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 * - 1.19 - Add a new ioctl to craete secondary kfd processes
48 * - 1.20 - Trap handler support for expert scheduling mode available
49 * - 1.21 - Debugger support to subscribe to LDS out-of-address exceptions
50 * - 1.22 - Add queue creation with metadata ring base address
51 * - 1.23 - Add profiler control ioctl to enable/disable profiler on a process
52 */
53#define KFD_IOCTL_MAJOR_VERSION 1
54#define KFD_IOCTL_MINOR_VERSION 23
55
56struct kfd_ioctl_get_version_args {
57 __u32 major_version; /* from KFD */
58 __u32 minor_version; /* from KFD */
59};
60
61/* For kfd_ioctl_create_queue_args.queue_type. */
62#define KFD_IOC_QUEUE_TYPE_COMPUTE 0x0
63#define KFD_IOC_QUEUE_TYPE_SDMA 0x1
64#define KFD_IOC_QUEUE_TYPE_COMPUTE_AQL 0x2
65#define KFD_IOC_QUEUE_TYPE_SDMA_XGMI 0x3
66#define KFD_IOC_QUEUE_TYPE_SDMA_BY_ENG_ID 0x4
67
68#define KFD_MAX_QUEUE_PERCENTAGE 100
69#define KFD_MAX_QUEUE_PRIORITY 15
70
71#define KFD_MIN_QUEUE_RING_SIZE 1024
72
73struct kfd_ioctl_create_queue_args {
74 __u64 ring_base_address; /* to KFD */
75 __u64 write_pointer_address; /* to KFD */
76 __u64 read_pointer_address; /* to KFD */
77 __u64 doorbell_offset; /* from KFD */
78
79 __u32 ring_size; /* to KFD */
80 __u32 gpu_id; /* to KFD */
81 __u32 queue_type; /* to KFD */
82 __u32 queue_percentage; /* to KFD */
83 __u32 queue_priority; /* to KFD */
84 __u32 queue_id; /* from KFD */
85
86 __u64 eop_buffer_address; /* to KFD */
87 __u64 eop_buffer_size; /* to KFD */
88 __u64 ctx_save_restore_address; /* to KFD */
89 __u32 ctx_save_restore_size; /* to KFD */
90 __u32 ctl_stack_size; /* to KFD */
91 __u32 sdma_engine_id; /* to KFD */
92 __u32 metadata_ring_size; /* to KFD */
93};
94
95struct kfd_ioctl_destroy_queue_args {
96 __u32 queue_id; /* to KFD */
97 __u32 pad;
98};
99
100struct kfd_ioctl_update_queue_args {
101 __u64 ring_base_address; /* to KFD */
102
103 __u32 queue_id; /* to KFD */
104 __u32 ring_size; /* to KFD */
105 __u32 queue_percentage; /* to KFD */
106 __u32 queue_priority; /* to KFD */
107};
108
109struct kfd_ioctl_set_cu_mask_args {
110 __u32 queue_id; /* to KFD */
111 __u32 num_cu_mask; /* to KFD */
112 __u64 cu_mask_ptr; /* to KFD */
113};
114
115struct kfd_ioctl_get_queue_wave_state_args {
116 __u64 ctl_stack_address; /* to KFD */
117 __u32 ctl_stack_used_size; /* from KFD */
118 __u32 save_area_used_size; /* from KFD */
119 __u32 queue_id; /* to KFD */
120 __u32 pad;
121};
122
123struct kfd_ioctl_get_available_memory_args {
124 __u64 available; /* from KFD */
125 __u32 gpu_id; /* to KFD */
126 __u32 pad;
127};
128
129struct kfd_dbg_device_info_entry {
130 __u64 exception_status;
131 __u64 lds_base;
132 __u64 lds_limit;
133 __u64 scratch_base;
134 __u64 scratch_limit;
135 __u64 gpuvm_base;
136 __u64 gpuvm_limit;
137 __u32 gpu_id;
138 __u32 location_id;
139 __u32 vendor_id;
140 __u32 device_id;
141 __u32 revision_id;
142 __u32 subsystem_vendor_id;
143 __u32 subsystem_device_id;
144 __u32 fw_version;
145 __u32 gfx_target_version;
146 __u32 simd_count;
147 __u32 max_waves_per_simd;
148 __u32 array_count;
149 __u32 simd_arrays_per_engine;
150 __u32 num_xcc;
151 __u32 capability;
152 __u32 debug_prop;
153 __u32 capability2;
154 __u32 pad;
155};
156
157/* For kfd_ioctl_set_memory_policy_args.default_policy and alternate_policy */
158#define KFD_IOC_CACHE_POLICY_COHERENT 0
159#define KFD_IOC_CACHE_POLICY_NONCOHERENT 1
160
161/* Misc. per process flags */
162#define KFD_PROC_FLAG_MFMA_HIGH_PRECISION (1 << 0)
163
164struct kfd_ioctl_set_memory_policy_args {
165 __u64 alternate_aperture_base; /* to KFD */
166 __u64 alternate_aperture_size; /* to KFD */
167
168 __u32 gpu_id; /* to KFD */
169 __u32 default_policy; /* to KFD */
170 __u32 alternate_policy; /* to KFD */
171 __u32 misc_process_flag; /* to KFD */
172};
173
174/*
175 * All counters are monotonic. They are used for profiling of compute jobs.
176 * The profiling is done by userspace.
177 *
178 * In case of GPU reset, the counter should not be affected.
179 */
180
181struct kfd_ioctl_get_clock_counters_args {
182 __u64 gpu_clock_counter; /* from KFD */
183 __u64 cpu_clock_counter; /* from KFD */
184 __u64 system_clock_counter; /* from KFD */
185 __u64 system_clock_freq; /* from KFD */
186
187 __u32 gpu_id; /* to KFD */
188 __u32 pad;
189};
190
191struct kfd_process_device_apertures {
192 __u64 lds_base; /* from KFD */
193 __u64 lds_limit; /* from KFD */
194 __u64 scratch_base; /* from KFD */
195 __u64 scratch_limit; /* from KFD */
196 __u64 gpuvm_base; /* from KFD */
197 __u64 gpuvm_limit; /* from KFD */
198 __u32 gpu_id; /* from KFD */
199 __u32 pad;
200};
201
202/*
203 * AMDKFD_IOC_GET_PROCESS_APERTURES is deprecated. Use
204 * AMDKFD_IOC_GET_PROCESS_APERTURES_NEW instead, which supports an
205 * unlimited number of GPUs.
206 */
207#define NUM_OF_SUPPORTED_GPUS 7
208struct kfd_ioctl_get_process_apertures_args {
209 struct kfd_process_device_apertures
210 process_apertures[NUM_OF_SUPPORTED_GPUS];/* from KFD */
211
212 /* from KFD, should be in the range [1 - NUM_OF_SUPPORTED_GPUS] */
213 __u32 num_of_nodes;
214 __u32 pad;
215};
216
217struct kfd_ioctl_get_process_apertures_new_args {
218 /* User allocated. Pointer to struct kfd_process_device_apertures
219 * filled in by Kernel
220 */
221 __u64 kfd_process_device_apertures_ptr;
222 /* to KFD - indicates amount of memory present in
223 * kfd_process_device_apertures_ptr
224 * from KFD - Number of entries filled by KFD.
