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/virtio_net.h>
18 #include <linux/mmu_context.h>
19 #include <linux/miscdevice.h>
20 #include <linux/mutex.h>
21 #include <linux/rcupdate.h>
22 #include <linux/poll.h>
23 #include <linux/file.h>
24 #include <linux/highmem.h>
25 #include <linux/slab.h>
26 #include <linux/kthread.h>
27 #include <linux/cgroup.h>
32 VHOST_MEMORY_MAX_NREGIONS = 64,
33 VHOST_MEMORY_F_LOG = 0x1,
36 static unsigned vhost_zcopy_mask __read_mostly;
38 #define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num])
39 #define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num])
41 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
44 struct vhost_poll *poll;
46 poll = container_of(pt, struct vhost_poll, table);
48 add_wait_queue(wqh, &poll->wait);
51 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
54 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
56 if (!((unsigned long)key & poll->mask))
59 vhost_poll_queue(poll);
63 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
65 INIT_LIST_HEAD(&work->node);
67 init_waitqueue_head(&work->done);
69 work->queue_seq = work->done_seq = 0;
72 /* Init poll structure */
73 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
74 unsigned long mask, struct vhost_dev *dev)
76 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
77 init_poll_funcptr(&poll->table, vhost_poll_func);
82 vhost_work_init(&poll->work, fn);
85 /* Start polling a file. We add ourselves to file's wait queue. The caller must
86 * keep a reference to a file until after vhost_poll_stop is called. */
87 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
95 mask = file->f_op->poll(file, &poll->table);
97 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
100 remove_wait_queue(poll->wqh, &poll->wait);
107 /* Stop polling a file. After this function returns, it becomes safe to drop the
108 * file reference. You must also flush afterwards. */
109 void vhost_poll_stop(struct vhost_poll *poll)
112 remove_wait_queue(poll->wqh, &poll->wait);
117 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
122 spin_lock_irq(&dev->work_lock);
123 left = seq - work->done_seq;
124 spin_unlock_irq(&dev->work_lock);
128 static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
133 spin_lock_irq(&dev->work_lock);
134 seq = work->queue_seq;
136 spin_unlock_irq(&dev->work_lock);
137 wait_event(work->done, vhost_work_seq_done(dev, work, seq));
138 spin_lock_irq(&dev->work_lock);
139 flushing = --work->flushing;
140 spin_unlock_irq(&dev->work_lock);
141 BUG_ON(flushing < 0);
144 /* Flush any work that has been scheduled. When calling this, don't hold any
145 * locks that are also used by the callback. */
146 void vhost_poll_flush(struct vhost_poll *poll)
148 vhost_work_flush(poll->dev, &poll->work);
151 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
155 spin_lock_irqsave(&dev->work_lock, flags);
156 if (list_empty(&work->node)) {
157 list_add_tail(&work->node, &dev->work_list);
159 wake_up_process(dev->worker);
161 spin_unlock_irqrestore(&dev->work_lock, flags);
164 void vhost_poll_queue(struct vhost_poll *poll)
166 vhost_work_queue(poll->dev, &poll->work);
169 static void vhost_vq_reset(struct vhost_dev *dev,
170 struct vhost_virtqueue *vq)
176 vq->last_avail_idx = 0;
178 vq->last_used_idx = 0;
179 vq->signalled_used = 0;
180 vq->signalled_used_valid = false;
182 vq->log_used = false;
183 vq->log_addr = -1ull;
186 vq->private_data = NULL;
188 vq->error_ctx = NULL;
199 static int vhost_worker(void *data)
201 struct vhost_dev *dev = data;
202 struct vhost_work *work = NULL;
203 unsigned uninitialized_var(seq);
204 mm_segment_t oldfs = get_fs();
210 /* mb paired w/ kthread_stop */
211 set_current_state(TASK_INTERRUPTIBLE);
213 spin_lock_irq(&dev->work_lock);
215 work->done_seq = seq;
217 wake_up_all(&work->done);
220 if (kthread_should_stop()) {
221 spin_unlock_irq(&dev->work_lock);
222 __set_current_state(TASK_RUNNING);
225 if (!list_empty(&dev->work_list)) {
226 work = list_first_entry(&dev->work_list,
227 struct vhost_work, node);
228 list_del_init(&work->node);
229 seq = work->queue_seq;
232 spin_unlock_irq(&dev->work_lock);
