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 static void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
65 INIT_LIST_HEAD(&work->node);
67 init_waitqueue_head(&work->done);
69 work->queue_seq = work->done_seq = 0;
72 /* Init poll structure */
73 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
74 unsigned long mask, struct vhost_dev *dev)
76 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
77 init_poll_funcptr(&poll->table, vhost_poll_func);
81 vhost_work_init(&poll->work, fn);
84 /* Start polling a file. We add ourselves to file's wait queue. The caller must
85 * keep a reference to a file until after vhost_poll_stop is called. */
86 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
89 mask = file->f_op->poll(file, &poll->table);
91 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
94 /* Stop polling a file. After this function returns, it becomes safe to drop the
95 * file reference. You must also flush afterwards. */
96 void vhost_poll_stop(struct vhost_poll *poll)
98 remove_wait_queue(poll->wqh, &poll->wait);
101 static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
107 spin_lock_irq(&dev->work_lock);
108 seq = work->queue_seq;
110 spin_unlock_irq(&dev->work_lock);
111 wait_event(work->done, ({
112 spin_lock_irq(&dev->work_lock);
113 left = seq - work->done_seq <= 0;
114 spin_unlock_irq(&dev->work_lock);
117 spin_lock_irq(&dev->work_lock);
118 flushing = --work->flushing;
119 spin_unlock_irq(&dev->work_lock);
120 BUG_ON(flushing < 0);
123 /* Flush any work that has been scheduled. When calling this, don't hold any
124 * locks that are also used by the callback. */
125 void vhost_poll_flush(struct vhost_poll *poll)
127 vhost_work_flush(poll->dev, &poll->work);
130 static inline void vhost_work_queue(struct vhost_dev *dev,
131 struct vhost_work *work)
135 spin_lock_irqsave(&dev->work_lock, flags);
136 if (list_empty(&work->node)) {
137 list_add_tail(&work->node, &dev->work_list);
139 wake_up_process(dev->worker);
141 spin_unlock_irqrestore(&dev->work_lock, flags);
144 void vhost_poll_queue(struct vhost_poll *poll)
146 vhost_work_queue(poll->dev, &poll->work);
149 static void vhost_vq_reset(struct vhost_dev *dev,
150 struct vhost_virtqueue *vq)
156 vq->last_avail_idx = 0;
158 vq->last_used_idx = 0;
161 vq->log_used = false;
162 vq->log_addr = -1ull;
165 vq->private_data = NULL;
167 vq->error_ctx = NULL;
175 static int vhost_worker(void *data)
177 struct vhost_dev *dev = data;
178 struct vhost_work *work = NULL;
179 unsigned uninitialized_var(seq);
182 /* mb paired w/ kthread_stop */
183 set_current_state(TASK_INTERRUPTIBLE);
185 spin_lock_irq(&dev->work_lock);
187 work->done_seq = seq;
189 wake_up_all(&work->done);
192 if (kthread_should_stop()) {
193 spin_unlock_irq(&dev->work_lock);
194 __set_current_state(TASK_RUNNING);
197 if (!list_empty(&dev->work_list)) {
198 work = list_first_entry(&dev->work_list,
199 struct vhost_work, node);
200 list_del_init(&work->node);
201 seq = work->queue_seq;
204 spin_unlock_irq(&dev->work_lock);
207 __set_current_state(TASK_RUNNING);
215 /* Helper to allocate iovec buffers for all vqs. */
216 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
219 for (i = 0; i < dev->nvqs; ++i) {
220 dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect *
221 UIO_MAXIOV, GFP_KERNEL);
222 dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV,
224 dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads *
225 UIO_MAXIOV, GFP_KERNEL);
227 if (!dev->vqs[i].indirect || !dev->vqs[i].log ||
233 for (; i >= 0; --i) {
234 kfree(dev->vqs[i].indirect);
235 kfree(dev->vqs[i].log);
236 kfree(dev->vqs[i].heads);
241 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
244 for (i = 0; i < dev->nvqs; ++i) {
245 kfree(dev->vqs[i].indirect);
246 dev->vqs[i].indirect = NULL;
247 kfree(dev->vqs[i].log);
248 dev->vqs[i].log = NULL;
249 kfree(dev->vqs[i].heads);
250 dev->vqs[i].heads = NULL;
254 long vhost_dev_init(struct vhost_dev *dev,
255 struct vhost_virtqueue *vqs, int nvqs)
261 mutex_init(&dev->mutex);
263 dev->log_file = NULL;
266 spin_lock_init(&dev->work_lock);
267 INIT_LIST_HEAD(&dev->work_list);
270 for (i = 0; i < dev->nvqs; ++i) {
271 dev->vqs[i].log = NULL;
272 dev->vqs[i].indirect = NULL;
273 dev->vqs[i].heads = NULL;
274 dev->vqs[i].dev = dev;
275 mutex_init(&dev->vqs[i].mutex);
276 vhost_vq_reset(dev, dev->vqs + i);