225 */
226 __u32 num_of_nodes;
227 __u32 pad;
228};
229
230#define MAX_ALLOWED_NUM_POINTS 100
231#define MAX_ALLOWED_AW_BUFF_SIZE 4096
232#define MAX_ALLOWED_WAC_BUFF_SIZE 128
233
234struct kfd_ioctl_dbg_register_args {
235 __u32 gpu_id; /* to KFD */
236 __u32 pad;
237};
238
239struct kfd_ioctl_dbg_unregister_args {
240 __u32 gpu_id; /* to KFD */
241 __u32 pad;
242};
243
244struct kfd_ioctl_dbg_address_watch_args {
245 __u64 content_ptr; /* a pointer to the actual content */
246 __u32 gpu_id; /* to KFD */
247 __u32 buf_size_in_bytes; /*including gpu_id and buf_size */
248};
249
250struct kfd_ioctl_dbg_wave_control_args {
251 __u64 content_ptr; /* a pointer to the actual content */
252 __u32 gpu_id; /* to KFD */
253 __u32 buf_size_in_bytes; /*including gpu_id and buf_size */
254};
255
256#define KFD_INVALID_FD 0xffffffff
257
258/* Matching HSA_EVENTTYPE */
259#define KFD_IOC_EVENT_SIGNAL 0
260#define KFD_IOC_EVENT_NODECHANGE 1
261#define KFD_IOC_EVENT_DEVICESTATECHANGE 2
262#define KFD_IOC_EVENT_HW_EXCEPTION 3
263#define KFD_IOC_EVENT_SYSTEM_EVENT 4
264#define KFD_IOC_EVENT_DEBUG_EVENT 5
265#define KFD_IOC_EVENT_PROFILE_EVENT 6
266#define KFD_IOC_EVENT_QUEUE_EVENT 7
267#define KFD_IOC_EVENT_MEMORY 8
268
269#define KFD_IOC_WAIT_RESULT_COMPLETE 0
270#define KFD_IOC_WAIT_RESULT_TIMEOUT 1
271#define KFD_IOC_WAIT_RESULT_FAIL 2
272
273#define KFD_SIGNAL_EVENT_LIMIT 4096
274
275/* For kfd_event_data.hw_exception_data.reset_type. */
276#define KFD_HW_EXCEPTION_WHOLE_GPU_RESET 0
277#define KFD_HW_EXCEPTION_PER_ENGINE_RESET 1
278
279/* For kfd_event_data.hw_exception_data.reset_cause. */
280#define KFD_HW_EXCEPTION_GPU_HANG 0
281#define KFD_HW_EXCEPTION_ECC 1
282
283/* For kfd_hsa_memory_exception_data.ErrorType */
284#define KFD_MEM_ERR_NO_RAS 0
285#define KFD_MEM_ERR_SRAM_ECC 1
286#define KFD_MEM_ERR_POISON_CONSUMED 2
287#define KFD_MEM_ERR_GPU_HANG 3
288
289struct kfd_ioctl_create_event_args {
290 __u64 event_page_offset; /* from KFD */
291 __u32 event_trigger_data; /* from KFD - signal events only */
292 __u32 event_type; /* to KFD */
293 __u32 auto_reset; /* to KFD */
294 __u32 node_id; /* to KFD - only valid for certain
295 event types */
296 __u32 event_id; /* from KFD */
297 __u32 event_slot_index; /* from KFD */
298};
299
300struct kfd_ioctl_destroy_event_args {
301 __u32 event_id; /* to KFD */
302 __u32 pad;
303};
304
305struct kfd_ioctl_set_event_args {
306 __u32 event_id; /* to KFD */
307 __u32 pad;
308};
309
310struct kfd_ioctl_reset_event_args {
311 __u32 event_id; /* to KFD */
312 __u32 pad;
313};
314
315struct kfd_memory_exception_failure {
316 __u32 NotPresent; /* Page not present or supervisor privilege */
317 __u32 ReadOnly; /* Write access to a read-only page */
318 __u32 NoExecute; /* Execute access to a page marked NX */
319 __u32 imprecise; /* Can't determine the exact fault address */
320};
321
322/* memory exception data */
323struct kfd_hsa_memory_exception_data {
324 struct kfd_memory_exception_failure failure;
325 __u64 va;
326 __u32 gpu_id;
327 __u32 ErrorType; /* 0 = no RAS error,
328 * 1 = ECC_SRAM,
329 * 2 = Link_SYNFLOOD (poison),
330 * 3 = GPU hang (not attributable to a specific cause),
331 * other values reserved
332 */
333};
334
335/* hw exception data */
336struct kfd_hsa_hw_exception_data {
337 __u32 reset_type;
338 __u32 reset_cause;
339 __u32 memory_lost;
340 __u32 gpu_id;
341};
342
343/* hsa signal event data */
344struct kfd_hsa_signal_event_data {
345 __u64 last_event_age; /* to and from KFD */
346};
347
348/* Event data */
349struct kfd_event_data {
350 union {
351 /* From KFD */
352 struct kfd_hsa_memory_exception_data memory_exception_data;
353 struct kfd_hsa_hw_exception_data hw_exception_data;
354 /* To and From KFD */
355 struct kfd_hsa_signal_event_data signal_event_data;
356 };
357 __u64 kfd_event_data_ext; /* pointer to an extension structure
358 for future exception types */
359 __u32 event_id; /* to KFD */
360 __u32 pad;
361};
362
363struct kfd_ioctl_wait_events_args {
364 __u64 events_ptr; /* pointed to struct
365 kfd_event_data array, to KFD */
366 __u32 num_events; /* to KFD */
367 __u32 wait_for_all; /* to KFD */
368 __u32 timeout; /* to KFD */
369 __u32 wait_result; /* from KFD */
370};
371
372struct kfd_ioctl_set_scratch_backing_va_args {
373 __u64 va_addr; /* to KFD */
374 __u32 gpu_id; /* to KFD */
375 __u32 pad;
376};
377
378struct kfd_ioctl_get_tile_config_args {
379 /* to KFD: pointer to tile array */
380 __u64 tile_config_ptr;
381 /* to KFD: pointer to macro tile array */
382 __u64 macro_tile_config_ptr;
383 /* to KFD: array size allocated by user mode
384 * from KFD: array size filled by kernel
385 */
386 __u32 num_tile_configs;
387 /* to KFD: array size allocated by user mode
388 * from KFD: array size filled by kernel
389 */
390 __u32 num_macro_tile_configs;
391
392 __u32 gpu_id; /* to KFD */
393 __u32 gb_addr_config; /* from KFD */
394 __u32 num_banks; /* from KFD */
395 __u32 num_ranks; /* from KFD */
396 /* struct size can be extended later if needed
397 * without breaking ABI compatibility
398 */
399};
400
401struct kfd_ioctl_set_trap_handler_args {
402 __u64 tba_addr; /* to KFD */
403 __u64 tma_addr; /* to KFD */
404 __u32 gpu_id; /* to KFD */
405 __u32 pad;
406};
407
408struct kfd_ioctl_acquire_vm_args {
409 __u32 drm_fd; /* to KFD */
410 __u32 gpu_id; /* to KFD */
411};
412
413/* Allocation flags: memory types */
414#define KFD_IOC_ALLOC_MEM_FLAGS_VRAM (1 << 0)
415#define KFD_IOC_ALLOC_MEM_FLAGS_GTT (1 << 1)
416#define KFD_IOC_ALLOC_MEM_FLAGS_USERPTR (1 << 2)
417#define KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL (1 << 3)
418#define KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP (1 << 4)
419/* Allocation flags: attributes/access options */
420#define KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE (1 << 31)
421#define KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE (1 << 30)
422#define KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC (1 << 29)
423#define KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE (1 << 28)
424#define KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM (1 << 27)
425#define KFD_IOC_ALLOC_MEM_FLAGS_COHERENT (1 << 26)
426#define KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED (1 << 25)
427#define KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT (1 << 24)
428#define KFD_IOC_ALLOC_MEM_FLAGS_CONTIGUOUS (1 << 23)
429
430/* Allocate memory for later SVM (shared virtual memory) mapping.
431 *
432 * @va_addr: virtual address of the memory to be allocated
433 * all later mappings on all GPUs will use this address
434 * @size: size in bytes
435 * @handle: buffer handle returned to user mode, used to refer to
436 * this allocation for mapping, unmapping and freeing
437 * @mmap_offset: for CPU-mapping the allocation by mmapping a render node
438 * for userptrs this is overloaded to specify the CPU address
439 * @gpu_id: device identifier
440 * @flags: memory type and attributes. See KFD_IOC_ALLOC_MEM_FLAGS above
441 */
442struct kfd_ioctl_alloc_memory_of_gpu_args {
443 __u64 va_addr; /* to KFD */
444 __u64 size; /* to KFD */
445 __u64 handle; /* from KFD */
446 __u64 mmap_offset; /* to KFD (userptr), from KFD (mmap offset) */
447 __u32 gpu_id; /* to KFD */
448 __u32 flags;
449};
450
451/* Free memory allocated with kfd_ioctl_alloc_memory_of_gpu
452 *
453 * @handle: memory handle returned by alloc
454 */
455struct kfd_ioctl_free_memory_of_gpu_args {
456 __u64 handle; /* to KFD */
457};
458
459/* Map memory to one or more GPUs
460 *
461 * @handle: memory handle returned by alloc
462 * @device_ids_array_ptr: array of gpu_ids (__u32 per device)
463 * @n_devices: number of devices in the array
464 * @n_success: number of devices mapped successfully
465 *
466 * @n_success returns information to the caller how many devices from
467 * the start of the array have mapped the buffer successfully. It can
468 * be passed into a subsequent retry call to skip those devices. For
469 * the first call the caller should initialize it to 0.
470 *
471 * If the ioctl completes with return code 0 (success), n_success ==
472 * n_devices.
473 */
474struct kfd_ioctl_map_memory_to_gpu_args {
475 __u64 handle; /* to KFD */
476 __u64 device_ids_array_ptr; /* to KFD */
477 __u32 n_devices; /* to KFD */
478 __u32 n_success; /* to/from KFD */
479};
480
481/* Unmap memory from one or more GPUs
482 *
483 * same arguments as for mapping
484 */
485struct kfd_ioctl_unmap_memory_from_gpu_args {
486 __u64 handle; /* to KFD */
487 __u64 device_ids_array_ptr; /* to KFD */
488 __u32 n_devices; /* to KFD */
489 __u32 n_success; /* to/from KFD */
490};
491
492/* Allocate GWS for specific queue
493 *
494 * @queue_id: queue's id that GWS is allocated for
495 * @num_gws: how many GWS to allocate
496 * @first_gws: index of the first GWS allocated.
497 * only support contiguous GWS allocation
498 */
499struct kfd_ioctl_alloc_queue_gws_args {
500 __u32 queue_id; /* to KFD */
501 __u32 num_gws; /* to KFD */
502 __u32 first_gws; /* from KFD */
503 __u32 pad;
504};
505
506struct kfd_ioctl_get_dmabuf_info_args {
507 __u64 size; /* from KFD */
508 __u64 metadata_ptr; /* to KFD */
509 __u32 metadata_size; /* to KFD (space allocated by user)
510 * from KFD (actual metadata size)
511 */
512 __u32 gpu_id; /* from KFD */
513 __u32 flags; /* from KFD (KFD_IOC_ALLOC_MEM_FLAGS) */
514 __u32 dmabuf_fd; /* to KFD */
515};
516
517struct kfd_ioctl_import_dmabuf_args {
518 __u64 va_addr; /* to KFD */
519 __u64 handle; /* from KFD */
520 __u32 gpu_id; /* to KFD */
521 __u32 dmabuf_fd; /* to KFD */
522};
523
524struct kfd_ioctl_export_dmabuf_args {
525 __u64 handle; /* to KFD */
526 __u32 flags; /* to KFD */
527 __u32 dmabuf_fd; /* from KFD */
528};
529
530/*
531 * KFD SMI(System Management Interface) events
532 */
533enum kfd_smi_event {
534 KFD_SMI_EVENT_NONE = 0, /* not used */
535 KFD_SMI_EVENT_VMFAULT = 1, /* event start counting at 1 */
536 KFD_SMI_EVENT_THERMAL_THROTTLE = 2,
537 KFD_SMI_EVENT_GPU_PRE_RESET = 3,
538 KFD_SMI_EVENT_GPU_POST_RESET = 4,
539 KFD_SMI_EVENT_MIGRATE_START = 5,
540 KFD_SMI_EVENT_MIGRATE_END = 6,
541 KFD_SMI_EVENT_PAGE_FAULT_START = 7,
542 KFD_SMI_EVENT_PAGE_FAULT_END = 8,
543 KFD_SMI_EVENT_QUEUE_EVICTION = 9,
544 KFD_SMI_EVENT_QUEUE_RESTORE = 10,
545 KFD_SMI_EVENT_UNMAP_FROM_GPU = 11,
546 KFD_SMI_EVENT_PROCESS_START = 12,
547 KFD_SMI_EVENT_PROCESS_END = 13,
548
549 /*
550 * max event number, as a flag bit to get events from all processes,
551 * this requires super user permission, otherwise will not be able to
552 * receive event from any process. Without this flag to receive events
553 * from same process.
