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 #define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num])
37 #define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num])
39 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
42 struct vhost_poll *poll;
44 poll = container_of(pt, struct vhost_poll, table);
46 add_wait_queue(wqh, &poll->wait);
49 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
52 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
54 if (!((unsigned long)key & poll->mask))
57 vhost_poll_queue(poll);
61 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
63 INIT_LIST_HEAD(&work->node);
65 init_waitqueue_head(&work->done);
67 work->queue_seq = work->done_seq = 0;
70 /* Init poll structure */
71 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
72 unsigned long mask, struct vhost_dev *dev)
74 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
75 init_poll_funcptr(&poll->table, vhost_poll_func);
80 vhost_work_init(&poll->work, fn);
83 /* Start polling a file. We add ourselves to file's wait queue. The caller must
84 * keep a reference to a file until after vhost_poll_stop is called. */
85 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
93 mask = file->f_op->poll(file, &poll->table);
95 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
98 remove_wait_queue(poll->wqh, &poll->wait);
105 /* Stop polling a file. After this function returns, it becomes safe to drop the
106 * file reference. You must also flush afterwards. */
107 void vhost_poll_stop(struct vhost_poll *poll)
110 remove_wait_queue(poll->wqh, &poll->wait);
115 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
120 spin_lock_irq(&dev->work_lock);
121 left = seq - work->done_seq;
122 spin_unlock_irq(&dev->work_lock);
126 static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
131 spin_lock_irq(&dev->work_lock);
132 seq = work->queue_seq;
134 spin_unlock_irq(&dev->work_lock);
135 wait_event(work->done, vhost_work_seq_done(dev, work, seq));
136 spin_lock_irq(&dev->work_lock);
137 flushing = --work->flushing;
138 spin_unlock_irq(&dev->work_lock);
139 BUG_ON(flushing < 0);
142 /* Flush any work that has been scheduled. When calling this, don't hold any
143 * locks that are also used by the callback. */
144 void vhost_poll_flush(struct vhost_poll *poll)
146 vhost_work_flush(poll->dev, &poll->work);
149 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
153 spin_lock_irqsave(&dev->work_lock, flags);
154 if (list_empty(&work->node)) {
155 list_add_tail(&work->node, &dev->work_list);
157 wake_up_process(dev->worker);
159 spin_unlock_irqrestore(&dev->work_lock, flags);
162 void vhost_poll_queue(struct vhost_poll *poll)
164 vhost_work_queue(poll->dev, &poll->work);
167 static void vhost_vq_reset(struct vhost_dev *dev,
168 struct vhost_virtqueue *vq)
174 vq->last_avail_idx = 0;
176 vq->last_used_idx = 0;
177 vq->signalled_used = 0;
178 vq->signalled_used_valid = false;
180 vq->log_used = false;
181 vq->log_addr = -1ull;
182 vq->private_data = NULL;
184 vq->error_ctx = NULL;
192 static int vhost_worker(void *data)
194 struct vhost_dev *dev = data;
195 struct vhost_work *work = NULL;
196 unsigned uninitialized_var(seq);
197 mm_segment_t oldfs = get_fs();
203 /* mb paired w/ kthread_stop */
204 set_current_state(TASK_INTERRUPTIBLE);
206 spin_lock_irq(&dev->work_lock);
208 work->done_seq = seq;
210 wake_up_all(&work->done);
213 if (kthread_should_stop()) {
214 spin_unlock_irq(&dev->work_lock);
215 __set_current_state(TASK_RUNNING);
218 if (!list_empty(&dev->work_list)) {
219 work = list_first_entry(&dev->work_list,
220 struct vhost_work, node);
221 list_del_init(&work->node);
222 seq = work->queue_seq;
225 spin_unlock_irq(&dev->work_lock);
228 __set_current_state(TASK_RUNNING);
241 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
251 /* Helper to allocate iovec buffers for all vqs. */
252 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