235 __set_current_state(TASK_RUNNING);
248 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
256 kfree(vq->ubuf_info);
257 vq->ubuf_info = NULL;
260 void vhost_enable_zcopy(int vq)
262 vhost_zcopy_mask |= 0x1 << vq;
265 /* Helper to allocate iovec buffers for all vqs. */
266 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
271 for (i = 0; i < dev->nvqs; ++i) {
272 dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect *
273 UIO_MAXIOV, GFP_KERNEL);
274 dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV,
276 dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads *
277 UIO_MAXIOV, GFP_KERNEL);
278 zcopy = vhost_zcopy_mask & (0x1 << i);
280 dev->vqs[i].ubuf_info =
281 kmalloc(sizeof *dev->vqs[i].ubuf_info *
282 UIO_MAXIOV, GFP_KERNEL);
283 if (!dev->vqs[i].indirect || !dev->vqs[i].log ||
284 !dev->vqs[i].heads ||
285 (zcopy && !dev->vqs[i].ubuf_info))
292 vhost_vq_free_iovecs(&dev->vqs[i]);
296 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
300 for (i = 0; i < dev->nvqs; ++i)
301 vhost_vq_free_iovecs(&dev->vqs[i]);
304 long vhost_dev_init(struct vhost_dev *dev,
305 struct vhost_virtqueue *vqs, int nvqs)
311 mutex_init(&dev->mutex);
313 dev->log_file = NULL;
316 spin_lock_init(&dev->work_lock);
317 INIT_LIST_HEAD(&dev->work_list);
320 for (i = 0; i < dev->nvqs; ++i) {
321 dev->vqs[i].log = NULL;
322 dev->vqs[i].indirect = NULL;
323 dev->vqs[i].heads = NULL;
324 dev->vqs[i].ubuf_info = NULL;
325 dev->vqs[i].dev = dev;
326 mutex_init(&dev->vqs[i].mutex);
327 vhost_vq_reset(dev, dev->vqs + i);
328 if (dev->vqs[i].handle_kick)
329 vhost_poll_init(&dev->vqs[i].poll,
330 dev->vqs[i].handle_kick, POLLIN, dev);
336 /* Caller should have device mutex */
337 long vhost_dev_check_owner(struct vhost_dev *dev)
339 /* Are you the owner? If not, I don't think you mean to do that */
340 return dev->mm == current->mm ? 0 : -EPERM;
343 struct vhost_attach_cgroups_struct {
344 struct vhost_work work;
345 struct task_struct *owner;
349 static void vhost_attach_cgroups_work(struct vhost_work *work)
351 struct vhost_attach_cgroups_struct *s;
353 s = container_of(work, struct vhost_attach_cgroups_struct, work);
354 s->ret = cgroup_attach_task_all(s->owner, current);
357 static int vhost_attach_cgroups(struct vhost_dev *dev)
359 struct vhost_attach_cgroups_struct attach;
361 attach.owner = current;
362 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
363 vhost_work_queue(dev, &attach.work);
364 vhost_work_flush(dev, &attach.work);
368 /* Caller should have device mutex */
369 static long vhost_dev_set_owner(struct vhost_dev *dev)
371 struct task_struct *worker;
374 /* Is there an owner already? */
380 /* No owner, become one */
381 dev->mm = get_task_mm(current);
382 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
383 if (IS_ERR(worker)) {
384 err = PTR_ERR(worker);
388 dev->worker = worker;
389 wake_up_process(worker); /* avoid contributing to loadavg */
391 err = vhost_attach_cgroups(dev);
395 err = vhost_dev_alloc_iovecs(dev);
401 kthread_stop(worker);
411 /* Caller should have device mutex */
412 long vhost_dev_reset_owner(struct vhost_dev *dev)
414 struct vhost_memory *memory;
416 /* Restore memory to default empty mapping. */
417 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
421 vhost_dev_cleanup(dev, true);
423 memory->nregions = 0;
424 RCU_INIT_POINTER(dev->memory, memory);
428 void vhost_dev_stop(struct vhost_dev *dev)
432 for (i = 0; i < dev->nvqs; ++i) {
433 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
434 vhost_poll_stop(&dev->vqs[i].poll);
435 vhost_poll_flush(&dev->vqs[i].poll);
440 /* Caller should have device mutex if and only if locked is set */
441 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
445 for (i = 0; i < dev->nvqs; ++i) {
446 if (dev->vqs[i].error_ctx)
447 eventfd_ctx_put(dev->vqs[i].error_ctx);
448 if (dev->vqs[i].error)
449 fput(dev->vqs[i].error);
450 if (dev->vqs[i].kick)
451 fput(dev->vqs[i].kick);
452 if (dev->vqs[i].call_ctx)
453 eventfd_ctx_put(dev->vqs[i].call_ctx);
454 if (dev->vqs[i].call)
455 fput(dev->vqs[i].call);
456 vhost_vq_reset(dev, dev->vqs + i);
458 vhost_dev_free_iovecs(dev);
460 eventfd_ctx_put(dev->log_ctx);
