blob: 33cafbb965209a7d069c95bcbc2a5c414dae507b [file] [log] [blame]
Andrew Lewyckyf3a39812015-05-10 12:15:46 +03001/*
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>
Ingo Molnar3f07c012017-02-08 18:51:30 +010026#include <linux/sched/signal.h>
Felix Kuehling9b56bb12017-10-27 19:35:19 -040027#include <linux/sched/mm.h>
Andrew Lewyckyf3a39812015-05-10 12:15:46 +030028#include <linux/uaccess.h>
Andrew Lewyckyf3a39812015-05-10 12:15:46 +030029#include <linux/mman.h>
30#include <linux/memory.h>
31#include "kfd_priv.h"
32#include "kfd_events.h"
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +030033#include <linux/device.h>
Andrew Lewyckyf3a39812015-05-10 12:15:46 +030034
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;
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +030049
50 /* Event */
51 struct kfd_event *event;
52 uint32_t input_index;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +030053};
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
Kent Russell79775b62017-08-15 23:00:05 -0400113 pr_debug("Allocated event signal slot in page %p, slot %d\n",
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300114 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
Kent Russell79775b62017-08-15 23:00:05 -0400158 pr_debug("Allocated new event signal page at %p, for process %p\n",
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300159 page, p);
Kent Russell79775b62017-08-15 23:00:05 -0400160 pr_debug("Page index is %d\n", page->page_index);
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300161
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);
Edward O'Callaghan991ca8e2016-05-01 00:06:27 +1000180 if (!ret) {
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300181 ret = allocate_signal_page(devkfd, p);
Edward O'Callaghan991ca8e2016-05-01 00:06:27 +1000182 if (ret)
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300183 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
Kent Russell8eabaf52017-08-15 23:00:04 -0400197 * and reused until it exits.
198 */
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300199}
200
201static struct signal_page *lookup_signal_page_by_index(struct kfd_process *p,
202 unsigned int page_index)
203{
204 struct signal_page *page;
205
206 /*
207 * This is safe because we don't delete signal pages until the
208 * process exits.
209 */
210 list_for_each_entry(page, &p->signal_event_pages, event_pages)
211 if (page->page_index == page_index)
212 return page;
213
214 return NULL;
215}
216
217/*
218 * Assumes that p->event_mutex is held and of course that p is not going
219 * away (current or locked).
220 */
221static struct kfd_event *lookup_event_by_id(struct kfd_process *p, uint32_t id)
222{
223 struct kfd_event *ev;
224
225 hash_for_each_possible(p->events, ev, events, id)
226 if (ev->event_id == id)
227 return ev;
228
229 return NULL;
230}
231
232static u32 make_signal_event_id(struct signal_page *page,
233 unsigned int signal_slot_index)
234{
235 return page->page_index |
236 (signal_slot_index << SIGNAL_EVENT_ID_SLOT_SHIFT);
237}
238
239/*
240 * Produce a kfd event id for a nonsignal event.
241 * These are arbitrary numbers, so we do a sequential search through
242 * the hash table for an unused number.
243 */
244static u32 make_nonsignal_event_id(struct kfd_process *p)
245{
246 u32 id;
247
248 for (id = p->next_nonsignal_event_id;
249 id < KFD_LAST_NONSIGNAL_EVENT_ID &&
Kent Russell4eacc26b2017-08-15 23:00:06 -0400250 lookup_event_by_id(p, id);
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300251 id++)
252 ;
253
254 if (id < KFD_LAST_NONSIGNAL_EVENT_ID) {
255
256 /*
257 * What if id == LAST_NONSIGNAL_EVENT_ID - 1?
258 * Then next_nonsignal_event_id = LAST_NONSIGNAL_EVENT_ID so
259 * the first loop fails immediately and we proceed with the
260 * wraparound loop below.
