master
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 * - 1.18 - Rename pad in set_memory_policy_args to misc_process_flag
47 */
48#define KFD_IOCTL_MAJOR_VERSION 1
49#define KFD_IOCTL_MINOR_VERSION 18
50
51struct kfd_ioctl_get_version_args {
52 __u32 major_version; /* from KFD */
53 __u32 minor_version; /* from KFD */
54};
55
56/* For kfd_ioctl_create_queue_args.queue_type. */
57#define KFD_IOC_QUEUE_TYPE_COMPUTE 0x0
58#define KFD_IOC_QUEUE_TYPE_SDMA 0x1
59#define KFD_IOC_QUEUE_TYPE_COMPUTE_AQL 0x2
60#define KFD_IOC_QUEUE_TYPE_SDMA_XGMI 0x3
61#define KFD_IOC_QUEUE_TYPE_SDMA_BY_ENG_ID 0x4
62
63#define KFD_MAX_QUEUE_PERCENTAGE 100
64#define KFD_MAX_QUEUE_PRIORITY 15
65
66#define KFD_MIN_QUEUE_RING_SIZE 1024
67
68struct kfd_ioctl_create_queue_args {
69 __u64 ring_base_address; /* to KFD */
70 __u64 write_pointer_address; /* from KFD */
71 __u64 read_pointer_address; /* from KFD */
72 __u64 doorbell_offset; /* from KFD */
73
74 __u32 ring_size; /* to KFD */
75 __u32 gpu_id; /* to KFD */
76 __u32 queue_type; /* to KFD */
77 __u32 queue_percentage; /* to KFD */
78 __u32 queue_priority; /* to KFD */
79 __u32 queue_id; /* from KFD */
80
81 __u64 eop_buffer_address; /* to KFD */
82 __u64 eop_buffer_size; /* to KFD */
83 __u64 ctx_save_restore_address; /* to KFD */
84 __u32 ctx_save_restore_size; /* to KFD */
85 __u32 ctl_stack_size; /* to KFD */
86 __u32 sdma_engine_id; /* to KFD */
87 __u32 pad;
88};
89
90struct kfd_ioctl_destroy_queue_args {
91 __u32 queue_id; /* to KFD */
92 __u32 pad;
93};
94
95struct kfd_ioctl_update_queue_args {
96 __u64 ring_base_address; /* to KFD */
97
98 __u32 queue_id; /* to KFD */
99 __u32 ring_size; /* to KFD */
100 __u32 queue_percentage; /* to KFD */
101 __u32 queue_priority; /* to KFD */
102};
103
104struct kfd_ioctl_set_cu_mask_args {
105 __u32 queue_id; /* to KFD */
106 __u32 num_cu_mask; /* to KFD */
107 __u64 cu_mask_ptr; /* to KFD */
108};
109
110struct kfd_ioctl_get_queue_wave_state_args {
111 __u64 ctl_stack_address; /* to KFD */
112 __u32 ctl_stack_used_size; /* from KFD */
113 __u32 save_area_used_size; /* from KFD */
114 __u32 queue_id; /* to KFD */
115 __u32 pad;
116};
117
118struct kfd_ioctl_get_available_memory_args {
119 __u64 available; /* from KFD */
120 __u32 gpu_id; /* to KFD */
121 __u32 pad;
122};
123
124struct kfd_dbg_device_info_entry {
125 __u64 exception_status;
126 __u64 lds_base;
127 __u64 lds_limit;
128 __u64 scratch_base;
129 __u64 scratch_limit;
130 __u64 gpuvm_base;
131 __u64 gpuvm_limit;
132 __u32 gpu_id;
133 __u32 location_id;
134 __u32 vendor_id;
135 __u32 device_id;
136 __u32 revision_id;
137 __u32 subsystem_vendor_id;
138 __u32 subsystem_device_id;
139 __u32 fw_version;
140 __u32 gfx_target_version;
141 __u32 simd_count;
142 __u32 max_waves_per_simd;
143 __u32 array_count;
144 __u32 simd_arrays_per_engine;
145 __u32 num_xcc;
146 __u32 capability;
147 __u32 debug_prop;
148};
149
150/* For kfd_ioctl_set_memory_policy_args.default_policy and alternate_policy */
151#define KFD_IOC_CACHE_POLICY_COHERENT 0
152#define KFD_IOC_CACHE_POLICY_NONCOHERENT 1
153
154/* Misc. per process flags */
155#define KFD_PROC_FLAG_MFMA_HIGH_PRECISION (1 << 0)
156
157struct kfd_ioctl_set_memory_policy_args {
158 __u64 alternate_aperture_base; /* to KFD */
159 __u64 alternate_aperture_size; /* to KFD */
160
161 __u32 gpu_id; /* to KFD */
162 __u32 default_policy; /* to KFD */
163 __u32 alternate_policy; /* to KFD */
164 __u32 misc_process_flag; /* to KFD */
165};
166
167/*
168 * All counters are monotonic. They are used for profiling of compute jobs.
169 * The profiling is done by userspace.
170 *
171 * In case of GPU reset, the counter should not be affected.
172 */
173
174struct kfd_ioctl_get_clock_counters_args {
175 __u64 gpu_clock_counter; /* from KFD */
176 __u64 cpu_clock_counter; /* from KFD */
177 __u64 system_clock_counter; /* from KFD */
178 __u64 system_clock_freq; /* from KFD */
179
180 __u32 gpu_id; /* to KFD */
181 __u32 pad;
182};
183
184struct kfd_process_device_apertures {
185 __u64 lds_base; /* from KFD */
186 __u64 lds_limit; /* from KFD */
187 __u64 scratch_base; /* from KFD */
188 __u64 scratch_limit; /* from KFD */
189 __u64 gpuvm_base; /* from KFD */
190 __u64 gpuvm_limit; /* from KFD */
191 __u32 gpu_id; /* from KFD */
192 __u32 pad;
193};
194
195/*
196 * AMDKFD_IOC_GET_PROCESS_APERTURES is deprecated. Use
197 * AMDKFD_IOC_GET_PROCESS_APERTURES_NEW instead, which supports an
198 * unlimited number of GPUs.
199 */
200#define NUM_OF_SUPPORTED_GPUS 7
201struct kfd_ioctl_get_process_apertures_args {
202 struct kfd_process_device_apertures
203 process_apertures[NUM_OF_SUPPORTED_GPUS];/* from KFD */
204
205 /* from KFD, should be in the range [1 - NUM_OF_SUPPORTED_GPUS] */
206 __u32 num_of_nodes;
207 __u32 pad;
208};
209
210struct kfd_ioctl_get_process_apertures_new_args {
211 /* User allocated. Pointer to struct kfd_process_device_apertures
212 * filled in by Kernel
213 */
214 __u64 kfd_process_device_apertures_ptr;
215 /* to KFD - indicates amount of memory present in
216 * kfd_process_device_apertures_ptr
217 * from KFD - Number of entries filled by KFD.
