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>
39 * --------------------------------------------------------------------
40 * irqfd: Allows an fd to be used to inject an interrupt to the guest
42 * Credit goes to Avi Kivity for the original idea.
43 * --------------------------------------------------------------------
47 /* Used for MSI fast-path */
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 */
54 struct work_struct inject;
55 /* Used for setup/shutdown */
56 struct eventfd_ctx *eventfd;
57 struct list_head list;
59 struct work_struct shutdown;
62 static struct workqueue_struct *irqfd_cleanup_wq;
65 irqfd_inject(struct work_struct *work)
67 struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
68 struct kvm *kvm = irqfd->kvm;
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);
75 * Race-free decouple logic (ordering is critical)
78 irqfd_shutdown(struct work_struct *work)
80 struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
84 * Synchronize with the wait-queue and unhook ourselves to prevent
87 eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
90 * We know no new events will be scheduled at this point, so block
91 * until all previously outstanding events have completed
93 flush_work(&irqfd->inject);
96 * It is now safe to release the object's resources
98 eventfd_ctx_put(irqfd->eventfd);
103 /* assumes kvm->irqfds.lock is held */
105 irqfd_is_active(struct _irqfd *irqfd)
107 return list_empty(&irqfd->list) ? false : true;
111 * Mark the irqfd as inactive and schedule it for removal
113 * assumes kvm->irqfds.lock is held
116 irqfd_deactivate(struct _irqfd *irqfd)
118 BUG_ON(!irqfd_is_active(irqfd));
120 list_del_init(&irqfd->list);
122 queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
126 * Called with wqh->lock held and interrupts disabled
129 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
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;
136 if (flags & POLLIN) {
138 irq = rcu_dereference(irqfd->irq_entry);
139 /* An event has been signaled, inject an interrupt */
141 kvm_set_msi(irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1);
143 schedule_work(&irqfd->inject);
147 if (flags & POLLHUP) {
148 /* The eventfd is closing, detach from KVM */
151 spin_lock_irqsave(&kvm->irqfds.lock, flags);
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
162 if (irqfd_is_active(irqfd))
163 irqfd_deactivate(irqfd);
165 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
172 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
175 struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
176 add_wait_queue(wqh, &irqfd->wait);
179 /* Must be called under irqfds.lock */
180 static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd,
181 struct kvm_irq_routing_table *irq_rt)
183 struct kvm_kernel_irq_routing_entry *e;
184 struct hlist_node *n;
186 if (irqfd->gsi >= irq_rt->nr_rt_entries) {
187 rcu_assign_pointer(irqfd->irq_entry, NULL);
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);
196 rcu_assign_pointer(irqfd->irq_entry, NULL);
201 kvm_irqfd_assign(struct kvm *kvm, int fd, int gsi)
203 struct kvm_irq_routing_table *irq_rt;
204 struct _irqfd *irqfd, *tmp;
205 struct file *file = NULL;
206 struct eventfd_ctx *eventfd = NULL;
210 irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
216 INIT_LIST_HEAD(&irqfd->list);
217 INIT_WORK(&irqfd->inject, irqfd_inject);
218 INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
220 file = eventfd_fget(fd);
226 eventfd = eventfd_ctx_fileget(file);
227 if (IS_ERR(eventfd)) {
228 ret = PTR_ERR(eventfd);
232 irqfd->eventfd = eventfd;
235 * Install our own custom wake-up handling so we are notified via
236 * a callback whenever someone signals the underlying eventfd
238 init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
239 init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
241 spin_lock_irq(&kvm->irqfds.lock);
244 list_for_each_entry(tmp, &kvm->irqfds.items, list) {
245 if (irqfd->eventfd != tmp->eventfd)
247 /* This fd is used for another irq already. */
249 spin_unlock_irq(&kvm->irqfds.lock);
253 irq_rt = rcu_dereference_protected(kvm->irq_routing,
254 lockdep_is_held(&kvm->irqfds.lock));
255 irqfd_update(kvm, irqfd, irq_rt);
257 events = file->f_op->poll(file, &irqfd->pt);
259 list_add_tail(&irqfd->list, &kvm->irqfds.items);
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.
