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 struct kfd_ioctl_set_cu_mask_args {
80 	__u32 queue_id;		/* to KFD */
81 	__u32 num_cu_mask;		/* to KFD */
82 	__u64 cu_mask_ptr;		/* to KFD */
83 };
84 
85 /* For kfd_ioctl_set_memory_policy_args.default_policy and alternate_policy */
86 #define KFD_IOC_CACHE_POLICY_COHERENT 0
87 #define KFD_IOC_CACHE_POLICY_NONCOHERENT 1
88 
89 struct kfd_ioctl_set_memory_policy_args {
90 	__u64 alternate_aperture_base;	/* to KFD */
91 	__u64 alternate_aperture_size;	/* to KFD */
92 
93 	__u32 gpu_id;			/* to KFD */
94 	__u32 default_policy;		/* to KFD */
95 	__u32 alternate_policy;		/* to KFD */
96 	__u32 pad;
97 };
98 
99 /*
100  * All counters are monotonic. They are used for profiling of compute jobs.
101  * The profiling is done by userspace.
102  *
103  * In case of GPU reset, the counter should not be affected.
104  */
105 
106 struct kfd_ioctl_get_clock_counters_args {
107 	__u64 gpu_clock_counter;	/* from KFD */
108 	__u64 cpu_clock_counter;	/* from KFD */
109 	__u64 system_clock_counter;	/* from KFD */
110 	__u64 system_clock_freq;	/* from KFD */
111 
112 	__u32 gpu_id;		/* to KFD */
113 	__u32 pad;
114 };
115 
116 struct kfd_process_device_apertures {
117 	__u64 lds_base;		/* from KFD */
118 	__u64 lds_limit;		/* from KFD */
119 	__u64 scratch_base;		/* from KFD */
120 	__u64 scratch_limit;		/* from KFD */
121 	__u64 gpuvm_base;		/* from KFD */
122 	__u64 gpuvm_limit;		/* from KFD */
123 	__u32 gpu_id;		/* from KFD */
124 	__u32 pad;
125 };
126 
127 /*
128  * AMDKFD_IOC_GET_PROCESS_APERTURES is deprecated. Use
129  * AMDKFD_IOC_GET_PROCESS_APERTURES_NEW instead, which supports an
130  * unlimited number of GPUs.
131  */
132 #define NUM_OF_SUPPORTED_GPUS 7
133 struct kfd_ioctl_get_process_apertures_args {
134 	struct kfd_process_device_apertures
135 			process_apertures[NUM_OF_SUPPORTED_GPUS];/* from KFD */
136 
137 	/* from KFD, should be in the range [1 - NUM_OF_SUPPORTED_GPUS] */
138 	__u32 num_of_nodes;
139 	__u32 pad;
140 };
141 
142 struct kfd_ioctl_get_process_apertures_new_args {
143 	/* User allocated. Pointer to struct kfd_process_device_apertures
144 	 * filled in by Kernel
145 	 */
146 	__u64 kfd_process_device_apertures_ptr;
147 	/* to KFD - indicates amount of memory present in
148 	 *  kfd_process_device_apertures_ptr
149 	 * from KFD - Number of entries filled by KFD.
150 	 */
151 	__u32 num_of_nodes;
152 	__u32 pad;
153 };
154 
155 #define MAX_ALLOWED_NUM_POINTS    100
156 #define MAX_ALLOWED_AW_BUFF_SIZE 4096
157 #define MAX_ALLOWED_WAC_BUFF_SIZE  128
158 
159 struct kfd_ioctl_dbg_register_args {
160 	__u32 gpu_id;		/* to KFD */
161 	__u32 pad;
162 };
163 
164 struct kfd_ioctl_dbg_unregister_args {
165 	__u32 gpu_id;		/* to KFD */
166 	__u32 pad;
167 };
168 
169 struct kfd_ioctl_dbg_address_watch_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 struct kfd_ioctl_dbg_wave_control_args {
176 	__u64 content_ptr;		/* a pointer to the actual content */
177 	__u32 gpu_id;		/* to KFD */
178 	__u32 buf_size_in_bytes;	/*including gpu_id and buf_size */
179 };
180 
181 /* Matching HSA_EVENTTYPE */
182 #define KFD_IOC_EVENT_SIGNAL			0
183 #define KFD_IOC_EVENT_NODECHANGE		1
184 #define KFD_IOC_EVENT_DEVICESTATECHANGE		2
185 #define KFD_IOC_EVENT_HW_EXCEPTION		3
186 #define KFD_IOC_EVENT_SYSTEM_EVENT		4
187 #define KFD_IOC_EVENT_DEBUG_EVENT		5
188 #define KFD_IOC_EVENT_PROFILE_EVENT		6
189 #define KFD_IOC_EVENT_QUEUE_EVENT		7
190 #define KFD_IOC_EVENT_MEMORY			8
191 
192 #define KFD_IOC_WAIT_RESULT_COMPLETE		0
193 #define KFD_IOC_WAIT_RESULT_TIMEOUT		1
194 #define KFD_IOC_WAIT_RESULT_FAIL		2
195 
196 #define KFD_SIGNAL_EVENT_LIMIT			4096
197 
198 /* For kfd_event_data.hw_exception_data.reset_type. */
199 #define KFD_HW_EXCEPTION_WHOLE_GPU_RESET	0
200 #define KFD_HW_EXCEPTION_PER_ENGINE_RESET	1
201 
202 /* For kfd_event_data.hw_exception_data.reset_cause. */
203 #define KFD_HW_EXCEPTION_GPU_HANG	0
204 #define KFD_HW_EXCEPTION_ECC		1
205 
206 
207 struct kfd_ioctl_create_event_args {
