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_device_queue_manager.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 bool event_age_enabled; /* set to true when last_event_age is non-zero */
45};
46
47/*
48 * Each signal event needs a 64-bit signal slot where the signaler will write
49 * a 1 before sending an interrupt. (This is needed because some interrupts
50 * do not contain enough spare data bits to identify an event.)
51 * We get whole pages and map them to the process VA.
52 * Individual signal events use their event_id as slot index.
53 */
54struct kfd_signal_page {
55 uint64_t *kernel_address;
56 uint64_t __user *user_address;
57 bool need_to_free_pages;
58};
59
60static uint64_t *page_slots(struct kfd_signal_page *page)
61{
62 return page->kernel_address;
63}
64
65static struct kfd_signal_page *allocate_signal_page(struct kfd_process *p)
66{
67 void *backing_store;
68 struct kfd_signal_page *page;
69
70 page = kzalloc(sizeof(*page), GFP_KERNEL);
71 if (!page)
72 return NULL;
73
74 backing_store = (void *) __get_free_pages(GFP_KERNEL,
75 get_order(KFD_SIGNAL_EVENT_LIMIT * 8));
76 if (!backing_store)
77 goto fail_alloc_signal_store;
78
79 /* Initialize all events to unsignaled */
80 memset(backing_store, (uint8_t) UNSIGNALED_EVENT_SLOT,
81 KFD_SIGNAL_EVENT_LIMIT * 8);
82
83 page->kernel_address = backing_store;
84 page->need_to_free_pages = true;
85 pr_debug("Allocated new event signal page at %p, for process %p\n",
86 page, p);
87
88 return page;
89
90fail_alloc_signal_store:
91 kfree(objp: page);
92 return NULL;
93}
94
95static int allocate_event_notification_slot(struct kfd_process *p,
96 struct kfd_event *ev,
97 const int *restore_id)
98{
99 int id;
100
101 if (!p->signal_page) {
102 p->signal_page = allocate_signal_page(p);
103 if (!p->signal_page)
104 return -ENOMEM;
105 /* Oldest user mode expects 256 event slots */
106 p->signal_mapped_size = 256*8;
107 }
108
109 if (restore_id) {
110 id = idr_alloc(&p->event_idr, ptr: ev, start: *restore_id, end: *restore_id + 1,
111 GFP_KERNEL);
112 } else {
113 /*
114 * Compatibility with old user mode: Only use signal slots
115 * user mode has mapped, may be less than
116 * KFD_SIGNAL_EVENT_LIMIT. This also allows future increase
117 * of the event limit without breaking user mode.
118 */
119 id = idr_alloc(&p->event_idr, ptr: ev, start: 0, end: p->signal_mapped_size / 8,
120 GFP_KERNEL);
121 }
122 if (id < 0)
123 return id;
124
125 ev->event_id = id;
126 page_slots(page: p->signal_page)[id] = UNSIGNALED_EVENT_SLOT;
127
128 return 0;
129}
130
131/*
132 * Assumes that p->event_mutex or rcu_readlock is held and of course that p is
133 * not going away.
134 */
135static struct kfd_event *lookup_event_by_id(struct kfd_process *p, uint32_t id)
136{
137 return idr_find(&p->event_idr, id);
138}
139
140/**
141 * lookup_signaled_event_by_partial_id - Lookup signaled event from partial ID
142 * @p: Pointer to struct kfd_process
143 * @id: ID to look up
144 * @bits: Number of valid bits in @id
145 *
146 * Finds the first signaled event with a matching partial ID. If no
147 * matching signaled event is found, returns NULL. In that case the
148 * caller should assume that the partial ID is invalid and do an
149 * exhaustive search of all siglaned events.
150 *
151 * If multiple events with the same partial ID signal at the same
152 * time, they will be found one interrupt at a time, not necessarily
153 * in the same order the interrupts occurred. As long as the number of
154 * interrupts is correct, all signaled events will be seen by the
155 * driver.
156 */
157static struct kfd_event *lookup_signaled_event_by_partial_id(
158 struct kfd_process *p, uint32_t id, uint32_t bits)
159{
160 struct kfd_event *ev;
161
162 if (!p->signal_page || id >= KFD_SIGNAL_EVENT_LIMIT)
163 return NULL;
164
165 /* Fast path for the common case that @id is not a partial ID
166 * and we only need a single lookup.
167 */
168 if (bits > 31 || (1U << bits) >= KFD_SIGNAL_EVENT_LIMIT) {
169 if (page_slots(page: p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
170 return NULL;
171
172 return idr_find(&p->event_idr, id);
173 }
174
175 /* General case for partial IDs: Iterate over all matching IDs
176 * and find the first one that has signaled.
177 */
178 for (ev = NULL; id < KFD_SIGNAL_EVENT_LIMIT && !ev; id += 1U << bits) {
179 if (page_slots(page: p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
180 continue;
181
182 ev = idr_find(&p->event_idr, id);
183 }
184
185 return ev;
186}
187
188static int create_signal_event(struct file *devkfd, struct kfd_process *p,
189 struct kfd_event *ev, const int *restore_id)
190{
191 int ret;
192
193 if (p->signal_mapped_size &&
194 p->signal_event_count == p->signal_mapped_size / 8) {
195 if (!p->signal_event_limit_reached) {
196 pr_debug("Signal event wasn't created because limit was reached\n");
197 p->signal_event_limit_reached = true;
198 }
199 return -ENOSPC;
200 }
201
202 ret = allocate_event_notification_slot(p, ev, restore_id);
203 if (ret) {
204 pr_warn("Signal event wasn't created because out of kernel memory\n");
205 return ret;
206 }
207
208 p->signal_event_count++;
209
210 ev->user_signal_address = &p->signal_page->user_address[ev->event_id];
211 pr_debug("Signal event number %zu created with id %d, address %p\n",
212 p->signal_event_count, ev->event_id,
213 ev->user_signal_address);
214
215 return 0;
216}
217
218static int create_other_event(struct kfd_process *p, struct kfd_event *ev, const int *restore_id)
219{
220 int id;
221
222 if (restore_id)
223 id = idr_alloc(&p->event_idr, ptr: ev, start: *restore_id, end: *restore_id + 1,
224 GFP_KERNEL);
225 else
226 /* Cast KFD_LAST_NONSIGNAL_EVENT to uint32_t. This allows an
227 * intentional integer overflow to -1 without a compiler
228 * warning. idr_alloc treats a negative value as "maximum
229 * signed integer".