256 for (i = 0; i < dev->nvqs; ++i) {
257 dev->vqs[i]->indirect = kmalloc(sizeof *dev->vqs[i]->indirect *
258 UIO_MAXIOV, GFP_KERNEL);
259 dev->vqs[i]->log = kmalloc(sizeof *dev->vqs[i]->log * UIO_MAXIOV,
261 dev->vqs[i]->heads = kmalloc(sizeof *dev->vqs[i]->heads *
262 UIO_MAXIOV, GFP_KERNEL);
263 if (!dev->vqs[i]->indirect || !dev->vqs[i]->log ||
271 vhost_vq_free_iovecs(dev->vqs[i]);
275 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
279 for (i = 0; i < dev->nvqs; ++i)
280 vhost_vq_free_iovecs(dev->vqs[i]);
283 long vhost_dev_init(struct vhost_dev *dev,
284 struct vhost_virtqueue **vqs, int nvqs)
290 mutex_init(&dev->mutex);
292 dev->log_file = NULL;
295 spin_lock_init(&dev->work_lock);
296 INIT_LIST_HEAD(&dev->work_list);
299 for (i = 0; i < dev->nvqs; ++i) {
300 dev->vqs[i]->log = NULL;
301 dev->vqs[i]->indirect = NULL;
302 dev->vqs[i]->heads = NULL;
303 dev->vqs[i]->dev = dev;
304 mutex_init(&dev->vqs[i]->mutex);
305 vhost_vq_reset(dev, dev->vqs[i]);
306 if (dev->vqs[i]->handle_kick)
307 vhost_poll_init(&dev->vqs[i]->poll,
308 dev->vqs[i]->handle_kick, POLLIN, dev);
314 /* Caller should have device mutex */
315 long vhost_dev_check_owner(struct vhost_dev *dev)
317 /* Are you the owner? If not, I don't think you mean to do that */
318 return dev->mm == current->mm ? 0 : -EPERM;
321 struct vhost_attach_cgroups_struct {
322 struct vhost_work work;
323 struct task_struct *owner;
327 static void vhost_attach_cgroups_work(struct vhost_work *work)
329 struct vhost_attach_cgroups_struct *s;
331 s = container_of(work, struct vhost_attach_cgroups_struct, work);
332 s->ret = cgroup_attach_task_all(s->owner, current);
335 static int vhost_attach_cgroups(struct vhost_dev *dev)
337 struct vhost_attach_cgroups_struct attach;
339 attach.owner = current;
340 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
341 vhost_work_queue(dev, &attach.work);
342 vhost_work_flush(dev, &attach.work);
346 /* Caller should have device mutex */
347 static long vhost_dev_set_owner(struct vhost_dev *dev)
349 struct task_struct *worker;
352 /* Is there an owner already? */
358 /* No owner, become one */
359 dev->mm = get_task_mm(current);
360 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
361 if (IS_ERR(worker)) {
362 err = PTR_ERR(worker);
366 dev->worker = worker;
367 wake_up_process(worker); /* avoid contributing to loadavg */
369 err = vhost_attach_cgroups(dev);
373 err = vhost_dev_alloc_iovecs(dev);
379 kthread_stop(worker);
389 /* Caller should have device mutex */
390 long vhost_dev_reset_owner(struct vhost_dev *dev)
392 struct vhost_memory *memory;
394 /* Restore memory to default empty mapping. */
395 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
399 vhost_dev_cleanup(dev, true);
401 memory->nregions = 0;
402 RCU_INIT_POINTER(dev->memory, memory);
406 void vhost_dev_stop(struct vhost_dev *dev)
410 for (i = 0; i < dev->nvqs; ++i) {
411 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
412 vhost_poll_stop(&dev->vqs[i]->poll);
413 vhost_poll_flush(&dev->vqs[i]->poll);
418 /* Caller should have device mutex if and only if locked is set */
419 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
423 for (i = 0; i < dev->nvqs; ++i) {
424 if (dev->vqs[i]->error_ctx)
425 eventfd_ctx_put(dev->vqs[i]->error_ctx);
426 if (dev->vqs[i]->error)
427 fput(dev->vqs[i]->error);
428 if (dev->vqs[i]->kick)
429 fput(dev->vqs[i]->kick);
430 if (dev->vqs[i]->call_ctx)
431 eventfd_ctx_put(dev->vqs[i]->call_ctx);
432 if (dev->vqs[i]->call)
433 fput(dev->vqs[i]->call);
434 vhost_vq_reset(dev, dev->vqs[i]);
436 vhost_dev_free_iovecs(dev);
438 eventfd_ctx_put(dev->log_ctx);
442 dev->log_file = NULL;
443 /* No one will access memory at this point */
444 kfree(rcu_dereference_protected(dev->memory,
446 lockdep_is_held(&dev->mutex)));
447 RCU_INIT_POINTER(dev->memory, NULL);
448 WARN_ON(!list_empty(&dev->work_list));
