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);
81 vhost_work_init(&poll->work, fn);
84 /* Start polling a file. We add ourselves to file's wait queue. The caller must
85 * keep a reference to a file until after vhost_poll_stop is called. */
86 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
90 mask = file->f_op->poll(file, &poll->table);
92 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
95 /* Stop polling a file. After this function returns, it becomes safe to drop the
96 * file reference. You must also flush afterwards. */
97 void vhost_poll_stop(struct vhost_poll *poll)
99 remove_wait_queue(poll->wqh, &poll->wait);
102 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
107 spin_lock_irq(&dev->work_lock);
108 left = seq - work->done_seq;
109 spin_unlock_irq(&dev->work_lock);
113 static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
118 spin_lock_irq(&dev->work_lock);
119 seq = work->queue_seq;
121 spin_unlock_irq(&dev->work_lock);
122 wait_event(work->done, vhost_work_seq_done(dev, work, seq));
123 spin_lock_irq(&dev->work_lock);
124 flushing = --work->flushing;
125 spin_unlock_irq(&dev->work_lock);
126 BUG_ON(flushing < 0);
129 /* Flush any work that has been scheduled. When calling this, don't hold any
130 * locks that are also used by the callback. */
131 void vhost_poll_flush(struct vhost_poll *poll)
133 vhost_work_flush(poll->dev, &poll->work);
136 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
140 spin_lock_irqsave(&dev->work_lock, flags);
141 if (list_empty(&work->node)) {
142 list_add_tail(&work->node, &dev->work_list);
144 wake_up_process(dev->worker);
146 spin_unlock_irqrestore(&dev->work_lock, flags);
149 void vhost_poll_queue(struct vhost_poll *poll)
151 vhost_work_queue(poll->dev, &poll->work);
154 static void vhost_vq_reset(struct vhost_dev *dev,
155 struct vhost_virtqueue *vq)
161 vq->last_avail_idx = 0;
163 vq->last_used_idx = 0;
164 vq->signalled_used = 0;
165 vq->signalled_used_valid = false;
167 vq->log_used = false;
168 vq->log_addr = -1ull;
171 vq->private_data = NULL;
173 vq->error_ctx = NULL;
184 static int vhost_worker(void *data)
186 struct vhost_dev *dev = data;
187 struct vhost_work *work = NULL;
188 unsigned uninitialized_var(seq);
189 mm_segment_t oldfs = get_fs();
195 /* mb paired w/ kthread_stop */
196 set_current_state(TASK_INTERRUPTIBLE);
198 spin_lock_irq(&dev->work_lock);
200 work->done_seq = seq;
202 wake_up_all(&work->done);
205 if (kthread_should_stop()) {
206 spin_unlock_irq(&dev->work_lock);
207 __set_current_state(TASK_RUNNING);
210 if (!list_empty(&dev->work_list)) {
211 work = list_first_entry(&dev->work_list,
212 struct vhost_work, node);
213 list_del_init(&work->node);
214 seq = work->queue_seq;
217 spin_unlock_irq(&dev->work_lock);
220 __set_current_state(TASK_RUNNING);
233 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
241 kfree(vq->ubuf_info);
242 vq->ubuf_info = NULL;
245 void vhost_enable_zcopy(int vq)
247 vhost_zcopy_mask |= 0x1 << vq;
250 /* Helper to allocate iovec buffers for all vqs. */
251 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 zcopy = vhost_zcopy_mask & (0x1 << i);
265 dev->vqs[i].ubuf_info =
266 kmalloc(sizeof *dev->vqs[i].ubuf_info *
267 UIO_MAXIOV, GFP_KERNEL);
268 if (!dev->vqs[i].indirect || !dev->vqs[i].log ||
269 !dev->vqs[i].heads ||
270 (zcopy && !dev->vqs[i].ubuf_info))
277 vhost_vq_free_iovecs(&dev->vqs[i]);
281 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
285 for (i = 0; i < dev->nvqs; ++i)
286 vhost_vq_free_iovecs(&dev->vqs[i]);
289 long vhost_dev_init(struct vhost_dev *dev,
290 struct vhost_virtqueue *vqs, int nvqs)
296 mutex_init(&dev->mutex);
298 dev->log_file = NULL;
