Loading...
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}
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/*
39 * --------------------------------------------------------------------
40 * irqfd: Allows an fd to be used to inject an interrupt to the guest
41 *
42 * Credit goes to Avi Kivity for the original idea.
43 * --------------------------------------------------------------------
44 */
45
46struct _irqfd {
47 /* Used for MSI fast-path */
48 struct kvm *kvm;
49 wait_queue_t wait;
50 /* Update side is protected by irqfds.lock */
51 struct kvm_kernel_irq_routing_entry __rcu *irq_entry;
52 /* Used for level IRQ fast-path */
53 int gsi;
54 struct work_struct inject;
55 /* Used for setup/shutdown */
56 struct eventfd_ctx *eventfd;
57 struct list_head list;
58 poll_table pt;
59 struct work_struct shutdown;
60};
61
62static struct workqueue_struct *irqfd_cleanup_wq;
63
64static void
65irqfd_inject(struct work_struct *work)
66{
67 struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
68 struct kvm *kvm = irqfd->kvm;
69
70 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1);
71 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0);
72}
73
74/*
75 * Race-free decouple logic (ordering is critical)
76 */
77static void
78irqfd_shutdown(struct work_struct *work)
79{
80 struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
81 u64 cnt;
82
83 /*
84 * Synchronize with the wait-queue and unhook ourselves to prevent
85 * further events.
86 */
87 eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
88
89 /*
90 * We know no new events will be scheduled at this point, so block
91 * until all previously outstanding events have completed
92 */
93 flush_work_sync(&irqfd->inject);
94
95 /*
96 * It is now safe to release the object's resources
97 */
98 eventfd_ctx_put(irqfd->eventfd);
99 kfree(irqfd);
100}
101
102
103/* assumes kvm->irqfds.lock is held */
104static bool
105irqfd_is_active(struct _irqfd *irqfd)
106{
107 return list_empty(&irqfd->list) ? false : true;
108}
109
110/*
111 * Mark the irqfd as inactive and schedule it for removal
112 *
113 * assumes kvm->irqfds.lock is held
114 */
115static void
116irqfd_deactivate(struct _irqfd *irqfd)
117{
118 BUG_ON(!irqfd_is_active(irqfd));
119
120 list_del_init(&irqfd->list);
121
122 queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
123}
124
125/*
126 * Called with wqh->lock held and interrupts disabled
127 */
128static int
129irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
130{
131 struct _irqfd *irqfd = container_of(wait, struct _irqfd, wait);
132 unsigned long flags = (unsigned long)key;
133 struct kvm_kernel_irq_routing_entry *irq;
134 struct kvm *kvm = irqfd->kvm;
135
136 if (flags & POLLIN) {
137 rcu_read_lock();
138 irq = rcu_dereference(irqfd->irq_entry);
139 /* An event has been signaled, inject an interrupt */
140 if (irq)
141 kvm_set_msi(irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1);
142 else
143 schedule_work(&irqfd->inject);
144 rcu_read_unlock();
145 }
146
147 if (flags & POLLHUP) {
148 /* The eventfd is closing, detach from KVM */
149 unsigned long flags;
150
151 spin_lock_irqsave(&kvm->irqfds.lock, flags);
152
153 /*
154 * We must check if someone deactivated the irqfd before
155 * we could acquire the irqfds.lock since the item is
156 * deactivated from the KVM side before it is unhooked from
157 * the wait-queue. If it is already deactivated, we can
158 * simply return knowing the other side will cleanup for us.
