2 * Copyright 2014 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
23 #ifndef KFD_IOCTL_H_INCLUDED
24 #define KFD_IOCTL_H_INCLUDED
27 #include <linux/ioctl.h>
30 * - 1.1 - initial version
31 * - 1.3 - Add SMI events support
32 * - 1.4 - Indicate new SRAM EDC bit in device properties
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
44 #define KFD_IOCTL_MAJOR_VERSION 1
45 #define KFD_IOCTL_MINOR_VERSION 14
47 struct kfd_ioctl_get_version_args {
48 __u32 major_version; /* from KFD */
49 __u32 minor_version; /* from KFD */
52 /* For kfd_ioctl_create_queue_args.queue_type. */
53 #define KFD_IOC_QUEUE_TYPE_COMPUTE 0x0
54 #define KFD_IOC_QUEUE_TYPE_SDMA 0x1
55 #define KFD_IOC_QUEUE_TYPE_COMPUTE_AQL 0x2
56 #define KFD_IOC_QUEUE_TYPE_SDMA_XGMI 0x3
58 #define KFD_MAX_QUEUE_PERCENTAGE 100
59 #define KFD_MAX_QUEUE_PRIORITY 15
61 struct kfd_ioctl_create_queue_args {
62 __u64 ring_base_address; /* to KFD */
63 __u64 write_pointer_address; /* from KFD */
64 __u64 read_pointer_address; /* from KFD */
65 __u64 doorbell_offset; /* from KFD */
67 __u32 ring_size; /* to KFD */
68 __u32 gpu_id; /* to KFD */
69 __u32 queue_type; /* to KFD */
70 __u32 queue_percentage; /* to KFD */
71 __u32 queue_priority; /* to KFD */
72 __u32 queue_id; /* from KFD */
74 __u64 eop_buffer_address; /* to KFD */
75 __u64 eop_buffer_size; /* to KFD */
76 __u64 ctx_save_restore_address; /* to KFD */
77 __u32 ctx_save_restore_size; /* to KFD */
78 __u32 ctl_stack_size; /* to KFD */
81 struct kfd_ioctl_destroy_queue_args {
82 __u32 queue_id; /* to KFD */
86 struct kfd_ioctl_update_queue_args {
87 __u64 ring_base_address; /* to KFD */
89 __u32 queue_id; /* to KFD */
90 __u32 ring_size; /* to KFD */
91 __u32 queue_percentage; /* to KFD */
92 __u32 queue_priority; /* to KFD */
95 struct kfd_ioctl_set_cu_mask_args {
96 __u32 queue_id; /* to KFD */
97 __u32 num_cu_mask; /* to KFD */
98 __u64 cu_mask_ptr; /* to KFD */
101 struct kfd_ioctl_get_queue_wave_state_args {
102 __u64 ctl_stack_address; /* to KFD */
103 __u32 ctl_stack_used_size; /* from KFD */
104 __u32 save_area_used_size; /* from KFD */
105 __u32 queue_id; /* to KFD */
109 struct kfd_ioctl_get_available_memory_args {
110 __u64 available; /* from KFD */
111 __u32 gpu_id; /* to KFD */
115 struct kfd_dbg_device_info_entry {
116 __u64 exception_status;
128 __u32 subsystem_vendor_id;
129 __u32 subsystem_device_id;
131 __u32 gfx_target_version;
133 __u32 max_waves_per_simd;
135 __u32 simd_arrays_per_engine;
141 /* For kfd_ioctl_set_memory_policy_args.default_policy and alternate_policy */
142 #define KFD_IOC_CACHE_POLICY_COHERENT 0
143 #define KFD_IOC_CACHE_POLICY_NONCOHERENT 1
145 struct kfd_ioctl_set_memory_policy_args {
146 __u64 alternate_aperture_base; /* to KFD */
147 __u64 alternate_aperture_size; /* to KFD */
149 __u32 gpu_id; /* to KFD */
150 __u32 default_policy; /* to KFD */
151 __u32 alternate_policy; /* to KFD */
156 * All counters are monotonic. They are used for profiling of compute jobs.
157 * The profiling is done by userspace.
159 * In case of GPU reset, the counter should not be affected.
162 struct kfd_ioctl_get_clock_counters_args {
163 __u64 gpu_clock_counter; /* from KFD */
164 __u64 cpu_clock_counter; /* from KFD */
165 __u64 system_clock_counter; /* from KFD */
166 __u64 system_clock_freq; /* from KFD */
168 __u32 gpu_id; /* to KFD */
172 struct kfd_process_device_apertures {
173 __u64 lds_base; /* from KFD */
174 __u64 lds_limit; /* from KFD */
175 __u64 scratch_base; /* from KFD */
176 __u64 scratch_limit; /* from KFD */
177 __u64 gpuvm_base; /* from KFD */
178 __u64 gpuvm_limit; /* from KFD */
179 __u32 gpu_id; /* from KFD */
184 * AMDKFD_IOC_GET_PROCESS_APERTURES is deprecated. Use
185 * AMDKFD_IOC_GET_PROCESS_APERTURES_NEW instead, which supports an
186 * unlimited number of GPUs.
188 #define NUM_OF_SUPPORTED_GPUS 7
189 struct kfd_ioctl_get_process_apertures_args {
190 struct kfd_process_device_apertures
191 process_apertures[NUM_OF_SUPPORTED_GPUS];/* from KFD */
193 /* from KFD, should be in the range [1 - NUM_OF_SUPPORTED_GPUS] */
198 struct kfd_ioctl_get_process_apertures_new_args {
199 /* User allocated. Pointer to struct kfd_process_device_apertures
200 * filled in by Kernel
202 __u64 kfd_process_device_apertures_ptr;
203 /* to KFD - indicates amount of memory present in
204 * kfd_process_device_apertures_ptr
205 * from KFD - Number of entries filled by KFD.
