Linux Audio

Check our new training course

Open-source upstreaming

Need help get the support for your hardware in upstream Linux?
Loading...
v5.4
  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_IRQFD
 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
 58/*
 59 * Since resampler irqfds share an IRQ source ID, we de-assert once
 60 * then notify all of the resampler irqfds using this GSI.  We can't
 61 * do multiple de-asserts or we risk racing with incoming re-asserts.
 62 */
 63static void
 64irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
 65{
 66	struct kvm_kernel_irqfd_resampler *resampler;
 67	struct kvm *kvm;
 68	struct kvm_kernel_irqfd *irqfd;
 69	int idx;
 70
 71	resampler = container_of(kian,
 72			struct kvm_kernel_irqfd_resampler, notifier);
 73	kvm = resampler->kvm;
 74
 75	kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 76		    resampler->notifier.gsi, 0, false);
 77
 78	idx = srcu_read_lock(&kvm->irq_srcu);
 79
 80	list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
 
 81		eventfd_signal(irqfd->resamplefd, 1);
 82
 83	srcu_read_unlock(&kvm->irq_srcu, idx);
 84}
 85
 86static void
 87irqfd_resampler_shutdown(struct kvm_kernel_irqfd *irqfd)
 88{
 89	struct kvm_kernel_irqfd_resampler *resampler = irqfd->resampler;
 90	struct kvm *kvm = resampler->kvm;
 91
 92	mutex_lock(&kvm->irqfds.resampler_lock);
 93
 94	list_del_rcu(&irqfd->resampler_link);
 95	synchronize_srcu(&kvm->irq_srcu);
 96
 97	if (list_empty(&resampler->list)) {
 98		list_del(&resampler->link);
 99		kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
100		kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
101			    resampler->notifier.gsi, 0, false);
102		kfree(resampler);
103	}
104
105	mutex_unlock(&kvm->irqfds.resampler_lock);
106}
107
108/*
109 * Race-free decouple logic (ordering is critical)
110 */
111static void
112irqfd_shutdown(struct work_struct *work)
113{
114	struct kvm_kernel_irqfd *irqfd =
115		container_of(work, struct kvm_kernel_irqfd, shutdown);
116	struct kvm *kvm = irqfd->kvm;
117	u64 cnt;
118
119	/* Make sure irqfd has been initalized in assign path. */
120	synchronize_srcu(&kvm->irq_srcu);
121
122	/*
123	 * Synchronize with the wait-queue and unhook ourselves to prevent
124	 * further events.
125	 */
126	eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
127
128	/*
129	 * We know no new events will be scheduled at this point, so block
130	 * until all previously outstanding events have completed
131	 */
132	flush_work(&irqfd->inject);
133
134	if (irqfd->resampler) {
135		irqfd_resampler_shutdown(irqfd);
136		eventfd_ctx_put(irqfd->resamplefd);
137	}
138
139	/*
140	 * It is now safe to release the object's resources
141	 */
142#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
143	irq_bypass_unregister_consumer(&irqfd->consumer);
144#endif
145	eventfd_ctx_put(irqfd->eventfd);
146	kfree(irqfd);
147}
148
149
150/* assumes kvm->irqfds.lock is held */
151static bool
152irqfd_is_active(struct kvm_kernel_irqfd *irqfd)
153{
154	return list_empty(&irqfd->list) ? false : true;
155}
156
157/*
158 * Mark the irqfd as inactive and schedule it for removal
159 *
160 * assumes kvm->irqfds.lock is held
161 */
162static void
163irqfd_deactivate(struct kvm_kernel_irqfd *irqfd)
164{
165	BUG_ON(!irqfd_is_active(irqfd));
166
167	list_del_init(&irqfd->list);
168
169	queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
170}
171
172int __attribute__((weak)) kvm_arch_set_irq_inatomic(
173				struct kvm_kernel_irq_routing_entry *irq,
174				struct kvm *kvm, int irq_source_id,
175				int level,
176				bool line_status)
177{
178	return -EWOULDBLOCK;
179}
180
181/*
182 * Called with wqh->lock held and interrupts disabled
183 */
184static int
185irqfd_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
186{
187	struct kvm_kernel_irqfd *irqfd =
188		container_of(wait, struct kvm_kernel_irqfd, wait);
189	__poll_t flags = key_to_poll(key);
190	struct kvm_kernel_irq_routing_entry irq;
191	struct kvm *kvm = irqfd->kvm;
192	unsigned seq;
193	int idx;
 
194
195	if (flags & EPOLLIN) {
 
 
 
196		idx = srcu_read_lock(&kvm->irq_srcu);
197		do {
198			seq = read_seqcount_begin(&irqfd->irq_entry_sc);
199			irq = irqfd->irq_entry;
200		} while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
201		/* An event has been signaled, inject an interrupt */
202		if (kvm_arch_set_irq_inatomic(&irq, kvm,
203					      KVM_USERSPACE_IRQ_SOURCE_ID, 1,
204					      false) == -EWOULDBLOCK)
205			schedule_work(&irqfd->inject);
206		srcu_read_unlock(&kvm->irq_srcu, idx);
 
207	}
208
209	if (flags & EPOLLHUP) {
210		/* The eventfd is closing, detach from KVM */
211		unsigned long iflags;
212
213		spin_lock_irqsave(&kvm->irqfds.lock, iflags);
214
215		/*
216		 * We must check if someone deactivated the irqfd before
217		 * we could acquire the irqfds.lock since the item is
218		 * deactivated from the KVM side before it is unhooked from
219		 * the wait-queue.  If it is already deactivated, we can
220		 * simply return knowing the other side will cleanup for us.
221		 * We cannot race against the irqfd going away since the
222		 * other side is required to acquire wqh->lock, which we hold
223		 */
224		if (irqfd_is_active(irqfd))
225			irqfd_deactivate(irqfd);
226
227		spin_unlock_irqrestore(&kvm->irqfds.lock, iflags);
228	}
229
230	return 0;
231}
232
233static void
234irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
235			poll_table *pt)
236{
237	struct kvm_kernel_irqfd *irqfd =
238		container_of(pt, struct kvm_kernel_irqfd, pt);
239	add_wait_queue(wqh, &irqfd->wait);
240}
241
242/* Must be called under irqfds.lock */
243static void irqfd_update(struct kvm *kvm, struct kvm_kernel_irqfd *irqfd)
244{
245	struct kvm_kernel_irq_routing_entry *e;
246	struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
247	int n_entries;
248
249	n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
250
251	write_seqcount_begin(&irqfd->irq_entry_sc);
252
253	e = entries;
254	if (n_entries == 1)
255		irqfd->irq_entry = *e;
256	else
257		irqfd->irq_entry.type = 0;
258
259	write_seqcount_end(&irqfd->irq_entry_sc);
260}
261
262#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
263void __attribute__((weak)) kvm_arch_irq_bypass_stop(
264				struct irq_bypass_consumer *cons)
265{
266}
267
268void __attribute__((weak)) kvm_arch_irq_bypass_start(
269				struct irq_bypass_consumer *cons)
270{
271}
272
273int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
274				struct kvm *kvm, unsigned int host_irq,
275				uint32_t guest_irq, bool set)
276{
277	return 0;
278}
 
