xref: /linux-6.15/include/uapi/linux/kfd_ioctl.h (revision 00a62703)
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 #define KFD_IOCTL_MAJOR_VERSION 1
30 #define KFD_IOCTL_MINOR_VERSION 1
31 
32 struct kfd_ioctl_get_version_args {
33 	__u32 major_version;	/* from KFD */
34 	__u32 minor_version;	/* from KFD */
35 };
36 
37 /* For kfd_ioctl_create_queue_args.queue_type. */
38 #define KFD_IOC_QUEUE_TYPE_COMPUTE	0
39 #define KFD_IOC_QUEUE_TYPE_SDMA		1
40 #define KFD_IOC_QUEUE_TYPE_COMPUTE_AQL	2
41 
42 #define KFD_MAX_QUEUE_PERCENTAGE	100
43 #define KFD_MAX_QUEUE_PRIORITY		15
44 
45 struct kfd_ioctl_create_queue_args {
46 	__u64 ring_base_address;	/* to KFD */
47 	__u64 write_pointer_address;	/* from KFD */
48 	__u64 read_pointer_address;	/* from KFD */
49 	__u64 doorbell_offset;	/* from KFD */
50 
51 	__u32 ring_size;		/* to KFD */
52 	__u32 gpu_id;		/* to KFD */
53 	__u32 queue_type;		/* to KFD */
54 	__u32 queue_percentage;	/* to KFD */
55 	__u32 queue_priority;	/* to KFD */
56 	__u32 queue_id;		/* from KFD */
57 
58 	__u64 eop_buffer_address;	/* to KFD */
59 	__u64 eop_buffer_size;	/* to KFD */
60 	__u64 ctx_save_restore_address; /* to KFD */
61 	__u32 ctx_save_restore_size;	/* to KFD */
62 	__u32 ctl_stack_size;		/* to KFD */
63 };
64 
65 struct kfd_ioctl_destroy_queue_args {
66 	__u32 queue_id;		/* to KFD */
67 	__u32 pad;
68 };
69 
70 struct kfd_ioctl_update_queue_args {
71 	__u64 ring_base_address;	/* to KFD */
72 
73 	__u32 queue_id;		/* to KFD */
74 	__u32 ring_size;		/* to KFD */
75 	__u32 queue_percentage;	/* to KFD */
76 	__u32 queue_priority;	/* to KFD */
77 };
78 
79 /* For kfd_ioctl_set_memory_policy_args.default_policy and alternate_policy */
80 #define KFD_IOC_CACHE_POLICY_COHERENT 0
81 #define KFD_IOC_CACHE_POLICY_NONCOHERENT 1
82 
83 struct kfd_ioctl_set_memory_policy_args {
84 	__u64 alternate_aperture_base;	/* to KFD */
85 	__u64 alternate_aperture_size;	/* to KFD */
86 
87 	__u32 gpu_id;			/* to KFD */
88 	__u32 default_policy;		/* to KFD */
89 	__u32 alternate_policy;		/* to KFD */
90 	__u32 pad;
91 };
92 
93 /*
94  * All counters are monotonic. They are used for profiling of compute jobs.
95  * The profiling is done by userspace.
96  *
97  * In case of GPU reset, the counter should not be affected.
98  */
99 
100 struct kfd_ioctl_get_clock_counters_args {
101 	__u64 gpu_clock_counter;	/* from KFD */
102 	__u64 cpu_clock_counter;	/* from KFD */
103 	__u64 system_clock_counter;	/* from KFD */
104 	__u64 system_clock_freq;	/* from KFD */
105 
106 	__u32 gpu_id;		/* to KFD */
107 	__u32 pad;
108 };
109 
110 struct kfd_process_device_apertures {
111 	__u64 lds_base;		/* from KFD */
112 	__u64 lds_limit;		/* from KFD */
113 	__u64 scratch_base;		/* from KFD */
114 	__u64 scratch_limit;		/* from KFD */
115 	__u64 gpuvm_base;		/* from KFD */
116 	__u64 gpuvm_limit;		/* from KFD */
117 	__u32 gpu_id;		/* from KFD */
118 	__u32 pad;
119 };
120 
121 /*
122  * AMDKFD_IOC_GET_PROCESS_APERTURES is deprecated. Use
123  * AMDKFD_IOC_GET_PROCESS_APERTURES_NEW instead, which supports an
124  * unlimited number of GPUs.
