Commit | Line | Data |
---|---|---|
4a488a7a OG |
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_PRIV_H_INCLUDED | |
24 | #define KFD_PRIV_H_INCLUDED | |
25 | ||
26 | #include <linux/hashtable.h> | |
27 | #include <linux/mmu_notifier.h> | |
28 | #include <linux/mutex.h> | |
29 | #include <linux/types.h> | |
30 | #include <linux/atomic.h> | |
31 | #include <linux/workqueue.h> | |
32 | #include <linux/spinlock.h> | |
19f6d2a6 | 33 | #include <linux/kfd_ioctl.h> |
482f0777 | 34 | #include <linux/idr.h> |
04ad47bd | 35 | #include <linux/kfifo.h> |
851a645e | 36 | #include <linux/seq_file.h> |
5ce10687 | 37 | #include <linux/kref.h> |
4a488a7a OG |
38 | #include <kgd_kfd_interface.h> |
39 | ||
e596b903 YZ |
40 | #include "amd_shared.h" |
41 | ||
5b5c4e40 EP |
42 | #define KFD_SYSFS_FILE_MODE 0444 |
43 | ||
a1235e10 OG |
44 | #define KFD_MMAP_DOORBELL_MASK 0x8000000000000ull |
45 | #define KFD_MMAP_EVENTS_MASK 0x4000000000000ull | |
46 | #define KFD_MMAP_RESERVED_MEM_MASK 0x2000000000000ull | |
f3a39818 | 47 | |
ed6e6a34 BG |
48 | /* |
49 | * When working with cp scheduler we should assign the HIQ manually or via | |
50 | * the radeon driver to a fixed hqd slot, here are the fixed HIQ hqd slot | |
51 | * definitions for Kaveri. In Kaveri only the first ME queues participates | |
52 | * in the cp scheduling taking that in mind we set the HIQ slot in the | |
53 | * second ME. | |
54 | */ | |
55 | #define KFD_CIK_HIQ_PIPE 4 | |
56 | #define KFD_CIK_HIQ_QUEUE 0 | |
57 | ||
5b5c4e40 EP |
58 | /* GPU ID hash width in bits */ |
59 | #define KFD_GPU_ID_HASH_WIDTH 16 | |
60 | ||
61 | /* Macro for allocating structures */ | |
62 | #define kfd_alloc_struct(ptr_to_struct) \ | |
63 | ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) | |
64 | ||
19f6d2a6 | 65 | #define KFD_MAX_NUM_OF_PROCESSES 512 |
b8cbab04 | 66 | #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 |
19f6d2a6 | 67 | |
373d7080 FK |
68 | /* |
69 | * Size of the per-process TBA+TMA buffer: 2 pages | |
70 | * | |
71 | * The first page is the TBA used for the CWSR ISA code. The second | |
72 | * page is used as TMA for daisy changing a user-mode trap handler. | |
73 | */ | |
74 | #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2) | |
75 | #define KFD_CWSR_TMA_OFFSET PAGE_SIZE | |
76 | ||
19f6d2a6 | 77 | /* |
b8cbab04 OG |
78 | * Kernel module parameter to specify maximum number of supported queues per |
79 | * device | |
19f6d2a6 | 80 | */ |
b8cbab04 | 81 | extern int max_num_of_queues_per_device; |
19f6d2a6 | 82 | |
b8cbab04 OG |
83 | #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096 |
84 | #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ | |
85 | (KFD_MAX_NUM_OF_PROCESSES * \ | |
86 | KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) | |
19f6d2a6 | 87 | |
ed6e6a34 BG |
88 | #define KFD_KERNEL_QUEUE_SIZE 2048 |
89 | ||
31c21fec BG |
90 | /* Kernel module parameter to specify the scheduling policy */ |
91 | extern int sched_policy; | |
92 | ||
a99c6d4f FK |
93 | /* |
94 | * Kernel module parameter to specify the maximum process | |
95 | * number per HW scheduler | |
96 | */ | |
97 | extern int hws_max_conc_proc; | |
98 | ||
373d7080 FK |
99 | extern int cwsr_enable; |
100 | ||
81663016 OG |
101 | /* |
102 | * Kernel module parameter to specify whether to send sigterm to HSA process on | |
103 | * unhandled exception | |
104 | */ | |
105 | extern int send_sigterm; | |
106 | ||
ebcfd1e2 FK |
107 | /* |
108 | * Ignore CRAT table during KFD initialization, can be used to work around | |
109 | * broken CRAT tables on some AMD systems | |
110 | */ | |
111 | extern int ignore_crat; | |
112 | ||
31c21fec BG |
113 | /** |
114 | * enum kfd_sched_policy | |
115 | * | |
116 | * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp) | |
117 | * scheduling. In this scheduling mode we're using the firmware code to | |
118 | * schedule the user mode queues and kernel queues such as HIQ and DIQ. | |
119 | * the HIQ queue is used as a special queue that dispatches the configuration | |
120 | * to the cp and the user mode queues list that are currently running. | |