554 */
555 KFD_SMI_EVENT_ALL_PROCESS = 64
556};
557
558/* The reason of the page migration event */
559enum KFD_MIGRATE_TRIGGERS {
560 KFD_MIGRATE_TRIGGER_PREFETCH, /* Prefetch to GPU VRAM or system memory */
561 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, /* GPU page fault recover */
562 KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU, /* CPU page fault recover */
563 KFD_MIGRATE_TRIGGER_TTM_EVICTION /* TTM eviction */
564};
565
566/* The reason of user queue evition event */
567enum KFD_QUEUE_EVICTION_TRIGGERS {
568 KFD_QUEUE_EVICTION_TRIGGER_SVM, /* SVM buffer migration */
569 KFD_QUEUE_EVICTION_TRIGGER_USERPTR, /* userptr movement */
570 KFD_QUEUE_EVICTION_TRIGGER_TTM, /* TTM move buffer */
571 KFD_QUEUE_EVICTION_TRIGGER_SUSPEND, /* GPU suspend */
572 KFD_QUEUE_EVICTION_CRIU_CHECKPOINT, /* CRIU checkpoint */
573 KFD_QUEUE_EVICTION_CRIU_RESTORE /* CRIU restore */
574};
575
576/* The reason of unmap buffer from GPU event */
577enum KFD_SVM_UNMAP_TRIGGERS {
578 KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY, /* MMU notifier CPU buffer movement */
579 KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE,/* MMU notifier page migration */
580 KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU /* Unmap to free the buffer */
581};
582
583#define KFD_SMI_EVENT_MASK_FROM_INDEX(i) (1ULL << ((i) - 1))
584#define KFD_SMI_EVENT_MSG_SIZE 96
585
586struct kfd_ioctl_smi_events_args {
587 __u32 gpuid; /* to KFD */
588 __u32 anon_fd; /* from KFD */
589};
590
591/*
592 * SVM event tracing via SMI system management interface
593 *
594 * Open event file descriptor
595 * use ioctl AMDKFD_IOC_SMI_EVENTS, pass in gpuid and return a anonymous file
596 * descriptor to receive SMI events.
597 * If calling with sudo permission, then file descriptor can be used to receive
598 * SVM events from all processes, otherwise, to only receive SVM events of same
599 * process.
600 *
601 * To enable the SVM event
602 * Write event file descriptor with KFD_SMI_EVENT_MASK_FROM_INDEX(event) bitmap
603 * mask to start record the event to the kfifo, use bitmap mask combination
604 * for multiple events. New event mask will overwrite the previous event mask.
605 * KFD_SMI_EVENT_MASK_FROM_INDEX(KFD_SMI_EVENT_ALL_PROCESS) bit requires sudo
606 * permisson to receive SVM events from all process.
607 *
608 * To receive the event
609 * Application can poll file descriptor to wait for the events, then read event
610 * from the file into a buffer. Each event is one line string message, starting
611 * with the event id, then the event specific information.
612 *
613 * To decode event information
614 * The following event format string macro can be used with sscanf to decode
615 * the specific event information.
616 * event triggers: the reason to generate the event, defined as enum for unmap,
617 * eviction and migrate events.
618 * node, from, to, prefetch_loc, preferred_loc: GPU ID, or 0 for system memory.
619 * addr: user mode address, in pages
620 * size: in pages
621 * pid: the process ID to generate the event
622 * ns: timestamp in nanosecond-resolution, starts at system boot time but
623 * stops during suspend
624 * migrate_update: GPU page fault is recovered by 'M' for migrate, 'U' for update
625 * rw: 'W' for write page fault, 'R' for read page fault
626 * rescheduled: 'R' if the queue restore failed and rescheduled to try again
627 * error_code: migrate failure error code, 0 if no error
628 */
629#define KFD_EVENT_FMT_UPDATE_GPU_RESET(reset_seq_num, reset_cause)\
630 "%x %s\n", (reset_seq_num), (reset_cause)
631
632#define KFD_EVENT_FMT_THERMAL_THROTTLING(bitmask, counter)\
633 "%llx:%llx\n", (bitmask), (counter)
634
635#define KFD_EVENT_FMT_VMFAULT(pid, task_name)\
636 "%x:%s\n", (pid), (task_name)
637
638#define KFD_EVENT_FMT_PAGEFAULT_START(ns, pid, addr, node, rw)\
639 "%lld -%d @%lx(%x) %c\n", (ns), (pid), (addr), (node), (rw)
640
641#define KFD_EVENT_FMT_PAGEFAULT_END(ns, pid, addr, node, migrate_update)\
642 "%lld -%d @%lx(%x) %c\n", (ns), (pid), (addr), (node), (migrate_update)
643
644#define KFD_EVENT_FMT_MIGRATE_START(ns, pid, start, size, from, to, prefetch_loc,\
645 preferred_loc, migrate_trigger)\
646 "%lld -%d @%lx(%lx) %x->%x %x:%x %d\n", (ns), (pid), (start), (size),\
647 (from), (to), (prefetch_loc), (preferred_loc), (migrate_trigger)
648
649#define KFD_EVENT_FMT_MIGRATE_END(ns, pid, start, size, from, to, migrate_trigger, error_code) \
650 "%lld -%d @%lx(%lx) %x->%x %d %d\n", (ns), (pid), (start), (size),\
651 (from), (to), (migrate_trigger), (error_code)
652
653#define KFD_EVENT_FMT_QUEUE_EVICTION(ns, pid, node, evict_trigger)\
654 "%lld -%d %x %d\n", (ns), (pid), (node), (evict_trigger)
655
656#define KFD_EVENT_FMT_QUEUE_RESTORE(ns, pid, node, rescheduled)\
657 "%lld -%d %x %c\n", (ns), (pid), (node), (rescheduled)
658
659#define KFD_EVENT_FMT_UNMAP_FROM_GPU(ns, pid, addr, size, node, unmap_trigger)\
660 "%lld -%d @%lx(%lx) %x %d\n", (ns), (pid), (addr), (size),\
661 (node), (unmap_trigger)
662
663#define KFD_EVENT_FMT_PROCESS(pid, task_name)\
664 "%x %s\n", (pid), (task_name)
665
666/**************************************************************************************************
667 * CRIU IOCTLs (Checkpoint Restore In Userspace)
668 *
669 * When checkpointing a process, the userspace application will perform:
670 * 1. PROCESS_INFO op to determine current process information. This pauses execution and evicts
671 * all the queues.
672 * 2. CHECKPOINT op to checkpoint process contents (BOs, queues, events, svm-ranges)
673 * 3. UNPAUSE op to un-evict all the queues
674 *
675 * When restoring a process, the CRIU userspace application will perform:
676 *
677 * 1. RESTORE op to restore process contents
678 * 2. RESUME op to start the process
679 *
680 * Note: Queues are forced into an evicted state after a successful PROCESS_INFO. User
681 * application needs to perform an UNPAUSE operation after calling PROCESS_INFO.
682 */
683
684enum kfd_criu_op {
685 KFD_CRIU_OP_PROCESS_INFO,
686 KFD_CRIU_OP_CHECKPOINT,
687 KFD_CRIU_OP_UNPAUSE,
688 KFD_CRIU_OP_RESTORE,
689 KFD_CRIU_OP_RESUME,
690};
691
692/**
693 * kfd_ioctl_criu_args - Arguments perform CRIU operation
694 * @devices: [in/out] User pointer to memory location for devices information.
695 * This is an array of type kfd_criu_device_bucket.
696 * @bos: [in/out] User pointer to memory location for BOs information
697 * This is an array of type kfd_criu_bo_bucket.
698 * @priv_data: [in/out] User pointer to memory location for private data
699 * @priv_data_size: [in/out] Size of priv_data in bytes
700 * @num_devices: [in/out] Number of GPUs used by process. Size of @devices array.
701 * @num_bos [in/out] Number of BOs used by process. Size of @bos array.
702 * @num_objects: [in/out] Number of objects used by process. Objects are opaque to
703 * user application.
