
 668
 669         for (ct = 0; ct < num_of_cache_types; ct++) {
 670                 cu_processor_id = gpu_processor_id;
 671                 for (i = 0; i < cu_info->num_shader_engines; i++) {
 672                         for (j = 0; j < cu_info->num_shader_arrays_per_engine;
 673                                 j++) {
 674                                 for (k = 0; k < cu_info->num_cu_per_sh;
 675                                         k += pcache_info[ct].num_cu_shared) {
 676
 677                                         ret = fill_in_pcache(pcache,
 678                                                 pcache_info,
 679                                                 cu_info,
 680                                                 mem_available,
 681                                                 cu_info->cu_bitmap[i][j],
 682                                                 ct,
 683                                                 cu_processor_id,
 684                                                 k);
 685
 686                                         if (ret < 0)
 687                                                 break;
 688
 689                                         if (!ret) {
 690                                                 pcache++;
 691                                                 (*num_of_entries)++;
 692                                                 mem_available -=
 693                                                         sizeof(*pcache);
 694                                                 (*size_filled) +=
 695                                                         sizeof(*pcache);
 696                                         }
 697
 698                                         /* Move to next CU block */
 699                                         cu_processor_id +=
 700                                                 pcache_info[ct].num_cu_shared;
 701                                 }
 702                         }
 703                 }
 704         }
 705
 706         pr_debug("Added [%d] GPU cache entries\n", *num_of_entries);
 707
 708         return 0;
 709 }
 710
 711 /*
 712  * kfd_create_crat_image_acpi - Allocates memory for CRAT image and
 713  * copies CRAT from ACPI (if available).
 714  * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
 715  *
 716  *      @crat_image: CRAT read from ACPI. If no CRAT in ACPI then
 717  *                   crat_image will be NULL
 718  *      @size: [OUT] size of crat_image
 719  *
 720  *      Return 0 if successful else return error code
 721  */
 722 int kfd_create_crat_image_acpi(void **crat_image, size_t *size)
 723 {
 724         struct acpi_table_header *crat_table;
 725         acpi_status status;
 726         void *pcrat_image;
 727
 728         if (!crat_image)
 729                 return -EINVAL;
 730
 731         *crat_image = NULL;
 732
 733         /* Fetch the CRAT table from ACPI */
 734         status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table);
 735         if (status == AE_NOT_FOUND) {
 736                 pr_info("CRAT table not found\n");
 737                 return -ENODATA;
 738         } else if (ACPI_FAILURE(status)) {
 739                 const char *err = acpi_format_exception(status);
 740
 741                 pr_err("CRAT table error: %s\n", err);
 742                 return -EINVAL;
 743         }
 744
 745         if (ignore_crat) {
 746                 pr_info("CRAT table disabled by module option\n");
 747                 return -ENODATA;
 748         }
 749
 750         pcrat_image = kmalloc(crat_table->length, GFP_KERNEL);
 751         if (!pcrat_image)
 752                 return -ENOMEM;
 753
 754         memcpy(pcrat_image, crat_table, crat_table->length);
 755
 756         *crat_image = pcrat_image;
 757         *size = crat_table->length;
 758
 759         return 0;
 760 }
 761
 762 /* Memory required to create Virtual CRAT.
 763  * Since there is no easy way to predict the amount of memory required, the
 764  * following amount are allocated for CPU and GPU Virtual CRAT. This is
 765  * expected to cover all known conditions. But to be safe additional check
 766  * is put in the code to ensure we don't overwrite.
