1 /* 2 * Copyright 2014 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 */ 22 23 #ifndef KFD_IOCTL_H_INCLUDED 24 #define KFD_IOCTL_H_INCLUDED 25 26 #include <drm/drm.h> 27 #include <linux/ioctl.h> 28 29 /* 30 * - 1.1 - initial version 31 * - 1.3 - Add SMI events support 32 * - 1.4 - Indicate new SRAM EDC bit in device properties 33 * - 1.5 - Add SVM API 34 * - 1.6 - Query clear flags in SVM get_attr API 35 * - 1.7 - Checkpoint Restore (CRIU) API 36 * - 1.8 - CRIU - Support for SDMA transfers with GTT BOs 37 * - 1.9 - Add available memory ioctl 38 * - 1.10 - Add SMI profiler event log 39 * - 1.11 - Add unified memory for ctx save/restore area 40 * - 1.12 - Add DMA buf export ioctl 41 * - 1.13 - Add debugger API 42 * - 1.14 - Update kfd_event_data 43 * - 1.15 - Enable managing mappings in compute VMs with GEM_VA ioctl 44 * - 1.16 - Add contiguous VRAM allocation flag 45 * - 1.17 - Add SDMA queue creation with target SDMA engine ID 46 */ 47 #define KFD_IOCTL_MAJOR_VERSION 1 48 #define KFD_IOCTL_MINOR_VERSION 17 49 50 struct kfd_ioctl_get_version_args { 51 __u32 major_version; /* from KFD */ 52 __u32 minor_version; /* from KFD */ 53 }; 54 55 /* For kfd_ioctl_create_queue_args.queue_type. */ 56 #define KFD_IOC_QUEUE_TYPE_COMPUTE 0x0 57 #define KFD_IOC_QUEUE_TYPE_SDMA 0x1 58 #define KFD_IOC_QUEUE_TYPE_COMPUTE_AQL 0x2 59 #define KFD_IOC_QUEUE_TYPE_SDMA_XGMI 0x3 60 #define KFD_IOC_QUEUE_TYPE_SDMA_BY_ENG_ID 0x4 61 62 #define KFD_MAX_QUEUE_PERCENTAGE 100 63 #define KFD_MAX_QUEUE_PRIORITY 15 64 65 struct kfd_ioctl_create_queue_args { 66 __u64 ring_base_address; /* to KFD */ 67 __u64 write_pointer_address; /* from KFD */ 68 __u64 read_pointer_address; /* from KFD */ 69 __u64 doorbell_offset; /* from KFD */ 70 71 __u32 ring_size; /* to KFD */ 72 __u32 gpu_id; /* to KFD */ 73 __u32 queue_type; /* to KFD */ 74 __u32 queue_percentage; /* to KFD */ 75 __u32 queue_priority; /* to KFD */ 76 __u32 queue_id; /* from KFD */ 77 78 __u64 eop_buffer_address; /* to KFD */ 79 __u64 eop_buffer_size; /* to KFD */ 80 __u64 ctx_save_restore_address; /* to KFD */ 81 __u32 ctx_save_restore_size; /* to KFD */ 82 __u32 ctl_stack_size; /* to KFD */ 83 __u32 sdma_engine_id; /* to KFD */ 84 __u32 pad; 85 }; 86 87 struct kfd_ioctl_destroy_queue_args { 88 __u32 queue_id; /* to KFD */ 89 __u32 pad; 90 }; 91 92 struct kfd_ioctl_update_queue_args { 93 __u64 ring_base_address; /* to KFD */ 94 95 __u32 queue_id; /* to KFD */ 96 __u32 ring_size; /* to KFD */ 97 __u32 queue_percentage; /* to KFD */ 98 __u32 queue_priority; /* to KFD */ 99 }; 100 101 struct kfd_ioctl_set_cu_mask_args { 102 __u32 queue_id; /* to KFD */ 103 __u32 num_cu_mask; /* to KFD */ 104 __u64 cu_mask_ptr; /* to KFD */ 105 }; 106 107 struct kfd_ioctl_get_queue_wave_state_args { 108 __u64 ctl_stack_address; /* to KFD */ 109 __u32 ctl_stack_used_size; /* from KFD */ 110 __u32 save_area_used_size; /* from KFD */ 111 __u32 queue_id; /* to KFD */ 112 __u32 pad; 113 }; 114 115 struct kfd_ioctl_get_available_memory_args { 116 __u64 available; /* from KFD */ 117 __u32 gpu_id; /* to KFD */ 118 __u32 pad; 119 }; 120 121 struct kfd_dbg_device_info_entry { 122 __u64 exception_status; 123 __u64 lds_base; 124 __u64 lds_limit; 125 __u64 scratch_base; 126 __u64 scratch_limit; 127 __u64 gpuvm_base; 128 __u64 gpuvm_limit; 129 __u32 gpu_id; 130 __u32 location_id; 131 __u32 vendor_id; 132 __u32 device_id; 133 __u32 revision_id; 134 __u32 subsystem_vendor_id; 135 __u32 subsystem_device_id; 136 __u32 fw_version; 137 __u32 gfx_target_version; 138 __u32 simd_count; 139 __u32 max_waves_per_simd; 140 __u32 array_count; 141 __u32 simd_arrays_per_engine; 142 __u32 num_xcc; 143 __u32 capability; 144 __u32 debug_prop; 145 }; 146 147 /* For kfd_ioctl_set_memory_policy_args.default_policy and alternate_policy */ 148 #define KFD_IOC_CACHE_POLICY_COHERENT 0 149 #define KFD_IOC_CACHE_POLICY_NONCOHERENT 1 150 151 struct kfd_ioctl_set_memory_policy_args { 152 __u64 alternate_aperture_base; /* to KFD */ 153 __u64 alternate_aperture_size; /* to KFD */ 154 155 __u32 gpu_id; /* to KFD */ 156 __u32 default_policy; /* to KFD */ 157 __u32 alternate_policy; /* to KFD */ 158 __u32 pad; 159 }; 160 161 /* 162 * All counters are monotonic. They are used for profiling of compute jobs. 163 * The profiling is done by userspace. 164 * 165 * In case of GPU reset, the counter should not be affected. 166 */ 167 168 struct kfd_ioctl_get_clock_counters_args { 169 __u64 gpu_clock_counter; /* from KFD */ 170 __u64 cpu_clock_counter; /* from KFD */ 171 __u64 system_clock_counter; /* from KFD */ 172 __u64 system_clock_freq; /* from KFD */ 173 174 __u32 gpu_id; /* to KFD */ 175 __u32 pad; 176 }; 177 178 struct kfd_process_device_apertures { 179 __u64 lds_base; /* from KFD */ 180 __u64 lds_limit; /* from KFD */ 181 __u64 scratch_base; /* from KFD */ 182 __u64 scratch_limit; /* from KFD */ 183 __u64 gpuvm_base; /* from KFD */ 184 __u64 gpuvm_limit; /* from KFD */ 185 __u32 gpu_id; /* from KFD */ 186 __u32 pad; 187 }; 188 189 /* 190 * AMDKFD_IOC_GET_PROCESS_APERTURES is deprecated. Use 191 * AMDKFD_IOC_GET_PROCESS_APERTURES_NEW instead, which supports an 192 * unlimited number of GPUs. 193 */ 194 #define NUM_OF_SUPPORTED_GPUS 7 195 struct kfd_ioctl_get_process_apertures_args { 196 struct kfd_process_device_apertures 197 process_apertures[NUM_OF_SUPPORTED_GPUS];/* from KFD */ 198 199 /* from KFD, should be in the range [1 - NUM_OF_SUPPORTED_GPUS] */ 200 __u32 num_of_nodes; 201 __u32 pad; 202 }; 203 204 struct kfd_ioctl_get_process_apertures_new_args { 205 /* User allocated. Pointer to struct kfd_process_device_apertures 206 * filled in by Kernel 207 */ 208 __u64 kfd_process_device_apertures_ptr; 209 /* to KFD - indicates amount of memory present in 210 * kfd_process_device_apertures_ptr 211 * from KFD - Number of entries filled by KFD. 212 */ 213 __u32 num_of_nodes; 214 __u32 pad; 215 }; 216 217 #define MAX_ALLOWED_NUM_POINTS 100 218 #define MAX_ALLOWED_AW_BUFF_SIZE 4096 219 #define MAX_ALLOWED_WAC_BUFF_SIZE 128 220 221 struct kfd_ioctl_dbg_register_args { 222 __u32 gpu_id; /* to KFD */ 223 __u32 pad; 224 }; 225 226 struct kfd_ioctl_dbg_unregister_args { 227 __u32 gpu_id; /* to KFD */ 228 __u32 pad; 229 }; 230 231 struct kfd_ioctl_dbg_address_watch_args { 232 __u64 content_ptr; /* a pointer to the actual content */ 233 __u32 gpu_id; /* to KFD */ 234 __u32 buf_size_in_bytes; /*including gpu_id and buf_size */ 235 }; 236 237 struct kfd_ioctl_dbg_wave_control_args { 238 __u64 content_ptr; /* a pointer to the actual content */ 239 __u32 gpu_id; /* to KFD */ 240 __u32 buf_size_in_bytes; /*including gpu_id and buf_size */ 241 }; 242 243 #define KFD_INVALID_FD 0xffffffff 244 245 /* Matching HSA_EVENTTYPE */ 246 #define KFD_IOC_EVENT_SIGNAL 0 247 #define KFD_IOC_EVENT_NODECHANGE 1 248 #define KFD_IOC_EVENT_DEVICESTATECHANGE 2 249 #define KFD_IOC_EVENT_HW_EXCEPTION 3 250 #define KFD_IOC_EVENT_SYSTEM_EVENT 4 251 #define KFD_IOC_EVENT_DEBUG_EVENT 5 252 #define KFD_IOC_EVENT_PROFILE_EVENT 6 253 #define KFD_IOC_EVENT_QUEUE_EVENT 7 254 #define KFD_IOC_EVENT_MEMORY 8 255 256 #define KFD_IOC_WAIT_RESULT_COMPLETE 0 257 #define KFD_IOC_WAIT_RESULT_TIMEOUT 1 258 #define KFD_IOC_WAIT_RESULT_FAIL 2 259 260 #define KFD_SIGNAL_EVENT_LIMIT 4096 261 262 /* For kfd_event_data.hw_exception_data.reset_type. */ 263 #define KFD_HW_EXCEPTION_WHOLE_GPU_RESET 0 264 #define KFD_HW_EXCEPTION_PER_ENGINE_RESET 1 265 266 /* For kfd_event_data.hw_exception_data.reset_cause. */ 267 #define KFD_HW_EXCEPTION_GPU_HANG 0 268 #define KFD_HW_EXCEPTION_ECC 1 269 270 /* For kfd_hsa_memory_exception_data.ErrorType */ 271 #define KFD_MEM_ERR_NO_RAS 0 272 #define KFD_MEM_ERR_SRAM_ECC 1 273 #define KFD_MEM_ERR_POISON_CONSUMED 2 274 #define KFD_MEM_ERR_GPU_HANG 3 275 276 struct kfd_ioctl_create_event_args { 277 __u64 event_page_offset; /* from