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;
153 vq->private_data = NULL;
155 vq->error_ctx = NULL;
163 static int vhost_worker(void *data)
165 struct vhost_dev *dev = data;
166 struct vhost_work *work = NULL;
167 unsigned uninitialized_var(seq);
170 /* mb paired w/ kthread_stop */
171 set_current_state(TASK_INTERRUPTIBLE);
173 spin_lock_irq(&dev->work_lock);
175 work->done_seq = seq;
177 wake_up_all(&work->done);
180 if (kthread_should_stop()) {
181 spin_unlock_irq(&dev->work_lock);
182 __set_current_state(TASK_RUNNING);
185 if (!list_empty(&dev->work_list)) {
186 work = list_first_entry(&dev->work_list,
187 struct vhost_work, node);
188 list_del_init(&work->node);
189 seq = work->queue_seq;
192 spin_unlock_irq(&dev->work_lock);
195 __set_current_state(TASK_RUNNING);
203 long vhost_dev_init(struct vhost_dev *dev,
204 struct vhost_virtqueue *vqs, int nvqs)
210 mutex_init(&dev->mutex);
212 dev->log_file = NULL;
215 spin_lock_init(&dev->work_lock);
216 INIT_LIST_HEAD(&dev->work_list);
219 for (i = 0; i < dev->nvqs; ++i) {
220 dev->vqs[i].dev = dev;
221 mutex_init(&dev->vqs[i].mutex);
222 vhost_vq_reset(dev, dev->vqs + i);
223 if (dev->vqs[i].handle_kick)
224 vhost_poll_init(&dev->vqs[i].poll,
225 dev->vqs[i].handle_kick, POLLIN, dev);
231 /* Caller should have device mutex */
232 long vhost_dev_check_owner(struct vhost_dev *dev)
234 /* Are you the owner? If not, I don't think you mean to do that */
235 return dev->mm == current->mm ? 0 : -EPERM;
238 /* Caller should have device mutex */
239 static long vhost_dev_set_owner(struct vhost_dev *dev)
241 struct task_struct *worker;
243 /* Is there an owner already? */
248 /* No owner, become one */
249 dev->mm = get_task_mm(current);
250 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
251 if (IS_ERR(worker)) {
252 err = PTR_ERR(worker);
256 dev->worker = worker;
257 err = cgroup_attach_task_current_cg(worker);
260 wake_up_process(worker); /* avoid contributing to loadavg */
264 kthread_stop(worker);
273 /* Caller should have device mutex */
274 long vhost_dev_reset_owner(struct vhost_dev *dev)
276 struct vhost_memory *memory;
278 /* Restore memory to default empty mapping. */
279 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
283 vhost_dev_cleanup(dev);
285 memory->nregions = 0;
286 dev->memory = memory;
290 /* Caller should have device mutex */
291 void vhost_dev_cleanup(struct vhost_dev *dev)
294 for (i = 0; i < dev->nvqs; ++i) {
295 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
296 vhost_poll_stop(&dev->vqs[i].poll);
297 vhost_poll_flush(&dev->vqs[i].poll);
299 if (dev->vqs[i].error_ctx)
300 eventfd_ctx_put(dev->vqs[i].error_ctx);
301 if (dev->vqs[i].error)
302 fput(dev->vqs[i].error);
303 if (dev->vqs[i].kick)
304 fput(dev->vqs[i].kick);
305 if (dev->vqs[i].call_ctx)
306 eventfd_ctx_put(dev->vqs[i].call_ctx);
307 if (dev->vqs[i].call)
308 fput(dev->vqs[i].call);
309 vhost_vq_reset(dev, dev->vqs + i);
312 eventfd_ctx_put(dev->log_ctx);
316 dev->log_file = NULL;
317 /* No one will access memory at this point */
324 WARN_ON(!list_empty(&dev->work_list));
325 kthread_stop(dev->worker);
328 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
330 u64 a = addr / VHOST_PAGE_SIZE / 8;
331 /* Make sure 64 bit math will not overflow. */
332 if (a > ULONG_MAX - (unsigned long)log_base ||
333 a + (unsigned long)log_base > ULONG_MAX)
336 return access_ok(VERIFY_WRITE, log_base + a,
337 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
340 /* Caller should have vq mutex and device mutex. */
341 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
349 for (i = 0; i < mem->nregions; ++i) {
350 struct vhost_memory_region *m = mem->regions + i;
351 unsigned long a = m->userspace_addr;
352 if (m->memory_size > ULONG_MAX)
354 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
357 else if (log_all && !log_access_ok(log_base,
365 /* Can we switch to this memory table? */
366 /* Caller should have device mutex but not vq mutex */
