Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * kvm eventfd support - use eventfd objects to signal various KVM events
4 *
5 * Copyright 2009 Novell. All Rights Reserved.
6 * Copyright 2010 Red Hat, Inc. and/or its affiliates.
7 *
8 * Author:
9 * Gregory Haskins <ghaskins@novell.com>
10 */
11
12#include <linux/kvm_host.h>
13#include <linux/kvm.h>
14#include <linux/kvm_irqfd.h>
15#include <linux/workqueue.h>
16#include <linux/syscalls.h>
17#include <linux/wait.h>
18#include <linux/poll.h>
19#include <linux/file.h>
20#include <linux/list.h>
21#include <linux/eventfd.h>
22#include <linux/kernel.h>
23#include <linux/srcu.h>
24#include <linux/slab.h>
25#include <linux/seqlock.h>
26#include <linux/irqbypass.h>
27#include <trace/events/kvm.h>
28
29#include <kvm/iodev.h>
30
31#ifdef CONFIG_HAVE_KVM_IRQCHIP
32
33static struct workqueue_struct *irqfd_cleanup_wq;
34
35bool __attribute__((weak))
36kvm_arch_irqfd_allowed(struct kvm *kvm, struct kvm_irqfd *args)
37{
38 return true;
39}
40
41static void
42irqfd_inject(struct work_struct *work)
43{
44 struct kvm_kernel_irqfd *irqfd =
45 container_of(work, struct kvm_kernel_irqfd, inject);
46 struct kvm *kvm = irqfd->kvm;
47
48 if (!irqfd->resampler) {
49 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
50 false);
51 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
52 false);
53 } else
54 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
55 irqfd->gsi, 1, false);
56}
57
58static void irqfd_resampler_notify(struct kvm_kernel_irqfd_resampler *resampler)
59{
60 struct kvm_kernel_irqfd *irqfd;
61
62 list_for_each_entry_srcu(irqfd, &resampler->list, resampler_link,
63 srcu_read_lock_held(&resampler->kvm->irq_srcu))
64 eventfd_signal(irqfd->resamplefd);
65}
66
67/*
68 * Since resampler irqfds share an IRQ source ID, we de-assert once
69 * then notify all of the resampler irqfds using this GSI. We can't
70 * do multiple de-asserts or we risk racing with incoming re-asserts.
71 */
72static void
73irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
74{
75 struct kvm_kernel_irqfd_resampler *resampler;
76 struct kvm *kvm;
77 int idx;
78
79 resampler = container_of(kian,
80 struct kvm_kernel_irqfd_resampler, notifier);
81 kvm = resampler->kvm;
82
83 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
84 resampler->notifier.gsi, 0, false);
85
86 idx = srcu_read_lock(&kvm->irq_srcu);
87 irqfd_resampler_notify(resampler);
88 srcu_read_unlock(&kvm->irq_srcu, idx);
89}
90
91static void
92irqfd_resampler_shutdown(struct kvm_kernel_irqfd *irqfd)
93{
94 struct kvm_kernel_irqfd_resampler *resampler = irqfd->resampler;
95 struct kvm *kvm = resampler->kvm;
96
97 mutex_lock(&kvm->irqfds.resampler_lock);
98
99 list_del_rcu(&irqfd->resampler_link);
100 synchronize_srcu(&kvm->irq_srcu);
101
102 if (list_empty(&resampler->list)) {
103 list_del_rcu(&resampler->link);
104 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
105 /*
106 * synchronize_srcu(&kvm->irq_srcu) already called
107 * in kvm_unregister_irq_ack_notifier().
108 */
109 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
110 resampler->notifier.gsi, 0, false);
111 kfree(resampler);
112 }
113
114 mutex_unlock(&kvm->irqfds.resampler_lock);
115}
116
117/*
118 * Race-free decouple logic (ordering is critical)
119 */
120static void
121irqfd_shutdown(struct work_struct *work)
122{
123 struct kvm_kernel_irqfd *irqfd =
124 container_of(work, struct kvm_kernel_irqfd, shutdown);
125 struct kvm *kvm = irqfd->kvm;
126 u64 cnt;
127
128 /* Make sure irqfd has been initialized in assign path. */
129 synchronize_srcu(&kvm->irq_srcu);
130
131 /*
132 * Synchronize with the wait-queue and unhook ourselves to prevent
133 * further events.
134 */
135 eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
136
137 /*
138 * We know no new events will be scheduled at this point, so block
139 * until all previously outstanding events have completed
140 */
141 flush_work(&irqfd->inject);
142
143 if (irqfd->resampler) {
144 irqfd_resampler_shutdown(irqfd);
145 eventfd_ctx_put(irqfd->resamplefd);
146 }
147
148 /*
149 * It is now safe to release the object's resources
150 */
151#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
152 irq_bypass_unregister_consumer(&irqfd->consumer);
153#endif
154 eventfd_ctx_put(irqfd->eventfd);
155 kfree(irqfd);
156}
157
158
159/* assumes kvm->irqfds.lock is held */
160static bool
161irqfd_is_active(struct kvm_kernel_irqfd *irqfd)
162{
163 return list_empty(&irqfd->list) ? false : true;
164}
165
166/*
167 * Mark the irqfd as inactive and schedule it for removal
168 *
169 * assumes kvm->irqfds.lock is held
170 */
171static void
172irqfd_deactivate(struct kvm_kernel_irqfd *irqfd)
173{
174 BUG_ON(!irqfd_is_active(irqfd));
175
176 list_del_init(&irqfd->list);
177
178 queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
179}
180
181int __attribute__((weak)) kvm_arch_set_irq_inatomic(
182 struct kvm_kernel_irq_routing_entry *irq,
183 struct kvm *kvm, int irq_source_id,
184 int level,
185 bool line_status)
186{
187 return -EWOULDBLOCK;
188}
189
190/*
191 * Called with wqh->lock held and interrupts disabled
192 */
193static int
194irqfd_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
195{
196 struct kvm_kernel_irqfd *irqfd =
197 container_of(wait, struct kvm_kernel_irqfd, wait);
198 __poll_t flags = key_to_poll(key);
199 struct kvm_kernel_irq_routing_entry irq;
200 struct kvm *kvm = irqfd->kvm;
201 unsigned seq;
202 int idx;
203 int ret = 0;
204
205 if (flags & EPOLLIN) {
206 u64 cnt;
207 eventfd_ctx_do_read(irqfd->eventfd, &cnt);
208
209 idx = srcu_read_lock(&kvm->irq_srcu);
210 do {
211 seq = read_seqcount_begin(&irqfd->irq_entry_sc);
212 irq = irqfd->irq_entry;
213 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
214 /* An event has been signaled, inject an interrupt */
215 if (kvm_arch_set_irq_inatomic(&irq, kvm,
216 KVM_USERSPACE_IRQ_SOURCE_ID, 1,
217 false) == -EWOULDBLOCK)
218 schedule_work(&irqfd->inject);
219 srcu_read_unlock(&kvm->irq_srcu, idx);
220 ret = 1;
221 }
222
223 if (flags & EPOLLHUP) {
224 /* The eventfd is closing, detach from KVM */
225 unsigned long iflags;
226
227 spin_lock_irqsave(&kvm->irqfds.lock, iflags);
228
229 /*
230 * We must check if someone deactivated the irqfd before
231 * we could acquire the irqfds.lock since the item is
232 * deactivated from the KVM side before it is unhooked from
233 * the wait-queue. If it is already deactivated, we can
234 * simply return knowing the other side will cleanup for us.
