Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
   1// SPDX-License-Identifier: GPL-2.0 OR MIT
   2/*
   3 * Copyright 2015-2022 Advanced Micro Devices, Inc.
   4 *
   5 * Permission is hereby granted, free of charge, to any person obtaining a
   6 * copy of this software and associated documentation files (the "Software"),
   7 * to deal in the Software without restriction, including without limitation
   8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9 * and/or sell copies of the Software, and to permit persons to whom the
  10 * Software is furnished to do so, subject to the following conditions:
  11 *
  12 * The above copyright notice and this permission notice shall be included in
  13 * all copies or substantial portions of the Software.
  14 *
  15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  21 * OTHER DEALINGS IN THE SOFTWARE.
  22 */
  23
  24#include <linux/pci.h>
  25#include <linux/acpi.h>
  26#include "kfd_crat.h"
  27#include "kfd_priv.h"
  28#include "kfd_topology.h"
  29#include "kfd_iommu.h"
  30#include "amdgpu.h"
  31#include "amdgpu_amdkfd.h"
  32
  33/* GPU Processor ID base for dGPUs for which VCRAT needs to be created.
  34 * GPU processor ID are expressed with Bit[31]=1.
  35 * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs
  36 * used in the CRAT.
  37 */
  38static uint32_t gpu_processor_id_low = 0x80001000;
  39
  40/* Return the next available gpu_processor_id and increment it for next GPU
  41 *	@total_cu_count - Total CUs present in the GPU including ones
  42 *			  masked off
  43 */
  44static inline unsigned int get_and_inc_gpu_processor_id(
  45				unsigned int total_cu_count)
  46{
  47	int current_id = gpu_processor_id_low;
  48
  49	gpu_processor_id_low += total_cu_count;
  50	return current_id;
  51}
  52
  53
  54static struct kfd_gpu_cache_info kaveri_cache_info[] = {
  55	{
  56		/* TCP L1 Cache per CU */
  57		.cache_size = 16,
  58		.cache_level = 1,
  59		.flags = (CRAT_CACHE_FLAGS_ENABLED |
  60				CRAT_CACHE_FLAGS_DATA_CACHE |
  61				CRAT_CACHE_FLAGS_SIMD_CACHE),
  62		.num_cu_shared = 1,
  63	},
  64	{
  65		/* Scalar L1 Instruction Cache (in SQC module) per bank */
  66		.cache_size = 16,
  67		.cache_level = 1,
  68		.flags = (CRAT_CACHE_FLAGS_ENABLED |
  69				CRAT_CACHE_FLAGS_INST_CACHE |
  70				CRAT_CACHE_FLAGS_SIMD_CACHE),
  71		.num_cu_shared = 2,
  72	},
  73	{
  74		/* Scalar L1 Data Cache (in SQC module) per bank */
  75		.cache_size = 8,
  76		.cache_level = 1,
  77		.flags = (CRAT_CACHE_FLAGS_ENABLED |
  78				CRAT_CACHE_FLAGS_DATA_CACHE |
  79				CRAT_CACHE_FLAGS_SIMD_CACHE),
  80		.num_cu_shared = 2,
  81	},
  82
  83	/* TODO: Add L2 Cache information */
  84};
  85
  86
  87static struct kfd_gpu_cache_info carrizo_cache_info[] = {
  88	{
  89		/* TCP L1 Cache per CU */
  90		.cache_size = 16,
  91		.cache_level = 1,
  92		.flags = (CRAT_CACHE_FLAGS_ENABLED |
  93				CRAT_CACHE_FLAGS_DATA_CACHE |
  94				CRAT_CACHE_FLAGS_SIMD_CACHE),
  95		.num_cu_shared = 1,
  96	},
  97	{
  98		/* Scalar L1 Instruction Cache (in SQC module) per bank */
  99		.cache_size = 8,
 100		.cache_level = 1,
 101		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 102				CRAT_CACHE_FLAGS_INST_CACHE |
 103				CRAT_CACHE_FLAGS_SIMD_CACHE),
 104		.num_cu_shared = 4,
 105	},
 106	{
 107		/* Scalar L1 Data Cache (in SQC module) per bank. */
 108		.cache_size = 4,
 109		.cache_level = 1,
 110		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 111				CRAT_CACHE_FLAGS_DATA_CACHE |
 112				CRAT_CACHE_FLAGS_SIMD_CACHE),
 113		.num_cu_shared = 4,
 114	},
 115
 116	/* TODO: Add L2 Cache information */
 117};
 118
 119#define hawaii_cache_info kaveri_cache_info
 120#define tonga_cache_info carrizo_cache_info
 121#define fiji_cache_info  carrizo_cache_info
 122#define polaris10_cache_info carrizo_cache_info
 123#define polaris11_cache_info carrizo_cache_info
 124#define polaris12_cache_info carrizo_cache_info
 125#define vegam_cache_info carrizo_cache_info
 126
 127/* NOTE: L1 cache information has been updated and L2/L3
 128 * cache information has been added for Vega10 and
 129 * newer ASICs. The unit for cache_size is KiB.
 130 * In future,  check & update cache details
 131 * for every new ASIC is required.
 132 */
 133
 134static struct kfd_gpu_cache_info vega10_cache_info[] = {
 135	{
 136		/* TCP L1 Cache per CU */
 137		.cache_size = 16,
 138		.cache_level = 1,
 139		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 140				CRAT_CACHE_FLAGS_DATA_CACHE |
 141				CRAT_CACHE_FLAGS_SIMD_CACHE),
 142		.num_cu_shared = 1,
 143	},
 144	{
 145		/* Scalar L1 Instruction Cache per SQC */
 146		.cache_size = 32,
 147		.cache_level = 1,
 148		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 149				CRAT_CACHE_FLAGS_INST_CACHE |
 150				CRAT_CACHE_FLAGS_SIMD_CACHE),
 151		.num_cu_shared = 3,
 152	},
 153	{
 154		/* Scalar L1 Data Cache per SQC */
 155		.cache_size = 16,
 156		.cache_level = 1,
 157		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 158				CRAT_CACHE_FLAGS_DATA_CACHE |
 159				CRAT_CACHE_FLAGS_SIMD_CACHE),
 160		.num_cu_shared = 3,
 161	},
 162	{
 163		/* L2 Data Cache per GPU (Total Tex Cache) */
 164		.cache_size = 4096,
 165		.cache_level = 2,
 166		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 167				CRAT_CACHE_FLAGS_DATA_CACHE |
 168				CRAT_CACHE_FLAGS_SIMD_CACHE),
 169		.num_cu_shared = 16,
 170	},
 171};
 172
 173static struct kfd_gpu_cache_info raven_cache_info[] = {
 174	{
 175		/* TCP L1 Cache per CU */
 176		.cache_size = 16,
 177		.cache_level = 1,
 178		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 179				CRAT_CACHE_FLAGS_DATA_CACHE |
 180				CRAT_CACHE_FLAGS_SIMD_CACHE),
 181		.num_cu_shared = 1,
 182	},
 183	{
 184		/* Scalar L1 Instruction Cache per SQC */
 185		.cache_size = 32,
 186		.cache_level = 1,
 187		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 188				CRAT_CACHE_FLAGS_INST_CACHE |
 189				CRAT_CACHE_FLAGS_SIMD_CACHE),
 190		.num_cu_shared = 3,
 191	},
 192	{
 193		/* Scalar L1 Data Cache per SQC */
 194		.cache_size = 16,
 195		.cache_level = 1,
 196		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 197				CRAT_CACHE_FLAGS_DATA_CACHE |
 198				CRAT_CACHE_FLAGS_SIMD_CACHE),
 199		.num_cu_shared = 3,
 200	},
 201	{
 202		/* L2 Data Cache per GPU (Total Tex Cache) */
 203		.cache_size = 1024,
 204		.cache_level = 2,
 205		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 206				CRAT_CACHE_FLAGS_DATA_CACHE |
 207				CRAT_CACHE_FLAGS_SIMD_CACHE),
 208		.num_cu_shared = 11,
 209	},
 210};
 211
 212static struct kfd_gpu_cache_info renoir_cache_info[] = {
 213	{
 214		/* TCP L1 Cache per CU */
 215		.cache_size = 16,
 216		.cache_level = 1,
 217		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 218				CRAT_CACHE_FLAGS_DATA_CACHE |
 219				CRAT_CACHE_FLAGS_SIMD_CACHE),
 220		.num_cu_shared = 1,
 221	},
 222	{
 223		/* Scalar L1 Instruction Cache per SQC */
 224		.cache_size = 32,
 225		.cache_level = 1,
 226		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 227				CRAT_CACHE_FLAGS_INST_CACHE |
 228				CRAT_CACHE_FLAGS_SIMD_CACHE),
 229		.num_cu_shared = 3,
 230	},
 231	{
 232		/* Scalar L1 Data Cache per SQC */
 233		.cache_size = 16,
 234		.cache_level = 1,
 235		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 236				CRAT_CACHE_FLAGS_DATA_CACHE |
 237				CRAT_CACHE_FLAGS_SIMD_CACHE),
 238		.num_cu_shared = 3,
 239	},
 240	{
 241		/* L2 Data Cache per GPU (Total Tex Cache) */
 242		.cache_size = 1024,
 243		.cache_level = 2,
 244		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 245				CRAT_CACHE_FLAGS_DATA_CACHE |
 246				CRAT_CACHE_FLAGS_SIMD_CACHE),
 247		.num_cu_shared = 8,
 248	},
 249};
 250
 251static struct kfd_gpu_cache_info vega12_cache_info[] = {
 252	{
 253		/* TCP L1 Cache per CU */
 254		.cache_size = 16,
 255		.cache_level = 1,
 256		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 257				CRAT_CACHE_FLAGS_DATA_CACHE |
 258				CRAT_CACHE_FLAGS_SIMD_CACHE),
 259		.num_cu_shared = 1,
 260	},
 261	{
 262		/* Scalar L1 Instruction Cache per SQC */
 263		.cache_size = 32,
 264		.cache_level = 1,
 265		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 266				CRAT_CACHE_FLAGS_INST_CACHE |
 267				CRAT_CACHE_FLAGS_SIMD_CACHE),
 268		.num_cu_shared = 3,
 269	},
 270	{
 271		/* Scalar L1 Data Cache per SQC */
 272		.cache_size = 16,
 273		.cache_level = 1,
 274		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 275				CRAT_CACHE_FLAGS_DATA_CACHE |
 276				CRAT_CACHE_FLAGS_SIMD_CACHE),
 277		.num_cu_shared = 3,
 278	},
 279	{
 280		/* L2 Data Cache per GPU (Total Tex Cache) */
 281		.cache_size = 2048,
 282		.cache_level = 2,
 283		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 284				CRAT_CACHE_FLAGS_DATA_CACHE |
 285				CRAT_CACHE_FLAGS_SIMD_CACHE),
 286		.num_cu_shared = 5,
 287	},
 288};
 289
 290static struct kfd_gpu_cache_info vega20_cache_info[] = {
 291	{
 292		/* TCP L1 Cache per CU */
 293		.cache_size = 16,
 294		.cache_level = 1,
 295		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 296				CRAT_CACHE_FLAGS_DATA_CACHE |
 297				CRAT_CACHE_FLAGS_SIMD_CACHE),
 298		.num_cu_shared = 1,
 299	},
 300	{
