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
510exit:
511 if (ret)
512 kfree(ev);
513
514 kfree(ev_priv);
515
516 mutex_unlock(&p->event_mutex);
517
518 return ret;
519}
520
521int kfd_criu_checkpoint_events(struct kfd_process *p,
522 uint8_t __user *user_priv_data,
523 uint64_t *priv_data_offset)
524{
525 struct kfd_criu_event_priv_data *ev_privs;
526 int i = 0;
527 int ret = 0;
528 struct kfd_event *ev;
529 uint32_t ev_id;
530
531 uint32_t num_events = kfd_get_num_events(p);
532
533 if (!num_events)
534 return 0;
535
536 ev_privs = kvzalloc(num_events * sizeof(*ev_privs), GFP_KERNEL);
537 if (!ev_privs)
538 return -ENOMEM;
539
540
541 idr_for_each_entry(&p->event_idr, ev, ev_id) {
542 struct kfd_criu_event_priv_data *ev_priv;
543
544 /*
545 * Currently, all events have same size of private_data, but the current ioctl's
546 * and CRIU plugin supports private_data of variable sizes
547 */
548 ev_priv = &ev_privs[i];
549
550 ev_priv->object_type = KFD_CRIU_OBJECT_TYPE_EVENT;
551
552 /* We store the user_handle with the first event */
553 if (i == 0 && p->signal_page)
554 ev_priv->user_handle = p->signal_handle;
555
556 ev_priv->event_id = ev->event_id;
557 ev_priv->auto_reset = ev->auto_reset;
558 ev_priv->type = ev->type;
559 ev_priv->signaled = ev->signaled;
560
561 if (ev_priv->type == KFD_EVENT_TYPE_MEMORY)
562 memcpy(&ev_priv->memory_exception_data,
563 &ev->memory_exception_data,
564 sizeof(struct kfd_hsa_memory_exception_data));
565 else if (ev_priv->type == KFD_EVENT_TYPE_HW_EXCEPTION)
566 memcpy(&ev_priv->hw_exception_data,
567 &ev->hw_exception_data,
568 sizeof(struct kfd_hsa_hw_exception_data));
569
570 pr_debug("Checkpointed event[%d] id = 0x%08x auto_reset = %x type = %x signaled = %x\n",
571 i,
572 ev_priv->event_id,
573 ev_priv->auto_reset,
574 ev_priv->type,
575 ev_priv->signaled);
576 i++;
577 }
578
579 ret = copy_to_user(user_priv_data + *priv_data_offset,
580 ev_privs, num_events * sizeof(*ev_privs));
581 if (ret) {
582 pr_err("Failed to copy events priv to user\n");
583 ret = -EFAULT;
584 }
585
586 *priv_data_offset += num_events * sizeof(*ev_privs);
587
588 kvfree(ev_privs);
589 return ret;
590}
591
592int kfd_get_num_events(struct kfd_process *p)
593{
594 struct kfd_event *ev;
595 uint32_t id;
596 u32 num_events = 0;
597
598 idr_for_each_entry(&p->event_idr, ev, id)
599 num_events++;
600
601 return num_events;
602}
603
604/* Assumes that p is current. */
605int kfd_event_destroy(struct kfd_process *p, uint32_t event_id)
606{
607 struct kfd_event *ev;
608 int ret = 0;
609
610 mutex_lock(&p->event_mutex);
611
612 ev = lookup_event_by_id(p, event_id);
613
614 if (ev)
615 destroy_event(p, ev);
616 else
617 ret = -EINVAL;
618
619 mutex_unlock(&p->event_mutex);
620 return ret;
621}
622
623static void set_event(struct kfd_event *ev)
624{
625 struct kfd_event_waiter *waiter;
626
627 /* Auto reset if the list is non-empty and we're waking
628 * someone. waitqueue_active is safe here because we're
629 * protected by the ev->lock, which is also held when
630 * updating the wait queues in kfd_wait_on_events.
