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/workqueue.h>
27#include <linux/syscalls.h>
28#include <linux/wait.h>
29#include <linux/poll.h>
30#include <linux/file.h>
31#include <linux/list.h>
32#include <linux/eventfd.h>
33#include <linux/kernel.h>
34#include <linux/slab.h>
35
36#include "iodev.h"
37
38#ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
39/*
40 * --------------------------------------------------------------------
41 * irqfd: Allows an fd to be used to inject an interrupt to the guest
42 *
43 * Credit goes to Avi Kivity for the original idea.
44 * --------------------------------------------------------------------
45 */
46
47/*
48 * Resampling irqfds are a special variety of irqfds used to emulate
49 * level triggered interrupts. The interrupt is asserted on eventfd
50 * trigger. On acknowledgement through the irq ack notifier, the
51 * interrupt is de-asserted and userspace is notified through the
52 * resamplefd. All resamplers on the same gsi are de-asserted
53 * together, so we don't need to track the state of each individual
54 * user. We can also therefore share the same irq source ID.
55 */
56struct _irqfd_resampler {
57 struct kvm *kvm;
58 /*
59 * List of resampling struct _irqfd objects sharing this gsi.
60 * RCU list modified under kvm->irqfds.resampler_lock
61 */
62 struct list_head list;
63 struct kvm_irq_ack_notifier notifier;
64 /*
65 * Entry in list of kvm->irqfd.resampler_list. Use for sharing
66 * resamplers among irqfds on the same gsi.
67 * Accessed and modified under kvm->irqfds.resampler_lock
68 */
69 struct list_head link;
70};
71
72struct _irqfd {
73 /* Used for MSI fast-path */
74 struct kvm *kvm;
75 wait_queue_t wait;
76 /* Update side is protected by irqfds.lock */
77 struct kvm_kernel_irq_routing_entry __rcu *irq_entry;
78 /* Used for level IRQ fast-path */
79 int gsi;
80 struct work_struct inject;
81 /* The resampler used by this irqfd (resampler-only) */
82 struct _irqfd_resampler *resampler;
83 /* Eventfd notified on resample (resampler-only) */
84 struct eventfd_ctx *resamplefd;
85 /* Entry in list of irqfds for a resampler (resampler-only) */
86 struct list_head resampler_link;
87 /* Used for setup/shutdown */
88 struct eventfd_ctx *eventfd;
89 struct list_head list;
90 poll_table pt;
91 struct work_struct shutdown;
92};
93
94static struct workqueue_struct *irqfd_cleanup_wq;
95
96static void
97irqfd_inject(struct work_struct *work)
98{
99 struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
100 struct kvm *kvm = irqfd->kvm;
101
102 if (!irqfd->resampler) {
103 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
104 false);
105 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
106 false);
107 } else
108 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
109 irqfd->gsi, 1, false);
110}
111
112/*
113 * Since resampler irqfds share an IRQ source ID, we de-assert once
114 * then notify all of the resampler irqfds using this GSI. We can't
115 * do multiple de-asserts or we risk racing with incoming re-asserts.
116 */
117static void
118irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
119{
120 struct _irqfd_resampler *resampler;
121 struct _irqfd *irqfd;
122
123 resampler = container_of(kian, struct _irqfd_resampler, notifier);
124
125 kvm_set_irq(resampler->kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
126 resampler->notifier.gsi, 0, false);
127
128 rcu_read_lock();
129
130 list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
131 eventfd_signal(irqfd->resamplefd, 1);
132
133 rcu_read_unlock();
134}
135
136static void
137irqfd_resampler_shutdown(struct _irqfd *irqfd)
138{
139 struct _irqfd_resampler *resampler = irqfd->resampler;
140 struct kvm *kvm = resampler->kvm;
141
142 mutex_lock(&kvm->irqfds.resampler_lock);
143
144 list_del_rcu(&irqfd->resampler_link);
145 synchronize_rcu();
146
147 if (list_empty(&resampler->list)) {
148 list_del(&resampler->link);
149 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
150 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
151 resampler->notifier.gsi, 0, false);
152 kfree(resampler);
153 }
154
155 mutex_unlock(&kvm->irqfds.resampler_lock);
156}
157
158/*
159 * Race-free decouple logic (ordering is critical)
160 */
161static void
162irqfd_shutdown(struct work_struct *work)
163{
164 struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
165 u64 cnt;
166
167 /*
168 * Synchronize with the wait-queue and unhook ourselves to prevent
169 * further events.