230 */
231 id = idr_alloc(&p->event_idr, ptr: ev, KFD_FIRST_NONSIGNAL_EVENT_ID,
232 end: (uint32_t)KFD_LAST_NONSIGNAL_EVENT_ID + 1,
233 GFP_KERNEL);
234
235 if (id < 0)
236 return id;
237 ev->event_id = id;
238
239 return 0;
240}
241
242int kfd_event_init_process(struct kfd_process *p)
243{
244 int id;
245
246 mutex_init(&p->event_mutex);
247 idr_init(idr: &p->event_idr);
248 p->signal_page = NULL;
249 p->signal_event_count = 1;
250 /* Allocate event ID 0. It is used for a fast path to ignore bogus events
251 * that are sent by the CP without a context ID
252 */
253 id = idr_alloc(&p->event_idr, NULL, start: 0, end: 1, GFP_KERNEL);
254 if (id < 0) {
255 idr_destroy(&p->event_idr);
256 mutex_destroy(lock: &p->event_mutex);
257 return id;
258 }
259 return 0;
260}
261
262static void destroy_event(struct kfd_process *p, struct kfd_event *ev)
263{
264 struct kfd_event_waiter *waiter;
265
266 /* Wake up pending waiters. They will return failure */
267 spin_lock(lock: &ev->lock);
268 list_for_each_entry(waiter, &ev->wq.head, wait.entry)
269 WRITE_ONCE(waiter->event, NULL);
270 wake_up_all(&ev->wq);
271 spin_unlock(lock: &ev->lock);
272
273 if (ev->type == KFD_EVENT_TYPE_SIGNAL ||
274 ev->type == KFD_EVENT_TYPE_DEBUG)
275 p->signal_event_count--;
276
277 idr_remove(&p->event_idr, id: ev->event_id);
278 kfree_rcu(ev, rcu);
279}
280
281static void destroy_events(struct kfd_process *p)
282{
283 struct kfd_event *ev;
284 uint32_t id;
285
286 idr_for_each_entry(&p->event_idr, ev, id)
287 if (ev)
288 destroy_event(p, ev);
289 idr_destroy(&p->event_idr);
290 mutex_destroy(lock: &p->event_mutex);
291}
292
293/*
294 * We assume that the process is being destroyed and there is no need to
295 * unmap the pages or keep bookkeeping data in order.
296 */
297static void shutdown_signal_page(struct kfd_process *p)
298{
299 struct kfd_signal_page *page = p->signal_page;
300
301 if (page) {
302 if (page->need_to_free_pages)
303 free_pages(addr: (unsigned long)page->kernel_address,
304 order: get_order(KFD_SIGNAL_EVENT_LIMIT * 8));
305 kfree(objp: page);
306 }
307}
308
309void kfd_event_free_process(struct kfd_process *p)
310{
311 destroy_events(p);
312 shutdown_signal_page(p);
313}
314
315static bool event_can_be_gpu_signaled(const struct kfd_event *ev)
316{
317 return ev->type == KFD_EVENT_TYPE_SIGNAL ||
318 ev->type == KFD_EVENT_TYPE_DEBUG;
319}
320
321static bool event_can_be_cpu_signaled(const struct kfd_event *ev)
322{
323 return ev->type == KFD_EVENT_TYPE_SIGNAL;
324}
325
326static int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
327 uint64_t size, uint64_t user_handle)
328{
329 struct kfd_signal_page *page;
330
331 if (p->signal_page)
332 return -EBUSY;
333
334 page = kzalloc(sizeof(*page), GFP_KERNEL);
335 if (!page)
336 return -ENOMEM;
337
338 /* Initialize all events to unsignaled */
339 memset(kernel_address, (uint8_t) UNSIGNALED_EVENT_SLOT,
340 KFD_SIGNAL_EVENT_LIMIT * 8);
341
342 page->kernel_address = kernel_address;
343
344 p->signal_page = page;
345 p->signal_mapped_size = size;
346 p->signal_handle = user_handle;
347 return 0;
348}
349
350int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset)
351{
352 struct kfd_node *kfd;
353 struct kfd_process_device *pdd;
354 void *mem, *kern_addr;
355 uint64_t size;
356 int err = 0;
357
358 if (p->signal_page) {
359 pr_err("Event page is already set\n");
360 return -EINVAL;
361 }
362
363 pdd = kfd_process_device_data_by_id(process: p, GET_GPU_ID(event_page_offset));
364 if (!pdd) {
365 pr_err("Getting device by id failed in %s\n", __func__);
366 return -EINVAL;
367 }
368 kfd = pdd->dev;
369
370 pdd = kfd_bind_process_to_device(dev: kfd, p);
371 if (IS_ERR(ptr: pdd))
372 return PTR_ERR(ptr: pdd);
373
374 mem = kfd_process_device_translate_handle(p: pdd,
375 GET_IDR_HANDLE(event_page_offset));
376 if (!mem) {
377 pr_err("Can't find BO, offset is 0x%llx\n", event_page_offset);
378 return -EINVAL;
379 }
380
381 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(mem, kptr: &kern_addr, size: &size);
382 if (err) {
383 pr_err("Failed to map event page to kernel\n");
384 return err;
385 }
386
387 err = kfd_event_page_set(p, kernel_address: kern_addr, size, user_handle: event_page_offset);
388 if (err) {
389 pr_err("Failed to set event page\n");
390 amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
391 return err;
392 }
393 return err;
394}
395
396int kfd_event_create(struct file *devkfd, struct kfd_process *p,