KFD */ 278 __u32 event_trigger_data; /* from KFD - signal events only */ 279 __u32 event_type; /* to KFD */ 280 __u32 auto_reset; /* to KFD */ 281 __u32 node_id; /* to KFD - only valid for certain 282 event types */ 283 __u32 event_id; /* from KFD */ 284 __u32 event_slot_index; /* from KFD */ 285 }; 286 287 struct kfd_ioctl_destroy_event_args { 288 __u32 event_id; /* to KFD */ 289 __u32 pad; 290 }; 291 292 struct kfd_ioctl_set_event_args { 293 __u32 event_id; /* to KFD */ 294 __u32 pad; 295 }; 296 297 struct kfd_ioctl_reset_event_args { 298 __u32 event_id; /* to KFD */ 299 __u32 pad; 300 }; 301 302 struct kfd_memory_exception_failure { 303 __u32 NotPresent; /* Page not present or supervisor privilege */ 304 __u32 ReadOnly; /* Write access to a read-only page */ 305 __u32 NoExecute; /* Execute access to a page marked NX */ 306 __u32 imprecise; /* Can't determine the exact fault address */ 307 }; 308 309 /* memory exception data */ 310 struct kfd_hsa_memory_exception_data { 311 struct kfd_memory_exception_failure failure; 312 __u64 va; 313 __u32 gpu_id; 314 __u32 ErrorType; /* 0 = no RAS error, 315 * 1 = ECC_SRAM, 316 * 2 = Link_SYNFLOOD (poison), 317 * 3 = GPU hang (not attributable to a specific cause), 318 * other values reserved 319 */ 320 }; 321 322 /* hw exception data */ 323 struct kfd_hsa_hw_exception_data { 324 __u32 reset_type; 325 __u32 reset_cause; 326 __u32 memory_lost; 327 __u32 gpu_id; 328 }; 329 330 /* hsa signal event data */ 331 struct kfd_hsa_signal_event_data { 332 __u64 last_event_age; /* to and from KFD */ 333 }; 334 335 /* Event data */ 336 struct kfd_event_data { 337 union { 338 /* From KFD */ 339 struct kfd_hsa_memory_exception_data memory_exception_data; 340 struct kfd_hsa_hw_exception_data hw_exception_data; 341 /* To and From KFD */ 342 struct kfd_hsa_signal_event_data signal_event_data; 343 }; 344 __u64 kfd_event_data_ext; /* pointer to an extension structure 345 for future exception types */ 346 __u32 event_id; /* to KFD */ 347 __u32 pad; 348 }; 349 350 struct kfd_ioctl_wait_events_args { 351 __u64 events_ptr; /* pointed to struct 352 kfd_event_data array, to KFD */ 353 __u32 num_events; /* to KFD */ 354 __u32 wait_for_all; /* to KFD */ 355 __u32 timeout; /* to KFD */ 356 __u32 wait_result; /* from KFD */ 357 }; 358 359 struct kfd_ioctl_set_scratch_backing_va_args { 360 __u64 va_addr; /* to KFD */ 361 __u32 gpu_id; /* to KFD */ 362 __u32 pad; 363 }; 364 365 struct kfd_ioctl_get_tile_config_args { 366 /* to KFD: pointer to tile array */ 367 __u64 tile_config_ptr; 368 /* to KFD: pointer to macro tile array */ 369 __u64 macro_tile_config_ptr; 370 /* to KFD: array size allocated by user mode 371 * from KFD: array size filled by kernel 372 */ 373 __u32 num_tile_configs; 374 /* to KFD: array size allocated by user mode 375 * from KFD: array size filled by kernel 376 */ 377 __u32 num_macro_tile_configs; 378 379 __u32 gpu_id; /* to KFD */ 380 __u32 gb_addr_config; /* from KFD */ 381 __u32 num_banks; /* from KFD */ 382 __u32 num_ranks; /* from KFD */ 383 /* struct size can be extended later if needed 384 * without breaking ABI compatibility 385 */ 386 }; 387 388 struct kfd_ioctl_set_trap_handler_args { 389 __u64 tba_addr; /* to KFD */ 390 __u64 tma_addr; /* to KFD */ 391 __u32 gpu_id; /* to KFD */ 392 __u32 pad; 393 }; 394 395 struct kfd_ioctl_acquire_vm_args { 396 __u32 drm_fd; /* to KFD */ 397 __u32 gpu_id; /* to KFD */ 398 }; 399 400 /* Allocation flags: memory types */ 401 #define KFD_IOC_ALLOC_MEM_FLAGS_VRAM (1 << 0) 402 #define KFD_IOC_ALLOC_MEM_FLAGS_GTT (1 << 1) 403 #define KFD_IOC_ALLOC_MEM_FLAGS_USERPTR (1 << 2) 404 #define KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL (1 << 3) 405 #define KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP (1 << 4) 406 /* Allocation flags: attributes/access options */ 407 #define KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE (1 << 31) 408 #define KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE (1 << 30) 409 #define KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC (1 << 29) 410 #define KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE (1 << 28) 411 #define KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM (1 << 27) 412 #define KFD_IOC_ALLOC_MEM_FLAGS_COHERENT (1 << 26) 413 #define KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED (1 << 25) 414 #define KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT (1 << 24) 415 #define KFD_IOC_ALLOC_MEM_FLAGS_CONTIGUOUS (1 << 23) 416 417 /* Allocate memory for later SVM (shared virtual memory) mapping. 418 * 419 * @va_addr: virtual address of the memory to be allocated 420 * all later mappings on all GPUs will use this address 421 * @size: size in bytes 422 * @handle: buffer handle returned to user mode, used to refer to 423 * this allocation for mapping, unmapping and freeing 424 * @mmap_offset: for CPU-mapping the allocation by mmapping a render node 425 * for userptrs this is overloaded to specify the CPU address 426 * @gpu_id: device identifier 427 * @flags: memory type and attributes. See KFD_IOC_ALLOC_MEM_FLAGS above 428 */ 429 struct kfd_ioctl_alloc_memory_of_gpu_args { 430 __u64 va_addr; /* to KFD */ 431 __u64 size; /* to KFD */ 432 __u64 handle; /* from KFD */ 433 __u64 mmap_offset; /* to KFD (userptr), from KFD (mmap offset) */ 434 __u32 gpu_id; /* to KFD */ 435 __u32 flags; 436 }; 437 438 /* Free memory allocated with kfd_ioctl_alloc_memory_of_gpu 439 * 440 * @handle: memory handle returned by alloc 441 */ 442 struct kfd_ioctl_free_memory_of_gpu_args { 443 __u64 handle; /* to KFD */ 444 }; 445 446 /* Map memory to one or more GPUs 447 * 448 * @handle: memory handle returned by alloc 449 * @device_ids_array_ptr: array of gpu_ids (__u32 per device) 450 * @n_devices: number of devices in the array 451 * @n_success: number of devices mapped successfully 452 * 453 * @n_success returns information to the caller how many devices from 454 * the start of the array have mapped the buffer successfully. It can 455 * be passed into a subsequent retry call to skip those devices. For 456 * the first call the caller should initialize it to 0. 457 * 458 * If the ioctl completes with return code 0 (success), n_success == 459 * n_devices. 460 */ 461 struct kfd_ioctl_map_memory_to_gpu_args { 462 __u64 handle; /* to KFD */ 463 __u64 device_ids_array_ptr; /* to KFD */ 464 __u32 n_devices; /* to KFD */ 465 __u32 n_success; /* to/from KFD */ 466 }; 467 468 /* Unmap memory from one or more GPUs 469 * 470 * same arguments as for mapping 471 */ 472 struct kfd_ioctl_unmap_memory_from_gpu_args { 473 __u64 handle; /* to KFD */ 474 __u64 device_ids_array_ptr; /* to KFD */ 475 __u32 n_devices; /* to KFD */ 476 __u32 n_success; /* to/from KFD */ 477 }; 478 479 /* Allocate GWS for specific queue 480 * 481 * @queue_id: queue's id that GWS is allocated for 482 * @num_gws: how many GWS to allocate 483 * @first_gws: index of the first GWS allocated. 