KVM: Distangle eventfd code from irqchip
[linux-3.10.git] / virt / kvm / eventfd.c
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 __KVM_HAVE_IOAPIC
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  */
56 struct _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
72 struct _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
94 static struct workqueue_struct *irqfd_cleanup_wq;
95
96 static void
97 irqfd_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                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0);
105         } else
106                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
107                             irqfd->gsi, 1);
108 }
109
110 /*
111  * Since resampler irqfds share an IRQ source ID, we de-assert once
112  * then notify all of the resampler irqfds using this GSI.  We can't
113  * do multiple de-asserts or we risk racing with incoming re-asserts.
114  */
115 static void
116 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
117 {
118         struct _irqfd_resampler *resampler;
119         struct _irqfd *irqfd;
120
121         resampler = container_of(kian, struct _irqfd_resampler, notifier);
122
123         kvm_set_irq(resampler->kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
124                     resampler->notifier.gsi, 0);
125
126         rcu_read_lock();
127
128         list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
129                 eventfd_signal(irqfd->resamplefd, 1);
130
131         rcu_read_unlock();
132 }
133
134 static void
135 irqfd_resampler_shutdown(struct _irqfd *irqfd)
136 {
137         struct _irqfd_resampler *resampler = irqfd->resampler;
138         struct kvm *kvm = resampler->kvm;
139
140         mutex_lock(&kvm->irqfds.resampler_lock);
141
142         list_del_rcu(&irqfd->resampler_link);
143         synchronize_rcu();
144
145         if (list_empty(&resampler->list)) {
146                 list_del(&resampler->link);
147                 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
148                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
149                             resampler->notifier.gsi, 0);
150                 kfree(resampler);
151         }
152
153         mutex_unlock(&kvm->irqfds.resampler_lock);
154 }
155
156 /*
157  * Race-free decouple logic (ordering is critical)
158  */
159 static void
160 irqfd_shutdown(struct work_struct *work)
161 {
162         struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
163         u64 cnt;
164
165         /*
166          * Synchronize with the wait-queue and unhook ourselves to prevent
167          * further events.
168          */
169         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
170
171         /*
172          * We know no new events will be scheduled at this point, so block
173          * until all previously outstanding events have completed
174          */
175         flush_work(&irqfd->inject);
176
177         if (irqfd->resampler) {
178                 irqfd_resampler_shutdown(irqfd);
179                 eventfd_ctx_put(irqfd->resamplefd);
180         }
181
182         /*
183          * It is now safe to release the object's resources
184          */
185         eventfd_ctx_put(irqfd->eventfd);
186         kfree(irqfd);
187 }
188
189
190 /* assumes kvm->irqfds.lock is held */
191 static bool
192 irqfd_is_active(struct _irqfd *irqfd)
193 {
194         return list_empty(&irqfd->list) ? false : true;
195 }
196
197 /*
198  * Mark the irqfd as inactive and schedule it for removal
199  *
200  * assumes kvm->irqfds.lock is held
201  */
202 static void
203 irqfd_deactivate(struct _irqfd *irqfd)
204 {
205         BUG_ON(!irqfd_is_active(irqfd));
206
207         list_del_init(&irqfd->list);
208
209         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
210 }
211
212 /*
213  * Called with wqh->lock held and interrupts disabled
214  */
215 static int
216 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
217 {
218         struct _irqfd *irqfd = container_of(wait, struct _irqfd, wait);
219         unsigned long flags = (unsigned long)key;
220         struct kvm_kernel_irq_routing_entry *irq;
221         struct kvm *kvm = irqfd->kvm;
222
223         if (flags & POLLIN) {
224                 rcu_read_lock();
225                 irq = rcu_dereference(irqfd->irq_entry);
226                 /* An event has been signaled, inject an interrupt */
227                 if (irq)
228                         kvm_set_msi(irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1);
229                 else
230                         schedule_work(&irqfd->inject);
231                 rcu_read_unlock();
232         }
233
234         if (flags & POLLHUP) {
235                 /* The eventfd is closing, detach from KVM */
236                 unsigned long flags;
237
238                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
239
240                 /*
241                  * We must check if someone deactivated the irqfd before
242                  * we could acquire the irqfds.lock since the item is
243                  * deactivated from the KVM side before it is unhooked from
244                  * the wait-queue.  If it is already deactivated, we can
245                  * simply return knowing the other side will cleanup for us.