 767  */
 768 #define VCRAT_SIZE_FOR_CPU      (2 * PAGE_SIZE)
 769 #define VCRAT_SIZE_FOR_GPU      (3 * PAGE_SIZE)
 770
 771 /* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node
 772  *
 773  *      @numa_node_id: CPU NUMA node id
 774  *      @avail_size: Available size in the memory
 775  *      @sub_type_hdr: Memory into which compute info will be filled in
 776  *
 777  *      Return 0 if successful else return -ve value
 778  */
 779 static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size,
 780                                 int proximity_domain,
 781                                 struct crat_subtype_computeunit *sub_type_hdr)
 782 {
 783         const struct cpumask *cpumask;
 784
 785         *avail_size -= sizeof(struct crat_subtype_computeunit);
 786         if (*avail_size < 0)
 787                 return -ENOMEM;
 788
 789         memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
 790
 791         /* Fill in subtype header data */
 792         sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
 793         sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
 794         sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
 795
 796         cpumask = cpumask_of_node(numa_node_id);
 797
 798         /* Fill in CU data */
 799         sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT;
 800         sub_type_hdr->proximity_domain = proximity_domain;
 801         sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id);
 802         if (sub_type_hdr->processor_id_low == -1)
 803                 return -EINVAL;
 804
 805         sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask);
 806
 807         return 0;
 808 }
 809
 810 /* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node
 811  *
 812  *      @numa_node_id: CPU NUMA node id
 813  *      @avail_size: Available size in the memory
 814  *      @sub_type_hdr: Memory into which compute info will be filled in
 815  *
 816  *      Return 0 if successful else return -ve value
 817  */
 818 static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size,
 819                         int proximity_domain,
 820                         struct crat_subtype_memory *sub_type_hdr)
 821 {
 822         uint64_t mem_in_bytes = 0;
 823         pg_data_t *pgdat;
 824         int zone_type;
 825
 826         *avail_size -= sizeof(struct crat_subtype_memory);
 827         if (*avail_size < 0)
 828                 return -ENOMEM;
 829
 830         memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
 831
 832         /* Fill in subtype header data */
 833         sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
 834         sub_type_hdr->length = sizeof(struct crat_subtype_memory);
 835         sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
 836
 837         /* Fill in Memory Subunit data */
 838
 839         /* Unlike si_meminfo, si_meminfo_node is not exported. So
 840          * the following lines are duplicated from si_meminfo_node
 841          * function
 842          */
 843         pgdat = NODE_DATA(numa_node_id);
 844         for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
 845                 mem_in_bytes += pgdat->node_zones[zone_type].managed_pages;
 846         mem_in_bytes <<= PAGE_SHIFT;
 847
 848         sub_type_hdr->length_low = lower_32_bits(mem_in_bytes);
 849         sub_type_hdr->length_high = upper_32_bits(mem_in_bytes);
 850         sub_type_hdr->proximity_domain = proximity_domain;
 851
 852         return 0;
 853 }
 854
 855 static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size,
 856                                 uint32_t *num_entries,
 857                                 struct crat_subtype_iolink *sub_type_hdr)
 858 {
 859         int nid;
 860         struct cpuinfo_x86 *c = &cpu_data(0);
 861         uint8_t link_type;
 862
 863         if (c->x86_vendor == X86_VENDOR_AMD)
 864                 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT;
 865         else
 866                 link_type = CRAT_IOLINK_TYPE_QPI_1_1;
 867
 868         *num_entries = 0;
 869
 870         /* Create IO links from this node to other CPU nodes */
 871         for_each_online_node(nid) {
 872                 if (nid == numa_node_id) /* node itself */
 873                         continue;
 874
 875                 *avail_size -= sizeof(struct crat_subtype_iolink);
 876                 if (*avail_size < 0)
 877                         return -ENOMEM;
 878
 879                 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
 880
 881                 /* Fill in subtype header data */
 882                 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
 883                 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
 884                 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
 885
 886                 /* Fill in IO link data */
 887                 sub_type_hdr->proximity_domain_from = numa_node_id;
 888                 sub_type_hdr->proximity_domain_to = nid;
 889                 sub_type_hdr->io_interface_type = link_type;
 890
 891                 (*num_entries)++;
 892                 sub_type_hdr++;
 893         }
 894
 895         return 0;
 896 }
 897
 898 /* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU
 899  *
 900  *      @pcrat_image: Fill in VCRAT for CPU
 901  *      @size:  [IN] allocated size of crat_image.
