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// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * kvm eventfd support - use eventfd objects to signal various KVM events
4 *
5 * Copyright 2009 Novell. All Rights Reserved.
6 * Copyright 2010 Red Hat, Inc. and/or its affiliates.
7 *
8 * Author:
9 * Gregory Haskins <ghaskins@novell.com>
10 */
11
12#include <linux/kvm_host.h>
13#include <linux/kvm.h>
14#include <linux/kvm_irqfd.h>
15#include <linux/workqueue.h>
16#include <linux/syscalls.h>
17#include <linux/wait.h>
18#include <linux/poll.h>
19#include <linux/file.h>
20#include <linux/list.h>
21#include <linux/eventfd.h>
22#include <linux/kernel.h>
23#include <linux/srcu.h>
24#include <linux/slab.h>
25#include <linux/seqlock.h>
26#include <linux/irqbypass.h>
27#include <trace/events/kvm.h>
28
29#include <kvm/iodev.h>
30
31#ifdef CONFIG_HAVE_KVM_IRQCHIP
32
33static struct workqueue_struct *irqfd_cleanup_wq;
34
35bool __attribute__((weak))
36kvm_arch_irqfd_allowed(struct kvm *kvm, struct kvm_irqfd *args)
37{
38 return true;
39}
40
41static void
42irqfd_inject(struct work_struct *work)
43{
44 struct kvm_kernel_irqfd *irqfd =
45 container_of(work, struct kvm_kernel_irqfd, inject);
46 struct kvm *kvm = irqfd->kvm;
47
48 if (!irqfd->resampler) {
49 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
50 false);
51 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
52 false);
53 } else
54 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
55 irqfd->gsi, 1, false);
56}
57
58static void irqfd_resampler_notify(struct kvm_kernel_irqfd_resampler *resampler)
59{
60 struct kvm_kernel_irqfd *irqfd;
61
62 list_for_each_entry_srcu(irqfd, &resampler->list, resampler_link,
63 srcu_read_lock_held(&resampler->kvm->irq_srcu))
64 eventfd_signal(irqfd->resamplefd);
65}
66
67/*
68 * Since resampler irqfds share an IRQ source ID, we de-assert once
69 * then notify all of the resampler irqfds using this GSI. We can't
70 * do multiple de-asserts or we risk racing with incoming re-asserts.
71 */
72static void
73irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
74{
75 struct kvm_kernel_irqfd_resampler *resampler;
76 struct kvm *kvm;
77 int idx;
78
79 resampler = container_of(kian,
80 struct kvm_kernel_irqfd_resampler, notifier);
81 kvm = resampler->kvm;
82
83 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
84 resampler->notifier.gsi, 0, false);
85
86 idx = srcu_read_lock(&kvm->irq_srcu);
87 irqfd_resampler_notify(resampler);
88 srcu_read_unlock(&kvm->irq_srcu, idx);
89}
90
91static void
92irqfd_resampler_shutdown(struct kvm_kernel_irqfd *irqfd)
93{
94 struct kvm_kernel_irqfd_resampler *resampler = irqfd->resampler;
95 struct kvm *kvm = resampler->kvm;
96
97 mutex_lock(&kvm->irqfds.resampler_lock);
98
99 list_del_rcu(&irqfd->resampler_link);
100
101 if (list_empty(&resampler->list)) {
102 list_del_rcu(&resampler->link);
103 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
104 /*
105 * synchronize_srcu_expedited(&kvm->irq_srcu) already called
106 * in kvm_unregister_irq_ack_notifier().
107 */
108 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
109 resampler->notifier.gsi, 0, false);
110 kfree(resampler);
111 } else {
112 synchronize_srcu_expedited(&kvm->irq_srcu);
113 }
114
115 mutex_unlock(&kvm->irqfds.resampler_lock);
116}
117
118/*
119 * Race-free decouple logic (ordering is critical)
120 */
121static void
122irqfd_shutdown(struct work_struct *work)
123{
124 struct kvm_kernel_irqfd *irqfd =
125 container_of(work, struct kvm_kernel_irqfd, shutdown);
126 struct kvm *kvm = irqfd->kvm;
127 u64 cnt;
128
129 /* Make sure irqfd has been initialized in assign path. */
130 synchronize_srcu_expedited(&kvm->irq_srcu);
131
132 /*
133 * Synchronize with the wait-queue and unhook ourselves to prevent
134 * further events.
