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/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/miscdevice.h>
19 #include <linux/mutex.h>
20 #include <linux/rcupdate.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/kthread.h>
26 #include <linux/cgroup.h>
28 #include <linux/net.h>
29 #include <linux/if_packet.h>
30 #include <linux/if_arp.h>
37 VHOST_MEMORY_MAX_NREGIONS = 64,
38 VHOST_MEMORY_F_LOG = 0x1,
41 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
44 struct vhost_poll *poll;
45 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 /* Init poll structure */
64 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
65 unsigned long mask, struct vhost_dev *dev)
67 struct vhost_work *work = &poll->work;
69 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
70 init_poll_funcptr(&poll->table, vhost_poll_func);
74 INIT_LIST_HEAD(&work->node);
76 init_waitqueue_head(&work->done);
78 work->queue_seq = work->done_seq = 0;
81 /* Start polling a file. We add ourselves to file's wait queue. The caller must
82 * keep a reference to a file until after vhost_poll_stop is called. */
83 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
86 mask = file->f_op->poll(file, &poll->table);
88 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
91 /* Stop polling a file. After this function returns, it becomes safe to drop the
92 * file reference. You must also flush afterwards. */
93 void vhost_poll_stop(struct vhost_poll *poll)
95 remove_wait_queue(poll->wqh, &poll->wait);
98 /* Flush any work that has been scheduled. When calling this, don't hold any
99 * locks that are also used by the callback. */
100 void vhost_poll_flush(struct vhost_poll *poll)
102 struct vhost_work *work = &poll->work;
107 spin_lock_irq(&poll->dev->work_lock);
108 seq = work->queue_seq;
110 spin_unlock_irq(&poll->dev->work_lock);
111 wait_event(work->done, ({
112 spin_lock_irq(&poll->dev->work_lock);
113 left = seq - work->done_seq <= 0;
114 spin_unlock_irq(&poll->dev->work_lock);
117 spin_lock_irq(&poll->dev->work_lock);
118 flushing = --work->flushing;
119 spin_unlock_irq(&poll->dev->work_lock);
120 BUG_ON(flushing < 0);
123 void vhost_poll_queue(struct vhost_poll *poll)
125 struct vhost_dev *dev = poll->dev;
126 struct vhost_work *work = &poll->work;
129 spin_lock_irqsave(&dev->work_lock, flags);
130 if (list_empty(&work->node)) {
131 list_add_tail(&work->node, &dev->work_list);
133 wake_up_process(dev->worker);
135 spin_unlock_irqrestore(&dev->work_lock, flags);
138 static void vhost_vq_reset(struct vhost_dev *dev,
139 struct vhost_virtqueue *vq)
145 vq->last_avail_idx = 0;
147 vq->last_used_idx = 0;
150 vq->log_used = false;
151 vq->log_addr = -1ull;
154 vq->private_data = NULL;
156 vq->error_ctx = NULL;
164 static int vhost_worker(void *data)
166 struct vhost_dev *dev = data;
167 struct vhost_work *work = NULL;
168 unsigned uninitialized_var(seq);
171 /* mb paired w/ kthread_stop */
172 set_current_state(TASK_INTERRUPTIBLE);
174 spin_lock_irq(&dev->work_lock);
176 work->done_seq = seq;
178 wake_up_all(&work->done);
181 if (kthread_should_stop()) {
182 spin_unlock_irq(&dev->work_lock);
183 __set_current_state(TASK_RUNNING);
186 if (!list_empty(&dev->work_list)) {
187 work = list_first_entry(&dev->work_list,
188 struct vhost_work, node);
189 list_del_init(&work->node);
190 seq = work->queue_seq;
193 spin_unlock_irq(&dev->work_lock);
196 __set_current_state(TASK_RUNNING);
204 long vhost_dev_init(struct vhost_dev *dev,
205 struct vhost_virtqueue *vqs, int nvqs)
211 mutex_init(&dev->mutex);
213 dev->log_file = NULL;
216 spin_lock_init(&dev->work_lock);
217 INIT_LIST_HEAD(&dev->work_list);
220 for (i = 0; i < dev->nvqs; ++i) {
221 dev->vqs[i].dev = dev;
222 mutex_init(&dev->vqs[i].mutex);
223 vhost_vq_reset(dev, dev->vqs + i);
224 if (dev->vqs[i].handle_kick)
225 vhost_poll_init(&dev->vqs[i].poll,
226 dev->vqs[i].handle_kick, POLLIN, dev);
232 /* Caller should have device mutex */
233 long vhost_dev_check_owner(struct vhost_dev *dev)
235 /* Are you the owner? If not, I don't think you mean to do that */
236 return dev->mm == current->mm ? 0 : -EPERM;
