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/srcu.h>
35 #include <linux/slab.h>
36 #include <linux/seqlock.h>
40 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
42 * --------------------------------------------------------------------
43 * irqfd: Allows an fd to be used to inject an interrupt to the guest
45 * Credit goes to Avi Kivity for the original idea.
46 * --------------------------------------------------------------------
50 * Resampling irqfds are a special variety of irqfds used to emulate
51 * level triggered interrupts. The interrupt is asserted on eventfd
52 * trigger. On acknowledgement through the irq ack notifier, the
53 * interrupt is de-asserted and userspace is notified through the
54 * resamplefd. All resamplers on the same gsi are de-asserted
55 * together, so we don't need to track the state of each individual
56 * user. We can also therefore share the same irq source ID.
58 struct _irqfd_resampler {
61 * List of resampling struct _irqfd objects sharing this gsi.
62 * RCU list modified under kvm->irqfds.resampler_lock
64 struct list_head list;
65 struct kvm_irq_ack_notifier notifier;
67 * Entry in list of kvm->irqfd.resampler_list. Use for sharing
68 * resamplers among irqfds on the same gsi.
69 * Accessed and modified under kvm->irqfds.resampler_lock
71 struct list_head link;
75 /* Used for MSI fast-path */
78 /* Update side is protected by irqfds.lock */
79 struct kvm_kernel_irq_routing_entry irq_entry;
80 seqcount_t irq_entry_sc;
81 /* Used for level IRQ fast-path */
83 struct work_struct inject;
84 /* The resampler used by this irqfd (resampler-only) */
85 struct _irqfd_resampler *resampler;
86 /* Eventfd notified on resample (resampler-only) */
87 struct eventfd_ctx *resamplefd;
88 /* Entry in list of irqfds for a resampler (resampler-only) */
89 struct list_head resampler_link;
90 /* Used for setup/shutdown */
91 struct eventfd_ctx *eventfd;
92 struct list_head list;
94 struct work_struct shutdown;
97 static struct workqueue_struct *irqfd_cleanup_wq;
100 irqfd_inject(struct work_struct *work)
102 struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
103 struct kvm *kvm = irqfd->kvm;
105 if (!irqfd->resampler) {
106 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
108 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
111 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
112 irqfd->gsi, 1, false);
116 * Since resampler irqfds share an IRQ source ID, we de-assert once
117 * then notify all of the resampler irqfds using this GSI. We can't
118 * do multiple de-asserts or we risk racing with incoming re-asserts.
121 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
123 struct _irqfd_resampler *resampler;
125 struct _irqfd *irqfd;
128 resampler = container_of(kian, struct _irqfd_resampler, notifier);
129 kvm = resampler->kvm;
131 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
132 resampler->notifier.gsi, 0, false);
134 idx = srcu_read_lock(&kvm->irq_srcu);
136 list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
137 eventfd_signal(irqfd->resamplefd, 1);
139 srcu_read_unlock(&kvm->irq_srcu, idx);
143 irqfd_resampler_shutdown(struct _irqfd *irqfd)
145 struct _irqfd_resampler *resampler = irqfd->resampler;
146 struct kvm *kvm = resampler->kvm;
148 mutex_lock(&kvm->irqfds.resampler_lock);
150 list_del_rcu(&irqfd->resampler_link);
151 synchronize_srcu(&kvm->irq_srcu);
153 if (list_empty(&resampler->list)) {
154 list_del(&resampler->link);
155 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
156 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
157 resampler->notifier.gsi, 0, false);
161 mutex_unlock(&kvm->irqfds.resampler_lock);
165 * Race-free decouple logic (ordering is critical)
168 irqfd_shutdown(struct work_struct *work)
170 struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
174 * Synchronize with the wait-queue and unhook ourselves to prevent
177 eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
180 * We know no new events will be scheduled at this point, so block
181 * until all previously outstanding events have completed
183 flush_work(&irqfd->inject);
185 if (irqfd->resampler) {
186 irqfd_resampler_shutdown(irqfd);
187 eventfd_ctx_put(irqfd->resamplefd);
191 * It is now safe to release the object's resources
193 eventfd_ctx_put(irqfd->eventfd);
198 /* assumes kvm->irqfds.lock is held */
200 irqfd_is_active(struct _irqfd *irqfd)
202 return list_empty(&irqfd->list) ? false : true;
206 * Mark the irqfd as inactive and schedule it for removal
208 * assumes kvm->irqfds.lock is held
211 irqfd_deactivate(struct _irqfd *irqfd)
213 BUG_ON(!irqfd_is_active(irqfd));
215 list_del_init(&irqfd->list);
217 queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
221 * Called with wqh->lock held and interrupts disabled
224 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
226 struct _irqfd *irqfd = container_of(wait, struct _irqfd, wait);
227 unsigned long flags = (unsigned long)key;
228 struct kvm_kernel_irq_routing_entry irq;
229 struct kvm *kvm = irqfd->kvm;
233 if (flags & POLLIN) {
234 idx = srcu_read_lock(&kvm->irq_srcu);
236 seq = read_seqcount_begin(&irqfd->irq_entry_sc);
237 irq = irqfd->irq_entry;
238 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
239 /* An event has been signaled, inject an interrupt */
240 if (irq.type == KVM_IRQ_ROUTING_MSI)
241 kvm_set_msi(&irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1,
244 schedule_work(&irqfd->inject);
245 srcu_read_unlock(&kvm->irq_srcu, idx);
248 if (flags & POLLHUP) {
249 /* The eventfd is closing, detach from KVM */
252 spin_lock_irqsave(&kvm->irqfds.lock, flags);
255 * We must check if someone deactivated the irqfd before
256 * we could acquire the irqfds.lock since the item is
257 * deactivated from the KVM side before it is unhooked from
258 * the wait-queue. If it is already deactivated, we can
259 * simply return knowing the other side will cleanup for us.
