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 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 */
325 WARN_ON(!list_empty(&dev->work_list));
326 kthread_stop(dev->worker);
329 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
331 u64 a = addr / VHOST_PAGE_SIZE / 8;
332 /* Make sure 64 bit math will not overflow. */
333 if (a > ULONG_MAX - (unsigned long)log_base ||
334 a + (unsigned long)log_base > ULONG_MAX)
337 return access_ok(VERIFY_WRITE, log_base + a,
338 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
341 /* Caller should have vq mutex and device mutex. */
342 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
350 for (i = 0; i < mem->nregions; ++i) {
351 struct vhost_memory_region *m = mem->regions + i;
352 unsigned long a = m->userspace_addr;
353 if (m->memory_size > ULONG_MAX)
355 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
358 else if (log_all && !log_access_ok(log_base,
366 /* Can we switch to this memory table? */
367 /* Caller should have device mutex but not vq mutex */
368 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
372 for (i = 0; i < d->nvqs; ++i) {
374 mutex_lock(&d->vqs[i].mutex);
375 /* If ring is inactive, will check when it's enabled. */
376 if (d->vqs[i].private_data)
377 ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
381 mutex_unlock(&d->vqs[i].mutex);
388 static int vq_access_ok(unsigned int num,
389 struct vring_desc __user *desc,
390 struct vring_avail __user *avail,
391 struct vring_used __user *used)
393 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
394 access_ok(VERIFY_READ, avail,
395 sizeof *avail + num * sizeof *avail->ring) &&
396 access_ok(VERIFY_WRITE, used,
397 sizeof *used + num * sizeof *used->ring);
400 /* Can we log writes? */
401 /* Caller should have device mutex but not vq mutex */
402 int vhost_log_access_ok(struct vhost_dev *dev)
404 return memory_access_ok(dev, dev->memory, 1);
407 /* Verify access for write logging. */
408 /* Caller should have vq mutex and device mutex */
409 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base)
411 return vq_memory_access_ok(log_base, vq->dev->memory,
412 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
413 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
415 vq->num * sizeof *vq->used->ring));
418 /* Can we start vq? */
419 /* Caller should have vq mutex and device mutex */
420 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
422 return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) &&
423 vq_log_access_ok(vq, vq->log_base);
426 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
428 struct vhost_memory mem, *newmem, *oldmem;
429 unsigned long size = offsetof(struct vhost_memory, regions);
430 if (copy_from_user(&mem, m, size))
434 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
436 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
440 memcpy(newmem, &mem, size);
441 if (copy_from_user(newmem->regions, m->regions,
442 mem.nregions * sizeof *m->regions)) {
447 if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL))) {
452 rcu_assign_pointer(d->memory, newmem);
458 static int init_used(struct vhost_virtqueue *vq,
459 struct vring_used __user *used)
461 int r = put_user(vq->used_flags, &used->flags);
464 return get_user(vq->last_used_idx, &used->idx);
467 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
469 struct file *eventfp, *filep = NULL,
470 *pollstart = NULL, *pollstop = NULL;
471 struct eventfd_ctx *ctx = NULL;
472 u32 __user *idxp = argp;
473 struct vhost_virtqueue *vq;
474 struct vhost_vring_state s;
475 struct vhost_vring_file f;
476 struct vhost_vring_addr a;
480 r = get_user(idx, idxp);
488 mutex_lock(&vq->mutex);
491 case VHOST_SET_VRING_NUM:
492 /* Resizing ring with an active backend?
493 * You don't want to do that. */
494 if (vq->private_data) {
498 if (copy_from_user(&s, argp, sizeof s)) {
502 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
508 case VHOST_SET_VRING_BASE:
509 /* Moving base with an active backend?
510 * You don't want to do that. */
511 if (vq->private_data) {
515 if (copy_from_user(&s, argp, sizeof s)) {
519 if (s.num > 0xffff) {
523 vq->last_avail_idx = s.num;
524 /* Forget the cached index value. */
525 vq->avail_idx = vq->last_avail_idx;
527 case VHOST_GET_VRING_BASE:
529 s.num = vq->last_avail_idx;
530 if (copy_to_user(argp, &s, sizeof s))
533 case VHOST_SET_VRING_ADDR:
534 if (copy_from_user(&a, argp, sizeof a)) {
538 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
542 /* For 32bit, verify that the top 32bits of the user
543 data are set to zero. */
544 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
545 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
546 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
550 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
551 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
552 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
557 /* We only verify access here if backend is configured.
