Commit | Line | Data |
---|---|---|
f3a39818 AL |
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 | #include <linux/mm_types.h> | |
24 | #include <linux/slab.h> | |
25 | #include <linux/types.h> | |
26 | #include <linux/sched.h> | |
27 | #include <linux/uaccess.h> | |
28 | #include <linux/mm.h> | |
29 | #include <linux/mman.h> | |
30 | #include <linux/memory.h> | |
31 | #include "kfd_priv.h" | |
32 | #include "kfd_events.h" | |
59d3e8be | 33 | #include <linux/device.h> |
f3a39818 AL |
34 | |
35 | /* | |
36 | * A task can only be on a single wait_queue at a time, but we need to support | |
37 | * waiting on multiple events (any/all). | |
38 | * Instead of each event simply having a wait_queue with sleeping tasks, it | |
39 | * has a singly-linked list of tasks. | |
40 | * A thread that wants to sleep creates an array of these, one for each event | |
41 | * and adds one to each event's waiter chain. | |
42 | */ | |
43 | struct kfd_event_waiter { | |
44 | struct list_head waiters; | |
45 | struct task_struct *sleeping_task; | |
46 | ||
47 | /* Transitions to true when the event this belongs to is signaled. */ | |
48 | bool activated; | |
59d3e8be AS |
49 | |
50 | /* Event */ | |
51 | struct kfd_event *event; | |
52 | uint32_t input_index; | |
f3a39818 AL |
53 | }; |
54 | ||
55 | /* | |
56 | * Over-complicated pooled allocator for event notification slots. | |
57 | * | |
58 | * Each signal event needs a 64-bit signal slot where the signaler will write | |
59 | * a 1 before sending an interrupt.l (This is needed because some interrupts | |
60 | * do not contain enough spare data bits to identify an event.) | |
61 | * We get whole pages from vmalloc and map them to the process VA. | |
62 | * Individual signal events are then allocated a slot in a page. | |
63 | */ | |
64 | ||
65 | struct signal_page { | |
66 | struct list_head event_pages; /* kfd_process.signal_event_pages */ | |
67 | uint64_t *kernel_address; | |
68 | uint64_t __user *user_address; | |
69 | uint32_t page_index; /* Index into the mmap aperture. */ | |
70 | unsigned int free_slots; | |
71 | unsigned long used_slot_bitmap[0]; | |
72 | }; | |
73 | ||
74 | #define SLOTS_PER_PAGE KFD_SIGNAL_EVENT_LIMIT | |
75 | #define SLOT_BITMAP_SIZE BITS_TO_LONGS(SLOTS_PER_PAGE) | |
76 | #define BITS_PER_PAGE (ilog2(SLOTS_PER_PAGE)+1) | |
77 | #define SIGNAL_PAGE_SIZE (sizeof(struct signal_page) + \ | |
78 | SLOT_BITMAP_SIZE * sizeof(long)) | |
79 | ||
80 | /* | |
81 | * For signal events, the event ID is used as the interrupt user data. | |
82 | * For SQ s_sendmsg interrupts, this is limited to 8 bits. | |
83 | */ | |
84 | ||
85 | #define INTERRUPT_DATA_BITS 8 | |
86 | #define SIGNAL_EVENT_ID_SLOT_SHIFT 0 | |
87 | ||
88 | static uint64_t *page_slots(struct signal_page *page) | |
89 | { | |
90 | return page->kernel_address; | |
91 | } | |
92 | ||
93 | static bool allocate_free_slot(struct kfd_process *process, | |
94 | struct signal_page **out_page, | |
95 | unsigned int *out_slot_index) | |
96 | { | |
97 | struct signal_page *page; | |
98 | ||
99 | list_for_each_entry(page, &process->signal_event_pages, event_pages) { | |
100 | if (page->free_slots > 0) { | |
101 | unsigned int slot = | |
102 | find_first_zero_bit(page->used_slot_bitmap, | |
103 | SLOTS_PER_PAGE); | |
104 | ||
105 | __set_bit(slot, page->used_slot_bitmap); | |
106 | page->free_slots--; | |
107 | ||
108 | page_slots(page)[slot] = UNSIGNALED_EVENT_SLOT; | |
109 | ||
110 | *out_page = page; | |
111 | *out_slot_index = slot; | |
112 | ||
113 | pr_debug("allocated event signal slot in page %p, slot %d\n", | |
114 | page, slot); | |
115 | ||
116 | return true; | |
117 | } | |
118 | } | |
119 | ||
120 | pr_debug("No free event signal slots were found for process %p\n", | |
121 | process); | |
122 | ||
123 | return false; | |
124 | } | |
125 | ||
126 | #define list_tail_entry(head, type, member) \ | |
127 | list_entry((head)->prev, type, member) | |
128 | ||