484 * only support contiguous GWS allocation 485 */ 486 struct kfd_ioctl_alloc_queue_gws_args { 487 __u32 queue_id; /* to KFD */ 488 __u32 num_gws; /* to KFD */ 489 __u32 first_gws; /* from KFD */ 490 __u32 pad; 491 }; 492 493 struct kfd_ioctl_get_dmabuf_info_args { 494 __u64 size; /* from KFD */ 495 __u64 metadata_ptr; /* to KFD */ 496 __u32 metadata_size; /* to KFD (space allocated by user) 497 * from KFD (actual metadata size) 498 */ 499 __u32 gpu_id; /* from KFD */ 500 __u32 flags; /* from KFD (KFD_IOC_ALLOC_MEM_FLAGS) */ 501 __u32 dmabuf_fd; /* to KFD */ 502 }; 503 504 struct kfd_ioctl_import_dmabuf_args { 505 __u64 va_addr; /* to KFD */ 506 __u64 handle; /* from KFD */ 507 __u32 gpu_id; /* to KFD */ 508 __u32 dmabuf_fd; /* to KFD */ 509 }; 510 511 struct kfd_ioctl_export_dmabuf_args { 512 __u64 handle; /* to KFD */ 513 __u32 flags; /* to KFD */ 514 __u32 dmabuf_fd; /* from KFD */ 515 }; 516 517 /* 518 * KFD SMI(System Management Interface) events 519 */ 520 enum kfd_smi_event { 521 KFD_SMI_EVENT_NONE = 0, /* not used */ 522 KFD_SMI_EVENT_VMFAULT = 1, /* event start counting at 1 */ 523 KFD_SMI_EVENT_THERMAL_THROTTLE = 2, 524 KFD_SMI_EVENT_GPU_PRE_RESET = 3, 525 KFD_SMI_EVENT_GPU_POST_RESET = 4, 526 KFD_SMI_EVENT_MIGRATE_START = 5, 527 KFD_SMI_EVENT_MIGRATE_END = 6, 528 KFD_SMI_EVENT_PAGE_FAULT_START = 7, 529 KFD_SMI_EVENT_PAGE_FAULT_END = 8, 530 KFD_SMI_EVENT_QUEUE_EVICTION = 9, 531 KFD_SMI_EVENT_QUEUE_RESTORE = 10, 532 KFD_SMI_EVENT_UNMAP_FROM_GPU = 11, 533 534 /* 535 * max event number, as a flag bit to get events from all processes, 536 * this requires super user permission, otherwise will not be able to 537 * receive event from any process. Without this flag to receive events 538 * from same process. 539 */ 540 KFD_SMI_EVENT_ALL_PROCESS = 64 541 }; 542 543 /* The reason of the page migration event */ 544 enum KFD_MIGRATE_TRIGGERS { 545 KFD_MIGRATE_TRIGGER_PREFETCH, /* Prefetch to GPU VRAM or system memory */ 546 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, /* GPU page fault recover */ 547 KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU, /* CPU page fault recover */ 548 KFD_MIGRATE_TRIGGER_TTM_EVICTION /* TTM eviction */ 549 }; 550 551 /* The reason of user queue evition event */ 552 enum KFD_QUEUE_EVICTION_TRIGGERS { 553 KFD_QUEUE_EVICTION_TRIGGER_SVM, /* SVM buffer migration */ 554 KFD_QUEUE_EVICTION_TRIGGER_USERPTR, /* userptr movement */ 555 KFD_QUEUE_EVICTION_TRIGGER_TTM, /* TTM move buffer */ 556 KFD_QUEUE_EVICTION_TRIGGER_SUSPEND, /* GPU suspend */ 557 KFD_QUEUE_EVICTION_CRIU_CHECKPOINT, /* CRIU checkpoint */ 558 KFD_QUEUE_EVICTION_CRIU_RESTORE /* CRIU restore */ 559 }; 560 561 /* The reason of unmap buffer from GPU event */ 562 enum KFD_SVM_UNMAP_TRIGGERS { 563 KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY, /* MMU notifier CPU buffer movement */ 564 KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE,/* MMU notifier page migration */ 565 KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU /* Unmap to free the buffer */ 566 }; 567 568 #define KFD_SMI_EVENT_MASK_FROM_INDEX(i) (1ULL << ((i) - 1)) 569 #define KFD_SMI_EVENT_MSG_SIZE 96 570 571 struct kfd_ioctl_smi_events_args { 572 __u32 gpuid; /* to KFD */ 573 __u32 anon_fd; /* from KFD */ 574 }; 575 576 /* 577 * SVM event tracing via SMI system management interface 578 * 579 * Open event file descriptor 580 * use ioctl AMDKFD_IOC_SMI_EVENTS, pass in gpuid and return a anonymous file 581 * descriptor to receive SMI events. 582 * If calling with sudo permission, then file descriptor can be used to receive 583 * SVM events from all processes, otherwise, to only receive SVM events of same 584 * process. 585 * 586 * To enable the SVM event 587 * Write event file descriptor with KFD_SMI_EVENT_MASK_FROM_INDEX(event) bitmap 588 * mask to start record the event to the kfifo, use bitmap mask combination 589 * for multiple events. New event mask will overwrite the previous event mask. 590 * KFD_SMI_EVENT_MASK_FROM_INDEX(KFD_SMI_EVENT_ALL_PROCESS) bit requires sudo 591 * permisson to receive SVM events from all process. 592 * 593 * To receive the event 594 * Application can poll file descriptor to wait for the events, then read event 595 * from the file into a buffer. Each event is one line string message, starting 596 * with the event id, then the event specific information. 597 * 598 * To decode event information 599 * The following event format string macro can be used with sscanf to decode 600 * the specific event information. 601 * event triggers: the reason to generate the event, defined as enum for unmap, 602 * eviction and migrate events. 603 * node, from, to, prefetch_loc, preferred_loc: GPU ID, or 0 for system memory. 604 * addr: user mode address, in pages 605 * size: in pages 606 * pid: the process ID to generate the event 607 * ns: timestamp in nanosecond-resolution, starts at system boot time but 608 * stops during suspend 609 * migrate_update: GPU page fault is recovered by 'M' for migrate, 'U' for update 610 * rw: 'W' for write page fault, 'R' for read page fault 611 * rescheduled: 'R' if the queue restore failed and rescheduled to try again 612 * error_code: migrate failure error code, 0 if no error 613 */ 614 #define KFD_EVENT_FMT_UPDATE_GPU_RESET(reset_seq_num, reset_cause)\ 615 "%x %s\n", (reset_seq_num), (reset_cause) 616 617 #define KFD_EVENT_FMT_THERMAL_THROTTLING(bitmask, counter)\ 618 "%llx:%llx\n", (bitmask), (counter) 619 620 #define KFD_EVENT_FMT_VMFAULT(pid, task_name)\ 621 "%x:%s\n", (pid), (task_name) 622 623 #define KFD_EVENT_FMT_PAGEFAULT_START(ns, pid, addr, node, rw)\ 624 "%lld -%d @%lx(%x) %c\n", (ns), (pid), (addr), (node), (rw) 625 626 #define KFD_EVENT_FMT_PAGEFAULT_END(ns, pid, addr, node, migrate_update)\ 627 "%lld -%d @%lx(%x) %c\n", (ns), (pid), (addr), (node), (migrate_update) 628 629 #define KFD_EVENT_FMT_MIGRATE_START(ns, pid, start, size, from, to, prefetch_loc,\ 630 preferred_loc, migrate_trigger)\ 631 "%lld -%d @%lx(%lx) %x->%x %x:%x %d\n", (ns), (pid), (start), (size),\ 632 (from), (to), (prefetch_loc), (preferred_loc), (migrate_trigger) 633 634 #define KFD_EVENT_FMT_MIGRATE_END(ns, pid, start, size, from, to, migrate_trigger, error_code) \ 635 "%lld -%d @%lx(%lx) %x->%x %d %d\n", (ns), (pid), (start), (size),\ 636 (from), (to), (migrate_trigger), (error_code) 637 638 #define KFD_EVENT_FMT_QUEUE_EVICTION(ns, pid, node, evict_trigger)\ 639 "%lld -%d %x %d\n", (ns), (pid), (node), (evict_trigger) 640 641 #define KFD_EVENT_FMT_QUEUE_RESTORE(ns, pid, node, rescheduled)\ 642 "%lld -%d %x %c\n", (ns), (pid), (node), (rescheduled) 643 644 #define KFD_EVENT_FMT_UNMAP_FROM_GPU(ns, pid, addr, size, node, unmap_trigger)\ 645 "%lld -%d @%lx(%lx) %x %d\n", (ns), (pid), (addr), (size),\ 646 (node), (unmap_trigger) 647 648 /************************************************************************************************** 649 * CRIU IOCTLs (Checkpoint Restore In Userspace) 650 * 651 * When checkpointing a process, the userspace application will perform: 652 * 1. PROCESS_INFO op to determine current process information. This pauses execution and evicts 653 * all the queues. 654 * 2. CHECKPOINT op to checkpoint process contents (BOs, queues, events, svm-ranges) 655 * 3. UNPAUSE op to un-evict all the queues 656 * 657 * When restoring a process, the CRIU userspace application will perform: 658 * 659 * 1. RESTORE op to restore process contents 660 * 2. RESUME op to start the process 661 * 662 * Note: Queues are forced into an evicted state after a successful PROCESS_INFO. User 663 * application needs to perform an UNPAUSE operation after calling PROCESS_INFO. 664 */ 665 666 enum kfd_criu_op { 667 KFD_CRIU_OP_PROCESS_INFO, 668 KFD_CRIU_OP_CHECKPOINT, 669 KFD_CRIU_OP_UNPAUSE, 670 KFD_CRIU_OP_RESTORE, 671 KFD_CRIU_OP_RESUME, 672 }; 673 674 /** 675 * kfd_ioctl_criu_args - Arguments perform CRIU operation 676 * @devices: [in/out] User pointer to memory location for devices information. 677 * This is an array of type kfd_criu_device_bucket. 678 * @bos: [in/out] User pointer to memory location for BOs information 679 * This is an array of type kfd_criu_bo_bucket. 680 * @priv_data: [in/out] User pointer to memory location for private data 681 * @priv_data_size: [in/out] Size of priv_data in bytes 682 * @num_devices: [in/out] Number of GPUs used by process. Size of @devices array. 683 * @num_bos [in/out] Number of BOs used by process. Size of @bos array. 684 * @num_objects: [in/out] Number of objects used by process. Objects are opaque to 685 * user application. 