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