129 | static bool allocate_signal_page(struct file *devkfd, struct kfd_process *p) | |
130 | { | |
131 | void *backing_store; | |
132 | struct signal_page *page; | |
133 | ||
134 | page = kzalloc(SIGNAL_PAGE_SIZE, GFP_KERNEL); | |
135 | if (!page) | |
136 | goto fail_alloc_signal_page; | |
137 | ||
138 | page->free_slots = SLOTS_PER_PAGE; | |
139 | ||
140 | backing_store = (void *) __get_free_pages(GFP_KERNEL | __GFP_ZERO, | |
141 | get_order(KFD_SIGNAL_EVENT_LIMIT * 8)); | |
142 | if (!backing_store) | |
143 | goto fail_alloc_signal_store; | |
144 | ||
145 | /* prevent user-mode info leaks */ | |
146 | memset(backing_store, (uint8_t) UNSIGNALED_EVENT_SLOT, | |
147 | KFD_SIGNAL_EVENT_LIMIT * 8); | |
148 | ||
149 | page->kernel_address = backing_store; | |
150 | ||
151 | if (list_empty(&p->signal_event_pages)) | |
152 | page->page_index = 0; | |
153 | else | |
154 | page->page_index = list_tail_entry(&p->signal_event_pages, | |
155 | struct signal_page, | |
156 | event_pages)->page_index + 1; | |
157 | ||
158 | pr_debug("allocated new event signal page at %p, for process %p\n", | |
159 | page, p); | |
160 | pr_debug("page index is %d\n", page->page_index); | |
161 | ||
162 | list_add(&page->event_pages, &p->signal_event_pages); | |
163 | ||
164 | return true; | |
165 | ||
166 | fail_alloc_signal_store: | |
167 | kfree(page); | |
168 | fail_alloc_signal_page: | |
169 | return false; | |
170 | } | |
171 | ||
172 | static bool allocate_event_notification_slot(struct file *devkfd, | |
173 | struct kfd_process *p, | |
174 | struct signal_page **page, | |
175 | unsigned int *signal_slot_index) | |
176 | { | |
177 | bool ret; | |
178 | ||
179 | ret = allocate_free_slot(p, page, signal_slot_index); | |
991ca8ee | 180 | if (!ret) { |
f3a39818 | 181 | ret = allocate_signal_page(devkfd, p); |
991ca8ee | 182 | if (ret) |
f3a39818 AL |
183 | ret = allocate_free_slot(p, page, signal_slot_index); |
184 | } | |
185 | ||
186 | return ret; | |
187 | } | |
188 | ||
189 | /* Assumes that the process's event_mutex is locked. */ | |
190 | static void release_event_notification_slot(struct signal_page *page, | |
191 | size_t slot_index) | |
192 | { | |
193 | __clear_bit(slot_index, page->used_slot_bitmap); | |
194 | page->free_slots++; | |
195 | ||
196 | /* We don't free signal pages, they are retained by the process | |
197 | * and reused until it exits. */ | |
198 | } | |
199 | ||
200 | static struct signal_page *lookup_signal_page_by_index(struct kfd_process *p, | |
201 | unsigned int page_index) | |
202 | { | |
203 | struct signal_page *page; | |
204 | ||
205 | /* | |
206 | * This is safe because we don't delete signal pages until the | |
207 | * process exits. | |
208 | */ | |
209 | list_for_each_entry(page, &p->signal_event_pages, event_pages) | |
210 | if (page->page_index == page_index) | |
211 | return page; | |
212 | ||
213 | return NULL; | |
214 | } | |
215 | ||
216 | /* | |
217 | * Assumes that p->event_mutex is held and of course that p is not going | |
218 | * away (current or locked). | |
219 | */ | |
220 | static struct kfd_event *lookup_event_by_id(struct kfd_process *p, uint32_t id) | |
221 | { | |
222 | struct kfd_event *ev; | |
223 | ||
224 | hash_for_each_possible(p->events, ev, events, id) | |
225 | if (ev->event_id == id) | |
226 | return ev; | |
227 | ||
228 | return NULL; | |
229 | } | |
230 | ||
231 | static u32 make_signal_event_id(struct signal_page *page, | |
232 | unsigned int signal_slot_index) | |
233 | { | |
234 | return page->page_index | | |
235 | (signal_slot_index << SIGNAL_EVENT_ID_SLOT_SHIFT); | |
236 | } | |
237 | ||
238 | /* | |
239 | * Produce a kfd event id for a nonsignal event. | |
240 | * These are arbitrary numbers, so we do a sequential search through | |
241 | * the hash table for an unused number. | |
242 | */ | |
243 | static u32 make_nonsignal_event_id(struct kfd_process *p) | |
244 | { | |
245 | u32 id; | |
246 | ||
247 | for (id = p->next_nonsignal_event_id; | |
