1 // SPDX-License-Identifier: GPL-2.0 OR MIT 2 /* 3 * Copyright 2014-2022 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/bsearch.h> 25 #include <linux/pci.h> 26 #include <linux/slab.h> 27 #include "kfd_priv.h" 28 #include "kfd_device_queue_manager.h" 29 #include "kfd_pm4_headers_vi.h" 30 #include "kfd_pm4_headers_aldebaran.h" 31 #include "cwsr_trap_handler.h" 32 #include "amdgpu_amdkfd.h" 33 #include "kfd_smi_events.h" 34 #include "kfd_svm.h" 35 #include "kfd_migrate.h" 36 #include "amdgpu.h" 37 #include "amdgpu_ip.h" 38 #include "amdgpu_xcp.h" 39 40 #define MQD_SIZE_ALIGNED 768 41 42 /* 43 * kfd_locked is used to lock the kfd driver during suspend or reset 44 * once locked, kfd driver will stop any further GPU execution. 45 * create process (open) will return -EAGAIN. 46 */ 47 static int kfd_locked; 48 49 #ifdef CONFIG_DRM_AMDGPU_CIK 50 extern const struct kfd2kgd_calls gfx_v7_kfd2kgd; 51 #endif 52 extern const struct kfd2kgd_calls gfx_v8_kfd2kgd; 53 extern const struct kfd2kgd_calls gfx_v9_kfd2kgd; 54 extern const struct kfd2kgd_calls arcturus_kfd2kgd; 55 extern const struct kfd2kgd_calls aldebaran_kfd2kgd; 56 extern const struct kfd2kgd_calls gc_9_4_3_kfd2kgd; 57 extern const struct kfd2kgd_calls gfx_v10_kfd2kgd; 58 extern const struct kfd2kgd_calls gfx_v10_3_kfd2kgd; 59 extern const struct kfd2kgd_calls gfx_v11_kfd2kgd; 60 extern const struct kfd2kgd_calls gfx_v12_kfd2kgd; 61 extern const struct kfd2kgd_calls gfx_v12_1_kfd2kgd; 62 63 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size, 64 unsigned int chunk_size); 65 static void kfd_gtt_sa_fini(struct kfd_dev *kfd); 66 67 static int kfd_resume(struct kfd_node *kfd); 68 69 static void kfd_device_info_set_sdma_info(struct kfd_dev *kfd) 70 { 71 uint32_t sdma_version = amdgpu_ip_version(kfd->adev, SDMA0_HWIP, 0); 72 73 switch (sdma_version) { 74 case IP_VERSION(4, 0, 0):/* VEGA10 */ 75 case IP_VERSION(4, 0, 1):/* VEGA12 */ 76 case IP_VERSION(4, 1, 0):/* RAVEN */ 77 case IP_VERSION(4, 1, 1):/* RAVEN */ 78 case IP_VERSION(4, 1, 2):/* RENOIR */ 79 case IP_VERSION(5, 2, 1):/* VANGOGH */ 80 case IP_VERSION(5, 2, 3):/* YELLOW_CARP */ 81 case IP_VERSION(5, 2, 6):/* GC 10.3.6 */ 82 case IP_VERSION(5, 2, 7):/* GC 10.3.7 */ 83 kfd->device_info.num_sdma_queues_per_engine = 2; 84 break; 85 case IP_VERSION(4, 2, 0):/* VEGA20 */ 86 case IP_VERSION(4, 2, 2):/* ARCTURUS */ 87 case IP_VERSION(4, 4, 0):/* ALDEBARAN */ 88 case IP_VERSION(4, 4, 2): 89 case IP_VERSION(4, 4, 5): 90 case IP_VERSION(4, 4, 4): 91 case IP_VERSION(5, 0, 0):/* NAVI10 */ 92 case IP_VERSION(5, 0, 1):/* CYAN_SKILLFISH */ 93 case IP_VERSION(5, 0, 2):/* NAVI14 */ 94 case IP_VERSION(5, 0, 5):/* NAVI12 */ 95 case IP_VERSION(5, 2, 0):/* SIENNA_CICHLID */ 96 case IP_VERSION(5, 2, 2):/* NAVY_FLOUNDER */ 97 case IP_VERSION(5, 2, 4):/* DIMGREY_CAVEFISH */ 98 case IP_VERSION(5, 2, 5):/* BEIGE_GOBY */ 99 kfd->device_info.num_sdma_queues_per_engine = 8; 100 break; 101 case IP_VERSION(6, 0, 0): 102 case IP_VERSION(6, 0, 1): 103 case IP_VERSION(6, 0, 2): 104 case IP_VERSION(6, 0, 3): 105 case IP_VERSION(6, 1, 0): 106 case IP_VERSION(6, 1, 1): 107 case IP_VERSION(6, 1, 2): 108 case IP_VERSION(6, 1, 3): 109 case IP_VERSION(6, 1, 4): 110 case IP_VERSION(6, 4, 0): 111 case IP_VERSION(7, 0, 0): 112 case IP_VERSION(7, 0, 1): 113 case IP_VERSION(7, 1, 0): 114 kfd->device_info.num_sdma_queues_per_engine = 8; 115 /* Reserve 1 for paging and 1 for gfx */ 116 kfd->device_info.num_reserved_sdma_queues_per_engine = 2; 117 break; 118 default: 119 dev_warn(kfd_device, 120 "Default sdma queue per engine(8) is set due to mismatch of sdma ip block(SDMA_HWIP:0x%x).\n", 121 sdma_version); 122 kfd->device_info.num_sdma_queues_per_engine = 8; 123 } 124 } 125 126 static void kfd_device_info_set_event_interrupt_class(struct kfd_dev *kfd) 127 { 128 uint32_t gc_version = KFD_GC_VERSION(kfd); 129 130 switch (gc_version) { 131 case IP_VERSION(9, 0, 1): /* VEGA10 */ 132 case IP_VERSION(9, 1, 0): /* RAVEN */ 133 case IP_VERSION(9, 2, 1): /* VEGA12 */ 134 case IP_VERSION(9, 2, 2): /* RAVEN */ 135 case IP_VERSION(9, 3, 0): /* RENOIR */ 136 case IP_VERSION(9, 4, 0): /* VEGA20 */ 137 case IP_VERSION(9, 4, 1): /* ARCTURUS */ 138 case IP_VERSION(9, 4, 2): /* ALDEBARAN */ 139 kfd->device_info.event_interrupt_class = &event_interrupt_class_v9; 140 break; 141 case IP_VERSION(9, 4, 3): /* GC 9.4.3 */ 142 case IP_VERSION(9, 4, 4): /* GC 9.4.4 */ 143 case IP_VERSION(9, 5, 0): /* GC 9.5.0 */ 144 kfd->device_info.event_interrupt_class = 145 &event_interrupt_class_v9_4_3; 146 break; 147 case IP_VERSION(10, 3, 1): /* VANGOGH */ 148 case IP_VERSION(10, 3, 3): /* YELLOW_CARP */ 149 case IP_VERSION(10, 3, 6): /* GC 10.3.6 */ 150 case IP_VERSION(10, 3, 7): /* GC 10.3.7 */ 151 case IP_VERSION(10, 1, 3): /* CYAN_SKILLFISH */ 152 case IP_VERSION(10, 1, 4): 153 case IP_VERSION(10, 1, 10): /* NAVI10 */ 154 case IP_VERSION(10, 1, 2): /* NAVI12 */ 155 case IP_VERSION(10, 1, 1): /* NAVI14 */ 156 case IP_VERSION(10, 3, 0): /* SIENNA_CICHLID */ 157 case IP_VERSION(10, 3, 2): /* NAVY_FLOUNDER */ 158 case IP_VERSION(10, 3, 4): /* DIMGREY_CAVEFISH */ 159 case IP_VERSION(10, 3, 5): /* BEIGE_GOBY */ 160 kfd->device_info.event_interrupt_class = &event_interrupt_class_v10; 161 break; 162 case IP_VERSION(11, 0, 0): 163 case IP_VERSION(11, 0, 1): 164 case IP_VERSION(11, 0, 2): 165 case IP_VERSION(11, 0, 3): 166 case IP_VERSION(11, 0, 4): 167 case IP_VERSION(11, 5, 0): 168 case IP_VERSION(11, 5, 1): 169 case IP_VERSION(11, 5, 2): 170 case IP_VERSION(11, 5, 3): 171 case IP_VERSION(11, 5, 4): 172 case IP_VERSION(11, 5, 6): 173 case IP_VERSION(11, 7, 0): 174 case IP_VERSION(11, 7, 1): 175 kfd->device_info.event_interrupt_class = &event_interrupt_class_v11; 176 break; 177 case IP_VERSION(12, 0, 0): 178 case IP_VERSION(12, 0, 1): 179 /* GFX12_TODO: Change to v12 version. */ 180 kfd->device_info.event_interrupt_class = &event_interrupt_class_v11; 181 break; 182 case IP_VERSION(12, 1, 0): 183 kfd->device_info.event_interrupt_class = 184 &event_interrupt_class_v12_1; 185 break; 186 default: 187 dev_warn(kfd_device, "v9 event interrupt handler is set due to " 188 "mismatch of gc ip block(GC_HWIP:0x%x).