301 spin_lock_init(&dev->work_lock);
302 INIT_LIST_HEAD(&dev->work_list);
305 for (i = 0; i < dev->nvqs; ++i) {
306 dev->vqs[i].log = NULL;
307 dev->vqs[i].indirect = NULL;
308 dev->vqs[i].heads = NULL;
309 dev->vqs[i].ubuf_info = NULL;
310 dev->vqs[i].dev = dev;
311 mutex_init(&dev->vqs[i].mutex);
312 vhost_vq_reset(dev, dev->vqs + i);
313 if (dev->vqs[i].handle_kick)
314 vhost_poll_init(&dev->vqs[i].poll,
315 dev->vqs[i].handle_kick, POLLIN, dev);
321 /* Caller should have device mutex */
322 long vhost_dev_check_owner(struct vhost_dev *dev)
324 /* Are you the owner? If not, I don't think you mean to do that */
325 return dev->mm == current->mm ? 0 : -EPERM;
328 struct vhost_attach_cgroups_struct {
329 struct vhost_work work;
330 struct task_struct *owner;
334 static void vhost_attach_cgroups_work(struct vhost_work *work)
336 struct vhost_attach_cgroups_struct *s;
338 s = container_of(work, struct vhost_attach_cgroups_struct, work);
339 s->ret = cgroup_attach_task_all(s->owner, current);
342 static int vhost_attach_cgroups(struct vhost_dev *dev)
344 struct vhost_attach_cgroups_struct attach;
346 attach.owner = current;
347 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
348 vhost_work_queue(dev, &attach.work);
349 vhost_work_flush(dev, &attach.work);
353 /* Caller should have device mutex */
354 static long vhost_dev_set_owner(struct vhost_dev *dev)
356 struct task_struct *worker;
359 /* Is there an owner already? */
365 /* No owner, become one */
366 dev->mm = get_task_mm(current);
367 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
368 if (IS_ERR(worker)) {
369 err = PTR_ERR(worker);
373 dev->worker = worker;
374 wake_up_process(worker); /* avoid contributing to loadavg */
376 err = vhost_attach_cgroups(dev);
380 err = vhost_dev_alloc_iovecs(dev);
386 kthread_stop(worker);
396 /* Caller should have device mutex */
397 long vhost_dev_reset_owner(struct vhost_dev *dev)
399 struct vhost_memory *memory;
401 /* Restore memory to default empty mapping. */
402 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
406 vhost_dev_cleanup(dev, true);
408 memory->nregions = 0;
409 RCU_INIT_POINTER(dev->memory, memory);
413 void vhost_dev_stop(struct vhost_dev *dev)
417 for (i = 0; i < dev->nvqs; ++i) {
418 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
419 vhost_poll_stop(&dev->vqs[i].poll);
420 vhost_poll_flush(&dev->vqs[i].poll);
425 /* Caller should have device mutex if and only if locked is set */
426 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
430 for (i = 0; i < dev->nvqs; ++i) {
431 if (dev->vqs[i].error_ctx)
432 eventfd_ctx_put(dev->vqs[i].error_ctx);
433 if (dev->vqs[i].error)
434 fput(dev->vqs[i].error);
435 if (dev->vqs[i].kick)
436 fput(dev->vqs[i].kick);
437 if (dev->vqs[i].call_ctx)
438 eventfd_ctx_put(dev->vqs[i].call_ctx);
439 if (dev->vqs[i].call)
440 fput(dev->vqs[i].call);
441 vhost_vq_reset(dev, dev->vqs + i);
443 vhost_dev_free_iovecs(dev);
445 eventfd_ctx_put(dev->log_ctx);
449 dev->log_file = NULL;
450 /* No one will access memory at this point */
451 kfree(rcu_dereference_protected(dev->memory,
453 lockdep_is_held(&dev->mutex)));
454 RCU_INIT_POINTER(dev->memory, NULL);
455 WARN_ON(!list_empty(&dev->work_list));
457 kthread_stop(dev->worker);
465 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
467 u64 a = addr / VHOST_PAGE_SIZE / 8;
469 /* Make sure 64 bit math will not overflow. */
470 if (a > ULONG_MAX - (unsigned long)log_base ||
471 a + (unsigned long)log_base > ULONG_MAX)
474 return access_ok(VERIFY_WRITE, log_base + a,
475 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
478 /* Caller should have vq mutex and device mutex. */
479 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