464 dev->log_file = NULL;
465 /* No one will access memory at this point */
466 kfree(rcu_dereference_protected(dev->memory,
468 lockdep_is_held(&dev->mutex)));
469 RCU_INIT_POINTER(dev->memory, NULL);
470 WARN_ON(!list_empty(&dev->work_list));
472 kthread_stop(dev->worker);
480 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
482 u64 a = addr / VHOST_PAGE_SIZE / 8;
484 /* Make sure 64 bit math will not overflow. */
485 if (a > ULONG_MAX - (unsigned long)log_base ||
486 a + (unsigned long)log_base > ULONG_MAX)
489 return access_ok(VERIFY_WRITE, log_base + a,
490 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
493 /* Caller should have vq mutex and device mutex. */
494 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
502 for (i = 0; i < mem->nregions; ++i) {
503 struct vhost_memory_region *m = mem->regions + i;
504 unsigned long a = m->userspace_addr;
505 if (m->memory_size > ULONG_MAX)
507 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
510 else if (log_all && !log_access_ok(log_base,
518 /* Can we switch to this memory table? */
519 /* Caller should have device mutex but not vq mutex */
520 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
525 for (i = 0; i < d->nvqs; ++i) {
527 mutex_lock(&d->vqs[i].mutex);
528 /* If ring is inactive, will check when it's enabled. */
529 if (d->vqs[i].private_data)
530 ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
534 mutex_unlock(&d->vqs[i].mutex);
541 static int vq_access_ok(struct vhost_dev *d, unsigned int num,
542 struct vring_desc __user *desc,
543 struct vring_avail __user *avail,
544 struct vring_used __user *used)
546 size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
547 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
548 access_ok(VERIFY_READ, avail,
549 sizeof *avail + num * sizeof *avail->ring + s) &&
550 access_ok(VERIFY_WRITE, used,
551 sizeof *used + num * sizeof *used->ring + s);
554 /* Can we log writes? */
555 /* Caller should have device mutex but not vq mutex */
556 int vhost_log_access_ok(struct vhost_dev *dev)
558 struct vhost_memory *mp;
560 mp = rcu_dereference_protected(dev->memory,
561 lockdep_is_held(&dev->mutex));
562 return memory_access_ok(dev, mp, 1);
565 /* Verify access for write logging. */
566 /* Caller should have vq mutex and device mutex */
567 static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
568 void __user *log_base)
570 struct vhost_memory *mp;
571 size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
573 mp = rcu_dereference_protected(vq->dev->memory,
574 lockdep_is_held(&vq->mutex));
575 return vq_memory_access_ok(log_base, mp,
576 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
577 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
579 vq->num * sizeof *vq->used->ring + s));
582 /* Can we start vq? */
583 /* Caller should have vq mutex and device mutex */
584 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
586 return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
587 vq_log_access_ok(vq->dev, vq, vq->log_base);
590 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
592 struct vhost_memory mem, *newmem, *oldmem;
593 unsigned long size = offsetof(struct vhost_memory, regions);
595 if (copy_from_user(&mem, m, size))
599 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
601 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
605 memcpy(newmem, &mem, size);
606 if (copy_from_user(newmem->regions, m->regions,
607 mem.nregions * sizeof *m->regions)) {
612 if (!memory_access_ok(d, newmem,
613 vhost_has_feature(d, VHOST_F_LOG_ALL))) {
617 oldmem = rcu_dereference_protected(d->memory,
618 lockdep_is_held(&d->mutex));
619 rcu_assign_pointer(d->memory, newmem);
625 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
627 struct file *eventfp, *filep = NULL;
628 bool pollstart = false, pollstop = false;
629 struct eventfd_ctx *ctx = NULL;
630 u32 __user *idxp = argp;
631 struct vhost_virtqueue *vq;
632 struct vhost_vring_state s;
633 struct vhost_vring_file f;
634 struct vhost_vring_addr a;
638 r = get_user(idx, idxp);
646 mutex_lock(&vq->mutex);
649 case VHOST_SET_VRING_NUM:
650 /* Resizing ring with an active backend?
