]> Gitweb @ Texas Instruments - Open Source Git Repositories - git.TI.com/gitweb - sitara-epos/sitara-epos-kernel.git/blob - virt/kvm/eventfd.c
am335x: lis33ldlh: Add accelerometer support to AM335X
[sitara-epos/sitara-epos-kernel.git] / virt / kvm / eventfd.c
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  */
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>
36 #include "iodev.h"
38 /*
39  * --------------------------------------------------------------------
40  * irqfd: Allows an fd to be used to inject an interrupt to the guest
41  *
42  * Credit goes to Avi Kivity for the original idea.
43  * --------------------------------------------------------------------
44  */
46 struct _irqfd {
47         /* Used for MSI fast-path */
48         struct kvm *kvm;
49         wait_queue_t wait;
50         /* Update side is protected by irqfds.lock */
51         struct kvm_kernel_irq_routing_entry __rcu *irq_entry;
52         /* Used for level IRQ fast-path */
53         int gsi;
54         struct work_struct inject;
55         /* Used for setup/shutdown */
56         struct eventfd_ctx *eventfd;
57         struct list_head list;
58         poll_table pt;
59         struct work_struct shutdown;
60 };
62 static struct workqueue_struct *irqfd_cleanup_wq;
64 static void
65 irqfd_inject(struct work_struct *work)
66 {
67         struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
68         struct kvm *kvm = irqfd->kvm;
70         kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1);
71         kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0);
72 }
74 /*
75  * Race-free decouple logic (ordering is critical)
76  */
77 static void
78 irqfd_shutdown(struct work_struct *work)
79 {
80         struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
81         u64 cnt;
83         /*
84          * Synchronize with the wait-queue and unhook ourselves to prevent
85          * further events.
86          */
87         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
89         /*
90          * We know no new events will be scheduled at this point, so block
91          * until all previously outstanding events have completed
92          */
93         flush_work_sync(&irqfd->inject);
95         /*
96          * It is now safe to release the object's resources
97          */
98         eventfd_ctx_put(irqfd->eventfd);
99         kfree(irqfd);
103 /* assumes kvm->irqfds.lock is held */
104 static bool
105 irqfd_is_active(struct _irqfd *irqfd)
107         return list_empty(&irqfd->list) ? false : true;
110 /*
111  * Mark the irqfd as inactive and schedule it for removal
112  *
113  * assumes kvm->irqfds.lock is held
114  */
115 static void
116 irqfd_deactivate(struct _irqfd *irqfd)
118         BUG_ON(!irqfd_is_active(irqfd));
120         list_del_init(&irqfd->list);
122         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
125 /*
126  * Called with wqh->lock held and interrupts disabled
127  */
128 static int
129 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
131         struct _irqfd *irqfd = container_of(wait, struct _irqfd, wait);
132         unsigned long flags = (unsigned long)key;
133         struct kvm_kernel_irq_routing_entry *irq;
134         struct kvm *kvm = irqfd->kvm;
136         if (flags & POLLIN) {
137                 rcu_read_lock();
138                 irq = rcu_dereference(irqfd->irq_entry);
139                 /* An event has been signaled, inject an interrupt */
140                 if (irq)
141                         kvm_set_msi(irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1);
142                 else
143                         schedule_work(&irqfd->inject);
144                 rcu_read_unlock();
145         }
147         if (flags & POLLHUP) {
148                 /* The eventfd is closing, detach from KVM */
149                 unsigned long flags;
151                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
153                 /*
154                  * We must check if someone deactivated the irqfd before
155                  * we could acquire the irqfds.lock since the item is
156                  * deactivated from the KVM side before it is unhooked from
157                  * the wait-queue.  If it is already deactivated, we can
158                  * simply return knowing the other side will cleanup for us.