 
 
 
 
 
 
279#endif
280
281static int
282kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
283{
284	struct kvm_kernel_irqfd *irqfd, *tmp;
285	struct fd f;
286	struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
287	int ret;
288	__poll_t events;
289	int idx;
290
291	if (!kvm_arch_intc_initialized(kvm))
292		return -EAGAIN;
293
294	if (!kvm_arch_irqfd_allowed(kvm, args))
295		return -EINVAL;
296
297	irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL_ACCOUNT);
298	if (!irqfd)
299		return -ENOMEM;
300
301	irqfd->kvm = kvm;
302	irqfd->gsi = args->gsi;
303	INIT_LIST_HEAD(&irqfd->list);
304	INIT_WORK(&irqfd->inject, irqfd_inject);
305	INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
306	seqcount_init(&irqfd->irq_entry_sc);
307
308	f = fdget(args->fd);
309	if (!f.file) {
310		ret = -EBADF;
311		goto out;
312	}
313
314	eventfd = eventfd_ctx_fileget(f.file);
315	if (IS_ERR(eventfd)) {
316		ret = PTR_ERR(eventfd);
317		goto fail;
318	}
319
320	irqfd->eventfd = eventfd;
321
322	if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
323		struct kvm_kernel_irqfd_resampler *resampler;
324
325		resamplefd = eventfd_ctx_fdget(args->resamplefd);
326		if (IS_ERR(resamplefd)) {
327			ret = PTR_ERR(resamplefd);
328			goto fail;
329		}
330
331		irqfd->resamplefd = resamplefd;
332		INIT_LIST_HEAD(&irqfd->resampler_link);
333
334		mutex_lock(&kvm->irqfds.resampler_lock);
335
336		list_for_each_entry(resampler,
337				    &kvm->irqfds.resampler_list, link) {
338			if (resampler->notifier.gsi == irqfd->gsi) {
339				irqfd->resampler = resampler;
340				break;
341			}
342		}
343
344		if (!irqfd->resampler) {
345			resampler = kzalloc(sizeof(*resampler),
346					    GFP_KERNEL_ACCOUNT);
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_srcu(&kvm->irq_srcu);
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	idx = srcu_read_lock(&kvm->irq_srcu);
391	irqfd_update(kvm, irqfd);
392
393	list_add_tail(&irqfd->list, &kvm->irqfds.items);
394
395	spin_unlock_irq(&kvm->irqfds.lock);
396
397	/*
398	 * Check if there was an event already pending on the eventfd
399	 * before we registered, and trigger it as if we didn't miss it.
400	 */
401	events = vfs_poll(f.file, &irqfd->pt);
402
403	if (events & EPOLLIN)
404		schedule_work(&irqfd->inject);
405
406#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
407	if (kvm_arch_has_irq_bypass()) {
408		irqfd->consumer.token = (void *)irqfd->eventfd;
409		irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
410		irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
411		irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
412		irqfd->consumer.start = kvm_arch_irq_bypass_start;
413		ret = irq_bypass_register_consumer(&irqfd->consumer);
414		if (ret)
415			pr_info("irq bypass consumer (token %p) registration fails: %d\n",
416				irqfd->consumer.token, ret);
417	}
418#endif
419
420	srcu_read_unlock(&kvm->irq_srcu, idx);
421
422	/*
423	 * do not drop the file until the irqfd is fully initialized, otherwise
424	 * we might race against the EPOLLHUP
425	 */
426	fdput(f);
427	return 0;
428
429fail:
430	if (irqfd->resampler)
431		irqfd_resampler_shutdown(irqfd);
432
433	if (resamplefd && !IS_ERR(resamplefd))
434		eventfd_ctx_put(resamplefd);
435
436	if (eventfd && !IS_ERR(eventfd))
437		eventfd_ctx_put(eventfd);
438
439	fdput(f);
440
441out:
442	kfree(irqfd);
443	return ret;
444}
445
446bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
447{
448	struct kvm_irq_ack_notifier *kian;
449	int gsi, idx;
450
451	idx = srcu_read_lock(&kvm->irq_srcu);
452	gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
453	if (gsi != -1)
454		hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
455					 link)
456			if (kian->gsi == gsi) {
457				srcu_read_unlock(&kvm->irq_srcu, idx);
458				return true;
459			}
460
461	srcu_read_unlock(&kvm->irq_srcu, idx);
462
463	return false;
464}
465EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
466
467void kvm_notify_acked_gsi(struct kvm *kvm, int gsi)