 301		/* Scalar L1 Instruction Cache per SQC */
 302		.cache_size = 32,
 303		.cache_level = 1,
 304		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 305				CRAT_CACHE_FLAGS_INST_CACHE |
 306				CRAT_CACHE_FLAGS_SIMD_CACHE),
 307		.num_cu_shared = 3,
 308	},
 309	{
 310		/* Scalar L1 Data Cache per SQC */
 311		.cache_size = 16,
 312		.cache_level = 1,
 313		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 314				CRAT_CACHE_FLAGS_DATA_CACHE |
 315				CRAT_CACHE_FLAGS_SIMD_CACHE),
 316		.num_cu_shared = 3,
 317	},
 318	{
 319		/* L2 Data Cache per GPU (Total Tex Cache) */
 320		.cache_size = 8192,
 321		.cache_level = 2,
 322		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 323				CRAT_CACHE_FLAGS_DATA_CACHE |
 324				CRAT_CACHE_FLAGS_SIMD_CACHE),
 325		.num_cu_shared = 16,
 326	},
 327};
 328
 329static struct kfd_gpu_cache_info aldebaran_cache_info[] = {
 330	{
 331		/* TCP L1 Cache per CU */
 332		.cache_size = 16,
 333		.cache_level = 1,
 334		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 335				CRAT_CACHE_FLAGS_DATA_CACHE |
 336				CRAT_CACHE_FLAGS_SIMD_CACHE),
 337		.num_cu_shared = 1,
 338	},
 339	{
 340		/* Scalar L1 Instruction Cache per SQC */
 341		.cache_size = 32,
 342		.cache_level = 1,
 343		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 344				CRAT_CACHE_FLAGS_INST_CACHE |
 345				CRAT_CACHE_FLAGS_SIMD_CACHE),
 346		.num_cu_shared = 2,
 347	},
 348	{
 349		/* Scalar L1 Data Cache per SQC */
 350		.cache_size = 16,
 351		.cache_level = 1,
 352		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 353				CRAT_CACHE_FLAGS_DATA_CACHE |
 354				CRAT_CACHE_FLAGS_SIMD_CACHE),
 355		.num_cu_shared = 2,
 356	},
 357	{
 358		/* L2 Data Cache per GPU (Total Tex Cache) */
 359		.cache_size = 8192,
 360		.cache_level = 2,
 361		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 362				CRAT_CACHE_FLAGS_DATA_CACHE |
 363				CRAT_CACHE_FLAGS_SIMD_CACHE),
 364		.num_cu_shared = 14,
 365	},
 366};
 367
 368static struct kfd_gpu_cache_info navi10_cache_info[] = {
 369	{
 370		/* TCP L1 Cache per CU */
 371		.cache_size = 16,
 372		.cache_level = 1,
 373		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 374				CRAT_CACHE_FLAGS_DATA_CACHE |
 375				CRAT_CACHE_FLAGS_SIMD_CACHE),
 376		.num_cu_shared = 1,
 377	},
 378	{
 379		/* Scalar L1 Instruction Cache per SQC */
 380		.cache_size = 32,
 381		.cache_level = 1,
 382		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 383				CRAT_CACHE_FLAGS_INST_CACHE |
 384				CRAT_CACHE_FLAGS_SIMD_CACHE),
 385		.num_cu_shared = 2,
 386	},
 387	{
 388		/* Scalar L1 Data Cache per SQC */
 389		.cache_size = 16,
 390		.cache_level = 1,
 391		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 392				CRAT_CACHE_FLAGS_DATA_CACHE |
 393				CRAT_CACHE_FLAGS_SIMD_CACHE),
 394		.num_cu_shared = 2,
 395	},
 396	{
 397		/* GL1 Data Cache per SA */
 398		.cache_size = 128,
 399		.cache_level = 1,
 400		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 401				CRAT_CACHE_FLAGS_DATA_CACHE |
 402				CRAT_CACHE_FLAGS_SIMD_CACHE),
 403		.num_cu_shared = 10,
 404	},
 405	{
 406		/* L2 Data Cache per GPU (Total Tex Cache) */
 407		.cache_size = 4096,
 408		.cache_level = 2,
 409		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 410				CRAT_CACHE_FLAGS_DATA_CACHE |
 411				CRAT_CACHE_FLAGS_SIMD_CACHE),
 412		.num_cu_shared = 10,
 413	},
 414};
 415
 416static struct kfd_gpu_cache_info vangogh_cache_info[] = {
 417	{
 418		/* TCP L1 Cache per CU */
 419		.cache_size = 16,
 420		.cache_level = 1,
 421		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 422				CRAT_CACHE_FLAGS_DATA_CACHE |
 423				CRAT_CACHE_FLAGS_SIMD_CACHE),
 424		.num_cu_shared = 1,
 425	},
 426	{
 427		/* Scalar L1 Instruction Cache per SQC */
 428		.cache_size = 32,
 429		.cache_level = 1,
 430		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 431				CRAT_CACHE_FLAGS_INST_CACHE |
 432				CRAT_CACHE_FLAGS_SIMD_CACHE),
 433		.num_cu_shared = 2,
 434	},
 435	{
 436		/* Scalar L1 Data Cache per SQC */
 437		.cache_size = 16,
 438		.cache_level = 1,
 439		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 440				CRAT_CACHE_FLAGS_DATA_CACHE |
 441				CRAT_CACHE_FLAGS_SIMD_CACHE),
 442		.num_cu_shared = 2,
 443	},
 444	{
 445		/* GL1 Data Cache per SA */
 446		.cache_size = 128,
 447		.cache_level = 1,
 448		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 449				CRAT_CACHE_FLAGS_DATA_CACHE |
 450				CRAT_CACHE_FLAGS_SIMD_CACHE),
 451		.num_cu_shared = 8,
 452	},
 453	{
 454		/* L2 Data Cache per GPU (Total Tex Cache) */
 455		.cache_size = 1024,
 456		.cache_level = 2,
 457		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 458				CRAT_CACHE_FLAGS_DATA_CACHE |
 459				CRAT_CACHE_FLAGS_SIMD_CACHE),
 460		.num_cu_shared = 8,
 461	},
 462};
 463
 464static struct kfd_gpu_cache_info navi14_cache_info[] = {
 465	{
 466		/* TCP L1 Cache per CU */
 467		.cache_size = 16,
 468		.cache_level = 1,
 469		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 470				CRAT_CACHE_FLAGS_DATA_CACHE |
 471				CRAT_CACHE_FLAGS_SIMD_CACHE),
 472		.num_cu_shared = 1,
 473	},
 474	{
 475		/* Scalar L1 Instruction Cache per SQC */
 476		.cache_size = 32,
 477		.cache_level = 1,
 478		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 479				CRAT_CACHE_FLAGS_INST_CACHE |
 480				CRAT_CACHE_FLAGS_SIMD_CACHE),
 481		.num_cu_shared = 2,
 482	},
 483	{
 484		/* Scalar L1 Data Cache per SQC */
 485		.cache_size = 16,
 486		.cache_level = 1,
 487		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 488				CRAT_CACHE_FLAGS_DATA_CACHE |
 489				CRAT_CACHE_FLAGS_SIMD_CACHE),
 490		.num_cu_shared = 2,
 491	},
 492	{
 493		/* GL1 Data Cache per SA */
 494		.cache_size = 128,
 495		.cache_level = 1,
 496		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 497				CRAT_CACHE_FLAGS_DATA_CACHE |
 498				CRAT_CACHE_FLAGS_SIMD_CACHE),
 499		.num_cu_shared = 12,
 500	},
 501	{
 502		/* L2 Data Cache per GPU (Total Tex Cache) */
 503		.cache_size = 2048,
 504		.cache_level = 2,
 505		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 506				CRAT_CACHE_FLAGS_DATA_CACHE |
 507				CRAT_CACHE_FLAGS_SIMD_CACHE),
 508		.num_cu_shared = 12,
 509	},
 510};
 511
 512static struct kfd_gpu_cache_info sienna_cichlid_cache_info[] = {
 513	{
 514		/* TCP L1 Cache per CU */
 515		.cache_size = 16,
 516		.cache_level = 1,
 517		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 518				CRAT_CACHE_FLAGS_DATA_CACHE |
 519				CRAT_CACHE_FLAGS_SIMD_CACHE),
 520		.num_cu_shared = 1,
 521	},
 522	{
 523		/* Scalar L1 Instruction Cache per SQC */
 524		.cache_size = 32,
 525		.cache_level = 1,
 526		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 527				CRAT_CACHE_FLAGS_INST_CACHE |
 528				CRAT_CACHE_FLAGS_SIMD_CACHE),
 529		.num_cu_shared = 2,
 530	},
 531	{
 532		/* Scalar L1 Data Cache per SQC */
 533		.cache_size = 16,
 534		.cache_level = 1,
 535		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 536				CRAT_CACHE_FLAGS_DATA_CACHE |
 537				CRAT_CACHE_FLAGS_SIMD_CACHE),
 538		.num_cu_shared = 2,
 539	},
 540	{
 541		/* GL1 Data Cache per SA */
 542		.cache_size = 128,
 543		.cache_level = 1,
 544		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 545				CRAT_CACHE_FLAGS_DATA_CACHE |
 546				CRAT_CACHE_FLAGS_SIMD_CACHE),
 547		.num_cu_shared = 10,
 548	},
 549	{
 550		/* L2 Data Cache per GPU (Total Tex Cache) */
 551		.cache_size = 4096,
 552		.cache_level = 2,
 553		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 554				CRAT_CACHE_FLAGS_DATA_CACHE |
 555				CRAT_CACHE_FLAGS_SIMD_CACHE),
 556		.num_cu_shared = 10,
 557	},
 558	{
 559		/* L3 Data Cache per GPU */
 560		.cache_size = 128*1024,
 561		.cache_level = 3,
 562		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 563				CRAT_CACHE_FLAGS_DATA_CACHE |
 564				CRAT_CACHE_FLAGS_SIMD_CACHE),
 565		.num_cu_shared = 10,
 566	},
 567};
 568
 569static struct kfd_gpu_cache_info navy_flounder_cache_info[] = {
 570	{
 571		/* TCP L1 Cache per CU */
 572		.cache_size = 16,
 573		.cache_level = 1,
 574		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 575				CRAT_CACHE_FLAGS_DATA_CACHE |
 576				CRAT_CACHE_FLAGS_SIMD_CACHE),
 577		.num_cu_shared = 1,
 578	},
 579	{
 580		/* Scalar L1 Instruction Cache per SQC */
 581		.cache_size = 32,
 582		.cache_level = 1,
 583		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 584				CRAT_CACHE_FLAGS_INST_CACHE |
 585				CRAT_CACHE_FLAGS_SIMD_CACHE),
 586		.num_cu_shared = 2,
 587	},
 588	{
 589		/* Scalar L1 Data Cache per SQC */
 590		.cache_size = 16,
 591		.cache_level = 1,
 592		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 593				CRAT_CACHE_FLAGS_DATA_CACHE |
 594				CRAT_CACHE_FLAGS_SIMD_CACHE),
 595		.num_cu_shared = 2,
 596	},
 597	{
 598		/* GL1 Data Cache per SA */
 599		.cache_size = 128,
 600		.cache_level = 1,
 601		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 602				CRAT_CACHE_FLAGS_DATA_CACHE |
 603				CRAT_CACHE_FLAGS_SIMD_CACHE),
 604		.num_cu_shared = 10,
 605	},
 606	{
 607		/* L2 Data Cache per GPU (Total Tex Cache) */
 608		.cache_size = 3072,
 609		.cache_level = 2,
 610		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 611				CRAT_CACHE_FLAGS_DATA_CACHE |
 612				CRAT_CACHE_FLAGS_SIMD_CACHE),