170 */
171 eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
172
173 /*
174 * We know no new events will be scheduled at this point, so block
175 * until all previously outstanding events have completed
176 */
177 flush_work(&irqfd->inject);
178
179 if (irqfd->resampler) {
180 irqfd_resampler_shutdown(irqfd);
181 eventfd_ctx_put(irqfd->resamplefd);
182 }
183
184 /*
185 * It is now safe to release the object's resources
186 */
187 eventfd_ctx_put(irqfd->eventfd);
188 kfree(irqfd);
189}
190
191
192/* assumes kvm->irqfds.lock is held */
193static bool
194irqfd_is_active(struct _irqfd *irqfd)
195{
196 return list_empty(&irqfd->list) ? false : true;
197}
198
199/*
200 * Mark the irqfd as inactive and schedule it for removal
201 *
202 * assumes kvm->irqfds.lock is held
203 */
204static void
205irqfd_deactivate(struct _irqfd *irqfd)
206{
207 BUG_ON(!irqfd_is_active(irqfd));
208
209 list_del_init(&irqfd->list);
210
211 queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
212}
213
214/*
215 * Called with wqh->lock held and interrupts disabled
216 */
217static int
218irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
219{
220 struct _irqfd *irqfd = container_of(wait, struct _irqfd, wait);
221 unsigned long flags = (unsigned long)key;
222 struct kvm_kernel_irq_routing_entry *irq;
223 struct kvm *kvm = irqfd->kvm;
224
225 if (flags & POLLIN) {
226 rcu_read_lock();
227 irq = rcu_dereference(irqfd->irq_entry);
228 /* An event has been signaled, inject an interrupt */
229 if (irq)
230 kvm_set_msi(irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1,
231 false);
232 else
233 schedule_work(&irqfd->inject);
234 rcu_read_unlock();
235 }
236
237 if (flags & POLLHUP) {
238 /* The eventfd is closing, detach from KVM */
239 unsigned long flags;
240
241 spin_lock_irqsave(&kvm->irqfds.lock, flags);
242
243 /*
244 * We must check if someone deactivated the irqfd before
245 * we could acquire the irqfds.lock since the item is
246 * deactivated from the KVM side before it is unhooked from
247 * the wait-queue. If it is already deactivated, we can
248 * simply return knowing the other side will cleanup for us.
249 * We cannot race against the irqfd going away since the
250 * other side is required to acquire wqh->lock, which we hold
251 */
252 if (irqfd_is_active(irqfd))
253 irqfd_deactivate(irqfd);
254
255 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
256 }
257
258 return 0;
259}
260
261static void
262irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
263 poll_table *pt)
264{
265 struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
266 add_wait_queue(wqh, &irqfd->wait);
267}
268
269/* Must be called under irqfds.lock */
270static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd,
271 struct kvm_irq_routing_table *irq_rt)
272{
273 struct kvm_kernel_irq_routing_entry *e;
274
275 if (irqfd->gsi >= irq_rt->nr_rt_entries) {
276 rcu_assign_pointer(irqfd->irq_entry, NULL);
277 return;
278 }
279
280 hlist_for_each_entry(e, &irq_rt->map[irqfd->gsi], link) {
281 /* Only fast-path MSI. */
282 if (e->type == KVM_IRQ_ROUTING_MSI)
283 rcu_assign_pointer(irqfd->irq_entry, e);
284 else
285 rcu_assign_pointer(irqfd->irq_entry, NULL);
286 }
287}
288
289static int
290kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
291{
292 struct kvm_irq_routing_table *irq_rt;
293 struct _irqfd *irqfd, *tmp;
294 struct fd f;
295 struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
296 int ret;
297 unsigned int events;
298
299 irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
300 if (!irqfd)
301 return -ENOMEM;
302
303 irqfd->kvm = kvm;
304 irqfd->gsi = args->gsi;
305 INIT_LIST_HEAD(&irqfd->list);
306 INIT_WORK(&irqfd->inject, irqfd_inject);
307 INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
308
309 f = fdget(args->fd);
310 if (!f.file) {
311 ret = -EBADF;
312 goto out;
313 }
314
315 eventfd = eventfd_ctx_fileget(f.file);
316 if (IS_ERR(eventfd)) {
317 ret = PTR_ERR(eventfd);
318 goto fail;
319 }
320
321 irqfd->eventfd = eventfd;
322
323 if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
324 struct _irqfd_resampler *resampler;
325
326 resamplefd = eventfd_ctx_fdget(args->resamplefd);
327 if (IS_ERR(resamplefd)) {