218 */
219 __u32 num_of_nodes;
220 __u32 pad;
221};
222
223#define MAX_ALLOWED_NUM_POINTS 100
224#define MAX_ALLOWED_AW_BUFF_SIZE 4096
225#define MAX_ALLOWED_WAC_BUFF_SIZE 128
226
227struct kfd_ioctl_dbg_register_args {
228 __u32 gpu_id; /* to KFD */
229 __u32 pad;
230};
231
232struct kfd_ioctl_dbg_unregister_args {
233 __u32 gpu_id; /* to KFD */
234 __u32 pad;
235};
236
237struct kfd_ioctl_dbg_address_watch_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
243struct kfd_ioctl_dbg_wave_control_args {
244 __u64 content_ptr; /* a pointer to the actual content */
245 __u32 gpu_id; /* to KFD */
246 __u32 buf_size_in_bytes; /*including gpu_id and buf_size */
247};
248
249#define KFD_INVALID_FD 0xffffffff
250
251/* Matching HSA_EVENTTYPE */
252#define KFD_IOC_EVENT_SIGNAL 0
253#define KFD_IOC_EVENT_NODECHANGE 1
254#define KFD_IOC_EVENT_DEVICESTATECHANGE 2
255#define KFD_IOC_EVENT_HW_EXCEPTION 3
256#define KFD_IOC_EVENT_SYSTEM_EVENT 4
257#define KFD_IOC_EVENT_DEBUG_EVENT 5
258#define KFD_IOC_EVENT_PROFILE_EVENT 6
259#define KFD_IOC_EVENT_QUEUE_EVENT 7
260#define KFD_IOC_EVENT_MEMORY 8
261
262#define KFD_IOC_WAIT_RESULT_COMPLETE 0
263#define KFD_IOC_WAIT_RESULT_TIMEOUT 1
264#define KFD_IOC_WAIT_RESULT_FAIL 2
265
266#define KFD_SIGNAL_EVENT_LIMIT 4096
267
268/* For kfd_event_data.hw_exception_data.reset_type. */
269#define KFD_HW_EXCEPTION_WHOLE_GPU_RESET 0
270#define KFD_HW_EXCEPTION_PER_ENGINE_RESET 1
271
272/* For kfd_event_data.hw_exception_data.reset_cause. */
273#define KFD_HW_EXCEPTION_GPU_HANG 0
274#define KFD_HW_EXCEPTION_ECC 1
275
276/* For kfd_hsa_memory_exception_data.ErrorType */
277#define KFD_MEM_ERR_NO_RAS 0
278#define KFD_MEM_ERR_SRAM_ECC 1
279#define KFD_MEM_ERR_POISON_CONSUMED 2
280#define KFD_MEM_ERR_GPU_HANG 3
281
282struct kfd_ioctl_create_event_args {
283 __u64 event_page_offset; /* from KFD */
284 __u32 event_trigger_data; /* from KFD - signal events only */
285 __u32 event_type; /* to KFD */
286 __u32 auto_reset; /* to KFD */
287 __u32 node_id; /* to KFD - only valid for certain
288 event types */
289 __u32 event_id; /* from KFD */
290 __u32 event_slot_index; /* from KFD */
291};
292
293struct kfd_ioctl_destroy_event_args {
294 __u32 event_id; /* to KFD */
295 __u32 pad;
296};
297
298struct kfd_ioctl_set_event_args {
299 __u32 event_id; /* to KFD */
300 __u32 pad;
301};
302
303struct kfd_ioctl_reset_event_args {
304 __u32 event_id; /* to KFD */
305 __u32 pad;
306};
307
308struct kfd_memory_exception_failure {
309 __u32 NotPresent; /* Page not present or supervisor privilege */
310 __u32 ReadOnly; /* Write access to a read-only page */
311 __u32 NoExecute; /* Execute access to a page marked NX */
312 __u32 imprecise; /* Can't determine the exact fault address */
313};
314
315/* memory exception data */
316struct kfd_hsa_memory_exception_data {
317 struct kfd_memory_exception_failure failure;
318 __u64 va;
319 __u32 gpu_id;
320 __u32 ErrorType; /* 0 = no RAS error,
321 * 1 = ECC_SRAM,
322 * 2 = Link_SYNFLOOD (poison),
323 * 3 = GPU hang (not attributable to a specific cause),
324 * other values reserved
325 */
326};
327
328/* hw exception data */
329struct kfd_hsa_hw_exception_data {
330 __u32 reset_type;
331 __u32 reset_cause;
332 __u32 memory_lost;
333 __u32 gpu_id;
334};
335
336/* hsa signal event data */
337struct kfd_hsa_signal_event_data {
338 __u64 last_event_age; /* to and from KFD */
339};
340
341/* Event data */
342struct kfd_event_data {
343 union {
344 /* From KFD */
345 struct kfd_hsa_memory_exception_data memory_exception_data;
346 struct kfd_hsa_hw_exception_data hw_exception_data;
347 /* To and From KFD */
348 struct kfd_hsa_signal_event_data signal_event_data;
349 };
350 __u64 kfd_event_data_ext; /* pointer to an extension structure
351 for future exception types */
352 __u32 event_id; /* to KFD */
353 __u32 pad;
354};
355
356struct kfd_ioctl_wait_events_args {
357 __u64 events_ptr; /* pointed to struct
358 kfd_event_data array, to KFD */
359 __u32 num_events; /* to KFD */
360 __u32 wait_for_all; /* to KFD */
361 __u32 timeout; /* to KFD */
362 __u32 wait_result; /* from KFD */
363};
364
365struct kfd_ioctl_set_scratch_backing_va_args {
366 __u64 va_addr; /* to KFD */
367 __u32 gpu_id; /* to KFD */
368 __u32 pad;
369};
370
371struct kfd_ioctl_get_tile_config_args {
372 /* to KFD: pointer to tile array */
373 __u64 tile_config_ptr;
374 /* to KFD: pointer to macro tile array */
375 __u64 macro_tile_config_ptr;
376 /* to KFD: array size allocated by user mode
377 * from KFD: array size filled by kernel
378 */
379 __u32 num_tile_configs;
380 /* to KFD: array size allocated by user mode
381 * from KFD: array size filled by kernel
382 */
383 __u32 num_macro_tile_configs;
384
385 __u32 gpu_id; /* to KFD */
386 __u32 gb_addr_config; /* from KFD */
387 __u32 num_banks; /* from KFD */
388 __u32 num_ranks; /* from KFD */
389 /* struct size can be extended later if needed
390 * without breaking ABI compatibility
391 */
392};
393
394struct kfd_ioctl_set_trap_handler_args {
395 __u64 tba_addr; /* to KFD */
396 __u64 tma_addr; /* to KFD */
397 __u32 gpu_id; /* to KFD */
398 __u32 pad;
399};
400
401struct kfd_ioctl_acquire_vm_args {
402 __u32 drm_fd; /* to KFD */
403 __u32 gpu_id; /* to KFD */
404};
405
406/* Allocation flags: memory types */
407#define KFD_IOC_ALLOC_MEM_FLAGS_VRAM (1 << 0)
408#define KFD_IOC_ALLOC_MEM_FLAGS_GTT (1 << 1)
409#define KFD_IOC_ALLOC_MEM_FLAGS_USERPTR (1 << 2)
410#define KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL (1 << 3)
411#define KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP (1 << 4)
412/* Allocation flags: attributes/access options */
413#define KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE (1 << 31)
414#define KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE (1 << 30)
415#define KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC (1 << 29)
416#define KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE (1 << 28)
417#define KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM (1 << 27)
418#define KFD_IOC_ALLOC_MEM_FLAGS_COHERENT (1 << 26)
419#define KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED (1 << 25)
420#define KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT (1 << 24)
421#define KFD_IOC_ALLOC_MEM_FLAGS_CONTIGUOUS (1 << 23)
422
423/* Allocate memory for later SVM (shared virtual memory) mapping.
424 *
425 * @va_addr: virtual address of the memory to be allocated
426 * all later mappings on all GPUs will use this address
427 * @size: size in bytes
428 * @handle: buffer handle returned to user mode, used to refer to
429 * this allocation for mapping, unmapping and freeing
430 * @mmap_offset: for CPU-mapping the allocation by mmapping a render node
431 * for userptrs this is overloaded to specify the CPU address
432 * @gpu_id: device identifier
433 * @flags: memory type and attributes. See KFD_IOC_ALLOC_MEM_FLAGS above
434 */
435struct kfd_ioctl_alloc_memory_of_gpu_args {
436 __u64 va_addr; /* to KFD */
437 __u64 size; /* to KFD */
438 __u64 handle; /* from KFD */
439 __u64 mmap_offset; /* to KFD (userptr), from KFD (mmap offset) */
440 __u32 gpu_id; /* to KFD */
441 __u32 flags;
442};
443
444/* Free memory allocated with kfd_ioctl_alloc_memory_of_gpu
445 *
446 * @handle: memory handle returned by alloc
447 */
448struct kfd_ioctl_free_memory_of_gpu_args {
449 __u64 handle; /* to KFD */
450};
451
452/* Map memory to one or more GPUs
453 *
454 * @handle: memory handle returned by alloc
455 * @device_ids_array_ptr: array of gpu_ids (__u32 per device)
456 * @n_devices: number of devices in the array
457 * @n_success: number of devices mapped successfully
458 *
459 * @n_success returns information to the caller how many devices from
460 * the start of the array have mapped the buffer successfully. It can
461 * be passed into a subsequent retry call to skip those devices. For
462 * the first call the caller should initialize it to 0.
463 *
464 * If the ioctl completes with return code 0 (success), n_success ==
465 * n_devices.
466 */
467struct kfd_ioctl_map_memory_to_gpu_args {
468 __u64 handle; /* to KFD */
469 __u64 device_ids_array_ptr; /* to KFD */
470 __u32 n_devices; /* to KFD */
471 __u32 n_success; /* to/from KFD */
472};
473
474/* Unmap memory from one or more GPUs
475 *
476 * same arguments as for mapping
477 */
478struct kfd_ioctl_unmap_memory_from_gpu_args {
479 __u64 handle; /* to KFD */
480 __u64 device_ids_array_ptr; /* to KFD */
481 __u32 n_devices; /* to KFD */
482 __u32 n_success; /* to/from KFD */
483};
484
485/* Allocate GWS for specific queue
486 *
487 * @queue_id: queue's id that GWS is allocated for
488 * @num_gws: how many GWS to allocate
489 * @first_gws: index of the first GWS allocated.
490 * only support contiguous GWS allocation
491 */
492struct kfd_ioctl_alloc_queue_gws_args {
493 __u32 queue_id; /* to KFD */
494 __u32 num_gws; /* to KFD */
495 __u32 first_gws; /* from KFD */
496 __u32 pad;
497};
498
499struct kfd_ioctl_get_dmabuf_info_args {
500 __u64 size; /* from KFD */
501 __u64 metadata_ptr; /* to KFD */
502 __u32 metadata_size; /* to KFD (space allocated by user)
503 * from KFD (actual metadata size)
504 */
505 __u32 gpu_id; /* from KFD */
506 __u32 flags; /* from KFD (KFD_IOC_ALLOC_MEM_FLAGS) */
507 __u32 dmabuf_fd; /* to KFD */
508};
509
510struct kfd_ioctl_import_dmabuf_args {
511 __u64 va_addr; /* to KFD */
512 __u64 handle; /* from KFD */
513 __u32 gpu_id; /* to KFD */
514 __u32 dmabuf_fd; /* to KFD */
515};
516
517struct kfd_ioctl_export_dmabuf_args {
518 __u64 handle; /* to KFD */
519 __u32 flags; /* to KFD */
520 __u32 dmabuf_fd; /* from KFD */
521};
522
523/*
524 * KFD SMI(System Management Interface) events
525 */
526enum kfd_smi_event {
527 KFD_SMI_EVENT_NONE = 0, /* not used */
528 KFD_SMI_EVENT_VMFAULT = 1, /* event start counting at 1 */
529 KFD_SMI_EVENT_THERMAL_THROTTLE = 2,
530 KFD_SMI_EVENT_GPU_PRE_RESET = 3,
531 KFD_SMI_EVENT_GPU_POST_RESET = 4,
532 KFD_SMI_EVENT_MIGRATE_START = 5,
533 KFD_SMI_EVENT_MIGRATE_END = 6,
534 KFD_SMI_EVENT_PAGE_FAULT_START = 7,
535 KFD_SMI_EVENT_PAGE_FAULT_END = 8,
536 KFD_SMI_EVENT_QUEUE_EVICTION = 9,
537 KFD_SMI_EVENT_QUEUE_RESTORE = 10,
538 KFD_SMI_EVENT_UNMAP_FROM_GPU = 11,
539 KFD_SMI_EVENT_PROCESS_START = 12,
540 KFD_SMI_EVENT_PROCESS_END = 13,
541
542 /*
543 * max event number, as a flag bit to get events from all processes,
544 * this requires super user permission, otherwise will not be able to
545 * receive event from any process. Without this flag to receive events
546 * from same process.