159 * We cannot race against the irqfd going away since the
160 * other side is required to acquire wqh->lock, which we hold
161 */
162 if (irqfd_is_active(irqfd))
163 irqfd_deactivate(irqfd);
164
165 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
166 }
167
168 return 0;
169}
170
171static void
172irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
173 poll_table *pt)
174{
175 struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
176 add_wait_queue(wqh, &irqfd->wait);
177}
178
179/* Must be called under irqfds.lock */
180static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd,
181 struct kvm_irq_routing_table *irq_rt)
182{
183 struct kvm_kernel_irq_routing_entry *e;
184 struct hlist_node *n;
185
186 if (irqfd->gsi >= irq_rt->nr_rt_entries) {
187 rcu_assign_pointer(irqfd->irq_entry, NULL);
188 return;
189 }
190
191 hlist_for_each_entry(e, n, &irq_rt->map[irqfd->gsi], link) {
192 /* Only fast-path MSI. */
193 if (e->type == KVM_IRQ_ROUTING_MSI)
194 rcu_assign_pointer(irqfd->irq_entry, e);
195 else
196 rcu_assign_pointer(irqfd->irq_entry, NULL);
197 }
198}
199
200static int
201kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
202{
203 struct kvm_irq_routing_table *irq_rt;
204 struct _irqfd *irqfd, *tmp;
205 struct file *file = NULL;
206 struct eventfd_ctx *eventfd = NULL;
207 int ret;
208 unsigned int events;
209
210 irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
211 if (!irqfd)
212 return -ENOMEM;
213
214 irqfd->kvm = kvm;
215 irqfd->gsi = args->gsi;
216 INIT_LIST_HEAD(&irqfd->list);
217 INIT_WORK(&irqfd->inject, irqfd_inject);
218 INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
219
220 file = eventfd_fget(args->fd);
221 if (IS_ERR(file)) {
222 ret = PTR_ERR(file);
223 goto fail;
224 }
225
226 eventfd = eventfd_ctx_fileget(file);
227 if (IS_ERR(eventfd)) {
228 ret = PTR_ERR(eventfd);
229 goto fail;
230 }
231
232 irqfd->eventfd = eventfd;
233
234 /*
235 * Install our own custom wake-up handling so we are notified via
236 * a callback whenever someone signals the underlying eventfd
237 */
238 init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
239 init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
240
241 spin_lock_irq(&kvm->irqfds.lock);
242
243 ret = 0;
244 list_for_each_entry(tmp, &kvm->irqfds.items, list) {
245 if (irqfd->eventfd != tmp->eventfd)
246 continue;
247 /* This fd is used for another irq already. */
248 ret = -EBUSY;
249 spin_unlock_irq(&kvm->irqfds.lock);
250 goto fail;
251 }
252
253 irq_rt = rcu_dereference_protected(kvm->irq_routing,
254 lockdep_is_held(&kvm->irqfds.lock));
255 irqfd_update(kvm, irqfd, irq_rt);
256
257 events = file->f_op->poll(file, &irqfd->pt);
258
259 list_add_tail(&irqfd->list, &kvm->irqfds.items);
260
261 /*
262 * Check if there was an event already pending on the eventfd
263 * before we registered, and trigger it as if we didn't miss it.
264 */
265 if (events & POLLIN)
266 schedule_work(&irqfd->inject);
267
268 spin_unlock_irq(&kvm->irqfds.lock);
269
270 /*
271 * do not drop the file until the irqfd is fully initialized, otherwise
272 * we might race against the POLLHUP
273 */
274 fput(file);
275
276 return 0;
277
278fail:
279 if (eventfd && !IS_ERR(eventfd))
280 eventfd_ctx_put(eventfd);
281
282 if (!IS_ERR(file))
283 fput(file);
284
285 kfree(irqfd);
286 return ret;
287}
288
289void
290kvm_eventfd_init(struct kvm *kvm)
291{
292 spin_lock_init(&kvm->irqfds.lock);
293 INIT_LIST_HEAD(&kvm->irqfds.items);
294 INIT_LIST_HEAD(&kvm->ioeventfds);
295}
296
297/*
298 * shutdown any irqfd's that match fd+gsi
299 */
300static int
301kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
302{
303 struct _irqfd *irqfd, *tmp;
304 struct eventfd_ctx *eventfd;
305
306 eventfd = eventfd_ctx_fdget(args->fd);
307 if (IS_ERR(eventfd))
308 return PTR_ERR(eventfd);
309
310 spin_lock_irq(&kvm->irqfds.lock);
311
312 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
313 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
314 /*
315 * This rcu_assign_pointer is needed for when
316 * another thread calls kvm_irq_routing_update before
317 * we flush workqueue below (we synchronize with
318 * kvm_irq_routing_update using irqfds.lock).
319 * It is paired with synchronize_rcu done by caller
320 * of that function.
321 */
322 rcu_assign_pointer(irqfd->irq_entry, NULL);
323 irqfd_deactivate(irqfd);
324 }
325 }
326
327 spin_unlock_irq(&kvm->irqfds.lock);
328 eventfd_ctx_put(eventfd);
329
330 /*
331 * Block until we know all outstanding shutdown jobs have completed
332 * so that we guarantee there will not be any more interrupts on this
333 * gsi once this deassign function returns.