248 | id < KFD_LAST_NONSIGNAL_EVENT_ID && | |
249 | lookup_event_by_id(p, id) != NULL; | |
250 | id++) | |
251 | ; | |
252 | ||
253 | if (id < KFD_LAST_NONSIGNAL_EVENT_ID) { | |
254 | ||
255 | /* | |
256 | * What if id == LAST_NONSIGNAL_EVENT_ID - 1? | |
257 | * Then next_nonsignal_event_id = LAST_NONSIGNAL_EVENT_ID so | |
258 | * the first loop fails immediately and we proceed with the | |
259 | * wraparound loop below. | |
260 | */ | |
261 | p->next_nonsignal_event_id = id + 1; | |
262 | ||
263 | return id; | |
264 | } | |
265 | ||
266 | for (id = KFD_FIRST_NONSIGNAL_EVENT_ID; | |
267 | id < KFD_LAST_NONSIGNAL_EVENT_ID && | |
268 | lookup_event_by_id(p, id) != NULL; | |
269 | id++) | |
270 | ; | |
271 | ||
272 | ||
273 | if (id < KFD_LAST_NONSIGNAL_EVENT_ID) { | |
274 | p->next_nonsignal_event_id = id + 1; | |
275 | return id; | |
276 | } | |
277 | ||
278 | p->next_nonsignal_event_id = KFD_FIRST_NONSIGNAL_EVENT_ID; | |
279 | return 0; | |
280 | } | |
281 | ||
282 | static struct kfd_event *lookup_event_by_page_slot(struct kfd_process *p, | |
283 | struct signal_page *page, | |
284 | unsigned int signal_slot) | |
285 | { | |
286 | return lookup_event_by_id(p, make_signal_event_id(page, signal_slot)); | |
287 | } | |
288 | ||
289 | static int create_signal_event(struct file *devkfd, | |
290 | struct kfd_process *p, | |
291 | struct kfd_event *ev) | |
292 | { | |
293 | if (p->signal_event_count == KFD_SIGNAL_EVENT_LIMIT) { | |
294 | pr_warn("amdkfd: Signal event wasn't created because limit was reached\n"); | |
295 | return -ENOMEM; | |
296 | } | |
297 | ||
298 | if (!allocate_event_notification_slot(devkfd, p, &ev->signal_page, | |
299 | &ev->signal_slot_index)) { | |
300 | pr_warn("amdkfd: Signal event wasn't created because out of kernel memory\n"); | |
301 | return -ENOMEM; | |
302 | } | |
303 | ||
304 | p->signal_event_count++; | |
305 | ||
306 | ev->user_signal_address = | |
307 | &ev->signal_page->user_address[ev->signal_slot_index]; | |
308 | ||
309 | ev->event_id = make_signal_event_id(ev->signal_page, | |
310 | ev->signal_slot_index); | |
311 | ||
312 | pr_debug("signal event number %zu created with id %d, address %p\n", | |
313 | p->signal_event_count, ev->event_id, | |
314 | ev->user_signal_address); | |
315 | ||
6235e15e OG |
316 | pr_debug("signal event number %zu created with id %d, address %p\n", |
317 | p->signal_event_count, ev->event_id, | |
318 | ev->user_signal_address); | |
319 | ||
f3a39818 AL |
320 | return 0; |
321 | } | |
322 | ||
323 | /* | |
324 | * No non-signal events are supported yet. | |
325 | * We create them as events that never signal. | |
326 | * Set event calls from user-mode are failed. | |
327 | */ | |
328 | static int create_other_event(struct kfd_process *p, struct kfd_event *ev) | |
329 | { | |
330 | ev->event_id = make_nonsignal_event_id(p); | |
331 | if (ev->event_id == 0) | |
332 | return -ENOMEM; | |
333 | ||
334 | return 0; | |
335 | } | |
336 | ||
337 | void kfd_event_init_process(struct kfd_process *p) | |
338 | { | |
339 | mutex_init(&p->event_mutex); | |
340 | hash_init(p->events); | |
341 | INIT_LIST_HEAD(&p->signal_event_pages); | |
342 | p->next_nonsignal_event_id = KFD_FIRST_NONSIGNAL_EVENT_ID; | |
343 | p->signal_event_count = 0; | |
344 | } | |
345 | ||
346 | static void destroy_event(struct kfd_process *p, struct kfd_event *ev) | |
347 | { | |
348 | if (ev->signal_page != NULL) { | |
349 | release_event_notification_slot(ev->signal_page, | |
350 | ev->signal_slot_index); | |
351 | p->signal_event_count--; | |
352 | } | |
353 | ||
354 | /* | |
355 | * Abandon the list of waiters. Individual waiting threads will | |
356 | * clean up their own data. | |
357 | */ | |
358 | list_del(&ev->waiters); | |
359 | ||
360 | hash_del(&ev->events); | |
361 | kfree(ev); | |
362 | } | |
363 | ||
364 | static void destroy_events(struct kfd_process *p) | |
365 | { | |
366 | struct kfd_event *ev; | |
367 | struct hlist_node *tmp; | |
