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 /* Fixme: Fake 32GB for 1PNPS1 mode bringup */ 34 #define DUMMY_VRAM_SIZE 31138512896 35 36 /* GPU Processor ID base for dGPUs for which VCRAT needs to be created. 37 * GPU processor ID are expressed with Bit[31]=1. 38 * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs 39 * used in the CRAT. 40 */ 41 static uint32_t gpu_processor_id_low = 0x80001000; 42 43 /* Return the next available gpu_processor_id and increment it for next GPU 44 * @total_cu_count - Total CUs present in the GPU including ones 45 * masked off 46 */ 47 static inline unsigned int get_and_inc_gpu_processor_id( 48 unsigned int total_cu_count) 49 { 50 int current_id = gpu_processor_id_low; 51 52 gpu_processor_id_low += total_cu_count; 53 return current_id; 54 } 55 56 57 static struct kfd_gpu_cache_info kaveri_cache_info[] = { 58 { 59 /* TCP L1 Cache per CU */ 60 .cache_size = 16, 61 .cache_level = 1, 62 .flags = (CRAT_CACHE_FLAGS_ENABLED | 63 CRAT_CACHE_FLAGS_DATA_CACHE | 64 CRAT_CACHE_FLAGS_SIMD_CACHE), 65 .num_cu_shared = 1, 66 }, 67 { 68 /* Scalar L1 Instruction Cache (in SQC module) per bank */ 69 .cache_size = 16, 70 .cache_level = 1, 71 .flags = (CRAT_CACHE_FLAGS_ENABLED | 72 CRAT_CACHE_FLAGS_INST_CACHE | 73 CRAT_CACHE_FLAGS_SIMD_CACHE), 74 .num_cu_shared = 2, 75 }, 76 { 77 /* Scalar L1 Data Cache (in SQC module) per bank */ 78 .cache_size = 8, 79 .cache_level = 1, 80 .flags = (CRAT_CACHE_FLAGS_ENABLED | 81 CRAT_CACHE_FLAGS_DATA_CACHE | 82 CRAT_CACHE_FLAGS_SIMD_CACHE), 83 .num_cu_shared = 2, 84 }, 85 86 /* TODO: Add L2 Cache information */ 87 }; 88 89 90 static struct kfd_gpu_cache_info carrizo_cache_info[] = { 91 { 92 /* TCP L1 Cache per CU */ 93 .cache_size = 16, 94 .cache_level = 1, 95 .flags = (CRAT_CACHE_FLAGS_ENABLED | 96 CRAT_CACHE_FLAGS_DATA_CACHE | 97 CRAT_CACHE_FLAGS_SIMD_CACHE), 98 .num_cu_shared = 1, 99 }, 100 { 101 /* Scalar L1 Instruction Cache (in SQC module) per bank */ 102 .cache_size = 8, 103 .cache_level = 1, 104 .flags = (CRAT_CACHE_FLAGS_ENABLED | 105 CRAT_CACHE_FLAGS_INST_CACHE | 106 CRAT_CACHE_FLAGS_SIMD_CACHE), 107 .num_cu_shared = 4, 108 }, 109 { 110 /* Scalar L1 Data Cache (in SQC module) per bank. */ 111 .cache_size = 4, 112 .cache_level = 1, 113 .flags = (CRAT_CACHE_FLAGS_ENABLED | 114 CRAT_CACHE_FLAGS_DATA_CACHE | 115 CRAT_CACHE_FLAGS_SIMD_CACHE), 116 .num_cu_shared = 4, 117 }, 118 119 /* TODO: Add L2 Cache information */ 120 }; 121 122 #define hawaii_cache_info kaveri_cache_info 123 #define tonga_cache_info carrizo_cache_info 124 #define fiji_cache_info carrizo_cache_info 125 #define polaris10_cache_info carrizo_cache_info 126 #define polaris11_cache_info carrizo_cache_info 127 #define polaris12_cache_info carrizo_cache_info 128 #define vegam_cache_info carrizo_cache_info 129 130 /* NOTE: L1 cache information has been updated and L2/L3 131 * cache information has been added for Vega10 and 132 * newer ASICs. The unit for cache_size is KiB. 133 * In future, check & update cache details 134 * for every new ASIC is required. 135 */ 136 137 static struct kfd_gpu_cache_info vega10_cache_info[] = { 138 { 139 /* TCP L1 Cache per CU */ 140 .cache_size = 16, 141 .cache_level = 1, 142 .flags = (CRAT_CACHE_FLAGS_ENABLED | 143 CRAT_CACHE_FLAGS_DATA_CACHE | 144 CRAT_CACHE_FLAGS_SIMD_CACHE), 145 .num_cu_shared = 1, 146 }, 147 { 148 /* Scalar L1 Instruction Cache per SQC */ 149 .cache_size = 32, 150 .cache_level = 1, 151 .flags = (CRAT_CACHE_FLAGS_ENABLED | 152 CRAT_CACHE_FLAGS_INST_CACHE | 153 CRAT_CACHE_FLAGS_SIMD_CACHE), 154 .num_cu_shared = 3, 155 }, 156 { 157 /* Scalar L1 Data Cache per SQC */ 158 .cache_size = 16, 159 .cache_level = 1, 160 .flags = (CRAT_CACHE_FLAGS_ENABLED | 161 CRAT_CACHE_FLAGS_DATA_CACHE | 162 CRAT_CACHE_FLAGS_SIMD_CACHE), 163 .num_cu_shared = 3, 164 }, 165 { 166 /* L2 Data Cache per GPU (Total Tex Cache) */ 167 .cache_size = 4096, 168 .cache_level = 2, 169 .flags = (CRAT_CACHE_FLAGS_ENABLED | 170 CRAT_CACHE_FLAGS_DATA_CACHE | 171 CRAT_CACHE_FLAGS_SIMD_CACHE), 172 .num_cu_shared = 16, 173 }, 174 }; 175 176 static struct kfd_gpu_cache_info raven_cache_info[] = { 177 { 178 /* TCP L1 Cache per CU */ 179 .cache_size = 16, 180 .cache_level = 1, 181 .flags = (CRAT_CACHE_FLAGS_ENABLED | 182 CRAT_CACHE_FLAGS_DATA_CACHE | 183 CRAT_CACHE_FLAGS_SIMD_CACHE), 184 .num_cu_shared = 1, 185 }, 186 { 187 /* Scalar L1 Instruction Cache per SQC */ 188 .cache_size = 32, 189 .cache_level = 1, 190 .flags = (CRAT_CACHE_FLAGS_ENABLED | 191 CRAT_CACHE_FLAGS_INST_CACHE | 192 CRAT_CACHE_FLAGS_SIMD_CACHE), 193 .num_cu_shared = 3, 194 }, 195 { 196 /* Scalar L1 Data Cache per SQC */ 197 .cache_size = 16, 198 .cache_level = 1, 199 .flags = (CRAT_CACHE_FLAGS_ENABLED | 200 CRAT_CACHE_FLAGS_DATA_CACHE | 201 CRAT_CACHE_FLAGS_SIMD_CACHE), 202 .num_cu_shared = 3, 203 }, 204 { 205 /* L2 Data Cache per GPU (Total Tex Cache) */ 206 .cache_size = 1024, 207 .cache_level = 2, 208 .flags = (CRAT_CACHE_FLAGS_ENABLED | 209 CRAT_CACHE_FLAGS_DATA_CACHE | 210 CRAT_CACHE_FLAGS_SIMD_CACHE), 211 .num_cu_shared = 11, 212 }, 213 }; 214 215 static struct kfd_gpu_cache_info renoir_cache_info[] = { 216 { 217 /* TCP L1 Cache per CU */ 218 .cache_size = 16, 219 .cache_level = 1, 220 .flags = (CRAT_CACHE_FLAGS_ENABLED | 221 CRAT_CACHE_FLAGS_DATA_CACHE | 222 CRAT_CACHE_FLAGS_SIMD_CACHE), 223 .num_cu_shared = 1, 224 }, 225 { 226 /* Scalar L1 Instruction Cache per SQC */ 227 .cache_size = 32, 228 .cache_level = 1, 229 .flags = (CRAT_CACHE_FLAGS_ENABLED | 230 CRAT_CACHE_FLAGS_INST_CACHE | 231 CRAT_CACHE_FLAGS_SIMD_CACHE), 232 .num_cu_shared = 3, 233 }, 234 { 235 /* Scalar L1 Data Cache per SQC */ 236 .cache_size = 16, 237 .cache_level = 1, 238 .flags = (CRAT_CACHE_FLAGS_ENABLED | 239 CRAT_CACHE_FLAGS_DATA_CACHE | 240 CRAT_CACHE_FLAGS_SIMD_CACHE), 241 .num_cu_shared = 3, 242 }, 243 { 244 /* L2 Data Cache per GPU (Total Tex Cache) */ 245 .cache_size = 1024, 246 .cache_level = 2, 247 .flags = (CRAT_CACHE_FLAGS_ENABLED | 248 CRAT_CACHE_FLAGS_DATA_CACHE | 249 CRAT_CACHE_FLAGS_SIMD_CACHE), 250 .num_cu_shared = 8, 251 }, 252 }; 253 254 static struct kfd_gpu_cache_info vega12_cache_info[] = { 255 { 256 /* TCP L1 Cache per CU */ 257 .cache_size = 16, 258 .cache_level = 1, 259 .flags = (CRAT_CACHE_FLAGS_ENABLED | 260 CRAT_CACHE_FLAGS_DATA_CACHE | 261 CRAT_CACHE_FLAGS_SIMD_CACHE), 262 .num_cu_shared = 1, 263 }, 264 { 265 /* Scalar L1 Instruction Cache per SQC */ 266 .cache_size = 32, 267 .cache_level = 1, 268 .flags = (CRAT_CACHE_FLAGS_ENABLED | 269 CRAT_CACHE_FLAGS_INST_CACHE | 270 CRAT_CACHE_FLAGS_SIMD_CACHE), 271 .num_cu_shared = 3, 272 }, 273 { 274 /* Scalar L1 Data Cache per SQC */ 275 .cache_size = 16, 276 .cache_level = 1, 277 .flags = (CRAT_CACHE_FLAGS_ENABLED | 278 CRAT_CACHE_FLAGS_DATA_CACHE | 279 CRAT_CACHE_FLAGS_SIMD_CACHE), 280 .num_cu_shared = 3, 281 }, 282 { 283 /* L2 Data Cache per GPU (Total Tex Cache) */ 284 .cache_size = 2048, 285 .cache_level = 2, 286 .flags = (CRAT_CACHE_FLAGS_ENABLED | 287 CRAT_CACHE_FLAGS_DATA_CACHE | 288 CRAT_CACHE_FLAGS_SIMD_CACHE), 289 .num_cu_shared = 5, 290 }, 291 }; 292 293 static struct kfd_gpu_cache_info vega20_cache_info[] = { 294 { 295 /* TCP L1 Cache per CU */ 296 .cache_size = 16, 297 .cache_level = 1, 298 .flags = (CRAT_CACHE_FLAGS_ENABLED | 299 CRAT_CACHE_FLAGS_DATA_CACHE | 300 CRAT_CACHE_FLAGS_SIMD_CACHE), 301 .num_cu_shared = 1, 302 }, 303 { 304 /* Scalar L1 Instruction Cache per SQC */ 305 .cache_size = 32, 306 .cache_level = 1, 307 .flags = (CRAT_CACHE_FLAGS_ENABLED | 308 CRAT_CACHE_FLAGS_INST_CACHE | 309 CRAT_CACHE_FLAGS_SIMD_CACHE), 310 .num_cu_shared = 3, 311 }, 312 { 313 /* Scalar L1 Data Cache per SQC */ 314 .cache_size = 16, 315 .cache_level = 1, 316 .flags = (CRAT_CACHE_FLAGS_ENABLED | 317 CRAT_CACHE_FLAGS_DATA_CACHE | 318 CRAT_CACHE_FLAGS_SIMD_CACHE), 319 .num_cu_shared = 3, 320 }, 321 { 322 /* L2 Data Cache per GPU (Total Tex Cache) */ 323 .cache_size = 8192, 324 .cache_level = 2, 325 .flags = (CRAT_CACHE_FLAGS_ENABLED | 326 CRAT_CACHE_FLAGS_DATA_CACHE | 327 CRAT_CACHE_FLAGS_SIMD_CACHE), 328 .num_cu_shared = 16, 329 }, 330 }; 331 332 static struct kfd_gpu_cache_info aldebaran_cache_info[] = { 333 { 334 /* TCP L1 Cache per CU */ 335 .cache_size = 16, 336 .cache_level = 1, 337 .flags = (CRAT_CACHE_FLAGS_ENABLED | 338 CRAT_CACHE_FLAGS_DATA_CACHE | 339 CRAT_CACHE_FLAGS_SIMD_CACHE), 340 .num_cu_shared = 1, 341 }, 342 { 343 /* Scalar L1 Instruction Cache per SQC */ 344 .cache_size = 32, 345 .cache_level = 1, 346 .flags = (CRAT_CACHE_FLAGS_ENABLED | 347 CRAT_CACHE_FLAGS_INST_CACHE | 348 CRAT_CACHE_FLAGS_SIMD_CACHE), 349 .num_cu_shared = 2, 350 }, 351 { 352 /* Scalar L1 Data Cache per SQC */ 353 .cache_size = 16, 354 .cache_level = 1, 355 .flags = (CRAT_CACHE_FLAGS_ENABLED | 356 CRAT_CACHE_FLAGS_DATA_CACHE | 357 CRAT_CACHE_FLAGS_SIMD_CACHE), 358 .num_cu_shared = 2, 359 }, 360 { 361 /* L2 Data Cache per GPU (Total Tex Cache) */ 362 .cache_size = 8192, 363 .cache_level = 2, 364 .flags = (CRAT_CACHE_FLAGS_ENABLED | 365 CRAT_CACHE_FLAGS_DATA_CACHE | 366 CRAT_CACHE_FLAGS_SIMD_CACHE), 367 .num_cu_shared = 14, 368 }, 369 }; 370 371 static struct kfd_gpu_cache_info navi10_cache_info[] = { 372 { 373 /* TCP L1 Cache per CU */ 374 .cache_size = 16, 375 .cache_level = 1, 376 .flags = (CRAT_CACHE_FLAGS_ENABLED | 377 CRAT_CACHE_FLAGS_DATA_CACHE | 378 CRAT_CACHE_FLAGS_SIMD_CACHE), 379 .num_cu_shared = 1, 380 }, 381 { 382 /* Scalar L1 Instruction Cache per SQC */ 383 .cache_size = 32, 384 .cache_level = 1, 385 .flags = (CRAT_CACHE_FLAGS_ENABLED | 386 CRAT_CACHE_FLAGS_INST_CACHE | 387 CRAT_CACHE_FLAGS_SIMD_CACHE), 388 .num_cu_shared = 2, 389 }, 390 { 391 /* Scalar L1 Data Cache per SQC */ 392 .cache_size = 16, 393 .cache_level = 1, 394 .flags = (CRAT_CACHE_FLAGS_ENABLED | 395 CRAT_CACHE_FLAGS_DATA_CACHE | 396 CRAT_CACHE_FLAGS_SIMD_CACHE), 397 .num_cu_shared = 2, 398 }, 399 { 400 /* GL1 Data Cache per SA */ 401 .cache_size = 128, 402 .cache_level = 1, 403 .flags = (CRAT_CACHE_FLAGS_ENABLED | 404 CRAT_CACHE_FLAGS_DATA_CACHE | 405 CRAT_CACHE_FLAGS_SIMD_CACHE), 406 .num_cu_shared = 10, 407 }, 408 { 409 /* L2 Data Cache per GPU (Total Tex Cache) */ 410 .cache_size = 4096, 411 .cache_level = 2, 412 .flags = (CRAT_CACHE_FLAGS_ENABLED | 413 CRAT_CACHE_FLAGS_DATA_CACHE | 414 CRAT_CACHE_FLAGS_SIMD_CACHE), 415 .num_cu_shared = 10, 416 }, 417 }; 418 419 static struct kfd_gpu_cache_info vangogh_cache_info[] = { 420 { 421 /* TCP L1 Cache per CU */ 422 .cache_size = 16, 423 .cache_level = 1, 424 .flags = (CRAT_CACHE_FLAGS_ENABLED | 425 CRAT_CACHE_FLAGS_DATA_CACHE | 426 CRAT_CACHE_FLAGS_SIMD_CACHE), 427 .num_cu_shared = 1, 428 }, 429 { 430 /* Scalar L1 Instruction Cache per SQC */ 431 .cache_size = 32, 432 .cache_level = 1, 433 .flags = (CRAT_CACHE_FLAGS_ENABLED | 434 CRAT_CACHE_FLAGS_INST_CACHE | 435 CRAT_CACHE_FLAGS_SIMD_CACHE), 436 .num_cu_shared = 2, 437 }, 438 { 439 /* Scalar L1 Data Cache per SQC */ 440 .cache_size = 16, 441 .cache_level = 1, 442 .flags = (CRAT_CACHE_FLAGS_ENABLED | 443 CRAT_CACHE_FLAGS_DATA_CACHE | 444 CRAT_CACHE_FLAGS_SIMD_CACHE), 445 .num_cu_shared = 2, 446 }, 447 { 448 /* GL1 Data Cache per SA */ 449 .cache_size = 128, 450 .cache_level = 1, 451 .flags = (CRAT_CACHE_FLAGS_ENABLED | 452 CRAT_CACHE_FLAGS_DATA_CACHE | 453 CRAT_CACHE_FLAGS_SIMD_CACHE), 454 .num_cu_shared = 8, 455 }, 456 { 457 /* L2 Data Cache per GPU (Total Tex Cache) */ 458 .cache_size = 1024, 459 .cache_level = 2, 460 .flags = (CRAT_CACHE_FLAGS_ENABLED | 461 CRAT_CACHE_FLAGS_DATA_CACHE | 462 CRAT_CACHE_FLAGS_SIMD_CACHE), 463 .num_cu_shared = 8, 464 }, 465 }; 466 467 static struct kfd_gpu_cache_info navi14_cache_info[] = { 468 { 469 /* TCP L1 Cache per CU */ 470 .cache_size = 16, 471 .cache_level = 1, 472 .flags = (CRAT_CACHE_FLAGS_ENABLED | 473 CRAT_CACHE_FLAGS_DATA_CACHE | 474 CRAT_CACHE_FLAGS_SIMD_CACHE), 475 .num_cu_shared = 1, 476 }, 477 { 478 /* Scalar L1 Instruction Cache per SQC */ 479 .cache_size = 32, 480 .cache_level = 1, 481 .flags = (CRAT_CACHE_FLAGS_ENABLED | 482 CRAT_CACHE_FLAGS_INST_CACHE | 483 CRAT_CACHE_FLAGS_SIMD_CACHE), 484 .num_cu_shared = 2, 485 }, 486 { 487 /* Scalar L1 Data Cache per SQC */ 488 .cache_size = 16, 489 .cache_level = 1, 490 .flags = (CRAT_CACHE_FLAGS_ENABLED | 491 CRAT_CACHE_FLAGS_DATA_CACHE | 492 CRAT_CACHE_FLAGS_SIMD_CACHE), 493 .num_cu_shared = 2, 494 }, 495 { 496 /* GL1 Data Cache per SA */ 497 .cache_size = 128, 498 .cache_level = 1, 499 .flags = (CRAT_CACHE_FLAGS_ENABLED | 500 CRAT_CACHE_FLAGS_DATA_CACHE | 501 CRAT_CACHE_FLAGS_SIMD_CACHE), 502 .num_cu_shared = 12, 503 }, 504 { 505 /* L2 Data Cache per GPU (Total Tex Cache) */ 506 .cache_size = 2048, 507 .cache_level = 2, 508 .flags = (CRAT_CACHE_FLAGS_ENABLED | 509 CRAT_CACHE_FLAGS_DATA_CACHE | 510 CRAT_CACHE_FLAGS_SIMD_CACHE), 511 .num_cu_shared = 12, 512 }, 513 }; 514 515 static struct kfd_gpu_cache_info sienna_cichlid_cache_info[] = { 516 { 517 /* TCP L1 Cache per CU */ 518 .cache_size = 16, 519 .cache_level = 1, 520 .flags = (CRAT_CACHE_FLAGS_ENABLED | 521 CRAT_CACHE_FLAGS_DATA_CACHE | 522 CRAT_CACHE_FLAGS_SIMD_CACHE), 523 .num_cu_shared = 1, 524 }, 525 { 526 /* Scalar L1 Instruction Cache per SQC */ 527 .cache_size = 32, 528 .cache_level = 1, 529 .flags = (CRAT_CACHE_FLAGS_ENABLED | 530 CRAT_CACHE_FLAGS_INST_CACHE | 531 CRAT_CACHE_FLAGS_SIMD_CACHE), 532 .num_cu_shared = 2, 533 }, 534 { 535 /* Scalar L1 Data Cache per SQC */ 536 .cache_size = 16, 537 .cache_level = 1, 538 .flags = (CRAT_CACHE_FLAGS_ENABLED | 539 CRAT_CACHE_FLAGS_DATA_CACHE | 540 CRAT_CACHE_FLAGS_SIMD_CACHE), 541 .num_cu_shared = 2, 542 }, 543 { 544 /* GL1 Data Cache per SA */ 545 .cache_size = 128, 546 .cache_level = 1, 547 .flags = (CRAT_CACHE_FLAGS_ENABLED | 548 CRAT_CACHE_FLAGS_DATA_CACHE | 549 CRAT_CACHE_FLAGS_SIMD_CACHE), 550 .num_cu_shared = 10, 551 }, 552 { 553 /* L2 Data Cache per GPU (Total Tex Cache) */ 554 .cache_size = 4096, 555 .cache_level = 2, 556 .flags = (CRAT_CACHE_FLAGS_ENABLED | 557 CRAT_CACHE_FLAGS_DATA_CACHE | 558 CRAT_CACHE_FLAGS_SIMD_CACHE), 559 .num_cu_shared = 10, 560 }, 561 { 562 /* L3 Data Cache per GPU */ 563 .cache_size = 128*1024, 564 .cache_level = 3, 565 .flags = (CRAT_CACHE_FLAGS_ENABLED | 566 CRAT_CACHE_FLAGS_DATA_CACHE | 567 CRAT_CACHE_FLAGS_SIMD_CACHE), 568 .num_cu_shared = 10, 569 }, 570 }; 571 572 static struct kfd_gpu_cache_info navy_flounder_cache_info[] = { 573 { 574 /* TCP L1 Cache per CU */ 575 .cache_size = 16, 576 .cache_level = 1, 577 .flags = (CRAT_CACHE_FLAGS_ENABLED | 578 CRAT_CACHE_FLAGS_DATA_CACHE | 579 CRAT_CACHE_FLAGS_SIMD_CACHE), 580 .num_cu_shared = 1, 581 }, 582 { 583 /* Scalar L1 Instruction Cache per SQC */ 584 .cache_size = 32, 585 .cache_level = 1, 586 .flags = (CRAT_CACHE_FLAGS_ENABLED | 587 CRAT_CACHE_FLAGS_INST_CACHE | 588 CRAT_CACHE_FLAGS_SIMD_CACHE), 589 .num_cu_shared = 2, 590 }, 591 { 592 /* Scalar L1 Data Cache per SQC */ 593 .cache_size = 16, 594 .cache_level = 1, 595 .flags = (CRAT_CACHE_FLAGS_ENABLED | 596 CRAT_CACHE_FLAGS_DATA_CACHE | 597 CRAT_CACHE_FLAGS_SIMD_CACHE), 598 .num_cu_shared = 2, 599 }, 600 { 601 /* GL1 Data Cache per SA */ 602 .cache_size = 128, 603 .cache_level = 1, 604 .flags = (CRAT_CACHE_FLAGS_ENABLED | 605 CRAT_CACHE_FLAGS_DATA_CACHE | 606 CRAT_CACHE_FLAGS_SIMD_CACHE), 607 .num_cu_shared = 10, 608 }, 609 { 610 /* L2 Data Cache per GPU (Total Tex Cache) */ 611 .cache_size = 3072, 612 .cache_level = 2, 613 .flags = (CRAT_CACHE_FLAGS_ENABLED | 614 CRAT_CACHE_FLAGS_DATA_CACHE | 615 CRAT_CACHE_FLAGS_SIMD_CACHE), 616 .num_cu_shared = 10, 617 }, 618 { 619 /* L3 Data Cache per GPU */ 620 .cache_size = 96*1024, 621 .cache_level = 3, 622 .flags = (CRAT_CACHE_FLAGS_ENABLED | 623 CRAT_CACHE_FLAGS_DATA_CACHE | 624 CRAT_CACHE_FLAGS_SIMD_CACHE), 625 .num_cu_shared = 10, 626 }, 627 }; 628 629 static struct kfd_gpu_cache_info dimgrey_cavefish_cache_info[] = { 630 { 631 /* TCP L1 Cache per CU */ 632 .cache_size = 16, 633 .cache_level = 1, 634 .flags = (CRAT_CACHE_FLAGS_ENABLED | 635 CRAT_CACHE_FLAGS_DATA_CACHE | 636 CRAT_CACHE_FLAGS_SIMD_CACHE), 637 .num_cu_shared = 1, 638 }, 639 { 640 /* Scalar L1 Instruction Cache per SQC */ 641 .cache_size = 32, 642 .cache_level = 1, 643 .flags = (CRAT_CACHE_FLAGS_ENABLED | 644 CRAT_CACHE_FLAGS_INST_CACHE | 645 CRAT_CACHE_FLAGS_SIMD_CACHE), 646 .num_cu_shared = 2, 647 }, 648 { 649 /* Scalar L1 Data Cache per SQC */ 650 .cache_size = 16, 651 .cache_level = 1, 652 .flags = (CRAT_CACHE_FLAGS_ENABLED | 653 CRAT_CACHE_FLAGS_DATA_CACHE | 654 CRAT_CACHE_FLAGS_SIMD_CACHE), 655 .num_cu_shared = 2, 656 }, 657 { 658 /* GL1 Data Cache per SA */ 659 .cache_size = 128, 660 .cache_level = 1, 661 .flags = (CRAT_CACHE_FLAGS_ENABLED | 662 CRAT_CACHE_FLAGS_DATA_CACHE | 663 CRAT_CACHE_FLAGS_SIMD_CACHE), 664 .num_cu_shared = 8, 665 }, 666 { 667 /* L2 Data Cache per GPU (Total Tex Cache) */ 668 .cache_size = 2048, 669 .cache_level = 2, 670 .flags = (CRAT_CACHE_FLAGS_ENABLED | 671 CRAT_CACHE_FLAGS_DATA_CACHE | 672 CRAT_CACHE_FLAGS_SIMD_CACHE), 673 .num_cu_shared = 8, 674 }, 675 { 676 /* L3 Data Cache per GPU */ 677 .cache_size = 32*1024, 678 .cache_level = 3, 679 .flags = (CRAT_CACHE_FLAGS_ENABLED | 680 CRAT_CACHE_FLAGS_DATA_CACHE | 681 CRAT_CACHE_FLAGS_SIMD_CACHE), 682 .num_cu_shared = 8, 683 }, 684 }; 685 686 static struct kfd_gpu_cache_info beige_goby_cache_info[] = { 687 { 688 /* TCP L1 Cache per CU */ 689 .cache_size = 16, 690 .cache_level = 1, 691 .flags = (CRAT_CACHE_FLAGS_ENABLED | 692 CRAT_CACHE_FLAGS_DATA_CACHE | 693 CRAT_CACHE_FLAGS_SIMD_CACHE), 694 .num_cu_shared = 1, 695 }, 696 { 697 /* Scalar L1 Instruction Cache per SQC */ 698 .cache_size = 32, 699 .cache_level = 1, 700 .flags = (CRAT_CACHE_FLAGS_ENABLED | 701 CRAT_CACHE_FLAGS_INST_CACHE | 702 CRAT_CACHE_FLAGS_SIMD_CACHE), 703 .num_cu_shared = 2, 704 }, 705 { 706 /* Scalar L1 Data Cache per SQC */ 707 .cache_size = 16, 708 .cache_level = 1, 709 .flags = (CRAT_CACHE_FLAGS_ENABLED | 710 CRAT_CACHE_FLAGS_DATA_CACHE | 711 CRAT_CACHE_FLAGS_SIMD_CACHE), 712 .num_cu_shared = 2, 713 }, 714 { 