468{
469	struct kvm_irq_ack_notifier *kian;
470
471	hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
472				 link)
473		if (kian->gsi == gsi)
474			kian->irq_acked(kian);
475}
476
477void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
478{
479	int gsi, idx;
480
481	trace_kvm_ack_irq(irqchip, pin);
482
483	idx = srcu_read_lock(&kvm->irq_srcu);
484	gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
485	if (gsi != -1)
486		kvm_notify_acked_gsi(kvm, gsi);
487	srcu_read_unlock(&kvm->irq_srcu, idx);
488}
489
490void kvm_register_irq_ack_notifier(struct kvm *kvm,
491				   struct kvm_irq_ack_notifier *kian)
492{
493	mutex_lock(&kvm->irq_lock);
494	hlist_add_head_rcu(&kian->link, &kvm->irq_ack_notifier_list);
495	mutex_unlock(&kvm->irq_lock);
496	kvm_arch_post_irq_ack_notifier_list_update(kvm);
497}
498
499void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
500				    struct kvm_irq_ack_notifier *kian)
501{
502	mutex_lock(&kvm->irq_lock);
503	hlist_del_init_rcu(&kian->link);
504	mutex_unlock(&kvm->irq_lock);
505	synchronize_srcu(&kvm->irq_srcu);
506	kvm_arch_post_irq_ack_notifier_list_update(kvm);
507}
508#endif
509
510void
511kvm_eventfd_init(struct kvm *kvm)
512{
513#ifdef CONFIG_HAVE_KVM_IRQFD
514	spin_lock_init(&kvm->irqfds.lock);
515	INIT_LIST_HEAD(&kvm->irqfds.items);
516	INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
517	mutex_init(&kvm->irqfds.resampler_lock);
518#endif
519	INIT_LIST_HEAD(&kvm->ioeventfds);
520}
521
522#ifdef CONFIG_HAVE_KVM_IRQFD
523/*
524 * shutdown any irqfd's that match fd+gsi
525 */
526static int
527kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
528{
529	struct kvm_kernel_irqfd *irqfd, *tmp;
530	struct eventfd_ctx *eventfd;
531
532	eventfd = eventfd_ctx_fdget(args->fd);
533	if (IS_ERR(eventfd))
534		return PTR_ERR(eventfd);
535
536	spin_lock_irq(&kvm->irqfds.lock);
537
538	list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
539		if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
540			/*
541			 * This clearing of irq_entry.type is needed for when
542			 * another thread calls kvm_irq_routing_update before
543			 * we flush workqueue below (we synchronize with
544			 * kvm_irq_routing_update using irqfds.lock).
545			 */
546			write_seqcount_begin(&irqfd->irq_entry_sc);
547			irqfd->irq_entry.type = 0;
548			write_seqcount_end(&irqfd->irq_entry_sc);
549			irqfd_deactivate(irqfd);
550		}
551	}
552
553	spin_unlock_irq(&kvm->irqfds.lock);
554	eventfd_ctx_put(eventfd);
555
556	/*
557	 * Block until we know all outstanding shutdown jobs have completed
558	 * so that we guarantee there will not be any more interrupts on this
559	 * gsi once this deassign function returns.
560	 */
561	flush_workqueue(irqfd_cleanup_wq);
562
563	return 0;
564}
565
566int
567kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
568{
569	if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
570		return -EINVAL;
571
572	if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
573		return kvm_irqfd_deassign(kvm, args);
574
575	return kvm_irqfd_assign(kvm, args);
576}
577
578/*
579 * This function is called as the kvm VM fd is being released. Shutdown all
580 * irqfds that still remain open
581 */
582void
583kvm_irqfd_release(struct kvm *kvm)
584{
585	struct kvm_kernel_irqfd *irqfd, *tmp;
586
587	spin_lock_irq(&kvm->irqfds.lock);
588
589	list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
590		irqfd_deactivate(irqfd);
591
592	spin_unlock_irq(&kvm->irqfds.lock);
593
594	/*
595	 * Block until we know all outstanding shutdown jobs have completed
596	 * since we do not take a kvm* reference.
597	 */
598	flush_workqueue(irqfd_cleanup_wq);
599
600}
601
602/*
603 * Take note of a change in irq routing.
604 * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
605 */
606void kvm_irq_routing_update(struct kvm *kvm)
607{
608	struct kvm_kernel_irqfd *irqfd;
609
610	spin_lock_irq(&kvm->irqfds.lock);
611
612	list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
 