651 * You don't want to do that. */
652 if (vq->private_data) {
656 if (copy_from_user(&s, argp, sizeof s)) {
660 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
666 case VHOST_SET_VRING_BASE:
667 /* Moving base with an active backend?
668 * You don't want to do that. */
669 if (vq->private_data) {
673 if (copy_from_user(&s, argp, sizeof s)) {
677 if (s.num > 0xffff) {
681 vq->last_avail_idx = s.num;
682 /* Forget the cached index value. */
683 vq->avail_idx = vq->last_avail_idx;
685 case VHOST_GET_VRING_BASE:
687 s.num = vq->last_avail_idx;
688 if (copy_to_user(argp, &s, sizeof s))
691 case VHOST_SET_VRING_ADDR:
692 if (copy_from_user(&a, argp, sizeof a)) {
696 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
700 /* For 32bit, verify that the top 32bits of the user
701 data are set to zero. */
702 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
703 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
704 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
708 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
709 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
710 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
715 /* We only verify access here if backend is configured.
716 * If it is not, we don't as size might not have been setup.
717 * We will verify when backend is configured. */
718 if (vq->private_data) {
719 if (!vq_access_ok(d, vq->num,
720 (void __user *)(unsigned long)a.desc_user_addr,
721 (void __user *)(unsigned long)a.avail_user_addr,
722 (void __user *)(unsigned long)a.used_user_addr)) {
727 /* Also validate log access for used ring if enabled. */
728 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
729 !log_access_ok(vq->log_base, a.log_guest_addr,
731 vq->num * sizeof *vq->used->ring)) {
737 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
738 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
739 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
740 vq->log_addr = a.log_guest_addr;
741 vq->used = (void __user *)(unsigned long)a.used_user_addr;
743 case VHOST_SET_VRING_KICK:
744 if (copy_from_user(&f, argp, sizeof f)) {
748 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
749 if (IS_ERR(eventfp)) {
750 r = PTR_ERR(eventfp);
753 if (eventfp != vq->kick) {
754 pollstop = (filep = vq->kick) != NULL;
755 pollstart = (vq->kick = eventfp) != NULL;
759 case VHOST_SET_VRING_CALL:
760 if (copy_from_user(&f, argp, sizeof f)) {
764 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
765 if (IS_ERR(eventfp)) {
766 r = PTR_ERR(eventfp);
769 if (eventfp != vq->call) {
773 vq->call_ctx = eventfp ?
774 eventfd_ctx_fileget(eventfp) : NULL;
778 case VHOST_SET_VRING_ERR:
779 if (copy_from_user(&f, argp, sizeof f)) {
783 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
784 if (IS_ERR(eventfp)) {
785 r = PTR_ERR(eventfp);
788 if (eventfp != vq->error) {
792 vq->error_ctx = eventfp ?
793 eventfd_ctx_fileget(eventfp) : NULL;
801 if (pollstop && vq->handle_kick)
802 vhost_poll_stop(&vq->poll);
805 eventfd_ctx_put(ctx);
809 if (pollstart && vq->handle_kick)
810 r = vhost_poll_start(&vq->poll, vq->kick);
812 mutex_unlock(&vq->mutex);
814 if (pollstop && vq->handle_kick)
815 vhost_poll_flush(&vq->poll);
819 /* Caller must have device mutex */
820 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
822 struct file *eventfp, *filep = NULL;
823 struct eventfd_ctx *ctx = NULL;
828 /* If you are not the owner, you can become one */
829 if (ioctl == VHOST_SET_OWNER) {
830 r = vhost_dev_set_owner(d);
834 /* You must be the owner to do anything else */
835 r = vhost_dev_check_owner(d);
840 case VHOST_SET_MEM_TABLE:
841 r = vhost_set_memory(d, argp);
843 case VHOST_SET_LOG_BASE:
844 if (copy_from_user(&p, argp, sizeof p)) {
848 if ((u64)(unsigned long)p != p) {
852 for (i = 0; i < d->nvqs; ++i) {
853 struct vhost_virtqueue *vq;
854 void __user *base = (void __user *)(unsigned long)p;
856 mutex_lock(&vq->mutex);
857 /* If ring is inactive, will check when it's enabled. */
858 if (vq->private_data && !vq_log_access_ok(d, vq, base))
862 mutex_unlock(&vq->mutex);
865 case VHOST_SET_LOG_FD:
866 r = get_user(fd, (int __user *)argp);
869 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
870 if (IS_ERR(eventfp)) {
871 r = PTR_ERR(eventfp);
874 if (eventfp != d->log_file) {
877 d->log_ctx = eventfp ?