367 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
371 for (i = 0; i < d->nvqs; ++i) {
373 mutex_lock(&d->vqs[i].mutex);
374 /* If ring is inactive, will check when it's enabled. */
375 if (d->vqs[i].private_data)
376 ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
380 mutex_unlock(&d->vqs[i].mutex);
387 static int vq_access_ok(unsigned int num,
388 struct vring_desc __user *desc,
389 struct vring_avail __user *avail,
390 struct vring_used __user *used)
392 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
393 access_ok(VERIFY_READ, avail,
394 sizeof *avail + num * sizeof *avail->ring) &&
395 access_ok(VERIFY_WRITE, used,
396 sizeof *used + num * sizeof *used->ring);
399 /* Can we log writes? */
400 /* Caller should have device mutex but not vq mutex */
401 int vhost_log_access_ok(struct vhost_dev *dev)
403 return memory_access_ok(dev, dev->memory, 1);
406 /* Verify access for write logging. */
407 /* Caller should have vq mutex and device mutex */
408 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base)
410 return vq_memory_access_ok(log_base, vq->dev->memory,
411 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
412 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
414 vq->num * sizeof *vq->used->ring));
417 /* Can we start vq? */
418 /* Caller should have vq mutex and device mutex */
419 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
421 return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) &&
422 vq_log_access_ok(vq, vq->log_base);
425 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
427 struct vhost_memory mem, *newmem, *oldmem;
428 unsigned long size = offsetof(struct vhost_memory, regions);
429 if (copy_from_user(&mem, m, size))
433 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
435 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
439 memcpy(newmem, &mem, size);
440 if (copy_from_user(newmem->regions, m->regions,
441 mem.nregions * sizeof *m->regions)) {
446 if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL))) {
451 rcu_assign_pointer(d->memory, newmem);
457 static int init_used(struct vhost_virtqueue *vq,
458 struct vring_used __user *used)
460 int r = put_user(vq->used_flags, &used->flags);
463 return get_user(vq->last_used_idx, &used->idx);
466 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
468 struct file *eventfp, *filep = NULL,
469 *pollstart = NULL, *pollstop = NULL;
470 struct eventfd_ctx *ctx = NULL;
471 u32 __user *idxp = argp;
472 struct vhost_virtqueue *vq;
473 struct vhost_vring_state s;
474 struct vhost_vring_file f;
475 struct vhost_vring_addr a;
479 r = get_user(idx, idxp);
487 mutex_lock(&vq->mutex);
490 case VHOST_SET_VRING_NUM:
491 /* Resizing ring with an active backend?
492 * You don't want to do that. */
493 if (vq->private_data) {
497 if (copy_from_user(&s, argp, sizeof s)) {
501 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
507 case VHOST_SET_VRING_BASE:
508 /* Moving base with an active backend?
509 * You don't want to do that. */
510 if (vq->private_data) {
514 if (copy_from_user(&s, argp, sizeof s)) {
518 if (s.num > 0xffff) {
522 vq->last_avail_idx = s.num;
523 /* Forget the cached index value. */
524 vq->avail_idx = vq->last_avail_idx;
526 case VHOST_GET_VRING_BASE:
528 s.num = vq->last_avail_idx;
529 if (copy_to_user(argp, &s, sizeof s))
532 case VHOST_SET_VRING_ADDR:
533 if (copy_from_user(&a, argp, sizeof a)) {
537 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
541 /* For 32bit, verify that the top 32bits of the user
542 data are set to zero. */
543 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
544 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
545 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
549 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
550 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
551 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
556 /* We only verify access here if backend is configured.