450 kthread_stop(dev->worker);
458 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
460 u64 a = addr / VHOST_PAGE_SIZE / 8;
462 /* Make sure 64 bit math will not overflow. */
463 if (a > ULONG_MAX - (unsigned long)log_base ||
464 a + (unsigned long)log_base > ULONG_MAX)
467 return access_ok(VERIFY_WRITE, log_base + a,
468 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
471 /* Caller should have vq mutex and device mutex. */
472 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
480 for (i = 0; i < mem->nregions; ++i) {
481 struct vhost_memory_region *m = mem->regions + i;
482 unsigned long a = m->userspace_addr;
483 if (m->memory_size > ULONG_MAX)
485 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
488 else if (log_all && !log_access_ok(log_base,
496 /* Can we switch to this memory table? */
497 /* Caller should have device mutex but not vq mutex */
498 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
503 for (i = 0; i < d->nvqs; ++i) {
505 mutex_lock(&d->vqs[i]->mutex);
506 /* If ring is inactive, will check when it's enabled. */
507 if (d->vqs[i]->private_data)
508 ok = vq_memory_access_ok(d->vqs[i]->log_base, mem,
512 mutex_unlock(&d->vqs[i]->mutex);
519 static int vq_access_ok(struct vhost_dev *d, unsigned int num,
520 struct vring_desc __user *desc,
521 struct vring_avail __user *avail,
522 struct vring_used __user *used)
524 size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
525 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
526 access_ok(VERIFY_READ, avail,
527 sizeof *avail + num * sizeof *avail->ring + s) &&
528 access_ok(VERIFY_WRITE, used,
529 sizeof *used + num * sizeof *used->ring + s);
532 /* Can we log writes? */
533 /* Caller should have device mutex but not vq mutex */
534 int vhost_log_access_ok(struct vhost_dev *dev)
536 struct vhost_memory *mp;
538 mp = rcu_dereference_protected(dev->memory,
539 lockdep_is_held(&dev->mutex));
540 return memory_access_ok(dev, mp, 1);
543 /* Verify access for write logging. */
544 /* Caller should have vq mutex and device mutex */
545 static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
546 void __user *log_base)
548 struct vhost_memory *mp;
549 size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
551 mp = rcu_dereference_protected(vq->dev->memory,
552 lockdep_is_held(&vq->mutex));
553 return vq_memory_access_ok(log_base, mp,
554 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
555 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
557 vq->num * sizeof *vq->used->ring + s));
560 /* Can we start vq? */
561 /* Caller should have vq mutex and device mutex */
562 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
564 return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
565 vq_log_access_ok(vq->dev, vq, vq->log_base);
568 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
570 struct vhost_memory mem, *newmem, *oldmem;
571 unsigned long size = offsetof(struct vhost_memory, regions);
573 if (copy_from_user(&mem, m, size))
577 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
579 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
583 memcpy(newmem, &mem, size);
584 if (copy_from_user(newmem->regions, m->regions,
585 mem.nregions * sizeof *m->regions)) {
590 if (!memory_access_ok(d, newmem,
591 vhost_has_feature(d, VHOST_F_LOG_ALL))) {
595 oldmem = rcu_dereference_protected(d->memory,
596 lockdep_is_held(&d->mutex));
597 rcu_assign_pointer(d->memory, newmem);
603 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
605 struct file *eventfp, *filep = NULL;
606 bool pollstart = false, pollstop = false;
607 struct eventfd_ctx *ctx = NULL;
608 u32 __user *idxp = argp;
609 struct vhost_virtqueue *vq;
610 struct vhost_vring_state s;
611 struct vhost_vring_file f;
612 struct vhost_vring_addr a;
616 r = get_user(idx, idxp);
624 mutex_lock(&vq->mutex);
627 case VHOST_SET_VRING_NUM:
628 /* Resizing ring with an active backend?