235 * We cannot race against the irqfd going away since the
236 * other side is required to acquire wqh->lock, which we hold
237 */
238 if (irqfd_is_active(irqfd))
239 irqfd_deactivate(irqfd);
240
241 spin_unlock_irqrestore(&kvm->irqfds.lock, iflags);
242 }
243
244 return ret;
245}
246
247static void
248irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
249 poll_table *pt)
250{
251 struct kvm_kernel_irqfd *irqfd =
252 container_of(pt, struct kvm_kernel_irqfd, pt);
253 add_wait_queue_priority(wqh, &irqfd->wait);
254}
255
256/* Must be called under irqfds.lock */
257static void irqfd_update(struct kvm *kvm, struct kvm_kernel_irqfd *irqfd)
258{
259 struct kvm_kernel_irq_routing_entry *e;
260 struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
261 int n_entries;
262
263 n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
264
265 write_seqcount_begin(&irqfd->irq_entry_sc);
266
267 e = entries;
268 if (n_entries == 1)
269 irqfd->irq_entry = *e;
270 else
271 irqfd->irq_entry.type = 0;
272
273 write_seqcount_end(&irqfd->irq_entry_sc);
274}
275
276#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
277void __attribute__((weak)) kvm_arch_irq_bypass_stop(
278 struct irq_bypass_consumer *cons)
279{
280}
281
282void __attribute__((weak)) kvm_arch_irq_bypass_start(
283 struct irq_bypass_consumer *cons)
284{
285}
286
287int __attribute__((weak)) kvm_arch_update_irqfd_routing(
288 struct kvm *kvm, unsigned int host_irq,
289 uint32_t guest_irq, bool set)
290{
291 return 0;
292}
293
294bool __attribute__((weak)) kvm_arch_irqfd_route_changed(
295 struct kvm_kernel_irq_routing_entry *old,
296 struct kvm_kernel_irq_routing_entry *new)
297{
298 return true;
299}
300#endif
301
302static int
303kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
304{
305 struct kvm_kernel_irqfd *irqfd, *tmp;
306 struct fd f;
307 struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
308 int ret;
309 __poll_t events;
310 int idx;
311
312 if (!kvm_arch_intc_initialized(kvm))
313 return -EAGAIN;
314
315 if (!kvm_arch_irqfd_allowed(kvm, args))
316 return -EINVAL;
317
318 irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL_ACCOUNT);
319 if (!irqfd)
320 return -ENOMEM;
321
322 irqfd->kvm = kvm;
323 irqfd->gsi = args->gsi;
324 INIT_LIST_HEAD(&irqfd->list);
325 INIT_WORK(&irqfd->inject, irqfd_inject);
326 INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
327 seqcount_spinlock_init(&irqfd->irq_entry_sc, &kvm->irqfds.lock);
328
329 f = fdget(args->fd);
330 if (!f.file) {
331 ret = -EBADF;
332 goto out;
333 }
334
335 eventfd = eventfd_ctx_fileget(f.file);
336 if (IS_ERR(eventfd)) {
337 ret = PTR_ERR(eventfd);
338 goto fail;
339 }
340
341 irqfd->eventfd = eventfd;
342
343 if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
344 struct kvm_kernel_irqfd_resampler *resampler;
345
346 resamplefd = eventfd_ctx_fdget(args->resamplefd);
347 if (IS_ERR(resamplefd)) {
348 ret = PTR_ERR(resamplefd);
349 goto fail;
350 }
351
352 irqfd->resamplefd = resamplefd;
353 INIT_LIST_HEAD(&irqfd->resampler_link);
354
355 mutex_lock(&kvm->irqfds.resampler_lock);
356
357 list_for_each_entry(resampler,
358 &kvm->irqfds.resampler_list, link) {
359 if (resampler->notifier.gsi == irqfd->gsi) {
360 irqfd->resampler = resampler;
361 break;
362 }
363 }
364
365 if (!irqfd->resampler) {
366 resampler = kzalloc(sizeof(*resampler),
367 GFP_KERNEL_ACCOUNT);
368 if (!resampler) {
369 ret = -ENOMEM;
370 mutex_unlock(&kvm->irqfds.resampler_lock);
371 goto fail;
372 }
373
374 resampler->kvm = kvm;
375 INIT_LIST_HEAD(&resampler->list);
376 resampler->notifier.gsi = irqfd->gsi;
377 resampler->notifier.irq_acked = irqfd_resampler_ack;
378 INIT_LIST_HEAD(&resampler->link);
379
380 list_add_rcu(&resampler->link, &kvm->irqfds.resampler_list);
381 kvm_register_irq_ack_notifier(kvm,
382 &resampler->notifier);
383 irqfd->resampler = resampler;
384 }
385
386 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
387 synchronize_srcu(&kvm->irq_srcu);
388
389 mutex_unlock(&kvm->irqfds.resampler_lock);
390 }
391
392 /*
393 * Install our own custom wake-up handling so we are notified via
394 * a callback whenever someone signals the underlying eventfd
395 */
396 init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
397 init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
398
399 spin_lock_irq(&kvm->irqfds.lock);
400
401 ret = 0;
402 list_for_each_entry(tmp, &kvm->irqfds.items, list) {
403 if (irqfd->eventfd != tmp->eventfd)
404 continue;
405 /* This fd is used for another irq already. */
406 ret = -EBUSY;
407 spin_unlock_irq(&kvm->irqfds.lock);
408 goto fail;
409 }
410
411 idx = srcu_read_lock(&kvm->irq_srcu);
412 irqfd_update(kvm, irqfd);
413
414 list_add_tail(&irqfd->list, &kvm->irqfds.items);
415
416 spin_unlock_irq(&kvm->irqfds.lock);
417
418 /*
419 * Check if there was an event already pending on the eventfd
420 * before we registered, and trigger it as if we didn't miss it.
