2 * kvm eventfd support - use eventfd objects to signal various KVM events
4 * Copyright 2009 Novell. All Rights Reserved.
5 * Copyright 2010 Red Hat, Inc. and/or its affiliates.
8 * Gregory Haskins <ghaskins@novell.com>
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.
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.
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.
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>
38 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
40 * --------------------------------------------------------------------
41 * irqfd: Allows an fd to be used to inject an interrupt to the guest
43 * Credit goes to Avi Kivity for the original idea.
44 * --------------------------------------------------------------------
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.
56 struct _irqfd_resampler {
59 * List of resampling struct _irqfd objects sharing this gsi.
60 * RCU list modified under kvm->irqfds.resampler_lock
62 struct list_head list;
63 struct kvm_irq_ack_notifier notifier;
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
69 struct list_head link;
73 /* Used for MSI fast-path */
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 */
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;
91 struct work_struct shutdown;
94 static struct workqueue_struct *irqfd_cleanup_wq;
97 irqfd_inject(struct work_struct *work)
99 struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
100 struct kvm *kvm = irqfd->kvm;
102 if (!irqfd->resampler) {
103 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
105 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
108 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
109 irqfd->gsi, 1, false);
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.
118 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
120 struct _irqfd_resampler *resampler;
121 struct _irqfd *irqfd;
123 resampler = container_of(kian, struct _irqfd_resampler, notifier);
125 kvm_set_irq(resampler->kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
126 resampler->notifier.gsi, 0, false);
130 list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
131 eventfd_signal(irqfd->resamplefd, 1);
137 irqfd_resampler_shutdown(struct _irqfd *irqfd)
139 struct _irqfd_resampler *resampler = irqfd->resampler;
140 struct kvm *kvm = resampler->kvm;
142 mutex_lock(&kvm->irqfds.resampler_lock);
144 list_del_rcu(&irqfd->resampler_link);
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);
155 mutex_unlock(&kvm->irqfds.resampler_lock);
159 * Race-free decouple logic (ordering is critical)
162 irqfd_shutdown(struct work_struct *work)
164 struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
168 * Synchronize with the wait-queue and unhook ourselves to prevent
171 eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
174 * We know no new events will be scheduled at this point, so block
175 * until all previously outstanding events have completed
177 flush_work(&irqfd->inject);
179 if (irqfd->resampler) {
180 irqfd_resampler_shutdown(irqfd);
181 eventfd_ctx_put(irqfd->resamplefd);
185 * It is now safe to release the object's resources
187 eventfd_ctx_put(irqfd->eventfd);
192 /* assumes kvm->irqfds.lock is held */
194 irqfd_is_active(struct _irqfd *irqfd)
196 return list_empty(&irqfd->list) ? false : true;
200 * Mark the irqfd as inactive and schedule it for removal
202 * assumes kvm->irqfds.lock is held
205 irqfd_deactivate(struct _irqfd *irqfd)
207 BUG_ON(!irqfd_is_active(irqfd));
209 list_del_init(&irqfd->list);
211 queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
215 * Called with wqh->lock held and interrupts disabled
218 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
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;
225 if (flags & POLLIN) {
227 irq = rcu_dereference(irqfd->irq_entry);
228 /* An event has been signaled, inject an interrupt */
230 kvm_set_msi(irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1,
233 schedule_work(&irqfd->inject);
237 if (flags & POLLHUP) {
238 /* The eventfd is closing, detach from KVM */
241 spin_lock_irqsave(&kvm->irqfds.lock, flags);
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
252 if (irqfd_is_active(irqfd))
253 irqfd_deactivate(irqfd);
255 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
262 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
265 struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
266 add_wait_queue(wqh, &irqfd->wait);
269 /* Must be called under irqfds.lock */
270 static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd,
271 struct kvm_irq_routing_table *irq_rt)
273 struct kvm_kernel_irq_routing_entry *e;
275 if (irqfd->gsi >= irq_rt->nr_rt_entries) {
276 rcu_assign_pointer(irqfd->irq_entry, NULL);
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);
285 rcu_assign_pointer(irqfd->irq_entry, NULL);
290 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
292 struct kvm_irq_routing_table *irq_rt;
293 struct _irqfd *irqfd, *tmp;