\n", gc_version); 189 kfd->device_info.event_interrupt_class = &event_interrupt_class_v9; 190 } 191 } 192 193 static void kfd_device_info_init(struct kfd_dev *kfd, 194 bool vf, uint32_t gfx_target_version) 195 { 196 uint32_t gc_version = KFD_GC_VERSION(kfd); 197 uint32_t asic_type = kfd->adev->asic_type; 198 199 kfd->device_info.max_pasid_bits = 16; 200 kfd->device_info.max_no_of_hqd = 24; 201 kfd->device_info.num_of_watch_points = 4; 202 kfd->device_info.mqd_size_aligned = MQD_SIZE_ALIGNED; 203 kfd->device_info.gfx_target_version = gfx_target_version; 204 205 if (KFD_IS_SOC15(kfd)) { 206 kfd->device_info.doorbell_size = 8; 207 kfd->device_info.ih_ring_entry_size = 8 * sizeof(uint32_t); 208 kfd->device_info.supports_cwsr = true; 209 210 kfd_device_info_set_sdma_info(kfd); 211 212 kfd_device_info_set_event_interrupt_class(kfd); 213 214 if (gc_version < IP_VERSION(11, 0, 0)) { 215 /* Navi2x+, Navi1x+ */ 216 if (gc_version == IP_VERSION(10, 3, 6)) 217 kfd->device_info.no_atomic_fw_version = 14; 218 else if (gc_version == IP_VERSION(10, 3, 7)) 219 kfd->device_info.no_atomic_fw_version = 3; 220 else if (gc_version >= IP_VERSION(10, 3, 0)) 221 kfd->device_info.no_atomic_fw_version = 92; 222 else if (gc_version >= IP_VERSION(10, 1, 1)) 223 kfd->device_info.no_atomic_fw_version = 145; 224 225 /* Navi1x+ */ 226 if (gc_version >= IP_VERSION(10, 1, 1)) 227 kfd->device_info.needs_pci_atomics = true; 228 } else if (gc_version < IP_VERSION(12, 0, 0)) { 229 /* 230 * PCIe atomics support acknowledgment in GFX11 RS64 CPFW requires 231 * MEC version >= 509. Prior RS64 CPFW versions (and all F32) require 232 * PCIe atomics support. 233 */ 234 kfd->device_info.needs_pci_atomics = true; 235 kfd->device_info.no_atomic_fw_version = kfd->adev->gfx.rs64_enable ? 509 : 0; 236 } else if (gc_version < IP_VERSION(13, 0, 0)) { 237 kfd->device_info.needs_pci_atomics = true; 238 kfd->device_info.no_atomic_fw_version = 2090; 239 } else { 240 kfd->device_info.needs_pci_atomics = true; 241 } 242 } else { 243 kfd->device_info.doorbell_size = 4; 244 kfd->device_info.ih_ring_entry_size = 4 * sizeof(uint32_t); 245 kfd->device_info.event_interrupt_class = &event_interrupt_class_cik; 246 kfd->device_info.num_sdma_queues_per_engine = 2; 247 248 if (asic_type != CHIP_KAVERI && 249 asic_type != CHIP_HAWAII && 250 asic_type != CHIP_TONGA) 251 kfd->device_info.supports_cwsr = true; 252 253 if (asic_type != CHIP_HAWAII && !vf) 254 kfd->device_info.needs_pci_atomics = true; 255 } 256 } 257 258 struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf) 259 { 260 struct kfd_dev *kfd = NULL; 261 const struct kfd2kgd_calls *f2g = NULL; 262 uint32_t gfx_target_version = 0; 263 264 switch (adev->asic_type) { 265 #ifdef CONFIG_DRM_AMDGPU_CIK 266 case CHIP_KAVERI: 267 gfx_target_version = 70000; 268 if (!vf) 269 f2g = &gfx_v7_kfd2kgd; 270 break; 271 #endif 272 case CHIP_CARRIZO: 273 gfx_target_version = 80001; 274 if (!vf) 275 f2g = &gfx_v8_kfd2kgd; 276 break; 277 #ifdef CONFIG_DRM_AMDGPU_CIK 278 case CHIP_HAWAII: 279 gfx_target_version = 70001; 280 if (!amdgpu_exp_hw_support) 281 pr_info( 282 "KFD support on Hawaii is experimental. See modparam exp_hw_support\n" 283 ); 284 else if (!vf) 285 f2g = &gfx_v7_kfd2kgd; 286 break; 287 #endif 288 case CHIP_TONGA: 289 gfx_target_version = 80002; 290 if (!vf) 291 f2g = &gfx_v8_kfd2kgd; 292 break; 293 case CHIP_FIJI: 294 case CHIP_POLARIS10: 295 gfx_target_version = 80003; 296 f2g = &gfx_v8_kfd2kgd; 297 break; 298 case CHIP_POLARIS11: 299 case CHIP_POLARIS12: 300 case CHIP_VEGAM: 301 gfx_target_version = 80003; 302 if (!vf) 303 f2g = &gfx_v8_kfd2kgd; 304 break; 305 default: 306 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 307 /* Vega 10 */ 308 case IP_VERSION(9, 0, 1): 309 gfx_target_version = 90000; 310 f2g = &gfx_v9_kfd2kgd; 311 break; 312 /* Raven */ 313 case IP_VERSION(9, 1, 0): 314 case IP_VERSION(9, 2, 2): 315 gfx_target_version = 90002; 316 if (!vf) 317 f2g = &gfx_v9_kfd2kgd; 318 break; 319 /* Vega12 */ 320 case IP_VERSION(9, 2, 1): 321 gfx_target_version = 90004; 322 if (!vf) 323 f2g = &gfx_v9_kfd2kgd; 324 break; 325 /* Renoir */ 326 case IP_VERSION(9, 3, 0): 327 gfx_target_version = 90012; 328 if (!vf) 329 f2g = &gfx_v9_kfd2kgd; 330 break; 331 /* Vega20 */ 332 case IP_VERSION(9, 4, 0): 333 gfx_target_version = 90006; 334 if (!vf) 335 f2g = &gfx_v9_kfd2kgd; 336 break; 337 /* Arcturus */ 338 case IP_VERSION(9, 4, 1): 339 gfx_target_version = 90008; 340 f2g = &arcturus_kfd2kgd; 341 break; 342 /* Aldebaran */ 343 case IP_VERSION(9, 4, 2): 344 gfx_target_version = 90010; 345 f2g = &aldebaran_kfd2kgd; 346 break; 347 case IP_VERSION(9, 4, 3): 348 case IP_VERSION(9, 4, 4): 349 gfx_target_version = 90402; 350 f2g = &gc_9_4_3_kfd2kgd; 351 break; 352 case IP_VERSION(9, 5, 0): 353 gfx_target_version = 90500; 354 f2g = &gc_9_4_3_kfd2kgd; 355 break; 356 /* Navi10 */ 357 case IP_VERSION(10, 1, 10): 358 gfx_target_version = 100100; 359 if (!vf) 360 f2g = &gfx_v10_kfd2kgd; 361 break; 362 /* Navi12 */ 363 case IP_VERSION(10, 1, 2): 364 gfx_target_version = 100101; 365 f2g = &gfx_v10_kfd2kgd; 366 break; 367 /* Navi14 */ 368 case IP_VERSION(10, 1, 1): 369 gfx_target_version = 100102; 370 if (!vf) 371 f2g = &gfx_v10_kfd2kgd; 372 break; 373 /* Cyan Skillfish */ 374 case IP_VERSION(10, 1, 3): 375 case IP_VERSION(10, 1, 4): 376 gfx_target_version = 100103; 377 if (!vf) 378 f2g = &gfx_v10_kfd2kgd; 379 break; 380 /* Sienna Cichlid */ 381 case IP_VERSION(10, 3, 0): 382 gfx_target_version = 100300; 383 f2g = &gfx_v10_3_kfd2kgd; 384 break; 385 /* Navy Flounder */ 386 case IP_VERSION(10, 3, 2): 387 gfx_target_version = 100301; 388 f2g = &gfx_v10_3_kfd2kgd; 389 break; 390 /* Van Gogh */ 391 case IP_VERSION(10, 3, 1): 392 gfx_target_version = 100303; 393 if (!vf) 394 f2g = &gfx_v10_3_kfd2kgd; 395 break; 396 /* Dimgrey Cavefish */ 397 case IP_VERSION(10, 3, 4): 398 gfx_target_version = 100302; 399 f2g = &gfx_v10_3_kfd2kgd; 400 break; 401 /* Beige Goby */ 402 case IP_VERSION(10, 3, 5): 403 gfx_target_version = 100304; 404 f2g = &gfx_v10_3_kfd2kgd; 405 break; 406 /* Yellow Carp */ 407 case IP_VERSION(10, 3, 3): 408 gfx_target_version = 100305; 409 if (!vf) 410 f2g = &gfx_v10_3_kfd2kgd; 411 break; 412 case IP_VERSION(10, 3, 6): 413 case IP_VERSION(10, 3, 7): 414 gfx_target_version = 100306; 415 if (!vf) 416 f2g = &gfx_v10_3_kfd2kgd; 417 break; 418 case IP_VERSION(11, 0, 0): 419 gfx_target_version = 110000; 420 f2g = &gfx_v11_kfd2kgd; 421 break; 422 case IP_VERSION(11, 0, 1): 423 case IP_VERSION(11, 0, 4): 424 gfx_target_version = 110003; 425 f2g = &gfx_v11_kfd2kgd; 426 break; 427 case IP_VERSION(11, 0, 2): 428 gfx_target_version = 110002; 429 f2g = &gfx_v11_kfd2kgd; 430 break; 431 case IP_VERSION(11, 0, 3): 432 /* Note: Compiler version is 11.0.1 while HW version is 11.0.3 */ 433 gfx_target_version = 110001; 434 f2g = &gfx_v11_kfd2kgd; 435 break; 436 case IP_VERSION(11, 5, 0): 437 gfx_target_version = 110500; 438 f2g = &gfx_v11_kfd2kgd; 439 break; 440 case IP_VERSION(11, 5, 1): 441 gfx_target_version = 110501; 442 f2g = &gfx_v11_kfd2kgd; 443 break; 444 case IP_VERSION(11, 5, 2): 445 gfx_target_version = 110502; 446 f2g = &gfx_v11_kfd2kgd; 447 break; 448 case IP_VERSION(11, 5, 3): 449 gfx_target_version = 110503; 450 f2g = &gfx_v11_kfd2kgd; 451 break; 452 case IP_VERSION(11, 5, 4): 453 case IP_VERSION(11, 5, 6): 454 gfx_target_version = 110504; 455 f2g = &gfx_v11_kfd2kgd; 456 break; 457 case IP_VERSION(11, 7, 0): 458 gfx_target_version = 110700; 459 f2g = &gfx_v11_kfd2kgd; 460 break; 461 case IP_VERSION(11, 7, 1): 462 gfx_target_version = 110701; 463 f2g = &gfx_v11_kfd2kgd; 464 break; 465 case IP_VERSION(12, 0, 0): 