878 eventfd_ctx_fileget(eventfp) : NULL;
881 for (i = 0; i < d->nvqs; ++i) {
882 mutex_lock(&d->vqs[i].mutex);
883 d->vqs[i].log_ctx = d->log_ctx;
884 mutex_unlock(&d->vqs[i].mutex);
887 eventfd_ctx_put(ctx);
899 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
900 __u64 addr, __u32 len)
902 struct vhost_memory_region *reg;
905 /* linear search is not brilliant, but we really have on the order of 6
906 * regions in practice */
907 for (i = 0; i < mem->nregions; ++i) {
908 reg = mem->regions + i;
909 if (reg->guest_phys_addr <= addr &&
910 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
916 /* TODO: This is really inefficient. We need something like get_user()
917 * (instruction directly accesses the data, with an exception table entry
918 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
920 static int set_bit_to_user(int nr, void __user *addr)
922 unsigned long log = (unsigned long)addr;
925 int bit = nr + (log % PAGE_SIZE) * 8;
928 r = get_user_pages_fast(log, 1, 1, &page);
932 base = kmap_atomic(page);
935 set_page_dirty_lock(page);
940 static int log_write(void __user *log_base,
941 u64 write_address, u64 write_length)
943 u64 write_page = write_address / VHOST_PAGE_SIZE;
948 write_length += write_address % VHOST_PAGE_SIZE;
950 u64 base = (u64)(unsigned long)log_base;
951 u64 log = base + write_page / 8;
952 int bit = write_page % 8;
953 if ((u64)(unsigned long)log != log)
955 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
958 if (write_length <= VHOST_PAGE_SIZE)
960 write_length -= VHOST_PAGE_SIZE;
966 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
967 unsigned int log_num, u64 len)
971 /* Make sure data written is seen before log. */
973 for (i = 0; i < log_num; ++i) {
974 u64 l = min(log[i].len, len);
975 r = log_write(vq->log_base, log[i].addr, l);
981 eventfd_signal(vq->log_ctx, 1);
985 /* Length written exceeds what we have stored. This is a bug. */
990 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
993 if (__put_user(vq->used_flags, &vq->used->flags) < 0)
995 if (unlikely(vq->log_used)) {
996 /* Make sure the flag is seen before log. */
998 /* Log used flag write. */
999 used = &vq->used->flags;
1000 log_write(vq->log_base, vq->log_addr +
1001 (used - (void __user *)vq->used),
1002 sizeof vq->used->flags);
1004 eventfd_signal(vq->log_ctx, 1);
1009 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1011 if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1013 if (unlikely(vq->log_used)) {
1015 /* Make sure the event is seen before log. */
1017 /* Log avail event write */
1018 used = vhost_avail_event(vq);
1019 log_write(vq->log_base, vq->log_addr +
1020 (used - (void __user *)vq->used),
1021 sizeof *vhost_avail_event(vq));
1023 eventfd_signal(vq->log_ctx, 1);
1028 int vhost_init_used(struct vhost_virtqueue *vq)
1031 if (!vq->private_data)
1034 r = vhost_update_used_flags(vq);
1037 vq->signalled_used_valid = false;
1038 return get_user(vq->last_used_idx, &vq->used->idx);
1041 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1042 struct iovec iov[], int iov_size)
1044 const struct vhost_memory_region *reg;
1045 struct vhost_memory *mem;
1052 mem = rcu_dereference(dev->memory);
1053 while ((u64)len > s) {
1055 if (unlikely(ret >= iov_size)) {
1059 reg = find_region(mem, addr, len);
1060 if (unlikely(!reg)) {