557 * If it is not, we don't as size might not have been setup.
558 * We will verify when backend is configured. */
559 if (vq->private_data) {
560 if (!vq_access_ok(vq->num,
561 (void __user *)(unsigned long)a.desc_user_addr,
562 (void __user *)(unsigned long)a.avail_user_addr,
563 (void __user *)(unsigned long)a.used_user_addr)) {
568 /* Also validate log access for used ring if enabled. */
569 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
570 !log_access_ok(vq->log_base, a.log_guest_addr,
572 vq->num * sizeof *vq->used->ring)) {
578 r = init_used(vq, (struct vring_used __user *)(unsigned long)
582 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
583 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
584 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
585 vq->log_addr = a.log_guest_addr;
586 vq->used = (void __user *)(unsigned long)a.used_user_addr;
588 case VHOST_SET_VRING_KICK:
589 if (copy_from_user(&f, argp, sizeof f)) {
593 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
594 if (IS_ERR(eventfp)) {
595 r = PTR_ERR(eventfp);
598 if (eventfp != vq->kick) {
599 pollstop = filep = vq->kick;
600 pollstart = vq->kick = eventfp;
604 case VHOST_SET_VRING_CALL:
605 if (copy_from_user(&f, argp, sizeof f)) {
609 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
610 if (IS_ERR(eventfp)) {
611 r = PTR_ERR(eventfp);
614 if (eventfp != vq->call) {
618 vq->call_ctx = eventfp ?
619 eventfd_ctx_fileget(eventfp) : NULL;
623 case VHOST_SET_VRING_ERR:
624 if (copy_from_user(&f, argp, sizeof f)) {
628 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
629 if (IS_ERR(eventfp)) {
630 r = PTR_ERR(eventfp);
633 if (eventfp != vq->error) {
637 vq->error_ctx = eventfp ?
638 eventfd_ctx_fileget(eventfp) : NULL;
646 if (pollstop && vq->handle_kick)
647 vhost_poll_stop(&vq->poll);
650 eventfd_ctx_put(ctx);
654 if (pollstart && vq->handle_kick)
655 vhost_poll_start(&vq->poll, vq->kick);
657 mutex_unlock(&vq->mutex);
659 if (pollstop && vq->handle_kick)
660 vhost_poll_flush(&vq->poll);
664 /* Caller must have device mutex */
665 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
667 void __user *argp = (void __user *)arg;
668 struct file *eventfp, *filep = NULL;
669 struct eventfd_ctx *ctx = NULL;
674 /* If you are not the owner, you can become one */
675 if (ioctl == VHOST_SET_OWNER) {
676 r = vhost_dev_set_owner(d);
680 /* You must be the owner to do anything else */
681 r = vhost_dev_check_owner(d);
686 case VHOST_SET_MEM_TABLE:
687 r = vhost_set_memory(d, argp);
689 case VHOST_SET_LOG_BASE:
690 if (copy_from_user(&p, argp, sizeof p)) {
694 if ((u64)(unsigned long)p != p) {
698 for (i = 0; i < d->nvqs; ++i) {
699 struct vhost_virtqueue *vq;
700 void __user *base = (void __user *)(unsigned long)p;
702 mutex_lock(&vq->mutex);
703 /* If ring is inactive, will check when it's enabled. */
704 if (vq->private_data && !vq_log_access_ok(vq, base))
708 mutex_unlock(&vq->mutex);
711 case VHOST_SET_LOG_FD:
712 r = get_user(fd, (int __user *)argp);
715 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
716 if (IS_ERR(eventfp)) {
717 r = PTR_ERR(eventfp);
720 if (eventfp != d->log_file) {
723 d->log_ctx = eventfp ?