487 for (i = 0; i < mem->nregions; ++i) {
488 struct vhost_memory_region *m = mem->regions + i;
489 unsigned long a = m->userspace_addr;
490 if (m->memory_size > ULONG_MAX)
492 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
495 else if (log_all && !log_access_ok(log_base,
503 /* Can we switch to this memory table? */
504 /* Caller should have device mutex but not vq mutex */
505 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
510 for (i = 0; i < d->nvqs; ++i) {
512 mutex_lock(&d->vqs[i].mutex);
513 /* If ring is inactive, will check when it's enabled. */
514 if (d->vqs[i].private_data)
515 ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
519 mutex_unlock(&d->vqs[i].mutex);
526 static int vq_access_ok(struct vhost_dev *d, unsigned int num,
527 struct vring_desc __user *desc,
528 struct vring_avail __user *avail,
529 struct vring_used __user *used)
531 size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
532 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
533 access_ok(VERIFY_READ, avail,
534 sizeof *avail + num * sizeof *avail->ring + s) &&
535 access_ok(VERIFY_WRITE, used,
536 sizeof *used + num * sizeof *used->ring + s);
539 /* Can we log writes? */
540 /* Caller should have device mutex but not vq mutex */
541 int vhost_log_access_ok(struct vhost_dev *dev)
543 struct vhost_memory *mp;
545 mp = rcu_dereference_protected(dev->memory,
546 lockdep_is_held(&dev->mutex));
547 return memory_access_ok(dev, mp, 1);
550 /* Verify access for write logging. */
551 /* Caller should have vq mutex and device mutex */
552 static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
553 void __user *log_base)
555 struct vhost_memory *mp;
556 size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
558 mp = rcu_dereference_protected(vq->dev->memory,
559 lockdep_is_held(&vq->mutex));
560 return vq_memory_access_ok(log_base, mp,
561 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
562 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
564 vq->num * sizeof *vq->used->ring + s));
567 /* Can we start vq? */
568 /* Caller should have vq mutex and device mutex */
569 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
571 return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
572 vq_log_access_ok(vq->dev, vq, vq->log_base);
575 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
577 struct vhost_memory mem, *newmem, *oldmem;
578 unsigned long size = offsetof(struct vhost_memory, regions);
580 if (copy_from_user(&mem, m, size))
584 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
586 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
590 memcpy(newmem, &mem, size);
591 if (copy_from_user(newmem->regions, m->regions,
592 mem.nregions * sizeof *m->regions)) {
597 if (!memory_access_ok(d, newmem,
598 vhost_has_feature(d, VHOST_F_LOG_ALL))) {
602 oldmem = rcu_dereference_protected(d->memory,
603 lockdep_is_held(&d->mutex));
604 rcu_assign_pointer(d->memory, newmem);
610 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
612 struct file *eventfp, *filep = NULL;
613 bool pollstart = false, pollstop = false;
614 struct eventfd_ctx *ctx = NULL;
615 u32 __user *idxp = argp;
616 struct vhost_virtqueue *vq;
617 struct vhost_vring_state s;
618 struct vhost_vring_file f;
619 struct vhost_vring_addr a;
623 r = get_user(idx, idxp);
631 mutex_lock(&vq->mutex);
634 case VHOST_SET_VRING_NUM:
635 /* Resizing ring with an active backend?
636 * You don't want to do that. */
637 if (vq->private_data) {
641 if (copy_from_user(&s, argp, sizeof s)) {
645 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
651 case VHOST_SET_VRING_BASE:
652 /* Moving base with an active backend?