715 /* GL1 Data Cache per SA */ 716 .cache_size = 128, 717 .cache_level = 1, 718 .flags = (CRAT_CACHE_FLAGS_ENABLED | 719 CRAT_CACHE_FLAGS_DATA_CACHE | 720 CRAT_CACHE_FLAGS_SIMD_CACHE), 721 .num_cu_shared = 8, 722 }, 723 { 724 /* L2 Data Cache per GPU (Total Tex Cache) */ 725 .cache_size = 1024, 726 .cache_level = 2, 727 .flags = (CRAT_CACHE_FLAGS_ENABLED | 728 CRAT_CACHE_FLAGS_DATA_CACHE | 729 CRAT_CACHE_FLAGS_SIMD_CACHE), 730 .num_cu_shared = 8, 731 }, 732 { 733 /* L3 Data Cache per GPU */ 734 .cache_size = 16*1024, 735 .cache_level = 3, 736 .flags = (CRAT_CACHE_FLAGS_ENABLED | 737 CRAT_CACHE_FLAGS_DATA_CACHE | 738 CRAT_CACHE_FLAGS_SIMD_CACHE), 739 .num_cu_shared = 8, 740 }, 741 }; 742 743 static struct kfd_gpu_cache_info yellow_carp_cache_info[] = { 744 { 745 /* TCP L1 Cache per CU */ 746 .cache_size = 16, 747 .cache_level = 1, 748 .flags = (CRAT_CACHE_FLAGS_ENABLED | 749 CRAT_CACHE_FLAGS_DATA_CACHE | 750 CRAT_CACHE_FLAGS_SIMD_CACHE), 751 .num_cu_shared = 1, 752 }, 753 { 754 /* Scalar L1 Instruction Cache per SQC */ 755 .cache_size = 32, 756 .cache_level = 1, 757 .flags = (CRAT_CACHE_FLAGS_ENABLED | 758 CRAT_CACHE_FLAGS_INST_CACHE | 759 CRAT_CACHE_FLAGS_SIMD_CACHE), 760 .num_cu_shared = 2, 761 }, 762 { 763 /* Scalar L1 Data Cache per SQC */ 764 .cache_size = 16, 765 .cache_level = 1, 766 .flags = (CRAT_CACHE_FLAGS_ENABLED | 767 CRAT_CACHE_FLAGS_DATA_CACHE | 768 CRAT_CACHE_FLAGS_SIMD_CACHE), 769 .num_cu_shared = 2, 770 }, 771 { 772 /* GL1 Data Cache per SA */ 773 .cache_size = 128, 774 .cache_level = 1, 775 .flags = (CRAT_CACHE_FLAGS_ENABLED | 776 CRAT_CACHE_FLAGS_DATA_CACHE | 777 CRAT_CACHE_FLAGS_SIMD_CACHE), 778 .num_cu_shared = 6, 779 }, 780 { 781 /* L2 Data Cache per GPU (Total Tex Cache) */ 782 .cache_size = 2048, 783 .cache_level = 2, 784 .flags = (CRAT_CACHE_FLAGS_ENABLED | 785 CRAT_CACHE_FLAGS_DATA_CACHE | 786 CRAT_CACHE_FLAGS_SIMD_CACHE), 787 .num_cu_shared = 6, 788 }, 789 }; 790 791 static struct kfd_gpu_cache_info gfx1037_cache_info[] = { 792 { 793 /* TCP L1 Cache per CU */ 794 .cache_size = 16, 795 .cache_level = 1, 796 .flags = (CRAT_CACHE_FLAGS_ENABLED | 797 CRAT_CACHE_FLAGS_DATA_CACHE | 798 CRAT_CACHE_FLAGS_SIMD_CACHE), 799 .num_cu_shared = 1, 800 }, 801 { 802 /* Scalar L1 Instruction Cache per SQC */ 803 .cache_size = 32, 804 .cache_level = 1, 805 .flags = (CRAT_CACHE_FLAGS_ENABLED | 806 CRAT_CACHE_FLAGS_INST_CACHE | 807 CRAT_CACHE_FLAGS_SIMD_CACHE), 808 .num_cu_shared = 2, 809 }, 810 { 811 /* Scalar L1 Data Cache per SQC */ 812 .cache_size = 16, 813 .cache_level = 1, 814 .flags = (CRAT_CACHE_FLAGS_ENABLED | 815 CRAT_CACHE_FLAGS_DATA_CACHE | 816 CRAT_CACHE_FLAGS_SIMD_CACHE), 817 .num_cu_shared = 2, 818 }, 819 { 820 /* GL1 Data Cache per SA */ 821 .cache_size = 128, 822 .cache_level = 1, 823 .flags = (CRAT_CACHE_FLAGS_ENABLED | 824 CRAT_CACHE_FLAGS_DATA_CACHE | 825 CRAT_CACHE_FLAGS_SIMD_CACHE), 826 .num_cu_shared = 2, 827 }, 828 { 829 /* L2 Data Cache per GPU (Total Tex Cache) */ 830 .cache_size = 256, 831 .cache_level = 2, 832 .flags = (CRAT_CACHE_FLAGS_ENABLED | 833 CRAT_CACHE_FLAGS_DATA_CACHE | 834 CRAT_CACHE_FLAGS_SIMD_CACHE), 835 .num_cu_shared = 2, 836 }, 837 }; 838 839 static struct kfd_gpu_cache_info gc_10_3_6_cache_info[] = { 840 { 841 /* TCP L1 Cache per CU */ 842 .cache_size = 16, 843 .cache_level = 1, 844 .flags = (CRAT_CACHE_FLAGS_ENABLED | 845 CRAT_CACHE_FLAGS_DATA_CACHE | 846 CRAT_CACHE_FLAGS_SIMD_CACHE), 847 .num_cu_shared = 1, 848 }, 849 { 850 /* Scalar L1 Instruction Cache per SQC */ 851 .cache_size = 32, 852 .cache_level = 1, 853 .flags = (CRAT_CACHE_FLAGS_ENABLED | 854 CRAT_CACHE_FLAGS_INST_CACHE | 855 CRAT_CACHE_FLAGS_SIMD_CACHE), 856 .num_cu_shared = 2, 857 }, 858 { 859 /* Scalar L1 Data Cache per SQC */ 860 .cache_size = 16, 861 .cache_level = 1, 862 .flags = (CRAT_CACHE_FLAGS_ENABLED | 863 CRAT_CACHE_FLAGS_DATA_CACHE | 864 CRAT_CACHE_FLAGS_SIMD_CACHE), 865 .num_cu_shared = 2, 866 }, 867 { 868 /* GL1 Data Cache per SA */ 869 .cache_size = 128, 870 .cache_level = 1, 871 .flags = (CRAT_CACHE_FLAGS_ENABLED | 872 CRAT_CACHE_FLAGS_DATA_CACHE | 873 CRAT_CACHE_FLAGS_SIMD_CACHE), 874 .num_cu_shared = 2, 875 }, 876 { 877 /* L2 Data Cache per GPU (Total Tex Cache) */ 878 .cache_size = 256, 879 .cache_level = 2, 880 .flags = (CRAT_CACHE_FLAGS_ENABLED | 881 CRAT_CACHE_FLAGS_DATA_CACHE | 882 CRAT_CACHE_FLAGS_SIMD_CACHE), 883 .num_cu_shared = 2, 884 }, 885 }; 886 887 static struct kfd_gpu_cache_info dummy_cache_info[] = { 888 { 889 /* TCP L1 Cache per CU */ 890 .cache_size = 16, 891 .cache_level = 1, 892 .flags = (CRAT_CACHE_FLAGS_ENABLED | 893 CRAT_CACHE_FLAGS_DATA_CACHE | 894 CRAT_CACHE_FLAGS_SIMD_CACHE), 895 .num_cu_shared = 1, 896 }, 897 { 898 /* Scalar L1 Instruction Cache per SQC */ 899 .cache_size = 32, 900 .cache_level = 1, 901 .flags = (CRAT_CACHE_FLAGS_ENABLED | 902 CRAT_CACHE_FLAGS_INST_CACHE | 903 CRAT_CACHE_FLAGS_SIMD_CACHE), 904 .num_cu_shared = 2, 905 }, 906 { 907 /* Scalar L1 Data Cache per SQC */ 908 .cache_size = 16, 909 .cache_level = 1, 910 .flags = (CRAT_CACHE_FLAGS_ENABLED | 911 CRAT_CACHE_FLAGS_DATA_CACHE | 912 CRAT_CACHE_FLAGS_SIMD_CACHE), 913 .num_cu_shared = 2, 914 }, 915 { 916 /* GL1 Data Cache per SA */ 917 .cache_size = 128, 918 .cache_level = 1, 919 .flags = (CRAT_CACHE_FLAGS_ENABLED | 920 CRAT_CACHE_FLAGS_DATA_CACHE | 921 CRAT_CACHE_FLAGS_SIMD_CACHE), 922 .num_cu_shared = 6, 923 }, 924 { 925 /* L2 Data Cache per GPU (Total Tex Cache) */ 926 .cache_size = 2048, 927 .cache_level = 2, 928 .flags = (CRAT_CACHE_FLAGS_ENABLED | 929 CRAT_CACHE_FLAGS_DATA_CACHE | 930 CRAT_CACHE_FLAGS_SIMD_CACHE), 931 .num_cu_shared = 6, 932 }, 933 }; 934 935 static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev, 936 struct crat_subtype_computeunit *cu) 937 { 938 dev->node_props.cpu_cores_count = cu->num_cpu_cores; 939 dev->node_props.cpu_core_id_base = cu->processor_id_low; 940 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT) 941 dev->node_props.capability |= HSA_CAP_ATS_PRESENT; 942 943 pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores, 944 cu->processor_id_low); 945 } 946 947 static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev, 948 struct crat_subtype_computeunit *cu) 949 { 950 dev->node_props.simd_id_base = cu->processor_id_low; 951 dev->node_props.simd_count = cu->num_simd_cores; 952 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb; 953 dev->node_props.max_waves_per_simd = cu->max_waves_simd; 954 dev->node_props.wave_front_size = cu->wave_front_size; 955 dev->node_props.array_count = cu->array_count; 956 dev->node_props.cu_per_simd_array = cu->num_cu_per_array; 957 dev->node_props.simd_per_cu = cu->num_simd_per_cu; 958 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu; 959 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE) 960 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE; 961 pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low); 962 } 963 964 /* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct 965 * topology device present in the device_list 966 */ 967 static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu, 968 struct list_head *device_list) 969 { 970 struct kfd_topology_device *dev; 971 972 pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n", 973 cu->proximity_domain, cu->hsa_capability); 974 list_for_each_entry(dev, device_list, list) { 975 if (cu->proximity_domain == dev->proximity_domain) { 976 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT) 977 kfd_populated_cu_info_cpu(dev, cu); 978 979 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT) 980 kfd_populated_cu_info_gpu(dev, cu); 981 break; 982 } 983 } 984 985 return 0; 986 } 987 988 static struct kfd_mem_properties * 989 find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width, 990 struct kfd_topology_device *dev) 991 { 992 struct kfd_mem_properties *props; 993 994 list_for_each_entry(props, &dev->mem_props, list) { 995 if (props->heap_type == heap_type 996 && props->flags == flags 997 && props->width == width) 998 return props; 999 } 1000 1001 return NULL; 1002 } 1003 /* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct 1004 * topology device present in the device_list 1005 */ 1006 static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem, 1007 struct list_head *device_list) 1008 { 1009 struct kfd_mem_properties *props; 1010 struct kfd_topology_device *dev; 1011 uint32_t heap_type; 1012 uint64_t size_in_bytes; 1013 uint32_t flags = 0; 1014 uint32_t width; 1015 1016 pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n", 1017 mem->proximity_domain); 1018 list_for_each_entry(dev, device_list, list) { 1019 if (mem->proximity_domain == dev->proximity_domain) { 1020 /* We're on GPU node */ 1021 if (dev->node_props.cpu_cores_count == 0) { 1022 /* APU */ 1023 if (mem->visibility_type == 0) 1024 heap_type = 1025 HSA_MEM_HEAP_TYPE_FB_PRIVATE; 1026 /* dGPU */ 1027 else 1028 heap_type = mem->visibility_type; 1029 } else 1030 heap_type = HSA_MEM_HEAP_TYPE_SYSTEM; 1031 