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_warn("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_page_offset <<= PAGE_SHIFT;
350 *event_slot_index = ev->event_id;
351 }
352 break;
353 default:
354 ret = create_other_event(p, ev);
355 break;
356 }
357
358 if (!ret) {
359 *event_id = ev->event_id;
360 *event_trigger_data = ev->event_id;
361 } else {
362 kfree(ev);
363 }
364
365 mutex_unlock(&p->event_mutex);
366
367 return ret;
368}
369
370/* Assumes that p is current. */
371int kfd_event_destroy(struct kfd_process *p, uint32_t event_id)
372{
373 struct kfd_event *ev;
374 int ret = 0;
375
376 mutex_lock(&p->event_mutex);
377
378 ev = lookup_event_by_id(p, event_id);
379
380 if (ev)
381 destroy_event(p, ev);
382 else
383 ret = -EINVAL;
384
385 mutex_unlock(&p->event_mutex);
386 return ret;
387}
388
389static void set_event(struct kfd_event *ev)
390{
391 struct kfd_event_waiter *waiter;
392
393 /* Auto reset if the list is non-empty and we're waking
394 * someone. waitqueue_active is safe here because we're
395 * protected by the p->event_mutex, which is also held when
396 * updating the wait queues in kfd_wait_on_events.
397 */
398 ev->signaled = !ev->auto_reset || !waitqueue_active(&ev->wq);
399
400 list_for_each_entry(waiter, &ev->wq.head, wait.entry)
401 waiter->activated = true;
402
403 wake_up_all(&ev->wq);
404}
405
406/* Assumes that p is current. */
407int kfd_set_event(struct kfd_process *p, uint32_t event_id)
408{
409 int ret = 0;
410 struct kfd_event *ev;
411
412 mutex_lock(&p->event_mutex);
413
414 ev = lookup_event_by_id(p, event_id);
415
416 if (ev && event_can_be_cpu_signaled(ev))
417 set_event(ev);
418 else
419 ret = -EINVAL;
420
421 mutex_unlock(&p->event_mutex);
422 return ret;
423}
424
425static void reset_event(struct kfd_event *ev)
426{
427 ev->signaled = false;
428}
429
430/* Assumes that p is current. */
431int kfd_reset_event(struct kfd_process *p, uint32_t event_id)
432{
433 int ret = 0;
434 struct kfd_event *ev;
435
436 mutex_lock(&p->event_mutex);
437
438 ev = lookup_event_by_id(p, event_id);
439
440 if (ev && event_can_be_cpu_signaled(ev))
441 reset_event(ev);
442 else
443 ret = -EINVAL;
444
445 mutex_unlock(&p->event_mutex);
446 return ret;
447
448}
449
450static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev)
451{
452 page_slots(p->signal_page)[ev->event_id] = UNSIGNALED_EVENT_SLOT;
453}
454
455static void set_event_from_interrupt(struct kfd_process *p,
456 struct kfd_event *ev)
457{
458 if (ev && event_can_be_gpu_signaled(ev)) {
459 acknowledge_signal(p, ev);
460 set_event(ev);
461 }
462}
463
464void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
465 uint32_t valid_id_bits)
466{
467 struct kfd_event *ev = NULL;
468
469 /*
470 * Because we are called from arbitrary context (workqueue) as opposed
471 * to process context, kfd_process could attempt to exit while we are
472 * running so the lookup function increments the process ref count.
473 */
474 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
475
476 if (!p)
477 return; /* Presumably process exited. */
478
479 mutex_lock(&p->event_mutex);
480
481 if (valid_id_bits)
482 ev = lookup_signaled_event_by_partial_id(p, partial_id,
483 valid_id_bits);
484 if (ev) {
485 set_event_from_interrupt(p, ev);
486 } else if (p->signal_page) {
487 /*
488 * Partial ID lookup failed. Assume that the event ID
489 * in the interrupt payload was invalid and do an
490 * exhaustive search of signaled events.