328 ret = PTR_ERR(resamplefd);
329 goto fail;
330 }
331
332 irqfd->resamplefd = resamplefd;
333 INIT_LIST_HEAD(&irqfd->resampler_link);
334
335 mutex_lock(&kvm->irqfds.resampler_lock);
336
337 list_for_each_entry(resampler,
338 &kvm->irqfds.resampler_list, link) {
339 if (resampler->notifier.gsi == irqfd->gsi) {
340 irqfd->resampler = resampler;
341 break;
342 }
343 }
344
345 if (!irqfd->resampler) {
346 resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
347 if (!resampler) {
348 ret = -ENOMEM;
349 mutex_unlock(&kvm->irqfds.resampler_lock);
350 goto fail;
351 }
352
353 resampler->kvm = kvm;
354 INIT_LIST_HEAD(&resampler->list);
355 resampler->notifier.gsi = irqfd->gsi;
356 resampler->notifier.irq_acked = irqfd_resampler_ack;
357 INIT_LIST_HEAD(&resampler->link);
358
359 list_add(&resampler->link, &kvm->irqfds.resampler_list);
360 kvm_register_irq_ack_notifier(kvm,
361 &resampler->notifier);
362 irqfd->resampler = resampler;
363 }
364
365 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
366 synchronize_rcu();
367
368 mutex_unlock(&kvm->irqfds.resampler_lock);
369 }
370
371 /*
372 * Install our own custom wake-up handling so we are notified via
373 * a callback whenever someone signals the underlying eventfd
374 */
375 init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
376 init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
377
378 spin_lock_irq(&kvm->irqfds.lock);
379
380 ret = 0;
381 list_for_each_entry(tmp, &kvm->irqfds.items, list) {
382 if (irqfd->eventfd != tmp->eventfd)
383 continue;
384 /* This fd is used for another irq already. */
385 ret = -EBUSY;
386 spin_unlock_irq(&kvm->irqfds.lock);
387 goto fail;
388 }
389
390 irq_rt = rcu_dereference_protected(kvm->irq_routing,
391 lockdep_is_held(&kvm->irqfds.lock));
392 irqfd_update(kvm, irqfd, irq_rt);
393
394 list_add_tail(&irqfd->list, &kvm->irqfds.items);
395
396 spin_unlock_irq(&kvm->irqfds.lock);
397
398 /*
399 * Check if there was an event already pending on the eventfd
400 * before we registered, and trigger it as if we didn't miss it.
401 */
402 events = f.file->f_op->poll(f.file, &irqfd->pt);
403
404 if (events & POLLIN)
405 schedule_work(&irqfd->inject);
406
407 /*
408 * do not drop the file until the irqfd is fully initialized, otherwise
409 * we might race against the POLLHUP
410 */
411 fdput(f);
412
413 return 0;
414
415fail:
416 if (irqfd->resampler)
417 irqfd_resampler_shutdown(irqfd);
418
419 if (resamplefd && !IS_ERR(resamplefd))
420 eventfd_ctx_put(resamplefd);
421
422 if (eventfd && !IS_ERR(eventfd))
423 eventfd_ctx_put(eventfd);
424
425 fdput(f);
426
427out:
428 kfree(irqfd);
429 return ret;
430}
431#endif
432
433void
434kvm_eventfd_init(struct kvm *kvm)
435{
436#ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
437 spin_lock_init(&kvm->irqfds.lock);
438 INIT_LIST_HEAD(&kvm->irqfds.items);
439 INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
440 mutex_init(&kvm->irqfds.resampler_lock);
441#endif
442 INIT_LIST_HEAD(&kvm->ioeventfds);
443}
444
445#ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
446/*
447 * shutdown any irqfd's that match fd+gsi
448 */
449static int
450kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
451{
452 struct _irqfd *irqfd, *tmp;
453 struct eventfd_ctx *eventfd;
454
455 eventfd = eventfd_ctx_fdget(args->fd);
456 if (IS_ERR(eventfd))
457 return PTR_ERR(eventfd);
458
459 spin_lock_irq(&kvm->irqfds.lock);
460
461 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
462 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
463 /*
464 * This rcu_assign_pointer is needed for when
465 * another thread calls kvm_irq_routing_update before
466 * we flush workqueue below (we synchronize with
467 * kvm_irq_routing_update using irqfds.lock).
468 * It is paired with synchronize_rcu done by caller
469 * of that function.
470 */
471 rcu_assign_pointer(irqfd->irq_entry, NULL);
472 irqfd_deactivate(irqfd);
473 }
474 }
475
476 spin_unlock_irq(&kvm->irqfds.lock);
477 eventfd_ctx_put(eventfd);
478
479 /*
480 * Block until we know all outstanding shutdown jobs have completed
481 * so that we guarantee there will not be any more interrupts on this
482 * gsi once this deassign function returns.