 
 
 
 
613		irqfd_update(kvm, irqfd);
614
615#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
616		if (irqfd->producer) {
 
617			int ret = kvm_arch_update_irqfd_routing(
618					irqfd->kvm, irqfd->producer->irq,
619					irqfd->gsi, 1);
620			WARN_ON(ret);
621		}
622#endif
623	}
624
625	spin_unlock_irq(&kvm->irqfds.lock);
626}
627
628/*
629 * create a host-wide workqueue for issuing deferred shutdown requests
630 * aggregated from all vm* instances. We need our own isolated
631 * queue to ease flushing work items when a VM exits.
632 */
633int kvm_irqfd_init(void)
634{
635	irqfd_cleanup_wq = alloc_workqueue("kvm-irqfd-cleanup", 0, 0);
636	if (!irqfd_cleanup_wq)
637		return -ENOMEM;
638
639	return 0;
640}
641
642void kvm_irqfd_exit(void)
643{
644	destroy_workqueue(irqfd_cleanup_wq);
645}
646#endif
647
648/*
649 * --------------------------------------------------------------------
650 * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
651 *
652 * userspace can register a PIO/MMIO address with an eventfd for receiving
653 * notification when the memory has been touched.
654 * --------------------------------------------------------------------
655 */
656
657struct _ioeventfd {
658	struct list_head     list;
659	u64                  addr;
660	int                  length;
661	struct eventfd_ctx  *eventfd;
662	u64                  datamatch;
663	struct kvm_io_device dev;
664	u8                   bus_idx;
665	bool                 wildcard;
666};
667
668static inline struct _ioeventfd *
669to_ioeventfd(struct kvm_io_device *dev)
670{
671	return container_of(dev, struct _ioeventfd, dev);
672}
673
674static void
675ioeventfd_release(struct _ioeventfd *p)
676{
677	eventfd_ctx_put(p->eventfd);
678	list_del(&p->list);
679	kfree(p);
680}
681
682static bool
683ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
684{
685	u64 _val;
686
687	if (addr != p->addr)
688		/* address must be precise for a hit */
689		return false;
690
691	if (!p->length)
692		/* length = 0 means only look at the address, so always a hit */
693		return true;
694
695	if (len != p->length)
696		/* address-range must be precise for a hit */
697		return false;
698
699	if (p->wildcard)
700		/* all else equal, wildcard is always a hit */
701		return true;
702
703	/* otherwise, we have to actually compare the data */
704
705	BUG_ON(!IS_ALIGNED((unsigned long)val, len));
706
707	switch (len) {
708	case 1:
709		_val = *(u8 *)val;
710		break;
711	case 2:
712		_val = *(u16 *)val;
713		break;
714	case 4:
715		_val = *(u32 *)val;
716		break;
717	case 8:
718		_val = *(u64 *)val;
719		break;
720	default:
721		return false;
722	}
723
724	return _val == p->datamatch ? true : false;
725}
726
727/* MMIO/PIO writes trigger an event if the addr/val match */
728static int
729ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
730		int len, const void *val)
731{
732	struct _ioeventfd *p = to_ioeventfd(this);
733
734	if (!ioeventfd_in_range(p, addr, len, val))
735		return -EOPNOTSUPP;
736
737	eventfd_signal(p->eventfd, 1);
738	return 0;
739}
740
741/*
742 * This function is called as KVM is completely shutting down.  We do not
743 * need to worry about locking just nuke anything we have as quickly as possible
744 */
745static void
746ioeventfd_destructor(struct kvm_io_device *this)
747{
748	struct _ioeventfd *p = to_ioeventfd(this);
749
750	ioeventfd_release(p);
751}
752
753static const struct kvm_io_device_ops ioeventfd_ops = {
754	.write      = ioeventfd_write,
755	.destructor = ioeventfd_destructor,
756};
757
758/* assumes kvm->slots_lock held */
759static bool
760ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
761{
762	struct _ioeventfd *_p;
763
764	list_for_each_entry(_p, &kvm->ioeventfds, list)
765		if (_p->bus_idx == p->bus_idx &&
766		    _p->addr == p->addr &&
767		    (!_p->length || !p->length ||
768		     (_p->length == p->length &&
769		      (_p->wildcard || p->wildcard ||
770		       _p->datamatch == p->datamatch))))
771			return true;
772
773	return false;
774}
775
776static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
777{
778	if (flags & KVM_IOEVENTFD_FLAG_PIO)
779		return KVM_PIO_BUS;
780	if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
781		return KVM_VIRTIO_CCW_NOTIFY_BUS;
782	return KVM_MMIO_BUS;
783}
784
785static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
786				enum kvm_bus bus_idx,
787				struct kvm_ioeventfd *args)
788{
789
790	struct eventfd_ctx *eventfd;
791	struct _ioeventfd *p;
792	int ret;
793
794	eventfd = eventfd_ctx_fdget(args->fd);
795	if (IS_ERR(eventfd))
796		return PTR_ERR(eventfd);
797
798	p = kzalloc(sizeof(*p), GFP_KERNEL_ACCOUNT);
799	if (!p) {
800		ret = -ENOMEM;
801		goto fail;
802	}
803
804	INIT_LIST_HEAD(&p->list);
805	p->addr    = args->addr;
806	p->bus_idx = bus_idx;
807	p->length  = args->len;
808	p->eventfd = eventfd;
809
810	/* The datamatch feature is optional, otherwise this is a wildcard */
811	if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
812		p->datamatch = args->datamatch;
813	else
814		p->wildcard = true;
815
816	mutex_lock(&kvm->slots_lock);
817
818	/* Verify that there isn't a match already */
819	if (ioeventfd_check_collision(kvm, p)) {
820		ret = -EEXIST;
821		goto unlock_fail;
822	}
823
824	kvm_iodevice_init(&p->dev, &ioeventfd_ops);
825
826	ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
827				      &p->dev);
828	if (ret < 0)
829		goto unlock_fail;
830
831	kvm_get_bus(kvm, bus_idx)->ioeventfd_count++;
832	list_add_tail(&p->list, &kvm->ioeventfds);
833
834	mutex_unlock(&kvm->slots_lock);
835
836	return 0;
837
838unlock_fail:
839	mutex_unlock(&kvm->slots_lock);
840
841fail:
842	kfree(p);
843	eventfd_ctx_put(eventfd);
844
845	return ret;
846}
847
848static int
849kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
850			   struct kvm_ioeventfd *args)
851{
852	struct _ioeventfd        *p, *tmp;
853	struct eventfd_ctx       *eventfd;
854	struct kvm_io_bus	 *bus;
855	int                       ret = -ENOENT;
 
856
857	eventfd = eventfd_ctx_fdget(args->fd);
858	if (IS_ERR(eventfd))
859		return PTR_ERR(eventfd);
860
 