397 uint32_t event_type, bool auto_reset, uint32_t node_id,
398 uint32_t *event_id, uint32_t *event_trigger_data,
399 uint64_t *event_page_offset, uint32_t *event_slot_index)
400{
401 int ret = 0;
402 struct kfd_event *ev = kzalloc(sizeof(*ev), GFP_KERNEL);
403
404 if (!ev)
405 return -ENOMEM;
406
407 ev->type = event_type;
408 ev->auto_reset = auto_reset;
409 ev->signaled = false;
410
411 spin_lock_init(&ev->lock);
412 init_waitqueue_head(&ev->wq);
413
414 *event_page_offset = 0;
415
416 mutex_lock(&p->event_mutex);
417
418 switch (event_type) {
419 case KFD_EVENT_TYPE_SIGNAL:
420 case KFD_EVENT_TYPE_DEBUG:
421 ret = create_signal_event(devkfd, p, ev, NULL);
422 if (!ret) {
423 *event_page_offset = KFD_MMAP_TYPE_EVENTS;
424 *event_slot_index = ev->event_id;
425 }
426 break;
427 default:
428 ret = create_other_event(p, ev, NULL);
429 break;
430 }
431
432 if (!ret) {
433 *event_id = ev->event_id;
434 *event_trigger_data = ev->event_id;
435 ev->event_age = 1;
436 } else {
437 kfree(objp: ev);
438 }
439
440 mutex_unlock(lock: &p->event_mutex);
441
442 return ret;
443}
444
445int kfd_criu_restore_event(struct file *devkfd,
446 struct kfd_process *p,
447 uint8_t __user *user_priv_ptr,
448 uint64_t *priv_data_offset,
449 uint64_t max_priv_data_size)
450{
451 struct kfd_criu_event_priv_data *ev_priv;
452 struct kfd_event *ev = NULL;
453 int ret = 0;
454
455 ev_priv = kmalloc(sizeof(*ev_priv), GFP_KERNEL);
456 if (!ev_priv)
457 return -ENOMEM;
458
459 ev = kzalloc(sizeof(*ev), GFP_KERNEL);
460 if (!ev) {
461 ret = -ENOMEM;
462 goto exit;
463 }
464
465 if (*priv_data_offset + sizeof(*ev_priv) > max_priv_data_size) {
466 ret = -EINVAL;
467 goto exit;
468 }
469
470 ret = copy_from_user(to: ev_priv, from: user_priv_ptr + *priv_data_offset, n: sizeof(*ev_priv));
471 if (ret) {
472 ret = -EFAULT;
473 goto exit;
474 }
475 *priv_data_offset += sizeof(*ev_priv);
476
477 if (ev_priv->user_handle) {
478 ret = kfd_kmap_event_page(p, event_page_offset: ev_priv->user_handle);
479 if (ret)
480 goto exit;
481 }
482
483 ev->type = ev_priv->type;
484 ev->auto_reset = ev_priv->auto_reset;
485 ev->signaled = ev_priv->signaled;
486
487 spin_lock_init(&ev->lock);
488 init_waitqueue_head(&ev->wq);
489
490 mutex_lock(&p->event_mutex);
491 switch (ev->type) {
492 case KFD_EVENT_TYPE_SIGNAL:
493 case KFD_EVENT_TYPE_DEBUG:
494 ret = create_signal_event(devkfd, p, ev, restore_id: &ev_priv->event_id);
495 break;
496 case KFD_EVENT_TYPE_MEMORY:
497 memcpy(&ev->memory_exception_data,
498 &ev_priv->memory_exception_data,
499 sizeof(struct kfd_hsa_memory_exception_data));
500
501 ret = create_other_event(p, ev, restore_id: &ev_priv->event_id);
502 break;
503 case KFD_EVENT_TYPE_HW_EXCEPTION:
504 memcpy(&ev->hw_exception_data,
505 &ev_priv->hw_exception_data,
506 sizeof(struct kfd_hsa_hw_exception_data));
507
508 ret = create_other_event(p, ev, restore_id: &ev_priv->event_id);
509 break;
510 }
511 mutex_unlock(lock: &p->event_mutex);
512
513exit:
514 if (ret)
515 kfree(objp: ev);
516
517 kfree(objp: ev_priv);
518
519 return ret;
520}
521
522int kfd_criu_checkpoint_events(struct kfd_process *p,
523 uint8_t __user *user_priv_data,
524 uint64_t *priv_data_offset)
525{
526 struct kfd_criu_event_priv_data *ev_privs;
527 int i = 0;
528 int ret = 0;
529 struct kfd_event *ev;
530 uint32_t ev_id;
531
532 uint32_t num_events = kfd_get_num_events(p);
533
534 if (!num_events)
535 return 0;
536
537 ev_privs = kvzalloc(num_events * sizeof(*ev_privs), GFP_KERNEL);
538 if (!ev_privs)
539 return -ENOMEM;
540
541
542 idr_for_each_entry(&p->event_idr, ev, ev_id) {
543 struct kfd_criu_event_priv_data *ev_priv;
544
545 /*
546 * Currently, all events have same size of private_data, but the current ioctl's
547 * and CRIU plugin supports private_data of variable sizes
548 */
549 ev_priv = &ev_privs[i];
550
551 ev_priv->object_type = KFD_CRIU_OBJECT_TYPE_EVENT;
552
553 /* We store the user_handle with the first event */
554 if (i == 0 && p->signal_page)
555 ev_priv->user_handle = p->signal_handle;
556
557 ev_priv->event_id = ev->event_id;
558 ev_priv->auto_reset = ev->auto_reset;
559 ev_priv->type = ev->type;
560 ev_priv->signaled = ev->signaled;
561
562 if (ev_priv->type == KFD_EVENT_TYPE_MEMORY)
563 memcpy(&ev_priv->memory_exception_data,
564 &ev->memory_exception_data,
565 sizeof(struct kfd_hsa_memory_exception_data));
566 else if (ev_priv->type == KFD_EVENT_TYPE_HW_EXCEPTION)