631 */
632 ev->signaled = !ev->auto_reset || !waitqueue_active(&ev->wq);
633
634 list_for_each_entry(waiter, &ev->wq.head, wait.entry)
635 WRITE_ONCE(waiter->activated, true);
636
637 wake_up_all(&ev->wq);
638}
639
640/* Assumes that p is current. */
641int kfd_set_event(struct kfd_process *p, uint32_t event_id)
642{
643 int ret = 0;
644 struct kfd_event *ev;
645
646 rcu_read_lock();
647
648 ev = lookup_event_by_id(p, event_id);
649 if (!ev) {
650 ret = -EINVAL;
651 goto unlock_rcu;
652 }
653 spin_lock(&ev->lock);
654
655 if (event_can_be_cpu_signaled(ev))
656 set_event(ev);
657 else
658 ret = -EINVAL;
659
660 spin_unlock(&ev->lock);
661unlock_rcu:
662 rcu_read_unlock();
663 return ret;
664}
665
666static void reset_event(struct kfd_event *ev)
667{
668 ev->signaled = false;
669}
670
671/* Assumes that p is current. */
672int kfd_reset_event(struct kfd_process *p, uint32_t event_id)
673{
674 int ret = 0;
675 struct kfd_event *ev;
676
677 rcu_read_lock();
678
679 ev = lookup_event_by_id(p, event_id);
680 if (!ev) {
681 ret = -EINVAL;
682 goto unlock_rcu;
683 }
684 spin_lock(&ev->lock);
685
686 if (event_can_be_cpu_signaled(ev))
687 reset_event(ev);
688 else
689 ret = -EINVAL;
690
691 spin_unlock(&ev->lock);
692unlock_rcu:
693 rcu_read_unlock();
694 return ret;
695
696}
697
698static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev)
699{
700 WRITE_ONCE(page_slots(p->signal_page)[ev->event_id], UNSIGNALED_EVENT_SLOT);
701}
702
703static void set_event_from_interrupt(struct kfd_process *p,
704 struct kfd_event *ev)
705{
706 if (ev && event_can_be_gpu_signaled(ev)) {
707 acknowledge_signal(p, ev);
708 spin_lock(&ev->lock);
709 set_event(ev);
710 spin_unlock(&ev->lock);
711 }
712}
713
714void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
715 uint32_t valid_id_bits)
716{
717 struct kfd_event *ev = NULL;
718
719 /*
720 * Because we are called from arbitrary context (workqueue) as opposed
721 * to process context, kfd_process could attempt to exit while we are
722 * running so the lookup function increments the process ref count.
723 */
724 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
725
726 if (!p)
727 return; /* Presumably process exited. */
728
729 rcu_read_lock();
730
731 if (valid_id_bits)
732 ev = lookup_signaled_event_by_partial_id(p, partial_id,
733 valid_id_bits);
734 if (ev) {
735 set_event_from_interrupt(p, ev);
736 } else if (p->signal_page) {
737 /*
738 * Partial ID lookup failed. Assume that the event ID
739 * in the interrupt payload was invalid and do an
740 * exhaustive search of signaled events.
741 */
742 uint64_t *slots = page_slots(p->signal_page);
743 uint32_t id;
744
745 if (valid_id_bits)
746 pr_debug_ratelimited("Partial ID invalid: %u (%u valid bits)\n",
747 partial_id, valid_id_bits);
748
749 if (p->signal_event_count < KFD_SIGNAL_EVENT_LIMIT / 64) {
750 /* With relatively few events, it's faster to
751 * iterate over the event IDR
752 */
753 idr_for_each_entry(&p->event_idr, ev, id) {
754 if (id >= KFD_SIGNAL_EVENT_LIMIT)
755 break;
756
757 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT)
758 set_event_from_interrupt(p, ev);
759 }
760 } else {
761 /* With relatively many events, it's faster to
762 * iterate over the signal slots and lookup
763 * only signaled events from the IDR.