 902  *              [OUT] actual size of data filled in crat_image
 903  */
 904 static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size)
 905 {
 906         struct crat_header *crat_table = (struct crat_header *)pcrat_image;
 907         struct acpi_table_header *acpi_table;
 908         acpi_status status;
 909         struct crat_subtype_generic *sub_type_hdr;
 910         int avail_size = *size;
 911         int numa_node_id;
 912         uint32_t entries = 0;
 913         int ret = 0;
 914
 915         if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_CPU)
 916                 return -EINVAL;
 917
 918         /* Fill in CRAT Header.
 919          * Modify length and total_entries as subunits are added.
 920          */
 921         avail_size -= sizeof(struct crat_header);
 922         if (avail_size < 0)
 923                 return -ENOMEM;
 924
 925         memset(crat_table, 0, sizeof(struct crat_header));
 926         memcpy(&crat_table->signature, CRAT_SIGNATURE,
 927                         sizeof(crat_table->signature));
 928         crat_table->length = sizeof(struct crat_header);
 929
 930         status = acpi_get_table("DSDT", 0, &acpi_table);
 931         if (status != AE_OK)
 932                 pr_warn("DSDT table not found for OEM information\n");
 933         else {
 934                 crat_table->oem_revision = acpi_table->revision;
 935                 memcpy(crat_table->oem_id, acpi_table->oem_id,
 936                                 CRAT_OEMID_LENGTH);
 937                 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id,
 938                                 CRAT_OEMTABLEID_LENGTH);
 939         }
 940         crat_table->total_entries = 0;
 941         crat_table->num_domains = 0;
 942
 943         sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
 944
 945         for_each_online_node(numa_node_id) {
 946                 if (kfd_numa_node_to_apic_id(numa_node_id) == -1)
 947                         continue;
 948
 949                 /* Fill in Subtype: Compute Unit */
 950                 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size,
 951                         crat_table->num_domains,
 952                         (struct crat_subtype_computeunit *)sub_type_hdr);
 953                 if (ret < 0)
 954                         return ret;
 955                 crat_table->length += sub_type_hdr->length;
 956                 crat_table->total_entries++;
 957
 958                 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
 959                         sub_type_hdr->length);
 960
 961                 /* Fill in Subtype: Memory */
 962                 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size,
 963                         crat_table->num_domains,
 964                         (struct crat_subtype_memory *)sub_type_hdr);
 965                 if (ret < 0)
 966                         return ret;
 967                 crat_table->length += sub_type_hdr->length;
 968                 crat_table->total_entries++;
 969
 970                 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
 971                         sub_type_hdr->length);
 972
 973                 /* Fill in Subtype: IO Link */
 974                 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size,
 975                                 &entries,
 976                                 (struct crat_subtype_iolink *)sub_type_hdr);
 977                 if (ret < 0)
 978                         return ret;
 979                 crat_table->length += (sub_type_hdr->length * entries);
 980                 crat_table->total_entries += entries;
 981
 982                 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
 983                                 sub_type_hdr->length * entries);
 984
 985                 crat_table->num_domains++;
 986         }
 987
 988         /* TODO: Add cache Subtype for CPU.
 989          * Currently, CPU cache information is available in function
 990          * detect_cache_attributes(cpu) defined in the file
 991          * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not
 992          * exported and to get the same information the code needs to be
 993          * duplicated.