334 */
335 flush_workqueue(irqfd_cleanup_wq);
336
337 return 0;
338}
339
340int
341kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
342{
343 if (args->flags & ~KVM_IRQFD_FLAG_DEASSIGN)
344 return -EINVAL;
345
346 if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
347 return kvm_irqfd_deassign(kvm, args);
348
349 return kvm_irqfd_assign(kvm, args);
350}
351
352/*
353 * This function is called as the kvm VM fd is being released. Shutdown all
354 * irqfds that still remain open
355 */
356void
357kvm_irqfd_release(struct kvm *kvm)
358{
359 struct _irqfd *irqfd, *tmp;
360
361 spin_lock_irq(&kvm->irqfds.lock);
362
363 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
364 irqfd_deactivate(irqfd);
365
366 spin_unlock_irq(&kvm->irqfds.lock);
367
368 /*
369 * Block until we know all outstanding shutdown jobs have completed
370 * since we do not take a kvm* reference.
371 */
372 flush_workqueue(irqfd_cleanup_wq);
373
374}
375
376/*
377 * Change irq_routing and irqfd.
378 * Caller must invoke synchronize_rcu afterwards.
379 */
380void kvm_irq_routing_update(struct kvm *kvm,
381 struct kvm_irq_routing_table *irq_rt)
382{
383 struct _irqfd *irqfd;
384
385 spin_lock_irq(&kvm->irqfds.lock);
386
387 rcu_assign_pointer(kvm->irq_routing, irq_rt);
388
389 list_for_each_entry(irqfd, &kvm->irqfds.items, list)
390 irqfd_update(kvm, irqfd, irq_rt);
391
392 spin_unlock_irq(&kvm->irqfds.lock);
393}
394
395/*
396 * create a host-wide workqueue for issuing deferred shutdown requests
397 * aggregated from all vm* instances. We need our own isolated single-thread
398 * queue to prevent deadlock against flushing the normal work-queue.
399 */
400static int __init irqfd_module_init(void)
401{
402 irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
403 if (!irqfd_cleanup_wq)
404 return -ENOMEM;
405
406 return 0;
407}
408
409static void __exit irqfd_module_exit(void)
410{
411 destroy_workqueue(irqfd_cleanup_wq);
412}
413
414module_init(irqfd_module_init);
415module_exit(irqfd_module_exit);
416
417/*
418 * --------------------------------------------------------------------
419 * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
420 *
421 * userspace can register a PIO/MMIO address with an eventfd for receiving
422 * notification when the memory has been touched.
423 * --------------------------------------------------------------------
424 */
425
426struct _ioeventfd {
427 struct list_head list;
428 u64 addr;
429 int length;
430 struct eventfd_ctx *eventfd;
431 u64 datamatch;
432 struct kvm_io_device dev;
433 bool wildcard;
434};
435
436static inline struct _ioeventfd *
437to_ioeventfd(struct kvm_io_device *dev)
438{
439 return container_of(dev, struct _ioeventfd, dev);
440}
441
442static void
443ioeventfd_release(struct _ioeventfd *p)
444{
445 eventfd_ctx_put(p->eventfd);
446 list_del(&p->list);
447 kfree(p);
448}
449
450static bool
451ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
452{
453 u64 _val;
454
455 if (!(addr == p->addr && len == p->length))
456 /* address-range must be precise for a hit */
457 return false;
458
459 if (p->wildcard)
460 /* all else equal, wildcard is always a hit */
461 return true;
462
463 /* otherwise, we have to actually compare the data */
464
465 BUG_ON(!IS_ALIGNED((unsigned long)val, len));
466
467 switch (len) {
468 case 1:
469 _val = *(u8 *)val;
470 break;
471 case 2:
472 _val = *(u16 *)val;
473 break;
474 case 4:
475 _val = *(u32 *)val;
476 break;
477 case 8:
478 _val = *(u64 *)val;
479 break;
480 default:
481 return false;
482 }
483
484 return _val == p->datamatch ? true : false;
485}
486
487/* MMIO/PIO writes trigger an event if the addr/val match */
488static int
489ioeventfd_write(struct kvm_io_device *this, gpa_t addr, int len,
490 const void *val)
491{
492 struct _ioeventfd *p = to_ioeventfd(this);
493
494 if (!ioeventfd_in_range(p, addr, len, val))
495 return -EOPNOTSUPP;
496
