Loading...
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/signal.h>
27#include <linux/sched/mm.h>
28#include <linux/uaccess.h>
29#include <linux/mman.h>
30#include <linux/memory.h>
31#include "kfd_priv.h"
32#include "kfd_events.h"
33#include "kfd_iommu.h"
34#include <linux/device.h>
35
36/*
37 * Wrapper around wait_queue_entry_t
38 */
39struct kfd_event_waiter {
40 wait_queue_entry_t wait;
41 struct kfd_event *event; /* Event to wait for */
42 bool activated; /* Becomes true when event is signaled */
43};
44
45/*
46 * Each signal event needs a 64-bit signal slot where the signaler will write
47 * a 1 before sending an interrupt. (This is needed because some interrupts
48 * do not contain enough spare data bits to identify an event.)
49 * We get whole pages and map them to the process VA.
50 * Individual signal events use their event_id as slot index.
51 */
52struct kfd_signal_page {
53 uint64_t *kernel_address;
54 uint64_t __user *user_address;
55 bool need_to_free_pages;
56};
57
58
59static uint64_t *page_slots(struct kfd_signal_page *page)
60{
61 return page->kernel_address;
62}
63
64static struct kfd_signal_page *allocate_signal_page(struct kfd_process *p)
65{
66 void *backing_store;
67 struct kfd_signal_page *page;
68
69 page = kzalloc(sizeof(*page), GFP_KERNEL);
70 if (!page)
71 return NULL;
72
73 backing_store = (void *) __get_free_pages(GFP_KERNEL,
74 get_order(KFD_SIGNAL_EVENT_LIMIT * 8));
75 if (!backing_store)
76 goto fail_alloc_signal_store;
77
78 /* Initialize all events to unsignaled */
79 memset(backing_store, (uint8_t) UNSIGNALED_EVENT_SLOT,
80 KFD_SIGNAL_EVENT_LIMIT * 8);
81
82 page->kernel_address = backing_store;
83 page->need_to_free_pages = true;
84 pr_debug("Allocated new event signal page at %p, for process %p\n",
85 page, p);
86
87 return page;
88
89fail_alloc_signal_store:
90 kfree(page);
91 return NULL;
92}
93
94static int allocate_event_notification_slot(struct kfd_process *p,
95 struct kfd_event *ev)
96{
97 int id;
98
99 if (!p->signal_page) {
100 p->signal_page = allocate_signal_page(p);
101 if (!p->signal_page)
102 return -ENOMEM;
103 /* Oldest user mode expects 256 event slots */
104 p->signal_mapped_size = 256*8;
105 }
106
107 /*
108 * Compatibility with old user mode: Only use signal slots
109 * user mode has mapped, may be less than
110 * KFD_SIGNAL_EVENT_LIMIT. This also allows future increase
111 * of the event limit without breaking user mode.
112 */
113 id = idr_alloc(&p->event_idr, ev, 0, p->signal_mapped_size / 8,
114 GFP_KERNEL);
115 if (id < 0)
116 return id;
117
118 ev->event_id = id;
119 page_slots(p->signal_page)[id] = UNSIGNALED_EVENT_SLOT;
120
121 return 0;
122}
123
124/*
125 * Assumes that p->event_mutex is held and of course that p is not going
126 * away (current or locked).
127 */
128static struct kfd_event *lookup_event_by_id(struct kfd_process *p, uint32_t id)
129{
130 return idr_find(&p->event_idr, id);
131}
132
133/**
134 * lookup_signaled_event_by_partial_id - Lookup signaled event from partial ID
135 * @p: Pointer to struct kfd_process
136 * @id: ID to look up
137 * @bits: Number of valid bits in @id
138 *
139 * Finds the first signaled event with a matching partial ID. If no
140 * matching signaled event is found, returns NULL. In that case the
141 * caller should assume that the partial ID is invalid and do an
142 * exhaustive search of all siglaned events.
143 *
144 * If multiple events with the same partial ID signal at the same
145 * time, they will be found one interrupt at a time, not necessarily
146 * in the same order the interrupts occurred. As long as the number of
147 * interrupts is correct, all signaled events will be seen by the
148 * driver.
149 */
150static struct kfd_event *lookup_signaled_event_by_partial_id(
151 struct kfd_process *p, uint32_t id, uint32_t bits)
152{
153 struct kfd_event *ev;
154
155 if (!p->signal_page || id >= KFD_SIGNAL_EVENT_LIMIT)
156 return NULL;
157
158 /* Fast path for the common case that @id is not a partial ID
159 * and we only need a single lookup.
