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/workqueue.h>
21 #include <linux/rcupdate.h>
22 #include <linux/poll.h>
23 #include <linux/file.h>
24 #include <linux/highmem.h>
26 #include <linux/net.h>
27 #include <linux/if_packet.h>
28 #include <linux/if_arp.h>
35 VHOST_MEMORY_MAX_NREGIONS = 64,
36 VHOST_MEMORY_F_LOG = 0x1,
39 static struct workqueue_struct *vhost_workqueue;
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;
55 poll = container_of(wait, struct vhost_poll, wait);
56 if (!((unsigned long)key & poll->mask))
59 queue_work(vhost_workqueue, &poll->work);
63 /* Init poll structure */
64 void vhost_poll_init(struct vhost_poll *poll, work_func_t func,
67 INIT_WORK(&poll->work, func);
68 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
69 init_poll_funcptr(&poll->table, vhost_poll_func);
73 /* Start polling a file. We add ourselves to file's wait queue. The caller must
74 * keep a reference to a file until after vhost_poll_stop is called. */
75 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
78 mask = file->f_op->poll(file, &poll->table);
80 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
83 /* Stop polling a file. After this function returns, it becomes safe to drop the
84 * file reference. You must also flush afterwards. */
85 void vhost_poll_stop(struct vhost_poll *poll)
87 remove_wait_queue(poll->wqh, &poll->wait);
90 /* Flush any work that has been scheduled. When calling this, don't hold any
91 * locks that are also used by the callback. */
92 void vhost_poll_flush(struct vhost_poll *poll)
94 flush_work(&poll->work);
97 void vhost_poll_queue(struct vhost_poll *poll)
99 queue_work(vhost_workqueue, &poll->work);
102 static void vhost_vq_reset(struct vhost_dev *dev,
103 struct vhost_virtqueue *vq)
109 vq->last_avail_idx = 0;
111 vq->last_used_idx = 0;
114 vq->log_used = false;
115 vq->log_addr = -1ull;
117 vq->private_data = NULL;
119 vq->error_ctx = NULL;
126 long vhost_dev_init(struct vhost_dev *dev,
127 struct vhost_virtqueue *vqs, int nvqs)
132 mutex_init(&dev->mutex);
134 dev->log_file = NULL;
138 for (i = 0; i < dev->nvqs; ++i) {
139 dev->vqs[i].dev = dev;
140 mutex_init(&dev->vqs[i].mutex);
141 vhost_vq_reset(dev, dev->vqs + i);
142 if (dev->vqs[i].handle_kick)
143 vhost_poll_init(&dev->vqs[i].poll,
144 dev->vqs[i].handle_kick,
150 /* Caller should have device mutex */
151 long vhost_dev_check_owner(struct vhost_dev *dev)
153 /* Are you the owner? If not, I don't think you mean to do that */
154 return dev->mm == current->mm ? 0 : -EPERM;
157 /* Caller should have device mutex */
158 static long vhost_dev_set_owner(struct vhost_dev *dev)
160 /* Is there an owner already? */
163 /* No owner, become one */
164 dev->mm = get_task_mm(current);
168 /* Caller should have device mutex */
169 long vhost_dev_reset_owner(struct vhost_dev *dev)
171 struct vhost_memory *memory;
173 /* Restore memory to default empty mapping. */
174 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
178 vhost_dev_cleanup(dev);
180 memory->nregions = 0;
181 dev->memory = memory;
185 /* Caller should have device mutex */
186 void vhost_dev_cleanup(struct vhost_dev *dev)
189 for (i = 0; i < dev->nvqs; ++i) {
190 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
191 vhost_poll_stop(&dev->vqs[i].poll);
192 vhost_poll_flush(&dev->vqs[i].poll);
194 if (dev->vqs[i].error_ctx)
195 eventfd_ctx_put(dev->vqs[i].error_ctx);
196 if (dev->vqs[i].error)
197 fput(dev->vqs[i].error);
198 if (dev->vqs[i].kick)
199 fput(dev->vqs[i].kick);
200 if (dev->vqs[i].call_ctx)
201 eventfd_ctx_put(dev->vqs[i].call_ctx);
202 if (dev->vqs[i].call)
203 fput(dev->vqs[i].call);
204 vhost_vq_reset(dev, dev->vqs + i);
207 eventfd_ctx_put(dev->log_ctx);
211 dev->log_file = NULL;
212 /* No one will access memory at this point */