704 * @pid: [in/out] PID of the process being checkpointed
705 * @op [in] Type of operation (kfd_criu_op)
706 *
707 * Return: 0 on success, -errno on failure
708 */
709struct kfd_ioctl_criu_args {
710 __u64 devices; /* Used during ops: CHECKPOINT, RESTORE */
711 __u64 bos; /* Used during ops: CHECKPOINT, RESTORE */
712 __u64 priv_data; /* Used during ops: CHECKPOINT, RESTORE */
713 __u64 priv_data_size; /* Used during ops: PROCESS_INFO, RESTORE */
714 __u32 num_devices; /* Used during ops: PROCESS_INFO, RESTORE */
715 __u32 num_bos; /* Used during ops: PROCESS_INFO, RESTORE */
716 __u32 num_objects; /* Used during ops: PROCESS_INFO, RESTORE */
717 __u32 pid; /* Used during ops: PROCESS_INFO, RESUME */
718 __u32 op;
719};
720
721struct kfd_criu_device_bucket {
722 __u32 user_gpu_id;
723 __u32 actual_gpu_id;
724 __u32 drm_fd;
725 __u32 pad;
726};
727
728struct kfd_criu_bo_bucket {
729 __u64 addr;
730 __u64 size;
731 __u64 offset;
732 __u64 restored_offset; /* During restore, updated offset for BO */
733 __u32 gpu_id; /* This is the user_gpu_id */
734 __u32 alloc_flags;
735 __u32 dmabuf_fd;
736 __u32 pad;
737};
738
739/* CRIU IOCTLs - END */
740/**************************************************************************************************/
741
742/* Register offset inside the remapped mmio page
743 */
744enum kfd_mmio_remap {
745 KFD_MMIO_REMAP_HDP_MEM_FLUSH_CNTL = 0,
746 KFD_MMIO_REMAP_HDP_REG_FLUSH_CNTL = 4,
747};
748
749/* Guarantee host access to memory */
750#define KFD_IOCTL_SVM_FLAG_HOST_ACCESS 0x00000001
751/* Fine grained coherency between all devices with access */
752#define KFD_IOCTL_SVM_FLAG_COHERENT 0x00000002
753/* Use any GPU in same hive as preferred device */
754#define KFD_IOCTL_SVM_FLAG_HIVE_LOCAL 0x00000004
755/* GPUs only read, allows replication */
756#define KFD_IOCTL_SVM_FLAG_GPU_RO 0x00000008
757/* Allow execution on GPU */
758#define KFD_IOCTL_SVM_FLAG_GPU_EXEC 0x00000010
759/* GPUs mostly read, may allow similar optimizations as RO, but writes fault */
760#define KFD_IOCTL_SVM_FLAG_GPU_READ_MOSTLY 0x00000020
761/* Keep GPU memory mapping always valid as if XNACK is disable */
762#define KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED 0x00000040
763/* Fine grained coherency between all devices using device-scope atomics */
764#define KFD_IOCTL_SVM_FLAG_EXT_COHERENT 0x00000080
765
766/**
767 * kfd_ioctl_svm_op - SVM ioctl operations
768 *
769 * @KFD_IOCTL_SVM_OP_SET_ATTR: Modify one or more attributes
770 * @KFD_IOCTL_SVM_OP_GET_ATTR: Query one or more attributes
771 */
772enum kfd_ioctl_svm_op {
773 KFD_IOCTL_SVM_OP_SET_ATTR,
774 KFD_IOCTL_SVM_OP_GET_ATTR
775};
776
777/** kfd_ioctl_svm_location - Enum for preferred and prefetch locations
778 *
779 * GPU IDs are used to specify GPUs as preferred and prefetch locations.
780 * Below definitions are used for system memory or for leaving the preferred
781 * location unspecified.
782 */
783enum kfd_ioctl_svm_location {
784 KFD_IOCTL_SVM_LOCATION_SYSMEM = 0,
785 KFD_IOCTL_SVM_LOCATION_UNDEFINED = 0xffffffff
786};
787
788/**
789 * kfd_ioctl_svm_attr_type - SVM attribute types
790 *
791 * @KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: gpuid of the preferred location, 0 for
792 * system memory
793 * @KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: gpuid of the prefetch location, 0 for
794 * system memory. Setting this triggers an
795 * immediate prefetch (migration).
796 * @KFD_IOCTL_SVM_ATTR_ACCESS:
797 * @KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
798 * @KFD_IOCTL_SVM_ATTR_NO_ACCESS: specify memory access for the gpuid given
799 * by the attribute value
800 * @KFD_IOCTL_SVM_ATTR_SET_FLAGS: bitmask of flags to set (see
801 * KFD_IOCTL_SVM_FLAG_...)
802 * @KFD_IOCTL_SVM_ATTR_CLR_FLAGS: bitmask of flags to clear
803 * @KFD_IOCTL_SVM_ATTR_GRANULARITY: migration granularity
804 * (log2 num pages)
805 */
806enum kfd_ioctl_svm_attr_type {
807 KFD_IOCTL_SVM_ATTR_PREFERRED_LOC,
808 KFD_IOCTL_SVM_ATTR_PREFETCH_LOC,
809 KFD_IOCTL_SVM_ATTR_ACCESS,
810 KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE,
811 KFD_IOCTL_SVM_ATTR_NO_ACCESS,
812 KFD_IOCTL_SVM_ATTR_SET_FLAGS,
813 KFD_IOCTL_SVM_ATTR_CLR_FLAGS,
814 KFD_IOCTL_SVM_ATTR_GRANULARITY
815};
816
817/**
818 * kfd_ioctl_svm_attribute - Attributes as pairs of type and value
819 *
820 * The meaning of the @value depends on the attribute type.
821 *
822 * @type: attribute type (see enum @kfd_ioctl_svm_attr_type)
823 * @value: attribute value
824 */
825struct kfd_ioctl_svm_attribute {
826 __u32 type;
827 __u32 value;
828};
829
830/**
831 * kfd_ioctl_svm_args - Arguments for SVM ioctl
832 *
833 * @op specifies the operation to perform (see enum
834 * @kfd_ioctl_svm_op). @start_addr and @size are common for all
835 * operations.
836 *
837 * A variable number of attributes can be given in @attrs.
838 * @nattr specifies the number of attributes. New attributes can be
839 * added in the future without breaking the ABI. If unknown attributes
840 * are given, the function returns -EINVAL.
841 *
842 * @KFD_IOCTL_SVM_OP_SET_ATTR sets attributes for a virtual address
843 * range. It may overlap existing virtual address ranges. If it does,
844 * the existing ranges will be split such that the attribute changes
845 * only apply to the specified address range.
846 *
847 * @KFD_IOCTL_SVM_OP_GET_ATTR returns the intersection of attributes
848 * over all memory in the given range and returns the result as the
849 * attribute value. If different pages have different preferred or
850 * prefetch locations, 0xffffffff will be returned for
851 * @KFD_IOCTL_SVM_ATTR_PREFERRED_LOC or
852 * @KFD_IOCTL_SVM_ATTR_PREFETCH_LOC resepctively. For
853 * @KFD_IOCTL_SVM_ATTR_SET_FLAGS, flags of all pages will be
854 * aggregated by bitwise AND. That means, a flag will be set in the
855 * output, if that flag is set for all pages in the range. For
856 * @KFD_IOCTL_SVM_ATTR_CLR_FLAGS, flags of all pages will be
857 * aggregated by bitwise NOR. That means, a flag will be set in the
858 * output, if that flag is clear for all pages in the range.
859 * The minimum migration granularity throughout the range will be
860 * returned for @KFD_IOCTL_SVM_ATTR_GRANULARITY.
861 *
862 * Querying of accessibility attributes works by initializing the
863 * attribute type to @KFD_IOCTL_SVM_ATTR_ACCESS and the value to the
864 * GPUID being queried. Multiple attributes can be given to allow
865 * querying multiple GPUIDs. The ioctl function overwrites the
866 * attribute type to indicate the access for the specified GPU.
867 */
868struct kfd_ioctl_svm_args {
869 __u64 start_addr;
870 __u64 size;
871 __u32 op;
872 __u32 nattr;
873 /* Variable length array of attributes */
874 struct kfd_ioctl_svm_attribute attrs[];
875};
876
877/**
878 * kfd_ioctl_set_xnack_mode_args - Arguments for set_xnack_mode
879 *
880 * @xnack_enabled: [in/out] Whether to enable XNACK mode for this process
881 *
882 * @xnack_enabled indicates whether recoverable page faults should be
883 * enabled for the current process. 0 means disabled, positive means
884 * enabled, negative means leave unchanged. If enabled, virtual address
885 * translations on GFXv9 and later AMD GPUs can return XNACK and retry
886 * the access until a valid PTE is available. This is used to implement
887 * device page faults.
888 *
889 * On output, @xnack_enabled returns the (new) current mode (0 or
890 * positive). Therefore, a negative input value can be used to query
891 * the current mode without changing it.
892 *
893 * The XNACK mode fundamentally changes the way SVM managed memory works
894 * in the driver, with subtle effects on application performance and
895 * functionality.
896 *
897 * Enabling XNACK mode requires shader programs to be compiled
898 * differently. Furthermore, not all GPUs support changing the mode
899 * per-process. Therefore changing the mode is only allowed while no
900 * user mode queues exist in the process. This ensure that no shader
901 * code is running that may be compiled for the wrong mode. And GPUs
902 * that cannot change to the requested mode will prevent the XNACK
903 * mode from occurring. All GPUs used by the process must be in the
904 * same XNACK mode.
905 *
906 * GFXv8 or older GPUs do not support 48 bit virtual addresses or SVM.
907 * Therefore those GPUs are not considered for the XNACK mode switch.