 994          */
 995
 996         *size = crat_table->length;
 997         pr_info("Virtual CRAT table created for CPU\n");
 998
 999         return 0;
1000 }
1001
1002 static int kfd_fill_gpu_memory_affinity(int *avail_size,
1003                 struct kfd_dev *kdev, uint8_t type, uint64_t size,
1004                 struct crat_subtype_memory *sub_type_hdr,
1005                 uint32_t proximity_domain,
1006                 const struct kfd_local_mem_info *local_mem_info)
1007 {
1008         *avail_size -= sizeof(struct crat_subtype_memory);
1009         if (*avail_size < 0)
1010                 return -ENOMEM;
1011
1012         memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1013         sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1014         sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1015         sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1016
1017         sub_type_hdr->proximity_domain = proximity_domain;
1018
1019         pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n",
1020                         type, size);
1021
1022         sub_type_hdr->length_low = lower_32_bits(size);
1023         sub_type_hdr->length_high = upper_32_bits(size);
1024
1025         sub_type_hdr->width = local_mem_info->vram_width;
1026         sub_type_hdr->visibility_type = type;
1027
1028         return 0;
1029 }
1030
1031 /* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU
1032  * to its NUMA node
1033  *      @avail_size: Available size in the memory
1034  *      @kdev - [IN] GPU device
1035  *      @sub_type_hdr: Memory into which io link info will be filled in
1036  *      @proximity_domain - proximity domain of the GPU node
1037  *
1038  *      Return 0 if successful else return -ve value
1039  */
1040 static int kfd_fill_gpu_direct_io_link(int *avail_size,
1041                         struct kfd_dev *kdev,
1042                         struct crat_subtype_iolink *sub_type_hdr,
1043                         uint32_t proximity_domain)
1044 {
1045         *avail_size -= sizeof(struct crat_subtype_iolink);
1046         if (*avail_size < 0)
1047                 return -ENOMEM;
1048
1049         memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1050
1051         /* Fill in subtype header data */
1052         sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1053         sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1054         sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1055
1056         /* Fill in IOLINK subtype.
1057          * TODO: Fill-in other fields of iolink subtype
1058          */
1059         sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS;
1060         sub_type_hdr->proximity_domain_from = proximity_domain;
1061 #ifdef CONFIG_NUMA
1062         if (kdev->pdev->dev.numa_node == NUMA_NO_NODE)
1063                 sub_type_hdr->proximity_domain_to = 0;
1064         else
1065                 sub_type_hdr->proximity_domain_to = kdev->pdev->dev.numa_node;
1066 #else
1067         sub_type_hdr->proximity_domain_to = 0;
1068 #endif
1069         return 0;
1070 }
1071
1072 /* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU
1073  *
1074  *      @pcrat_image: Fill in VCRAT for GPU
1075  *      @size:  [IN] allocated size of crat_image.
1076  *              [OUT] actual size of data filled in crat_image
1077  */
1078 static int kfd_create_vcrat_image_gpu(void *pcrat_image,
1079                                       size_t *size, struct kfd_dev *kdev,
1080                                       uint32_t proximity_domain)
1081 {
1082         struct crat_header *crat_table = (struct crat_header *)pcrat_image;
1083         struct crat_subtype_generic *sub_type_hdr;
1084         struct crat_subtype_computeunit *cu;
1085         struct kfd_cu_info cu_info;
1086         int avail_size = *size;
1087         uint32_t total_num_of_cu;
1088         int num_of_cache_entries = 0;
1089         int cache_mem_filled = 0;
1090         int ret = 0;
1091         struct kfd_local_mem_info local_mem_info;
1092
1093         if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU)
1094                 return -EINVAL;
1095
1096         /* Fill the CRAT Header.
1097          * Modify length and total_entries as subunits are added.
