xref: /linux-6.15/include/uapi/linux/kfd_ioctl.h (revision d7cfea33)
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  */
39 #define KFD_IOCTL_MAJOR_VERSION 1
40 #define KFD_IOCTL_MINOR_VERSION 9
41 
42 struct kfd_ioctl_get_version_args {
43 	__u32 major_version;	/* from KFD */
44 	__u32 minor_version;	/* from KFD */
45 };
46 
47 /* For kfd_ioctl_create_queue_args.queue_type. */
48 #define KFD_IOC_QUEUE_TYPE_COMPUTE		0x0
49 #define KFD_IOC_QUEUE_TYPE_SDMA			0x1
50 #define KFD_IOC_QUEUE_TYPE_COMPUTE_AQL		0x2
51 #define KFD_IOC_QUEUE_TYPE_SDMA_XGMI		0x3
52 
53 #define KFD_MAX_QUEUE_PERCENTAGE	100
54 #define KFD_MAX_QUEUE_PRIORITY		15
55 
56 struct kfd_ioctl_create_queue_args {
57 	__u64 ring_base_address;	/* to KFD */
58 	__u64 write_pointer_address;	/* from KFD */
59 	__u64 read_pointer_address;	/* from KFD */
60 	__u64 doorbell_offset;	/* from KFD */
61 
62 	__u32 ring_size;		/* to KFD */
63 	__u32 gpu_id;		/* to KFD */
64 	__u32 queue_type;		/* to KFD */
65 	__u32 queue_percentage;	/* to KFD */
66 	__u32 queue_priority;	/* to KFD */
67 	__u32 queue_id;		/* from KFD */
68 
69 	__u64 eop_buffer_address;	/* to KFD */
70 	__u64 eop_buffer_size;	/* to KFD */
71 	__u64 ctx_save_restore_address; /* to KFD */
72 	__u32 ctx_save_restore_size;	/* to KFD */
73 	__u32 ctl_stack_size;		/* to KFD */
74 };
75 
76 struct kfd_ioctl_destroy_queue_args {
77 	__u32 queue_id;		/* to KFD */
78 	__u32 pad;
79 };
80 
81 struct kfd_ioctl_update_queue_args {
82 	__u64 ring_base_address;	/* to KFD */
83 
84 	__u32 queue_id;		/* to KFD */
85 	__u32 ring_size;		/* to KFD */
86 	__u32 queue_percentage;	/* to KFD */
87 	__u32 queue_priority;	/* to KFD */
88 };
89 
90 struct kfd_ioctl_set_cu_mask_args {
91 	__u32 queue_id;		/* to KFD */
92 	__u32 num_cu_mask;		/* to KFD */
93 	__u64 cu_mask_ptr;		/* to KFD */
94 };
95 
96 struct kfd_ioctl_get_queue_wave_state_args {
97 	__u64 ctl_stack_address;	/* to KFD */
98 	__u32 ctl_stack_used_size;	/* from KFD */
99 	__u32 save_area_used_size;	/* from KFD */
100 	__u32 queue_id;			/* to KFD */
101 	__u32 pad;
102 };
103 
104 struct kfd_ioctl_get_available_memory_args {
105 	__u64 available;	/* from KFD */
106 	__u32 gpu_id;		/* to KFD */
107 	__u32 pad;
108 };
109 
110 /* For kfd_ioctl_set_memory_policy_args.default_policy and alternate_policy */
111 #define KFD_IOC_CACHE_POLICY_COHERENT 0
112 #define KFD_IOC_CACHE_POLICY_NONCOHERENT 1
113 
114 struct kfd_ioctl_set_memory_policy_args {
115 	__u64 alternate_aperture_base;	/* to KFD */
116 	__u64 alternate_aperture_size;	/* to KFD */
117 
118 	__u32 gpu_id;			/* to KFD */
119 	__u32 default_policy;		/* to KFD */
120 	__u32 alternate_policy;		/* to KFD */
121 	__u32 pad;
122 };
123 
124 /*
125  * All counters are monotonic. They are used for profiling of compute jobs.
126  * The profiling is done by userspace.
127  *
128  * In case of GPU reset, the counter should not be affected.
129  */
130 
131 struct kfd_ioctl_get_clock_counters_args {
132 	__u64 gpu_clock_counter;	/* from KFD */
133 	__u64 cpu_clock_counter;	/* from KFD */
134 	__u64 system_clock_counter;	/* from KFD */
135 	__u64 system_clock_freq;	/* from KFD */
136 
137 	__u32 gpu_id;		/* to KFD */
138 	__u32 pad;
139 };
140 
141 struct kfd_process_device_apertures {
142 	__u64 lds_base;		/* from KFD */
143 	__u64 lds_limit;		/* from KFD */
144 	__u64 scratch_base;		/* from KFD */
145 	__u64 scratch_limit;		/* from KFD */
146 	__u64 gpuvm_base;		/* from KFD */
147 	__u64 gpuvm_limit;		/* from KFD */
148 	__u32 gpu_id;		/* from KFD */
149 	__u32 pad;
150 };
151 
152 /*
153  * AMDKFD_IOC_GET_PROCESS_APERTURES is deprecated. Use
154  * AMDKFD_IOC_GET_PROCESS_APERTURES_NEW instead, which supports an
155  * unlimited number of GPUs.