686 * @pid: [in/out] PID of the process being checkpointed 687 * @op [in] Type of operation (kfd_criu_op) 688 * 689 * Return: 0 on success, -errno on failure 690 */ 691 struct kfd_ioctl_criu_args { 692 __u64 devices; /* Used during ops: CHECKPOINT, RESTORE */ 693 __u64 bos; /* Used during ops: CHECKPOINT, RESTORE */ 694 __u64 priv_data; /* Used during ops: CHECKPOINT, RESTORE */ 695 __u64 priv_data_size; /* Used during ops: PROCESS_INFO, RESTORE */ 696 __u32 num_devices; /* Used during ops: PROCESS_INFO, RESTORE */ 697 __u32 num_bos; /* Used during ops: PROCESS_INFO, RESTORE */ 698 __u32 num_objects; /* Used during ops: PROCESS_INFO, RESTORE */ 699 __u32 pid; /* Used during ops: PROCESS_INFO, RESUME */ 700 __u32 op; 701 }; 702 703 struct kfd_criu_device_bucket { 704 __u32 user_gpu_id; 705 __u32 actual_gpu_id; 706 __u32 drm_fd; 707 __u32 pad; 708 }; 709 710 struct kfd_criu_bo_bucket { 711 __u64 addr; 712 __u64 size; 713 __u64 offset; 714 __u64 restored_offset; /* During restore, updated offset for BO */ 715 __u32 gpu_id; /* This is the user_gpu_id */ 716 __u32 alloc_flags; 717 __u32 dmabuf_fd; 718 __u32 pad; 719 }; 720 721 /* CRIU IOCTLs - END */ 722 /**************************************************************************************************/ 723 724 /* Register offset inside the remapped mmio page 725 */ 726 enum kfd_mmio_remap { 727 KFD_MMIO_REMAP_HDP_MEM_FLUSH_CNTL = 0, 728 KFD_MMIO_REMAP_HDP_REG_FLUSH_CNTL = 4, 729 }; 730 731 /* Guarantee host access to memory */ 732 #define KFD_IOCTL_SVM_FLAG_HOST_ACCESS 0x00000001 733 /* Fine grained coherency between all devices with access */ 734 #define KFD_IOCTL_SVM_FLAG_COHERENT 0x00000002 735 /* Use any GPU in same hive as preferred device */ 736 #define KFD_IOCTL_SVM_FLAG_HIVE_LOCAL 0x00000004 737 /* GPUs only read, allows replication */ 738 #define KFD_IOCTL_SVM_FLAG_GPU_RO 0x00000008 739 /* Allow execution on GPU */ 740 #define KFD_IOCTL_SVM_FLAG_GPU_EXEC 0x00000010 741 /* GPUs mostly read, may allow similar optimizations as RO, but writes fault */ 742 #define KFD_IOCTL_SVM_FLAG_GPU_READ_MOSTLY 0x00000020 743 /* Keep GPU memory mapping always valid as if XNACK is disable */ 744 #define KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED 0x00000040 745 /* Fine grained coherency between all devices using device-scope atomics */ 746 #define KFD_IOCTL_SVM_FLAG_EXT_COHERENT 0x00000080 747 748 /** 749 * kfd_ioctl_svm_op - SVM ioctl operations 750 * 751 * @KFD_IOCTL_SVM_OP_SET_ATTR: Modify one or more attributes 752 * @KFD_IOCTL_SVM_OP_GET_ATTR: Query one or more attributes 753 */ 754 enum kfd_ioctl_svm_op { 755 KFD_IOCTL_SVM_OP_SET_ATTR, 756 KFD_IOCTL_SVM_OP_GET_ATTR 757 }; 758 759 /** kfd_ioctl_svm_location - Enum for preferred and prefetch locations 760 * 761 * GPU IDs are used to specify GPUs as preferred and prefetch locations. 762 * Below definitions are used for system memory or for leaving the preferred 763 * location unspecified. 764 */ 765 enum kfd_ioctl_svm_location { 766 KFD_IOCTL_SVM_LOCATION_SYSMEM = 0, 767 KFD_IOCTL_SVM_LOCATION_UNDEFINED = 0xffffffff 768 }; 769 770 /** 771 * kfd_ioctl_svm_attr_type - SVM attribute types 772 * 773 * @KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: gpuid of the preferred location, 0 for 774 * system memory 775 * @KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: gpuid of the prefetch location, 0 for 776 * system memory. Setting this triggers an 777 * immediate prefetch (migration). 778 * @KFD_IOCTL_SVM_ATTR_ACCESS: 779 * @KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 780 * @KFD_IOCTL_SVM_ATTR_NO_ACCESS: specify memory access for the gpuid given 781 * by the attribute value 782 * @KFD_IOCTL_SVM_ATTR_SET_FLAGS: bitmask of flags to set (see 783 * KFD_IOCTL_SVM_FLAG_...) 784 * @KFD_IOCTL_SVM_ATTR_CLR_FLAGS: bitmask of flags to clear 785 * @KFD_IOCTL_SVM_ATTR_GRANULARITY: migration granularity 786 * (log2 num pages) 787 */ 788 enum kfd_ioctl_svm_attr_type { 789 KFD_IOCTL_SVM_ATTR_PREFERRED_LOC, 790 KFD_IOCTL_SVM_ATTR_PREFETCH_LOC, 791 KFD_IOCTL_SVM_ATTR_ACCESS, 792 KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE, 793 KFD_IOCTL_SVM_ATTR_NO_ACCESS, 794 KFD_IOCTL_SVM_ATTR_SET_FLAGS, 795 KFD_IOCTL_SVM_ATTR_CLR_FLAGS, 796 KFD_IOCTL_SVM_ATTR_GRANULARITY 797 }; 798 799 /** 800 * kfd_ioctl_svm_attribute - Attributes as pairs of type and value 801 * 802 * The meaning of the @value depends on the attribute type. 803 * 804 * @type: attribute type (see enum @kfd_ioctl_svm_attr_type) 805 * @value: attribute value 806 */ 807 struct kfd_ioctl_svm_attribute { 808 __u32 type; 809 __u32 value; 810 }; 811 812 /** 813 * kfd_ioctl_svm_args - Arguments for SVM ioctl 814 * 815 * @op specifies the operation to perform (see enum 816 * @kfd_ioctl_svm_op). @start_addr and @size are common for all 817 * operations. 818 * 819 * A variable number of attributes can be given in @attrs. 820 * @nattr specifies the number of attributes. New attributes can be 821 * added in the future without breaking the ABI. If unknown attributes 822 * are given, the function returns -EINVAL. 823 * 824 * @KFD_IOCTL_SVM_OP_SET_ATTR sets attributes for a virtual address 825 * range. It may overlap existing virtual address ranges. If it does, 826 * the existing ranges will be split such that the attribute changes 827 * only apply to the specified address range. 828 * 829 * @KFD_IOCTL_SVM_OP_GET_ATTR returns the intersection of attributes 830 * over all memory in the given range and returns the result as the 831 * attribute value. If different pages have different preferred or 832 * prefetch locations, 0xffffffff will be returned for 833 * @KFD_IOCTL_SVM_ATTR_PREFERRED_LOC or 834 * @KFD_IOCTL_SVM_ATTR_PREFETCH_LOC resepctively. For 835 * @KFD_IOCTL_SVM_ATTR_SET_FLAGS, flags of all pages will be 836 * aggregated by bitwise AND. That means, a flag will be set in the 837 * output, if that flag is set for all pages in the range. For 838 * @KFD_IOCTL_SVM_ATTR_CLR_FLAGS, flags of all pages will be 839 * aggregated by bitwise NOR. That means, a flag will be set in the 840 * output, if that flag is clear for all pages in the range. 841 * The minimum migration granularity throughout the range will be 842 * returned for @KFD_IOCTL_SVM_ATTR_GRANULARITY. 843 * 844 * Querying of accessibility attributes works by initializing the 845 * attribute type to @KFD_IOCTL_SVM_ATTR_ACCESS and the value to the 846 * GPUID being queried. Multiple attributes can be given to allow 847 * querying multiple GPUIDs. The ioctl function overwrites the 848 * attribute type to indicate the access for the specified GPU. 849 */ 850 struct kfd_ioctl_svm_args { 851 __u64 start_addr; 852 __u64 size; 853 __u32 op; 854 __u32 nattr; 855 /* Variable length array of attributes */ 856 struct kfd_ioctl_svm_attribute attrs[]; 857 }; 858 859 /** 860 * kfd_ioctl_set_xnack_mode_args - Arguments for set_xnack_mode 861 * 862 * @xnack_enabled: [in/out] Whether to enable XNACK mode for this process 863 * 864 * @xnack_enabled indicates whether recoverable page faults should be 865 * enabled for the current process. 0 means disabled, positive means 866 * enabled, negative means leave unchanged. If enabled, virtual address 867 * translations on GFXv9 and later AMD GPUs can return XNACK and retry 868 * the access until a valid PTE is available. This is used to implement 869 * device page faults. 870 * 871 * On output, @xnack_enabled returns the (new) current mode (0 or 872 * positive). Therefore, a negative input value can be used to query 873 * the current mode without changing it. 874 * 875 * The XNACK mode fundamentally changes the way SVM managed memory works 876 * in the driver, with subtle effects on application performance and 877 * functionality. 