567 memcpy(&ev_priv->hw_exception_data,
568 &ev->hw_exception_data,
569 sizeof(struct kfd_hsa_hw_exception_data));
570
571 pr_debug("Checkpointed event[%d] id = 0x%08x auto_reset = %x type = %x signaled = %x\n",
572 i,
573 ev_priv->event_id,
574 ev_priv->auto_reset,
575 ev_priv->type,
576 ev_priv->signaled);
577 i++;
578 }
579
580 ret = copy_to_user(to: user_priv_data + *priv_data_offset,
581 from: ev_privs, n: num_events * sizeof(*ev_privs));
582 if (ret) {
583 pr_err("Failed to copy events priv to user\n");
584 ret = -EFAULT;
585 }
586
587 *priv_data_offset += num_events * sizeof(*ev_privs);
588
589 kvfree(addr: ev_privs);
590 return ret;
591}
592
593int kfd_get_num_events(struct kfd_process *p)
594{
595 struct kfd_event *ev;
596 uint32_t id;
597 u32 num_events = 0;
598
599 idr_for_each_entry(&p->event_idr, ev, id)
600 num_events++;
601
602 return num_events;
603}
604
605/* Assumes that p is current. */
606int kfd_event_destroy(struct kfd_process *p, uint32_t event_id)
607{
608 struct kfd_event *ev;
609 int ret = 0;
610
611 mutex_lock(&p->event_mutex);
612
613 ev = lookup_event_by_id(p, id: event_id);
614
615 if (ev)
616 destroy_event(p, ev);
617 else
618 ret = -EINVAL;
619
620 mutex_unlock(lock: &p->event_mutex);
621 return ret;
622}
623
624static void set_event(struct kfd_event *ev)
625{
626 struct kfd_event_waiter *waiter;
627
628 /* Auto reset if the list is non-empty and we're waking
629 * someone. waitqueue_active is safe here because we're
630 * protected by the ev->lock, which is also held when
631 * updating the wait queues in kfd_wait_on_events.
632 */
633 ev->signaled = !ev->auto_reset || !waitqueue_active(wq_head: &ev->wq);
634 if (!(++ev->event_age)) {
635 /* Never wrap back to reserved/default event age 0/1 */
636 ev->event_age = 2;
637 WARN_ONCE(1, "event_age wrap back!");
638 }
639
640 list_for_each_entry(waiter, &ev->wq.head, wait.entry)
641 WRITE_ONCE(waiter->activated, true);
642
643 wake_up_all(&ev->wq);
644}
645
646/* Assumes that p is current. */
647int kfd_set_event(struct kfd_process *p, uint32_t event_id)
648{
649 int ret = 0;
650 struct kfd_event *ev;
651
652 rcu_read_lock();
653
654 ev = lookup_event_by_id(p, id: event_id);
655 if (!ev) {
656 ret = -EINVAL;
657 goto unlock_rcu;
658 }
659 spin_lock(lock: &ev->lock);
660
661 if (event_can_be_cpu_signaled(ev))
662 set_event(ev);
663 else
664 ret = -EINVAL;
665
666 spin_unlock(lock: &ev->lock);
667unlock_rcu:
668 rcu_read_unlock();
669 return ret;
670}
671
672static void reset_event(struct kfd_event *ev)
673{
674 ev->signaled = false;
675}
676
677/* Assumes that p is current. */
678int kfd_reset_event(struct kfd_process *p, uint32_t event_id)
679{
680 int ret = 0;
681 struct kfd_event *ev;
682
683 rcu_read_lock();
684
685 ev = lookup_event_by_id(p, id: event_id);
686 if (!ev) {
687 ret = -EINVAL;
688 goto unlock_rcu;
689 }
690 spin_lock(lock: &ev->lock);
691
692 if (event_can_be_cpu_signaled(ev))
693 reset_event(ev);
694 else
695 ret = -EINVAL;
696
697 spin_unlock(lock: &ev->lock);
698unlock_rcu:
699 rcu_read_unlock();
700 return ret;
701
702}
703
704static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev)
705{
706 WRITE_ONCE(page_slots(p->signal_page)[ev->event_id], UNSIGNALED_EVENT_SLOT);
707}
708
709static void set_event_from_interrupt(struct kfd_process *p,
710 struct kfd_event *ev)
711{
712 if (ev && event_can_be_gpu_signaled(ev)) {
713 acknowledge_signal(p, ev);
714 spin_lock(lock: &ev->lock);
715 set_event(ev);
716 spin_unlock(lock: &ev->lock);
717 }
718}
719
720void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
721 uint32_t valid_id_bits)
722{
723 struct kfd_event *ev = NULL;
724
725 /*
726 * Because we are called from arbitrary context (workqueue) as opposed
727 * to process context, kfd_process could attempt to exit while we are
728 * running so the lookup function increments the process ref count.
729 */
730 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid, NULL);
731
732 if (!p)
733 return; /* Presumably process exited. */
734
735 rcu_read_lock();
736
737 if (valid_id_bits)
738 ev = lookup_signaled_event_by_partial_id(p, id: partial_id,
739 bits: valid_id_bits);
740 if (ev) {
741 set_event_from_interrupt(p, ev);
742 } else if (p->signal_page) {
743 /*
744 * Partial ID lookup failed. Assume that the event ID
745 * in the interrupt payload was invalid and do an
746 * exhaustive search of signaled events.