156  */
157 #define NUM_OF_SUPPORTED_GPUS 7
158 struct kfd_ioctl_get_process_apertures_args {
159 	struct kfd_process_device_apertures
160 			process_apertures[NUM_OF_SUPPORTED_GPUS];/* from KFD */
161 
162 	/* from KFD, should be in the range [1 - NUM_OF_SUPPORTED_GPUS] */
163 	__u32 num_of_nodes;
164 	__u32 pad;
165 };
166 
167 struct kfd_ioctl_get_process_apertures_new_args {
168 	/* User allocated. Pointer to struct kfd_process_device_apertures
169 	 * filled in by Kernel
170 	 */
171 	__u64 kfd_process_device_apertures_ptr;
172 	/* to KFD - indicates amount of memory present in
173 	 *  kfd_process_device_apertures_ptr
174 	 * from KFD - Number of entries filled by KFD.
175 	 */
176 	__u32 num_of_nodes;
177 	__u32 pad;
178 };
179 
180 #define MAX_ALLOWED_NUM_POINTS    100
181 #define MAX_ALLOWED_AW_BUFF_SIZE 4096
182 #define MAX_ALLOWED_WAC_BUFF_SIZE  128
183 
184 struct kfd_ioctl_dbg_register_args {
185 	__u32 gpu_id;		/* to KFD */
186 	__u32 pad;
187 };
188 
189 struct kfd_ioctl_dbg_unregister_args {
190 	__u32 gpu_id;		/* to KFD */
191 	__u32 pad;
192 };
193 
194 struct kfd_ioctl_dbg_address_watch_args {
195 	__u64 content_ptr;		/* a pointer to the actual content */
196 	__u32 gpu_id;		/* to KFD */
197 	__u32 buf_size_in_bytes;	/*including gpu_id and buf_size */
198 };
199 
200 struct kfd_ioctl_dbg_wave_control_args {
201 	__u64 content_ptr;		/* a pointer to the actual content */
202 	__u32 gpu_id;		/* to KFD */
203 	__u32 buf_size_in_bytes;	/*including gpu_id and buf_size */
204 };
205 
206 #define KFD_INVALID_FD     0xffffffff
207 
208 /* Matching HSA_EVENTTYPE */
209 #define KFD_IOC_EVENT_SIGNAL			0
210 #define KFD_IOC_EVENT_NODECHANGE		1
211 #define KFD_IOC_EVENT_DEVICESTATECHANGE		2
212 #define KFD_IOC_EVENT_HW_EXCEPTION		3
213 #define KFD_IOC_EVENT_SYSTEM_EVENT		4
214 #define KFD_IOC_EVENT_DEBUG_EVENT		5
215 #define KFD_IOC_EVENT_PROFILE_EVENT		6
216 #define KFD_IOC_EVENT_QUEUE_EVENT		7
217 #define KFD_IOC_EVENT_MEMORY			8
218 
219 #define KFD_IOC_WAIT_RESULT_COMPLETE		0
220 #define KFD_IOC_WAIT_RESULT_TIMEOUT		1
221 #define KFD_IOC_WAIT_RESULT_FAIL		2
222 
223 #define KFD_SIGNAL_EVENT_LIMIT			4096
224 
225 /* For kfd_event_data.hw_exception_data.reset_type. */
226 #define KFD_HW_EXCEPTION_WHOLE_GPU_RESET	0
227 #define KFD_HW_EXCEPTION_PER_ENGINE_RESET	1
228 
229 /* For kfd_event_data.hw_exception_data.reset_cause. */
230 #define KFD_HW_EXCEPTION_GPU_HANG	0
231 #define KFD_HW_EXCEPTION_ECC		1
232 
233 /* For kfd_hsa_memory_exception_data.ErrorType */
234 #define KFD_MEM_ERR_NO_RAS		0
235 #define KFD_MEM_ERR_SRAM_ECC		1
236 #define KFD_MEM_ERR_POISON_CONSUMED	2
237 #define KFD_MEM_ERR_GPU_HANG		3
238 
239 struct kfd_ioctl_create_event_args {
240 	__u64 event_page_offset;	/* from KFD */
241 	__u32 event_trigger_data;	/* from KFD - signal events only */
242 	__u32 event_type;		/* to KFD */
243 	__u32 auto_reset;		/* to KFD */
244 	__u32 node_id;		/* to KFD - only valid for certain
245 							event types */
246 	__u32 event_id;		/* from KFD */
247 	__u32 event_slot_index;	/* from KFD */
248 };
249 
250 struct kfd_ioctl_destroy_event_args {
251 	__u32 event_id;		/* to KFD */
252 	__u32 pad;
253 };
254 
255 struct kfd_ioctl_set_event_args {
256 	__u32 event_id;		/* to KFD */
257 	__u32 pad;
258 };
259 
260 struct kfd_ioctl_reset_event_args {
261 	__u32 event_id;		/* to KFD */
262 	__u32 pad;
263 };
264 
265 struct kfd_memory_exception_failure {
266 	__u32 NotPresent;	/* Page not present or supervisor privilege */
267 	__u32 ReadOnly;	/* Write access to a read-only page */