 613		.num_cu_shared = 10,
 614	},
 615	{
 616		/* L3 Data Cache per GPU */
 617		.cache_size = 96*1024,
 618		.cache_level = 3,
 619		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 620				CRAT_CACHE_FLAGS_DATA_CACHE |
 621				CRAT_CACHE_FLAGS_SIMD_CACHE),
 622		.num_cu_shared = 10,
 623	},
 624};
 625
 626static struct kfd_gpu_cache_info dimgrey_cavefish_cache_info[] = {
 627	{
 628		/* TCP L1 Cache per CU */
 629		.cache_size = 16,
 630		.cache_level = 1,
 631		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 632				CRAT_CACHE_FLAGS_DATA_CACHE |
 633				CRAT_CACHE_FLAGS_SIMD_CACHE),
 634		.num_cu_shared = 1,
 635	},
 636	{
 637		/* Scalar L1 Instruction Cache per SQC */
 638		.cache_size = 32,
 639		.cache_level = 1,
 640		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 641				CRAT_CACHE_FLAGS_INST_CACHE |
 642				CRAT_CACHE_FLAGS_SIMD_CACHE),
 643		.num_cu_shared = 2,
 644	},
 645	{
 646		/* Scalar L1 Data Cache per SQC */
 647		.cache_size = 16,
 648		.cache_level = 1,
 649		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 650				CRAT_CACHE_FLAGS_DATA_CACHE |
 651				CRAT_CACHE_FLAGS_SIMD_CACHE),
 652		.num_cu_shared = 2,
 653	},
 654	{
 655		/* GL1 Data Cache per SA */
 656		.cache_size = 128,
 657		.cache_level = 1,
 658		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 659				CRAT_CACHE_FLAGS_DATA_CACHE |
 660				CRAT_CACHE_FLAGS_SIMD_CACHE),
 661		.num_cu_shared = 8,
 662	},
 663	{
 664		/* L2 Data Cache per GPU (Total Tex Cache) */
 665		.cache_size = 2048,
 666		.cache_level = 2,
 667		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 668				CRAT_CACHE_FLAGS_DATA_CACHE |
 669				CRAT_CACHE_FLAGS_SIMD_CACHE),
 670		.num_cu_shared = 8,
 671	},
 672	{
 673		/* L3 Data Cache per GPU */
 674		.cache_size = 32*1024,
 675		.cache_level = 3,
 676		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 677				CRAT_CACHE_FLAGS_DATA_CACHE |
 678				CRAT_CACHE_FLAGS_SIMD_CACHE),
 679		.num_cu_shared = 8,
 680	},
 681};
 682
 683static struct kfd_gpu_cache_info beige_goby_cache_info[] = {
 684	{
 685		/* TCP L1 Cache per CU */
 686		.cache_size = 16,
 687		.cache_level = 1,
 688		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 689				CRAT_CACHE_FLAGS_DATA_CACHE |
 690				CRAT_CACHE_FLAGS_SIMD_CACHE),
 691		.num_cu_shared = 1,
 692	},
 693	{
 694		/* Scalar L1 Instruction Cache per SQC */
 695		.cache_size = 32,
 696		.cache_level = 1,
 697		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 698				CRAT_CACHE_FLAGS_INST_CACHE |
 699				CRAT_CACHE_FLAGS_SIMD_CACHE),
 700		.num_cu_shared = 2,
 701	},
 702	{
 703		/* Scalar L1 Data Cache per SQC */
 704		.cache_size = 16,
 705		.cache_level = 1,
 706		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 707				CRAT_CACHE_FLAGS_DATA_CACHE |
 708				CRAT_CACHE_FLAGS_SIMD_CACHE),
 709		.num_cu_shared = 2,
 710	},
 711	{
 712		/* GL1 Data Cache per SA */
 713		.cache_size = 128,
 714		.cache_level = 1,
 715		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 716				CRAT_CACHE_FLAGS_DATA_CACHE |
 717				CRAT_CACHE_FLAGS_SIMD_CACHE),
 718		.num_cu_shared = 8,
 719	},
 720	{
 721		/* L2 Data Cache per GPU (Total Tex Cache) */
 722		.cache_size = 1024,
 723		.cache_level = 2,
 724		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 725				CRAT_CACHE_FLAGS_DATA_CACHE |
 726				CRAT_CACHE_FLAGS_SIMD_CACHE),
 727		.num_cu_shared = 8,
 728	},
 729	{
 730		/* L3 Data Cache per GPU */
 731		.cache_size = 16*1024,
 732		.cache_level = 3,
 733		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 734				CRAT_CACHE_FLAGS_DATA_CACHE |
 735				CRAT_CACHE_FLAGS_SIMD_CACHE),
 736		.num_cu_shared = 8,
 737	},
 738};
 739
 740static struct kfd_gpu_cache_info yellow_carp_cache_info[] = {
 741	{
 742		/* TCP L1 Cache per CU */
 743		.cache_size = 16,
 744		.cache_level = 1,
 745		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 746				CRAT_CACHE_FLAGS_DATA_CACHE |
 747				CRAT_CACHE_FLAGS_SIMD_CACHE),
 748		.num_cu_shared = 1,
 749	},
 750	{
 751		/* Scalar L1 Instruction Cache per SQC */
 752		.cache_size = 32,
 753		.cache_level = 1,
 754		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 755				CRAT_CACHE_FLAGS_INST_CACHE |
 756				CRAT_CACHE_FLAGS_SIMD_CACHE),
 757		.num_cu_shared = 2,
 758	},
 759	{
 760		/* Scalar L1 Data Cache per SQC */
 761		.cache_size = 16,
 762		.cache_level = 1,
 763		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 764				CRAT_CACHE_FLAGS_DATA_CACHE |
 765				CRAT_CACHE_FLAGS_SIMD_CACHE),
 766		.num_cu_shared = 2,
 767	},
 768	{
 769		/* GL1 Data Cache per SA */
 770		.cache_size = 128,
 771		.cache_level = 1,
 772		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 773				CRAT_CACHE_FLAGS_DATA_CACHE |
 774				CRAT_CACHE_FLAGS_SIMD_CACHE),
 775		.num_cu_shared = 6,
 776	},
 777	{
 778		/* L2 Data Cache per GPU (Total Tex Cache) */
 779		.cache_size = 2048,
 780		.cache_level = 2,
 781		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 782				CRAT_CACHE_FLAGS_DATA_CACHE |
 783				CRAT_CACHE_FLAGS_SIMD_CACHE),
 784		.num_cu_shared = 6,
 785	},
 786};
 787
 788static struct kfd_gpu_cache_info gfx1037_cache_info[] = {
 789	{
 790		/* TCP L1 Cache per CU */
 791		.cache_size = 16,
 792		.cache_level = 1,
 793		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 794				CRAT_CACHE_FLAGS_DATA_CACHE |
 795				CRAT_CACHE_FLAGS_SIMD_CACHE),
 796		.num_cu_shared = 1,
 797	},
 798	{
 799		/* Scalar L1 Instruction Cache per SQC */
 800		.cache_size = 32,
 801		.cache_level = 1,
 802		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 803				CRAT_CACHE_FLAGS_INST_CACHE |
 804				CRAT_CACHE_FLAGS_SIMD_CACHE),
 805		.num_cu_shared = 2,
 806	},
 807	{
 808		/* Scalar L1 Data Cache per SQC */
 809		.cache_size = 16,
 810		.cache_level = 1,
 811		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 812				CRAT_CACHE_FLAGS_DATA_CACHE |
 813				CRAT_CACHE_FLAGS_SIMD_CACHE),
 814		.num_cu_shared = 2,
 815	},
 816	{
 817		/* GL1 Data Cache per SA */
 818		.cache_size = 128,
 819		.cache_level = 1,
 820		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 821				CRAT_CACHE_FLAGS_DATA_CACHE |
 822				CRAT_CACHE_FLAGS_SIMD_CACHE),
 823		.num_cu_shared = 2,
 824	},
 825	{
 826		/* L2 Data Cache per GPU (Total Tex Cache) */
 827		.cache_size = 256,
 828		.cache_level = 2,
 829		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 830				CRAT_CACHE_FLAGS_DATA_CACHE |
 831				CRAT_CACHE_FLAGS_SIMD_CACHE),
 832		.num_cu_shared = 2,
 833	},
 834};
 835
 836static struct kfd_gpu_cache_info gc_10_3_6_cache_info[] = {
 837	{
 838		/* TCP L1 Cache per CU */
 839		.cache_size = 16,
 840		.cache_level = 1,
 841		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 842			  CRAT_CACHE_FLAGS_DATA_CACHE |
 843			  CRAT_CACHE_FLAGS_SIMD_CACHE),
 844		.num_cu_shared = 1,
 845	},
 846	{
 847		/* Scalar L1 Instruction Cache per SQC */
 848		.cache_size = 32,
 849		.cache_level = 1,
 850		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 851			  CRAT_CACHE_FLAGS_INST_CACHE |
 852			  CRAT_CACHE_FLAGS_SIMD_CACHE),
 853		.num_cu_shared = 2,
 854	},
 855	{
 856		/* Scalar L1 Data Cache per SQC */
 857		.cache_size = 16,
 858		.cache_level = 1,
 859		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 860			  CRAT_CACHE_FLAGS_DATA_CACHE |
 861			  CRAT_CACHE_FLAGS_SIMD_CACHE),
 862		.num_cu_shared = 2,
 863	},
 864	{
 865		/* GL1 Data Cache per SA */
 866		.cache_size = 128,
 867		.cache_level = 1,
 868		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 869			  CRAT_CACHE_FLAGS_DATA_CACHE |
 870			  CRAT_CACHE_FLAGS_SIMD_CACHE),
 871		.num_cu_shared = 2,
 872	},
 873	{
 874		/* L2 Data Cache per GPU (Total Tex Cache) */
 875		.cache_size = 256,
 876		.cache_level = 2,
 877		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 878			  CRAT_CACHE_FLAGS_DATA_CACHE |
 879			  CRAT_CACHE_FLAGS_SIMD_CACHE),
 880		.num_cu_shared = 2,
 881	},
 882};
 883
 884static struct kfd_gpu_cache_info dummy_cache_info[] = {
 885	{
 886		/* TCP L1 Cache per CU */
 887		.cache_size = 16,
 888		.cache_level = 1,
 889		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 890				CRAT_CACHE_FLAGS_DATA_CACHE |
 891				CRAT_CACHE_FLAGS_SIMD_CACHE),
 892		.num_cu_shared = 1,
 893	},
 894	{
 895		/* Scalar L1 Instruction Cache per SQC */
 896		.cache_size = 32,
 897		.cache_level = 1,
 898		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 899				CRAT_CACHE_FLAGS_INST_CACHE |
 900				CRAT_CACHE_FLAGS_SIMD_CACHE),
 901		.num_cu_shared = 2,
 902	},
 903	{
 904		/* Scalar L1 Data Cache per SQC */
 905		.cache_size = 16,
 906		.cache_level = 1,
 907		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 908				CRAT_CACHE_FLAGS_DATA_CACHE |
 909				CRAT_CACHE_FLAGS_SIMD_CACHE),
 910		.num_cu_shared = 2,
 911	},
 912	{
 913		/* GL1 Data Cache per SA */
 914		.cache_size = 128,
 915		.cache_level = 1,
 916		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 917				CRAT_CACHE_FLAGS_DATA_CACHE |
 918				CRAT_CACHE_FLAGS_SIMD_CACHE),
 919		.num_cu_shared = 6,
 920	},
 921	{
 922		/* L2 Data Cache per GPU (Total Tex Cache) */
 923		.cache_size = 2048,
 924		.cache_level = 2,
 925		.flags = (CRAT_CACHE_FLAGS_ENABLED |
 926				CRAT_CACHE_FLAGS_DATA_CACHE |