125  */
126 #define NUM_OF_SUPPORTED_GPUS 7
127 struct kfd_ioctl_get_process_apertures_args {
128 	struct kfd_process_device_apertures
129 			process_apertures[NUM_OF_SUPPORTED_GPUS];/* from KFD */
130 
131 	/* from KFD, should be in the range [1 - NUM_OF_SUPPORTED_GPUS] */
132 	__u32 num_of_nodes;
133 	__u32 pad;
134 };
135 
136 struct kfd_ioctl_get_process_apertures_new_args {
137 	/* User allocated. Pointer to struct kfd_process_device_apertures
138 	 * filled in by Kernel
139 	 */
140 	__u64 kfd_process_device_apertures_ptr;
141 	/* to KFD - indicates amount of memory present in
142 	 *  kfd_process_device_apertures_ptr
143 	 * from KFD - Number of entries filled by KFD.
144 	 */
145 	__u32 num_of_nodes;
146 	__u32 pad;
147 };
148 
149 #define MAX_ALLOWED_NUM_POINTS    100
150 #define MAX_ALLOWED_AW_BUFF_SIZE 4096
151 #define MAX_ALLOWED_WAC_BUFF_SIZE  128
152 
153 struct kfd_ioctl_dbg_register_args {
154 	__u32 gpu_id;		/* to KFD */
155 	__u32 pad;
156 };
157 
158 struct kfd_ioctl_dbg_unregister_args {
159 	__u32 gpu_id;		/* to KFD */
160 	__u32 pad;
161 };
162 
163 struct kfd_ioctl_dbg_address_watch_args {
164 	__u64 content_ptr;		/* a pointer to the actual content */
165 	__u32 gpu_id;		/* to KFD */
166 	__u32 buf_size_in_bytes;	/*including gpu_id and buf_size */
167 };
168 
169 struct kfd_ioctl_dbg_wave_control_args {
170 	__u64 content_ptr;		/* a pointer to the actual content */
171 	__u32 gpu_id;		/* to KFD */
172 	__u32 buf_size_in_bytes;	/*including gpu_id and buf_size */
173 };
174 
175 /* Matching HSA_EVENTTYPE */
176 #define KFD_IOC_EVENT_SIGNAL			0
177 #define KFD_IOC_EVENT_NODECHANGE		1
178 #define KFD_IOC_EVENT_DEVICESTATECHANGE		2
179 #define KFD_IOC_EVENT_HW_EXCEPTION		3
180 #define KFD_IOC_EVENT_SYSTEM_EVENT		4
181 #define KFD_IOC_EVENT_DEBUG_EVENT		5
182 #define KFD_IOC_EVENT_PROFILE_EVENT		6
183 #define KFD_IOC_EVENT_QUEUE_EVENT		7
184 #define KFD_IOC_EVENT_MEMORY			8
185 
186 #define KFD_IOC_WAIT_RESULT_COMPLETE		0
187 #define KFD_IOC_WAIT_RESULT_TIMEOUT		1
188 #define KFD_IOC_WAIT_RESULT_FAIL		2
189 
190 #define KFD_SIGNAL_EVENT_LIMIT			4096
191 
192 struct kfd_ioctl_create_event_args {
193 	__u64 event_page_offset;	/* from KFD */
194 	__u32 event_trigger_data;	/* from KFD - signal events only */
195 	__u32 event_type;		/* to KFD */
196 	__u32 auto_reset;		/* to KFD */
197 	__u32 node_id;		/* to KFD - only valid for certain
198 							event types */
199 	__u32 event_id;		/* from KFD */
200 	__u32 event_slot_index;	/* from KFD */
201 };
202 
203 struct kfd_ioctl_destroy_event_args {
204 	__u32 event_id;		/* to KFD */
205 	__u32 pad;
206 };
207 
208 struct kfd_ioctl_set_event_args {
209 	__u32 event_id;		/* to KFD */
210 	__u32 pad;
211 };
212 
213 struct kfd_ioctl_reset_event_args {
214 	__u32 event_id;		/* to KFD */
215 	__u32 pad;
216 };
217 
218 struct kfd_memory_exception_failure {
219 	__u32 NotPresent;	/* Page not present or supervisor privilege */
