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