266 schedule_work(&irqfd->inject);
268 spin_unlock_irq(&kvm->irqfds.lock);
271 * do not drop the file until the irqfd is fully initialized, otherwise
272 * we might race against the POLLHUP
279 if (eventfd && !IS_ERR(eventfd))
280 eventfd_ctx_put(eventfd);
290 kvm_eventfd_init(struct kvm *kvm)
292 spin_lock_init(&kvm->irqfds.lock);
293 INIT_LIST_HEAD(&kvm->irqfds.items);
294 INIT_LIST_HEAD(&kvm->ioeventfds);
298 * shutdown any irqfd's that match fd+gsi
301 kvm_irqfd_deassign(struct kvm *kvm, int fd, int gsi)
303 struct _irqfd *irqfd, *tmp;
304 struct eventfd_ctx *eventfd;
306 eventfd = eventfd_ctx_fdget(fd);
308 return PTR_ERR(eventfd);
310 spin_lock_irq(&kvm->irqfds.lock);
312 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
313 if (irqfd->eventfd == eventfd && irqfd->gsi == gsi) {
315 * This rcu_assign_pointer is needed for when
316 * another thread calls kvm_irqfd_update before
317 * we flush workqueue below.
318 * It is paired with synchronize_rcu done by caller
321 rcu_assign_pointer(irqfd->irq_entry, NULL);
322 irqfd_deactivate(irqfd);
326 spin_unlock_irq(&kvm->irqfds.lock);
327 eventfd_ctx_put(eventfd);
330 * Block until we know all outstanding shutdown jobs have completed
331 * so that we guarantee there will not be any more interrupts on this
332 * gsi once this deassign function returns.
334 flush_workqueue(irqfd_cleanup_wq);
340 kvm_irqfd(struct kvm *kvm, int fd, int gsi, int flags)
342 if (flags & KVM_IRQFD_FLAG_DEASSIGN)
343 return kvm_irqfd_deassign(kvm, fd, gsi);
345 return kvm_irqfd_assign(kvm, fd, gsi);
349 * This function is called as the kvm VM fd is being released. Shutdown all
350 * irqfds that still remain open
353 kvm_irqfd_release(struct kvm *kvm)
355 struct _irqfd *irqfd, *tmp;
357 spin_lock_irq(&kvm->irqfds.lock);
359 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
360 irqfd_deactivate(irqfd);
362 spin_unlock_irq(&kvm->irqfds.lock);
365 * Block until we know all outstanding shutdown jobs have completed
366 * since we do not take a kvm* reference.
368 flush_workqueue(irqfd_cleanup_wq);
373 * Change irq_routing and irqfd.
374 * Caller must invoke synchronize_rcu afterwards.
376 void kvm_irq_routing_update(struct kvm *kvm,
377 struct kvm_irq_routing_table *irq_rt)
379 struct _irqfd *irqfd;
381 spin_lock_irq(&kvm->irqfds.lock);
383 rcu_assign_pointer(kvm->irq_routing, irq_rt);
385 list_for_each_entry(irqfd, &kvm->irqfds.items, list)
386 irqfd_update(kvm, irqfd, irq_rt);
388 spin_unlock_irq(&kvm->irqfds.lock);
392 * create a host-wide workqueue for issuing deferred shutdown requests
393 * aggregated from all vm* instances. We need our own isolated single-thread
394 * queue to prevent deadlock against flushing the normal work-queue.
396 static int __init irqfd_module_init(void)
398 irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
399 if (!irqfd_cleanup_wq)
405 static void __exit irqfd_module_exit(void)
407 destroy_workqueue(irqfd_cleanup_wq);
410 module_init(irqfd_module_init);
411 module_exit(irqfd_module_exit);
414 * --------------------------------------------------------------------
415 * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
417 * userspace can register a PIO/MMIO address with an eventfd for receiving
418 * notification when the memory has been touched.