764 */
765 for (id = 1; id < KFD_SIGNAL_EVENT_LIMIT; id++)
766 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT) {
767 ev = lookup_event_by_id(p, id);
768 set_event_from_interrupt(p, ev);
769 }
770 }
771 }
772
773 rcu_read_unlock();
774 kfd_unref_process(p);
775}
776
777static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events)
778{
779 struct kfd_event_waiter *event_waiters;
780 uint32_t i;
781
782 event_waiters = kmalloc_array(num_events,
783 sizeof(struct kfd_event_waiter),
784 GFP_KERNEL);
785 if (!event_waiters)
786 return NULL;
787
788 for (i = 0; (event_waiters) && (i < num_events) ; i++) {
789 init_wait(&event_waiters[i].wait);
790 event_waiters[i].activated = false;
791 }
792
793 return event_waiters;
794}
795
796static int init_event_waiter(struct kfd_process *p,
797 struct kfd_event_waiter *waiter,
798 uint32_t event_id)
799{
800 struct kfd_event *ev = lookup_event_by_id(p, event_id);
801
802 if (!ev)
803 return -EINVAL;
804
805 spin_lock(&ev->lock);
806 waiter->event = ev;
807 waiter->activated = ev->signaled;
808 ev->signaled = ev->signaled && !ev->auto_reset;
809 if (!waiter->activated)
810 add_wait_queue(&ev->wq, &waiter->wait);
811 spin_unlock(&ev->lock);
812
813 return 0;
814}
815
816/* test_event_condition - Test condition of events being waited for
817 * @all: Return completion only if all events have signaled
818 * @num_events: Number of events to wait for
819 * @event_waiters: Array of event waiters, one per event
820 *
821 * Returns KFD_IOC_WAIT_RESULT_COMPLETE if all (or one) event(s) have
822 * signaled. Returns KFD_IOC_WAIT_RESULT_TIMEOUT if no (or not all)
823 * events have signaled. Returns KFD_IOC_WAIT_RESULT_FAIL if any of
824 * the events have been destroyed.
825 */
826static uint32_t test_event_condition(bool all, uint32_t num_events,
827 struct kfd_event_waiter *event_waiters)
828{
829 uint32_t i;
830 uint32_t activated_count = 0;
831
832 for (i = 0; i < num_events; i++) {
833 if (!READ_ONCE(event_waiters[i].event))
834 return KFD_IOC_WAIT_RESULT_FAIL;
835
836 if (READ_ONCE(event_waiters[i].activated)) {
837 if (!all)
838 return KFD_IOC_WAIT_RESULT_COMPLETE;
839
840 activated_count++;
841 }
842 }
843
844 return activated_count == num_events ?
845 KFD_IOC_WAIT_RESULT_COMPLETE : KFD_IOC_WAIT_RESULT_TIMEOUT;
846}
847
848/*
849 * Copy event specific data, if defined.
850 * Currently only memory exception events have additional data to copy to user
851 */
852static int copy_signaled_event_data(uint32_t num_events,
853 struct kfd_event_waiter *event_waiters,
854 struct kfd_event_data __user *data)
855{
856 struct kfd_hsa_memory_exception_data *src;
857 struct kfd_hsa_memory_exception_data __user *dst;
858 struct kfd_event_waiter *waiter;
859 struct kfd_event *event;
860 uint32_t i;
861
862 for (i = 0; i < num_events; i++) {
863 waiter = &event_waiters[i];
864 event = waiter->event;
865 if (!event)
866 return -EINVAL; /* event was destroyed */
867 if (waiter->activated && event->type == KFD_EVENT_TYPE_MEMORY) {
868 dst = &data[i].memory_exception_data;
869 src = &event->memory_exception_data;
870 if (copy_to_user(dst, src,
871 sizeof(struct kfd_hsa_memory_exception_data)))
872 return -EFAULT;
873 }
874 }
875
876 return 0;
877}
878
879static long user_timeout_to_jiffies(uint32_t user_timeout_ms)
880{
881 if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE)
882 return 0;
883
884 if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE)
885 return MAX_SCHEDULE_TIMEOUT;
886
887 /*
888 * msecs_to_jiffies interprets all values above 2^31-1 as infinite,
889 * but we consider them finite.
890 * This hack is wrong, but nobody is likely to notice.