497 eventfd_signal(p->eventfd, 1);
498 return 0;
499}
500
501/*
502 * This function is called as KVM is completely shutting down. We do not
503 * need to worry about locking just nuke anything we have as quickly as possible
504 */
505static void
506ioeventfd_destructor(struct kvm_io_device *this)
507{
508 struct _ioeventfd *p = to_ioeventfd(this);
509
510 ioeventfd_release(p);
511}
512
513static const struct kvm_io_device_ops ioeventfd_ops = {
514 .write = ioeventfd_write,
515 .destructor = ioeventfd_destructor,
516};
517
518/* assumes kvm->slots_lock held */
519static bool
520ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
521{
522 struct _ioeventfd *_p;
523
524 list_for_each_entry(_p, &kvm->ioeventfds, list)
525 if (_p->addr == p->addr && _p->length == p->length &&
526 (_p->wildcard || p->wildcard ||
527 _p->datamatch == p->datamatch))
528 return true;
529
530 return false;
531}
532
533static int
534kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
535{
536 int pio = args->flags & KVM_IOEVENTFD_FLAG_PIO;
537 enum kvm_bus bus_idx = pio ? KVM_PIO_BUS : KVM_MMIO_BUS;
538 struct _ioeventfd *p;
539 struct eventfd_ctx *eventfd;
540 int ret;
541
542 /* must be natural-word sized */
543 switch (args->len) {
544 case 1:
545 case 2:
546 case 4:
547 case 8:
548 break;
549 default:
550 return -EINVAL;
551 }
552
553 /* check for range overflow */
554 if (args->addr + args->len < args->addr)
555 return -EINVAL;
556
557 /* check for extra flags that we don't understand */
558 if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
559 return -EINVAL;
560
561 eventfd = eventfd_ctx_fdget(args->fd);
562 if (IS_ERR(eventfd))
563 return PTR_ERR(eventfd);
564
565 p = kzalloc(sizeof(*p), GFP_KERNEL);
566 if (!p) {
567 ret = -ENOMEM;
568 goto fail;
569 }
570
571 INIT_LIST_HEAD(&p->list);
572 p->addr = args->addr;
573 p->length = args->len;
574 p->eventfd = eventfd;
575
576 /* The datamatch feature is optional, otherwise this is a wildcard */
577 if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
578 p->datamatch = args->datamatch;
579 else
580 p->wildcard = true;
581
582 mutex_lock(&kvm->slots_lock);
583
584 /* Verify that there isn't a match already */
585 if (ioeventfd_check_collision(kvm, p)) {
586 ret = -EEXIST;
587 goto unlock_fail;
588 }
589
590 kvm_iodevice_init(&p->dev, &ioeventfd_ops);
591
592 ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
593 &p->dev);
594 if (ret < 0)
595 goto unlock_fail;
596
597 list_add_tail(&p->list, &kvm->ioeventfds);
598
599 mutex_unlock(&kvm->slots_lock);
600
601 return 0;
602
603unlock_fail:
604 mutex_unlock(&kvm->slots_lock);
605
606fail:
607 kfree(p);
608 eventfd_ctx_put(eventfd);
609
610 return ret;
611}
612
613static int
614kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
615{
616 int pio = args->flags & KVM_IOEVENTFD_FLAG_PIO;
617 enum kvm_bus bus_idx = pio ? KVM_PIO_BUS : KVM_MMIO_BUS;
618 struct _ioeventfd *p, *tmp;
619 struct eventfd_ctx *eventfd;
620 int ret = -ENOENT;
621
622 eventfd = eventfd_ctx_fdget(args->fd);
623 if (IS_ERR(eventfd))
624 return PTR_ERR(eventfd);
625
626 mutex_lock(&kvm->slots_lock);
627
628 list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
629 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
630
631 if (p->eventfd != eventfd ||
632 p->addr != args->addr ||
633 p->length != args->len ||
634 p->wildcard != wildcard)
635 continue;
636
637 if (!p->wildcard && p->datamatch != args->datamatch)
638 continue;
639
640 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
641 ioeventfd_release(p);
642 ret = 0;
643 break;
644 }
645
646 mutex_unlock(&kvm->slots_lock);
647
648 eventfd_ctx_put(eventfd);
649
650 return ret;
651}
652
653int
654kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
655{
656 if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
657 return kvm_deassign_ioeventfd(kvm, args);
658
659 return kvm_assign_ioeventfd(kvm, args);
660}