466 gfx_target_version = 120000; 467 f2g = &gfx_v12_kfd2kgd; 468 break; 469 case IP_VERSION(12, 0, 1): 470 gfx_target_version = 120001; 471 f2g = &gfx_v12_kfd2kgd; 472 break; 473 case IP_VERSION(12, 1, 0): 474 gfx_target_version = 120500; 475 f2g = &gfx_v12_1_kfd2kgd; 476 break; 477 default: 478 break; 479 } 480 break; 481 } 482 483 if (!f2g) { 484 if (amdgpu_ip_version(adev, GC_HWIP, 0)) 485 dev_info(kfd_device, 486 "GC IP %06x %s not supported in kfd\n", 487 amdgpu_ip_version(adev, GC_HWIP, 0), 488 vf ? "VF" : ""); 489 else 490 dev_info(kfd_device, "%s %s not supported in kfd\n", 491 amdgpu_asic_name[adev->asic_type], vf ? "VF" : ""); 492 return NULL; 493 } 494 495 kfd = kzalloc_obj(*kfd); 496 if (!kfd) 497 return NULL; 498 499 kfd->adev = adev; 500 kfd_device_info_init(kfd, vf, gfx_target_version); 501 kfd->init_complete = false; 502 kfd->kfd2kgd = f2g; 503 atomic_set(&kfd->compute_profile, 0); 504 505 mutex_init(&kfd->doorbell_mutex); 506 507 ida_init(&kfd->doorbell_ida); 508 atomic_set(&kfd->kfd_processes_count, 0); 509 510 return kfd; 511 } 512 513 static void kfd_cwsr_init(struct kfd_dev *kfd) 514 { 515 if (cwsr_enable && kfd->device_info.supports_cwsr) { 516 if (KFD_GC_VERSION(kfd) < IP_VERSION(9, 0, 1)) { 517 BUILD_BUG_ON(sizeof(cwsr_trap_gfx8_hex) 518 > KFD_CWSR_TMA_OFFSET); 519 kfd->cwsr_isa = cwsr_trap_gfx8_hex; 520 kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx8_hex); 521 } else if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 1)) { 522 BUILD_BUG_ON(sizeof(cwsr_trap_arcturus_hex) 523 > KFD_CWSR_TMA_OFFSET); 524 kfd->cwsr_isa = cwsr_trap_arcturus_hex; 525 kfd->cwsr_isa_size = sizeof(cwsr_trap_arcturus_hex); 526 } else if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 2)) { 527 BUILD_BUG_ON(sizeof(cwsr_trap_aldebaran_hex) 528 > KFD_CWSR_TMA_OFFSET); 529 kfd->cwsr_isa = cwsr_trap_aldebaran_hex; 530 kfd->cwsr_isa_size = sizeof(cwsr_trap_aldebaran_hex); 531 } else if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3) || 532 KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 4)) { 533 BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_4_3_hex) 534 > KFD_CWSR_TMA_OFFSET); 535 kfd->cwsr_isa = cwsr_trap_gfx9_4_3_hex; 536 kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx9_4_3_hex); 537 } else if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 5, 0)) { 538 BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_5_0_hex) > PAGE_SIZE); 539 kfd->cwsr_isa = cwsr_trap_gfx9_5_0_hex; 540 kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx9_5_0_hex); 541 } else if (KFD_GC_VERSION(kfd) < IP_VERSION(10, 1, 1)) { 542 BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_hex) 543 > KFD_CWSR_TMA_OFFSET); 544 kfd->cwsr_isa = cwsr_trap_gfx9_hex; 545 kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx9_hex); 546 } else if (KFD_GC_VERSION(kfd) < IP_VERSION(10, 3, 0)) { 547 BUILD_BUG_ON(sizeof(cwsr_trap_nv1x_hex) 548 > KFD_CWSR_TMA_OFFSET); 549 kfd->cwsr_isa = cwsr_trap_nv1x_hex; 550 kfd->cwsr_isa_size = sizeof(cwsr_trap_nv1x_hex); 551 } else if (KFD_GC_VERSION(kfd) < IP_VERSION(11, 0, 0)) { 552 BUILD_BUG_ON(sizeof(cwsr_trap_gfx10_hex) 553 > KFD_CWSR_TMA_OFFSET); 554 kfd->cwsr_isa = cwsr_trap_gfx10_hex; 555 kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx10_hex); 556 } else if (KFD_GC_VERSION(kfd) < IP_VERSION(12, 0, 0)) { 557 /* The gfx11 cwsr trap handler must fit inside a single 558 page. */ 559 BUILD_BUG_ON(sizeof(cwsr_trap_gfx11_hex) > PAGE_SIZE); 560 kfd->cwsr_isa = cwsr_trap_gfx11_hex; 561 kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx11_hex); 562 } else if (KFD_GC_VERSION(kfd) < IP_VERSION(12, 1, 0)) { 563 BUILD_BUG_ON(sizeof(cwsr_trap_gfx12_hex) 564 > KFD_CWSR_TMA_OFFSET); 565 kfd->cwsr_isa = cwsr_trap_gfx12_hex; 566 kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx12_hex); 567 } else { 568 BUILD_BUG_ON(sizeof(cwsr_trap_gfx12_1_0_hex) 569 > KFD_CWSR_TMA_OFFSET); 570 kfd->cwsr_isa = cwsr_trap_gfx12_1_0_hex; 571 kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx12_1_0_hex); 572 } 573 574 kfd->cwsr_enabled = true; 575 } 576 } 577 578 static int kfd_gws_init(struct kfd_node *node) 579 { 580 int ret = 0; 581 struct kfd_dev *kfd = node->kfd; 582 uint32_t mes_rev = node->adev->mes.sched_version & AMDGPU_MES_VERSION_MASK; 583 584 if (node->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) 585 return 0; 586 587 if (hws_gws_support || (KFD_IS_SOC15(node) && 588 ((KFD_GC_VERSION(node) == IP_VERSION(9, 0, 1) 589 && kfd->mec2_fw_version >= 0x81b3) || 590 (KFD_GC_VERSION(node) <= IP_VERSION(9, 4, 0) 591 && kfd->mec2_fw_version >= 0x1b3) || 592 (KFD_GC_VERSION(node) == IP_VERSION(9, 4, 1) 593 && kfd->mec2_fw_version >= 0x30) || 594 (KFD_GC_VERSION(node) == IP_VERSION(9, 4, 2) 595 && kfd->mec2_fw_version >= 0x28) || 596 (KFD_GC_VERSION(node) == IP_VERSION(9, 4, 3) || 597 KFD_GC_VERSION(node) == IP_VERSION(9, 4, 4)) || 598 (KFD_GC_VERSION(node) == IP_VERSION(9, 5, 0)) || 599 (KFD_GC_VERSION(node) >= IP_VERSION(10, 3, 0) 600 && KFD_GC_VERSION(node) < IP_VERSION(11, 0, 0) 601 && kfd->mec2_fw_version >= 0x6b) || 602 (KFD_GC_VERSION(node) >= IP_VERSION(11, 0, 0) 603 && KFD_GC_VERSION(node) < IP_VERSION(12, 0, 0) 604 && mes_rev >= 68) || 605 (KFD_GC_VERSION(node) >= IP_VERSION(12, 0, 0))))) { 606 if (KFD_GC_VERSION(node) >= IP_VERSION(12, 0, 0)) 607 node->adev->gds.gws_size = 64; 608 ret = amdgpu_amdkfd_alloc_gws(node->adev, 609 node->adev->gds.gws_size, &node->gws); 610 } 611 612 return ret; 613 } 614 615 static void kfd_smi_init(struct kfd_node *dev) 616 { 617 INIT_LIST_HEAD(&dev->smi_clients); 618 spin_lock_init(&dev->smi_lock); 619 } 620 621 static int kfd_init_node(struct kfd_node *node) 622 { 623 int err = -1; 624 625 if (kfd_interrupt_init(node)) { 626 dev_err(kfd_device, "Error initializing interrupts\n"); 627 goto kfd_interrupt_error; 628 } 629 630 node->dqm = device_queue_manager_init(node); 631 if (!node->dqm) { 632 dev_err(kfd_device, "Error initializing queue manager\n"); 633 goto device_queue_manager_error; 634 } 635 636 if (kfd_gws_init(node)) { 637 dev_err(kfd_device, "Could not allocate %d gws\n", 638 node->adev->gds.gws_size); 639 goto gws_error; 640 } 641 642 if (kfd_resume(node)) 643 goto kfd_resume_error; 644 645 if (kfd_topology_add_device(node)) { 646 dev_err(kfd_device, "Error adding device to topology\n"); 647 goto kfd_topology_add_device_error; 648 } 649 650 kfd_smi_init(node); 651 652 return 0; 653 654 kfd_topology_add_device_error: 655 kfd_resume_error: 656 gws_error: 657 device_queue_manager_uninit(node->dqm); 658 device_queue_manager_error: 659 kfd_interrupt_exit(node); 660 kfd_interrupt_error: 661 if (node->gws) 662 amdgpu_amdkfd_free_gws(node->adev, node->gws); 663 664 /* Cleanup the node memory here */ 665 kfree(node); 666 return err; 667 } 668 669 static void kfd_cleanup_nodes(struct kfd_dev *kfd, unsigned int num_nodes) 670 { 671 struct kfd_node *knode; 672 unsigned int i; 673 674 /* 675 * flush_work ensures that there are no outstanding 676 * work-queue items that will access interrupt_ring. New work items 677 * can't be created because we stopped interrupt handling above. 