121 | * the DIQ queue is a debugging queue that dispatches debugging commands to the | |
122 | * firmware. | |
123 | * in this scheduling mode user mode queues over subscription feature is | |
124 | * enabled. | |
125 | * | |
126 | * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over | |
127 | * subscription feature disabled. | |
128 | * | |
129 | * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly | |
130 | * set the command processor registers and sets the queues "manually". This | |
131 | * mode is used *ONLY* for debugging proposes. | |
132 | * | |
133 | */ | |
134 | enum kfd_sched_policy { | |
135 | KFD_SCHED_POLICY_HWS = 0, | |
136 | KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION, | |
137 | KFD_SCHED_POLICY_NO_HWS | |
138 | }; | |
139 | ||
ed6e6a34 BG |
140 | enum cache_policy { |
141 | cache_policy_coherent, | |
142 | cache_policy_noncoherent | |
143 | }; | |
144 | ||
f3a39818 AL |
145 | struct kfd_event_interrupt_class { |
146 | bool (*interrupt_isr)(struct kfd_dev *dev, | |
147 | const uint32_t *ih_ring_entry); | |
148 | void (*interrupt_wq)(struct kfd_dev *dev, | |
149 | const uint32_t *ih_ring_entry); | |
150 | }; | |
151 | ||
4a488a7a | 152 | struct kfd_device_info { |
e596b903 | 153 | enum amd_asic_type asic_family; |
f3a39818 | 154 | const struct kfd_event_interrupt_class *event_interrupt_class; |
4a488a7a | 155 | unsigned int max_pasid_bits; |
992839ad | 156 | unsigned int max_no_of_hqd; |
4a488a7a | 157 | size_t ih_ring_entry_size; |
f7c826ad | 158 | uint8_t num_of_watch_points; |
19f6d2a6 | 159 | uint16_t mqd_size_aligned; |
373d7080 | 160 | bool supports_cwsr; |
64d1c3a4 | 161 | bool needs_iommu_device; |
3ee2d00c | 162 | bool needs_pci_atomics; |
4a488a7a OG |
163 | }; |
164 | ||
36b5c08f OG |
165 | struct kfd_mem_obj { |
166 | uint32_t range_start; | |
167 | uint32_t range_end; | |
168 | uint64_t gpu_addr; | |
169 | uint32_t *cpu_ptr; | |
170 | }; | |
171 | ||
44008d7a YZ |
172 | struct kfd_vmid_info { |
173 | uint32_t first_vmid_kfd; | |
174 | uint32_t last_vmid_kfd; | |
175 | uint32_t vmid_num_kfd; | |
176 | }; | |
177 | ||
4a488a7a OG |
178 | struct kfd_dev { |
179 | struct kgd_dev *kgd; | |
180 | ||
181 | const struct kfd_device_info *device_info; | |
182 | struct pci_dev *pdev; | |
183 | ||
184 | unsigned int id; /* topology stub index */ | |
185 | ||
19f6d2a6 OG |
186 | phys_addr_t doorbell_base; /* Start of actual doorbells used by |
187 | * KFD. It is aligned for mapping | |
188 | * into user mode | |
189 | */ | |
190 | size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell | |
191 | * to HW doorbell, GFX reserved some | |
192 | * at the start) | |
193 | */ | |
19f6d2a6 OG |
194 | u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells |
195 | * page used by kernel queue | |
196 | */ | |
197 | ||
4a488a7a | 198 | struct kgd2kfd_shared_resources shared_resources; |
44008d7a | 199 | struct kfd_vmid_info vm_info; |
4a488a7a | 200 | |
cea405b1 XZ |
201 | const struct kfd2kgd_calls *kfd2kgd; |
202 | struct mutex doorbell_mutex; | |
f761d8bd JP |
203 | DECLARE_BITMAP(doorbell_available_index, |
204 | KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); | |
cea405b1 | 205 | |
36b5c08f OG |
206 | void *gtt_mem; |
207 | uint64_t gtt_start_gpu_addr; | |
208 | void *gtt_start_cpu_ptr; | |
209 | void *gtt_sa_bitmap; | |
210 | struct mutex gtt_sa_lock; | |
211 | unsigned int gtt_sa_chunk_size; | |
212 | unsigned int gtt_sa_num_of_chunks; | |
213 | ||
2249d558 | 214 | /* Interrupts */ |
04ad47bd | 215 | struct kfifo ih_fifo; |
48e876a2 | 216 | struct workqueue_struct *ih_wq; |
2249d558 AL |
217 | struct work_struct interrupt_work; |
218 | spinlock_t interrupt_lock; | |
219 | ||
ed6e6a34 BG |
220 | /* QCM Device instance */ |
221 | struct device_queue_manager *dqm; | |
4a488a7a | 222 | |
ed6e6a34 | 223 | bool init_complete; |
2249d558 AL |
224 | /* |
225 | * Interrupts of interest to KFD are copied | |
226 | * from the HW ring into a SW ring. | |
227 | */ | |
228 | bool interrupts_active; | |
fbeb661b YS |
229 | |
230 | /* Debug manager */ | |