135 */
136 eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
137
138 /*
139 * We know no new events will be scheduled at this point, so block
140 * until all previously outstanding events have completed
141 */
142 flush_work(&irqfd->inject);
143
144 if (irqfd->resampler) {
145 irqfd_resampler_shutdown(irqfd);
146 eventfd_ctx_put(irqfd->resamplefd);
147 }
148
149 /*
150 * It is now safe to release the object's resources
151 */
152#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
153 irq_bypass_unregister_consumer(&irqfd->consumer);
154#endif
155 eventfd_ctx_put(irqfd->eventfd);
156 kfree(irqfd);
157}
158
159
160/* assumes kvm->irqfds.lock is held */
161static bool
162irqfd_is_active(struct kvm_kernel_irqfd *irqfd)
163{
164 return list_empty(&irqfd->list) ? false : true;
165}
166
167/*
168 * Mark the irqfd as inactive and schedule it for removal
169 *
170 * assumes kvm->irqfds.lock is held
171 */
172static void
173irqfd_deactivate(struct kvm_kernel_irqfd *irqfd)
174{
175 BUG_ON(!irqfd_is_active(irqfd));
176
177 list_del_init(&irqfd->list);
178
179 queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
180}
181
182int __attribute__((weak)) kvm_arch_set_irq_inatomic(
183 struct kvm_kernel_irq_routing_entry *irq,
184 struct kvm *kvm, int irq_source_id,
185 int level,
186 bool line_status)
187{
188 return -EWOULDBLOCK;
189}
190
191/*
192 * Called with wqh->lock held and interrupts disabled
193 */
194static int
195irqfd_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
196{
197 struct kvm_kernel_irqfd *irqfd =
198 container_of(wait, struct kvm_kernel_irqfd, wait);
199 __poll_t flags = key_to_poll(key);
200 struct kvm_kernel_irq_routing_entry irq;
201 struct kvm *kvm = irqfd->kvm;
202 unsigned seq;
203 int idx;
204 int ret = 0;
205
206 if (flags & EPOLLIN) {
207 u64 cnt;
208 eventfd_ctx_do_read(irqfd->eventfd, &cnt);
209
210 idx = srcu_read_lock(&kvm->irq_srcu);
211 do {
212 seq = read_seqcount_begin(&irqfd->irq_entry_sc);
213 irq = irqfd->irq_entry;
214 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
215 /* An event has been signaled, inject an interrupt */
216 if (kvm_arch_set_irq_inatomic(&irq, kvm,
217 KVM_USERSPACE_IRQ_SOURCE_ID, 1,
218 false) == -EWOULDBLOCK)
219 schedule_work(&irqfd->inject);
220 srcu_read_unlock(&kvm->irq_srcu, idx);
221 ret = 1;
222 }
223
224 if (flags & EPOLLHUP) {
225 /* The eventfd is closing, detach from KVM */
226 unsigned long iflags;
227
228 spin_lock_irqsave(&kvm->irqfds.lock, iflags);
229
230 /*
231 * We must check if someone deactivated the irqfd before
232 * we could acquire the irqfds.lock since the item is
233 * deactivated from the KVM side before it is unhooked from
234 * the wait-queue. If it is already deactivated, we can
235 * simply return knowing the other side will cleanup for us.