239 /* Caller should have device mutex */
240 static long vhost_dev_set_owner(struct vhost_dev *dev)
242 struct task_struct *worker;
244 /* Is there an owner already? */
249 /* No owner, become one */
250 dev->mm = get_task_mm(current);
251 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
252 if (IS_ERR(worker)) {
253 err = PTR_ERR(worker);
257 dev->worker = worker;
258 err = cgroup_attach_task_current_cg(worker);
261 wake_up_process(worker); /* avoid contributing to loadavg */
265 kthread_stop(worker);
274 /* Caller should have device mutex */
275 long vhost_dev_reset_owner(struct vhost_dev *dev)
277 struct vhost_memory *memory;
279 /* Restore memory to default empty mapping. */
280 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
284 vhost_dev_cleanup(dev);
286 memory->nregions = 0;
287 RCU_INIT_POINTER(dev->memory, memory);
291 /* Caller should have device mutex */
292 void vhost_dev_cleanup(struct vhost_dev *dev)
295 for (i = 0; i < dev->nvqs; ++i) {
296 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
297 vhost_poll_stop(&dev->vqs[i].poll);
298 vhost_poll_flush(&dev->vqs[i].poll);
300 if (dev->vqs[i].error_ctx)
301 eventfd_ctx_put(dev->vqs[i].error_ctx);
302 if (dev->vqs[i].error)
303 fput(dev->vqs[i].error);
304 if (dev->vqs[i].kick)
305 fput(dev->vqs[i].kick);
306 if (dev->vqs[i].call_ctx)
307 eventfd_ctx_put(dev->vqs[i].call_ctx);
308 if (dev->vqs[i].call)
309 fput(dev->vqs[i].call);
310 vhost_vq_reset(dev, dev->vqs + i);
313 eventfd_ctx_put(dev->log_ctx);
317 dev->log_file = NULL;
318 /* No one will access memory at this point */
319 kfree(rcu_dereference_protected(dev->memory,
320 lockdep_is_held(&dev->mutex)));
321 RCU_INIT_POINTER(dev->memory, NULL);
326 WARN_ON(!list_empty(&dev->work_list));
328 kthread_stop(dev->worker);
333 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
335 u64 a = addr / VHOST_PAGE_SIZE / 8;
336 /* Make sure 64 bit math will not overflow. */
337 if (a > ULONG_MAX - (unsigned long)log_base ||
338 a + (unsigned long)log_base > ULONG_MAX)
341 return access_ok(VERIFY_WRITE, log_base + a,
342 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
345 /* Caller should have vq mutex and device mutex. */
346 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
354 for (i = 0; i < mem->nregions; ++i) {
355 struct vhost_memory_region *m = mem->regions + i;
356 unsigned long a = m->userspace_addr;
357 if (m->memory_size > ULONG_MAX)
359 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
362 else if (log_all && !log_access_ok(log_base,
370 /* Can we switch to this memory table? */
371 /* Caller should have device mutex but not vq mutex */
372 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
376 for (i = 0; i < d->nvqs; ++i) {
378 mutex_lock(&d->vqs[i].mutex);
379 /* If ring is inactive, will check when it's enabled. */
380 if (d->vqs[i].private_data)
381 ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
385 mutex_unlock(&d->vqs[i].mutex);
392 static int vq_access_ok(unsigned int num,
393 struct vring_desc __user *desc,
394 struct vring_avail __user *avail,
395 struct vring_used __user *used)
397 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
398 access_ok(VERIFY_READ, avail,
399 sizeof *avail + num * sizeof *avail->ring) &&
400 access_ok(VERIFY_WRITE, used,
401 sizeof *used + num * sizeof *used->ring);
404 /* Can we log writes? */
405 /* Caller should have device mutex but not vq mutex */
406 int vhost_log_access_ok(struct vhost_dev *dev)
408 struct vhost_memory *mp;
410 mp = rcu_dereference_protected(dev->memory,
411 lockdep_is_held(&dev->mutex));
412 return memory_access_ok(dev, mp, 1);
415 /* Verify access for write logging. */
416 /* Caller should have vq mutex and device mutex */
417 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base)
419 struct vhost_memory *mp;
421 mp = rcu_dereference_protected(vq->dev->memory,
422 lockdep_is_held(&vq->mutex));
423 return vq_memory_access_ok(log_base, mp,
424 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
425 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
427 vq->num * sizeof *vq->used->ring));
430 /* Can we start vq? */
431 /* Caller should have vq mutex and device mutex */
