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 "amdgpu.h" 30 #include "amdgpu_amdkfd.h" 31 #include "amdgpu_xgmi.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 */ 38 static 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 */ 44 static 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 54 static struct kfd_gpu_cache_info kaveri_cache_info[] = { 55 { 56 /* TCP L1 Cache per CU */ 57 .cache_size = 16, 58 .cache_level = 1, 59 .cache_line_size = 64, 60 .flags = (CRAT_CACHE_FLAGS_ENABLED | 61 CRAT_CACHE_FLAGS_DATA_CACHE | 62 CRAT_CACHE_FLAGS_SIMD_CACHE), 63 .num_cu_shared = 1, 64 }, 65 { 66 /* Scalar L1 Instruction Cache (in SQC module) per bank */ 67 .cache_size = 16, 68 .cache_level = 1, 69 .cache_line_size = 64, 70 .flags = (CRAT_CACHE_FLAGS_ENABLED | 71 CRAT_CACHE_FLAGS_INST_CACHE | 72 CRAT_CACHE_FLAGS_SIMD_CACHE), 73 .num_cu_shared = 2, 74 }, 75 { 76 /* Scalar L1 Data Cache (in SQC module) per bank */ 77 .cache_size = 8, 78 .cache_level = 1, 79 .cache_line_size = 64, 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 .cache_line_size = 64, 96 .flags = (CRAT_CACHE_FLAGS_ENABLED | 97 CRAT_CACHE_FLAGS_DATA_CACHE | 98 CRAT_CACHE_FLAGS_SIMD_CACHE), 99 .num_cu_shared = 1, 100 }, 101 { 102 /* Scalar L1 Instruction Cache (in SQC module) per bank */ 103 .cache_size = 32, 104 .cache_level = 1, 105 .cache_line_size = 64, 106 .flags = (CRAT_CACHE_FLAGS_ENABLED | 107 CRAT_CACHE_FLAGS_INST_CACHE | 108 CRAT_CACHE_FLAGS_SIMD_CACHE), 109 .num_cu_shared = 4, 110 }, 111 { 112 /* Scalar L1 Data Cache (in SQC module) per bank. */ 113 .cache_size = 16, 114 .cache_level = 1, 115 .cache_line_size = 64, 116 .flags = (CRAT_CACHE_FLAGS_ENABLED | 117 CRAT_CACHE_FLAGS_DATA_CACHE | 118 CRAT_CACHE_FLAGS_SIMD_CACHE), 119 .num_cu_shared = 4, 120 }, 121 122 /* TODO: Add L2 Cache information */ 123 }; 124 125 #define hawaii_cache_info kaveri_cache_info 126 #define tonga_cache_info carrizo_cache_info 127 #define fiji_cache_info carrizo_cache_info 128 #define polaris10_cache_info carrizo_cache_info 129 #define polaris11_cache_info carrizo_cache_info 130 #define polaris12_cache_info carrizo_cache_info 131 #define vegam_cache_info carrizo_cache_info 132 133 /* NOTE: L1 cache information has been updated and L2/L3 134 * cache information has been added for Vega10 and 135 * newer ASICs. The unit for cache_size is KiB. 136 * In future, check & update cache details 137 * for every new ASIC is required. 138 */ 139 140 static struct kfd_gpu_cache_info vega10_cache_info[] = { 141 { 142 /* TCP L1 Cache per CU */ 143 .cache_size = 16, 144 .cache_level = 1, 145 .cache_line_size = 64, 146 .flags = (CRAT_CACHE_FLAGS_ENABLED | 147 CRAT_CACHE_FLAGS_DATA_CACHE | 148 CRAT_CACHE_FLAGS_SIMD_CACHE), 149 .num_cu_shared = 1, 150 }, 151 { 152 /* Scalar L1 Instruction Cache per SQC */ 153 .cache_size = 32, 154 .cache_level = 1, 155 .cache_line_size = 64, 156 .flags = (CRAT_CACHE_FLAGS_ENABLED | 157 CRAT_CACHE_FLAGS_INST_CACHE | 158 CRAT_CACHE_FLAGS_SIMD_CACHE), 159 .num_cu_shared = 3, 160 }, 161 { 162 /* Scalar L1 Data Cache per SQC */ 163 .cache_size = 16, 164 .cache_level = 1, 165 .cache_line_size = 64, 166 .flags = (CRAT_CACHE_FLAGS_ENABLED | 167 CRAT_CACHE_FLAGS_DATA_CACHE | 168 CRAT_CACHE_FLAGS_SIMD_CACHE), 169 .num_cu_shared = 3, 170 }, 171 { 172 /* L2 Data Cache per GPU (Total Tex Cache) */ 173 .cache_size = 4096, 174 .cache_level = 2, 175 .cache_line_size = 64, 176 .flags = (CRAT_CACHE_FLAGS_ENABLED | 177 CRAT_CACHE_FLAGS_DATA_CACHE | 178 CRAT_CACHE_FLAGS_SIMD_CACHE), 179 .num_cu_shared = 16, 180 }, 181 }; 182 183 static struct kfd_gpu_cache_info raven_cache_info[] = { 184 { 185 /* TCP L1 Cache per CU */ 186 .cache_size = 16, 187 .cache_level = 1, 188 .cache_line_size = 64, 189 .flags = (CRAT_CACHE_FLAGS_ENABLED | 190 CRAT_CACHE_FLAGS_DATA_CACHE | 191 CRAT_CACHE_FLAGS_SIMD_CACHE), 192 .num_cu_shared = 1, 193 }, 194 { 195 /* Scalar L1 Instruction Cache per SQC */ 196 .cache_size = 32, 197 .cache_level = 1, 198 .cache_line_size = 64, 199 .flags = (CRAT_CACHE_FLAGS_ENABLED | 200 CRAT_CACHE_FLAGS_INST_CACHE | 201 CRAT_CACHE_FLAGS_SIMD_CACHE), 202 .num_cu_shared = 3, 203 }, 204 { 205 /* Scalar L1 Data Cache per SQC */ 206 .cache_size = 16, 207 .cache_level = 1, 208 .cache_line_size = 64, 209 .flags = (CRAT_CACHE_FLAGS_ENABLED | 210 CRAT_CACHE_FLAGS_DATA_CACHE | 211 CRAT_CACHE_FLAGS_SIMD_CACHE), 212 .num_cu_shared = 3, 213 }, 214 { 215 /* L2 Data Cache per GPU (Total Tex Cache) */ 216 .cache_size = 1024, 217 .cache_level = 2, 218 .cache_line_size = 64, 219 .flags = (CRAT_CACHE_FLAGS_ENABLED | 220 CRAT_CACHE_FLAGS_DATA_CACHE | 221 CRAT_CACHE_FLAGS_SIMD_CACHE), 222 .num_cu_shared = 11, 223 }, 224 }; 225 226 static struct kfd_gpu_cache_info renoir_cache_info[] = { 227 { 228 /* TCP L1 Cache per CU */ 229 .cache_size = 16, 230 .cache_level = 1, 231 .cache_line_size = 64, 232 .flags = (CRAT_CACHE_FLAGS_ENABLED | 233 CRAT_CACHE_FLAGS_DATA_CACHE | 234 CRAT_CACHE_FLAGS_SIMD_CACHE), 235 .num_cu_shared = 1, 236 }, 237 { 238 /* Scalar L1 Instruction Cache per SQC */ 239 .cache_size = 32, 240 .cache_level = 1, 241 .cache_line_size = 64, 242 .flags = (CRAT_CACHE_FLAGS_ENABLED | 243 CRAT_CACHE_FLAGS_INST_CACHE | 244 CRAT_CACHE_FLAGS_SIMD_CACHE), 245 .num_cu_shared = 3, 246 }, 247 { 248 /* Scalar L1 Data Cache per SQC */ 249 .cache_size = 16, 250 .cache_level = 1, 251 .cache_line_size = 64, 252 .flags = (CRAT_CACHE_FLAGS_ENABLED | 253 CRAT_CACHE_FLAGS_DATA_CACHE | 254 CRAT_CACHE_FLAGS_SIMD_CACHE), 255 .num_cu_shared = 3, 256 }, 257 { 258 /* L2 Data Cache per GPU (Total Tex Cache) */ 259 .cache_size = 1024, 260 .cache_level = 2, 261 .cache_line_size = 64, 262 .flags = (CRAT_CACHE_FLAGS_ENABLED | 263 CRAT_CACHE_FLAGS_DATA_CACHE | 264 CRAT_CACHE_FLAGS_SIMD_CACHE), 265 .num_cu_shared = 8, 266 }, 267 }; 268 269 static struct kfd_gpu_cache_info vega12_cache_info[] = { 270 { 271 /* TCP L1 Cache per CU */ 272 .cache_size = 16, 273 .cache_level = 1, 274 .cache_line_size = 64, 275 .flags = (CRAT_CACHE_FLAGS_ENABLED | 276 CRAT_CACHE_FLAGS_DATA_CACHE | 277 CRAT_CACHE_FLAGS_SIMD_CACHE), 278 .num_cu_shared = 1, 279 }, 280 { 281 /* Scalar L1 Instruction Cache per SQC */ 282 .cache_size = 32, 283 .cache_level = 1, 284 .cache_line_size = 64, 285 .flags = (CRAT_CACHE_FLAGS_ENABLED | 286 CRAT_CACHE_FLAGS_INST_CACHE | 287 CRAT_CACHE_FLAGS_SIMD_CACHE), 288 .num_cu_shared = 3, 289 }, 290 { 291 /* Scalar L1 Data Cache per SQC */ 292 .cache_size = 16, 293 .cache_level = 1, 294 .cache_line_size = 64, 295 .flags = (CRAT_CACHE_FLAGS_ENABLED | 296 CRAT_CACHE_FLAGS_DATA_CACHE | 297 CRAT_CACHE_FLAGS_SIMD_CACHE), 298 .num_cu_shared = 3, 299 }, 300 { 301 /* L2 Data Cache per GPU (Total Tex Cache) */ 302 .cache_size = 2048, 303 .cache_level = 2, 304 .cache_line_size = 64, 305 .flags = (CRAT_CACHE_FLAGS_ENABLED | 306 CRAT_CACHE_FLAGS_DATA_CACHE | 307 CRAT_CACHE_FLAGS_SIMD_CACHE), 308 .num_cu_shared = 5, 309 }, 310 }; 311 312 static struct kfd_gpu_cache_info vega20_cache_info[] = { 313 { 314 /* TCP L1 Cache per CU */ 315 .cache_size = 16, 316 .cache_level = 1, 317 .cache_line_size = 64, 318 .flags = (CRAT_CACHE_FLAGS_ENABLED | 319 CRAT_CACHE_FLAGS_DATA_CACHE | 320 CRAT_CACHE_FLAGS_SIMD_CACHE), 321 .num_cu_shared = 1, 322 }, 323 { 324 /* Scalar L1 Instruction Cache per SQC */ 325 .cache_size = 32, 326 .cache_level = 1, 327 .cache_line_size = 64, 328 .flags = (CRAT_CACHE_FLAGS_ENABLED | 329 CRAT_CACHE_FLAGS_INST_CACHE | 330 CRAT_CACHE_FLAGS_SIMD_CACHE), 331 .num_cu_shared = 3, 332 }, 333 { 334 /* Scalar L1 Data Cache per SQC */ 335 .cache_size = 16, 336 .cache_level = 1, 337 .cache_line_size = 64, 338 .flags = (CRAT_CACHE_FLAGS_ENABLED | 339 CRAT_CACHE_FLAGS_DATA_CACHE | 340 CRAT_CACHE_FLAGS_SIMD_CACHE), 341 .num_cu_shared = 3, 342 }, 343 { 344 /* L2 Data Cache per GPU (Total Tex Cache) */ 345 .cache_size = 8192, 346 .cache_level = 2, 347 .cache_line_size = 64, 348 .flags = (CRAT_CACHE_FLAGS_ENABLED | 349 CRAT_CACHE_FLAGS_DATA_CACHE | 350 CRAT_CACHE_FLAGS_SIMD_CACHE), 351 .num_cu_shared = 16, 352 }, 353 }; 354 355 static struct kfd_gpu_cache_info aldebaran_cache_info[] = { 356 { 357 /* TCP L1 Cache per CU */ 358 .cache_size = 16, 359 .cache_level = 1, 360 .cache_line_size = 64, 361 .flags = (CRAT_CACHE_FLAGS_ENABLED | 362 CRAT_CACHE_FLAGS_DATA_CACHE | 363 CRAT_CACHE_FLAGS_SIMD_CACHE), 364 .num_cu_shared = 1, 365 }, 366 { 367 /* Scalar L1 Instruction Cache per SQC */ 368 .cache_size = 32, 369 .cache_level = 1, 370 .cache_line_size = 64, 371 .flags = (CRAT_CACHE_FLAGS_ENABLED | 372 CRAT_CACHE_FLAGS_INST_CACHE | 373 CRAT_CACHE_FLAGS_SIMD_CACHE), 374 .num_cu_shared = 2, 375 }, 376 { 377 /* Scalar L1 Data Cache per SQC */ 378 .cache_size = 16, 379 .cache_level = 1, 380 .cache_line_size = 64, 381 .flags = (CRAT_CACHE_FLAGS_ENABLED | 382 CRAT_CACHE_FLAGS_DATA_CACHE | 383 CRAT_CACHE_FLAGS_SIMD_CACHE), 384 .num_cu_shared = 2, 385 }, 386 { 387 /* L2 Data Cache per GPU (Total Tex Cache) */ 388 .cache_size = 8192, 389 .cache_level = 2, 390 .cache_line_size = 128, 391 .flags = (CRAT_CACHE_FLAGS_ENABLED | 392 CRAT_CACHE_FLAGS_DATA_CACHE | 393 CRAT_CACHE_FLAGS_SIMD_CACHE), 394 .num_cu_shared = 14, 395 }, 396 }; 397 398 static struct kfd_gpu_cache_info navi10_cache_info[] = { 399 { 400 /* TCP L1 Cache per CU */ 401 .cache_size = 16, 402 .cache_level = 1, 403 .cache_line_size = 128, 404 .flags = (CRAT_CACHE_FLAGS_ENABLED | 405 CRAT_CACHE_FLAGS_DATA_CACHE | 406 CRAT_CACHE_FLAGS_SIMD_CACHE), 407 .num_cu_shared = 1, 408 }, 409 { 410 /* Scalar L1 Instruction Cache per SQC */ 411 .cache_size = 32, 412 .cache_level = 1, 413 .cache_line_size = 64, 414 .flags = (CRAT_CACHE_FLAGS_ENABLED | 415 CRAT_CACHE_FLAGS_INST_CACHE | 416 CRAT_CACHE_FLAGS_SIMD_CACHE), 417 .num_cu_shared = 2, 418 }, 419 { 420 /* Scalar L1 Data Cache per SQC */ 421 .cache_size = 16, 422 .cache_level = 1, 423 .cache_line_size = 64, 424 .flags = (CRAT_CACHE_FLAGS_ENABLED | 425 CRAT_CACHE_FLAGS_DATA_CACHE | 426 CRAT_CACHE_FLAGS_SIMD_CACHE), 427 .num_cu_shared = 2, 428 }, 429 { 430 /* GL1 Data Cache per SA */ 431 .cache_size = 128, 432 .cache_level = 1, 433 .cache_line_size = 128, 434 .flags = (CRAT_CACHE_FLAGS_ENABLED | 435 CRAT_CACHE_FLAGS_DATA_CACHE | 436 CRAT_CACHE_FLAGS_SIMD_CACHE), 437 .num_cu_shared = 10, 438 }, 439 { 440 /* L2 Data Cache per GPU (Total Tex Cache) */ 441 .cache_size = 4096, 442 .cache_level = 2, 443 .cache_line_size = 128, 444 .flags = (CRAT_CACHE_FLAGS_ENABLED | 445 CRAT_CACHE_FLAGS_DATA_CACHE | 446 CRAT_CACHE_FLAGS_SIMD_CACHE), 447 .num_cu_shared = 10, 448 }, 449 }; 450 451 static struct kfd_gpu_cache_info vangogh_cache_info[] = { 452 { 453 /* TCP L1 Cache per CU */ 454 .cache_size = 16, 455 .cache_level = 1, 456 .cache_line_size = 128, 457 .flags = (CRAT_CACHE_FLAGS_ENABLED | 458 CRAT_CACHE_FLAGS_DATA_CACHE | 459 CRAT_CACHE_FLAGS_SIMD_CACHE), 460 .num_cu_shared = 1, 461 }, 462 { 463 /* Scalar L1 Instruction Cache per SQC */ 464 .cache_size = 32, 465 .cache_level = 1, 466 .cache_line_size = 64, 467 .flags = (CRAT_CACHE_FLAGS_ENABLED | 468 CRAT_CACHE_FLAGS_INST_CACHE | 469 CRAT_CACHE_FLAGS_SIMD_CACHE), 470 .num_cu_shared = 2, 471 }, 472 { 473 /* Scalar L1 Data Cache per SQC */ 474 .cache_size = 16, 475 .cache_level = 1, 476 .cache_line_size = 64, 477 .flags = (CRAT_CACHE_FLAGS_ENABLED | 478 CRAT_CACHE_FLAGS_DATA_CACHE | 479 CRAT_CACHE_FLAGS_SIMD_CACHE), 480 .num_cu_shared = 2, 481 }, 482 { 483 /* GL1 Data Cache per SA */ 484 .cache_size = 128, 485 .cache_level = 1, 486 .cache_line_size = 128, 487 .flags = (CRAT_CACHE_FLAGS_ENABLED | 488 CRAT_CACHE_FLAGS_DATA_CACHE | 489 CRAT_CACHE_FLAGS_SIMD_CACHE), 490 .num_cu_shared = 8, 491 }, 492 { 493 /* L2 Data Cache per GPU (Total Tex Cache) */ 494 .cache_size = 1024, 495 .cache_level = 2, 496 .cache_line_size = 128, 497 .flags = (CRAT_CACHE_FLAGS_ENABLED | 498 CRAT_CACHE_FLAGS_DATA_CACHE | 499 CRAT_CACHE_FLAGS_SIMD_CACHE), 500 .num_cu_shared = 8, 501 }, 502 }; 503 504 static struct kfd_gpu_cache_info navi14_cache_info[] = { 505 { 506 /* TCP L1 Cache per CU */ 507 .cache_size = 16, 508 .cache_level = 1, 509 .cache_line_size = 128, 510 .flags = (CRAT_CACHE_FLAGS_ENABLED | 511 CRAT_CACHE_FLAGS_DATA_CACHE | 512 CRAT_CACHE_FLAGS_SIMD_CACHE), 513 .num_cu_shared = 1, 514 }, 515 { 516 /* Scalar L1 Instruction Cache per SQC */ 517 .cache_size = 32, 518 .cache_level = 1, 519 .cache_line_size = 64, 520 .flags = (CRAT_CACHE_FLAGS_ENABLED | 521 CRAT_CACHE_FLAGS_INST_CACHE | 522 CRAT_CACHE_FLAGS_SIMD_CACHE), 523 .num_cu_shared = 2, 524 }, 525 { 526 /* Scalar L1 Data Cache per SQC */ 527 .cache_size = 16, 528 .cache_level = 1, 529 .cache_line_size = 64, 530 .flags = (CRAT_CACHE_FLAGS_ENABLED | 531 CRAT_CACHE_FLAGS_DATA_CACHE | 532 CRAT_CACHE_FLAGS_SIMD_CACHE), 533 .num_cu_shared = 2, 534 }, 535 { 536 /* GL1 Data Cache per SA */ 537 .cache_size = 128, 538 .cache_level = 1, 539 .cache_line_size = 128, 540 .flags = (CRAT_CACHE_FLAGS_ENABLED | 541 CRAT_CACHE_FLAGS_DATA_CACHE | 542 CRAT_CACHE_FLAGS_SIMD_CACHE), 543 .num_cu_shared = 12, 544 }, 545 { 546 /* L2 Data Cache per GPU (Total Tex Cache) */ 547 .cache_size = 2048, 548 .cache_level = 2, 549 .cache_line_size = 128, 550 .flags = (CRAT_CACHE_FLAGS_ENABLED | 551 CRAT_CACHE_FLAGS_DATA_CACHE | 552 CRAT_CACHE_FLAGS_SIMD_CACHE), 553 .num_cu_shared = 12, 554 }, 555 }; 556 557 static struct kfd_gpu_cache_info sienna_cichlid_cache_info[] = { 558 { 559 /* TCP L1 Cache per CU */ 560 .cache_size = 16, 561 .cache_level = 1, 562 .cache_line_size = 128, 563 .flags = (CRAT_CACHE_FLAGS_ENABLED | 564 CRAT_CACHE_FLAGS_DATA_CACHE | 565 CRAT_CACHE_FLAGS_SIMD_CACHE), 566 .num_cu_shared = 1, 567 }, 568 { 569 /* Scalar L1 Instruction Cache per SQC */ 570 .cache_size = 32, 571 .cache_level = 1, 572 .cache_line_size = 64, 573 .flags = (CRAT_CACHE_FLAGS_ENABLED | 574 CRAT_CACHE_FLAGS_INST_CACHE | 575 CRAT_CACHE_FLAGS_SIMD_CACHE), 576 .num_cu_shared = 2, 577 }, 578 { 579 /* Scalar L1 Data Cache per SQC */ 580 .cache_size = 16, 581 .cache_level = 1, 582 .cache_line_size = 64, 583 .flags = (CRAT_CACHE_FLAGS_ENABLED | 584 CRAT_CACHE_FLAGS_DATA_CACHE | 585 CRAT_CACHE_FLAGS_SIMD_CACHE), 586 .num_cu_shared = 2, 587 }, 588 { 589 /* GL1 Data Cache per SA */ 590 .cache_size = 128, 591 .cache_level = 1, 592 .cache_line_size = 128, 593 .flags = (CRAT_CACHE_FLAGS_ENABLED | 594 CRAT_CACHE_FLAGS_DATA_CACHE | 595 CRAT_CACHE_FLAGS_SIMD_CACHE), 596 .num_cu_shared = 10, 597 }, 598 { 599 /* L2 Data Cache per GPU (Total Tex Cache) */ 600 .cache_size = 4096, 601 .cache_level = 2, 602 .cache_line_size = 128, 603 .flags = (CRAT_CACHE_FLAGS_ENABLED | 604 CRAT_CACHE_FLAGS_DATA_CACHE | 605 CRAT_CACHE_FLAGS_SIMD_CACHE), 606 .num_cu_shared = 10, 607 }, 608 { 609 /* L3 Data Cache per GPU */ 610 .cache_size = 128*1024, 611 .cache_level = 3, 612 .cache_line_size = 64, 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 }; 619 620 static struct kfd_gpu_cache_info navy_flounder_cache_info[] = { 621 { 622 /* TCP L1 Cache per CU */ 623 .cache_size = 16, 624 .cache_level = 1, 625 .cache_line_size = 128, 626 .flags = (CRAT_CACHE_FLAGS_ENABLED | 627 CRAT_CACHE_FLAGS_DATA_CACHE | 628 CRAT_CACHE_FLAGS_SIMD_CACHE), 629 .num_cu_shared = 1, 630 }, 631 { 632 /* Scalar L1 Instruction Cache per SQC */ 633 .cache_size = 32, 634 .cache_level = 1, 635 .cache_line_size = 64, 636 .flags = (CRAT_CACHE_FLAGS_ENABLED | 637 CRAT_CACHE_FLAGS_INST_CACHE | 638 CRAT_CACHE_FLAGS_SIMD_CACHE), 639 .num_cu_shared = 2, 640 }, 641 { 642 /* Scalar L1 Data Cache per SQC */ 643 .cache_size = 16, 644 .cache_level = 1, 645 .cache_line_size = 64, 646 .flags = (CRAT_CACHE_FLAGS_ENABLED | 647 CRAT_CACHE_FLAGS_DATA_CACHE | 648 CRAT_CACHE_FLAGS_SIMD_CACHE), 649 .num_cu_shared = 2, 650 }, 651 { 652 /* GL1 Data Cache per SA */ 653 .cache_size = 128, 654 .cache_level = 1, 655 .cache_line_size = 128, 656 .flags = (CRAT_CACHE_FLAGS_ENABLED | 657 CRAT_CACHE_FLAGS_DATA_CACHE | 658 CRAT_CACHE_FLAGS_SIMD_CACHE), 659 .num_cu_shared = 10, 660 }, 661 { 662 /* L2 Data Cache per GPU (Total Tex Cache) */ 663 .cache_size = 3072, 664 .cache_level = 2, 665 .cache_line_size = 128, 666 .flags = (CRAT_CACHE_FLAGS_ENABLED | 667 CRAT_CACHE_FLAGS_DATA_CACHE | 668 CRAT_CACHE_FLAGS_SIMD_CACHE), 669 .num_cu_shared = 10, 670 }, 671 { 672 /* L3 Data Cache per GPU */ 673 .cache_size = 96*1024, 674 .cache_level = 3, 675 .cache_line_size = 64, 676 .flags = (CRAT_CACHE_FLAGS_ENABLED | 677 CRAT_CACHE_FLAGS_DATA_CACHE | 678 CRAT_CACHE_FLAGS_SIMD_CACHE), 679 .num_cu_shared = 10, 680 }, 681 }; 682 683 static struct kfd_gpu_cache_info dimgrey_cavefish_cache_info[] = { 684 { 685 /* TCP L1 Cache per CU */ 686 .cache_size = 16, 687 .cache_level = 1, 688 .cache_line_size = 128, 689 .flags = (CRAT_CACHE_FLAGS_ENABLED | 690 CRAT_CACHE_FLAGS_DATA_CACHE | 691 CRAT_CACHE_FLAGS_SIMD_CACHE), 692 .num_cu_shared = 1, 693 }, 694 { 695 /* Scalar L1 Instruction Cache per SQC */ 696 .cache_size = 32, 697 .cache_level = 1, 698 .cache_line_size = 64, 699 .flags = (CRAT_CACHE_FLAGS_ENABLED | 700 CRAT_CACHE_FLAGS_INST_CACHE | 701 CRAT_CACHE_FLAGS_SIMD_CACHE), 702 .num_cu_shared = 2, 703 }, 704 { 705 /* Scalar L1 Data Cache per SQC */ 706 .cache_size = 16, 707 .cache_level = 1, 708 .cache_line_size = 64, 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 .cache_line_size = 128, 719 .flags = (CRAT_CACHE_FLAGS_ENABLED | 720 CRAT_CACHE_FLAGS_DATA_CACHE | 721 CRAT_CACHE_FLAGS_SIMD_CACHE), 722 .num_cu_shared = 8, 723 }, 724 { 725 /* L2 Data Cache per GPU (Total Tex Cache) */ 726 .cache_size = 2048, 727 .cache_level = 2, 728 .cache_line_size = 128, 729 .flags = (CRAT_CACHE_FLAGS_ENABLED | 730 CRAT_CACHE_FLAGS_DATA_CACHE | 731 CRAT_CACHE_FLAGS_SIMD_CACHE), 732 .num_cu_shared = 8, 733 }, 734 { 735 /* L3 Data Cache per GPU */ 736 .cache_size = 32*1024, 737 .cache_level = 3, 738 .cache_line_size = 64, 739 .flags = (CRAT_CACHE_FLAGS_ENABLED | 740 CRAT_CACHE_FLAGS_DATA_CACHE | 741 CRAT_CACHE_FLAGS_SIMD_CACHE), 742 .num_cu_shared = 8, 743 }, 744 }; 745 746 static struct kfd_gpu_cache_info beige_goby_cache_info[] = { 747 { 748 /* TCP L1 Cache per CU */ 749 .cache_size = 16, 750 .cache_level = 1, 751 .cache_line_size = 128, 752 .flags = (CRAT_CACHE_FLAGS_ENABLED | 753 CRAT_CACHE_FLAGS_DATA_CACHE | 754 CRAT_CACHE_FLAGS_SIMD_CACHE), 755 .num_cu_shared = 1, 756 }, 757 { 758 /* Scalar L1 Instruction Cache per SQC */ 759 .cache_size = 32, 760 .cache_level = 1, 761 .cache_line_size = 64, 762 .flags = (CRAT_CACHE_FLAGS_ENABLED | 763 CRAT_CACHE_FLAGS_INST_CACHE | 764 CRAT_CACHE_FLAGS_SIMD_CACHE), 765 .num_cu_shared = 2, 766 }, 767 { 768 /* Scalar L1 Data Cache per SQC */ 769 .cache_size = 16, 770 .cache_level = 1, 771 .cache_line_size = 64, 772 .flags = (CRAT_CACHE_FLAGS_ENABLED | 773 CRAT_CACHE_FLAGS_DATA_CACHE | 774 CRAT_CACHE_FLAGS_SIMD_CACHE), 775 .num_cu_shared = 2, 776 }, 777 { 778 /* GL1 Data Cache per SA */ 779 .cache_size = 128, 780 .cache_level = 1, 781 .cache_line_size = 128, 782 .flags = (CRAT_CACHE_FLAGS_ENABLED | 783 CRAT_CACHE_FLAGS_DATA_CACHE | 784 CRAT_CACHE_FLAGS_SIMD_CACHE), 785 .num_cu_shared = 8, 786 }, 787 { 788 /* L2 Data Cache per GPU (Total Tex Cache) */ 789 .cache_size = 1024, 790 .cache_level = 2, 791 .cache_line_size = 128, 792 .flags = (CRAT_CACHE_FLAGS_ENABLED | 793 CRAT_CACHE_FLAGS_DATA_CACHE | 794 CRAT_CACHE_FLAGS_SIMD_CACHE), 795 .num_cu_shared = 8, 796 }, 797 { 798 /* L3 Data Cache per GPU */ 799 .cache_size = 16*1024, 800 .cache_level = 3, 801 .cache_line_size = 64, 802 .flags = (CRAT_CACHE_FLAGS_ENABLED | 803 CRAT_CACHE_FLAGS_DATA_CACHE | 804 CRAT_CACHE_FLAGS_SIMD_CACHE), 805 .num_cu_shared = 8, 806 }, 807 }; 808 809 static struct kfd_gpu_cache_info yellow_carp_cache_info[] = { 810 { 811 /* TCP L1 Cache per CU */ 812 .cache_size = 16, 813 .cache_level = 1, 814 .cache_line_size = 128, 815 .flags = (CRAT_CACHE_FLAGS_ENABLED | 816 CRAT_CACHE_FLAGS_DATA_CACHE | 817 CRAT_CACHE_FLAGS_SIMD_CACHE), 818 .num_cu_shared = 1, 819 }, 820 { 821 /* Scalar L1 Instruction Cache per SQC */ 822 .cache_size = 32, 823 .cache_level = 1, 824 .cache_line_size = 64, 825 .flags = (CRAT_CACHE_FLAGS_ENABLED | 826 CRAT_CACHE_FLAGS_INST_CACHE | 827 CRAT_CACHE_FLAGS_SIMD_CACHE), 828 .num_cu_shared = 2, 829 }, 830 { 831 /* Scalar L1 Data Cache per SQC */ 832 .cache_size = 16, 833 .cache_level = 1, 834 .cache_line_size = 64, 835 .flags = (CRAT_CACHE_FLAGS_ENABLED | 836 CRAT_CACHE_FLAGS_DATA_CACHE | 837 CRAT_CACHE_FLAGS_SIMD_CACHE), 838 .num_cu_shared = 2, 839 }, 840 { 841 /* GL1 Data Cache per SA */ 842 .cache_size = 128, 843 .cache_level = 1, 844 .cache_line_size = 128, 845 .flags = (CRAT_CACHE_FLAGS_ENABLED | 846 CRAT_CACHE_FLAGS_DATA_CACHE | 847 CRAT_CACHE_FLAGS_SIMD_CACHE), 848 .num_cu_shared = 6, 849 }, 850 { 851 /* L2 Data Cache per GPU (Total Tex Cache) */ 852 .cache_size = 2048, 853 .cache_level = 2, 854 .cache_line_size = 128, 855 .flags = (CRAT_CACHE_FLAGS_ENABLED | 856 CRAT_CACHE_FLAGS_DATA_CACHE | 857 CRAT_CACHE_FLAGS_SIMD_CACHE), 858 .num_cu_shared = 6, 859 }, 860 }; 861 862 static struct kfd_gpu_cache_info gfx1037_cache_info[] = { 863 { 864 /* TCP L1 Cache per CU */ 865 .cache_size = 16, 866 .cache_level = 1, 867 .cache_line_size = 128, 868 .flags = (CRAT_CACHE_FLAGS_ENABLED | 869 CRAT_CACHE_FLAGS_DATA_CACHE | 870 CRAT_CACHE_FLAGS_SIMD_CACHE), 871 .num_cu_shared = 1, 872 }, 873 { 874 /* Scalar L1 Instruction Cache per SQC */ 875 .cache_size = 32, 876 .cache_level = 1, 877 .cache_line_size = 64, 878 .flags = (CRAT_CACHE_FLAGS_ENABLED | 879 CRAT_CACHE_FLAGS_INST_CACHE | 880 CRAT_CACHE_FLAGS_SIMD_CACHE), 881 .num_cu_shared = 2, 882 }, 883 { 884 /* Scalar L1 Data Cache per SQC */ 885 .cache_size = 16, 886 .cache_level = 1, 887 .cache_line_size = 64, 888 .flags = (CRAT_CACHE_FLAGS_ENABLED | 889 CRAT_CACHE_FLAGS_DATA_CACHE | 890 CRAT_CACHE_FLAGS_SIMD_CACHE), 891 .num_cu_shared = 2, 892 }, 893 { 894 /* GL1 Data Cache per SA */ 895 .cache_size = 128, 896 .cache_level = 1, 897 .cache_line_size = 128, 898 .flags = (CRAT_CACHE_FLAGS_ENABLED | 899 CRAT_CACHE_FLAGS_DATA_CACHE | 900 CRAT_CACHE_FLAGS_SIMD_CACHE), 901 .num_cu_shared = 2, 902 }, 903 { 904 /* L2 Data Cache per GPU (Total Tex Cache) */ 905 .cache_size = 256, 906 .cache_level = 2, 907 .cache_line_size = 128, 908 .flags = (CRAT_CACHE_FLAGS_ENABLED | 909 CRAT_CACHE_FLAGS_DATA_CACHE | 910 CRAT_CACHE_FLAGS_SIMD_CACHE), 911 .num_cu_shared = 2, 912 }, 913 }; 914 915 static struct kfd_gpu_cache_info gc_10_3_6_cache_info[] = { 916 { 917 /* TCP L1 Cache per CU */ 918 .cache_size = 16, 919 .cache_level = 1, 920 .cache_line_size = 128, 921 .flags = (CRAT_CACHE_FLAGS_ENABLED | 922 CRAT_CACHE_FLAGS_DATA_CACHE | 923 CRAT_CACHE_FLAGS_SIMD_CACHE), 924 .num_cu_shared = 1, 925 }, 926 { 927 /* Scalar L1 Instruction Cache per SQC */ 928 .cache_size = 32, 929 .cache_level = 1, 930 .cache_line_size = 64, 931 .flags = (CRAT_CACHE_FLAGS_ENABLED | 932 CRAT_CACHE_FLAGS_INST_CACHE | 933 CRAT_CACHE_FLAGS_SIMD_CACHE), 934 .num_cu_shared = 2, 935 }, 936 { 937 /* Scalar L1 Data Cache per SQC */ 938 .cache_size = 16, 939 .cache_level = 1, 940 .cache_line_size = 64, 941 .flags = (CRAT_CACHE_FLAGS_ENABLED | 942 CRAT_CACHE_FLAGS_DATA_CACHE | 943 CRAT_CACHE_FLAGS_SIMD_CACHE), 944 .num_cu_shared = 2, 945 }, 946 { 947 /* GL1 Data Cache per SA */ 948 .cache_size = 128, 949 .cache_level = 1, 950 .cache_line_size = 128, 951 .flags = (CRAT_CACHE_FLAGS_ENABLED | 952 CRAT_CACHE_FLAGS_DATA_CACHE | 953 CRAT_CACHE_FLAGS_SIMD_CACHE), 954 .num_cu_shared = 2, 955 }, 956 { 957 /* L2 Data Cache per GPU (Total Tex Cache) */ 958 .cache_size = 256, 959 .cache_level = 2, 960 .cache_line_size = 128, 961 .flags = (CRAT_CACHE_FLAGS_ENABLED | 962 CRAT_CACHE_FLAGS_DATA_CACHE | 963 CRAT_CACHE_FLAGS_SIMD_CACHE), 964 .num_cu_shared = 2, 965 }, 966 }; 967 968 static struct kfd_gpu_cache_info dummy_cache_info[] = { 969 { 970 /* TCP L1 Cache per CU */ 971 .cache_size = 16, 972 .cache_level = 1, 973 .cache_line_size = 64, 974 .flags = (CRAT_CACHE_FLAGS_ENABLED | 975 CRAT_CACHE_FLAGS_DATA_CACHE | 976 CRAT_CACHE_FLAGS_SIMD_CACHE), 977 .num_cu_shared = 1, 978 }, 979 { 980 /* Scalar L1 Instruction Cache per SQC */ 981 .cache_size = 32, 982 .cache_level = 1, 983 .cache_line_size = 64, 984 .flags = (CRAT_CACHE_FLAGS_ENABLED | 985 CRAT_CACHE_FLAGS_INST_CACHE | 986 CRAT_CACHE_FLAGS_SIMD_CACHE), 987 .num_cu_shared = 2, 988 }, 989 { 990 /* Scalar L1 Data Cache per SQC */ 991 .cache_size = 16, 992 .cache_level = 1, 993 .cache_line_size = 64, 994 .flags = (CRAT_CACHE_FLAGS_ENABLED | 995 CRAT_CACHE_FLAGS_DATA_CACHE | 996 CRAT_CACHE_FLAGS_SIMD_CACHE), 997 .num_cu_shared = 2, 998 }, 999 { 1000 /* GL1 Data Cache per SA */ 1001 .cache_size = 128, 1002 .cache_level = 1, 1003 .cache_line_size = 64, 1004 .flags = (CRAT_CACHE_FLAGS_ENABLED | 1005 CRAT_CACHE_FLAGS_DATA_CACHE | 1006 CRAT_CACHE_FLAGS_SIMD_CACHE), 1007 .num_cu_shared = 6, 1008 }, 1009 { 1010 /* L2 Data Cache per GPU (Total Tex Cache) */ 1011 .cache_size = 2048, 1012 .cache_level = 2, 1013 .cache_line_size = 64, 1014 .flags = (CRAT_CACHE_FLAGS_ENABLED | 1015 CRAT_CACHE_FLAGS_DATA_CACHE | 1016 CRAT_CACHE_FLAGS_SIMD_CACHE), 1017 .num_cu_shared = 6, 1018 }, 1019 }; 1020 1021 static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev, 1022 struct crat_subtype_computeunit *cu) 1023 { 1024 dev->node_props.cpu_cores_count = cu->num_cpu_cores; 1025 dev->node_props.cpu_core_id_base = cu->processor_id_low; 1026 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT) 1027 dev->node_props.capability |= HSA_CAP_ATS_PRESENT; 1028 1029 pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores, 1030 cu->processor_id_low); 1031 } 1032 1033 static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev, 1034 struct crat_subtype_computeunit *cu) 1035 { 1036 dev->node_props.simd_id_base = cu->processor_id_low; 1037 dev->node_props.simd_count = cu->num_simd_cores; 1038 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb; 1039 dev->node_props.max_waves_per_simd = cu->max_waves_simd; 1040 dev->node_props.wave_front_size = cu->wave_front_size; 1041 dev->node_props.array_count = cu->array_count; 1042 dev->node_props.cu_per_simd_array = cu->num_cu_per_array; 1043 dev->node_props.simd_per_cu = cu->num_simd_per_cu; 1044 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu; 1045 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE) 1046 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE; 1047 pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low); 1048 } 1049 1050 /* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct 1051 * topology device present in the device_list 1052 */ 1053 static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu, 1054 struct list_head *device_list) 1055 { 1056 struct kfd_topology_device *dev; 1057 1058 pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n", 1059 cu->proximity_domain, cu->hsa_capability); 1060 list_for_each_entry(dev, device_list, list) { 1061 if (cu->proximity_domain == dev->proximity_domain) { 1062 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT) 1063 kfd_populated_cu_info_cpu(dev, cu); 1064 1065 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT) 1066 kfd_populated_cu_info_gpu(dev, cu); 1067 break; 1068 } 1069 } 1070 1071 return 0; 1072 } 1073 1074 static struct kfd_mem_properties * 1075 find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width, 1076 struct kfd_topology_device *dev) 1077 { 1078 struct kfd_mem_properties *props; 1079 1080 list_for_each_entry(props, &dev->mem_props, list) { 1081 if (props->heap_type == heap_type 1082 && props->flags == flags 1083 && props->width == width) 1084 return props; 1085 } 1086 1087 return NULL; 1088 } 1089 /* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct 1090 * topology device present in the device_list 1091 */ 1092 static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem, 1093 struct list_head *device_list) 1094 { 1095 struct kfd_mem_properties *props; 1096 struct kfd_topology_device *dev; 1097 uint32_t heap_type; 1098 uint64_t size_in_bytes; 1099 uint32_t flags = 0; 1100 uint32_t width; 1101 1102 pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n", 