 927				CRAT_CACHE_FLAGS_SIMD_CACHE),
 928		.num_cu_shared = 6,
 929	},
 930};
 931
 932static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
 933		struct crat_subtype_computeunit *cu)
 934{
 935	dev->node_props.cpu_cores_count = cu->num_cpu_cores;
 936	dev->node_props.cpu_core_id_base = cu->processor_id_low;
 937	if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
 938		dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
 939
 940	pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
 941			cu->processor_id_low);
 942}
 943
 944static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
 945		struct crat_subtype_computeunit *cu)
 946{
 947	dev->node_props.simd_id_base = cu->processor_id_low;
 948	dev->node_props.simd_count = cu->num_simd_cores;
 949	dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
 950	dev->node_props.max_waves_per_simd = cu->max_waves_simd;
 951	dev->node_props.wave_front_size = cu->wave_front_size;
 952	dev->node_props.array_count = cu->array_count;
 953	dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
 954	dev->node_props.simd_per_cu = cu->num_simd_per_cu;
 955	dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
 956	if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
 957		dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
 958	pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low);
 959}
 960
 961/* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct
 962 * topology device present in the device_list
 963 */
 964static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
 965				struct list_head *device_list)
 966{
 967	struct kfd_topology_device *dev;
 968
 969	pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
 970			cu->proximity_domain, cu->hsa_capability);
 971	list_for_each_entry(dev, device_list, list) {
 972		if (cu->proximity_domain == dev->proximity_domain) {
 973			if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
 974				kfd_populated_cu_info_cpu(dev, cu);
 975
 976			if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
 977				kfd_populated_cu_info_gpu(dev, cu);
 978			break;
 979		}
 980	}
 981
 982	return 0;
 983}
 984
 985static struct kfd_mem_properties *
 986find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width,
 987		struct kfd_topology_device *dev)
 988{
 989	struct kfd_mem_properties *props;
 990
 991	list_for_each_entry(props, &dev->mem_props, list) {
 992		if (props->heap_type == heap_type
 993				&& props->flags == flags
 994				&& props->width == width)
 995			return props;
 996	}
 997
 998	return NULL;
 999}
1000/* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
1001 * topology device present in the device_list
1002 */
1003static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
1004				struct list_head *device_list)
1005{
1006	struct kfd_mem_properties *props;
1007	struct kfd_topology_device *dev;
1008	uint32_t heap_type;
1009	uint64_t size_in_bytes;
1010	uint32_t flags = 0;
1011	uint32_t width;
1012
1013	pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
1014			mem->proximity_domain);
1015	list_for_each_entry(dev, device_list, list) {
1016		if (mem->proximity_domain == dev->proximity_domain) {
1017			/* We're on GPU node */
1018			if (dev->node_props.cpu_cores_count == 0) {
1019				/* APU */
1020				if (mem->visibility_type == 0)
1021					heap_type =
1022						HSA_MEM_HEAP_TYPE_FB_PRIVATE;
1023				/* dGPU */
1024				else
1025					heap_type = mem->visibility_type;
1026			} else
1027				heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
1028
1029			if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
1030				flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
1031			if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
1032				flags |= HSA_MEM_FLAGS_NON_VOLATILE;
1033
1034			size_in_bytes =
1035				((uint64_t)mem->length_high << 32) +
1036							mem->length_low;
1037			width = mem->width;
1038
1039			/* Multiple banks of the same type are aggregated into
1040			 * one. User mode doesn't care about multiple physical
1041			 * memory segments. It's managed as a single virtual
1042			 * heap for user mode.
1043			 */
1044			props = find_subtype_mem(heap_type, flags, width, dev);
1045			if (props) {
1046				props->size_in_bytes += size_in_bytes;
1047				break;
1048			}
1049
1050			props = kfd_alloc_struct(props);
1051			if (!props)
1052				return -ENOMEM;
1053
1054			props->heap_type = heap_type;
1055			props->flags = flags;
1056			props->size_in_bytes = size_in_bytes;
1057			props->width = width;
1058
1059			dev->node_props.mem_banks_count++;
1060			list_add_tail(&props->list, &dev->mem_props);
1061
1062			break;
1063		}
1064	}
1065
1066	return 0;
1067}
1068
1069/* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct
1070 * topology device present in the device_list
1071 */
1072static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache,
1073			struct list_head *device_list)
1074{
1075	struct kfd_cache_properties *props;
1076	struct kfd_topology_device *dev;
1077	uint32_t id;
1078	uint32_t total_num_of_cu;
1079
1080	id = cache->processor_id_low;
1081
1082	pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id);
1083	list_for_each_entry(dev, device_list, list) {
1084		total_num_of_cu = (dev->node_props.array_count *
1085					dev->node_props.cu_per_simd_array);
1086
1087		/* Cache infomration in CRAT doesn't have proximity_domain
1088		 * information as it is associated with a CPU core or GPU
1089		 * Compute Unit. So map the cache using CPU core Id or SIMD
1090		 * (GPU) ID.
1091		 * TODO: This works because currently we can safely assume that
1092		 *  Compute Units are parsed before caches are parsed. In
1093		 *  future, remove this dependency
1094		 */
1095		if ((id >= dev->node_props.cpu_core_id_base &&
1096			id <= dev->node_props.cpu_core_id_base +
1097				dev->node_props.cpu_cores_count) ||
1098			(id >= dev->node_props.simd_id_base &&
1099			id < dev->node_props.simd_id_base +
1100				total_num_of_cu)) {
1101			props = kfd_alloc_struct(props);
1102			if (!props)
1103				return -ENOMEM;
1104
1105			props->processor_id_low = id;
1106			props->cache_level = cache->cache_level;
1107			props->cache_size = cache->cache_size;
1108			props->cacheline_size = cache->cache_line_size;
1109			props->cachelines_per_tag = cache->lines_per_tag;
1110			props->cache_assoc = cache->associativity;
1111			props->cache_latency = cache->cache_latency;
1112
1113			memcpy(props->sibling_map, cache->sibling_map,
1114					CRAT_SIBLINGMAP_SIZE);
1115
1116			/* set the sibling_map_size as 32 for CRAT from ACPI */
1117			props->sibling_map_size = CRAT_SIBLINGMAP_SIZE;
1118
1119			if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1120				props->cache_type |= HSA_CACHE_TYPE_DATA;
1121			if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
1122				props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1123			if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1124				props->cache_type |= HSA_CACHE_TYPE_CPU;
1125			if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1126				props->cache_type |= HSA_CACHE_TYPE_HSACU;
1127
1128			dev->node_props.caches_count++;
1129			list_add_tail(&props->list, &dev->cache_props);
1130
1131			break;
1132		}
1133	}
1134
1135	return 0;
1136}
1137
1138/* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct
1139 * topology device present in the device_list
1140 */
1141static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
1142					struct list_head *device_list)
1143{
1144	struct kfd_iolink_properties *props = NULL, *props2;
1145	struct kfd_topology_device *dev, *to_dev;
1146	uint32_t id_from;
1147	uint32_t id_to;
1148
1149	id_from = iolink->proximity_domain_from;
1150	id_to = iolink->proximity_domain_to;
1151
1152	pr_debug("Found IO link entry in CRAT table with id_from=%d, id_to %d\n",
1153			id_from, id_to);
1154	list_for_each_entry(dev, device_list, list) {
1155		if (id_from == dev->proximity_domain) {
1156			props = kfd_alloc_struct(props);
1157			if (!props)
1158				return -ENOMEM;
1159
1160			props->node_from = id_from;
1161			props->node_to = id_to;
1162			props->ver_maj = iolink->version_major;
1163			props->ver_min = iolink->version_minor;
1164			props->iolink_type = iolink->io_interface_type;
1165
1166			if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1167				props->weight = 20;
1168			else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1169				props->weight = 15 * iolink->num_hops_xgmi;
1170			else
1171				props->weight = node_distance(id_from, id_to);
1172
1173			props->min_latency = iolink->minimum_latency;
1174			props->max_latency = iolink->maximum_latency;
1175			props->min_bandwidth = iolink->minimum_bandwidth_mbs;
1176			props->max_bandwidth = iolink->maximum_bandwidth_mbs;
1177			props->rec_transfer_size =
1178					iolink->recommended_transfer_size;
1179
1180			dev->node_props.io_links_count++;
1181			list_add_tail(&props->list, &dev->io_link_props);
1182			break;
1183		}
1184	}
1185
1186	/* CPU topology is created before GPUs are detected, so CPU->GPU
1187	 * links are not built at that time. If a PCIe type is discovered, it
1188	 * means a GPU is detected and we are adding GPU->CPU to the topology.