558 * If it is not, we don't as size might not have been setup.
559 * We will verify when backend is configured. */
560 if (vq->private_data) {
561 if (!vq_access_ok(vq->num,
562 (void __user *)(unsigned long)a.desc_user_addr,
563 (void __user *)(unsigned long)a.avail_user_addr,
564 (void __user *)(unsigned long)a.used_user_addr)) {
569 /* Also validate log access for used ring if enabled. */
570 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
571 !log_access_ok(vq->log_base, a.log_guest_addr,
573 vq->num * sizeof *vq->used->ring)) {
579 r = init_used(vq, (struct vring_used __user *)(unsigned long)
583 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
584 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
585 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
586 vq->log_addr = a.log_guest_addr;
587 vq->used = (void __user *)(unsigned long)a.used_user_addr;
589 case VHOST_SET_VRING_KICK:
590 if (copy_from_user(&f, argp, sizeof f)) {
594 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
595 if (IS_ERR(eventfp)) {
596 r = PTR_ERR(eventfp);
599 if (eventfp != vq->kick) {
600 pollstop = filep = vq->kick;
601 pollstart = vq->kick = eventfp;
605 case VHOST_SET_VRING_CALL:
606 if (copy_from_user(&f, argp, sizeof f)) {
610 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
611 if (IS_ERR(eventfp)) {
612 r = PTR_ERR(eventfp);
615 if (eventfp != vq->call) {
619 vq->call_ctx = eventfp ?
620 eventfd_ctx_fileget(eventfp) : NULL;
624 case VHOST_SET_VRING_ERR:
625 if (copy_from_user(&f, argp, sizeof f)) {
629 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
630 if (IS_ERR(eventfp)) {
631 r = PTR_ERR(eventfp);
634 if (eventfp != vq->error) {
638 vq->error_ctx = eventfp ?
639 eventfd_ctx_fileget(eventfp) : NULL;
647 if (pollstop && vq->handle_kick)
648 vhost_poll_stop(&vq->poll);
651 eventfd_ctx_put(ctx);
655 if (pollstart && vq->handle_kick)
656 vhost_poll_start(&vq->poll, vq->kick);
658 mutex_unlock(&vq->mutex);
660 if (pollstop && vq->handle_kick)
661 vhost_poll_flush(&vq->poll);
665 /* Caller must have device mutex */
666 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
668 void __user *argp = (void __user *)arg;
669 struct file *eventfp, *filep = NULL;
670 struct eventfd_ctx *ctx = NULL;
675 /* If you are not the owner, you can become one */
676 if (ioctl == VHOST_SET_OWNER) {
677 r = vhost_dev_set_owner(d);
681 /* You must be the owner to do anything else */
682 r = vhost_dev_check_owner(d);
687 case VHOST_SET_MEM_TABLE:
688 r = vhost_set_memory(d, argp);
690 case VHOST_SET_LOG_BASE:
691 if (copy_from_user(&p, argp, sizeof p)) {
695 if ((u64)(unsigned long)p != p) {
699 for (i = 0; i < d->nvqs; ++i) {
700 struct vhost_virtqueue *vq;
701 void __user *base = (void __user *)(unsigned long)p;
703 mutex_lock(&vq->mutex);
704 /* If ring is inactive, will check when it's enabled. */
705 if (vq->private_data && !vq_log_access_ok(vq, base))
709 mutex_unlock(&vq->mutex);
712 case VHOST_SET_LOG_FD:
713 r = get_user(fd, (int __user *)argp);
716 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
717 if (IS_ERR(eventfp)) {
718 r = PTR_ERR(eventfp);
721 if (eventfp != d->log_file) {
724 d->log_ctx = eventfp ?