236 * We cannot race against the irqfd going away since the
237 * other side is required to acquire wqh->lock, which we hold
238 */
239 if (irqfd_is_active(irqfd))
240 irqfd_deactivate(irqfd);
241
242 spin_unlock_irqrestore(&kvm->irqfds.lock, iflags);
243 }
244
245 return ret;
246}
247
248static void
249irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
250 poll_table *pt)
251{
252 struct kvm_kernel_irqfd *irqfd =
253 container_of(pt, struct kvm_kernel_irqfd, pt);
254 add_wait_queue_priority(wqh, &irqfd->wait);
255}
256
257/* Must be called under irqfds.lock */
258static void irqfd_update(struct kvm *kvm, struct kvm_kernel_irqfd *irqfd)
259{
260 struct kvm_kernel_irq_routing_entry *e;
261 struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
262 int n_entries;
263
264 n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
265
266 write_seqcount_begin(&irqfd->irq_entry_sc);
267
268 e = entries;
269 if (n_entries == 1)
270 irqfd->irq_entry = *e;
271 else
272 irqfd->irq_entry.type = 0;
273
274 write_seqcount_end(&irqfd->irq_entry_sc);
275}
276
277#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
278void __attribute__((weak)) kvm_arch_irq_bypass_stop(
279 struct irq_bypass_consumer *cons)
280{
281}
282
283void __attribute__((weak)) kvm_arch_irq_bypass_start(
284 struct irq_bypass_consumer *cons)
285{
286}
287
288int __attribute__((weak)) kvm_arch_update_irqfd_routing(
289 struct kvm *kvm, unsigned int host_irq,
290 uint32_t guest_irq, bool set)
291{
292 return 0;
293}
294
295bool __attribute__((weak)) kvm_arch_irqfd_route_changed(
296 struct kvm_kernel_irq_routing_entry *old,
297 struct kvm_kernel_irq_routing_entry *new)
298{
299 return true;
300}
301#endif
302
303static int
304kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
305{
306 struct kvm_kernel_irqfd *irqfd, *tmp;
307 struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
308 int ret;
309 __poll_t events;
310 int idx;
311
312 if (!kvm_arch_intc_initialized(kvm))
313 return -EAGAIN;
314
315 if (!kvm_arch_irqfd_allowed(kvm, args))
316 return -EINVAL;
317
318 irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL_ACCOUNT);
319 if (!irqfd)
320 return -ENOMEM;
321
322 irqfd->kvm = kvm;
323 irqfd->gsi = args->gsi;
324 INIT_LIST_HEAD(&irqfd->list);
325 INIT_WORK(&irqfd->inject, irqfd_inject);
326 INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
327 seqcount_spinlock_init(&irqfd->irq_entry_sc, &kvm->irqfds.lock);
328
329 CLASS(fd, f)(args->fd);
330 if (fd_empty(f)) {
331 ret = -EBADF;
332 goto out;
333 }
334
335 eventfd = eventfd_ctx_fileget(fd_file(f));
336 if (IS_ERR(eventfd)) {
337 ret = PTR_ERR(eventfd);
338 goto out;
339 }
340
341 irqfd->eventfd = eventfd;
342
343 if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
344 struct kvm_kernel_irqfd_resampler *resampler;
345
346 resamplefd = eventfd_ctx_fdget(args->resamplefd);
347 if (IS_ERR(resamplefd)) {
348 ret = PTR_ERR(resamplefd);
349 goto fail;
350 }
351
352 irqfd->resamplefd = resamplefd;
353 INIT_LIST_HEAD(&irqfd->resampler_link);
354
355 mutex_lock(&kvm->irqfds.resampler_lock);
356
357 list_for_each_entry(resampler,
358 &kvm->irqfds.resampler_list, link) {
359 if (resampler->notifier.gsi == irqfd->gsi) {
360 irqfd->resampler = resampler;
361 break;
362 }
363 }
364
365 if (!irqfd->resampler) {
366 resampler = kzalloc(sizeof(*resampler),
367 GFP_KERNEL_ACCOUNT);
368 if (!resampler) {
369 ret = -ENOMEM;
370 mutex_unlock(&kvm->irqfds.resampler_lock);
371 goto fail;
372 }
373
374 resampler->kvm = kvm;
375 INIT_LIST_HEAD(&resampler->list);
376 resampler->notifier.gsi = irqfd->gsi;
377 resampler->notifier.irq_acked = irqfd_resampler_ack;
378 INIT_LIST_HEAD(&resampler->link);
379
380 list_add_rcu(&resampler->link, &kvm->irqfds.resampler_list);
381 kvm_register_irq_ack_notifier(kvm,
382 &resampler->notifier);
383 irqfd->resampler = resampler;
384 }
385
386 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
387 synchronize_srcu_expedited(&kvm->irq_srcu);
388
389 mutex_unlock(&kvm->irqfds.resampler_lock);
390 }
391
392 /*
393 * Install our own custom wake-up handling so we are notified via
394 * a callback whenever someone signals the underlying eventfd
395 */
396 init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
397 init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
398
399 spin_lock_irq(&kvm->irqfds.lock);
400
401 ret = 0;
402 list_for_each_entry(tmp, &kvm->irqfds.items, list) {
403 if (irqfd->eventfd != tmp->eventfd)
404 continue;
405 /* This fd is used for another irq already. */
406 ret = -EBUSY;
407 spin_unlock_irq(&kvm->irqfds.lock);
408 goto fail;
409 }
410
411 idx = srcu_read_lock(&kvm->irq_srcu);
412 irqfd_update(kvm, irqfd);
413
414 list_add_tail(&irqfd->list, &kvm->irqfds.items);
415
416 spin_unlock_irq(&kvm->irqfds.lock);
417
418 /*
419 * Check if there was an event already pending on the eventfd
420 * before we registered, and trigger it as if we didn't miss it.
