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.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 == false) {
181 ret = allocate_signal_page(devkfd, p);
182 if (ret == true)
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}
1// SPDX-License-Identifier: GPL-2.0 OR MIT
2/*
3 * Copyright 2014-2022 Advanced Micro Devices, Inc.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21 * OTHER DEALINGS IN THE SOFTWARE.
22 */
23
24#include <linux/mm_types.h>
25#include <linux/slab.h>
26#include <linux/types.h>
27#include <linux/sched/signal.h>
28#include <linux/sched/mm.h>
29#include <linux/uaccess.h>
30#include <linux/mman.h>
31#include <linux/memory.h>
32#include "kfd_priv.h"
33#include "kfd_events.h"
34#include "kfd_iommu.h"
35#include <linux/device.h>
36
37/*
38 * Wrapper around wait_queue_entry_t
39 */
40struct kfd_event_waiter {
41 wait_queue_entry_t wait;
42 struct kfd_event *event; /* Event to wait for */
43 bool activated; /* Becomes true when event is signaled */
44};
45
46/*
47 * Each signal event needs a 64-bit signal slot where the signaler will write
48 * a 1 before sending an interrupt. (This is needed because some interrupts
49 * do not contain enough spare data bits to identify an event.)
50 * We get whole pages and map them to the process VA.
51 * Individual signal events use their event_id as slot index.
52 */
53struct kfd_signal_page {
54 uint64_t *kernel_address;
55 uint64_t __user *user_address;
56 bool need_to_free_pages;
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 const int *restore_id)
97{
98 int id;
99
100 if (!p->signal_page) {
101 p->signal_page = allocate_signal_page(p);
102 if (!p->signal_page)
103 return -ENOMEM;
104 /* Oldest user mode expects 256 event slots */
105 p->signal_mapped_size = 256*8;
106 }
107
108 if (restore_id) {
109 id = idr_alloc(&p->event_idr, ev, *restore_id, *restore_id + 1,
110 GFP_KERNEL);
111 } else {
112 /*
113 * Compatibility with old user mode: Only use signal slots
114 * user mode has mapped, may be less than
115 * KFD_SIGNAL_EVENT_LIMIT. This also allows future increase
116 * of the event limit without breaking user mode.
117 */
118 id = idr_alloc(&p->event_idr, ev, 0, p->signal_mapped_size / 8,
119 GFP_KERNEL);
120 }
121 if (id < 0)
122 return id;
123
124 ev->event_id = id;
125 page_slots(p->signal_page)[id] = UNSIGNALED_EVENT_SLOT;
126
127 return 0;
128}
129
130/*
131 * Assumes that p->event_mutex or rcu_readlock is held and of course that p is
132 * not going away.
133 */
134static struct kfd_event *lookup_event_by_id(struct kfd_process *p, uint32_t id)
135{
136 return idr_find(&p->event_idr, id);
137}
138
139/**
140 * lookup_signaled_event_by_partial_id - Lookup signaled event from partial ID
141 * @p: Pointer to struct kfd_process
142 * @id: ID to look up
143 * @bits: Number of valid bits in @id
144 *
145 * Finds the first signaled event with a matching partial ID. If no
146 * matching signaled event is found, returns NULL. In that case the
147 * caller should assume that the partial ID is invalid and do an
148 * exhaustive search of all siglaned events.
149 *
150 * If multiple events with the same partial ID signal at the same
151 * time, they will be found one interrupt at a time, not necessarily
152 * in the same order the interrupts occurred. As long as the number of
153 * interrupts is correct, all signaled events will be seen by the
154 * driver.
155 */
156static struct kfd_event *lookup_signaled_event_by_partial_id(
157 struct kfd_process *p, uint32_t id, uint32_t bits)
158{
159 struct kfd_event *ev;
160
161 if (!p->signal_page || id >= KFD_SIGNAL_EVENT_LIMIT)
162 return NULL;
163
164 /* Fast path for the common case that @id is not a partial ID
165 * and we only need a single lookup.
166 */
167 if (bits > 31 || (1U << bits) >= KFD_SIGNAL_EVENT_LIMIT) {
168 if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
169 return NULL;
170
171 return idr_find(&p->event_idr, id);
172 }
173
174 /* General case for partial IDs: Iterate over all matching IDs
175 * and find the first one that has signaled.
176 */
177 for (ev = NULL; id < KFD_SIGNAL_EVENT_LIMIT && !ev; id += 1U << bits) {
178 if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
179 continue;
180
181 ev = idr_find(&p->event_idr, id);
182 }
183
184 return ev;
185}
186
187static int create_signal_event(struct file *devkfd, struct kfd_process *p,
188 struct kfd_event *ev, const int *restore_id)
189{
190 int ret;
191
192 if (p->signal_mapped_size &&
193 p->signal_event_count == p->signal_mapped_size / 8) {
194 if (!p->signal_event_limit_reached) {
195 pr_debug("Signal event wasn't created because limit was reached\n");
196 p->signal_event_limit_reached = true;
197 }
198 return -ENOSPC;
199 }
200
201 ret = allocate_event_notification_slot(p, ev, restore_id);
202 if (ret) {
203 pr_warn("Signal event wasn't created because out of kernel memory\n");
204 return ret;
205 }
206
207 p->signal_event_count++;
208
209 ev->user_signal_address = &p->signal_page->user_address[ev->event_id];
210 pr_debug("Signal event number %zu created with id %d, address %p\n",
211 p->signal_event_count, ev->event_id,
212 ev->user_signal_address);
213
214 return 0;
215}
216
217static int create_other_event(struct kfd_process *p, struct kfd_event *ev, const int *restore_id)
218{
219 int id;
220
221 if (restore_id)
222 id = idr_alloc(&p->event_idr, ev, *restore_id, *restore_id + 1,
223 GFP_KERNEL);
224 else
225 /* Cast KFD_LAST_NONSIGNAL_EVENT to uint32_t. This allows an
226 * intentional integer overflow to -1 without a compiler
227 * warning. idr_alloc treats a negative value as "maximum
228 * signed integer".