368 | unsigned int hash_bkt; | |
369 | ||
370 | hash_for_each_safe(p->events, hash_bkt, tmp, ev, events) | |
371 | destroy_event(p, ev); | |
372 | } | |
373 | ||
374 | /* | |
375 | * We assume that the process is being destroyed and there is no need to | |
376 | * unmap the pages or keep bookkeeping data in order. | |
377 | */ | |
378 | static void shutdown_signal_pages(struct kfd_process *p) | |
379 | { | |
380 | struct signal_page *page, *tmp; | |
381 | ||
382 | list_for_each_entry_safe(page, tmp, &p->signal_event_pages, | |
383 | event_pages) { | |
384 | free_pages((unsigned long)page->kernel_address, | |
385 | get_order(KFD_SIGNAL_EVENT_LIMIT * 8)); | |
386 | kfree(page); | |
387 | } | |
388 | } | |
389 | ||
390 | void kfd_event_free_process(struct kfd_process *p) | |
391 | { | |
392 | destroy_events(p); | |
393 | shutdown_signal_pages(p); | |
394 | } | |
395 | ||
396 | static bool event_can_be_gpu_signaled(const struct kfd_event *ev) | |
397 | { | |
398 | return ev->type == KFD_EVENT_TYPE_SIGNAL || | |
399 | ev->type == KFD_EVENT_TYPE_DEBUG; | |
400 | } | |
401 | ||
402 | static bool event_can_be_cpu_signaled(const struct kfd_event *ev) | |
403 | { | |
404 | return ev->type == KFD_EVENT_TYPE_SIGNAL; | |
405 | } | |
406 | ||
407 | int kfd_event_create(struct file *devkfd, struct kfd_process *p, | |
408 | uint32_t event_type, bool auto_reset, uint32_t node_id, | |
409 | uint32_t *event_id, uint32_t *event_trigger_data, | |
410 | uint64_t *event_page_offset, uint32_t *event_slot_index) | |
411 | { | |
412 | int ret = 0; | |
413 | struct kfd_event *ev = kzalloc(sizeof(*ev), GFP_KERNEL); | |
414 | ||
415 | if (!ev) | |
416 | return -ENOMEM; | |
417 | ||
418 | ev->type = event_type; | |
419 | ev->auto_reset = auto_reset; | |
420 | ev->signaled = false; | |
421 | ||
422 | INIT_LIST_HEAD(&ev->waiters); | |
423 | ||
424 | *event_page_offset = 0; | |
425 | ||
426 | mutex_lock(&p->event_mutex); | |
427 | ||
428 | switch (event_type) { | |
429 | case KFD_EVENT_TYPE_SIGNAL: | |
430 | case KFD_EVENT_TYPE_DEBUG: | |
431 | ret = create_signal_event(devkfd, p, ev); | |
432 | if (!ret) { | |
433 | *event_page_offset = (ev->signal_page->page_index | | |
434 | KFD_MMAP_EVENTS_MASK); | |
435 | *event_page_offset <<= PAGE_SHIFT; | |
436 | *event_slot_index = ev->signal_slot_index; | |
437 | } | |
438 | break; | |
439 | default: | |
440 | ret = create_other_event(p, ev); | |
441 | break; | |
442 | } | |
443 | ||
444 | if (!ret) { | |
445 | hash_add(p->events, &ev->events, ev->event_id); | |
446 | ||
447 | *event_id = ev->event_id; | |
448 | *event_trigger_data = ev->event_id; | |
449 | } else { | |
450 | kfree(ev); | |
451 | } | |
452 | ||
453 | mutex_unlock(&p->event_mutex); | |
454 | ||
455 | return ret; | |
456 | } | |
457 | ||
458 | /* Assumes that p is current. */ | |
459 | int kfd_event_destroy(struct kfd_process *p, uint32_t event_id) | |
460 | { | |
461 | struct kfd_event *ev; | |
462 | int ret = 0; | |
463 | ||
464 | mutex_lock(&p->event_mutex); | |
465 | ||
466 | ev = lookup_event_by_id(p, event_id); | |
467 | ||
468 | if (ev) | |
469 | destroy_event(p, ev); | |
470 | else | |
471 | ret = -EINVAL; | |
472 | ||
473 | mutex_unlock(&p->event_mutex); | |
474 | return ret; | |
475 | } | |
476 | ||
477 | static void set_event(struct kfd_event *ev) | |
478 | { | |
479 | struct kfd_event_waiter *waiter; | |
480 | struct kfd_event_waiter *next; | |
481 | ||
482 | /* Auto reset if the list is non-empty and we're waking someone. */ | |
483 | ev->signaled = !ev->auto_reset || list_empty(&ev->waiters); | |
484 | ||
485 | list_for_each_entry_safe(waiter, next, &ev->waiters, waiters) { | |
486 | waiter->activated = true; | |
487 | ||
488 | /* _init because free_waiters will call list_del */ | |
489 | list_del_init(&waiter->waiters); | |
490 | ||
491 | wake_up_process(waiter->sleeping_task); | |
492 | } | |
493 | } | |
494 | ||
495 | /* Assumes that p is current. */ | |
496 | int kfd_set_event(struct kfd_process *p, uint32_t event_id) | |