231 | struct kfd_dbgmgr *dbgmgr; | |
373d7080 | 232 | |
a99c6d4f FK |
233 | /* Maximum process number mapped to HW scheduler */ |
234 | unsigned int max_proc_per_quantum; | |
235 | ||
373d7080 FK |
236 | /* CWSR */ |
237 | bool cwsr_enabled; | |
238 | const void *cwsr_isa; | |
239 | unsigned int cwsr_isa_size; | |
4a488a7a OG |
240 | }; |
241 | ||
242 | /* KGD2KFD callbacks */ | |
243 | void kgd2kfd_exit(void); | |
cea405b1 XZ |
244 | struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, |
245 | struct pci_dev *pdev, const struct kfd2kgd_calls *f2g); | |
4a488a7a | 246 | bool kgd2kfd_device_init(struct kfd_dev *kfd, |
cea405b1 | 247 | const struct kgd2kfd_shared_resources *gpu_resources); |
4a488a7a OG |
248 | void kgd2kfd_device_exit(struct kfd_dev *kfd); |
249 | ||
19f6d2a6 OG |
250 | enum kfd_mempool { |
251 | KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, | |
252 | KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, | |
253 | KFD_MEMPOOL_FRAMEBUFFER = 3, | |
254 | }; | |
255 | ||
4a488a7a OG |
256 | /* Character device interface */ |
257 | int kfd_chardev_init(void); | |
258 | void kfd_chardev_exit(void); | |
259 | struct device *kfd_chardev(void); | |
260 | ||
241f24f8 | 261 | /** |
7da2bcf8 | 262 | * enum kfd_unmap_queues_filter |
241f24f8 | 263 | * |
7da2bcf8 | 264 | * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. |
241f24f8 | 265 | * |
7da2bcf8 | 266 | * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the |
241f24f8 BG |
267 | * running queues list. |
268 | * | |
7da2bcf8 | 269 | * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to |
241f24f8 BG |
270 | * specific process. |
271 | * | |
272 | */ | |
7da2bcf8 YZ |
273 | enum kfd_unmap_queues_filter { |
274 | KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, | |
275 | KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, | |
276 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, | |
277 | KFD_UNMAP_QUEUES_FILTER_BY_PASID | |
241f24f8 | 278 | }; |
19f6d2a6 | 279 | |
ed8aab45 BG |
280 | /** |
281 | * enum kfd_queue_type | |
282 | * | |
283 | * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. | |
284 | * | |
285 | * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type. | |
286 | * | |
287 | * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. | |
288 | * | |
289 | * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. | |
290 | */ | |
291 | enum kfd_queue_type { | |
292 | KFD_QUEUE_TYPE_COMPUTE, | |
293 | KFD_QUEUE_TYPE_SDMA, | |
294 | KFD_QUEUE_TYPE_HIQ, | |
295 | KFD_QUEUE_TYPE_DIQ | |
296 | }; | |
297 | ||
6e99df57 BG |
298 | enum kfd_queue_format { |
299 | KFD_QUEUE_FORMAT_PM4, | |
300 | KFD_QUEUE_FORMAT_AQL | |
301 | }; | |
302 | ||
ed8aab45 BG |
303 | /** |
304 | * struct queue_properties | |
305 | * | |
306 | * @type: The queue type. | |
307 | * | |
308 | * @queue_id: Queue identifier. | |
309 | * | |
310 | * @queue_address: Queue ring buffer address. | |
311 | * | |
312 | * @queue_size: Queue ring buffer size. | |
313 | * | |
314 | * @priority: Defines the queue priority relative to other queues in the | |
315 | * process. | |
316 | * This is just an indication and HW scheduling may override the priority as | |
317 | * necessary while keeping the relative prioritization. | |
318 | * the priority granularity is from 0 to f which f is the highest priority. | |
319 | * currently all queues are initialized with the highest priority. | |
320 | * | |
321 | * @queue_percent: This field is partially implemented and currently a zero in | |
322 | * this field defines that the queue is non active. | |
323 | * | |
324 | * @read_ptr: User space address which points to the number of dwords the | |
325 | * cp read from the ring buffer. This field updates automatically by the H/W. | |
326 | * | |
327 | * @write_ptr: Defines the number of dwords written to the ring buffer. | |
328 | * | |
329 | * @doorbell_ptr: This field aim is to notify the H/W of new packet written to | |
8eabaf54 KR |
330 | * the queue ring buffer. This field should be similar to write_ptr and the |
331 | * user should update this field after he updated the write_ptr. | |