1103 mem->proximity_domain); 1104 list_for_each_entry(dev, device_list, list) { 1105 if (mem->proximity_domain == dev->proximity_domain) { 1106 /* We're on GPU node */ 1107 if (dev->node_props.cpu_cores_count == 0) { 1108 /* APU */ 1109 if (mem->visibility_type == 0) 1110 heap_type = 1111 HSA_MEM_HEAP_TYPE_FB_PRIVATE; 1112 /* dGPU */ 1113 else 1114 heap_type = mem->visibility_type; 1115 } else 1116 heap_type = HSA_MEM_HEAP_TYPE_SYSTEM; 1117 1118 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE) 1119 flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE; 1120 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE) 1121 flags |= HSA_MEM_FLAGS_NON_VOLATILE; 1122 1123 size_in_bytes = 1124 ((uint64_t)mem->length_high << 32) + 1125 mem->length_low; 1126 width = mem->width; 1127 1128 /* Multiple banks of the same type are aggregated into 1129 * one. User mode doesn't care about multiple physical 1130 * memory segments. It's managed as a single virtual 1131 * heap for user mode. 1132 */ 1133 props = find_subtype_mem(heap_type, flags, width, dev); 1134 if (props) { 1135 props->size_in_bytes += size_in_bytes; 1136 break; 1137 } 1138 1139 props = kfd_alloc_struct(props); 1140 if (!props) 1141 return -ENOMEM; 1142 1143 props->heap_type = heap_type; 1144 props->flags = flags; 1145 props->size_in_bytes = size_in_bytes; 1146 props->width = width; 1147 1148 dev->node_props.mem_banks_count++; 1149 list_add_tail(&props->list, &dev->mem_props); 1150 1151 break; 1152 } 1153 } 1154 1155 return 0; 1156 } 1157 1158 /* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct 1159 * topology device present in the device_list 1160 */ 1161 static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache, 1162 struct list_head *device_list) 1163 { 1164 struct kfd_cache_properties *props; 1165 struct kfd_topology_device *dev; 1166 uint32_t id; 1167 uint32_t total_num_of_cu; 1168 1169 id = cache->processor_id_low; 1170 1171 pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id); 1172 list_for_each_entry(dev, device_list, list) { 1173 total_num_of_cu = (dev->node_props.array_count * 1174 dev->node_props.cu_per_simd_array); 1175 1176 /* Cache infomration in CRAT doesn't have proximity_domain 1177 * information as it is associated with a CPU core or GPU 1178 * Compute Unit. So map the cache using CPU core Id or SIMD 1179 * (GPU) ID. 1180 * TODO: This works because currently we can safely assume that 1181 * Compute Units are parsed before caches are parsed. In 1182 * future, remove this dependency 1183 */ 1184 if ((id >= dev->node_props.cpu_core_id_base && 1185 id <= dev->node_props.cpu_core_id_base + 1186 dev->node_props.cpu_cores_count) || 1187 (id >= dev->node_props.simd_id_base && 1188 id < dev->node_props.simd_id_base + 1189 total_num_of_cu)) { 1190 props = kfd_alloc_struct(props); 1191 if (!props) 1192 return -ENOMEM; 1193 1194 props->processor_id_low = id; 1195 props->cache_level = cache->cache_level; 1196 props->cache_size = cache->cache_size; 1197 props->cacheline_size = cache->cache_line_size; 1198 props->cachelines_per_tag = cache->lines_per_tag; 1199 props->cache_assoc = cache->associativity; 1200 props->cache_latency = cache->cache_latency; 1201 1202 memcpy(props->sibling_map, cache->sibling_map, 1203 CRAT_SIBLINGMAP_SIZE); 1204 1205 /* set the sibling_map_size as 32 for CRAT from ACPI */ 1206 props->sibling_map_size = CRAT_SIBLINGMAP_SIZE; 1207 1208 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE) 1209 props->cache_type |= HSA_CACHE_TYPE_DATA; 1210 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE) 1211 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION; 1212 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE) 1213 props->cache_type |= HSA_CACHE_TYPE_CPU; 1214 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE) 1215 props->cache_type |= HSA_CACHE_TYPE_HSACU; 1216 1217 dev->node_props.caches_count++; 1218 list_add_tail(&props->list, &dev->cache_props); 1219 1220 break; 1221 } 1222 } 1223 1224 return 0; 1225 } 1226 1227 /* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct 1228 * topology device present in the device_list 1229 */ 1230 static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink, 1231 struct list_head *device_list) 1232 { 1233 struct kfd_iolink_properties *props = NULL, *props2; 1234 struct kfd_topology_device *dev, *to_dev; 1235 uint32_t id_from; 1236 uint32_t id_to; 1237 1238 id_from = iolink->proximity_domain_from; 1239 id_to = iolink->proximity_domain_to; 1240 1241 pr_debug("Found IO link entry in CRAT table with id_from=%d, id_to %d\n", 1242 id_from, id_to); 1243 list_for_each_entry(dev, device_list, list) { 1244 if (id_from == dev->proximity_domain) { 1245 props = kfd_alloc_struct(props); 1246 if (!props) 1247 return -ENOMEM; 1248 1249 props->node_from = id_from; 1250 props->node_to = id_to; 1251 props->ver_maj = iolink->version_major; 1252 props->ver_min = iolink->version_minor; 1253 props->iolink_type = iolink->io_interface_type; 1254 1255 if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) 1256 props->weight = 20; 1257 else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI) 1258 props->weight = iolink->weight_xgmi; 1259 else 1260 props->weight = node_distance(id_from, id_to); 1261 1262 props->min_latency = iolink->minimum_latency; 1263 props->max_latency = iolink->maximum_latency; 1264 props->min_bandwidth = iolink->minimum_bandwidth_mbs; 1265 props->max_bandwidth = iolink->maximum_bandwidth_mbs; 1266 props->rec_transfer_size = 1267 iolink->recommended_transfer_size; 1268 1269 dev->node_props.io_links_count++; 1270 list_add_tail(&props->list, &dev->io_link_props); 1271 break; 1272 } 1273 } 1274 1275 /* CPU topology is created before GPUs are detected, so CPU->GPU 1276 * links are not built at that time. If a PCIe type is discovered, it 1277 * means a GPU is detected and we are adding GPU->CPU to the topology. 1278 * At this time, also add the corresponded CPU->GPU link if GPU 1279 * is large bar. 1280 * For xGMI, we only added the link with one direction in the crat 1281 * table, add corresponded reversed direction link now. 1282 */ 1283 if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) { 1284 to_dev = kfd_topology_device_by_proximity_domain_no_lock(id_to); 1285 if (!to_dev) 1286 return -ENODEV; 1287 /* same everything but the other direction */ 1288 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL); 1289 if (!props2) 1290 return -ENOMEM; 1291 1292 props2->node_from = id_to; 1293 props2->node_to = id_from; 1294 props2->kobj = NULL; 1295 to_dev->node_props.io_links_count++; 1296 list_add_tail(&props2->list, &to_dev->io_link_props); 1297 } 1298 1299 return 0; 1300 } 1301 1302 /* kfd_parse_subtype - parse subtypes and attach it to correct topology device 1303 * present in the device_list 1304 * @sub_type_hdr - subtype section of crat_image 1305 * @device_list - list of topology devices present in this crat_image 1306 */ 1307 static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr, 1308 struct list_head *device_list) 1309 { 1310 struct crat_subtype_computeunit *cu; 1311 struct crat_subtype_memory *mem; 1312 struct crat_subtype_cache *cache; 1313 struct crat_subtype_iolink *iolink; 1314 int ret = 0; 1315 1316 switch (sub_type_hdr->type) { 1317 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY: 1318 cu = (struct crat_subtype_computeunit *)sub_type_hdr; 1319 ret = kfd_parse_subtype_cu(cu, device_list); 1320 break; 1321 case CRAT_SUBTYPE_MEMORY_AFFINITY: 1322 mem = (struct crat_subtype_memory *)sub_type_hdr; 1323 ret = kfd_parse_subtype_mem(mem, device_list); 1324 break; 1325 case CRAT_SUBTYPE_CACHE_AFFINITY: 1326 cache = (struct crat_subtype_cache *)sub_type_hdr; 1327 ret = kfd_parse_subtype_cache(cache, device_list); 1328 break; 1329 case CRAT_SUBTYPE_TLB_AFFINITY: 1330 /* 1331 * For now, nothing to do here 1332 */ 1333 pr_debug("Found TLB entry in CRAT table (not processing)\n"); 1334 break; 1335 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY: 1336 /* 1337 * For now, nothing to do here 1338 */ 1339 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n"); 1340 break; 1341 case CRAT_SUBTYPE_IOLINK_AFFINITY: 1342 iolink = (struct crat_subtype_iolink *)sub_type_hdr; 1343 ret = kfd_parse_subtype_iolink(iolink, device_list); 1344 break; 1345 default: 1346 pr_warn("Unknown subtype %d in CRAT\n", 1347 sub_type_hdr->type); 1348 } 1349 1350 return ret; 1351 } 1352 1353 /* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT 1354 * create a kfd_topology_device and add in to device_list. Also parse 1355 * CRAT subtypes and attach it to appropriate kfd_topology_device 1356 * @crat_image - input image containing CRAT 1357 * @device_list - [OUT] list of kfd_topology_device generated after 1358 * parsing crat_image 1359 * @proximity_domain - Proximity domain of the first device in the table 1360 * 1361 * Return - 0 if successful else -ve value 1362 */ 1363 int kfd_parse_crat_table(void *crat_image, struct list_head *device_list, 1364 uint32_t proximity_domain) 1365 { 1366 struct kfd_topology_device *top_dev = NULL; 1367 struct crat_subtype_generic *sub_type_hdr; 1368 uint16_t node_id; 1369 int ret = 0; 1370 struct crat_header *crat_table = (struct crat_header *)crat_image; 1371 uint16_t num_nodes; 1372 uint32_t image_len; 