295 struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
299 irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
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);
315 eventfd = eventfd_ctx_fileget(f.file);
316 if (IS_ERR(eventfd)) {
317 ret = PTR_ERR(eventfd);
321 irqfd->eventfd = eventfd;
323 if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
324 struct _irqfd_resampler *resampler;
326 resamplefd = eventfd_ctx_fdget(args->resamplefd);
327 if (IS_ERR(resamplefd)) {
328 ret = PTR_ERR(resamplefd);
332 irqfd->resamplefd = resamplefd;
333 INIT_LIST_HEAD(&irqfd->resampler_link);
335 mutex_lock(&kvm->irqfds.resampler_lock);
337 list_for_each_entry(resampler,
338 &kvm->irqfds.resampler_list, link) {
339 if (resampler->notifier.gsi == irqfd->gsi) {
340 irqfd->resampler = resampler;
345 if (!irqfd->resampler) {
346 resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
349 mutex_unlock(&kvm->irqfds.resampler_lock);
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);
359 list_add(&resampler->link, &kvm->irqfds.resampler_list);
360 kvm_register_irq_ack_notifier(kvm,
361 &resampler->notifier);
362 irqfd->resampler = resampler;
365 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
368 mutex_unlock(&kvm->irqfds.resampler_lock);
372 * Install our own custom wake-up handling so we are notified via
373 * a callback whenever someone signals the underlying eventfd
375 init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
376 init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
378 spin_lock_irq(&kvm->irqfds.lock);
381 list_for_each_entry(tmp, &kvm->irqfds.items, list) {
382 if (irqfd->eventfd != tmp->eventfd)
384 /* This fd is used for another irq already. */
386 spin_unlock_irq(&kvm->irqfds.lock);
390 irq_rt = rcu_dereference_protected(kvm->irq_routing,
391 lockdep_is_held(&kvm->irqfds.lock));
392 irqfd_update(kvm, irqfd, irq_rt);
394 list_add_tail(&irqfd->list, &kvm->irqfds.items);
396 spin_unlock_irq(&kvm->irqfds.lock);
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.
402 events = f.file->f_op->poll(f.file, &irqfd->pt);
405 schedule_work(&irqfd->inject);
408 * do not drop the file until the irqfd is fully initialized, otherwise
409 * we might race against the POLLHUP
416 if (irqfd->resampler)
417 irqfd_resampler_shutdown(irqfd);
419 if (resamplefd && !IS_ERR(resamplefd))
420 eventfd_ctx_put(resamplefd);
422 if (eventfd && !IS_ERR(eventfd))
423 eventfd_ctx_put(eventfd);
434 kvm_eventfd_init(struct kvm *kvm)
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);
442 INIT_LIST_HEAD(&kvm->ioeventfds);
445 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
447 * shutdown any irqfd's that match fd+gsi
450 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
452 struct _irqfd *irqfd, *tmp;
453 struct eventfd_ctx *eventfd;
455 eventfd = eventfd_ctx_fdget(args->fd);
457 return PTR_ERR(eventfd);
459 spin_lock_irq(&kvm->irqfds.lock);
461 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
462 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
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
471 rcu_assign_pointer(irqfd->irq_entry, NULL);
472 irqfd_deactivate(irqfd);
476 spin_unlock_irq(&kvm->irqfds.lock);
477 eventfd_ctx_put(eventfd);
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.
484 flush_workqueue(irqfd_cleanup_wq);
490 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
492 if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
495 if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
496 return kvm_irqfd_deassign(kvm, args);
498 return kvm_irqfd_assign(kvm, args);
502 * This function is called as the kvm VM fd is being released. Shutdown all
503 * irqfds that still remain open
506 kvm_irqfd_release(struct kvm *kvm)
508 struct _irqfd *irqfd, *tmp;
510 spin_lock_irq(&kvm->irqfds.lock);
512 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
513 irqfd_deactivate(irqfd);
515 spin_unlock_irq(&kvm->irqfds.lock);
518 * Block until we know all outstanding shutdown jobs have completed
519 * since we do not take a kvm* reference.
521 flush_workqueue(irqfd_cleanup_wq);
526 * Change irq_routing and irqfd.
527 * Caller must invoke synchronize_rcu afterwards.
529 void kvm_irq_routing_update(struct kvm *kvm,
530 struct kvm_irq_routing_table *irq_rt)
532 struct _irqfd *irqfd;
534 spin_lock_irq(&kvm->irqfds.lock);
536 rcu_assign_pointer(kvm->irq_routing, irq_rt);
538 list_for_each_entry(irqfd, &kvm->irqfds.items, list)
539 irqfd_update(kvm, irqfd, irq_rt);
541 spin_unlock_irq(&kvm->irqfds.lock);
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.
549 int kvm_irqfd_init(void)
551 irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
552 if (!irqfd_cleanup_wq)
558 void kvm_irqfd_exit(void)
560 destroy_workqueue(irqfd_cleanup_wq);
565 * --------------------------------------------------------------------
566 * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
568 * userspace can register a PIO/MMIO address with an eventfd for receiving
569 * notification when the memory has been touched.