261 */
262 p->next_nonsignal_event_id = id + 1;
263
264 return id;
265 }
266
267 for (id = KFD_FIRST_NONSIGNAL_EVENT_ID;
268 id < KFD_LAST_NONSIGNAL_EVENT_ID &&
Kent Russell4eacc26b2017-08-15 23:00:06 -0400269 lookup_event_by_id(p, id);
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300270 id++)
271 ;
272
273
274 if (id < KFD_LAST_NONSIGNAL_EVENT_ID) {
275 p->next_nonsignal_event_id = id + 1;
276 return id;
277 }
278
279 p->next_nonsignal_event_id = KFD_FIRST_NONSIGNAL_EVENT_ID;
280 return 0;
281}
282
283static struct kfd_event *lookup_event_by_page_slot(struct kfd_process *p,
284 struct signal_page *page,
285 unsigned int signal_slot)
286{
287 return lookup_event_by_id(p, make_signal_event_id(page, signal_slot));
288}
289
290static int create_signal_event(struct file *devkfd,
291 struct kfd_process *p,
292 struct kfd_event *ev)
293{
294 if (p->signal_event_count == KFD_SIGNAL_EVENT_LIMIT) {
Felix Kuehlingc9861692017-09-20 18:10:22 -0400295 if (!p->signal_event_limit_reached) {
296 pr_warn("Signal event wasn't created because limit was reached\n");
297 p->signal_event_limit_reached = true;
298 }
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300299 return -ENOMEM;
300 }
301
302 if (!allocate_event_notification_slot(devkfd, p, &ev->signal_page,
303 &ev->signal_slot_index)) {
Kent Russell79775b62017-08-15 23:00:05 -0400304 pr_warn("Signal event wasn't created because out of kernel memory\n");
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300305 return -ENOMEM;
306 }
307
308 p->signal_event_count++;
309
310 ev->user_signal_address =
311 &ev->signal_page->user_address[ev->signal_slot_index];
312
313 ev->event_id = make_signal_event_id(ev->signal_page,
314 ev->signal_slot_index);
315
Kent Russell79775b62017-08-15 23:00:05 -0400316 pr_debug("Signal event number %zu created with id %d, address %p\n",
Oded Gabbay6235e152015-04-30 18:05:36 +0300317 p->signal_event_count, ev->event_id,
318 ev->user_signal_address);
319
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300320 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{
Kent Russell4eacc26b2017-08-15 23:00:06 -0400348 if (ev->signal_page) {
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300349 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;
Kent Russell8eabaf52017-08-15 23:00:04 -0400587 unsigned int i;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300588
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
Sean Keely1f9d09b2017-10-27 19:35:20 -0400620static int init_event_waiter_get_status(struct kfd_process *p,
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300621 struct kfd_event_waiter *waiter,
622 uint32_t event_id,
623 uint32_t input_index)
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300624{
625 struct kfd_event *ev = lookup_event_by_id(p, event_id);
626
627 if (!ev)
628 return -EINVAL;
629
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300630 waiter->event = ev;
631 waiter->input_index = input_index;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300632 waiter->activated = ev->signaled;
633 ev->signaled = ev->signaled && !ev->auto_reset;
634
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300635 return 0;
636}
637
Sean Keely1f9d09b2017-10-27 19:35:20 -0400638static void init_event_waiter_add_to_waitlist(struct kfd_event_waiter *waiter)
639{
640 struct kfd_event *ev = waiter->event;
641
642 /* Only add to the wait list if we actually need to
643 * wait on this event.
644 */
645 if (!waiter->activated)
646 list_add(&waiter->waiters, &ev->waiters);
647}
648
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300649static bool test_event_condition(bool all, uint32_t num_events,
650 struct kfd_event_waiter *event_waiters)
651{
652 uint32_t i;
653 uint32_t activated_count = 0;
654
655 for (i = 0; i < num_events; i++) {
656 if (event_waiters[i].activated) {
657 if (!all)
658 return true;
659
660 activated_count++;
661 }
662 }
663
664 return activated_count == num_events;
665}
666
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300667/*
668 * Copy event specific data, if defined.