891 */
892 user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF);
893
894 return msecs_to_jiffies(user_timeout_ms) + 1;
895}
896
897static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters)
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 spin_unlock(&waiters[i].event->lock);
907 }
908
909 kfree(waiters);
910}
911
912int kfd_wait_on_events(struct kfd_process *p,
913 uint32_t num_events, void __user *data,
914 bool all, uint32_t user_timeout_ms,
915 uint32_t *wait_result)
916{
917 struct kfd_event_data __user *events =
918 (struct kfd_event_data __user *) data;
919 uint32_t i;
920 int ret = 0;
921
922 struct kfd_event_waiter *event_waiters = NULL;
923 long timeout = user_timeout_to_jiffies(user_timeout_ms);
924
925 event_waiters = alloc_event_waiters(num_events);
926 if (!event_waiters) {
927 ret = -ENOMEM;
928 goto out;
929 }
930
931 /* Use p->event_mutex here to protect against concurrent creation and
932 * destruction of events while we initialize event_waiters.
933 */
934 mutex_lock(&p->event_mutex);
935
936 for (i = 0; i < num_events; i++) {
937 struct kfd_event_data event_data;
938
939 if (copy_from_user(&event_data, &events[i],
940 sizeof(struct kfd_event_data))) {
941 ret = -EFAULT;
942 goto out_unlock;
943 }
944
945 ret = init_event_waiter(p, &event_waiters[i],
946 event_data.event_id);
947 if (ret)
948 goto out_unlock;
949 }
950
951 /* Check condition once. */
952 *wait_result = test_event_condition(all, num_events, event_waiters);
953 if (*wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) {
954 ret = copy_signaled_event_data(num_events,
955 event_waiters, events);
956 goto out_unlock;
957 } else if (WARN_ON(*wait_result == KFD_IOC_WAIT_RESULT_FAIL)) {
958 /* This should not happen. Events shouldn't be
959 * destroyed while we're holding the event_mutex
960 */
961 goto out_unlock;
962 }
963
964 mutex_unlock(&p->event_mutex);
965
966 while (true) {
967 if (fatal_signal_pending(current)) {
968 ret = -EINTR;
969 break;
970 }
971
972 if (signal_pending(current)) {
973 /*
974 * This is wrong when a nonzero, non-infinite timeout
975 * is specified. We need to use
976 * ERESTARTSYS_RESTARTBLOCK, but struct restart_block
977 * contains a union with data for each user and it's
978 * in generic kernel code that I don't want to
979 * touch yet.
980 */
981 ret = -ERESTARTSYS;
982 break;
983 }
984
985 /* Set task state to interruptible sleep before
986 * checking wake-up conditions. A concurrent wake-up
987 * will put the task back into runnable state. In that
988 * case schedule_timeout will not put the task to
989 * sleep and we'll get a chance to re-check the
990 * updated conditions almost immediately. Otherwise,
991 * this race condition would lead to a soft hang or a
992 * very long sleep.
993 */
994 set_current_state(TASK_INTERRUPTIBLE);
995
996 *wait_result = test_event_condition(all, num_events,
997 event_waiters);
998 if (*wait_result != KFD_IOC_WAIT_RESULT_TIMEOUT)
999 break;
1000
1001 if (timeout <= 0)
1002 break;
1003
1004 timeout = schedule_timeout(timeout);
1005 }
1006 __set_current_state(TASK_RUNNING);
1007
1008 mutex_lock(&p->event_mutex);
1009 /* copy_signaled_event_data may sleep. So this has to happen
1010 * after the task state is set back to RUNNING.
1011 *
1012 * The event may also have been destroyed after signaling. So
1013 * copy_signaled_event_data also must confirm that the event
1014 * still exists. Therefore this must be under the p->event_mutex
1015 * which is also held when events are destroyed.