160 */
161 if (bits > 31 || (1U << bits) >= KFD_SIGNAL_EVENT_LIMIT) {
162 if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
163 return NULL;
164
165 return idr_find(&p->event_idr, id);
166 }
167
168 /* General case for partial IDs: Iterate over all matching IDs
169 * and find the first one that has signaled.
170 */
171 for (ev = NULL; id < KFD_SIGNAL_EVENT_LIMIT && !ev; id += 1U << bits) {
172 if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
173 continue;
174
175 ev = idr_find(&p->event_idr, id);
176 }
177
178 return ev;
179}
180
181static int create_signal_event(struct file *devkfd,
182 struct kfd_process *p,
183 struct kfd_event *ev)
184{
185 int ret;
186
187 if (p->signal_mapped_size &&
188 p->signal_event_count == p->signal_mapped_size / 8) {
189 if (!p->signal_event_limit_reached) {
190 pr_debug("Signal event wasn't created because limit was reached\n");
191 p->signal_event_limit_reached = true;
192 }
193 return -ENOSPC;
194 }
195
196 ret = allocate_event_notification_slot(p, ev);
197 if (ret) {
198 pr_warn("Signal event wasn't created because out of kernel memory\n");
199 return ret;
200 }
201
202 p->signal_event_count++;
203
204 ev->user_signal_address = &p->signal_page->user_address[ev->event_id];
205 pr_debug("Signal event number %zu created with id %d, address %p\n",
206 p->signal_event_count, ev->event_id,
207 ev->user_signal_address);
208
209 return 0;
210}
211
212static int create_other_event(struct kfd_process *p, struct kfd_event *ev)
213{
214 /* Cast KFD_LAST_NONSIGNAL_EVENT to uint32_t. This allows an
215 * intentional integer overflow to -1 without a compiler
216 * warning. idr_alloc treats a negative value as "maximum
217 * signed integer".
218 */
219 int id = idr_alloc(&p->event_idr, ev, KFD_FIRST_NONSIGNAL_EVENT_ID,
220 (uint32_t)KFD_LAST_NONSIGNAL_EVENT_ID + 1,
221 GFP_KERNEL);
222
223 if (id < 0)
224 return id;
225 ev->event_id = id;
226
227 return 0;
228}
229
230void kfd_event_init_process(struct kfd_process *p)
231{
232 mutex_init(&p->event_mutex);
233 idr_init(&p->event_idr);
234 p->signal_page = NULL;
235 p->signal_event_count = 0;
236}
237
238static void destroy_event(struct kfd_process *p, struct kfd_event *ev)
239{
240 struct kfd_event_waiter *waiter;
241
242 /* Wake up pending waiters. They will return failure */
243 list_for_each_entry(waiter, &ev->wq.head, wait.entry)
244 waiter->event = NULL;
245 wake_up_all(&ev->wq);
246
247 if (ev->type == KFD_EVENT_TYPE_SIGNAL ||
248 ev->type == KFD_EVENT_TYPE_DEBUG)
249 p->signal_event_count--;
250
251 idr_remove(&p->event_idr, ev->event_id);
252 kfree(ev);
253}
254
255static void destroy_events(struct kfd_process *p)
256{
257 struct kfd_event *ev;
258 uint32_t id;
259
260 idr_for_each_entry(&p->event_idr, ev, id)
261 destroy_event(p, ev);
262 idr_destroy(&p->event_idr);
263}
264
265/*
266 * We assume that the process is being destroyed and there is no need to
267 * unmap the pages or keep bookkeeping data in order.