211 #define MAX_ALLOWED_NUM_POINTS 100
212 #define MAX_ALLOWED_AW_BUFF_SIZE 4096
213 #define MAX_ALLOWED_WAC_BUFF_SIZE 128
215 struct kfd_ioctl_dbg_register_args {
216 __u32 gpu_id; /* to KFD */
220 struct kfd_ioctl_dbg_unregister_args {
221 __u32 gpu_id; /* to KFD */
225 struct kfd_ioctl_dbg_address_watch_args {
226 __u64 content_ptr; /* a pointer to the actual content */
227 __u32 gpu_id; /* to KFD */
228 __u32 buf_size_in_bytes; /*including gpu_id and buf_size */
231 struct kfd_ioctl_dbg_wave_control_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 */
237 #define KFD_INVALID_FD 0xffffffff
239 /* Matching HSA_EVENTTYPE */
240 #define KFD_IOC_EVENT_SIGNAL 0
241 #define KFD_IOC_EVENT_NODECHANGE 1
242 #define KFD_IOC_EVENT_DEVICESTATECHANGE 2
243 #define KFD_IOC_EVENT_HW_EXCEPTION 3
244 #define KFD_IOC_EVENT_SYSTEM_EVENT 4
245 #define KFD_IOC_EVENT_DEBUG_EVENT 5
246 #define KFD_IOC_EVENT_PROFILE_EVENT 6
247 #define KFD_IOC_EVENT_QUEUE_EVENT 7
248 #define KFD_IOC_EVENT_MEMORY 8
250 #define KFD_IOC_WAIT_RESULT_COMPLETE 0
251 #define KFD_IOC_WAIT_RESULT_TIMEOUT 1
252 #define KFD_IOC_WAIT_RESULT_FAIL 2
254 #define KFD_SIGNAL_EVENT_LIMIT 4096
256 /* For kfd_event_data.hw_exception_data.reset_type. */
257 #define KFD_HW_EXCEPTION_WHOLE_GPU_RESET 0
258 #define KFD_HW_EXCEPTION_PER_ENGINE_RESET 1
260 /* For kfd_event_data.hw_exception_data.reset_cause. */
261 #define KFD_HW_EXCEPTION_GPU_HANG 0
262 #define KFD_HW_EXCEPTION_ECC 1
264 /* For kfd_hsa_memory_exception_data.ErrorType */
265 #define KFD_MEM_ERR_NO_RAS 0
266 #define KFD_MEM_ERR_SRAM_ECC 1
267 #define KFD_MEM_ERR_POISON_CONSUMED 2
268 #define KFD_MEM_ERR_GPU_HANG 3
270 struct kfd_ioctl_create_event_args {
271 __u64 event_page_offset; /* from KFD */
272 __u32 event_trigger_data; /* from KFD - signal events only */
273 __u32 event_type; /* to KFD */
274 __u32 auto_reset; /* to KFD */
275 __u32 node_id; /* to KFD - only valid for certain
277 __u32 event_id; /* from KFD */
278 __u32 event_slot_index; /* from KFD */
281 struct kfd_ioctl_destroy_event_args {
282 __u32 event_id; /* to KFD */
286 struct kfd_ioctl_set_event_args {
287 __u32 event_id; /* to KFD */
291 struct kfd_ioctl_reset_event_args {
292 __u32 event_id; /* to KFD */
296 struct kfd_memory_exception_failure {
297 __u32 NotPresent; /* Page not present or supervisor privilege */
298 __u32 ReadOnly; /* Write access to a read-only page */
299 __u32 NoExecute; /* Execute access to a page marked NX */
300 __u32 imprecise; /* Can't determine the exact fault address */
303 /* memory exception data */
304 struct kfd_hsa_memory_exception_data {
305 struct kfd_memory_exception_failure failure;
308 __u32 ErrorType; /* 0 = no RAS error,
310 * 2 = Link_SYNFLOOD (poison),
311 * 3 = GPU hang (not attributable to a specific cause),
312 * other values reserved
316 /* hw exception data */
317 struct kfd_hsa_hw_exception_data {
324 /* hsa signal event data */
325 struct kfd_hsa_signal_event_data {
326 __u64 last_event_age; /* to and from KFD */
330 struct kfd_event_data {
333 struct kfd_hsa_memory_exception_data memory_exception_data;
334 struct kfd_hsa_hw_exception_data hw_exception_data;
335 /* To and From KFD */
336 struct kfd_hsa_signal_event_data signal_event_data;
338 __u64 kfd_event_data_ext; /* pointer to an extension structure
339 for future exception types */
340 __u32 event_id; /* to KFD */
344 struct kfd_ioctl_wait_events_args {
345 __u64 events_ptr; /* pointed to struct
346 kfd_event_data array, to KFD */
347 __u32 num_events; /* to KFD */
348 __u32 wait_for_all; /* to KFD */
349 __u32 timeout; /* to KFD */
350 __u32 wait_result; /* from KFD */
353 struct kfd_ioctl_set_scratch_backing_va_args {
354 __u64 va_addr; /* to KFD */
355 __u32 gpu_id; /* to KFD */
359 struct kfd_ioctl_get_tile_config_args {
360 /* to KFD: pointer to tile array */
361 __u64 tile_config_ptr;
362 /* to KFD: pointer to macro tile array */
363 __u64 macro_tile_config_ptr;
364 /* to KFD: array size allocated by user mode
365 * from KFD: array size filled by kernel
367 __u32 num_tile_configs;
368 /* to KFD: array size allocated by user mode
369 * from KFD: array size filled by kernel
371 __u32 num_macro_tile_configs;
373 __u32 gpu_id; /* to KFD */
374 __u32 gb_addr_config; /* from KFD */
375 __u32 num_banks; /* from KFD */
376 __u32 num_ranks; /* from KFD */
377 /* struct size can be extended later if needed
378 * without breaking ABI compatibility
382 struct kfd_ioctl_set_trap_handler_args {
383 __u64 tba_addr; /* to KFD */
384 __u64 tma_addr; /* to KFD */
385 __u32 gpu_id; /* to KFD */
389 struct kfd_ioctl_acquire_vm_args {
390 __u32 drm_fd; /* to KFD */
391 __u32 gpu_id; /* to KFD */
394 /* Allocation flags: memory types */
395 #define KFD_IOC_ALLOC_MEM_FLAGS_VRAM (1 << 0)
396 #define KFD_IOC_ALLOC_MEM_FLAGS_GTT (1 << 1)
397 #define KFD_IOC_ALLOC_MEM_FLAGS_USERPTR (1 << 2)
398 #define KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL (1 << 3)
399 #define KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP (1 << 4)
400 /* Allocation flags: attributes/access options */
401 #define KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE (1 << 31)
402 #define KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE (1 << 30)
403 #define KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC (1 << 29)
404 #define KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE (1 << 28)
405 #define KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM (1 << 27)
406 #define KFD_IOC_ALLOC_MEM_FLAGS_COHERENT (1 << 26)
407 #define KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED (1 << 25)
408 #define KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT (1 << 24)
410 /* Allocate memory for later SVM (shared virtual memory) mapping.
412 * @va_addr: virtual address of the memory to be allocated
413 * all later mappings on all GPUs will use this address
414 * @size: size in bytes
415 * @handle: buffer handle returned to user mode, used to refer to
416 * this allocation for mapping, unmapping and freeing
417 * @mmap_offset: for CPU-mapping the allocation by mmapping a render node
418 * for userptrs this is overloaded to specify the CPU address
419 * @gpu_id: device identifier
420 * @flags: memory type and attributes. See KFD_IOC_ALLOC_MEM_FLAGS above
422 struct kfd_ioctl_alloc_memory_of_gpu_args {
423 __u64 va_addr; /* to KFD */
424 __u64 size; /* to KFD */
425 __u64 handle; /* from KFD */
426 __u64 mmap_offset; /* to KFD (userptr), from KFD (mmap offset) */
427 __u32 gpu_id; /* to KFD */
431 /* Free memory allocated with kfd_ioctl_alloc_memory_of_gpu
433 * @handle: memory handle returned by alloc
435 struct kfd_ioctl_free_memory_of_gpu_args {
436 __u64 handle; /* to KFD */
439 /* Map memory to one or more GPUs
441 * @handle: memory handle returned by alloc
442 * @device_ids_array_ptr: array of gpu_ids (__u32 per device)
443 * @n_devices: number of devices in the array
444 * @n_success: number of devices mapped successfully
446 * @n_success returns information to the caller how many devices from
447 * the start of the array have mapped the buffer successfully. It can
448 * be passed into a subsequent retry call to skip those devices. For
449 * the first call the caller should initialize it to 0.
451 * If the ioctl completes with return code 0 (success), n_success ==
454 struct kfd_ioctl_map_memory_to_gpu_args {
455 __u64 handle; /* to KFD */
456 __u64 device_ids_array_ptr; /* to KFD */
457 __u32 n_devices; /* to KFD */
458 __u32 n_success; /* to/from KFD */
461 /* Unmap memory from one or more GPUs
463 * same arguments as for mapping
465 struct kfd_ioctl_unmap_memory_from_gpu_args {
466 __u64 handle; /* to KFD */
467 __u64 device_ids_array_ptr; /* to KFD */
468 __u32 n_devices; /* to KFD */
469 __u32 n_success; /* to/from KFD */
472 /* Allocate GWS for specific queue
474 * @queue_id: queue's id that GWS is allocated for
475 * @num_gws: how many GWS to allocate
476 * @first_gws: index of the first GWS allocated.