747 */
748 uint64_t *slots = page_slots(page: p->signal_page);
749 uint32_t id;
750
751 if (valid_id_bits)
752 pr_debug_ratelimited("Partial ID invalid: %u (%u valid bits)\n",
753 partial_id, valid_id_bits);
754
755 if (p->signal_event_count < KFD_SIGNAL_EVENT_LIMIT / 64) {
756 /* With relatively few events, it's faster to
757 * iterate over the event IDR
758 */
759 idr_for_each_entry(&p->event_idr, ev, id) {
760 if (id >= KFD_SIGNAL_EVENT_LIMIT)
761 break;
762
763 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT)
764 set_event_from_interrupt(p, ev);
765 }
766 } else {
767 /* With relatively many events, it's faster to
768 * iterate over the signal slots and lookup
769 * only signaled events from the IDR.
770 */
771 for (id = 1; id < KFD_SIGNAL_EVENT_LIMIT; id++)
772 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT) {
773 ev = lookup_event_by_id(p, id);
774 set_event_from_interrupt(p, ev);
775 }
776 }
777 }
778
779 rcu_read_unlock();
780 kfd_unref_process(p);
781}
782
783static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events)
784{
785 struct kfd_event_waiter *event_waiters;
786 uint32_t i;
787
788 event_waiters = kcalloc(num_events, sizeof(struct kfd_event_waiter),
789 GFP_KERNEL);
790 if (!event_waiters)
791 return NULL;
792
793 for (i = 0; i < num_events; i++)
794 init_wait(&event_waiters[i].wait);
795
796 return event_waiters;
797}
798
799static int init_event_waiter(struct kfd_process *p,
800 struct kfd_event_waiter *waiter,
801 struct kfd_event_data *event_data)
802{
803 struct kfd_event *ev = lookup_event_by_id(p, id: event_data->event_id);
804
805 if (!ev)
806 return -EINVAL;
807
808 spin_lock(lock: &ev->lock);
809 waiter->event = ev;
810 waiter->activated = ev->signaled;
811 ev->signaled = ev->signaled && !ev->auto_reset;
812
813 /* last_event_age = 0 reserved for backward compatible */
814 if (waiter->event->type == KFD_EVENT_TYPE_SIGNAL &&
815 event_data->signal_event_data.last_event_age) {
816 waiter->event_age_enabled = true;
817 if (ev->event_age != event_data->signal_event_data.last_event_age)
818 waiter->activated = true;
819 }
820
821 if (!waiter->activated)
822 add_wait_queue(wq_head: &ev->wq, wq_entry: &waiter->wait);
823 spin_unlock(lock: &ev->lock);
824
825 return 0;
826}
827
828/* test_event_condition - Test condition of events being waited for
829 * @all: Return completion only if all events have signaled
830 * @num_events: Number of events to wait for
831 * @event_waiters: Array of event waiters, one per event
832 *
833 * Returns KFD_IOC_WAIT_RESULT_COMPLETE if all (or one) event(s) have
834 * signaled. Returns KFD_IOC_WAIT_RESULT_TIMEOUT if no (or not all)
835 * events have signaled. Returns KFD_IOC_WAIT_RESULT_FAIL if any of
836 * the events have been destroyed.
837 */
838static uint32_t test_event_condition(bool all, uint32_t num_events,
839 struct kfd_event_waiter *event_waiters)
840{
841 uint32_t i;
842 uint32_t activated_count = 0;
843
844 for (i = 0; i < num_events; i++) {
845 if (!READ_ONCE(event_waiters[i].event))
846 return KFD_IOC_WAIT_RESULT_FAIL;
847
848 if (READ_ONCE(event_waiters[i].activated)) {
849 if (!all)
850 return KFD_IOC_WAIT_RESULT_COMPLETE;
851
852 activated_count++;
853 }
854 }
855
856 return activated_count == num_events ?
857 KFD_IOC_WAIT_RESULT_COMPLETE : KFD_IOC_WAIT_RESULT_TIMEOUT;
858}
859
860/*
861 * Copy event specific data, if defined.
862 * Currently only memory exception events have additional data to copy to user
863 */
864static int copy_signaled_event_data(uint32_t num_events,
865 struct kfd_event_waiter *event_waiters,
866 struct kfd_event_data __user *data)
867{
868 void *src;
869 void __user *dst;
870 struct kfd_event_waiter *waiter;
871 struct kfd_event *event;
872 uint32_t i, size = 0;
873
874 for (i = 0; i < num_events; i++) {
875 waiter = &event_waiters[i];
876 event = waiter->event;
877 if (!event)
878 return -EINVAL; /* event was destroyed */
879 if (waiter->activated) {
880 if (event->type == KFD_EVENT_TYPE_MEMORY) {
881 dst = &data[i].memory_exception_data;
882 src = &event->memory_exception_data;
883 size = sizeof(struct kfd_hsa_memory_exception_data);
884 } else if (event->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
885 dst = &data[i].memory_exception_data;
886 src = &event->hw_exception_data;
887 size = sizeof(struct kfd_hsa_hw_exception_data);
888 } else if (event->type == KFD_EVENT_TYPE_SIGNAL &&
889 waiter->event_age_enabled) {
890 dst = &data[i].signal_event_data.last_event_age;
891 src = &event->event_age;
892 size = sizeof(u64);
893 }
894 if (size && copy_to_user(to: dst, from: src, n: size))
895 return -EFAULT;
896 }
897 }
898
899 return 0;
900}
901
902static long user_timeout_to_jiffies(uint32_t user_timeout_ms)
903{
904 if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE)
905 return 0;
906
907 if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE)
908 return MAX_SCHEDULE_TIMEOUT;
909
910 /*
911 * msecs_to_jiffies interprets all values above 2^31-1 as infinite,
912 * but we consider them finite.