1032 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE) 1033 flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE; 1034 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE) 1035 flags |= HSA_MEM_FLAGS_NON_VOLATILE; 1036 1037 size_in_bytes = 1038 ((uint64_t)mem->length_high << 32) + 1039 mem->length_low; 1040 width = mem->width; 1041 1042 /* Multiple banks of the same type are aggregated into 1043 * one. User mode doesn't care about multiple physical 1044 * memory segments. It's managed as a single virtual 1045 * heap for user mode. 1046 */ 1047 props = find_subtype_mem(heap_type, flags, width, dev); 1048 if (props) { 1049 props->size_in_bytes += size_in_bytes; 1050 break; 1051 } 1052 1053 props = kfd_alloc_struct(props); 1054 if (!props) 1055 return -ENOMEM; 1056 1057 props->heap_type = heap_type; 1058 props->flags = flags; 1059 if (size_in_bytes == 0) 1060 size_in_bytes = DUMMY_VRAM_SIZE; /* Fixme: TBD */ 1061 props->size_in_bytes = size_in_bytes; 1062 props->width = width; 1063 1064 dev->node_props.mem_banks_count++; 1065 list_add_tail(&props->list, &dev->mem_props); 1066 1067 break; 1068 } 1069 } 1070 1071 return 0; 1072 } 1073 1074 /* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct 1075 * topology device present in the device_list 1076 */ 1077 static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache, 1078 struct list_head *device_list) 1079 { 1080 struct kfd_cache_properties *props; 1081 struct kfd_topology_device *dev; 1082 uint32_t id; 1083 uint32_t total_num_of_cu; 1084 1085 id = cache->processor_id_low; 1086 1087 pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id); 1088 list_for_each_entry(dev, device_list, list) { 1089 total_num_of_cu = (dev->node_props.array_count * 1090 dev->node_props.cu_per_simd_array); 1091 1092 /* Cache infomration in CRAT doesn't have proximity_domain 1093 * information as it is associated with a CPU core or GPU 1094 * Compute Unit. So map the cache using CPU core Id or SIMD 1095 * (GPU) ID. 1096 * TODO: This works because currently we can safely assume that 1097 * Compute Units are parsed before caches are parsed. In 1098 * future, remove this dependency 1099 */ 1100 if ((id >= dev->node_props.cpu_core_id_base && 1101 id <= dev->node_props.cpu_core_id_base + 1102 dev->node_props.cpu_cores_count) || 1103 (id >= dev->node_props.simd_id_base && 1104 id < dev->node_props.simd_id_base + 1105 total_num_of_cu)) { 1106 props = kfd_alloc_struct(props); 1107 if (!props) 1108 return -ENOMEM; 1109 1110 props->processor_id_low = id; 1111 props->cache_level = cache->cache_level; 1112 props->cache_size = cache->cache_size; 1113 props->cacheline_size = cache->cache_line_size; 1114 props->cachelines_per_tag = cache->lines_per_tag; 1115 props->cache_assoc = cache->associativity; 1116 props->cache_latency = cache->cache_latency; 1117 1118 memcpy(props->sibling_map, cache->sibling_map, 1119 CRAT_SIBLINGMAP_SIZE); 1120 1121 /* set the sibling_map_size as 32 for CRAT from ACPI */ 1122 props->sibling_map_size = CRAT_SIBLINGMAP_SIZE; 1123 1124 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE) 1125 props->cache_type |= HSA_CACHE_TYPE_DATA; 1126 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE) 1127 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION; 1128 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE) 1129 props->cache_type |= HSA_CACHE_TYPE_CPU; 1130 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE) 1131 props->cache_type |= HSA_CACHE_TYPE_HSACU; 1132 1133 dev->node_props.caches_count++; 1134 list_add_tail(&props->list, &dev->cache_props); 1135 1136 break; 1137 } 1138 } 1139 1140 return 0; 1141 } 1142 1143 /* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct 1144 * topology device present in the device_list 1145 */ 1146 static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink, 1147 struct list_head *device_list) 1148 { 1149 struct kfd_iolink_properties *props = NULL, *props2; 1150 struct kfd_topology_device *dev, *to_dev; 1151 uint32_t id_from; 1152 uint32_t id_to; 1153 1154 id_from = iolink->proximity_domain_from; 1155 id_to = iolink->proximity_domain_to; 1156 1157 pr_debug("Found IO link entry in CRAT table with id_from=%d, id_to %d\n", 1158 id_from, id_to); 1159 list_for_each_entry(dev, device_list, list) { 1160 if (id_from == dev->proximity_domain) { 1161 props = kfd_alloc_struct(props); 1162 if (!props) 1163 return -ENOMEM; 1164 1165 props->node_from = id_from; 1166 props->node_to = id_to; 1167 props->ver_maj = iolink->version_major; 1168 props->ver_min = iolink->version_minor; 1169 props->iolink_type = iolink->io_interface_type; 1170 1171 if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) 1172 props->weight = 20; 1173 else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI) 1174 props->weight = iolink->weight_xgmi; 1175 else 1176 props->weight = node_distance(id_from, id_to); 1177 1178 props->min_latency = iolink->minimum_latency; 1179 props->max_latency = iolink->maximum_latency; 1180 props->min_bandwidth = iolink->minimum_bandwidth_mbs; 1181 props->max_bandwidth = iolink->maximum_bandwidth_mbs; 1182 props->rec_transfer_size = 1183 iolink->recommended_transfer_size; 1184 1185 dev->node_props.io_links_count++; 1186 list_add_tail(&props->list, &dev->io_link_props); 1187 break; 1188 } 1189 } 1190 1191 /* CPU topology is created before GPUs are detected, so CPU->GPU 1192 * links are not built at that time. If a PCIe type is discovered, it 1193 * means a GPU is detected and we are adding GPU->CPU to the topology. 1194 * At this time, also add the corresponded CPU->GPU link if GPU 1195 * is large bar. 1196 * For xGMI, we only added the link with one direction in the crat 1197 * table, add corresponded reversed direction link now. 1198 */ 1199 if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) { 1200 to_dev = kfd_topology_device_by_proximity_domain_no_lock(id_to); 1201 if (!to_dev) 1202 return -ENODEV; 1203 /* same everything but the other direction */ 1204 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL); 1205 if (!props2) 1206 return -ENOMEM; 1207 1208 props2->node_from = id_to; 1209 props2->node_to = id_from; 1210 props2->kobj = NULL; 1211 to_dev->node_props.io_links_count++; 1212 list_add_tail(&props2->list, &to_dev->io_link_props); 1213 } 1214 1215 return 0; 1216 } 1217 1218 /* kfd_parse_subtype - parse subtypes and attach it to correct topology device 1219 * present in the device_list 1220 * @sub_type_hdr - subtype section of crat_image 1221 * @device_list - list of topology devices present in this crat_image 1222 */ 1223 static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr, 1224 struct list_head *device_list) 1225 { 1226 struct crat_subtype_computeunit *cu; 1227 struct crat_subtype_memory *mem; 1228 struct crat_subtype_cache *cache; 1229 struct crat_subtype_iolink *iolink; 1230 int ret = 0; 1231 1232 switch (sub_type_hdr->type) { 1233 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY: 1234 cu = (struct crat_subtype_computeunit *)sub_type_hdr; 1235 ret = kfd_parse_subtype_cu(cu, device_list); 1236 break; 1237 case CRAT_SUBTYPE_MEMORY_AFFINITY: 1238 mem = (struct crat_subtype_memory *)sub_type_hdr; 1239 ret = kfd_parse_subtype_mem(mem, device_list); 1240 break; 1241 case CRAT_SUBTYPE_CACHE_AFFINITY: 1242 cache = (struct crat_subtype_cache *)sub_type_hdr; 1243 ret = kfd_parse_subtype_cache(cache, device_list); 1244 break; 1245 case CRAT_SUBTYPE_TLB_AFFINITY: 1246 /* 1247 * For now, nothing to do here 1248 */ 1249 pr_debug("Found TLB entry in CRAT table (not processing)\n"); 1250 break; 1251 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY: 1252 /* 1253 * For now, nothing to do here 1254 */ 1255 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n"); 1256 break; 1257 case CRAT_SUBTYPE_IOLINK_AFFINITY: 1258 iolink = (struct crat_subtype_iolink *)sub_type_hdr; 1259 ret = kfd_parse_subtype_iolink(iolink, device_list); 1260 break; 1261 default: 1262 pr_warn("Unknown subtype %d in CRAT\n", 1263 sub_type_hdr->type); 1264 } 1265 1266 return ret; 1267 } 1268 1269 /* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT 1270 * create a kfd_topology_device and add in to device_list. Also parse 1271 * CRAT subtypes and attach it to appropriate kfd_topology_device 1272 * @crat_image - input image containing CRAT 1273 * @device_list - [OUT] list of kfd_topology_device generated after 1274 * parsing crat_image 1275 * @proximity_domain - Proximity domain of the first device in the table 1276 * 1277 * Return - 0 if successful else -ve value 1278 */ 1279 int kfd_parse_crat_table(void *crat_image, struct list_head *device_list, 1280 uint32_t proximity_domain) 1281 { 1282 struct kfd_topology_device *top_dev = NULL; 1283 struct crat_subtype_generic *sub_type_hdr; 1284 uint16_t node_id; 1285 int ret = 0; 1286 struct crat_header *crat_table = (struct crat_header *)crat_image; 1287 uint16_t num_nodes; 1288 uint32_t image_len; 1289 1290 if (!crat_image) 