ed8aab45 BG |
332 | * |
333 | * @doorbell_off: The doorbell offset in the doorbell pci-bar. | |
334 | * | |
8eabaf54 KR |
335 | * @is_interop: Defines if this is a interop queue. Interop queue means that |
336 | * the queue can access both graphics and compute resources. | |
ed8aab45 BG |
337 | * |
338 | * @is_active: Defines if the queue is active or not. | |
339 | * | |
340 | * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid | |
341 | * of the queue. | |
342 | * | |
343 | * This structure represents the queue properties for each queue no matter if | |
344 | * it's user mode or kernel mode queue. | |
345 | * | |
346 | */ | |
347 | struct queue_properties { | |
348 | enum kfd_queue_type type; | |
6e99df57 | 349 | enum kfd_queue_format format; |
ed8aab45 BG |
350 | unsigned int queue_id; |
351 | uint64_t queue_address; | |
352 | uint64_t queue_size; | |
353 | uint32_t priority; | |
354 | uint32_t queue_percent; | |
355 | uint32_t *read_ptr; | |
356 | uint32_t *write_ptr; | |
5cd78de5 | 357 | uint32_t __iomem *doorbell_ptr; |
ed8aab45 BG |
358 | uint32_t doorbell_off; |
359 | bool is_interop; | |
360 | bool is_active; | |
361 | /* Not relevant for user mode queues in cp scheduling */ | |
362 | unsigned int vmid; | |
77669eb8 BG |
363 | /* Relevant only for sdma queues*/ |
364 | uint32_t sdma_engine_id; | |
365 | uint32_t sdma_queue_id; | |
366 | uint32_t sdma_vm_addr; | |
ff3d04a1 BG |
367 | /* Relevant only for VI */ |
368 | uint64_t eop_ring_buffer_address; | |
369 | uint32_t eop_ring_buffer_size; | |
370 | uint64_t ctx_save_restore_area_address; | |
371 | uint32_t ctx_save_restore_area_size; | |
373d7080 FK |
372 | uint32_t ctl_stack_size; |
373 | uint64_t tba_addr; | |
374 | uint64_t tma_addr; | |
ed8aab45 BG |
375 | }; |
376 | ||
377 | /** | |
378 | * struct queue | |
379 | * | |
380 | * @list: Queue linked list. | |
381 | * | |
382 | * @mqd: The queue MQD. | |
383 | * | |
384 | * @mqd_mem_obj: The MQD local gpu memory object. | |
385 | * | |
386 | * @gart_mqd_addr: The MQD gart mc address. | |
387 | * | |
388 | * @properties: The queue properties. | |
389 | * | |
390 | * @mec: Used only in no cp scheduling mode and identifies to micro engine id | |
8eabaf54 | 391 | * that the queue should be execute on. |
ed8aab45 | 392 | * |
8eabaf54 KR |
393 | * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe |
394 | * id. | |
ed8aab45 BG |
395 | * |
396 | * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. | |
397 | * | |
398 | * @process: The kfd process that created this queue. | |
399 | * | |
400 | * @device: The kfd device that created this queue. | |
401 | * | |
402 | * This structure represents user mode compute queues. | |
403 | * It contains all the necessary data to handle such queues. | |
404 | * | |
405 | */ | |
406 | ||
407 | struct queue { | |
408 | struct list_head list; | |
409 | void *mqd; | |
410 | struct kfd_mem_obj *mqd_mem_obj; | |
411 | uint64_t gart_mqd_addr; | |
412 | struct queue_properties properties; | |
413 | ||
414 | uint32_t mec; | |
415 | uint32_t pipe; | |
416 | uint32_t queue; | |
417 | ||
77669eb8 BG |
418 | unsigned int sdma_id; |
419 | ||
ed8aab45 BG |
420 | struct kfd_process *process; |
421 | struct kfd_dev *device; | |
422 | }; | |
423 | ||
6e99df57 BG |
424 | /* |
425 | * Please read the kfd_mqd_manager.h description. | |
426 | */ | |
427 | enum KFD_MQD_TYPE { | |
85d258f9 BG |
428 | KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */ |
429 | KFD_MQD_TYPE_HIQ, /* for hiq */ | |
430 | KFD_MQD_TYPE_CP, /* for cp queues and diq */ | |
431 | KFD_MQD_TYPE_SDMA, /* for sdma queues */ | |
6e99df57 BG |
432 | KFD_MQD_TYPE_MAX |
433 | }; | |
434 | ||
241f24f8 BG |
435 | struct scheduling_resources { |
436 | unsigned int vmid_mask; | |
437 | enum kfd_queue_type type; | |
438 | uint64_t queue_mask; | |
439 | uint64_t gws_mask; | |
440 | uint32_t oac_mask; | |
441 | uint32_t gds_heap_base; | |
442 | uint32_t gds_heap_size; | |
443 | }; | |
444 | ||
445 | struct process_queue_manager { | |
446 | /* data */ | |
447 | struct kfd_process *process; | |