1189	 * At this time, also add the corresponded CPU->GPU link if GPU
1190	 * is large bar.
1191	 * For xGMI, we only added the link with one direction in the crat
1192	 * table, add corresponded reversed direction link now.
1193	 */
1194	if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) {
1195		to_dev = kfd_topology_device_by_proximity_domain_no_lock(id_to);
1196		if (!to_dev)
1197			return -ENODEV;
1198		/* same everything but the other direction */
1199		props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL);
1200		if (!props2)
1201			return -ENOMEM;
1202
1203		props2->node_from = id_to;
1204		props2->node_to = id_from;
1205		props2->kobj = NULL;
1206		to_dev->node_props.io_links_count++;
1207		list_add_tail(&props2->list, &to_dev->io_link_props);
1208	}
1209
1210	return 0;
1211}
1212
1213/* kfd_parse_subtype - parse subtypes and attach it to correct topology device
1214 * present in the device_list
1215 *	@sub_type_hdr - subtype section of crat_image
1216 *	@device_list - list of topology devices present in this crat_image
1217 */
1218static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr,
1219				struct list_head *device_list)
1220{
1221	struct crat_subtype_computeunit *cu;
1222	struct crat_subtype_memory *mem;
1223	struct crat_subtype_cache *cache;
1224	struct crat_subtype_iolink *iolink;
1225	int ret = 0;
1226
1227	switch (sub_type_hdr->type) {
1228	case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
1229		cu = (struct crat_subtype_computeunit *)sub_type_hdr;
1230		ret = kfd_parse_subtype_cu(cu, device_list);
1231		break;
1232	case CRAT_SUBTYPE_MEMORY_AFFINITY:
1233		mem = (struct crat_subtype_memory *)sub_type_hdr;
1234		ret = kfd_parse_subtype_mem(mem, device_list);
1235		break;
1236	case CRAT_SUBTYPE_CACHE_AFFINITY:
1237		cache = (struct crat_subtype_cache *)sub_type_hdr;
1238		ret = kfd_parse_subtype_cache(cache, device_list);
1239		break;
1240	case CRAT_SUBTYPE_TLB_AFFINITY:
1241		/*
1242		 * For now, nothing to do here
1243		 */
1244		pr_debug("Found TLB entry in CRAT table (not processing)\n");
1245		break;
1246	case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
1247		/*
1248		 * For now, nothing to do here
1249		 */
1250		pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n");
1251		break;
1252	case CRAT_SUBTYPE_IOLINK_AFFINITY:
1253		iolink = (struct crat_subtype_iolink *)sub_type_hdr;
1254		ret = kfd_parse_subtype_iolink(iolink, device_list);
1255		break;
1256	default:
1257		pr_warn("Unknown subtype %d in CRAT\n",
1258				sub_type_hdr->type);
1259	}
1260
1261	return ret;
1262}
1263
1264/* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT
1265 * create a kfd_topology_device and add in to device_list. Also parse
1266 * CRAT subtypes and attach it to appropriate kfd_topology_device
1267 *	@crat_image - input image containing CRAT
1268 *	@device_list - [OUT] list of kfd_topology_device generated after
1269 *		       parsing crat_image
1270 *	@proximity_domain - Proximity domain of the first device in the table
1271 *
1272 *	Return - 0 if successful else -ve value
1273 */
1274int kfd_parse_crat_table(void *crat_image, struct list_head *device_list,
1275			 uint32_t proximity_domain)
1276{
1277	struct kfd_topology_device *top_dev = NULL;
1278	struct crat_subtype_generic *sub_type_hdr;
1279	uint16_t node_id;
1280	int ret = 0;
1281	struct crat_header *crat_table = (struct crat_header *)crat_image;
1282	uint16_t num_nodes;
1283	uint32_t image_len;
1284
1285	if (!crat_image)
1286		return -EINVAL;
1287
1288	if (!list_empty(device_list)) {
1289		pr_warn("Error device list should be empty\n");
1290		return -EINVAL;
1291	}
1292
1293	num_nodes = crat_table->num_domains;
1294	image_len = crat_table->length;
1295
1296	pr_debug("Parsing CRAT table with %d nodes\n", num_nodes);
1297
1298	for (node_id = 0; node_id < num_nodes; node_id++) {
1299		top_dev = kfd_create_topology_device(device_list);
1300		if (!top_dev)
1301			break;
1302		top_dev->proximity_domain = proximity_domain++;
1303	}
1304
1305	if (!top_dev) {
1306		ret = -ENOMEM;
1307		goto err;
1308	}
1309
1310	memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH);
1311	memcpy(top_dev->oem_table_id, crat_table->oem_table_id,
1312			CRAT_OEMTABLEID_LENGTH);
1313	top_dev->oem_revision = crat_table->oem_revision;
1314
1315	sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1316	while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
1317			((char *)crat_image) + image_len) {
1318		if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
1319			ret = kfd_parse_subtype(sub_type_hdr, device_list);
1320			if (ret)
1321				break;
1322		}
1323
1324		sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1325				sub_type_hdr->length);
1326	}
1327
1328err:
1329	if (ret)
1330		kfd_release_topology_device_list(device_list);
1331
1332	return ret;
1333}
1334
1335
1336static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
1337						   struct kfd_gpu_cache_info *pcache_info)
1338{
1339	struct amdgpu_device *adev = kdev->adev;
1340	int i = 0;
1341
1342	/* TCP L1 Cache per CU */
1343	if (adev->gfx.config.gc_tcp_l1_size) {
1344		pcache_info[i].cache_size = adev->gfx.config.gc_tcp_l1_size;
1345		pcache_info[i].cache_level = 1;
1346		pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1347					CRAT_CACHE_FLAGS_DATA_CACHE |
1348					CRAT_CACHE_FLAGS_SIMD_CACHE);
1349		pcache_info[0].num_cu_shared = adev->gfx.config.gc_num_tcp_per_wpg / 2;
1350		i++;
1351	}
1352	/* Scalar L1 Instruction Cache per SQC */
1353	if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) {
1354		pcache_info[i].cache_size =
1355			adev->gfx.config.gc_l1_instruction_cache_size_per_sqc;
1356		pcache_info[i].cache_level = 1;
1357		pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1358					CRAT_CACHE_FLAGS_INST_CACHE |
1359					CRAT_CACHE_FLAGS_SIMD_CACHE);
1360		pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1361		i++;
1362	}
1363	/* Scalar L1 Data Cache per SQC */
1364	if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) {
1365		pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc;
1366		pcache_info[i].cache_level = 1;
1367		pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1368					CRAT_CACHE_FLAGS_DATA_CACHE |
1369					CRAT_CACHE_FLAGS_SIMD_CACHE);
1370		pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1371		i++;
1372	}
1373	/* GL1 Data Cache per SA */
1374	if (adev->gfx.config.gc_gl1c_per_sa &&
1375	    adev->gfx.config.gc_gl1c_size_per_instance) {
1376		pcache_info[i].cache_size = adev->gfx.config.gc_gl1c_per_sa *
1377			adev->gfx.config.gc_gl1c_size_per_instance;
1378		pcache_info[i].cache_level = 1;
1379		pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1380					CRAT_CACHE_FLAGS_DATA_CACHE |
1381					CRAT_CACHE_FLAGS_SIMD_CACHE);
1382		pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1383		i++;
1384	}
1385	/* L2 Data Cache per GPU (Total Tex Cache) */
1386	if (adev->gfx.config.gc_gl2c_per_gpu) {
1387		pcache_info[i].cache_size = adev->gfx.config.gc_gl2c_per_gpu;
1388		pcache_info[i].cache_level = 2;
1389		pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1390					CRAT_CACHE_FLAGS_DATA_CACHE |
1391					CRAT_CACHE_FLAGS_SIMD_CACHE);
1392		pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1393		i++;
1394	}
1395	/* L3 Data Cache per GPU */
1396	if (adev->gmc.mall_size) {
1397		pcache_info[i].cache_size = adev->gmc.mall_size / 1024;
1398		pcache_info[i].cache_level = 3;
1399		pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1400					CRAT_CACHE_FLAGS_DATA_CACHE |
1401					CRAT_CACHE_FLAGS_SIMD_CACHE);
1402		pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1403		i++;
1404	}
1405	return i;
1406}
1407
1408int kfd_get_gpu_cache_info(struct kfd_dev *kdev, struct kfd_gpu_cache_info **pcache_info)
1409{
1410	int num_of_cache_types = 0;
1411
1412	switch (kdev->adev->asic_type) {
1413	case CHIP_KAVERI:
1414		*pcache_info = kaveri_cache_info;
1415		num_of_cache_types = ARRAY_SIZE(kaveri_cache_info);
1416		break;
1417	case CHIP_HAWAII:
1418		*pcache_info = hawaii_cache_info;
1419		num_of_cache_types = ARRAY_SIZE(hawaii_cache_info);
1420		break;
1421	case CHIP_CARRIZO:
1422		*pcache_info = carrizo_cache_info;
1423		num_of_cache_types = ARRAY_SIZE(carrizo_cache_info);
1424		break;
1425	case CHIP_TONGA:
1426		*pcache_info = tonga_cache_info;
1427		num_of_cache_types = ARRAY_SIZE(tonga_cache_info);
1428		break;
1429	case CHIP_FIJI:
1430		*pcache_info = fiji_cache_info;
1431		num_of_cache_types = ARRAY_SIZE(fiji_cache_info);
1432		break;
1433	case CHIP_POLARIS10:
1434		*pcache_info = polaris10_cache_info;
1435		num_of_cache_types = ARRAY_SIZE(polaris10_cache_info);
1436		break;
1437	case CHIP_POLARIS11:
1438		*pcache_info = polaris11_cache_info;
1439		num_of_cache_types = ARRAY_SIZE(polaris11_cache_info);
1440		break;
1441	case CHIP_POLARIS12:
1442		*pcache_info = polaris12_cache_info;
1443		num_of_cache_types = ARRAY_SIZE(polaris12_cache_info);
1444		break;
1445	case CHIP_VEGAM:
1446		*pcache_info = vegam_cache_info;
1447		num_of_cache_types = ARRAY_SIZE(vegam_cache_info);
1448		break;
1449	default:
1450		switch (KFD_GC_VERSION(kdev)) {
1451		case IP_VERSION(9, 0, 1):
1452			*pcache_info = vega10_cache_info;
1453			num_of_cache_types = ARRAY_SIZE(vega10_cache_info);
1454			break;
1455		case IP_VERSION(9, 2, 1):
1456			*pcache_info = vega12_cache_info;
1457			num_of_cache_types = ARRAY_SIZE(vega12_cache_info);