421 */
422 events = vfs_poll(fd_file(f), &irqfd->pt);
423
424 if (events & EPOLLIN)
425 schedule_work(&irqfd->inject);
426
427#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
428 if (kvm_arch_has_irq_bypass()) {
429 irqfd->consumer.token = (void *)irqfd->eventfd;
430 irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
431 irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
432 irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
433 irqfd->consumer.start = kvm_arch_irq_bypass_start;
434 ret = irq_bypass_register_consumer(&irqfd->consumer);
435 if (ret)
436 pr_info("irq bypass consumer (token %p) registration fails: %d\n",
437 irqfd->consumer.token, ret);
438 }
439#endif
440
441 srcu_read_unlock(&kvm->irq_srcu, idx);
442 return 0;
443
444fail:
445 if (irqfd->resampler)
446 irqfd_resampler_shutdown(irqfd);
447
448 if (resamplefd && !IS_ERR(resamplefd))
449 eventfd_ctx_put(resamplefd);
450
451 if (eventfd && !IS_ERR(eventfd))
452 eventfd_ctx_put(eventfd);
453
454out:
455 kfree(irqfd);
456 return ret;
457}
458
459bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
460{
461 struct kvm_irq_ack_notifier *kian;
462 int gsi, idx;
463
464 idx = srcu_read_lock(&kvm->irq_srcu);
465 gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
466 if (gsi != -1)
467 hlist_for_each_entry_srcu(kian, &kvm->irq_ack_notifier_list,
468 link, srcu_read_lock_held(&kvm->irq_srcu))
469 if (kian->gsi == gsi) {
470 srcu_read_unlock(&kvm->irq_srcu, idx);
471 return true;
472 }
473
474 srcu_read_unlock(&kvm->irq_srcu, idx);
475
476 return false;
477}
478EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
479
480void kvm_notify_acked_gsi(struct kvm *kvm, int gsi)
481{
482 struct kvm_irq_ack_notifier *kian;
483
484 hlist_for_each_entry_srcu(kian, &kvm->irq_ack_notifier_list,
485 link, srcu_read_lock_held(&kvm->irq_srcu))
486 if (kian->gsi == gsi)
487 kian->irq_acked(kian);
488}
489
490void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
491{
492 int gsi, idx;
493
494 trace_kvm_ack_irq(irqchip, pin);
495
496 idx = srcu_read_lock(&kvm->irq_srcu);
497 gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
498 if (gsi != -1)
499 kvm_notify_acked_gsi(kvm, gsi);
500 srcu_read_unlock(&kvm->irq_srcu, idx);
501}
502
503void kvm_register_irq_ack_notifier(struct kvm *kvm,
504 struct kvm_irq_ack_notifier *kian)
505{
506 mutex_lock(&kvm->irq_lock);
507 hlist_add_head_rcu(&kian->link, &kvm->irq_ack_notifier_list);
508 mutex_unlock(&kvm->irq_lock);
509 kvm_arch_post_irq_ack_notifier_list_update(kvm);
510}
511
512void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
513 struct kvm_irq_ack_notifier *kian)
514{
515 mutex_lock(&kvm->irq_lock);
516 hlist_del_init_rcu(&kian->link);
517 mutex_unlock(&kvm->irq_lock);
518 synchronize_srcu_expedited(&kvm->irq_srcu);
519 kvm_arch_post_irq_ack_notifier_list_update(kvm);
520}
521
522/*
523 * shutdown any irqfd's that match fd+gsi
524 */
525static int
526kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
527{
528 struct kvm_kernel_irqfd *irqfd, *tmp;
529 struct eventfd_ctx *eventfd;
530
531 eventfd = eventfd_ctx_fdget(args->fd);
532 if (IS_ERR(eventfd))
533 return PTR_ERR(eventfd);
534
535 spin_lock_irq(&kvm->irqfds.lock);
536
537 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
538 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
539 /*
540 * This clearing of irq_entry.type is needed for when
541 * another thread calls kvm_irq_routing_update before
542 * we flush workqueue below (we synchronize with
543 * kvm_irq_routing_update using irqfds.lock).