268 */
269static void shutdown_signal_page(struct kfd_process *p)
270{
271 struct kfd_signal_page *page = p->signal_page;
272
273 if (page) {
274 if (page->need_to_free_pages)
275 free_pages((unsigned long)page->kernel_address,
276 get_order(KFD_SIGNAL_EVENT_LIMIT * 8));
277 kfree(page);
278 }
279}
280
281void kfd_event_free_process(struct kfd_process *p)
282{
283 destroy_events(p);
284 shutdown_signal_page(p);
285}
286
287static bool event_can_be_gpu_signaled(const struct kfd_event *ev)
288{
289 return ev->type == KFD_EVENT_TYPE_SIGNAL ||
290 ev->type == KFD_EVENT_TYPE_DEBUG;
291}
292
293static bool event_can_be_cpu_signaled(const struct kfd_event *ev)
294{
295 return ev->type == KFD_EVENT_TYPE_SIGNAL;
296}
297
298int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
299 uint64_t size)
300{
301 struct kfd_signal_page *page;
302
303 if (p->signal_page)
304 return -EBUSY;
305
306 page = kzalloc(sizeof(*page), GFP_KERNEL);
307 if (!page)
308 return -ENOMEM;
309
310 /* Initialize all events to unsignaled */
311 memset(kernel_address, (uint8_t) UNSIGNALED_EVENT_SLOT,
312 KFD_SIGNAL_EVENT_LIMIT * 8);
313
314 page->kernel_address = kernel_address;
315
316 p->signal_page = page;
317 p->signal_mapped_size = size;
318
319 return 0;
320}
321
322int kfd_event_create(struct file *devkfd, struct kfd_process *p,
323 uint32_t event_type, bool auto_reset, uint32_t node_id,
324 uint32_t *event_id, uint32_t *event_trigger_data,
325 uint64_t *event_page_offset, uint32_t *event_slot_index)
326{
327 int ret = 0;
328 struct kfd_event *ev = kzalloc(sizeof(*ev), GFP_KERNEL);
329
330 if (!ev)
331 return -ENOMEM;
332
333 ev->type = event_type;
334 ev->auto_reset = auto_reset;
335 ev->signaled = false;
336
337 init_waitqueue_head(&ev->wq);
338
339 *event_page_offset = 0;
340
341 mutex_lock(&p->event_mutex);
342
343 switch (event_type) {
344 case KFD_EVENT_TYPE_SIGNAL:
345 case KFD_EVENT_TYPE_DEBUG:
346 ret = create_signal_event(devkfd, p, ev);
347 if (!ret) {
348 *event_page_offset = KFD_MMAP_TYPE_EVENTS;
349 *event_slot_index = ev->event_id;
350 }
351 break;
352 default:
353 ret = create_other_event(p, ev);
354 break;
355 }
356
357 if (!ret) {
358 *event_id = ev->event_id;
359 *event_trigger_data = ev->event_id;
360 } else {
361 kfree(ev);
362 }
363
364 mutex_unlock(&p->event_mutex);
365
366 return ret;
367}
368
369/* Assumes that p is current. */
370int kfd_event_destroy(struct kfd_process *p, uint32_t event_id)
371{
372 struct kfd_event *ev;
373 int ret = 0;
374
375 mutex_lock(&p->event_mutex);
376
377 ev = lookup_event_by_id(p, event_id);
378
379 if (ev)
380 destroy_event(p, ev);
381 else
382 ret = -EINVAL;
383
384 mutex_unlock(&p->event_mutex);
385 return ret;
386}
387
388static void set_event(struct kfd_event *ev)
389{
390 struct kfd_event_waiter *waiter;
391
392 /* Auto reset if the list is non-empty and we're waking
393 * someone. waitqueue_active is safe here because we're
394 * protected by the p->event_mutex, which is also held when
395 * updating the wait queues in kfd_wait_on_events.
396 */
397 ev->signaled = !ev->auto_reset || !waitqueue_active(&ev->wq);
398
399 list_for_each_entry(waiter, &ev->wq.head, wait.entry)
400 waiter->activated = true;
401
402 wake_up_all(&ev->wq);
403}
404
405/* Assumes that p is current. */
406int kfd_set_event(struct kfd_process *p, uint32_t event_id)
407{
408 int ret = 0;
409 struct kfd_event *ev;
410
411 mutex_lock(&p->event_mutex);
412
413 ev = lookup_event_by_id(p, event_id);
414
415 if (ev && event_can_be_cpu_signaled(ev))
416 set_event(ev);
417 else
418 ret = -EINVAL;
419
420 mutex_unlock(&p->event_mutex);
421 return ret;
422}
423
424static void reset_event(struct kfd_event *ev)
425{
426 ev->signaled = false;
427}
428
429/* Assumes that p is current. */
430int kfd_reset_event(struct kfd_process *p, uint32_t event_id)
431{
432 int ret = 0;
433 struct kfd_event *ev;
434
435 mutex_lock(&p->event_mutex);
436
437 ev = lookup_event_by_id(p, event_id);
438
439 if (ev && event_can_be_cpu_signaled(ev))
440 reset_event(ev);
441 else
442 ret = -EINVAL;
443
444 mutex_unlock(&p->event_mutex);
445 return ret;
446
447}
448
449static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev)
450{
451 page_slots(p->signal_page)[ev->event_id] = UNSIGNALED_EVENT_SLOT;
452}
453
454static void set_event_from_interrupt(struct kfd_process *p,
455 struct kfd_event *ev)
456{
457 if (ev && event_can_be_gpu_signaled(ev)) {
458 acknowledge_signal(p, ev);
459 set_event(ev);
460 }
461}
462
463void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
464 uint32_t valid_id_bits)
465{
466 struct kfd_event *ev = NULL;
467
468 /*
469 * Because we are called from arbitrary context (workqueue) as opposed
470 * to process context, kfd_process could attempt to exit while we are
471 * running so the lookup function increments the process ref count.