570 * --------------------------------------------------------------------
574 struct list_head list;
577 struct eventfd_ctx *eventfd;
579 struct kvm_io_device dev;
584 static inline struct _ioeventfd *
585 to_ioeventfd(struct kvm_io_device *dev)
587 return container_of(dev, struct _ioeventfd, dev);
591 ioeventfd_release(struct _ioeventfd *p)
593 eventfd_ctx_put(p->eventfd);
599 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
604 /* address must be precise for a hit */
608 /* length = 0 means only look at the address, so always a hit */
611 if (len != p->length)
612 /* address-range must be precise for a hit */
616 /* all else equal, wildcard is always a hit */
619 /* otherwise, we have to actually compare the data */
621 BUG_ON(!IS_ALIGNED((unsigned long)val, len));
640 return _val == p->datamatch ? true : false;
643 /* MMIO/PIO writes trigger an event if the addr/val match */
645 ioeventfd_write(struct kvm_io_device *this, gpa_t addr, int len,
648 struct _ioeventfd *p = to_ioeventfd(this);
650 if (!ioeventfd_in_range(p, addr, len, val))
653 eventfd_signal(p->eventfd, 1);
658 * This function is called as KVM is completely shutting down. We do not
659 * need to worry about locking just nuke anything we have as quickly as possible
662 ioeventfd_destructor(struct kvm_io_device *this)
664 struct _ioeventfd *p = to_ioeventfd(this);
666 ioeventfd_release(p);
669 static const struct kvm_io_device_ops ioeventfd_ops = {
670 .write = ioeventfd_write,
671 .destructor = ioeventfd_destructor,
674 /* assumes kvm->slots_lock held */
676 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
678 struct _ioeventfd *_p;
680 list_for_each_entry(_p, &kvm->ioeventfds, list)
681 if (_p->bus_idx == p->bus_idx &&
682 _p->addr == p->addr &&
683 (!_p->length || !p->length ||
684 (_p->length == p->length &&
685 (_p->wildcard || p->wildcard ||
686 _p->datamatch == p->datamatch))))
692 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
694 if (flags & KVM_IOEVENTFD_FLAG_PIO)
696 if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
697 return KVM_VIRTIO_CCW_NOTIFY_BUS;
702 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
704 enum kvm_bus bus_idx;
705 struct _ioeventfd *p;
706 struct eventfd_ctx *eventfd;
709 bus_idx = ioeventfd_bus_from_flags(args->flags);
710 /* must be natural-word sized, or 0 to ignore length */
722 /* check for range overflow */
723 if (args->addr + args->len < args->addr)
726 /* check for extra flags that we don't understand */
727 if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
730 /* ioeventfd with no length can't be combined with DATAMATCH */
732 args->flags & (KVM_IOEVENTFD_FLAG_PIO |
733 KVM_IOEVENTFD_FLAG_DATAMATCH))
736 eventfd = eventfd_ctx_fdget(args->fd);
738 return PTR_ERR(eventfd);
740 p = kzalloc(sizeof(*p), GFP_KERNEL);
746 INIT_LIST_HEAD(&p->list);
747 p->addr = args->addr;
748 p->bus_idx = bus_idx;
749 p->length = args->len;
750 p->eventfd = eventfd;
752 /* The datamatch feature is optional, otherwise this is a wildcard */
753 if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
754 p->datamatch = args->datamatch;
758 mutex_lock(&kvm->slots_lock);
760 /* Verify that there isn't a match already */
761 if (ioeventfd_check_collision(kvm, p)) {
766 kvm_iodevice_init(&p->dev, &ioeventfd_ops);
768 ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
773 kvm->buses[bus_idx]->ioeventfd_count++;
774 list_add_tail(&p->list, &kvm->ioeventfds);
776 mutex_unlock(&kvm->slots_lock);
781 mutex_unlock(&kvm->slots_lock);
785 eventfd_ctx_put(eventfd);
791 kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
793 enum kvm_bus bus_idx;
794 struct _ioeventfd *p, *tmp;
795 struct eventfd_ctx *eventfd;
798 bus_idx = ioeventfd_bus_from_flags(args->flags);
799 eventfd = eventfd_ctx_fdget(args->fd);
801 return PTR_ERR(eventfd);
803 mutex_lock(&kvm->slots_lock);
805 list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
806 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
808 if (p->bus_idx != bus_idx ||
809 p->eventfd != eventfd ||
810 p->addr != args->addr ||
811 p->length != args->len ||
812 p->wildcard != wildcard)
815 if (!p->wildcard && p->datamatch != args->datamatch)
818 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
819 kvm->buses[bus_idx]->ioeventfd_count--;
820 ioeventfd_release(p);
825 mutex_unlock(&kvm->slots_lock);
827 eventfd_ctx_put(eventfd);
833 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
835 if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
836 return kvm_deassign_ioeventfd(kvm, args);
838 return kvm_assign_ioeventfd(kvm, args);