260 * We cannot race against the irqfd going away since the
261 * other side is required to acquire wqh->lock, which we hold
263 if (irqfd_is_active(irqfd))
264 irqfd_deactivate(irqfd);
266 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
273 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
276 struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
277 add_wait_queue(wqh, &irqfd->wait);
280 /* Must be called under irqfds.lock */
281 static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd)
283 struct kvm_kernel_irq_routing_entry *e;
284 struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
287 n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
289 write_seqcount_begin(&irqfd->irq_entry_sc);
291 irqfd->irq_entry.type = 0;
294 for (i = 0; i < n_entries; ++i, ++e) {
295 /* Only fast-path MSI. */
296 if (e->type == KVM_IRQ_ROUTING_MSI)
297 irqfd->irq_entry = *e;
300 write_seqcount_end(&irqfd->irq_entry_sc);
304 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
306 struct _irqfd *irqfd, *tmp;
308 struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
313 irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
318 irqfd->gsi = args->gsi;
319 INIT_LIST_HEAD(&irqfd->list);
320 INIT_WORK(&irqfd->inject, irqfd_inject);
321 INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
322 seqcount_init(&irqfd->irq_entry_sc);
330 eventfd = eventfd_ctx_fileget(f.file);
331 if (IS_ERR(eventfd)) {
332 ret = PTR_ERR(eventfd);
336 irqfd->eventfd = eventfd;
338 if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
339 struct _irqfd_resampler *resampler;
341 resamplefd = eventfd_ctx_fdget(args->resamplefd);
342 if (IS_ERR(resamplefd)) {
343 ret = PTR_ERR(resamplefd);
347 irqfd->resamplefd = resamplefd;
348 INIT_LIST_HEAD(&irqfd->resampler_link);
350 mutex_lock(&kvm->irqfds.resampler_lock);
352 list_for_each_entry(resampler,
353 &kvm->irqfds.resampler_list, link) {
354 if (resampler->notifier.gsi == irqfd->gsi) {
355 irqfd->resampler = resampler;
360 if (!irqfd->resampler) {
361 resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
364 mutex_unlock(&kvm->irqfds.resampler_lock);
368 resampler->kvm = kvm;
369 INIT_LIST_HEAD(&resampler->list);
370 resampler->notifier.gsi = irqfd->gsi;
371 resampler->notifier.irq_acked = irqfd_resampler_ack;
372 INIT_LIST_HEAD(&resampler->link);
374 list_add(&resampler->link, &kvm->irqfds.resampler_list);
375 kvm_register_irq_ack_notifier(kvm,
376 &resampler->notifier);
377 irqfd->resampler = resampler;
380 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
381 synchronize_srcu(&kvm->irq_srcu);
383 mutex_unlock(&kvm->irqfds.resampler_lock);
387 * Install our own custom wake-up handling so we are notified via
388 * a callback whenever someone signals the underlying eventfd
390 init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
391 init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
393 spin_lock_irq(&kvm->irqfds.lock);
396 list_for_each_entry(tmp, &kvm->irqfds.items, list) {
397 if (irqfd->eventfd != tmp->eventfd)
399 /* This fd is used for another irq already. */
401 spin_unlock_irq(&kvm->irqfds.lock);
405 idx = srcu_read_lock(&kvm->irq_srcu);
406 irqfd_update(kvm, irqfd);
407 srcu_read_unlock(&kvm->irq_srcu, idx);
409 list_add_tail(&irqfd->list, &kvm->irqfds.items);
411 spin_unlock_irq(&kvm->irqfds.lock);
414 * Check if there was an event already pending on the eventfd
415 * before we registered, and trigger it as if we didn't miss it.