229 */
230 id = idr_alloc(&p->event_idr, ev, KFD_FIRST_NONSIGNAL_EVENT_ID,
231 (uint32_t)KFD_LAST_NONSIGNAL_EVENT_ID + 1,
232 GFP_KERNEL);
233
234 if (id < 0)
235 return id;
236 ev->event_id = id;
237
238 return 0;
239}
240
241int kfd_event_init_process(struct kfd_process *p)
242{
243 int id;
244
245 mutex_init(&p->event_mutex);
246 idr_init(&p->event_idr);
247 p->signal_page = NULL;
248 p->signal_event_count = 1;
249 /* Allocate event ID 0. It is used for a fast path to ignore bogus events
250 * that are sent by the CP without a context ID
251 */
252 id = idr_alloc(&p->event_idr, NULL, 0, 1, GFP_KERNEL);
253 if (id < 0) {
254 idr_destroy(&p->event_idr);
255 mutex_destroy(&p->event_mutex);
256 return id;
257 }
258 return 0;
259}
260
261static void destroy_event(struct kfd_process *p, struct kfd_event *ev)
262{
263 struct kfd_event_waiter *waiter;
264
265 /* Wake up pending waiters. They will return failure */
266 spin_lock(&ev->lock);
267 list_for_each_entry(waiter, &ev->wq.head, wait.entry)
268 WRITE_ONCE(waiter->event, NULL);
269 wake_up_all(&ev->wq);
270 spin_unlock(&ev->lock);
271
272 if (ev->type == KFD_EVENT_TYPE_SIGNAL ||
273 ev->type == KFD_EVENT_TYPE_DEBUG)
274 p->signal_event_count--;
275
276 idr_remove(&p->event_idr, ev->event_id);
277 kfree_rcu(ev, rcu);
278}
279
280static void destroy_events(struct kfd_process *p)
281{
282 struct kfd_event *ev;
283 uint32_t id;
284
285 idr_for_each_entry(&p->event_idr, ev, id)
286 if (ev)
287 destroy_event(p, ev);
288 idr_destroy(&p->event_idr);
289 mutex_destroy(&p->event_mutex);
290}
291
292/*
293 * We assume that the process is being destroyed and there is no need to
294 * unmap the pages or keep bookkeeping data in order.
295 */
296static void shutdown_signal_page(struct kfd_process *p)
297{
298 struct kfd_signal_page *page = p->signal_page;
299
300 if (page) {
301 if (page->need_to_free_pages)
302 free_pages((unsigned long)page->kernel_address,
303 get_order(KFD_SIGNAL_EVENT_LIMIT * 8));
304 kfree(page);
305 }
306}
307
308void kfd_event_free_process(struct kfd_process *p)
309{
310 destroy_events(p);
311 shutdown_signal_page(p);
312}
313
314static bool event_can_be_gpu_signaled(const struct kfd_event *ev)
315{
316 return ev->type == KFD_EVENT_TYPE_SIGNAL ||
317 ev->type == KFD_EVENT_TYPE_DEBUG;
318}
319
320static bool event_can_be_cpu_signaled(const struct kfd_event *ev)
321{
322 return ev->type == KFD_EVENT_TYPE_SIGNAL;
323}
324
325static int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
326 uint64_t size, uint64_t user_handle)
327{
328 struct kfd_signal_page *page;
329
330 if (p->signal_page)
331 return -EBUSY;
332
333 page = kzalloc(sizeof(*page), GFP_KERNEL);
334 if (!page)
335 return -ENOMEM;
336
337 /* Initialize all events to unsignaled */
338 memset(kernel_address, (uint8_t) UNSIGNALED_EVENT_SLOT,
339 KFD_SIGNAL_EVENT_LIMIT * 8);
340
341 page->kernel_address = kernel_address;
342
343 p->signal_page = page;
344 p->signal_mapped_size = size;
345 p->signal_handle = user_handle;
346 return 0;
347}
348
349int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset)
350{
351 struct kfd_dev *kfd;
352 struct kfd_process_device *pdd;
353 void *mem, *kern_addr;
354 uint64_t size;
355 int err = 0;
356
357 if (p->signal_page) {
358 pr_err("Event page is already set\n");
359 return -EINVAL;
360 }
361
362 pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(event_page_offset));
363 if (!pdd) {
364 pr_err("Getting device by id failed in %s\n", __func__);
365 return -EINVAL;
366 }
367 kfd = pdd->dev;
368
369 pdd = kfd_bind_process_to_device(kfd, p);
370 if (IS_ERR(pdd))
371 return PTR_ERR(pdd);
372
373 mem = kfd_process_device_translate_handle(pdd,
374 GET_IDR_HANDLE(event_page_offset));
375 if (!mem) {
376 pr_err("Can't find BO, offset is 0x%llx\n", event_page_offset);
377 return -EINVAL;
378 }
379
380 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(mem, &kern_addr, &size);
381 if (err) {
382 pr_err("Failed to map event page to kernel\n");
383 return err;
384 }
385
386 err = kfd_event_page_set(p, kern_addr, size, event_page_offset);
387 if (err) {
388 pr_err("Failed to set event page\n");
389 amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
390 return err;
391 }
392 return err;
393}
394