1065 size = reg->memory_size - addr + reg->guest_phys_addr;
1066 _iov->iov_len = min((u64)len - s, size);
1067 _iov->iov_base = (void __user *)(unsigned long)
1068 (reg->userspace_addr + addr - reg->guest_phys_addr);
1078 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1079 * function returns the next descriptor in the chain,
1080 * or -1U if we're at the end. */
1081 static unsigned next_desc(struct vring_desc *desc)
1085 /* If this descriptor says it doesn't chain, we're done. */
1086 if (!(desc->flags & VRING_DESC_F_NEXT))
1089 /* Check they're not leading us off end of descriptors. */
1091 /* Make sure compiler knows to grab that: we don't want it changing! */
1092 /* We will use the result as an index in an array, so most
1093 * architectures only need a compiler barrier here. */
1094 read_barrier_depends();
1099 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1100 struct iovec iov[], unsigned int iov_size,
1101 unsigned int *out_num, unsigned int *in_num,
1102 struct vhost_log *log, unsigned int *log_num,
1103 struct vring_desc *indirect)
1105 struct vring_desc desc;
1106 unsigned int i = 0, count, found = 0;
1110 if (unlikely(indirect->len % sizeof desc)) {
1111 vq_err(vq, "Invalid length in indirect descriptor: "
1112 "len 0x%llx not multiple of 0x%zx\n",
1113 (unsigned long long)indirect->len,
1118 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1120 if (unlikely(ret < 0)) {
1121 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1125 /* We will use the result as an address to read from, so most
1126 * architectures only need a compiler barrier here. */
1127 read_barrier_depends();
1129 count = indirect->len / sizeof desc;
1130 /* Buffers are chained via a 16 bit next field, so
1131 * we can have at most 2^16 of these. */
1132 if (unlikely(count > USHRT_MAX + 1)) {
1133 vq_err(vq, "Indirect buffer length too big: %d\n",
1139 unsigned iov_count = *in_num + *out_num;
1140 if (unlikely(++found > count)) {
1141 vq_err(vq, "Loop detected: last one at %u "
1142 "indirect size %u\n",
1146 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1147 vq->indirect, sizeof desc))) {
1148 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1149 i, (size_t)indirect->addr + i * sizeof desc);
1152 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1153 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1154 i, (size_t)indirect->addr + i * sizeof desc);
1158 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1159 iov_size - iov_count);
1160 if (unlikely(ret < 0)) {
1161 vq_err(vq, "Translation failure %d indirect idx %d\n",
1165 /* If this is an input descriptor, increment that count. */
1166 if (desc.flags & VRING_DESC_F_WRITE) {
1168 if (unlikely(log)) {
1169 log[*log_num].addr = desc.addr;
1170 log[*log_num].len = desc.len;
1174 /* If it's an output descriptor, they're all supposed
1175 * to come before any input descriptors. */
1176 if (unlikely(*in_num)) {
1177 vq_err(vq, "Indirect descriptor "
1178 "has out after in: idx %d\n", i);
1183 } while ((i = next_desc(&desc)) != -1);
1187 /* This looks in the virtqueue and for the first available buffer, and converts
1188 * it to an iovec for convenient access. Since descriptors consist of some
1189 * number of output then some number of input descriptors, it's actually two
1190 * iovecs, but we pack them into one and note how many of each there were.