417 events = f.file->f_op->poll(f.file, &irqfd->pt);
420 schedule_work(&irqfd->inject);
423 * do not drop the file until the irqfd is fully initialized, otherwise
424 * we might race against the POLLHUP
431 if (irqfd->resampler)
432 irqfd_resampler_shutdown(irqfd);
434 if (resamplefd && !IS_ERR(resamplefd))
435 eventfd_ctx_put(resamplefd);
437 if (eventfd && !IS_ERR(eventfd))
438 eventfd_ctx_put(eventfd);
449 kvm_eventfd_init(struct kvm *kvm)
451 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
452 spin_lock_init(&kvm->irqfds.lock);
453 INIT_LIST_HEAD(&kvm->irqfds.items);
454 INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
455 mutex_init(&kvm->irqfds.resampler_lock);
457 INIT_LIST_HEAD(&kvm->ioeventfds);
460 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
462 * shutdown any irqfd's that match fd+gsi
465 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
467 struct _irqfd *irqfd, *tmp;
468 struct eventfd_ctx *eventfd;
470 eventfd = eventfd_ctx_fdget(args->fd);
472 return PTR_ERR(eventfd);
474 spin_lock_irq(&kvm->irqfds.lock);
476 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
477 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
479 * This clearing of irq_entry.type is needed for when
480 * another thread calls kvm_irq_routing_update before
481 * we flush workqueue below (we synchronize with
482 * kvm_irq_routing_update using irqfds.lock).
484 write_seqcount_begin(&irqfd->irq_entry_sc);
485 irqfd->irq_entry.type = 0;
486 write_seqcount_end(&irqfd->irq_entry_sc);
487 irqfd_deactivate(irqfd);
491 spin_unlock_irq(&kvm->irqfds.lock);
492 eventfd_ctx_put(eventfd);
495 * Block until we know all outstanding shutdown jobs have completed
496 * so that we guarantee there will not be any more interrupts on this
497 * gsi once this deassign function returns.
499 flush_workqueue(irqfd_cleanup_wq);
505 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
507 if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
510 if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
511 return kvm_irqfd_deassign(kvm, args);
513 return kvm_irqfd_assign(kvm, args);
517 * This function is called as the kvm VM fd is being released. Shutdown all
518 * irqfds that still remain open
521 kvm_irqfd_release(struct kvm *kvm)
523 struct _irqfd *irqfd, *tmp;
525 spin_lock_irq(&kvm->irqfds.lock);
527 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
528 irqfd_deactivate(irqfd);
530 spin_unlock_irq(&kvm->irqfds.lock);
533 * Block until we know all outstanding shutdown jobs have completed
534 * since we do not take a kvm* reference.
536 flush_workqueue(irqfd_cleanup_wq);
541 * Take note of a change in irq routing.
542 * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
544 void kvm_irq_routing_update(struct kvm *kvm)
546 struct _irqfd *irqfd;
548 spin_lock_irq(&kvm->irqfds.lock);
550 list_for_each_entry(irqfd, &kvm->irqfds.items, list)
551 irqfd_update(kvm, irqfd);
553 spin_unlock_irq(&kvm->irqfds.lock);
557 * create a host-wide workqueue for issuing deferred shutdown requests
558 * aggregated from all vm* instances. We need our own isolated single-thread
559 * queue to prevent deadlock against flushing the normal work-queue.
561 int kvm_irqfd_init(void)
563 irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
564 if (!irqfd_cleanup_wq)
570 void kvm_irqfd_exit(void)
572 destroy_workqueue(irqfd_cleanup_wq);
577 * --------------------------------------------------------------------
578 * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
580 * userspace can register a PIO/MMIO address with an eventfd for receiving
581 * notification when the memory has been touched.