477 * only support contiguous GWS allocation
479 struct kfd_ioctl_alloc_queue_gws_args {
480 __u32 queue_id; /* to KFD */
481 __u32 num_gws; /* to KFD */
482 __u32 first_gws; /* from KFD */
486 struct kfd_ioctl_get_dmabuf_info_args {
487 __u64 size; /* from KFD */
488 __u64 metadata_ptr; /* to KFD */
489 __u32 metadata_size; /* to KFD (space allocated by user)
490 * from KFD (actual metadata size)
492 __u32 gpu_id; /* from KFD */
493 __u32 flags; /* from KFD (KFD_IOC_ALLOC_MEM_FLAGS) */
494 __u32 dmabuf_fd; /* to KFD */
497 struct kfd_ioctl_import_dmabuf_args {
498 __u64 va_addr; /* to KFD */
499 __u64 handle; /* from KFD */
500 __u32 gpu_id; /* to KFD */
501 __u32 dmabuf_fd; /* to KFD */
504 struct kfd_ioctl_export_dmabuf_args {
505 __u64 handle; /* to KFD */
506 __u32 flags; /* to KFD */
507 __u32 dmabuf_fd; /* from KFD */
511 * KFD SMI(System Management Interface) events
514 KFD_SMI_EVENT_NONE = 0, /* not used */
515 KFD_SMI_EVENT_VMFAULT = 1, /* event start counting at 1 */
516 KFD_SMI_EVENT_THERMAL_THROTTLE = 2,
517 KFD_SMI_EVENT_GPU_PRE_RESET = 3,
518 KFD_SMI_EVENT_GPU_POST_RESET = 4,
519 KFD_SMI_EVENT_MIGRATE_START = 5,
520 KFD_SMI_EVENT_MIGRATE_END = 6,
521 KFD_SMI_EVENT_PAGE_FAULT_START = 7,
522 KFD_SMI_EVENT_PAGE_FAULT_END = 8,
523 KFD_SMI_EVENT_QUEUE_EVICTION = 9,
524 KFD_SMI_EVENT_QUEUE_RESTORE = 10,
525 KFD_SMI_EVENT_UNMAP_FROM_GPU = 11,
528 * max event number, as a flag bit to get events from all processes,
529 * this requires super user permission, otherwise will not be able to
530 * receive event from any process. Without this flag to receive events
533 KFD_SMI_EVENT_ALL_PROCESS = 64
536 enum KFD_MIGRATE_TRIGGERS {
537 KFD_MIGRATE_TRIGGER_PREFETCH,
538 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
539 KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU,
540 KFD_MIGRATE_TRIGGER_TTM_EVICTION
543 enum KFD_QUEUE_EVICTION_TRIGGERS {
544 KFD_QUEUE_EVICTION_TRIGGER_SVM,
545 KFD_QUEUE_EVICTION_TRIGGER_USERPTR,
546 KFD_QUEUE_EVICTION_TRIGGER_TTM,
547 KFD_QUEUE_EVICTION_TRIGGER_SUSPEND,
548 KFD_QUEUE_EVICTION_CRIU_CHECKPOINT,
549 KFD_QUEUE_EVICTION_CRIU_RESTORE
552 enum KFD_SVM_UNMAP_TRIGGERS {
553 KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY,
554 KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE,
555 KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU
558 #define KFD_SMI_EVENT_MASK_FROM_INDEX(i) (1ULL << ((i) - 1))
559 #define KFD_SMI_EVENT_MSG_SIZE 96
561 struct kfd_ioctl_smi_events_args {
562 __u32 gpuid; /* to KFD */
563 __u32 anon_fd; /* from KFD */
566 /**************************************************************************************************
567 * CRIU IOCTLs (Checkpoint Restore In Userspace)
569 * When checkpointing a process, the userspace application will perform:
570 * 1. PROCESS_INFO op to determine current process information. This pauses execution and evicts
572 * 2. CHECKPOINT op to checkpoint process contents (BOs, queues, events, svm-ranges)
573 * 3. UNPAUSE op to un-evict all the queues
575 * When restoring a process, the CRIU userspace application will perform:
577 * 1. RESTORE op to restore process contents
578 * 2. RESUME op to start the process
580 * Note: Queues are forced into an evicted state after a successful PROCESS_INFO. User
581 * application needs to perform an UNPAUSE operation after calling PROCESS_INFO.
585 KFD_CRIU_OP_PROCESS_INFO,
586 KFD_CRIU_OP_CHECKPOINT,
593 * kfd_ioctl_criu_args - Arguments perform CRIU operation
594 * @devices: [in/out] User pointer to memory location for devices information.
595 * This is an array of type kfd_criu_device_bucket.
596 * @bos: [in/out] User pointer to memory location for BOs information
597 * This is an array of type kfd_criu_bo_bucket.
598 * @priv_data: [in/out] User pointer to memory location for private data
599 * @priv_data_size: [in/out] Size of priv_data in bytes
600 * @num_devices: [in/out] Number of GPUs used by process. Size of @devices array.
601 * @num_bos [in/out] Number of BOs used by process. Size of @bos array.
602 * @num_objects: [in/out] Number of objects used by process. Objects are opaque to
604 * @pid: [in/out] PID of the process being checkpointed
605 * @op [in] Type of operation (kfd_criu_op)
607 * Return: 0 on success, -errno on failure
609 struct kfd_ioctl_criu_args {
610 __u64 devices; /* Used during ops: CHECKPOINT, RESTORE */
611 __u64 bos; /* Used during ops: CHECKPOINT, RESTORE */
612 __u64 priv_data; /* Used during ops: CHECKPOINT, RESTORE */
613 __u64 priv_data_size; /* Used during ops: PROCESS_INFO, RESTORE */
614 __u32 num_devices; /* Used during ops: PROCESS_INFO, RESTORE */
615 __u32 num_bos; /* Used during ops: PROCESS_INFO, RESTORE */
616 __u32 num_objects; /* Used during ops: PROCESS_INFO, RESTORE */
617 __u32 pid; /* Used during ops: PROCESS_INFO, RESUME */
621 struct kfd_criu_device_bucket {
628 struct kfd_criu_bo_bucket {
632 __u64 restored_offset; /* During restore, updated offset for BO */
633 __u32 gpu_id; /* This is the user_gpu_id */
639 /* CRIU IOCTLs - END */
640 /**************************************************************************************************/
642 /* Register offset inside the remapped mmio page
644 enum kfd_mmio_remap {
645 KFD_MMIO_REMAP_HDP_MEM_FLUSH_CNTL = 0,
646 KFD_MMIO_REMAP_HDP_REG_FLUSH_CNTL = 4,
649 /* Guarantee host access to memory */
650 #define KFD_IOCTL_SVM_FLAG_HOST_ACCESS 0x00000001
651 /* Fine grained coherency between all devices with access */
652 #define KFD_IOCTL_SVM_FLAG_COHERENT 0x00000002
653 /* Use any GPU in same hive as preferred device */
654 #define KFD_IOCTL_SVM_FLAG_HIVE_LOCAL 0x00000004
655 /* GPUs only read, allows replication */
656 #define KFD_IOCTL_SVM_FLAG_GPU_RO 0x00000008
657 /* Allow execution on GPU */
658 #define KFD_IOCTL_SVM_FLAG_GPU_EXEC 0x00000010
659 /* GPUs mostly read, may allow similar optimizations as RO, but writes fault */
660 #define KFD_IOCTL_SVM_FLAG_GPU_READ_MOSTLY 0x00000020
661 /* Keep GPU memory mapping always valid as if XNACK is disable */
662 #define KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED 0x00000040
663 /* Fine grained coherency between all devices using device-scope atomics */
664 #define KFD_IOCTL_SVM_FLAG_EXT_COHERENT 0x00000080
667 * kfd_ioctl_svm_op - SVM ioctl operations
669 * @KFD_IOCTL_SVM_OP_SET_ATTR: Modify one or more attributes
670 * @KFD_IOCTL_SVM_OP_GET_ATTR: Query one or more attributes
672 enum kfd_ioctl_svm_op {
673 KFD_IOCTL_SVM_OP_SET_ATTR,
674 KFD_IOCTL_SVM_OP_GET_ATTR
677 /** kfd_ioctl_svm_location - Enum for preferred and prefetch locations
679 * GPU IDs are used to specify GPUs as preferred and prefetch locations.