913 * This hack is wrong, but nobody is likely to notice.
914 */
915 user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF);
916
917 return msecs_to_jiffies(m: user_timeout_ms) + 1;
918}
919
920static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters,
921 bool undo_auto_reset)
922{
923 uint32_t i;
924
925 for (i = 0; i < num_events; i++)
926 if (waiters[i].event) {
927 spin_lock(lock: &waiters[i].event->lock);
928 remove_wait_queue(wq_head: &waiters[i].event->wq,
929 wq_entry: &waiters[i].wait);
930 if (undo_auto_reset && waiters[i].activated &&
931 waiters[i].event && waiters[i].event->auto_reset)
932 set_event(waiters[i].event);
933 spin_unlock(lock: &waiters[i].event->lock);
934 }
935
936 kfree(objp: waiters);
937}
938
939int kfd_wait_on_events(struct kfd_process *p,
940 uint32_t num_events, void __user *data,
941 bool all, uint32_t *user_timeout_ms,
942 uint32_t *wait_result)
943{
944 struct kfd_event_data __user *events =
945 (struct kfd_event_data __user *) data;
946 uint32_t i;
947 int ret = 0;
948
949 struct kfd_event_waiter *event_waiters = NULL;
950 long timeout = user_timeout_to_jiffies(user_timeout_ms: *user_timeout_ms);
951
952 event_waiters = alloc_event_waiters(num_events);
953 if (!event_waiters) {
954 ret = -ENOMEM;
955 goto out;
956 }
957
958 /* Use p->event_mutex here to protect against concurrent creation and
959 * destruction of events while we initialize event_waiters.
960 */
961 mutex_lock(&p->event_mutex);
962
963 for (i = 0; i < num_events; i++) {
964 struct kfd_event_data event_data;
965
966 if (copy_from_user(to: &event_data, from: &events[i],
967 n: sizeof(struct kfd_event_data))) {
968 ret = -EFAULT;
969 goto out_unlock;
970 }
971
972 ret = init_event_waiter(p, waiter: &event_waiters[i], event_data: &event_data);
973 if (ret)
974 goto out_unlock;
975 }
976
977 /* Check condition once. */
978 *wait_result = test_event_condition(all, num_events, event_waiters);
979 if (*wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) {
980 ret = copy_signaled_event_data(num_events,
981 event_waiters, data: events);
982 goto out_unlock;
983 } else if (WARN_ON(*wait_result == KFD_IOC_WAIT_RESULT_FAIL)) {
984 /* This should not happen. Events shouldn't be
985 * destroyed while we're holding the event_mutex
986 */
987 goto out_unlock;
988 }
989
990 mutex_unlock(lock: &p->event_mutex);
991
992 while (true) {
993 if (fatal_signal_pending(current)) {
994 ret = -EINTR;
995 break;
996 }
997
998 if (signal_pending(current)) {
999 ret = -ERESTARTSYS;
1000 if (*user_timeout_ms != KFD_EVENT_TIMEOUT_IMMEDIATE &&
1001 *user_timeout_ms != KFD_EVENT_TIMEOUT_INFINITE)
1002 *user_timeout_ms = jiffies_to_msecs(
1003 max(0l, timeout-1));
1004 break;
1005 }
1006
1007 /* Set task state to interruptible sleep before
1008 * checking wake-up conditions. A concurrent wake-up
1009 * will put the task back into runnable state. In that
1010 * case schedule_timeout will not put the task to
1011 * sleep and we'll get a chance to re-check the
1012 * updated conditions almost immediately. Otherwise,
1013 * this race condition would lead to a soft hang or a
1014 * very long sleep.
1015 */
1016 set_current_state(TASK_INTERRUPTIBLE);
1017
1018 *wait_result = test_event_condition(all, num_events,
1019 event_waiters);
1020 if (*wait_result != KFD_IOC_WAIT_RESULT_TIMEOUT)
1021 break;
1022
1023 if (timeout <= 0)
1024 break;
1025
1026 timeout = schedule_timeout(timeout);
1027 }
1028 __set_current_state(TASK_RUNNING);
1029
1030 mutex_lock(&p->event_mutex);
1031 /* copy_signaled_event_data may sleep. So this has to happen
1032 * after the task state is set back to RUNNING.
1033 *
1034 * The event may also have been destroyed after signaling. So
1035 * copy_signaled_event_data also must confirm that the event
1036 * still exists. Therefore this must be under the p->event_mutex
1037 * which is also held when events are destroyed.
1038 */
1039 if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE)
1040 ret = copy_signaled_event_data(num_events,
1041 event_waiters, data: events);
1042
1043out_unlock:
1044 free_waiters(num_events, waiters: event_waiters, undo_auto_reset: ret == -ERESTARTSYS);
1045 mutex_unlock(lock: &p->event_mutex);
1046out:
1047 if (ret)
1048 *wait_result = KFD_IOC_WAIT_RESULT_FAIL;
1049 else if (*wait_result == KFD_IOC_WAIT_RESULT_FAIL)
1050 ret = -EIO;
1051
1052 return ret;
1053}
1054
1055int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma)
1056{
1057 unsigned long pfn;
1058 struct kfd_signal_page *page;
1059 int ret;
1060
1061 /* check required size doesn't exceed the allocated size */
1062 if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) <
1063 get_order(size: vma->vm_end - vma->vm_start)) {
1064 pr_err("Event page mmap requested illegal size\n");
1065 return -EINVAL;
1066 }
1067
1068 page = p->signal_page;
1069 if (!page) {
1070 /* Probably KFD bug, but mmap is user-accessible. */
1071 pr_debug("Signal page could not be found\n");
1072 return -EINVAL;
1073 }
1074
1075 pfn = __pa(page->kernel_address);
1076 pfn >>= PAGE_SHIFT;
1077
1078 vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE
1079 | VM_DONTDUMP | VM_PFNMAP);
1080
1081 pr_debug("Mapping signal page\n");
1082 pr_debug(" start user address == 0x%08lx\n", vma->vm_start);
1083 pr_debug(" end user address == 0x%08lx\n", vma->vm_end);
1084 pr_debug(" pfn == 0x%016lX\n", pfn);
1085 pr_debug(" vm_flags == 0x%08lX\n", vma->vm_flags);
1086 pr_debug(" size == 0x%08lX\n",
1087 vma->vm_end - vma->vm_start);
1088
1089 page->user_address = (uint64_t __user *)vma->vm_start;
1090
1091 /* mapping the page to user process */
1092 ret = remap_pfn_range(vma, addr: vma->vm_start, pfn,
1093 size: vma->vm_end - vma->vm_start, pgprot: vma->vm_page_prot);
1094 if (!ret)
1095 p->signal_mapped_size = vma->vm_end - vma->vm_start;
1096
1097 return ret;
1098}
1099
1100/*
1101 * Assumes that p is not going away.
