1 /* Copyright (C) 2009 Red Hat, Inc.
2 * Copyright (C) 2006 Rusty Russell IBM Corporation
4 * Author: Michael S. Tsirkin <mst@redhat.com>
6 * Inspiration, some code, and most witty comments come from
7 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
9 * This work is licensed under the terms of the GNU GPL, version 2.
11 * Generic code for virtio server in host kernel.
14 #include <linux/eventfd.h>
15 #include <linux/vhost.h>
16 #include <linux/uio.h>
18 #include <linux/mmu_context.h>
19 #include <linux/miscdevice.h>
20 #include <linux/mutex.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/vmalloc.h>
26 #include <linux/kthread.h>
27 #include <linux/cgroup.h>
28 #include <linux/module.h>
29 #include <linux/sort.h>
33 static ushort max_mem_regions = 64;
34 module_param(max_mem_regions, ushort, 0444);
35 MODULE_PARM_DESC(max_mem_regions,
36 "Maximum number of memory regions in memory map. (default: 64)");
39 VHOST_MEMORY_F_LOG = 0x1,
42 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
43 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
45 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
46 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
48 vq->user_be = !virtio_legacy_is_little_endian();
51 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
56 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
61 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
63 struct vhost_vring_state s;
68 if (copy_from_user(&s, argp, sizeof(s)))
71 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
72 s.num != VHOST_VRING_BIG_ENDIAN)
75 if (s.num == VHOST_VRING_BIG_ENDIAN)
76 vhost_enable_cross_endian_big(vq);
78 vhost_enable_cross_endian_little(vq);
83 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
86 struct vhost_vring_state s = {
91 if (copy_to_user(argp, &s, sizeof(s)))
97 static void vhost_init_is_le(struct vhost_virtqueue *vq)
99 /* Note for legacy virtio: user_be is initialized at reset time
100 * according to the host endianness. If userspace does not set an
101 * explicit endianness, the default behavior is native endian, as
102 * expected by legacy virtio.
104 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
107 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
111 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
116 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
122 static void vhost_init_is_le(struct vhost_virtqueue *vq)
124 if (vhost_has_feature(vq, VIRTIO_F_VERSION_1))
127 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
129 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
131 vq->is_le = virtio_legacy_is_little_endian();
134 struct vhost_flush_struct {
135 struct vhost_work work;
136 struct completion wait_event;
139 static void vhost_flush_work(struct vhost_work *work)
141 struct vhost_flush_struct *s;
143 s = container_of(work, struct vhost_flush_struct, work);
144 complete(&s->wait_event);
147 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
150 struct vhost_poll *poll;
152 poll = container_of(pt, struct vhost_poll, table);
154 add_wait_queue(wqh, &poll->wait);
157 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
160 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
162 if (!((unsigned long)key & poll->mask))
165 vhost_poll_queue(poll);
169 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
171 clear_bit(VHOST_WORK_QUEUED, &work->flags);
173 init_waitqueue_head(&work->done);
175 EXPORT_SYMBOL_GPL(vhost_work_init);
177 /* Init poll structure */
178 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
179 unsigned long mask, struct vhost_dev *dev)
181 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
182 init_poll_funcptr(&poll->table, vhost_poll_func);
187 vhost_work_init(&poll->work, fn);
189 EXPORT_SYMBOL_GPL(vhost_poll_init);
191 /* Start polling a file. We add ourselves to file's wait queue. The caller must
192 * keep a reference to a file until after vhost_poll_stop is called. */
193 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
201 mask = file->f_op->poll(file, &poll->table);
203 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
204 if (mask & POLLERR) {
206 remove_wait_queue(poll->wqh, &poll->wait);
212 EXPORT_SYMBOL_GPL(vhost_poll_start);
214 /* Stop polling a file. After this function returns, it becomes safe to drop the
215 * file reference. You must also flush afterwards. */
216 void vhost_poll_stop(struct vhost_poll *poll)
219 remove_wait_queue(poll->wqh, &poll->wait);
223 EXPORT_SYMBOL_GPL(vhost_poll_stop);
225 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
227 struct vhost_flush_struct flush;
230 init_completion(&flush.wait_event);
231 vhost_work_init(&flush.work, vhost_flush_work);
233 vhost_work_queue(dev, &flush.work);
234 wait_for_completion(&flush.wait_event);
237 EXPORT_SYMBOL_GPL(vhost_work_flush);
239 /* Flush any work that has been scheduled. When calling this, don't hold any
240 * locks that are also used by the callback. */
241 void vhost_poll_flush(struct vhost_poll *poll)
243 vhost_work_flush(poll->dev, &poll->work);
245 EXPORT_SYMBOL_GPL(vhost_poll_flush);
247 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
252 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
253 /* We can only add the work to the list after we're
254 * sure it was not in the list.