908 *
909 * Return: 0 on success, -errno on failure
910 */
911struct kfd_ioctl_set_xnack_mode_args {
912 __s32 xnack_enabled;
913};
914
915/* Wave launch override modes */
916enum kfd_dbg_trap_override_mode {
917 KFD_DBG_TRAP_OVERRIDE_OR = 0,
918 KFD_DBG_TRAP_OVERRIDE_REPLACE = 1
919};
920
921/* Wave launch overrides */
922enum kfd_dbg_trap_mask {
923 KFD_DBG_TRAP_MASK_FP_INVALID = 1,
924 KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL = 2,
925 KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO = 4,
926 KFD_DBG_TRAP_MASK_FP_OVERFLOW = 8,
927 KFD_DBG_TRAP_MASK_FP_UNDERFLOW = 16,
928 KFD_DBG_TRAP_MASK_FP_INEXACT = 32,
929 KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO = 64,
930 KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH = 128,
931 KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION = 256,
932 KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START = (1 << 30),
933 KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END = (1 << 31)
934};
935
936/* Wave launch modes */
937enum kfd_dbg_trap_wave_launch_mode {
938 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_NORMAL = 0,
939 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_HALT = 1,
940 KFD_DBG_TRAP_WAVE_LAUNCH_MODE_DEBUG = 3
941};
942
943/* Address watch modes */
944enum kfd_dbg_trap_address_watch_mode {
945 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_READ = 0,
946 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_NONREAD = 1,
947 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_ATOMIC = 2,
948 KFD_DBG_TRAP_ADDRESS_WATCH_MODE_ALL = 3
949};
950
951/* Additional wave settings */
952enum kfd_dbg_trap_flags {
953 KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP = 1,
954 KFD_DBG_TRAP_FLAG_SINGLE_ALU_OP = 2,
955 KFD_DBG_TRAP_FLAG_LDS_OUT_OF_ADDR_RANGE = 4
956};
957
958/* Trap exceptions */
959enum kfd_dbg_trap_exception_code {
960 EC_NONE = 0,
961 /* per queue */
962 EC_QUEUE_WAVE_ABORT = 1,
963 EC_QUEUE_WAVE_TRAP = 2,
964 EC_QUEUE_WAVE_MATH_ERROR = 3,
965 EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION = 4,
966 EC_QUEUE_WAVE_MEMORY_VIOLATION = 5,
967 EC_QUEUE_WAVE_APERTURE_VIOLATION = 6,
968 EC_QUEUE_PACKET_DISPATCH_DIM_INVALID = 16,
969 EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID = 17,
970 EC_QUEUE_PACKET_DISPATCH_CODE_INVALID = 18,
971 EC_QUEUE_PACKET_RESERVED = 19,
972 EC_QUEUE_PACKET_UNSUPPORTED = 20,
973 EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID = 21,
974 EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID = 22,
975 EC_QUEUE_PACKET_VENDOR_UNSUPPORTED = 23,
976 EC_QUEUE_PREEMPTION_ERROR = 30,
977 EC_QUEUE_NEW = 31,
978 /* per device */
979 EC_DEVICE_QUEUE_DELETE = 32,
980 EC_DEVICE_MEMORY_VIOLATION = 33,
981 EC_DEVICE_RAS_ERROR = 34,
982 EC_DEVICE_FATAL_HALT = 35,
983 EC_DEVICE_NEW = 36,
984 /* per process */
985 EC_PROCESS_RUNTIME = 48,
986 EC_PROCESS_DEVICE_REMOVE = 49,
987 EC_MAX
988};
989
990/* Mask generated by ecode in kfd_dbg_trap_exception_code */
991#define KFD_EC_MASK(ecode) (1ULL << (ecode - 1))
992
993/* Masks for exception code type checks below */
994#define KFD_EC_MASK_QUEUE (KFD_EC_MASK(EC_QUEUE_WAVE_ABORT) | \
995 KFD_EC_MASK(EC_QUEUE_WAVE_TRAP) | \
996 KFD_EC_MASK(EC_QUEUE_WAVE_MATH_ERROR) | \
997 KFD_EC_MASK(EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION) | \
998 KFD_EC_MASK(EC_QUEUE_WAVE_MEMORY_VIOLATION) | \
999 KFD_EC_MASK(EC_QUEUE_WAVE_APERTURE_VIOLATION) | \
1000 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_DIM_INVALID) | \
1001 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID) | \
1002 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_CODE_INVALID) | \
1003 KFD_EC_MASK(EC_QUEUE_PACKET_RESERVED) | \
1004 KFD_EC_MASK(EC_QUEUE_PACKET_UNSUPPORTED) | \
1005 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) | \
1006 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) | \
1007 KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED) | \
1008 KFD_EC_MASK(EC_QUEUE_PREEMPTION_ERROR) | \
1009 KFD_EC_MASK(EC_QUEUE_NEW))
1010#define KFD_EC_MASK_DEVICE (KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE) | \
1011 KFD_EC_MASK(EC_DEVICE_RAS_ERROR) | \
1012 KFD_EC_MASK(EC_DEVICE_FATAL_HALT) | \
1013 KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION) | \
1014 KFD_EC_MASK(EC_DEVICE_NEW))
1015#define KFD_EC_MASK_PROCESS (KFD_EC_MASK(EC_PROCESS_RUNTIME) | \
1016 KFD_EC_MASK(EC_PROCESS_DEVICE_REMOVE))
1017#define KFD_EC_MASK_PACKET (KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_DIM_INVALID) | \
1018 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID) | \
1019 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_CODE_INVALID) | \
1020 KFD_EC_MASK(EC_QUEUE_PACKET_RESERVED) | \
1021 KFD_EC_MASK(EC_QUEUE_PACKET_UNSUPPORTED) | \
1022 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) | \
1023 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) | \
1024 KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED))
1025
1026/* Checks for exception code types for KFD search */
1027#define KFD_DBG_EC_IS_VALID(ecode) (ecode > EC_NONE && ecode < EC_MAX)
1028#define KFD_DBG_EC_TYPE_IS_QUEUE(ecode) \
1029 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_QUEUE))
1030#define KFD_DBG_EC_TYPE_IS_DEVICE(ecode) \
1031 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_DEVICE))
1032#define KFD_DBG_EC_TYPE_IS_PROCESS(ecode) \
1033 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_PROCESS))
1034#define KFD_DBG_EC_TYPE_IS_PACKET(ecode) \
1035 (KFD_DBG_EC_IS_VALID(ecode) && !!(KFD_EC_MASK(ecode) & KFD_EC_MASK_PACKET))
1036
1037
1038/* Runtime enable states */
1039enum kfd_dbg_runtime_state {
1040 DEBUG_RUNTIME_STATE_DISABLED = 0,
1041 DEBUG_RUNTIME_STATE_ENABLED = 1,
1042 DEBUG_RUNTIME_STATE_ENABLED_BUSY = 2,
1043 DEBUG_RUNTIME_STATE_ENABLED_ERROR = 3
1044};
1045
1046/* Runtime enable status */
1047struct kfd_runtime_info {
1048 __u64 r_debug;
1049 __u32 runtime_state;
1050 __u32 ttmp_setup;
1051};
1052
1053/* Enable modes for runtime enable */
1054#define KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK 1
1055#define KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK 2
1056
1057/**
1058 * kfd_ioctl_runtime_enable_args - Arguments for runtime enable
1059 *
1060 * Coordinates debug exception signalling and debug device enablement with runtime.
1061 *
1062 * @r_debug - pointer to user struct for sharing information between ROCr and the debuggger
1063 * @mode_mask - mask to set mode
1064 * KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK - enable runtime for debugging, otherwise disable
1065 * KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK - enable trap temporary setup (ignore on disable)
1066 * @capabilities_mask - mask to notify runtime on what KFD supports
1067 *
1068 * Return - 0 on SUCCESS.
1069 * - EBUSY if runtime enable call already pending.
1070 * - EEXIST if user queues already active prior to call.
1071 * If process is debug enabled, runtime enable will enable debug devices and
1072 * wait for debugger process to send runtime exception EC_PROCESS_RUNTIME
1073 * to unblock - see kfd_ioctl_dbg_trap_args.
1074 *
1075 */
1076struct kfd_ioctl_runtime_enable_args {
1077 __u64 r_debug;
1078 __u32 mode_mask;
1079 __u32 capabilities_mask;
1080};
1081
1082/* Queue information */
1083struct kfd_queue_snapshot_entry {
1084 __u64 exception_status;
1085 __u64 ring_base_address;
1086 __u64 write_pointer_address;
1087 __u64 read_pointer_address;
1088 __u64 ctx_save_restore_address;
1089 __u32 queue_id;
1090 __u32 gpu_id;
1091 __u32 ring_size;
1092 __u32 queue_type;
1093 __u32 ctx_save_restore_area_size;
1094 __u32 reserved;
1095};
1096
1097/* Queue status return for suspend/resume */
1098#define KFD_DBG_QUEUE_ERROR_BIT 30
1099#define KFD_DBG_QUEUE_INVALID_BIT 31
1100#define KFD_DBG_QUEUE_ERROR_MASK (1 << KFD_DBG_QUEUE_ERROR_BIT)
1101#define KFD_DBG_QUEUE_INVALID_MASK (1 << KFD_DBG_QUEUE_INVALID_BIT)
1102
1103/* Context save area header information */
1104struct kfd_context_save_area_header {
1105 struct {
1106 __u32 control_stack_offset;
1107 __u32 control_stack_size;
1108 __u32 wave_state_offset;
1109 __u32 wave_state_size;
1110 } wave_state;
1111 __u32 debug_offset;
1112 __u32 debug_size;
1113 __u64 err_payload_addr;
1114 __u32 err_event_id;
1115 __u32 reserved1;
1116};
1117
1118/*
1119 * Debug operations
1120 *
1121 * For specifics on usage and return values, see documentation per operation
1122 * below. Otherwise, generic error returns apply:
1123 * - ESRCH if the process to debug does not exist.
1124 *
1125 * - EINVAL (with KFD_IOC_DBG_TRAP_ENABLE exempt) if operation
1126 * KFD_IOC_DBG_TRAP_ENABLE has not succeeded prior.