220 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
222 u64 a = addr / VHOST_PAGE_SIZE / 8;
223 /* Make sure 64 bit math will not overflow. */
224 if (a > ULONG_MAX - (unsigned long)log_base ||
225 a + (unsigned long)log_base > ULONG_MAX)
228 return access_ok(VERIFY_WRITE, log_base + a,
229 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
232 /* Caller should have vq mutex and device mutex. */
233 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
237 for (i = 0; i < mem->nregions; ++i) {
238 struct vhost_memory_region *m = mem->regions + i;
239 unsigned long a = m->userspace_addr;
240 if (m->memory_size > ULONG_MAX)
242 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
245 else if (log_all && !log_access_ok(log_base,
253 /* Can we switch to this memory table? */
254 /* Caller should have device mutex but not vq mutex */
255 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
259 for (i = 0; i < d->nvqs; ++i) {
261 mutex_lock(&d->vqs[i].mutex);
262 /* If ring is inactive, will check when it's enabled. */
263 if (d->vqs[i].private_data)
264 ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
268 mutex_unlock(&d->vqs[i].mutex);
275 static int vq_access_ok(unsigned int num,
276 struct vring_desc __user *desc,
277 struct vring_avail __user *avail,
278 struct vring_used __user *used)
280 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
281 access_ok(VERIFY_READ, avail,
282 sizeof *avail + num * sizeof *avail->ring) &&
283 access_ok(VERIFY_WRITE, used,
284 sizeof *used + num * sizeof *used->ring);
287 /* Can we log writes? */
288 /* Caller should have device mutex but not vq mutex */
289 int vhost_log_access_ok(struct vhost_dev *dev)
291 return memory_access_ok(dev, dev->memory, 1);
294 /* Verify access for write logging. */
295 /* Caller should have vq mutex and device mutex */
296 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base)
298 return vq_memory_access_ok(log_base, vq->dev->memory,
299 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
300 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
302 vq->num * sizeof *vq->used->ring));
305 /* Can we start vq? */
306 /* Caller should have vq mutex and device mutex */
307 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
309 return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) &&
310 vq_log_access_ok(vq, vq->log_base);
313 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
315 struct vhost_memory mem, *newmem, *oldmem;
316 unsigned long size = offsetof(struct vhost_memory, regions);
318 r = copy_from_user(&mem, m, size);
323 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
325 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
329 memcpy(newmem, &mem, size);
330 r = copy_from_user(newmem->regions, m->regions,
331 mem.nregions * sizeof *m->regions);
337 if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL)))
340 rcu_assign_pointer(d->memory, newmem);
346 static int init_used(struct vhost_virtqueue *vq,
347 struct vring_used __user *used)
349 int r = put_user(vq->used_flags, &used->flags);
352 return get_user(vq->last_used_idx, &used->idx);
355 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
357 struct file *eventfp, *filep = NULL,
358 *pollstart = NULL, *pollstop = NULL;
359 struct eventfd_ctx *ctx = NULL;
360 u32 __user *idxp = argp;
361 struct vhost_virtqueue *vq;
362 struct vhost_vring_state s;
363 struct vhost_vring_file f;
364 struct vhost_vring_addr a;
368 r = get_user(idx, idxp);
376 mutex_lock(&vq->mutex);
379 case VHOST_SET_VRING_NUM:
380 /* Resizing ring with an active backend?
381 * You don't want to do that. */
382 if (vq->private_data) {
386 r = copy_from_user(&s, argp, sizeof s);
389 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
395 case VHOST_SET_VRING_BASE:
396 /* Moving base with an active backend?