419 * --------------------------------------------------------------------
423 struct list_head list;
426 struct eventfd_ctx *eventfd;
428 struct kvm_io_device dev;
432 static inline struct _ioeventfd *
433 to_ioeventfd(struct kvm_io_device *dev)
435 return container_of(dev, struct _ioeventfd, dev);
439 ioeventfd_release(struct _ioeventfd *p)
441 eventfd_ctx_put(p->eventfd);
447 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
451 if (!(addr == p->addr && len == p->length))
452 /* address-range must be precise for a hit */
456 /* all else equal, wildcard is always a hit */
459 /* otherwise, we have to actually compare the data */
461 BUG_ON(!IS_ALIGNED((unsigned long)val, len));
480 return _val == p->datamatch ? true : false;
483 /* MMIO/PIO writes trigger an event if the addr/val match */
485 ioeventfd_write(struct kvm_io_device *this, gpa_t addr, int len,
488 struct _ioeventfd *p = to_ioeventfd(this);
490 if (!ioeventfd_in_range(p, addr, len, val))
493 eventfd_signal(p->eventfd, 1);
498 * This function is called as KVM is completely shutting down. We do not
499 * need to worry about locking just nuke anything we have as quickly as possible
502 ioeventfd_destructor(struct kvm_io_device *this)
504 struct _ioeventfd *p = to_ioeventfd(this);
506 ioeventfd_release(p);
509 static const struct kvm_io_device_ops ioeventfd_ops = {
510 .write = ioeventfd_write,
511 .destructor = ioeventfd_destructor,
514 /* assumes kvm->slots_lock held */
516 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
518 struct _ioeventfd *_p;
520 list_for_each_entry(_p, &kvm->ioeventfds, list)
521 if (_p->addr == p->addr && _p->length == p->length &&
522 (_p->wildcard || p->wildcard ||
523 _p->datamatch == p->datamatch))
530 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
532 int pio = args->flags & KVM_IOEVENTFD_FLAG_PIO;
533 enum kvm_bus bus_idx = pio ? KVM_PIO_BUS : KVM_MMIO_BUS;
534 struct _ioeventfd *p;
535 struct eventfd_ctx *eventfd;
538 /* must be natural-word sized */
549 /* check for range overflow */
550 if (args->addr + args->len < args->addr)
553 /* check for extra flags that we don't understand */
554 if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
557 eventfd = eventfd_ctx_fdget(args->fd);
559 return PTR_ERR(eventfd);
561 p = kzalloc(sizeof(*p), GFP_KERNEL);
567 INIT_LIST_HEAD(&p->list);
568 p->addr = args->addr;
569 p->length = args->len;
570 p->eventfd = eventfd;
572 /* The datamatch feature is optional, otherwise this is a wildcard */
573 if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
574 p->datamatch = args->datamatch;
578 mutex_lock(&kvm->slots_lock);
580 /* Verify that there isnt a match already */
581 if (ioeventfd_check_collision(kvm, p)) {
586 kvm_iodevice_init(&p->dev, &ioeventfd_ops);
588 ret = kvm_io_bus_register_dev(kvm, bus_idx, &p->dev);
592 list_add_tail(&p->list, &kvm->ioeventfds);
594 mutex_unlock(&kvm->slots_lock);
599 mutex_unlock(&kvm->slots_lock);
603 eventfd_ctx_put(eventfd);
609 kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
611 int pio = args->flags & KVM_IOEVENTFD_FLAG_PIO;
612 enum kvm_bus bus_idx = pio ? KVM_PIO_BUS : KVM_MMIO_BUS;
613 struct _ioeventfd *p, *tmp;
614 struct eventfd_ctx *eventfd;
617 eventfd = eventfd_ctx_fdget(args->fd);
619 return PTR_ERR(eventfd);
621 mutex_lock(&kvm->slots_lock);
623 list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
624 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
626 if (p->eventfd != eventfd ||
627 p->addr != args->addr ||
628 p->length != args->len ||
629 p->wildcard != wildcard)
632 if (!p->wildcard && p->datamatch != args->datamatch)
635 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
636 ioeventfd_release(p);
641 mutex_unlock(&kvm->slots_lock);
643 eventfd_ctx_put(eventfd);
649 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
651 if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
652 return kvm_deassign_ioeventfd(kvm, args);
654 return kvm_assign_ioeventfd(kvm, args);