395int kfd_event_create(struct file *devkfd, struct kfd_process *p,
396 uint32_t event_type, bool auto_reset, uint32_t node_id,
397 uint32_t *event_id, uint32_t *event_trigger_data,
398 uint64_t *event_page_offset, uint32_t *event_slot_index)
399{
400 int ret = 0;
401 struct kfd_event *ev = kzalloc(sizeof(*ev), GFP_KERNEL);
402
403 if (!ev)
404 return -ENOMEM;
405
406 ev->type = event_type;
407 ev->auto_reset = auto_reset;
408 ev->signaled = false;
409
410 spin_lock_init(&ev->lock);
411 init_waitqueue_head(&ev->wq);
412
413 *event_page_offset = 0;
414
415 mutex_lock(&p->event_mutex);
416
417 switch (event_type) {
418 case KFD_EVENT_TYPE_SIGNAL:
419 case KFD_EVENT_TYPE_DEBUG:
420 ret = create_signal_event(devkfd, p, ev, NULL);
421 if (!ret) {
422 *event_page_offset = KFD_MMAP_TYPE_EVENTS;
423 *event_slot_index = ev->event_id;
424 }
425 break;
426 default:
427 ret = create_other_event(p, ev, NULL);
428 break;
429 }
430
431 if (!ret) {
432 *event_id = ev->event_id;
433 *event_trigger_data = ev->event_id;
434 } else {
435 kfree(ev);
436 }
437
438 mutex_unlock(&p->event_mutex);
439
440 return ret;
441}
442
443int kfd_criu_restore_event(struct file *devkfd,
444 struct kfd_process *p,
445 uint8_t __user *user_priv_ptr,
446 uint64_t *priv_data_offset,
447 uint64_t max_priv_data_size)
448{
449 struct kfd_criu_event_priv_data *ev_priv;
450 struct kfd_event *ev = NULL;
451 int ret = 0;
452
453 ev_priv = kmalloc(sizeof(*ev_priv), GFP_KERNEL);
454 if (!ev_priv)
455 return -ENOMEM;
456
457 ev = kzalloc(sizeof(*ev), GFP_KERNEL);
458 if (!ev) {
459 ret = -ENOMEM;
460 goto exit;
461 }
462
463 if (*priv_data_offset + sizeof(*ev_priv) > max_priv_data_size) {
464 ret = -EINVAL;
465 goto exit;
466 }
467
468 ret = copy_from_user(ev_priv, user_priv_ptr + *priv_data_offset, sizeof(*ev_priv));
469 if (ret) {
470 ret = -EFAULT;
471 goto exit;
472 }
473 *priv_data_offset += sizeof(*ev_priv);
474
475 if (ev_priv->user_handle) {
476 ret = kfd_kmap_event_page(p, ev_priv->user_handle);
477 if (ret)
478 goto exit;
479 }
480
481 ev->type = ev_priv->type;
482 ev->auto_reset = ev_priv->auto_reset;
483 ev->signaled = ev_priv->signaled;
484
485 spin_lock_init(&ev->lock);
486 init_waitqueue_head(&ev->wq);
487
488 mutex_lock(&p->event_mutex);
489 switch (ev->type) {
490 case KFD_EVENT_TYPE_SIGNAL:
491 case KFD_EVENT_TYPE_DEBUG:
492 ret = create_signal_event(devkfd, p, ev, &ev_priv->event_id);
493 break;
494 case KFD_EVENT_TYPE_MEMORY:
495 memcpy(&ev->memory_exception_data,
496 &ev_priv->memory_exception_data,
497 sizeof(struct kfd_hsa_memory_exception_data));
498
499 ret = create_other_event(p, ev, &ev_priv->event_id);
500 break;
501 case KFD_EVENT_TYPE_HW_EXCEPTION:
502 memcpy(&ev->hw_exception_data,
503 &ev_priv->hw_exception_data,
504 sizeof(struct kfd_hsa_hw_exception_data));
505
506 ret = create_other_event(p, ev, &ev_priv->event_id);
507 break;
508 }
509 mutex_unlock(&p->event_mutex);
510
511exit:
512 if (ret)
513 kfree(ev);
514
515 kfree(ev_priv);
516
517 return ret;
518}
519
520int kfd_criu_checkpoint_events(struct kfd_process *p,
521 uint8_t __user *user_priv_data,
522 uint64_t *priv_data_offset)
523{
524 struct kfd_criu_event_priv_data *ev_privs;
525 int i = 0;
526 int ret = 0;
527 struct kfd_event *ev;
528 uint32_t ev_id;
529
530 uint32_t num_events = kfd_get_num_events(p);
531
532 if (!num_events)
533 return 0;
534
535 ev_privs = kvzalloc(num_events * sizeof(*ev_privs), GFP_KERNEL);
536 if (!ev_privs)
537 return -ENOMEM;
538
539
540 idr_for_each_entry(&p->event_idr, ev, ev_id) {
541 struct kfd_criu_event_priv_data *ev_priv;
542
543 /*
544 * Currently, all events have same size of private_data, but the current ioctl's
545 * and CRIU plugin supports private_data of variable sizes
546 */
547 ev_priv = &ev_privs[i];
548
549 ev_priv->object_type = KFD_CRIU_OBJECT_TYPE_EVENT;
550
551 /* We store the user_handle with the first event */
552 if (i == 0 && p->signal_page)
553 ev_priv->user_handle = p->signal_handle;
554
555 ev_priv->event_id = ev->event_id;
556 ev_priv->auto_reset = ev->auto_reset;
557 ev_priv->type = ev->type;
558 ev_priv->signaled = ev->signaled;
559
560 if (ev_priv->type == KFD_EVENT_TYPE_MEMORY)
561 memcpy(&ev_priv->memory_exception_data,
562 &ev->memory_exception_data,