397 * You don't want to do that. */
398 if (vq->private_data) {
402 r = copy_from_user(&s, argp, sizeof s);
405 if (s.num > 0xffff) {
409 vq->last_avail_idx = s.num;
410 /* Forget the cached index value. */
411 vq->avail_idx = vq->last_avail_idx;
413 case VHOST_GET_VRING_BASE:
415 s.num = vq->last_avail_idx;
416 r = copy_to_user(argp, &s, sizeof s);
418 case VHOST_SET_VRING_ADDR:
419 r = copy_from_user(&a, argp, sizeof a);
422 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
426 /* For 32bit, verify that the top 32bits of the user
427 data are set to zero. */
428 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
429 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
430 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
434 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
435 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
436 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
441 /* We only verify access here if backend is configured.
442 * If it is not, we don't as size might not have been setup.
443 * We will verify when backend is configured. */
444 if (vq->private_data) {
445 if (!vq_access_ok(vq->num,
446 (void __user *)(unsigned long)a.desc_user_addr,
447 (void __user *)(unsigned long)a.avail_user_addr,
448 (void __user *)(unsigned long)a.used_user_addr)) {
453 /* Also validate log access for used ring if enabled. */
454 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
455 !log_access_ok(vq->log_base, a.log_guest_addr,
457 vq->num * sizeof *vq->used->ring)) {
463 r = init_used(vq, (struct vring_used __user *)(unsigned long)
467 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
468 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
469 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
470 vq->log_addr = a.log_guest_addr;
471 vq->used = (void __user *)(unsigned long)a.used_user_addr;
473 case VHOST_SET_VRING_KICK:
474 r = copy_from_user(&f, argp, sizeof f);
477 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
479 return PTR_ERR(eventfp);
480 if (eventfp != vq->kick) {
481 pollstop = filep = vq->kick;
482 pollstart = vq->kick = eventfp;
486 case VHOST_SET_VRING_CALL:
487 r = copy_from_user(&f, argp, sizeof f);
490 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
492 return PTR_ERR(eventfp);
493 if (eventfp != vq->call) {
497 vq->call_ctx = eventfp ?
498 eventfd_ctx_fileget(eventfp) : NULL;
502 case VHOST_SET_VRING_ERR:
503 r = copy_from_user(&f, argp, sizeof f);
506 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
508 return PTR_ERR(eventfp);
509 if (eventfp != vq->error) {
513 vq->error_ctx = eventfp ?
514 eventfd_ctx_fileget(eventfp) : NULL;
522 if (pollstop && vq->handle_kick)
523 vhost_poll_stop(&vq->poll);
526 eventfd_ctx_put(ctx);
530 if (pollstart && vq->handle_kick)
531 vhost_poll_start(&vq->poll, vq->kick);
533 mutex_unlock(&vq->mutex);
535 if (pollstop && vq->handle_kick)
536 vhost_poll_flush(&vq->poll);
540 /* Caller must have device mutex */
541 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
543 void __user *argp = (void __user *)arg;
544 struct file *eventfp, *filep = NULL;
545 struct eventfd_ctx *ctx = NULL;
550 /* If you are not the owner, you can become one */
551 if (ioctl == VHOST_SET_OWNER) {
552 r = vhost_dev_set_owner(d);
556 /* You must be the owner to do anything else */
557 r = vhost_dev_check_owner(d);
562 case VHOST_SET_MEM_TABLE:
563 r = vhost_set_memory(d, argp);
565 case VHOST_SET_LOG_BASE:
566 r = copy_from_user(&p, argp, sizeof p);
569 if ((u64)(unsigned long)p != p) {
573 for (i = 0; i < d->nvqs; ++i) {
574 struct vhost_virtqueue *vq;
575 void __user *base = (void __user *)(unsigned long)p;
577 mutex_lock(&vq->mutex);
578 /* If ring is inactive, will check when it's enabled. */
579 if (vq->private_data && !vq_log_access_ok(vq, base))
583 mutex_unlock(&vq->mutex);
586 case VHOST_SET_LOG_FD:
587 r = get_user(fd, (int __user *)argp);
590 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
591 if (IS_ERR(eventfp)) {
592 r = PTR_ERR(eventfp);
595 if (eventfp != d->log_file) {
598 d->log_ctx = eventfp ?