432 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
434 return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) &&
435 vq_log_access_ok(vq, vq->log_base);
438 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
440 struct vhost_memory mem, *newmem, *oldmem;
441 unsigned long size = offsetof(struct vhost_memory, regions);
442 if (copy_from_user(&mem, m, size))
446 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
448 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
452 memcpy(newmem, &mem, size);
453 if (copy_from_user(newmem->regions, m->regions,
454 mem.nregions * sizeof *m->regions)) {
459 if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL))) {
463 oldmem = rcu_dereference_protected(d->memory,
464 lockdep_is_held(&d->mutex));
465 rcu_assign_pointer(d->memory, newmem);
471 static int init_used(struct vhost_virtqueue *vq,
472 struct vring_used __user *used)
474 int r = put_user(vq->used_flags, &used->flags);
477 return get_user(vq->last_used_idx, &used->idx);
480 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
482 struct file *eventfp, *filep = NULL,
483 *pollstart = NULL, *pollstop = NULL;
484 struct eventfd_ctx *ctx = NULL;
485 u32 __user *idxp = argp;
486 struct vhost_virtqueue *vq;
487 struct vhost_vring_state s;
488 struct vhost_vring_file f;
489 struct vhost_vring_addr a;
493 r = get_user(idx, idxp);
501 mutex_lock(&vq->mutex);
504 case VHOST_SET_VRING_NUM:
505 /* Resizing ring with an active backend?
506 * You don't want to do that. */
507 if (vq->private_data) {
511 if (copy_from_user(&s, argp, sizeof s)) {
515 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
521 case VHOST_SET_VRING_BASE:
522 /* Moving base with an active backend?
523 * You don't want to do that. */
524 if (vq->private_data) {
528 if (copy_from_user(&s, argp, sizeof s)) {
532 if (s.num > 0xffff) {
536 vq->last_avail_idx = s.num;
537 /* Forget the cached index value. */
538 vq->avail_idx = vq->last_avail_idx;
540 case VHOST_GET_VRING_BASE:
542 s.num = vq->last_avail_idx;
543 if (copy_to_user(argp, &s, sizeof s))
546 case VHOST_SET_VRING_ADDR:
547 if (copy_from_user(&a, argp, sizeof a)) {
551 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
555 /* For 32bit, verify that the top 32bits of the user
556 data are set to zero. */
557 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
558 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
559 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
563 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
564 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
565 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
570 /* We only verify access here if backend is configured.
571 * If it is not, we don't as size might not have been setup.
572 * We will verify when backend is configured. */
573 if (vq->private_data) {
574 if (!vq_access_ok(vq->num,
575 (void __user *)(unsigned long)a.desc_user_addr,
576 (void __user *)(unsigned long)a.avail_user_addr,
577 (void __user *)(unsigned long)a.used_user_addr)) {
582 /* Also validate log access for used ring if enabled. */
583 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
584 !log_access_ok(vq->log_base, a.log_guest_addr,
586 vq->num * sizeof *vq->used->ring)) {
592 r = init_used(vq, (struct vring_used __user *)(unsigned long)
596 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
597 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
598 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
599 vq->log_addr = a.log_guest_addr;
600 vq->used = (void __user *)(unsigned long)a.used_user_addr;
602 case VHOST_SET_VRING_KICK:
603 if (copy_from_user(&f, argp, sizeof f)) {
607 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
608 if (IS_ERR(eventfp)) {
609 r = PTR_ERR(eventfp);
612 if (eventfp != vq->kick) {
613 pollstop = filep = vq->kick;
614 pollstart = vq->kick = eventfp;
618 case VHOST_SET_VRING_CALL:
619 if (copy_from_user(&f, argp, sizeof f)) {
623 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
624 if (IS_ERR(eventfp)) {
625 r = PTR_ERR(eventfp);
628 if (eventfp != vq->call) {
632 vq->call_ctx = eventfp ?