1016 */
1017 if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE)
1018 ret = copy_signaled_event_data(num_events,
1019 event_waiters, events);
1020
1021out_unlock:
1022 free_waiters(num_events, event_waiters);
1023 mutex_unlock(&p->event_mutex);
1024out:
1025 if (ret)
1026 *wait_result = KFD_IOC_WAIT_RESULT_FAIL;
1027 else if (*wait_result == KFD_IOC_WAIT_RESULT_FAIL)
1028 ret = -EIO;
1029
1030 return ret;
1031}
1032
1033int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma)
1034{
1035 unsigned long pfn;
1036 struct kfd_signal_page *page;
1037 int ret;
1038
1039 /* check required size doesn't exceed the allocated size */
1040 if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) <
1041 get_order(vma->vm_end - vma->vm_start)) {
1042 pr_err("Event page mmap requested illegal size\n");
1043 return -EINVAL;
1044 }
1045
1046 page = p->signal_page;
1047 if (!page) {
1048 /* Probably KFD bug, but mmap is user-accessible. */
1049 pr_debug("Signal page could not be found\n");
1050 return -EINVAL;
1051 }
1052
1053 pfn = __pa(page->kernel_address);
1054 pfn >>= PAGE_SHIFT;
1055
1056 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE
1057 | VM_DONTDUMP | VM_PFNMAP;
1058
1059 pr_debug("Mapping signal page\n");
1060 pr_debug(" start user address == 0x%08lx\n", vma->vm_start);
1061 pr_debug(" end user address == 0x%08lx\n", vma->vm_end);
1062 pr_debug(" pfn == 0x%016lX\n", pfn);
1063 pr_debug(" vm_flags == 0x%08lX\n", vma->vm_flags);
1064 pr_debug(" size == 0x%08lX\n",
1065 vma->vm_end - vma->vm_start);
1066
1067 page->user_address = (uint64_t __user *)vma->vm_start;
1068
1069 /* mapping the page to user process */
1070 ret = remap_pfn_range(vma, vma->vm_start, pfn,
1071 vma->vm_end - vma->vm_start, vma->vm_page_prot);
1072 if (!ret)
1073 p->signal_mapped_size = vma->vm_end - vma->vm_start;
1074
1075 return ret;
1076}
1077
1078/*
1079 * Assumes that p is not going away.
1080 */
1081static void lookup_events_by_type_and_signal(struct kfd_process *p,
1082 int type, void *event_data)
1083{
1084 struct kfd_hsa_memory_exception_data *ev_data;
1085 struct kfd_event *ev;
1086 uint32_t id;
1087 bool send_signal = true;
1088
1089 ev_data = (struct kfd_hsa_memory_exception_data *) event_data;
1090
1091 rcu_read_lock();
1092
1093 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1094 idr_for_each_entry_continue(&p->event_idr, ev, id)
1095 if (ev->type == type) {
1096 send_signal = false;
1097 dev_dbg(kfd_device,
1098 "Event found: id %X type %d",
1099 ev->event_id, ev->type);
1100 spin_lock(&ev->lock);
1101 set_event(ev);
1102 if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data)
1103 ev->memory_exception_data = *ev_data;
1104 spin_unlock(&ev->lock);
1105 }
1106
1107 if (type == KFD_EVENT_TYPE_MEMORY) {
1108 dev_warn(kfd_device,
1109 "Sending SIGSEGV to process %d (pasid 0x%x)",
1110 p->lead_thread->pid, p->pasid);
1111 send_sig(SIGSEGV, p->lead_thread, 0);
1112 }
1113
1114 /* Send SIGTERM no event of type "type" has been found*/
1115 if (send_signal) {
1116 if (send_sigterm) {
1117 dev_warn(kfd_device,
1118 "Sending SIGTERM to process %d (pasid 0x%x)",
1119 p->lead_thread->pid, p->pasid);
1120 send_sig(SIGTERM, p->lead_thread, 0);
1121 } else {
1122 dev_err(kfd_device,
1123 "Process %d (pasid 0x%x) got unhandled exception",
1124 p->lead_thread->pid, p->pasid);
1125 }
1126 }
1127
1128 rcu_read_unlock();
1129}
1130
1131#ifdef KFD_SUPPORT_IOMMU_V2
1132void kfd_signal_iommu_event(struct kfd_dev *dev, u32 pasid,
1133 unsigned long address, bool is_write_requested,
1134 bool is_execute_requested)
1135{
1136 struct kfd_hsa_memory_exception_data memory_exception_data;
1137 struct vm_area_struct *vma;
1138 int user_gpu_id;
1139
1140 /*
1141 * Because we are called from arbitrary context (workqueue) as opposed
1142 * to process context, kfd_process could attempt to exit while we are
1143 * running so the lookup function increments the process ref count.
1144 */
1145 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1146 struct mm_struct *mm;
1147
1148 if (!p)
1149 return; /* Presumably process exited. */
1150
1151 /* Take a safe reference to the mm_struct, which may otherwise
1152 * disappear even while the kfd_process is still referenced.