1373 1374 if (!crat_image) 1375 return -EINVAL; 1376 1377 if (!list_empty(device_list)) { 1378 pr_warn("Error device list should be empty\n"); 1379 return -EINVAL; 1380 } 1381 1382 num_nodes = crat_table->num_domains; 1383 image_len = crat_table->length; 1384 1385 pr_debug("Parsing CRAT table with %d nodes\n", num_nodes); 1386 1387 for (node_id = 0; node_id < num_nodes; node_id++) { 1388 top_dev = kfd_create_topology_device(device_list); 1389 if (!top_dev) 1390 break; 1391 top_dev->proximity_domain = proximity_domain++; 1392 } 1393 1394 if (!top_dev) { 1395 ret = -ENOMEM; 1396 goto err; 1397 } 1398 1399 memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH); 1400 memcpy(top_dev->oem_table_id, crat_table->oem_table_id, 1401 CRAT_OEMTABLEID_LENGTH); 1402 top_dev->oem_revision = crat_table->oem_revision; 1403 1404 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1); 1405 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) < 1406 ((char *)crat_image) + image_len) { 1407 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) { 1408 ret = kfd_parse_subtype(sub_type_hdr, device_list); 1409 if (ret) 1410 break; 1411 } 1412 1413 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1414 sub_type_hdr->length); 1415 } 1416 1417 err: 1418 if (ret) 1419 kfd_release_topology_device_list(device_list); 1420 1421 return ret; 1422 } 1423 1424 1425 static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev, 1426 struct kfd_gpu_cache_info *pcache_info) 1427 { 1428 struct amdgpu_device *adev = kdev->adev; 1429 int i = 0; 1430 1431 /* TCP L1 Cache per CU */ 1432 if (adev->gfx.config.gc_tcp_l1_size) { 1433 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_l1_size; 1434 pcache_info[i].cache_level = 1; 1435 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1436 CRAT_CACHE_FLAGS_DATA_CACHE | 1437 CRAT_CACHE_FLAGS_SIMD_CACHE); 1438 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_tcp_per_wpg / 2; 1439 pcache_info[i].cache_line_size = adev->gfx.config.gc_tcp_cache_line_size; 1440 i++; 1441 } 1442 /* Scalar L1 Instruction Cache per SQC */ 1443 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) { 1444 pcache_info[i].cache_size = 1445 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc; 1446 pcache_info[i].cache_level = 1; 1447 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1448 CRAT_CACHE_FLAGS_INST_CACHE | 1449 CRAT_CACHE_FLAGS_SIMD_CACHE); 1450 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2; 1451 pcache_info[i].cache_line_size = adev->gfx.config.gc_instruction_cache_line_size; 1452 i++; 1453 } 1454 /* Scalar L1 Data Cache per SQC */ 1455 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) { 1456 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc; 1457 pcache_info[i].cache_level = 1; 1458 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1459 CRAT_CACHE_FLAGS_DATA_CACHE | 1460 CRAT_CACHE_FLAGS_SIMD_CACHE); 1461 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2; 1462 pcache_info[i].cache_line_size = adev->gfx.config.gc_scalar_data_cache_line_size; 1463 i++; 1464 } 1465 /* GL1 Data Cache per SA */ 1466 if (adev->gfx.config.gc_gl1c_per_sa && 1467 adev->gfx.config.gc_gl1c_size_per_instance) { 1468 pcache_info[i].cache_size = adev->gfx.config.gc_gl1c_per_sa * 1469 adev->gfx.config.gc_gl1c_size_per_instance; 1470 pcache_info[i].cache_level = 1; 1471 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1472 CRAT_CACHE_FLAGS_DATA_CACHE | 1473 CRAT_CACHE_FLAGS_SIMD_CACHE); 1474 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh; 1475 pcache_info[i].cache_line_size = 0; 1476 i++; 1477 } 1478 /* L2 Data Cache per GPU (Total Tex Cache) */ 1479 if (adev->gfx.config.gc_gl2c_per_gpu) { 1480 pcache_info[i].cache_size = adev->gfx.config.gc_gl2c_per_gpu; 1481 pcache_info[i].cache_level = 2; 1482 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1483 CRAT_CACHE_FLAGS_DATA_CACHE | 1484 CRAT_CACHE_FLAGS_SIMD_CACHE); 1485 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh; 1486 pcache_info[i].cache_line_size = adev->gfx.config.gc_tcc_cache_line_size; 1487 i++; 1488 } 1489 /* L3 Data Cache per GPU */ 1490 if (adev->gmc.mall_size) { 1491 pcache_info[i].cache_size = adev->gmc.mall_size / 1024; 1492 pcache_info[i].cache_level = 3; 1493 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1494 CRAT_CACHE_FLAGS_DATA_CACHE | 1495 CRAT_CACHE_FLAGS_SIMD_CACHE); 1496 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh; 1497 pcache_info[i].cache_line_size = 0; 1498 i++; 1499 } 1500 return i; 1501 } 1502 1503 static int kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev *kdev, 1504 struct kfd_gpu_cache_info *pcache_info) 1505 { 1506 struct amdgpu_device *adev = kdev->adev; 1507 int i = 0; 1508 1509 /* TCP L1 Cache per CU */ 1510 if (adev->gfx.config.gc_tcp_size_per_cu) { 1511 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_size_per_cu; 1512 pcache_info[i].cache_level = 1; 1513 /* Cacheline size not available in IP discovery for gc943,gc944 */ 1514 pcache_info[i].cache_line_size = 128; 1515 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1516 CRAT_CACHE_FLAGS_DATA_CACHE | 1517 CRAT_CACHE_FLAGS_SIMD_CACHE); 1518 pcache_info[i].num_cu_shared = 1; 1519 i++; 1520 } 1521 /* Scalar L1 Instruction Cache per SQC */ 1522 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) { 1523 pcache_info[i].cache_size = 1524 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc; 1525 pcache_info[i].cache_level = 1; 1526 pcache_info[i].cache_line_size = 64; 1527 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1528 CRAT_CACHE_FLAGS_INST_CACHE | 1529 CRAT_CACHE_FLAGS_SIMD_CACHE); 1530 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_cu_per_sqc; 1531 i++; 1532 } 1533 /* Scalar L1 Data Cache per SQC */ 1534 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) { 1535 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc; 1536 pcache_info[i].cache_level = 1; 1537 pcache_info[i].cache_line_size = 64; 1538 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1539 CRAT_CACHE_FLAGS_DATA_CACHE | 1540 CRAT_CACHE_FLAGS_SIMD_CACHE); 1541 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_cu_per_sqc; 1542 i++; 1543 } 1544 /* L2 Data Cache per GPU (Total Tex Cache) */ 1545 if (adev->gfx.config.gc_tcc_size) { 1546 pcache_info[i].cache_size = adev->gfx.config.gc_tcc_size; 1547 pcache_info[i].cache_level = 2; 1548 pcache_info[i].cache_line_size = 128; 1549 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1550 CRAT_CACHE_FLAGS_DATA_CACHE | 1551 CRAT_CACHE_FLAGS_SIMD_CACHE); 1552 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh; 1553 i++; 1554 } 1555 /* L3 Data Cache per GPU */ 1556 if (adev->gmc.mall_size) { 1557 pcache_info[i].cache_size = adev->gmc.mall_size / 1024; 1558 pcache_info[i].cache_level = 3; 1559 pcache_info[i].cache_line_size = 64; 1560 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | 1561 CRAT_CACHE_FLAGS_DATA_CACHE | 1562 CRAT_CACHE_FLAGS_SIMD_CACHE); 1563 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh; 1564 i++; 1565 } 1566 return i; 1567 } 1568 1569 int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pcache_info) 1570 { 1571 int num_of_cache_types = 0; 1572 1573 switch (kdev->adev->asic_type) { 1574 case CHIP_KAVERI: 1575 *pcache_info = kaveri_cache_info; 1576 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info); 1577 break; 1578 case CHIP_HAWAII: 1579 *pcache_info = hawaii_cache_info; 1580 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info); 1581 break; 1582 case CHIP_CARRIZO: 1583 *pcache_info = carrizo_cache_info; 1584 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info); 1585 break; 1586 case CHIP_TONGA: 1587 *pcache_info = tonga_cache_info; 1588 num_of_cache_types = ARRAY_SIZE(tonga_cache_info); 1589 break; 1590 case CHIP_FIJI: 1591 *pcache_info = fiji_cache_info; 1592 num_of_cache_types = ARRAY_SIZE(fiji_cache_info); 1593 break; 1594 case CHIP_POLARIS10: 1595 *pcache_info = polaris10_cache_info; 1596 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info); 1597 break; 1598 case CHIP_POLARIS11: 1599 *pcache_info = polaris11_cache_info; 1600 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info); 1601 break; 1602 case CHIP_POLARIS12: 1603 *pcache_info = polaris12_cache_info; 1604 num_of_cache_types = ARRAY_SIZE(polaris12_cache_info); 1605 break; 1606 case CHIP_VEGAM: 1607 *pcache_info = vegam_cache_info; 1608 num_of_cache_types = ARRAY_SIZE(vegam_cache_info); 1609 break; 1610 default: 1611 switch (KFD_GC_VERSION(kdev)) { 1612 case IP_VERSION(9, 0, 1): 1613 *pcache_info = vega10_cache_info; 1614 num_of_cache_types = ARRAY_SIZE(vega10_cache_info); 1615 break; 1616 case IP_VERSION(9, 2, 1): 1617 *pcache_info = vega12_cache_info; 1618 num_of_cache_types = ARRAY_SIZE(vega12_cache_info); 1619 break; 1620 case IP_VERSION(9, 4, 0): 1621 case IP_VERSION(9, 4, 1): 1622 *pcache_info = vega20_cache_info; 1623 num_of_cache_types = ARRAY_SIZE(vega20_cache_info); 1624 break; 1625 case IP_VERSION(9, 4, 2): 1626 *pcache_info = aldebaran_cache_info; 1627 num_of_cache_types = ARRAY_SIZE(aldebaran_cache_info); 1628 break; 1629 case IP_VERSION(9, 4, 3): 1630 case IP_VERSION(9, 4, 4): 1631 num_of_cache_types = 1632 kfd_fill_gpu_cache_info_from_gfx_config_v2(kdev->kfd, 1633 *pcache_info); 1634 break; 1635 case IP_VERSION(9, 1, 0): 1636 case