653 * You don't want to do that. */
654 if (vq->private_data) {
658 if (copy_from_user(&s, argp, sizeof s)) {
662 if (s.num > 0xffff) {
666 vq->last_avail_idx = s.num;
667 /* Forget the cached index value. */
668 vq->avail_idx = vq->last_avail_idx;
670 case VHOST_GET_VRING_BASE:
672 s.num = vq->last_avail_idx;
673 if (copy_to_user(argp, &s, sizeof s))
676 case VHOST_SET_VRING_ADDR:
677 if (copy_from_user(&a, argp, sizeof a)) {
681 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
685 /* For 32bit, verify that the top 32bits of the user
686 data are set to zero. */
687 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
688 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
689 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
693 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
694 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
695 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
700 /* We only verify access here if backend is configured.
701 * If it is not, we don't as size might not have been setup.
702 * We will verify when backend is configured. */
703 if (vq->private_data) {
704 if (!vq_access_ok(d, vq->num,
705 (void __user *)(unsigned long)a.desc_user_addr,
706 (void __user *)(unsigned long)a.avail_user_addr,
707 (void __user *)(unsigned long)a.used_user_addr)) {
712 /* Also validate log access for used ring if enabled. */
713 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
714 !log_access_ok(vq->log_base, a.log_guest_addr,
716 vq->num * sizeof *vq->used->ring)) {
722 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
723 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
724 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
725 vq->log_addr = a.log_guest_addr;
726 vq->used = (void __user *)(unsigned long)a.used_user_addr;
728 case VHOST_SET_VRING_KICK:
729 if (copy_from_user(&f, argp, sizeof f)) {
733 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
734 if (IS_ERR(eventfp)) {
735 r = PTR_ERR(eventfp);
738 if (eventfp != vq->kick) {
739 pollstop = (filep = vq->kick) != NULL;
740 pollstart = (vq->kick = eventfp) != NULL;
744 case VHOST_SET_VRING_CALL:
745 if (copy_from_user(&f, argp, sizeof f)) {
749 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
750 if (IS_ERR(eventfp)) {
751 r = PTR_ERR(eventfp);
754 if (eventfp != vq->call) {
758 vq->call_ctx = eventfp ?
759 eventfd_ctx_fileget(eventfp) : NULL;
763 case VHOST_SET_VRING_ERR:
764 if (copy_from_user(&f, argp, sizeof f)) {
768 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
769 if (IS_ERR(eventfp)) {
770 r = PTR_ERR(eventfp);
773 if (eventfp != vq->error) {
777 vq->error_ctx = eventfp ?
778 eventfd_ctx_fileget(eventfp) : NULL;
786 if (pollstop && vq->handle_kick)
787 vhost_poll_stop(&vq->poll);
790 eventfd_ctx_put(ctx);
794 if (pollstart && vq->handle_kick)
795 vhost_poll_start(&vq->poll, vq->kick);
797 mutex_unlock(&vq->mutex);
799 if (pollstop && vq->handle_kick)
800 vhost_poll_flush(&vq->poll);
804 /* Caller must have device mutex */
805 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
807 struct file *eventfp, *filep = NULL;
808 struct eventfd_ctx *ctx = NULL;
813 /* If you are not the owner, you can become one */
814 if (ioctl == VHOST_SET_OWNER) {
815 r = vhost_dev_set_owner(d);
819 /* You must be the owner to do anything else */
820 r = vhost_dev_check_owner(d);
825 case VHOST_SET_MEM_TABLE:
826 r = vhost_set_memory(d, argp);
828 case VHOST_SET_LOG_BASE:
829 if (copy_from_user(&p, argp, sizeof p)) {
833 if ((u64)(unsigned long)p != p) {
837 for (i = 0; i < d->nvqs; ++i) {
838 struct vhost_virtqueue *vq;
839 void __user *base = (void __user *)(unsigned long)p;
841 mutex_lock(&vq->mutex);
842 /* If ring is inactive, will check when it's enabled. */
843 if (vq->private_data && !vq_log_access_ok(d, vq, base))
847 mutex_unlock(&vq->mutex);
850 case VHOST_SET_LOG_FD:
851 r = get_user(fd, (int __user *)argp);
854 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
855 if (IS_ERR(eventfp)) {
856 r = PTR_ERR(eventfp);
859 if (eventfp != d->log_file) {
862 d->log_ctx = eventfp ?
