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