421 */
422 events = vfs_poll(f.file, &irqfd->pt);
423
424 if (events & EPOLLIN)
425 schedule_work(&irqfd->inject);
426
427#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
428 if (kvm_arch_has_irq_bypass()) {
429 irqfd->consumer.token = (void *)irqfd->eventfd;
430 irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
431 irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
432 irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
433 irqfd->consumer.start = kvm_arch_irq_bypass_start;
434 ret = irq_bypass_register_consumer(&irqfd->consumer);
435 if (ret)
436 pr_info("irq bypass consumer (token %p) registration fails: %d\n",
437 irqfd->consumer.token, ret);
438 }
439#endif
440
441 srcu_read_unlock(&kvm->irq_srcu, idx);
442
443 /*
444 * do not drop the file until the irqfd is fully initialized, otherwise
445 * we might race against the EPOLLHUP
446 */
447 fdput(f);
448 return 0;
449
450fail:
451 if (irqfd->resampler)
452 irqfd_resampler_shutdown(irqfd);
453
454 if (resamplefd && !IS_ERR(resamplefd))
455 eventfd_ctx_put(resamplefd);
456
457 if (eventfd && !IS_ERR(eventfd))
458 eventfd_ctx_put(eventfd);
459
460 fdput(f);
461
462out:
463 kfree(irqfd);
464 return ret;
465}
466
467bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
468{
469 struct kvm_irq_ack_notifier *kian;
470 int gsi, idx;
471
472 idx = srcu_read_lock(&kvm->irq_srcu);
473 gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
474 if (gsi != -1)
475 hlist_for_each_entry_srcu(kian, &kvm->irq_ack_notifier_list,
476 link, srcu_read_lock_held(&kvm->irq_srcu))
477 if (kian->gsi == gsi) {
478 srcu_read_unlock(&kvm->irq_srcu, idx);
479 return true;
480 }
481
482 srcu_read_unlock(&kvm->irq_srcu, idx);
483
484 return false;
485}
486EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
487
488void kvm_notify_acked_gsi(struct kvm *kvm, int gsi)
489{
490 struct kvm_irq_ack_notifier *kian;
491
492 hlist_for_each_entry_srcu(kian, &kvm->irq_ack_notifier_list,
493 link, srcu_read_lock_held(&kvm->irq_srcu))
494 if (kian->gsi == gsi)
495 kian->irq_acked(kian);
496}
497
498void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
499{
500 int gsi, idx;
501
502 trace_kvm_ack_irq(irqchip, pin);
503
504 idx = srcu_read_lock(&kvm->irq_srcu);
505 gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
506 if (gsi != -1)
507 kvm_notify_acked_gsi(kvm, gsi);
508 srcu_read_unlock(&kvm->irq_srcu, idx);
509}
510
511void kvm_register_irq_ack_notifier(struct kvm *kvm,
512 struct kvm_irq_ack_notifier *kian)
513{
514 mutex_lock(&kvm->irq_lock);
515 hlist_add_head_rcu(&kian->link, &kvm->irq_ack_notifier_list);
516 mutex_unlock(&kvm->irq_lock);
517 kvm_arch_post_irq_ack_notifier_list_update(kvm);
518}
519
520void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
521 struct kvm_irq_ack_notifier *kian)
522{
523 mutex_lock(&kvm->irq_lock);
524 hlist_del_init_rcu(&kian->link);
525 mutex_unlock(&kvm->irq_lock);
526 synchronize_srcu(&kvm->irq_srcu);
527 kvm_arch_post_irq_ack_notifier_list_update(kvm);
528}
529
530/*
531 * shutdown any irqfd's that match fd+gsi
532 */
533static int
534kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
535{
536 struct kvm_kernel_irqfd *irqfd, *tmp;
537 struct eventfd_ctx *eventfd;
538
539 eventfd = eventfd_ctx_fdget(args->fd);
540 if (IS_ERR(eventfd))
541 return PTR_ERR(eventfd);
542
543 spin_lock_irq(&kvm->irqfds.lock);
544
545 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
546 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
547 /*
548 * This clearing of irq_entry.type is needed for when
549 * another thread calls kvm_irq_routing_update before
550 * we flush workqueue below (we synchronize with
551 * kvm_irq_routing_update using irqfds.lock).
552 */
553 write_seqcount_begin(&irqfd->irq_entry_sc);
554 irqfd->irq_entry.type = 0;
555 write_seqcount_end(&irqfd->irq_entry_sc);
556 irqfd_deactivate(irqfd);
557 }
558 }
559
560 spin_unlock_irq(&kvm->irqfds.lock);
561 eventfd_ctx_put(eventfd);
562
563 /*
564 * Block until we know all outstanding shutdown jobs have completed
565 * so that we guarantee there will not be any more interrupts on this
566 * gsi once this deassign function returns.
567 */
568 flush_workqueue(irqfd_cleanup_wq);
569
570 return 0;
571}
572
573int
574kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
575{
576 if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
577 return -EINVAL;
578
579 if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
580 return kvm_irqfd_deassign(kvm, args);
581
582 return kvm_irqfd_assign(kvm, args);
583}
584
585/*
586 * This function is called as the kvm VM fd is being released. Shutdown all
587 * irqfds that still remain open
588 */
589void
590kvm_irqfd_release(struct kvm *kvm)
591{
592 struct kvm_kernel_irqfd *irqfd, *tmp;
593
594 spin_lock_irq(&kvm->irqfds.lock);
595
596 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
597 irqfd_deactivate(irqfd);
598
599 spin_unlock_irq(&kvm->irqfds.lock);
600
601 /*
602 * Block until we know all outstanding shutdown jobs have completed
603 * since we do not take a kvm* reference.
604 */
605 flush_workqueue(irqfd_cleanup_wq);
606
607}
608
609/*
610 * Take note of a change in irq routing.
611 * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
612 */
613void kvm_irq_routing_update(struct kvm *kvm)
614{
615 struct kvm_kernel_irqfd *irqfd;
616
617 spin_lock_irq(&kvm->irqfds.lock);
618
619 list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
620#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
621 /* Under irqfds.lock, so can read irq_entry safely */
622 struct kvm_kernel_irq_routing_entry old = irqfd->irq_entry;
623#endif
624
625 irqfd_update(kvm, irqfd);
626
627#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
628 if (irqfd->producer &&
629 kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry)) {
630 int ret = kvm_arch_update_irqfd_routing(
631 irqfd->kvm, irqfd->producer->irq,
632 irqfd->gsi, 1);
633 WARN_ON(ret);
634 }
635#endif
636 }
637
638 spin_unlock_irq(&kvm->irqfds.lock);
639}
640
641bool kvm_notify_irqfd_resampler(struct kvm *kvm,
642 unsigned int irqchip,
643 unsigned int pin)
644{
645 struct kvm_kernel_irqfd_resampler *resampler;
646 int gsi, idx;
647
648 idx = srcu_read_lock(&kvm->irq_srcu);
649 gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
650 if (gsi != -1) {
651 list_for_each_entry_srcu(resampler,
652 &kvm->irqfds.resampler_list, link,
653 srcu_read_lock_held(&kvm->irq_srcu)) {
654 if (resampler->notifier.gsi == gsi) {
655 irqfd_resampler_notify(resampler);
656 srcu_read_unlock(&kvm->irq_srcu, idx);
657 return true;
658 }
659 }
660 }
661 srcu_read_unlock(&kvm->irq_srcu, idx);
662
663 return false;
664}
665
666/*
667 * create a host-wide workqueue for issuing deferred shutdown requests
668 * aggregated from all vm* instances. We need our own isolated
669 * queue to ease flushing work items when a VM exits.
670 */
671int kvm_irqfd_init(void)
672{
673 irqfd_cleanup_wq = alloc_workqueue("kvm-irqfd-cleanup", 0, 0);
674 if (!irqfd_cleanup_wq)
675 return -ENOMEM;
676
677 return 0;
678}
679
680void kvm_irqfd_exit(void)
681{
682 destroy_workqueue(irqfd_cleanup_wq);
683}
684#endif
685
686/*
687 * --------------------------------------------------------------------
688 * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
689 *
690 * userspace can register a PIO/MMIO address with an eventfd for receiving
691 * notification when the memory has been touched.