246                  * We cannot race against the irqfd going away since the
247                  * other side is required to acquire wqh->lock, which we hold
248                  */
249                 if (irqfd_is_active(irqfd))
250                         irqfd_deactivate(irqfd);
251
252                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
253         }
254
255         return 0;
256 }
257
258 static void
259 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
260                         poll_table *pt)
261 {
262         struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
263         add_wait_queue(wqh, &irqfd->wait);
264 }
265
266 /* Must be called under irqfds.lock */
267 static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd,
268                          struct kvm_irq_routing_table *irq_rt)
269 {
270         struct kvm_kernel_irq_routing_entry *e;
271         struct hlist_node *n;
272
273         if (irqfd->gsi >= irq_rt->nr_rt_entries) {
274                 rcu_assign_pointer(irqfd->irq_entry, NULL);
275                 return;
276         }
277
278         hlist_for_each_entry(e, n, &irq_rt->map[irqfd->gsi], link) {
279                 /* Only fast-path MSI. */
280                 if (e->type == KVM_IRQ_ROUTING_MSI)
281                         rcu_assign_pointer(irqfd->irq_entry, e);
282                 else
283                         rcu_assign_pointer(irqfd->irq_entry, NULL);
284         }
285 }
286
287 static int
288 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
289 {
290         struct kvm_irq_routing_table *irq_rt;
291         struct _irqfd *irqfd, *tmp;
292         struct file *file = NULL;
293         struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
294         int ret;
295         unsigned int events;
296
297         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
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
307         file = eventfd_fget(args->fd);
308         if (IS_ERR(file)) {
309                 ret = PTR_ERR(file);
310                 goto fail;
311         }
312
313         eventfd = eventfd_ctx_fileget(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 _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, list) {
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_rcu();
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         irq_rt = rcu_dereference_protected(kvm->irq_routing,
389                                            lockdep_is_held(&kvm->irqfds.lock));
390         irqfd_update(kvm, irqfd, irq_rt);
391
392         events = file->f_op->poll(file, &irqfd->pt);
393
394         list_add_tail(&irqfd->list, &kvm->irqfds.items);
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         if (events & POLLIN)
401                 schedule_work(&irqfd->inject);
402
403         spin_unlock_irq(&kvm->irqfds.lock);
404
405         /*
406          * do not drop the file until the irqfd is fully initialized, otherwise
407          * we might race against the POLLHUP
408          */
409         fput(file);
410
411         return 0;
412
413 fail:
414         if (irqfd->resampler)
415                 irqfd_resampler_shutdown(irqfd);
416
417         if (resamplefd && !IS_ERR(resamplefd))
418                 eventfd_ctx_put(resamplefd);
419
420         if (eventfd && !IS_ERR(eventfd))
421                 eventfd_ctx_put(eventfd);
422
423         if (!IS_ERR(file))
424                 fput(file);
425
426         kfree(irqfd);
427         return ret;
428 }
429 #endif
430
431 void
432 kvm_eventfd_init(struct kvm *kvm)
433 {
434 #ifdef __KVM_HAVE_IOAPIC
435         spin_lock_init(&kvm->irqfds.lock);
436         INIT_LIST_HEAD(&kvm->irqfds.items);
437         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
438         mutex_init(&kvm->irqfds.resampler_lock);
439 #endif
440         INIT_LIST_HEAD(&kvm->ioeventfds);
441 }
442
443 #ifdef __KVM_HAVE_IOAPIC
444 /*
445  * shutdown any irqfd's that match fd+gsi
446  */
447 static int
448 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
449 {
450         struct _irqfd *irqfd, *tmp;
451         struct eventfd_ctx *eventfd;
452
453         eventfd = eventfd_ctx_fdget(args->fd);
454         if (IS_ERR(eventfd))
455                 return PTR_ERR(eventfd);
456
457         spin_lock_irq(&kvm->irqfds.lock);
458
459         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
460                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
461                         /*
462                          * This rcu_assign_pointer is needed for when
463                          * another thread calls kvm_irq_routing_update before
464                          * we flush workqueue below (we synchronize with
465                          * kvm_irq_routing_update using irqfds.lock).