547 */
548 KFD_SMI_EVENT_ALL_PROCESS = 64
549};
550
551/* The reason of the page migration event */
552enum KFD_MIGRATE_TRIGGERS {
553 KFD_MIGRATE_TRIGGER_PREFETCH, /* Prefetch to GPU VRAM or system memory */
554 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, /* GPU page fault recover */
555 KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU, /* CPU page fault recover */
556 KFD_MIGRATE_TRIGGER_TTM_EVICTION /* TTM eviction */
557};
558
559/* The reason of user queue evition event */
560enum KFD_QUEUE_EVICTION_TRIGGERS {
561 KFD_QUEUE_EVICTION_TRIGGER_SVM, /* SVM buffer migration */
562 KFD_QUEUE_EVICTION_TRIGGER_USERPTR, /* userptr movement */
563 KFD_QUEUE_EVICTION_TRIGGER_TTM, /* TTM move buffer */
564 KFD_QUEUE_EVICTION_TRIGGER_SUSPEND, /* GPU suspend */
565 KFD_QUEUE_EVICTION_CRIU_CHECKPOINT, /* CRIU checkpoint */
566 KFD_QUEUE_EVICTION_CRIU_RESTORE /* CRIU restore */
567};
568
569/* The reason of unmap buffer from GPU event */
570enum KFD_SVM_UNMAP_TRIGGERS {
571 KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY, /* MMU notifier CPU buffer movement */
572 KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE,/* MMU notifier page migration */
573 KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU /* Unmap to free the buffer */
574};
575
576#define KFD_SMI_EVENT_MASK_FROM_INDEX(i) (1ULL << ((i) - 1))
577#define KFD_SMI_EVENT_MSG_SIZE 96
578
579struct kfd_ioctl_smi_events_args {
580 __u32 gpuid; /* to KFD */
581 __u32 anon_fd; /* from KFD */
582};
583
584/*
585 * SVM event tracing via SMI system management interface
586 *
587 * Open event file descriptor
588 * use ioctl AMDKFD_IOC_SMI_EVENTS, pass in gpuid and return a anonymous file
589 * descriptor to receive SMI events.
590 * If calling with sudo permission, then file descriptor can be used to receive
591 * SVM events from all processes, otherwise, to only receive SVM events of same
592 * process.
593 *
594 * To enable the SVM event
595 * Write event file descriptor with KFD_SMI_EVENT_MASK_FROM_INDEX(event) bitmap
596 * mask to start record the event to the kfifo, use bitmap mask combination
597 * for multiple events. New event mask will overwrite the previous event mask.
598 * KFD_SMI_EVENT_MASK_FROM_INDEX(KFD_SMI_EVENT_ALL_PROCESS) bit requires sudo
599 * permisson to receive SVM events from all process.
600 *
601 * To receive the event
602 * Application can poll file descriptor to wait for the events, then read event
603 * from the file into a buffer. Each event is one line string message, starting
604 * with the event id, then the event specific information.
605 *
606 * To decode event information
607 * The following event format string macro can be used with sscanf to decode
608 * the specific event information.
609 * event triggers: the reason to generate the event, defined as enum for unmap,
610 * eviction and migrate events.
611 * node, from, to, prefetch_loc, preferred_loc: GPU ID, or 0 for system memory.
612 * addr: user mode address, in pages
613 * size: in pages
614 * pid: the process ID to generate the event
615 * ns: timestamp in nanosecond-resolution, starts at system boot time but
616 * stops during suspend
617 * migrate_update: GPU page fault is recovered by 'M' for migrate, 'U' for update
618 * rw: 'W' for write page fault, 'R' for read page fault
619 * rescheduled: 'R' if the queue restore failed and rescheduled to try again
620 * error_code: migrate failure error code, 0 if no error
621 */
622#define KFD_EVENT_FMT_UPDATE_GPU_RESET(reset_seq_num, reset_cause)\
623 "%x %s\n", (reset_seq_num), (reset_cause)
624
625#define KFD_EVENT_FMT_THERMAL_THROTTLING(bitmask, counter)\
626 "%llx:%llx\n", (bitmask), (counter)
627
628#define KFD_EVENT_FMT_VMFAULT(pid, task_name)\
629 "%x:%s\n", (pid), (task_name)
630
631#define KFD_EVENT_FMT_PAGEFAULT_START(ns, pid, addr, node, rw)\
632 "%lld -%d @%lx(%x) %c\n", (ns), (pid), (addr), (node), (rw)
633
634#define KFD_EVENT_FMT_PAGEFAULT_END(ns, pid, addr, node, migrate_update)\
635 "%lld -%d @%lx(%x) %c\n", (ns), (pid), (addr), (node), (migrate_update)
636
637#define KFD_EVENT_FMT_MIGRATE_START(ns, pid, start, size, from, to, prefetch_loc,\
638 preferred_loc, migrate_trigger)\
639 "%lld -%d @%lx(%lx) %x->%x %x:%x %d\n", (ns), (pid), (start), (size),\
640 (from), (to), (prefetch_loc), (preferred_loc), (migrate_trigger)
641
642#define KFD_EVENT_FMT_MIGRATE_END(ns, pid, start, size, from, to, migrate_trigger, error_code) \
643 "%lld -%d @%lx(%lx) %x->%x %d %d\n", (ns), (pid), (start), (size),\
644 (from), (to), (migrate_trigger), (error_code)
645
646#define KFD_EVENT_FMT_QUEUE_EVICTION(ns, pid, node, evict_trigger)\
647 "%lld -%d %x %d\n", (ns), (pid), (node), (evict_trigger)
648
649#define KFD_EVENT_FMT_QUEUE_RESTORE(ns, pid, node, rescheduled)\
650 "%lld -%d %x %c\n", (ns), (pid), (node), (rescheduled)
651
652#define KFD_EVENT_FMT_UNMAP_FROM_GPU(ns, pid, addr, size, node, unmap_trigger)\
653 "%lld -%d @%lx(%lx) %x %d\n", (ns), (pid), (addr), (size),\
654 (node), (unmap_trigger)
655
656#define KFD_EVENT_FMT_PROCESS(pid, task_name)\
657 "%x %s\n", (pid), (task_name)
658
659/**************************************************************************************************
660 * CRIU IOCTLs (Checkpoint Restore In Userspace)
661 *
662 * When checkpointing a process, the userspace application will perform:
663 * 1. PROCESS_INFO op to determine current process information. This pauses execution and evicts
664 * all the queues.
665 * 2. CHECKPOINT op to checkpoint process contents (BOs, queues, events, svm-ranges)
666 * 3. UNPAUSE op to un-evict all the queues
667 *
668 * When restoring a process, the CRIU userspace application will perform:
669 *
670 * 1. RESTORE op to restore process contents
671 * 2. RESUME op to start the process
672 *
673 * Note: Queues are forced into an evicted state after a successful PROCESS_INFO. User
674 * application needs to perform an UNPAUSE operation after calling PROCESS_INFO.
675 */
676
677enum kfd_criu_op {
678 KFD_CRIU_OP_PROCESS_INFO,
679 KFD_CRIU_OP_CHECKPOINT,
680 KFD_CRIU_OP_UNPAUSE,
681 KFD_CRIU_OP_RESTORE,
682 KFD_CRIU_OP_RESUME,
683};
684
685/**
686 * kfd_ioctl_criu_args - Arguments perform CRIU operation
687 * @devices: [in/out] User pointer to memory location for devices information.
688 * This is an array of type kfd_criu_device_bucket.
689 * @bos: [in/out] User pointer to memory location for BOs information
690 * This is an array of type kfd_criu_bo_bucket.
691 * @priv_data: [in/out] User pointer to memory location for private data
692 * @priv_data_size: [in/out] Size of priv_data in bytes
693 * @num_devices: [in/out] Number of GPUs used by process. Size of @devices array.
694 * @num_bos [in/out] Number of BOs used by process. Size of @bos array.
695 * @num_objects: [in/out] Number of objects used by process. Objects are opaque to
696 * user application.