599 eventfd_ctx_fileget(eventfp) : NULL;
602 for (i = 0; i < d->nvqs; ++i) {
603 mutex_lock(&d->vqs[i].mutex);
604 d->vqs[i].log_ctx = d->log_ctx;
605 mutex_unlock(&d->vqs[i].mutex);
608 eventfd_ctx_put(ctx);
613 r = vhost_set_vring(d, ioctl, argp);
620 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
621 __u64 addr, __u32 len)
623 struct vhost_memory_region *reg;
625 /* linear search is not brilliant, but we really have on the order of 6
626 * regions in practice */
627 for (i = 0; i < mem->nregions; ++i) {
628 reg = mem->regions + i;
629 if (reg->guest_phys_addr <= addr &&
630 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
636 /* TODO: This is really inefficient. We need something like get_user()
637 * (instruction directly accesses the data, with an exception table entry
638 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
640 static int set_bit_to_user(int nr, void __user *addr)
642 unsigned long log = (unsigned long)addr;
645 int bit = nr + (log % PAGE_SIZE) * 8;
647 r = get_user_pages_fast(log, 1, 1, &page);
650 base = kmap_atomic(page, KM_USER0);
652 kunmap_atomic(base, KM_USER0);
653 set_page_dirty_lock(page);
658 static int log_write(void __user *log_base,
659 u64 write_address, u64 write_length)
664 write_address /= VHOST_PAGE_SIZE;
666 u64 base = (u64)(unsigned long)log_base;
667 u64 log = base + write_address / 8;
668 int bit = write_address % 8;
669 if ((u64)(unsigned long)log != log)
671 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
674 if (write_length <= VHOST_PAGE_SIZE)
676 write_length -= VHOST_PAGE_SIZE;
677 write_address += VHOST_PAGE_SIZE;
682 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
683 unsigned int log_num, u64 len)
687 /* Make sure data written is seen before log. */
689 for (i = 0; i < log_num; ++i) {
690 u64 l = min(log[i].len, len);
691 r = log_write(vq->log_base, log[i].addr, l);
699 eventfd_signal(vq->log_ctx, 1);
700 /* Length written exceeds what we have stored. This is a bug. */
705 int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
706 struct iovec iov[], int iov_size)
708 const struct vhost_memory_region *reg;
709 struct vhost_memory *mem;
716 mem = rcu_dereference(dev->memory);
717 while ((u64)len > s) {
719 if (ret >= iov_size) {
723 reg = find_region(mem, addr, len);
729 size = reg->memory_size - addr + reg->guest_phys_addr;
730 _iov->iov_len = min((u64)len, size);
731 _iov->iov_base = (void *)(unsigned long)
732 (reg->userspace_addr + addr - reg->guest_phys_addr);
742 /* Each buffer in the virtqueues is actually a chain of descriptors. This
743 * function returns the next descriptor in the chain,
744 * or -1U if we're at the end. */
745 static unsigned next_desc(struct vring_desc *desc)
749 /* If this descriptor says it doesn't chain, we're done. */
750 if (!(desc->flags & VRING_DESC_F_NEXT))
753 /* Check they're not leading us off end of descriptors. */
755 /* Make sure compiler knows to grab that: we don't want it changing! */
756 /* We will use the result as an index in an array, so most
757 * architectures only need a compiler barrier here. */
758 read_barrier_depends();
763 static unsigned get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
764 struct iovec iov[], unsigned int iov_size,
765 unsigned int *out_num, unsigned int *in_num,
766 struct vhost_log *log, unsigned int *log_num,
767 struct vring_desc *indirect)
769 struct vring_desc desc;
770 unsigned int i = 0, count, found = 0;
774 if (indirect->len % sizeof desc) {
775 vq_err(vq, "Invalid length in indirect descriptor: "
776 "len 0x%llx not multiple of 0x%zx\n",
777 (unsigned long long)indirect->len,
782 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
783 ARRAY_SIZE(vq->indirect));
785 vq_err(vq, "Translation failure %d in indirect.\n", ret);