1192 * This function returns the descriptor number found, or vq->num (which is
1193 * never a valid descriptor number) if none was found. A negative code is
1194 * returned on error. */
1195 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1196 struct iovec iov[], unsigned int iov_size,
1197 unsigned int *out_num, unsigned int *in_num,
1198 struct vhost_log *log, unsigned int *log_num)
1200 struct vring_desc desc;
1201 unsigned int i, head, found = 0;
1205 /* Check it isn't doing very strange things with descriptor numbers. */
1206 last_avail_idx = vq->last_avail_idx;
1207 if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1208 vq_err(vq, "Failed to access avail idx at %p\n",
1213 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1214 vq_err(vq, "Guest moved used index from %u to %u",
1215 last_avail_idx, vq->avail_idx);
1219 /* If there's nothing new since last we looked, return invalid. */
1220 if (vq->avail_idx == last_avail_idx)
1223 /* Only get avail ring entries after they have been exposed by guest. */
1226 /* Grab the next descriptor number they're advertising, and increment
1227 * the index we've seen. */
1228 if (unlikely(__get_user(head,
1229 &vq->avail->ring[last_avail_idx % vq->num]))) {
1230 vq_err(vq, "Failed to read head: idx %d address %p\n",
1232 &vq->avail->ring[last_avail_idx % vq->num]);
1236 /* If their number is silly, that's an error. */
1237 if (unlikely(head >= vq->num)) {
1238 vq_err(vq, "Guest says index %u > %u is available",
1243 /* When we start there are none of either input nor output. */
1244 *out_num = *in_num = 0;
1250 unsigned iov_count = *in_num + *out_num;
1251 if (unlikely(i >= vq->num)) {
1252 vq_err(vq, "Desc index is %u > %u, head = %u",
1256 if (unlikely(++found > vq->num)) {
1257 vq_err(vq, "Loop detected: last one at %u "
1258 "vq size %u head %u\n",
1262 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1263 if (unlikely(ret)) {
1264 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1268 if (desc.flags & VRING_DESC_F_INDIRECT) {
1269 ret = get_indirect(dev, vq, iov, iov_size,
1271 log, log_num, &desc);
1272 if (unlikely(ret < 0)) {
1273 vq_err(vq, "Failure detected "
1274 "in indirect descriptor at idx %d\n", i);
1280 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1281 iov_size - iov_count);
1282 if (unlikely(ret < 0)) {
1283 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1287 if (desc.flags & VRING_DESC_F_WRITE) {
1288 /* If this is an input descriptor,
1289 * increment that count. */
1291 if (unlikely(log)) {
1292 log[*log_num].addr = desc.addr;
1293 log[*log_num].len = desc.len;
1297 /* If it's an output descriptor, they're all supposed
1298 * to come before any input descriptors. */
1299 if (unlikely(*in_num)) {
1300 vq_err(vq, "Descriptor has out after in: "
1306 } while ((i = next_desc(&desc)) != -1);
1308 /* On success, increment avail index. */
1309 vq->last_avail_idx++;
1311 /* Assume notifications from guest are disabled at this point,
1312 * if they aren't we would need to update avail_event index. */
1313 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1317 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1318 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1320 vq->last_avail_idx -= n;
1323 /* After we've used one of their buffers, we tell them about it. We'll then
1324 * want to notify the guest, using eventfd. */
1325 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1327 struct vring_used_elem __user *used;
1329 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1330 * next entry in that used ring. */
1331 used = &vq->used->ring[vq->last_used_idx % vq->num];
1332 if (__put_user(head, &used->id)) {
1333 vq_err(vq, "Failed to write used id");
1336 if (__put_user(len, &used->len)) {
1337 vq_err(vq, "Failed to write used len");
1340 /* Make sure buffer is written before we update index. */
1342 if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1343 vq_err(vq, "Failed to increment used idx");
1346 if (unlikely(vq->log_used)) {
1347 /* Make sure data is seen before log. */
1349 /* Log used ring entry write. */
1350 log_write(vq->log_base,
1352 ((void __user *)used - (void __user *)vq->used),
1354 /* Log used index update. */
1355 log_write(vq->log_base,
1356 vq->log_addr + offsetof(struct vring_used, idx),
1357 sizeof vq->used->idx);
1359 eventfd_signal(vq->log_ctx, 1);
1361 vq->last_used_idx++;
1362 /* If the driver never bothers to signal in a very long while,
1363 * used index might wrap around. If that happens, invalidate
1364 * signalled_used index we stored. TODO: make sure driver
1365 * signals at least once in 2^16 and remove this. */
1366 if (unlikely(vq->last_used_idx == vq->signalled_used))
1367 vq->signalled_used_valid = false;
1371 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1372 struct vring_used_elem *heads,
1375 struct vring_used_elem __user *used;
1379 start = vq->last_used_idx % vq->num;
1380 used = vq->used->ring + start;
1381 if (__copy_to_user(used, heads, count * sizeof *used)) {