697 * @pid: [in/out] PID of the process being checkpointed
698 * @op [in] Type of operation (kfd_criu_op)
699 *
700 * Return: 0 on success, -errno on failure
701 */
702struct kfd_ioctl_criu_args {
703 __u64 devices; /* Used during ops: CHECKPOINT, RESTORE */
704 __u64 bos; /* Used during ops: CHECKPOINT, RESTORE */
705 __u64 priv_data; /* Used during ops: CHECKPOINT, RESTORE */
706 __u64 priv_data_size; /* Used during ops: PROCESS_INFO, RESTORE */
707 __u32 num_devices; /* Used during ops: PROCESS_INFO, RESTORE */
708 __u32 num_bos; /* Used during ops: PROCESS_INFO, RESTORE */
709 __u32 num_objects; /* Used during ops: PROCESS_INFO, RESTORE */
710 __u32 pid; /* Used during ops: PROCESS_INFO, RESUME */
711 __u32 op;
712};
713
714struct kfd_criu_device_bucket {
715 __u32 user_gpu_id;
716 __u32 actual_gpu_id;
717 __u32 drm_fd;
718 __u32 pad;
719};
720
721struct kfd_criu_bo_bucket {
722 __u64 addr;
723 __u64 size;
724 __u64 offset;
725 __u64 restored_offset; /* During restore, updated offset for BO */
726 __u32 gpu_id; /* This is the user_gpu_id */
727 __u32 alloc_flags;
728 __u32 dmabuf_fd;
729 __u32 pad;
730};
731
732/* CRIU IOCTLs - END */
733/**************************************************************************************************/
734
735/* Register offset inside the remapped mmio page
736 */
737enum kfd_mmio_remap {
738 KFD_MMIO_REMAP_HDP_MEM_FLUSH_CNTL = 0,
739 KFD_MMIO_REMAP_HDP_REG_FLUSH_CNTL = 4,
740};
741
742/* Guarantee host access to memory */
743#define KFD_IOCTL_SVM_FLAG_HOST_ACCESS 0x00000001
744/* Fine grained coherency between all devices with access */
745#define KFD_IOCTL_SVM_FLAG_COHERENT 0x00000002
746/* Use any GPU in same hive as preferred device */
747#define KFD_IOCTL_SVM_FLAG_HIVE_LOCAL 0x00000004
748/* GPUs only read, allows replication */
749#define KFD_IOCTL_SVM_FLAG_GPU_RO 0x00000008
750/* Allow execution on GPU */
751#define KFD_IOCTL_SVM_FLAG_GPU_EXEC 0x00000010
752/* GPUs mostly read, may allow similar optimizations as RO, but writes fault */
753#define KFD_IOCTL_SVM_FLAG_GPU_READ_MOSTLY 0x00000020
754/* Keep GPU memory mapping always valid as if XNACK is disable */
755#define KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED 0x00000040
756/* Fine grained coherency between all devices using device-scope atomics */
757#define KFD_IOCTL_SVM_FLAG_EXT_COHERENT 0x00000080
758
759/**
760 * kfd_ioctl_svm_op - SVM ioctl operations
761 *
762 * @KFD_IOCTL_SVM_OP_SET_ATTR: Modify one or more attributes
763 * @KFD_IOCTL_SVM_OP_GET_ATTR: Query one or more attributes
764 */
765enum kfd_ioctl_svm_op {
766 KFD_IOCTL_SVM_OP_SET_ATTR,
767 KFD_IOCTL_SVM_OP_GET_ATTR
768};
769
770/** kfd_ioctl_svm_location - Enum for preferred and prefetch locations
771 *
772 * GPU IDs are used to specify GPUs as preferred and prefetch locations.
773 * Below definitions are used for system memory or for leaving the preferred
774 * location unspecified.
775 */
776enum kfd_ioctl_svm_location {
777 KFD_IOCTL_SVM_LOCATION_SYSMEM = 0,
778 KFD_IOCTL_SVM_LOCATION_UNDEFINED = 0xffffffff
779};
780
781/**
782 * kfd_ioctl_svm_attr_type - SVM attribute types
783 *
784 * @KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: gpuid of the preferred location, 0 for
785 * system memory
786 * @KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: gpuid of the prefetch location, 0 for
787 * system memory. Setting this triggers an
788 * immediate prefetch (migration).
789 * @KFD_IOCTL_SVM_ATTR_ACCESS:
790 * @KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
791 * @KFD_IOCTL_SVM_ATTR_NO_ACCESS: specify memory access for the gpuid given
792 * by the attribute value
793 * @KFD_IOCTL_SVM_ATTR_SET_FLAGS: bitmask of flags to set (see
794 * KFD_IOCTL_SVM_FLAG_...)
795 * @KFD_IOCTL_SVM_ATTR_CLR_FLAGS: bitmask of flags to clear
796 * @KFD_IOCTL_SVM_ATTR_GRANULARITY: migration granularity
797 * (log2 num pages)
798 */
799enum kfd_ioctl_svm_attr_type {
800 KFD_IOCTL_SVM_ATTR_PREFERRED_LOC,
801 KFD_IOCTL_SVM_ATTR_PREFETCH_LOC,
802 KFD_IOCTL_SVM_ATTR_ACCESS,
803 KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE,
804 KFD_IOCTL_SVM_ATTR_NO_ACCESS,
805 KFD_IOCTL_SVM_ATTR_SET_FLAGS,
806 KFD_IOCTL_SVM_ATTR_CLR_FLAGS,
807 KFD_IOCTL_SVM_ATTR_GRANULARITY
808};
809
810/**
811 * kfd_ioctl_svm_attribute - Attributes as pairs of type and value
812 *
813 * The meaning of the @value depends on the attribute type.
814 *
815 * @type: attribute type (see enum @kfd_ioctl_svm_attr_type)
816 * @value: attribute value
817 */
818struct kfd_ioctl_svm_attribute {
819 __u32 type;
820 __u32 value;
821};
822
823/**
824 * kfd_ioctl_svm_args - Arguments for SVM ioctl
825 *
826 * @op specifies the operation to perform (see enum
827 * @kfd_ioctl_svm_op). @start_addr and @size are common for all
828 * operations.
829 *
830 * A variable number of attributes can be given in @attrs.
831 * @nattr specifies the number of attributes. New attributes can be
832 * added in the future without breaking the ABI. If unknown attributes
833 * are given, the function returns -EINVAL.
834 *
835 * @KFD_IOCTL_SVM_OP_SET_ATTR sets attributes for a virtual address
836 * range. It may overlap existing virtual address ranges. If it does,
837 * the existing ranges will be split such that the attribute changes
838 * only apply to the specified address range.
839 *
840 * @KFD_IOCTL_SVM_OP_GET_ATTR returns the intersection of attributes
841 * over all memory in the given range and returns the result as the
842 * attribute value. If different pages have different preferred or
843 * prefetch locations, 0xffffffff will be returned for
844 * @KFD_IOCTL_SVM_ATTR_PREFERRED_LOC or
845 * @KFD_IOCTL_SVM_ATTR_PREFETCH_LOC resepctively. For
846 * @KFD_IOCTL_SVM_ATTR_SET_FLAGS, flags of all pages will be
847 * aggregated by bitwise AND. That means, a flag will be set in the
848 * output, if that flag is set for all pages in the range. For
849 * @KFD_IOCTL_SVM_ATTR_CLR_FLAGS, flags of all pages will be
850 * aggregated by bitwise NOR. That means, a flag will be set in the
851 * output, if that flag is clear for all pages in the range.
852 * The minimum migration granularity throughout the range will be
853 * returned for @KFD_IOCTL_SVM_ATTR_GRANULARITY.
854 *
855 * Querying of accessibility attributes works by initializing the
856 * attribute type to @KFD_IOCTL_SVM_ATTR_ACCESS and the value to the
857 * GPUID being queried. Multiple attributes can be given to allow
858 * querying multiple GPUIDs. The ioctl function overwrites the
859 * attribute type to indicate the access for the specified GPU.
860 */
861struct kfd_ioctl_svm_args {
862 __u64 start_addr;
863 __u64 size;
864 __u32 op;
865 __u32 nattr;
866 /* Variable length array of attributes */
867 struct kfd_ioctl_svm_attribute attrs[];
868};
869
870/**
871 * kfd_ioctl_set_xnack_mode_args - Arguments for set_xnack_mode
872 *
873 * @xnack_enabled: [in/out] Whether to enable XNACK mode for this process
874 *
875 * @xnack_enabled indicates whether recoverable page faults should be
876 * enabled for the current process. 0 means disabled, positive means
877 * enabled, negative means leave unchanged. If enabled, virtual address
878 * translations on GFXv9 and later AMD GPUs can return XNACK and retry
879 * the access until a valid PTE is available. This is used to implement
880 * device page faults.
881 *
882 * On output, @xnack_enabled returns the (new) current mode (0 or
883 * positive). Therefore, a negative input value can be used to query
884 * the current mode without changing it.
885 *
886 * The XNACK mode fundamentally changes the way SVM managed memory works
887 * in the driver, with subtle effects on application performance and
888 * functionality.
889 *
890 * Enabling XNACK mode requires shader programs to be compiled
891 * differently. Furthermore, not all GPUs support changing the mode
892 * per-process. Therefore changing the mode is only allowed while no
893 * user mode queues exist in the process. This ensure that no shader
894 * code is running that may be compiled for the wrong mode. And GPUs
895 * that cannot change to the requested mode will prevent the XNACK
896 * mode from occurring. All GPUs used by the process must be in the
897 * same XNACK mode.