1102 */
1103static void lookup_events_by_type_and_signal(struct kfd_process *p,
1104 int type, void *event_data)
1105{
1106 struct kfd_hsa_memory_exception_data *ev_data;
1107 struct kfd_event *ev;
1108 uint32_t id;
1109 bool send_signal = true;
1110
1111 ev_data = (struct kfd_hsa_memory_exception_data *) event_data;
1112
1113 rcu_read_lock();
1114
1115 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1116 idr_for_each_entry_continue(&p->event_idr, ev, id)
1117 if (ev->type == type) {
1118 send_signal = false;
1119 dev_dbg(kfd_device,
1120 "Event found: id %X type %d",
1121 ev->event_id, ev->type);
1122 spin_lock(lock: &ev->lock);
1123 set_event(ev);
1124 if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data)
1125 ev->memory_exception_data = *ev_data;
1126 spin_unlock(lock: &ev->lock);
1127 }
1128
1129 if (type == KFD_EVENT_TYPE_MEMORY) {
1130 dev_warn(kfd_device,
1131 "Sending SIGSEGV to process pid %d",
1132 p->lead_thread->pid);
1133 send_sig(SIGSEGV, p->lead_thread, 0);
1134 }
1135
1136 /* Send SIGTERM no event of type "type" has been found*/
1137 if (send_signal) {
1138 if (send_sigterm) {
1139 dev_warn(kfd_device,
1140 "Sending SIGTERM to process pid %d",
1141 p->lead_thread->pid);
1142 send_sig(SIGTERM, p->lead_thread, 0);
1143 } else {
1144 dev_err(kfd_device,
1145 "Process pid %d got unhandled exception",
1146 p->lead_thread->pid);
1147 }
1148 }
1149
1150 rcu_read_unlock();
1151}
1152
1153void kfd_signal_hw_exception_event(u32 pasid)
1154{
1155 /*
1156 * Because we are called from arbitrary context (workqueue) as opposed
1157 * to process context, kfd_process could attempt to exit while we are
1158 * running so the lookup function increments the process ref count.
1159 */
1160 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid, NULL);
1161
1162 if (!p)
1163 return; /* Presumably process exited. */
1164
1165 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL);
1166 kfd_unref_process(p);
1167}
1168
1169void kfd_signal_vm_fault_event_with_userptr(struct kfd_process *p, uint64_t gpu_va)
1170{
1171 struct kfd_process_device *pdd;
1172 struct kfd_hsa_memory_exception_data exception_data;
1173 int i;
1174
1175 memset(&exception_data, 0, sizeof(exception_data));
1176 exception_data.va = gpu_va;
1177 exception_data.failure.NotPresent = 1;
1178
1179 // Send VM seg fault to all kfd process device
1180 for (i = 0; i < p->n_pdds; i++) {
1181 pdd = p->pdds[i];
1182 exception_data.gpu_id = pdd->user_gpu_id;
1183 kfd_evict_process_device(pdd);
1184 kfd_signal_vm_fault_event(pdd, NULL, data: &exception_data);
1185 }
1186}
1187
1188void kfd_signal_vm_fault_event(struct kfd_process_device *pdd,
1189 struct kfd_vm_fault_info *info,
1190 struct kfd_hsa_memory_exception_data *data)
1191{
1192 struct kfd_event *ev;
1193 uint32_t id;
1194 struct kfd_process *p = pdd->process;
1195 struct kfd_hsa_memory_exception_data memory_exception_data;
1196 int user_gpu_id;
1197
1198 user_gpu_id = kfd_process_get_user_gpu_id(p, actual_gpu_id: pdd->dev->id);
1199 if (unlikely(user_gpu_id == -EINVAL)) {
1200 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n",
1201 pdd->dev->id);
1202 return;
1203 }
1204
1205 /* SoC15 chips and onwards will pass in data from now on. */
1206 if (!data) {
1207 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1208 memory_exception_data.gpu_id = user_gpu_id;
1209 memory_exception_data.failure.imprecise = true;
1210
1211 /* Set failure reason */
1212 if (info) {
1213 memory_exception_data.va = (info->page_addr) <<
1214 PAGE_SHIFT;
1215 memory_exception_data.failure.NotPresent =
1216 info->prot_valid ? 1 : 0;
1217 memory_exception_data.failure.NoExecute =
1218 info->prot_exec ? 1 : 0;
1219 memory_exception_data.failure.ReadOnly =
1220 info->prot_write ? 1 : 0;
1221 memory_exception_data.failure.imprecise = 0;
1222 }
1223 }
1224
1225 rcu_read_lock();
1226
1227 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1228 idr_for_each_entry_continue(&p->event_idr, ev, id)
1229 if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1230 spin_lock(lock: &ev->lock);
1231 ev->memory_exception_data = data ? *data :
1232 memory_exception_data;
1233 set_event(ev);
1234 spin_unlock(lock: &ev->lock);
1235 }
1236
1237 rcu_read_unlock();
1238}
1239
1240void kfd_signal_reset_event(struct kfd_node *dev)
1241{
1242 struct kfd_hsa_hw_exception_data hw_exception_data;
1243 struct kfd_hsa_memory_exception_data memory_exception_data;
1244 struct kfd_process *p;
1245 struct kfd_event *ev;
1246 unsigned int temp;
1247 uint32_t id, idx;
1248 int reset_cause = atomic_read(v: &dev->sram_ecc_flag) ?