863 eventfd_ctx_fileget(eventfp) : NULL;
866 for (i = 0; i < d->nvqs; ++i) {
867 mutex_lock(&d->vqs[i].mutex);
868 d->vqs[i].log_ctx = d->log_ctx;
869 mutex_unlock(&d->vqs[i].mutex);
872 eventfd_ctx_put(ctx);
884 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
885 __u64 addr, __u32 len)
887 struct vhost_memory_region *reg;
890 /* linear search is not brilliant, but we really have on the order of 6
891 * regions in practice */
892 for (i = 0; i < mem->nregions; ++i) {
893 reg = mem->regions + i;
894 if (reg->guest_phys_addr <= addr &&
895 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
901 /* TODO: This is really inefficient. We need something like get_user()
902 * (instruction directly accesses the data, with an exception table entry
903 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
905 static int set_bit_to_user(int nr, void __user *addr)
907 unsigned long log = (unsigned long)addr;
910 int bit = nr + (log % PAGE_SIZE) * 8;
913 r = get_user_pages_fast(log, 1, 1, &page);
917 base = kmap_atomic(page);
920 set_page_dirty_lock(page);
925 static int log_write(void __user *log_base,
926 u64 write_address, u64 write_length)
928 u64 write_page = write_address / VHOST_PAGE_SIZE;
933 write_length += write_address % VHOST_PAGE_SIZE;
935 u64 base = (u64)(unsigned long)log_base;
936 u64 log = base + write_page / 8;
937 int bit = write_page % 8;
938 if ((u64)(unsigned long)log != log)
940 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
943 if (write_length <= VHOST_PAGE_SIZE)
945 write_length -= VHOST_PAGE_SIZE;
951 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
952 unsigned int log_num, u64 len)
956 /* Make sure data written is seen before log. */
958 for (i = 0; i < log_num; ++i) {
959 u64 l = min(log[i].len, len);
960 r = log_write(vq->log_base, log[i].addr, l);
966 eventfd_signal(vq->log_ctx, 1);
970 /* Length written exceeds what we have stored. This is a bug. */
975 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
978 if (__put_user(vq->used_flags, &vq->used->flags) < 0)
980 if (unlikely(vq->log_used)) {
981 /* Make sure the flag is seen before log. */
983 /* Log used flag write. */
984 used = &vq->used->flags;
985 log_write(vq->log_base, vq->log_addr +
986 (used - (void __user *)vq->used),
987 sizeof vq->used->flags);
989 eventfd_signal(vq->log_ctx, 1);
994 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
996 if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
998 if (unlikely(vq->log_used)) {
1000 /* Make sure the event is seen before log. */
1002 /* Log avail event write */
1003 used = vhost_avail_event(vq);
1004 log_write(vq->log_base, vq->log_addr +
1005 (used - (void __user *)vq->used),
1006 sizeof *vhost_avail_event(vq));
1008 eventfd_signal(vq->log_ctx, 1);
1013 int vhost_init_used(struct vhost_virtqueue *vq)
1016 if (!vq->private_data)
1019 r = vhost_update_used_flags(vq);
1022 vq->signalled_used_valid = false;
1023 return get_user(vq->last_used_idx, &vq->used->idx);
1026 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1027 struct iovec iov[], int iov_size)
1029 const struct vhost_memory_region *reg;
1030 struct vhost_memory *mem;
1037 mem = rcu_dereference(dev->memory);
1038 while ((u64)len > s) {
1040 if (unlikely(ret >= iov_size)) {
1044 reg = find_region(mem, addr, len);
1045 if (unlikely(!reg)) {
1050 size = reg->memory_size - addr + reg->guest_phys_addr;
1051 _iov->iov_len = min((u64)len - s, size);
1052 _iov->iov_base = (void __user *)(unsigned long)
1053 (reg->userspace_addr + addr - reg->guest_phys_addr);
1063 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1064 * function returns the next descriptor in the chain,
1065 * or -1U if we're at the end. */
1066 static unsigned next_desc(struct vring_desc *desc)
1070 /* If this descriptor says it doesn't chain, we're done. */
1071 if (!(desc->flags & VRING_DESC_F_NEXT))
1074 /* Check they're not leading us off end of descriptors. */
1076 /* Make sure compiler knows to grab that: we don't want it changing! */
1077 /* We will use the result as an index in an array, so most
1078 * architectures only need a compiler barrier here. */
1079 read_barrier_depends();