257 llist_add(&work->node, &dev->work_list);
258 wake_up_process(dev->worker);
261 EXPORT_SYMBOL_GPL(vhost_work_queue);
263 /* A lockless hint for busy polling code to exit the loop */
264 bool vhost_has_work(struct vhost_dev *dev)
266 return !llist_empty(&dev->work_list);
268 EXPORT_SYMBOL_GPL(vhost_has_work);
270 void vhost_poll_queue(struct vhost_poll *poll)
272 vhost_work_queue(poll->dev, &poll->work);
274 EXPORT_SYMBOL_GPL(vhost_poll_queue);
276 static void vhost_vq_reset(struct vhost_dev *dev,
277 struct vhost_virtqueue *vq)
283 vq->last_avail_idx = 0;
285 vq->last_used_idx = 0;
286 vq->signalled_used = 0;
287 vq->signalled_used_valid = false;
289 vq->log_used = false;
290 vq->log_addr = -1ull;
291 vq->private_data = NULL;
292 vq->acked_features = 0;
294 vq->error_ctx = NULL;
301 vhost_reset_is_le(vq);
302 vhost_disable_cross_endian(vq);
303 vq->busyloop_timeout = 0;
306 static int vhost_worker(void *data)
308 struct vhost_dev *dev = data;
309 struct vhost_work *work, *work_next;
310 struct llist_node *node;
311 mm_segment_t oldfs = get_fs();
317 /* mb paired w/ kthread_stop */
318 set_current_state(TASK_INTERRUPTIBLE);
320 if (kthread_should_stop()) {
321 __set_current_state(TASK_RUNNING);
325 node = llist_del_all(&dev->work_list);
329 node = llist_reverse_order(node);
330 /* make sure flag is seen after deletion */
332 llist_for_each_entry_safe(work, work_next, node, node) {
333 clear_bit(VHOST_WORK_QUEUED, &work->flags);
334 __set_current_state(TASK_RUNNING);
345 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
355 /* Helper to allocate iovec buffers for all vqs. */
356 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
358 struct vhost_virtqueue *vq;
361 for (i = 0; i < dev->nvqs; ++i) {
363 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
365 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
366 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
367 if (!vq->indirect || !vq->log || !vq->heads)
374 vhost_vq_free_iovecs(dev->vqs[i]);
378 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
382 for (i = 0; i < dev->nvqs; ++i)
383 vhost_vq_free_iovecs(dev->vqs[i]);
386 void vhost_dev_init(struct vhost_dev *dev,
387 struct vhost_virtqueue **vqs, int nvqs)
389 struct vhost_virtqueue *vq;
394 mutex_init(&dev->mutex);
396 dev->log_file = NULL;
400 init_llist_head(&dev->work_list);
403 for (i = 0; i < dev->nvqs; ++i) {
409 mutex_init(&vq->mutex);
410 vhost_vq_reset(dev, vq);
412 vhost_poll_init(&vq->poll, vq->handle_kick,
416 EXPORT_SYMBOL_GPL(vhost_dev_init);
418 /* Caller should have device mutex */
419 long vhost_dev_check_owner(struct vhost_dev *dev)
421 /* Are you the owner? If not, I don't think you mean to do that */
422 return dev->mm == current->mm ? 0 : -EPERM;
424 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
426 struct vhost_attach_cgroups_struct {
427 struct vhost_work work;
428 struct task_struct *owner;
432 static void vhost_attach_cgroups_work(struct vhost_work *work)
434 struct vhost_attach_cgroups_struct *s;
436 s = container_of(work, struct vhost_attach_cgroups_struct, work);
437 s->ret = cgroup_attach_task_all(s->owner, current);
440 static int vhost_attach_cgroups(struct vhost_dev *dev)
442 struct vhost_attach_cgroups_struct attach;
444 attach.owner = current;
445 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
446 vhost_work_queue(dev, &attach.work);
447 vhost_work_flush(dev, &attach.work);
451 /* Caller should have device mutex */
452 bool vhost_dev_has_owner(struct vhost_dev *dev)
456 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
458 /* Caller should have device mutex */
459 long vhost_dev_set_owner(struct vhost_dev *dev)
461 struct task_struct *worker;
464 /* Is there an owner already? */
465 if (vhost_dev_has_owner(dev)) {
470 /* No owner, become one */
471 dev->mm = get_task_mm(current);
472 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
473 if (IS_ERR(worker)) {
474 err = PTR_ERR(worker);
478 dev->worker = worker;
479 wake_up_process(worker); /* avoid contributing to loadavg */
481 err = vhost_attach_cgroups(dev);
485 err = vhost_dev_alloc_iovecs(dev);
491 kthread_stop(worker);
500 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
502 struct vhost_memory *vhost_dev_reset_owner_prepare(void)
504 return kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
506 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
508 /* Caller should have device mutex */
509 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_memory *memory)
513 vhost_dev_cleanup(dev, true);
515 /* Restore memory to default empty mapping. */
516 memory->nregions = 0;
517 dev->memory = memory;
518 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
519 * VQs aren't running.