IP_VERSION(9, 2, 2): 1637 *pcache_info = raven_cache_info; 1638 num_of_cache_types = ARRAY_SIZE(raven_cache_info); 1639 break; 1640 case IP_VERSION(9, 3, 0): 1641 *pcache_info = renoir_cache_info; 1642 num_of_cache_types = ARRAY_SIZE(renoir_cache_info); 1643 break; 1644 case IP_VERSION(10, 1, 10): 1645 case IP_VERSION(10, 1, 2): 1646 case IP_VERSION(10, 1, 3): 1647 case IP_VERSION(10, 1, 4): 1648 *pcache_info = navi10_cache_info; 1649 num_of_cache_types = ARRAY_SIZE(navi10_cache_info); 1650 break; 1651 case IP_VERSION(10, 1, 1): 1652 *pcache_info = navi14_cache_info; 1653 num_of_cache_types = ARRAY_SIZE(navi14_cache_info); 1654 break; 1655 case IP_VERSION(10, 3, 0): 1656 *pcache_info = sienna_cichlid_cache_info; 1657 num_of_cache_types = ARRAY_SIZE(sienna_cichlid_cache_info); 1658 break; 1659 case IP_VERSION(10, 3, 2): 1660 *pcache_info = navy_flounder_cache_info; 1661 num_of_cache_types = ARRAY_SIZE(navy_flounder_cache_info); 1662 break; 1663 case IP_VERSION(10, 3, 4): 1664 *pcache_info = dimgrey_cavefish_cache_info; 1665 num_of_cache_types = ARRAY_SIZE(dimgrey_cavefish_cache_info); 1666 break; 1667 case IP_VERSION(10, 3, 1): 1668 *pcache_info = vangogh_cache_info; 1669 num_of_cache_types = ARRAY_SIZE(vangogh_cache_info); 1670 break; 1671 case IP_VERSION(10, 3, 5): 1672 *pcache_info = beige_goby_cache_info; 1673 num_of_cache_types = ARRAY_SIZE(beige_goby_cache_info); 1674 break; 1675 case IP_VERSION(10, 3, 3): 1676 *pcache_info = yellow_carp_cache_info; 1677 num_of_cache_types = ARRAY_SIZE(yellow_carp_cache_info); 1678 break; 1679 case IP_VERSION(10, 3, 6): 1680 *pcache_info = gc_10_3_6_cache_info; 1681 num_of_cache_types = ARRAY_SIZE(gc_10_3_6_cache_info); 1682 break; 1683 case IP_VERSION(10, 3, 7): 1684 *pcache_info = gfx1037_cache_info; 1685 num_of_cache_types = ARRAY_SIZE(gfx1037_cache_info); 1686 break; 1687 case IP_VERSION(11, 0, 0): 1688 case IP_VERSION(11, 0, 1): 1689 case IP_VERSION(11, 0, 2): 1690 case IP_VERSION(11, 0, 3): 1691 case IP_VERSION(11, 0, 4): 1692 case IP_VERSION(11, 5, 0): 1693 case IP_VERSION(11, 5, 1): 1694 case IP_VERSION(11, 5, 2): 1695 case IP_VERSION(12, 0, 0): 1696 case IP_VERSION(12, 0, 1): 1697 num_of_cache_types = 1698 kfd_fill_gpu_cache_info_from_gfx_config(kdev->kfd, *pcache_info); 1699 break; 1700 default: 1701 *pcache_info = dummy_cache_info; 1702 num_of_cache_types = ARRAY_SIZE(dummy_cache_info); 1703 pr_warn("dummy cache info is used temporarily and real cache info need update later.\n"); 1704 break; 1705 } 1706 } 1707 return num_of_cache_types; 1708 } 1709 1710 /* Memory required to create Virtual CRAT. 1711 * Since there is no easy way to predict the amount of memory required, the 1712 * following amount is allocated for GPU Virtual CRAT. This is 1713 * expected to cover all known conditions. But to be safe additional check 1714 * is put in the code to ensure we don't overwrite. 1715 */ 1716 #define VCRAT_SIZE_FOR_GPU (4 * PAGE_SIZE) 1717 1718 /* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node 1719 * 1720 * @numa_node_id: CPU NUMA node id 1721 * @avail_size: Available size in the memory 1722 * @sub_type_hdr: Memory into which compute info will be filled in 1723 * 1724 * Return 0 if successful else return -ve value 1725 */ 1726 static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size, 1727 int proximity_domain, 1728 struct crat_subtype_computeunit *sub_type_hdr) 1729 { 1730 const struct cpumask *cpumask; 1731 1732 *avail_size -= sizeof(struct crat_subtype_computeunit); 1733 if (*avail_size < 0) 1734 return -ENOMEM; 1735 1736 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit)); 1737 1738 /* Fill in subtype header data */ 1739 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY; 1740 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit); 1741 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 1742 1743 cpumask = cpumask_of_node(numa_node_id); 1744 1745 /* Fill in CU data */ 1746 sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT; 1747 sub_type_hdr->proximity_domain = proximity_domain; 1748 sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id); 1749 if (sub_type_hdr->processor_id_low == -1) 1750 return -EINVAL; 1751 1752 sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask); 1753 1754 return 0; 1755 } 1756 1757 /* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node 1758 * 1759 * @numa_node_id: CPU NUMA node id 1760 * @avail_size: Available size in the memory 1761 * @sub_type_hdr: Memory into which compute info will be filled in 1762 * 1763 * Return 0 if successful else return -ve value 1764 */ 1765 static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size, 1766 int proximity_domain, 1767 struct crat_subtype_memory *sub_type_hdr) 1768 { 1769 uint64_t mem_in_bytes = 0; 1770 pg_data_t *pgdat; 1771 int zone_type; 1772 1773 *avail_size -= sizeof(struct crat_subtype_memory); 1774 if (*avail_size < 0) 1775 return -ENOMEM; 1776 1777 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory)); 1778 1779 /* Fill in subtype header data */ 1780 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY; 1781 sub_type_hdr->length = sizeof(struct crat_subtype_memory); 1782 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 1783 1784 /* Fill in Memory Subunit data */ 1785 1786 /* Unlike si_meminfo, si_meminfo_node is not exported. So 1787 * the following lines are duplicated from si_meminfo_node 1788 * function 1789 */ 1790 pgdat = NODE_DATA(numa_node_id); 1791 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) 1792 mem_in_bytes += zone_managed_pages(&pgdat->node_zones[zone_type]); 1793 mem_in_bytes <<= PAGE_SHIFT; 1794 1795 sub_type_hdr->length_low = lower_32_bits(mem_in_bytes); 1796 sub_type_hdr->length_high = upper_32_bits(mem_in_bytes); 1797 sub_type_hdr->proximity_domain = proximity_domain; 1798 1799 return 0; 1800 } 1801 1802 #ifdef CONFIG_X86_64 1803 static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size, 1804 uint32_t *num_entries, 1805 struct crat_subtype_iolink *sub_type_hdr) 1806 { 1807 int nid; 1808 struct cpuinfo_x86 *c = &cpu_data(0); 1809 uint8_t link_type; 1810 1811 if (c->x86_vendor == X86_VENDOR_AMD) 1812 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT; 1813 else 1814 link_type = CRAT_IOLINK_TYPE_QPI_1_1; 1815 1816 *num_entries = 0; 1817 1818 /* Create IO links from this node to other CPU nodes */ 1819 for_each_online_node(nid) { 1820 if (nid == numa_node_id) /* node itself */ 1821 continue; 1822 1823 *avail_size -= sizeof(struct crat_subtype_iolink); 1824 if (*avail_size < 0) 1825 return -ENOMEM; 1826 1827 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink)); 1828 1829 /* Fill in subtype header data */ 1830 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY; 1831 sub_type_hdr->length = sizeof(struct crat_subtype_iolink); 1832 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 1833 1834 /* Fill in IO link data */ 1835 sub_type_hdr->proximity_domain_from = numa_node_id; 1836 sub_type_hdr->proximity_domain_to = nid; 1837 sub_type_hdr->io_interface_type = link_type; 1838 1839 (*num_entries)++; 1840 sub_type_hdr++; 1841 } 1842 1843 return 0; 1844 } 1845 #endif 1846 1847 /* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU 1848 * 1849 * @pcrat_image: Fill in VCRAT for CPU 1850 * @size: [IN] allocated size of crat_image. 1851 * [OUT] actual size of data filled in crat_image 1852 */ 1853 static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size) 1854 { 1855 struct crat_header *crat_table = (struct crat_header *)pcrat_image; 1856 struct acpi_table_header *acpi_table; 1857 acpi_status status; 1858 struct crat_subtype_generic *sub_type_hdr; 1859 int avail_size = *size; 1860 int numa_node_id; 1861 #ifdef CONFIG_X86_64 1862 uint32_t entries = 0; 1863 #endif 1864 int ret = 0; 1865 1866 if (!pcrat_image) 1867 return -EINVAL; 1868 1869 /* Fill in CRAT Header. 1870 * Modify length and total_entries as subunits are added. 1871 */ 1872 avail_size -= sizeof(struct crat_header); 1873 if (avail_size < 0) 1874 return -ENOMEM; 1875 1876 memset(crat_table, 0, sizeof(struct crat_header)); 1877 memcpy(&crat_table->signature, CRAT_SIGNATURE, 1878 sizeof(crat_table->signature)); 1879 crat_table->length = sizeof(struct crat_header); 1880 1881 status = acpi_get_table("DSDT", 0, &acpi_table); 1882 if (status != AE_OK) 1883 pr_warn("DSDT table not found for OEM information\n"); 1884 else { 1885 crat_table->oem_revision = acpi_table->revision; 1886 memcpy(crat_table->oem_id, acpi_table->oem_id, 1887 CRAT_OEMID_LENGTH); 1888 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id, 1889 CRAT_OEMTABLEID_LENGTH); 1890 acpi_put_table(acpi_table); 1891 } 1892 crat_table->total_entries = 0; 1893 crat_table->num_domains = 0; 1894 1895 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1); 1896 1897 for_each_online_node(numa_node_id) { 1898 if (kfd_numa_node_to_apic_id(numa_node_id) == -1) 1899 continue; 1900 1901 /* Fill in Subtype: Compute Unit */ 1902 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size, 1903 crat_table->num_domains, 1904 (struct crat_subtype_computeunit *)sub_type_hdr); 1905 if (ret < 0) 1906 return ret; 1907 crat_table->length += sub_type_hdr->length; 1908 crat_table->total_entries++; 1909 1910 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1911 sub_type_hdr->length); 1912 1913 /* Fill in Subtype: Memory */ 1914 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size, 1915 crat_table->num_domains, 1916 (struct crat_subtype_memory *)sub_type_hdr); 1917 if (ret < 0) 1918 return ret; 1919 crat_table->length += sub_type_hdr->length; 1920 crat_table->total_entries++; 1921 1922 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1923 sub_type_hdr->length); 1924 1925 /* Fill in Subtype: IO Link */ 1926 #ifdef CONFIG_X86_64 1927 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size, 1928 &entries, 1929 (struct crat_subtype_iolink *)sub_type_hdr); 1930 if (ret < 0) 1931 return ret; 1932 1933 if (entries) { 1934 crat_table->length += (sub_type_hdr->length * entries); 1935 crat_table->total_entries += entries; 1936 1937 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 1938 sub_type_hdr->length * entries); 1939 } 1940 #else 1941 pr_info("IO link not available for non x86 platforms\n"); 1942 #endif 1943 1944 crat_table->num_domains++; 1945 } 1946 1947 /* TODO: Add cache Subtype for CPU. 1948 * Currently, CPU cache information is available in function 1949 * detect_cache_attributes(cpu) defined in the file 1950 * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not 1951 * exported and to get the same information the code needs to be 1952 * duplicated. 