678 */ 679 flush_workqueue(kfd->ih_wq); 680 destroy_workqueue(kfd->ih_wq); 681 682 for (i = 0; i < num_nodes; i++) { 683 knode = kfd->nodes[i]; 684 device_queue_manager_uninit(knode->dqm); 685 kfd_interrupt_exit(knode); 686 kfd_topology_remove_device(knode); 687 if (knode->gws) 688 amdgpu_amdkfd_free_gws(knode->adev, knode->gws); 689 kfree(knode); 690 kfd->nodes[i] = NULL; 691 } 692 } 693 694 static void kfd_setup_interrupt_bitmap(struct kfd_node *node, 695 unsigned int kfd_node_idx) 696 { 697 struct amdgpu_device *adev = node->adev; 698 uint32_t xcc_mask = node->xcc_mask; 699 uint32_t xcc, mapped_xcc; 700 uint32_t bitmap; 701 /* 702 * Interrupt bitmap is setup for processing interrupts from 703 * different XCDs and AIDs. 704 * Interrupt bitmap is defined as follows: 705 * 1. Bits 0-15 - correspond to the NodeId field. 706 * Each bit corresponds to NodeId number. For example, if 707 * a KFD node has interrupt bitmap set to 0x7, then this 708 * KFD node will process interrupts with NodeId = 0, 1 and 2 709 * in the IH cookie. 710 * 2. Bits 16-31 - unused. 711 * 712 * Please note that the kfd_node_idx argument passed to this 713 * function is not related to NodeId field received in the 714 * IH cookie. 715 * 716 * In CPX mode, a KFD node will process an interrupt if: 717 * - the Node Id matches the corresponding bit set in 718 * Bits 0-15. 719 * - AND VMID reported in the interrupt lies within the 720 * VMID range of the node. 721 */ 722 switch (KFD_GC_VERSION(node)) { 723 case IP_VERSION(12, 1, 0): 724 for_each_inst(xcc, xcc_mask) { 725 mapped_xcc = GET_INST(GC, xcc); 726 bitmap = 0x2 | (0x4 << (mapped_xcc % 4)); 727 if (mapped_xcc/4) 728 bitmap = bitmap << 8; 729 node->interrupt_bitmap |= bitmap; 730 } 731 break; 732 default: 733 for_each_inst(xcc, xcc_mask) { 734 mapped_xcc = GET_INST(GC, xcc); 735 node->interrupt_bitmap |= (mapped_xcc % 2 ? 5 : 3) << (4 * (mapped_xcc / 2)); 736 } 737 break; 738 } 739 dev_info(kfd_device, "Node: %d, interrupt_bitmap: %x\n", kfd_node_idx, 740 node->interrupt_bitmap); 741 } 742 743 bool kgd2kfd_device_init(struct kfd_dev *kfd, 744 const struct kgd2kfd_shared_resources *gpu_resources) 745 { 746 unsigned int size, map_process_packet_size, i; 747 struct kfd_node *node; 748 uint32_t first_vmid_kfd, last_vmid_kfd, vmid_num_kfd; 749 unsigned int max_proc_per_quantum; 750 int partition_mode; 751 int xcp_idx; 752 753 kfd->profiler_process = NULL; 754 mutex_init(&kfd->profiler_lock); 755 756 kfd->mec_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev, 757 KGD_ENGINE_MEC1); 758 kfd->mec2_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev, 759 KGD_ENGINE_MEC2); 760 kfd->sdma_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev, 761 KGD_ENGINE_SDMA1); 762 kfd->shared_resources = *gpu_resources; 763 764 kfd->num_nodes = amdgpu_xcp_get_num_xcp(kfd->adev->xcp_mgr); 765 766 if (kfd->num_nodes == 0) { 767 dev_err(kfd_device, 768 "KFD num nodes cannot be 0, num_xcc_in_node: %d\n", 769 kfd->adev->gfx.num_xcc_per_xcp); 770 goto out; 771 } 772 773 /* Allow BIF to recode atomics to PCIe 3.0 AtomicOps. 774 * 32 and 64-bit requests are possible and must be 775 * supported. 776 */ 777 kfd->pci_atomic_requested = amdgpu_amdkfd_have_atomics_support(kfd->adev); 778 if (!kfd->pci_atomic_requested && 779 kfd->device_info.needs_pci_atomics && 780 (!kfd->device_info.no_atomic_fw_version || 781 kfd->mec_fw_version < kfd->device_info.no_atomic_fw_version)) { 782 dev_info(kfd_device, 783 "skipped device %x:%x, PCI rejects atomics %d<%d\n", 784 kfd->adev->pdev->vendor, kfd->adev->pdev->device, 785 kfd->mec_fw_version, 786 kfd->device_info.no_atomic_fw_version); 787 return false; 788 } 789 790 first_vmid_kfd = ffs(gpu_resources->compute_vmid_bitmap)-1; 791 last_vmid_kfd = fls(gpu_resources->compute_vmid_bitmap)-1; 792 vmid_num_kfd = last_vmid_kfd - first_vmid_kfd + 1; 793 794 /* For multi-partition capable GPUs, we need special handling for VMIDs 795 * depending on partition mode. 796 * In CPX mode, the VMID range needs to be shared between XCDs. 797 * Additionally, there are 13 VMIDs (3-15) available for KFD. To 798 * divide them equally, we change starting VMID to 4 and not use 799 * VMID 3. 800 * If the VMID range changes for multi-partition capable GPUs, then 801 * this code MUST be revisited. 802 */ 803 if (kfd->adev->xcp_mgr && (KFD_GC_VERSION(kfd) != IP_VERSION(12, 1, 0))) { 804 partition_mode = amdgpu_xcp_query_partition_mode(kfd->adev->xcp_mgr, 805 AMDGPU_XCP_FL_LOCKED); 806 if (partition_mode == AMDGPU_CPX_PARTITION_MODE && 807 kfd->num_nodes != 1) { 808 vmid_num_kfd /= 2; 809 first_vmid_kfd = last_vmid_kfd + 1 - vmid_num_kfd*2; 810 } 811 } 812 813 /* Verify module parameters regarding mapped process number*/ 814 if (hws_max_conc_proc >= 0) 815 max_proc_per_quantum = min((u32)hws_max_conc_proc, vmid_num_kfd); 816 else 817 max_proc_per_quantum = vmid_num_kfd; 818 819 /* calculate max size of mqds needed for queues */ 820 size = max_num_of_queues_per_device * 821 kfd->device_info.mqd_size_aligned; 822 823 /* 824 * calculate max size of runlist packet. 825 * There can be only 2 packets at once 826 */ 827 map_process_packet_size = KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 2) ? 828 sizeof(struct pm4_mes_map_process_aldebaran) : 829 sizeof(struct pm4_mes_map_process); 830 size += (KFD_MAX_NUM_OF_PROCESSES * map_process_packet_size + 831 max_num_of_queues_per_device * sizeof(struct pm4_mes_map_queues) 832 + sizeof(struct pm4_mes_runlist)) * 2; 833 834 /* Add size of HIQ & DIQ */ 835 size += KFD_KERNEL_QUEUE_SIZE * 2; 836 837 /* add another 512KB for all other allocations on gart (HPD, fences) */ 838 size += 512 * 1024; 839 840 if (amdgpu_amdkfd_alloc_kernel_mem( 841 kfd->adev, size, AMDGPU_GEM_DOMAIN_GTT, 842 &kfd->gtt_mem, 843 &kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr, 844 false)) { 845 dev_err(kfd_device, "Could not allocate %d bytes\n", size); 846 goto alloc_kernel_mem_failure; 847 } 848 849 dev_info(kfd_device, "Allocated %d bytes on gart\n", size); 850 851 /* Initialize GTT sa with 512 byte chunk size */ 852 if (kfd_gtt_sa_init(kfd, size, 512) != 0) { 853 dev_err(kfd_device, "Error initializing gtt sub-allocator\n"); 854 goto kfd_gtt_sa_init_error; 855 } 856 857 if (kfd_doorbell_init(kfd)) { 858 dev_err(kfd_device, 859 "Error initializing doorbell aperture\n"); 860 goto kfd_doorbell_error; 861 } 862 863 if (amdgpu_use_xgmi_p2p) 864 kfd->hive_id = kfd->adev->gmc.xgmi.hive_id; 865 866 /* 867 * For multi-partition capable GPUs, the KFD abstracts all partitions 868 * within a socket as xGMI connected in the topology so assign a unique 869 * hive id per device based on the pci device location if device is in 