789 /* We will use the result as an address to read from, so most
790 * architectures only need a compiler barrier here. */
791 read_barrier_depends();
793 count = indirect->len / sizeof desc;
794 /* Buffers are chained via a 16 bit next field, so
795 * we can have at most 2^16 of these. */
796 if (count > USHORT_MAX + 1) {
797 vq_err(vq, "Indirect buffer length too big: %d\n",
803 unsigned iov_count = *in_num + *out_num;
804 if (++found > count) {
805 vq_err(vq, "Loop detected: last one at %u "
806 "indirect size %u\n",
810 if (memcpy_fromiovec((unsigned char *)&desc, vq->indirect,
812 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
813 i, (size_t)indirect->addr + i * sizeof desc);
816 if (desc.flags & VRING_DESC_F_INDIRECT) {
817 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
818 i, (size_t)indirect->addr + i * sizeof desc);
822 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
823 iov_size - iov_count);
825 vq_err(vq, "Translation failure %d indirect idx %d\n",
829 /* If this is an input descriptor, increment that count. */
830 if (desc.flags & VRING_DESC_F_WRITE) {
833 log[*log_num].addr = desc.addr;
834 log[*log_num].len = desc.len;
838 /* If it's an output descriptor, they're all supposed
839 * to come before any input descriptors. */
841 vq_err(vq, "Indirect descriptor "
842 "has out after in: idx %d\n", i);
847 } while ((i = next_desc(&desc)) != -1);
851 /* This looks in the virtqueue and for the first available buffer, and converts
852 * it to an iovec for convenient access. Since descriptors consist of some
853 * number of output then some number of input descriptors, it's actually two
854 * iovecs, but we pack them into one and note how many of each there were.
856 * This function returns the descriptor number found, or vq->num (which
857 * is never a valid descriptor number) if none was found. */
858 unsigned vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
859 struct iovec iov[], unsigned int iov_size,
860 unsigned int *out_num, unsigned int *in_num,
861 struct vhost_log *log, unsigned int *log_num)
863 struct vring_desc desc;
864 unsigned int i, head, found = 0;
868 /* Check it isn't doing very strange things with descriptor numbers. */
869 last_avail_idx = vq->last_avail_idx;
870 if (get_user(vq->avail_idx, &vq->avail->idx)) {
871 vq_err(vq, "Failed to access avail idx at %p\n",
876 if ((u16)(vq->avail_idx - last_avail_idx) > vq->num) {
877 vq_err(vq, "Guest moved used index from %u to %u",
878 last_avail_idx, vq->avail_idx);
882 /* If there's nothing new since last we looked, return invalid. */
883 if (vq->avail_idx == last_avail_idx)
886 /* Only get avail ring entries after they have been exposed by guest. */
889 /* Grab the next descriptor number they're advertising, and increment
890 * the index we've seen. */
891 if (get_user(head, &vq->avail->ring[last_avail_idx % vq->num])) {
892 vq_err(vq, "Failed to read head: idx %d address %p\n",
894 &vq->avail->ring[last_avail_idx % vq->num]);
898 /* If their number is silly, that's an error. */
899 if (head >= vq->num) {
900 vq_err(vq, "Guest says index %u > %u is available",
905 /* When we start there are none of either input nor output. */
906 *out_num = *in_num = 0;
912 unsigned iov_count = *in_num + *out_num;
914 vq_err(vq, "Desc index is %u > %u, head = %u",
918 if (++found > vq->num) {
919 vq_err(vq, "Loop detected: last one at %u "
920 "vq size %u head %u\n",
924 ret = copy_from_user(&desc, vq->desc + i, sizeof desc);
926 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
930 if (desc.flags & VRING_DESC_F_INDIRECT) {
931 ret = get_indirect(dev, vq, iov, iov_size,
933 log, log_num, &desc);
935 vq_err(vq, "Failure detected "
936 "in indirect descriptor at idx %d\n", i);
942 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