521 for (i = 0; i < dev->nvqs; ++i)
522 dev->vqs[i]->memory = memory;
524 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
526 void vhost_dev_stop(struct vhost_dev *dev)
530 for (i = 0; i < dev->nvqs; ++i) {
531 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
532 vhost_poll_stop(&dev->vqs[i]->poll);
533 vhost_poll_flush(&dev->vqs[i]->poll);
537 EXPORT_SYMBOL_GPL(vhost_dev_stop);
539 /* Caller should have device mutex if and only if locked is set */
540 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
544 for (i = 0; i < dev->nvqs; ++i) {
545 if (dev->vqs[i]->error_ctx)
546 eventfd_ctx_put(dev->vqs[i]->error_ctx);
547 if (dev->vqs[i]->error)
548 fput(dev->vqs[i]->error);
549 if (dev->vqs[i]->kick)
550 fput(dev->vqs[i]->kick);
551 if (dev->vqs[i]->call_ctx)
552 eventfd_ctx_put(dev->vqs[i]->call_ctx);
553 if (dev->vqs[i]->call)
554 fput(dev->vqs[i]->call);
555 vhost_vq_reset(dev, dev->vqs[i]);
557 vhost_dev_free_iovecs(dev);
559 eventfd_ctx_put(dev->log_ctx);
563 dev->log_file = NULL;
564 /* No one will access memory at this point */
567 WARN_ON(!llist_empty(&dev->work_list));
569 kthread_stop(dev->worker);
576 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
578 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
580 u64 a = addr / VHOST_PAGE_SIZE / 8;
582 /* Make sure 64 bit math will not overflow. */
583 if (a > ULONG_MAX - (unsigned long)log_base ||
584 a + (unsigned long)log_base > ULONG_MAX)
587 return access_ok(VERIFY_WRITE, log_base + a,
588 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
591 /* Caller should have vq mutex and device mutex. */
592 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
600 for (i = 0; i < mem->nregions; ++i) {
601 struct vhost_memory_region *m = mem->regions + i;
602 unsigned long a = m->userspace_addr;
603 if (m->memory_size > ULONG_MAX)
605 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
608 else if (log_all && !log_access_ok(log_base,
616 /* Can we switch to this memory table? */
617 /* Caller should have device mutex but not vq mutex */
618 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
623 for (i = 0; i < d->nvqs; ++i) {
627 mutex_lock(&d->vqs[i]->mutex);
628 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
629 /* If ring is inactive, will check when it's enabled. */
630 if (d->vqs[i]->private_data)
631 ok = vq_memory_access_ok(d->vqs[i]->log_base, mem, log);
634 mutex_unlock(&d->vqs[i]->mutex);
641 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
642 struct vring_desc __user *desc,
643 struct vring_avail __user *avail,
644 struct vring_used __user *used)
646 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
647 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
648 access_ok(VERIFY_READ, avail,
649 sizeof *avail + num * sizeof *avail->ring + s) &&
650 access_ok(VERIFY_WRITE, used,
651 sizeof *used + num * sizeof *used->ring + s);
654 /* Can we log writes? */
655 /* Caller should have device mutex but not vq mutex */
656 int vhost_log_access_ok(struct vhost_dev *dev)
658 return memory_access_ok(dev, dev->memory, 1);
660 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
662 /* Verify access for write logging. */
663 /* Caller should have vq mutex and device mutex */
664 static int vq_log_access_ok(struct vhost_virtqueue *vq,
665 void __user *log_base)
667 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
669 return vq_memory_access_ok(log_base, vq->memory,
670 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
671 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
673 vq->num * sizeof *vq->used->ring + s));
676 /* Can we start vq? */
677 /* Caller should have vq mutex and device mutex */
678 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
680 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
681 vq_log_access_ok(vq, vq->log_base);
683 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
685 static int vhost_memory_reg_sort_cmp(const void *p1, const void *p2)
687 const struct vhost_memory_region *r1 = p1, *r2 = p2;
688 if (r1->guest_phys_addr < r2->guest_phys_addr)
690 if (r1->guest_phys_addr > r2->guest_phys_addr)
695 static void *vhost_kvzalloc(unsigned long size)
697 void *n = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
704 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
706 struct vhost_memory mem, *newmem, *oldmem;
707 unsigned long size = offsetof(struct vhost_memory, regions);
710 if (copy_from_user(&mem, m, size))
714 if (mem.nregions > max_mem_regions)
716 newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
720 memcpy(newmem, &mem, size);
721 if (copy_from_user(newmem->regions, m->regions,
722 mem.nregions * sizeof *m->regions)) {
726 sort(newmem->regions, newmem->nregions, sizeof(*newmem->regions),
727 vhost_memory_reg_sort_cmp, NULL);
729 if (!memory_access_ok(d, newmem, 0)) {
736 /* All memory accesses are done under some VQ mutex. */
737 for (i = 0; i < d->nvqs; ++i) {
738 mutex_lock(&d->vqs[i]->mutex);
739 d->vqs[i]->memory = newmem;
740 mutex_unlock(&d->vqs[i]->mutex);
746 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
748 struct file *eventfp, *filep = NULL;
749 bool pollstart = false, pollstop = false;
750 struct eventfd_ctx *ctx = NULL;
751 u32 __user *idxp = argp;
752 struct vhost_virtqueue *vq;
753 struct vhost_vring_state s;
754 struct vhost_vring_file f;
755 struct vhost_vring_addr a;
759 r = get_user(idx, idxp);
767 mutex_lock(&vq->mutex);
770 case VHOST_SET_VRING_NUM:
771 /* Resizing ring with an active backend?