220 	__u32 ReadOnly;	/* Write access to a read-only page */
221 	__u32 NoExecute;	/* Execute access to a page marked NX */
222 	__u32 pad;
223 };
224 
225 /* memory exception data*/
226 struct kfd_hsa_memory_exception_data {
227 	struct kfd_memory_exception_failure failure;
228 	__u64 va;
229 	__u32 gpu_id;
230 	__u32 pad;
231 };
232 
233 /* Event data*/
234 struct kfd_event_data {
235 	union {
236 		struct kfd_hsa_memory_exception_data memory_exception_data;
237 	};				/* From KFD */
238 	__u64 kfd_event_data_ext;	/* pointer to an extension structure
239 					   for future exception types */
240 	__u32 event_id;		/* to KFD */
241 	__u32 pad;
242 };
243 
244 struct kfd_ioctl_wait_events_args {
245 	__u64 events_ptr;		/* pointed to struct
246 					   kfd_event_data array, to KFD */
247 	__u32 num_events;		/* to KFD */
248 	__u32 wait_for_all;		/* to KFD */
249 	__u32 timeout;		/* to KFD */
250 	__u32 wait_result;		/* from KFD */
251 };
252 
253 struct kfd_ioctl_set_scratch_backing_va_args {
254 	__u64 va_addr;	/* to KFD */
255 	__u32 gpu_id;	/* to KFD */
256 	__u32 pad;
257 };
258 
259 struct kfd_ioctl_get_tile_config_args {
260 	/* to KFD: pointer to tile array */
261 	__u64 tile_config_ptr;
262 	/* to KFD: pointer to macro tile array */
263 	__u64 macro_tile_config_ptr;
264 	/* to KFD: array size allocated by user mode
265 	 * from KFD: array size filled by kernel
266 	 */
267 	__u32 num_tile_configs;
268 	/* to KFD: array size allocated by user mode
269 	 * from KFD: array size filled by kernel
270 	 */
271 	__u32 num_macro_tile_configs;
272 
273 	__u32 gpu_id;		/* to KFD */
274 	__u32 gb_addr_config;	/* from KFD */
275 	__u32 num_banks;		/* from KFD */
276 	__u32 num_ranks;		/* from KFD */
277 	/* struct size can be extended later if needed
278 	 * without breaking ABI compatibility
279 	 */
280 };
281 
282 struct kfd_ioctl_set_trap_handler_args {
283 	__u64 tba_addr;		/* to KFD */
284 	__u64 tma_addr;		/* to KFD */
285 	__u32 gpu_id;		/* to KFD */
286 	__u32 pad;
287 };
288 
289 struct kfd_ioctl_acquire_vm_args {
290 	__u32 drm_fd;	/* to KFD */
291 	__u32 gpu_id;	/* to KFD */
292 };
293 
294 /* Allocation flags: memory types */
295 #define KFD_IOC_ALLOC_MEM_FLAGS_VRAM		(1 << 0)
296 #define KFD_IOC_ALLOC_MEM_FLAGS_GTT		(1 << 1)
297 #define KFD_IOC_ALLOC_MEM_FLAGS_USERPTR		(1 << 2)
298 #define KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL	(1 << 3)
299 /* Allocation flags: attributes/access options */
300 #define KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE	(1 << 31)
301 #define KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE	(1 << 30)
302 #define KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC		(1 << 29)
303 #define KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE	(1 << 28)
304 #define KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM	(1 << 27)
305 #define KFD_IOC_ALLOC_MEM_FLAGS_COHERENT	(1 << 26)
306 
307 /* Allocate memory for later SVM (shared virtual memory) mapping.