497 | { | |
498 | int ret = 0; | |
499 | struct kfd_event *ev; | |
500 | ||
501 | mutex_lock(&p->event_mutex); | |
502 | ||
503 | ev = lookup_event_by_id(p, event_id); | |
504 | ||
505 | if (ev && event_can_be_cpu_signaled(ev)) | |
506 | set_event(ev); | |
507 | else | |
508 | ret = -EINVAL; | |
509 | ||
510 | mutex_unlock(&p->event_mutex); | |
511 | return ret; | |
512 | } | |
513 | ||
514 | static void reset_event(struct kfd_event *ev) | |
515 | { | |
516 | ev->signaled = false; | |
517 | } | |
518 | ||
519 | /* Assumes that p is current. */ | |
520 | int kfd_reset_event(struct kfd_process *p, uint32_t event_id) | |
521 | { | |
522 | int ret = 0; | |
523 | struct kfd_event *ev; | |
524 | ||
525 | mutex_lock(&p->event_mutex); | |
526 | ||
527 | ev = lookup_event_by_id(p, event_id); | |
528 | ||
529 | if (ev && event_can_be_cpu_signaled(ev)) | |
530 | reset_event(ev); | |
531 | else | |
532 | ret = -EINVAL; | |
533 | ||
534 | mutex_unlock(&p->event_mutex); | |
535 | return ret; | |
536 | ||
537 | } | |
538 | ||
539 | static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev) | |
540 | { | |
541 | page_slots(ev->signal_page)[ev->signal_slot_index] = | |
542 | UNSIGNALED_EVENT_SLOT; | |
543 | } | |
544 | ||
545 | static bool is_slot_signaled(struct signal_page *page, unsigned int index) | |
546 | { | |
547 | return page_slots(page)[index] != UNSIGNALED_EVENT_SLOT; | |
548 | } | |
549 | ||
550 | static void set_event_from_interrupt(struct kfd_process *p, | |
551 | struct kfd_event *ev) | |
552 | { | |
553 | if (ev && event_can_be_gpu_signaled(ev)) { | |
554 | acknowledge_signal(p, ev); | |
555 | set_event(ev); | |
556 | } | |
557 | } | |
558 | ||
559 | void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id, | |
560 | uint32_t valid_id_bits) | |
561 | { | |
562 | struct kfd_event *ev; | |
563 | ||
564 | /* | |
565 | * Because we are called from arbitrary context (workqueue) as opposed | |
566 | * to process context, kfd_process could attempt to exit while we are | |
567 | * running so the lookup function returns a locked process. | |
568 | */ | |
569 | struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); | |
570 | ||
571 | if (!p) | |
572 | return; /* Presumably process exited. */ | |
573 | ||
574 | mutex_lock(&p->event_mutex); | |
575 | ||
576 | if (valid_id_bits >= INTERRUPT_DATA_BITS) { | |
577 | /* Partial ID is a full ID. */ | |
578 | ev = lookup_event_by_id(p, partial_id); | |
579 | set_event_from_interrupt(p, ev); | |
580 | } else { | |
581 | /* | |
582 | * Partial ID is in fact partial. For now we completely | |
583 | * ignore it, but we could use any bits we did receive to | |
584 | * search faster. | |
585 | */ | |
586 | struct signal_page *page; | |
587 | unsigned i; | |
588 | ||
589 | list_for_each_entry(page, &p->signal_event_pages, event_pages) | |
590 | for (i = 0; i < SLOTS_PER_PAGE; i++) | |
591 | if (is_slot_signaled(page, i)) { | |
592 | ev = lookup_event_by_page_slot(p, | |
593 | page, i); | |
594 | set_event_from_interrupt(p, ev); | |
595 | } | |
596 | } | |
597 | ||
598 | mutex_unlock(&p->event_mutex); | |
599 | mutex_unlock(&p->mutex); | |
600 | } | |
601 | ||
602 | static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events) | |
603 | { | |
604 | struct kfd_event_waiter *event_waiters; | |
605 | uint32_t i; | |
606 | ||
607 | event_waiters = kmalloc_array(num_events, | |
608 | sizeof(struct kfd_event_waiter), | |
609 | GFP_KERNEL); | |
610 | ||
611 | for (i = 0; (event_waiters) && (i < num_events) ; i++) { | |
612 | INIT_LIST_HEAD(&event_waiters[i].waiters); | |
613 | event_waiters[i].sleeping_task = current; | |
614 | event_waiters[i].activated = false; | |
615 | } | |
616 | ||
617 | return event_waiters; | |
618 | } | |
619 | ||
620 | static int init_event_waiter(struct kfd_process *p, | |
59d3e8be AS |
621 | struct kfd_event_waiter *waiter, |
622 | uint32_t event_id, | |
623 | uint32_t input_index) | |
f3a39818 AL |