472 */
473 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
474
475 if (!p)
476 return; /* Presumably process exited. */
477
478 mutex_lock(&p->event_mutex);
479
480 if (valid_id_bits)
481 ev = lookup_signaled_event_by_partial_id(p, partial_id,
482 valid_id_bits);
483 if (ev) {
484 set_event_from_interrupt(p, ev);
485 } else if (p->signal_page) {
486 /*
487 * Partial ID lookup failed. Assume that the event ID
488 * in the interrupt payload was invalid and do an
489 * exhaustive search of signaled events.
490 */
491 uint64_t *slots = page_slots(p->signal_page);
492 uint32_t id;
493
494 if (valid_id_bits)
495 pr_debug_ratelimited("Partial ID invalid: %u (%u valid bits)\n",
496 partial_id, valid_id_bits);
497
498 if (p->signal_event_count < KFD_SIGNAL_EVENT_LIMIT / 64) {
499 /* With relatively few events, it's faster to
500 * iterate over the event IDR
501 */
502 idr_for_each_entry(&p->event_idr, ev, id) {
503 if (id >= KFD_SIGNAL_EVENT_LIMIT)
504 break;
505
506 if (slots[id] != UNSIGNALED_EVENT_SLOT)
507 set_event_from_interrupt(p, ev);
508 }
509 } else {
510 /* With relatively many events, it's faster to
511 * iterate over the signal slots and lookup
512 * only signaled events from the IDR.
513 */
514 for (id = 0; id < KFD_SIGNAL_EVENT_LIMIT; id++)
515 if (slots[id] != UNSIGNALED_EVENT_SLOT) {
516 ev = lookup_event_by_id(p, id);
517 set_event_from_interrupt(p, ev);
518 }
519 }
520 }
521
522 mutex_unlock(&p->event_mutex);
523 kfd_unref_process(p);
524}
525
526static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events)
527{
528 struct kfd_event_waiter *event_waiters;
529 uint32_t i;
530
531 event_waiters = kmalloc_array(num_events,
532 sizeof(struct kfd_event_waiter),
533 GFP_KERNEL);
534
535 for (i = 0; (event_waiters) && (i < num_events) ; i++) {
536 init_wait(&event_waiters[i].wait);
537 event_waiters[i].activated = false;
538 }
539
540 return event_waiters;
541}
542
543static int init_event_waiter_get_status(struct kfd_process *p,
544 struct kfd_event_waiter *waiter,
545 uint32_t event_id)
546{
547 struct kfd_event *ev = lookup_event_by_id(p, event_id);
548
549 if (!ev)
550 return -EINVAL;
551
552 waiter->event = ev;
553 waiter->activated = ev->signaled;
554 ev->signaled = ev->signaled && !ev->auto_reset;
555
556 return 0;
557}
558
559static void init_event_waiter_add_to_waitlist(struct kfd_event_waiter *waiter)
560{
561 struct kfd_event *ev = waiter->event;
562
563 /* Only add to the wait list if we actually need to
564 * wait on this event.
565 */
566 if (!waiter->activated)
567 add_wait_queue(&ev->wq, &waiter->wait);
568}
569
570/* test_event_condition - Test condition of events being waited for
571 * @all: Return completion only if all events have signaled
572 * @num_events: Number of events to wait for
573 * @event_waiters: Array of event waiters, one per event
574 *
575 * Returns KFD_IOC_WAIT_RESULT_COMPLETE if all (or one) event(s) have
576 * signaled. Returns KFD_IOC_WAIT_RESULT_TIMEOUT if no (or not all)
577 * events have signaled. Returns KFD_IOC_WAIT_RESULT_FAIL if any of
578 * the events have been destroyed.
579 */
580static uint32_t test_event_condition(bool all, uint32_t num_events,
581 struct kfd_event_waiter *event_waiters)
582{
583 uint32_t i;
584 uint32_t activated_count = 0;
585
586 for (i = 0; i < num_events; i++) {
587 if (!event_waiters[i].event)
588 return KFD_IOC_WAIT_RESULT_FAIL;
589
590 if (event_waiters[i].activated) {
591 if (!all)
592 return KFD_IOC_WAIT_RESULT_COMPLETE;
593
594 activated_count++;
595 }
596 }
597
598 return activated_count == num_events ?