491 */
492 uint64_t *slots = page_slots(p->signal_page);
493 uint32_t id;
494
495 if (valid_id_bits)
496 pr_debug_ratelimited("Partial ID invalid: %u (%u valid bits)\n",
497 partial_id, valid_id_bits);
498
499 if (p->signal_event_count < KFD_SIGNAL_EVENT_LIMIT / 64) {
500 /* With relatively few events, it's faster to
501 * iterate over the event IDR
502 */
503 idr_for_each_entry(&p->event_idr, ev, id) {
504 if (id >= KFD_SIGNAL_EVENT_LIMIT)
505 break;
506
507 if (slots[id] != UNSIGNALED_EVENT_SLOT)
508 set_event_from_interrupt(p, ev);
509 }
510 } else {
511 /* With relatively many events, it's faster to
512 * iterate over the signal slots and lookup
513 * only signaled events from the IDR.
514 */
515 for (id = 0; id < KFD_SIGNAL_EVENT_LIMIT; id++)
516 if (slots[id] != UNSIGNALED_EVENT_SLOT) {
517 ev = lookup_event_by_id(p, id);
518 set_event_from_interrupt(p, ev);
519 }
520 }
521 }
522
523 mutex_unlock(&p->event_mutex);
524 kfd_unref_process(p);
525}
526
527static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events)
528{
529 struct kfd_event_waiter *event_waiters;
530 uint32_t i;
531
532 event_waiters = kmalloc_array(num_events,
533 sizeof(struct kfd_event_waiter),
534 GFP_KERNEL);
535
536 for (i = 0; (event_waiters) && (i < num_events) ; i++) {
537 init_wait(&event_waiters[i].wait);
538 event_waiters[i].activated = false;
539 }
540
541 return event_waiters;
542}
543
544static int init_event_waiter_get_status(struct kfd_process *p,
545 struct kfd_event_waiter *waiter,
546 uint32_t event_id)
547{
548 struct kfd_event *ev = lookup_event_by_id(p, event_id);
549
550 if (!ev)
551 return -EINVAL;
552
553 waiter->event = ev;
554 waiter->activated = ev->signaled;
555 ev->signaled = ev->signaled && !ev->auto_reset;
556
557 return 0;
558}
559
560static void init_event_waiter_add_to_waitlist(struct kfd_event_waiter *waiter)
561{
562 struct kfd_event *ev = waiter->event;
563
564 /* Only add to the wait list if we actually need to
565 * wait on this event.
566 */
567 if (!waiter->activated)
568 add_wait_queue(&ev->wq, &waiter->wait);
569}
570
571/* test_event_condition - Test condition of events being waited for
572 * @all: Return completion only if all events have signaled
573 * @num_events: Number of events to wait for
574 * @event_waiters: Array of event waiters, one per event
575 *
576 * Returns KFD_IOC_WAIT_RESULT_COMPLETE if all (or one) event(s) have
577 * signaled. Returns KFD_IOC_WAIT_RESULT_TIMEOUT if no (or not all)
578 * events have signaled. Returns KFD_IOC_WAIT_RESULT_FAIL if any of
579 * the events have been destroyed.
580 */
581static uint32_t test_event_condition(bool all, uint32_t num_events,
582 struct kfd_event_waiter *event_waiters)
583{
584 uint32_t i;
585 uint32_t activated_count = 0;
586
587 for (i = 0; i < num_events; i++) {
588 if (!event_waiters[i].event)
589 return KFD_IOC_WAIT_RESULT_FAIL;
590
591 if (event_waiters[i].activated) {
592 if (!all)
593 return KFD_IOC_WAIT_RESULT_COMPLETE;
594
595 activated_count++;
596 }
597 }
598
599 return activated_count == num_events ?
600 KFD_IOC_WAIT_RESULT_COMPLETE : KFD_IOC_WAIT_RESULT_TIMEOUT;
601}
602
603/*
604 * Copy event specific data, if defined.