772 * You don't want to do that. */
773 if (vq->private_data) {
777 if (copy_from_user(&s, argp, sizeof s)) {
781 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
787 case VHOST_SET_VRING_BASE:
788 /* Moving base with an active backend?
789 * You don't want to do that. */
790 if (vq->private_data) {
794 if (copy_from_user(&s, argp, sizeof s)) {
798 if (s.num > 0xffff) {
802 vq->last_avail_idx = s.num;
803 /* Forget the cached index value. */
804 vq->avail_idx = vq->last_avail_idx;
806 case VHOST_GET_VRING_BASE:
808 s.num = vq->last_avail_idx;
809 if (copy_to_user(argp, &s, sizeof s))
812 case VHOST_SET_VRING_ADDR:
813 if (copy_from_user(&a, argp, sizeof a)) {
817 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
821 /* For 32bit, verify that the top 32bits of the user
822 data are set to zero. */
823 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
824 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
825 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
830 /* Make sure it's safe to cast pointers to vring types. */
831 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
832 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
833 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
834 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
835 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
840 /* We only verify access here if backend is configured.
841 * If it is not, we don't as size might not have been setup.
842 * We will verify when backend is configured. */
843 if (vq->private_data) {
844 if (!vq_access_ok(vq, vq->num,
845 (void __user *)(unsigned long)a.desc_user_addr,
846 (void __user *)(unsigned long)a.avail_user_addr,
847 (void __user *)(unsigned long)a.used_user_addr)) {
852 /* Also validate log access for used ring if enabled. */
853 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
854 !log_access_ok(vq->log_base, a.log_guest_addr,
856 vq->num * sizeof *vq->used->ring)) {
862 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
863 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
864 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
865 vq->log_addr = a.log_guest_addr;
866 vq->used = (void __user *)(unsigned long)a.used_user_addr;
868 case VHOST_SET_VRING_KICK:
869 if (copy_from_user(&f, argp, sizeof f)) {
873 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
874 if (IS_ERR(eventfp)) {
875 r = PTR_ERR(eventfp);
878 if (eventfp != vq->kick) {
879 pollstop = (filep = vq->kick) != NULL;
880 pollstart = (vq->kick = eventfp) != NULL;
884 case VHOST_SET_VRING_CALL:
885 if (copy_from_user(&f, argp, sizeof f)) {
889 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
890 if (IS_ERR(eventfp)) {
891 r = PTR_ERR(eventfp);
894 if (eventfp != vq->call) {
898 vq->call_ctx = eventfp ?
899 eventfd_ctx_fileget(eventfp) : NULL;
903 case VHOST_SET_VRING_ERR:
904 if (copy_from_user(&f, argp, sizeof f)) {
908 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
909 if (IS_ERR(eventfp)) {
910 r = PTR_ERR(eventfp);
913 if (eventfp != vq->error) {
917 vq->error_ctx = eventfp ?