159                  * We cannot race against the irqfd going away since the
160                  * other side is required to acquire wqh->lock, which we hold
161                  */
162                 if (irqfd_is_active(irqfd))
163                         irqfd_deactivate(irqfd);
165                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
166         }
168         return 0;
171 static void
172 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
173                         poll_table *pt)
175         struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
176         add_wait_queue(wqh, &irqfd->wait);
179 /* Must be called under irqfds.lock */
180 static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd,
181                          struct kvm_irq_routing_table *irq_rt)
183         struct kvm_kernel_irq_routing_entry *e;
184         struct hlist_node *n;
186         if (irqfd->gsi >= irq_rt->nr_rt_entries) {
187                 rcu_assign_pointer(irqfd->irq_entry, NULL);
188                 return;
189         }
191         hlist_for_each_entry(e, n, &irq_rt->map[irqfd->gsi], link) {
192                 /* Only fast-path MSI. */
193                 if (e->type == KVM_IRQ_ROUTING_MSI)
194                         rcu_assign_pointer(irqfd->irq_entry, e);
195                 else
196                         rcu_assign_pointer(irqfd->irq_entry, NULL);
197         }
200 static int
201 kvm_irqfd_assign(struct kvm *kvm, int fd, int gsi)
203         struct kvm_irq_routing_table *irq_rt;
204         struct _irqfd *irqfd, *tmp;
205         struct file *file = NULL;
206         struct eventfd_ctx *eventfd = NULL;
207         int ret;
208         unsigned int events;
210         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
211         if (!irqfd)
212                 return -ENOMEM;
214         irqfd->kvm = kvm;
215         irqfd->gsi = gsi;
216         INIT_LIST_HEAD(&irqfd->list);
217         INIT_WORK(&irqfd->inject, irqfd_inject);
218         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
220         file = eventfd_fget(fd);
221         if (IS_ERR(file)) {
222                 ret = PTR_ERR(file);
223                 goto fail;
224         }
226         eventfd = eventfd_ctx_fileget(file);
227         if (IS_ERR(eventfd)) {
228                 ret = PTR_ERR(eventfd);
229                 goto fail;
230         }
232         irqfd->eventfd = eventfd;
234         /*
235          * Install our own custom wake-up handling so we are notified via
236          * a callback whenever someone signals the underlying eventfd
237          */
238         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
239         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
241         spin_lock_irq(&kvm->irqfds.lock);
243         ret = 0;
244         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
245                 if (irqfd->eventfd != tmp->eventfd)
246                         continue;
247                 /* This fd is used for another irq already. */
248                 ret = -EBUSY;
249                 spin_unlock_irq(&kvm->irqfds.lock);
250                 goto fail;
251         }
253         irq_rt = rcu_dereference_protected(kvm->irq_routing,
254                                            lockdep_is_held(&kvm->irqfds.lock));
255         irqfd_update(kvm, irqfd, irq_rt);
257         events = file->f_op->poll(file, &irqfd->pt);
259         list_add_tail(&irqfd->list, &kvm->irqfds.items);
261         /*
262          * Check if there was an event already pending on the eventfd
263          * before we registered, and trigger it as if we didn't miss it.
264          */
265         if (events & POLLIN)
266                 schedule_work(&irqfd->inject);
268         spin_unlock_irq(&kvm->irqfds.lock);
270         /*
271          * do not drop the file until the irqfd is fully initialized, otherwise
272          * we might race against the POLLHUP
273          */
274         fput(file);
276         return 0;
278 fail:
279         if (eventfd && !IS_ERR(eventfd))
280                 eventfd_ctx_put(eventfd);
282         if (!IS_ERR(file))
283                 fput(file);
285         kfree(irqfd);
286         return ret;
289 void
290 kvm_eventfd_init(struct kvm *kvm)
292         spin_lock_init(&kvm->irqfds.lock);
293         INIT_LIST_HEAD(&kvm->irqfds.items);
294         INIT_LIST_HEAD(&kvm->ioeventfds);
297 /*
298  * shutdown any irqfd's that match fd+gsi
299  */
300 static int
301 kvm_irqfd_deassign(struct kvm *kvm, int fd, int gsi)
303         struct _irqfd *irqfd, *tmp;
304         struct eventfd_ctx *eventfd;
306         eventfd = eventfd_ctx_fdget(fd);
307         if (IS_ERR(eventfd))
308                 return PTR_ERR(eventfd);
310         spin_lock_irq(&kvm->irqfds.lock);
312         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
313                 if (irqfd->eventfd == eventfd && irqfd->gsi == gsi) {
314                         /*
315                          * This rcu_assign_pointer is needed for when
316                          * another thread calls kvm_irq_routing_update before
317                          * we flush workqueue below (we synchronize with
318                          * kvm_irq_routing_update using irqfds.lock).
319                          * It is paired with synchronize_rcu done by caller
320                          * of that function.