208 	__u64 event_page_offset;	/* from KFD */
209 	__u32 event_trigger_data;	/* from KFD - signal events only */
210 	__u32 event_type;		/* to KFD */
211 	__u32 auto_reset;		/* to KFD */
212 	__u32 node_id;		/* to KFD - only valid for certain
213 							event types */
214 	__u32 event_id;		/* from KFD */
215 	__u32 event_slot_index;	/* from KFD */
216 };
217 
218 struct kfd_ioctl_destroy_event_args {
219 	__u32 event_id;		/* to KFD */
220 	__u32 pad;
221 };
222 
223 struct kfd_ioctl_set_event_args {
224 	__u32 event_id;		/* to KFD */
225 	__u32 pad;
226 };
227 
228 struct kfd_ioctl_reset_event_args {
229 	__u32 event_id;		/* to KFD */
230 	__u32 pad;
231 };
232 
233 struct kfd_memory_exception_failure {
234 	__u32 NotPresent;	/* Page not present or supervisor privilege */
235 	__u32 ReadOnly;	/* Write access to a read-only page */
236 	__u32 NoExecute;	/* Execute access to a page marked NX */
237 	__u32 imprecise;	/* Can't determine the	exact fault address */
238 };
239 
240 /* memory exception data*/
241 struct kfd_hsa_memory_exception_data {
242 	struct kfd_memory_exception_failure failure;
243 	__u64 va;
244 	__u32 gpu_id;
245 	__u32 pad;
246 };
247 
248 /* hw exception data */
249 struct kfd_hsa_hw_exception_data {
250 	__u32 reset_type;
251 	__u32 reset_cause;
252 	__u32 memory_lost;
253 	__u32 gpu_id;
254 };
255 
256 /* Event data */
257 struct kfd_event_data {
258 	union {
259 		struct kfd_hsa_memory_exception_data memory_exception_data;
260 		struct kfd_hsa_hw_exception_data hw_exception_data;
261 	};				/* From KFD */
262 	__u64 kfd_event_data_ext;	/* pointer to an extension structure
263 					   for future exception types */
264 	__u32 event_id;		/* to KFD */
265 	__u32 pad;
266 };
267 
268 struct kfd_ioctl_wait_events_args {
269 	__u64 events_ptr;		/* pointed to struct
270 					   kfd_event_data array, to KFD */
271 	__u32 num_events;		/* to KFD */
272 	__u32 wait_for_all;		/* to KFD */
273 	__u32 timeout;		/* to KFD */
274 	__u32 wait_result;		/* from KFD */
275 };
276 
277 struct kfd_ioctl_set_scratch_backing_va_args {
278 	__u64 va_addr;	/* to KFD */
279 	__u32 gpu_id;	/* to KFD */
280 	__u32 pad;
281 };
282 
283 struct kfd_ioctl_get_tile_config_args {
284 	/* to KFD: pointer to tile array */
285 	__u64 tile_config_ptr;
286 	/* to KFD: pointer to macro tile array */
287 	__u64 macro_tile_config_ptr;
288 	/* to KFD: array size allocated by user mode
289 	 * from KFD: array size filled by kernel
290 	 */
291 	__u32 num_tile_configs;
292 	/* to KFD: array size allocated by user mode
293 	 * from KFD: array size filled by kernel
294 	 */
295 	__u32 num_macro_tile_configs;
296 
297 	__u32 gpu_id;		/* to KFD */
298 	__u32 gb_addr_config;	/* from KFD */
299 	__u32 num_banks;		/* from KFD */
300 	__u32 num_ranks;		/* from KFD */
301 	/* struct size can be extended later if needed
302 	 * without breaking ABI compatibility
303 	 */
304 };
305 
306 struct kfd_ioctl_set_trap_handler_args {
307 	__u64 tba_addr;		/* to KFD */
308 	__u64 tma_addr;		/* to KFD */
309 	__u32 gpu_id;		/* to KFD */
310 	__u32 pad;
311 };
312 
313 struct kfd_ioctl_acquire_vm_args {
314 	__u32 drm_fd;	/* to KFD */
315 	__u32 gpu_id;	/* to KFD */
316 };
317 
318 /* Allocation flags: memory types */
319 #define KFD_IOC_ALLOC_MEM_FLAGS_VRAM		(1 << 0)
320 #define KFD_IOC_ALLOC_MEM_FLAGS_GTT		(1 << 1)
321 #define KFD_IOC_ALLOC_MEM_FLAGS_USERPTR		(1 << 2)
322 #define KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL	(1 << 3)
323 /* Allocation flags: attributes/access options */
324 #define KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE	(1 << 31)
325 #define KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE	(1 << 30)
326 #define KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC		(1 << 29)
327 #define KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE	(1 << 28)
328 #define KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM	(1 << 27)
329 #define KFD_IOC_ALLOC_MEM_FLAGS_COHERENT	(1 << 26)
330 
331 /* Allocate memory for later SVM (shared virtual memory) mapping.