563 sizeof(struct kfd_hsa_memory_exception_data));
564 else if (ev_priv->type == KFD_EVENT_TYPE_HW_EXCEPTION)
565 memcpy(&ev_priv->hw_exception_data,
566 &ev->hw_exception_data,
567 sizeof(struct kfd_hsa_hw_exception_data));
568
569 pr_debug("Checkpointed event[%d] id = 0x%08x auto_reset = %x type = %x signaled = %x\n",
570 i,
571 ev_priv->event_id,
572 ev_priv->auto_reset,
573 ev_priv->type,
574 ev_priv->signaled);
575 i++;
576 }
577
578 ret = copy_to_user(user_priv_data + *priv_data_offset,
579 ev_privs, num_events * sizeof(*ev_privs));
580 if (ret) {
581 pr_err("Failed to copy events priv to user\n");
582 ret = -EFAULT;
583 }
584
585 *priv_data_offset += num_events * sizeof(*ev_privs);
586
587 kvfree(ev_privs);
588 return ret;
589}
590
591int kfd_get_num_events(struct kfd_process *p)
592{
593 struct kfd_event *ev;
594 uint32_t id;
595 u32 num_events = 0;
596
597 idr_for_each_entry(&p->event_idr, ev, id)
598 num_events++;
599
600 return num_events;
601}
602
603/* Assumes that p is current. */
604int kfd_event_destroy(struct kfd_process *p, uint32_t event_id)
605{
606 struct kfd_event *ev;
607 int ret = 0;
608
609 mutex_lock(&p->event_mutex);
610
611 ev = lookup_event_by_id(p, event_id);
612
613 if (ev)
614 destroy_event(p, ev);
615 else
616 ret = -EINVAL;
617
618 mutex_unlock(&p->event_mutex);
619 return ret;
620}
621
622static void set_event(struct kfd_event *ev)
623{
624 struct kfd_event_waiter *waiter;
625
626 /* Auto reset if the list is non-empty and we're waking
627 * someone. waitqueue_active is safe here because we're
628 * protected by the ev->lock, which is also held when
629 * updating the wait queues in kfd_wait_on_events.
630 */
631 ev->signaled = !ev->auto_reset || !waitqueue_active(&ev->wq);
632
633 list_for_each_entry(waiter, &ev->wq.head, wait.entry)
634 WRITE_ONCE(waiter->activated, true);
635
636 wake_up_all(&ev->wq);
637}
638
639/* Assumes that p is current. */
640int kfd_set_event(struct kfd_process *p, uint32_t event_id)
641{
642 int ret = 0;
643 struct kfd_event *ev;
644
645 rcu_read_lock();
646
647 ev = lookup_event_by_id(p, event_id);
648 if (!ev) {
649 ret = -EINVAL;
650 goto unlock_rcu;
651 }
652 spin_lock(&ev->lock);
653
654 if (event_can_be_cpu_signaled(ev))
655 set_event(ev);
656 else
657 ret = -EINVAL;
658
659 spin_unlock(&ev->lock);
660unlock_rcu:
661 rcu_read_unlock();
662 return ret;
663}
664
665static void reset_event(struct kfd_event *ev)
666{
667 ev->signaled = false;
668}
669
670/* Assumes that p is current. */
671int kfd_reset_event(struct kfd_process *p, uint32_t event_id)
672{
673 int ret = 0;
674 struct kfd_event *ev;
675
676 rcu_read_lock();
677
678 ev = lookup_event_by_id(p, event_id);
679 if (!ev) {
680 ret = -EINVAL;
681 goto unlock_rcu;
682 }
683 spin_lock(&ev->lock);
684
685 if (event_can_be_cpu_signaled(ev))
686 reset_event(ev);
687 else
688 ret = -EINVAL;
689
690 spin_unlock(&ev->lock);
691unlock_rcu:
692 rcu_read_unlock();
693 return ret;
694
695}
696
697static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev)
698{
699 WRITE_ONCE(page_slots(p->signal_page)[ev->event_id], UNSIGNALED_EVENT_SLOT);
700}
701
702static void set_event_from_interrupt(struct kfd_process *p,
703 struct kfd_event *ev)
704{
705 if (ev && event_can_be_gpu_signaled(ev)) {
706 acknowledge_signal(p, ev);
707 spin_lock(&ev->lock);
708 set_event(ev);
709 spin_unlock(&ev->lock);
710 }
711}
712
713void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
714 uint32_t valid_id_bits)
715{
716 struct kfd_event *ev = NULL;
717
718 /*
719 * Because we are called from arbitrary context (workqueue) as opposed
720 * to process context, kfd_process could attempt to exit while we are
721 * running so the lookup function increments the process ref count.
722 */
723 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
724
725 if (!p)
726 return; /* Presumably process exited. */
727
728 rcu_read_lock();
729
730 if (valid_id_bits)
731 ev = lookup_signaled_event_by_partial_id(p, partial_id,
732 valid_id_bits);
733 if (ev) {
734 set_event_from_interrupt(p, ev);
735 } else if (p->signal_page) {
736 /*
737 * Partial ID lookup failed. Assume that the event ID
738 * in the interrupt payload was invalid and do an
739 * exhaustive search of signaled events.