605 * Currently only memory exception events have additional data to copy to user
606 */
607static int copy_signaled_event_data(uint32_t num_events,
608 struct kfd_event_waiter *event_waiters,
609 struct kfd_event_data __user *data)
610{
611 struct kfd_hsa_memory_exception_data *src;
612 struct kfd_hsa_memory_exception_data __user *dst;
613 struct kfd_event_waiter *waiter;
614 struct kfd_event *event;
615 uint32_t i;
616
617 for (i = 0; i < num_events; i++) {
618 waiter = &event_waiters[i];
619 event = waiter->event;
620 if (waiter->activated && event->type == KFD_EVENT_TYPE_MEMORY) {
621 dst = &data[i].memory_exception_data;
622 src = &event->memory_exception_data;
623 if (copy_to_user(dst, src,
624 sizeof(struct kfd_hsa_memory_exception_data)))
625 return -EFAULT;
626 }
627 }
628
629 return 0;
630
631}
632
633
634
635static long user_timeout_to_jiffies(uint32_t user_timeout_ms)
636{
637 if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE)
638 return 0;
639
640 if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE)
641 return MAX_SCHEDULE_TIMEOUT;
642
643 /*
644 * msecs_to_jiffies interprets all values above 2^31-1 as infinite,
645 * but we consider them finite.
646 * This hack is wrong, but nobody is likely to notice.
647 */
648 user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF);
649
650 return msecs_to_jiffies(user_timeout_ms) + 1;
651}
652
653static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters)
654{
655 uint32_t i;
656
657 for (i = 0; i < num_events; i++)
658 if (waiters[i].event)
659 remove_wait_queue(&waiters[i].event->wq,
660 &waiters[i].wait);
661
662 kfree(waiters);
663}
664
665int kfd_wait_on_events(struct kfd_process *p,
666 uint32_t num_events, void __user *data,
667 bool all, uint32_t user_timeout_ms,
668 uint32_t *wait_result)
669{
670 struct kfd_event_data __user *events =
671 (struct kfd_event_data __user *) data;
672 uint32_t i;
673 int ret = 0;
674
675 struct kfd_event_waiter *event_waiters = NULL;
676 long timeout = user_timeout_to_jiffies(user_timeout_ms);
677
678 event_waiters = alloc_event_waiters(num_events);
679 if (!event_waiters) {
680 ret = -ENOMEM;
681 goto out;
682 }
683
684 mutex_lock(&p->event_mutex);
685
686 for (i = 0; i < num_events; i++) {
687 struct kfd_event_data event_data;
688
689 if (copy_from_user(&event_data, &events[i],
690 sizeof(struct kfd_event_data))) {
691 ret = -EFAULT;
692 goto out_unlock;
693 }
694
695 ret = init_event_waiter_get_status(p, &event_waiters[i],
696 event_data.event_id);
697 if (ret)
698 goto out_unlock;
699 }
700
701 /* Check condition once. */
702 *wait_result = test_event_condition(all, num_events, event_waiters);
703 if (*wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) {
704 ret = copy_signaled_event_data(num_events,
705 event_waiters, events);
706 goto out_unlock;
707 } else if (WARN_ON(*wait_result == KFD_IOC_WAIT_RESULT_FAIL)) {
708 /* This should not happen. Events shouldn't be
709 * destroyed while we're holding the event_mutex
710 */
711 goto out_unlock;
712 }
713
714 /* Add to wait lists if we need to wait. */
715 for (i = 0; i < num_events; i++)
716 init_event_waiter_add_to_waitlist(&event_waiters[i]);
717
718 mutex_unlock(&p->event_mutex);
719
720 while (true) {
721 if (fatal_signal_pending(current)) {
722 ret = -EINTR;
723 break;
724 }
725
726 if (signal_pending(current)) {
727 /*
728 * This is wrong when a nonzero, non-infinite timeout
729 * is specified. We need to use
730 * ERESTARTSYS_RESTARTBLOCK, but struct restart_block
731 * contains a union with data for each user and it's
732 * in generic kernel code that I don't want to
733 * touch yet.