680 * Below definitions are used for system memory or for leaving the preferred
681 * location unspecified.
683 enum kfd_ioctl_svm_location {
684 KFD_IOCTL_SVM_LOCATION_SYSMEM = 0,
685 KFD_IOCTL_SVM_LOCATION_UNDEFINED = 0xffffffff
689 * kfd_ioctl_svm_attr_type - SVM attribute types
691 * @KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: gpuid of the preferred location, 0 for
693 * @KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: gpuid of the prefetch location, 0 for
694 * system memory. Setting this triggers an
695 * immediate prefetch (migration).
696 * @KFD_IOCTL_SVM_ATTR_ACCESS:
697 * @KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
698 * @KFD_IOCTL_SVM_ATTR_NO_ACCESS: specify memory access for the gpuid given
699 * by the attribute value
700 * @KFD_IOCTL_SVM_ATTR_SET_FLAGS: bitmask of flags to set (see
701 * KFD_IOCTL_SVM_FLAG_...)
702 * @KFD_IOCTL_SVM_ATTR_CLR_FLAGS: bitmask of flags to clear
703 * @KFD_IOCTL_SVM_ATTR_GRANULARITY: migration granularity
706 enum kfd_ioctl_svm_attr_type {
707 KFD_IOCTL_SVM_ATTR_PREFERRED_LOC,
708 KFD_IOCTL_SVM_ATTR_PREFETCH_LOC,
709 KFD_IOCTL_SVM_ATTR_ACCESS,
710 KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE,
711 KFD_IOCTL_SVM_ATTR_NO_ACCESS,
712 KFD_IOCTL_SVM_ATTR_SET_FLAGS,
713 KFD_IOCTL_SVM_ATTR_CLR_FLAGS,
714 KFD_IOCTL_SVM_ATTR_GRANULARITY
718 * kfd_ioctl_svm_attribute - Attributes as pairs of type and value
720 * The meaning of the @value depends on the attribute type.
722 * @type: attribute type (see enum @kfd_ioctl_svm_attr_type)
723 * @value: attribute value
725 struct kfd_ioctl_svm_attribute {
731 * kfd_ioctl_svm_args - Arguments for SVM ioctl
733 * @op specifies the operation to perform (see enum
734 * @kfd_ioctl_svm_op). @start_addr and @size are common for all
737 * A variable number of attributes can be given in @attrs.
738 * @nattr specifies the number of attributes. New attributes can be
739 * added in the future without breaking the ABI. If unknown attributes
740 * are given, the function returns -EINVAL.
742 * @KFD_IOCTL_SVM_OP_SET_ATTR sets attributes for a virtual address
743 * range. It may overlap existing virtual address ranges. If it does,
744 * the existing ranges will be split such that the attribute changes
745 * only apply to the specified address range.
747 * @KFD_IOCTL_SVM_OP_GET_ATTR returns the intersection of attributes
748 * over all memory in the given range and returns the result as the
749 * attribute value. If different pages have different preferred or
750 * prefetch locations, 0xffffffff will be returned for
751 * @KFD_IOCTL_SVM_ATTR_PREFERRED_LOC or
752 * @KFD_IOCTL_SVM_ATTR_PREFETCH_LOC resepctively. For
753 * @KFD_IOCTL_SVM_ATTR_SET_FLAGS, flags of all pages will be
754 * aggregated by bitwise AND. That means, a flag will be set in the
755 * output, if that flag is set for all pages in the range. For
756 * @KFD_IOCTL_SVM_ATTR_CLR_FLAGS, flags of all pages will be
757 * aggregated by bitwise NOR. That means, a flag will be set in the
758 * output, if that flag is clear for all pages in the range.
759 * The minimum migration granularity throughout the range will be
760 * returned for @KFD_IOCTL_SVM_ATTR_GRANULARITY.
762 * Querying of accessibility attributes works by initializing the
763 * attribute type to @KFD_IOCTL_SVM_ATTR_ACCESS and the value to the
764 * GPUID being queried. Multiple attributes can be given to allow
765 * querying multiple GPUIDs. The ioctl function overwrites the
766 * attribute type to indicate the access for the specified GPU.
768 struct kfd_ioctl_svm_args {
773 /* Variable length array of attributes */
774 struct kfd_ioctl_svm_attribute attrs[];
778 * kfd_ioctl_set_xnack_mode_args - Arguments for set_xnack_mode
780 * @xnack_enabled: [in/out] Whether to enable XNACK mode for this process
782 * @xnack_enabled indicates whether recoverable page faults should be
783 * enabled for the current process. 0 means disabled, positive means
784 * enabled, negative means leave unchanged. If enabled, virtual address
785 * translations on GFXv9 and later AMD GPUs can return XNACK and retry
786 * the access until a valid PTE is available. This is used to implement
787 * device page faults.
789 * On output, @xnack_enabled returns the (new) current mode (0 or
790 * positive). Therefore, a negative input value can be used to query
791 * the current mode without changing it.
793 * The XNACK mode fundamentally changes the way SVM managed memory works
794 * in the driver, with subtle effects on application performance and
797 * Enabling XNACK mode requires shader programs to be compiled
798 * differently. Furthermore, not all GPUs support changing the mode
799 * per-process. Therefore changing the mode is only allowed while no
800 * user mode queues exist in the process. This ensure that no shader
801 * code is running that may be compiled for the wrong mode. And GPUs
802 * that cannot change to the requested mode will prevent the XNACK
803 * mode from occurring. All GPUs used by the process must be in the
806 * GFXv8 or older GPUs do not support 48 bit virtual addresses or SVM.
807 * Therefore those GPUs are not considered for the XNACK mode switch.