740 */
741 uint64_t *slots = page_slots(p->signal_page);
742 uint32_t id;
743
744 if (valid_id_bits)
745 pr_debug_ratelimited("Partial ID invalid: %u (%u valid bits)\n",
746 partial_id, valid_id_bits);
747
748 if (p->signal_event_count < KFD_SIGNAL_EVENT_LIMIT / 64) {
749 /* With relatively few events, it's faster to
750 * iterate over the event IDR
751 */
752 idr_for_each_entry(&p->event_idr, ev, id) {
753 if (id >= KFD_SIGNAL_EVENT_LIMIT)
754 break;
755
756 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT)
757 set_event_from_interrupt(p, ev);
758 }
759 } else {
760 /* With relatively many events, it's faster to
761 * iterate over the signal slots and lookup
762 * only signaled events from the IDR.
763 */
764 for (id = 1; id < KFD_SIGNAL_EVENT_LIMIT; id++)
765 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT) {
766 ev = lookup_event_by_id(p, id);
767 set_event_from_interrupt(p, ev);
768 }
769 }
770 }
771
772 rcu_read_unlock();
773 kfd_unref_process(p);
774}
775
776static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events)
777{
778 struct kfd_event_waiter *event_waiters;
779 uint32_t i;
780
781 event_waiters = kmalloc_array(num_events,
782 sizeof(struct kfd_event_waiter),
783 GFP_KERNEL);
784 if (!event_waiters)
785 return NULL;
786
787 for (i = 0; (event_waiters) && (i < num_events) ; i++) {
788 init_wait(&event_waiters[i].wait);
789 event_waiters[i].activated = false;
790 }
791
792 return event_waiters;
793}
794
795static int init_event_waiter(struct kfd_process *p,
796 struct kfd_event_waiter *waiter,
797 uint32_t event_id)
798{
799 struct kfd_event *ev = lookup_event_by_id(p, event_id);
800
801 if (!ev)
802 return -EINVAL;
803
804 spin_lock(&ev->lock);
805 waiter->event = ev;
806 waiter->activated = ev->signaled;
807 ev->signaled = ev->signaled && !ev->auto_reset;
808 if (!waiter->activated)
809 add_wait_queue(&ev->wq, &waiter->wait);
810 spin_unlock(&ev->lock);
811
812 return 0;
813}
814
815/* test_event_condition - Test condition of events being waited for
816 * @all: Return completion only if all events have signaled
817 * @num_events: Number of events to wait for
818 * @event_waiters: Array of event waiters, one per event
819 *
820 * Returns KFD_IOC_WAIT_RESULT_COMPLETE if all (or one) event(s) have
821 * signaled. Returns KFD_IOC_WAIT_RESULT_TIMEOUT if no (or not all)
822 * events have signaled. Returns KFD_IOC_WAIT_RESULT_FAIL if any of
823 * the events have been destroyed.
824 */
825static uint32_t test_event_condition(bool all, uint32_t num_events,
826 struct kfd_event_waiter *event_waiters)
827{
828 uint32_t i;
829 uint32_t activated_count = 0;
830
831 for (i = 0; i < num_events; i++) {
832 if (!READ_ONCE(event_waiters[i].event))
833 return KFD_IOC_WAIT_RESULT_FAIL;
834
835 if (READ_ONCE(event_waiters[i].activated)) {
836 if (!all)
837 return KFD_IOC_WAIT_RESULT_COMPLETE;
838
839 activated_count++;
840 }
841 }
842
843 return activated_count == num_events ?
844 KFD_IOC_WAIT_RESULT_COMPLETE : KFD_IOC_WAIT_RESULT_TIMEOUT;
845}
846
847/*
848 * Copy event specific data, if defined.
849 * Currently only memory exception events have additional data to copy to user
850 */
851static int copy_signaled_event_data(uint32_t num_events,
852 struct kfd_event_waiter *event_waiters,
853 struct kfd_event_data __user *data)
854{
855 struct kfd_hsa_memory_exception_data *src;
856 struct kfd_hsa_memory_exception_data __user *dst;
857 struct kfd_event_waiter *waiter;
858 struct kfd_event *event;
859 uint32_t i;
860
861 for (i = 0; i < num_events; i++) {
862 waiter = &event_waiters[i];
863 event = waiter->event;
864 if (!event)
865 return -EINVAL; /* event was destroyed */
866 if (waiter->activated && event->type == KFD_EVENT_TYPE_MEMORY) {
867 dst = &data[i].memory_exception_data;
868 src = &event->memory_exception_data;
869 if (copy_to_user(dst, src,
870 sizeof(struct kfd_hsa_memory_exception_data)))
871 return -EFAULT;
872 }
873 }
874
875 return 0;
876}
877
878static long user_timeout_to_jiffies(uint32_t user_timeout_ms)
879{
880 if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE)
881 return 0;
882
883 if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE)
884 return MAX_SCHEDULE_TIMEOUT;
885
886 /*
887 * msecs_to_jiffies interprets all values above 2^31-1 as infinite,
888 * but we consider them finite.
889 * This hack is wrong, but nobody is likely to notice.