898 *
899 * GFXv8 or older GPUs do not support 48 bit virtual addresses or SVM.
900 * Therefore those GPUs are not considered for the XNACK mode switch.
901 *
902 * Return: 0 on success, -errno on failure
903 */
904struct kfd_ioctl_set_xnack_mode_args {
905 __s32 xnack_enabled;
906};
907
908/* Wave launch override modes */
909enum kfd_dbg_trap_override_mode {
910 KFD_DBG_TRAP_OVERRIDE_OR = 0,
911 KFD_DBG_TRAP_OVERRIDE_REPLACE = 1
912};
913
914/* Wave launch overrides */
915enum kfd_dbg_trap_mask {
916 KFD_DBG_TRAP_MASK_FP_INVALID = 1,
917 KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL = 2,
918 KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO = 4,
919 KFD_DBG_TRAP_MASK_FP_OVERFLOW = 8,
920 KFD_DBG_TRAP_MASK_FP_UNDERFLOW = 16,
921 KFD_DBG_TRAP_MASK_FP_INEXACT = 32,
922 KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO = 64,
923 KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH = 128,
924 KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION = 256,
925 KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START = (1 << 30),
926 KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END = (1 << 31)
927};
928
929/* Wave launch modes */
930enum kfd_dbg_trap_wave_launch_mode {
931 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_NORMAL = 0,
932 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_HALT = 1,
933 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_DEBUG = 3
934};
935
936/* Address watch modes */
937enum kfd_dbg_trap_address_watch_mode {
938 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_READ = 0,
939 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_NONREAD = 1,
940 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_ATOMIC = 2,
941 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_ALL = 3
942};
943
944/* Additional wave settings */
945enum kfd_dbg_trap_flags {
946 KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP = 1,
947 KFD_DBG_TRAP_FLAG_SINGLE_ALU_OP = 2,
948};
949
950/* Trap exceptions */
951enum kfd_dbg_trap_exception_code {
952 EC_NONE = 0,
953 /* per queue */
954 EC_QUEUE_WAVE_ABORT = 1,
955 EC_QUEUE_WAVE_TRAP = 2,
956 EC_QUEUE_WAVE_MATH_ERROR = 3,
957 EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION = 4,
958 EC_QUEUE_WAVE_MEMORY_VIOLATION = 5,
959 EC_QUEUE_WAVE_APERTURE_VIOLATION = 6,
960 EC_QUEUE_PACKET_DISPATCH_DIM_INVALID = 16,
961 EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID = 17,
962 EC_QUEUE_PACKET_DISPATCH_CODE_INVALID = 18,
963 EC_QUEUE_PACKET_RESERVED = 19,
964 EC_QUEUE_PACKET_UNSUPPORTED = 20,
965 EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID = 21,
966 EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID = 22,
967 EC_QUEUE_PACKET_VENDOR_UNSUPPORTED = 23,
968 EC_QUEUE_PREEMPTION_ERROR = 30,
969 EC_QUEUE_NEW = 31,
970 /* per device */
971 EC_DEVICE_QUEUE_DELETE = 32,
972 EC_DEVICE_MEMORY_VIOLATION = 33,
973 EC_DEVICE_RAS_ERROR = 34,
974 EC_DEVICE_FATAL_HALT = 35,
975 EC_DEVICE_NEW = 36,
976 /* per process */
977 EC_PROCESS_RUNTIME = 48,
978 EC_PROCESS_DEVICE_REMOVE = 49,
979 EC_MAX
980};
981
982/* Mask generated by ecode in kfd_dbg_trap_exception_code */
983#define KFD_EC_MASK(ecode) (1ULL << (ecode - 1))
984
985/* Masks for exception code type checks below */
986#define KFD_EC_MASK_QUEUE (KFD_EC_MASK(EC_QUEUE_WAVE_ABORT) | \
987 KFD_EC_MASK(EC_QUEUE_WAVE_TRAP) | \
988 KFD_EC_MASK(EC_QUEUE_WAVE_MATH_ERROR) | \
989 KFD_EC_MASK(EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION) | \
990 KFD_EC_MASK(EC_QUEUE_WAVE_MEMORY_VIOLATION) | \
991 KFD_EC_MASK(EC_QUEUE_WAVE_APERTURE_VIOLATION) | \
992 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_DIM_INVALID) | \
993 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID) | \
994 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_CODE_INVALID) | \
995 KFD_EC_MASK(EC_QUEUE_PACKET_RESERVED) | \
996 KFD_EC_MASK(EC_QUEUE_PACKET_UNSUPPORTED) | \
997 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) | \
998 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) | \
999 KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED) | \
1000 KFD_EC_MASK(EC_QUEUE_PREEMPTION_ERROR) | \
1001 KFD_EC_MASK(EC_QUEUE_NEW))
1002#define KFD_EC_MASK_DEVICE (KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE) | \
1003 KFD_EC_MASK(EC_DEVICE_RAS_ERROR) | \
1004 KFD_EC_MASK(EC_DEVICE_FATAL_HALT) | \
1005 KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION) | \
1006 KFD_EC_MASK(EC_DEVICE_NEW))
1007#define KFD_EC_MASK_PROCESS (KFD_EC_MASK(EC_PROCESS_RUNTIME) | \
1008 KFD_EC_MASK(EC_PROCESS_DEVICE_REMOVE))
1009#define KFD_EC_MASK_PACKET (KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_DIM_INVALID) | \
1010 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID) | \
1011 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_CODE_INVALID) | \
1012 KFD_EC_MASK(EC_QUEUE_PACKET_RESERVED) | \
1013 KFD_EC_MASK(EC_QUEUE_PACKET_UNSUPPORTED) | \
1014 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) | \
1015 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) | \
1016 KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED))
1017
1018/* Checks for exception code types for KFD search */
1019#define KFD_DBG_EC_IS_VALID(ecode) (ecode > EC_NONE && ecode < EC_MAX)
1020#define KFD_DBG_EC_TYPE_IS_QUEUE(ecode) \
1021 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_QUEUE))
1022#define KFD_DBG_EC_TYPE_IS_DEVICE(ecode) \
1023 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_DEVICE))
1024#define KFD_DBG_EC_TYPE_IS_PROCESS(ecode) \
1025 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_PROCESS))
1026#define KFD_DBG_EC_TYPE_IS_PACKET(ecode) \
1027 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_PACKET))
1028
1029
1030/* Runtime enable states */
1031enum kfd_dbg_runtime_state {
1032 DEBUG_RUNTIME_STATE_DISABLED = 0,
1033 DEBUG_RUNTIME_STATE_ENABLED = 1,
1034 DEBUG_RUNTIME_STATE_ENABLED_BUSY = 2,
1035 DEBUG_RUNTIME_STATE_ENABLED_ERROR = 3
1036};
1037
1038/* Runtime enable status */
1039struct kfd_runtime_info {
1040 __u64 r_debug;
1041 __u32 runtime_state;
1042 __u32 ttmp_setup;
1043};
1044
1045/* Enable modes for runtime enable */
1046#define KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK 1
1047#define KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK 2
1048
1049/**
1050 * kfd_ioctl_runtime_enable_args - Arguments for runtime enable
1051 *
1052 * Coordinates debug exception signalling and debug device enablement with runtime.
1053 *
1054 * @r_debug - pointer to user struct for sharing information between ROCr and the debuggger
1055 * @mode_mask - mask to set mode
1056 * KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK - enable runtime for debugging, otherwise disable
1057 * KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK - enable trap temporary setup (ignore on disable)
1058 * @capabilities_mask - mask to notify runtime on what KFD supports
1059 *
1060 * Return - 0 on SUCCESS.
1061 * - EBUSY if runtime enable call already pending.
1062 * - EEXIST if user queues already active prior to call.
1063 * If process is debug enabled, runtime enable will enable debug devices and
1064 * wait for debugger process to send runtime exception EC_PROCESS_RUNTIME
1065 * to unblock - see kfd_ioctl_dbg_trap_args.
1066 *
1067 */
1068struct kfd_ioctl_runtime_enable_args {
1069 __u64 r_debug;
1070 __u32 mode_mask;
1071 __u32 capabilities_mask;
1072};
1073
1074/* Queue information */
1075struct kfd_queue_snapshot_entry {
1076 __u64 exception_status;
1077 __u64 ring_base_address;
1078 __u64 write_pointer_address;
1079 __u64 read_pointer_address;
1080 __u64 ctx_save_restore_address;
1081 __u32 queue_id;
1082 __u32 gpu_id;
1083 __u32 ring_size;
1084 __u32 queue_type;
1085 __u32 ctx_save_restore_area_size;
1086 __u32 reserved;
1087};
1088
1089/* Queue status return for suspend/resume */
1090#define KFD_DBG_QUEUE_ERROR_BIT 30
1091#define KFD_DBG_QUEUE_INVALID_BIT 31
1092#define KFD_DBG_QUEUE_ERROR_MASK (1 << KFD_DBG_QUEUE_ERROR_BIT)
1093#define KFD_DBG_QUEUE_INVALID_MASK (1 << KFD_DBG_QUEUE_INVALID_BIT)
1094
1095/* Context save area header information */
1096struct kfd_context_save_area_header {
1097 struct {
1098 __u32 control_stack_offset;
1099 __u32 control_stack_size;
1100 __u32 wave_state_offset;
1101 __u32 wave_state_size;
1102 } wave_state;
1103 __u32 debug_offset;
1104 __u32 debug_size;
1105 __u64 err_payload_addr;
1106 __u32 err_event_id;
1107 __u32 reserved1;
1108};
1109
1110/*
1111 * Debug operations
1112 *
1113 * For specifics on usage and return values, see documentation per operation
1114 * below. Otherwise, generic error returns apply:
1115 * - ESRCH if the process to debug does not exist.