669 * Currently only memory exception events have additional data to copy to user
670 */
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400671static int copy_signaled_event_data(uint32_t num_events,
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300672 struct kfd_event_waiter *event_waiters,
673 struct kfd_event_data __user *data)
674{
675 struct kfd_hsa_memory_exception_data *src;
676 struct kfd_hsa_memory_exception_data __user *dst;
677 struct kfd_event_waiter *waiter;
678 struct kfd_event *event;
679 uint32_t i;
680
681 for (i = 0; i < num_events; i++) {
682 waiter = &event_waiters[i];
683 event = waiter->event;
684 if (waiter->activated && event->type == KFD_EVENT_TYPE_MEMORY) {
685 dst = &data[waiter->input_index].memory_exception_data;
686 src = &event->memory_exception_data;
687 if (copy_to_user(dst, src,
688 sizeof(struct kfd_hsa_memory_exception_data)))
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400689 return -EFAULT;
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300690 }
691 }
692
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400693 return 0;
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300694
695}
696
697
698
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300699static long user_timeout_to_jiffies(uint32_t user_timeout_ms)
700{
701 if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE)
702 return 0;
703
704 if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE)
705 return MAX_SCHEDULE_TIMEOUT;
706
707 /*
708 * msecs_to_jiffies interprets all values above 2^31-1 as infinite,
709 * but we consider them finite.
710 * This hack is wrong, but nobody is likely to notice.
711 */
712 user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF);
713
714 return msecs_to_jiffies(user_timeout_ms) + 1;
715}
716
717static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters)
718{
719 uint32_t i;
720
721 for (i = 0; i < num_events; i++)
722 list_del(&waiters[i].waiters);
723
724 kfree(waiters);
725}
726
727int kfd_wait_on_events(struct kfd_process *p,
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300728 uint32_t num_events, void __user *data,
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300729 bool all, uint32_t user_timeout_ms,
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400730 uint32_t *wait_result)
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300731{
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300732 struct kfd_event_data __user *events =
733 (struct kfd_event_data __user *) data;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300734 uint32_t i;
735 int ret = 0;
Sean Keely1f9d09b2017-10-27 19:35:20 -0400736
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300737 struct kfd_event_waiter *event_waiters = NULL;
738 long timeout = user_timeout_to_jiffies(user_timeout_ms);
739
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400740 event_waiters = alloc_event_waiters(num_events);
741 if (!event_waiters) {
742 ret = -ENOMEM;
743 goto out;
744 }
745
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300746 mutex_lock(&p->event_mutex);
747
Sean Keely1f9d09b2017-10-27 19:35:20 -0400748 /* Set to something unreasonable - this is really
749 * just a bool for now.
750 */
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400751 *wait_result = KFD_IOC_WAIT_RESULT_TIMEOUT;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300752
753 for (i = 0; i < num_events; i++) {
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300754 struct kfd_event_data event_data;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300755
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300756 if (copy_from_user(&event_data, &events[i],
Pan Bian8bf79382016-12-01 16:10:42 +0800757 sizeof(struct kfd_event_data))) {
758 ret = -EFAULT;
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400759 goto out_unlock;
Pan Bian8bf79382016-12-01 16:10:42 +0800760 }
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300761
Sean Keely1f9d09b2017-10-27 19:35:20 -0400762 ret = init_event_waiter_get_status(p, &event_waiters[i],
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300763 event_data.event_id, i);
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300764 if (ret)
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400765 goto out_unlock;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300766 }
767
Sean Keely1f9d09b2017-10-27 19:35:20 -0400768 /* Check condition once. */
769 if (test_event_condition(all, num_events, event_waiters)) {
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400770 *wait_result = KFD_IOC_WAIT_RESULT_COMPLETE;
771 ret = copy_signaled_event_data(num_events,
772 event_waiters, events);
773 goto out_unlock;
Sean Keely1f9d09b2017-10-27 19:35:20 -0400774 } else {
775 /* Add to wait lists if we need to wait. */
776 for (i = 0; i < num_events; i++)
777 init_event_waiter_add_to_waitlist(&event_waiters[i]);
778 }
779
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300780 mutex_unlock(&p->event_mutex);
781
782 while (true) {
783 if (fatal_signal_pending(current)) {
784 ret = -EINTR;
785 break;
786 }
787
788 if (signal_pending(current)) {
789 /*
790 * This is wrong when a nonzero, non-infinite timeout
791 * is specified. We need to use
792 * ERESTARTSYS_RESTARTBLOCK, but struct restart_block
793 * contains a union with data for each user and it's
794 * in generic kernel code that I don't want to
795 * touch yet.