321                          */
322                         rcu_assign_pointer(irqfd->irq_entry, NULL);
323                         irqfd_deactivate(irqfd);
324                 }
325         }
327         spin_unlock_irq(&kvm->irqfds.lock);
328         eventfd_ctx_put(eventfd);
330         /*
331          * Block until we know all outstanding shutdown jobs have completed
332          * so that we guarantee there will not be any more interrupts on this
333          * gsi once this deassign function returns.
334          */
335         flush_workqueue(irqfd_cleanup_wq);
337         return 0;
340 int
341 kvm_irqfd(struct kvm *kvm, int fd, int gsi, int flags)
343         if (flags & KVM_IRQFD_FLAG_DEASSIGN)
344                 return kvm_irqfd_deassign(kvm, fd, gsi);
346         return kvm_irqfd_assign(kvm, fd, gsi);
349 /*
350  * This function is called as the kvm VM fd is being released. Shutdown all
351  * irqfds that still remain open
352  */
353 void
354 kvm_irqfd_release(struct kvm *kvm)
356         struct _irqfd *irqfd, *tmp;
358         spin_lock_irq(&kvm->irqfds.lock);
360         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
361                 irqfd_deactivate(irqfd);
363         spin_unlock_irq(&kvm->irqfds.lock);
365         /*
366          * Block until we know all outstanding shutdown jobs have completed
367          * since we do not take a kvm* reference.
368          */
369         flush_workqueue(irqfd_cleanup_wq);
373 /*
374  * Change irq_routing and irqfd.
375  * Caller must invoke synchronize_rcu afterwards.
376  */
377 void kvm_irq_routing_update(struct kvm *kvm,
378                             struct kvm_irq_routing_table *irq_rt)
380         struct _irqfd *irqfd;
382         spin_lock_irq(&kvm->irqfds.lock);
384         rcu_assign_pointer(kvm->irq_routing, irq_rt);
386         list_for_each_entry(irqfd, &kvm->irqfds.items, list)
387                 irqfd_update(kvm, irqfd, irq_rt);
389         spin_unlock_irq(&kvm->irqfds.lock);
392 /*
393  * create a host-wide workqueue for issuing deferred shutdown requests
394  * aggregated from all vm* instances. We need our own isolated single-thread
395  * queue to prevent deadlock against flushing the normal work-queue.
396  */
397 static int __init irqfd_module_init(void)
399         irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
400         if (!irqfd_cleanup_wq)
401                 return -ENOMEM;
403         return 0;
406 static void __exit irqfd_module_exit(void)
408         destroy_workqueue(irqfd_cleanup_wq);
411 module_init(irqfd_module_init);
412 module_exit(irqfd_module_exit);
414 /*
415  * --------------------------------------------------------------------
416  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
417  *
418  * userspace can register a PIO/MMIO address with an eventfd for receiving
419  * notification when the memory has been touched.
420  * --------------------------------------------------------------------
421  */
423 struct _ioeventfd {
424         struct list_head     list;
425         u64                  addr;
426         int                  length;
427         struct eventfd_ctx  *eventfd;
428         u64                  datamatch;
429         struct kvm_io_device dev;
430         bool                 wildcard;
431 };
433 static inline struct _ioeventfd *
434 to_ioeventfd(struct kvm_io_device *dev)
436         return container_of(dev, struct _ioeventfd, dev);
439 static void
440 ioeventfd_release(struct _ioeventfd *p)
442         eventfd_ctx_put(p->eventfd);
443         list_del(&p->list);
444         kfree(p);
447 static bool
448 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
450         u64 _val;
452         if (!(addr == p->addr && len == p->length))
453                 /* address-range must be precise for a hit */
454                 return false;
456         if (p->wildcard)
457                 /* all else equal, wildcard is always a hit */
458                 return true;
460         /* otherwise, we have to actually compare the data */
462         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
464         switch (len) {
465         case 1:
466                 _val = *(u8 *)val;
467                 break;
468         case 2:
469                 _val = *(u16 *)val;
470                 break;
471         case 4:
472                 _val = *(u32 *)val;
473                 break;
474         case 8:
475                 _val = *(u64 *)val;
476                 break;
477         default:
478                 return false;
479         }
481         return _val == p->datamatch ? true : false;
484 /* MMIO/PIO writes trigger an event if the addr/val match */
485 static int
486 ioeventfd_write(struct kvm_io_device *this, gpa_t addr, int len,
487                 const void *val)
489         struct _ioeventfd *p = to_ioeventfd(this);
491         if (!ioeventfd_in_range(p, addr, len, val))
492                 return -EOPNOTSUPP;