544 */
545 write_seqcount_begin(&irqfd->irq_entry_sc);
546 irqfd->irq_entry.type = 0;
547 write_seqcount_end(&irqfd->irq_entry_sc);
548 irqfd_deactivate(irqfd);
549 }
550 }
551
552 spin_unlock_irq(&kvm->irqfds.lock);
553 eventfd_ctx_put(eventfd);
554
555 /*
556 * Block until we know all outstanding shutdown jobs have completed
557 * so that we guarantee there will not be any more interrupts on this
558 * gsi once this deassign function returns.
559 */
560 flush_workqueue(irqfd_cleanup_wq);
561
562 return 0;
563}
564
565int
566kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
567{
568 if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
569 return -EINVAL;
570
571 if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
572 return kvm_irqfd_deassign(kvm, args);
573
574 return kvm_irqfd_assign(kvm, args);
575}
576
577/*
578 * This function is called as the kvm VM fd is being released. Shutdown all
579 * irqfds that still remain open
580 */
581void
582kvm_irqfd_release(struct kvm *kvm)
583{
584 struct kvm_kernel_irqfd *irqfd, *tmp;
585
586 spin_lock_irq(&kvm->irqfds.lock);
587
588 list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
589 irqfd_deactivate(irqfd);
590
591 spin_unlock_irq(&kvm->irqfds.lock);
592
593 /*
594 * Block until we know all outstanding shutdown jobs have completed
595 * since we do not take a kvm* reference.
596 */
597 flush_workqueue(irqfd_cleanup_wq);
598
599}
600
601/*
602 * Take note of a change in irq routing.
603 * Caller must invoke synchronize_srcu_expedited(&kvm->irq_srcu) afterwards.
604 */
605void kvm_irq_routing_update(struct kvm *kvm)
606{
607 struct kvm_kernel_irqfd *irqfd;
608
609 spin_lock_irq(&kvm->irqfds.lock);
610
611 list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
612#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
613 /* Under irqfds.lock, so can read irq_entry safely */
614 struct kvm_kernel_irq_routing_entry old = irqfd->irq_entry;
615#endif
616
617 irqfd_update(kvm, irqfd);
618
619#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
620 if (irqfd->producer &&
621 kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry)) {
622 int ret = kvm_arch_update_irqfd_routing(
623 irqfd->kvm, irqfd->producer->irq,
624 irqfd->gsi, 1);
625 WARN_ON(ret);
626 }
627#endif
628 }
629
630 spin_unlock_irq(&kvm->irqfds.lock);
631}
632
633bool kvm_notify_irqfd_resampler(struct kvm *kvm,
634 unsigned int irqchip,
635 unsigned int pin)
636{
637 struct kvm_kernel_irqfd_resampler *resampler;
638 int gsi, idx;
639
640 idx = srcu_read_lock(&kvm->irq_srcu);
641 gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
642 if (gsi != -1) {
643 list_for_each_entry_srcu(resampler,
644 &kvm->irqfds.resampler_list, link,
645 srcu_read_lock_held(&kvm->irq_srcu)) {
646 if (resampler->notifier.gsi == gsi) {
647 irqfd_resampler_notify(resampler);
648 srcu_read_unlock(&kvm->irq_srcu, idx);
649 return true;
650 }
651 }
652 }
653 srcu_read_unlock(&kvm->irq_srcu, idx);
654
655 return false;
656}
657
658/*
659 * create a host-wide workqueue for issuing deferred shutdown requests
660 * aggregated from all vm* instances. We need our own isolated
661 * queue to ease flushing work items when a VM exits.