629 * You don't want to do that. */
630 if (vq->private_data) {
634 if (copy_from_user(&s, argp, sizeof s)) {
638 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
644 case VHOST_SET_VRING_BASE:
645 /* Moving base with an active backend?
646 * You don't want to do that. */
647 if (vq->private_data) {
651 if (copy_from_user(&s, argp, sizeof s)) {
655 if (s.num > 0xffff) {
659 vq->last_avail_idx = s.num;
660 /* Forget the cached index value. */
661 vq->avail_idx = vq->last_avail_idx;
663 case VHOST_GET_VRING_BASE:
665 s.num = vq->last_avail_idx;
666 if (copy_to_user(argp, &s, sizeof s))
669 case VHOST_SET_VRING_ADDR:
670 if (copy_from_user(&a, argp, sizeof a)) {
674 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
678 /* For 32bit, verify that the top 32bits of the user
679 data are set to zero. */
680 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
681 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
682 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
686 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
687 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
688 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
693 /* We only verify access here if backend is configured.
694 * If it is not, we don't as size might not have been setup.
695 * We will verify when backend is configured. */
696 if (vq->private_data) {
697 if (!vq_access_ok(d, vq->num,
698 (void __user *)(unsigned long)a.desc_user_addr,
699 (void __user *)(unsigned long)a.avail_user_addr,
700 (void __user *)(unsigned long)a.used_user_addr)) {
705 /* Also validate log access for used ring if enabled. */
706 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
707 !log_access_ok(vq->log_base, a.log_guest_addr,
709 vq->num * sizeof *vq->used->ring)) {
715 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
716 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
717 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
718 vq->log_addr = a.log_guest_addr;
719 vq->used = (void __user *)(unsigned long)a.used_user_addr;
721 case VHOST_SET_VRING_KICK:
722 if (copy_from_user(&f, argp, sizeof f)) {
726 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
727 if (IS_ERR(eventfp)) {
728 r = PTR_ERR(eventfp);
731 if (eventfp != vq->kick) {
732 pollstop = (filep = vq->kick) != NULL;
733 pollstart = (vq->kick = eventfp) != NULL;
737 case VHOST_SET_VRING_CALL:
738 if (copy_from_user(&f, argp, sizeof f)) {
742 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
743 if (IS_ERR(eventfp)) {
744 r = PTR_ERR(eventfp);
747 if (eventfp != vq->call) {
751 vq->call_ctx = eventfp ?
752 eventfd_ctx_fileget(eventfp) : NULL;
756 case VHOST_SET_VRING_ERR:
757 if (copy_from_user(&f, argp, sizeof f)) {
761 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
762 if (IS_ERR(eventfp)) {
763 r = PTR_ERR(eventfp);
766 if (eventfp != vq->error) {
770 vq->error_ctx = eventfp ?
771 eventfd_ctx_fileget(eventfp) : NULL;
779 if (pollstop && vq->handle_kick)
780 vhost_poll_stop(&vq->poll);
783 eventfd_ctx_put(ctx);
787 if (pollstart && vq->handle_kick)
788 r = vhost_poll_start(&vq->poll, vq->kick);
790 mutex_unlock(&vq->mutex);
792 if (pollstop && vq->handle_kick)
793 vhost_poll_flush(&vq->poll);
797 /* Caller must have device mutex */
798 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
800 struct file *eventfp, *filep = NULL;
801 struct eventfd_ctx *ctx = NULL;
806 /* If you are not the owner, you can become one */
807 if (ioctl == VHOST_SET_OWNER) {
808 r = vhost_dev_set_owner(d);
812 /* You must be the owner to do anything else */
813 r = vhost_dev_check_owner(d);
818 case VHOST_SET_MEM_TABLE:
819 r = vhost_set_memory(d, argp);
821 case VHOST_SET_LOG_BASE:
822 if (copy_from_user(&p, argp, sizeof p)) {
826 if ((u64)(unsigned long)p != p) {
830 for (i = 0; i < d->nvqs; ++i) {
831 struct vhost_virtqueue *vq;
832 void __user *base = (void __user *)(unsigned long)p;
834 mutex_lock(&vq->mutex);
835 /* If ring is inactive, will check when it's enabled. */
836 if (vq->private_data && !vq_log_access_ok(d, vq, base))
840 mutex_unlock(&vq->mutex);
843 case VHOST_SET_LOG_FD:
844 r = get_user(fd, (int __user *)argp);
847 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
848 if (IS_ERR(eventfp)) {
849 r = PTR_ERR(eventfp);
852 if (eventfp != d->log_file) {
855 d->log_ctx = eventfp ?