1291 return -EINVAL; 1292 1293 if (!list_empty(device_list)) { 1294 pr_warn("Error device list should be empty\n"); 1295 return -EINVAL; 1296 } 1297 1298 num_nodes = crat_table->num_domains; 1299 image_len = crat_table->length; 1300 1301 pr_debug("Parsing CRAT table with %d nodes\n", num_nodes); 1302 1303 for (node_id = 0; node_id < num_nodes; node_id++) { 1304 top_dev = kfd_create_topology_device(device_list); 1305 if (!top_dev) 1306 break; 1307 top_dev->proximity_domain = proximity_domain++; 1308 } 1309 1310 if (!top_dev) { 1311 ret = -ENOMEM; 1312 goto err; 1313 } 1314 1315 memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH); 1316 memcpy(top_dev->oem_table_id, crat_table->oem_table_id, 1317 CRAT_OEMTABLEID_LENGTH); 1318 top_dev->oem_revision = crat_table->oem_revision; 1319 1320 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1); 1321 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) < 1322 ((char *)crat_image) + image_len) { 1323 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) { 1324 ret = kfd_parse_subtype(sub_type_hdr, device_list); 1325 if (ret) 1326 break; 1327 } 1328 1329 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1330 sub_type_hdr->length); 1331 } 1332 1333 err: 1334 if (ret) 1335 kfd_release_topology_device_list(device_list); 1336 1337 return ret; 1338 } 1339 1340 1341 static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev, 1342 struct kfd_gpu_cache_info *pcache_info) 1343 { 1344 struct amdgpu_device *adev = kdev->adev; 1345 int i = 0; 1346 1347 /* TCP L1 Cache per CU */ 1348 if (adev->gfx.config.gc_tcp_l1_size) { 1349 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_l1_size; 1350 pcache_info[i].cache_level = 1; 1351 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1352 CRAT_CACHE_FLAGS_DATA_CACHE | 1353 CRAT_CACHE_FLAGS_SIMD_CACHE); 1354 pcache_info[0].num_cu_shared = adev->gfx.config.gc_num_tcp_per_wpg / 2; 1355 i++; 1356 } 1357 /* Scalar L1 Instruction Cache per SQC */ 1358 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) { 1359 pcache_info[i].cache_size = 1360 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc; 1361 pcache_info[i].cache_level = 1; 1362 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1363 CRAT_CACHE_FLAGS_INST_CACHE | 1364 CRAT_CACHE_FLAGS_SIMD_CACHE); 1365 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2; 1366 i++; 1367 } 1368 /* Scalar L1 Data Cache per SQC */ 1369 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) { 1370 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc; 1371 pcache_info[i].cache_level = 1; 1372 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1373 CRAT_CACHE_FLAGS_DATA_CACHE | 1374 CRAT_CACHE_FLAGS_SIMD_CACHE); 1375 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2; 1376 i++; 1377 } 1378 /* GL1 Data Cache per SA */ 1379 if (adev->gfx.config.gc_gl1c_per_sa && 1380 adev->gfx.config.gc_gl1c_size_per_instance) { 1381 pcache_info[i].cache_size = adev->gfx.config.gc_gl1c_per_sa * 1382 adev->gfx.config.gc_gl1c_size_per_instance; 1383 pcache_info[i].cache_level = 1; 1384 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1385 CRAT_CACHE_FLAGS_DATA_CACHE | 1386 CRAT_CACHE_FLAGS_SIMD_CACHE); 1387 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh; 1388 i++; 1389 } 1390 /* L2 Data Cache per GPU (Total Tex Cache) */ 1391 if (adev->gfx.config.gc_gl2c_per_gpu) { 1392 pcache_info[i].cache_size = adev->gfx.config.gc_gl2c_per_gpu; 1393 pcache_info[i].cache_level = 2; 1394 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1395 CRAT_CACHE_FLAGS_DATA_CACHE | 1396 CRAT_CACHE_FLAGS_SIMD_CACHE); 1397 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh; 1398 i++; 1399 } 1400 /* L3 Data Cache per GPU */ 1401 if (adev->gmc.mall_size) { 1402 pcache_info[i].cache_size = adev->gmc.mall_size / 1024; 1403 pcache_info[i].cache_level = 3; 1404 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1405 CRAT_CACHE_FLAGS_DATA_CACHE | 1406 CRAT_CACHE_FLAGS_SIMD_CACHE); 1407 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh; 1408 i++; 1409 } 1410 return i; 1411 } 1412 1413 int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pcache_info) 1414 { 1415 int num_of_cache_types = 0; 1416 1417 switch (kdev->adev->asic_type) { 1418 case CHIP_KAVERI: 1419 *pcache_info = kaveri_cache_info; 1420 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info); 1421 break; 1422 case CHIP_HAWAII: 1423 *pcache_info = hawaii_cache_info; 1424 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info); 1425 break; 1426 case CHIP_CARRIZO: 1427 *pcache_info = carrizo_cache_info; 1428 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info); 1429 break; 1430 case CHIP_TONGA: 1431 *pcache_info = tonga_cache_info; 1432 num_of_cache_types = ARRAY_SIZE(tonga_cache_info); 1433 break; 1434 case CHIP_FIJI: 1435 *pcache_info = fiji_cache_info; 1436 num_of_cache_types = ARRAY_SIZE(fiji_cache_info); 1437 break; 1438 case CHIP_POLARIS10: 1439 *pcache_info = polaris10_cache_info; 1440 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info); 1441 break; 1442 case CHIP_POLARIS11: 1443 *pcache_info = polaris11_cache_info; 1444 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info); 1445 break; 1446 case CHIP_POLARIS12: 1447 *pcache_info = polaris12_cache_info; 1448 num_of_cache_types = ARRAY_SIZE(polaris12_cache_info); 1449 break; 1450 case CHIP_VEGAM: 1451 *pcache_info = vegam_cache_info; 1452 num_of_cache_types = ARRAY_SIZE(vegam_cache_info); 1453 break; 1454 default: 1455 switch (KFD_GC_VERSION(kdev)) { 1456 case IP_VERSION(9, 0, 1): 1457 *pcache_info = vega10_cache_info; 1458 num_of_cache_types = ARRAY_SIZE(vega10_cache_info); 1459 break; 1460 case IP_VERSION(9, 2, 1): 1461 *pcache_info = vega12_cache_info; 1462 num_of_cache_types = ARRAY_SIZE(vega12_cache_info); 1463 break; 1464 case IP_VERSION(9, 4, 0): 1465 case IP_VERSION(9, 4, 1): 1466 *pcache_info = vega20_cache_info; 1467 num_of_cache_types = ARRAY_SIZE(vega20_cache_info); 1468 break; 1469 case IP_VERSION(9, 4, 2): 1470 case IP_VERSION(9, 4, 3): 1471 *pcache_info = aldebaran_cache_info; 1472 num_of_cache_types = ARRAY_SIZE(aldebaran_cache_info); 1473 break; 1474 case IP_VERSION(9, 1, 0): 1475 case IP_VERSION(9, 2, 2): 1476 *pcache_info = raven_cache_info; 1477 num_of_cache_types = ARRAY_SIZE(raven_cache_info); 1478 break; 1479 case IP_VERSION(9, 3, 0): 1480 *pcache_info = renoir_cache_info; 1481 num_of_cache_types = ARRAY_SIZE(renoir_cache_info); 1482 break; 1483 case IP_VERSION(10, 1, 10): 1484 case IP_VERSION(10, 1, 2): 1485 case IP_VERSION(10, 1, 3): 1486 case IP_VERSION(10, 1, 4): 1487 *pcache_info = navi10_cache_info; 1488 num_of_cache_types = ARRAY_SIZE(navi10_cache_info); 1489 break; 1490 case IP_VERSION(10, 1, 1): 1491 *pcache_info = navi14_cache_info; 1492 num_of_cache_types = ARRAY_SIZE(navi14_cache_info); 1493 break; 1494 case IP_VERSION(10, 3, 0): 1495 *pcache_info = sienna_cichlid_cache_info; 1496 num_of_cache_types = ARRAY_SIZE(sienna_cichlid_cache_info); 1497 break; 1498 case IP_VERSION(10, 3, 2): 1499 *pcache_info = navy_flounder_cache_info; 1500 num_of_cache_types = ARRAY_SIZE(navy_flounder_cache_info); 1501 break; 1502 case IP_VERSION(10, 3, 4): 1503 *pcache_info = dimgrey_cavefish_cache_info; 1504 num_of_cache_types = ARRAY_SIZE(dimgrey_cavefish_cache_info); 1505 break; 1506 case IP_VERSION(10, 3, 1): 1507 *pcache_info = vangogh_cache_info; 1508 num_of_cache_types = ARRAY_SIZE(vangogh_cache_info); 1509 break; 1510 case IP_VERSION(10, 3, 5): 1511 *pcache_info = beige_goby_cache_info; 1512 num_of_cache_types = ARRAY_SIZE(beige_goby_cache_info); 1513 break; 1514 case IP_VERSION(10, 3, 3): 1515 *pcache_info = yellow_carp_cache_info; 1516 num_of_cache_types = ARRAY_SIZE(yellow_carp_cache_info); 1517 break; 1518 case IP_VERSION(10, 3, 6): 1519 *pcache_info = gc_10_3_6_cache_info; 1520 num_of_cache_types = ARRAY_SIZE(gc_10_3_6_cache_info); 1521 break; 1522 case IP_VERSION(10, 3, 7): 1523 *pcache_info = gfx1037_cache_info; 1524 num_of_cache_types = ARRAY_SIZE(gfx1037_cache_info); 1525 break; 1526 case IP_VERSION(11, 0, 0): 1527 case IP_VERSION(11, 0, 1): 1528 case IP_VERSION(11, 0, 2): 1529 case IP_VERSION(11, 0, 3): 1530 case IP_VERSION(11, 0, 4): 1531 num_of_cache_types = 1532 kfd_fill_gpu_cache_info_from_gfx_config(kdev->kfd, *pcache_info); 1533 break; 1534 default: 1535 *pcache_info = dummy_cache_info; 1536 num_of_cache_types = ARRAY_SIZE(dummy_cache_info); 1537 pr_warn("dummy cache info is used temporarily and real cache info need update later.\n"); 1538 break; 1539 } 1540 } 1541 return num_of_cache_types; 1542 } 1543 1544 static bool kfd_ignore_crat(void) 1545 { 1546 bool ret; 1547 1548 if (ignore_crat) 1549 return true; 1550 1551 #ifndef KFD_SUPPORT_IOMMU_V2 1552 ret = true; 1553 #else 1554 ret = false; 1555 #endif 1556 1557 return ret; 1558 } 1559 1560 /* 1561 * kfd_create_crat_image_acpi - Allocates memory for CRAT image and 1562 * copies CRAT from ACPI (if available). 