662 */
663int kvm_irqfd_init(void)
664{
665 irqfd_cleanup_wq = alloc_workqueue("kvm-irqfd-cleanup", 0, 0);
666 if (!irqfd_cleanup_wq)
667 return -ENOMEM;
668
669 return 0;
670}
671
672void kvm_irqfd_exit(void)
673{
674 destroy_workqueue(irqfd_cleanup_wq);
675}
676#endif
677
678/*
679 * --------------------------------------------------------------------
680 * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
681 *
682 * userspace can register a PIO/MMIO address with an eventfd for receiving
683 * notification when the memory has been touched.
684 * --------------------------------------------------------------------
685 */
686
687struct _ioeventfd {
688 struct list_head list;
689 u64 addr;
690 int length;
691 struct eventfd_ctx *eventfd;
692 u64 datamatch;
693 struct kvm_io_device dev;
694 u8 bus_idx;
695 bool wildcard;
696};
697
698static inline struct _ioeventfd *
699to_ioeventfd(struct kvm_io_device *dev)
700{
701 return container_of(dev, struct _ioeventfd, dev);
702}
703
704static void
705ioeventfd_release(struct _ioeventfd *p)
706{
707 eventfd_ctx_put(p->eventfd);
708 list_del(&p->list);
709 kfree(p);
710}
711
712static bool
713ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
714{
715 u64 _val;
716
717 if (addr != p->addr)
718 /* address must be precise for a hit */
719 return false;
720
721 if (!p->length)
722 /* length = 0 means only look at the address, so always a hit */
723 return true;
724
725 if (len != p->length)
726 /* address-range must be precise for a hit */
727 return false;
728
729 if (p->wildcard)
730 /* all else equal, wildcard is always a hit */
731 return true;
732
733 /* otherwise, we have to actually compare the data */
734
735 BUG_ON(!IS_ALIGNED((unsigned long)val, len));
736
737 switch (len) {
738 case 1:
739 _val = *(u8 *)val;
740 break;
741 case 2:
742 _val = *(u16 *)val;
743 break;
744 case 4:
745 _val = *(u32 *)val;
746 break;
747 case 8:
748 _val = *(u64 *)val;
749 break;
750 default:
751 return false;
752 }
753
754 return _val == p->datamatch;
755}
756
757/* MMIO/PIO writes trigger an event if the addr/val match */
758static int
759ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
760 int len, const void *val)
761{
762 struct _ioeventfd *p = to_ioeventfd(this);
763
764 if (!ioeventfd_in_range(p, addr, len, val))
765 return -EOPNOTSUPP;
766
767 eventfd_signal(p->eventfd);
768 return 0;
769}
770
771/*
772 * This function is called as KVM is completely shutting down. We do not
773 * need to worry about locking just nuke anything we have as quickly as possible
774 */
775static void
776ioeventfd_destructor(struct kvm_io_device *this)
777{
778 struct _ioeventfd *p = to_ioeventfd(this);
779
780 ioeventfd_release(p);
781}
782
783static const struct kvm_io_device_ops ioeventfd_ops = {
784 .write = ioeventfd_write,
785 .destructor = ioeventfd_destructor,
786};
787
788/* assumes kvm->slots_lock held */
789static bool
790ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
791{
792 struct _ioeventfd *_p;
793
794 list_for_each_entry(_p, &kvm->ioeventfds, list)
795 if (_p->bus_idx == p->bus_idx &&
796 _p->addr == p->addr &&
797 (!_p->length || !p->length ||
798 (_p->length == p->length &&
799 (_p->wildcard || p->wildcard ||
800 _p->datamatch == p->datamatch))))
801 return true;
802
803 return false;
804}
805
806static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
807{
808 if (flags & KVM_IOEVENTFD_FLAG_PIO)
809 return KVM_PIO_BUS;
810 if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
811 return KVM_VIRTIO_CCW_NOTIFY_BUS;
812 return KVM_MMIO_BUS;
813}
814
815static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
816 enum kvm_bus bus_idx,
817 struct kvm_ioeventfd *args)
818{
819
820 struct eventfd_ctx *eventfd;
821 struct _ioeventfd *p;
822 int ret;
823
824 eventfd = eventfd_ctx_fdget(args->fd);
825 if (IS_ERR(eventfd))
826 return PTR_ERR(eventfd);
827
828 p = kzalloc(sizeof(*p), GFP_KERNEL_ACCOUNT);
829 if (!p) {
830 ret = -ENOMEM;
831 goto fail;
832 }
833
834 INIT_LIST_HEAD(&p->list);
835 p->addr = args->addr;
836 p->bus_idx = bus_idx;