268 	__u32 NoExecute;	/* Execute access to a page marked NX */
269 	__u32 imprecise;	/* Can't determine the	exact fault address */
270 };
271 
272 /* memory exception data */
273 struct kfd_hsa_memory_exception_data {
274 	struct kfd_memory_exception_failure failure;
275 	__u64 va;
276 	__u32 gpu_id;
277 	__u32 ErrorType; /* 0 = no RAS error,
278 			  * 1 = ECC_SRAM,
279 			  * 2 = Link_SYNFLOOD (poison),
280 			  * 3 = GPU hang (not attributable to a specific cause),
281 			  * other values reserved
282 			  */
283 };
284 
285 /* hw exception data */
286 struct kfd_hsa_hw_exception_data {
287 	__u32 reset_type;
288 	__u32 reset_cause;
289 	__u32 memory_lost;
290 	__u32 gpu_id;
291 };
292 
293 /* Event data */
294 struct kfd_event_data {
295 	union {
296 		struct kfd_hsa_memory_exception_data memory_exception_data;
297 		struct kfd_hsa_hw_exception_data hw_exception_data;
298 	};				/* From KFD */
299 	__u64 kfd_event_data_ext;	/* pointer to an extension structure
300 					   for future exception types */
301 	__u32 event_id;		/* to KFD */
302 	__u32 pad;
303 };
304 
305 struct kfd_ioctl_wait_events_args {
306 	__u64 events_ptr;		/* pointed to struct
307 					   kfd_event_data array, to KFD */
308 	__u32 num_events;		/* to KFD */
309 	__u32 wait_for_all;		/* to KFD */
310 	__u32 timeout;		/* to KFD */
311 	__u32 wait_result;		/* from KFD */
312 };
313 
314 struct kfd_ioctl_set_scratch_backing_va_args {
315 	__u64 va_addr;	/* to KFD */
316 	__u32 gpu_id;	/* to KFD */
317 	__u32 pad;
318 };
319 
320 struct kfd_ioctl_get_tile_config_args {
321 	/* to KFD: pointer to tile array */
322 	__u64 tile_config_ptr;
323 	/* to KFD: pointer to macro tile array */
324 	__u64 macro_tile_config_ptr;
325 	/* to KFD: array size allocated by user mode
326 	 * from KFD: array size filled by kernel
327 	 */
328 	__u32 num_tile_configs;
329 	/* to KFD: array size allocated by user mode
330 	 * from KFD: array size filled by kernel
331 	 */
332 	__u32 num_macro_tile_configs;
333 
334 	__u32 gpu_id;		/* to KFD */
335 	__u32 gb_addr_config;	/* from KFD */
336 	__u32 num_banks;		/* from KFD */
337 	__u32 num_ranks;		/* from KFD */
338 	/* struct size can be extended later if needed
339 	 * without breaking ABI compatibility
340 	 */
341 };
342 
343 struct kfd_ioctl_set_trap_handler_args {
344 	__u64 tba_addr;		/* to KFD */
345 	__u64 tma_addr;		/* to KFD */
346 	__u32 gpu_id;		/* to KFD */
347 	__u32 pad;
348 };
349 
350 struct kfd_ioctl_acquire_vm_args {
351 	__u32 drm_fd;	/* to KFD */
352 	__u32 gpu_id;	/* to KFD */
353 };
354 
355 /* Allocation flags: memory types */
356 #define KFD_IOC_ALLOC_MEM_FLAGS_VRAM		(1 << 0)
357 #define KFD_IOC_ALLOC_MEM_FLAGS_GTT		(1 << 1)
358 #define KFD_IOC_ALLOC_MEM_FLAGS_USERPTR		(1 << 2)
359 #define KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL	(1 << 3)
360 #define KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP	(1 << 4)
361 /* Allocation flags: attributes/access options */
362 #define KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE	(1 << 31)
363 #define KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE	(1 << 30)
364 #define KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC		(1 << 29)
365 #define KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE	(1 << 28)
366 #define KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM	(1 << 27)
367 #define KFD_IOC_ALLOC_MEM_FLAGS_COHERENT	(1 << 26)
368 #define KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED	(1 << 25)
369 
370 /* Allocate memory for later SVM (shared virtual memory) mapping.