724 eventfd_ctx_fileget(eventfp) : NULL;
727 for (i = 0; i < d->nvqs; ++i) {
728 mutex_lock(&d->vqs[i].mutex);
729 d->vqs[i].log_ctx = d->log_ctx;
730 mutex_unlock(&d->vqs[i].mutex);
733 eventfd_ctx_put(ctx);
738 r = vhost_set_vring(d, ioctl, argp);
745 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
746 __u64 addr, __u32 len)
748 struct vhost_memory_region *reg;
750 /* linear search is not brilliant, but we really have on the order of 6
751 * regions in practice */
752 for (i = 0; i < mem->nregions; ++i) {
753 reg = mem->regions + i;
754 if (reg->guest_phys_addr <= addr &&
755 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
761 /* TODO: This is really inefficient. We need something like get_user()
762 * (instruction directly accesses the data, with an exception table entry
763 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
765 static int set_bit_to_user(int nr, void __user *addr)
767 unsigned long log = (unsigned long)addr;
770 int bit = nr + (log % PAGE_SIZE) * 8;
772 r = get_user_pages_fast(log, 1, 1, &page);
776 base = kmap_atomic(page, KM_USER0);
778 kunmap_atomic(base, KM_USER0);
779 set_page_dirty_lock(page);
784 static int log_write(void __user *log_base,
785 u64 write_address, u64 write_length)
790 write_address /= VHOST_PAGE_SIZE;
792 u64 base = (u64)(unsigned long)log_base;
793 u64 log = base + write_address / 8;
794 int bit = write_address % 8;
795 if ((u64)(unsigned long)log != log)
797 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
800 if (write_length <= VHOST_PAGE_SIZE)
802 write_length -= VHOST_PAGE_SIZE;
803 write_address += VHOST_PAGE_SIZE;
808 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
809 unsigned int log_num, u64 len)
813 /* Make sure data written is seen before log. */
815 for (i = 0; i < log_num; ++i) {
816 u64 l = min(log[i].len, len);
817 r = log_write(vq->log_base, log[i].addr, l);
825 eventfd_signal(vq->log_ctx, 1);
826 /* Length written exceeds what we have stored. This is a bug. */
831 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
832 struct iovec iov[], int iov_size)
834 const struct vhost_memory_region *reg;
835 struct vhost_memory *mem;
842 mem = rcu_dereference(dev->memory);
843 while ((u64)len > s) {
845 if (unlikely(ret >= iov_size)) {
849 reg = find_region(mem, addr, len);
850 if (unlikely(!reg)) {
855 size = reg->memory_size - addr + reg->guest_phys_addr;
856 _iov->iov_len = min((u64)len, size);
857 _iov->iov_base = (void __user *)(unsigned long)
858 (reg->userspace_addr + addr - reg->guest_phys_addr);
868 /* Each buffer in the virtqueues is actually a chain of descriptors. This
869 * function returns the next descriptor in the chain,
870 * or -1U if we're at the end. */
871 static unsigned next_desc(struct vring_desc *desc)
875 /* If this descriptor says it doesn't chain, we're done. */
876 if (!(desc->flags & VRING_DESC_F_NEXT))
879 /* Check they're not leading us off end of descriptors. */
881 /* Make sure compiler knows to grab that: we don't want it changing! */
882 /* We will use the result as an index in an array, so most
883 * architectures only need a compiler barrier here. */
884 read_barrier_depends();
889 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
890 struct iovec iov[], unsigned int iov_size,
891 unsigned int *out_num, unsigned int *in_num,
892 struct vhost_log *log, unsigned int *log_num,
893 struct vring_desc *indirect)
895 struct vring_desc desc;
896 unsigned int i = 0, count, found = 0;
900 if (unlikely(indirect->len % sizeof desc)) {
901 vq_err(vq, "Invalid length in indirect descriptor: "
902 "len 0x%llx not multiple of 0x%zx\n",
903 (unsigned long long)indirect->len,
908 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
909 ARRAY_SIZE(vq->indirect));
910 if (unlikely(ret < 0)) {
911 vq_err(vq, "Translation failure %d in indirect.\n", ret);
915 /* We will use the result as an address to read from, so most