308  *
309  * @va_addr:     virtual address of the memory to be allocated
310  *               all later mappings on all GPUs will use this address
311  * @size:        size in bytes
312  * @handle:      buffer handle returned to user mode, used to refer to
313  *               this allocation for mapping, unmapping and freeing
314  * @mmap_offset: for CPU-mapping the allocation by mmapping a render node
315  *               for userptrs this is overloaded to specify the CPU address
316  * @gpu_id:      device identifier
317  * @flags:       memory type and attributes. See KFD_IOC_ALLOC_MEM_FLAGS above
318  */
319 struct kfd_ioctl_alloc_memory_of_gpu_args {
320 	__u64 va_addr;		/* to KFD */
321 	__u64 size;		/* to KFD */
322 	__u64 handle;		/* from KFD */
323 	__u64 mmap_offset;	/* to KFD (userptr), from KFD (mmap offset) */
324 	__u32 gpu_id;		/* to KFD */
325 	__u32 flags;
326 };
327 
328 /* Free memory allocated with kfd_ioctl_alloc_memory_of_gpu
329  *
330  * @handle: memory handle returned by alloc
331  */
332 struct kfd_ioctl_free_memory_of_gpu_args {
333 	__u64 handle;		/* to KFD */
334 };
335 
336 /* Map memory to one or more GPUs
337  *
338  * @handle:                memory handle returned by alloc
339  * @device_ids_array_ptr:  array of gpu_ids (__u32 per device)
340  * @n_devices:             number of devices in the array
341  * @n_success:             number of devices mapped successfully
342  *
343  * @n_success returns information to the caller how many devices from
344  * the start of the array have mapped the buffer successfully. It can
345  * be passed into a subsequent retry call to skip those devices. For
346  * the first call the caller should initialize it to 0.
347  *
348  * If the ioctl completes with return code 0 (success), n_success ==
349  * n_devices.
350  */
351 struct kfd_ioctl_map_memory_to_gpu_args {
352 	__u64 handle;			/* to KFD */
353 	__u64 device_ids_array_ptr;	/* to KFD */
354 	__u32 n_devices;		/* to KFD */
355 	__u32 n_success;		/* to/from KFD */
356 };
357 
358 /* Unmap memory from one or more GPUs
359  *
360  * same arguments as for mapping
361  */
362 struct kfd_ioctl_unmap_memory_from_gpu_args {
363 	__u64 handle;			/* to KFD */
364 	__u64 device_ids_array_ptr;	/* to KFD */
365 	__u32 n_devices;		/* to KFD */
366 	__u32 n_success;		/* to/from KFD */
367 };
368 
369 #define AMDKFD_IOCTL_BASE 'K'
370 #define AMDKFD_IO(nr)			_IO(AMDKFD_IOCTL_BASE, nr)
371 #define AMDKFD_IOR(nr, type)		_IOR(AMDKFD_IOCTL_BASE, nr, type)
372 #define AMDKFD_IOW(nr, type)		_IOW(AMDKFD_IOCTL_BASE, nr, type)
373 #define AMDKFD_IOWR(nr, type)		_IOWR(AMDKFD_IOCTL_BASE, nr, type)
374 
375 #define AMDKFD_IOC_GET_VERSION			\
376 		AMDKFD_IOR(0x01, struct kfd_ioctl_get_version_args)
377 
378 #define AMDKFD_IOC_CREATE_QUEUE			\