599 KFD_IOC_WAIT_RESULT_COMPLETE : KFD_IOC_WAIT_RESULT_TIMEOUT;
600}
601
602/*
603 * Copy event specific data, if defined.
604 * Currently only memory exception events have additional data to copy to user
605 */
606static int copy_signaled_event_data(uint32_t num_events,
607 struct kfd_event_waiter *event_waiters,
608 struct kfd_event_data __user *data)
609{
610 struct kfd_hsa_memory_exception_data *src;
611 struct kfd_hsa_memory_exception_data __user *dst;
612 struct kfd_event_waiter *waiter;
613 struct kfd_event *event;
614 uint32_t i;
615
616 for (i = 0; i < num_events; i++) {
617 waiter = &event_waiters[i];
618 event = waiter->event;
619 if (waiter->activated && event->type == KFD_EVENT_TYPE_MEMORY) {
620 dst = &data[i].memory_exception_data;
621 src = &event->memory_exception_data;
622 if (copy_to_user(dst, src,
623 sizeof(struct kfd_hsa_memory_exception_data)))
624 return -EFAULT;
625 }
626 }
627
628 return 0;
629
630}
631
632
633
634static long user_timeout_to_jiffies(uint32_t user_timeout_ms)
635{
636 if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE)
637 return 0;
638
639 if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE)
640 return MAX_SCHEDULE_TIMEOUT;
641
642 /*
643 * msecs_to_jiffies interprets all values above 2^31-1 as infinite,
644 * but we consider them finite.
645 * This hack is wrong, but nobody is likely to notice.
646 */
647 user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF);
648
649 return msecs_to_jiffies(user_timeout_ms) + 1;
650}
651
652static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters)
653{
654 uint32_t i;
655
656 for (i = 0; i < num_events; i++)
657 if (waiters[i].event)
658 remove_wait_queue(&waiters[i].event->wq,
659 &waiters[i].wait);
660
661 kfree(waiters);
662}
663
664int kfd_wait_on_events(struct kfd_process *p,
665 uint32_t num_events, void __user *data,
666 bool all, uint32_t user_timeout_ms,
667 uint32_t *wait_result)
668{
669 struct kfd_event_data __user *events =
670 (struct kfd_event_data __user *) data;
671 uint32_t i;
672 int ret = 0;
673
674 struct kfd_event_waiter *event_waiters = NULL;
675 long timeout = user_timeout_to_jiffies(user_timeout_ms);
676
677 event_waiters = alloc_event_waiters(num_events);
678 if (!event_waiters) {
679 ret = -ENOMEM;
680 goto out;
681 }
682
683 mutex_lock(&p->event_mutex);
684
685 for (i = 0; i < num_events; i++) {
686 struct kfd_event_data event_data;
687
688 if (copy_from_user(&event_data, &events[i],
689 sizeof(struct kfd_event_data))) {
690 ret = -EFAULT;
691 goto out_unlock;
692 }
693
694 ret = init_event_waiter_get_status(p, &event_waiters[i],
695 event_data.event_id);
696 if (ret)
697 goto out_unlock;
698 }
699
700 /* Check condition once. */
701 *wait_result = test_event_condition(all, num_events, event_waiters);
702 if (*wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) {
703 ret = copy_signaled_event_data(num_events,
704 event_waiters, events);
705 goto out_unlock;
706 } else if (WARN_ON(*wait_result == KFD_IOC_WAIT_RESULT_FAIL)) {
707 /* This should not happen. Events shouldn't be
708 * destroyed while we're holding the event_mutex
709 */
710 goto out_unlock;
711 }
712
713 /* Add to wait lists if we need to wait. */
714 for (i = 0; i < num_events; i++)
715 init_event_waiter_add_to_waitlist(&event_waiters[i]);
716
717 mutex_unlock(&p->event_mutex);
718
719 while (true) {
720 if (fatal_signal_pending(current)) {
721 ret = -EINTR;
722 break;
723 }
724
725 if (signal_pending(current)) {
726 /*
727 * This is wrong when a nonzero, non-infinite timeout
728 * is specified. We need to use
729 * ERESTARTSYS_RESTARTBLOCK, but struct restart_block
730 * contains a union with data for each user and it's
731 * in generic kernel code that I don't want to
732 * touch yet.
733 */
734 ret = -ERESTARTSYS;
735 break;
736 }
737
738 /* Set task state to interruptible sleep before
739 * checking wake-up conditions. A concurrent wake-up
740 * will put the task back into runnable state. In that
741 * case schedule_timeout will not put the task to
742 * sleep and we'll get a chance to re-check the
743 * updated conditions almost immediately. Otherwise,
744 * this race condition would lead to a soft hang or a
745 * very long sleep.