1116 *
1117 * - EINVAL (with KFD_IOC_DBG_TRAP_ENABLE exempt) if operation
1118 * KFD_IOC_DBG_TRAP_ENABLE has not succeeded prior.
1119 * Also returns this error if GPU hardware scheduling is not supported.
1120 *
1121 * - EPERM (with KFD_IOC_DBG_TRAP_DISABLE exempt) if target process is not
1122 * PTRACE_ATTACHED. KFD_IOC_DBG_TRAP_DISABLE is exempt to allow
1123 * clean up of debug mode as long as process is debug enabled.
1124 *
1125 * - EACCES if any DBG_HW_OP (debug hardware operation) is requested when
1126 * AMDKFD_IOC_RUNTIME_ENABLE has not succeeded prior.
1127 *
1128 * - ENODEV if any GPU does not support debugging on a DBG_HW_OP call.
1129 *
1130 * - Other errors may be returned when a DBG_HW_OP occurs while the GPU
1131 * is in a fatal state.
1132 *
1133 */
1134enum kfd_dbg_trap_operations {
1135 KFD_IOC_DBG_TRAP_ENABLE = 0,
1136 KFD_IOC_DBG_TRAP_DISABLE = 1,
1137 KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT = 2,
1138 KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED = 3,
1139 KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE = 4, /* DBG_HW_OP */
1140 KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE = 5, /* DBG_HW_OP */
1141 KFD_IOC_DBG_TRAP_SUSPEND_QUEUES = 6, /* DBG_HW_OP */
1142 KFD_IOC_DBG_TRAP_RESUME_QUEUES = 7, /* DBG_HW_OP */
1143 KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH = 8, /* DBG_HW_OP */
1144 KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH = 9, /* DBG_HW_OP */
1145 KFD_IOC_DBG_TRAP_SET_FLAGS = 10,
1146 KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT = 11,
1147 KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO = 12,
1148 KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT = 13,
1149 KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT = 14
1150};
1151
1152/**
1153 * kfd_ioctl_dbg_trap_enable_args
1154 *
1155 * Arguments for KFD_IOC_DBG_TRAP_ENABLE.
1156 *
1157 * Enables debug session for target process. Call @op KFD_IOC_DBG_TRAP_DISABLE in
1158 * kfd_ioctl_dbg_trap_args to disable debug session.
1159 *
1160 * @exception_mask (IN) - exceptions to raise to the debugger
1161 * @rinfo_ptr (IN) - pointer to runtime info buffer (see kfd_runtime_info)
1162 * @rinfo_size (IN/OUT) - size of runtime info buffer in bytes
1163 * @dbg_fd (IN) - fd the KFD will nofify the debugger with of raised
1164 * exceptions set in exception_mask.
1165 *
1166 * Generic errors apply (see kfd_dbg_trap_operations).
1167 * Return - 0 on SUCCESS.
1168 * Copies KFD saved kfd_runtime_info to @rinfo_ptr on enable.
1169 * Size of kfd_runtime saved by the KFD returned to @rinfo_size.
1170 * - EBADF if KFD cannot get a reference to dbg_fd.
1171 * - EFAULT if KFD cannot copy runtime info to rinfo_ptr.
1172 * - EINVAL if target process is already debug enabled.
1173 *
1174 */
1175struct kfd_ioctl_dbg_trap_enable_args {
1176 __u64 exception_mask;
1177 __u64 rinfo_ptr;
1178 __u32 rinfo_size;
1179 __u32 dbg_fd;
1180};
1181
1182/**
1183 * kfd_ioctl_dbg_trap_send_runtime_event_args
1184 *
1185 *
1186 * Arguments for KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT.
1187 * Raises exceptions to runtime.
1188 *
1189 * @exception_mask (IN) - exceptions to raise to runtime
1190 * @gpu_id (IN) - target device id
1191 * @queue_id (IN) - target queue id
1192 *
1193 * Generic errors apply (see kfd_dbg_trap_operations).
1194 * Return - 0 on SUCCESS.
1195 * - ENODEV if gpu_id not found.
1196 * If exception_mask contains EC_PROCESS_RUNTIME, unblocks pending
1197 * AMDKFD_IOC_RUNTIME_ENABLE call - see kfd_ioctl_runtime_enable_args.
1198 * All other exceptions are raised to runtime through err_payload_addr.
1199 * See kfd_context_save_area_header.
1200 */
1201struct kfd_ioctl_dbg_trap_send_runtime_event_args {
1202 __u64 exception_mask;
1203 __u32 gpu_id;
1204 __u32 queue_id;
1205};
1206
1207/**
1208 * kfd_ioctl_dbg_trap_set_exceptions_enabled_args
1209 *
1210 * Arguments for KFD_IOC_SET_EXCEPTIONS_ENABLED
1211 * Set new exceptions to be raised to the debugger.
1212 *
1213 * @exception_mask (IN) - new exceptions to raise the debugger
1214 *
1215 * Generic errors apply (see kfd_dbg_trap_operations).
1216 * Return - 0 on SUCCESS.
1217 */
1218struct kfd_ioctl_dbg_trap_set_exceptions_enabled_args {
1219 __u64 exception_mask;
1220};
1221
1222/**
1223 * kfd_ioctl_dbg_trap_set_wave_launch_override_args
1224 *
1225 * Arguments for KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE
1226 * Enable HW exceptions to raise trap.
1227 *
1228 * @override_mode (IN) - see kfd_dbg_trap_override_mode
1229 * @enable_mask (IN/OUT) - reference kfd_dbg_trap_mask.
1230 * IN is the override modes requested to be enabled.
1231 * OUT is referenced in Return below.
1232 * @support_request_mask (IN/OUT) - reference kfd_dbg_trap_mask.
1233 * IN is the override modes requested for support check.
1234 * OUT is referenced in Return below.
1235 *
1236 * Generic errors apply (see kfd_dbg_trap_operations).
1237 * Return - 0 on SUCCESS.
1238 * Previous enablement is returned in @enable_mask.
1239 * Actual override support is returned in @support_request_mask.
1240 * - EINVAL if override mode is not supported.
1241 * - EACCES if trap support requested is not actually supported.
1242 * i.e. enable_mask (IN) is not a subset of support_request_mask (OUT).
1243 * Otherwise it is considered a generic error (see kfd_dbg_trap_operations).
1244 */
1245struct kfd_ioctl_dbg_trap_set_wave_launch_override_args {
1246 __u32 override_mode;
1247 __u32 enable_mask;
1248 __u32 support_request_mask;
1249 __u32 pad;
1250};
1251
1252/**
1253 * kfd_ioctl_dbg_trap_set_wave_launch_mode_args
1254 *
1255 * Arguments for KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE
1256 * Set wave launch mode.
1257 *
1258 * @mode (IN) - see kfd_dbg_trap_wave_launch_mode
1259 *
1260 * Generic errors apply (see kfd_dbg_trap_operations).
1261 * Return - 0 on SUCCESS.
1262 */
1263struct kfd_ioctl_dbg_trap_set_wave_launch_mode_args {
1264 __u32 launch_mode;
1265 __u32 pad;
1266};
1267
1268/**
1269 * kfd_ioctl_dbg_trap_suspend_queues_ags
1270 *
1271 * Arguments for KFD_IOC_DBG_TRAP_SUSPEND_QUEUES
1272 * Suspend queues.
1273 *
1274 * @exception_mask (IN) - raised exceptions to clear
1275 * @queue_array_ptr (IN) - pointer to array of queue ids (u32 per queue id)
1276 * to suspend
1277 * @num_queues (IN) - number of queues to suspend in @queue_array_ptr
1278 * @grace_period (IN) - wave time allowance before preemption
1279 * per 1K GPU clock cycle unit
1280 *
1281 * Generic errors apply (see kfd_dbg_trap_operations).
1282 * Destruction of a suspended queue is blocked until the queue is
1283 * resumed. This allows the debugger to access queue information and
1284 * the its context save area without running into a race condition on
1285 * queue destruction.
1286 * Automatically copies per queue context save area header information
1287 * into the save area base
1288 * (see kfd_queue_snapshot_entry and kfd_context_save_area_header).
1289 *
1290 * Return - Number of queues suspended on SUCCESS.
1291 * . KFD_DBG_QUEUE_ERROR_MASK and KFD_DBG_QUEUE_INVALID_MASK masked
1292 * for each queue id in @queue_array_ptr array reports unsuccessful
1293 * suspend reason.
1294 * KFD_DBG_QUEUE_ERROR_MASK = HW failure.
1295 * KFD_DBG_QUEUE_INVALID_MASK = queue does not exist, is new or
1296 * is being destroyed.
1297 */
1298struct kfd_ioctl_dbg_trap_suspend_queues_args {
1299 __u64 exception_mask;
1300 __u64 queue_array_ptr;
1301 __u32 num_queues;
1302 __u32 grace_period;
1303};
1304
1305/**
1306 * kfd_ioctl_dbg_trap_resume_queues_args
1307 *
1308 * Arguments for KFD_IOC_DBG_TRAP_RESUME_QUEUES
1309 * Resume queues.