 
861	mutex_lock(&kvm->slots_lock);
862
863	list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
864		bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
865
866		if (p->bus_idx != bus_idx ||
867		    p->eventfd != eventfd  ||
868		    p->addr != args->addr  ||
869		    p->length != args->len ||
870		    p->wildcard != wildcard)
871			continue;
872
873		if (!p->wildcard && p->datamatch != args->datamatch)
874			continue;
875
876		kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
877		bus = kvm_get_bus(kvm, bus_idx);
878		if (bus)
879			bus->ioeventfd_count--;
880		ioeventfd_release(p);
881		ret = 0;
882		break;
883	}
884
885	mutex_unlock(&kvm->slots_lock);
886
887	eventfd_ctx_put(eventfd);
888
889	return ret;
890}
891
892static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
893{
894	enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
895	int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
896
897	if (!args->len && bus_idx == KVM_MMIO_BUS)
898		kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
899
900	return ret;
901}
902
903static int
904kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
905{
906	enum kvm_bus              bus_idx;
907	int ret;
908
909	bus_idx = ioeventfd_bus_from_flags(args->flags);
910	/* must be natural-word sized, or 0 to ignore length */
911	switch (args->len) {
912	case 0:
913	case 1:
914	case 2:
915	case 4:
916	case 8:
917		break;
918	default:
919		return -EINVAL;
920	}
921
922	/* check for range overflow */
923	if (args->addr + args->len < args->addr)
924		return -EINVAL;
925
926	/* check for extra flags that we don't understand */
927	if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
928		return -EINVAL;
929
930	/* ioeventfd with no length can't be combined with DATAMATCH */
931	if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
932		return -EINVAL;
933
934	ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
935	if (ret)
936		goto fail;
937
938	/* When length is ignored, MMIO is also put on a separate bus, for
939	 * faster lookups.
940	 */
941	if (!args->len && bus_idx == KVM_MMIO_BUS) {
942		ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
943		if (ret < 0)
944			goto fast_fail;
945	}
946
947	return 0;
948
949fast_fail:
950	kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
951fail:
952	return ret;
953}
954
955int
956kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
957{
958	if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
959		return kvm_deassign_ioeventfd(kvm, args);
960
961	return kvm_assign_ioeventfd(kvm, args);
962}
v6.2
  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_IRQFD
 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
 58/*
 59 * Since resampler irqfds share an IRQ source ID, we de-assert once
 60 * then notify all of the resampler irqfds using this GSI.  We can't
 61 * do multiple de-asserts or we risk racing with incoming re-asserts.
 62 */
 63static void
 64irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
 65{
 66	struct kvm_kernel_irqfd_resampler *resampler;
 67	struct kvm *kvm;
 68	struct kvm_kernel_irqfd *irqfd;
 69	int idx;
 70
 71	resampler = container_of(kian,
 72			struct kvm_kernel_irqfd_resampler, notifier);
 73	kvm = resampler->kvm;
 74
 75	kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 76		    resampler->notifier.gsi, 0, false);
 77
 78	idx = srcu_read_lock(&kvm->irq_srcu);
 79
 80	list_for_each_entry_srcu(irqfd, &resampler->list, resampler_link,
 81	    srcu_read_lock_held(&kvm->irq_srcu))
 82		eventfd_signal(irqfd->resamplefd, 1);
 83
 84	srcu_read_unlock(&kvm->irq_srcu, idx);
 85}
 86
 87static void
 88irqfd_resampler_shutdown(struct kvm_kernel_irqfd *irqfd)
 89{
 90	struct kvm_kernel_irqfd_resampler *resampler = irqfd->resampler;
 91	struct kvm *kvm = resampler->kvm;
 92
 93	mutex_lock(&kvm->irqfds.resampler_lock);
 94
 95	list_del_rcu(&irqfd->resampler_link);
 96	synchronize_srcu(&kvm->irq_srcu);
 97
 98	if (list_empty(&resampler->list)) {
 99		list_del(&resampler->link);
100		kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
101		kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
102			    resampler->notifier.gsi, 0, false);
103		kfree(resampler);
104	}
105
106	mutex_unlock(&kvm->irqfds.resampler_lock);
107}
108
109/*
110 * Race-free decouple logic (ordering is critical)
111 */
112static void
113irqfd_shutdown(struct work_struct *work)
114{
115	struct kvm_kernel_irqfd *irqfd =
116		container_of(work, struct kvm_kernel_irqfd, shutdown);
117	struct kvm *kvm = irqfd->kvm;
118	u64 cnt;
119
120	/* Make sure irqfd has been initialized in assign path. */
121	synchronize_srcu(&kvm->irq_srcu);
122
123	/*
124	 * Synchronize with the wait-queue and unhook ourselves to prevent
125	 * further events.
126	 */
127	eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
128
129	/*
130	 * We know no new events will be scheduled at this point, so block
131	 * until all previously outstanding events have completed
132	 */
133	flush_work(&irqfd->inject);
134
135	if (irqfd->resampler) {
136		irqfd_resampler_shutdown(irqfd);
137		eventfd_ctx_put(irqfd->resamplefd);
138	}
139
140	/*
141	 * It is now safe to release the object's resources
142	 */
143#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
144	irq_bypass_unregister_consumer(&irqfd->consumer);
145#endif
146	eventfd_ctx_put(irqfd->eventfd);
147	kfree(irqfd);
148}
149
150
151/* assumes kvm->irqfds.lock is held */
152static bool
153irqfd_is_active(struct kvm_kernel_irqfd *irqfd)
154{
155	return list_empty(&irqfd->list) ? false : true;
156}
157
158/*
159 * Mark the irqfd as inactive and schedule it for removal
160 *
161 * assumes kvm->irqfds.lock is held
162 */
163static void
164irqfd_deactivate(struct kvm_kernel_irqfd *irqfd)
165{
166	BUG_ON(!irqfd_is_active(irqfd));
167
168	list_del_init(&irqfd->list);
169
170	queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
171}
172
173int __attribute__((weak)) kvm_arch_set_irq_inatomic(
174				struct kvm_kernel_irq_routing_entry *irq,
175				struct kvm *kvm, int irq_source_id,
176				int level,
177				bool line_status)
178{
179	return -EWOULDBLOCK;
180}
181
182/*
183 * Called with wqh->lock held and interrupts disabled
184 */
185static int
186irqfd_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
187{
188	struct kvm_kernel_irqfd *irqfd =
189		container_of(wait, struct kvm_kernel_irqfd, wait);
190	__poll_t flags = key_to_poll(key);
191	struct kvm_kernel_irq_routing_entry irq;