1953 */ 1954 1955 *size = crat_table->length; 1956 pr_info("Virtual CRAT table created for CPU\n"); 1957 1958 return 0; 1959 } 1960 1961 static int kfd_fill_gpu_memory_affinity(int *avail_size, 1962 struct kfd_node *kdev, uint8_t type, uint64_t size, 1963 struct crat_subtype_memory *sub_type_hdr, 1964 uint32_t proximity_domain, 1965 const struct kfd_local_mem_info *local_mem_info) 1966 { 1967 *avail_size -= sizeof(struct crat_subtype_memory); 1968 if (*avail_size < 0) 1969 return -ENOMEM; 1970 1971 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory)); 1972 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY; 1973 sub_type_hdr->length = sizeof(struct crat_subtype_memory); 1974 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED; 1975 1976 sub_type_hdr->proximity_domain = proximity_domain; 1977 1978 pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n", 1979 type, size); 1980 1981 sub_type_hdr->length_low = lower_32_bits(size); 1982 sub_type_hdr->length_high = upper_32_bits(size); 1983 1984 sub_type_hdr->width = local_mem_info->vram_width; 1985 sub_type_hdr->visibility_type = type; 1986 1987 return 0; 1988 } 1989 1990 #ifdef CONFIG_ACPI_NUMA 1991 static void kfd_find_numa_node_in_srat(struct kfd_node *kdev) 1992 { 1993 struct acpi_table_header *table_header = NULL; 1994 struct acpi_subtable_header *sub_header = NULL; 1995 unsigned long table_end, subtable_len; 1996 u32 pci_id = pci_domain_nr(kdev->adev->pdev->bus) << 16 | 1997 pci_dev_id(kdev->adev->pdev); 1998 u32 bdf; 1999 acpi_status status; 2000 struct acpi_srat_cpu_affinity *cpu; 2001 struct acpi_srat_generic_affinity *gpu; 2002 int pxm = 0, max_pxm = 0; 2003 int numa_node = NUMA_NO_NODE; 2004 bool found = false; 2005 2006 /* Fetch the SRAT table from ACPI */ 2007 status = acpi_get_table(ACPI_SIG_SRAT, 0, &table_header); 2008 if (status == AE_NOT_FOUND) { 2009 pr_warn("SRAT table not found\n"); 2010 return; 2011 } else if (ACPI_FAILURE(status)) { 2012 const char *err = acpi_format_exception(status); 2013 pr_err("SRAT table error: %s\n", err); 2014 return; 2015 } 2016 2017 table_end = (unsigned long)table_header + table_header->length; 2018 2019 /* Parse all entries looking for a match. */ 2020 sub_header = (struct acpi_subtable_header *) 2021 ((unsigned long)table_header + 2022 sizeof(struct acpi_table_srat)); 2023 subtable_len = sub_header->length; 2024 2025 while (((unsigned long)sub_header) + subtable_len < table_end) { 2026 /* 2027 * If length is 0, break from this loop to avoid 2028 * infinite loop. 2029 */ 2030 if (subtable_len == 0) { 2031 pr_err("SRAT invalid zero length\n"); 2032 break; 2033 } 2034 2035 switch (sub_header->type) { 2036 case ACPI_SRAT_TYPE_CPU_AFFINITY: 2037 cpu = (struct acpi_srat_cpu_affinity *)sub_header; 2038 pxm = *((u32 *)cpu->proximity_domain_hi) << 8 | 2039 cpu->proximity_domain_lo; 2040 if (pxm > max_pxm) 2041 max_pxm = pxm; 2042 break; 2043 case ACPI_SRAT_TYPE_GENERIC_AFFINITY: 2044 gpu = (struct acpi_srat_generic_affinity *)sub_header; 2045 bdf = *((u16 *)(&gpu->device_handle[0])) << 16 | 2046 *((u16 *)(&gpu->device_handle[2])); 2047 if (bdf == pci_id) { 2048 found = true; 2049 numa_node = pxm_to_node(gpu->proximity_domain); 2050 } 2051 break; 2052 default: 2053 break; 2054 } 2055 2056 if (found) 2057 break; 2058 2059 sub_header = (struct acpi_subtable_header *) 2060 ((unsigned long)sub_header + subtable_len); 2061 subtable_len = sub_header->length; 2062 } 2063 2064 acpi_put_table(table_header); 2065 2066 /* Workaround bad cpu-gpu binding case */ 2067 if (found && (numa_node < 0 || 2068 numa_node > pxm_to_node(max_pxm))) 2069 numa_node = 0; 2070 2071 if (numa_node != NUMA_NO_NODE) 2072 set_dev_node(&kdev->adev->pdev->dev, numa_node); 2073 } 2074 #endif 2075 2076 #define KFD_CRAT_INTRA_SOCKET_WEIGHT 13 2077 #define KFD_CRAT_XGMI_WEIGHT 15 2078 2079 /* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU 2080 * to its NUMA node 2081 * @avail_size: Available size in the memory 2082 * @kdev - [IN] GPU device 2083 * @sub_type_hdr: Memory into which io link info will be filled in 2084 * @proximity_domain - proximity domain of the GPU node 2085 * 2086 * Return 0 if successful else return -ve value 2087 */ 2088 static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size, 2089 struct kfd_node *kdev, 2090 struct crat_subtype_iolink *sub_type_hdr, 2091 uint32_t proximity_domain) 2092 { 2093 *avail_size -= sizeof(struct crat_subtype_iolink); 2094 if (*avail_size < 0) 2095 return -ENOMEM; 2096 2097 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink)); 2098 2099 /* Fill in subtype header data */ 2100 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY; 2101 sub_type_hdr->length = sizeof(struct crat_subtype_iolink); 2102 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED; 2103 if (kfd_dev_is_large_bar(kdev)) 2104 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL; 2105 2106 /* Fill in IOLINK subtype. 2107 * TODO: Fill-in other fields of iolink subtype 2108 */ 2109 if (kdev->adev->gmc.xgmi.connected_to_cpu || 2110 (KFD_GC_VERSION(kdev) == IP_VERSION(9, 4, 3) && 2111 kdev->adev->smuio.funcs->get_pkg_type(kdev->adev) == 2112 AMDGPU_PKG_TYPE_APU)) { 2113 bool ext_cpu = KFD_GC_VERSION(kdev) != IP_VERSION(9, 4, 3); 2114 int mem_bw = 819200, weight = ext_cpu ? KFD_CRAT_XGMI_WEIGHT : 2115 KFD_CRAT_INTRA_SOCKET_WEIGHT; 2116 uint32_t bandwidth = ext_cpu ? amdgpu_amdkfd_get_xgmi_bandwidth_mbytes( 2117 kdev->adev, NULL, true) : mem_bw; 2118 2119 /* 2120 * with host gpu xgmi link, host can access gpu memory whether 2121 * or not pcie bar type is large, so always create bidirectional 2122 * io link. 2123 */ 2124 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL; 2125 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI; 2126 sub_type_hdr->weight_xgmi = weight; 2127 sub_type_hdr->minimum_bandwidth_mbs = bandwidth; 2128 sub_type_hdr->maximum_bandwidth_mbs = bandwidth; 2129 } else { 2130 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS; 2131 sub_type_hdr->minimum_bandwidth_mbs = 2132 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, true); 2133 sub_type_hdr->maximum_bandwidth_mbs = 2134 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, false); 2135 } 2136 2137 sub_type_hdr->proximity_domain_from = proximity_domain; 2138 2139 #ifdef CONFIG_ACPI_NUMA 2140 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE && 2141 num_possible_nodes() > 1) 2142 kfd_find_numa_node_in_srat(kdev); 2143 #endif 2144 #ifdef CONFIG_NUMA 2145 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE) 2146 sub_type_hdr->proximity_domain_to = 0; 2147 else 2148 sub_type_hdr->proximity_domain_to = kdev->adev->pdev->dev.numa_node; 2149 #else 2150 sub_type_hdr->proximity_domain_to = 0; 2151 #endif 2152 return 0; 2153 } 2154 2155 static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size, 2156 struct kfd_node *kdev, 2157 struct kfd_node *peer_kdev, 2158 struct crat_subtype_iolink *sub_type_hdr, 2159 uint32_t proximity_domain_from, 2160 uint32_t proximity_domain_to) 2161 { 2162 bool use_ta_info = kdev->kfd->num_nodes == 1; 2163 2164 *avail_size -= sizeof(struct crat_subtype_iolink); 2165 if (*avail_size < 0) 2166 return -ENOMEM; 2167 2168 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink)); 2169 2170 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY; 2171 sub_type_hdr->length = sizeof(struct crat_subtype_iolink); 2172 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED | 2173 CRAT_IOLINK_FLAGS_BI_DIRECTIONAL; 2174 2175 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI; 2176 sub_type_hdr->proximity_domain_from = proximity_domain_from; 2177 sub_type_hdr->proximity_domain_to = proximity_domain_to; 2178 2179 if (use_ta_info) { 2180 sub_type_hdr->weight_xgmi = KFD_CRAT_XGMI_WEIGHT * 2181 amdgpu_amdkfd_get_xgmi_hops_count(kdev->adev, peer_kdev->adev); 2182 sub_type_hdr->maximum_bandwidth_mbs = 2183 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, 2184 peer_kdev->adev, false); 2185 sub_type_hdr->minimum_bandwidth_mbs = sub_type_hdr->maximum_bandwidth_mbs ? 