918 eventfd_ctx_fileget(eventfp) : NULL;
922 case VHOST_SET_VRING_ENDIAN:
923 r = vhost_set_vring_endian(vq, argp);
925 case VHOST_GET_VRING_ENDIAN:
926 r = vhost_get_vring_endian(vq, idx, argp);
928 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
929 if (copy_from_user(&s, argp, sizeof(s))) {
933 vq->busyloop_timeout = s.num;
935 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
937 s.num = vq->busyloop_timeout;
938 if (copy_to_user(argp, &s, sizeof(s)))
945 if (pollstop && vq->handle_kick)
946 vhost_poll_stop(&vq->poll);
949 eventfd_ctx_put(ctx);
953 if (pollstart && vq->handle_kick)
954 r = vhost_poll_start(&vq->poll, vq->kick);
956 mutex_unlock(&vq->mutex);
958 if (pollstop && vq->handle_kick)
959 vhost_poll_flush(&vq->poll);
962 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
964 /* Caller must have device mutex */
965 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
967 struct file *eventfp, *filep = NULL;
968 struct eventfd_ctx *ctx = NULL;
973 /* If you are not the owner, you can become one */
974 if (ioctl == VHOST_SET_OWNER) {
975 r = vhost_dev_set_owner(d);
979 /* You must be the owner to do anything else */
980 r = vhost_dev_check_owner(d);
985 case VHOST_SET_MEM_TABLE:
986 r = vhost_set_memory(d, argp);
988 case VHOST_SET_LOG_BASE:
989 if (copy_from_user(&p, argp, sizeof p)) {
993 if ((u64)(unsigned long)p != p) {
997 for (i = 0; i < d->nvqs; ++i) {
998 struct vhost_virtqueue *vq;
999 void __user *base = (void __user *)(unsigned long)p;
1001 mutex_lock(&vq->mutex);
1002 /* If ring is inactive, will check when it's enabled. */
1003 if (vq->private_data && !vq_log_access_ok(vq, base))
1006 vq->log_base = base;
1007 mutex_unlock(&vq->mutex);
1010 case VHOST_SET_LOG_FD:
1011 r = get_user(fd, (int __user *)argp);
1014 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1015 if (IS_ERR(eventfp)) {
1016 r = PTR_ERR(eventfp);
1019 if (eventfp != d->log_file) {
1020 filep = d->log_file;
1021 d->log_file = eventfp;
1023 d->log_ctx = eventfp ?
1024 eventfd_ctx_fileget(eventfp) : NULL;
1027 for (i = 0; i < d->nvqs; ++i) {
1028 mutex_lock(&d->vqs[i]->mutex);
1029 d->vqs[i]->log_ctx = d->log_ctx;
1030 mutex_unlock(&d->vqs[i]->mutex);
1033 eventfd_ctx_put(ctx);
1044 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1046 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
1047 __u64 addr, __u32 len)
1049 const struct vhost_memory_region *reg;
1050 int start = 0, end = mem->nregions;
1052 while (start < end) {
1053 int slot = start + (end - start) / 2;
1054 reg = mem->regions + slot;
1055 if (addr >= reg->guest_phys_addr)
1061 reg = mem->regions + start;
1062 if (addr >= reg->guest_phys_addr &&
1063 reg->guest_phys_addr + reg->memory_size > addr)
1068 /* TODO: This is really inefficient. We need something like get_user()
1069 * (instruction directly accesses the data, with an exception table entry
1070 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1072 static int set_bit_to_user(int nr, void __user *addr)
1074 unsigned long log = (unsigned long)addr;
1077 int bit = nr + (log % PAGE_SIZE) * 8;
1080 r = get_user_pages_fast(log, 1, 1, &page);
1084 base = kmap_atomic(page);
1086 kunmap_atomic(base);
1087 set_page_dirty_lock(page);
1092 static int log_write(void __user *log_base,
1093 u64 write_address, u64 write_length)
1095 u64 write_page = write_address / VHOST_PAGE_SIZE;
1100 write_length += write_address % VHOST_PAGE_SIZE;
1102 u64 base = (u64)(unsigned long)log_base;
1103 u64 log = base + write_page / 8;
1104 int bit = write_page % 8;
1105 if ((u64)(unsigned long)log != log)
1107 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1110 if (write_length <= VHOST_PAGE_SIZE)
1112 write_length -= VHOST_PAGE_SIZE;
1118 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1119 unsigned int log_num, u64 len)
1123 /* Make sure data written is seen before log. */
1125 for (i = 0; i < log_num; ++i) {
1126 u64 l = min(log[i].len, len);
1127 r = log_write(vq->log_base, log[i].addr, l);
1133 eventfd_signal(vq->log_ctx, 1);
1137 /* Length written exceeds what we have stored. This is a bug. */
1141 EXPORT_SYMBOL_GPL(vhost_log_write);
1143 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1146 if (__put_user(cpu_to_vhost16(vq, vq->used_flags), &vq->used->flags) < 0)
1148 if (unlikely(vq->log_used)) {
1149 /* Make sure the flag is seen before log. */
1151 /* Log used flag write. */
1152 used = &vq->used->flags;
1153 log_write(vq->log_base, vq->log_addr +
1154 (used - (void __user *)vq->used),
1155 sizeof vq->used->flags);
1157 eventfd_signal(vq->log_ctx, 1);
1162 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1164 if (__put_user(cpu_to_vhost16(vq, vq->avail_idx), vhost_avail_event(vq)))
1166 if (unlikely(vq->log_used)) {