192	struct kvm *kvm = irqfd->kvm;
193	unsigned seq;
194	int idx;
195	int ret = 0;
196
197	if (flags & EPOLLIN) {
198		u64 cnt;
199		eventfd_ctx_do_read(irqfd->eventfd, &cnt);
200
201		idx = srcu_read_lock(&kvm->irq_srcu);
202		do {
203			seq = read_seqcount_begin(&irqfd->irq_entry_sc);
204			irq = irqfd->irq_entry;
205		} while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
206		/* An event has been signaled, inject an interrupt */
207		if (kvm_arch_set_irq_inatomic(&irq, kvm,
208					      KVM_USERSPACE_IRQ_SOURCE_ID, 1,
209					      false) == -EWOULDBLOCK)
210			schedule_work(&irqfd->inject);
211		srcu_read_unlock(&kvm->irq_srcu, idx);
212		ret = 1;
213	}
214
215	if (flags & EPOLLHUP) {
216		/* The eventfd is closing, detach from KVM */
217		unsigned long iflags;
218
219		spin_lock_irqsave(&kvm->irqfds.lock, iflags);
220
221		/*
222		 * We must check if someone deactivated the irqfd before
223		 * we could acquire the irqfds.lock since the item is
224		 * deactivated from the KVM side before it is unhooked from
225		 * the wait-queue.  If it is already deactivated, we can
226		 * simply return knowing the other side will cleanup for us.
227		 * We cannot race against the irqfd going away since the
228		 * other side is required to acquire wqh->lock, which we hold
229		 */
230		if (irqfd_is_active(irqfd))
231			irqfd_deactivate(irqfd);
232
233		spin_unlock_irqrestore(&kvm->irqfds.lock, iflags);
234	}
235
236	return ret;
237}
238
239static void
240irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
241			poll_table *pt)
242{
243	struct kvm_kernel_irqfd *irqfd =
244		container_of(pt, struct kvm_kernel_irqfd, pt);
245	add_wait_queue_priority(wqh, &irqfd->wait);
246}
247
248/* Must be called under irqfds.lock */
249static void irqfd_update(struct kvm *kvm, struct kvm_kernel_irqfd *irqfd)
250{
251	struct kvm_kernel_irq_routing_entry *e;
252	struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
253	int n_entries;
254
255	n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
256
257	write_seqcount_begin(&irqfd->irq_entry_sc);
258
259	e = entries;
260	if (n_entries == 1)
261		irqfd->irq_entry = *e;
262	else
263		irqfd->irq_entry.type = 0;
264
265	write_seqcount_end(&irqfd->irq_entry_sc);
266}
267
268#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
269void __attribute__((weak)) kvm_arch_irq_bypass_stop(
270				struct irq_bypass_consumer *cons)
271{
272}
273
274void __attribute__((weak)) kvm_arch_irq_bypass_start(
275				struct irq_bypass_consumer *cons)
276{
277}
278
279int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
280				struct kvm *kvm, unsigned int host_irq,
281				uint32_t guest_irq, bool set)
282{
283	return 0;
284}
285
286bool __attribute__((weak)) kvm_arch_irqfd_route_changed(
287				struct kvm_kernel_irq_routing_entry *old,
288				struct kvm_kernel_irq_routing_entry *new)
289{
290	return true;
291}
292#endif
293
294static int
295kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
296{
297	struct kvm_kernel_irqfd *irqfd, *tmp;
298	struct fd f;
299	struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
300	int ret;
301	__poll_t events;
302	int idx;
303
304	if (!kvm_arch_intc_initialized(kvm))
305		return -EAGAIN;
306
307	if (!kvm_arch_irqfd_allowed(kvm, args))
308		return -EINVAL;
309
310	irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL_ACCOUNT);
311	if (!irqfd)
312		return -ENOMEM;
313
314	irqfd->kvm = kvm;
315	irqfd->gsi = args->gsi;
316	INIT_LIST_HEAD(&irqfd->list);
317	INIT_WORK(&irqfd->inject, irqfd_inject);
318	INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
319	seqcount_spinlock_init(&irqfd->irq_entry_sc, &kvm->irqfds.lock);
320
321	f = fdget(args->fd);
322	if (!f.file) {
323		ret = -EBADF;
324		goto out;
325	}
326
327	eventfd = eventfd_ctx_fileget(f.file);
328	if (IS_ERR(eventfd)) {
329		ret = PTR_ERR(eventfd);
330		goto fail;
331	}
332
333	irqfd->eventfd = eventfd;
334
335	if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
336		struct kvm_kernel_irqfd_resampler *resampler;
337
338		resamplefd = eventfd_ctx_fdget(args->resamplefd);
339		if (IS_ERR(resamplefd)) {
340			ret = PTR_ERR(resamplefd);
341			goto fail;
342		}
343
344		irqfd->resamplefd = resamplefd;
345		INIT_LIST_HEAD(&irqfd->resampler_link);
346
347		mutex_lock(&kvm->irqfds.resampler_lock);
348
349		list_for_each_entry(resampler,
350				    &kvm->irqfds.resampler_list, link) {
351			if (resampler->notifier.gsi == irqfd->gsi) {
352				irqfd->resampler = resampler;
353				break;
354			}
355		}
356
357		if (!irqfd->resampler) {
358			resampler = kzalloc(sizeof(*resampler),
359					    GFP_KERNEL_ACCOUNT);
360			if (!resampler) {
361				ret = -ENOMEM;
362				mutex_unlock(&kvm->irqfds.resampler_lock);
363				goto fail;
364			}
365
366			resampler->kvm = kvm;
367			INIT_LIST_HEAD(&resampler->list);
368			resampler->notifier.gsi = irqfd->gsi;
369			resampler->notifier.irq_acked = irqfd_resampler_ack;
370			INIT_LIST_HEAD(&resampler->link);
371
372			list_add(&resampler->link, &kvm->irqfds.resampler_list);
373			kvm_register_irq_ack_notifier(kvm,
374						      &resampler->notifier);
375			irqfd->resampler = resampler;
376		}
377
378		list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
379		synchronize_srcu(&kvm->irq_srcu);
380
381		mutex_unlock(&kvm->irqfds.resampler_lock);
382	}
383
384	/*
385	 * Install our own custom wake-up handling so we are notified via
386	 * a callback whenever someone signals the underlying eventfd
387	 */
388	init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
389	init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
390
391	spin_lock_irq(&kvm->irqfds.lock);
392
393	ret = 0;
394	list_for_each_entry(tmp, &kvm->irqfds.items, list) {
395		if (irqfd->eventfd != tmp->eventfd)
396			continue;
397		/* This fd is used for another irq already. */
398		ret = -EBUSY;
399		spin_unlock_irq(&kvm->irqfds.lock);
400		goto fail;
401	}
402
403	idx = srcu_read_lock(&kvm->irq_srcu);
404	irqfd_update(kvm, irqfd);
405
406	list_add_tail(&irqfd->list, &kvm->irqfds.items);
407
408	spin_unlock_irq(&kvm->irqfds.lock);
409
410	/*
411	 * Check if there was an event already pending on the eventfd
412	 * before we registered, and trigger it as if we didn't miss it.
413	 */
414	events = vfs_poll(f.file, &irqfd->pt);
415
416	if (events & EPOLLIN)
417		schedule_work(&irqfd->inject);
418
419#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
420	if (kvm_arch_has_irq_bypass()) {