725 eventfd_ctx_fileget(eventfp) : NULL;
728 for (i = 0; i < d->nvqs; ++i) {
729 mutex_lock(&d->vqs[i].mutex);
730 d->vqs[i].log_ctx = d->log_ctx;
731 mutex_unlock(&d->vqs[i].mutex);
734 eventfd_ctx_put(ctx);
739 r = vhost_set_vring(d, ioctl, argp);
746 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
747 __u64 addr, __u32 len)
749 struct vhost_memory_region *reg;
751 /* linear search is not brilliant, but we really have on the order of 6
752 * regions in practice */
753 for (i = 0; i < mem->nregions; ++i) {
754 reg = mem->regions + i;
755 if (reg->guest_phys_addr <= addr &&
756 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
762 /* TODO: This is really inefficient. We need something like get_user()
763 * (instruction directly accesses the data, with an exception table entry
764 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
766 static int set_bit_to_user(int nr, void __user *addr)
768 unsigned long log = (unsigned long)addr;
771 int bit = nr + (log % PAGE_SIZE) * 8;
773 r = get_user_pages_fast(log, 1, 1, &page);
777 base = kmap_atomic(page, KM_USER0);
779 kunmap_atomic(base, KM_USER0);
780 set_page_dirty_lock(page);
785 static int log_write(void __user *log_base,
786 u64 write_address, u64 write_length)
791 write_address /= VHOST_PAGE_SIZE;
793 u64 base = (u64)(unsigned long)log_base;
794 u64 log = base + write_address / 8;
795 int bit = write_address % 8;
796 if ((u64)(unsigned long)log != log)
798 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
801 if (write_length <= VHOST_PAGE_SIZE)
803 write_length -= VHOST_PAGE_SIZE;
804 write_address += VHOST_PAGE_SIZE;
809 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
810 unsigned int log_num, u64 len)
814 /* Make sure data written is seen before log. */
816 for (i = 0; i < log_num; ++i) {
817 u64 l = min(log[i].len, len);
818 r = log_write(vq->log_base, log[i].addr, l);
826 eventfd_signal(vq->log_ctx, 1);
827 /* Length written exceeds what we have stored. This is a bug. */
832 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
833 struct iovec iov[], int iov_size)
835 const struct vhost_memory_region *reg;
836 struct vhost_memory *mem;
843 mem = rcu_dereference(dev->memory);
844 while ((u64)len > s) {
846 if (unlikely(ret >= iov_size)) {
850 reg = find_region(mem, addr, len);
851 if (unlikely(!reg)) {
856 size = reg->memory_size - addr + reg->guest_phys_addr;
857 _iov->iov_len = min((u64)len, size);
858 _iov->iov_base = (void __user *)(unsigned long)
859 (reg->userspace_addr + addr - reg->guest_phys_addr);
869 /* Each buffer in the virtqueues is actually a chain of descriptors. This
870 * function returns the next descriptor in the chain,
871 * or -1U if we're at the end. */
872 static unsigned next_desc(struct vring_desc *desc)
876 /* If this descriptor says it doesn't chain, we're done. */
877 if (!(desc->flags & VRING_DESC_F_NEXT))
880 /* Check they're not leading us off end of descriptors. */
882 /* Make sure compiler knows to grab that: we don't want it changing! */
883 /* We will use the result as an index in an array, so most
884 * architectures only need a compiler barrier here. */
885 read_barrier_depends();
890 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
891 struct iovec iov[], unsigned int iov_size,
892 unsigned int *out_num, unsigned int *in_num,
893 struct vhost_log *log, unsigned int *log_num,
894 struct vring_desc *indirect)
896 struct vring_desc desc;
897 unsigned int i = 0, count, found = 0;
901 if (unlikely(indirect->len % sizeof desc)) {
902 vq_err(vq, "Invalid length in indirect descriptor: "
903 "len 0x%llx not multiple of 0x%zx\n",
904 (unsigned long long)indirect->len,
909 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
910 ARRAY_SIZE(vq->indirect));
911 if (unlikely(ret < 0)) {
912 vq_err(vq, "Translation failure %d in indirect.\n", ret);
916 /* We will use the result as an address to read from, so most
917 * architectures only need a compiler barrier here. */
918 read_barrier_depends();
920 count = indirect->len / sizeof desc;
921 /* Buffers are chained via a 16 bit next field, so
922 * we can have at most 2^16 of these. */
923 if (unlikely(count > USHRT_MAX + 1)) {
924 vq_err(vq, "Indirect buffer length too big: %d\n",
930 unsigned iov_count = *in_num + *out_num;
931 if (unlikely(++found > count)) {
932 vq_err(vq, "Loop detected: last one at %u "
933 "indirect size %u\n",
937 if (unlikely(memcpy_fromiovec((unsigned char *)&desc, vq->indirect,
939 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
940 i, (size_t)indirect->addr + i * sizeof desc);
943 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
944 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
945 i, (size_t)indirect->addr + i * sizeof desc);
949 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
950 iov_size - iov_count);
951 if (unlikely(ret < 0)) {
952 vq_err(vq, "Translation failure %d indirect idx %d\n",
956 /* If this is an input descriptor, increment that count. */
957 if (desc.flags & VRING_DESC_F_WRITE) {
960 log[*log_num].addr = desc.addr;
961 log[*log_num].len = desc.len;
965 /* If it's an output descriptor, they're all supposed
966 * to come before any input descriptors. */
967 if (unlikely(*in_num)) {
968 vq_err(vq, "Indirect descriptor "
969 "has out after in: idx %d\n", i);
974 } while ((i = next_desc(&desc)) != -1);
978 /* This looks in the virtqueue and for the first available buffer, and converts
979 * it to an iovec for convenient access. Since descriptors consist of some
980 * number of output then some number of input descriptors, it's actually two
981 * iovecs, but we pack them into one and note how many of each there were.