746 */
747 set_current_state(TASK_INTERRUPTIBLE);
748
749 *wait_result = test_event_condition(all, num_events,
750 event_waiters);
751 if (*wait_result != KFD_IOC_WAIT_RESULT_TIMEOUT)
752 break;
753
754 if (timeout <= 0)
755 break;
756
757 timeout = schedule_timeout(timeout);
758 }
759 __set_current_state(TASK_RUNNING);
760
761 /* copy_signaled_event_data may sleep. So this has to happen
762 * after the task state is set back to RUNNING.
763 */
764 if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE)
765 ret = copy_signaled_event_data(num_events,
766 event_waiters, events);
767
768 mutex_lock(&p->event_mutex);
769out_unlock:
770 free_waiters(num_events, event_waiters);
771 mutex_unlock(&p->event_mutex);
772out:
773 if (ret)
774 *wait_result = KFD_IOC_WAIT_RESULT_FAIL;
775 else if (*wait_result == KFD_IOC_WAIT_RESULT_FAIL)
776 ret = -EIO;
777
778 return ret;
779}
780
781int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma)
782{
783 unsigned long pfn;
784 struct kfd_signal_page *page;
785 int ret;
786
787 /* check required size doesn't exceed the allocated size */
788 if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) <
789 get_order(vma->vm_end - vma->vm_start)) {
790 pr_err("Event page mmap requested illegal size\n");
791 return -EINVAL;
792 }
793
794 page = p->signal_page;
795 if (!page) {
796 /* Probably KFD bug, but mmap is user-accessible. */
797 pr_debug("Signal page could not be found\n");
798 return -EINVAL;
799 }
800
801 pfn = __pa(page->kernel_address);
802 pfn >>= PAGE_SHIFT;
803
804 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE
805 | VM_DONTDUMP | VM_PFNMAP;
806
807 pr_debug("Mapping signal page\n");
808 pr_debug(" start user address == 0x%08lx\n", vma->vm_start);
809 pr_debug(" end user address == 0x%08lx\n", vma->vm_end);
810 pr_debug(" pfn == 0x%016lX\n", pfn);
811 pr_debug(" vm_flags == 0x%08lX\n", vma->vm_flags);
812 pr_debug(" size == 0x%08lX\n",
813 vma->vm_end - vma->vm_start);
814
815 page->user_address = (uint64_t __user *)vma->vm_start;
816
817 /* mapping the page to user process */
818 ret = remap_pfn_range(vma, vma->vm_start, pfn,
819 vma->vm_end - vma->vm_start, vma->vm_page_prot);
820 if (!ret)
821 p->signal_mapped_size = vma->vm_end - vma->vm_start;
822
823 return ret;
824}
825
826/*
827 * Assumes that p->event_mutex is held and of course
828 * that p is not going away (current or locked).
829 */
830static void lookup_events_by_type_and_signal(struct kfd_process *p,
831 int type, void *event_data)
832{
833 struct kfd_hsa_memory_exception_data *ev_data;
834 struct kfd_event *ev;
835 uint32_t id;
836 bool send_signal = true;
837
838 ev_data = (struct kfd_hsa_memory_exception_data *) event_data;
839
840 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
841 idr_for_each_entry_continue(&p->event_idr, ev, id)
842 if (ev->type == type) {
843 send_signal = false;
844 dev_dbg(kfd_device,
845 "Event found: id %X type %d",
846 ev->event_id, ev->type);
847 set_event(ev);
848 if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data)
849 ev->memory_exception_data = *ev_data;
850 }
851
852 if (type == KFD_EVENT_TYPE_MEMORY) {
853 dev_warn(kfd_device,
854 "Sending SIGSEGV to process %d (pasid 0x%x)",
855 p->lead_thread->pid, p->pasid);
856 send_sig(SIGSEGV, p->lead_thread, 0);
857 }
858
859 /* Send SIGTERM no event of type "type" has been found*/
860 if (send_signal) {
861 if (send_sigterm) {
862 dev_warn(kfd_device,
863 "Sending SIGTERM to process %d (pasid 0x%x)",
864 p->lead_thread->pid, p->pasid);
865 send_sig(SIGTERM, p->lead_thread, 0);
866 } else {
867 dev_err(kfd_device,
868 "Process %d (pasid 0x%x) got unhandled exception",
869 p->lead_thread->pid, p->pasid);
870 }
871 }
872}
873
874#ifdef KFD_SUPPORT_IOMMU_V2
875void kfd_signal_iommu_event(struct kfd_dev *dev, unsigned int pasid,
876 unsigned long address, bool is_write_requested,
877 bool is_execute_requested)
878{
879 struct kfd_hsa_memory_exception_data memory_exception_data;
880 struct vm_area_struct *vma;
881
882 /*
883 * Because we are called from arbitrary context (workqueue) as opposed
884 * to process context, kfd_process could attempt to exit while we are
885 * running so the lookup function increments the process ref count.