1458			break;
1459		case IP_VERSION(9, 4, 0):
1460		case IP_VERSION(9, 4, 1):
1461			*pcache_info = vega20_cache_info;
1462			num_of_cache_types = ARRAY_SIZE(vega20_cache_info);
1463			break;
1464		case IP_VERSION(9, 4, 2):
1465			*pcache_info = aldebaran_cache_info;
1466			num_of_cache_types = ARRAY_SIZE(aldebaran_cache_info);
1467			break;
1468		case IP_VERSION(9, 1, 0):
1469		case IP_VERSION(9, 2, 2):
1470			*pcache_info = raven_cache_info;
1471			num_of_cache_types = ARRAY_SIZE(raven_cache_info);
1472			break;
1473		case IP_VERSION(9, 3, 0):
1474			*pcache_info = renoir_cache_info;
1475			num_of_cache_types = ARRAY_SIZE(renoir_cache_info);
1476			break;
1477		case IP_VERSION(10, 1, 10):
1478		case IP_VERSION(10, 1, 2):
1479		case IP_VERSION(10, 1, 3):
1480		case IP_VERSION(10, 1, 4):
1481			*pcache_info = navi10_cache_info;
1482			num_of_cache_types = ARRAY_SIZE(navi10_cache_info);
1483			break;
1484		case IP_VERSION(10, 1, 1):
1485			*pcache_info = navi14_cache_info;
1486			num_of_cache_types = ARRAY_SIZE(navi14_cache_info);
1487			break;
1488		case IP_VERSION(10, 3, 0):
1489			*pcache_info = sienna_cichlid_cache_info;
1490			num_of_cache_types = ARRAY_SIZE(sienna_cichlid_cache_info);
1491			break;
1492		case IP_VERSION(10, 3, 2):
1493			*pcache_info = navy_flounder_cache_info;
1494			num_of_cache_types = ARRAY_SIZE(navy_flounder_cache_info);
1495			break;
1496		case IP_VERSION(10, 3, 4):
1497			*pcache_info = dimgrey_cavefish_cache_info;
1498			num_of_cache_types = ARRAY_SIZE(dimgrey_cavefish_cache_info);
1499			break;
1500		case IP_VERSION(10, 3, 1):
1501			*pcache_info = vangogh_cache_info;
1502			num_of_cache_types = ARRAY_SIZE(vangogh_cache_info);
1503			break;
1504		case IP_VERSION(10, 3, 5):
1505			*pcache_info = beige_goby_cache_info;
1506			num_of_cache_types = ARRAY_SIZE(beige_goby_cache_info);
1507			break;
1508		case IP_VERSION(10, 3, 3):
1509			*pcache_info = yellow_carp_cache_info;
1510			num_of_cache_types = ARRAY_SIZE(yellow_carp_cache_info);
1511			break;
1512		case IP_VERSION(10, 3, 6):
1513			*pcache_info = gc_10_3_6_cache_info;
1514			num_of_cache_types = ARRAY_SIZE(gc_10_3_6_cache_info);
1515			break;
1516		case IP_VERSION(10, 3, 7):
1517			*pcache_info = gfx1037_cache_info;
1518			num_of_cache_types = ARRAY_SIZE(gfx1037_cache_info);
1519			break;
1520		case IP_VERSION(11, 0, 0):
1521		case IP_VERSION(11, 0, 1):
1522		case IP_VERSION(11, 0, 2):
1523		case IP_VERSION(11, 0, 3):
1524		case IP_VERSION(11, 0, 4):
1525			num_of_cache_types =
1526				kfd_fill_gpu_cache_info_from_gfx_config(kdev, *pcache_info);
1527			break;
1528		default:
1529			*pcache_info = dummy_cache_info;
1530			num_of_cache_types = ARRAY_SIZE(dummy_cache_info);
1531			pr_warn("dummy cache info is used temporarily and real cache info need update later.\n");
1532			break;
1533		}
1534	}
1535	return num_of_cache_types;
1536}
1537
1538static bool kfd_ignore_crat(void)
1539{
1540	bool ret;
1541
1542	if (ignore_crat)
1543		return true;
1544
1545#ifndef KFD_SUPPORT_IOMMU_V2
1546	ret = true;
1547#else
1548	ret = false;
1549#endif
1550
1551	return ret;
1552}
1553
1554/*
1555 * kfd_create_crat_image_acpi - Allocates memory for CRAT image and
1556 * copies CRAT from ACPI (if available).
1557 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
1558 *
1559 *	@crat_image: CRAT read from ACPI. If no CRAT in ACPI then
1560 *		     crat_image will be NULL
1561 *	@size: [OUT] size of crat_image
1562 *
1563 *	Return 0 if successful else return error code
1564 */
1565int kfd_create_crat_image_acpi(void **crat_image, size_t *size)
1566{
1567	struct acpi_table_header *crat_table;
1568	acpi_status status;
1569	void *pcrat_image;
1570	int rc = 0;
1571
1572	if (!crat_image)
1573		return -EINVAL;
1574
1575	*crat_image = NULL;
1576
1577	if (kfd_ignore_crat()) {
1578		pr_info("CRAT table disabled by module option\n");
1579		return -ENODATA;
1580	}
1581
1582	/* Fetch the CRAT table from ACPI */
1583	status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table);
1584	if (status == AE_NOT_FOUND) {
1585		pr_info("CRAT table not found\n");
1586		return -ENODATA;
1587	} else if (ACPI_FAILURE(status)) {
1588		const char *err = acpi_format_exception(status);
1589
1590		pr_err("CRAT table error: %s\n", err);
1591		return -EINVAL;
1592	}
1593
1594	pcrat_image = kvmalloc(crat_table->length, GFP_KERNEL);
1595	if (!pcrat_image) {
1596		rc = -ENOMEM;
1597		goto out;
1598	}
1599
1600	memcpy(pcrat_image, crat_table, crat_table->length);
1601	*crat_image = pcrat_image;
1602	*size = crat_table->length;
1603out:
1604	acpi_put_table(crat_table);
1605	return rc;
1606}
1607
1608/* Memory required to create Virtual CRAT.
1609 * Since there is no easy way to predict the amount of memory required, the
1610 * following amount is allocated for GPU Virtual CRAT. This is
1611 * expected to cover all known conditions. But to be safe additional check
1612 * is put in the code to ensure we don't overwrite.
1613 */
1614#define VCRAT_SIZE_FOR_GPU	(4 * PAGE_SIZE)
1615
1616/* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node
1617 *
1618 *	@numa_node_id: CPU NUMA node id
1619 *	@avail_size: Available size in the memory
1620 *	@sub_type_hdr: Memory into which compute info will be filled in
1621 *
1622 *	Return 0 if successful else return -ve value
1623 */
1624static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size,
1625				int proximity_domain,
1626				struct crat_subtype_computeunit *sub_type_hdr)
1627{
1628	const struct cpumask *cpumask;
1629
1630	*avail_size -= sizeof(struct crat_subtype_computeunit);
1631	if (*avail_size < 0)
1632		return -ENOMEM;
1633
1634	memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
1635
1636	/* Fill in subtype header data */
1637	sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
1638	sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
1639	sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1640
1641	cpumask = cpumask_of_node(numa_node_id);
1642
1643	/* Fill in CU data */
1644	sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT;
1645	sub_type_hdr->proximity_domain = proximity_domain;
1646	sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id);
1647	if (sub_type_hdr->processor_id_low == -1)
1648		return -EINVAL;
1649
1650	sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask);
1651
1652	return 0;
1653}
1654
1655/* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node
1656 *
1657 *	@numa_node_id: CPU NUMA node id
1658 *	@avail_size: Available size in the memory
1659 *	@sub_type_hdr: Memory into which compute info will be filled in
1660 *
1661 *	Return 0 if successful else return -ve value
1662 */
1663static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size,
1664			int proximity_domain,
1665			struct crat_subtype_memory *sub_type_hdr)
1666{
1667	uint64_t mem_in_bytes = 0;
1668	pg_data_t *pgdat;
1669	int zone_type;
1670
1671	*avail_size -= sizeof(struct crat_subtype_memory);
1672	if (*avail_size < 0)
1673		return -ENOMEM;
1674
1675	memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1676
1677	/* Fill in subtype header data */
1678	sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1679	sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1680	sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1681
1682	/* Fill in Memory Subunit data */
1683
1684	/* Unlike si_meminfo, si_meminfo_node is not exported. So
1685	 * the following lines are duplicated from si_meminfo_node
1686	 * function
1687	 */
1688	pgdat = NODE_DATA(numa_node_id);
1689	for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
1690		mem_in_bytes += zone_managed_pages(&pgdat->node_zones[zone_type]);
1691	mem_in_bytes <<= PAGE_SHIFT;
1692
1693	sub_type_hdr->length_low = lower_32_bits(mem_in_bytes);
1694	sub_type_hdr->length_high = upper_32_bits(mem_in_bytes);
1695	sub_type_hdr->proximity_domain = proximity_domain;
1696
1697	return 0;
1698}
1699
1700#ifdef CONFIG_X86_64
1701static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size,
1702				uint32_t *num_entries,
1703				struct crat_subtype_iolink *sub_type_hdr)
1704{
1705	int nid;
1706	struct cpuinfo_x86 *c = &cpu_data(0);
1707	uint8_t link_type;
1708
1709	if (c->x86_vendor == X86_VENDOR_AMD)
1710		link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT;
1711	else
1712		link_type = CRAT_IOLINK_TYPE_QPI_1_1;
1713
1714	*num_entries = 0;
1715
1716	/* Create IO links from this node to other CPU nodes */
1717	for_each_online_node(nid) {
1718		if (nid == numa_node_id) /* node itself */
1719			continue;
1720
1721		*avail_size -= sizeof(struct crat_subtype_iolink);
1722		if (*avail_size < 0)
1723			return -ENOMEM;
1724
1725		memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1726
1727		/* Fill in subtype header data */
1728		sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1729		sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1730		sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1731
1732		/* Fill in IO link data */
1733		sub_type_hdr->proximity_domain_from = numa_node_id;
1734		sub_type_hdr->proximity_domain_to = nid;
1735		sub_type_hdr->io_interface_type = link_type;
1736
1737		(*num_entries)++;
1738		sub_type_hdr++;
1739	}
1740
1741	return 0;
1742}
1743#endif
1744
1745/* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU
1746 *
1747 *	@pcrat_image: Fill in VCRAT for CPU
1748 *	@size:	[IN] allocated size of crat_image.