332  *
333  * @va_addr:     virtual address of the memory to be allocated
334  *               all later mappings on all GPUs will use this address
335  * @size:        size in bytes
336  * @handle:      buffer handle returned to user mode, used to refer to
337  *               this allocation for mapping, unmapping and freeing
338  * @mmap_offset: for CPU-mapping the allocation by mmapping a render node
339  *               for userptrs this is overloaded to specify the CPU address
340  * @gpu_id:      device identifier
341  * @flags:       memory type and attributes. See KFD_IOC_ALLOC_MEM_FLAGS above
342  */
343 struct kfd_ioctl_alloc_memory_of_gpu_args {
344 	__u64 va_addr;		/* to KFD */
345 	__u64 size;		/* to KFD */
346 	__u64 handle;		/* from KFD */
347 	__u64 mmap_offset;	/* to KFD (userptr), from KFD (mmap offset) */
348 	__u32 gpu_id;		/* to KFD */
349 	__u32 flags;
350 };
351 
352 /* Free memory allocated with kfd_ioctl_alloc_memory_of_gpu
353  *
354  * @handle: memory handle returned by alloc
355  */
356 struct kfd_ioctl_free_memory_of_gpu_args {
357 	__u64 handle;		/* to KFD */
358 };
359 
360 /* Map memory to one or more GPUs
361  *
362  * @handle:                memory handle returned by alloc
363  * @device_ids_array_ptr:  array of gpu_ids (__u32 per device)
364  * @n_devices:             number of devices in the array
365  * @n_success:             number of devices mapped successfully
366  *
367  * @n_success returns information to the caller how many devices from
368  * the start of the array have mapped the buffer successfully. It can
369  * be passed into a subsequent retry call to skip those devices. For
370  * the first call the caller should initialize it to 0.
371  *
372  * If the ioctl completes with return code 0 (success), n_success ==
373  * n_devices.
374  */
375 struct kfd_ioctl_map_memory_to_gpu_args {
376 	__u64 handle;			/* to KFD */
377 	__u64 device_ids_array_ptr;	/* to KFD */
378 	__u32 n_devices;		/* to KFD */
379 	__u32 n_success;		/* to/from KFD */
380 };
381 
382 /* Unmap memory from one or more GPUs
383  *
384  * same arguments as for mapping
385  */
386 struct kfd_ioctl_unmap_memory_from_gpu_args {
387 	__u64 handle;			/* to KFD */
388 	__u64 device_ids_array_ptr;	/* to KFD */
389 	__u32 n_devices;		/* to KFD */
390 	__u32 n_success;		/* to/from KFD */
391 };
392 
393 #define AMDKFD_IOCTL_BASE 'K'
394 #define AMDKFD_IO(nr)			_IO(AMDKFD_IOCTL_BASE, nr)
395 #define AMDKFD_IOR(nr, type)		_IOR(AMDKFD_IOCTL_BASE, nr, type)
396 #define AMDKFD_IOW(nr, type)		_IOW(AMDKFD_IOCTL_BASE, nr, type)
397 #define AMDKFD_IOWR(nr, type)		_IOWR(AMDKFD_IOCTL_BASE, nr, type)
398 
399 #define AMDKFD_IOC_GET_VERSION			\
400 		AMDKFD_IOR(0x01, struct kfd_ioctl_get_version_args)
401 
402 #define AMDKFD_IOC_CREATE_QUEUE			\
403 		AMDKFD_IOWR(0x02, struct kfd_ioctl_create_queue_args)