870 * PCIe mode. 871 */ 872 if (!kfd->hive_id && kfd->num_nodes > 1) 873 kfd->hive_id = pci_dev_id(kfd->adev->pdev); 874 875 kfd->noretry = kfd->adev->gmc.noretry; 876 877 kfd_cwsr_init(kfd); 878 879 dev_info(kfd_device, "Total number of KFD nodes to be created: %d\n", 880 kfd->num_nodes); 881 882 /* Allocate the KFD nodes */ 883 for (i = 0, xcp_idx = 0; i < kfd->num_nodes; i++) { 884 node = kzalloc_obj(struct kfd_node); 885 if (!node) 886 goto node_alloc_error; 887 888 node->node_id = i; 889 node->adev = kfd->adev; 890 node->kfd = kfd; 891 node->kfd2kgd = kfd->kfd2kgd; 892 node->vm_info.vmid_num_kfd = vmid_num_kfd; 893 node->xcp = amdgpu_get_next_xcp(kfd->adev->xcp_mgr, &xcp_idx); 894 /* TODO : Check if error handling is needed */ 895 if (node->xcp) { 896 amdgpu_xcp_get_inst_details(node->xcp, AMDGPU_XCP_GFX, 897 &node->xcc_mask); 898 ++xcp_idx; 899 } else { 900 node->xcc_mask = 901 (1U << NUM_XCC(kfd->adev->gfx.xcc_mask)) - 1; 902 } 903 904 if (node->xcp) { 905 dev_info(kfd_device, "KFD node %d partition %d size %lldM\n", 906 node->node_id, node->xcp->mem_id, 907 KFD_XCP_MEMORY_SIZE(node->adev, node->node_id) >> 20); 908 } 909 910 if (partition_mode == AMDGPU_CPX_PARTITION_MODE && 911 kfd->num_nodes != 1 && 912 (KFD_GC_VERSION(kfd) != IP_VERSION(12, 1, 0))) { 913 /* For multi-partition capable GPUs and CPX mode, first 914 * XCD gets VMID range 4-9 and second XCD gets VMID 915 * range 10-15. 916 */ 917 918 node->vm_info.first_vmid_kfd = (i%2 == 0) ? 919 first_vmid_kfd : 920 first_vmid_kfd+vmid_num_kfd; 921 node->vm_info.last_vmid_kfd = (i%2 == 0) ? 922 last_vmid_kfd-vmid_num_kfd : 923 last_vmid_kfd; 924 node->compute_vmid_bitmap = 925 ((0x1 << (node->vm_info.last_vmid_kfd + 1)) - 1) - 926 ((0x1 << (node->vm_info.first_vmid_kfd)) - 1); 927 } else { 928 node->vm_info.first_vmid_kfd = first_vmid_kfd; 929 node->vm_info.last_vmid_kfd = last_vmid_kfd; 930 node->compute_vmid_bitmap = 931 gpu_resources->compute_vmid_bitmap; 932 } 933 934 node->max_proc_per_quantum = max_proc_per_quantum; 935 atomic_set(&node->sram_ecc_flag, 0); 936 937 amdgpu_amdkfd_get_local_mem_info(kfd->adev, 938 &node->local_mem_info, node->xcp); 939 940 if (kfd->adev->xcp_mgr) 941 kfd_setup_interrupt_bitmap(node, i); 942 943 /* Initialize the KFD node */ 944 if (kfd_init_node(node)) { 945 dev_err(kfd_device, "Error initializing KFD node\n"); 946 goto node_init_error; 947 } 948 949 spin_lock_init(&node->watch_points_lock); 950 951 kfd->nodes[i] = node; 952 } 953 954 svm_range_set_max_pages(kfd->adev); 955 956 /* Release pairs with the acquire in kfd_init_apertures(). */ 957 smp_store_release(&kfd->init_complete, true); 958 dev_info(kfd_device, "added device %x:%x\n", kfd->adev->pdev->vendor, 959 kfd->adev->pdev->device); 960 961 pr_debug("Starting kfd with the following scheduling policy %d\n", 962 node->dqm->sched_policy); 963 964 goto out; 965 966 node_init_error: 967 node_alloc_error: 968 kfd_cleanup_nodes(kfd, i); 969 kfd_doorbell_fini(kfd); 970 kfd_doorbell_error: 971 kfd_gtt_sa_fini(kfd); 972 kfd_gtt_sa_init_error: 973 amdgpu_amdkfd_free_kernel_mem(kfd->adev, &kfd->gtt_mem); 974 alloc_kernel_mem_failure: 975 dev_err(kfd_device, 976 "device %x:%x NOT added due to errors\n", 977 kfd->adev->pdev->vendor, kfd->adev->pdev->device); 978 out: 979 return kfd->init_complete; 980 } 981 982 void kgd2kfd_device_fini(struct kfd_dev *kfd) 983 { 984 if (kfd->init_complete) { 985 /* Cleanup KFD nodes */ 986 kfd_cleanup_nodes(kfd, kfd->num_nodes); 987 /* Cleanup common/shared resources */ 988 kfd_doorbell_fini(kfd); 989 ida_destroy(&kfd->doorbell_ida); 990 kfd_gtt_sa_fini(kfd); 991 amdgpu_amdkfd_free_kernel_mem(kfd->adev, &kfd->gtt_mem); 992 mutex_destroy(&kfd->profiler_lock); 993 } 994 995 kfree(kfd); 996 } 997 998 void kgd2kfd_device_exit(struct kfd_dev *kfd) 999 { 1000 kgd2kfd_device_fini(kfd); 1001 1002 /* after remove a kfd device unlock kfd driver */ 1003 kgd2kfd_unlock_kfd(NULL); 1004 } 1005 1006 int kgd2kfd_pre_reset(struct kfd_dev *kfd, 1007 struct amdgpu_reset_context *reset_context) 1008 { 1009 struct kfd_node *node; 1010 int i; 1011 1012 if (!kfd->init_complete) 1013 return 0; 1014 1015 for (i = 0; i < kfd->num_nodes; i++) { 1016 node = kfd->nodes[i]; 1017 kfd_smi_event_update_gpu_reset(node, false, reset_context); 1018 } 1019 1020 kgd2kfd_suspend(kfd, true); 1021 1022 for (i = 0; i < kfd->num_nodes; i++) 1023 kfd_signal_reset_event(kfd->nodes[i]); 1024 1025 return 0; 1026 } 1027 1028 /* 1029 * Fix me. KFD won't be able to resume existing process for now. 1030 * We will keep all existing process in a evicted state and 1031 * wait the process to be terminated. 1032 */ 1033 1034 int kgd2kfd_post_reset(struct kfd_dev *kfd) 1035 { 1036 int ret; 1037 struct kfd_node *node; 1038 int i; 1039 1040 if (!kfd->init_complete) 1041 return 0; 1042 1043 for (i = 0; i < kfd->num_nodes; i++) 1044 kfd_dqm_set_queues_as_reset(kfd->nodes[i]->dqm); 1045 1046 for (i = 0; i < kfd->num_nodes; i++) { 1047 ret = kfd_resume(kfd->nodes[i]); 1048 if (ret) 1049 return ret; 1050 } 1051 1052 mutex_lock(&kfd_processes_mutex); 1053 --kfd_locked; 1054 mutex_unlock(&kfd_processes_mutex); 1055 1056 for (i = 0; i < kfd->num_nodes; i++) { 1057 node = kfd->nodes[i]; 1058 atomic_set(&node->sram_ecc_flag, 0); 1059 kfd_smi_event_update_gpu_reset(node, true, NULL); 1060 } 1061 1062 return 0; 1063 } 1064 1065 bool kfd_is_locked(struct kfd_dev *kfd) 1066 { 1067 uint8_t id = 0; 1068 struct kfd_node *dev; 1069 1070 lockdep_assert_held(&kfd_processes_mutex); 1071 1072 /* check reset/suspend lock */ 1073 if (kfd_locked > 0) 1074 return true; 1075 1076 if (kfd) 1077 return kfd->kfd_dev_lock > 0; 1078 1079 /* check lock on all cgroup accessible devices */ 1080 while (kfd_topology_enum_kfd_devices(id++, &dev) == 0) { 1081 if (!dev || kfd_devcgroup_check_permission(dev)) 1082 continue; 1083 1084 if (dev->kfd->kfd_dev_lock > 0) 1085 return true; 1086 } 1087 1088 return false; 1089 } 1090 1091 void kgd2kfd_suspend(struct kfd_dev *kfd, bool suspend_proc) 1092 { 1093 struct kfd_node *node; 1094 int i; 1095 1096 if (!kfd->init_complete) 1097 return; 1098 1099 if (suspend_proc) 1100 kgd2kfd_suspend_process(kfd); 1101 1102 for (i = 0; i < kfd->num_nodes; i++) { 1103 node = kfd->nodes[i]; 1104 node->dqm->ops.stop(node->dqm); 1105 } 1106 } 1107 1108 int kgd2kfd_resume(struct kfd_dev *kfd, bool resume_proc) 1109 { 1110 int ret = 0, i; 1111 1112 if (!kfd->init_complete) 1113 return 0; 1114 1115 for (i = 0; i < kfd->num_nodes; i++) { 1116 ret = kfd_resume(kfd->nodes[i]); 1117 if (ret) 1118 return ret; 1119 } 1120 1121 if (resume_proc) 1122 ret = kgd2kfd_resume_process(kfd); 1123 1124 return ret; 1125 } 1126 1127 void kgd2kfd_suspend_process(struct kfd_dev *kfd) 1128 { 1129 if (!kfd->init_complete) 1130 return; 1131 1132 mutex_lock(&kfd_processes_mutex); 1133 /* For first KFD device suspend all the KFD processes */ 1134 if (++kfd_locked == 1) 1135 