983 * This function returns the descriptor number found, or vq->num (which is
984 * never a valid descriptor number) if none was found. A negative code is
985 * returned on error. */
986 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
987 struct iovec iov[], unsigned int iov_size,
988 unsigned int *out_num, unsigned int *in_num,
989 struct vhost_log *log, unsigned int *log_num)
991 struct vring_desc desc;
992 unsigned int i, head, found = 0;
996 /* Check it isn't doing very strange things with descriptor numbers. */
997 last_avail_idx = vq->last_avail_idx;
998 if (unlikely(get_user(vq->avail_idx, &vq->avail->idx))) {
999 vq_err(vq, "Failed to access avail idx at %p\n",
1004 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1005 vq_err(vq, "Guest moved used index from %u to %u",
1006 last_avail_idx, vq->avail_idx);
1010 /* If there's nothing new since last we looked, return invalid. */
1011 if (vq->avail_idx == last_avail_idx)
1014 /* Only get avail ring entries after they have been exposed by guest. */
1017 /* Grab the next descriptor number they're advertising, and increment
1018 * the index we've seen. */
1019 if (unlikely(get_user(head,
1020 &vq->avail->ring[last_avail_idx % vq->num]))) {
1021 vq_err(vq, "Failed to read head: idx %d address %p\n",
1023 &vq->avail->ring[last_avail_idx % vq->num]);
1027 /* If their number is silly, that's an error. */
1028 if (unlikely(head >= vq->num)) {
1029 vq_err(vq, "Guest says index %u > %u is available",
1034 /* When we start there are none of either input nor output. */
1035 *out_num = *in_num = 0;
1041 unsigned iov_count = *in_num + *out_num;
1042 if (unlikely(i >= vq->num)) {
1043 vq_err(vq, "Desc index is %u > %u, head = %u",
1047 if (unlikely(++found > vq->num)) {
1048 vq_err(vq, "Loop detected: last one at %u "
1049 "vq size %u head %u\n",
1053 ret = copy_from_user(&desc, vq->desc + i, sizeof desc);
1054 if (unlikely(ret)) {
1055 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1059 if (desc.flags & VRING_DESC_F_INDIRECT) {
1060 ret = get_indirect(dev, vq, iov, iov_size,
1062 log, log_num, &desc);
1063 if (unlikely(ret < 0)) {
1064 vq_err(vq, "Failure detected "
1065 "in indirect descriptor at idx %d\n", i);
1071 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1072 iov_size - iov_count);
1073 if (unlikely(ret < 0)) {
1074 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1078 if (desc.flags & VRING_DESC_F_WRITE) {
1079 /* If this is an input descriptor,
1080 * increment that count. */
1082 if (unlikely(log)) {
1083 log[*log_num].addr = desc.addr;
1084 log[*log_num].len = desc.len;
1088 /* If it's an output descriptor, they're all supposed
1089 * to come before any input descriptors. */
1090 if (unlikely(*in_num)) {
1091 vq_err(vq, "Descriptor has out after in: "
1097 } while ((i = next_desc(&desc)) != -1);
1099 /* On success, increment avail index. */
1100 vq->last_avail_idx++;
1104 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1105 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1107 vq->last_avail_idx -= n;
1110 /* After we've used one of their buffers, we tell them about it. We'll then
1111 * want to notify the guest, using eventfd. */
1112 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1114 struct vring_used_elem __user *used;
1116 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1117 * next entry in that used ring. */
1118 used = &vq->used->ring[vq->last_used_idx % vq->num];
1119 if (put_user(head, &used->id)) {
1120 vq_err(vq, "Failed to write used id");
1123 if (put_user(len, &used->len)) {
1124 vq_err(vq, "Failed to write used len");
1127 /* Make sure buffer is written before we update index. */
1129 if (put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1130 vq_err(vq, "Failed to increment used idx");
1133 if (unlikely(vq->log_used)) {
1134 /* Make sure data is seen before log. */