1168 /* Make sure the event is seen before log. */
1170 /* Log avail event write */
1171 used = vhost_avail_event(vq);
1172 log_write(vq->log_base, vq->log_addr +
1173 (used - (void __user *)vq->used),
1174 sizeof *vhost_avail_event(vq));
1176 eventfd_signal(vq->log_ctx, 1);
1181 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1183 __virtio16 last_used_idx;
1185 bool is_le = vq->is_le;
1187 if (!vq->private_data) {
1188 vhost_reset_is_le(vq);
1192 vhost_init_is_le(vq);
1194 r = vhost_update_used_flags(vq);
1197 vq->signalled_used_valid = false;
1198 if (!access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1202 r = __get_user(last_used_idx, &vq->used->idx);
1205 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1211 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1213 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1214 struct iovec iov[], int iov_size)
1216 const struct vhost_memory_region *reg;
1217 struct vhost_memory *mem;
1223 while ((u64)len > s) {
1225 if (unlikely(ret >= iov_size)) {
1229 reg = find_region(mem, addr, len);
1230 if (unlikely(!reg)) {
1235 size = reg->memory_size - addr + reg->guest_phys_addr;
1236 _iov->iov_len = min((u64)len - s, size);
1237 _iov->iov_base = (void __user *)(unsigned long)
1238 (reg->userspace_addr + addr - reg->guest_phys_addr);
1247 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1248 * function returns the next descriptor in the chain,
1249 * or -1U if we're at the end. */
1250 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1254 /* If this descriptor says it doesn't chain, we're done. */
1255 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1258 /* Check they're not leading us off end of descriptors. */
1259 next = vhost16_to_cpu(vq, desc->next);
1260 /* Make sure compiler knows to grab that: we don't want it changing! */
1261 /* We will use the result as an index in an array, so most
1262 * architectures only need a compiler barrier here. */
1263 read_barrier_depends();
1268 static int get_indirect(struct vhost_virtqueue *vq,
1269 struct iovec iov[], unsigned int iov_size,
1270 unsigned int *out_num, unsigned int *in_num,
1271 struct vhost_log *log, unsigned int *log_num,
1272 struct vring_desc *indirect)
1274 struct vring_desc desc;
1275 unsigned int i = 0, count, found = 0;
1276 u32 len = vhost32_to_cpu(vq, indirect->len);
1277 struct iov_iter from;
1281 if (unlikely(len % sizeof desc)) {
1282 vq_err(vq, "Invalid length in indirect descriptor: "
1283 "len 0x%llx not multiple of 0x%zx\n",
1284 (unsigned long long)len,
1289 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1291 if (unlikely(ret < 0)) {
1292 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1295 iov_iter_init(&from, READ, vq->indirect, ret, len);
1297 /* We will use the result as an address to read from, so most
1298 * architectures only need a compiler barrier here. */
1299 read_barrier_depends();
1301 count = len / sizeof desc;
1302 /* Buffers are chained via a 16 bit next field, so
1303 * we can have at most 2^16 of these. */
1304 if (unlikely(count > USHRT_MAX + 1)) {
1305 vq_err(vq, "Indirect buffer length too big: %d\n",
1311 unsigned iov_count = *in_num + *out_num;
1312 if (unlikely(++found > count)) {
1313 vq_err(vq, "Loop detected: last one at %u "
1314 "indirect size %u\n",
1318 if (unlikely(copy_from_iter(&desc, sizeof(desc), &from) !=
1320 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1321 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1324 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1325 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1326 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1330 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1331 vhost32_to_cpu(vq, desc.len), iov + iov_count,
1332 iov_size - iov_count);
1333 if (unlikely(ret < 0)) {
1334 vq_err(vq, "Translation failure %d indirect idx %d\n",
1338 /* If this is an input descriptor, increment that count. */
1339 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1341 if (unlikely(log)) {
1342 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1343 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1347 /* If it's an output descriptor, they're all supposed
1348 * to come before any input descriptors. */
1349 if (unlikely(*in_num)) {
1350 vq_err(vq, "Indirect descriptor "
1351 "has out after in: idx %d\n", i);
1356 } while ((i = next_desc(vq, &desc)) != -1);
1360 /* This looks in the virtqueue and for the first available buffer, and converts
1361 * it to an iovec for convenient access. Since descriptors consist of some
1362 * number of output then some number of input descriptors, it's actually two
1363 * iovecs, but we pack them into one and note how many of each there were.