692 * --------------------------------------------------------------------
693 */
694
695struct _ioeventfd {
696 struct list_head list;
697 u64 addr;
698 int length;
699 struct eventfd_ctx *eventfd;
700 u64 datamatch;
701 struct kvm_io_device dev;
702 u8 bus_idx;
703 bool wildcard;
704};
705
706static inline struct _ioeventfd *
707to_ioeventfd(struct kvm_io_device *dev)
708{
709 return container_of(dev, struct _ioeventfd, dev);
710}
711
712static void
713ioeventfd_release(struct _ioeventfd *p)
714{
715 eventfd_ctx_put(p->eventfd);
716 list_del(&p->list);
717 kfree(p);
718}
719
720static bool
721ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
722{
723 u64 _val;
724
725 if (addr != p->addr)
726 /* address must be precise for a hit */
727 return false;
728
729 if (!p->length)
730 /* length = 0 means only look at the address, so always a hit */
731 return true;
732
733 if (len != p->length)
734 /* address-range must be precise for a hit */
735 return false;
736
737 if (p->wildcard)
738 /* all else equal, wildcard is always a hit */
739 return true;
740
741 /* otherwise, we have to actually compare the data */
742
743 BUG_ON(!IS_ALIGNED((unsigned long)val, len));
744
745 switch (len) {
746 case 1:
747 _val = *(u8 *)val;
748 break;
749 case 2:
750 _val = *(u16 *)val;
751 break;
752 case 4:
753 _val = *(u32 *)val;
754 break;
755 case 8:
756 _val = *(u64 *)val;
757 break;
758 default:
759 return false;
760 }
761
762 return _val == p->datamatch;
763}
764
765/* MMIO/PIO writes trigger an event if the addr/val match */
766static int
767ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
768 int len, const void *val)
769{
770 struct _ioeventfd *p = to_ioeventfd(this);
771
772 if (!ioeventfd_in_range(p, addr, len, val))
773 return -EOPNOTSUPP;
774
775 eventfd_signal(p->eventfd);
776 return 0;
777}
778
779/*
780 * This function is called as KVM is completely shutting down. We do not
781 * need to worry about locking just nuke anything we have as quickly as possible
782 */
783static void
784ioeventfd_destructor(struct kvm_io_device *this)
785{
786 struct _ioeventfd *p = to_ioeventfd(this);
787
788 ioeventfd_release(p);
789}
790
791static const struct kvm_io_device_ops ioeventfd_ops = {
792 .write = ioeventfd_write,
793 .destructor = ioeventfd_destructor,
794};
795
796/* assumes kvm->slots_lock held */
797static bool
798ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
799{
800 struct _ioeventfd *_p;
801
802 list_for_each_entry(_p, &kvm->ioeventfds, list)
803 if (_p->bus_idx == p->bus_idx &&
804 _p->addr == p->addr &&
805 (!_p->length || !p->length ||
806 (_p->length == p->length &&
807 (_p->wildcard || p->wildcard ||
808 _p->datamatch == p->datamatch))))
809 return true;
810
811 return false;
812}
813
814static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
815{
816 if (flags & KVM_IOEVENTFD_FLAG_PIO)
817 return KVM_PIO_BUS;
818 if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
819 return KVM_VIRTIO_CCW_NOTIFY_BUS;
820 return KVM_MMIO_BUS;
821}
822
823static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
824 enum kvm_bus bus_idx,
825 struct kvm_ioeventfd *args)
826{
827
828 struct eventfd_ctx *eventfd;
829 struct _ioeventfd *p;
830 int ret;
831
832 eventfd = eventfd_ctx_fdget(args->fd);
833 if (IS_ERR(eventfd))
834 return PTR_ERR(eventfd);
835
836 p = kzalloc(sizeof(*p), GFP_KERNEL_ACCOUNT);
837 if (!p) {
838 ret = -ENOMEM;
839 goto fail;
840 }
841
842 INIT_LIST_HEAD(&p->list);
843 p->addr = args->addr;
844 p->bus_idx = bus_idx;
845 p->length = args->len;
846 p->eventfd = eventfd;
847
848 /* The datamatch feature is optional, otherwise this is a wildcard */
849 if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
850 p->datamatch = args->datamatch;
851 else
852 p->wildcard = true;
853
854 mutex_lock(&kvm->slots_lock);
855
856 /* Verify that there isn't a match already */
857 if (ioeventfd_check_collision(kvm, p)) {
858 ret = -EEXIST;
859 goto unlock_fail;
860 }
861
862 kvm_iodevice_init(&p->dev, &ioeventfd_ops);
863
864 ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
865 &p->dev);
866 if (ret < 0)
867 goto unlock_fail;
868
869 kvm_get_bus(kvm, bus_idx)->ioeventfd_count++;
870 list_add_tail(&p->list, &kvm->ioeventfds);
871
872 mutex_unlock(&kvm->slots_lock);
873
874 return 0;
875
876unlock_fail:
877 mutex_unlock(&kvm->slots_lock);
878 kfree(p);
879
880fail:
881 eventfd_ctx_put(eventfd);
882
883 return ret;
884}
885
886static int
887kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
888 struct kvm_ioeventfd *args)
889{
890 struct _ioeventfd *p;
891 struct eventfd_ctx *eventfd;
892 struct kvm_io_bus *bus;
893 int ret = -ENOENT;
894 bool wildcard;
895
896 eventfd = eventfd_ctx_fdget(args->fd);
897 if (IS_ERR(eventfd))
898 return PTR_ERR(eventfd);
899
900 wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
901
902 mutex_lock(&kvm->slots_lock);
903
904 list_for_each_entry(p, &kvm->ioeventfds, list) {
905 if (p->bus_idx != bus_idx ||
906 p->eventfd != eventfd ||
907 p->addr != args->addr ||
908 p->length != args->len ||
909 p->wildcard != wildcard)
910 continue;
911
912 if (!p->wildcard && p->datamatch != args->datamatch)
913 continue;
914
915 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
916 bus = kvm_get_bus(kvm, bus_idx);
917 if (bus)
918 bus->ioeventfd_count--;
919 ret = 0;
920 break;
921 }
922
923 mutex_unlock(&kvm->slots_lock);
924
925 eventfd_ctx_put(eventfd);
926
927 return ret;
928}
929
930static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
931{
932 enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
933 int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
934
935 if (!args->len && bus_idx == KVM_MMIO_BUS)
936 kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
937
938 return ret;
939}
940
941static int
942kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
943{
944 enum kvm_bus bus_idx;
945 int ret;
946
947 bus_idx = ioeventfd_bus_from_flags(args->flags);
948 /* must be natural-word sized, or 0 to ignore length */
949 switch (args->len) {
950 case 0:
951 case 1:
952 case 2:
953 case 4:
954 case 8:
955 break;
956 default:
957 return -EINVAL;
958 }
959
960 /* check for range overflow */
961 if (args->addr + args->len < args->addr)
962 return -EINVAL;
963
964 /* check for extra flags that we don't understand */
965 if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
966 return -EINVAL;
967
968 /* ioeventfd with no length can't be combined with DATAMATCH */
969 if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
970 return -EINVAL;
971
972 ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
973 if (ret)
974 goto fail;
975
976 /* When length is ignored, MMIO is also put on a separate bus, for
977 * faster lookups.