1136 /* Log used ring entry write. */
1137 log_write(vq->log_base,
1139 ((void __user *)used - (void __user *)vq->used),
1141 /* Log used index update. */
1142 log_write(vq->log_base,
1143 vq->log_addr + offsetof(struct vring_used, idx),
1144 sizeof vq->used->idx);
1146 eventfd_signal(vq->log_ctx, 1);
1148 vq->last_used_idx++;
1152 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1153 struct vring_used_elem *heads,
1156 struct vring_used_elem __user *used;
1159 start = vq->last_used_idx % vq->num;
1160 used = vq->used->ring + start;
1161 if (copy_to_user(used, heads, count * sizeof *used)) {
1162 vq_err(vq, "Failed to write used");
1165 if (unlikely(vq->log_used)) {
1166 /* Make sure data is seen before log. */
1168 /* Log used ring entry write. */
1169 log_write(vq->log_base,
1171 ((void __user *)used - (void __user *)vq->used),
1172 count * sizeof *used);
1174 vq->last_used_idx += count;
1178 /* After we've used one of their buffers, we tell them about it. We'll then
1179 * want to notify the guest, using eventfd. */
1180 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1185 start = vq->last_used_idx % vq->num;
1186 n = vq->num - start;
1188 r = __vhost_add_used_n(vq, heads, n);
1194 r = __vhost_add_used_n(vq, heads, count);
1196 /* Make sure buffer is written before we update index. */
1198 if (put_user(vq->last_used_idx, &vq->used->idx)) {
1199 vq_err(vq, "Failed to increment used idx");
1202 if (unlikely(vq->log_used)) {
1203 /* Log used index update. */
1204 log_write(vq->log_base,
1205 vq->log_addr + offsetof(struct vring_used, idx),
1206 sizeof vq->used->idx);
1208 eventfd_signal(vq->log_ctx, 1);
1213 /* This actually signals the guest, using eventfd. */
1214 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1217 /* Flush out used index updates. This is paired
1218 * with the barrier that the Guest executes when enabling
1222 if (get_user(flags, &vq->avail->flags)) {
1223 vq_err(vq, "Failed to get flags");
1227 /* If they don't want an interrupt, don't signal, unless empty. */
1228 if ((flags & VRING_AVAIL_F_NO_INTERRUPT) &&
1229 (vq->avail_idx != vq->last_avail_idx ||
1230 !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY)))
1233 /* Signal the Guest tell them we used something up. */
1235 eventfd_signal(vq->call_ctx, 1);
1238 /* And here's the combo meal deal. Supersize me! */
1239 void vhost_add_used_and_signal(struct vhost_dev *dev,
1240 struct vhost_virtqueue *vq,
1241 unsigned int head, int len)
1243 vhost_add_used(vq, head, len);
1244 vhost_signal(dev, vq);
1247 /* multi-buffer version of vhost_add_used_and_signal */
1248 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1249 struct vhost_virtqueue *vq,
1250 struct vring_used_elem *heads, unsigned count)
1252 vhost_add_used_n(vq, heads, count);
1253 vhost_signal(dev, vq);
1256 /* OK, now we need to know about added descriptors. */
1257 bool vhost_enable_notify(struct vhost_virtqueue *vq)
1261 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1263 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1264 r = put_user(vq->used_flags, &vq->used->flags);
1266 vq_err(vq, "Failed to enable notification at %p: %d\n",
1267 &vq->used->flags, r);
1270 /* They could have slipped one in as we were doing that: make
1271 * sure it's written, then check again. */
1273 r = get_user(avail_idx, &vq->avail->idx);
1275 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1276 &vq->avail->idx, r);
1280 return avail_idx != vq->avail_idx;
1283 /* We don't need to be notified again. */
1284 void vhost_disable_notify(struct vhost_virtqueue *vq)
1287 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1289 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1290 r = put_user(vq->used_flags, &vq->used->flags);
1292 vq_err(vq, "Failed to enable notification at %p: %d\n",
1293 &vq->used->flags, r);