kfd_suspend_all_processes(); 1136 mutex_unlock(&kfd_processes_mutex); 1137 } 1138 1139 int kgd2kfd_resume_process(struct kfd_dev *kfd) 1140 { 1141 int ret = 0; 1142 1143 if (!kfd->init_complete) 1144 return 0; 1145 1146 mutex_lock(&kfd_processes_mutex); 1147 if (--kfd_locked == 0) 1148 ret = kfd_resume_all_processes(); 1149 WARN_ONCE(kfd_locked < 0, "KFD suspend / resume ref. error"); 1150 mutex_unlock(&kfd_processes_mutex); 1151 1152 return ret; 1153 } 1154 1155 static int kfd_resume(struct kfd_node *node) 1156 { 1157 int err = 0; 1158 1159 err = node->dqm->ops.start(node->dqm); 1160 if (err) 1161 dev_err(kfd_device, 1162 "Error starting queue manager for device %x:%x\n", 1163 node->adev->pdev->vendor, node->adev->pdev->device); 1164 1165 return err; 1166 } 1167 1168 /* This is called directly from KGD at ISR. */ 1169 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) 1170 { 1171 uint32_t patched_ihre[KFD_MAX_RING_ENTRY_SIZE], i; 1172 bool is_patched = false; 1173 unsigned long flags; 1174 struct kfd_node *node; 1175 1176 if (!kfd->init_complete) 1177 return; 1178 1179 if (kfd->device_info.ih_ring_entry_size > sizeof(patched_ihre)) { 1180 dev_err_once(kfd_device, "Ring entry too small\n"); 1181 return; 1182 } 1183 1184 for (i = 0; i < kfd->num_nodes; i++) { 1185 /* Race if another thread in b/w 1186 * kfd_cleanup_nodes and kfree(kfd), 1187 * when kfd->nodes[i] = NULL 1188 */ 1189 if (kfd->nodes[i]) 1190 node = kfd->nodes[i]; 1191 else 1192 return; 1193 1194 spin_lock_irqsave(&node->interrupt_lock, flags); 1195 1196 if (node->interrupts_active 1197 && interrupt_is_wanted(node, ih_ring_entry, 1198 patched_ihre, &is_patched) 1199 && enqueue_ih_ring_entry(node, 1200 is_patched ? patched_ihre : ih_ring_entry)) { 1201 queue_work(node->kfd->ih_wq, &node->interrupt_work); 1202 spin_unlock_irqrestore(&node->interrupt_lock, flags); 1203 return; 1204 } 1205 spin_unlock_irqrestore(&node->interrupt_lock, flags); 1206 } 1207 1208 } 1209 1210 int kgd2kfd_quiesce_mm(struct mm_struct *mm, uint32_t trigger) 1211 { 1212 struct kfd_process *p; 1213 int r; 1214 1215 /* Because we are called from arbitrary context (workqueue) as opposed 1216 * to process context, kfd_process could attempt to exit while we are 1217 * running so the lookup function increments the process ref count. 1218 */ 1219 p = kfd_lookup_process_by_mm(mm); 1220 if (!p) 1221 return -ESRCH; 1222 1223 WARN(debug_evictions, "Evicting pid %d", p->lead_thread->pid); 1224 r = kfd_process_evict_queues(p, trigger); 1225 1226 kfd_unref_process(p); 1227 return r; 1228 } 1229 1230 int kgd2kfd_resume_mm(struct mm_struct *mm) 1231 { 1232 struct kfd_process *p; 1233 int r; 1234 1235 /* Because we are called from arbitrary context (workqueue) as opposed 1236 * to process context, kfd_process could attempt to exit while we are 1237 * running so the lookup function increments the process ref count. 1238 */ 1239 p = kfd_lookup_process_by_mm(mm); 1240 if (!p) 1241 return -ESRCH; 1242 1243 r = kfd_process_restore_queues(p); 1244 1245 kfd_unref_process(p); 1246 return r; 1247 } 1248 1249 /** kgd2kfd_schedule_evict_and_restore_process - Schedules work queue that will 1250 * prepare for safe eviction of KFD BOs that belong to the specified 1251 * process. 1252 * 1253 * @mm: mm_struct that identifies a group of KFD processes 1254 * @context_id: an id that identifies a specific KFD context in the above kfd process group 1255 * @fence: eviction fence attached to KFD process BOs 1256 * 1257 */ 1258 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm, 1259 u16 context_id, struct dma_fence *fence) 1260 { 1261 struct kfd_process *p; 1262 unsigned long active_time; 1263 unsigned long delay_jiffies = msecs_to_jiffies(PROCESS_ACTIVE_TIME_MS); 1264 1265 if (!fence) 1266 return -EINVAL; 1267 1268 if (dma_fence_is_signaled(fence)) 1269 return 0; 1270 1271 p = kfd_lookup_process_by_id(mm, context_id); 1272 if (!p) 1273 return -ENODEV; 1274 1275 if (fence->seqno == p->last_eviction_seqno) 1276 goto out; 1277 1278 p->last_eviction_seqno = fence->seqno; 1279 1280 /* Avoid KFD process starvation. Wait for at least 1281 * PROCESS_ACTIVE_TIME_MS before evicting the process again 1282 */ 1283 active_time = get_jiffies_64() - p->last_restore_timestamp; 1284 if (delay_jiffies > active_time) 1285 delay_jiffies -= active_time; 1286 else 1287 delay_jiffies = 0; 1288 1289 /* During process initialization eviction_work.dwork is initialized 1290 * to kfd_evict_bo_worker 1291 */ 1292 WARN(debug_evictions, "Scheduling eviction of pid %d in %ld jiffies", 1293 p->lead_thread->pid, delay_jiffies); 1294 schedule_delayed_work(&p->eviction_work, delay_jiffies); 1295 out: 1296 kfd_unref_process(p); 1297 return 0; 1298 } 1299 1300 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size, 1301 unsigned int chunk_size) 1302 { 1303 if (WARN_ON(buf_size < chunk_size)) 1304 return -EINVAL; 1305 if (WARN_ON(buf_size == 0)) 1306 return -EINVAL; 1307 if (WARN_ON(chunk_size == 0)) 1308 return -EINVAL; 1309 1310 kfd->gtt_sa_chunk_size = chunk_size; 1311 kfd->gtt_sa_num_of_chunks = buf_size / chunk_size; 1312 1313 kfd->gtt_sa_bitmap = bitmap_zalloc(kfd->gtt_sa_num_of_chunks, 1314 GFP_KERNEL); 1315 if (!kfd->gtt_sa_bitmap) 1316 return -ENOMEM; 1317 1318 pr_debug("gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n", 1319 kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap); 1320 1321 mutex_init(&kfd->gtt_sa_lock); 1322 1323 return 0; 1324 } 1325 1326 static void kfd_gtt_sa_fini(struct kfd_dev *kfd) 1327 { 1328 mutex_destroy(&kfd->gtt_sa_lock); 1329 bitmap_free(kfd->gtt_sa_bitmap); 1330 } 1331 1332 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr, 1333 unsigned int bit_num, 1334 unsigned int chunk_size) 1335 { 1336 return start_addr + bit_num * chunk_size; 1337 } 1338 1339 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr, 1340 unsigned int bit_num, 1341 unsigned int chunk_size) 1342 { 1343 return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size); 1344 } 1345 1346 int kfd_gtt_sa_allocate(struct kfd_node *node, unsigned int size, 1347 struct kfd_mem_obj **mem_obj) 1348 { 1349 unsigned int found, start_search, cur_size; 1350 struct kfd_dev *kfd = node->kfd; 1351 1352 if (size == 0) 1353 return -EINVAL; 1354 1355 if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size) 1356 return -ENOMEM; 1357 1358 *mem_obj = kzalloc_obj(struct kfd_mem_obj); 1359 if (!