379 		AMDKFD_IOWR(0x02, struct kfd_ioctl_create_queue_args)
380 
381 #define AMDKFD_IOC_DESTROY_QUEUE		\
382 		AMDKFD_IOWR(0x03, struct kfd_ioctl_destroy_queue_args)
383 
384 #define AMDKFD_IOC_SET_MEMORY_POLICY		\
385 		AMDKFD_IOW(0x04, struct kfd_ioctl_set_memory_policy_args)
386 
387 #define AMDKFD_IOC_GET_CLOCK_COUNTERS		\
388 		AMDKFD_IOWR(0x05, struct kfd_ioctl_get_clock_counters_args)
389 
390 #define AMDKFD_IOC_GET_PROCESS_APERTURES	\
391 		AMDKFD_IOR(0x06, struct kfd_ioctl_get_process_apertures_args)
392 
393 #define AMDKFD_IOC_UPDATE_QUEUE			\
394 		AMDKFD_IOW(0x07, struct kfd_ioctl_update_queue_args)
395 
396 #define AMDKFD_IOC_CREATE_EVENT			\
397 		AMDKFD_IOWR(0x08, struct kfd_ioctl_create_event_args)
398 
399 #define AMDKFD_IOC_DESTROY_EVENT		\
400 		AMDKFD_IOW(0x09, struct kfd_ioctl_destroy_event_args)
401 
402 #define AMDKFD_IOC_SET_EVENT			\
403 		AMDKFD_IOW(0x0A, struct kfd_ioctl_set_event_args)
404 
405 #define AMDKFD_IOC_RESET_EVENT			\
406 		AMDKFD_IOW(0x0B, struct kfd_ioctl_reset_event_args)
407 
408 #define AMDKFD_IOC_WAIT_EVENTS			\
409 		AMDKFD_IOWR(0x0C, struct kfd_ioctl_wait_events_args)
410 
411 #define AMDKFD_IOC_DBG_REGISTER			\
412 		AMDKFD_IOW(0x0D, struct kfd_ioctl_dbg_register_args)
413 
414 #define AMDKFD_IOC_DBG_UNREGISTER		\
415 		AMDKFD_IOW(0x0E, struct kfd_ioctl_dbg_unregister_args)
416 
417 #define AMDKFD_IOC_DBG_ADDRESS_WATCH		\
418 		AMDKFD_IOW(0x0F, struct kfd_ioctl_dbg_address_watch_args)
419 
420 #define AMDKFD_IOC_DBG_WAVE_CONTROL		\
421 		AMDKFD_IOW(0x10, struct kfd_ioctl_dbg_wave_control_args)
422 
423 #define AMDKFD_IOC_SET_SCRATCH_BACKING_VA	\
424 		AMDKFD_IOWR(0x11, struct kfd_ioctl_set_scratch_backing_va_args)
425 
426 #define AMDKFD_IOC_GET_TILE_CONFIG                                      \
427 		AMDKFD_IOWR(0x12, struct kfd_ioctl_get_tile_config_args)
428 
429 #define AMDKFD_IOC_SET_TRAP_HANDLER		\
430 		AMDKFD_IOW(0x13, struct kfd_ioctl_set_trap_handler_args)
431 
432 #define AMDKFD_IOC_GET_PROCESS_APERTURES_NEW	\
433 		AMDKFD_IOWR(0x14,		\
434 			struct kfd_ioctl_get_process_apertures_new_args)
435 
436 #define AMDKFD_IOC_ACQUIRE_VM			\
437 		AMDKFD_IOW(0x15, struct kfd_ioctl_acquire_vm_args)
438 
439 #define AMDKFD_IOC_ALLOC_MEMORY_OF_GPU		\
440 		AMDKFD_IOWR(0x16, struct kfd_ioctl_alloc_memory_of_gpu_args)
441 
442 #define AMDKFD_IOC_FREE_MEMORY_OF_GPU		\
443 		AMDKFD_IOW(0x17, struct kfd_ioctl_free_memory_of_gpu_args)
444 
445 #define AMDKFD_IOC_MAP_MEMORY_TO_GPU		\
446 		AMDKFD_IOWR(0x18, struct kfd_ioctl_map_memory_to_gpu_args)
447 
448 #define AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU	\
449 		AMDKFD_IOWR(0x19, struct kfd_ioctl_unmap_memory_from_gpu_args)
450 
451 #define AMDKFD_COMMAND_START		0x01
452 #define AMDKFD_COMMAND_END		0x1A
453 
454 #endif
455