371  *
372  * @va_addr:     virtual address of the memory to be allocated
373  *               all later mappings on all GPUs will use this address
374  * @size:        size in bytes
375  * @handle:      buffer handle returned to user mode, used to refer to
376  *               this allocation for mapping, unmapping and freeing
377  * @mmap_offset: for CPU-mapping the allocation by mmapping a render node
378  *               for userptrs this is overloaded to specify the CPU address
379  * @gpu_id:      device identifier
380  * @flags:       memory type and attributes. See KFD_IOC_ALLOC_MEM_FLAGS above
381  */
382 struct kfd_ioctl_alloc_memory_of_gpu_args {
383 	__u64 va_addr;		/* to KFD */
384 	__u64 size;		/* to KFD */
385 	__u64 handle;		/* from KFD */
386 	__u64 mmap_offset;	/* to KFD (userptr), from KFD (mmap offset) */
387 	__u32 gpu_id;		/* to KFD */
388 	__u32 flags;
389 };
390 
391 /* Free memory allocated with kfd_ioctl_alloc_memory_of_gpu
392  *
393  * @handle: memory handle returned by alloc
394  */
395 struct kfd_ioctl_free_memory_of_gpu_args {
396 	__u64 handle;		/* to KFD */
397 };
398 
399 /* Map memory to one or more GPUs
400  *
401  * @handle:                memory handle returned by alloc
402  * @device_ids_array_ptr:  array of gpu_ids (__u32 per device)
403  * @n_devices:             number of devices in the array
404  * @n_success:             number of devices mapped successfully
405  *
406  * @n_success returns information to the caller how many devices from
407  * the start of the array have mapped the buffer successfully. It can
408  * be passed into a subsequent retry call to skip those devices. For
409  * the first call the caller should initialize it to 0.
410  *
411  * If the ioctl completes with return code 0 (success), n_success ==
412  * n_devices.
413  */
414 struct kfd_ioctl_map_memory_to_gpu_args {
415 	__u64 handle;			/* to KFD */
416 	__u64 device_ids_array_ptr;	/* to KFD */
417 	__u32 n_devices;		/* to KFD */
418 	__u32 n_success;		/* to/from KFD */
419 };
420 
421 /* Unmap memory from one or more GPUs
422  *
423  * same arguments as for mapping
424  */
425 struct kfd_ioctl_unmap_memory_from_gpu_args {
426 	__u64 handle;			/* to KFD */
427 	__u64 device_ids_array_ptr;	/* to KFD */
428 	__u32 n_devices;		/* to KFD */
429 	__u32 n_success;		/* to/from KFD */
430 };
431 
432 /* Allocate GWS for specific queue
433  *
434  * @queue_id:    queue's id that GWS is allocated for
435  * @num_gws:     how many GWS to allocate
436  * @first_gws:   index of the first GWS allocated.
437  *               only support contiguous GWS allocation
438  */
439 struct kfd_ioctl_alloc_queue_gws_args {
440 	__u32 queue_id;		/* to KFD */
441 	__u32 num_gws;		/* to KFD */
442 	__u32 first_gws;	/* from KFD */
443 	__u32 pad;
444 };
445 
446 struct kfd_ioctl_get_dmabuf_info_args {
447 	__u64 size;		/* from KFD */
448 	__u64 metadata_ptr;	/* to KFD */
449 	__u32 metadata_size;	/* to KFD (space allocated by user)
450 				 * from KFD (actual metadata size)
451 				 */
452 	__u32 gpu_id;	/* from KFD */
453 	__u32 flags;		/* from KFD (KFD_IOC_ALLOC_MEM_FLAGS) */
454 	__u32 dmabuf_fd;	/* to KFD */
455 };
456 
457 struct kfd_ioctl_import_dmabuf_args {
458 	__u64 va_addr;	/* to KFD */
459 	__u64 handle;	/* from KFD */
460 	__u32 gpu_id;	/* to KFD */
461 	__u32 dmabuf_fd;	/* to KFD */
462 };
463 
464 /*
465  * KFD SMI(System Management Interface) events
466  */
467 enum kfd_smi_event {
468 	KFD_SMI_EVENT_NONE = 0, /* not used */
469 	KFD_SMI_EVENT_VMFAULT = 1, /* event start counting at 1 */
470 	KFD_SMI_EVENT_THERMAL_THROTTLE = 2,
471 	KFD_SMI_EVENT_GPU_PRE_RESET = 3,
472 	KFD_SMI_EVENT_GPU_POST_RESET = 4,
473 	KFD_SMI_EVENT_MIGRATE_START = 5,
474 	KFD_SMI_EVENT_MIGRATE_END = 6,
475 	KFD_SMI_EVENT_PAGE_FAULT_START = 7,
476 	KFD_SMI_EVENT_PAGE_FAULT_END = 8,
477 	KFD_SMI_EVENT_QUEUE_EVICTION = 9,
478 	KFD_SMI_EVENT_QUEUE_RESTORE = 10,
479 	KFD_SMI_EVENT_UNMAP_FROM_GPU = 11,
480 
481 	/*
482 	 * max event number, as a flag bit to get events from all processes,
483 	 * this requires super user permission, otherwise will not be able to
484 	 * receive event from any process. Without this flag to receive events
485 	 * from same process.