1749 *		[OUT] actual size of data filled in crat_image
1750 */
1751static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size)
1752{
1753	struct crat_header *crat_table = (struct crat_header *)pcrat_image;
1754	struct acpi_table_header *acpi_table;
1755	acpi_status status;
1756	struct crat_subtype_generic *sub_type_hdr;
1757	int avail_size = *size;
1758	int numa_node_id;
1759#ifdef CONFIG_X86_64
1760	uint32_t entries = 0;
1761#endif
1762	int ret = 0;
1763
1764	if (!pcrat_image)
1765		return -EINVAL;
1766
1767	/* Fill in CRAT Header.
1768	 * Modify length and total_entries as subunits are added.
1769	 */
1770	avail_size -= sizeof(struct crat_header);
1771	if (avail_size < 0)
1772		return -ENOMEM;
1773
1774	memset(crat_table, 0, sizeof(struct crat_header));
1775	memcpy(&crat_table->signature, CRAT_SIGNATURE,
1776			sizeof(crat_table->signature));
1777	crat_table->length = sizeof(struct crat_header);
1778
1779	status = acpi_get_table("DSDT", 0, &acpi_table);
1780	if (status != AE_OK)
1781		pr_warn("DSDT table not found for OEM information\n");
1782	else {
1783		crat_table->oem_revision = acpi_table->revision;
1784		memcpy(crat_table->oem_id, acpi_table->oem_id,
1785				CRAT_OEMID_LENGTH);
1786		memcpy(crat_table->oem_table_id, acpi_table->oem_table_id,
1787				CRAT_OEMTABLEID_LENGTH);
1788		acpi_put_table(acpi_table);
1789	}
1790	crat_table->total_entries = 0;
1791	crat_table->num_domains = 0;
1792
1793	sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1794
1795	for_each_online_node(numa_node_id) {
1796		if (kfd_numa_node_to_apic_id(numa_node_id) == -1)
1797			continue;
1798
1799		/* Fill in Subtype: Compute Unit */
1800		ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size,
1801			crat_table->num_domains,
1802			(struct crat_subtype_computeunit *)sub_type_hdr);
1803		if (ret < 0)
1804			return ret;
1805		crat_table->length += sub_type_hdr->length;
1806		crat_table->total_entries++;
1807
1808		sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1809			sub_type_hdr->length);
1810
1811		/* Fill in Subtype: Memory */
1812		ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size,
1813			crat_table->num_domains,
1814			(struct crat_subtype_memory *)sub_type_hdr);
1815		if (ret < 0)
1816			return ret;
1817		crat_table->length += sub_type_hdr->length;
1818		crat_table->total_entries++;
1819
1820		sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1821			sub_type_hdr->length);
1822
1823		/* Fill in Subtype: IO Link */
1824#ifdef CONFIG_X86_64
1825		ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size,
1826				&entries,
1827				(struct crat_subtype_iolink *)sub_type_hdr);
1828		if (ret < 0)
1829			return ret;
1830
1831		if (entries) {
1832			crat_table->length += (sub_type_hdr->length * entries);
1833			crat_table->total_entries += entries;
1834
1835			sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1836					sub_type_hdr->length * entries);
1837		}
1838#else
1839		pr_info("IO link not available for non x86 platforms\n");
1840#endif
1841
1842		crat_table->num_domains++;
1843	}
1844
1845	/* TODO: Add cache Subtype for CPU.
1846	 * Currently, CPU cache information is available in function
1847	 * detect_cache_attributes(cpu) defined in the file
1848	 * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not
1849	 * exported and to get the same information the code needs to be
1850	 * duplicated.
1851	 */
1852
1853	*size = crat_table->length;
1854	pr_info("Virtual CRAT table created for CPU\n");
1855
1856	return 0;
1857}
1858
1859static int kfd_fill_gpu_memory_affinity(int *avail_size,
1860		struct kfd_dev *kdev, uint8_t type, uint64_t size,
1861		struct crat_subtype_memory *sub_type_hdr,
1862		uint32_t proximity_domain,
1863		const struct kfd_local_mem_info *local_mem_info)
1864{
1865	*avail_size -= sizeof(struct crat_subtype_memory);
1866	if (*avail_size < 0)
1867		return -ENOMEM;
1868
1869	memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1870	sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1871	sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1872	sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1873
1874	sub_type_hdr->proximity_domain = proximity_domain;
1875
1876	pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n",
1877			type, size);
1878
1879	sub_type_hdr->length_low = lower_32_bits(size);
1880	sub_type_hdr->length_high = upper_32_bits(size);
1881
1882	sub_type_hdr->width = local_mem_info->vram_width;
1883	sub_type_hdr->visibility_type = type;
1884
1885	return 0;
1886}
1887
1888#ifdef CONFIG_ACPI_NUMA
1889static void kfd_find_numa_node_in_srat(struct kfd_dev *kdev)
1890{
1891	struct acpi_table_header *table_header = NULL;
1892	struct acpi_subtable_header *sub_header = NULL;
1893	unsigned long table_end, subtable_len;
1894	u32 pci_id = pci_domain_nr(kdev->adev->pdev->bus) << 16 |
1895			pci_dev_id(kdev->adev->pdev);
1896	u32 bdf;
1897	acpi_status status;
1898	struct acpi_srat_cpu_affinity *cpu;
1899	struct acpi_srat_generic_affinity *gpu;
1900	int pxm = 0, max_pxm = 0;
1901	int numa_node = NUMA_NO_NODE;
1902	bool found = false;
1903
1904	/* Fetch the SRAT table from ACPI */
1905	status = acpi_get_table(ACPI_SIG_SRAT, 0, &table_header);
1906	if (status == AE_NOT_FOUND) {
1907		pr_warn("SRAT table not found\n");
1908		return;
1909	} else if (ACPI_FAILURE(status)) {
1910		const char *err = acpi_format_exception(status);
1911		pr_err("SRAT table error: %s\n", err);
1912		return;
1913	}
1914
1915	table_end = (unsigned long)table_header + table_header->length;
1916
1917	/* Parse all entries looking for a match. */
1918	sub_header = (struct acpi_subtable_header *)
1919			((unsigned long)table_header +
1920			sizeof(struct acpi_table_srat));
1921	subtable_len = sub_header->length;
1922
1923	while (((unsigned long)sub_header) + subtable_len  < table_end) {
1924		/*
1925		 * If length is 0, break from this loop to avoid
1926		 * infinite loop.
1927		 */
1928		if (subtable_len == 0) {
1929			pr_err("SRAT invalid zero length\n");
1930			break;
1931		}
1932
1933		switch (sub_header->type) {
1934		case ACPI_SRAT_TYPE_CPU_AFFINITY:
1935			cpu = (struct acpi_srat_cpu_affinity *)sub_header;
1936			pxm = *((u32 *)cpu->proximity_domain_hi) << 8 |
1937					cpu->proximity_domain_lo;
1938			if (pxm > max_pxm)
1939				max_pxm = pxm;
1940			break;
1941		case ACPI_SRAT_TYPE_GENERIC_AFFINITY:
1942			gpu = (struct acpi_srat_generic_affinity *)sub_header;
1943			bdf = *((u16 *)(&gpu->device_handle[0])) << 16 |
1944					*((u16 *)(&gpu->device_handle[2]));
1945			if (bdf == pci_id) {
1946				found = true;
1947				numa_node = pxm_to_node(gpu->proximity_domain);
1948			}
1949			break;
1950		default:
1951			break;
1952		}
1953
1954		if (found)
1955			break;
1956
1957		sub_header = (struct acpi_subtable_header *)
1958				((unsigned long)sub_header + subtable_len);
1959		subtable_len = sub_header->length;
1960	}
1961
1962	acpi_put_table(table_header);
1963
1964	/* Workaround bad cpu-gpu binding case */
1965	if (found && (numa_node < 0 ||
1966			numa_node > pxm_to_node(max_pxm)))
1967		numa_node = 0;
1968
1969	if (numa_node != NUMA_NO_NODE)
1970		set_dev_node(&kdev->adev->pdev->dev, numa_node);
1971}
1972#endif
1973
1974/* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU
1975 * to its NUMA node
1976 *	@avail_size: Available size in the memory
1977 *	@kdev - [IN] GPU device
1978 *	@sub_type_hdr: Memory into which io link info will be filled in
1979 *	@proximity_domain - proximity domain of the GPU node
1980 *
1981 *	Return 0 if successful else return -ve value
1982 */
1983static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
1984			struct kfd_dev *kdev,
1985			struct crat_subtype_iolink *sub_type_hdr,
1986			uint32_t proximity_domain)
1987{
1988	*avail_size -= sizeof(struct crat_subtype_iolink);
1989	if (*avail_size < 0)
1990		return -ENOMEM;
1991
1992	memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1993
1994	/* Fill in subtype header data */
1995	sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1996	sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1997	sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1998	if (kfd_dev_is_large_bar(kdev))
1999		sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2000
2001	/* Fill in IOLINK subtype.
2002	 * TODO: Fill-in other fields of iolink subtype
2003	 */
2004	if (kdev->adev->gmc.xgmi.connected_to_cpu) {
2005		/*
2006		 * with host gpu xgmi link, host can access gpu memory whether
2007		 * or not pcie bar type is large, so always create bidirectional
2008		 * io link.
2009		 */
2010		sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2011		sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2012		sub_type_hdr->num_hops_xgmi = 1;
2013		if (KFD_GC_VERSION(kdev) == IP_VERSION(9, 4, 2)) {
2014			sub_type_hdr->minimum_bandwidth_mbs =
2015					amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(
2016							kdev->adev, NULL, true);
2017			sub_type_hdr->maximum_bandwidth_mbs =
2018					sub_type_hdr->minimum_bandwidth_mbs;
2019		}
2020	} else {
2021		sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS;
2022		sub_type_hdr->minimum_bandwidth_mbs =
2023				amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, true);
2024		sub_type_hdr->maximum_bandwidth_mbs =
2025				amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, false);
2026	}
2027
2028	sub_type_hdr->proximity_domain_from = proximity_domain;
2029
2030#ifdef CONFIG_ACPI_NUMA
2031	if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE)
2032		kfd_find_numa_node_in_srat(kdev);
2033#endif
2034#ifdef CONFIG_NUMA
2035	if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE)
2036		sub_type_hdr->proximity_domain_to = 0;
2037	else
2038		sub_type_hdr->proximity_domain_to = kdev->adev->pdev->dev.numa_node;
2039#else
2040	sub_type_hdr->proximity_domain_to = 0;
2041#endif
2042	return 0;
2043}
2044
2045static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size,
2046			struct kfd_dev *kdev,
2047			struct kfd_dev *peer_kdev,
2048			struct crat_subtype_iolink *sub_type_hdr,
2049			uint32_t proximity_domain_from,
2050			uint32_t proximity_domain_to)
2051{
2052	*avail_size -= sizeof(struct crat_subtype_iolink);
2053	if (*avail_size < 0)
2054		return -ENOMEM;
2055
2056	memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
2057
2058	sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
2059	sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
2060	sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED |
2061			       CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2062
2063	sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2064	sub_type_hdr->proximity_domain_from = proximity_domain_from;
2065	sub_type_hdr->proximity_domain_to = proximity_domain_to;
2066	sub_type_hdr->num_hops_xgmi =
2067		amdgpu_amdkfd_get_xgmi_hops_count(kdev->adev, peer_kdev->adev);
2068	sub_type_hdr->maximum_bandwidth_mbs =
2069		amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, peer_kdev->adev, false);
2070	sub_type_hdr->minimum_bandwidth_mbs = sub_type_hdr->maximum_bandwidth_mbs ?