494         eventfd_signal(p->eventfd, 1);
495         return 0;
498 /*
499  * This function is called as KVM is completely shutting down.  We do not
500  * need to worry about locking just nuke anything we have as quickly as possible
501  */
502 static void
503 ioeventfd_destructor(struct kvm_io_device *this)
505         struct _ioeventfd *p = to_ioeventfd(this);
507         ioeventfd_release(p);
510 static const struct kvm_io_device_ops ioeventfd_ops = {
511         .write      = ioeventfd_write,
512         .destructor = ioeventfd_destructor,
513 };
515 /* assumes kvm->slots_lock held */
516 static bool
517 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
519         struct _ioeventfd *_p;
521         list_for_each_entry(_p, &kvm->ioeventfds, list)
522                 if (_p->addr == p->addr && _p->length == p->length &&
523                     (_p->wildcard || p->wildcard ||
524                      _p->datamatch == p->datamatch))
525                         return true;
527         return false;
530 static int
531 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
533         int                       pio = args->flags & KVM_IOEVENTFD_FLAG_PIO;
534         enum kvm_bus              bus_idx = pio ? KVM_PIO_BUS : KVM_MMIO_BUS;
535         struct _ioeventfd        *p;
536         struct eventfd_ctx       *eventfd;
537         int                       ret;
539         /* must be natural-word sized */
540         switch (args->len) {
541         case 1:
542         case 2:
543         case 4:
544         case 8:
545                 break;
546         default:
547                 return -EINVAL;
548         }
550         /* check for range overflow */
551         if (args->addr + args->len < args->addr)
552                 return -EINVAL;
554         /* check for extra flags that we don't understand */
555         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
556                 return -EINVAL;
558         eventfd = eventfd_ctx_fdget(args->fd);
559         if (IS_ERR(eventfd))
560                 return PTR_ERR(eventfd);
562         p = kzalloc(sizeof(*p), GFP_KERNEL);
563         if (!p) {
564                 ret = -ENOMEM;
565                 goto fail;
566         }
568         INIT_LIST_HEAD(&p->list);
569         p->addr    = args->addr;
570         p->length  = args->len;
571         p->eventfd = eventfd;
573         /* The datamatch feature is optional, otherwise this is a wildcard */
574         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
575                 p->datamatch = args->datamatch;
576         else
577                 p->wildcard = true;
579         mutex_lock(&kvm->slots_lock);
581         /* Verify that there isn't a match already */
582         if (ioeventfd_check_collision(kvm, p)) {
583                 ret = -EEXIST;
584                 goto unlock_fail;
585         }
587         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
589         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
590                                       &p->dev);
591         if (ret < 0)
592                 goto unlock_fail;
594         list_add_tail(&p->list, &kvm->ioeventfds);
596         mutex_unlock(&kvm->slots_lock);
598         return 0;
600 unlock_fail:
601         mutex_unlock(&kvm->slots_lock);
603 fail:
604         kfree(p);
605         eventfd_ctx_put(eventfd);
607         return ret;
610 static int
611 kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
613         int                       pio = args->flags & KVM_IOEVENTFD_FLAG_PIO;
614         enum kvm_bus              bus_idx = pio ? KVM_PIO_BUS : KVM_MMIO_BUS;
615         struct _ioeventfd        *p, *tmp;
616         struct eventfd_ctx       *eventfd;
617         int                       ret = -ENOENT;
619         eventfd = eventfd_ctx_fdget(args->fd);
620         if (IS_ERR(eventfd))
621                 return PTR_ERR(eventfd);
623         mutex_lock(&kvm->slots_lock);
625         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
626                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
628                 if (p->eventfd != eventfd  ||
629                     p->addr != args->addr  ||
630                     p->length != args->len ||
631                     p->wildcard != wildcard)
632                         continue;
634                 if (!p->wildcard && p->datamatch != args->datamatch)
635                         continue;
637                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
638                 ioeventfd_release(p);
639                 ret = 0;
640                 break;
641         }
643         mutex_unlock(&kvm->slots_lock);
645         eventfd_ctx_put(eventfd);
647         return ret;
650 int
651 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
653         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
654                 return kvm_deassign_ioeventfd(kvm, args);
656         return kvm_assign_ioeventfd(kvm, args);