(*mem_obj)) 1360 return -ENOMEM; 1361 1362 pr_debug("Allocated mem_obj = %p for size = %d\n", *mem_obj, size); 1363 1364 start_search = 0; 1365 1366 mutex_lock(&kfd->gtt_sa_lock); 1367 1368 kfd_gtt_restart_search: 1369 /* Find the first chunk that is free */ 1370 found = find_next_zero_bit(kfd->gtt_sa_bitmap, 1371 kfd->gtt_sa_num_of_chunks, 1372 start_search); 1373 1374 pr_debug("Found = %d\n", found); 1375 1376 /* If there wasn't any free chunk, bail out */ 1377 if (found == kfd->gtt_sa_num_of_chunks) 1378 goto kfd_gtt_no_free_chunk; 1379 1380 /* Update fields of mem_obj */ 1381 (*mem_obj)->range_start = found; 1382 (*mem_obj)->range_end = found; 1383 (*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr( 1384 kfd->gtt_start_gpu_addr, 1385 found, 1386 kfd->gtt_sa_chunk_size); 1387 (*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr( 1388 kfd->gtt_start_cpu_ptr, 1389 found, 1390 kfd->gtt_sa_chunk_size); 1391 1392 pr_debug("gpu_addr = %p, cpu_addr = %p\n", 1393 (uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr); 1394 1395 /* If we need only one chunk, mark it as allocated and get out */ 1396 if (size <= kfd->gtt_sa_chunk_size) { 1397 pr_debug("Single bit\n"); 1398 __set_bit(found, kfd->gtt_sa_bitmap); 1399 goto kfd_gtt_out; 1400 } 1401 1402 /* Otherwise, try to see if we have enough contiguous chunks */ 1403 cur_size = size - kfd->gtt_sa_chunk_size; 1404 do { 1405 (*mem_obj)->range_end = 1406 find_next_zero_bit(kfd->gtt_sa_bitmap, 1407 kfd->gtt_sa_num_of_chunks, ++found); 1408 /* 1409 * If next free chunk is not contiguous than we need to 1410 * restart our search from the last free chunk we found (which 1411 * wasn't contiguous to the previous ones 1412 */ 1413 if ((*mem_obj)->range_end != found) { 1414 start_search = found; 1415 goto kfd_gtt_restart_search; 1416 } 1417 1418 /* 1419 * If we reached end of buffer, bail out with error 1420 */ 1421 if (found == kfd->gtt_sa_num_of_chunks) 1422 goto kfd_gtt_no_free_chunk; 1423 1424 /* Check if we don't need another chunk */ 1425 if (cur_size <= kfd->gtt_sa_chunk_size) 1426 cur_size = 0; 1427 else 1428 cur_size -= kfd->gtt_sa_chunk_size; 1429 1430 } while (cur_size > 0); 1431 1432 pr_debug("range_start = %d, range_end = %d\n", 1433 (*mem_obj)->range_start, (*mem_obj)->range_end); 1434 1435 /* Mark the chunks as allocated */ 1436 bitmap_set(kfd->gtt_sa_bitmap, (*mem_obj)->range_start, 1437 (*mem_obj)->range_end - (*mem_obj)->range_start + 1); 1438 1439 kfd_gtt_out: 1440 mutex_unlock(&kfd->gtt_sa_lock); 1441 return 0; 1442 1443 kfd_gtt_no_free_chunk: 1444 pr_debug("Allocation failed with mem_obj = %p\n", *mem_obj); 1445 mutex_unlock(&kfd->gtt_sa_lock); 1446 kfree(*mem_obj); 1447 return -ENOMEM; 1448 } 1449 1450 int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj) 1451 { 1452 struct kfd_dev *kfd = node->kfd; 1453 1454 /* Act like kfree when trying to free a NULL object */ 1455 if (!mem_obj) 1456 return 0; 1457 1458 pr_debug("Free mem_obj = %p, range_start = %d, range_end = %d\n", 1459 mem_obj, mem_obj->range_start, mem_obj->range_end); 1460 1461 mutex_lock(&kfd->gtt_sa_lock); 1462 1463 /* Mark the chunks as free */ 1464 bitmap_clear(kfd->gtt_sa_bitmap, mem_obj->range_start, 1465 mem_obj->range_end - mem_obj->range_start + 1); 1466 1467 mutex_unlock(&kfd->gtt_sa_lock); 1468 1469 kfree(mem_obj); 1470 return 0; 1471 } 1472 1473 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd) 1474 { 1475 /* 1476 * TODO: Currently update SRAM ECC flag for first node. 1477 * This needs to be updated later when we can 1478 * identify SRAM ECC error on other nodes also. 1479 */ 1480 if (kfd) 1481 atomic_inc(&kfd->nodes[0]->sram_ecc_flag); 1482 } 1483 1484 void kfd_inc_compute_active(struct kfd_node *node) 1485 { 1486 if (atomic_inc_return(&node->kfd->compute_profile) == 1) 1487 amdgpu_amdkfd_set_compute_idle(node->adev, false); 1488 } 1489 1490 void kfd_dec_compute_active(struct kfd_node *node) 1491 { 1492 int count = atomic_dec_return(&node->kfd->compute_profile); 1493 1494 if (count == 0) 1495 amdgpu_amdkfd_set_compute_idle(node->adev, true); 1496 WARN_ONCE(count < 0, "Compute profile ref. count error"); 1497 } 1498 1499 static bool kfd_compute_active(struct kfd_node *node) 1500 { 1501 if (atomic_read(&node->kfd->compute_profile)) 1502 return true; 1503 return false; 1504 } 1505 1506 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask) 1507 { 1508 /* 1509 * TODO: For now, raise the throttling event only on first node. 1510 * This will need to change after we are able to determine 1511 * which node raised the throttling event. 1512 */ 1513 if (kfd && kfd->init_complete) 1514 kfd_smi_event_update_thermal_throttling(kfd->nodes[0], 1515 throttle_bitmask); 1516 } 1517 1518 /* kfd_get_num_sdma_engines returns the number of PCIe optimized SDMA and 1519 * kfd_get_num_xgmi_sdma_engines returns the number of XGMI SDMA. 1520 * When the device has more than two engines, we reserve two for PCIe to enable 1521 * full-duplex and the rest are used as XGMI. 1522 */ 1523 unsigned int kfd_get_num_sdma_engines(struct kfd_node *node) 1524 { 1525 /* If XGMI is not supported, all SDMA engines are PCIe. 1526 * Also, on GC 12.1, all SDMA engines are the same. 1527 */ 1528 if (!node->adev->gmc.xgmi.supported || 1529 KFD_GC_VERSION(node->kfd) == IP_VERSION(12, 1, 0)) 1530 return node->adev->sdma.num_instances/(int)node->kfd->num_nodes; 1531 1532 return min(node->adev->sdma.num_instances/(int)node->kfd->num_nodes, 2); 1533 } 1534 1535 unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *node) 1536 { 1537 /* After reserved for PCIe, the rest of engines are XGMI */ 1538 return node->adev->sdma.num_instances/(int)node->kfd->num_nodes - 1539 kfd_get_num_sdma_engines(node); 1540 } 1541 1542 int kgd2kfd_check_and_lock_kfd(struct kfd_dev *kfd) 1543 { 1544 struct kfd_process *p; 1545 int r = 0, temp, idx; 1546 1547 mutex_lock(&kfd_processes_mutex); 1548 1549 /* kfd_processes_count is per kfd_dev, return -EBUSY without 1550 * further check 1551 */ 1552 if (!!atomic_read(&kfd->kfd_processes_count)) { 1553 pr_debug("process_wq_release not finished\n"); 1554 r = -EBUSY; 1555 goto out; 1556 } 1557 1558 if (hash_empty(kfd_processes_table) && !kfd_is_locked(kfd)) 1559 goto out; 1560 1561 /* fail under system reset/resume or kfd device is partition switching. */ 1562 if (kfd_is_locked(kfd)) { 1563 r = -EBUSY; 1564 goto out; 1565 } 1566 1567 /* 1568 * ensure all running processes are cgroup excluded from device before mode switch. 1569 * i.e. no pdd was created on the process socket. 