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