421		irqfd->consumer.token = (void *)irqfd->eventfd;
422		irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
423		irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
424		irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
425		irqfd->consumer.start = kvm_arch_irq_bypass_start;
426		ret = irq_bypass_register_consumer(&irqfd->consumer);
427		if (ret)
428			pr_info("irq bypass consumer (token %p) registration fails: %d\n",
429				irqfd->consumer.token, ret);
430	}
431#endif
432
433	srcu_read_unlock(&kvm->irq_srcu, idx);
434
435	/*
436	 * do not drop the file until the irqfd is fully initialized, otherwise
437	 * we might race against the EPOLLHUP
438	 */
439	fdput(f);
440	return 0;
441
442fail:
443	if (irqfd->resampler)
444		irqfd_resampler_shutdown(irqfd);
445
446	if (resamplefd && !IS_ERR(resamplefd))
447		eventfd_ctx_put(resamplefd);
448
449	if (eventfd && !IS_ERR(eventfd))
450		eventfd_ctx_put(eventfd);
451
452	fdput(f);
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(&kvm->irq_srcu);
519	kvm_arch_post_irq_ack_notifier_list_update(kvm);
520}
521#endif
522
523void
524kvm_eventfd_init(struct kvm *kvm)
525{
526#ifdef CONFIG_HAVE_KVM_IRQFD
527	spin_lock_init(&kvm->irqfds.lock);
528	INIT_LIST_HEAD(&kvm->irqfds.items);
529	INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
530	mutex_init(&kvm->irqfds.resampler_lock);
531#endif
532	INIT_LIST_HEAD(&kvm->ioeventfds);
533}
534
535#ifdef CONFIG_HAVE_KVM_IRQFD
536/*
537 * shutdown any irqfd's that match fd+gsi
538 */
539static int
540kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
541{
542	struct kvm_kernel_irqfd *irqfd, *tmp;
543	struct eventfd_ctx *eventfd;
544
545	eventfd = eventfd_ctx_fdget(args->fd);
546	if (IS_ERR(eventfd))
547		return PTR_ERR(eventfd);
548
549	spin_lock_irq(&kvm->irqfds.lock);
550
551	list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
552		if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
553			/*
554			 * This clearing of irq_entry.type is needed for when
555			 * another thread calls kvm_irq_routing_update before
556			 * we flush workqueue below (we synchronize with
557			 * kvm_irq_routing_update using irqfds.lock).
558			 */
559			write_seqcount_begin(&irqfd->irq_entry_sc);
560			irqfd->irq_entry.type = 0;
561			write_seqcount_end(&irqfd->irq_entry_sc);
562			irqfd_deactivate(irqfd);
563		}
564	}
565
566	spin_unlock_irq(&kvm->irqfds.lock);
567	eventfd_ctx_put(eventfd);
568
569	/*
570	 * Block until we know all outstanding shutdown jobs have completed
571	 * so that we guarantee there will not be any more interrupts on this
572	 * gsi once this deassign function returns.
573	 */
574	flush_workqueue(irqfd_cleanup_wq);
575
576	return 0;
577}
578
579int
580kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
581{
582	if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
583		return -EINVAL;
584
585	if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
586		return kvm_irqfd_deassign(kvm, args);
587
588	return kvm_irqfd_assign(kvm, args);
589}
590
591/*
592 * This function is called as the kvm VM fd is being released. Shutdown all
593 * irqfds that still remain open
594 */
595void
596kvm_irqfd_release(struct kvm *kvm)
597{
598	struct kvm_kernel_irqfd *irqfd, *tmp;
599
600	spin_lock_irq(&kvm->irqfds.lock);
601
602	list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
603		irqfd_deactivate(irqfd);
604
605	spin_unlock_irq(&kvm->irqfds.lock);
606
607	/*
608	 * Block until we know all outstanding shutdown jobs have completed
609	 * since we do not take a kvm* reference.
610	 */
611	flush_workqueue(irqfd_cleanup_wq);
612
613}
614
615/*
616 * Take note of a change in irq routing.
617 * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
618 */
619void kvm_irq_routing_update(struct kvm *kvm)
620{
621	struct kvm_kernel_irqfd *irqfd;
622
623	spin_lock_irq(&kvm->irqfds.lock);
624
625	list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
626#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
627		/* Under irqfds.lock, so can read irq_entry safely */
628		struct kvm_kernel_irq_routing_entry old = irqfd->irq_entry;
629#endif
630
631		irqfd_update(kvm, irqfd);
632
633#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
634		if (irqfd->producer &&
635		    kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry)) {
636			int ret = kvm_arch_update_irqfd_routing(
637					irqfd->kvm, irqfd->producer->irq,
638					irqfd->gsi, 1);
639			WARN_ON(ret);
640		}
641#endif
642	}
643
644	spin_unlock_irq(&kvm->irqfds.lock);
645}
646
647/*
648 * create a host-wide workqueue for issuing deferred shutdown requests
649 * aggregated from all vm* instances. We need our own isolated
650 * queue to ease flushing work items when a VM exits.
651 */
652int kvm_irqfd_init(void)
653{
654	irqfd_cleanup_wq = alloc_workqueue("kvm-irqfd-cleanup", 0, 0);
655	if (!irqfd_cleanup_wq)
656		return -ENOMEM;
657
658	return 0;
659}
660
661void kvm_irqfd_exit(void)
662{
663	destroy_workqueue(irqfd_cleanup_wq);
664}
665#endif
666
667/*
668 * --------------------------------------------------------------------
669 * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
670 *
671 * userspace can register a PIO/MMIO address with an eventfd for receiving
672 * notification when the memory has been touched.
673 * --------------------------------------------------------------------
674 */
675
676struct _ioeventfd {
677	struct list_head     list;
678	u64                  addr;
679	int                  length;
680	struct eventfd_ctx  *eventfd;
681	u64                  datamatch;
682	struct kvm_io_device dev;
683	u8                   bus_idx;
684	bool                 wildcard;
685};
686
687static inline struct _ioeventfd *
688to_ioeventfd(struct kvm_io_device *dev)
689{
690	return container_of(dev, struct _ioeventfd, dev);
691}
692
693static void
694ioeventfd_release(struct _ioeventfd *p)
695{
696	eventfd_ctx_put(p->eventfd);
697	list_del(&p->list);
698	kfree(p);
699}
700
701static bool
702ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
703{
704	u64 _val;
705
706	if (addr != p->addr)
707		/* address must be precise for a hit */
708		return false;
709
710	if (!p->length)
711		/* length = 0 means only look at the address, so always a hit */
712		return true;
713
714	if (len != p->length)
715		/* address-range must be precise for a hit */
716		return false;
717
718	if (p->wildcard)
719		/* all else equal, wildcard is always a hit */