486 	 */
487 	KFD_SMI_EVENT_ALL_PROCESS = 64
488 };
489 
490 enum KFD_MIGRATE_TRIGGERS {
491 	KFD_MIGRATE_TRIGGER_PREFETCH,
492 	KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
493 	KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU,
494 	KFD_MIGRATE_TRIGGER_TTM_EVICTION
495 };
496 
497 enum KFD_QUEUE_EVICTION_TRIGGERS {
498 	KFD_QUEUE_EVICTION_TRIGGER_SVM,
499 	KFD_QUEUE_EVICTION_TRIGGER_USERPTR,
500 	KFD_QUEUE_EVICTION_TRIGGER_TTM,
501 	KFD_QUEUE_EVICTION_TRIGGER_SUSPEND,
502 	KFD_QUEUE_EVICTION_CRIU_CHECKPOINT,
503 	KFD_QUEUE_EVICTION_CRIU_RESTORE
504 };
505 
506 enum KFD_SVM_UNMAP_TRIGGERS {
507 	KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY,
508 	KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE,
509 	KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU
510 };
511 
512 #define KFD_SMI_EVENT_MASK_FROM_INDEX(i) (1ULL << ((i) - 1))
513 #define KFD_SMI_EVENT_MSG_SIZE	96
514 
515 struct kfd_ioctl_smi_events_args {
516 	__u32 gpuid;	/* to KFD */
517 	__u32 anon_fd;	/* from KFD */
518 };
519 
520 /**************************************************************************************************
521  * CRIU IOCTLs (Checkpoint Restore In Userspace)
522  *
523  * When checkpointing a process, the userspace application will perform:
524  * 1. PROCESS_INFO op to determine current process information. This pauses execution and evicts
525  *    all the queues.
526  * 2. CHECKPOINT op to checkpoint process contents (BOs, queues, events, svm-ranges)
527  * 3. UNPAUSE op to un-evict all the queues
528  *
529  * When restoring a process, the CRIU userspace application will perform:
530  *
531  * 1. RESTORE op to restore process contents
532  * 2. RESUME op to start the process
533  *
534  * Note: Queues are forced into an evicted state after a successful PROCESS_INFO. User
535  * application needs to perform an UNPAUSE operation after calling PROCESS_INFO.
536  */
537 
538 enum kfd_criu_op {
539 	KFD_CRIU_OP_PROCESS_INFO,
540 	KFD_CRIU_OP_CHECKPOINT,
541 	KFD_CRIU_OP_UNPAUSE,
542 	KFD_CRIU_OP_RESTORE,
543 	KFD_CRIU_OP_RESUME,
544 };
545 
546 /**
547  * kfd_ioctl_criu_args - Arguments perform CRIU operation
548  * @devices:		[in/out] User pointer to memory location for devices information.
549  * 			This is an array of type kfd_criu_device_bucket.
550  * @bos:		[in/out] User pointer to memory location for BOs information
551  * 			This is an array of type kfd_criu_bo_bucket.
552  * @priv_data:		[in/out] User pointer to memory location for private data
553  * @priv_data_size:	[in/out] Size of priv_data in bytes
554  * @num_devices:	[in/out] Number of GPUs used by process. Size of @devices array.
555  * @num_bos		[in/out] Number of BOs used by process. Size of @bos array.
556  * @num_objects:	[in/out] Number of objects used by process. Objects are opaque to
557  *				 user application.
558  * @pid:		[in/out] PID of the process being checkpointed
559  * @op			[in] Type of operation (kfd_criu_op)
560  *
561  * Return: 0 on success, -errno on failure
562  */
563 struct kfd_ioctl_criu_args {
564 	__u64 devices;		/* Used during ops: CHECKPOINT, RESTORE */
565 	__u64 bos;		/* Used during ops: CHECKPOINT, RESTORE */
566 	__u64 priv_data;	/* Used during ops: CHECKPOINT, RESTORE */
567 	__u64 priv_data_size;	/* Used during ops: PROCESS_INFO, RESTORE */
568 	__u32 num_devices;	/* Used during ops: PROCESS_INFO, RESTORE */
569 	__u32 num_bos;		/* Used during ops: PROCESS_INFO, RESTORE */
570 	__u32 num_objects;	/* Used during ops: PROCESS_INFO, RESTORE */
571 	__u32 pid;		/* Used during ops: PROCESS_INFO, RESUME */
572 	__u32 op;
573 };
574 
575 struct kfd_criu_device_bucket {
576 	__u32 user_gpu_id;
577 	__u32 actual_gpu_id;
578 	__u32 drm_fd;
579 	__u32 pad;
580 };
581 
582 struct kfd_criu_bo_bucket {
583 	__u64 addr;
584 	__u64 size;
585 	__u64 offset;
586 	__u64 restored_offset;    /* During restore, updated offset for BO */
587 	__u32 gpu_id;             /* This is the user_gpu_id */
588 	__u32 alloc_flags;
589 	__u32 dmabuf_fd;
590 	__u32 pad;
591 };
592 
593 /* CRIU IOCTLs - END */