483 */
484 flush_workqueue(irqfd_cleanup_wq);
485
486 return 0;
487}
488
489int
490kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
491{
492 if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
493 return -EINVAL;
494
495 if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
496 return kvm_irqfd_deassign(kvm, args);
497
498 return kvm_irqfd_assign(kvm, args);
499}
500
501/*
502 * This function is called as the kvm VM fd is being released. Shutdown all
503 * irqfds that still remain open
504 */
505void
506kvm_irqfd_release(struct kvm *kvm)
507{
508 struct _irqfd *irqfd, *tmp;
509
510 spin_lock_irq(&kvm->irqfds.lock);
511
512 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
513 irqfd_deactivate(irqfd);
514
515 spin_unlock_irq(&kvm->irqfds.lock);
516
517 /*
518 * Block until we know all outstanding shutdown jobs have completed
519 * since we do not take a kvm* reference.
520 */
521 flush_workqueue(irqfd_cleanup_wq);
522
523}
524
525/*
526 * Change irq_routing and irqfd.
527 * Caller must invoke synchronize_rcu afterwards.
528 */
529void kvm_irq_routing_update(struct kvm *kvm,
530 struct kvm_irq_routing_table *irq_rt)
531{
532 struct _irqfd *irqfd;
533
534 spin_lock_irq(&kvm->irqfds.lock);
535
536 rcu_assign_pointer(kvm->irq_routing, irq_rt);
537
538 list_for_each_entry(irqfd, &kvm->irqfds.items, list)
539 irqfd_update(kvm, irqfd, irq_rt);
540
541 spin_unlock_irq(&kvm->irqfds.lock);
542}
543
544/*
545 * create a host-wide workqueue for issuing deferred shutdown requests
546 * aggregated from all vm* instances. We need our own isolated single-thread
547 * queue to prevent deadlock against flushing the normal work-queue.
548 */
549int kvm_irqfd_init(void)
550{
551 irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
552 if (!irqfd_cleanup_wq)
553 return -ENOMEM;
554
555 return 0;
556}
557
558void kvm_irqfd_exit(void)
559{
560 destroy_workqueue(irqfd_cleanup_wq);
561}
562#endif
563
564/*
565 * --------------------------------------------------------------------
566 * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
567 *
568 * userspace can register a PIO/MMIO address with an eventfd for receiving
569 * notification when the memory has been touched.
570 * --------------------------------------------------------------------
571 */
572
573struct _ioeventfd {
574 struct list_head list;
575 u64 addr;
576 int length;
577 struct eventfd_ctx *eventfd;
578 u64 datamatch;
579 struct kvm_io_device dev;
580 u8 bus_idx;
581 bool wildcard;
582};
583
584static inline struct _ioeventfd *
585to_ioeventfd(struct kvm_io_device *dev)
586{
587 return container_of(dev, struct _ioeventfd, dev);
588}
589
590static void
591ioeventfd_release(struct _ioeventfd *p)
592{
593 eventfd_ctx_put(p->eventfd);
594 list_del(&p->list);
595 kfree(p);
596}
597
598static bool
599ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
600{
601 u64 _val;
602
603 if (!(addr == p->addr && len == p->length))
604 /* address-range must be precise for a hit */
605 return false;
606
607 if (p->wildcard)
608 /* all else equal, wildcard is always a hit */
609 return true;
610
611 /* otherwise, we have to actually compare the data */
612
613 BUG_ON(!IS_ALIGNED((unsigned long)val, len));
614
615 switch (len) {
616 case 1:
617 _val = *(u8 *)val;
618 break;
619 case 2:
620 _val = *(u16 *)val;
621 break;
622 case 4:
623 _val = *(u32 *)val;
624 break;
625 case 8:
626 _val = *(u64 *)val;
627 break;
628 default:
629 return false;
630 }
631
632 return _val == p->datamatch ? true : false;
633}
634
635/* MMIO/PIO writes trigger an event if the addr/val match */
636static int
637ioeventfd_write(struct kvm_io_device *this, gpa_t addr, int len,
638 const void *val)
639{
640 struct _ioeventfd *p = to_ioeventfd(this);
641
642 if (!ioeventfd_in_range(p, addr, len, val))
643 return -EOPNOTSUPP;
644
645 eventfd_signal(p->eventfd, 1);
646 return 0;
647}
648
649/*
650 * This function is called as KVM is completely shutting down. We do not
651 * need to worry about locking just nuke anything we have as quickly as possible
652 */