1563 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory 1564 * 1565 * @crat_image: CRAT read from ACPI. If no CRAT in ACPI then 1566 * crat_image will be NULL 1567 * @size: [OUT] size of crat_image 1568 * 1569 * Return 0 if successful else return error code 1570 */ 1571 int kfd_create_crat_image_acpi(void **crat_image, size_t *size) 1572 { 1573 struct acpi_table_header *crat_table; 1574 acpi_status status; 1575 void *pcrat_image; 1576 int rc = 0; 1577 1578 if (!crat_image) 1579 return -EINVAL; 1580 1581 *crat_image = NULL; 1582 1583 if (kfd_ignore_crat()) { 1584 pr_info("CRAT table disabled by module option\n"); 1585 return -ENODATA; 1586 } 1587 1588 /* Fetch the CRAT table from ACPI */ 1589 status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table); 1590 if (status == AE_NOT_FOUND) { 1591 pr_info("CRAT table not found\n"); 1592 return -ENODATA; 1593 } else if (ACPI_FAILURE(status)) { 1594 const char *err = acpi_format_exception(status); 1595 1596 pr_err("CRAT table error: %s\n", err); 1597 return -EINVAL; 1598 } 1599 1600 pcrat_image = kvmalloc(crat_table->length, GFP_KERNEL); 1601 if (!pcrat_image) { 1602 rc = -ENOMEM; 1603 goto out; 1604 } 1605 1606 memcpy(pcrat_image, crat_table, crat_table->length); 1607 *crat_image = pcrat_image; 1608 *size = crat_table->length; 1609 out: 1610 acpi_put_table(crat_table); 1611 return rc; 1612 } 1613 1614 /* Memory required to create Virtual CRAT. 1615 * Since there is no easy way to predict the amount of memory required, the 1616 * following amount is allocated for GPU Virtual CRAT. This is 1617 * expected to cover all known conditions. But to be safe additional check 1618 * is put in the code to ensure we don't overwrite. 1619 */ 1620 #define VCRAT_SIZE_FOR_GPU (4 * PAGE_SIZE) 1621 1622 /* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node 1623 * 1624 * @numa_node_id: CPU NUMA node id 1625 * @avail_size: Available size in the memory 1626 * @sub_type_hdr: Memory into which compute info will be filled in 1627 * 1628 * Return 0 if successful else return -ve value 1629 */ 1630 static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size, 1631 int proximity_domain, 1632 struct crat_subtype_computeunit *sub_type_hdr) 1633 { 1634 const struct cpumask *cpumask; 1635 1636 *avail_size -= sizeof(struct crat_subtype_computeunit); 1637 if (*avail_size < 0) 1638 return -ENOMEM; 1639 1640 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit)); 1641 1642 /* Fill in subtype header data */ 1643 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY; 1644 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit); 1645 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 1646 1647 cpumask = cpumask_of_node(numa_node_id); 1648 1649 /* Fill in CU data */ 1650 sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT; 1651 sub_type_hdr->proximity_domain = proximity_domain; 1652 sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id); 1653 if (sub_type_hdr->processor_id_low == -1) 1654 return -EINVAL; 1655 1656 sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask); 1657 1658 return 0; 1659 } 1660 1661 /* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node 1662 * 1663 * @numa_node_id: CPU NUMA node id 1664 * @avail_size: Available size in the memory 1665 * @sub_type_hdr: Memory into which compute info will be filled in 1666 * 1667 * Return 0 if successful else return -ve value 1668 */ 1669 static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size, 1670 int proximity_domain, 1671 struct crat_subtype_memory *sub_type_hdr) 1672 { 1673 uint64_t mem_in_bytes = 0; 1674 pg_data_t *pgdat; 1675 int zone_type; 1676 1677 *avail_size -= sizeof(struct crat_subtype_memory); 1678 if (*avail_size < 0) 1679 return -ENOMEM; 1680 1681 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory)); 1682 1683 /* Fill in subtype header data */ 1684 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY; 1685 sub_type_hdr->length = sizeof(struct crat_subtype_memory); 1686 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 1687 1688 /* Fill in Memory Subunit data */ 1689 1690 /* Unlike si_meminfo, si_meminfo_node is not exported. So 1691 * the following lines are duplicated from si_meminfo_node 1692 * function 1693 */ 1694 pgdat = NODE_DATA(numa_node_id); 1695 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) 1696 mem_in_bytes += zone_managed_pages(&pgdat->node_zones[zone_type]); 1697 mem_in_bytes <<= PAGE_SHIFT; 1698 1699 sub_type_hdr->length_low = lower_32_bits(mem_in_bytes); 1700 sub_type_hdr->length_high = upper_32_bits(mem_in_bytes); 1701 sub_type_hdr->proximity_domain = proximity_domain; 1702 1703 return 0; 1704 } 1705 1706 #ifdef CONFIG_X86_64 1707 static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size, 1708 uint32_t *num_entries, 1709 struct crat_subtype_iolink *sub_type_hdr) 1710 { 1711 int nid; 1712 struct cpuinfo_x86 *c = &cpu_data(0); 1713 uint8_t link_type; 1714 1715 if (c->x86_vendor == X86_VENDOR_AMD) 1716 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT; 1717 else 1718 link_type = CRAT_IOLINK_TYPE_QPI_1_1; 1719 1720 *num_entries = 0; 1721 1722 /* Create IO links from this node to other CPU nodes */ 1723 for_each_online_node(nid) { 1724 if (nid == numa_node_id) /* node itself */ 1725 continue; 1726 1727 *avail_size -= sizeof(struct crat_subtype_iolink); 1728 if (*avail_size < 0) 1729 return -ENOMEM; 1730 1731 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink)); 1732 1733 /* Fill in subtype header data */ 1734 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY; 1735 sub_type_hdr->length = sizeof(struct crat_subtype_iolink); 1736 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 1737 1738 /* Fill in IO link data */ 1739 sub_type_hdr->proximity_domain_from = numa_node_id; 1740 sub_type_hdr->proximity_domain_to = nid; 1741 sub_type_hdr->io_interface_type = link_type; 1742 1743 (*num_entries)++; 1744 sub_type_hdr++; 1745 } 1746 1747 return 0; 1748 } 1749 #endif 1750 1751 /* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU 1752 * 1753 * @pcrat_image: Fill in VCRAT for CPU 1754 * @size: [IN] allocated size of crat_image. 1755 * [OUT] actual size of data filled in crat_image 1756 */ 1757 static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size) 1758 { 1759 struct crat_header *crat_table = (struct crat_header *)pcrat_image; 1760 struct acpi_table_header *acpi_table; 1761 acpi_status status; 1762 struct crat_subtype_generic *sub_type_hdr; 1763 int avail_size = *size; 1764 int numa_node_id; 1765 #ifdef CONFIG_X86_64 1766 uint32_t entries = 0; 1767 #endif 1768 int ret = 0; 1769 1770 if (!pcrat_image) 1771 return -EINVAL; 1772 1773 /* Fill in CRAT Header. 1774 * Modify length and total_entries as subunits are added. 1775 */ 1776 avail_size -= sizeof(struct crat_header); 1777 if (avail_size < 0) 1778 return -ENOMEM; 1779 1780 memset(crat_table, 0, sizeof(struct crat_header)); 1781 memcpy(&crat_table->signature, CRAT_SIGNATURE, 1782 sizeof(crat_table->signature)); 1783 crat_table->length = sizeof(struct crat_header); 1784 1785 status = acpi_get_table("DSDT", 0, &acpi_table); 1786 if (status != AE_OK) 1787 pr_warn("DSDT table not found for OEM information\n"); 1788 else { 1789 crat_table->oem_revision = acpi_table->revision; 1790 memcpy(crat_table->oem_id, acpi_table->oem_id, 1791 CRAT_OEMID_LENGTH); 1792 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id, 1793 CRAT_OEMTABLEID_LENGTH); 1794 acpi_put_table(acpi_table); 1795 } 1796 crat_table->total_entries = 0; 1797 crat_table->num_domains = 0; 1798 1799 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1); 1800 1801 for_each_online_node(numa_node_id) { 1802 if (kfd_numa_node_to_apic_id(numa_node_id) == -1) 1803 continue; 1804 1805 /* Fill in Subtype: Compute Unit */ 1806 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size, 1807 crat_table->num_domains, 1808 (struct crat_subtype_computeunit *)sub_type_hdr); 1809 if (ret < 0) 1810 return ret; 1811 crat_table->length += sub_type_hdr->length; 1812 crat_table->total_entries++; 1813 1814 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1815 sub_type_hdr->length); 1816 1817 /* Fill in Subtype: Memory */ 1818 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size, 1819 crat_table->num_domains, 1820 (struct crat_subtype_memory *)sub_type_hdr); 1821 if (ret < 0) 1822 return ret; 1823 crat_table->length += sub_type_hdr->length; 1824 crat_table->total_entries++; 1825 1826 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1827 sub_type_hdr->length); 1828 1829 /* Fill in Subtype: IO Link */ 1830 #ifdef CONFIG_X86_64 1831 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size, 1832 &entries, 1833 (struct crat_subtype_iolink *)sub_type_hdr); 1834 if (ret < 0) 1835 return ret; 1836 1837 if (entries) { 1838 crat_table->length += (sub_type_hdr->length * entries); 1839 crat_table->total_entries += entries; 1840 1841 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1842 sub_type_hdr->length * entries); 1843 } 1844 #else 1845 pr_info("IO link not available for non x86 platforms\n"); 1846 #endif 1847 1848 crat_table->num_domains++; 1849 } 1850 1851 /* TODO: Add cache Subtype for CPU. 1852 * Currently, CPU cache information is available in function 1853 * detect_cache_attributes(cpu) defined in the file 1854 * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not 1855 * exported and to get the same information the code needs to be 1856 * duplicated. 