1365 * This function returns the descriptor number found, or vq->num (which is
1366 * never a valid descriptor number) if none was found. A negative code is
1367 * returned on error. */
1368 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1369 struct iovec iov[], unsigned int iov_size,
1370 unsigned int *out_num, unsigned int *in_num,
1371 struct vhost_log *log, unsigned int *log_num)
1373 struct vring_desc desc;
1374 unsigned int i, head, found = 0;
1376 __virtio16 avail_idx;
1377 __virtio16 ring_head;
1380 /* Check it isn't doing very strange things with descriptor numbers. */
1381 last_avail_idx = vq->last_avail_idx;
1382 if (unlikely(__get_user(avail_idx, &vq->avail->idx))) {
1383 vq_err(vq, "Failed to access avail idx at %p\n",
1387 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1389 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1390 vq_err(vq, "Guest moved used index from %u to %u",
1391 last_avail_idx, vq->avail_idx);
1395 /* If there's nothing new since last we looked, return invalid. */
1396 if (vq->avail_idx == last_avail_idx)
1399 /* Only get avail ring entries after they have been exposed by guest. */
1402 /* Grab the next descriptor number they're advertising, and increment
1403 * the index we've seen. */
1404 if (unlikely(__get_user(ring_head,
1405 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
1406 vq_err(vq, "Failed to read head: idx %d address %p\n",
1408 &vq->avail->ring[last_avail_idx % vq->num]);
1412 head = vhost16_to_cpu(vq, ring_head);
1414 /* If their number is silly, that's an error. */
1415 if (unlikely(head >= vq->num)) {
1416 vq_err(vq, "Guest says index %u > %u is available",
1421 /* When we start there are none of either input nor output. */
1422 *out_num = *in_num = 0;
1428 unsigned iov_count = *in_num + *out_num;
1429 if (unlikely(i >= vq->num)) {
1430 vq_err(vq, "Desc index is %u > %u, head = %u",
1434 if (unlikely(++found > vq->num)) {
1435 vq_err(vq, "Loop detected: last one at %u "
1436 "vq size %u head %u\n",
1440 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1441 if (unlikely(ret)) {
1442 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1446 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
1447 ret = get_indirect(vq, iov, iov_size,
1449 log, log_num, &desc);
1450 if (unlikely(ret < 0)) {
1451 vq_err(vq, "Failure detected "
1452 "in indirect descriptor at idx %d\n", i);
1458 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1459 vhost32_to_cpu(vq, desc.len), iov + iov_count,
1460 iov_size - iov_count);
1461 if (unlikely(ret < 0)) {
1462 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1466 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1467 /* If this is an input descriptor,
1468 * increment that count. */
1470 if (unlikely(log)) {
1471 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1472 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1476 /* If it's an output descriptor, they're all supposed
1477 * to come before any input descriptors. */
1478 if (unlikely(*in_num)) {
1479 vq_err(vq, "Descriptor has out after in: "
1485 } while ((i = next_desc(vq, &desc)) != -1);
1487 /* On success, increment avail index. */
1488 vq->last_avail_idx++;
1490 /* Assume notifications from guest are disabled at this point,
1491 * if they aren't we would need to update avail_event index. */
1492 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1495 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1497 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1498 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1500 vq->last_avail_idx -= n;
1502 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1504 /* After we've used one of their buffers, we tell them about it. We'll then
1505 * want to notify the guest, using eventfd. */
1506 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1508 struct vring_used_elem heads = {
1509 cpu_to_vhost32(vq, head),
1510 cpu_to_vhost32(vq, len)
1513 return vhost_add_used_n(vq, &heads, 1);
1515 EXPORT_SYMBOL_GPL(vhost_add_used);
1517 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1518 struct vring_used_elem *heads,
1521 struct vring_used_elem __user *used;
1525 start = vq->last_used_idx & (vq->num - 1);
1526 used = vq->used->ring + start;
1528 if (__put_user(heads[0].id, &used->id)) {
1529 vq_err(vq, "Failed to write used id");
1532 if (__put_user(heads[0].len, &used->len)) {
1533 vq_err(vq, "Failed to write used len");
1536 } else if (__copy_to_user(used, heads, count * sizeof *used)) {
1537 vq_err(vq, "Failed to write used");
1540 if (unlikely(vq->log_used)) {
1541 /* Make sure data is seen before log. */
1543 /* Log used ring entry write. */
1544 log_write(vq->log_base,
1546 ((void __user *)used - (void __user *)vq->used),
1547 count * sizeof *used);
1549 old = vq->last_used_idx;
1550 new = (vq->last_used_idx += count);
1551 /* If the driver never bothers to signal in a very long while,