1127 * Also returns this error if GPU hardware scheduling is not supported.
1128 *
1129 * - EPERM (with KFD_IOC_DBG_TRAP_DISABLE exempt) if target process is not
1130 * PTRACE_ATTACHED. KFD_IOC_DBG_TRAP_DISABLE is exempt to allow
1131 * clean up of debug mode as long as process is debug enabled.
1132 *
1133 * - EACCES if any DBG_HW_OP (debug hardware operation) is requested when
1134 * AMDKFD_IOC_RUNTIME_ENABLE has not succeeded prior.
1135 *
1136 * - ENODEV if any GPU does not support debugging on a DBG_HW_OP call.
1137 *
1138 * - Other errors may be returned when a DBG_HW_OP occurs while the GPU
1139 * is in a fatal state.
1140 *
1141 */
1142enum kfd_dbg_trap_operations {
1143 KFD_IOC_DBG_TRAP_ENABLE = 0,
1144 KFD_IOC_DBG_TRAP_DISABLE = 1,
1145 KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT = 2,
1146 KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED = 3,
1147 KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE = 4, /* DBG_HW_OP */
1148 KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE = 5, /* DBG_HW_OP */
1149 KFD_IOC_DBG_TRAP_SUSPEND_QUEUES = 6, /* DBG_HW_OP */
1150 KFD_IOC_DBG_TRAP_RESUME_QUEUES = 7, /* DBG_HW_OP */
1151 KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH = 8, /* DBG_HW_OP */
1152 KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH = 9, /* DBG_HW_OP */
1153 KFD_IOC_DBG_TRAP_SET_FLAGS = 10,
1154 KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT = 11,
1155 KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO = 12,
1156 KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT = 13,
1157 KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT = 14
1158};
1159
1160/**
1161 * kfd_ioctl_dbg_trap_enable_args
1162 *
1163 * Arguments for KFD_IOC_DBG_TRAP_ENABLE.
1164 *
1165 * Enables debug session for target process. Call @op KFD_IOC_DBG_TRAP_DISABLE in
1166 * kfd_ioctl_dbg_trap_args to disable debug session.
1167 *
1168 * @exception_mask (IN) - exceptions to raise to the debugger
1169 * @rinfo_ptr (IN) - pointer to runtime info buffer (see kfd_runtime_info)
1170 * @rinfo_size (IN/OUT) - size of runtime info buffer in bytes
1171 * @dbg_fd (IN) - fd the KFD will nofify the debugger with of raised
1172 * exceptions set in exception_mask.
1173 *
1174 * Generic errors apply (see kfd_dbg_trap_operations).
1175 * Return - 0 on SUCCESS.
1176 * Copies KFD saved kfd_runtime_info to @rinfo_ptr on enable.
1177 * Size of kfd_runtime saved by the KFD returned to @rinfo_size.
1178 * - EBADF if KFD cannot get a reference to dbg_fd.
1179 * - EFAULT if KFD cannot copy runtime info to rinfo_ptr.
1180 * - EINVAL if target process is already debug enabled.
1181 *
1182 */
1183struct kfd_ioctl_dbg_trap_enable_args {
1184 __u64 exception_mask;
1185 __u64 rinfo_ptr;
1186 __u32 rinfo_size;
1187 __u32 dbg_fd;
1188};
1189
1190/**
1191 * kfd_ioctl_dbg_trap_send_runtime_event_args
1192 *
1193 *
1194 * Arguments for KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT.
1195 * Raises exceptions to runtime.
1196 *
1197 * @exception_mask (IN) - exceptions to raise to runtime
1198 * @gpu_id (IN) - target device id
1199 * @queue_id (IN) - target queue id
1200 *
1201 * Generic errors apply (see kfd_dbg_trap_operations).
1202 * Return - 0 on SUCCESS.
1203 * - ENODEV if gpu_id not found.
1204 * If exception_mask contains EC_PROCESS_RUNTIME, unblocks pending
1205 * AMDKFD_IOC_RUNTIME_ENABLE call - see kfd_ioctl_runtime_enable_args.
1206 * All other exceptions are raised to runtime through err_payload_addr.
1207 * See kfd_context_save_area_header.
1208 */
1209struct kfd_ioctl_dbg_trap_send_runtime_event_args {
1210 __u64 exception_mask;
1211 __u32 gpu_id;
1212 __u32 queue_id;
1213};
1214
1215/**
1216 * kfd_ioctl_dbg_trap_set_exceptions_enabled_args
1217 *
1218 * Arguments for KFD_IOC_SET_EXCEPTIONS_ENABLED
1219 * Set new exceptions to be raised to the debugger.
1220 *
1221 * @exception_mask (IN) - new exceptions to raise the debugger
1222 *
1223 * Generic errors apply (see kfd_dbg_trap_operations).
1224 * Return - 0 on SUCCESS.
1225 */
1226struct kfd_ioctl_dbg_trap_set_exceptions_enabled_args {
1227 __u64 exception_mask;
1228};
1229
1230/**
1231 * kfd_ioctl_dbg_trap_set_wave_launch_override_args
1232 *
1233 * Arguments for KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE
1234 * Enable HW exceptions to raise trap.
1235 *
1236 * @override_mode (IN) - see kfd_dbg_trap_override_mode
1237 * @enable_mask (IN/OUT) - reference kfd_dbg_trap_mask.
1238 * IN is the override modes requested to be enabled.
1239 * OUT is referenced in Return below.
1240 * @support_request_mask (IN/OUT) - reference kfd_dbg_trap_mask.
1241 * IN is the override modes requested for support check.
1242 * OUT is referenced in Return below.
1243 *
1244 * Generic errors apply (see kfd_dbg_trap_operations).
1245 * Return - 0 on SUCCESS.
1246 * Previous enablement is returned in @enable_mask.
1247 * Actual override support is returned in @support_request_mask.
1248 * - EINVAL if override mode is not supported.
1249 * - EACCES if trap support requested is not actually supported.
1250 * i.e. enable_mask (IN) is not a subset of support_request_mask (OUT).
1251 * Otherwise it is considered a generic error (see kfd_dbg_trap_operations).
1252 */
1253struct kfd_ioctl_dbg_trap_set_wave_launch_override_args {
1254 __u32 override_mode;
1255 __u32 enable_mask;
1256 __u32 support_request_mask;
1257 __u32 pad;
1258};
1259
1260/**
1261 * kfd_ioctl_dbg_trap_set_wave_launch_mode_args
1262 *
1263 * Arguments for KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE
1264 * Set wave launch mode.
1265 *
1266 * @mode (IN) - see kfd_dbg_trap_wave_launch_mode
1267 *
1268 * Generic errors apply (see kfd_dbg_trap_operations).
1269 * Return - 0 on SUCCESS.
1270 */
1271struct kfd_ioctl_dbg_trap_set_wave_launch_mode_args {
1272 __u32 launch_mode;
1273 __u32 pad;
1274};
1275
1276/**
1277 * kfd_ioctl_dbg_trap_suspend_queues_ags
1278 *
1279 * Arguments for KFD_IOC_DBG_TRAP_SUSPEND_QUEUES
1280 * Suspend queues.
1281 *
1282 * @exception_mask (IN) - raised exceptions to clear
1283 * @queue_array_ptr (IN) - pointer to array of queue ids (u32 per queue id)
1284 * to suspend
1285 * @num_queues (IN) - number of queues to suspend in @queue_array_ptr
1286 * @grace_period (IN) - wave time allowance before preemption
1287 * per 1K GPU clock cycle unit
1288 *
1289 * Generic errors apply (see kfd_dbg_trap_operations).
1290 * Destruction of a suspended queue is blocked until the queue is
1291 * resumed. This allows the debugger to access queue information and
1292 * the its context save area without running into a race condition on
1293 * queue destruction.
1294 * Automatically copies per queue context save area header information
1295 * into the save area base
1296 * (see kfd_queue_snapshot_entry and kfd_context_save_area_header).
1297 *
1298 * Return - Number of queues suspended on SUCCESS.
1299 * . KFD_DBG_QUEUE_ERROR_MASK and KFD_DBG_QUEUE_INVALID_MASK masked
1300 * for each queue id in @queue_array_ptr array reports unsuccessful
1301 * suspend reason.
1302 * KFD_DBG_QUEUE_ERROR_MASK = HW failure.
1303 * KFD_DBG_QUEUE_INVALID_MASK = queue does not exist, is new or
1304 * is being destroyed.
1305 */
1306struct kfd_ioctl_dbg_trap_suspend_queues_args {
1307 __u64 exception_mask;
1308 __u64 queue_array_ptr;
1309 __u32 num_queues;
1310 __u32 grace_period;
1311};
1312
1313/**
1314 * kfd_ioctl_dbg_trap_resume_queues_args
1315 *
1316 * Arguments for KFD_IOC_DBG_TRAP_RESUME_QUEUES
1317 * Resume queues.
1318 *
1319 * @queue_array_ptr (IN) - pointer to array of queue ids (u32 per queue id)
1320 * to resume
1321 * @num_queues (IN) - number of queues to resume in @queue_array_ptr
1322 *
1323 * Generic errors apply (see kfd_dbg_trap_operations).
1324 * Return - Number of queues resumed on SUCCESS.
1325 * KFD_DBG_QUEUE_ERROR_MASK and KFD_DBG_QUEUE_INVALID_MASK mask
1326 * for each queue id in @queue_array_ptr array reports unsuccessful
1327 * resume reason.
1328 * KFD_DBG_QUEUE_ERROR_MASK = HW failure.
1329 * KFD_DBG_QUEUE_INVALID_MASK = queue does not exist.