241f24f8 BG |
448 | struct list_head queues; |
449 | unsigned long *queue_slot_bitmap; | |
450 | }; | |
451 | ||
452 | struct qcm_process_device { | |
453 | /* The Device Queue Manager that owns this data */ | |
454 | struct device_queue_manager *dqm; | |
455 | struct process_queue_manager *pqm; | |
241f24f8 BG |
456 | /* Queues list */ |
457 | struct list_head queues_list; | |
458 | struct list_head priv_queue_list; | |
459 | ||
460 | unsigned int queue_count; | |
461 | unsigned int vmid; | |
462 | bool is_debug; | |
9fd3f1bf FK |
463 | |
464 | /* This flag tells if we should reset all wavefronts on | |
465 | * process termination | |
466 | */ | |
467 | bool reset_wavefronts; | |
468 | ||
241f24f8 BG |
469 | /* |
470 | * All the memory management data should be here too | |
471 | */ | |
472 | uint64_t gds_context_area; | |
473 | uint32_t sh_mem_config; | |
474 | uint32_t sh_mem_bases; | |
475 | uint32_t sh_mem_ape1_base; | |
476 | uint32_t sh_mem_ape1_limit; | |
477 | uint32_t page_table_base; | |
478 | uint32_t gds_size; | |
479 | uint32_t num_gws; | |
480 | uint32_t num_oac; | |
6a1c9510 | 481 | uint32_t sh_hidden_private_base; |
373d7080 FK |
482 | |
483 | /* CWSR memory */ | |
484 | void *cwsr_kaddr; | |
485 | uint64_t tba_addr; | |
486 | uint64_t tma_addr; | |
241f24f8 BG |
487 | }; |
488 | ||
733fa1f7 YZ |
489 | |
490 | enum kfd_pdd_bound { | |
491 | PDD_UNBOUND = 0, | |
492 | PDD_BOUND, | |
493 | PDD_BOUND_SUSPENDED, | |
494 | }; | |
495 | ||
19f6d2a6 OG |
496 | /* Data that is per-process-per device. */ |
497 | struct kfd_process_device { | |
498 | /* | |
499 | * List of all per-device data for a process. | |
500 | * Starts from kfd_process.per_device_data. | |
501 | */ | |
502 | struct list_head per_device_list; | |
503 | ||
504 | /* The device that owns this data. */ | |
505 | struct kfd_dev *dev; | |
506 | ||
9fd3f1bf FK |
507 | /* The process that owns this kfd_process_device. */ |
508 | struct kfd_process *process; | |
19f6d2a6 | 509 | |
45102048 BG |
510 | /* per-process-per device QCM data structure */ |
511 | struct qcm_process_device qpd; | |
512 | ||
19f6d2a6 OG |
513 | /*Apertures*/ |
514 | uint64_t lds_base; | |
515 | uint64_t lds_limit; | |
516 | uint64_t gpuvm_base; | |
517 | uint64_t gpuvm_limit; | |
518 | uint64_t scratch_base; | |
519 | uint64_t scratch_limit; | |
520 | ||
9fd3f1bf FK |
521 | /* Flag used to tell the pdd has dequeued from the dqm. |
522 | * This is used to prevent dev->dqm->ops.process_termination() from | |
523 | * being called twice when it is already called in IOMMU callback | |
524 | * function. | |
a82918f1 | 525 | */ |
9fd3f1bf | 526 | bool already_dequeued; |
64d1c3a4 FK |
527 | |
528 | /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ | |
529 | enum kfd_pdd_bound bound; | |
19f6d2a6 OG |
530 | }; |
531 | ||
52a5fdce AS |
532 | #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) |
533 | ||
4a488a7a OG |
534 | /* Process data */ |
535 | struct kfd_process { | |
19f6d2a6 OG |
536 | /* |
537 | * kfd_process are stored in an mm_struct*->kfd_process* | |
538 | * hash table (kfd_processes in kfd_process.c) | |
539 | */ | |
540 | struct hlist_node kfd_processes; | |
541 | ||
9b56bb11 FK |
542 | /* |
543 | * Opaque pointer to mm_struct. We don't hold a reference to | |
544 | * it so it should never be dereferenced from here. This is | |
545 | * only used for looking up processes by their mm. | |
546 | */ | |
547 | void *mm; | |
19f6d2a6 | 548 | |
5ce10687 FK |
549 | struct kref ref; |
550 | struct work_struct release_work; | |
551 | ||
19f6d2a6 OG |
552 | struct mutex mutex; |
553 | ||
554 | /* | |
555 | * In any process, the thread that started main() is the lead | |
556 | * thread and outlives the rest. | |
557 | * It is here because amd_iommu_bind_pasid wants a task_struct. | |
894a8293 FK |
558 | * It can also be used for safely getting a reference to the |
559 | * mm_struct of the process. | |
19f6d2a6 OG |
560 | */ |
561 | struct task_struct *lead_thread; | |
562 | ||
563 | /* We want to receive a notification when the mm_struct is destroyed */ | |