878 * 879 * Enabling XNACK mode requires shader programs to be compiled 880 * differently. Furthermore, not all GPUs support changing the mode 881 * per-process. Therefore changing the mode is only allowed while no 882 * user mode queues exist in the process. This ensure that no shader 883 * code is running that may be compiled for the wrong mode. And GPUs 884 * that cannot change to the requested mode will prevent the XNACK 885 * mode from occurring. All GPUs used by the process must be in the 886 * same XNACK mode. 887 * 888 * GFXv8 or older GPUs do not support 48 bit virtual addresses or SVM. 889 * Therefore those GPUs are not considered for the XNACK mode switch. 890 * 891 * Return: 0 on success, -errno on failure 892 */ 893 struct kfd_ioctl_set_xnack_mode_args { 894 __s32 xnack_enabled; 895 }; 896 897 /* Wave launch override modes */ 898 enum kfd_dbg_trap_override_mode { 899 KFD_DBG_TRAP_OVERRIDE_OR = 0, 900 KFD_DBG_TRAP_OVERRIDE_REPLACE = 1 901 }; 902 903 /* Wave launch overrides */ 904 enum kfd_dbg_trap_mask { 905 KFD_DBG_TRAP_MASK_FP_INVALID = 1, 906 KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL = 2, 907 KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO = 4, 908 KFD_DBG_TRAP_MASK_FP_OVERFLOW = 8, 909 KFD_DBG_TRAP_MASK_FP_UNDERFLOW = 16, 910 KFD_DBG_TRAP_MASK_FP_INEXACT = 32, 911 KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO = 64, 912 KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH = 128, 913 KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION = 256, 914 KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START = (1 << 30), 915 KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END = (1 << 31) 916 }; 917 918 /* Wave launch modes */ 919 enum kfd_dbg_trap_wave_launch_mode { 920 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_NORMAL = 0, 921 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_HALT = 1, 922 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_DEBUG = 3 923 }; 924 925 /* Address watch modes */ 926 enum kfd_dbg_trap_address_watch_mode { 927 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_READ = 0, 928 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_NONREAD = 1, 929 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_ATOMIC = 2, 930 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_ALL = 3 931 }; 932 933 /* Additional wave settings */ 934 enum kfd_dbg_trap_flags { 935 KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP = 1, 936 KFD_DBG_TRAP_FLAG_SINGLE_ALU_OP = 2, 937 }; 938 939 /* Trap exceptions */ 940 enum kfd_dbg_trap_exception_code { 941 EC_NONE = 0, 942 /* per queue */ 943 EC_QUEUE_WAVE_ABORT = 1, 944 EC_QUEUE_WAVE_TRAP = 2, 945 EC_QUEUE_WAVE_MATH_ERROR = 3, 946 EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION = 4, 947 EC_QUEUE_WAVE_MEMORY_VIOLATION = 5, 948 EC_QUEUE_WAVE_APERTURE_VIOLATION = 6, 949 EC_QUEUE_PACKET_DISPATCH_DIM_INVALID = 16, 950 EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID = 17, 951 EC_QUEUE_PACKET_DISPATCH_CODE_INVALID = 18, 952 EC_QUEUE_PACKET_RESERVED = 19, 953 EC_QUEUE_PACKET_UNSUPPORTED = 20, 954 EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID = 21, 955 EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID = 22, 956 EC_QUEUE_PACKET_VENDOR_UNSUPPORTED = 23, 957 EC_QUEUE_PREEMPTION_ERROR = 30, 958 EC_QUEUE_NEW = 31, 959 /* per device */ 960 EC_DEVICE_QUEUE_DELETE = 32, 961 EC_DEVICE_MEMORY_VIOLATION = 33, 962 EC_DEVICE_RAS_ERROR = 34, 963 EC_DEVICE_FATAL_HALT = 35, 964 EC_DEVICE_NEW = 36, 965 /* per process */ 966 EC_PROCESS_RUNTIME = 48, 967 EC_PROCESS_DEVICE_REMOVE = 49, 968 EC_MAX 969 }; 970 971 /* Mask generated by ecode in kfd_dbg_trap_exception_code */ 972 #define KFD_EC_MASK(ecode) (1ULL << (ecode - 1)) 973 974 /* Masks for exception code type checks below */ 975 #define KFD_EC_MASK_QUEUE (KFD_EC_MASK(EC_QUEUE_WAVE_ABORT) | \ 976 KFD_EC_MASK(EC_QUEUE_WAVE_TRAP) | \ 977 KFD_EC_MASK(EC_QUEUE_WAVE_MATH_ERROR) | \ 978 KFD_EC_MASK(EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION) | \ 979 KFD_EC_MASK(EC_QUEUE_WAVE_MEMORY_VIOLATION) | \ 980 KFD_EC_MASK(EC_QUEUE_WAVE_APERTURE_VIOLATION) | \ 981 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_DIM_INVALID) | \ 982 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID) | \ 983 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_CODE_INVALID) | \ 984 KFD_EC_MASK(EC_QUEUE_PACKET_RESERVED) | \ 985 KFD_EC_MASK(EC_QUEUE_PACKET_UNSUPPORTED) | \ 986 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) | \ 987 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) | \ 988 KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED) | \ 989 KFD_EC_MASK(EC_QUEUE_PREEMPTION_ERROR) | \ 990 KFD_EC_MASK(EC_QUEUE_NEW)) 991 #define KFD_EC_MASK_DEVICE (KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE) | \ 992 KFD_EC_MASK(EC_DEVICE_RAS_ERROR) | \ 993 KFD_EC_MASK(EC_DEVICE_FATAL_HALT) | \ 994 KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION) | \ 995 KFD_EC_MASK(EC_DEVICE_NEW)) 996 #define KFD_EC_MASK_PROCESS (KFD_EC_MASK(EC_PROCESS_RUNTIME) | \ 997 KFD_EC_MASK(EC_PROCESS_DEVICE_REMOVE)) 998 #define KFD_EC_MASK_PACKET (KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_DIM_INVALID) | \ 999 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID) | \ 1000 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_CODE_INVALID) | \ 1001 KFD_EC_MASK(EC_QUEUE_PACKET_RESERVED) | \ 1002 KFD_EC_MASK(EC_QUEUE_PACKET_UNSUPPORTED) | \ 1003 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) | \ 1004 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) | \ 1005 KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED)) 1006 1007 /* Checks for exception code types for KFD search */ 1008 #define KFD_DBG_EC_IS_VALID(ecode) (ecode > EC_NONE && ecode < EC_MAX) 1009 #define KFD_DBG_EC_TYPE_IS_QUEUE(ecode) \ 1010 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_QUEUE)) 1011 #define KFD_DBG_EC_TYPE_IS_DEVICE(ecode) \ 1012 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_DEVICE)) 1013 #define KFD_DBG_EC_TYPE_IS_PROCESS(ecode) \ 1014 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_PROCESS)) 1015 #define KFD_DBG_EC_TYPE_IS_PACKET(ecode) \ 1016 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_PACKET)) 1017 1018 1019 /* Runtime enable states */ 1020 enum kfd_dbg_runtime_state { 1021 DEBUG_RUNTIME_STATE_DISABLED = 0, 1022 DEBUG_RUNTIME_STATE_ENABLED = 1, 1023 DEBUG_RUNTIME_STATE_ENABLED_BUSY = 2, 1024 DEBUG_RUNTIME_STATE_ENABLED_ERROR = 3 1025 }; 1026 1027 /* Runtime enable status */ 1028 struct kfd_runtime_info { 1029 __u64 r_debug; 1030 __u32 runtime_state; 1031 __u32 ttmp_setup; 1032 }; 1033 1034 /* Enable modes for runtime enable */ 1035 #define KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK 1 1036 #define KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK 2 1037 1038 /** 1039 * kfd_ioctl_runtime_enable_args - Arguments for runtime enable 1040 * 1041 * Coordinates debug exception signalling and debug device enablement with runtime. 1042 * 1043 * @r_debug - pointer to user struct for sharing information between ROCr and the debuggger 1044 * @mode_mask - mask to set mode 1045 * KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK - enable runtime for debugging, otherwise disable 1046 * KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK - enable trap temporary setup (ignore on disable) 1047 * @capabilities_mask - mask to notify runtime on what KFD supports 1048 * 1049 * Return - 0 on SUCCESS. 1050 * - EBUSY if runtime enable call already pending. 