978 */
979 if (!args->len && bus_idx == KVM_MMIO_BUS) {
980 ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
981 if (ret < 0)
982 goto fast_fail;
983 }
984
985 return 0;
986
987fast_fail:
988 kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
989fail:
990 return ret;
991}
992
993int
994kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
995{
996 if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
997 return kvm_deassign_ioeventfd(kvm, args);
998
999 return kvm_assign_ioeventfd(kvm, args);
1000}
1001
1002void
1003kvm_eventfd_init(struct kvm *kvm)
1004{
1005#ifdef CONFIG_HAVE_KVM_IRQCHIP
1006 spin_lock_init(&kvm->irqfds.lock);
1007 INIT_LIST_HEAD(&kvm->irqfds.items);
1008 INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
1009 mutex_init(&kvm->irqfds.resampler_lock);
1010#endif
1011 INIT_LIST_HEAD(&kvm->ioeventfds);
1012}
1/*
2 * kvm eventfd support - use eventfd objects to signal various KVM events
3 *
4 * Copyright 2009 Novell. All Rights Reserved.
5 * Copyright 2010 Red Hat, Inc. and/or its affiliates.
6 *
7 * Author:
8 * Gregory Haskins <ghaskins@novell.com>
9 *
10 * This file is free software; you can redistribute it and/or modify
11 * it under the terms of version 2 of the GNU General Public License
12 * as published by the Free Software Foundation.
13 *
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
18 *
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software Foundation,
21 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA.
22 */
23
24#include <linux/kvm_host.h>
25#include <linux/kvm.h>
26#include <linux/kvm_irqfd.h>
27#include <linux/workqueue.h>
28#include <linux/syscalls.h>
29#include <linux/wait.h>
30#include <linux/poll.h>
31#include <linux/file.h>
32#include <linux/list.h>
33#include <linux/eventfd.h>
34#include <linux/kernel.h>
35#include <linux/srcu.h>
36#include <linux/slab.h>
37#include <linux/seqlock.h>
38#include <linux/irqbypass.h>
39#include <trace/events/kvm.h>
40
41#include <kvm/iodev.h>
42
43#ifdef CONFIG_HAVE_KVM_IRQFD
44
45static struct workqueue_struct *irqfd_cleanup_wq;
46
47static void
48irqfd_inject(struct work_struct *work)
49{
50 struct kvm_kernel_irqfd *irqfd =
51 container_of(work, struct kvm_kernel_irqfd, inject);
52 struct kvm *kvm = irqfd->kvm;
53
54 if (!irqfd->resampler) {
55 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
56 false);
57 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
58 false);
59 } else
60 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
61 irqfd->gsi, 1, false);
62}
63
64/*
65 * Since resampler irqfds share an IRQ source ID, we de-assert once
66 * then notify all of the resampler irqfds using this GSI. We can't
67 * do multiple de-asserts or we risk racing with incoming re-asserts.
68 */
69static void
70irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
71{
72 struct kvm_kernel_irqfd_resampler *resampler;
73 struct kvm *kvm;
74 struct kvm_kernel_irqfd *irqfd;
75 int idx;
76
77 resampler = container_of(kian,
78 struct kvm_kernel_irqfd_resampler, notifier);
79 kvm = resampler->kvm;
80
81 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
82 resampler->notifier.gsi, 0, false);
83
84 idx = srcu_read_lock(&kvm->irq_srcu);
85
86 list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
87 eventfd_signal(irqfd->resamplefd, 1);
88
89 srcu_read_unlock(&kvm->irq_srcu, idx);
90}
91
92static void
93irqfd_resampler_shutdown(struct kvm_kernel_irqfd *irqfd)
94{
95 struct kvm_kernel_irqfd_resampler *resampler = irqfd->resampler;
96 struct kvm *kvm = resampler->kvm;
97
98 mutex_lock(&kvm->irqfds.resampler_lock);
99
100 list_del_rcu(&irqfd->resampler_link);
101 synchronize_srcu(&kvm->irq_srcu);
102
103 if (list_empty(&resampler->list)) {
104 list_del(&resampler->link);
105 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
106 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
107 resampler->notifier.gsi, 0, false);
108 kfree(resampler);
109 }
110
111 mutex_unlock(&kvm->irqfds.resampler_lock);
112}
113
114/*
115 * Race-free decouple logic (ordering is critical)
116 */
117static void
118irqfd_shutdown(struct work_struct *work)
119{
120 struct kvm_kernel_irqfd *irqfd =
121 container_of(work, struct kvm_kernel_irqfd, shutdown);
122 u64 cnt;
123
124 /*
125 * Synchronize with the wait-queue and unhook ourselves to prevent
126 * further events.
127 */
128 eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
129
130 /*
131 * We know no new events will be scheduled at this point, so block
132 * until all previously outstanding events have completed
133 */
134 flush_work(&irqfd->inject);
135
136 if (irqfd->resampler) {
137 irqfd_resampler_shutdown(irqfd);
138 eventfd_ctx_put(irqfd->resamplefd);
139 }
140
141 /*
142 * It is now safe to release the object's resources
143 */
144#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
145 irq_bypass_unregister_consumer(&irqfd->consumer);
146#endif
147 eventfd_ctx_put(irqfd->eventfd);
148 kfree(irqfd);
149}
150
151
152/* assumes kvm->irqfds.lock is held */
153static bool
154irqfd_is_active(struct kvm_kernel_irqfd *irqfd)
155{
156 return list_empty(&irqfd->list) ? false : true;
157}
158
159/*
160 * Mark the irqfd as inactive and schedule it for removal
161 *
162 * assumes kvm->irqfds.lock is held
163 */
164static void
165irqfd_deactivate(struct kvm_kernel_irqfd *irqfd)
166{
167 BUG_ON(!irqfd_is_active(irqfd));
168
169 list_del_init(&irqfd->list);
170
171 queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
172}
173
174int __attribute__((weak)) kvm_arch_set_irq_inatomic(
175 struct kvm_kernel_irq_routing_entry *irq,
176 struct kvm *kvm, int irq_source_id,
177 int level,
178 bool line_status)
179{
180 return -EWOULDBLOCK;
181}
182
183/*
184 * Called with wqh->lock held and interrupts disabled
185 */
186static int
187irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
188{
189 struct kvm_kernel_irqfd *irqfd =
190 container_of(wait, struct kvm_kernel_irqfd, wait);
191 unsigned long flags = (unsigned long)key;
192 struct kvm_kernel_irq_routing_entry irq;
193 struct kvm *kvm = irqfd->kvm;
194 unsigned seq;
195 int idx;
196
197 if (flags & POLLIN) {
198 idx = srcu_read_lock(&kvm->irq_srcu);
199 do {
200 seq = read_seqcount_begin(&irqfd->irq_entry_sc);
201 irq = irqfd->irq_entry;
202 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
203 /* An event has been signaled, inject an interrupt */
204 if (kvm_arch_set_irq_inatomic(&irq, kvm,
205 KVM_USERSPACE_IRQ_SOURCE_ID, 1,
206 false) == -EWOULDBLOCK)
207 schedule_work(&irqfd->inject);
208 srcu_read_unlock(&kvm->irq_srcu, idx);
209 }
210
211 if (flags & POLLHUP) {
212 /* The eventfd is closing, detach from KVM */
213 unsigned long flags;
214
215 spin_lock_irqsave(&kvm->irqfds.lock, flags);
216
217 /*
218 * We must check if someone deactivated the irqfd before
219 * we could acquire the irqfds.lock since the item is
220 * deactivated from the KVM side before it is unhooked from
221 * the wait-queue. If it is already deactivated, we can
222 * simply return knowing the other side will cleanup for us.