916 * architectures only need a compiler barrier here. */
917 read_barrier_depends();
919 count = indirect->len / sizeof desc;
920 /* Buffers are chained via a 16 bit next field, so
921 * we can have at most 2^16 of these. */
922 if (unlikely(count > USHRT_MAX + 1)) {
923 vq_err(vq, "Indirect buffer length too big: %d\n",
929 unsigned iov_count = *in_num + *out_num;
930 if (unlikely(++found > count)) {
931 vq_err(vq, "Loop detected: last one at %u "
932 "indirect size %u\n",
936 if (unlikely(memcpy_fromiovec((unsigned char *)&desc, vq->indirect,
938 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
939 i, (size_t)indirect->addr + i * sizeof desc);
942 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
943 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
944 i, (size_t)indirect->addr + i * sizeof desc);
948 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
949 iov_size - iov_count);
950 if (unlikely(ret < 0)) {
951 vq_err(vq, "Translation failure %d indirect idx %d\n",
955 /* If this is an input descriptor, increment that count. */
956 if (desc.flags & VRING_DESC_F_WRITE) {
959 log[*log_num].addr = desc.addr;
960 log[*log_num].len = desc.len;
964 /* If it's an output descriptor, they're all supposed
965 * to come before any input descriptors. */
966 if (unlikely(*in_num)) {
967 vq_err(vq, "Indirect descriptor "
968 "has out after in: idx %d\n", i);
973 } while ((i = next_desc(&desc)) != -1);
977 /* This looks in the virtqueue and for the first available buffer, and converts
978 * it to an iovec for convenient access. Since descriptors consist of some
979 * number of output then some number of input descriptors, it's actually two
980 * iovecs, but we pack them into one and note how many of each there were.
982 * This function returns the descriptor number found, or vq->num (which is
983 * never a valid descriptor number) if none was found. A negative code is
984 * returned on error. */
985 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
986 struct iovec iov[], unsigned int iov_size,
987 unsigned int *out_num, unsigned int *in_num,
988 struct vhost_log *log, unsigned int *log_num)
990 struct vring_desc desc;
991 unsigned int i, head, found = 0;
995 /* Check it isn't doing very strange things with descriptor numbers. */
996 last_avail_idx = vq->last_avail_idx;
997 if (unlikely(get_user(vq->avail_idx, &vq->avail->idx))) {
998 vq_err(vq, "Failed to access avail idx at %p\n",
1003 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1004 vq_err(vq, "Guest moved used index from %u to %u",
1005 last_avail_idx, vq->avail_idx);
1009 /* If there's nothing new since last we looked, return invalid. */
1010 if (vq->avail_idx == last_avail_idx)
1013 /* Only get avail ring entries after they have been exposed by guest. */
1016 /* Grab the next descriptor number they're advertising, and increment
1017 * the index we've seen. */
1018 if (unlikely(get_user(head,
1019 &vq->avail->ring[last_avail_idx % vq->num]))) {
1020 vq_err(vq, "Failed to read head: idx %d address %p\n",
1022 &vq->avail->ring[last_avail_idx % vq->num]);
1026 /* If their number is silly, that's an error. */
1027 if (unlikely(head >= vq->num)) {
1028 vq_err(vq, "Guest says index %u > %u is available",
1033 /* When we start there are none of either input nor output. */
1034 *out_num = *in_num = 0;
1040 unsigned iov_count = *in_num + *out_num;
1041 if (unlikely(i >= vq->num)) {
1042 vq_err(vq, "Desc index is %u > %u, head = %u",
1046 if (unlikely(++found > vq->num)) {
1047 vq_err(vq, "Loop detected: last one at %u "
1048 "vq size %u head %u\n",
1052 ret = copy_from_user(&desc, vq->desc + i, sizeof desc);
1053 if (unlikely(ret)) {
1054 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1058 if (desc.flags & VRING_DESC_F_INDIRECT) {
1059 ret = get_indirect(dev, vq, iov, iov_size,
1061 log, log_num, &desc);
1062 if (unlikely(ret < 0)) {
1063 vq_err(vq, "Failure detected "
1064 "in indirect descriptor at idx %d\n", i);