734 */
735 ret = -ERESTARTSYS;
736 break;
737 }
738
739 /* Set task state to interruptible sleep before
740 * checking wake-up conditions. A concurrent wake-up
741 * will put the task back into runnable state. In that
742 * case schedule_timeout will not put the task to
743 * sleep and we'll get a chance to re-check the
744 * updated conditions almost immediately. Otherwise,
745 * this race condition would lead to a soft hang or a
746 * very long sleep.
747 */
748 set_current_state(TASK_INTERRUPTIBLE);
749
750 *wait_result = test_event_condition(all, num_events,
751 event_waiters);
752 if (*wait_result != KFD_IOC_WAIT_RESULT_TIMEOUT)
753 break;
754
755 if (timeout <= 0)
756 break;
757
758 timeout = schedule_timeout(timeout);
759 }
760 __set_current_state(TASK_RUNNING);
761
762 /* copy_signaled_event_data may sleep. So this has to happen
763 * after the task state is set back to RUNNING.
764 */
765 if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE)
766 ret = copy_signaled_event_data(num_events,
767 event_waiters, events);
768
769 mutex_lock(&p->event_mutex);
770out_unlock:
771 free_waiters(num_events, event_waiters);
772 mutex_unlock(&p->event_mutex);
773out:
774 if (ret)
775 *wait_result = KFD_IOC_WAIT_RESULT_FAIL;
776 else if (*wait_result == KFD_IOC_WAIT_RESULT_FAIL)
777 ret = -EIO;
778
779 return ret;
780}
781
782int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma)
783{
784 unsigned long pfn;
785 struct kfd_signal_page *page;
786 int ret;
787
788 /* check required size doesn't exceed the allocated size */
789 if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) <
790 get_order(vma->vm_end - vma->vm_start)) {
791 pr_err("Event page mmap requested illegal size\n");
792 return -EINVAL;
793 }
794
795 page = p->signal_page;
796 if (!page) {
797 /* Probably KFD bug, but mmap is user-accessible. */
798 pr_debug("Signal page could not be found\n");
799 return -EINVAL;
800 }
801
802 pfn = __pa(page->kernel_address);
803 pfn >>= PAGE_SHIFT;
804
805 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE
806 | VM_DONTDUMP | VM_PFNMAP;
807
808 pr_debug("Mapping signal page\n");
809 pr_debug(" start user address == 0x%08lx\n", vma->vm_start);
810 pr_debug(" end user address == 0x%08lx\n", vma->vm_end);
811 pr_debug(" pfn == 0x%016lX\n", pfn);
812 pr_debug(" vm_flags == 0x%08lX\n", vma->vm_flags);
813 pr_debug(" size == 0x%08lX\n",
814 vma->vm_end - vma->vm_start);
815
816 page->user_address = (uint64_t __user *)vma->vm_start;
817
818 /* mapping the page to user process */
819 ret = remap_pfn_range(vma, vma->vm_start, pfn,
820 vma->vm_end - vma->vm_start, vma->vm_page_prot);
821 if (!ret)
822 p->signal_mapped_size = vma->vm_end - vma->vm_start;
823
824 return ret;
825}
826
827/*
828 * Assumes that p->event_mutex is held and of course
829 * that p is not going away (current or locked).