1310 *
1311 * @queue_array_ptr (IN) - pointer to array of queue ids (u32 per queue id)
1312 * to resume
1313 * @num_queues (IN) - number of queues to resume in @queue_array_ptr
1314 *
1315 * Generic errors apply (see kfd_dbg_trap_operations).
1316 * Return - Number of queues resumed on SUCCESS.
1317 * KFD_DBG_QUEUE_ERROR_MASK and KFD_DBG_QUEUE_INVALID_MASK mask
1318 * for each queue id in @queue_array_ptr array reports unsuccessful
1319 * resume reason.
1320 * KFD_DBG_QUEUE_ERROR_MASK = HW failure.
1321 * KFD_DBG_QUEUE_INVALID_MASK = queue does not exist.
1322 */
1323struct kfd_ioctl_dbg_trap_resume_queues_args {
1324 __u64 queue_array_ptr;
1325 __u32 num_queues;
1326 __u32 pad;
1327};
1328
1329/**
1330 * kfd_ioctl_dbg_trap_set_node_address_watch_args
1331 *
1332 * Arguments for KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH
1333 * Sets address watch for device.
1334 *
1335 * @address (IN) - watch address to set
1336 * @mode (IN) - see kfd_dbg_trap_address_watch_mode
1337 * @mask (IN) - watch address mask
1338 * @gpu_id (IN) - target gpu to set watch point
1339 * @id (OUT) - watch id allocated
1340 *
1341 * Generic errors apply (see kfd_dbg_trap_operations).
1342 * Return - 0 on SUCCESS.
1343 * Allocated watch ID returned to @id.
1344 * - ENODEV if gpu_id not found.
1345 * - ENOMEM if watch IDs can be allocated
1346 */
1347struct kfd_ioctl_dbg_trap_set_node_address_watch_args {
1348 __u64 address;
1349 __u32 mode;
1350 __u32 mask;
1351 __u32 gpu_id;
1352 __u32 id;
1353};
1354
1355/**
1356 * kfd_ioctl_dbg_trap_clear_node_address_watch_args
1357 *
1358 * Arguments for KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH
1359 * Clear address watch for device.
1360 *
1361 * @gpu_id (IN) - target device to clear watch point
1362 * @id (IN) - allocated watch id to clear
1363 *
1364 * Generic errors apply (see kfd_dbg_trap_operations).
1365 * Return - 0 on SUCCESS.
1366 * - ENODEV if gpu_id not found.
1367 * - EINVAL if watch ID has not been allocated.
1368 */
1369struct kfd_ioctl_dbg_trap_clear_node_address_watch_args {
1370 __u32 gpu_id;
1371 __u32 id;
1372};
1373
1374/**
1375 * kfd_ioctl_dbg_trap_set_flags_args
1376 *
1377 * Arguments for KFD_IOC_DBG_TRAP_SET_FLAGS
1378 * Sets flags for wave behaviour.
1379 *
1380 * @flags (IN/OUT) - IN = flags to enable, OUT = flags previously enabled
1381 *
1382 * Generic errors apply (see kfd_dbg_trap_operations).
1383 * Return - 0 on SUCCESS.
1384 * - EACCESS if any debug device does not allow flag options.
1385 */
1386struct kfd_ioctl_dbg_trap_set_flags_args {
1387 __u32 flags;
1388 __u32 pad;
1389};
1390
1391/**
1392 * kfd_ioctl_dbg_trap_query_debug_event_args
1393 *
1394 * Arguments for KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT
1395 *
1396 * Find one or more raised exceptions. This function can return multiple
1397 * exceptions from a single queue or a single device with one call. To find
1398 * all raised exceptions, this function must be called repeatedly until it
1399 * returns -EAGAIN. Returned exceptions can optionally be cleared by
1400 * setting the corresponding bit in the @exception_mask input parameter.
1401 * However, clearing an exception prevents retrieving further information
1402 * about it with KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO.
1403 *
1404 * @exception_mask (IN/OUT) - exception to clear (IN) and raised (OUT)
1405 * @gpu_id (OUT) - gpu id of exceptions raised
1406 * @queue_id (OUT) - queue id of exceptions raised
1407 *
1408 * Generic errors apply (see kfd_dbg_trap_operations).
1409 * Return - 0 on raised exception found
1410 * Raised exceptions found are returned in @exception mask
1411 * with reported source id returned in @gpu_id or @queue_id.
1412 * - EAGAIN if no raised exception has been found
1413 */
1414struct kfd_ioctl_dbg_trap_query_debug_event_args {
1415 __u64 exception_mask;
1416 __u32 gpu_id;
1417 __u32 queue_id;
1418};
1419
1420/**
1421 * kfd_ioctl_dbg_trap_query_exception_info_args
1422 *
1423 * Arguments KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO
1424 * Get additional info on raised exception.
1425 *
1426 * @info_ptr (IN) - pointer to exception info buffer to copy to
1427 * @info_size (IN/OUT) - exception info buffer size (bytes)
1428 * @source_id (IN) - target gpu or queue id
1429 * @exception_code (IN) - target exception
1430 * @clear_exception (IN) - clear raised @exception_code exception
1431 * (0 = false, 1 = true)
1432 *
1433 * Generic errors apply (see kfd_dbg_trap_operations).
1434 * Return - 0 on SUCCESS.
1435 * If @exception_code is EC_DEVICE_MEMORY_VIOLATION, copy @info_size(OUT)
1436 * bytes of memory exception data to @info_ptr.
1437 * If @exception_code is EC_PROCESS_RUNTIME, copy saved
1438 * kfd_runtime_info to @info_ptr.
1439 * Actual required @info_ptr size (bytes) is returned in @info_size.
1440 */
1441struct kfd_ioctl_dbg_trap_query_exception_info_args {
1442 __u64 info_ptr;
1443 __u32 info_size;
1444 __u32 source_id;
1445 __u32 exception_code;
1446 __u32 clear_exception;
1447};
1448
1449/**
1450 * kfd_ioctl_dbg_trap_get_queue_snapshot_args
1451 *
1452 * Arguments KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT
1453 * Get queue information.
1454 *
1455 * @exception_mask (IN) - exceptions raised to clear
1456 * @snapshot_buf_ptr (IN) - queue snapshot entry buffer (see kfd_queue_snapshot_entry)
1457 * @num_queues (IN/OUT) - number of queue snapshot entries
1458 * The debugger specifies the size of the array allocated in @num_queues.
1459 * KFD returns the number of queues that actually existed. If this is
1460 * larger than the size specified by the debugger, KFD will not overflow
1461 * the array allocated by the debugger.
1462 *
1463 * @entry_size (IN/OUT) - size per entry in bytes
1464 * The debugger specifies sizeof(struct kfd_queue_snapshot_entry) in
1465 * @entry_size. KFD returns the number of bytes actually populated per
1466 * entry. The debugger should use the KFD_IOCTL_MINOR_VERSION to determine,
1467 * which fields in struct kfd_queue_snapshot_entry are valid. This allows
1468 * growing the ABI in a backwards compatible manner.
1469 * Note that entry_size(IN) should still be used to stride the snapshot buffer in the
1470 * event that it's larger than actual kfd_queue_snapshot_entry.
1471 *
1472 * Generic errors apply (see kfd_dbg_trap_operations).
1473 * Return - 0 on SUCCESS.
1474 * Copies @num_queues(IN) queue snapshot entries of size @entry_size(IN)
1475 * into @snapshot_buf_ptr if @num_queues(IN) > 0.
1476 * Otherwise return @num_queues(OUT) queue snapshot entries that exist.
1477 */
1478struct kfd_ioctl_dbg_trap_queue_snapshot_args {
1479 __u64 exception_mask;
1480 __u64 snapshot_buf_ptr;
1481 __u32 num_queues;
1482 __u32 entry_size;
1483};
1484
1485/**
1486 * kfd_ioctl_dbg_trap_get_device_snapshot_args
1487 *
1488 * Arguments for KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT
1489 * Get device information.
1490 *
1491 * @exception_mask (IN) - exceptions raised to clear
1492 * @snapshot_buf_ptr (IN) - pointer to snapshot buffer (see kfd_dbg_device_info_entry)
1493 * @num_devices (IN/OUT) - number of debug devices to snapshot
1494 * The debugger specifies the size of the array allocated in @num_devices.
1495 * KFD returns the number of devices that actually existed. If this is
1496 * larger than the size specified by the debugger, KFD will not overflow
1497 * the array allocated by the debugger.
1498 *
1499 * @entry_size (IN/OUT) - size per entry in bytes
1500 * The debugger specifies sizeof(struct kfd_dbg_device_info_entry) in
1501 * @entry_size. KFD returns the number of bytes actually populated. The
1502 * debugger should use KFD_IOCTL_MINOR_VERSION to determine, which fields
1503 * in struct kfd_dbg_device_info_entry are valid. This allows growing the
1504 * ABI in a backwards compatible manner.
1505 * Note that entry_size(IN) should still be used to stride the snapshot buffer in the
1506 * event that it's larger than actual kfd_dbg_device_info_entry.
1507 *
1508 * Generic errors apply (see kfd_dbg_trap_operations).
1509 * Return - 0 on SUCCESS.