1084 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1085 struct iovec iov[], unsigned int iov_size,
1086 unsigned int *out_num, unsigned int *in_num,
1087 struct vhost_log *log, unsigned int *log_num,
1088 struct vring_desc *indirect)
1090 struct vring_desc desc;
1091 unsigned int i = 0, count, found = 0;
1095 if (unlikely(indirect->len % sizeof desc)) {
1096 vq_err(vq, "Invalid length in indirect descriptor: "
1097 "len 0x%llx not multiple of 0x%zx\n",
1098 (unsigned long long)indirect->len,
1103 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1105 if (unlikely(ret < 0)) {
1106 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1110 /* We will use the result as an address to read from, so most
1111 * architectures only need a compiler barrier here. */
1112 read_barrier_depends();
1114 count = indirect->len / sizeof desc;
1115 /* Buffers are chained via a 16 bit next field, so
1116 * we can have at most 2^16 of these. */
1117 if (unlikely(count > USHRT_MAX + 1)) {
1118 vq_err(vq, "Indirect buffer length too big: %d\n",
1124 unsigned iov_count = *in_num + *out_num;
1125 if (unlikely(++found > count)) {
1126 vq_err(vq, "Loop detected: last one at %u "
1127 "indirect size %u\n",
1131 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1132 vq->indirect, sizeof desc))) {
1133 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1134 i, (size_t)indirect->addr + i * sizeof desc);
1137 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1138 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1139 i, (size_t)indirect->addr + i * sizeof desc);
1143 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1144 iov_size - iov_count);
1145 if (unlikely(ret < 0)) {
1146 vq_err(vq, "Translation failure %d indirect idx %d\n",
1150 /* If this is an input descriptor, increment that count. */
1151 if (desc.flags & VRING_DESC_F_WRITE) {
1153 if (unlikely(log)) {
1154 log[*log_num].addr = desc.addr;
1155 log[*log_num].len = desc.len;
1159 /* If it's an output descriptor, they're all supposed
1160 * to come before any input descriptors. */
1161 if (unlikely(*in_num)) {
1162 vq_err(vq, "Indirect descriptor "
1163 "has out after in: idx %d\n", i);
1168 } while ((i = next_desc(&desc)) != -1);
1172 /* This looks in the virtqueue and for the first available buffer, and converts
1173 * it to an iovec for convenient access. Since descriptors consist of some
1174 * number of output then some number of input descriptors, it's actually two
1175 * iovecs, but we pack them into one and note how many of each there were.
1177 * This function returns the descriptor number found, or vq->num (which is
1178 * never a valid descriptor number) if none was found. A negative code is
1179 * returned on error. */
1180 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1181 struct iovec iov[], unsigned int iov_size,
1182 unsigned int *out_num, unsigned int *in_num,
1183 struct vhost_log *log, unsigned int *log_num)
1185 struct vring_desc desc;
1186 unsigned int i, head, found = 0;
1190 /* Check it isn't doing very strange things with descriptor numbers. */
1191 last_avail_idx = vq->last_avail_idx;
1192 if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1193 vq_err(vq, "Failed to access avail idx at %p\n",
1198 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1199 vq_err(vq, "Guest moved used index from %u to %u",
1200 last_avail_idx, vq->avail_idx);
1204 /* If there's nothing new since last we looked, return invalid. */
1205 if (vq->avail_idx == last_avail_idx)
1208 /* Only get avail ring entries after they have been exposed by guest. */
1211 /* Grab the next descriptor number they're advertising, and increment
1212 * the index we've seen. */
1213 if (unlikely(__get_user(head,
1214 &vq->avail->ring[last_avail_idx % vq->num]))) {
1215 vq_err(vq, "Failed to read head: idx %d address %p\n",
1217 &vq->avail->ring[last_avail_idx % vq->num]);
1221 /* If their number is silly, that's an error. */
1222 if (unlikely(head >= vq->num)) {
1223 vq_err(vq, "Guest says index %u > %u is available",
1228 /* When we start there are none of either input nor output. */
1229 *out_num = *in_num = 0;
1235 unsigned iov_count = *in_num + *out_num;
1236 if (unlikely(i >= vq->num)) {