830 */
831static void lookup_events_by_type_and_signal(struct kfd_process *p,
832 int type, void *event_data)
833{
834 struct kfd_hsa_memory_exception_data *ev_data;
835 struct kfd_event *ev;
836 uint32_t id;
837 bool send_signal = true;
838
839 ev_data = (struct kfd_hsa_memory_exception_data *) event_data;
840
841 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
842 idr_for_each_entry_continue(&p->event_idr, ev, id)
843 if (ev->type == type) {
844 send_signal = false;
845 dev_dbg(kfd_device,
846 "Event found: id %X type %d",
847 ev->event_id, ev->type);
848 set_event(ev);
849 if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data)
850 ev->memory_exception_data = *ev_data;
851 }
852
853 if (type == KFD_EVENT_TYPE_MEMORY) {
854 dev_warn(kfd_device,
855 "Sending SIGSEGV to HSA Process with PID %d ",
856 p->lead_thread->pid);
857 send_sig(SIGSEGV, p->lead_thread, 0);
858 }
859
860 /* Send SIGTERM no event of type "type" has been found*/
861 if (send_signal) {
862 if (send_sigterm) {
863 dev_warn(kfd_device,
864 "Sending SIGTERM to HSA Process with PID %d ",
865 p->lead_thread->pid);
866 send_sig(SIGTERM, p->lead_thread, 0);
867 } else {
868 dev_err(kfd_device,
869 "HSA Process (PID %d) got unhandled exception",
870 p->lead_thread->pid);
871 }
872 }
873}
874
875#ifdef KFD_SUPPORT_IOMMU_V2
876void kfd_signal_iommu_event(struct kfd_dev *dev, unsigned int pasid,
877 unsigned long address, bool is_write_requested,
878 bool is_execute_requested)
879{
880 struct kfd_hsa_memory_exception_data memory_exception_data;
881 struct vm_area_struct *vma;
882
883 /*
884 * Because we are called from arbitrary context (workqueue) as opposed
885 * to process context, kfd_process could attempt to exit while we are
886 * running so the lookup function increments the process ref count.
887 */
888 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
889 struct mm_struct *mm;
890
891 if (!p)
892 return; /* Presumably process exited. */
893
894 /* Take a safe reference to the mm_struct, which may otherwise
895 * disappear even while the kfd_process is still referenced.
896 */
897 mm = get_task_mm(p->lead_thread);
898 if (!mm) {
899 kfd_unref_process(p);
900 return; /* Process is exiting */
901 }
902
903 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
904
905 down_read(&mm->mmap_sem);
906 vma = find_vma(mm, address);
907
908 memory_exception_data.gpu_id = dev->id;
909 memory_exception_data.va = address;
910 /* Set failure reason */
911 memory_exception_data.failure.NotPresent = 1;
912 memory_exception_data.failure.NoExecute = 0;
913 memory_exception_data.failure.ReadOnly = 0;
914 if (vma && address >= vma->vm_start) {
915 memory_exception_data.failure.NotPresent = 0;
916
917 if (is_write_requested && !(vma->vm_flags & VM_WRITE))
918 memory_exception_data.failure.ReadOnly = 1;
919 else
920 memory_exception_data.failure.ReadOnly = 0;
921
922 if (is_execute_requested && !(vma->vm_flags & VM_EXEC))
923 memory_exception_data.failure.NoExecute = 1;
924 else
925 memory_exception_data.failure.NoExecute = 0;
926 }
927
928 up_read(&mm->mmap_sem);
929 mmput(mm);
930
931 pr_debug("notpresent %d, noexecute %d, readonly %d\n",
932 memory_exception_data.failure.NotPresent,
933 memory_exception_data.failure.NoExecute,
934 memory_exception_data.failure.ReadOnly);
935
936 /* Workaround on Raven to not kill the process when memory is freed
937 * before IOMMU is able to finish processing all the excessive PPRs
938 */
939 if (dev->device_info->asic_family != CHIP_RAVEN) {
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 = 1;
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_process *p;
1015 struct kfd_event *ev;
1016 unsigned int temp;
1017 uint32_t id, idx;
1018
1019 /* Whole gpu reset caused by GPU hang and memory is lost */
1020 memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1021 hw_exception_data.gpu_id = dev->id;
1022 hw_exception_data.memory_lost = 1;
1023
1024 idx = srcu_read_lock(&kfd_processes_srcu);
1025 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1026 mutex_lock(&p->event_mutex);
1027 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1028 idr_for_each_entry_continue(&p->event_idr, ev, id)
1029 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1030 ev->hw_exception_data = hw_exception_data;
1031 set_event(ev);
1032 }
1033 mutex_unlock(&p->event_mutex);
1034 }
1035 srcu_read_unlock(&kfd_processes_srcu, idx);
1036}
1037