2071		amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, NULL, true) : 0;
2072
2073	return 0;
2074}
2075
2076/* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU
2077 *
2078 *	@pcrat_image: Fill in VCRAT for GPU
2079 *	@size:	[IN] allocated size of crat_image.
2080 *		[OUT] actual size of data filled in crat_image
2081 */
2082static int kfd_create_vcrat_image_gpu(void *pcrat_image,
2083				      size_t *size, struct kfd_dev *kdev,
2084				      uint32_t proximity_domain)
2085{
2086	struct crat_header *crat_table = (struct crat_header *)pcrat_image;
2087	struct crat_subtype_generic *sub_type_hdr;
2088	struct kfd_local_mem_info local_mem_info;
2089	struct kfd_topology_device *peer_dev;
2090	struct crat_subtype_computeunit *cu;
2091	struct kfd_cu_info cu_info;
2092	int avail_size = *size;
2093	uint32_t total_num_of_cu;
2094	uint32_t nid = 0;
2095	int ret = 0;
2096
2097	if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU)
2098		return -EINVAL;
2099
2100	/* Fill the CRAT Header.
2101	 * Modify length and total_entries as subunits are added.
2102	 */
2103	avail_size -= sizeof(struct crat_header);
2104	if (avail_size < 0)
2105		return -ENOMEM;
2106
2107	memset(crat_table, 0, sizeof(struct crat_header));
2108
2109	memcpy(&crat_table->signature, CRAT_SIGNATURE,
2110			sizeof(crat_table->signature));
2111	/* Change length as we add more subtypes*/
2112	crat_table->length = sizeof(struct crat_header);
2113	crat_table->num_domains = 1;
2114	crat_table->total_entries = 0;
2115
2116	/* Fill in Subtype: Compute Unit
2117	 * First fill in the sub type header and then sub type data
2118	 */
2119	avail_size -= sizeof(struct crat_subtype_computeunit);
2120	if (avail_size < 0)
2121		return -ENOMEM;
2122
2123	sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1);
2124	memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
2125
2126	sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
2127	sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
2128	sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
2129
2130	/* Fill CU subtype data */
2131	cu = (struct crat_subtype_computeunit *)sub_type_hdr;
2132	cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT;
2133	cu->proximity_domain = proximity_domain;
2134
2135	amdgpu_amdkfd_get_cu_info(kdev->adev, &cu_info);
2136	cu->num_simd_per_cu = cu_info.simd_per_cu;
2137	cu->num_simd_cores = cu_info.simd_per_cu * cu_info.cu_active_number;
2138	cu->max_waves_simd = cu_info.max_waves_per_simd;
2139
2140	cu->wave_front_size = cu_info.wave_front_size;
2141	cu->array_count = cu_info.num_shader_arrays_per_engine *
2142		cu_info.num_shader_engines;
2143	total_num_of_cu = (cu->array_count * cu_info.num_cu_per_sh);
2144	cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu);
2145	cu->num_cu_per_array = cu_info.num_cu_per_sh;
2146	cu->max_slots_scatch_cu = cu_info.max_scratch_slots_per_cu;
2147	cu->num_banks = cu_info.num_shader_engines;
2148	cu->lds_size_in_kb = cu_info.lds_size;
2149
2150	cu->hsa_capability = 0;
2151
2152	/* Check if this node supports IOMMU. During parsing this flag will
2153	 * translate to HSA_CAP_ATS_PRESENT
2154	 */
2155	if (!kfd_iommu_check_device(kdev))
2156		cu->hsa_capability |= CRAT_CU_FLAGS_IOMMU_PRESENT;
2157
2158	crat_table->length += sub_type_hdr->length;
2159	crat_table->total_entries++;
2160
2161	/* Fill in Subtype: Memory. Only on systems with large BAR (no
2162	 * private FB), report memory as public. On other systems
2163	 * report the total FB size (public+private) as a single
2164	 * private heap.
2165	 */
2166	local_mem_info = kdev->local_mem_info;
2167	sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2168			sub_type_hdr->length);
2169
2170	if (debug_largebar)
2171		local_mem_info.local_mem_size_private = 0;
2172
2173	if (local_mem_info.local_mem_size_private == 0)
2174		ret = kfd_fill_gpu_memory_affinity(&avail_size,
2175				kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC,
2176				local_mem_info.local_mem_size_public,
2177				(struct crat_subtype_memory *)sub_type_hdr,
2178				proximity_domain,
2179				&local_mem_info);
2180	else
2181		ret = kfd_fill_gpu_memory_affinity(&avail_size,
2182				kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE,
2183				local_mem_info.local_mem_size_public +
2184				local_mem_info.local_mem_size_private,
2185				(struct crat_subtype_memory *)sub_type_hdr,
2186				proximity_domain,
2187				&local_mem_info);
2188	if (ret < 0)
2189		return ret;
2190
2191	crat_table->length += sizeof(struct crat_subtype_memory);
2192	crat_table->total_entries++;
2193
2194	/* Fill in Subtype: IO_LINKS
2195	 *  Only direct links are added here which is Link from GPU to
2196	 *  its NUMA node. Indirect links are added by userspace.
2197	 */
2198	sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2199		sub_type_hdr->length);
2200	ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev,
2201		(struct crat_subtype_iolink *)sub_type_hdr, proximity_domain);
2202
2203	if (ret < 0)
2204		return ret;
2205
2206	crat_table->length += sub_type_hdr->length;
2207	crat_table->total_entries++;
2208
2209
2210	/* Fill in Subtype: IO_LINKS
2211	 * Direct links from GPU to other GPUs through xGMI.
2212	 * We will loop GPUs that already be processed (with lower value
2213	 * of proximity_domain), add the link for the GPUs with same
2214	 * hive id (from this GPU to other GPU) . The reversed iolink
2215	 * (from other GPU to this GPU) will be added
2216	 * in kfd_parse_subtype_iolink.
2217	 */
2218	if (kdev->hive_id) {
2219		for (nid = 0; nid < proximity_domain; ++nid) {
2220			peer_dev = kfd_topology_device_by_proximity_domain_no_lock(nid);
2221			if (!peer_dev->gpu)
2222				continue;
2223			if (peer_dev->gpu->hive_id != kdev->hive_id)
2224				continue;
2225			sub_type_hdr = (typeof(sub_type_hdr))(
2226				(char *)sub_type_hdr +
2227				sizeof(struct crat_subtype_iolink));
2228			ret = kfd_fill_gpu_xgmi_link_to_gpu(
2229				&avail_size, kdev, peer_dev->gpu,
2230				(struct crat_subtype_iolink *)sub_type_hdr,
2231				proximity_domain, nid);
2232			if (ret < 0)
2233				return ret;
2234			crat_table->length += sub_type_hdr->length;
2235			crat_table->total_entries++;
2236		}
2237	}
2238	*size = crat_table->length;
2239	pr_info("Virtual CRAT table created for GPU\n");
2240
2241	return ret;
2242}
2243
2244/* kfd_create_crat_image_virtual - Allocates memory for CRAT image and
2245 *		creates a Virtual CRAT (VCRAT) image
2246 *
2247 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
2248 *
2249 *	@crat_image: VCRAT image created because ACPI does not have a
2250 *		     CRAT for this device
2251 *	@size: [OUT] size of virtual crat_image
2252 *	@flags:	COMPUTE_UNIT_CPU - Create VCRAT for CPU device
2253 *		COMPUTE_UNIT_GPU - Create VCRAT for GPU
2254 *		(COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU
2255 *			-- this option is not currently implemented.
2256 *			The assumption is that all AMD APUs will have CRAT
2257 *	@kdev: Valid kfd_device required if flags contain COMPUTE_UNIT_GPU
2258 *
2259 *	Return 0 if successful else return -ve value
2260 */
2261int kfd_create_crat_image_virtual(void **crat_image, size_t *size,
2262				  int flags, struct kfd_dev *kdev,
2263				  uint32_t proximity_domain)
2264{
2265	void *pcrat_image = NULL;
2266	int ret = 0, num_nodes;
2267	size_t dyn_size;
2268
2269	if (!crat_image)
2270		return -EINVAL;
2271
2272	*crat_image = NULL;
2273
2274	/* Allocate the CPU Virtual CRAT size based on the number of online
2275	 * nodes. Allocate VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image.
2276	 * This should cover all the current conditions. A check is put not
2277	 * to overwrite beyond allocated size for GPUs
2278	 */
2279	switch (flags) {
2280	case COMPUTE_UNIT_CPU:
2281		num_nodes = num_online_nodes();
2282		dyn_size = sizeof(struct crat_header) +
2283			num_nodes * (sizeof(struct crat_subtype_computeunit) +
2284			sizeof(struct crat_subtype_memory) +
2285			(num_nodes - 1) * sizeof(struct crat_subtype_iolink));
2286		pcrat_image = kvmalloc(dyn_size, GFP_KERNEL);
2287		if (!pcrat_image)
2288			return -ENOMEM;
2289		*size = dyn_size;
2290		pr_debug("CRAT size is %ld", dyn_size);
2291		ret = kfd_create_vcrat_image_cpu(pcrat_image, size);
2292		break;
2293	case COMPUTE_UNIT_GPU:
2294		if (!kdev)
2295			return -EINVAL;
2296		pcrat_image = kvmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL);
2297		if (!pcrat_image)
2298			return -ENOMEM;
2299		*size = VCRAT_SIZE_FOR_GPU;
2300		ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev,
2301						 proximity_domain);
2302		break;
2303	case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU):
2304		/* TODO: */
2305		ret = -EINVAL;
2306		pr_err("VCRAT not implemented for APU\n");
2307		break;
2308	default:
2309		ret = -EINVAL;
2310	}
2311
2312	if (!ret)
2313		*crat_image = pcrat_image;
2314	else
2315		kvfree(pcrat_image);
2316
2317	return ret;
2318}
2319
2320
2321/* kfd_destroy_crat_image
2322 *
2323 *	@crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..)
2324 *
2325 */
2326void kfd_destroy_crat_image(void *crat_image)
2327{
2328	kvfree(crat_image);
2329}