594 /**************************************************************************************************/
595 
596 /* Register offset inside the remapped mmio page
597  */
598 enum kfd_mmio_remap {
599 	KFD_MMIO_REMAP_HDP_MEM_FLUSH_CNTL = 0,
600 	KFD_MMIO_REMAP_HDP_REG_FLUSH_CNTL = 4,
601 };
602 
603 /* Guarantee host access to memory */
604 #define KFD_IOCTL_SVM_FLAG_HOST_ACCESS 0x00000001
605 /* Fine grained coherency between all devices with access */
606 #define KFD_IOCTL_SVM_FLAG_COHERENT    0x00000002
607 /* Use any GPU in same hive as preferred device */
608 #define KFD_IOCTL_SVM_FLAG_HIVE_LOCAL  0x00000004
609 /* GPUs only read, allows replication */
610 #define KFD_IOCTL_SVM_FLAG_GPU_RO      0x00000008
611 /* Allow execution on GPU */
612 #define KFD_IOCTL_SVM_FLAG_GPU_EXEC    0x00000010
613 /* GPUs mostly read, may allow similar optimizations as RO, but writes fault */
614 #define KFD_IOCTL_SVM_FLAG_GPU_READ_MOSTLY     0x00000020
615 
616 /**
617  * kfd_ioctl_svm_op - SVM ioctl operations
618  *
619  * @KFD_IOCTL_SVM_OP_SET_ATTR: Modify one or more attributes
620  * @KFD_IOCTL_SVM_OP_GET_ATTR: Query one or more attributes
621  */
622 enum kfd_ioctl_svm_op {
623 	KFD_IOCTL_SVM_OP_SET_ATTR,
624 	KFD_IOCTL_SVM_OP_GET_ATTR
625 };
626 
627 /** kfd_ioctl_svm_location - Enum for preferred and prefetch locations
628  *
629  * GPU IDs are used to specify GPUs as preferred and prefetch locations.
630  * Below definitions are used for system memory or for leaving the preferred
631  * location unspecified.
632  */
633 enum kfd_ioctl_svm_location {
634 	KFD_IOCTL_SVM_LOCATION_SYSMEM = 0,
635 	KFD_IOCTL_SVM_LOCATION_UNDEFINED = 0xffffffff
636 };
637 
638 /**
639  * kfd_ioctl_svm_attr_type - SVM attribute types
640  *
641  * @KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: gpuid of the preferred location, 0 for
642  *                                    system memory
643  * @KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: gpuid of the prefetch location, 0 for
644  *                                   system memory. Setting this triggers an
645  *                                   immediate prefetch (migration).
646  * @KFD_IOCTL_SVM_ATTR_ACCESS:
647  * @KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
648  * @KFD_IOCTL_SVM_ATTR_NO_ACCESS: specify memory access for the gpuid given
649  *                                by the attribute value
650  * @KFD_IOCTL_SVM_ATTR_SET_FLAGS: bitmask of flags to set (see
651  *                                KFD_IOCTL_SVM_FLAG_...)
652  * @KFD_IOCTL_SVM_ATTR_CLR_FLAGS: bitmask of flags to clear
653  * @KFD_IOCTL_SVM_ATTR_GRANULARITY: migration granularity
654  *                                  (log2 num pages)
655  */
656 enum kfd_ioctl_svm_attr_type {
657 	KFD_IOCTL_SVM_ATTR_PREFERRED_LOC,
658 	KFD_IOCTL_SVM_ATTR_PREFETCH_LOC,
659 	KFD_IOCTL_SVM_ATTR_ACCESS,
660 	KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE,
661 	KFD_IOCTL_SVM_ATTR_NO_ACCESS,
662 	KFD_IOCTL_SVM_ATTR_SET_FLAGS,
663 	KFD_IOCTL_SVM_ATTR_CLR_FLAGS,
664 	KFD_IOCTL_SVM_ATTR_GRANULARITY
665 };
666 
667 /**
668  * kfd_ioctl_svm_attribute - Attributes as pairs of type and value
669  *
670  * The meaning of the @value depends on the attribute type.
671  *
672  * @type: attribute type (see enum @kfd_ioctl_svm_attr_type)
673  * @value: attribute value
674  */
675 struct kfd_ioctl_svm_attribute {
676 	__u32 type;
677 	__u32 value;
678 };
679 
680 /**
681  * kfd_ioctl_svm_args - Arguments for SVM ioctl
682  *
683  * @op specifies the operation to perform (see enum
684  * @kfd_ioctl_svm_op).  @start_addr and @size are common for all
685  * operations.
686  *
687  * A variable number of attributes can be given in @attrs.
688  * @nattr specifies the number of attributes. New attributes can be
689  * added in the future without breaking the ABI. If unknown attributes
690  * are given, the function returns -EINVAL.
691  *
692  * @KFD_IOCTL_SVM_OP_SET_ATTR sets attributes for a virtual address
693  * range. It may overlap existing virtual address ranges. If it does,
694  * the existing ranges will be split such that the attribute changes
695  * only apply to the specified address range.