1153 */
1154 mm = get_task_mm(p->lead_thread);
1155 if (!mm) {
1156 kfd_unref_process(p);
1157 return; /* Process is exiting */
1158 }
1159
1160 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1161 if (unlikely(user_gpu_id == -EINVAL)) {
1162 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1163 return;
1164 }
1165 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1166
1167 mmap_read_lock(mm);
1168 vma = find_vma(mm, address);
1169
1170 memory_exception_data.gpu_id = user_gpu_id;
1171 memory_exception_data.va = address;
1172 /* Set failure reason */
1173 memory_exception_data.failure.NotPresent = 1;
1174 memory_exception_data.failure.NoExecute = 0;
1175 memory_exception_data.failure.ReadOnly = 0;
1176 if (vma && address >= vma->vm_start) {
1177 memory_exception_data.failure.NotPresent = 0;
1178
1179 if (is_write_requested && !(vma->vm_flags & VM_WRITE))
1180 memory_exception_data.failure.ReadOnly = 1;
1181 else
1182 memory_exception_data.failure.ReadOnly = 0;
1183
1184 if (is_execute_requested && !(vma->vm_flags & VM_EXEC))
1185 memory_exception_data.failure.NoExecute = 1;
1186 else
1187 memory_exception_data.failure.NoExecute = 0;
1188 }
1189
1190 mmap_read_unlock(mm);
1191 mmput(mm);
1192
1193 pr_debug("notpresent %d, noexecute %d, readonly %d\n",
1194 memory_exception_data.failure.NotPresent,
1195 memory_exception_data.failure.NoExecute,
1196 memory_exception_data.failure.ReadOnly);
1197
1198 /* Workaround on Raven to not kill the process when memory is freed
1199 * before IOMMU is able to finish processing all the excessive PPRs
1200 */
1201
1202 if (KFD_GC_VERSION(dev) != IP_VERSION(9, 1, 0) &&
1203 KFD_GC_VERSION(dev) != IP_VERSION(9, 2, 2) &&
1204 KFD_GC_VERSION(dev) != IP_VERSION(9, 3, 0))
1205 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY,
1206 &memory_exception_data);
1207
1208 kfd_unref_process(p);
1209}
1210#endif /* KFD_SUPPORT_IOMMU_V2 */
1211
1212void kfd_signal_hw_exception_event(u32 pasid)
1213{
1214 /*
1215 * Because we are called from arbitrary context (workqueue) as opposed
1216 * to process context, kfd_process could attempt to exit while we are
1217 * running so the lookup function increments the process ref count.
1218 */
1219 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1220
1221 if (!p)
1222 return; /* Presumably process exited. */
1223
1224 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL);
1225 kfd_unref_process(p);
1226}
1227
1228void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid,
1229 struct kfd_vm_fault_info *info)
1230{
1231 struct kfd_event *ev;
1232 uint32_t id;
1233 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1234 struct kfd_hsa_memory_exception_data memory_exception_data;
1235 int user_gpu_id;
1236
1237 if (!p)
1238 return; /* Presumably process exited. */
1239
1240 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1241 if (unlikely(user_gpu_id == -EINVAL)) {
1242 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1243 return;
1244 }
1245
1246 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1247 memory_exception_data.gpu_id = user_gpu_id;
1248 memory_exception_data.failure.imprecise = true;
1249 /* Set failure reason */
1250 if (info) {
1251 memory_exception_data.va = (info->page_addr) << PAGE_SHIFT;
1252 memory_exception_data.failure.NotPresent =
1253 info->prot_valid ? 1 : 0;
1254 memory_exception_data.failure.NoExecute =
1255 info->prot_exec ? 1 : 0;
1256 memory_exception_data.failure.ReadOnly =
1257 info->prot_write ? 1 : 0;
1258 memory_exception_data.failure.imprecise = 0;
1259 }
1260
1261 rcu_read_lock();
1262
1263 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1264 idr_for_each_entry_continue(&p->event_idr, ev, id)
1265 if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1266 spin_lock(&ev->lock);
1267 ev->memory_exception_data = memory_exception_data;
1268 set_event(ev);
1269 spin_unlock(&ev->lock);
1270 }
1271
1272 rcu_read_unlock();
1273 kfd_unref_process(p);
1274}
1275
1276void kfd_signal_reset_event(struct kfd_dev *dev)
1277{
1278 struct kfd_hsa_hw_exception_data hw_exception_data;
1279 struct kfd_hsa_memory_exception_data memory_exception_data;
1280 struct kfd_process *p;
1281 struct kfd_event *ev;
1282 unsigned int temp;
1283 uint32_t id, idx;
1284 int reset_cause = atomic_read(&dev->sram_ecc_flag) ?