633 eventfd_ctx_fileget(eventfp) : NULL;
637 case VHOST_SET_VRING_ERR:
638 if (copy_from_user(&f, argp, sizeof f)) {
642 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
643 if (IS_ERR(eventfp)) {
644 r = PTR_ERR(eventfp);
647 if (eventfp != vq->error) {
651 vq->error_ctx = eventfp ?
652 eventfd_ctx_fileget(eventfp) : NULL;
660 if (pollstop && vq->handle_kick)
661 vhost_poll_stop(&vq->poll);
664 eventfd_ctx_put(ctx);
668 if (pollstart && vq->handle_kick)
669 vhost_poll_start(&vq->poll, vq->kick);
671 mutex_unlock(&vq->mutex);
673 if (pollstop && vq->handle_kick)
674 vhost_poll_flush(&vq->poll);
678 /* Caller must have device mutex */
679 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
681 void __user *argp = (void __user *)arg;
682 struct file *eventfp, *filep = NULL;
683 struct eventfd_ctx *ctx = NULL;
688 /* If you are not the owner, you can become one */
689 if (ioctl == VHOST_SET_OWNER) {
690 r = vhost_dev_set_owner(d);
694 /* You must be the owner to do anything else */
695 r = vhost_dev_check_owner(d);
700 case VHOST_SET_MEM_TABLE:
701 r = vhost_set_memory(d, argp);
703 case VHOST_SET_LOG_BASE:
704 if (copy_from_user(&p, argp, sizeof p)) {
708 if ((u64)(unsigned long)p != p) {
712 for (i = 0; i < d->nvqs; ++i) {
713 struct vhost_virtqueue *vq;
714 void __user *base = (void __user *)(unsigned long)p;
716 mutex_lock(&vq->mutex);
717 /* If ring is inactive, will check when it's enabled. */
718 if (vq->private_data && !vq_log_access_ok(vq, base))
722 mutex_unlock(&vq->mutex);
725 case VHOST_SET_LOG_FD:
726 r = get_user(fd, (int __user *)argp);
729 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
730 if (IS_ERR(eventfp)) {
731 r = PTR_ERR(eventfp);
734 if (eventfp != d->log_file) {
737 d->log_ctx = eventfp ?
738 eventfd_ctx_fileget(eventfp) : NULL;
741 for (i = 0; i < d->nvqs; ++i) {
742 mutex_lock(&d->vqs[i].mutex);
743 d->vqs[i].log_ctx = d->log_ctx;
744 mutex_unlock(&d->vqs[i].mutex);
747 eventfd_ctx_put(ctx);
752 r = vhost_set_vring(d, ioctl, argp);
759 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
760 __u64 addr, __u32 len)
762 struct vhost_memory_region *reg;
764 /* linear search is not brilliant, but we really have on the order of 6
765 * regions in practice */
766 for (i = 0; i < mem->nregions; ++i) {
767 reg = mem->regions + i;
768 if (reg->guest_phys_addr <= addr &&
769 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
775 /* TODO: This is really inefficient. We need something like get_user()
776 * (instruction directly accesses the data, with an exception table entry
777 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
779 static int set_bit_to_user(int nr, void __user *addr)
781 unsigned long log = (unsigned long)addr;
784 int bit = nr + (log % PAGE_SIZE) * 8;
786 r = get_user_pages_fast(log, 1, 1, &page);
790 base = kmap_atomic(page, KM_USER0);
792 kunmap_atomic(base, KM_USER0);
793 set_page_dirty_lock(page);
798 static int log_write(void __user *log_base,
799 u64 write_address, u64 write_length)
804 write_address /= VHOST_PAGE_SIZE;
806 u64 base = (u64)(unsigned long)log_base;
807 u64 log = base + write_address / 8;
808 int bit = write_address % 8;
809 if ((u64)(unsigned long)log != log)
811 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
814 if (write_length <= VHOST_PAGE_SIZE)
816 write_length -= VHOST_PAGE_SIZE;
817 write_address += VHOST_PAGE_SIZE;
822 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
823 unsigned int log_num, u64 len)
827 /* Make sure data written is seen before log. */
829 for (i = 0; i < log_num; ++i) {
830 u64 l = min(log[i].len, len);
831 r = log_write(vq->log_base, log[i].addr, l);
839 eventfd_signal(vq->log_ctx, 1);