1382 vq_err(vq, "Failed to write used");
1385 if (unlikely(vq->log_used)) {
1386 /* Make sure data is seen before log. */
1388 /* Log used ring entry write. */
1389 log_write(vq->log_base,
1391 ((void __user *)used - (void __user *)vq->used),
1392 count * sizeof *used);
1394 old = vq->last_used_idx;
1395 new = (vq->last_used_idx += count);
1396 /* If the driver never bothers to signal in a very long while,
1397 * used index might wrap around. If that happens, invalidate
1398 * signalled_used index we stored. TODO: make sure driver
1399 * signals at least once in 2^16 and remove this. */
1400 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1401 vq->signalled_used_valid = false;
1405 /* After we've used one of their buffers, we tell them about it. We'll then
1406 * want to notify the guest, using eventfd. */
1407 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1412 start = vq->last_used_idx % vq->num;
1413 n = vq->num - start;
1415 r = __vhost_add_used_n(vq, heads, n);
1421 r = __vhost_add_used_n(vq, heads, count);
1423 /* Make sure buffer is written before we update index. */
1425 if (put_user(vq->last_used_idx, &vq->used->idx)) {
1426 vq_err(vq, "Failed to increment used idx");
1429 if (unlikely(vq->log_used)) {
1430 /* Log used index update. */
1431 log_write(vq->log_base,
1432 vq->log_addr + offsetof(struct vring_used, idx),
1433 sizeof vq->used->idx);
1435 eventfd_signal(vq->log_ctx, 1);
1440 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1442 __u16 old, new, event;
1444 /* Flush out used index updates. This is paired
1445 * with the barrier that the Guest executes when enabling
1449 if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1450 unlikely(vq->avail_idx == vq->last_avail_idx))
1453 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1455 if (__get_user(flags, &vq->avail->flags)) {
1456 vq_err(vq, "Failed to get flags");
1459 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1461 old = vq->signalled_used;
1462 v = vq->signalled_used_valid;
1463 new = vq->signalled_used = vq->last_used_idx;
1464 vq->signalled_used_valid = true;
1469 if (get_user(event, vhost_used_event(vq))) {
1470 vq_err(vq, "Failed to get used event idx");
1473 return vring_need_event(event, new, old);
1476 /* This actually signals the guest, using eventfd. */
1477 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1479 /* Signal the Guest tell them we used something up. */
1480 if (vq->call_ctx && vhost_notify(dev, vq))
1481 eventfd_signal(vq->call_ctx, 1);
1484 /* And here's the combo meal deal. Supersize me! */
1485 void vhost_add_used_and_signal(struct vhost_dev *dev,
1486 struct vhost_virtqueue *vq,
1487 unsigned int head, int len)
1489 vhost_add_used(vq, head, len);
1490 vhost_signal(dev, vq);
1493 /* multi-buffer version of vhost_add_used_and_signal */
1494 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1495 struct vhost_virtqueue *vq,
1496 struct vring_used_elem *heads, unsigned count)
1498 vhost_add_used_n(vq, heads, count);
1499 vhost_signal(dev, vq);
1502 /* OK, now we need to know about added descriptors. */
1503 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1508 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1510 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1511 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1512 r = vhost_update_used_flags(vq);
1514 vq_err(vq, "Failed to enable notification at %p: %d\n",
1515 &vq->used->flags, r);
1519 r = vhost_update_avail_event(vq, vq->avail_idx);
1521 vq_err(vq, "Failed to update avail event index at %p: %d\n",
1522 vhost_avail_event(vq), r);
1526 /* They could have slipped one in as we were doing that: make
1527 * sure it's written, then check again. */
1529 r = __get_user(avail_idx, &vq->avail->idx);
1531 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1532 &vq->avail->idx, r);
1536 return avail_idx != vq->avail_idx;
1539 /* We don't need to be notified again. */
1540 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1544 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1546 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1547 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1548 r = vhost_update_used_flags(vq);
1550 vq_err(vq, "Failed to enable notification at %p: %d\n",
1551 &vq->used->flags, r);
1555 static void vhost_zerocopy_done_signal(struct kref *kref)
1557 struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
1559 wake_up(&ubufs->wait);
1562 struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
1565 struct vhost_ubuf_ref *ubufs;
1566 /* No zero copy backend? Nothing to count. */
1569 ubufs = kmalloc(sizeof *ubufs, GFP_KERNEL);
1571 return ERR_PTR(-ENOMEM);
1572 kref_init(&ubufs->kref);
1573 init_waitqueue_head(&ubufs->wait);
1578 void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
1580 kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1583 void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
1585 kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1586 wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));