809 * Return: 0 on success, -errno on failure
811 struct kfd_ioctl_set_xnack_mode_args {
815 /* Wave launch override modes */
816 enum kfd_dbg_trap_override_mode {
817 KFD_DBG_TRAP_OVERRIDE_OR = 0,
818 KFD_DBG_TRAP_OVERRIDE_REPLACE = 1
821 /* Wave launch overrides */
822 enum kfd_dbg_trap_mask {
823 KFD_DBG_TRAP_MASK_FP_INVALID = 1,
824 KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL = 2,
825 KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO = 4,
826 KFD_DBG_TRAP_MASK_FP_OVERFLOW = 8,
827 KFD_DBG_TRAP_MASK_FP_UNDERFLOW = 16,
828 KFD_DBG_TRAP_MASK_FP_INEXACT = 32,
829 KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO = 64,
830 KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH = 128,
831 KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION = 256,
832 KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START = (1 << 30),
833 KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END = (1 << 31)
836 /* Wave launch modes */
837 enum kfd_dbg_trap_wave_launch_mode {
838 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_NORMAL = 0,
839 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_HALT = 1,
840 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_DEBUG = 3
843 /* Address watch modes */
844 enum kfd_dbg_trap_address_watch_mode {
845 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_READ = 0,
846 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_NONREAD = 1,
847 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_ATOMIC = 2,
848 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_ALL = 3
851 /* Additional wave settings */
852 enum kfd_dbg_trap_flags {
853 KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP = 1,
856 /* Trap exceptions */
857 enum kfd_dbg_trap_exception_code {
860 EC_QUEUE_WAVE_ABORT = 1,
861 EC_QUEUE_WAVE_TRAP = 2,
862 EC_QUEUE_WAVE_MATH_ERROR = 3,
863 EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION = 4,
864 EC_QUEUE_WAVE_MEMORY_VIOLATION = 5,
865 EC_QUEUE_WAVE_APERTURE_VIOLATION = 6,
866 EC_QUEUE_PACKET_DISPATCH_DIM_INVALID = 16,
867 EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID = 17,
868 EC_QUEUE_PACKET_DISPATCH_CODE_INVALID = 18,
869 EC_QUEUE_PACKET_RESERVED = 19,
870 EC_QUEUE_PACKET_UNSUPPORTED = 20,
871 EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID = 21,
872 EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID = 22,
873 EC_QUEUE_PACKET_VENDOR_UNSUPPORTED = 23,
874 EC_QUEUE_PREEMPTION_ERROR = 30,
877 EC_DEVICE_QUEUE_DELETE = 32,
878 EC_DEVICE_MEMORY_VIOLATION = 33,
879 EC_DEVICE_RAS_ERROR = 34,
880 EC_DEVICE_FATAL_HALT = 35,
883 EC_PROCESS_RUNTIME = 48,
884 EC_PROCESS_DEVICE_REMOVE = 49,
888 /* Mask generated by ecode in kfd_dbg_trap_exception_code */
889 #define KFD_EC_MASK(ecode) (1ULL << (ecode - 1))
891 /* Masks for exception code type checks below */
892 #define KFD_EC_MASK_QUEUE (KFD_EC_MASK(EC_QUEUE_WAVE_ABORT) | \
893 KFD_EC_MASK(EC_QUEUE_WAVE_TRAP) | \
894 KFD_EC_MASK(EC_QUEUE_WAVE_MATH_ERROR) | \
895 KFD_EC_MASK(EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION) | \
896 KFD_EC_MASK(EC_QUEUE_WAVE_MEMORY_VIOLATION) | \
897 KFD_EC_MASK(EC_QUEUE_WAVE_APERTURE_VIOLATION) | \
898 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_DIM_INVALID) | \
899 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID) | \
900 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_CODE_INVALID) | \
901 KFD_EC_MASK(EC_QUEUE_PACKET_RESERVED) | \
902 KFD_EC_MASK(EC_QUEUE_PACKET_UNSUPPORTED) | \
903 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) | \
904 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) | \
905 KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED) | \
906 KFD_EC_MASK(EC_QUEUE_PREEMPTION_ERROR) | \
907 KFD_EC_MASK(EC_QUEUE_NEW))
908 #define KFD_EC_MASK_DEVICE (KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE) | \
909 KFD_EC_MASK(EC_DEVICE_RAS_ERROR) | \
910 KFD_EC_MASK(EC_DEVICE_FATAL_HALT) | \
911 KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION) | \
912 KFD_EC_MASK(EC_DEVICE_NEW))
913 #define KFD_EC_MASK_PROCESS (KFD_EC_MASK(EC_PROCESS_RUNTIME) | \
914 KFD_EC_MASK(EC_PROCESS_DEVICE_REMOVE))
916 /* Checks for exception code types for KFD search */
917 #define KFD_DBG_EC_TYPE_IS_QUEUE(ecode) \
918 (!!(KFD_EC_MASK(ecode) & KFD_EC_MASK_QUEUE))
919 #define KFD_DBG_EC_TYPE_IS_DEVICE(ecode) \
920 (!!(KFD_EC_MASK(ecode) & KFD_EC_MASK_DEVICE))
921 #define KFD_DBG_EC_TYPE_IS_PROCESS(ecode) \
922 (!!(KFD_EC_MASK(ecode) & KFD_EC_MASK_PROCESS))
925 /* Runtime enable states */
926 enum kfd_dbg_runtime_state {
927 DEBUG_RUNTIME_STATE_DISABLED = 0,
928 DEBUG_RUNTIME_STATE_ENABLED = 1,
929 DEBUG_RUNTIME_STATE_ENABLED_BUSY = 2,
930 DEBUG_RUNTIME_STATE_ENABLED_ERROR = 3
933 /* Runtime enable status */
934 struct kfd_runtime_info {
940 /* Enable modes for runtime enable */
941 #define KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK 1
942 #define KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK 2
945 * kfd_ioctl_runtime_enable_args - Arguments for runtime enable
947 * Coordinates debug exception signalling and debug device enablement with runtime.
949 * @r_debug - pointer to user struct for sharing information between ROCr and the debuggger
950 * @mode_mask - mask to set mode
951 * KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK - enable runtime for debugging, otherwise disable
952 * KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK - enable trap temporary setup (ignore on disable)
953 * @capabilities_mask - mask to notify runtime on what KFD supports
955 * Return - 0 on SUCCESS.
956 * - EBUSY if runtime enable call already pending.
957 * - EEXIST if user queues already active prior to call.
958 * If process is debug enabled, runtime enable will enable debug devices and
959 * wait for debugger process to send runtime exception EC_PROCESS_RUNTIME
960 * to unblock - see kfd_ioctl_dbg_trap_args.
963 struct kfd_ioctl_runtime_enable_args {
966 __u32 capabilities_mask;
969 /* Queue information */
970 struct kfd_queue_snapshot_entry {
971 __u64 exception_status;
972 __u64 ring_base_address;
973 __u64 write_pointer_address;
974 __u64 read_pointer_address;
975 __u64 ctx_save_restore_address;
980 __u32 ctx_save_restore_area_size;
984 /* Queue status return for suspend/resume */
985 #define KFD_DBG_QUEUE_ERROR_BIT 30
986 #define KFD_DBG_QUEUE_INVALID_BIT 31
987 #define KFD_DBG_QUEUE_ERROR_MASK (1 << KFD_DBG_QUEUE_ERROR_BIT)
988 #define KFD_DBG_QUEUE_INVALID_MASK (1 << KFD_DBG_QUEUE_INVALID_BIT)
990 /* Context save area header information */
991 struct kfd_context_save_area_header {
993 __u32 control_stack_offset;
994 __u32 control_stack_size;
995 __u32 wave_state_offset;
996 __u32 wave_state_size;
1000 __u64 err_payload_addr;
1008 * For specifics on usage and return values, see documentation per operation
1009 * below. Otherwise, generic error returns apply:
1010 * - ESRCH if the process to debug does not exist.
1012 * - EINVAL (with KFD_IOC_DBG_TRAP_ENABLE exempt) if operation
1013 * KFD_IOC_DBG_TRAP_ENABLE has not succeeded prior.
1014 * Also returns this error if GPU hardware scheduling is not supported.
1016 * - EPERM (with KFD_IOC_DBG_TRAP_DISABLE exempt) if target process is not
1017 * PTRACE_ATTACHED. KFD_IOC_DBG_TRAP_DISABLE is exempt to allow
1018 * clean up of debug mode as long as process is debug enabled.
1020 * - EACCES if any DBG_HW_OP (debug hardware operation) is requested when
1021 * AMDKFD_IOC_RUNTIME_ENABLE has not succeeded prior.
1023 * - ENODEV if any GPU does not support debugging on a DBG_HW_OP call.
1025 * - Other errors may be returned when a DBG_HW_OP occurs while the GPU
1026 * is in a fatal state.
1029 enum kfd_dbg_trap_operations {
1030 KFD_IOC_DBG_TRAP_ENABLE = 0,
1031 KFD_IOC_DBG_TRAP_DISABLE = 1,
1032 KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT = 2,
1033 KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED = 3,
1034 KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE = 4, /* DBG_HW_OP */
1035 KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE = 5, /* DBG_HW_OP */
1036 KFD_IOC_DBG_TRAP_SUSPEND_QUEUES = 6, /* DBG_HW_OP */
1037 KFD_IOC_DBG_TRAP_RESUME_QUEUES = 7, /* DBG_HW_OP */
1038 KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH = 8, /* DBG_HW_OP */
1039 KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH = 9, /* DBG_HW_OP */
1040 KFD_IOC_DBG_TRAP_SET_FLAGS = 10,
1041 KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT = 11,
1042 KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO = 12,
1043 KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT = 13,
1044 KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT = 14
1048 * kfd_ioctl_dbg_trap_enable_args
1050 * Arguments for KFD_IOC_DBG_TRAP_ENABLE.