856 eventfd_ctx_fileget(eventfp) : NULL;
859 for (i = 0; i < d->nvqs; ++i) {
860 mutex_lock(&d->vqs[i]->mutex);
861 d->vqs[i]->log_ctx = d->log_ctx;
862 mutex_unlock(&d->vqs[i]->mutex);
865 eventfd_ctx_put(ctx);
877 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
878 __u64 addr, __u32 len)
880 struct vhost_memory_region *reg;
883 /* linear search is not brilliant, but we really have on the order of 6
884 * regions in practice */
885 for (i = 0; i < mem->nregions; ++i) {
886 reg = mem->regions + i;
887 if (reg->guest_phys_addr <= addr &&
888 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
894 /* TODO: This is really inefficient. We need something like get_user()
895 * (instruction directly accesses the data, with an exception table entry
896 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
898 static int set_bit_to_user(int nr, void __user *addr)
900 unsigned long log = (unsigned long)addr;
903 int bit = nr + (log % PAGE_SIZE) * 8;
906 r = get_user_pages_fast(log, 1, 1, &page);
910 base = kmap_atomic(page);
913 set_page_dirty_lock(page);
918 static int log_write(void __user *log_base,
919 u64 write_address, u64 write_length)
921 u64 write_page = write_address / VHOST_PAGE_SIZE;
926 write_length += write_address % VHOST_PAGE_SIZE;
928 u64 base = (u64)(unsigned long)log_base;
929 u64 log = base + write_page / 8;
930 int bit = write_page % 8;
931 if ((u64)(unsigned long)log != log)
933 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
936 if (write_length <= VHOST_PAGE_SIZE)
938 write_length -= VHOST_PAGE_SIZE;
944 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
945 unsigned int log_num, u64 len)
949 /* Make sure data written is seen before log. */
951 for (i = 0; i < log_num; ++i) {
952 u64 l = min(log[i].len, len);
953 r = log_write(vq->log_base, log[i].addr, l);
959 eventfd_signal(vq->log_ctx, 1);
963 /* Length written exceeds what we have stored. This is a bug. */
968 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
971 if (__put_user(vq->used_flags, &vq->used->flags) < 0)
973 if (unlikely(vq->log_used)) {
974 /* Make sure the flag is seen before log. */
976 /* Log used flag write. */
977 used = &vq->used->flags;
978 log_write(vq->log_base, vq->log_addr +
979 (used - (void __user *)vq->used),
980 sizeof vq->used->flags);
982 eventfd_signal(vq->log_ctx, 1);
987 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
989 if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
991 if (unlikely(vq->log_used)) {
993 /* Make sure the event is seen before log. */
995 /* Log avail event write */
996 used = vhost_avail_event(vq);
997 log_write(vq->log_base, vq->log_addr +
998 (used - (void __user *)vq->used),
999 sizeof *vhost_avail_event(vq));
1001 eventfd_signal(vq->log_ctx, 1);
1006 int vhost_init_used(struct vhost_virtqueue *vq)
1009 if (!vq->private_data)
1012 r = vhost_update_used_flags(vq);
1015 vq->signalled_used_valid = false;
1016 return get_user(vq->last_used_idx, &vq->used->idx);
1019 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1020 struct iovec iov[], int iov_size)
1022 const struct vhost_memory_region *reg;
1023 struct vhost_memory *mem;
1030 mem = rcu_dereference(dev->memory);
1031 while ((u64)len > s) {
1033 if (unlikely(ret >= iov_size)) {
1037 reg = find_region(mem, addr, len);
1038 if (unlikely(!reg)) {
1043 size = reg->memory_size - addr + reg->guest_phys_addr;