1285 KFD_HW_EXCEPTION_ECC :
1286 KFD_HW_EXCEPTION_GPU_HANG;
1287
1288 /* Whole gpu reset caused by GPU hang and memory is lost */
1289 memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1290 hw_exception_data.memory_lost = 1;
1291 hw_exception_data.reset_cause = reset_cause;
1292
1293 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1294 memory_exception_data.ErrorType = KFD_MEM_ERR_SRAM_ECC;
1295 memory_exception_data.failure.imprecise = true;
1296
1297 idx = srcu_read_lock(&kfd_processes_srcu);
1298 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1299 int user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1300
1301 if (unlikely(user_gpu_id == -EINVAL)) {
1302 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1303 continue;
1304 }
1305
1306 rcu_read_lock();
1307
1308 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1309 idr_for_each_entry_continue(&p->event_idr, ev, id) {
1310 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1311 spin_lock(&ev->lock);
1312 ev->hw_exception_data = hw_exception_data;
1313 ev->hw_exception_data.gpu_id = user_gpu_id;
1314 set_event(ev);
1315 spin_unlock(&ev->lock);
1316 }
1317 if (ev->type == KFD_EVENT_TYPE_MEMORY &&
1318 reset_cause == KFD_HW_EXCEPTION_ECC) {
1319 spin_lock(&ev->lock);
1320 ev->memory_exception_data = memory_exception_data;
1321 ev->memory_exception_data.gpu_id = user_gpu_id;
1322 set_event(ev);
1323 spin_unlock(&ev->lock);
1324 }
1325 }
1326
1327 rcu_read_unlock();
1328 }
1329 srcu_read_unlock(&kfd_processes_srcu, idx);
1330}
1331
1332void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid)
1333{
1334 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1335 struct kfd_hsa_memory_exception_data memory_exception_data;
1336 struct kfd_hsa_hw_exception_data hw_exception_data;
1337 struct kfd_event *ev;
1338 uint32_t id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1339 int user_gpu_id;
1340
1341 if (!p)
1342 return; /* Presumably process exited. */
1343
1344 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1345 if (unlikely(user_gpu_id == -EINVAL)) {
1346 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1347 return;
1348 }
1349
1350 memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1351 hw_exception_data.gpu_id = user_gpu_id;
1352 hw_exception_data.memory_lost = 1;
1353 hw_exception_data.reset_cause = KFD_HW_EXCEPTION_ECC;
1354
1355 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1356 memory_exception_data.ErrorType = KFD_MEM_ERR_POISON_CONSUMED;
1357 memory_exception_data.gpu_id = user_gpu_id;
1358 memory_exception_data.failure.imprecise = true;
1359
1360 rcu_read_lock();
1361
1362 idr_for_each_entry_continue(&p->event_idr, ev, id) {
1363 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1364 spin_lock(&ev->lock);
1365 ev->hw_exception_data = hw_exception_data;
1366 set_event(ev);
1367 spin_unlock(&ev->lock);
1368 }
1369
1370 if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1371 spin_lock(&ev->lock);
1372 ev->memory_exception_data = memory_exception_data;
1373 set_event(ev);
1374 spin_unlock(&ev->lock);
1375 }
1376 }
1377
1378 rcu_read_unlock();
1379
1380 /* user application will handle SIGBUS signal */
1381 send_sig(SIGBUS, p->lead_thread, 0);
1382
1383 kfd_unref_process(p);
1384}
1385

source code of linux/drivers/gpu/drm/amd/amdkfd/kfd_events.c