1052 * Enables debug session for target process. Call @op KFD_IOC_DBG_TRAP_DISABLE in
1053 * kfd_ioctl_dbg_trap_args to disable debug session.
1055 * @exception_mask (IN) - exceptions to raise to the debugger
1056 * @rinfo_ptr (IN) - pointer to runtime info buffer (see kfd_runtime_info)
1057 * @rinfo_size (IN/OUT) - size of runtime info buffer in bytes
1058 * @dbg_fd (IN) - fd the KFD will nofify the debugger with of raised
1059 * exceptions set in exception_mask.
1061 * Generic errors apply (see kfd_dbg_trap_operations).
1062 * Return - 0 on SUCCESS.
1063 * Copies KFD saved kfd_runtime_info to @rinfo_ptr on enable.
1064 * Size of kfd_runtime saved by the KFD returned to @rinfo_size.
1065 * - EBADF if KFD cannot get a reference to dbg_fd.
1066 * - EFAULT if KFD cannot copy runtime info to rinfo_ptr.
1067 * - EINVAL if target process is already debug enabled.
1070 struct kfd_ioctl_dbg_trap_enable_args {
1071 __u64 exception_mask;
1078 * kfd_ioctl_dbg_trap_send_runtime_event_args
1081 * Arguments for KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT.
1082 * Raises exceptions to runtime.
1084 * @exception_mask (IN) - exceptions to raise to runtime
1085 * @gpu_id (IN) - target device id
1086 * @queue_id (IN) - target queue id
1088 * Generic errors apply (see kfd_dbg_trap_operations).
1089 * Return - 0 on SUCCESS.
1090 * - ENODEV if gpu_id not found.
1091 * If exception_mask contains EC_PROCESS_RUNTIME, unblocks pending
1092 * AMDKFD_IOC_RUNTIME_ENABLE call - see kfd_ioctl_runtime_enable_args.
1093 * All other exceptions are raised to runtime through err_payload_addr.
1094 * See kfd_context_save_area_header.
1096 struct kfd_ioctl_dbg_trap_send_runtime_event_args {
1097 __u64 exception_mask;
1103 * kfd_ioctl_dbg_trap_set_exceptions_enabled_args
1105 * Arguments for KFD_IOC_SET_EXCEPTIONS_ENABLED
1106 * Set new exceptions to be raised to the debugger.
1108 * @exception_mask (IN) - new exceptions to raise the debugger
1110 * Generic errors apply (see kfd_dbg_trap_operations).
1111 * Return - 0 on SUCCESS.
1113 struct kfd_ioctl_dbg_trap_set_exceptions_enabled_args {
1114 __u64 exception_mask;
1118 * kfd_ioctl_dbg_trap_set_wave_launch_override_args
1120 * Arguments for KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE
1121 * Enable HW exceptions to raise trap.
1123 * @override_mode (IN) - see kfd_dbg_trap_override_mode
1124 * @enable_mask (IN/OUT) - reference kfd_dbg_trap_mask.
1125 * IN is the override modes requested to be enabled.
1126 * OUT is referenced in Return below.
1127 * @support_request_mask (IN/OUT) - reference kfd_dbg_trap_mask.
1128 * IN is the override modes requested for support check.
1129 * OUT is referenced in Return below.
1131 * Generic errors apply (see kfd_dbg_trap_operations).
1132 * Return - 0 on SUCCESS.
1133 * Previous enablement is returned in @enable_mask.
1134 * Actual override support is returned in @support_request_mask.
1135 * - EINVAL if override mode is not supported.
1136 * - EACCES if trap support requested is not actually supported.
1137 * i.e. enable_mask (IN) is not a subset of support_request_mask (OUT).
1138 * Otherwise it is considered a generic error (see kfd_dbg_trap_operations).
1140 struct kfd_ioctl_dbg_trap_set_wave_launch_override_args {
1141 __u32 override_mode;
1143 __u32 support_request_mask;
1148 * kfd_ioctl_dbg_trap_set_wave_launch_mode_args
1150 * Arguments for KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE
1151 * Set wave launch mode.
1153 * @mode (IN) - see kfd_dbg_trap_wave_launch_mode
1155 * Generic errors apply (see kfd_dbg_trap_operations).
1156 * Return - 0 on SUCCESS.
1158 struct kfd_ioctl_dbg_trap_set_wave_launch_mode_args {
1164 * kfd_ioctl_dbg_trap_suspend_queues_ags
1166 * Arguments for KFD_IOC_DBG_TRAP_SUSPEND_QUEUES
1169 * @exception_mask (IN) - raised exceptions to clear
1170 * @queue_array_ptr (IN) - pointer to array of queue ids (u32 per queue id)
1172 * @num_queues (IN) - number of queues to suspend in @queue_array_ptr
1173 * @grace_period (IN) - wave time allowance before preemption
1174 * per 1K GPU clock cycle unit
1176 * Generic errors apply (see kfd_dbg_trap_operations).
1177 * Destruction of a suspended queue is blocked until the queue is
1178 * resumed. This allows the debugger to access queue information and
1179 * the its context save area without running into a race condition on
1180 * queue destruction.
1181 * Automatically copies per queue context save area header information
1182 * into the save area base
1183 * (see kfd_queue_snapshot_entry and kfd_context_save_area_header).
1185 * Return - Number of queues suspended on SUCCESS.
1186 * . KFD_DBG_QUEUE_ERROR_MASK and KFD_DBG_QUEUE_INVALID_MASK masked
1187 * for each queue id in @queue_array_ptr array reports unsuccessful
1189 * KFD_DBG_QUEUE_ERROR_MASK = HW failure.
1190 * KFD_DBG_QUEUE_INVALID_MASK = queue does not exist, is new or
1191 * is being destroyed.
1193 struct kfd_ioctl_dbg_trap_suspend_queues_args {
1194 __u64 exception_mask;
1195 __u64 queue_array_ptr;
1201 * kfd_ioctl_dbg_trap_resume_queues_args
1203 * Arguments for KFD_IOC_DBG_TRAP_RESUME_QUEUES
1206 * @queue_array_ptr (IN) - pointer to array of queue ids (u32 per queue id)
1208 * @num_queues (IN) - number of queues to resume in @queue_array_ptr
1210 * Generic errors apply (see kfd_dbg_trap_operations).
1211 * Return - Number of queues resumed on SUCCESS.
1212 * KFD_DBG_QUEUE_ERROR_MASK and KFD_DBG_QUEUE_INVALID_MASK mask
1213 * for each queue id in @queue_array_ptr array reports unsuccessful
1215 * KFD_DBG_QUEUE_ERROR_MASK = HW failure.
1216 * KFD_DBG_QUEUE_INVALID_MASK = queue does not exist.
1218 struct kfd_ioctl_dbg_trap_resume_queues_args {
1219 __u64 queue_array_ptr;
1225 * kfd_ioctl_dbg_trap_set_node_address_watch_args
1227 * Arguments for KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH
1228 * Sets address watch for device.
1230 * @address (IN) - watch address to set
1231 * @mode (IN) - see kfd_dbg_trap_address_watch_mode
1232 * @mask (IN) - watch address mask
1233 * @gpu_id (IN) - target gpu to set watch point
1234 * @id (OUT) - watch id allocated
1236 * Generic errors apply (see kfd_dbg_trap_operations).
1237 * Return - 0 on SUCCESS.
1238 * Allocated watch ID returned to @id.
1239 * - ENODEV if gpu_id not found.