1044 _iov->iov_len = min((u64)len - s, size);
1045 _iov->iov_base = (void __user *)(unsigned long)
1046 (reg->userspace_addr + addr - reg->guest_phys_addr);
1056 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1057 * function returns the next descriptor in the chain,
1058 * or -1U if we're at the end. */
1059 static unsigned next_desc(struct vring_desc *desc)
1063 /* If this descriptor says it doesn't chain, we're done. */
1064 if (!(desc->flags & VRING_DESC_F_NEXT))
1067 /* Check they're not leading us off end of descriptors. */
1069 /* Make sure compiler knows to grab that: we don't want it changing! */
1070 /* We will use the result as an index in an array, so most
1071 * architectures only need a compiler barrier here. */
1072 read_barrier_depends();
1077 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1078 struct iovec iov[], unsigned int iov_size,
1079 unsigned int *out_num, unsigned int *in_num,
1080 struct vhost_log *log, unsigned int *log_num,
1081 struct vring_desc *indirect)
1083 struct vring_desc desc;
1084 unsigned int i = 0, count, found = 0;
1088 if (unlikely(indirect->len % sizeof desc)) {
1089 vq_err(vq, "Invalid length in indirect descriptor: "
1090 "len 0x%llx not multiple of 0x%zx\n",
1091 (unsigned long long)indirect->len,
1096 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1098 if (unlikely(ret < 0)) {
1099 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1103 /* We will use the result as an address to read from, so most
1104 * architectures only need a compiler barrier here. */
1105 read_barrier_depends();
1107 count = indirect->len / sizeof desc;
1108 /* Buffers are chained via a 16 bit next field, so
1109 * we can have at most 2^16 of these. */
1110 if (unlikely(count > USHRT_MAX + 1)) {
1111 vq_err(vq, "Indirect buffer length too big: %d\n",
1117 unsigned iov_count = *in_num + *out_num;
1118 if (unlikely(++found > count)) {
1119 vq_err(vq, "Loop detected: last one at %u "
1120 "indirect size %u\n",
1124 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1125 vq->indirect, sizeof desc))) {
1126 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1127 i, (size_t)indirect->addr + i * sizeof desc);
1130 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1131 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1132 i, (size_t)indirect->addr + i * sizeof desc);
1136 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1137 iov_size - iov_count);
1138 if (unlikely(ret < 0)) {
1139 vq_err(vq, "Translation failure %d indirect idx %d\n",
1143 /* If this is an input descriptor, increment that count. */
1144 if (desc.flags & VRING_DESC_F_WRITE) {
1146 if (unlikely(log)) {
1147 log[*log_num].addr = desc.addr;
1148 log[*log_num].len = desc.len;
1152 /* If it's an output descriptor, they're all supposed
1153 * to come before any input descriptors. */
1154 if (unlikely(*in_num)) {
1155 vq_err(vq, "Indirect descriptor "
1156 "has out after in: idx %d\n", i);
1161 } while ((i = next_desc(&desc)) != -1);
1165 /* This looks in the virtqueue and for the first available buffer, and converts
1166 * it to an iovec for convenient access. Since descriptors consist of some
1167 * number of output then some number of input descriptors, it's actually two
1168 * iovecs, but we pack them into one and note how many of each there were.