1237 vq_err(vq, "Desc index is %u > %u, head = %u",
1241 if (unlikely(++found > vq->num)) {
1242 vq_err(vq, "Loop detected: last one at %u "
1243 "vq size %u head %u\n",
1247 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1248 if (unlikely(ret)) {
1249 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1253 if (desc.flags & VRING_DESC_F_INDIRECT) {
1254 ret = get_indirect(dev, vq, iov, iov_size,
1256 log, log_num, &desc);
1257 if (unlikely(ret < 0)) {
1258 vq_err(vq, "Failure detected "
1259 "in indirect descriptor at idx %d\n", i);
1265 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1266 iov_size - iov_count);
1267 if (unlikely(ret < 0)) {
1268 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1272 if (desc.flags & VRING_DESC_F_WRITE) {
1273 /* If this is an input descriptor,
1274 * increment that count. */
1276 if (unlikely(log)) {
1277 log[*log_num].addr = desc.addr;
1278 log[*log_num].len = desc.len;
1282 /* If it's an output descriptor, they're all supposed
1283 * to come before any input descriptors. */
1284 if (unlikely(*in_num)) {
1285 vq_err(vq, "Descriptor has out after in: "
1291 } while ((i = next_desc(&desc)) != -1);
1293 /* On success, increment avail index. */
1294 vq->last_avail_idx++;
1296 /* Assume notifications from guest are disabled at this point,
1297 * if they aren't we would need to update avail_event index. */
1298 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1302 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1303 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1305 vq->last_avail_idx -= n;
1308 /* After we've used one of their buffers, we tell them about it. We'll then
1309 * want to notify the guest, using eventfd. */
1310 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1312 struct vring_used_elem __user *used;
1314 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1315 * next entry in that used ring. */
1316 used = &vq->used->ring[vq->last_used_idx % vq->num];
1317 if (__put_user(head, &used->id)) {
1318 vq_err(vq, "Failed to write used id");
1321 if (__put_user(len, &used->len)) {
1322 vq_err(vq, "Failed to write used len");
1325 /* Make sure buffer is written before we update index. */
1327 if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1328 vq_err(vq, "Failed to increment used idx");
1331 if (unlikely(vq->log_used)) {
1332 /* Make sure data is seen before log. */
1334 /* Log used ring entry write. */
1335 log_write(vq->log_base,
1337 ((void __user *)used - (void __user *)vq->used),
1339 /* Log used index update. */
1340 log_write(vq->log_base,
1341 vq->log_addr + offsetof(struct vring_used, idx),
1342 sizeof vq->used->idx);
1344 eventfd_signal(vq->log_ctx, 1);
1346 vq->last_used_idx++;
1347 /* If the driver never bothers to signal in a very long while,
1348 * used index might wrap around. If that happens, invalidate
1349 * signalled_used index we stored. TODO: make sure driver
1350 * signals at least once in 2^16 and remove this. */
1351 if (unlikely(vq->last_used_idx == vq->signalled_used))
1352 vq->signalled_used_valid = false;
1356 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1357 struct vring_used_elem *heads,
1360 struct vring_used_elem __user *used;
1364 start = vq->last_used_idx % vq->num;
1365 used = vq->used->ring + start;
1366 if (__copy_to_user(used, heads, count * sizeof *used)) {
1367 vq_err(vq, "Failed to write used");
1370 if (unlikely(vq->log_used)) {
1371 /* Make sure data is seen before log. */
1373 /* Log used ring entry write. */
1374 log_write(vq->log_base,
1376 ((void __user *)used - (void __user *)vq->used),
1377 count * sizeof *used);
1379 old = vq->last_used_idx;
1380 new = (vq->last_used_idx += count);
1381 /* If the driver never bothers to signal in a very long while,
1382 * used index might wrap around. If that happens, invalidate
1383 * signalled_used index we stored. TODO: make sure driver
1384 * signals at least once in 2^16 and remove this. */
1385 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1386 vq->signalled_used_valid = false;
1390 /* After we've used one of their buffers, we tell them about it. We'll then
1391 * want to notify the guest, using eventfd. */