890 */
891 user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF);
892
893 return msecs_to_jiffies(user_timeout_ms) + 1;
894}
895
896static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters,
897 bool undo_auto_reset)
898{
899 uint32_t i;
900
901 for (i = 0; i < num_events; i++)
902 if (waiters[i].event) {
903 spin_lock(&waiters[i].event->lock);
904 remove_wait_queue(&waiters[i].event->wq,
905 &waiters[i].wait);
906 if (undo_auto_reset && waiters[i].activated &&
907 waiters[i].event && waiters[i].event->auto_reset)
908 set_event(waiters[i].event);
909 spin_unlock(&waiters[i].event->lock);
910 }
911
912 kfree(waiters);
913}
914
915int kfd_wait_on_events(struct kfd_process *p,
916 uint32_t num_events, void __user *data,
917 bool all, uint32_t *user_timeout_ms,
918 uint32_t *wait_result)
919{
920 struct kfd_event_data __user *events =
921 (struct kfd_event_data __user *) data;
922 uint32_t i;
923 int ret = 0;
924
925 struct kfd_event_waiter *event_waiters = NULL;
926 long timeout = user_timeout_to_jiffies(*user_timeout_ms);
927
928 event_waiters = alloc_event_waiters(num_events);
929 if (!event_waiters) {
930 ret = -ENOMEM;
931 goto out;
932 }
933
934 /* Use p->event_mutex here to protect against concurrent creation and
935 * destruction of events while we initialize event_waiters.
936 */
937 mutex_lock(&p->event_mutex);
938
939 for (i = 0; i < num_events; i++) {
940 struct kfd_event_data event_data;
941
942 if (copy_from_user(&event_data, &events[i],
943 sizeof(struct kfd_event_data))) {
944 ret = -EFAULT;
945 goto out_unlock;
946 }
947
948 ret = init_event_waiter(p, &event_waiters[i],
949 event_data.event_id);
950 if (ret)
951 goto out_unlock;
952 }
953
954 /* Check condition once. */
955 *wait_result = test_event_condition(all, num_events, event_waiters);
956 if (*wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) {
957 ret = copy_signaled_event_data(num_events,
958 event_waiters, events);
959 goto out_unlock;
960 } else if (WARN_ON(*wait_result == KFD_IOC_WAIT_RESULT_FAIL)) {
961 /* This should not happen. Events shouldn't be
962 * destroyed while we're holding the event_mutex
963 */
964 goto out_unlock;
965 }
966
967 mutex_unlock(&p->event_mutex);
968
969 while (true) {
970 if (fatal_signal_pending(current)) {
971 ret = -EINTR;
972 break;
973 }
974
975 if (signal_pending(current)) {
976 ret = -ERESTARTSYS;
977 if (*user_timeout_ms != KFD_EVENT_TIMEOUT_IMMEDIATE &&
978 *user_timeout_ms != KFD_EVENT_TIMEOUT_INFINITE)
979 *user_timeout_ms = jiffies_to_msecs(
980 max(0l, timeout-1));
981 break;
982 }
983
984 /* Set task state to interruptible sleep before
985 * checking wake-up conditions. A concurrent wake-up
986 * will put the task back into runnable state. In that
987 * case schedule_timeout will not put the task to
988 * sleep and we'll get a chance to re-check the
989 * updated conditions almost immediately. Otherwise,
990 * this race condition would lead to a soft hang or a
991 * very long sleep.
992 */
993 set_current_state(TASK_INTERRUPTIBLE);
994
995 *wait_result = test_event_condition(all, num_events,
996 event_waiters);
997 if (*wait_result != KFD_IOC_WAIT_RESULT_TIMEOUT)
998 break;
999
1000 if (timeout <= 0)
1001 break;
1002
1003 timeout = schedule_timeout(timeout);
1004 }
1005 __set_current_state(TASK_RUNNING);
1006
1007 mutex_lock(&p->event_mutex);
1008 /* copy_signaled_event_data may sleep. So this has to happen
1009 * after the task state is set back to RUNNING.
1010 *
1011 * The event may also have been destroyed after signaling. So
1012 * copy_signaled_event_data also must confirm that the event
1013 * still exists. Therefore this must be under the p->event_mutex
1014 * which is also held when events are destroyed.
1015 */
1016 if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE)
1017 ret = copy_signaled_event_data(num_events,
1018 event_waiters, events);
1019
1020out_unlock:
1021 free_waiters(num_events, event_waiters, ret == -ERESTARTSYS);
1022 mutex_unlock(&p->event_mutex);
1023out:
1024 if (ret)
1025 *wait_result = KFD_IOC_WAIT_RESULT_FAIL;
1026 else if (*wait_result == KFD_IOC_WAIT_RESULT_FAIL)
1027 ret = -EIO;
1028
1029 return ret;
1030}
1031
1032int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma)
1033{
1034 unsigned long pfn;
1035 struct kfd_signal_page *page;
1036 int ret;
1037
1038 /* check required size doesn't exceed the allocated size */
1039 if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) <
1040 get_order(vma->vm_end - vma->vm_start)) {
1041 pr_err("Event page mmap requested illegal size\n");
1042 return -EINVAL;
1043 }
1044
1045 page = p->signal_page;
1046 if (!page) {
1047 /* Probably KFD bug, but mmap is user-accessible. */
1048 pr_debug("Signal page could not be found\n");
1049 return -EINVAL;
1050 }
1051
1052 pfn = __pa(page->kernel_address);
1053 pfn >>= PAGE_SHIFT;
1054
1055 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE
1056 | VM_DONTDUMP | VM_PFNMAP;
1057
1058 pr_debug("Mapping signal page\n");
1059 pr_debug(" start user address == 0x%08lx\n", vma->vm_start);
1060 pr_debug(" end user address == 0x%08lx\n", vma->vm_end);
1061 pr_debug(" pfn == 0x%016lX\n", pfn);
1062 pr_debug(" vm_flags == 0x%08lX\n", vma->vm_flags);
1063 pr_debug(" size == 0x%08lX\n",
1064 vma->vm_end - vma->vm_start);
1065
1066 page->user_address = (uint64_t __user *)vma->vm_start;
1067
1068 /* mapping the page to user process */
1069 ret = remap_pfn_range(vma, vma->vm_start, pfn,
1070 vma->vm_end - vma->vm_start, vma->vm_page_prot);
1071 if (!ret)
1072 p->signal_mapped_size = vma->vm_end - vma->vm_start;
1073
1074 return ret;
1075}
1076
1077/*
1078 * Assumes that p is not going away.