404 
405 #define AMDKFD_IOC_DESTROY_QUEUE		\
406 		AMDKFD_IOWR(0x03, struct kfd_ioctl_destroy_queue_args)
407 
408 #define AMDKFD_IOC_SET_MEMORY_POLICY		\
409 		AMDKFD_IOW(0x04, struct kfd_ioctl_set_memory_policy_args)
410 
411 #define AMDKFD_IOC_GET_CLOCK_COUNTERS		\
412 		AMDKFD_IOWR(0x05, struct kfd_ioctl_get_clock_counters_args)
413 
414 #define AMDKFD_IOC_GET_PROCESS_APERTURES	\
415 		AMDKFD_IOR(0x06, struct kfd_ioctl_get_process_apertures_args)
416 
417 #define AMDKFD_IOC_UPDATE_QUEUE			\
418 		AMDKFD_IOW(0x07, struct kfd_ioctl_update_queue_args)
419 
420 #define AMDKFD_IOC_CREATE_EVENT			\
421 		AMDKFD_IOWR(0x08, struct kfd_ioctl_create_event_args)
422 
423 #define AMDKFD_IOC_DESTROY_EVENT		\
424 		AMDKFD_IOW(0x09, struct kfd_ioctl_destroy_event_args)
425 
426 #define AMDKFD_IOC_SET_EVENT			\
427 		AMDKFD_IOW(0x0A, struct kfd_ioctl_set_event_args)
428 
429 #define AMDKFD_IOC_RESET_EVENT			\
430 		AMDKFD_IOW(0x0B, struct kfd_ioctl_reset_event_args)
431 
432 #define AMDKFD_IOC_WAIT_EVENTS			\
433 		AMDKFD_IOWR(0x0C, struct kfd_ioctl_wait_events_args)
434 
435 #define AMDKFD_IOC_DBG_REGISTER			\
436 		AMDKFD_IOW(0x0D, struct kfd_ioctl_dbg_register_args)
437 
438 #define AMDKFD_IOC_DBG_UNREGISTER		\
439 		AMDKFD_IOW(0x0E, struct kfd_ioctl_dbg_unregister_args)
440 
441 #define AMDKFD_IOC_DBG_ADDRESS_WATCH		\
442 		AMDKFD_IOW(0x0F, struct kfd_ioctl_dbg_address_watch_args)
443 
444 #define AMDKFD_IOC_DBG_WAVE_CONTROL		\
445 		AMDKFD_IOW(0x10, struct kfd_ioctl_dbg_wave_control_args)
446 
447 #define AMDKFD_IOC_SET_SCRATCH_BACKING_VA	\
448 		AMDKFD_IOWR(0x11, struct kfd_ioctl_set_scratch_backing_va_args)
449 
450 #define AMDKFD_IOC_GET_TILE_CONFIG                                      \
451 		AMDKFD_IOWR(0x12, struct kfd_ioctl_get_tile_config_args)
452 
453 #define AMDKFD_IOC_SET_TRAP_HANDLER		\
454 		AMDKFD_IOW(0x13, struct kfd_ioctl_set_trap_handler_args)
455 
456 #define AMDKFD_IOC_GET_PROCESS_APERTURES_NEW	\
457 		AMDKFD_IOWR(0x14,		\
458 			struct kfd_ioctl_get_process_apertures_new_args)
459 
460 #define AMDKFD_IOC_ACQUIRE_VM			\
461 		AMDKFD_IOW(0x15, struct kfd_ioctl_acquire_vm_args)
462 
463 #define AMDKFD_IOC_ALLOC_MEMORY_OF_GPU		\
464 		AMDKFD_IOWR(0x16, struct kfd_ioctl_alloc_memory_of_gpu_args)
465 
466 #define AMDKFD_IOC_FREE_MEMORY_OF_GPU		\
467 		AMDKFD_IOW(0x17, struct kfd_ioctl_free_memory_of_gpu_args)
468 
469 #define AMDKFD_IOC_MAP_MEMORY_TO_GPU		\
470 		AMDKFD_IOWR(0x18, struct kfd_ioctl_map_memory_to_gpu_args)
471 
472 #define AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU	\
473 		AMDKFD_IOWR(0x19, struct kfd_ioctl_unmap_memory_from_gpu_args)
474 
475 #define AMDKFD_IOC_SET_CU_MASK		\
476 		AMDKFD_IOW(0x1A, struct kfd_ioctl_set_cu_mask_args)
477 
478 #define AMDKFD_COMMAND_START		0x01
479 #define AMDKFD_COMMAND_END		0x1B
480 
481 #endif
482