564 | struct mmu_notifier mmu_notifier; | |
565 | ||
566 | /* Use for delayed freeing of kfd_process structure */ | |
567 | struct rcu_head rcu; | |
568 | ||
569 | unsigned int pasid; | |
a91e70e3 | 570 | unsigned int doorbell_index; |
19f6d2a6 OG |
571 | |
572 | /* | |
573 | * List of kfd_process_device structures, | |
574 | * one for each device the process is using. | |
575 | */ | |
576 | struct list_head per_device_data; | |
577 | ||
45102048 BG |
578 | struct process_queue_manager pqm; |
579 | ||
19f6d2a6 OG |
580 | /*Is the user space process 32 bit?*/ |
581 | bool is_32bit_user_mode; | |
f3a39818 AL |
582 | |
583 | /* Event-related data */ | |
584 | struct mutex event_mutex; | |
482f0777 FK |
585 | /* Event ID allocator and lookup */ |
586 | struct idr event_idr; | |
50cb7dd9 FK |
587 | /* Event page */ |
588 | struct kfd_signal_page *signal_page; | |
b9a5d0a5 | 589 | size_t signal_mapped_size; |
f3a39818 | 590 | size_t signal_event_count; |
c986169f | 591 | bool signal_event_limit_reached; |
4a488a7a OG |
592 | }; |
593 | ||
64d1c3a4 FK |
594 | #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ |
595 | extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); | |
596 | extern struct srcu_struct kfd_processes_srcu; | |
597 | ||
76baee6c OG |
598 | /** |
599 | * Ioctl function type. | |
600 | * | |
601 | * \param filep pointer to file structure. | |
602 | * \param p amdkfd process pointer. | |
603 | * \param data pointer to arg that was copied from user. | |
604 | */ | |
605 | typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, | |
606 | void *data); | |
607 | ||
608 | struct amdkfd_ioctl_desc { | |
609 | unsigned int cmd; | |
610 | int flags; | |
611 | amdkfd_ioctl_t *func; | |
612 | unsigned int cmd_drv; | |
613 | const char *name; | |
614 | }; | |
615 | ||
19f6d2a6 OG |
616 | void kfd_process_create_wq(void); |
617 | void kfd_process_destroy_wq(void); | |
373d7080 | 618 | struct kfd_process *kfd_create_process(struct file *filep); |
19f6d2a6 | 619 | struct kfd_process *kfd_get_process(const struct task_struct *); |
f3a39818 | 620 | struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid); |
abb208a8 | 621 | void kfd_unref_process(struct kfd_process *p); |
19f6d2a6 | 622 | |
64c7f8cf | 623 | struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, |
733fa1f7 | 624 | struct kfd_process *p); |
19f6d2a6 | 625 | struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, |
093c7d8c AS |
626 | struct kfd_process *p); |
627 | struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, | |
628 | struct kfd_process *p); | |
19f6d2a6 | 629 | |
373d7080 FK |
630 | int kfd_reserved_mem_mmap(struct kfd_process *process, |
631 | struct vm_area_struct *vma); | |
632 | ||
775921ed | 633 | /* Process device data iterator */ |
8eabaf54 KR |
634 | struct kfd_process_device *kfd_get_first_process_device_data( |
635 | struct kfd_process *p); | |
636 | struct kfd_process_device *kfd_get_next_process_device_data( | |
637 | struct kfd_process *p, | |
775921ed AS |
638 | struct kfd_process_device *pdd); |
639 | bool kfd_has_process_device_data(struct kfd_process *p); | |
640 | ||
19f6d2a6 OG |
641 | /* PASIDs */ |
642 | int kfd_pasid_init(void); | |
643 | void kfd_pasid_exit(void); | |
644 | bool kfd_set_pasid_limit(unsigned int new_limit); | |
645 | unsigned int kfd_get_pasid_limit(void); | |
646 | unsigned int kfd_pasid_alloc(void); | |
647 | void kfd_pasid_free(unsigned int pasid); | |
648 | ||
649 | /* Doorbells */ | |
735df2ba FK |
650 | int kfd_doorbell_init(struct kfd_dev *kfd); |
651 | void kfd_doorbell_fini(struct kfd_dev *kfd); | |
19f6d2a6 OG |
652 | int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma); |
653 | u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, | |
654 | unsigned int *doorbell_off); | |
655 | void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); | |
656 | u32 read_kernel_doorbell(u32 __iomem *db); | |
657 | void write_kernel_doorbell(u32 __iomem *db, u32 value); | |
658 | unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd, | |
659 | struct kfd_process *process, | |