886 */
887 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
888 struct mm_struct *mm;
889
890 if (!p)
891 return; /* Presumably process exited. */
892
893 /* Take a safe reference to the mm_struct, which may otherwise
894 * disappear even while the kfd_process is still referenced.
895 */
896 mm = get_task_mm(p->lead_thread);
897 if (!mm) {
898 kfd_unref_process(p);
899 return; /* Process is exiting */
900 }
901
902 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
903
904 mmap_read_lock(mm);
905 vma = find_vma(mm, address);
906
907 memory_exception_data.gpu_id = dev->id;
908 memory_exception_data.va = address;
909 /* Set failure reason */
910 memory_exception_data.failure.NotPresent = 1;
911 memory_exception_data.failure.NoExecute = 0;
912 memory_exception_data.failure.ReadOnly = 0;
913 if (vma && address >= vma->vm_start) {
914 memory_exception_data.failure.NotPresent = 0;
915
916 if (is_write_requested && !(vma->vm_flags & VM_WRITE))
917 memory_exception_data.failure.ReadOnly = 1;
918 else
919 memory_exception_data.failure.ReadOnly = 0;
920
921 if (is_execute_requested && !(vma->vm_flags & VM_EXEC))
922 memory_exception_data.failure.NoExecute = 1;
923 else
924 memory_exception_data.failure.NoExecute = 0;
925 }
926
927 mmap_read_unlock(mm);
928 mmput(mm);
929
930 pr_debug("notpresent %d, noexecute %d, readonly %d\n",
931 memory_exception_data.failure.NotPresent,
932 memory_exception_data.failure.NoExecute,
933 memory_exception_data.failure.ReadOnly);
934
935 /* Workaround on Raven to not kill the process when memory is freed
936 * before IOMMU is able to finish processing all the excessive PPRs
937 */
938 if (dev->device_info->asic_family != CHIP_RAVEN &&
939 dev->device_info->asic_family != CHIP_RENOIR) {
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 }
948
949 kfd_unref_process(p);
950}
951#endif /* KFD_SUPPORT_IOMMU_V2 */
952
953void kfd_signal_hw_exception_event(unsigned int pasid)
954{
955 /*
956 * Because we are called from arbitrary context (workqueue) as opposed
957 * to process context, kfd_process could attempt to exit while we are
958 * running so the lookup function increments the process ref count.
959 */
960 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
961
962 if (!p)
963 return; /* Presumably process exited. */
964
965 mutex_lock(&p->event_mutex);
966
967 /* Lookup events by type and signal them */
968 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL);
969
970 mutex_unlock(&p->event_mutex);
971 kfd_unref_process(p);
972}
973
974void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid,
975 struct kfd_vm_fault_info *info)
976{
977 struct kfd_event *ev;
978 uint32_t id;
979 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
980 struct kfd_hsa_memory_exception_data memory_exception_data;
981
982 if (!p)
983 return; /* Presumably process exited. */
984 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
985 memory_exception_data.gpu_id = dev->id;
986 memory_exception_data.failure.imprecise = true;
987 /* Set failure reason */
988 if (info) {
989 memory_exception_data.va = (info->page_addr) << PAGE_SHIFT;
990 memory_exception_data.failure.NotPresent =
991 info->prot_valid ? 1 : 0;
992 memory_exception_data.failure.NoExecute =
993 info->prot_exec ? 1 : 0;
994 memory_exception_data.failure.ReadOnly =
995 info->prot_write ? 1 : 0;
996 memory_exception_data.failure.imprecise = 0;
997 }
998 mutex_lock(&p->event_mutex);
999
1000 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1001 idr_for_each_entry_continue(&p->event_idr, ev, id)
1002 if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1003 ev->memory_exception_data = memory_exception_data;
1004 set_event(ev);
1005 }
1006
1007 mutex_unlock(&p->event_mutex);
1008 kfd_unref_process(p);
1009}
1010
1011void kfd_signal_reset_event(struct kfd_dev *dev)
1012{
1013 struct kfd_hsa_hw_exception_data hw_exception_data;
1014 struct kfd_hsa_memory_exception_data memory_exception_data;
1015 struct kfd_process *p;
1016 struct kfd_event *ev;
1017 unsigned int temp;
1018 uint32_t id, idx;
1019 int reset_cause = atomic_read(&dev->sram_ecc_flag) ?