1552 * used index might wrap around. If that happens, invalidate
1553 * signalled_used index we stored. TODO: make sure driver
1554 * signals at least once in 2^16 and remove this. */
1555 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1556 vq->signalled_used_valid = false;
1560 /* After we've used one of their buffers, we tell them about it. We'll then
1561 * want to notify the guest, using eventfd. */
1562 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1567 start = vq->last_used_idx & (vq->num - 1);
1568 n = vq->num - start;
1570 r = __vhost_add_used_n(vq, heads, n);
1576 r = __vhost_add_used_n(vq, heads, count);
1578 /* Make sure buffer is written before we update index. */
1580 if (__put_user(cpu_to_vhost16(vq, vq->last_used_idx), &vq->used->idx)) {
1581 vq_err(vq, "Failed to increment used idx");
1584 if (unlikely(vq->log_used)) {
1585 /* Log used index update. */
1586 log_write(vq->log_base,
1587 vq->log_addr + offsetof(struct vring_used, idx),
1588 sizeof vq->used->idx);
1590 eventfd_signal(vq->log_ctx, 1);
1594 EXPORT_SYMBOL_GPL(vhost_add_used_n);
1596 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1601 /* Flush out used index updates. This is paired
1602 * with the barrier that the Guest executes when enabling
1606 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1607 unlikely(vq->avail_idx == vq->last_avail_idx))
1610 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1612 if (__get_user(flags, &vq->avail->flags)) {
1613 vq_err(vq, "Failed to get flags");
1616 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
1618 old = vq->signalled_used;
1619 v = vq->signalled_used_valid;
1620 new = vq->signalled_used = vq->last_used_idx;
1621 vq->signalled_used_valid = true;
1626 if (__get_user(event, vhost_used_event(vq))) {
1627 vq_err(vq, "Failed to get used event idx");
1630 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
1633 /* This actually signals the guest, using eventfd. */
1634 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1636 /* Signal the Guest tell them we used something up. */
1637 if (vq->call_ctx && vhost_notify(dev, vq))
1638 eventfd_signal(vq->call_ctx, 1);
1640 EXPORT_SYMBOL_GPL(vhost_signal);
1642 /* And here's the combo meal deal. Supersize me! */
1643 void vhost_add_used_and_signal(struct vhost_dev *dev,
1644 struct vhost_virtqueue *vq,
1645 unsigned int head, int len)
1647 vhost_add_used(vq, head, len);
1648 vhost_signal(dev, vq);
1650 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1652 /* multi-buffer version of vhost_add_used_and_signal */
1653 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1654 struct vhost_virtqueue *vq,
1655 struct vring_used_elem *heads, unsigned count)
1657 vhost_add_used_n(vq, heads, count);
1658 vhost_signal(dev, vq);
1660 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1662 /* return true if we're sure that avaiable ring is empty */
1663 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1665 __virtio16 avail_idx;
1668 r = __get_user(avail_idx, &vq->avail->idx);
1672 return vhost16_to_cpu(vq, avail_idx) == vq->avail_idx;
1674 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
1676 /* OK, now we need to know about added descriptors. */
1677 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1679 __virtio16 avail_idx;
1682 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1684 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1685 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1686 r = vhost_update_used_flags(vq);
1688 vq_err(vq, "Failed to enable notification at %p: %d\n",
1689 &vq->used->flags, r);
1693 r = vhost_update_avail_event(vq, vq->avail_idx);
1695 vq_err(vq, "Failed to update avail event index at %p: %d\n",
1696 vhost_avail_event(vq), r);
1700 /* They could have slipped one in as we were doing that: make
1701 * sure it's written, then check again. */
1703 r = __get_user(avail_idx, &vq->avail->idx);
1705 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1706 &vq->avail->idx, r);
1710 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
1712 EXPORT_SYMBOL_GPL(vhost_enable_notify);
1714 /* We don't need to be notified again. */
1715 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1719 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1721 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1722 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1723 r = vhost_update_used_flags(vq);
1725 vq_err(vq, "Failed to enable notification at %p: %d\n",
1726 &vq->used->flags, r);
1729 EXPORT_SYMBOL_GPL(vhost_disable_notify);
1731 static int __init vhost_init(void)
1736 static void __exit vhost_exit(void)
1740 module_init(vhost_init);
1741 module_exit(vhost_exit);
1743 MODULE_VERSION("0.0.1");
1744 MODULE_LICENSE("GPL v2");
1745 MODULE_AUTHOR("Michael S. Tsirkin");
1746 MODULE_DESCRIPTION("Host kernel accelerator for virtio");