1070 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1071 iov_size - iov_count);
1072 if (unlikely(ret < 0)) {
1073 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1077 if (desc.flags & VRING_DESC_F_WRITE) {
1078 /* If this is an input descriptor,
1079 * increment that count. */
1081 if (unlikely(log)) {
1082 log[*log_num].addr = desc.addr;
1083 log[*log_num].len = desc.len;
1087 /* If it's an output descriptor, they're all supposed
1088 * to come before any input descriptors. */
1089 if (unlikely(*in_num)) {
1090 vq_err(vq, "Descriptor has out after in: "
1096 } while ((i = next_desc(&desc)) != -1);
1098 /* On success, increment avail index. */
1099 vq->last_avail_idx++;
1103 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1104 void vhost_discard_vq_desc(struct vhost_virtqueue *vq)
1106 vq->last_avail_idx--;
1109 /* After we've used one of their buffers, we tell them about it. We'll then
1110 * want to notify the guest, using eventfd. */
1111 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1113 struct vring_used_elem __user *used;
1115 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1116 * next entry in that used ring. */
1117 used = &vq->used->ring[vq->last_used_idx % vq->num];
1118 if (put_user(head, &used->id)) {
1119 vq_err(vq, "Failed to write used id");
1122 if (put_user(len, &used->len)) {
1123 vq_err(vq, "Failed to write used len");
1126 /* Make sure buffer is written before we update index. */
1128 if (put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1129 vq_err(vq, "Failed to increment used idx");
1132 if (unlikely(vq->log_used)) {
1133 /* Make sure data is seen before log. */
1135 /* Log used ring entry write. */
1136 log_write(vq->log_base,
1138 ((void __user *)used - (void __user *)vq->used),
1140 /* Log used index update. */
1141 log_write(vq->log_base,
1142 vq->log_addr + offsetof(struct vring_used, idx),
1143 sizeof vq->used->idx);
1145 eventfd_signal(vq->log_ctx, 1);
1147 vq->last_used_idx++;
1151 /* This actually signals the guest, using eventfd. */
1152 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1155 /* Flush out used index updates. This is paired
1156 * with the barrier that the Guest executes when enabling
1160 if (get_user(flags, &vq->avail->flags)) {
1161 vq_err(vq, "Failed to get flags");
1165 /* If they don't want an interrupt, don't signal, unless empty. */
1166 if ((flags & VRING_AVAIL_F_NO_INTERRUPT) &&
1167 (vq->avail_idx != vq->last_avail_idx ||
1168 !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY)))
1171 /* Signal the Guest tell them we used something up. */
1173 eventfd_signal(vq->call_ctx, 1);
1176 /* And here's the combo meal deal. Supersize me! */
1177 void vhost_add_used_and_signal(struct vhost_dev *dev,
1178 struct vhost_virtqueue *vq,
1179 unsigned int head, int len)
1181 vhost_add_used(vq, head, len);
1182 vhost_signal(dev, vq);
1185 /* OK, now we need to know about added descriptors. */
1186 bool vhost_enable_notify(struct vhost_virtqueue *vq)
1190 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1192 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1193 r = put_user(vq->used_flags, &vq->used->flags);
1195 vq_err(vq, "Failed to enable notification at %p: %d\n",
1196 &vq->used->flags, r);
1199 /* They could have slipped one in as we were doing that: make
1200 * sure it's written, then check again. */
1202 r = get_user(avail_idx, &vq->avail->idx);
1204 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1205 &vq->avail->idx, r);
1209 return avail_idx != vq->last_avail_idx;
1212 /* We don't need to be notified again. */
1213 void vhost_disable_notify(struct vhost_virtqueue *vq)
1216 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1218 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1219 r = put_user(vq->used_flags, &vq->used->flags);
1221 vq_err(vq, "Failed to enable notification at %p: %d\n",
1222 &vq->used->flags, r);