720		return true;
721
722	/* otherwise, we have to actually compare the data */
723
724	BUG_ON(!IS_ALIGNED((unsigned long)val, len));
725
726	switch (len) {
727	case 1:
728		_val = *(u8 *)val;
729		break;
730	case 2:
731		_val = *(u16 *)val;
732		break;
733	case 4:
734		_val = *(u32 *)val;
735		break;
736	case 8:
737		_val = *(u64 *)val;
738		break;
739	default:
740		return false;
741	}
742
743	return _val == p->datamatch;
744}
745
746/* MMIO/PIO writes trigger an event if the addr/val match */
747static int
748ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
749		int len, const void *val)
750{
751	struct _ioeventfd *p = to_ioeventfd(this);
752
753	if (!ioeventfd_in_range(p, addr, len, val))
754		return -EOPNOTSUPP;
755
756	eventfd_signal(p->eventfd, 1);
757	return 0;
758}
759
760/*
761 * This function is called as KVM is completely shutting down.  We do not
762 * need to worry about locking just nuke anything we have as quickly as possible
763 */
764static void
765ioeventfd_destructor(struct kvm_io_device *this)
766{
767	struct _ioeventfd *p = to_ioeventfd(this);
768
769	ioeventfd_release(p);
770}
771
772static const struct kvm_io_device_ops ioeventfd_ops = {
773	.write      = ioeventfd_write,
774	.destructor = ioeventfd_destructor,
775};
776
777/* assumes kvm->slots_lock held */
778static bool
779ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
780{
781	struct _ioeventfd *_p;
782
783	list_for_each_entry(_p, &kvm->ioeventfds, list)
784		if (_p->bus_idx == p->bus_idx &&
785		    _p->addr == p->addr &&
786		    (!_p->length || !p->length ||
787		     (_p->length == p->length &&
788		      (_p->wildcard || p->wildcard ||
789		       _p->datamatch == p->datamatch))))
790			return true;
791
792	return false;
793}
794
795static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
796{
797	if (flags & KVM_IOEVENTFD_FLAG_PIO)
798		return KVM_PIO_BUS;
799	if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
800		return KVM_VIRTIO_CCW_NOTIFY_BUS;
801	return KVM_MMIO_BUS;
802}
803
804static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
805				enum kvm_bus bus_idx,
806				struct kvm_ioeventfd *args)
807{
808
809	struct eventfd_ctx *eventfd;
810	struct _ioeventfd *p;
811	int ret;
812
813	eventfd = eventfd_ctx_fdget(args->fd);
814	if (IS_ERR(eventfd))
815		return PTR_ERR(eventfd);
816
817	p = kzalloc(sizeof(*p), GFP_KERNEL_ACCOUNT);
818	if (!p) {
819		ret = -ENOMEM;
820		goto fail;
821	}
822
823	INIT_LIST_HEAD(&p->list);
824	p->addr    = args->addr;
825	p->bus_idx = bus_idx;
826	p->length  = args->len;
827	p->eventfd = eventfd;
828
829	/* The datamatch feature is optional, otherwise this is a wildcard */
830	if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
831		p->datamatch = args->datamatch;
832	else
833		p->wildcard = true;
834
835	mutex_lock(&kvm->slots_lock);
836
837	/* Verify that there isn't a match already */
838	if (ioeventfd_check_collision(kvm, p)) {
839		ret = -EEXIST;
840		goto unlock_fail;
841	}
842
843	kvm_iodevice_init(&p->dev, &ioeventfd_ops);
844
845	ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
846				      &p->dev);
847	if (ret < 0)
848		goto unlock_fail;
849
850	kvm_get_bus(kvm, bus_idx)->ioeventfd_count++;
851	list_add_tail(&p->list, &kvm->ioeventfds);
852
853	mutex_unlock(&kvm->slots_lock);
854
855	return 0;
856
857unlock_fail:
858	mutex_unlock(&kvm->slots_lock);
859
860fail:
861	kfree(p);
862	eventfd_ctx_put(eventfd);
863
864	return ret;
865}
866
867static int
868kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
869			   struct kvm_ioeventfd *args)
870{
871	struct _ioeventfd        *p, *tmp;
872	struct eventfd_ctx       *eventfd;
873	struct kvm_io_bus	 *bus;
874	int                       ret = -ENOENT;
875	bool                      wildcard;
876
877	eventfd = eventfd_ctx_fdget(args->fd);
878	if (IS_ERR(eventfd))
879		return PTR_ERR(eventfd);
880
881	wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
882
883	mutex_lock(&kvm->slots_lock);
884
885	list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 
886
887		if (p->bus_idx != bus_idx ||
888		    p->eventfd != eventfd  ||
889		    p->addr != args->addr  ||
890		    p->length != args->len ||
891		    p->wildcard != wildcard)
892			continue;
893
894		if (!p->wildcard && p->datamatch != args->datamatch)
895			continue;
896
897		kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
898		bus = kvm_get_bus(kvm, bus_idx);
899		if (bus)
900			bus->ioeventfd_count--;
901		ioeventfd_release(p);
902		ret = 0;
903		break;
904	}
905
906	mutex_unlock(&kvm->slots_lock);
907
908	eventfd_ctx_put(eventfd);
909
910	return ret;
911}
912
913static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
914{
915	enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
916	int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
917
918	if (!args->len && bus_idx == KVM_MMIO_BUS)
919		kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
920
921	return ret;
922}
923
924static int
925kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
926{
927	enum kvm_bus              bus_idx;
928	int ret;
929
930	bus_idx = ioeventfd_bus_from_flags(args->flags);
931	/* must be natural-word sized, or 0 to ignore length */
932	switch (args->len) {
933	case 0:
934	case 1:
935	case 2:
936	case 4:
937	case 8:
938		break;
939	default:
940		return -EINVAL;
941	}
942
943	/* check for range overflow */
944	if (args->addr + args->len < args->addr)
945		return -EINVAL;
946
947	/* check for extra flags that we don't understand */
948	if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
949		return -EINVAL;
950
951	/* ioeventfd with no length can't be combined with DATAMATCH */
952	if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
953		return -EINVAL;
954
955	ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
956	if (ret)
957		goto fail;
958
959	/* When length is ignored, MMIO is also put on a separate bus, for
960	 * faster lookups.
961	 */
962	if (!args->len && bus_idx == KVM_MMIO_BUS) {
963		ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
964		if (ret < 0)
965			goto fast_fail;
966	}
967
968	return 0;
969
970fast_fail:
971	kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
972fail:
973	return ret;
974}
975
976int
977kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
978{
979	if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
980		return kvm_deassign_ioeventfd(kvm, args);
981
982	return kvm_assign_ioeventfd(kvm, args);
983}