1510 * Copies @num_devices(IN) device snapshot entries of size @entry_size(IN)
1511 * into @snapshot_buf_ptr if @num_devices(IN) > 0.
1512 * Otherwise return @num_devices(OUT) queue snapshot entries that exist.
1513 */
1514struct kfd_ioctl_dbg_trap_device_snapshot_args {
1515 __u64 exception_mask;
1516 __u64 snapshot_buf_ptr;
1517 __u32 num_devices;
1518 __u32 entry_size;
1519};
1520
1521/**
1522 * kfd_ioctl_dbg_trap_args
1523 *
1524 * Arguments to debug target process.
1525 *
1526 * @pid - target process to debug
1527 * @op - debug operation (see kfd_dbg_trap_operations)
1528 *
1529 * @op determines which union struct args to use.
1530 * Refer to kern docs for each kfd_ioctl_dbg_trap_*_args struct.
1531 */
1532struct kfd_ioctl_dbg_trap_args {
1533 __u32 pid;
1534 __u32 op;
1535
1536 union {
1537 struct kfd_ioctl_dbg_trap_enable_args enable;
1538 struct kfd_ioctl_dbg_trap_send_runtime_event_args send_runtime_event;
1539 struct kfd_ioctl_dbg_trap_set_exceptions_enabled_args set_exceptions_enabled;
1540 struct kfd_ioctl_dbg_trap_set_wave_launch_override_args launch_override;
1541 struct kfd_ioctl_dbg_trap_set_wave_launch_mode_args launch_mode;
1542 struct kfd_ioctl_dbg_trap_suspend_queues_args suspend_queues;
1543 struct kfd_ioctl_dbg_trap_resume_queues_args resume_queues;
1544 struct kfd_ioctl_dbg_trap_set_node_address_watch_args set_node_address_watch;
1545 struct kfd_ioctl_dbg_trap_clear_node_address_watch_args clear_node_address_watch;
1546 struct kfd_ioctl_dbg_trap_set_flags_args set_flags;
1547 struct kfd_ioctl_dbg_trap_query_debug_event_args query_debug_event;
1548 struct kfd_ioctl_dbg_trap_query_exception_info_args query_exception_info;
1549 struct kfd_ioctl_dbg_trap_queue_snapshot_args queue_snapshot;
1550 struct kfd_ioctl_dbg_trap_device_snapshot_args device_snapshot;
1551 };
1552};
1553
1554#define AMDKFD_IOCTL_BASE 'K'
1555#define AMDKFD_IO(nr) _IO(AMDKFD_IOCTL_BASE, nr)
1556#define AMDKFD_IOR(nr, type) _IOR(AMDKFD_IOCTL_BASE, nr, type)
1557#define AMDKFD_IOW(nr, type) _IOW(AMDKFD_IOCTL_BASE, nr, type)
1558#define AMDKFD_IOWR(nr, type) _IOWR(AMDKFD_IOCTL_BASE, nr, type)
1559
1560#define AMDKFD_IOC_GET_VERSION \
1561 AMDKFD_IOR(0x01, struct kfd_ioctl_get_version_args)
1562
1563#define AMDKFD_IOC_CREATE_QUEUE \
1564 AMDKFD_IOWR(0x02, struct kfd_ioctl_create_queue_args)
1565
1566#define AMDKFD_IOC_DESTROY_QUEUE \
1567 AMDKFD_IOWR(0x03, struct kfd_ioctl_destroy_queue_args)
1568
1569#define AMDKFD_IOC_SET_MEMORY_POLICY \
1570 AMDKFD_IOW(0x04, struct kfd_ioctl_set_memory_policy_args)
1571
1572#define AMDKFD_IOC_GET_CLOCK_COUNTERS \
1573 AMDKFD_IOWR(0x05, struct kfd_ioctl_get_clock_counters_args)
1574
1575#define AMDKFD_IOC_GET_PROCESS_APERTURES \
1576 AMDKFD_IOR(0x06, struct kfd_ioctl_get_process_apertures_args)
1577
1578#define AMDKFD_IOC_UPDATE_QUEUE \
1579 AMDKFD_IOW(0x07, struct kfd_ioctl_update_queue_args)
1580
1581#define AMDKFD_IOC_CREATE_EVENT \
1582 AMDKFD_IOWR(0x08, struct kfd_ioctl_create_event_args)
1583
1584#define AMDKFD_IOC_DESTROY_EVENT \
1585 AMDKFD_IOW(0x09, struct kfd_ioctl_destroy_event_args)
1586
1587#define AMDKFD_IOC_SET_EVENT \
1588 AMDKFD_IOW(0x0A, struct kfd_ioctl_set_event_args)
1589
1590#define AMDKFD_IOC_RESET_EVENT \
1591 AMDKFD_IOW(0x0B, struct kfd_ioctl_reset_event_args)
1592
1593#define AMDKFD_IOC_WAIT_EVENTS \
1594 AMDKFD_IOWR(0x0C, struct kfd_ioctl_wait_events_args)
1595
1596#define AMDKFD_IOC_DBG_REGISTER_DEPRECATED \
1597 AMDKFD_IOW(0x0D, struct kfd_ioctl_dbg_register_args)
1598
1599#define AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED \
1600 AMDKFD_IOW(0x0E, struct kfd_ioctl_dbg_unregister_args)
1601
1602#define AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED \
1603 AMDKFD_IOW(0x0F, struct kfd_ioctl_dbg_address_watch_args)
1604
1605#define AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED \
1606 AMDKFD_IOW(0x10, struct kfd_ioctl_dbg_wave_control_args)
1607
1608#define AMDKFD_IOC_SET_SCRATCH_BACKING_VA \
1609 AMDKFD_IOWR(0x11, struct kfd_ioctl_set_scratch_backing_va_args)
1610
1611#define AMDKFD_IOC_GET_TILE_CONFIG \
1612 AMDKFD_IOWR(0x12, struct kfd_ioctl_get_tile_config_args)
1613
1614#define AMDKFD_IOC_SET_TRAP_HANDLER \
1615 AMDKFD_IOW(0x13, struct kfd_ioctl_set_trap_handler_args)
1616
1617#define AMDKFD_IOC_GET_PROCESS_APERTURES_NEW \
1618 AMDKFD_IOWR(0x14, \
1619 struct kfd_ioctl_get_process_apertures_new_args)
1620
1621#define AMDKFD_IOC_ACQUIRE_VM \
1622 AMDKFD_IOW(0x15, struct kfd_ioctl_acquire_vm_args)
1623
1624#define AMDKFD_IOC_ALLOC_MEMORY_OF_GPU \
1625 AMDKFD_IOWR(0x16, struct kfd_ioctl_alloc_memory_of_gpu_args)
1626
1627#define AMDKFD_IOC_FREE_MEMORY_OF_GPU \
1628 AMDKFD_IOW(0x17, struct kfd_ioctl_free_memory_of_gpu_args)
1629
1630#define AMDKFD_IOC_MAP_MEMORY_TO_GPU \
1631 AMDKFD_IOWR(0x18, struct kfd_ioctl_map_memory_to_gpu_args)
1632
1633#define AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU \
1634 AMDKFD_IOWR(0x19, struct kfd_ioctl_unmap_memory_from_gpu_args)
1635
1636#define AMDKFD_IOC_SET_CU_MASK \
1637 AMDKFD_IOW(0x1A, struct kfd_ioctl_set_cu_mask_args)
1638
1639#define AMDKFD_IOC_GET_QUEUE_WAVE_STATE \
1640 AMDKFD_IOWR(0x1B, struct kfd_ioctl_get_queue_wave_state_args)
1641
1642#define AMDKFD_IOC_GET_DMABUF_INFO \
1643 AMDKFD_IOWR(0x1C, struct kfd_ioctl_get_dmabuf_info_args)
1644
1645#define AMDKFD_IOC_IMPORT_DMABUF \
1646 AMDKFD_IOWR(0x1D, struct kfd_ioctl_import_dmabuf_args)
1647
1648#define AMDKFD_IOC_ALLOC_QUEUE_GWS \
1649 AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
1650
1651#define AMDKFD_IOC_SMI_EVENTS \
1652 AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
1653
1654#define AMDKFD_IOC_SVM AMDKFD_IOWR(0x20, struct kfd_ioctl_svm_args)
1655
1656#define AMDKFD_IOC_SET_XNACK_MODE \
1657 AMDKFD_IOWR(0x21, struct kfd_ioctl_set_xnack_mode_args)
1658
1659#define AMDKFD_IOC_CRIU_OP \
1660 AMDKFD_IOWR(0x22, struct kfd_ioctl_criu_args)
1661
1662#define AMDKFD_IOC_AVAILABLE_MEMORY \
1663 AMDKFD_IOWR(0x23, struct kfd_ioctl_get_available_memory_args)
1664
1665#define AMDKFD_IOC_EXPORT_DMABUF \
1666 AMDKFD_IOWR(0x24, struct kfd_ioctl_export_dmabuf_args)
1667
1668#define AMDKFD_IOC_RUNTIME_ENABLE \
1669 AMDKFD_IOWR(0x25, struct kfd_ioctl_runtime_enable_args)
1670
1671#define AMDKFD_IOC_DBG_TRAP \
1672 AMDKFD_IOWR(0x26, struct kfd_ioctl_dbg_trap_args)
1673
1674#define AMDKFD_COMMAND_START 0x01
1675#define AMDKFD_COMMAND_END 0x27
1676
1677#endif