837 p->length = args->len;
838 p->eventfd = eventfd;
839
840 /* The datamatch feature is optional, otherwise this is a wildcard */
841 if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
842 p->datamatch = args->datamatch;
843 else
844 p->wildcard = true;
845
846 mutex_lock(&kvm->slots_lock);
847
848 /* Verify that there isn't a match already */
849 if (ioeventfd_check_collision(kvm, p)) {
850 ret = -EEXIST;
851 goto unlock_fail;
852 }
853
854 kvm_iodevice_init(&p->dev, &ioeventfd_ops);
855
856 ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
857 &p->dev);
858 if (ret < 0)
859 goto unlock_fail;
860
861 kvm_get_bus(kvm, bus_idx)->ioeventfd_count++;
862 list_add_tail(&p->list, &kvm->ioeventfds);
863
864 mutex_unlock(&kvm->slots_lock);
865
866 return 0;
867
868unlock_fail:
869 mutex_unlock(&kvm->slots_lock);
870 kfree(p);
871
872fail:
873 eventfd_ctx_put(eventfd);
874
875 return ret;
876}
877
878static int
879kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
880 struct kvm_ioeventfd *args)
881{
882 struct _ioeventfd *p;
883 struct eventfd_ctx *eventfd;
884 struct kvm_io_bus *bus;
885 int ret = -ENOENT;
886 bool wildcard;
887
888 eventfd = eventfd_ctx_fdget(args->fd);
889 if (IS_ERR(eventfd))
890 return PTR_ERR(eventfd);
891
892 wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
893
894 mutex_lock(&kvm->slots_lock);
895
896 list_for_each_entry(p, &kvm->ioeventfds, list) {
897 if (p->bus_idx != bus_idx ||
898 p->eventfd != eventfd ||
899 p->addr != args->addr ||
900 p->length != args->len ||
901 p->wildcard != wildcard)
902 continue;
903
904 if (!p->wildcard && p->datamatch != args->datamatch)
905 continue;
906
907 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
908 bus = kvm_get_bus(kvm, bus_idx);
909 if (bus)
910 bus->ioeventfd_count--;
911 ret = 0;
912 break;
913 }
914
915 mutex_unlock(&kvm->slots_lock);
916
917 eventfd_ctx_put(eventfd);
918
919 return ret;
920}
921
922static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
923{
924 enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
925 int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
926
927 if (!args->len && bus_idx == KVM_MMIO_BUS)
928 kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
929
930 return ret;
931}
932
933static int
934kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
935{
936 enum kvm_bus bus_idx;
937 int ret;
938
939 bus_idx = ioeventfd_bus_from_flags(args->flags);
940 /* must be natural-word sized, or 0 to ignore length */
941 switch (args->len) {
942 case 0:
943 case 1:
944 case 2:
945 case 4:
946 case 8:
947 break;
948 default:
949 return -EINVAL;
950 }
951
952 /* check for range overflow */
953 if (args->addr + args->len < args->addr)
954 return -EINVAL;
955
956 /* check for extra flags that we don't understand */
957 if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
958 return -EINVAL;
959
960 /* ioeventfd with no length can't be combined with DATAMATCH */
961 if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
962 return -EINVAL;
963
964 ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
965 if (ret)
966 goto fail;
967
968 /* When length is ignored, MMIO is also put on a separate bus, for
969 * faster lookups.
970 */
971 if (!args->len && bus_idx == KVM_MMIO_BUS) {
972 ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
973 if (ret < 0)
974 goto fast_fail;
975 }
976
977 return 0;
978
979fast_fail:
980 kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
981fail:
982 return ret;
983}
984
985int
986kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
987{
988 if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
989 return kvm_deassign_ioeventfd(kvm, args);
990
991 return kvm_assign_ioeventfd(kvm, args);
992}
993
994void
995kvm_eventfd_init(struct kvm *kvm)
996{
997#ifdef CONFIG_HAVE_KVM_IRQCHIP
998 spin_lock_init(&kvm->irqfds.lock);
999 INIT_LIST_HEAD(&kvm->irqfds.items);
1000 INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
1001 mutex_init(&kvm->irqfds.resampler_lock);
1002#endif
1003 INIT_LIST_HEAD(&kvm->ioeventfds);
1004}