466                          * It is paired with synchronize_rcu done by caller
467                          * of that function.
468                          */
469                         rcu_assign_pointer(irqfd->irq_entry, NULL);
470                         irqfd_deactivate(irqfd);
471                 }
472         }
473
474         spin_unlock_irq(&kvm->irqfds.lock);
475         eventfd_ctx_put(eventfd);
476
477         /*
478          * Block until we know all outstanding shutdown jobs have completed
479          * so that we guarantee there will not be any more interrupts on this
480          * gsi once this deassign function returns.
481          */
482         flush_workqueue(irqfd_cleanup_wq);
483
484         return 0;
485 }
486
487 int
488 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
489 {
490         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
491                 return -EINVAL;
492
493         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
494                 return kvm_irqfd_deassign(kvm, args);
495
496         return kvm_irqfd_assign(kvm, args);
497 }
498
499 /*
500  * This function is called as the kvm VM fd is being released. Shutdown all
501  * irqfds that still remain open
502  */
503 void
504 kvm_irqfd_release(struct kvm *kvm)
505 {
506         struct _irqfd *irqfd, *tmp;
507
508         spin_lock_irq(&kvm->irqfds.lock);
509
510         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
511                 irqfd_deactivate(irqfd);
512
513         spin_unlock_irq(&kvm->irqfds.lock);
514
515         /*
516          * Block until we know all outstanding shutdown jobs have completed
517          * since we do not take a kvm* reference.
518          */
519         flush_workqueue(irqfd_cleanup_wq);
520
521 }
522
523 /*
524  * Change irq_routing and irqfd.
525  * Caller must invoke synchronize_rcu afterwards.
526  */
527 void kvm_irq_routing_update(struct kvm *kvm,
528                             struct kvm_irq_routing_table *irq_rt)
529 {
530         struct _irqfd *irqfd;
531
532         spin_lock_irq(&kvm->irqfds.lock);
533
534         rcu_assign_pointer(kvm->irq_routing, irq_rt);
535
536         list_for_each_entry(irqfd, &kvm->irqfds.items, list)
537                 irqfd_update(kvm, irqfd, irq_rt);
538
539         spin_unlock_irq(&kvm->irqfds.lock);
540 }
541
542 /*
543  * create a host-wide workqueue for issuing deferred shutdown requests
544  * aggregated from all vm* instances. We need our own isolated single-thread
545  * queue to prevent deadlock against flushing the normal work-queue.
546  */
547 static int __init irqfd_module_init(void)
548 {
549         irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
550         if (!irqfd_cleanup_wq)
551                 return -ENOMEM;
552
553         return 0;
554 }
555
556 static void __exit irqfd_module_exit(void)
557 {
558         destroy_workqueue(irqfd_cleanup_wq);
559 }
560
561 module_init(irqfd_module_init);
562 module_exit(irqfd_module_exit);
563 #endif
564
565 /*
566  * --------------------------------------------------------------------
567  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
568  *
569  * userspace can register a PIO/MMIO address with an eventfd for receiving
570  * notification when the memory has been touched.