1098          */
1099         avail_size -= sizeof(struct crat_header);
1100         if (avail_size < 0)
1101                 return -ENOMEM;
1102
1103         memset(crat_table, 0, sizeof(struct crat_header));
1104
1105         memcpy(&crat_table->signature, CRAT_SIGNATURE,
1106                         sizeof(crat_table->signature));
1107         /* Change length as we add more subtypes*/
1108         crat_table->length = sizeof(struct crat_header);
1109         crat_table->num_domains = 1;
1110         crat_table->total_entries = 0;
1111
1112         /* Fill in Subtype: Compute Unit
1113          * First fill in the sub type header and then sub type data
1114          */
1115         avail_size -= sizeof(struct crat_subtype_computeunit);
1116         if (avail_size < 0)
1117                 return -ENOMEM;
1118
1119         sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1);
1120         memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
1121
1122         sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
1123         sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
1124         sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1125
1126         /* Fill CU subtype data */
1127         cu = (struct crat_subtype_computeunit *)sub_type_hdr;
1128         cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT;
1129         cu->proximity_domain = proximity_domain;
1130
1131         kdev->kfd2kgd->get_cu_info(kdev->kgd, &cu_info);
1132         cu->num_simd_per_cu = cu_info.simd_per_cu;
1133         cu->num_simd_cores = cu_info.simd_per_cu * cu_info.cu_active_number;
1134         cu->max_waves_simd = cu_info.max_waves_per_simd;
1135
1136         cu->wave_front_size = cu_info.wave_front_size;
1137         cu->array_count = cu_info.num_shader_arrays_per_engine *
1138                 cu_info.num_shader_engines;
1139         total_num_of_cu = (cu->array_count * cu_info.num_cu_per_sh);
1140         cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu);
1141         cu->num_cu_per_array = cu_info.num_cu_per_sh;
1142         cu->max_slots_scatch_cu = cu_info.max_scratch_slots_per_cu;
1143         cu->num_banks = cu_info.num_shader_engines;
1144         cu->lds_size_in_kb = cu_info.lds_size;
1145
1146         cu->hsa_capability = 0;
1147
1148         /* Check if this node supports IOMMU. During parsing this flag will
1149          * translate to HSA_CAP_ATS_PRESENT
1150          */
1151         if (!kfd_iommu_check_device(kdev))
1152                 cu->hsa_capability |= CRAT_CU_FLAGS_IOMMU_PRESENT;
1153
1154         crat_table->length += sub_type_hdr->length;
1155         crat_table->total_entries++;
1156
1157         /* Fill in Subtype: Memory. Only on systems with large BAR (no
1158          * private FB), report memory as public. On other systems
1159          * report the total FB size (public+private) as a single
1160          * private heap.
1161          */
1162         kdev->kfd2kgd->get_local_mem_info(kdev->kgd, &local_mem_info);
1163         sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1164                         sub_type_hdr->length);
1165
1166         if (debug_largebar)
1167                 local_mem_info.local_mem_size_private = 0;
1168
1169         if (local_mem_info.local_mem_size_private == 0)
1170                 ret = kfd_fill_gpu_memory_affinity(&avail_size,
1171                                 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC,
1172                                 local_mem_info.local_mem_size_public,
1173                                 (struct crat_subtype_memory *)sub_type_hdr,
1174                                 proximity_domain,
1175                                 &local_mem_info);
1176         else
1177                 ret = kfd_fill_gpu_memory_affinity(&avail_size,
1178                                 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE,
1179                                 local_mem_info.local_mem_size_public +
1180                                 local_mem_info.local_mem_size_private,
1181                                 (struct crat_subtype_memory *)sub_type_hdr,
1182                                 proximity_domain,
1183                                 &local_mem_info);
1184         if (ret < 0)
1185                 return ret;
1186
1187         crat_table->length += sizeof(struct crat_subtype_memory);
1188         crat_table->total_entries++;
1189
1190         /* TODO: Fill in cache information. This information is NOT readily
1191          * available in KGD
1192          */
1193         sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1194                 sub_type_hdr->length);
1195         ret = kfd_fill_gpu_cache_info(kdev, cu->processor_id_low,
1196                                 avail_size,
1197                                 &cu_info,
1198                                 (struct crat_subtype_cache *)sub_type_hdr,
1199                                 &cache_mem_filled,
1200                                 &num_of_cache_entries);
1201
1202         if (ret < 0)
1203                 return ret;
1204
1205         crat_table->length += cache_mem_filled;
1206         crat_table->total_entries += num_of_cache_entries;
1207         avail_size -= cache_mem_filled;
1208
1209         /* Fill in Subtype: IO_LINKS
1210          *  Only direct links are added here which is Link from GPU to
1211          *  to its NUMA node. Indirect links are added by userspace.