1392 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1397 start = vq->last_used_idx % vq->num;
1398 n = vq->num - start;
1400 r = __vhost_add_used_n(vq, heads, n);
1406 r = __vhost_add_used_n(vq, heads, count);
1408 /* Make sure buffer is written before we update index. */
1410 if (put_user(vq->last_used_idx, &vq->used->idx)) {
1411 vq_err(vq, "Failed to increment used idx");
1414 if (unlikely(vq->log_used)) {
1415 /* Log used index update. */
1416 log_write(vq->log_base,
1417 vq->log_addr + offsetof(struct vring_used, idx),
1418 sizeof vq->used->idx);
1420 eventfd_signal(vq->log_ctx, 1);
1425 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1427 __u16 old, new, event;
1429 /* Flush out used index updates. This is paired
1430 * with the barrier that the Guest executes when enabling
1434 if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1435 unlikely(vq->avail_idx == vq->last_avail_idx))
1438 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1440 if (__get_user(flags, &vq->avail->flags)) {
1441 vq_err(vq, "Failed to get flags");
1444 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1446 old = vq->signalled_used;
1447 v = vq->signalled_used_valid;
1448 new = vq->signalled_used = vq->last_used_idx;
1449 vq->signalled_used_valid = true;
1454 if (get_user(event, vhost_used_event(vq))) {
1455 vq_err(vq, "Failed to get used event idx");
1458 return vring_need_event(event, new, old);
1461 /* This actually signals the guest, using eventfd. */
1462 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1464 /* Signal the Guest tell them we used something up. */
1465 if (vq->call_ctx && vhost_notify(dev, vq))
1466 eventfd_signal(vq->call_ctx, 1);
1469 /* And here's the combo meal deal. Supersize me! */
1470 void vhost_add_used_and_signal(struct vhost_dev *dev,
1471 struct vhost_virtqueue *vq,
1472 unsigned int head, int len)
1474 vhost_add_used(vq, head, len);
1475 vhost_signal(dev, vq);
1478 /* multi-buffer version of vhost_add_used_and_signal */
1479 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1480 struct vhost_virtqueue *vq,
1481 struct vring_used_elem *heads, unsigned count)
1483 vhost_add_used_n(vq, heads, count);
1484 vhost_signal(dev, vq);
1487 /* OK, now we need to know about added descriptors. */
1488 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1493 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1495 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1496 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1497 r = vhost_update_used_flags(vq);
1499 vq_err(vq, "Failed to enable notification at %p: %d\n",
1500 &vq->used->flags, r);
1504 r = vhost_update_avail_event(vq, vq->avail_idx);
1506 vq_err(vq, "Failed to update avail event index at %p: %d\n",
1507 vhost_avail_event(vq), r);
1511 /* They could have slipped one in as we were doing that: make
1512 * sure it's written, then check again. */
1514 r = __get_user(avail_idx, &vq->avail->idx);
1516 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1517 &vq->avail->idx, r);
1521 return avail_idx != vq->avail_idx;
1524 /* We don't need to be notified again. */
1525 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1529 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1531 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1532 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1533 r = vhost_update_used_flags(vq);
1535 vq_err(vq, "Failed to enable notification at %p: %d\n",
1536 &vq->used->flags, r);
1540 static void vhost_zerocopy_done_signal(struct kref *kref)
1542 struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
1544 wake_up(&ubufs->wait);
1547 struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
1550 struct vhost_ubuf_ref *ubufs;
1551 /* No zero copy backend? Nothing to count. */
1554 ubufs = kmalloc(sizeof *ubufs, GFP_KERNEL);
1556 return ERR_PTR(-ENOMEM);
1557 kref_init(&ubufs->kref);
1558 init_waitqueue_head(&ubufs->wait);
1563 void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
1565 kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1568 void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
1570 kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1571 wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));