1330 */
1331struct kfd_ioctl_dbg_trap_resume_queues_args {
1332 __u64 queue_array_ptr;
1333 __u32 num_queues;
1334 __u32 pad;
1335};
1336
1337/**
1338 * kfd_ioctl_dbg_trap_set_node_address_watch_args
1339 *
1340 * Arguments for KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH
1341 * Sets address watch for device.
1342 *
1343 * @address (IN) - watch address to set
1344 * @mode (IN) - see kfd_dbg_trap_address_watch_mode
1345 * @mask (IN) - watch address mask
1346 * @gpu_id (IN) - target gpu to set watch point
1347 * @id (OUT) - watch id allocated
1348 *
1349 * Generic errors apply (see kfd_dbg_trap_operations).
1350 * Return - 0 on SUCCESS.
1351 * Allocated watch ID returned to @id.
1352 * - ENODEV if gpu_id not found.
1353 * - ENOMEM if watch IDs can be allocated
1354 */
1355struct kfd_ioctl_dbg_trap_set_node_address_watch_args {
1356 __u64 address;
1357 __u32 mode;
1358 __u32 mask;
1359 __u32 gpu_id;
1360 __u32 id;
1361};
1362
1363/**
1364 * kfd_ioctl_dbg_trap_clear_node_address_watch_args
1365 *
1366 * Arguments for KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH
1367 * Clear address watch for device.
1368 *
1369 * @gpu_id (IN) - target device to clear watch point
1370 * @id (IN) - allocated watch id to clear
1371 *
1372 * Generic errors apply (see kfd_dbg_trap_operations).
1373 * Return - 0 on SUCCESS.
1374 * - ENODEV if gpu_id not found.
1375 * - EINVAL if watch ID has not been allocated.
1376 */
1377struct kfd_ioctl_dbg_trap_clear_node_address_watch_args {
1378 __u32 gpu_id;
1379 __u32 id;
1380};
1381
1382/**
1383 * kfd_ioctl_dbg_trap_set_flags_args
1384 *
1385 * Arguments for KFD_IOC_DBG_TRAP_SET_FLAGS
1386 * Sets flags for wave behaviour.
1387 *
1388 * @flags (IN/OUT) - IN = flags to enable, OUT = flags previously enabled
1389 *
1390 * Generic errors apply (see kfd_dbg_trap_operations).
1391 * Return - 0 on SUCCESS.
1392 * - EACCESS if any debug device does not allow flag options.
1393 */
1394struct kfd_ioctl_dbg_trap_set_flags_args {
1395 __u32 flags;
1396 __u32 pad;
1397};
1398
1399/**
1400 * kfd_ioctl_dbg_trap_query_debug_event_args
1401 *
1402 * Arguments for KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT
1403 *
1404 * Find one or more raised exceptions. This function can return multiple
1405 * exceptions from a single queue or a single device with one call. To find
1406 * all raised exceptions, this function must be called repeatedly until it
1407 * returns -EAGAIN. Returned exceptions can optionally be cleared by
1408 * setting the corresponding bit in the @exception_mask input parameter.
1409 * However, clearing an exception prevents retrieving further information
1410 * about it with KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO.
1411 *
1412 * @exception_mask (IN/OUT) - exception to clear (IN) and raised (OUT)
1413 * @gpu_id (OUT) - gpu id of exceptions raised
1414 * @queue_id (OUT) - queue id of exceptions raised
1415 *
1416 * Generic errors apply (see kfd_dbg_trap_operations).
1417 * Return - 0 on raised exception found
1418 * Raised exceptions found are returned in @exception mask
1419 * with reported source id returned in @gpu_id or @queue_id.
1420 * - EAGAIN if no raised exception has been found
1421 */
1422struct kfd_ioctl_dbg_trap_query_debug_event_args {
1423 __u64 exception_mask;
1424 __u32 gpu_id;
1425 __u32 queue_id;
1426};
1427
1428/**
1429 * kfd_ioctl_dbg_trap_query_exception_info_args
1430 *
1431 * Arguments KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO
1432 * Get additional info on raised exception.
1433 *
1434 * @info_ptr (IN) - pointer to exception info buffer to copy to
1435 * @info_size (IN/OUT) - exception info buffer size (bytes)
1436 * @source_id (IN) - target gpu or queue id
1437 * @exception_code (IN) - target exception
1438 * @clear_exception (IN) - clear raised @exception_code exception
1439 * (0 = false, 1 = true)
1440 *
1441 * Generic errors apply (see kfd_dbg_trap_operations).
1442 * Return - 0 on SUCCESS.
1443 * If @exception_code is EC_DEVICE_MEMORY_VIOLATION, copy @info_size(OUT)
1444 * bytes of memory exception data to @info_ptr.
1445 * If @exception_code is EC_PROCESS_RUNTIME, copy saved
1446 * kfd_runtime_info to @info_ptr.
1447 * Actual required @info_ptr size (bytes) is returned in @info_size.
1448 */
1449struct kfd_ioctl_dbg_trap_query_exception_info_args {
1450 __u64 info_ptr;
1451 __u32 info_size;
1452 __u32 source_id;
1453 __u32 exception_code;
1454 __u32 clear_exception;
1455};
1456
1457/**
1458 * kfd_ioctl_dbg_trap_get_queue_snapshot_args
1459 *
1460 * Arguments KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT
1461 * Get queue information.
1462 *
1463 * @exception_mask (IN) - exceptions raised to clear
1464 * @snapshot_buf_ptr (IN) - queue snapshot entry buffer (see kfd_queue_snapshot_entry)
1465 * @num_queues (IN/OUT) - number of queue snapshot entries
1466 * The debugger specifies the size of the array allocated in @num_queues.
1467 * KFD returns the number of queues that actually existed. If this is
1468 * larger than the size specified by the debugger, KFD will not overflow
1469 * the array allocated by the debugger.
1470 *
1471 * @entry_size (IN/OUT) - size per entry in bytes
1472 * The debugger specifies sizeof(struct kfd_queue_snapshot_entry) in
1473 * @entry_size. KFD returns the number of bytes actually populated per
1474 * entry. The debugger should use the KFD_IOCTL_MINOR_VERSION to determine,
1475 * which fields in struct kfd_queue_snapshot_entry are valid. This allows
1476 * growing the ABI in a backwards compatible manner.
1477 * Note that entry_size(IN) should still be used to stride the snapshot buffer in the
1478 * event that it's larger than actual kfd_queue_snapshot_entry.
1479 *
1480 * Generic errors apply (see kfd_dbg_trap_operations).
1481 * Return - 0 on SUCCESS.
1482 * Copies @num_queues(IN) queue snapshot entries of size @entry_size(IN)
1483 * into @snapshot_buf_ptr if @num_queues(IN) > 0.
1484 * Otherwise return @num_queues(OUT) queue snapshot entries that exist.
1485 */
1486struct kfd_ioctl_dbg_trap_queue_snapshot_args {
1487 __u64 exception_mask;
1488 __u64 snapshot_buf_ptr;
1489 __u32 num_queues;
1490 __u32 entry_size;
1491};
1492
1493/**
1494 * kfd_ioctl_dbg_trap_get_device_snapshot_args
1495 *
1496 * Arguments for KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT
1497 * Get device information.
1498 *
1499 * @exception_mask (IN) - exceptions raised to clear
1500 * @snapshot_buf_ptr (IN) - pointer to snapshot buffer (see kfd_dbg_device_info_entry)
1501 * @num_devices (IN/OUT) - number of debug devices to snapshot
1502 * The debugger specifies the size of the array allocated in @num_devices.
1503 * KFD returns the number of devices that actually existed. If this is
1504 * larger than the size specified by the debugger, KFD will not overflow
1505 * the array allocated by the debugger.
1506 *
1507 * @entry_size (IN/OUT) - size per entry in bytes
1508 * The debugger specifies sizeof(struct kfd_dbg_device_info_entry) in
1509 * @entry_size. KFD returns the number of bytes actually populated. The
1510 * debugger should use KFD_IOCTL_MINOR_VERSION to determine, which fields
1511 * in struct kfd_dbg_device_info_entry are valid. This allows growing the
1512 * ABI in a backwards compatible manner.
1513 * Note that entry_size(IN) should still be used to stride the snapshot buffer in the
1514 * event that it's larger than actual kfd_dbg_device_info_entry.
1515 *
1516 * Generic errors apply (see kfd_dbg_trap_operations).
1517 * Return - 0 on SUCCESS.
1518 * Copies @num_devices(IN) device snapshot entries of size @entry_size(IN)
1519 * into @snapshot_buf_ptr if @num_devices(IN) > 0.
1520 * Otherwise return @num_devices(OUT) queue snapshot entries that exist.
1521 */
1522struct kfd_ioctl_dbg_trap_device_snapshot_args {
1523 __u64 exception_mask;
1524 __u64 snapshot_buf_ptr;
1525 __u32 num_devices;
1526 __u32 entry_size;
1527};
1528
1529/**
1530 * kfd_ioctl_dbg_trap_args
1531 *
1532 * Arguments to debug target process.
1533 *
1534 * @pid - target process to debug
1535 * @op - debug operation (see kfd_dbg_trap_operations)
1536 *
1537 * @op determines which union struct args to use.
1538 * Refer to kern docs for each kfd_ioctl_dbg_trap_*_args struct.