1170 * This function returns the descriptor number found, or vq->num (which is
1171 * never a valid descriptor number) if none was found. A negative code is
1172 * returned on error. */
1173 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1174 struct iovec iov[], unsigned int iov_size,
1175 unsigned int *out_num, unsigned int *in_num,
1176 struct vhost_log *log, unsigned int *log_num)
1178 struct vring_desc desc;
1179 unsigned int i, head, found = 0;
1183 /* Check it isn't doing very strange things with descriptor numbers. */
1184 last_avail_idx = vq->last_avail_idx;
1185 if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1186 vq_err(vq, "Failed to access avail idx at %p\n",
1191 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1192 vq_err(vq, "Guest moved used index from %u to %u",
1193 last_avail_idx, vq->avail_idx);
1197 /* If there's nothing new since last we looked, return invalid. */
1198 if (vq->avail_idx == last_avail_idx)
1201 /* Only get avail ring entries after they have been exposed by guest. */
1204 /* Grab the next descriptor number they're advertising, and increment
1205 * the index we've seen. */
1206 if (unlikely(__get_user(head,
1207 &vq->avail->ring[last_avail_idx % vq->num]))) {
1208 vq_err(vq, "Failed to read head: idx %d address %p\n",
1210 &vq->avail->ring[last_avail_idx % vq->num]);
1214 /* If their number is silly, that's an error. */
1215 if (unlikely(head >= vq->num)) {
1216 vq_err(vq, "Guest says index %u > %u is available",
1221 /* When we start there are none of either input nor output. */
1222 *out_num = *in_num = 0;
1228 unsigned iov_count = *in_num + *out_num;
1229 if (unlikely(i >= vq->num)) {
1230 vq_err(vq, "Desc index is %u > %u, head = %u",
1234 if (unlikely(++found > vq->num)) {
1235 vq_err(vq, "Loop detected: last one at %u "
1236 "vq size %u head %u\n",
1240 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1241 if (unlikely(ret)) {
1242 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1246 if (desc.flags & VRING_DESC_F_INDIRECT) {
1247 ret = get_indirect(dev, vq, iov, iov_size,
1249 log, log_num, &desc);
1250 if (unlikely(ret < 0)) {
1251 vq_err(vq, "Failure detected "
1252 "in indirect descriptor at idx %d\n", i);
1258 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1259 iov_size - iov_count);
1260 if (unlikely(ret < 0)) {
1261 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1265 if (desc.flags & VRING_DESC_F_WRITE) {
1266 /* If this is an input descriptor,
1267 * increment that count. */
1269 if (unlikely(log)) {
1270 log[*log_num].addr = desc.addr;
1271 log[*log_num].len = desc.len;
1275 /* If it's an output descriptor, they're all supposed
1276 * to come before any input descriptors. */
1277 if (unlikely(*in_num)) {
1278 vq_err(vq, "Descriptor has out after in: "
1284 } while ((i = next_desc(&desc)) != -1);
1286 /* On success, increment avail index. */
1287 vq->last_avail_idx++;
1289 /* Assume notifications from guest are disabled at this point,
1290 * if they aren't we would need to update avail_event index. */
1291 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1295 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1296 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1298 vq->last_avail_idx -= n;
1301 /* After we've used one of their buffers, we tell them about it. We'll then
1302 * want to notify the guest, using eventfd. */
1303 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1305 struct vring_used_elem __user *used;
1307 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1308 * next entry in that used ring. */
1309 used = &vq->used->ring[vq->last_used_idx % vq->num];
1310 if (__put_user(head, &used->id)) {
1311 vq_err(vq, "Failed to write used id");
1314 if (__put_user(len, &used->len)) {
1315 vq_err(vq, "Failed to write used len");
1318 /* Make sure buffer is written before we update index. */
1320 if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1321 vq_err(vq, "Failed to increment used idx");
1324 if (unlikely(vq->log_used)) {
1325 /* Make sure data is seen before log. */
1327 /* Log used ring entry write. */
1328 log_write(vq->log_base,
1330 ((void __user *)used - (void __user *)vq->used),
1332 /* Log used index update. */
1333 log_write(vq->log_base,
1334 vq->log_addr + offsetof(struct vring_used, idx),
1335 sizeof vq->used->idx);
1337 eventfd_signal(vq->log_ctx, 1);
1339 vq->last_used_idx++;
1340 /* If the driver never bothers to signal in a very long while,
1341 * used index might wrap around. If that happens, invalidate
1342 * signalled_used index we stored. TODO: make sure driver
1343 * signals at least once in 2^16 and remove this. */