1570 */ 1571 idx = srcu_read_lock(&kfd_processes_srcu); 1572 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) { 1573 int i; 1574 1575 for (i = 0; i < p->n_pdds; i++) { 1576 if (p->pdds[i]->dev->kfd != kfd) 1577 continue; 1578 1579 r = -EBUSY; 1580 goto proc_check_unlock; 1581 } 1582 } 1583 1584 proc_check_unlock: 1585 srcu_read_unlock(&kfd_processes_srcu, idx); 1586 out: 1587 if (!r) 1588 ++kfd->kfd_dev_lock; 1589 mutex_unlock(&kfd_processes_mutex); 1590 1591 return r; 1592 } 1593 1594 /* unlock a kfd dev or kfd driver */ 1595 void kgd2kfd_unlock_kfd(struct kfd_dev *kfd) 1596 { 1597 mutex_lock(&kfd_processes_mutex); 1598 if (kfd) 1599 --kfd->kfd_dev_lock; 1600 else 1601 --kfd_locked; 1602 mutex_unlock(&kfd_processes_mutex); 1603 } 1604 1605 int kgd2kfd_start_sched(struct kfd_dev *kfd, uint32_t node_id) 1606 { 1607 struct kfd_node *node; 1608 int ret; 1609 1610 if (!kfd->init_complete) 1611 return 0; 1612 1613 if (node_id >= kfd->num_nodes) { 1614 dev_warn(kfd->adev->dev, "Invalid node ID: %u exceeds %u\n", 1615 node_id, kfd->num_nodes - 1); 1616 return -EINVAL; 1617 } 1618 node = kfd->nodes[node_id]; 1619 1620 ret = node->dqm->ops.unhalt(node->dqm); 1621 if (ret) 1622 dev_err(kfd_device, "Error in starting scheduler\n"); 1623 1624 return ret; 1625 } 1626 1627 int kgd2kfd_start_sched_all_nodes(struct kfd_dev *kfd) 1628 { 1629 struct kfd_node *node; 1630 int i, r; 1631 1632 if (!kfd->init_complete) 1633 return 0; 1634 1635 for (i = 0; i < kfd->num_nodes; i++) { 1636 node = kfd->nodes[i]; 1637 r = node->dqm->ops.unhalt(node->dqm); 1638 if (r) { 1639 dev_err(kfd_device, "Error in starting scheduler\n"); 1640 return r; 1641 } 1642 } 1643 return 0; 1644 } 1645 1646 int kgd2kfd_stop_sched(struct kfd_dev *kfd, uint32_t node_id) 1647 { 1648 struct kfd_node *node; 1649 1650 if (!kfd->init_complete) 1651 return 0; 1652 1653 if (node_id >= kfd->num_nodes) { 1654 dev_warn(kfd->adev->dev, "Invalid node ID: %u exceeds %u\n", 1655 node_id, kfd->num_nodes - 1); 1656 return -EINVAL; 1657 } 1658 1659 node = kfd->nodes[node_id]; 1660 return node->dqm->ops.halt(node->dqm); 1661 } 1662 1663 int kgd2kfd_stop_sched_all_nodes(struct kfd_dev *kfd) 1664 { 1665 struct kfd_node *node; 1666 int i, r; 1667 1668 if (!kfd->init_complete) 1669 return 0; 1670 1671 for (i = 0; i < kfd->num_nodes; i++) { 1672 node = kfd->nodes[i]; 1673 r = node->dqm->ops.halt(node->dqm); 1674 if (r) 1675 return r; 1676 } 1677 return 0; 1678 } 1679 1680 int amdgpu_amdkfd_stop_sched_all(struct amdgpu_device *adev) 1681 { 1682 if (!adev->kfd.init_complete) 1683 return 0; 1684 1685 return kgd2kfd_stop_sched_all_nodes(adev->kfd.dev); 1686 } 1687 1688 int amdgpu_amdkfd_start_sched_all(struct amdgpu_device *adev) 1689 { 1690 if (!adev->kfd.init_complete) 1691 return 0; 1692 1693 return kgd2kfd_start_sched_all_nodes(adev->kfd.dev); 1694 } 1695 1696 bool kgd2kfd_compute_active(struct kfd_dev *kfd, uint32_t node_id) 1697 { 1698 struct kfd_node *node; 1699 1700 if (!kfd->init_complete) 1701 return false; 1702 1703 if (node_id >= kfd->num_nodes) { 1704 dev_warn(kfd->adev->dev, "Invalid node ID: %u exceeds %u\n", 1705 node_id, kfd->num_nodes - 1); 1706 return false; 1707 } 1708 1709 node = kfd->nodes[node_id]; 1710 1711 return kfd_compute_active(node); 1712 } 1713 1714 /** 1715 * kgd2kfd_vmfault_fast_path() - KFD vm page fault interrupt handling fast path for gmc v9 1716 * @adev: amdgpu device 1717 * @entry: vm fault interrupt vector 1718 * @retry_fault: if this is retry fault 1719 * 1720 * retry fault - 1721 * with CAM enabled, adev primary ring 1722 * | gmc_v9_0_process_interrupt() 1723 * adev soft_ring 1724 * | gmc_v9_0_process_interrupt() worker failed to recover page fault 1725 * KFD node ih_fifo 1726 * | KFD interrupt_wq worker 1727 * kfd_signal_vm_fault_event 1728 * 1729 * without CAM, adev primary ring1 1730 * | gmc_v9_0_process_interrupt worker failed to recvoer page fault 1731 * KFD node ih_fifo 1732 * | KFD interrupt_wq worker 1733 * kfd_signal_vm_fault_event 1734 * 1735 * no-retry fault - 1736 * adev primary ring 1737 * | gmc_v9_0_process_interrupt() 1738 * KFD node ih_fifo 1739 * | KFD interrupt_wq worker 1740 * kfd_signal_vm_fault_event 1741 * 1742 * fast path - After kfd_signal_vm_fault_event, gmc_v9_0_process_interrupt drop the page fault 1743 * of same process, don't copy interrupt to KFD node ih_fifo. 1744 * With gdb debugger enabled, need convert the retry fault to no-retry fault for 1745 * debugger, cannot use the fast path. 1746 * 1747 * Return: 1748 * true - use the fast path to handle this fault 1749 * false - use normal path to handle it 1750 */ 1751 bool kgd2kfd_vmfault_fast_path(struct amdgpu_device *adev, struct amdgpu_iv_entry *entry, 1752 bool retry_fault) 1753 { 1754 struct kfd_process *p; 1755 u32 cam_index; 1756 u32 src_data_idx; 1757 1758 src_data_idx = (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(12, 1, 0)) ? 1759 3 : 2; 1760 1761 if (entry->ih == &adev->irq.ih_soft || entry->ih == &adev->irq.ih1) { 1762 p = kfd_lookup_process_by_pasid(entry->pasid, NULL); 1763 if (!p) 1764 return true; 1765 1766 if (p->gpu_page_fault && !p->debug_trap_enabled) { 1767 if (retry_fault && adev->irq.retry_cam_enabled) { 1768 cam_index = entry->src_data[src_data_idx] & 0x3ff; 1769 1770 WDOORBELL32(adev->irq.retry_cam_doorbell_index, cam_index); 1771 } 1772 1773 kfd_unref_process(p); 1774 return true; 1775 } 1776 1777 /* 1778 * This is the first page fault, set flag and then signal user space 1779 */ 1780 p->gpu_page_fault = true; 1781 kfd_unref_process(p); 1782 } 1783 return false; 1784 } 1785 1786 /** kgd2kfd_teardown_processes - gracefully tear down existing 1787 * kfd processes that use adev 1788 * 1789 * @adev: amdgpu_device where kfd processes run on and will be 1790 * teardown 1791 * 1792 */ 1793 void kgd2kfd_teardown_processes(struct amdgpu_device *adev) 1794 { 1795 struct hlist_node *p_temp; 1796 struct kfd_process *p; 1797 struct kfd_node *dev; 1798 unsigned int temp; 1799 1800 mutex_lock(&kfd_processes_mutex); 1801 1802 if (hash_empty(kfd_processes_table)) { 1803 mutex_unlock(&kfd_processes_mutex); 1804 return; 1805 } 1806 1807 hash_for_each_safe(kfd_processes_table, temp, p_temp, p, kfd_processes) { 1808 for (int i = 0; i < p->n_pdds; i++) { 1809 dev = p->pdds[i]->dev; 1810 if (dev->adev == adev) 1811 kfd_signal_process_terminate_event(p); 1812 } 1813 } 1814 1815 mutex_unlock(&kfd_processes_mutex); 1816 1817 /* wait all kfd processes use adev terminate */ 1818 while (!!atomic_read(&adev->kfd.dev->kfd_processes_count)) 1819 cond_resched(); 1820 } 1821 1822 int kgd2kfd_reset_mes_queue(struct kfd_dev *kfd, uint32_t node_id, 1823 int queue_type, int pipe, int queue, 1824 unsigned int db) 1825 { 1826 struct kfd_node *node; 1827 int ret; 1828 1829 if (!kfd->init_complete) 1830 return 0; 1831 1832 if (node_id >= kfd->num_nodes) { 1833 dev_warn(kfd->adev->dev, "Invalid node ID: %u exceeds %u\n", 1834 node_id, kfd->num_nodes - 1); 1835 return -EINVAL; 1836 } 1837 node = kfd->nodes[node_id]; 1838 1839 ret = kfd_reset_queue_mes(node->dqm, queue_type, pipe, queue, db); 1840 if (ret) 1841 dev_err(kfd_device, "Error resetting queue\n"); 1842 1843 return ret; 1844 } 1845 1846 #if defined(CONFIG_DEBUG_FS) 1847 1848 /* This function will send a package to HIQ to hang the HWS 1849 * which will trigger a GPU reset and bring the HWS back to normal state 1850 */ 1851 int kfd_debugfs_hang_hws(struct kfd_node *dev) 1852 { 1853 if (dev->dqm->sched_policy != KFD_SCHED_POLICY_HWS) { 1854 pr_err("HWS is not enabled"); 1855 return -EINVAL; 1856 } 1857 1858 if (dev->kfd->shared_resources.enable_mes) { 1859 dev_err(dev->adev->dev, "Inducing MES hang is not supported\n"); 1860 return -EINVAL; 1861 } 1862 1863 return dqm_debugfs_hang_hws(dev->dqm); 1864 } 1865 1866 #endif 1867