2186 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, NULL, true) : 0; 2187 } else { 2188 bool is_single_hop = kdev->kfd == peer_kdev->kfd; 2189 int weight = is_single_hop ? KFD_CRAT_INTRA_SOCKET_WEIGHT : 2190 (2 * KFD_CRAT_INTRA_SOCKET_WEIGHT) + KFD_CRAT_XGMI_WEIGHT; 2191 int mem_bw = 819200; 2192 2193 sub_type_hdr->weight_xgmi = weight; 2194 sub_type_hdr->maximum_bandwidth_mbs = is_single_hop ? mem_bw : 0; 2195 sub_type_hdr->minimum_bandwidth_mbs = is_single_hop ? mem_bw : 0; 2196 } 2197 2198 return 0; 2199 } 2200 2201 /* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU 2202 * 2203 * @pcrat_image: Fill in VCRAT for GPU 2204 * @size: [IN] allocated size of crat_image. 2205 * [OUT] actual size of data filled in crat_image 2206 */ 2207 static int kfd_create_vcrat_image_gpu(void *pcrat_image, 2208 size_t *size, struct kfd_node *kdev, 2209 uint32_t proximity_domain) 2210 { 2211 struct crat_header *crat_table = (struct crat_header *)pcrat_image; 2212 struct amdgpu_gfx_config *gfx_info = &kdev->adev->gfx.config; 2213 struct amdgpu_cu_info *cu_info = &kdev->adev->gfx.cu_info; 2214 struct crat_subtype_generic *sub_type_hdr; 2215 struct kfd_local_mem_info local_mem_info; 2216 struct kfd_topology_device *peer_dev; 2217 struct crat_subtype_computeunit *cu; 2218 int avail_size = *size; 2219 uint32_t total_num_of_cu; 2220 uint32_t nid = 0; 2221 int ret = 0; 2222 2223 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU) 2224 return -EINVAL; 2225 2226 /* Fill the CRAT Header. 2227 * Modify length and total_entries as subunits are added. 2228 */ 2229 avail_size -= sizeof(struct crat_header); 2230 memset(crat_table, 0, sizeof(struct crat_header)); 2231 2232 memcpy(&crat_table->signature, CRAT_SIGNATURE, 2233 sizeof(crat_table->signature)); 2234 /* Change length as we add more subtypes*/ 2235 crat_table->length = sizeof(struct crat_header); 2236 crat_table->num_domains = 1; 2237 crat_table->total_entries = 0; 2238 2239 /* Fill in Subtype: Compute Unit 2240 * First fill in the sub type header and then sub type data 2241 */ 2242 avail_size -= sizeof(struct crat_subtype_computeunit); 2243 sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1); 2244 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit)); 2245 2246 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY; 2247 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit); 2248 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; 2249 2250 /* Fill CU subtype data */ 2251 cu = (struct crat_subtype_computeunit *)sub_type_hdr; 2252 cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT; 2253 cu->proximity_domain = proximity_domain; 2254 2255 cu->num_simd_per_cu = cu_info->simd_per_cu; 2256 cu->num_simd_cores = cu_info->simd_per_cu * 2257 (cu_info->number / kdev->kfd->num_nodes); 2258 cu->max_waves_simd = cu_info->max_waves_per_simd; 2259 2260 cu->wave_front_size = cu_info->wave_front_size; 2261 cu->array_count = gfx_info->max_sh_per_se * 2262 gfx_info->max_shader_engines; 2263 total_num_of_cu = (cu->array_count * gfx_info->max_cu_per_sh); 2264 cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu); 2265 cu->num_cu_per_array = gfx_info->max_cu_per_sh; 2266 cu->max_slots_scatch_cu = cu_info->max_scratch_slots_per_cu; 2267 cu->num_banks = gfx_info->max_shader_engines; 2268 cu->lds_size_in_kb = cu_info->lds_size; 2269 2270 cu->hsa_capability = 0; 2271 2272 crat_table->length += sub_type_hdr->length; 2273 crat_table->total_entries++; 2274 2275 /* Fill in Subtype: Memory. Only on systems with large BAR (no 2276 * private FB), report memory as public. On other systems 2277 * report the total FB size (public+private) as a single 2278 * private heap. 2279 */ 2280 local_mem_info = kdev->local_mem_info; 2281 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 2282 sub_type_hdr->length); 2283 2284 if (kdev->adev->debug_largebar) 2285 local_mem_info.local_mem_size_private = 0; 2286 2287 if (local_mem_info.local_mem_size_private == 0) 2288 ret = kfd_fill_gpu_memory_affinity(&avail_size, 2289 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC, 2290 local_mem_info.local_mem_size_public, 2291 (struct crat_subtype_memory *)sub_type_hdr, 2292 proximity_domain, 2293 &local_mem_info); 2294 else 2295 ret = kfd_fill_gpu_memory_affinity(&avail_size, 2296 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE, 2297 local_mem_info.local_mem_size_public + 2298 local_mem_info.local_mem_size_private, 2299 (struct crat_subtype_memory *)sub_type_hdr, 2300 proximity_domain, 2301 &local_mem_info); 2302 if (ret < 0) 2303 return ret; 2304 2305 crat_table->length += sizeof(struct crat_subtype_memory); 2306 crat_table->total_entries++; 2307 2308 /* Fill in Subtype: IO_LINKS 2309 * Only direct links are added here which is Link from GPU to 2310 * its NUMA node. Indirect links are added by userspace. 2311 */ 2312 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + 2313 sub_type_hdr->length); 2314 ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev, 2315 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain); 2316 2317 if (ret < 0) 2318 return ret; 2319 2320 crat_table->length += sub_type_hdr->length; 2321 crat_table->total_entries++; 2322 2323 2324 /* Fill in Subtype: IO_LINKS 2325 * Direct links from GPU to other GPUs through xGMI. 2326 * We will loop GPUs that already be processed (with lower value 2327 * of proximity_domain), add the link for the GPUs with same 2328 * hive id (from this GPU to other GPU) . The reversed iolink 2329 * (from other GPU to this GPU) will be added 2330 * in kfd_parse_subtype_iolink. 2331 */ 2332 if (kdev->kfd->hive_id) { 2333 for (nid = 0; nid < proximity_domain; ++nid) { 2334 peer_dev = kfd_topology_device_by_proximity_domain_no_lock(nid); 2335 if (!peer_dev->gpu) 2336 continue; 2337 if (peer_dev->gpu->kfd->hive_id != kdev->kfd->hive_id) 2338 continue; 2339 if (!amdgpu_xgmi_get_is_sharing_enabled(kdev->adev, peer_dev->gpu->adev)) 2340 continue; 2341 sub_type_hdr = (typeof(sub_type_hdr))( 2342 (char *)sub_type_hdr + 2343 sizeof(struct crat_subtype_iolink)); 2344 ret = kfd_fill_gpu_xgmi_link_to_gpu( 2345 &avail_size, kdev, peer_dev->gpu, 2346 (struct crat_subtype_iolink *)sub_type_hdr, 2347 proximity_domain, nid); 2348 if (ret < 0) 2349 return ret; 2350 crat_table->length += sub_type_hdr->length; 2351 crat_table->total_entries++; 2352 } 2353 } 2354 *size = crat_table->length; 2355 pr_info("Virtual CRAT table created for GPU\n"); 2356 2357 return ret; 2358 } 2359 2360 /* kfd_create_crat_image_virtual - Allocates memory for CRAT image and 2361 * creates a Virtual CRAT (VCRAT) image 2362 * 2363 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory 2364 * 2365 * @crat_image: VCRAT image created because ACPI does not have a 2366 * CRAT for this device 2367 * @size: [OUT] size of virtual crat_image 2368 * @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device 2369 * COMPUTE_UNIT_GPU - Create VCRAT for GPU 2370 * (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU 2371 * -- this option is not currently implemented. 2372 * The assumption is that all AMD APUs will have CRAT 2373 * @kdev: Valid kfd_node required if flags contain COMPUTE_UNIT_GPU 2374 * 2375 * Return 0 if successful else return -ve value 2376 */ 2377 int kfd_create_crat_image_virtual(void **crat_image, size_t *size, 2378 int flags, struct kfd_node *kdev, 2379 uint32_t proximity_domain) 2380 { 2381 void *pcrat_image = NULL; 2382 int ret = 0, num_nodes; 2383 size_t dyn_size; 2384 2385 if (!crat_image) 2386 return -EINVAL; 2387 2388 *crat_image = NULL; 2389 2390 /* Allocate the CPU Virtual CRAT size based on the number of online 2391 * nodes. Allocate VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image. 2392 * This should cover all the current conditions. A check is put not 2393 * to overwrite beyond allocated size for GPUs 2394 */ 2395 switch (flags) { 2396 case COMPUTE_UNIT_CPU: 2397 num_nodes = num_online_nodes(); 2398 dyn_size = sizeof(struct crat_header) + 2399 num_nodes * (sizeof(struct crat_subtype_computeunit) + 2400 sizeof(struct crat_subtype_memory) + 2401 (num_nodes - 1) * sizeof(struct crat_subtype_iolink)); 2402 pcrat_image = kvmalloc(dyn_size, GFP_KERNEL); 2403 if (!pcrat_image) 2404 return -ENOMEM; 2405 *size = dyn_size; 2406 pr_debug("CRAT size is %ld", dyn_size); 2407 ret = kfd_create_vcrat_image_cpu(pcrat_image, size); 2408 break; 2409 case COMPUTE_UNIT_GPU: 2410 if (!kdev) 2411 return -EINVAL; 2412 pcrat_image = kvmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL); 2413 if (!pcrat_image) 2414 return -ENOMEM; 2415 *size = VCRAT_SIZE_FOR_GPU; 2416 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev, 2417 proximity_domain); 2418 break; 2419 case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU): 2420 /* TODO: */ 2421 ret = -EINVAL; 2422 pr_err("VCRAT not implemented for APU\n"); 2423 break; 2424 default: 2425 ret = -EINVAL; 2426 } 2427 2428 if (!ret) 2429 *crat_image = pcrat_image; 2430 else 2431 kvfree(pcrat_image); 2432 2433 return ret; 2434 } 2435 2436 2437 /* kfd_destroy_crat_image 2438 * 2439 * @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..) 2440 * 2441 */ 2442 void kfd_destroy_crat_image(void *crat_image) 2443 { 2444 kvfree(crat_image); 2445 } 2446