653static void
654ioeventfd_destructor(struct kvm_io_device *this)
655{
656 struct _ioeventfd *p = to_ioeventfd(this);
657
658 ioeventfd_release(p);
659}
660
661static const struct kvm_io_device_ops ioeventfd_ops = {
662 .write = ioeventfd_write,
663 .destructor = ioeventfd_destructor,
664};
665
666/* assumes kvm->slots_lock held */
667static bool
668ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
669{
670 struct _ioeventfd *_p;
671
672 list_for_each_entry(_p, &kvm->ioeventfds, list)
673 if (_p->bus_idx == p->bus_idx &&
674 _p->addr == p->addr && _p->length == p->length &&
675 (_p->wildcard || p->wildcard ||
676 _p->datamatch == p->datamatch))
677 return true;
678
679 return false;
680}
681
682static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
683{
684 if (flags & KVM_IOEVENTFD_FLAG_PIO)
685 return KVM_PIO_BUS;
686 if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
687 return KVM_VIRTIO_CCW_NOTIFY_BUS;
688 return KVM_MMIO_BUS;
689}
690
691static int
692kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
693{
694 enum kvm_bus bus_idx;
695 struct _ioeventfd *p;
696 struct eventfd_ctx *eventfd;
697 int ret;
698
699 bus_idx = ioeventfd_bus_from_flags(args->flags);
700 /* must be natural-word sized */
701 switch (args->len) {
702 case 1:
703 case 2:
704 case 4:
705 case 8:
706 break;
707 default:
708 return -EINVAL;
709 }
710
711 /* check for range overflow */
712 if (args->addr + args->len < args->addr)
713 return -EINVAL;
714
715 /* check for extra flags that we don't understand */
716 if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
717 return -EINVAL;
718
719 eventfd = eventfd_ctx_fdget(args->fd);
720 if (IS_ERR(eventfd))
721 return PTR_ERR(eventfd);
722
723 p = kzalloc(sizeof(*p), GFP_KERNEL);
724 if (!p) {
725 ret = -ENOMEM;
726 goto fail;
727 }
728
729 INIT_LIST_HEAD(&p->list);
730 p->addr = args->addr;
731 p->bus_idx = bus_idx;
732 p->length = args->len;
733 p->eventfd = eventfd;
734
735 /* The datamatch feature is optional, otherwise this is a wildcard */
736 if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
737 p->datamatch = args->datamatch;
738 else
739 p->wildcard = true;
740
741 mutex_lock(&kvm->slots_lock);
742
743 /* Verify that there isn't a match already */
744 if (ioeventfd_check_collision(kvm, p)) {
745 ret = -EEXIST;
746 goto unlock_fail;
747 }
748
749 kvm_iodevice_init(&p->dev, &ioeventfd_ops);
750
751 ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
752 &p->dev);
753 if (ret < 0)
754 goto unlock_fail;
755
756 kvm->buses[bus_idx]->ioeventfd_count++;
757 list_add_tail(&p->list, &kvm->ioeventfds);
758
759 mutex_unlock(&kvm->slots_lock);
760
761 return 0;
762
763unlock_fail:
764 mutex_unlock(&kvm->slots_lock);
765
766fail:
767 kfree(p);
768 eventfd_ctx_put(eventfd);
769
770 return ret;
771}
772
773static int
774kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
775{
776 enum kvm_bus bus_idx;
777 struct _ioeventfd *p, *tmp;
778 struct eventfd_ctx *eventfd;
779 int ret = -ENOENT;
780
781 bus_idx = ioeventfd_bus_from_flags(args->flags);
782 eventfd = eventfd_ctx_fdget(args->fd);
783 if (IS_ERR(eventfd))
784 return PTR_ERR(eventfd);
785
786 mutex_lock(&kvm->slots_lock);
787
788 list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
789 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
790
791 if (p->bus_idx != bus_idx ||
792 p->eventfd != eventfd ||
793 p->addr != args->addr ||
794 p->length != args->len ||
795 p->wildcard != wildcard)
796 continue;
797
798 if (!p->wildcard && p->datamatch != args->datamatch)
799 continue;
800
801 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
802 kvm->buses[bus_idx]->ioeventfd_count--;
803 ioeventfd_release(p);
804 ret = 0;
805 break;
806 }
807
808 mutex_unlock(&kvm->slots_lock);
809
810 eventfd_ctx_put(eventfd);
811
812 return ret;
813}
814
815int
816kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
817{
818 if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
819 return kvm_deassign_ioeventfd(kvm, args);
820
821 return kvm_assign_ioeventfd(kvm, args);
822}