796 */
797 ret = -ERESTARTSYS;
798 break;
799 }
800
Sean Keelyd9aeec42017-10-27 19:35:21 -0400801 /* Set task state to interruptible sleep before
802 * checking wake-up conditions. A concurrent wake-up
803 * will put the task back into runnable state. In that
804 * case schedule_timeout will not put the task to
805 * sleep and we'll get a chance to re-check the
806 * updated conditions almost immediately. Otherwise,
807 * this race condition would lead to a soft hang or a
808 * very long sleep.
809 */
810 set_current_state(TASK_INTERRUPTIBLE);
811
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300812 if (test_event_condition(all, num_events, event_waiters)) {
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400813 *wait_result = KFD_IOC_WAIT_RESULT_COMPLETE;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300814 break;
815 }
816
817 if (timeout <= 0) {
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400818 *wait_result = KFD_IOC_WAIT_RESULT_TIMEOUT;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300819 break;
820 }
821
Sean Keelyd9aeec42017-10-27 19:35:21 -0400822 timeout = schedule_timeout(timeout);
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300823 }
824 __set_current_state(TASK_RUNNING);
825
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400826 /* copy_signaled_event_data may sleep. So this has to happen
827 * after the task state is set back to RUNNING.
828 */
829 if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE)
830 ret = copy_signaled_event_data(num_events,
831 event_waiters, events);
832
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300833 mutex_lock(&p->event_mutex);
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400834out_unlock:
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300835 free_waiters(num_events, event_waiters);
836 mutex_unlock(&p->event_mutex);
Felix Kuehlingfdf0c832017-10-27 19:35:22 -0400837out:
838 if (ret)
839 *wait_result = KFD_IOC_WAIT_RESULT_FAIL;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300840
841 return ret;
842}
843
844int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma)
845{
846
847 unsigned int page_index;
848 unsigned long pfn;
849 struct signal_page *page;
850
851 /* check required size is logical */
852 if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) !=
853 get_order(vma->vm_end - vma->vm_start)) {
Kent Russell79775b62017-08-15 23:00:05 -0400854 pr_err("Event page mmap requested illegal size\n");
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300855 return -EINVAL;
856 }
857
858 page_index = vma->vm_pgoff;
859
860 page = lookup_signal_page_by_index(p, page_index);
861 if (!page) {
862 /* Probably KFD bug, but mmap is user-accessible. */
Kent Russell79775b62017-08-15 23:00:05 -0400863 pr_debug("Signal page could not be found for page_index %u\n",
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300864 page_index);
865 return -EINVAL;
866 }
867
868 pfn = __pa(page->kernel_address);
869 pfn >>= PAGE_SHIFT;
870
871 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE
872 | VM_DONTDUMP | VM_PFNMAP;
873
Kent Russell79775b62017-08-15 23:00:05 -0400874 pr_debug("Mapping signal page\n");
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300875 pr_debug(" start user address == 0x%08lx\n", vma->vm_start);
876 pr_debug(" end user address == 0x%08lx\n", vma->vm_end);
877 pr_debug(" pfn == 0x%016lX\n", pfn);
878 pr_debug(" vm_flags == 0x%08lX\n", vma->vm_flags);
879 pr_debug(" size == 0x%08lX\n",
880 vma->vm_end - vma->vm_start);
881
882 page->user_address = (uint64_t __user *)vma->vm_start;
883
884 /* mapping the page to user process */
885 return remap_pfn_range(vma, vma->vm_start, pfn,
886 vma->vm_end - vma->vm_start, vma->vm_page_prot);
887}
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300888
889/*
890 * Assumes that p->event_mutex is held and of course
891 * that p is not going away (current or locked).