1079 */
1080static void lookup_events_by_type_and_signal(struct kfd_process *p,
1081 int type, void *event_data)
1082{
1083 struct kfd_hsa_memory_exception_data *ev_data;
1084 struct kfd_event *ev;
1085 uint32_t id;
1086 bool send_signal = true;
1087
1088 ev_data = (struct kfd_hsa_memory_exception_data *) event_data;
1089
1090 rcu_read_lock();
1091
1092 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1093 idr_for_each_entry_continue(&p->event_idr, ev, id)
1094 if (ev->type == type) {
1095 send_signal = false;
1096 dev_dbg(kfd_device,
1097 "Event found: id %X type %d",
1098 ev->event_id, ev->type);
1099 spin_lock(&ev->lock);
1100 set_event(ev);
1101 if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data)
1102 ev->memory_exception_data = *ev_data;
1103 spin_unlock(&ev->lock);
1104 }
1105
1106 if (type == KFD_EVENT_TYPE_MEMORY) {
1107 dev_warn(kfd_device,
1108 "Sending SIGSEGV to process %d (pasid 0x%x)",
1109 p->lead_thread->pid, p->pasid);
1110 send_sig(SIGSEGV, p->lead_thread, 0);
1111 }
1112
1113 /* Send SIGTERM no event of type "type" has been found*/
1114 if (send_signal) {
1115 if (send_sigterm) {
1116 dev_warn(kfd_device,
1117 "Sending SIGTERM to process %d (pasid 0x%x)",
1118 p->lead_thread->pid, p->pasid);
1119 send_sig(SIGTERM, p->lead_thread, 0);
1120 } else {
1121 dev_err(kfd_device,
1122 "Process %d (pasid 0x%x) got unhandled exception",
1123 p->lead_thread->pid, p->pasid);
1124 }
1125 }
1126
1127 rcu_read_unlock();
1128}
1129
1130#ifdef KFD_SUPPORT_IOMMU_V2
1131void kfd_signal_iommu_event(struct kfd_dev *dev, u32 pasid,
1132 unsigned long address, bool is_write_requested,
1133 bool is_execute_requested)
1134{
1135 struct kfd_hsa_memory_exception_data memory_exception_data;
1136 struct vm_area_struct *vma;
1137 int user_gpu_id;
1138
1139 /*
1140 * Because we are called from arbitrary context (workqueue) as opposed
1141 * to process context, kfd_process could attempt to exit while we are
1142 * running so the lookup function increments the process ref count.
1143 */
1144 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1145 struct mm_struct *mm;
1146
1147 if (!p)
1148 return; /* Presumably process exited. */
1149
1150 /* Take a safe reference to the mm_struct, which may otherwise
1151 * disappear even while the kfd_process is still referenced.
1152 */
1153 mm = get_task_mm(p->lead_thread);
1154 if (!mm) {
1155 kfd_unref_process(p);
1156 return; /* Process is exiting */
1157 }
1158
1159 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1160 if (unlikely(user_gpu_id == -EINVAL)) {
1161 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1162 return;
1163 }
1164 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1165
1166 mmap_read_lock(mm);
1167 vma = find_vma(mm, address);
1168
1169 memory_exception_data.gpu_id = user_gpu_id;
1170 memory_exception_data.va = address;
1171 /* Set failure reason */
1172 memory_exception_data.failure.NotPresent = 1;
1173 memory_exception_data.failure.NoExecute = 0;
1174 memory_exception_data.failure.ReadOnly = 0;
1175 if (vma && address >= vma->vm_start) {
1176 memory_exception_data.failure.NotPresent = 0;
1177
1178 if (is_write_requested && !(vma->vm_flags & VM_WRITE))
1179 memory_exception_data.failure.ReadOnly = 1;
1180 else
1181 memory_exception_data.failure.ReadOnly = 0;
1182
1183 if (is_execute_requested && !(vma->vm_flags & VM_EXEC))
1184 memory_exception_data.failure.NoExecute = 1;
1185 else
1186 memory_exception_data.failure.NoExecute = 0;
1187 }
1188
1189 mmap_read_unlock(mm);
1190 mmput(mm);
1191
1192 pr_debug("notpresent %d, noexecute %d, readonly %d\n",
1193 memory_exception_data.failure.NotPresent,
1194 memory_exception_data.failure.NoExecute,
1195 memory_exception_data.failure.ReadOnly);
1196
1197 /* Workaround on Raven to not kill the process when memory is freed
1198 * before IOMMU is able to finish processing all the excessive PPRs
1199 */
1200
1201 if (KFD_GC_VERSION(dev) != IP_VERSION(9, 1, 0) &&
1202 KFD_GC_VERSION(dev) != IP_VERSION(9, 2, 2) &&
1203 KFD_GC_VERSION(dev) != IP_VERSION(9, 3, 0))
1204 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY,
1205 &memory_exception_data);
1206
1207 kfd_unref_process(p);
1208}
1209#endif /* KFD_SUPPORT_IOMMU_V2 */
1210
1211void kfd_signal_hw_exception_event(u32 pasid)
1212{
1213 /*
1214 * Because we are called from arbitrary context (workqueue) as opposed
1215 * to process context, kfd_process could attempt to exit while we are
1216 * running so the lookup function increments the process ref count.