696  *
697  * @KFD_IOCTL_SVM_OP_GET_ATTR returns the intersection of attributes
698  * over all memory in the given range and returns the result as the
699  * attribute value. If different pages have different preferred or
700  * prefetch locations, 0xffffffff will be returned for
701  * @KFD_IOCTL_SVM_ATTR_PREFERRED_LOC or
702  * @KFD_IOCTL_SVM_ATTR_PREFETCH_LOC resepctively. For
703  * @KFD_IOCTL_SVM_ATTR_SET_FLAGS, flags of all pages will be
704  * aggregated by bitwise AND. That means, a flag will be set in the
705  * output, if that flag is set for all pages in the range. For
706  * @KFD_IOCTL_SVM_ATTR_CLR_FLAGS, flags of all pages will be
707  * aggregated by bitwise NOR. That means, a flag will be set in the
708  * output, if that flag is clear for all pages in the range.
709  * The minimum migration granularity throughout the range will be
710  * returned for @KFD_IOCTL_SVM_ATTR_GRANULARITY.
711  *
712  * Querying of accessibility attributes works by initializing the
713  * attribute type to @KFD_IOCTL_SVM_ATTR_ACCESS and the value to the
714  * GPUID being queried. Multiple attributes can be given to allow
715  * querying multiple GPUIDs. The ioctl function overwrites the
716  * attribute type to indicate the access for the specified GPU.
717  */
718 struct kfd_ioctl_svm_args {
719 	__u64 start_addr;
720 	__u64 size;
721 	__u32 op;
722 	__u32 nattr;
723 	/* Variable length array of attributes */
724 	struct kfd_ioctl_svm_attribute attrs[];
725 };
726 
727 /**
728  * kfd_ioctl_set_xnack_mode_args - Arguments for set_xnack_mode
729  *
730  * @xnack_enabled:       [in/out] Whether to enable XNACK mode for this process
731  *
732  * @xnack_enabled indicates whether recoverable page faults should be
733  * enabled for the current process. 0 means disabled, positive means
734  * enabled, negative means leave unchanged. If enabled, virtual address
735  * translations on GFXv9 and later AMD GPUs can return XNACK and retry
736  * the access until a valid PTE is available. This is used to implement
737  * device page faults.
738  *
739  * On output, @xnack_enabled returns the (new) current mode (0 or
740  * positive). Therefore, a negative input value can be used to query
741  * the current mode without changing it.
742  *
743  * The XNACK mode fundamentally changes the way SVM managed memory works
744  * in the driver, with subtle effects on application performance and
745  * functionality.
746  *
747  * Enabling XNACK mode requires shader programs to be compiled
748  * differently. Furthermore, not all GPUs support changing the mode
749  * per-process. Therefore changing the mode is only allowed while no
750  * user mode queues exist in the process. This ensure that no shader
751  * code is running that may be compiled for the wrong mode. And GPUs
752  * that cannot change to the requested mode will prevent the XNACK
753  * mode from occurring. All GPUs used by the process must be in the
754  * same XNACK mode.
755  *
756  * GFXv8 or older GPUs do not support 48 bit virtual addresses or SVM.
757  * Therefore those GPUs are not considered for the XNACK mode switch.
758  *
759  * Return: 0 on success, -errno on failure
760  */
761 struct kfd_ioctl_set_xnack_mode_args {
762 	__s32 xnack_enabled;
763 };
764 
765 #define AMDKFD_IOCTL_BASE 'K'
766 #define AMDKFD_IO(nr)			_IO(AMDKFD_IOCTL_BASE, nr)
767 #define AMDKFD_IOR(nr, type)		_IOR(AMDKFD_IOCTL_BASE, nr, type)
768 #define AMDKFD_IOW(nr, type)		_IOW(AMDKFD_IOCTL_BASE, nr, type)
769 #define AMDKFD_IOWR(nr, type)		_IOWR(AMDKFD_IOCTL_BASE, nr, type)
770 
771 #define AMDKFD_IOC_GET_VERSION			\
772 		AMDKFD_IOR(0x01, struct kfd_ioctl_get_version_args)
773 
774 #define AMDKFD_IOC_CREATE_QUEUE			\
775 		AMDKFD_IOWR(0x02, struct kfd_ioctl_create_queue_args)
776 
777 #define AMDKFD_IOC_DESTROY_QUEUE		\