840 /* Length written exceeds what we have stored. This is a bug. */
845 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
846 struct iovec iov[], int iov_size)
848 const struct vhost_memory_region *reg;
849 struct vhost_memory *mem;
856 mem = rcu_dereference(dev->memory);
857 while ((u64)len > s) {
859 if (unlikely(ret >= iov_size)) {
863 reg = find_region(mem, addr, len);
864 if (unlikely(!reg)) {
869 size = reg->memory_size - addr + reg->guest_phys_addr;
870 _iov->iov_len = min((u64)len, size);
871 _iov->iov_base = (void __user *)(unsigned long)
872 (reg->userspace_addr + addr - reg->guest_phys_addr);
882 /* Each buffer in the virtqueues is actually a chain of descriptors. This
883 * function returns the next descriptor in the chain,
884 * or -1U if we're at the end. */
885 static unsigned next_desc(struct vring_desc *desc)
889 /* If this descriptor says it doesn't chain, we're done. */
890 if (!(desc->flags & VRING_DESC_F_NEXT))
893 /* Check they're not leading us off end of descriptors. */
895 /* Make sure compiler knows to grab that: we don't want it changing! */
896 /* We will use the result as an index in an array, so most
897 * architectures only need a compiler barrier here. */
898 read_barrier_depends();
903 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
904 struct iovec iov[], unsigned int iov_size,
905 unsigned int *out_num, unsigned int *in_num,
906 struct vhost_log *log, unsigned int *log_num,
907 struct vring_desc *indirect)
909 struct vring_desc desc;
910 unsigned int i = 0, count, found = 0;
914 if (unlikely(indirect->len % sizeof desc)) {
915 vq_err(vq, "Invalid length in indirect descriptor: "
916 "len 0x%llx not multiple of 0x%zx\n",
917 (unsigned long long)indirect->len,
922 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
923 ARRAY_SIZE(vq->indirect));
924 if (unlikely(ret < 0)) {
925 vq_err(vq, "Translation failure %d in indirect.\n", ret);
929 /* We will use the result as an address to read from, so most
930 * architectures only need a compiler barrier here. */
931 read_barrier_depends();
933 count = indirect->len / sizeof desc;
934 /* Buffers are chained via a 16 bit next field, so
935 * we can have at most 2^16 of these. */
936 if (unlikely(count > USHRT_MAX + 1)) {
937 vq_err(vq, "Indirect buffer length too big: %d\n",
943 unsigned iov_count = *in_num + *out_num;
944 if (unlikely(++found > count)) {
945 vq_err(vq, "Loop detected: last one at %u "
946 "indirect size %u\n",
950 if (unlikely(memcpy_fromiovec((unsigned char *)&desc, vq->indirect,
952 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
953 i, (size_t)indirect->addr + i * sizeof desc);
956 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
957 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
958 i, (size_t)indirect->addr + i * sizeof desc);
962 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
963 iov_size - iov_count);
964 if (unlikely(ret < 0)) {
965 vq_err(vq, "Translation failure %d indirect idx %d\n",
969 /* If this is an input descriptor, increment that count. */
970 if (desc.flags & VRING_DESC_F_WRITE) {
973 log[*log_num].addr = desc.addr;
974 log[*log_num].len = desc.len;
978 /* If it's an output descriptor, they're all supposed
979 * to come before any input descriptors. */
980 if (unlikely(*in_num)) {
981 vq_err(vq, "Indirect descriptor "
982 "has out after in: idx %d\n", i);
987 } while ((i = next_desc(&desc)) != -1);
991 /* This looks in the virtqueue and for the first available buffer, and converts
992 * it to an iovec for convenient access. Since descriptors consist of some
993 * number of output then some number of input descriptors, it's actually two
994 * iovecs, but we pack them into one and note how many of each there were.