1344 if (unlikely(vq->last_used_idx == vq->signalled_used))
1345 vq->signalled_used_valid = false;
1349 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1350 struct vring_used_elem *heads,
1353 struct vring_used_elem __user *used;
1357 start = vq->last_used_idx % vq->num;
1358 used = vq->used->ring + start;
1359 if (__copy_to_user(used, heads, count * sizeof *used)) {
1360 vq_err(vq, "Failed to write used");
1363 if (unlikely(vq->log_used)) {
1364 /* Make sure data is seen before log. */
1366 /* Log used ring entry write. */
1367 log_write(vq->log_base,
1369 ((void __user *)used - (void __user *)vq->used),
1370 count * sizeof *used);
1372 old = vq->last_used_idx;
1373 new = (vq->last_used_idx += count);
1374 /* If the driver never bothers to signal in a very long while,
1375 * used index might wrap around. If that happens, invalidate
1376 * signalled_used index we stored. TODO: make sure driver
1377 * signals at least once in 2^16 and remove this. */
1378 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1379 vq->signalled_used_valid = false;
1383 /* After we've used one of their buffers, we tell them about it. We'll then
1384 * want to notify the guest, using eventfd. */
1385 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1390 start = vq->last_used_idx % vq->num;
1391 n = vq->num - start;
1393 r = __vhost_add_used_n(vq, heads, n);
1399 r = __vhost_add_used_n(vq, heads, count);
1401 /* Make sure buffer is written before we update index. */
1403 if (put_user(vq->last_used_idx, &vq->used->idx)) {
1404 vq_err(vq, "Failed to increment used idx");
1407 if (unlikely(vq->log_used)) {
1408 /* Log used index update. */
1409 log_write(vq->log_base,
1410 vq->log_addr + offsetof(struct vring_used, idx),
1411 sizeof vq->used->idx);
1413 eventfd_signal(vq->log_ctx, 1);
1418 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1420 __u16 old, new, event;
1422 /* Flush out used index updates. This is paired
1423 * with the barrier that the Guest executes when enabling
1427 if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1428 unlikely(vq->avail_idx == vq->last_avail_idx))
1431 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1433 if (__get_user(flags, &vq->avail->flags)) {
1434 vq_err(vq, "Failed to get flags");
1437 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1439 old = vq->signalled_used;
1440 v = vq->signalled_used_valid;
1441 new = vq->signalled_used = vq->last_used_idx;
1442 vq->signalled_used_valid = true;
1447 if (get_user(event, vhost_used_event(vq))) {
1448 vq_err(vq, "Failed to get used event idx");
1451 return vring_need_event(event, new, old);
1454 /* This actually signals the guest, using eventfd. */
1455 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1457 /* Signal the Guest tell them we used something up. */
1458 if (vq->call_ctx && vhost_notify(dev, vq))
1459 eventfd_signal(vq->call_ctx, 1);
1462 /* And here's the combo meal deal. Supersize me! */
1463 void vhost_add_used_and_signal(struct vhost_dev *dev,
1464 struct vhost_virtqueue *vq,
1465 unsigned int head, int len)
1467 vhost_add_used(vq, head, len);
1468 vhost_signal(dev, vq);
1471 /* multi-buffer version of vhost_add_used_and_signal */
1472 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1473 struct vhost_virtqueue *vq,
1474 struct vring_used_elem *heads, unsigned count)
1476 vhost_add_used_n(vq, heads, count);
1477 vhost_signal(dev, vq);
1480 /* OK, now we need to know about added descriptors. */
1481 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1486 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1488 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1489 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1490 r = vhost_update_used_flags(vq);
1492 vq_err(vq, "Failed to enable notification at %p: %d\n",
1493 &vq->used->flags, r);
1497 r = vhost_update_avail_event(vq, vq->avail_idx);
1499 vq_err(vq, "Failed to update avail event index at %p: %d\n",
1500 vhost_avail_event(vq), r);
1504 /* They could have slipped one in as we were doing that: make
1505 * sure it's written, then check again. */
1507 r = __get_user(avail_idx, &vq->avail->idx);
1509 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1510 &vq->avail->idx, r);
1514 return avail_idx != vq->avail_idx;
1517 /* We don't need to be notified again. */
1518 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1522 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1524 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1525 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1526 r = vhost_update_used_flags(vq);
1528 vq_err(vq, "Failed to enable notification at %p: %d\n",
1529 &vq->used->flags, r);