1240 * - ENOMEM if watch IDs can be allocated
1242 struct kfd_ioctl_dbg_trap_set_node_address_watch_args {
1251 * kfd_ioctl_dbg_trap_clear_node_address_watch_args
1253 * Arguments for KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH
1254 * Clear address watch for device.
1256 * @gpu_id (IN) - target device to clear watch point
1257 * @id (IN) - allocated watch id to clear
1259 * Generic errors apply (see kfd_dbg_trap_operations).
1260 * Return - 0 on SUCCESS.
1261 * - ENODEV if gpu_id not found.
1262 * - EINVAL if watch ID has not been allocated.
1264 struct kfd_ioctl_dbg_trap_clear_node_address_watch_args {
1270 * kfd_ioctl_dbg_trap_set_flags_args
1272 * Arguments for KFD_IOC_DBG_TRAP_SET_FLAGS
1273 * Sets flags for wave behaviour.
1275 * @flags (IN/OUT) - IN = flags to enable, OUT = flags previously enabled
1277 * Generic errors apply (see kfd_dbg_trap_operations).
1278 * Return - 0 on SUCCESS.
1279 * - EACCESS if any debug device does not allow flag options.
1281 struct kfd_ioctl_dbg_trap_set_flags_args {
1287 * kfd_ioctl_dbg_trap_query_debug_event_args
1289 * Arguments for KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT
1291 * Find one or more raised exceptions. This function can return multiple
1292 * exceptions from a single queue or a single device with one call. To find
1293 * all raised exceptions, this function must be called repeatedly until it
1294 * returns -EAGAIN. Returned exceptions can optionally be cleared by
1295 * setting the corresponding bit in the @exception_mask input parameter.
1296 * However, clearing an exception prevents retrieving further information
1297 * about it with KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO.
1299 * @exception_mask (IN/OUT) - exception to clear (IN) and raised (OUT)
1300 * @gpu_id (OUT) - gpu id of exceptions raised
1301 * @queue_id (OUT) - queue id of exceptions raised
1303 * Generic errors apply (see kfd_dbg_trap_operations).
1304 * Return - 0 on raised exception found
1305 * Raised exceptions found are returned in @exception mask
1306 * with reported source id returned in @gpu_id or @queue_id.
1307 * - EAGAIN if no raised exception has been found
1309 struct kfd_ioctl_dbg_trap_query_debug_event_args {
1310 __u64 exception_mask;
1316 * kfd_ioctl_dbg_trap_query_exception_info_args
1318 * Arguments KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO
1319 * Get additional info on raised exception.
1321 * @info_ptr (IN) - pointer to exception info buffer to copy to
1322 * @info_size (IN/OUT) - exception info buffer size (bytes)
1323 * @source_id (IN) - target gpu or queue id
1324 * @exception_code (IN) - target exception
1325 * @clear_exception (IN) - clear raised @exception_code exception
1326 * (0 = false, 1 = true)
1328 * Generic errors apply (see kfd_dbg_trap_operations).
1329 * Return - 0 on SUCCESS.
1330 * If @exception_code is EC_DEVICE_MEMORY_VIOLATION, copy @info_size(OUT)
1331 * bytes of memory exception data to @info_ptr.
1332 * If @exception_code is EC_PROCESS_RUNTIME, copy saved
1333 * kfd_runtime_info to @info_ptr.
1334 * Actual required @info_ptr size (bytes) is returned in @info_size.
1336 struct kfd_ioctl_dbg_trap_query_exception_info_args {
1340 __u32 exception_code;
1341 __u32 clear_exception;
1345 * kfd_ioctl_dbg_trap_get_queue_snapshot_args
1347 * Arguments KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT
1348 * Get queue information.
1350 * @exception_mask (IN) - exceptions raised to clear
1351 * @snapshot_buf_ptr (IN) - queue snapshot entry buffer (see kfd_queue_snapshot_entry)
1352 * @num_queues (IN/OUT) - number of queue snapshot entries
1353 * The debugger specifies the size of the array allocated in @num_queues.
1354 * KFD returns the number of queues that actually existed. If this is
1355 * larger than the size specified by the debugger, KFD will not overflow
1356 * the array allocated by the debugger.
1358 * @entry_size (IN/OUT) - size per entry in bytes
1359 * The debugger specifies sizeof(struct kfd_queue_snapshot_entry) in
1360 * @entry_size. KFD returns the number of bytes actually populated per
1361 * entry. The debugger should use the KFD_IOCTL_MINOR_VERSION to determine,
1362 * which fields in struct kfd_queue_snapshot_entry are valid. This allows
1363 * growing the ABI in a backwards compatible manner.
1364 * Note that entry_size(IN) should still be used to stride the snapshot buffer in the
1365 * event that it's larger than actual kfd_queue_snapshot_entry.
1367 * Generic errors apply (see kfd_dbg_trap_operations).
1368 * Return - 0 on SUCCESS.
1369 * Copies @num_queues(IN) queue snapshot entries of size @entry_size(IN)
1370 * into @snapshot_buf_ptr if @num_queues(IN) > 0.
1371 * Otherwise return @num_queues(OUT) queue snapshot entries that exist.
1373 struct kfd_ioctl_dbg_trap_queue_snapshot_args {
1374 __u64 exception_mask;
1375 __u64 snapshot_buf_ptr;
1381 * kfd_ioctl_dbg_trap_get_device_snapshot_args
1383 * Arguments for KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT
1384 * Get device information.
1386 * @exception_mask (IN) - exceptions raised to clear
1387 * @snapshot_buf_ptr (IN) - pointer to snapshot buffer (see kfd_dbg_device_info_entry)
1388 * @num_devices (IN/OUT) - number of debug devices to snapshot
1389 * The debugger specifies the size of the array allocated in @num_devices.
1390 * KFD returns the number of devices that actually existed. If this is
1391 * larger than the size specified by the debugger, KFD will not overflow
1392 * the array allocated by the debugger.
1394 * @entry_size (IN/OUT) - size per entry in bytes
1395 * The debugger specifies sizeof(struct kfd_dbg_device_info_entry) in
1396 * @entry_size. KFD returns the number of bytes actually populated. The
1397 * debugger should use KFD_IOCTL_MINOR_VERSION to determine, which fields
1398 * in struct kfd_dbg_device_info_entry are valid. This allows growing the
1399 * ABI in a backwards compatible manner.
1400 * Note that entry_size(IN) should still be used to stride the snapshot buffer in the
1401 * event that it's larger than actual kfd_dbg_device_info_entry.
1403 * Generic errors apply (see kfd_dbg_trap_operations).
1404 * Return - 0 on SUCCESS.
1405 * Copies @num_devices(IN) device snapshot entries of size @entry_size(IN)
1406 * into @snapshot_buf_ptr if @num_devices(IN) > 0.
1407 * Otherwise return @num_devices(OUT) queue snapshot entries that exist.
1409 struct kfd_ioctl_dbg_trap_device_snapshot_args {
1410 __u64 exception_mask;
1411 __u64 snapshot_buf_ptr;
1417 * kfd_ioctl_dbg_trap_args
1419 * Arguments to debug target process.
1421 * @pid - target process to debug
1422 * @op - debug operation (see kfd_dbg_trap_operations)
1424 * @op determines which union struct args to use.
1425 * Refer to kern docs for each kfd_ioctl_dbg_trap_*_args struct.