996 * This function returns the descriptor number found, or vq->num (which is
997 * never a valid descriptor number) if none was found. A negative code is
998 * returned on error. */
999 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1000 struct iovec iov[], unsigned int iov_size,
1001 unsigned int *out_num, unsigned int *in_num,
1002 struct vhost_log *log, unsigned int *log_num)
1004 struct vring_desc desc;
1005 unsigned int i, head, found = 0;
1009 /* Check it isn't doing very strange things with descriptor numbers. */
1010 last_avail_idx = vq->last_avail_idx;
1011 if (unlikely(get_user(vq->avail_idx, &vq->avail->idx))) {
1012 vq_err(vq, "Failed to access avail idx at %p\n",
1017 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1018 vq_err(vq, "Guest moved used index from %u to %u",
1019 last_avail_idx, vq->avail_idx);
1023 /* If there's nothing new since last we looked, return invalid. */
1024 if (vq->avail_idx == last_avail_idx)
1027 /* Only get avail ring entries after they have been exposed by guest. */
1030 /* Grab the next descriptor number they're advertising, and increment
1031 * the index we've seen. */
1032 if (unlikely(get_user(head,
1033 &vq->avail->ring[last_avail_idx % vq->num]))) {
1034 vq_err(vq, "Failed to read head: idx %d address %p\n",
1036 &vq->avail->ring[last_avail_idx % vq->num]);
1040 /* If their number is silly, that's an error. */
1041 if (unlikely(head >= vq->num)) {
1042 vq_err(vq, "Guest says index %u > %u is available",
1047 /* When we start there are none of either input nor output. */
1048 *out_num = *in_num = 0;
1054 unsigned iov_count = *in_num + *out_num;
1055 if (unlikely(i >= vq->num)) {
1056 vq_err(vq, "Desc index is %u > %u, head = %u",
1060 if (unlikely(++found > vq->num)) {
1061 vq_err(vq, "Loop detected: last one at %u "
1062 "vq size %u head %u\n",
1066 ret = copy_from_user(&desc, vq->desc + i, sizeof desc);
1067 if (unlikely(ret)) {
1068 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1072 if (desc.flags & VRING_DESC_F_INDIRECT) {
1073 ret = get_indirect(dev, vq, iov, iov_size,
1075 log, log_num, &desc);
1076 if (unlikely(ret < 0)) {
1077 vq_err(vq, "Failure detected "
1078 "in indirect descriptor at idx %d\n", i);
1084 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1085 iov_size - iov_count);
1086 if (unlikely(ret < 0)) {
1087 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1091 if (desc.flags & VRING_DESC_F_WRITE) {
1092 /* If this is an input descriptor,
1093 * increment that count. */
1095 if (unlikely(log)) {
1096 log[*log_num].addr = desc.addr;
1097 log[*log_num].len = desc.len;
1101 /* If it's an output descriptor, they're all supposed
1102 * to come before any input descriptors. */
1103 if (unlikely(*in_num)) {
1104 vq_err(vq, "Descriptor has out after in: "
1110 } while ((i = next_desc(&desc)) != -1);
1112 /* On success, increment avail index. */
1113 vq->last_avail_idx++;
1117 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1118 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1120 vq->last_avail_idx -= n;
1123 /* After we've used one of their buffers, we tell them about it. We'll then
1124 * want to notify the guest, using eventfd. */
1125 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1127 struct vring_used_elem __user *used;
1129 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1130 * next entry in that used ring. */
1131 used = &vq->used->ring[vq->last_used_idx % vq->num];
1132 if (put_user(head, &used->id)) {
1133 vq_err(vq, "Failed to write used id");
1136 if (put_user(len, &used->len)) {
1137 vq_err(vq, "Failed to write used len");
1140 /* Make sure buffer is written before we update index. */
1142 if (put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1143 vq_err(vq, "Failed to increment used idx");
1146 if (unlikely(vq->log_used)) {
1147 /* Make sure data is seen before log. */