892 */
893static void lookup_events_by_type_and_signal(struct kfd_process *p,
894 int type, void *event_data)
895{
896 struct kfd_hsa_memory_exception_data *ev_data;
897 struct kfd_event *ev;
898 int bkt;
899 bool send_signal = true;
900
901 ev_data = (struct kfd_hsa_memory_exception_data *) event_data;
902
903 hash_for_each(p->events, bkt, ev, events)
904 if (ev->type == type) {
905 send_signal = false;
906 dev_dbg(kfd_device,
907 "Event found: id %X type %d",
908 ev->event_id, ev->type);
909 set_event(ev);
910 if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data)
911 ev->memory_exception_data = *ev_data;
912 }
913
914 /* Send SIGTERM no event of type "type" has been found*/
915 if (send_signal) {
Oded Gabbay81663012014-12-24 13:30:52 +0200916 if (send_sigterm) {
917 dev_warn(kfd_device,
918 "Sending SIGTERM to HSA Process with PID %d ",
919 p->lead_thread->pid);
920 send_sig(SIGTERM, p->lead_thread, 0);
921 } else {
922 dev_err(kfd_device,
923 "HSA Process (PID %d) got unhandled exception",
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300924 p->lead_thread->pid);
Oded Gabbay81663012014-12-24 13:30:52 +0200925 }
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300926 }
927}
928
929void kfd_signal_iommu_event(struct kfd_dev *dev, unsigned int pasid,
930 unsigned long address, bool is_write_requested,
931 bool is_execute_requested)
932{
933 struct kfd_hsa_memory_exception_data memory_exception_data;
934 struct vm_area_struct *vma;
935
936 /*
937 * Because we are called from arbitrary context (workqueue) as opposed
938 * to process context, kfd_process could attempt to exit while we are
939 * running so the lookup function returns a locked process.
940 */
941 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
Felix Kuehling9b56bb12017-10-27 19:35:19 -0400942 struct mm_struct *mm;
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300943
944 if (!p)
945 return; /* Presumably process exited. */
946
Felix Kuehling9b56bb12017-10-27 19:35:19 -0400947 /* Take a safe reference to the mm_struct, which may otherwise
948 * disappear even while the kfd_process is still referenced.
949 */
950 mm = get_task_mm(p->lead_thread);
951 if (!mm) {
952 mutex_unlock(&p->mutex);
953 return; /* Process is exiting */
954 }
955
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300956 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
957
Felix Kuehling9b56bb12017-10-27 19:35:19 -0400958 down_read(&mm->mmap_sem);
959 vma = find_vma(mm, address);
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300960
961 memory_exception_data.gpu_id = dev->id;
962 memory_exception_data.va = address;
963 /* Set failure reason */
964 memory_exception_data.failure.NotPresent = 1;
965 memory_exception_data.failure.NoExecute = 0;
966 memory_exception_data.failure.ReadOnly = 0;
967 if (vma) {
968 if (vma->vm_start > address) {
969 memory_exception_data.failure.NotPresent = 1;
970 memory_exception_data.failure.NoExecute = 0;
971 memory_exception_data.failure.ReadOnly = 0;
972 } else {
973 memory_exception_data.failure.NotPresent = 0;
974 if (is_write_requested && !(vma->vm_flags & VM_WRITE))
975 memory_exception_data.failure.ReadOnly = 1;
976 else
977 memory_exception_data.failure.ReadOnly = 0;
978 if (is_execute_requested && !(vma->vm_flags & VM_EXEC))
979 memory_exception_data.failure.NoExecute = 1;
980 else
981 memory_exception_data.failure.NoExecute = 0;
982 }
983 }
984
Felix Kuehling9b56bb12017-10-27 19:35:19 -0400985 up_read(&mm->mmap_sem);
986 mmput(mm);
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300987
988 mutex_lock(&p->event_mutex);
989
990 /* Lookup events by type and signal them */
991 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY,
992 &memory_exception_data);
993
994 mutex_unlock(&p->event_mutex);
995 mutex_unlock(&p->mutex);
996}
Alexey Skidanov930c5ff2014-11-25 10:34:31 +0200997
998void kfd_signal_hw_exception_event(unsigned int pasid)
999{
1000 /*
1001 * Because we are called from arbitrary context (workqueue) as opposed
1002 * to process context, kfd_process could attempt to exit while we are
1003 * running so the lookup function returns a locked process.
1004 */
1005 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1006
1007 if (!p)
1008 return; /* Presumably process exited. */
1009
1010 mutex_lock(&p->event_mutex);
1011
1012 /* Lookup events by type and signal them */
1013 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL);
1014
1015 mutex_unlock(&p->event_mutex);
1016 mutex_unlock(&p->mutex);
1017}