660 | unsigned int queue_id); | |
a91e70e3 FK |
661 | phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, |
662 | struct kfd_process *process); | |
663 | int kfd_alloc_process_doorbells(struct kfd_process *process); | |
664 | void kfd_free_process_doorbells(struct kfd_process *process); | |
19f6d2a6 | 665 | |
6e81090b OG |
666 | /* GTT Sub-Allocator */ |
667 | ||
668 | int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, | |
669 | struct kfd_mem_obj **mem_obj); | |
670 | ||
671 | int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); | |
672 | ||
4a488a7a OG |
673 | extern struct device *kfd_device; |
674 | ||
5b5c4e40 EP |
675 | /* Topology */ |
676 | int kfd_topology_init(void); | |
677 | void kfd_topology_shutdown(void); | |
678 | int kfd_topology_add_device(struct kfd_dev *gpu); | |
679 | int kfd_topology_remove_device(struct kfd_dev *gpu); | |
3a87177e HK |
680 | struct kfd_topology_device *kfd_topology_device_by_proximity_domain( |
681 | uint32_t proximity_domain); | |
5b5c4e40 EP |
682 | struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); |
683 | struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); | |
6d82eb0e | 684 | int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); |
520b8fb7 | 685 | int kfd_numa_node_to_apic_id(int numa_node_id); |
5b5c4e40 | 686 | |
4a488a7a | 687 | /* Interrupts */ |
2249d558 AL |
688 | int kfd_interrupt_init(struct kfd_dev *dev); |
689 | void kfd_interrupt_exit(struct kfd_dev *dev); | |
b3f5e6b4 | 690 | void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); |
2249d558 AL |
691 | bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); |
692 | bool interrupt_is_wanted(struct kfd_dev *dev, const uint32_t *ih_ring_entry); | |
4a488a7a OG |
693 | |
694 | /* Power Management */ | |
b3f5e6b4 AL |
695 | void kgd2kfd_suspend(struct kfd_dev *kfd); |
696 | int kgd2kfd_resume(struct kfd_dev *kfd); | |
4a488a7a | 697 | |
19f6d2a6 OG |
698 | /* amdkfd Apertures */ |
699 | int kfd_init_apertures(struct kfd_process *process); | |
700 | ||
ed6e6a34 | 701 | /* Queue Context Management */ |
e88a614c | 702 | int init_queue(struct queue **q, const struct queue_properties *properties); |
ed6e6a34 | 703 | void uninit_queue(struct queue *q); |
45102048 | 704 | void print_queue_properties(struct queue_properties *q); |
ed6e6a34 BG |
705 | void print_queue(struct queue *q); |
706 | ||
64c7f8cf BG |
707 | struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type, |
708 | struct kfd_dev *dev); | |
4b8f589b BG |
709 | struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, |
710 | struct kfd_dev *dev); | |
ee04955a FK |
711 | struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, |
712 | struct kfd_dev *dev); | |
4b8f589b BG |
713 | struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, |
714 | struct kfd_dev *dev); | |
ee04955a FK |
715 | struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, |
716 | struct kfd_dev *dev); | |
64c7f8cf BG |
717 | struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); |
718 | void device_queue_manager_uninit(struct device_queue_manager *dqm); | |
241f24f8 BG |
719 | struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, |
720 | enum kfd_queue_type type); | |
721 | void kernel_queue_uninit(struct kernel_queue *kq); | |
722 | ||
45102048 BG |
723 | /* Process Queue Manager */ |
724 | struct process_queue_node { | |
725 | struct queue *q; | |
726 | struct kernel_queue *kq; | |
727 | struct list_head process_queue_list; | |
728 | }; | |
729 | ||
9fd3f1bf FK |
730 | void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); |
731 | void kfd_process_dequeue_from_all_devices(struct kfd_process *p); | |
45102048 BG |
732 | int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); |
733 | void pqm_uninit(struct process_queue_manager *pqm); | |
734 | int pqm_create_queue(struct process_queue_manager *pqm, | |
735 | struct kfd_dev *dev, | |
736 | struct file *f, | |
737 | struct queue_properties *properties, | |
45102048 BG |
738 | unsigned int *qid); |
739 | int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); | |