1020 KFD_HW_EXCEPTION_ECC :
1021 KFD_HW_EXCEPTION_GPU_HANG;
1022
1023 /* Whole gpu reset caused by GPU hang and memory is lost */
1024 memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1025 hw_exception_data.gpu_id = dev->id;
1026 hw_exception_data.memory_lost = 1;
1027 hw_exception_data.reset_cause = reset_cause;
1028
1029 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1030 memory_exception_data.ErrorType = KFD_MEM_ERR_SRAM_ECC;
1031 memory_exception_data.gpu_id = dev->id;
1032 memory_exception_data.failure.imprecise = true;
1033
1034 idx = srcu_read_lock(&kfd_processes_srcu);
1035 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1036 mutex_lock(&p->event_mutex);
1037 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1038 idr_for_each_entry_continue(&p->event_idr, ev, id) {
1039 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1040 ev->hw_exception_data = hw_exception_data;
1041 set_event(ev);
1042 }
1043 if (ev->type == KFD_EVENT_TYPE_MEMORY &&
1044 reset_cause == KFD_HW_EXCEPTION_ECC) {
1045 ev->memory_exception_data = memory_exception_data;
1046 set_event(ev);
1047 }
1048 }
1049 mutex_unlock(&p->event_mutex);
1050 }
1051 srcu_read_unlock(&kfd_processes_srcu, idx);
1052}
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"
33#include <linux/device.h>
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 */
43struct 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;
49
50 /* Event */
51 struct kfd_event *event;
52 uint32_t input_index;
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
65struct 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
88static uint64_t *page_slots(struct signal_page *page)
89{
90 return page->kernel_address;
91}
92
93static 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
129static 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
166fail_alloc_signal_store:
167 kfree(page);
168fail_alloc_signal_page:
169 return false;
170}
171
172static 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);
180 if (!ret) {
181 ret = allocate_signal_page(devkfd, p);
182 if (ret)
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. */
190static 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
200static 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 */
220static 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
231static 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 */
243static 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
282static 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
289static 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
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
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 */
328static 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
337void 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
346static 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
364static 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 */
378static 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
390void kfd_event_free_process(struct kfd_process *p)
391{
392 destroy_events(p);
393 shutdown_signal_pages(p);
394}
395
396static 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
402static bool event_can_be_cpu_signaled(const struct kfd_event *ev)
403{
404 return ev->type == KFD_EVENT_TYPE_SIGNAL;
405}
406
407int 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. */
459int 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
477static 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. */
496int 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
514static void reset_event(struct kfd_event *ev)
515{
516 ev->signaled = false;
517}
518
519/* Assumes that p is current. */
520int 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
539static 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
545static bool is_slot_signaled(struct signal_page *page, unsigned int index)
546{
547 return page_slots(page)[index] != UNSIGNALED_EVENT_SLOT;
548}
549
550static 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
559void 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
602static 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
620static int init_event_waiter(struct kfd_process *p,
621 struct kfd_event_waiter *waiter,
622 uint32_t event_id,
623 uint32_t input_index)
624{
625 struct kfd_event *ev = lookup_event_by_id(p, event_id);
626
627 if (!ev)
628 return -EINVAL;
629
630 waiter->event = ev;
631 waiter->input_index = input_index;
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
640static 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
658/*
659 * Copy event specific data, if defined.
660 * Currently only memory exception events have additional data to copy to user
661 */
662static 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
690static 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
708static 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
718int kfd_wait_on_events(struct kfd_process *p,
719 uint32_t num_events, void __user *data,
720 bool all, uint32_t user_timeout_ms,
721 enum kfd_event_wait_result *wait_result)
722{
723 struct kfd_event_data __user *events =
724 (struct kfd_event_data __user *) data;
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++) {
739 struct kfd_event_data event_data;
740
741 if (copy_from_user(&event_data, &events[i],
742 sizeof(struct kfd_event_data)))
743 goto fail;
744
745 ret = init_event_waiter(p, &event_waiters[i],
746 event_data.event_id, i);
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)) {
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;
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
796fail:
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
807int 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}
851
852/*
853 * Assumes that p->event_mutex is held and of course
854 * that p is not going away (current or locked).
855 */
856static 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) {
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",
887 p->lead_thread->pid);
888 }
889 }
890}
891
892void 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}
949
950void 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}