1051 * - EEXIST if user queues already active prior to call. 1052 * If process is debug enabled, runtime enable will enable debug devices and 1053 * wait for debugger process to send runtime exception EC_PROCESS_RUNTIME 1054 * to unblock - see kfd_ioctl_dbg_trap_args. 1055 * 1056 */ 1057 struct kfd_ioctl_runtime_enable_args { 1058 __u64 r_debug; 1059 __u32 mode_mask; 1060 __u32 capabilities_mask; 1061 }; 1062 1063 /* Queue information */ 1064 struct kfd_queue_snapshot_entry { 1065 __u64 exception_status; 1066 __u64 ring_base_address; 1067 __u64 write_pointer_address; 1068 __u64 read_pointer_address; 1069 __u64 ctx_save_restore_address; 1070 __u32 queue_id; 1071 __u32 gpu_id; 1072 __u32 ring_size; 1073 __u32 queue_type; 1074 __u32 ctx_save_restore_area_size; 1075 __u32 reserved; 1076 }; 1077 1078 /* Queue status return for suspend/resume */ 1079 #define KFD_DBG_QUEUE_ERROR_BIT 30 1080 #define KFD_DBG_QUEUE_INVALID_BIT 31 1081 #define KFD_DBG_QUEUE_ERROR_MASK (1 << KFD_DBG_QUEUE_ERROR_BIT) 1082 #define KFD_DBG_QUEUE_INVALID_MASK (1 << KFD_DBG_QUEUE_INVALID_BIT) 1083 1084 /* Context save area header information */ 1085 struct kfd_context_save_area_header { 1086 struct { 1087 __u32 control_stack_offset; 1088 __u32 control_stack_size; 1089 __u32 wave_state_offset; 1090 __u32 wave_state_size; 1091 } wave_state; 1092 __u32 debug_offset; 1093 __u32 debug_size; 1094 __u64 err_payload_addr; 1095 __u32 err_event_id; 1096 __u32 reserved1; 1097 }; 1098 1099 /* 1100 * Debug operations 1101 * 1102 * For specifics on usage and return values, see documentation per operation 1103 * below. Otherwise, generic error returns apply: 1104 * - ESRCH if the process to debug does not exist. 1105 * 1106 * - EINVAL (with KFD_IOC_DBG_TRAP_ENABLE exempt) if operation 1107 * KFD_IOC_DBG_TRAP_ENABLE has not succeeded prior. 1108 * Also returns this error if GPU hardware scheduling is not supported. 1109 * 1110 * - EPERM (with KFD_IOC_DBG_TRAP_DISABLE exempt) if target process is not 1111 * PTRACE_ATTACHED. KFD_IOC_DBG_TRAP_DISABLE is exempt to allow 1112 * clean up of debug mode as long as process is debug enabled. 1113 * 1114 * - EACCES if any DBG_HW_OP (debug hardware operation) is requested when 1115 * AMDKFD_IOC_RUNTIME_ENABLE has not succeeded prior. 1116 * 1117 * - ENODEV if any GPU does not support debugging on a DBG_HW_OP call. 1118 * 1119 * - Other errors may be returned when a DBG_HW_OP occurs while the GPU 1120 * is in a fatal state. 1121 * 1122 */ 1123 enum kfd_dbg_trap_operations { 1124 KFD_IOC_DBG_TRAP_ENABLE = 0, 1125 KFD_IOC_DBG_TRAP_DISABLE = 1, 1126 KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT = 2, 1127 KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED = 3, 1128 KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE = 4, /* DBG_HW_OP */ 1129 KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE = 5, /* DBG_HW_OP */ 1130 KFD_IOC_DBG_TRAP_SUSPEND_QUEUES = 6, /* DBG_HW_OP */ 1131 KFD_IOC_DBG_TRAP_RESUME_QUEUES = 7, /* DBG_HW_OP */ 1132 KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH = 8, /* DBG_HW_OP */ 1133 KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH = 9, /* DBG_HW_OP */ 1134 KFD_IOC_DBG_TRAP_SET_FLAGS = 10, 1135 KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT = 11, 1136 KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO = 12, 1137 KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT = 13, 1138 KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT = 14 1139 }; 1140 1141 /** 1142 * kfd_ioctl_dbg_trap_enable_args 1143 * 1144 * Arguments for KFD_IOC_DBG_TRAP_ENABLE. 1145 * 1146 * Enables debug session for target process. Call @op KFD_IOC_DBG_TRAP_DISABLE in 1147 * kfd_ioctl_dbg_trap_args to disable debug session. 1148 * 1149 * @exception_mask (IN) - exceptions to raise to the debugger 1150 * @rinfo_ptr (IN) - pointer to runtime info buffer (see kfd_runtime_info) 1151 * @rinfo_size (IN/OUT) - size of runtime info buffer in bytes 1152 * @dbg_fd (IN) - fd the KFD will nofify the debugger with of raised 1153 * exceptions set in exception_mask. 1154 * 1155 * Generic errors apply (see kfd_dbg_trap_operations). 1156 * Return - 0 on SUCCESS. 1157 * Copies KFD saved kfd_runtime_info to @rinfo_ptr on enable. 1158 * Size of kfd_runtime saved by the KFD returned to @rinfo_size. 1159 * - EBADF if KFD cannot get a reference to dbg_fd. 1160 * - EFAULT if KFD cannot copy runtime info to rinfo_ptr. 1161 * - EINVAL if target process is already debug enabled. 1162 * 1163 */ 1164 struct kfd_ioctl_dbg_trap_enable_args { 1165 __u64 exception_mask; 1166 __u64 rinfo_ptr; 1167 __u32 rinfo_size; 1168 __u32 dbg_fd; 1169 }; 1170 1171 /** 1172 * kfd_ioctl_dbg_trap_send_runtime_event_args 1173 * 1174 * 1175 * Arguments for KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT. 1176 * Raises exceptions to runtime. 1177 * 1178 * @exception_mask (IN) - exceptions to raise to runtime 1179 * @gpu_id (IN) - target device id 1180 * @queue_id (IN) - target queue id 1181 * 1182 * Generic errors apply (see kfd_dbg_trap_operations). 1183 * Return - 0 on SUCCESS. 1184 * - ENODEV if gpu_id not found. 1185 * If exception_mask contains EC_PROCESS_RUNTIME, unblocks pending 1186 * AMDKFD_IOC_RUNTIME_ENABLE call - see kfd_ioctl_runtime_enable_args. 1187 * All other exceptions are raised to runtime through err_payload_addr. 1188 * See kfd_context_save_area_header. 1189 */ 1190 struct kfd_ioctl_dbg_trap_send_runtime_event_args { 1191 __u64 exception_mask; 1192 __u32 gpu_id; 1193 __u32 queue_id; 1194 }; 1195 1196 /** 1197 * kfd_ioctl_dbg_trap_set_exceptions_enabled_args 1198 * 1199 * Arguments for KFD_IOC_SET_EXCEPTIONS_ENABLED 1200 * Set new exceptions to be raised to the debugger. 1201 * 1202 * @exception_mask (IN) - new exceptions to raise the debugger 1203 * 1204 * Generic errors apply (see kfd_dbg_trap_operations). 1205 * Return - 0 on SUCCESS. 1206 */ 1207 struct kfd_ioctl_dbg_trap_set_exceptions_enabled_args { 1208 __u64 exception_mask; 1209 }; 1210 1211 /** 1212 * kfd_ioctl_dbg_trap_set_wave_launch_override_args 1213 * 1214 * Arguments for KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE 1215 * Enable HW exceptions to raise trap. 1216 * 1217 * @override_mode (IN) - see kfd_dbg_trap_override_mode 1218 * @enable_mask (IN/OUT) - reference kfd_dbg_trap_mask. 1219 * IN is the override modes requested to be enabled. 1220 * OUT is referenced in Return below. 1221 * @support_request_mask (IN/OUT) - reference kfd_dbg_trap_mask. 1222 * IN is the override modes requested for support check. 1223 * OUT is referenced in Return below. 1224 * 1225 * Generic errors apply (see kfd_dbg_trap_operations). 1226 * Return - 0 on SUCCESS. 1227 * Previous enablement is returned in @enable_mask. 1228 * Actual override support is returned in @support_request_mask. 1229 * - EINVAL if override mode is not supported. 1230 * - EACCES if trap support requested is not actually supported. 1231 * i.e. enable_mask (IN) is not a subset of support_request_mask (OUT). 1232 * Otherwise it is considered a generic error (see kfd_dbg_trap_operations). 1233 */ 1234 struct kfd_ioctl_dbg_trap_set_wave_launch_override_args { 1235 __u32 override_mode; 1236 __u32 enable_mask; 1237 __u32 support_request_mask; 1238 __u32 pad; 1239 }; 1240 1241 /** 1242 * kfd_ioctl_dbg_trap_set_wave_launch_mode_args 1243 * 1244 * Arguments for KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE 1245 * Set wave launch mode. 1246 * 1247 * @mode (IN) - see kfd_dbg_trap_wave_launch_mode 1248 * 1249 * Generic errors apply (see kfd_dbg_trap_operations). 1250 * Return - 0 on SUCCESS. 1251 */ 1252 struct kfd_ioctl_dbg_trap_set_wave_launch_mode_args { 1253 __u32 launch_mode; 1254 __u32 pad; 1255 }; 1256 1257 /** 1258 * kfd_ioctl_dbg_trap_suspend_queues_ags 1259 * 1260 * Arguments for KFD_IOC_DBG_TRAP_SUSPEND_QUEUES 1261 * Suspend queues. 1262 * 1263 * @exception_mask (IN) - raised exceptions to clear 1264 * @queue_array_ptr (IN) - pointer to array of queue ids (u32 per queue id) 1265 * to suspend 1266 * @num_queues (IN) - number of queues to suspend in @queue_array_ptr 1267 * @grace_period (IN) - wave time allowance before preemption 1268 * per 1K GPU clock cycle unit 1269 * 1270 * Generic errors apply (see kfd_dbg_trap_operations). 