740 | int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, | |
741 | struct queue_properties *p); | |
fbeb661b YS |
742 | struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, |
743 | unsigned int qid); | |
45102048 | 744 | |
788bf83d YS |
745 | int amdkfd_fence_wait_timeout(unsigned int *fence_addr, |
746 | unsigned int fence_value, | |
8c72c3d7 | 747 | unsigned int timeout_ms); |
788bf83d | 748 | |
ed6e6a34 BG |
749 | /* Packet Manager */ |
750 | ||
64c7f8cf BG |
751 | #define KFD_FENCE_COMPLETED (100) |
752 | #define KFD_FENCE_INIT (10) | |
241f24f8 | 753 | |
ed6e6a34 BG |
754 | struct packet_manager { |
755 | struct device_queue_manager *dqm; | |
756 | struct kernel_queue *priv_queue; | |
757 | struct mutex lock; | |
758 | bool allocated; | |
759 | struct kfd_mem_obj *ib_buffer_obj; | |
851a645e | 760 | unsigned int ib_size_bytes; |
ed6e6a34 BG |
761 | }; |
762 | ||
64c7f8cf BG |
763 | int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); |
764 | void pm_uninit(struct packet_manager *pm); | |
765 | int pm_send_set_resources(struct packet_manager *pm, | |
766 | struct scheduling_resources *res); | |
767 | int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); | |
768 | int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, | |
769 | uint32_t fence_value); | |
770 | ||
771 | int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, | |
7da2bcf8 | 772 | enum kfd_unmap_queues_filter mode, |
64c7f8cf BG |
773 | uint32_t filter_param, bool reset, |
774 | unsigned int sdma_engine); | |
775 | ||
241f24f8 BG |
776 | void pm_release_ib(struct packet_manager *pm); |
777 | ||
19f6d2a6 | 778 | uint64_t kfd_get_number_elems(struct kfd_dev *kfd); |
19f6d2a6 | 779 | |
f3a39818 AL |
780 | /* Events */ |
781 | extern const struct kfd_event_interrupt_class event_interrupt_class_cik; | |
930c5ff4 | 782 | extern const struct kfd_device_global_init_class device_global_init_class_cik; |
f3a39818 | 783 | |
f3a39818 AL |
784 | void kfd_event_init_process(struct kfd_process *p); |
785 | void kfd_event_free_process(struct kfd_process *p); | |
786 | int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); | |
787 | int kfd_wait_on_events(struct kfd_process *p, | |
59d3e8be | 788 | uint32_t num_events, void __user *data, |
f3a39818 | 789 | bool all, uint32_t user_timeout_ms, |
fdf0c833 | 790 | uint32_t *wait_result); |
f3a39818 AL |
791 | void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id, |
792 | uint32_t valid_id_bits); | |
59d3e8be AS |
793 | void kfd_signal_iommu_event(struct kfd_dev *dev, |
794 | unsigned int pasid, unsigned long address, | |
795 | bool is_write_requested, bool is_execute_requested); | |
930c5ff4 | 796 | void kfd_signal_hw_exception_event(unsigned int pasid); |
f3a39818 AL |
797 | int kfd_set_event(struct kfd_process *p, uint32_t event_id); |
798 | int kfd_reset_event(struct kfd_process *p, uint32_t event_id); | |
799 | int kfd_event_create(struct file *devkfd, struct kfd_process *p, | |
800 | uint32_t event_type, bool auto_reset, uint32_t node_id, | |
801 | uint32_t *event_id, uint32_t *event_trigger_data, | |
802 | uint64_t *event_page_offset, uint32_t *event_slot_index); | |
803 | int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); | |
804 | ||
c3447e81 BG |
805 | int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p); |
806 | ||
851a645e FK |
807 | /* Debugfs */ |
808 | #if defined(CONFIG_DEBUG_FS) | |
809 | ||
810 | void kfd_debugfs_init(void); | |
811 | void kfd_debugfs_fini(void); | |
812 | int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); | |
813 | int pqm_debugfs_mqds(struct seq_file *m, void *data); | |
814 | int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); | |
815 | int dqm_debugfs_hqds(struct seq_file *m, void *data); | |
816 | int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); | |
817 | int pm_debugfs_runlist(struct seq_file *m, void *data); | |
818 | ||
819 | #else | |
820 | ||
821 | static inline void kfd_debugfs_init(void) {} | |
822 | static inline void kfd_debugfs_fini(void) {} | |
823 | ||
824 | #endif | |
825 | ||
4a488a7a | 826 | #endif |