624 | { |
625 | struct kfd_event *ev = lookup_event_by_id(p, event_id); | |
626 | ||
627 | if (!ev) | |
628 | return -EINVAL; | |
629 | ||
59d3e8be AS |
630 | waiter->event = ev; |
631 | waiter->input_index = input_index; | |
f3a39818 AL |
632 | waiter->activated = ev->signaled; |
633 | ev->signaled = ev->signaled && !ev->auto_reset; | |
634 | ||
635 | list_add(&waiter->waiters, &ev->waiters); | |
636 | ||
637 | return 0; | |
638 | } | |
639 | ||
640 | static bool test_event_condition(bool all, uint32_t num_events, | |
641 | struct kfd_event_waiter *event_waiters) | |
642 | { | |
643 | uint32_t i; | |
644 | uint32_t activated_count = 0; | |
645 | ||
646 | for (i = 0; i < num_events; i++) { | |
647 | if (event_waiters[i].activated) { | |
648 | if (!all) | |
649 | return true; | |
650 | ||
651 | activated_count++; | |
652 | } | |
653 | } | |
654 | ||
655 | return activated_count == num_events; | |
656 | } | |
657 | ||
59d3e8be AS |
658 | /* |
659 | * Copy event specific data, if defined. | |
660 | * Currently only memory exception events have additional data to copy to user | |
661 | */ | |
662 | static bool copy_signaled_event_data(uint32_t num_events, | |
663 | struct kfd_event_waiter *event_waiters, | |
664 | struct kfd_event_data __user *data) | |
665 | { | |
666 | struct kfd_hsa_memory_exception_data *src; | |
667 | struct kfd_hsa_memory_exception_data __user *dst; | |
668 | struct kfd_event_waiter *waiter; | |
669 | struct kfd_event *event; | |
670 | uint32_t i; | |
671 | ||
672 | for (i = 0; i < num_events; i++) { | |
673 | waiter = &event_waiters[i]; | |
674 | event = waiter->event; | |
675 | if (waiter->activated && event->type == KFD_EVENT_TYPE_MEMORY) { | |
676 | dst = &data[waiter->input_index].memory_exception_data; | |
677 | src = &event->memory_exception_data; | |
678 | if (copy_to_user(dst, src, | |
679 | sizeof(struct kfd_hsa_memory_exception_data))) | |
680 | return false; | |
681 | } | |
682 | } | |
683 | ||
684 | return true; | |
685 | ||
686 | } | |
687 | ||
688 | ||
689 | ||
f3a39818 AL |
690 | static long user_timeout_to_jiffies(uint32_t user_timeout_ms) |
691 | { | |
692 | if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE) | |
693 | return 0; | |
694 | ||
695 | if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE) | |
696 | return MAX_SCHEDULE_TIMEOUT; | |
697 | ||
698 | /* | |
699 | * msecs_to_jiffies interprets all values above 2^31-1 as infinite, | |
700 | * but we consider them finite. | |
701 | * This hack is wrong, but nobody is likely to notice. | |
702 | */ | |
703 | user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF); | |
704 | ||
705 | return msecs_to_jiffies(user_timeout_ms) + 1; | |
706 | } | |
707 | ||
708 | static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters) | |
709 | { | |
710 | uint32_t i; | |
711 | ||
712 | for (i = 0; i < num_events; i++) | |
713 | list_del(&waiters[i].waiters); | |
714 | ||
715 | kfree(waiters); | |
716 | } | |
717 | ||
718 | int kfd_wait_on_events(struct kfd_process *p, | |
59d3e8be | 719 | uint32_t num_events, void __user *data, |
f3a39818 AL |
720 | bool all, uint32_t user_timeout_ms, |
721 | enum kfd_event_wait_result *wait_result) | |
722 | { | |
59d3e8be AS |
723 | struct kfd_event_data __user *events = |
724 | (struct kfd_event_data __user *) data; | |
f3a39818 AL |
725 | uint32_t i; |
726 | int ret = 0; | |
727 | struct kfd_event_waiter *event_waiters = NULL; | |
728 | long timeout = user_timeout_to_jiffies(user_timeout_ms); | |
729 | ||
730 | mutex_lock(&p->event_mutex); | |
731 | ||
732 | event_waiters = alloc_event_waiters(num_events); | |
733 | if (!event_waiters) { | |
734 | ret = -ENOMEM; | |
735 | goto fail; | |
736 | } | |
737 | ||
738 | for (i = 0; i < num_events; i++) { | |
59d3e8be | 739 | struct kfd_event_data event_data; |
f3a39818 | 740 | |
59d3e8be AS |
741 | if (copy_from_user(&event_data, &events[i], |
742 | sizeof(struct kfd_event_data))) | |
f3a39818 AL |