943 iov_size - iov_count);
945 vq_err(vq, "Translation failure %d descriptor idx %d\n",
949 if (desc.flags & VRING_DESC_F_WRITE) {
950 /* If this is an input descriptor,
951 * increment that count. */
954 log[*log_num].addr = desc.addr;
955 log[*log_num].len = desc.len;
959 /* If it's an output descriptor, they're all supposed
960 * to come before any input descriptors. */
962 vq_err(vq, "Descriptor has out after in: "
968 } while ((i = next_desc(&desc)) != -1);
970 /* On success, increment avail index. */
971 vq->last_avail_idx++;
975 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
976 void vhost_discard_vq_desc(struct vhost_virtqueue *vq)
978 vq->last_avail_idx--;
981 /* After we've used one of their buffers, we tell them about it. We'll then
982 * want to notify the guest, using eventfd. */
983 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
985 struct vring_used_elem *used;
987 /* The virtqueue contains a ring of used buffers. Get a pointer to the
988 * next entry in that used ring. */
989 used = &vq->used->ring[vq->last_used_idx % vq->num];
990 if (put_user(head, &used->id)) {
991 vq_err(vq, "Failed to write used id");
994 if (put_user(len, &used->len)) {
995 vq_err(vq, "Failed to write used len");
998 /* Make sure buffer is written before we update index. */
1000 if (put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1001 vq_err(vq, "Failed to increment used idx");
1004 if (unlikely(vq->log_used)) {
1005 /* Make sure data is seen before log. */
1007 log_write(vq->log_base, vq->log_addr + sizeof *vq->used->ring *
1008 (vq->last_used_idx % vq->num),
1009 sizeof *vq->used->ring);
1010 log_write(vq->log_base, vq->log_addr, sizeof *vq->used->ring);
1012 eventfd_signal(vq->log_ctx, 1);
1014 vq->last_used_idx++;
1018 /* This actually signals the guest, using eventfd. */
1019 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1022 if (get_user(flags, &vq->avail->flags)) {
1023 vq_err(vq, "Failed to get flags");
1027 /* If they don't want an interrupt, don't signal, unless empty. */
1028 if ((flags & VRING_AVAIL_F_NO_INTERRUPT) &&
1029 (vq->avail_idx != vq->last_avail_idx ||
1030 !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY)))
1033 /* Signal the Guest tell them we used something up. */
1035 eventfd_signal(vq->call_ctx, 1);
1038 /* And here's the combo meal deal. Supersize me! */
1039 void vhost_add_used_and_signal(struct vhost_dev *dev,
1040 struct vhost_virtqueue *vq,
1041 unsigned int head, int len)
1043 vhost_add_used(vq, head, len);
1044 vhost_signal(dev, vq);
1047 /* OK, now we need to know about added descriptors. */
1048 bool vhost_enable_notify(struct vhost_virtqueue *vq)
1052 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1054 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1055 r = put_user(vq->used_flags, &vq->used->flags);
1057 vq_err(vq, "Failed to enable notification at %p: %d\n",
1058 &vq->used->flags, r);
1061 /* They could have slipped one in as we were doing that: make
1062 * sure it's written, then check again. */
1064 r = get_user(avail_idx, &vq->avail->idx);
1066 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1067 &vq->avail->idx, r);
1071 return avail_idx != vq->last_avail_idx;
1074 /* We don't need to be notified again. */
1075 void vhost_disable_notify(struct vhost_virtqueue *vq)
1078 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1080 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1081 r = put_user(vq->used_flags, &vq->used->flags);
1083 vq_err(vq, "Failed to enable notification at %p: %d\n",
1084 &vq->used->flags, r);
1087 int vhost_init(void)
1089 vhost_workqueue = create_singlethread_workqueue("vhost");
1090 if (!vhost_workqueue)
1095 void vhost_cleanup(void)
1097 destroy_workqueue(vhost_workqueue);