223 * We cannot race against the irqfd going away since the
224 * other side is required to acquire wqh->lock, which we hold
225 */
226 if (irqfd_is_active(irqfd))
227 irqfd_deactivate(irqfd);
228
229 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
230 }
231
232 return 0;
233}
234
235static void
236irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
237 poll_table *pt)
238{
239 struct kvm_kernel_irqfd *irqfd =
240 container_of(pt, struct kvm_kernel_irqfd, pt);
241 add_wait_queue(wqh, &irqfd->wait);
242}
243
244/* Must be called under irqfds.lock */
245static void irqfd_update(struct kvm *kvm, struct kvm_kernel_irqfd *irqfd)
246{
247 struct kvm_kernel_irq_routing_entry *e;
248 struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
249 int n_entries;
250
251 n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
252
253 write_seqcount_begin(&irqfd->irq_entry_sc);
254
255 e = entries;
256 if (n_entries == 1)
257 irqfd->irq_entry = *e;
258 else
259 irqfd->irq_entry.type = 0;
260
261 write_seqcount_end(&irqfd->irq_entry_sc);
262}
263
264#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
265void __attribute__((weak)) kvm_arch_irq_bypass_stop(
266 struct irq_bypass_consumer *cons)
267{
268}
269
270void __attribute__((weak)) kvm_arch_irq_bypass_start(
271 struct irq_bypass_consumer *cons)
272{
273}
274
275int __attribute__((weak)) kvm_arch_update_irqfd_routing(
276 struct kvm *kvm, unsigned int host_irq,
277 uint32_t guest_irq, bool set)
278{
279 return 0;
280}
281#endif
282
283static int
284kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
285{
286 struct kvm_kernel_irqfd *irqfd, *tmp;
287 struct fd f;
288 struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
289 int ret;
290 unsigned int events;
291 int idx;
292
293 if (!kvm_arch_intc_initialized(kvm))
294 return -EAGAIN;
295
296 irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
297 if (!irqfd)
298 return -ENOMEM;
299
300 irqfd->kvm = kvm;
301 irqfd->gsi = args->gsi;
302 INIT_LIST_HEAD(&irqfd->list);
303 INIT_WORK(&irqfd->inject, irqfd_inject);
304 INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
305 seqcount_init(&irqfd->irq_entry_sc);
306
307 f = fdget(args->fd);
308 if (!f.file) {
309 ret = -EBADF;
310 goto out;
311 }
312
313 eventfd = eventfd_ctx_fileget(f.file);
314 if (IS_ERR(eventfd)) {
315 ret = PTR_ERR(eventfd);
316 goto fail;
317 }
318
319 irqfd->eventfd = eventfd;
320
321 if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
322 struct kvm_kernel_irqfd_resampler *resampler;
323
324 resamplefd = eventfd_ctx_fdget(args->resamplefd);
325 if (IS_ERR(resamplefd)) {
326 ret = PTR_ERR(resamplefd);
327 goto fail;
328 }
329
330 irqfd->resamplefd = resamplefd;
331 INIT_LIST_HEAD(&irqfd->resampler_link);
332
333 mutex_lock(&kvm->irqfds.resampler_lock);
334
335 list_for_each_entry(resampler,
336 &kvm->irqfds.resampler_list, link) {
337 if (resampler->notifier.gsi == irqfd->gsi) {
338 irqfd->resampler = resampler;
339 break;
340 }
341 }
342
343 if (!irqfd->resampler) {
344 resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
345 if (!resampler) {
346 ret = -ENOMEM;
347 mutex_unlock(&kvm->irqfds.resampler_lock);
348 goto fail;
349 }
350
351 resampler->kvm = kvm;
352 INIT_LIST_HEAD(&resampler->list);
353 resampler->notifier.gsi = irqfd->gsi;
354 resampler->notifier.irq_acked = irqfd_resampler_ack;
355 INIT_LIST_HEAD(&resampler->link);
356
357 list_add(&resampler->link, &kvm->irqfds.resampler_list);
358 kvm_register_irq_ack_notifier(kvm,
359 &resampler->notifier);
360 irqfd->resampler = resampler;
361 }
362
363 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
364 synchronize_srcu(&kvm->irq_srcu);
365
366 mutex_unlock(&kvm->irqfds.resampler_lock);
367 }
368
369 /*
370 * Install our own custom wake-up handling so we are notified via
371 * a callback whenever someone signals the underlying eventfd
372 */
373 init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
374 init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
375
376 spin_lock_irq(&kvm->irqfds.lock);
377
378 ret = 0;
379 list_for_each_entry(tmp, &kvm->irqfds.items, list) {
380 if (irqfd->eventfd != tmp->eventfd)
381 continue;
382 /* This fd is used for another irq already. */
383 ret = -EBUSY;
384 spin_unlock_irq(&kvm->irqfds.lock);
385 goto fail;
386 }
387
388 idx = srcu_read_lock(&kvm->irq_srcu);
389 irqfd_update(kvm, irqfd);
390 srcu_read_unlock(&kvm->irq_srcu, idx);
391
392 list_add_tail(&irqfd->list, &kvm->irqfds.items);
393
394 spin_unlock_irq(&kvm->irqfds.lock);
395
396 /*
397 * Check if there was an event already pending on the eventfd
398 * before we registered, and trigger it as if we didn't miss it.