743 | goto fail; |
744 | ||
59d3e8be AS |
745 | ret = init_event_waiter(p, &event_waiters[i], |
746 | event_data.event_id, i); | |
f3a39818 AL |
747 | if (ret) |
748 | goto fail; | |
749 | } | |
750 | ||
751 | mutex_unlock(&p->event_mutex); | |
752 | ||
753 | while (true) { | |
754 | if (fatal_signal_pending(current)) { | |
755 | ret = -EINTR; | |
756 | break; | |
757 | } | |
758 | ||
759 | if (signal_pending(current)) { | |
760 | /* | |
761 | * This is wrong when a nonzero, non-infinite timeout | |
762 | * is specified. We need to use | |
763 | * ERESTARTSYS_RESTARTBLOCK, but struct restart_block | |
764 | * contains a union with data for each user and it's | |
765 | * in generic kernel code that I don't want to | |
766 | * touch yet. | |
767 | */ | |
768 | ret = -ERESTARTSYS; | |
769 | break; | |
770 | } | |
771 | ||
772 | if (test_event_condition(all, num_events, event_waiters)) { | |
59d3e8be AS |
773 | if (copy_signaled_event_data(num_events, |
774 | event_waiters, events)) | |
775 | *wait_result = KFD_WAIT_COMPLETE; | |
776 | else | |
777 | *wait_result = KFD_WAIT_ERROR; | |
f3a39818 AL |
778 | break; |
779 | } | |
780 | ||
781 | if (timeout <= 0) { | |
782 | *wait_result = KFD_WAIT_TIMEOUT; | |
783 | break; | |
784 | } | |
785 | ||
786 | timeout = schedule_timeout_interruptible(timeout); | |
787 | } | |
788 | __set_current_state(TASK_RUNNING); | |
789 | ||
790 | mutex_lock(&p->event_mutex); | |
791 | free_waiters(num_events, event_waiters); | |
792 | mutex_unlock(&p->event_mutex); | |
793 | ||
794 | return ret; | |
795 | ||
796 | fail: | |
797 | if (event_waiters) | |
798 | free_waiters(num_events, event_waiters); | |
799 | ||
800 | mutex_unlock(&p->event_mutex); | |
801 | ||
802 | *wait_result = KFD_WAIT_ERROR; | |
803 | ||
804 | return ret; | |
805 | } | |
806 | ||
807 | int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma) | |
808 | { | |
809 | ||
810 | unsigned int page_index; | |
811 | unsigned long pfn; | |
812 | struct signal_page *page; | |
813 | ||
814 | /* check required size is logical */ | |
815 | if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) != | |
816 | get_order(vma->vm_end - vma->vm_start)) { | |
817 | pr_err("amdkfd: event page mmap requested illegal size\n"); | |
818 | return -EINVAL; | |
819 | } | |
820 | ||
821 | page_index = vma->vm_pgoff; | |
822 | ||
823 | page = lookup_signal_page_by_index(p, page_index); | |
824 | if (!page) { | |
825 | /* Probably KFD bug, but mmap is user-accessible. */ | |
826 | pr_debug("signal page could not be found for page_index %u\n", | |
827 | page_index); | |
828 | return -EINVAL; | |
829 | } | |
830 | ||
831 | pfn = __pa(page->kernel_address); | |
832 | pfn >>= PAGE_SHIFT; | |
833 | ||
834 | vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE | |
835 | | VM_DONTDUMP | VM_PFNMAP; | |
836 | ||
837 | pr_debug("mapping signal page\n"); | |
838 | pr_debug(" start user address == 0x%08lx\n", vma->vm_start); | |
839 | pr_debug(" end user address == 0x%08lx\n", vma->vm_end); | |
840 | pr_debug(" pfn == 0x%016lX\n", pfn); | |
841 | pr_debug(" vm_flags == 0x%08lX\n", vma->vm_flags); | |
842 | pr_debug(" size == 0x%08lX\n", | |
843 | vma->vm_end - vma->vm_start); | |
844 | ||
845 | page->user_address = (uint64_t __user *)vma->vm_start; | |
846 | ||
847 | /* mapping the page to user process */ | |
848 | return remap_pfn_range(vma, vma->vm_start, pfn, | |
849 | vma->vm_end - vma->vm_start, vma->vm_page_prot); | |
850 | } | |
59d3e8be AS |
851 | |
852 | /* | |
853 | * Assumes that p->event_mutex is held and of course | |
854 | * that p is not going away (current or locked). | |
855 | */ | |
856 | static void lookup_events_by_type_and_signal(struct kfd_process *p, | |
857 | int type, void *event_data) | |
858 | { | |
859 | struct kfd_hsa_memory_exception_data *ev_data; | |
860 | struct kfd_event *ev; | |
861 | int bkt; | |
862 | bool send_signal = true; | |
863 | ||
864 | ev_data = (struct kfd_hsa_memory_exception_data *) event_data; | |