1217 */
1218 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1219
1220 if (!p)
1221 return; /* Presumably process exited. */
1222
1223 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL);
1224 kfd_unref_process(p);
1225}
1226
1227void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid,
1228 struct kfd_vm_fault_info *info)
1229{
1230 struct kfd_event *ev;
1231 uint32_t id;
1232 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1233 struct kfd_hsa_memory_exception_data memory_exception_data;
1234 int user_gpu_id;
1235
1236 if (!p)
1237 return; /* Presumably process exited. */
1238
1239 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1240 if (unlikely(user_gpu_id == -EINVAL)) {
1241 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1242 return;
1243 }
1244
1245 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1246 memory_exception_data.gpu_id = user_gpu_id;
1247 memory_exception_data.failure.imprecise = true;
1248 /* Set failure reason */
1249 if (info) {
1250 memory_exception_data.va = (info->page_addr) << PAGE_SHIFT;
1251 memory_exception_data.failure.NotPresent =
1252 info->prot_valid ? 1 : 0;
1253 memory_exception_data.failure.NoExecute =
1254 info->prot_exec ? 1 : 0;
1255 memory_exception_data.failure.ReadOnly =
1256 info->prot_write ? 1 : 0;
1257 memory_exception_data.failure.imprecise = 0;
1258 }
1259
1260 rcu_read_lock();
1261
1262 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1263 idr_for_each_entry_continue(&p->event_idr, ev, id)
1264 if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1265 spin_lock(&ev->lock);
1266 ev->memory_exception_data = memory_exception_data;
1267 set_event(ev);
1268 spin_unlock(&ev->lock);
1269 }
1270
1271 rcu_read_unlock();
1272 kfd_unref_process(p);
1273}
1274
1275void kfd_signal_reset_event(struct kfd_dev *dev)
1276{
1277 struct kfd_hsa_hw_exception_data hw_exception_data;
1278 struct kfd_hsa_memory_exception_data memory_exception_data;
1279 struct kfd_process *p;
1280 struct kfd_event *ev;
1281 unsigned int temp;
1282 uint32_t id, idx;
1283 int reset_cause = atomic_read(&dev->sram_ecc_flag) ?
1284 KFD_HW_EXCEPTION_ECC :
1285 KFD_HW_EXCEPTION_GPU_HANG;
1286
1287 /* Whole gpu reset caused by GPU hang and memory is lost */
1288 memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1289 hw_exception_data.memory_lost = 1;
1290 hw_exception_data.reset_cause = reset_cause;
1291
1292 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1293 memory_exception_data.ErrorType = KFD_MEM_ERR_SRAM_ECC;
1294 memory_exception_data.failure.imprecise = true;
1295
1296 idx = srcu_read_lock(&kfd_processes_srcu);
1297 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1298 int user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1299
1300 if (unlikely(user_gpu_id == -EINVAL)) {
1301 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1302 continue;
1303 }
1304
1305 rcu_read_lock();
1306
1307 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1308 idr_for_each_entry_continue(&p->event_idr, ev, id) {
1309 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1310 spin_lock(&ev->lock);
1311 ev->hw_exception_data = hw_exception_data;
1312 ev->hw_exception_data.gpu_id = user_gpu_id;
1313 set_event(ev);
1314 spin_unlock(&ev->lock);
1315 }
1316 if (ev->type == KFD_EVENT_TYPE_MEMORY &&
1317 reset_cause == KFD_HW_EXCEPTION_ECC) {
1318 spin_lock(&ev->lock);
1319 ev->memory_exception_data = memory_exception_data;
1320 ev->memory_exception_data.gpu_id = user_gpu_id;
1321 set_event(ev);
1322 spin_unlock(&ev->lock);
1323 }
1324 }
1325
1326 rcu_read_unlock();
1327 }
1328 srcu_read_unlock(&kfd_processes_srcu, idx);
1329}
1330
1331void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid)
1332{
1333 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1334 struct kfd_hsa_memory_exception_data memory_exception_data;
1335 struct kfd_hsa_hw_exception_data hw_exception_data;
1336 struct kfd_event *ev;
1337 uint32_t id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1338 int user_gpu_id;
1339
1340 if (!p)
1341 return; /* Presumably process exited. */
1342
1343 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1344 if (unlikely(user_gpu_id == -EINVAL)) {
1345 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1346 return;
1347 }
1348
1349 memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1350 hw_exception_data.gpu_id = user_gpu_id;
1351 hw_exception_data.memory_lost = 1;
1352 hw_exception_data.reset_cause = KFD_HW_EXCEPTION_ECC;
1353
1354 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1355 memory_exception_data.ErrorType = KFD_MEM_ERR_POISON_CONSUMED;
1356 memory_exception_data.gpu_id = user_gpu_id;
1357 memory_exception_data.failure.imprecise = true;
1358
1359 rcu_read_lock();
1360
1361 idr_for_each_entry_continue(&p->event_idr, ev, id) {
1362 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1363 spin_lock(&ev->lock);
1364 ev->hw_exception_data = hw_exception_data;
1365 set_event(ev);
1366 spin_unlock(&ev->lock);
1367 }
1368
1369 if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1370 spin_lock(&ev->lock);
1371 ev->memory_exception_data = memory_exception_data;
1372 set_event(ev);
1373 spin_unlock(&ev->lock);
1374 }
1375 }
1376
1377 rcu_read_unlock();
1378
1379 /* user application will handle SIGBUS signal */
1380 send_sig(SIGBUS, p->lead_thread, 0);
1381
1382 kfd_unref_process(p);
1383}