1539 */
1540struct kfd_ioctl_dbg_trap_args {
1541 __u32 pid;
1542 __u32 op;
1543
1544 union {
1545 struct kfd_ioctl_dbg_trap_enable_args enable;
1546 struct kfd_ioctl_dbg_trap_send_runtime_event_args send_runtime_event;
1547 struct kfd_ioctl_dbg_trap_set_exceptions_enabled_args set_exceptions_enabled;
1548 struct kfd_ioctl_dbg_trap_set_wave_launch_override_args launch_override;
1549 struct kfd_ioctl_dbg_trap_set_wave_launch_mode_args launch_mode;
1550 struct kfd_ioctl_dbg_trap_suspend_queues_args suspend_queues;
1551 struct kfd_ioctl_dbg_trap_resume_queues_args resume_queues;
1552 struct kfd_ioctl_dbg_trap_set_node_address_watch_args set_node_address_watch;
1553 struct kfd_ioctl_dbg_trap_clear_node_address_watch_args clear_node_address_watch;
1554 struct kfd_ioctl_dbg_trap_set_flags_args set_flags;
1555 struct kfd_ioctl_dbg_trap_query_debug_event_args query_debug_event;
1556 struct kfd_ioctl_dbg_trap_query_exception_info_args query_exception_info;
1557 struct kfd_ioctl_dbg_trap_queue_snapshot_args queue_snapshot;
1558 struct kfd_ioctl_dbg_trap_device_snapshot_args device_snapshot;
1559 };
1560};
1561
1562#define KFD_IOC_PROFILER_VERSION_NUM 1
1563enum kfd_profiler_ops {
1564 KFD_IOC_PROFILER_PMC = 0,
1565 KFD_IOC_PROFILER_VERSION = 2,
1566 KFD_IOC_PROFILER_PTL_CONTROL = 3,
1567};
1568
1569/**
1570 * Enables/Disables GPU Specific profiler settings
1571 */
1572struct kfd_ioctl_pmc_settings {
1573 __u32 gpu_id; /* This is the user_gpu_id */
1574 __u32 lock; /* Lock GPU for Profiling */
1575 __u32 perfcount_enable; /* Force Perfcount Enable for queues on GPU */
1576};
1577
1578struct kfd_ioctl_ptl_control {
1579 __u32 gpu_id; /* user_gpu_id */
1580 __u32 enable; /* set 1 to enable PTL, set 0 to disable PTL */
1581};
1582
1583struct kfd_ioctl_profiler_args {
1584 __u32 op; /* kfd_profiler_op */
1585 union {
1586 struct kfd_ioctl_pmc_settings pmc;
1587 struct kfd_ioctl_ptl_control ptl;
1588 __u32 version; /* KFD_IOC_PROFILER_VERSION_NUM */
1589 };
1590};
1591
1592#define AMDKFD_IOCTL_BASE 'K'
1593#define AMDKFD_IO(nr) _IO(AMDKFD_IOCTL_BASE, nr)
1594#define AMDKFD_IOR(nr, type) _IOR(AMDKFD_IOCTL_BASE, nr, type)
1595#define AMDKFD_IOW(nr, type) _IOW(AMDKFD_IOCTL_BASE, nr, type)
1596#define AMDKFD_IOWR(nr, type) _IOWR(AMDKFD_IOCTL_BASE, nr, type)
1597
1598#define AMDKFD_IOC_GET_VERSION \
1599 AMDKFD_IOR(0x01, struct kfd_ioctl_get_version_args)
1600
1601#define AMDKFD_IOC_CREATE_QUEUE \
1602 AMDKFD_IOWR(0x02, struct kfd_ioctl_create_queue_args)
1603
1604#define AMDKFD_IOC_DESTROY_QUEUE \
1605 AMDKFD_IOWR(0x03, struct kfd_ioctl_destroy_queue_args)
1606
1607#define AMDKFD_IOC_SET_MEMORY_POLICY \
1608 AMDKFD_IOW(0x04, struct kfd_ioctl_set_memory_policy_args)
1609
1610#define AMDKFD_IOC_GET_CLOCK_COUNTERS \
1611 AMDKFD_IOWR(0x05, struct kfd_ioctl_get_clock_counters_args)
1612
1613#define AMDKFD_IOC_GET_PROCESS_APERTURES \
1614 AMDKFD_IOR(0x06, struct kfd_ioctl_get_process_apertures_args)
1615
1616#define AMDKFD_IOC_UPDATE_QUEUE \
1617 AMDKFD_IOW(0x07, struct kfd_ioctl_update_queue_args)
1618
1619#define AMDKFD_IOC_CREATE_EVENT \
1620 AMDKFD_IOWR(0x08, struct kfd_ioctl_create_event_args)
1621
1622#define AMDKFD_IOC_DESTROY_EVENT \
1623 AMDKFD_IOW(0x09, struct kfd_ioctl_destroy_event_args)
1624
1625#define AMDKFD_IOC_SET_EVENT \
1626 AMDKFD_IOW(0x0A, struct kfd_ioctl_set_event_args)
1627
1628#define AMDKFD_IOC_RESET_EVENT \
1629 AMDKFD_IOW(0x0B, struct kfd_ioctl_reset_event_args)
1630
1631#define AMDKFD_IOC_WAIT_EVENTS \
1632 AMDKFD_IOWR(0x0C, struct kfd_ioctl_wait_events_args)
1633
1634#define AMDKFD_IOC_DBG_REGISTER_DEPRECATED \
1635 AMDKFD_IOW(0x0D, struct kfd_ioctl_dbg_register_args)
1636
1637#define AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED \
1638 AMDKFD_IOW(0x0E, struct kfd_ioctl_dbg_unregister_args)
1639
1640#define AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED \
1641 AMDKFD_IOW(0x0F, struct kfd_ioctl_dbg_address_watch_args)
1642
1643#define AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED \
1644 AMDKFD_IOW(0x10, struct kfd_ioctl_dbg_wave_control_args)
1645
1646#define AMDKFD_IOC_SET_SCRATCH_BACKING_VA \
1647 AMDKFD_IOWR(0x11, struct kfd_ioctl_set_scratch_backing_va_args)
1648
1649#define AMDKFD_IOC_GET_TILE_CONFIG \
1650 AMDKFD_IOWR(0x12, struct kfd_ioctl_get_tile_config_args)
1651
1652#define AMDKFD_IOC_SET_TRAP_HANDLER \
1653 AMDKFD_IOW(0x13, struct kfd_ioctl_set_trap_handler_args)
1654
1655#define AMDKFD_IOC_GET_PROCESS_APERTURES_NEW \
1656 AMDKFD_IOWR(0x14, \
1657 struct kfd_ioctl_get_process_apertures_new_args)
1658
1659#define AMDKFD_IOC_ACQUIRE_VM \
1660 AMDKFD_IOW(0x15, struct kfd_ioctl_acquire_vm_args)
1661
1662#define AMDKFD_IOC_ALLOC_MEMORY_OF_GPU \
1663 AMDKFD_IOWR(0x16, struct kfd_ioctl_alloc_memory_of_gpu_args)
1664
1665#define AMDKFD_IOC_FREE_MEMORY_OF_GPU \
1666 AMDKFD_IOW(0x17, struct kfd_ioctl_free_memory_of_gpu_args)
1667
1668#define AMDKFD_IOC_MAP_MEMORY_TO_GPU \
1669 AMDKFD_IOWR(0x18, struct kfd_ioctl_map_memory_to_gpu_args)
1670
1671#define AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU \
1672 AMDKFD_IOWR(0x19, struct kfd_ioctl_unmap_memory_from_gpu_args)
1673
1674#define AMDKFD_IOC_SET_CU_MASK \
1675 AMDKFD_IOW(0x1A, struct kfd_ioctl_set_cu_mask_args)
1676
1677#define AMDKFD_IOC_GET_QUEUE_WAVE_STATE \
1678 AMDKFD_IOWR(0x1B, struct kfd_ioctl_get_queue_wave_state_args)
1679
1680#define AMDKFD_IOC_GET_DMABUF_INFO \
1681 AMDKFD_IOWR(0x1C, struct kfd_ioctl_get_dmabuf_info_args)
1682
1683#define AMDKFD_IOC_IMPORT_DMABUF \
1684 AMDKFD_IOWR(0x1D, struct kfd_ioctl_import_dmabuf_args)
1685
1686#define AMDKFD_IOC_ALLOC_QUEUE_GWS \
1687 AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
1688
1689#define AMDKFD_IOC_SMI_EVENTS \
1690 AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
1691
1692#define AMDKFD_IOC_SVM AMDKFD_IOWR(0x20, struct kfd_ioctl_svm_args)
1693
1694#define AMDKFD_IOC_SET_XNACK_MODE \
1695 AMDKFD_IOWR(0x21, struct kfd_ioctl_set_xnack_mode_args)
1696
1697#define AMDKFD_IOC_CRIU_OP \
1698 AMDKFD_IOWR(0x22, struct kfd_ioctl_criu_args)
1699
1700#define AMDKFD_IOC_AVAILABLE_MEMORY \
1701 AMDKFD_IOWR(0x23, struct kfd_ioctl_get_available_memory_args)
1702
1703#define AMDKFD_IOC_EXPORT_DMABUF \
1704 AMDKFD_IOWR(0x24, struct kfd_ioctl_export_dmabuf_args)
1705
1706#define AMDKFD_IOC_RUNTIME_ENABLE \
1707 AMDKFD_IOWR(0x25, struct kfd_ioctl_runtime_enable_args)
1708
1709#define AMDKFD_IOC_DBG_TRAP \
1710 AMDKFD_IOWR(0x26, struct kfd_ioctl_dbg_trap_args)
1711
1712#define AMDKFD_IOC_CREATE_PROCESS \
1713 AMDKFD_IO(0x27)
1714
1715#define AMDKFD_IOC_PROFILER \
1716 AMDKFD_IOWR(0x28, struct kfd_ioctl_profiler_args)
1717
1718#define AMDKFD_COMMAND_START 0x01
1719#define AMDKFD_COMMAND_END 0x29
1720
1721#endif