1857 */ 1858 1859 *size = crat_table->length; 1860 pr_info("Virtual CRAT table created for CPU\n"); 1861 1862 return 0; 1863 } 1864 1865 static int kfd_fill_gpu_memory_affinity(int *avail_size, 1866 struct kfd_node *kdev, uint8_t type, uint64_t size, 1867 struct crat_subtype_memory *sub_type_hdr, 1868 uint32_t proximity_domain, 1869 const struct kfd_local_mem_info *local_mem_info) 1870 { 1871 *avail_size -= sizeof(struct crat_subtype_memory); 1872 if (*avail_size < 0) 1873 return -ENOMEM; 1874 1875 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory)); 1876 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY; 1877 sub_type_hdr->length = sizeof(struct crat_subtype_memory); 1878 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED; 1879 1880 sub_type_hdr->proximity_domain = proximity_domain; 1881 1882 pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n", 1883 type, size); 1884 1885 sub_type_hdr->length_low = lower_32_bits(size); 1886 sub_type_hdr->length_high = upper_32_bits(size); 1887 1888 sub_type_hdr->width = local_mem_info->vram_width; 1889 sub_type_hdr->visibility_type = type; 1890 1891 return 0; 1892 } 1893 1894 #ifdef CONFIG_ACPI_NUMA 1895 static void kfd_find_numa_node_in_srat(struct kfd_node *kdev) 1896 { 1897 struct acpi_table_header *table_header = NULL; 1898 struct acpi_subtable_header *sub_header = NULL; 1899 unsigned long table_end, subtable_len; 1900 u32 pci_id = pci_domain_nr(kdev->adev->pdev->bus) << 16 | 1901 pci_dev_id(kdev->adev->pdev); 1902 u32 bdf; 1903 acpi_status status; 1904 struct acpi_srat_cpu_affinity *cpu; 1905 struct acpi_srat_generic_affinity *gpu; 1906 int pxm = 0, max_pxm = 0; 1907 int numa_node = NUMA_NO_NODE; 1908 bool found = false; 1909 1910 /* Fetch the SRAT table from ACPI */ 1911 status = acpi_get_table(ACPI_SIG_SRAT, 0, &table_header); 1912 if (status == AE_NOT_FOUND) { 1913 pr_warn("SRAT table not found\n"); 1914 return; 1915 } else if (ACPI_FAILURE(status)) { 1916 const char *err = acpi_format_exception(status); 1917 pr_err("SRAT table error: %s\n", err); 1918 return; 1919 } 1920 1921 table_end = (unsigned long)table_header + table_header->length; 1922 1923 /* Parse all entries looking for a match. */ 1924 sub_header = (struct acpi_subtable_header *) 1925 ((unsigned long)table_header + 1926 sizeof(struct acpi_table_srat)); 1927 subtable_len = sub_header->length; 1928 1929 while (((unsigned long)sub_header) + subtable_len < table_end) { 1930 /* 1931 * If length is 0, break from this loop to avoid 1932 * infinite loop. 1933 */ 1934 if (subtable_len == 0) { 1935 pr_err("SRAT invalid zero length\n"); 1936 break; 1937 } 1938 1939 switch (sub_header->type) { 1940 case ACPI_SRAT_TYPE_CPU_AFFINITY: 1941 cpu = (struct acpi_srat_cpu_affinity *)sub_header; 1942 pxm = *((u32 *)cpu->proximity_domain_hi) << 8 | 1943 cpu->proximity_domain_lo; 1944 if (pxm > max_pxm) 1945 max_pxm = pxm; 1946 break; 1947 case ACPI_SRAT_TYPE_GENERIC_AFFINITY: 1948 gpu = (struct acpi_srat_generic_affinity *)sub_header; 1949 bdf = *((u16 *)(&gpu->device_handle[0])) << 16 | 1950 *((u16 *)(&gpu->device_handle[2])); 1951 if (bdf == pci_id) { 1952 found = true; 1953 numa_node = pxm_to_node(gpu->proximity_domain); 1954 } 1955 break; 1956 default: 1957 break; 1958 } 1959 1960 if (found) 1961 break; 1962 1963 sub_header = (struct acpi_subtable_header *) 1964 ((unsigned long)sub_header + subtable_len); 1965 subtable_len = sub_header->length; 1966 } 1967 1968 acpi_put_table(table_header); 1969 1970 /* Workaround bad cpu-gpu binding case */ 1971 if (found && (numa_node < 0 || 1972 numa_node > pxm_to_node(max_pxm))) 1973 numa_node = 0; 1974 1975 if (numa_node != NUMA_NO_NODE) 1976 set_dev_node(&kdev->adev->pdev->dev, numa_node); 1977 } 1978 #endif 1979 1980 #define KFD_CRAT_INTRA_SOCKET_WEIGHT 13 1981 #define KFD_CRAT_XGMI_WEIGHT 15 1982 1983 /* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU 1984 * to its NUMA node 1985 * @avail_size: Available size in the memory 1986 * @kdev - [IN] GPU device 1987 * @sub_type_hdr: Memory into which io link info will be filled in 1988 * @proximity_domain - proximity domain of the GPU node 1989 * 1990 * Return 0 if successful else return -ve value 1991 */ 1992 static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size, 1993 struct kfd_node *kdev, 1994 struct crat_subtype_iolink *sub_type_hdr, 1995 uint32_t proximity_domain) 1996 { 1997 *avail_size -= sizeof(struct crat_subtype_iolink); 1998 if (*avail_size < 0) 1999 return -ENOMEM; 2000 2001 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink)); 2002 2003 /* Fill in subtype header data */ 2004 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY; 2005 sub_type_hdr->length = sizeof(struct crat_subtype_iolink); 2006 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED; 2007 if (kfd_dev_is_large_bar(kdev)) 2008 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL; 2009 2010 /* Fill in IOLINK subtype. 2011 * TODO: Fill-in other fields of iolink subtype 2012 */ 2013 if (kdev->adev->gmc.xgmi.connected_to_cpu || 2014 (KFD_GC_VERSION(kdev) == IP_VERSION(9, 4, 3) && 2015 kdev->adev->smuio.funcs->get_pkg_type(kdev->adev) == 2016 AMDGPU_PKG_TYPE_APU)) { 2017 bool ext_cpu = KFD_GC_VERSION(kdev) != IP_VERSION(9, 4, 3); 2018 int mem_bw = 819200, weight = ext_cpu ? KFD_CRAT_XGMI_WEIGHT : 2019 KFD_CRAT_INTRA_SOCKET_WEIGHT; 2020 uint32_t bandwidth = ext_cpu ? amdgpu_amdkfd_get_xgmi_bandwidth_mbytes( 2021 kdev->adev, NULL, true) : mem_bw; 2022 2023 /* 2024 * with host gpu xgmi link, host can access gpu memory whether 2025 * or not pcie bar type is large, so always create bidirectional 2026 * io link. 2027 */ 2028 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL; 2029 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI; 2030 sub_type_hdr->weight_xgmi = weight; 2031 sub_type_hdr->minimum_bandwidth_mbs = bandwidth; 2032 sub_type_hdr->maximum_bandwidth_mbs = bandwidth; 2033 } else { 2034 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS; 2035 sub_type_hdr->minimum_bandwidth_mbs = 2036 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, true); 2037 sub_type_hdr->maximum_bandwidth_mbs = 2038 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, false); 2039 } 2040 2041 sub_type_hdr->proximity_domain_from = proximity_domain; 2042 2043 #ifdef CONFIG_ACPI_NUMA 2044 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE && 2045 num_possible_nodes() > 1) 2046 kfd_find_numa_node_in_srat(kdev); 2047 #endif 2048 #ifdef CONFIG_NUMA 2049 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE) 2050 sub_type_hdr->proximity_domain_to = 0; 2051 else 2052 sub_type_hdr->proximity_domain_to = kdev->adev->pdev->dev.numa_node; 2053 #else 2054 sub_type_hdr->proximity_domain_to = 0; 2055 #endif 2056 return 0; 2057 } 2058 2059 static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size, 2060 struct kfd_node *kdev, 2061 struct kfd_node *peer_kdev, 2062 struct crat_subtype_iolink *sub_type_hdr, 2063 uint32_t proximity_domain_from, 2064 uint32_t proximity_domain_to) 2065 { 2066 bool use_ta_info = kdev->kfd->num_nodes == 1; 2067 2068 *avail_size -= sizeof(struct crat_subtype_iolink); 2069 if (*avail_size < 0) 2070 return -ENOMEM; 2071 2072 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink)); 2073 2074 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY; 2075 sub_type_hdr->length = sizeof(struct crat_subtype_iolink); 2076 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED | 2077 CRAT_IOLINK_FLAGS_BI_DIRECTIONAL; 2078 2079 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI; 2080 sub_type_hdr->proximity_domain_from = proximity_domain_from; 2081 sub_type_hdr->proximity_domain_to = proximity_domain_to; 2082 2083 if (use_ta_info) { 2084 sub_type_hdr->weight_xgmi = KFD_CRAT_XGMI_WEIGHT * 2085 amdgpu_amdkfd_get_xgmi_hops_count(kdev->adev, peer_kdev->adev); 2086 sub_type_hdr->maximum_bandwidth_mbs = 2087 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, 2088 peer_kdev->adev, false); 2089 sub_type_hdr->minimum_bandwidth_mbs = sub_type_hdr->maximum_bandwidth_mbs ? 2090 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, NULL, true) : 0; 2091 } else { 2092 bool is_single_hop = kdev->kfd == peer_kdev->kfd; 2093 int weight = is_single_hop ? KFD_CRAT_INTRA_SOCKET_WEIGHT : 2094 (2 * KFD_CRAT_INTRA_SOCKET_WEIGHT) + KFD_CRAT_XGMI_WEIGHT; 2095 int mem_bw = 819200; 2096 2097 sub_type_hdr->weight_xgmi = weight; 2098 sub_type_hdr->maximum_bandwidth_mbs = is_single_hop ? mem_bw : 0; 2099 sub_type_hdr->minimum_bandwidth_mbs = is_single_hop ? mem_bw : 0; 2100 } 2101 2102 return 0; 2103 } 2104 2105 /* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU 2106 * 2107 * @pcrat_image: Fill in VCRAT for GPU 2108 * @size: [IN] allocated size of crat_image. 