1271 * Destruction of a suspended queue is blocked until the queue is 1272 * resumed. This allows the debugger to access queue information and 1273 * the its context save area without running into a race condition on 1274 * queue destruction. 1275 * Automatically copies per queue context save area header information 1276 * into the save area base 1277 * (see kfd_queue_snapshot_entry and kfd_context_save_area_header). 1278 * 1279 * Return - Number of queues suspended on SUCCESS. 1280 * . KFD_DBG_QUEUE_ERROR_MASK and KFD_DBG_QUEUE_INVALID_MASK masked 1281 * for each queue id in @queue_array_ptr array reports unsuccessful 1282 * suspend reason. 1283 * KFD_DBG_QUEUE_ERROR_MASK = HW failure. 1284 * KFD_DBG_QUEUE_INVALID_MASK = queue does not exist, is new or 1285 * is being destroyed. 1286 */ 1287 struct kfd_ioctl_dbg_trap_suspend_queues_args { 1288 __u64 exception_mask; 1289 __u64 queue_array_ptr; 1290 __u32 num_queues; 1291 __u32 grace_period; 1292 }; 1293 1294 /** 1295 * kfd_ioctl_dbg_trap_resume_queues_args 1296 * 1297 * Arguments for KFD_IOC_DBG_TRAP_RESUME_QUEUES 1298 * Resume queues. 1299 * 1300 * @queue_array_ptr (IN) - pointer to array of queue ids (u32 per queue id) 1301 * to resume 1302 * @num_queues (IN) - number of queues to resume in @queue_array_ptr 1303 * 1304 * Generic errors apply (see kfd_dbg_trap_operations). 1305 * Return - Number of queues resumed on SUCCESS. 1306 * KFD_DBG_QUEUE_ERROR_MASK and KFD_DBG_QUEUE_INVALID_MASK mask 1307 * for each queue id in @queue_array_ptr array reports unsuccessful 1308 * resume reason. 1309 * KFD_DBG_QUEUE_ERROR_MASK = HW failure. 1310 * KFD_DBG_QUEUE_INVALID_MASK = queue does not exist. 1311 */ 1312 struct kfd_ioctl_dbg_trap_resume_queues_args { 1313 __u64 queue_array_ptr; 1314 __u32 num_queues; 1315 __u32 pad; 1316 }; 1317 1318 /** 1319 * kfd_ioctl_dbg_trap_set_node_address_watch_args 1320 * 1321 * Arguments for KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH 1322 * Sets address watch for device. 1323 * 1324 * @address (IN) - watch address to set 1325 * @mode (IN) - see kfd_dbg_trap_address_watch_mode 1326 * @mask (IN) - watch address mask 1327 * @gpu_id (IN) - target gpu to set watch point 1328 * @id (OUT) - watch id allocated 1329 * 1330 * Generic errors apply (see kfd_dbg_trap_operations). 1331 * Return - 0 on SUCCESS. 1332 * Allocated watch ID returned to @id. 1333 * - ENODEV if gpu_id not found. 1334 * - ENOMEM if watch IDs can be allocated 1335 */ 1336 struct kfd_ioctl_dbg_trap_set_node_address_watch_args { 1337 __u64 address; 1338 __u32 mode; 1339 __u32 mask; 1340 __u32 gpu_id; 1341 __u32 id; 1342 }; 1343 1344 /** 1345 * kfd_ioctl_dbg_trap_clear_node_address_watch_args 1346 * 1347 * Arguments for KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH 1348 * Clear address watch for device. 1349 * 1350 * @gpu_id (IN) - target device to clear watch point 1351 * @id (IN) - allocated watch id to clear 1352 * 1353 * Generic errors apply (see kfd_dbg_trap_operations). 1354 * Return - 0 on SUCCESS. 1355 * - ENODEV if gpu_id not found. 1356 * - EINVAL if watch ID has not been allocated. 1357 */ 1358 struct kfd_ioctl_dbg_trap_clear_node_address_watch_args { 1359 __u32 gpu_id; 1360 __u32 id; 1361 }; 1362 1363 /** 1364 * kfd_ioctl_dbg_trap_set_flags_args 1365 * 1366 * Arguments for KFD_IOC_DBG_TRAP_SET_FLAGS 1367 * Sets flags for wave behaviour. 1368 * 1369 * @flags (IN/OUT) - IN = flags to enable, OUT = flags previously enabled 1370 * 1371 * Generic errors apply (see kfd_dbg_trap_operations). 1372 * Return - 0 on SUCCESS. 1373 * - EACCESS if any debug device does not allow flag options. 1374 */ 1375 struct kfd_ioctl_dbg_trap_set_flags_args { 1376 __u32 flags; 1377 __u32 pad; 1378 }; 1379 1380 /** 1381 * kfd_ioctl_dbg_trap_query_debug_event_args 1382 * 1383 * Arguments for KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT 1384 * 1385 * Find one or more raised exceptions. This function can return multiple 1386 * exceptions from a single queue or a single device with one call. To find 1387 * all raised exceptions, this function must be called repeatedly until it 1388 * returns -EAGAIN. Returned exceptions can optionally be cleared by 1389 * setting the corresponding bit in the @exception_mask input parameter. 1390 * However, clearing an exception prevents retrieving further information 1391 * about it with KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO. 1392 * 1393 * @exception_mask (IN/OUT) - exception to clear (IN) and raised (OUT) 1394 * @gpu_id (OUT) - gpu id of exceptions raised 1395 * @queue_id (OUT) - queue id of exceptions raised 1396 * 1397 * Generic errors apply (see kfd_dbg_trap_operations). 1398 * Return - 0 on raised exception found 1399 * Raised exceptions found are returned in @exception mask 1400 * with reported source id returned in @gpu_id or @queue_id. 1401 * - EAGAIN if no raised exception has been found 1402 */ 1403 struct kfd_ioctl_dbg_trap_query_debug_event_args { 1404 __u64 exception_mask; 1405 __u32 gpu_id; 1406 __u32 queue_id; 1407 }; 1408 1409 /** 1410 * kfd_ioctl_dbg_trap_query_exception_info_args 1411 * 1412 * Arguments KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO 1413 * Get additional info on raised exception. 1414 * 1415 * @info_ptr (IN) - pointer to exception info buffer to copy to 1416 * @info_size (IN/OUT) - exception info buffer size (bytes) 1417 * @source_id (IN) - target gpu or queue id 1418 * @exception_code (IN) - target exception 1419 * @clear_exception (IN) - clear raised @exception_code exception 1420 * (0 = false, 1 = true) 1421 * 1422 * Generic errors apply (see kfd_dbg_trap_operations). 1423 * Return - 0 on SUCCESS. 1424 * If @exception_code is EC_DEVICE_MEMORY_VIOLATION, copy @info_size(OUT) 1425 * bytes of memory exception data to @info_ptr. 1426 * If @exception_code is EC_PROCESS_RUNTIME, copy saved 1427 * kfd_runtime_info to @info_ptr. 1428 * Actual required @info_ptr size (bytes) is returned in @info_size. 1429 */ 1430 struct kfd_ioctl_dbg_trap_query_exception_info_args { 1431 __u64 info_ptr; 1432 __u32 info_size; 1433 __u32 source_id; 1434 __u32 exception_code; 1435 __u32 clear_exception; 1436 }; 1437 1438 /** 1439 * kfd_ioctl_dbg_trap_get_queue_snapshot_args 1440 * 1441 * Arguments KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT 1442 * Get queue information. 1443 * 1444 * @exception_mask (IN) - exceptions raised to clear 1445 * @snapshot_buf_ptr (IN) - queue snapshot entry buffer (see kfd_queue_snapshot_entry) 1446 * @num_queues (IN/OUT) - number of queue snapshot entries 1447 * The debugger specifies the size of the array allocated in @num_queues. 1448 * KFD returns the number of queues that actually existed. If this is 1449 * larger than the size specified by the debugger, KFD will not overflow 1450 * the array allocated by the debugger. 1451 * 1452 * @entry_size (IN/OUT) - size per entry in bytes 1453 * The debugger specifies sizeof(struct kfd_queue_snapshot_entry) in 1454 * @entry_size. KFD returns the number of bytes actually populated per 1455 * entry. The debugger should use the KFD_IOCTL_MINOR_VERSION to determine, 1456 * which fields in struct kfd_queue_snapshot_entry are valid. This allows 1457 * growing the ABI in a backwards compatible manner. 1458 * Note that entry_size(IN) should still be used to stride the snapshot buffer in the 1459 * event that it's larger than actual kfd_queue_snapshot_entry. 1460 * 1461 * Generic errors apply (see kfd_dbg_trap_operations). 1462 * Return - 0 on SUCCESS. 1463 * Copies @num_queues(IN) queue snapshot entries of size @entry_size(IN) 1464 * into @snapshot_buf_ptr if @num_queues(IN) > 0. 1465 * Otherwise return @num_queues(OUT) queue snapshot entries that exist. 