1212          */
1213         sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1214                 cache_mem_filled);
1215         ret = kfd_fill_gpu_direct_io_link(&avail_size, kdev,
1216                 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain);
1217
1218         if (ret < 0)
1219                 return ret;
1220
1221         crat_table->length += sub_type_hdr->length;
1222         crat_table->total_entries++;
1223
1224         *size = crat_table->length;
1225         pr_info("Virtual CRAT table created for GPU\n");
1226
1227         return ret;
1228 }
1229
1230 /* kfd_create_crat_image_virtual - Allocates memory for CRAT image and
1231  *              creates a Virtual CRAT (VCRAT) image
1232  *
1233  * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
1234  *
1235  *      @crat_image: VCRAT image created because ACPI does not have a
1236  *                   CRAT for this device
1237  *      @size: [OUT] size of virtual crat_image
1238  *      @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device
1239  *              COMPUTE_UNIT_GPU - Create VCRAT for GPU
1240  *              (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU
1241  *                      -- this option is not currently implemented.
1242  *                      The assumption is that all AMD APUs will have CRAT
1243  *      @kdev: Valid kfd_device required if flags contain COMPUTE_UNIT_GPU
1244  *
1245  *      Return 0 if successful else return -ve value
1246  */
1247 int kfd_create_crat_image_virtual(void **crat_image, size_t *size,
1248                                   int flags, struct kfd_dev *kdev,
1249                                   uint32_t proximity_domain)
1250 {
1251         void *pcrat_image = NULL;
1252         int ret = 0;
1253
1254         if (!crat_image)
1255                 return -EINVAL;
1256
1257         *crat_image = NULL;
1258
1259         /* Allocate one VCRAT_SIZE_FOR_CPU for CPU virtual CRAT image and
1260          * VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image. This should cover
1261          * all the current conditions. A check is put not to overwrite beyond
1262          * allocated size
1263          */
1264         switch (flags) {
1265         case COMPUTE_UNIT_CPU:
1266                 pcrat_image = kmalloc(VCRAT_SIZE_FOR_CPU, GFP_KERNEL);
1267                 if (!pcrat_image)
1268                         return -ENOMEM;
1269                 *size = VCRAT_SIZE_FOR_CPU;
1270                 ret = kfd_create_vcrat_image_cpu(pcrat_image, size);
1271                 break;
1272         case COMPUTE_UNIT_GPU:
1273                 if (!kdev)
1274                         return -EINVAL;
1275                 pcrat_image = kmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL);
1276                 if (!pcrat_image)
1277                         return -ENOMEM;
1278                 *size = VCRAT_SIZE_FOR_GPU;
1279                 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev,
1280                                                  proximity_domain);
1281                 break;
1282         case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU):
1283                 /* TODO: */
1284                 ret = -EINVAL;
1285                 pr_err("VCRAT not implemented for APU\n");
1286                 break;
1287         default:
1288                 ret = -EINVAL;
1289         }
1290
1291         if (!ret)
1292                 *crat_image = pcrat_image;
1293         else
1294                 kfree(pcrat_image);
1295
1296         return ret;
1297 }
1298
1299
1300 /* kfd_destroy_crat_image
1301  *
1302  *      @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..)
1303  *
1304  */
1305 void kfd_destroy_crat_image(void *crat_image)
1306 {
1307         kfree(crat_image);
1308 }