865 | ||
866 | hash_for_each(p->events, bkt, ev, events) | |
867 | if (ev->type == type) { | |
868 | send_signal = false; | |
869 | dev_dbg(kfd_device, | |
870 | "Event found: id %X type %d", | |
871 | ev->event_id, ev->type); | |
872 | set_event(ev); | |
873 | if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data) | |
874 | ev->memory_exception_data = *ev_data; | |
875 | } | |
876 | ||
877 | /* Send SIGTERM no event of type "type" has been found*/ | |
878 | if (send_signal) { | |
81663016 OG |
879 | if (send_sigterm) { |
880 | dev_warn(kfd_device, | |
881 | "Sending SIGTERM to HSA Process with PID %d ", | |
882 | p->lead_thread->pid); | |
883 | send_sig(SIGTERM, p->lead_thread, 0); | |
884 | } else { | |
885 | dev_err(kfd_device, | |
886 | "HSA Process (PID %d) got unhandled exception", | |
59d3e8be | 887 | p->lead_thread->pid); |
81663016 | 888 | } |
59d3e8be AS |
889 | } |
890 | } | |
891 | ||
892 | void kfd_signal_iommu_event(struct kfd_dev *dev, unsigned int pasid, | |
893 | unsigned long address, bool is_write_requested, | |
894 | bool is_execute_requested) | |
895 | { | |
896 | struct kfd_hsa_memory_exception_data memory_exception_data; | |
897 | struct vm_area_struct *vma; | |
898 | ||
899 | /* | |
900 | * Because we are called from arbitrary context (workqueue) as opposed | |
901 | * to process context, kfd_process could attempt to exit while we are | |
902 | * running so the lookup function returns a locked process. | |
903 | */ | |
904 | struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); | |
905 | ||
906 | if (!p) | |
907 | return; /* Presumably process exited. */ | |
908 | ||
909 | memset(&memory_exception_data, 0, sizeof(memory_exception_data)); | |
910 | ||
911 | down_read(&p->mm->mmap_sem); | |
912 | vma = find_vma(p->mm, address); | |
913 | ||
914 | memory_exception_data.gpu_id = dev->id; | |
915 | memory_exception_data.va = address; | |
916 | /* Set failure reason */ | |
917 | memory_exception_data.failure.NotPresent = 1; | |
918 | memory_exception_data.failure.NoExecute = 0; | |
919 | memory_exception_data.failure.ReadOnly = 0; | |
920 | if (vma) { | |
921 | if (vma->vm_start > address) { | |
922 | memory_exception_data.failure.NotPresent = 1; | |
923 | memory_exception_data.failure.NoExecute = 0; | |
924 | memory_exception_data.failure.ReadOnly = 0; | |
925 | } else { | |
926 | memory_exception_data.failure.NotPresent = 0; | |
927 | if (is_write_requested && !(vma->vm_flags & VM_WRITE)) | |
928 | memory_exception_data.failure.ReadOnly = 1; | |
929 | else | |
930 | memory_exception_data.failure.ReadOnly = 0; | |
931 | if (is_execute_requested && !(vma->vm_flags & VM_EXEC)) | |
932 | memory_exception_data.failure.NoExecute = 1; | |
933 | else | |
934 | memory_exception_data.failure.NoExecute = 0; | |
935 | } | |
936 | } | |
937 | ||
938 | up_read(&p->mm->mmap_sem); | |
939 | ||
940 | mutex_lock(&p->event_mutex); | |
941 | ||
942 | /* Lookup events by type and signal them */ | |
943 | lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY, | |
944 | &memory_exception_data); | |
945 | ||
946 | mutex_unlock(&p->event_mutex); | |
947 | mutex_unlock(&p->mutex); | |
948 | } | |
930c5ff4 AS |
949 | |
950 | void kfd_signal_hw_exception_event(unsigned int pasid) | |
951 | { | |
952 | /* | |
953 | * Because we are called from arbitrary context (workqueue) as opposed | |
954 | * to process context, kfd_process could attempt to exit while we are | |
955 | * running so the lookup function returns a locked process. | |
956 | */ | |
957 | struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); | |
958 | ||
959 | if (!p) | |
960 | return; /* Presumably process exited. */ | |
961 | ||
962 | mutex_lock(&p->event_mutex); | |
963 | ||
964 | /* Lookup events by type and signal them */ | |
965 | lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL); | |
966 | ||
967 | mutex_unlock(&p->event_mutex); | |
968 | mutex_unlock(&p->mutex); | |
969 | } |