2109 * [OUT] actual size of data filled in crat_image 2110 */ 2111 static int kfd_create_vcrat_image_gpu(void *pcrat_image, 2112 size_t *size, struct kfd_node *kdev, 2113 uint32_t proximity_domain) 2114 { 2115 struct crat_header *crat_table = (struct crat_header *)pcrat_image; 2116 struct crat_subtype_generic *sub_type_hdr; 2117 struct kfd_local_mem_info local_mem_info; 2118 struct kfd_topology_device *peer_dev; 2119 struct crat_subtype_computeunit *cu; 2120 struct kfd_cu_info cu_info; 2121 int avail_size = *size; 2122 uint32_t total_num_of_cu; 2123 uint32_t nid = 0; 2124 int ret = 0; 2125 2126 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU) 2127 return -EINVAL; 2128 2129 /* Fill the CRAT Header. 2130 * Modify length and total_entries as subunits are added. 2131 */ 2132 avail_size -= sizeof(struct crat_header); 2133 if (avail_size < 0) 2134 return -ENOMEM; 2135 2136 memset(crat_table, 0, sizeof(struct crat_header)); 2137 2138 memcpy(&crat_table->signature, CRAT_SIGNATURE, 2139 sizeof(crat_table->signature)); 2140 /* Change length as we add more subtypes*/ 2141 crat_table->length = sizeof(struct crat_header); 2142 crat_table->num_domains = 1; 2143 crat_table->total_entries = 0; 2144 2145 /* Fill in Subtype: Compute Unit 2146 * First fill in the sub type header and then sub type data 2147 */ 2148 avail_size -= sizeof(struct crat_subtype_computeunit); 2149 if (avail_size < 0) 2150 return -ENOMEM; 2151 2152 sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1); 2153 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit)); 2154 2155 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY; 2156 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit); 2157 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 2158 2159 /* Fill CU subtype data */ 2160 cu = (struct crat_subtype_computeunit *)sub_type_hdr; 2161 cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT; 2162 cu->proximity_domain = proximity_domain; 2163 2164 amdgpu_amdkfd_get_cu_info(kdev->adev, &cu_info); 2165 cu->num_simd_per_cu = cu_info.simd_per_cu; 2166 cu->num_simd_cores = cu_info.simd_per_cu * cu_info.cu_active_number; 2167 cu->max_waves_simd = cu_info.max_waves_per_simd; 2168 2169 cu->wave_front_size = cu_info.wave_front_size; 2170 cu->array_count = cu_info.num_shader_arrays_per_engine * 2171 cu_info.num_shader_engines; 2172 total_num_of_cu = (cu->array_count * cu_info.num_cu_per_sh); 2173 cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu); 2174 cu->num_cu_per_array = cu_info.num_cu_per_sh; 2175 cu->max_slots_scatch_cu = cu_info.max_scratch_slots_per_cu; 2176 cu->num_banks = cu_info.num_shader_engines; 2177 cu->lds_size_in_kb = cu_info.lds_size; 2178 2179 cu->hsa_capability = 0; 2180 2181 /* Check if this node supports IOMMU. During parsing this flag will 2182 * translate to HSA_CAP_ATS_PRESENT 2183 */ 2184 if (!kfd_iommu_check_device(kdev->kfd)) 2185 cu->hsa_capability |= CRAT_CU_FLAGS_IOMMU_PRESENT; 2186 2187 crat_table->length += sub_type_hdr->length; 2188 crat_table->total_entries++; 2189 2190 /* Fill in Subtype: Memory. Only on systems with large BAR (no 2191 * private FB), report memory as public. On other systems 2192 * report the total FB size (public+private) as a single 2193 * private heap. 2194 */ 2195 local_mem_info = kdev->local_mem_info; 2196 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 2197 sub_type_hdr->length); 2198 2199 if (debug_largebar) 2200 local_mem_info.local_mem_size_private = 0; 2201 2202 if (local_mem_info.local_mem_size_private == 0) 2203 ret = kfd_fill_gpu_memory_affinity(&avail_size, 2204 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC, 2205 local_mem_info.local_mem_size_public, 2206 (struct crat_subtype_memory *)sub_type_hdr, 2207 proximity_domain, 2208 &local_mem_info); 2209 else 2210 ret = kfd_fill_gpu_memory_affinity(&avail_size, 2211 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE, 2212 local_mem_info.local_mem_size_public + 2213 local_mem_info.local_mem_size_private, 2214 (struct crat_subtype_memory *)sub_type_hdr, 2215 proximity_domain, 2216 &local_mem_info); 2217 if (ret < 0) 2218 return ret; 2219 2220 crat_table->length += sizeof(struct crat_subtype_memory); 2221 crat_table->total_entries++; 2222 2223 /* Fill in Subtype: IO_LINKS 2224 * Only direct links are added here which is Link from GPU to 2225 * its NUMA node. Indirect links are added by userspace. 2226 */ 2227 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 2228 sub_type_hdr->length); 2229 ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev, 2230 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain); 2231 2232 if (ret < 0) 2233 return ret; 2234 2235 crat_table->length += sub_type_hdr->length; 2236 crat_table->total_entries++; 2237 2238 2239 /* Fill in Subtype: IO_LINKS 2240 * Direct links from GPU to other GPUs through xGMI. 2241 * We will loop GPUs that already be processed (with lower value 2242 * of proximity_domain), add the link for the GPUs with same 2243 * hive id (from this GPU to other GPU) . The reversed iolink 2244 * (from other GPU to this GPU) will be added 2245 * in kfd_parse_subtype_iolink. 2246 */ 2247 if (kdev->kfd->hive_id) { 2248 for (nid = 0; nid < proximity_domain; ++nid) { 2249 peer_dev = kfd_topology_device_by_proximity_domain_no_lock(nid); 2250 if (!peer_dev->gpu) 2251 continue; 2252 if (peer_dev->gpu->kfd->hive_id != kdev->kfd->hive_id) 2253 continue; 2254 sub_type_hdr = (typeof(sub_type_hdr))( 2255 (char *)sub_type_hdr + 2256 sizeof(struct crat_subtype_iolink)); 2257 ret = kfd_fill_gpu_xgmi_link_to_gpu( 2258 &avail_size, kdev, peer_dev->gpu, 2259 (struct crat_subtype_iolink *)sub_type_hdr, 2260 proximity_domain, nid); 2261 if (ret < 0) 2262 return ret; 2263 crat_table->length += sub_type_hdr->length; 2264 crat_table->total_entries++; 2265 } 2266 } 2267 *size = crat_table->length; 2268 pr_info("Virtual CRAT table created for GPU\n"); 2269 2270 return ret; 2271 } 2272 2273 /* kfd_create_crat_image_virtual - Allocates memory for CRAT image and 2274 * creates a Virtual CRAT (VCRAT) image 2275 * 2276 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory 2277 * 2278 * @crat_image: VCRAT image created because ACPI does not have a 2279 * CRAT for this device 2280 * @size: [OUT] size of virtual crat_image 2281 * @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device 2282 * COMPUTE_UNIT_GPU - Create VCRAT for GPU 2283 * (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU 2284 * -- this option is not currently implemented. 2285 * The assumption is that all AMD APUs will have CRAT 2286 * @kdev: Valid kfd_node required if flags contain COMPUTE_UNIT_GPU 2287 * 2288 * Return 0 if successful else return -ve value 2289 */ 2290 int kfd_create_crat_image_virtual(void **crat_image, size_t *size, 2291 int flags, struct kfd_node *kdev, 2292 uint32_t proximity_domain) 2293 { 2294 void *pcrat_image = NULL; 2295 int ret = 0, num_nodes; 2296 size_t dyn_size; 2297 2298 if (!crat_image) 2299 return -EINVAL; 2300 2301 *crat_image = NULL; 2302 2303 /* Allocate the CPU Virtual CRAT size based on the number of online 2304 * nodes. Allocate VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image. 2305 * This should cover all the current conditions. A check is put not 2306 * to overwrite beyond allocated size for GPUs 2307 */ 2308 switch (flags) { 2309 case COMPUTE_UNIT_CPU: 2310 num_nodes = num_online_nodes(); 2311 dyn_size = sizeof(struct crat_header) + 2312 num_nodes * (sizeof(struct crat_subtype_computeunit) + 2313 sizeof(struct crat_subtype_memory) + 2314 (num_nodes - 1) * sizeof(struct crat_subtype_iolink)); 2315 pcrat_image = kvmalloc(dyn_size, GFP_KERNEL); 2316 if (!pcrat_image) 2317 return -ENOMEM; 2318 *size = dyn_size; 2319 pr_debug("CRAT size is %ld", dyn_size); 2320 ret = kfd_create_vcrat_image_cpu(pcrat_image, size); 2321 break; 2322 case COMPUTE_UNIT_GPU: 2323 if (!kdev) 2324 return -EINVAL; 2325 pcrat_image = kvmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL); 2326 if (!pcrat_image) 2327 return -ENOMEM; 2328 *size = VCRAT_SIZE_FOR_GPU; 2329 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev, 2330 proximity_domain); 2331 break; 2332 case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU): 2333 /* TODO: */ 2334 ret = -EINVAL; 2335 pr_err("VCRAT not implemented for APU\n"); 2336 break; 2337 default: 2338 ret = -EINVAL; 2339 } 2340 2341 if (!ret) 2342 *crat_image = pcrat_image; 2343 else 2344 kvfree(pcrat_image); 2345 2346 return ret; 2347 } 2348 2349 2350 /* kfd_destroy_crat_image 2351 * 2352 * @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..) 2353 * 2354 */ 2355 void kfd_destroy_crat_image(void *crat_image) 2356 { 2357 kvfree(crat_image); 2358 } 2359