571  * --------------------------------------------------------------------
572  */
573
574 struct _ioeventfd {
575         struct list_head     list;
576         u64                  addr;
577         int                  length;
578         struct eventfd_ctx  *eventfd;
579         u64                  datamatch;
580         struct kvm_io_device dev;
581         bool                 wildcard;
582 };
583
584 static inline struct _ioeventfd *
585 to_ioeventfd(struct kvm_io_device *dev)
586 {
587         return container_of(dev, struct _ioeventfd, dev);
588 }
589
590 static void
591 ioeventfd_release(struct _ioeventfd *p)
592 {
593         eventfd_ctx_put(p->eventfd);
594         list_del(&p->list);
595         kfree(p);
596 }
597
598 static bool
599 ioeventfd_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 */
636 static int
637 ioeventfd_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  */
653 static void
654 ioeventfd_destructor(struct kvm_io_device *this)
655 {
656         struct _ioeventfd *p = to_ioeventfd(this);
657
658         ioeventfd_release(p);
659 }
660
661 static const struct kvm_io_device_ops ioeventfd_ops = {
662         .write      = ioeventfd_write,
663         .destructor = ioeventfd_destructor,
664 };
665
666 /* assumes kvm->slots_lock held */
667 static bool
668 ioeventfd_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->addr == p->addr && _p->length == p->length &&
674                     (_p->wildcard || p->wildcard ||
675                      _p->datamatch == p->datamatch))
676                         return true;
677
678         return false;
679 }
680
681 static int
682 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
683 {
684         int                       pio = args->flags & KVM_IOEVENTFD_FLAG_PIO;
685         enum kvm_bus              bus_idx = pio ? KVM_PIO_BUS : KVM_MMIO_BUS;
686         struct _ioeventfd        *p;
687         struct eventfd_ctx       *eventfd;
688         int                       ret;
689
690         /* must be natural-word sized */
691         switch (args->len) {
692         case 1:
693         case 2:
694         case 4:
695         case 8:
696                 break;
697         default:
698                 return -EINVAL;
699         }
700
701         /* check for range overflow */
702         if (args->addr + args->len < args->addr)
703                 return -EINVAL;
704
705         /* check for extra flags that we don't understand */
706         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
707                 return -EINVAL;
708
709         eventfd = eventfd_ctx_fdget(args->fd);
710         if (IS_ERR(eventfd))
711                 return PTR_ERR(eventfd);
712
713         p = kzalloc(sizeof(*p), GFP_KERNEL);
714         if (!p) {
715                 ret = -ENOMEM;
716                 goto fail;
717         }
718
719         INIT_LIST_HEAD(&p->list);
720         p->addr    = args->addr;
721         p->length  = args->len;
722         p->eventfd = eventfd;
723
724         /* The datamatch feature is optional, otherwise this is a wildcard */
725         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
726                 p->datamatch = args->datamatch;
727         else
728                 p->wildcard = true;
729
730         mutex_lock(&kvm->slots_lock);
731
732         /* Verify that there isn't a match already */
733         if (ioeventfd_check_collision(kvm, p)) {
734                 ret = -EEXIST;
735                 goto unlock_fail;
736         }
737
738         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
739
740         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
741                                       &p->dev);
742         if (ret < 0)
743                 goto unlock_fail;
744
745         list_add_tail(&p->list, &kvm->ioeventfds);
746
747         mutex_unlock(&kvm->slots_lock);
748
749         return 0;
750
751 unlock_fail:
752         mutex_unlock(&kvm->slots_lock);
753
754 fail:
755         kfree(p);
756         eventfd_ctx_put(eventfd);
757
758         return ret;
759 }
760
761 static int
762 kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
763 {
764         int                       pio = args->flags & KVM_IOEVENTFD_FLAG_PIO;
765         enum kvm_bus              bus_idx = pio ? KVM_PIO_BUS : KVM_MMIO_BUS;
766         struct _ioeventfd        *p, *tmp;
767         struct eventfd_ctx       *eventfd;
768         int                       ret = -ENOENT;
769
770         eventfd = eventfd_ctx_fdget(args->fd);
771         if (IS_ERR(eventfd))
772                 return PTR_ERR(eventfd);
773
774         mutex_lock(&kvm->slots_lock);
775
776         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
777                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
778
779                 if (p->eventfd != eventfd  ||
780                     p->addr != args->addr  ||
781                     p->length != args->len ||
782                     p->wildcard != wildcard)
783                         continue;
784
785                 if (!p->wildcard && p->datamatch != args->datamatch)
786                         continue;
787
788                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
789                 ioeventfd_release(p);
790                 ret = 0;
791                 break;
792         }
793
794         mutex_unlock(&kvm->slots_lock);
795
796         eventfd_ctx_put(eventfd);
797
798         return ret;
799 }
800
801 int
802 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
803 {
804         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
805                 return kvm_deassign_ioeventfd(kvm, args);
806
807         return kvm_assign_ioeventfd(kvm, args);
808 }