778 		AMDKFD_IOWR(0x03, struct kfd_ioctl_destroy_queue_args)
779 
780 #define AMDKFD_IOC_SET_MEMORY_POLICY		\
781 		AMDKFD_IOW(0x04, struct kfd_ioctl_set_memory_policy_args)
782 
783 #define AMDKFD_IOC_GET_CLOCK_COUNTERS		\
784 		AMDKFD_IOWR(0x05, struct kfd_ioctl_get_clock_counters_args)
785 
786 #define AMDKFD_IOC_GET_PROCESS_APERTURES	\
787 		AMDKFD_IOR(0x06, struct kfd_ioctl_get_process_apertures_args)
788 
789 #define AMDKFD_IOC_UPDATE_QUEUE			\
790 		AMDKFD_IOW(0x07, struct kfd_ioctl_update_queue_args)
791 
792 #define AMDKFD_IOC_CREATE_EVENT			\
793 		AMDKFD_IOWR(0x08, struct kfd_ioctl_create_event_args)
794 
795 #define AMDKFD_IOC_DESTROY_EVENT		\
796 		AMDKFD_IOW(0x09, struct kfd_ioctl_destroy_event_args)
797 
798 #define AMDKFD_IOC_SET_EVENT			\
799 		AMDKFD_IOW(0x0A, struct kfd_ioctl_set_event_args)
800 
801 #define AMDKFD_IOC_RESET_EVENT			\
802 		AMDKFD_IOW(0x0B, struct kfd_ioctl_reset_event_args)
803 
804 #define AMDKFD_IOC_WAIT_EVENTS			\
805 		AMDKFD_IOWR(0x0C, struct kfd_ioctl_wait_events_args)
806 
807 #define AMDKFD_IOC_DBG_REGISTER_DEPRECATED	\
808 		AMDKFD_IOW(0x0D, struct kfd_ioctl_dbg_register_args)
809 
810 #define AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED	\
811 		AMDKFD_IOW(0x0E, struct kfd_ioctl_dbg_unregister_args)
812 
813 #define AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED	\
814 		AMDKFD_IOW(0x0F, struct kfd_ioctl_dbg_address_watch_args)
815 
816 #define AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED	\
817 		AMDKFD_IOW(0x10, struct kfd_ioctl_dbg_wave_control_args)
818 
819 #define AMDKFD_IOC_SET_SCRATCH_BACKING_VA	\
820 		AMDKFD_IOWR(0x11, struct kfd_ioctl_set_scratch_backing_va_args)
821 
822 #define AMDKFD_IOC_GET_TILE_CONFIG                                      \
823 		AMDKFD_IOWR(0x12, struct kfd_ioctl_get_tile_config_args)
824 
825 #define AMDKFD_IOC_SET_TRAP_HANDLER		\
826 		AMDKFD_IOW(0x13, struct kfd_ioctl_set_trap_handler_args)
827 
828 #define AMDKFD_IOC_GET_PROCESS_APERTURES_NEW	\
829 		AMDKFD_IOWR(0x14,		\
830 			struct kfd_ioctl_get_process_apertures_new_args)
831 
832 #define AMDKFD_IOC_ACQUIRE_VM			\
833 		AMDKFD_IOW(0x15, struct kfd_ioctl_acquire_vm_args)
834 
835 #define AMDKFD_IOC_ALLOC_MEMORY_OF_GPU		\
836 		AMDKFD_IOWR(0x16, struct kfd_ioctl_alloc_memory_of_gpu_args)
837 
838 #define AMDKFD_IOC_FREE_MEMORY_OF_GPU		\
839 		AMDKFD_IOW(0x17, struct kfd_ioctl_free_memory_of_gpu_args)
840 
841 #define AMDKFD_IOC_MAP_MEMORY_TO_GPU		\
842 		AMDKFD_IOWR(0x18, struct kfd_ioctl_map_memory_to_gpu_args)
843 
844 #define AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU	\
845 		AMDKFD_IOWR(0x19, struct kfd_ioctl_unmap_memory_from_gpu_args)
846 
847 #define AMDKFD_IOC_SET_CU_MASK		\
848 		AMDKFD_IOW(0x1A, struct kfd_ioctl_set_cu_mask_args)
849 
850 #define AMDKFD_IOC_GET_QUEUE_WAVE_STATE		\
851 		AMDKFD_IOWR(0x1B, struct kfd_ioctl_get_queue_wave_state_args)
852 
853 #define AMDKFD_IOC_GET_DMABUF_INFO		\
854 		AMDKFD_IOWR(0x1C, struct kfd_ioctl_get_dmabuf_info_args)
855 
856 #define AMDKFD_IOC_IMPORT_DMABUF		\
857 		AMDKFD_IOWR(0x1D, struct kfd_ioctl_import_dmabuf_args)
858 
859 #define AMDKFD_IOC_ALLOC_QUEUE_GWS		\
860 		AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
861 
862 #define AMDKFD_IOC_SMI_EVENTS			\
863 		AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
864 
865 #define AMDKFD_IOC_SVM	AMDKFD_IOWR(0x20, struct kfd_ioctl_svm_args)
866 
867 #define AMDKFD_IOC_SET_XNACK_MODE		\
868 		AMDKFD_IOWR(0x21, struct kfd_ioctl_set_xnack_mode_args)
869 
870 #define AMDKFD_IOC_CRIU_OP			\
871 		AMDKFD_IOWR(0x22, struct kfd_ioctl_criu_args)
872 
873 #define AMDKFD_IOC_AVAILABLE_MEMORY		\
874 		AMDKFD_IOWR(0x23, struct kfd_ioctl_get_available_memory_args)
875 
876 #define AMDKFD_COMMAND_START		0x01
877 #define AMDKFD_COMMAND_END		0x24
878 
879 #endif
880