1149 /* Log used ring entry write. */
1150 log_write(vq->log_base,
1152 ((void __user *)used - (void __user *)vq->used),
1154 /* Log used index update. */
1155 log_write(vq->log_base,
1156 vq->log_addr + offsetof(struct vring_used, idx),
1157 sizeof vq->used->idx);
1159 eventfd_signal(vq->log_ctx, 1);
1161 vq->last_used_idx++;
1165 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1166 struct vring_used_elem *heads,
1169 struct vring_used_elem __user *used;
1172 start = vq->last_used_idx % vq->num;
1173 used = vq->used->ring + start;
1174 if (copy_to_user(used, heads, count * sizeof *used)) {
1175 vq_err(vq, "Failed to write used");
1178 if (unlikely(vq->log_used)) {
1179 /* Make sure data is seen before log. */
1181 /* Log used ring entry write. */
1182 log_write(vq->log_base,
1184 ((void __user *)used - (void __user *)vq->used),
1185 count * sizeof *used);
1187 vq->last_used_idx += count;
1191 /* After we've used one of their buffers, we tell them about it. We'll then
1192 * want to notify the guest, using eventfd. */
1193 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1198 start = vq->last_used_idx % vq->num;
1199 n = vq->num - start;
1201 r = __vhost_add_used_n(vq, heads, n);
1207 r = __vhost_add_used_n(vq, heads, count);
1209 /* Make sure buffer is written before we update index. */
1211 if (put_user(vq->last_used_idx, &vq->used->idx)) {
1212 vq_err(vq, "Failed to increment used idx");
1215 if (unlikely(vq->log_used)) {
1216 /* Log used index update. */
1217 log_write(vq->log_base,
1218 vq->log_addr + offsetof(struct vring_used, idx),
1219 sizeof vq->used->idx);
1221 eventfd_signal(vq->log_ctx, 1);
1226 /* This actually signals the guest, using eventfd. */
1227 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1230 /* Flush out used index updates. This is paired
1231 * with the barrier that the Guest executes when enabling
1235 if (get_user(flags, &vq->avail->flags)) {
1236 vq_err(vq, "Failed to get flags");
1240 /* If they don't want an interrupt, don't signal, unless empty. */
1241 if ((flags & VRING_AVAIL_F_NO_INTERRUPT) &&
1242 (vq->avail_idx != vq->last_avail_idx ||
1243 !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY)))
1246 /* Signal the Guest tell them we used something up. */
1248 eventfd_signal(vq->call_ctx, 1);
1251 /* And here's the combo meal deal. Supersize me! */
1252 void vhost_add_used_and_signal(struct vhost_dev *dev,
1253 struct vhost_virtqueue *vq,
1254 unsigned int head, int len)
1256 vhost_add_used(vq, head, len);
1257 vhost_signal(dev, vq);
1260 /* multi-buffer version of vhost_add_used_and_signal */
1261 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1262 struct vhost_virtqueue *vq,
1263 struct vring_used_elem *heads, unsigned count)
1265 vhost_add_used_n(vq, heads, count);
1266 vhost_signal(dev, vq);
1269 /* OK, now we need to know about added descriptors. */
1270 bool vhost_enable_notify(struct vhost_virtqueue *vq)
1274 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1276 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1277 r = put_user(vq->used_flags, &vq->used->flags);
1279 vq_err(vq, "Failed to enable notification at %p: %d\n",
1280 &vq->used->flags, r);
1283 /* They could have slipped one in as we were doing that: make
1284 * sure it's written, then check again. */
1286 r = get_user(avail_idx, &vq->avail->idx);
1288 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1289 &vq->avail->idx, r);
1293 return avail_idx != vq->avail_idx;
1296 /* We don't need to be notified again. */
1297 void vhost_disable_notify(struct vhost_virtqueue *vq)
1300 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1302 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1303 r = put_user(vq->used_flags, &vq->used->flags);
1305 vq_err(vq, "Failed to enable notification at %p: %d\n",
1306 &vq->used->flags, r);