399 */
400 events = f.file->f_op->poll(f.file, &irqfd->pt);
401
402 if (events & POLLIN)
403 schedule_work(&irqfd->inject);
404
405 /*
406 * do not drop the file until the irqfd is fully initialized, otherwise
407 * we might race against the POLLHUP
408 */
409 fdput(f);
410#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
411 if (kvm_arch_has_irq_bypass()) {
412 irqfd->consumer.token = (void *)irqfd->eventfd;
413 irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
414 irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
415 irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
416 irqfd->consumer.start = kvm_arch_irq_bypass_start;
417 ret = irq_bypass_register_consumer(&irqfd->consumer);
418 if (ret)
419 pr_info("irq bypass consumer (token %p) registration fails: %d\n",
420 irqfd->consumer.token, ret);
421 }
422#endif
423
424 return 0;
425
426fail:
427 if (irqfd->resampler)
428 irqfd_resampler_shutdown(irqfd);
429
430 if (resamplefd && !IS_ERR(resamplefd))
431 eventfd_ctx_put(resamplefd);
432
433 if (eventfd && !IS_ERR(eventfd))
434 eventfd_ctx_put(eventfd);
435
436 fdput(f);
437
438out:
439 kfree(irqfd);
440 return ret;
441}
442
443bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
444{
445 struct kvm_irq_ack_notifier *kian;
446 int gsi, idx;
447
448 idx = srcu_read_lock(&kvm->irq_srcu);
449 gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
450 if (gsi != -1)
451 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
452 link)
453 if (kian->gsi == gsi) {
454 srcu_read_unlock(&kvm->irq_srcu, idx);
455 return true;
456 }
457
458 srcu_read_unlock(&kvm->irq_srcu, idx);
459
460 return false;
461}
462EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
463
464void kvm_notify_acked_gsi(struct kvm *kvm, int gsi)
465{
466 struct kvm_irq_ack_notifier *kian;
467
468 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
469 link)
470 if (kian->gsi == gsi)
471 kian->irq_acked(kian);
472}
473
474void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
475{
476 int gsi, idx;
477
478 trace_kvm_ack_irq(irqchip, pin);
479
480 idx = srcu_read_lock(&kvm->irq_srcu);
481 gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
482 if (gsi != -1)
483 kvm_notify_acked_gsi(kvm, gsi);
484 srcu_read_unlock(&kvm->irq_srcu, idx);
485}
486
487void kvm_register_irq_ack_notifier(struct kvm *kvm,
488 struct kvm_irq_ack_notifier *kian)
489{
490 mutex_lock(&kvm->irq_lock);
491 hlist_add_head_rcu(&kian->link, &kvm->irq_ack_notifier_list);
492 mutex_unlock(&kvm->irq_lock);
493 kvm_vcpu_request_scan_ioapic(kvm);
494}
495
496void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
497 struct kvm_irq_ack_notifier *kian)
498{
499 mutex_lock(&kvm->irq_lock);
500 hlist_del_init_rcu(&kian->link);
501 mutex_unlock(&kvm->irq_lock);
502 synchronize_srcu(&kvm->irq_srcu);
503 kvm_vcpu_request_scan_ioapic(kvm);
504}
505#endif
506
507void
508kvm_eventfd_init(struct kvm *kvm)
509{
510#ifdef CONFIG_HAVE_KVM_IRQFD
511 spin_lock_init(&kvm->irqfds.lock);
512 INIT_LIST_HEAD(&kvm->irqfds.items);
513 INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
514 mutex_init(&kvm->irqfds.resampler_lock);
515#endif
516 INIT_LIST_HEAD(&kvm->ioeventfds);
517}
518
519#ifdef CONFIG_HAVE_KVM_IRQFD
520/*
521 * shutdown any irqfd's that match fd+gsi
522 */
523static int
524kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
525{
526 struct kvm_kernel_irqfd *irqfd, *tmp;
527 struct eventfd_ctx *eventfd;
528
529 eventfd = eventfd_ctx_fdget(args->fd);
530 if (IS_ERR(eventfd))
531 return PTR_ERR(eventfd);
532
533 spin_lock_irq(&kvm->irqfds.lock);
534
535 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
536 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
537 /*
538 * This clearing of irq_entry.type is needed for when
539 * another thread calls kvm_irq_routing_update before
540 * we flush workqueue below (we synchronize with
541 * kvm_irq_routing_update using irqfds.lock).
542 */
543 write_seqcount_begin(&irqfd->irq_entry_sc);
544 irqfd->irq_entry.type = 0;
545 write_seqcount_end(&irqfd->irq_entry_sc);
546 irqfd_deactivate(irqfd);
547 }
548 }
549
550 spin_unlock_irq(&kvm->irqfds.lock);
551 eventfd_ctx_put(eventfd);
552
553 /*
554 * Block until we know all outstanding shutdown jobs have completed
555 * so that we guarantee there will not be any more interrupts on this
556 * gsi once this deassign function returns.
557 */
558 flush_workqueue(irqfd_cleanup_wq);
559
560 return 0;
561}
562
563int
564kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
565{
566 if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
567 return -EINVAL;
568
569 if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
570 return kvm_irqfd_deassign(kvm, args);
571
572 return kvm_irqfd_assign(kvm, args);
573}
574
575/*
576 * This function is called as the kvm VM fd is being released. Shutdown all
577 * irqfds that still remain open
578 */
579void
580kvm_irqfd_release(struct kvm *kvm)
581{
582 struct kvm_kernel_irqfd *irqfd, *tmp;
583
584 spin_lock_irq(&kvm->irqfds.lock);
585
586 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
587 irqfd_deactivate(irqfd);
588
589 spin_unlock_irq(&kvm->irqfds.lock);
590
591 /*
592 * Block until we know all outstanding shutdown jobs have completed
593 * since we do not take a kvm* reference.
594 */
595 flush_workqueue(irqfd_cleanup_wq);
596
597}
598
599/*
600 * Take note of a change in irq routing.
601 * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
602 */
603void kvm_irq_routing_update(struct kvm *kvm)
604{
605 struct kvm_kernel_irqfd *irqfd;
606
607 spin_lock_irq(&kvm->irqfds.lock);
608
609 list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
610 irqfd_update(kvm, irqfd);
611
612#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
613 if (irqfd->producer) {
614 int ret = kvm_arch_update_irqfd_routing(
615 irqfd->kvm, irqfd->producer->irq,
616 irqfd->gsi, 1);
617 WARN_ON(ret);
618 }
619#endif
620 }
621
622 spin_unlock_irq(&kvm->irqfds.lock);
623}
624
625/*
626 * create a host-wide workqueue for issuing deferred shutdown requests
627 * aggregated from all vm* instances. We need our own isolated
628 * queue to ease flushing work items when a VM exits.
629 */
630int kvm_irqfd_init(void)
631{
632 irqfd_cleanup_wq = alloc_workqueue("kvm-irqfd-cleanup", 0, 0);
633 if (!irqfd_cleanup_wq)
634 return -ENOMEM;
635
636 return 0;
637}
638
639void kvm_irqfd_exit(void)
640{
641 destroy_workqueue(irqfd_cleanup_wq);
642}
643#endif
644
645/*
646 * --------------------------------------------------------------------
647 * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
648 *
649 * userspace can register a PIO/MMIO address with an eventfd for receiving
650 * notification when the memory has been touched.