1466 */ 1467 struct kfd_ioctl_dbg_trap_queue_snapshot_args { 1468 __u64 exception_mask; 1469 __u64 snapshot_buf_ptr; 1470 __u32 num_queues; 1471 __u32 entry_size; 1472 }; 1473 1474 /** 1475 * kfd_ioctl_dbg_trap_get_device_snapshot_args 1476 * 1477 * Arguments for KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT 1478 * Get device information. 1479 * 1480 * @exception_mask (IN) - exceptions raised to clear 1481 * @snapshot_buf_ptr (IN) - pointer to snapshot buffer (see kfd_dbg_device_info_entry) 1482 * @num_devices (IN/OUT) - number of debug devices to snapshot 1483 * The debugger specifies the size of the array allocated in @num_devices. 1484 * KFD returns the number of devices that actually existed. If this is 1485 * larger than the size specified by the debugger, KFD will not overflow 1486 * the array allocated by the debugger. 1487 * 1488 * @entry_size (IN/OUT) - size per entry in bytes 1489 * The debugger specifies sizeof(struct kfd_dbg_device_info_entry) in 1490 * @entry_size. KFD returns the number of bytes actually populated. The 1491 * debugger should use KFD_IOCTL_MINOR_VERSION to determine, which fields 1492 * in struct kfd_dbg_device_info_entry are valid. This allows growing the 1493 * ABI in a backwards compatible manner. 1494 * Note that entry_size(IN) should still be used to stride the snapshot buffer in the 1495 * event that it's larger than actual kfd_dbg_device_info_entry. 1496 * 1497 * Generic errors apply (see kfd_dbg_trap_operations). 1498 * Return - 0 on SUCCESS. 1499 * Copies @num_devices(IN) device snapshot entries of size @entry_size(IN) 1500 * into @snapshot_buf_ptr if @num_devices(IN) > 0. 1501 * Otherwise return @num_devices(OUT) queue snapshot entries that exist. 1502 */ 1503 struct kfd_ioctl_dbg_trap_device_snapshot_args { 1504 __u64 exception_mask; 1505 __u64 snapshot_buf_ptr; 1506 __u32 num_devices; 1507 __u32 entry_size; 1508 }; 1509 1510 /** 1511 * kfd_ioctl_dbg_trap_args 1512 * 1513 * Arguments to debug target process. 1514 * 1515 * @pid - target process to debug 1516 * @op - debug operation (see kfd_dbg_trap_operations) 1517 * 1518 * @op determines which union struct args to use. 1519 * Refer to kern docs for each kfd_ioctl_dbg_trap_*_args struct. 1520 */ 1521 struct kfd_ioctl_dbg_trap_args { 1522 __u32 pid; 1523 __u32 op; 1524 1525 union { 1526 struct kfd_ioctl_dbg_trap_enable_args enable; 1527 struct kfd_ioctl_dbg_trap_send_runtime_event_args send_runtime_event; 1528 struct kfd_ioctl_dbg_trap_set_exceptions_enabled_args set_exceptions_enabled; 1529 struct kfd_ioctl_dbg_trap_set_wave_launch_override_args launch_override; 1530 struct kfd_ioctl_dbg_trap_set_wave_launch_mode_args launch_mode; 1531 struct kfd_ioctl_dbg_trap_suspend_queues_args suspend_queues; 1532 struct kfd_ioctl_dbg_trap_resume_queues_args resume_queues; 1533 struct kfd_ioctl_dbg_trap_set_node_address_watch_args set_node_address_watch; 1534 struct kfd_ioctl_dbg_trap_clear_node_address_watch_args clear_node_address_watch; 1535 struct kfd_ioctl_dbg_trap_set_flags_args set_flags; 1536 struct kfd_ioctl_dbg_trap_query_debug_event_args query_debug_event; 1537 struct kfd_ioctl_dbg_trap_query_exception_info_args query_exception_info; 1538 struct kfd_ioctl_dbg_trap_queue_snapshot_args queue_snapshot; 1539 struct kfd_ioctl_dbg_trap_device_snapshot_args device_snapshot; 1540 }; 1541 }; 1542 1543 #define AMDKFD_IOCTL_BASE 'K' 1544 #define AMDKFD_IO(nr) _IO(AMDKFD_IOCTL_BASE, nr) 1545 #define AMDKFD_IOR(nr, type) _IOR(AMDKFD_IOCTL_BASE, nr, type) 1546 #define AMDKFD_IOW(nr, type) _IOW(AMDKFD_IOCTL_BASE, nr, type) 1547 #define AMDKFD_IOWR(nr, type) _IOWR(AMDKFD_IOCTL_BASE, nr, type) 1548 1549 #define AMDKFD_IOC_GET_VERSION \ 1550 AMDKFD_IOR(0x01, struct kfd_ioctl_get_version_args) 1551 1552 #define AMDKFD_IOC_CREATE_QUEUE \ 1553 AMDKFD_IOWR(0x02, struct kfd_ioctl_create_queue_args) 1554 1555 #define AMDKFD_IOC_DESTROY_QUEUE \ 1556 AMDKFD_IOWR(0x03, struct kfd_ioctl_destroy_queue_args) 1557 1558 #define AMDKFD_IOC_SET_MEMORY_POLICY \ 1559 AMDKFD_IOW(0x04, struct kfd_ioctl_set_memory_policy_args) 1560 1561 #define AMDKFD_IOC_GET_CLOCK_COUNTERS \ 1562 AMDKFD_IOWR(0x05, struct kfd_ioctl_get_clock_counters_args) 1563 1564 #define AMDKFD_IOC_GET_PROCESS_APERTURES \ 1565 AMDKFD_IOR(0x06, struct kfd_ioctl_get_process_apertures_args) 1566 1567 #define AMDKFD_IOC_UPDATE_QUEUE \ 1568 AMDKFD_IOW(0x07, struct kfd_ioctl_update_queue_args) 1569 1570 #define AMDKFD_IOC_CREATE_EVENT \ 1571 AMDKFD_IOWR(0x08, struct kfd_ioctl_create_event_args) 1572 1573 #define AMDKFD_IOC_DESTROY_EVENT \ 1574 AMDKFD_IOW(0x09, struct kfd_ioctl_destroy_event_args) 1575 1576 #define AMDKFD_IOC_SET_EVENT \ 1577 AMDKFD_IOW(0x0A, struct kfd_ioctl_set_event_args) 1578 1579 #define AMDKFD_IOC_RESET_EVENT \ 1580 AMDKFD_IOW(0x0B, struct kfd_ioctl_reset_event_args) 1581 1582 #define AMDKFD_IOC_WAIT_EVENTS \ 1583 AMDKFD_IOWR(0x0C, struct kfd_ioctl_wait_events_args) 1584 1585 #define AMDKFD_IOC_DBG_REGISTER_DEPRECATED \ 1586 AMDKFD_IOW(0x0D, struct kfd_ioctl_dbg_register_args) 1587 1588 #define AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED \ 1589 AMDKFD_IOW(0x0E, struct kfd_ioctl_dbg_unregister_args) 1590 1591 #define AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED \ 1592 AMDKFD_IOW(0x0F, struct kfd_ioctl_dbg_address_watch_args) 1593 1594 #define AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED \ 1595 AMDKFD_IOW(0x10, struct kfd_ioctl_dbg_wave_control_args) 1596 1597 #define AMDKFD_IOC_SET_SCRATCH_BACKING_VA \ 1598 AMDKFD_IOWR(0x11, struct kfd_ioctl_set_scratch_backing_va_args) 1599 1600 #define AMDKFD_IOC_GET_TILE_CONFIG \ 1601 AMDKFD_IOWR(0x12, struct kfd_ioctl_get_tile_config_args) 1602 1603 #define AMDKFD_IOC_SET_TRAP_HANDLER \ 1604 AMDKFD_IOW(0x13, struct kfd_ioctl_set_trap_handler_args) 1605 1606 #define AMDKFD_IOC_GET_PROCESS_APERTURES_NEW \ 1607 AMDKFD_IOWR(0x14, \ 1608 struct kfd_ioctl_get_process_apertures_new_args) 1609 1610 #define AMDKFD_IOC_ACQUIRE_VM \ 1611 AMDKFD_IOW(0x15, struct kfd_ioctl_acquire_vm_args) 1612 1613 #define AMDKFD_IOC_ALLOC_MEMORY_OF_GPU \ 1614 AMDKFD_IOWR(0x16, struct kfd_ioctl_alloc_memory_of_gpu_args) 1615 1616 #define AMDKFD_IOC_FREE_MEMORY_OF_GPU \ 1617 AMDKFD_IOW(0x17, struct kfd_ioctl_free_memory_of_gpu_args) 1618 1619 #define AMDKFD_IOC_MAP_MEMORY_TO_GPU \ 1620 AMDKFD_IOWR(0x18, struct kfd_ioctl_map_memory_to_gpu_args) 1621 1622 #define AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU \ 1623 AMDKFD_IOWR(0x19, struct kfd_ioctl_unmap_memory_from_gpu_args) 1624 1625 #define AMDKFD_IOC_SET_CU_MASK \ 1626 AMDKFD_IOW(0x1A, struct kfd_ioctl_set_cu_mask_args) 1627 1628 #define AMDKFD_IOC_GET_QUEUE_WAVE_STATE \ 1629 AMDKFD_IOWR(0x1B, struct kfd_ioctl_get_queue_wave_state_args) 1630 1631 #define AMDKFD_IOC_GET_DMABUF_INFO \ 1632 AMDKFD_IOWR(0x1C, struct kfd_ioctl_get_dmabuf_info_args) 1633 1634 #define AMDKFD_IOC_IMPORT_DMABUF \ 1635 AMDKFD_IOWR(0x1D, struct kfd_ioctl_import_dmabuf_args) 1636 1637 #define AMDKFD_IOC_ALLOC_QUEUE_GWS \ 1638 AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args) 1639 1640 #define AMDKFD_IOC_SMI_EVENTS \ 1641 AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args) 1642 1643 #define AMDKFD_IOC_SVM AMDKFD_IOWR(0x20, struct kfd_ioctl_svm_args) 1644 1645 #define AMDKFD_IOC_SET_XNACK_MODE \ 1646 AMDKFD_IOWR(0x21, struct kfd_ioctl_set_xnack_mode_args) 1647 1648 #define AMDKFD_IOC_CRIU_OP \ 1649 AMDKFD_IOWR(0x22, struct kfd_ioctl_criu_args) 1650 1651 #define AMDKFD_IOC_AVAILABLE_MEMORY \ 1652 AMDKFD_IOWR(0x23, struct kfd_ioctl_get_available_memory_args) 1653 1654 #define AMDKFD_IOC_EXPORT_DMABUF \ 1655 AMDKFD_IOWR(0x24, struct kfd_ioctl_export_dmabuf_args) 1656 1657 #define AMDKFD_IOC_RUNTIME_ENABLE \ 1658 AMDKFD_IOWR(0x25, struct kfd_ioctl_runtime_enable_args) 1659 1660 #define AMDKFD_IOC_DBG_TRAP \ 1661 AMDKFD_IOWR(0x26, struct kfd_ioctl_dbg_trap_args) 1662 1663 #define AMDKFD_COMMAND_START 0x01 1664 #define AMDKFD_COMMAND_END 0x27 1665 1666 #endif 1667