582 * --------------------------------------------------------------------
586 struct list_head list;
589 struct eventfd_ctx *eventfd;
591 struct kvm_io_device dev;
596 static inline struct _ioeventfd *
597 to_ioeventfd(struct kvm_io_device *dev)
599 return container_of(dev, struct _ioeventfd, dev);
603 ioeventfd_release(struct _ioeventfd *p)
605 eventfd_ctx_put(p->eventfd);
611 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
616 /* address must be precise for a hit */
620 /* length = 0 means only look at the address, so always a hit */
623 if (len != p->length)
624 /* address-range must be precise for a hit */
628 /* all else equal, wildcard is always a hit */
631 /* otherwise, we have to actually compare the data */
633 BUG_ON(!IS_ALIGNED((unsigned long)val, len));
652 return _val == p->datamatch ? true : false;
655 /* MMIO/PIO writes trigger an event if the addr/val match */
657 ioeventfd_write(struct kvm_io_device *this, gpa_t addr, int len,
660 struct _ioeventfd *p = to_ioeventfd(this);
662 if (!ioeventfd_in_range(p, addr, len, val))
665 eventfd_signal(p->eventfd, 1);
670 * This function is called as KVM is completely shutting down. We do not
671 * need to worry about locking just nuke anything we have as quickly as possible
674 ioeventfd_destructor(struct kvm_io_device *this)
676 struct _ioeventfd *p = to_ioeventfd(this);
678 ioeventfd_release(p);
681 static const struct kvm_io_device_ops ioeventfd_ops = {
682 .write = ioeventfd_write,
683 .destructor = ioeventfd_destructor,
686 /* assumes kvm->slots_lock held */
688 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
690 struct _ioeventfd *_p;
692 list_for_each_entry(_p, &kvm->ioeventfds, list)
693 if (_p->bus_idx == p->bus_idx &&
694 _p->addr == p->addr &&
695 (!_p->length || !p->length ||
696 (_p->length == p->length &&
697 (_p->wildcard || p->wildcard ||
698 _p->datamatch == p->datamatch))))
704 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
706 if (flags & KVM_IOEVENTFD_FLAG_PIO)
708 if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
709 return KVM_VIRTIO_CCW_NOTIFY_BUS;
714 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
716 enum kvm_bus bus_idx;
717 struct _ioeventfd *p;
718 struct eventfd_ctx *eventfd;
721 bus_idx = ioeventfd_bus_from_flags(args->flags);
722 /* must be natural-word sized, or 0 to ignore length */
734 /* check for range overflow */
735 if (args->addr + args->len < args->addr)
738 /* check for extra flags that we don't understand */
739 if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
742 /* ioeventfd with no length can't be combined with DATAMATCH */
744 args->flags & (KVM_IOEVENTFD_FLAG_PIO |
745 KVM_IOEVENTFD_FLAG_DATAMATCH))
748 eventfd = eventfd_ctx_fdget(args->fd);
750 return PTR_ERR(eventfd);
752 p = kzalloc(sizeof(*p), GFP_KERNEL);
758 INIT_LIST_HEAD(&p->list);
759 p->addr = args->addr;
760 p->bus_idx = bus_idx;
761 p->length = args->len;
762 p->eventfd = eventfd;
764 /* The datamatch feature is optional, otherwise this is a wildcard */
765 if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
766 p->datamatch = args->datamatch;
770 mutex_lock(&kvm->slots_lock);
772 /* Verify that there isn't a match already */
773 if (ioeventfd_check_collision(kvm, p)) {
778 kvm_iodevice_init(&p->dev, &ioeventfd_ops);
780 ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
785 /* When length is ignored, MMIO is also put on a separate bus, for
788 if (!args->len && !(args->flags & KVM_IOEVENTFD_FLAG_PIO)) {
789 ret = kvm_io_bus_register_dev(kvm, KVM_FAST_MMIO_BUS,
790 p->addr, 0, &p->dev);
795 kvm->buses[bus_idx]->ioeventfd_count++;
796 list_add_tail(&p->list, &kvm->ioeventfds);
798 mutex_unlock(&kvm->slots_lock);
803 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
805 mutex_unlock(&kvm->slots_lock);
809 eventfd_ctx_put(eventfd);
815 kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
817 enum kvm_bus bus_idx;
818 struct _ioeventfd *p, *tmp;
819 struct eventfd_ctx *eventfd;
822 bus_idx = ioeventfd_bus_from_flags(args->flags);
823 eventfd = eventfd_ctx_fdget(args->fd);
825 return PTR_ERR(eventfd);
827 mutex_lock(&kvm->slots_lock);
829 list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
830 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
832 if (p->bus_idx != bus_idx ||
833 p->eventfd != eventfd ||
834 p->addr != args->addr ||
835 p->length != args->len ||
836 p->wildcard != wildcard)
839 if (!p->wildcard && p->datamatch != args->datamatch)
842 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
844 kvm_io_bus_unregister_dev(kvm, KVM_FAST_MMIO_BUS,
847 kvm->buses[bus_idx]->ioeventfd_count--;
848 ioeventfd_release(p);
853 mutex_unlock(&kvm->slots_lock);
855 eventfd_ctx_put(eventfd);
861 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
863 if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
864 return kvm_deassign_ioeventfd(kvm, args);
866 return kvm_assign_ioeventfd(kvm, args);