651 * --------------------------------------------------------------------
652 */
653
654struct _ioeventfd {
655 struct list_head list;
656 u64 addr;
657 int length;
658 struct eventfd_ctx *eventfd;
659 u64 datamatch;
660 struct kvm_io_device dev;
661 u8 bus_idx;
662 bool wildcard;
663};
664
665static inline struct _ioeventfd *
666to_ioeventfd(struct kvm_io_device *dev)
667{
668 return container_of(dev, struct _ioeventfd, dev);
669}
670
671static void
672ioeventfd_release(struct _ioeventfd *p)
673{
674 eventfd_ctx_put(p->eventfd);
675 list_del(&p->list);
676 kfree(p);
677}
678
679static bool
680ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
681{
682 u64 _val;
683
684 if (addr != p->addr)
685 /* address must be precise for a hit */
686 return false;
687
688 if (!p->length)
689 /* length = 0 means only look at the address, so always a hit */
690 return true;
691
692 if (len != p->length)
693 /* address-range must be precise for a hit */
694 return false;
695
696 if (p->wildcard)
697 /* all else equal, wildcard is always a hit */
698 return true;
699
700 /* otherwise, we have to actually compare the data */
701
702 BUG_ON(!IS_ALIGNED((unsigned long)val, len));
703
704 switch (len) {
705 case 1:
706 _val = *(u8 *)val;
707 break;
708 case 2:
709 _val = *(u16 *)val;
710 break;
711 case 4:
712 _val = *(u32 *)val;
713 break;
714 case 8:
715 _val = *(u64 *)val;
716 break;
717 default:
718 return false;
719 }
720
721 return _val == p->datamatch ? true : false;
722}
723
724/* MMIO/PIO writes trigger an event if the addr/val match */
725static int
726ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
727 int len, const void *val)
728{
729 struct _ioeventfd *p = to_ioeventfd(this);
730
731 if (!ioeventfd_in_range(p, addr, len, val))
732 return -EOPNOTSUPP;
733
734 eventfd_signal(p->eventfd, 1);
735 return 0;
736}
737
738/*
739 * This function is called as KVM is completely shutting down. We do not
740 * need to worry about locking just nuke anything we have as quickly as possible
741 */
742static void
743ioeventfd_destructor(struct kvm_io_device *this)
744{
745 struct _ioeventfd *p = to_ioeventfd(this);
746
747 ioeventfd_release(p);
748}
749
750static const struct kvm_io_device_ops ioeventfd_ops = {
751 .write = ioeventfd_write,
752 .destructor = ioeventfd_destructor,
753};
754
755/* assumes kvm->slots_lock held */
756static bool
757ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
758{
759 struct _ioeventfd *_p;
760
761 list_for_each_entry(_p, &kvm->ioeventfds, list)
762 if (_p->bus_idx == p->bus_idx &&
763 _p->addr == p->addr &&
764 (!_p->length || !p->length ||
765 (_p->length == p->length &&
766 (_p->wildcard || p->wildcard ||
767 _p->datamatch == p->datamatch))))
768 return true;
769
770 return false;
771}
772
773static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
774{
775 if (flags & KVM_IOEVENTFD_FLAG_PIO)
776 return KVM_PIO_BUS;
777 if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
778 return KVM_VIRTIO_CCW_NOTIFY_BUS;
779 return KVM_MMIO_BUS;
780}
781
782static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
783 enum kvm_bus bus_idx,
784 struct kvm_ioeventfd *args)
785{
786
787 struct eventfd_ctx *eventfd;
788 struct _ioeventfd *p;
789 int ret;
790
791 eventfd = eventfd_ctx_fdget(args->fd);
792 if (IS_ERR(eventfd))
793 return PTR_ERR(eventfd);
794
795 p = kzalloc(sizeof(*p), GFP_KERNEL);
796 if (!p) {
797 ret = -ENOMEM;
798 goto fail;
799 }
800
801 INIT_LIST_HEAD(&p->list);
802 p->addr = args->addr;
803 p->bus_idx = bus_idx;
804 p->length = args->len;
805 p->eventfd = eventfd;
806
807 /* The datamatch feature is optional, otherwise this is a wildcard */
808 if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
809 p->datamatch = args->datamatch;
810 else
811 p->wildcard = true;
812
813 mutex_lock(&kvm->slots_lock);
814
815 /* Verify that there isn't a match already */
816 if (ioeventfd_check_collision(kvm, p)) {
817 ret = -EEXIST;
818 goto unlock_fail;
819 }
820
821 kvm_iodevice_init(&p->dev, &ioeventfd_ops);
822
823 ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
824 &p->dev);
825 if (ret < 0)
826 goto unlock_fail;
827
828 kvm->buses[bus_idx]->ioeventfd_count++;
829 list_add_tail(&p->list, &kvm->ioeventfds);
830
831 mutex_unlock(&kvm->slots_lock);
832
833 return 0;
834
835unlock_fail:
836 mutex_unlock(&kvm->slots_lock);
837
838fail:
839 kfree(p);
840 eventfd_ctx_put(eventfd);
841
842 return ret;
843}
844
845static int
846kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
847 struct kvm_ioeventfd *args)
848{
849 struct _ioeventfd *p, *tmp;
850 struct eventfd_ctx *eventfd;
851 int ret = -ENOENT;
852
853 eventfd = eventfd_ctx_fdget(args->fd);
854 if (IS_ERR(eventfd))
855 return PTR_ERR(eventfd);
856
857 mutex_lock(&kvm->slots_lock);
858
859 list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
860 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
861
862 if (p->bus_idx != bus_idx ||
863 p->eventfd != eventfd ||
864 p->addr != args->addr ||
865 p->length != args->len ||
866 p->wildcard != wildcard)
867 continue;
868
869 if (!p->wildcard && p->datamatch != args->datamatch)
870 continue;
871
872 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
873 if (kvm->buses[bus_idx])
874 kvm->buses[bus_idx]->ioeventfd_count--;
875 ioeventfd_release(p);
876 ret = 0;
877 break;
878 }
879
880 mutex_unlock(&kvm->slots_lock);
881
882 eventfd_ctx_put(eventfd);
883
884 return ret;
885}
886
887static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
888{
889 enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
890 int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
891
892 if (!args->len && bus_idx == KVM_MMIO_BUS)
893 kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
894
895 return ret;
896}
897
898static int
899kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
900{
901 enum kvm_bus bus_idx;
902 int ret;
903
904 bus_idx = ioeventfd_bus_from_flags(args->flags);
905 /* must be natural-word sized, or 0 to ignore length */
906 switch (args->len) {
907 case 0:
908 case 1:
909 case 2:
910 case 4:
911 case 8:
912 break;
913 default:
914 return -EINVAL;
915 }
916
917 /* check for range overflow */
918 if (args->addr + args->len < args->addr)
919 return -EINVAL;
920
921 /* check for extra flags that we don't understand */
922 if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
923 return -EINVAL;
924
925 /* ioeventfd with no length can't be combined with DATAMATCH */
926 if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
927 return -EINVAL;
928
929 ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
930 if (ret)
931 goto fail;
932
933 /* When length is ignored, MMIO is also put on a separate bus, for
934 * faster lookups.
935 */
936 if (!args->len && bus_idx == KVM_MMIO_BUS) {
937 ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
938 if (ret < 0)
939 goto fast_fail;
940 }
941
942 return 0;
943
944fast_fail:
945 kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
946fail:
947 return ret;
948}
949
950int
951kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
952{
953 if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
954 return kvm_deassign_ioeventfd(kvm, args);
955
956 return kvm_assign_ioeventfd(kvm, args);
957}