1249 KFD_HW_EXCEPTION_ECC :
1250 KFD_HW_EXCEPTION_GPU_HANG;
1251
1252 /* Whole gpu reset caused by GPU hang and memory is lost */
1253 memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1254 hw_exception_data.memory_lost = 1;
1255 hw_exception_data.reset_cause = reset_cause;
1256
1257 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1258 memory_exception_data.ErrorType = KFD_MEM_ERR_SRAM_ECC;
1259 memory_exception_data.failure.imprecise = true;
1260
1261 idx = srcu_read_lock(ssp: &kfd_processes_srcu);
1262 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1263 int user_gpu_id = kfd_process_get_user_gpu_id(p, actual_gpu_id: dev->id);
1264 struct kfd_process_device *pdd = kfd_get_process_device_data(dev, p);
1265
1266 if (unlikely(user_gpu_id == -EINVAL)) {
1267 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1268 continue;
1269 }
1270
1271 if (unlikely(!pdd)) {
1272 WARN_ONCE(1, "Could not get device data from process pid:%d\n",
1273 p->lead_thread->pid);
1274 continue;
1275 }
1276
1277 if (dev->dqm->detect_hang_count && !pdd->has_reset_queue)
1278 continue;
1279
1280 if (dev->dqm->detect_hang_count) {
1281 struct amdgpu_task_info *ti;
1282 struct amdgpu_fpriv *drv_priv;
1283
1284 if (unlikely(amdgpu_file_to_fpriv(pdd->drm_file, &drv_priv))) {
1285 WARN_ONCE(1, "Could not get vm for device %x from pid:%d\n",
1286 dev->id, p->lead_thread->pid);
1287 continue;
1288 }
1289
1290 ti = amdgpu_vm_get_task_info_vm(vm: &drv_priv->vm);
1291 if (ti) {
1292 dev_err(dev->adev->dev,
1293 "Queues reset on process %s tid %d thread %s pid %d\n",
1294 ti->process_name, ti->tgid, ti->task.comm, ti->task.pid);
1295 amdgpu_vm_put_task_info(task_info: ti);
1296 }
1297 }
1298
1299 rcu_read_lock();
1300
1301 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1302 idr_for_each_entry_continue(&p->event_idr, ev, id) {
1303 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1304 spin_lock(lock: &ev->lock);
1305 ev->hw_exception_data = hw_exception_data;
1306 ev->hw_exception_data.gpu_id = user_gpu_id;
1307 set_event(ev);
1308 spin_unlock(lock: &ev->lock);
1309 }
1310 if (ev->type == KFD_EVENT_TYPE_MEMORY &&
1311 reset_cause == KFD_HW_EXCEPTION_ECC) {
1312 spin_lock(lock: &ev->lock);
1313 ev->memory_exception_data = memory_exception_data;
1314 ev->memory_exception_data.gpu_id = user_gpu_id;
1315 set_event(ev);
1316 spin_unlock(lock: &ev->lock);
1317 }
1318 }
1319
1320 rcu_read_unlock();
1321 }
1322 srcu_read_unlock(ssp: &kfd_processes_srcu, idx);
1323}
1324
1325void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid)
1326{
1327 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid, NULL);
1328 struct kfd_hsa_memory_exception_data memory_exception_data;
1329 struct kfd_hsa_hw_exception_data hw_exception_data;
1330 struct kfd_event *ev;
1331 uint32_t id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1332 int user_gpu_id;
1333
1334 if (!p) {
1335 dev_warn(dev->adev->dev, "Not find process with pasid:%d\n", pasid);
1336 return; /* Presumably process exited. */
1337 }
1338
1339 user_gpu_id = kfd_process_get_user_gpu_id(p, actual_gpu_id: dev->id);
1340 if (unlikely(user_gpu_id == -EINVAL)) {
1341 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1342 kfd_unref_process(p);
1343 return;
1344 }
1345
1346 memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1347 hw_exception_data.gpu_id = user_gpu_id;
1348 hw_exception_data.memory_lost = 1;
1349 hw_exception_data.reset_cause = KFD_HW_EXCEPTION_ECC;
1350
1351 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1352 memory_exception_data.ErrorType = KFD_MEM_ERR_POISON_CONSUMED;
1353 memory_exception_data.gpu_id = user_gpu_id;
1354 memory_exception_data.failure.imprecise = true;
1355
1356 rcu_read_lock();
1357
1358 idr_for_each_entry_continue(&p->event_idr, ev, id) {
1359 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1360 spin_lock(lock: &ev->lock);
1361 ev->hw_exception_data = hw_exception_data;
1362 set_event(ev);
1363 spin_unlock(lock: &ev->lock);
1364 }
1365
1366 if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1367 spin_lock(lock: &ev->lock);
1368 ev->memory_exception_data = memory_exception_data;
1369 set_event(ev);
1370 spin_unlock(lock: &ev->lock);
1371 }
1372 }
1373
1374 dev_warn(dev->adev->dev, "Send SIGBUS to process %s(pasid:%d)\n",
1375 p->lead_thread->comm, pasid);
1376 rcu_read_unlock();
1377
1378 /* user application will handle SIGBUS signal */
1379 send_sig(SIGBUS, p->lead_thread, 0);
1380
1381 kfd_unref_process(p);
1382}
1383

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