1427 struct kfd_ioctl_dbg_trap_args {
1432 struct kfd_ioctl_dbg_trap_enable_args enable;
1433 struct kfd_ioctl_dbg_trap_send_runtime_event_args send_runtime_event;
1434 struct kfd_ioctl_dbg_trap_set_exceptions_enabled_args set_exceptions_enabled;
1435 struct kfd_ioctl_dbg_trap_set_wave_launch_override_args launch_override;
1436 struct kfd_ioctl_dbg_trap_set_wave_launch_mode_args launch_mode;
1437 struct kfd_ioctl_dbg_trap_suspend_queues_args suspend_queues;
1438 struct kfd_ioctl_dbg_trap_resume_queues_args resume_queues;
1439 struct kfd_ioctl_dbg_trap_set_node_address_watch_args set_node_address_watch;
1440 struct kfd_ioctl_dbg_trap_clear_node_address_watch_args clear_node_address_watch;
1441 struct kfd_ioctl_dbg_trap_set_flags_args set_flags;
1442 struct kfd_ioctl_dbg_trap_query_debug_event_args query_debug_event;
1443 struct kfd_ioctl_dbg_trap_query_exception_info_args query_exception_info;
1444 struct kfd_ioctl_dbg_trap_queue_snapshot_args queue_snapshot;
1445 struct kfd_ioctl_dbg_trap_device_snapshot_args device_snapshot;
1449 #define AMDKFD_IOCTL_BASE 'K'
1450 #define AMDKFD_IO(nr) _IO(AMDKFD_IOCTL_BASE, nr)
1451 #define AMDKFD_IOR(nr, type) _IOR(AMDKFD_IOCTL_BASE, nr, type)
1452 #define AMDKFD_IOW(nr, type) _IOW(AMDKFD_IOCTL_BASE, nr, type)
1453 #define AMDKFD_IOWR(nr, type) _IOWR(AMDKFD_IOCTL_BASE, nr, type)
1455 #define AMDKFD_IOC_GET_VERSION \
1456 AMDKFD_IOR(0x01, struct kfd_ioctl_get_version_args)
1458 #define AMDKFD_IOC_CREATE_QUEUE \
1459 AMDKFD_IOWR(0x02, struct kfd_ioctl_create_queue_args)
1461 #define AMDKFD_IOC_DESTROY_QUEUE \
1462 AMDKFD_IOWR(0x03, struct kfd_ioctl_destroy_queue_args)
1464 #define AMDKFD_IOC_SET_MEMORY_POLICY \
1465 AMDKFD_IOW(0x04, struct kfd_ioctl_set_memory_policy_args)
1467 #define AMDKFD_IOC_GET_CLOCK_COUNTERS \
1468 AMDKFD_IOWR(0x05, struct kfd_ioctl_get_clock_counters_args)
1470 #define AMDKFD_IOC_GET_PROCESS_APERTURES \
1471 AMDKFD_IOR(0x06, struct kfd_ioctl_get_process_apertures_args)
1473 #define AMDKFD_IOC_UPDATE_QUEUE \
1474 AMDKFD_IOW(0x07, struct kfd_ioctl_update_queue_args)
1476 #define AMDKFD_IOC_CREATE_EVENT \
1477 AMDKFD_IOWR(0x08, struct kfd_ioctl_create_event_args)
1479 #define AMDKFD_IOC_DESTROY_EVENT \
1480 AMDKFD_IOW(0x09, struct kfd_ioctl_destroy_event_args)
1482 #define AMDKFD_IOC_SET_EVENT \
1483 AMDKFD_IOW(0x0A, struct kfd_ioctl_set_event_args)
1485 #define AMDKFD_IOC_RESET_EVENT \
1486 AMDKFD_IOW(0x0B, struct kfd_ioctl_reset_event_args)
1488 #define AMDKFD_IOC_WAIT_EVENTS \
1489 AMDKFD_IOWR(0x0C, struct kfd_ioctl_wait_events_args)
1491 #define AMDKFD_IOC_DBG_REGISTER_DEPRECATED \
1492 AMDKFD_IOW(0x0D, struct kfd_ioctl_dbg_register_args)
1494 #define AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED \
1495 AMDKFD_IOW(0x0E, struct kfd_ioctl_dbg_unregister_args)
1497 #define AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED \
1498 AMDKFD_IOW(0x0F, struct kfd_ioctl_dbg_address_watch_args)
1500 #define AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED \
1501 AMDKFD_IOW(0x10, struct kfd_ioctl_dbg_wave_control_args)
1503 #define AMDKFD_IOC_SET_SCRATCH_BACKING_VA \
1504 AMDKFD_IOWR(0x11, struct kfd_ioctl_set_scratch_backing_va_args)
1506 #define AMDKFD_IOC_GET_TILE_CONFIG \
1507 AMDKFD_IOWR(0x12, struct kfd_ioctl_get_tile_config_args)
1509 #define AMDKFD_IOC_SET_TRAP_HANDLER \
1510 AMDKFD_IOW(0x13, struct kfd_ioctl_set_trap_handler_args)
1512 #define AMDKFD_IOC_GET_PROCESS_APERTURES_NEW \
1514 struct kfd_ioctl_get_process_apertures_new_args)
1516 #define AMDKFD_IOC_ACQUIRE_VM \
1517 AMDKFD_IOW(0x15, struct kfd_ioctl_acquire_vm_args)
1519 #define AMDKFD_IOC_ALLOC_MEMORY_OF_GPU \
1520 AMDKFD_IOWR(0x16, struct kfd_ioctl_alloc_memory_of_gpu_args)
1522 #define AMDKFD_IOC_FREE_MEMORY_OF_GPU \
1523 AMDKFD_IOW(0x17, struct kfd_ioctl_free_memory_of_gpu_args)
1525 #define AMDKFD_IOC_MAP_MEMORY_TO_GPU \
1526 AMDKFD_IOWR(0x18, struct kfd_ioctl_map_memory_to_gpu_args)
1528 #define AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU \
1529 AMDKFD_IOWR(0x19, struct kfd_ioctl_unmap_memory_from_gpu_args)
1531 #define AMDKFD_IOC_SET_CU_MASK \
1532 AMDKFD_IOW(0x1A, struct kfd_ioctl_set_cu_mask_args)
1534 #define AMDKFD_IOC_GET_QUEUE_WAVE_STATE \
1535 AMDKFD_IOWR(0x1B, struct kfd_ioctl_get_queue_wave_state_args)
1537 #define AMDKFD_IOC_GET_DMABUF_INFO \
1538 AMDKFD_IOWR(0x1C, struct kfd_ioctl_get_dmabuf_info_args)
1540 #define AMDKFD_IOC_IMPORT_DMABUF \
1541 AMDKFD_IOWR(0x1D, struct kfd_ioctl_import_dmabuf_args)
1543 #define AMDKFD_IOC_ALLOC_QUEUE_GWS \
1544 AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
1546 #define AMDKFD_IOC_SMI_EVENTS \
1547 AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
1549 #define AMDKFD_IOC_SVM AMDKFD_IOWR(0x20, struct kfd_ioctl_svm_args)
1551 #define AMDKFD_IOC_SET_XNACK_MODE \
1552 AMDKFD_IOWR(0x21, struct kfd_ioctl_set_xnack_mode_args)
1554 #define AMDKFD_IOC_CRIU_OP \
1555 AMDKFD_IOWR(0x22, struct kfd_ioctl_criu_args)
1557 #define AMDKFD_IOC_AVAILABLE_MEMORY \
1558 AMDKFD_IOWR(0x23, struct kfd_ioctl_get_available_memory_args)
1560 #define AMDKFD_IOC_EXPORT_DMABUF \
1561 AMDKFD_IOWR(0x24, struct kfd_ioctl_export_dmabuf_args)
1563 #define AMDKFD_IOC_RUNTIME_ENABLE \
1564 AMDKFD_IOWR(0x25, struct kfd_ioctl_runtime_enable_args)
1566 #define AMDKFD_IOC_DBG_TRAP \
1567 AMDKFD_IOWR(0x26, struct kfd_ioctl_dbg_trap_args)
1569 #define AMDKFD_COMMAND_START 0x01
1570 #define AMDKFD_COMMAND_END 0x27