277 if (dev->vqs[i].handle_kick)
278 vhost_poll_init(&dev->vqs[i].poll,
279 dev->vqs[i].handle_kick, POLLIN, dev);
285 /* Caller should have device mutex */
286 long vhost_dev_check_owner(struct vhost_dev *dev)
288 /* Are you the owner? If not, I don't think you mean to do that */
289 return dev->mm == current->mm ? 0 : -EPERM;
292 struct vhost_attach_cgroups_struct {
293 struct vhost_work work;
294 struct task_struct *owner;
298 static void vhost_attach_cgroups_work(struct vhost_work *work)
300 struct vhost_attach_cgroups_struct *s;
301 s = container_of(work, struct vhost_attach_cgroups_struct, work);
302 s->ret = cgroup_attach_task_all(s->owner, current);
305 static int vhost_attach_cgroups(struct vhost_dev *dev)
307 struct vhost_attach_cgroups_struct attach;
308 attach.owner = current;
309 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
310 vhost_work_queue(dev, &attach.work);
311 vhost_work_flush(dev, &attach.work);
315 /* Caller should have device mutex */
316 static long vhost_dev_set_owner(struct vhost_dev *dev)
318 struct task_struct *worker;
320 /* Is there an owner already? */
325 /* No owner, become one */
326 dev->mm = get_task_mm(current);
327 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
328 if (IS_ERR(worker)) {
329 err = PTR_ERR(worker);
333 dev->worker = worker;
334 wake_up_process(worker); /* avoid contributing to loadavg */
336 err = vhost_attach_cgroups(dev);
340 err = vhost_dev_alloc_iovecs(dev);
346 kthread_stop(worker);
356 /* Caller should have device mutex */
357 long vhost_dev_reset_owner(struct vhost_dev *dev)
359 struct vhost_memory *memory;
361 /* Restore memory to default empty mapping. */
362 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
366 vhost_dev_cleanup(dev);
368 memory->nregions = 0;
369 dev->memory = memory;
373 /* Caller should have device mutex */
374 void vhost_dev_cleanup(struct vhost_dev *dev)
377 for (i = 0; i < dev->nvqs; ++i) {
378 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
379 vhost_poll_stop(&dev->vqs[i].poll);
380 vhost_poll_flush(&dev->vqs[i].poll);
382 if (dev->vqs[i].error_ctx)
383 eventfd_ctx_put(dev->vqs[i].error_ctx);
384 if (dev->vqs[i].error)
385 fput(dev->vqs[i].error);
386 if (dev->vqs[i].kick)
387 fput(dev->vqs[i].kick);
388 if (dev->vqs[i].call_ctx)
389 eventfd_ctx_put(dev->vqs[i].call_ctx);
390 if (dev->vqs[i].call)
391 fput(dev->vqs[i].call);
392 vhost_vq_reset(dev, dev->vqs + i);
394 vhost_dev_free_iovecs(dev);
396 eventfd_ctx_put(dev->log_ctx);
400 dev->log_file = NULL;
401 /* No one will access memory at this point */
408 WARN_ON(!list_empty(&dev->work_list));
410 kthread_stop(dev->worker);
415 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
417 u64 a = addr / VHOST_PAGE_SIZE / 8;
418 /* Make sure 64 bit math will not overflow. */
419 if (a > ULONG_MAX - (unsigned long)log_base ||
420 a + (unsigned long)log_base > ULONG_MAX)
423 return access_ok(VERIFY_WRITE, log_base + a,
424 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
427 /* Caller should have vq mutex and device mutex. */
428 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
436 for (i = 0; i < mem->nregions; ++i) {
437 struct vhost_memory_region *m = mem->regions + i;
438 unsigned long a = m->userspace_addr;
439 if (m->memory_size > ULONG_MAX)
441 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
444 else if (log_all && !log_access_ok(log_base,
452 /* Can we switch to this memory table? */
453 /* Caller should have device mutex but not vq mutex */
454 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
458 for (i = 0; i < d->nvqs; ++i) {
460 mutex_lock(&d->vqs[i].mutex);
461 /* If ring is inactive, will check when it's enabled. */
462 if (d->vqs[i].private_data)
463 ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
467 mutex_unlock(&d->vqs[i].mutex);
474 static int vq_access_ok(unsigned int num,
475 struct vring_desc __user *desc,
476 struct vring_avail __user *avail,
477 struct vring_used __user *used)
479 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
480 access_ok(VERIFY_READ, avail,
481 sizeof *avail + num * sizeof *avail->ring) &&
482 access_ok(VERIFY_WRITE, used,
483 sizeof *used + num * sizeof *used->ring);
486 /* Can we log writes? */
487 /* Caller should have device mutex but not vq mutex */
488 int vhost_log_access_ok(struct vhost_dev *dev)
490 return memory_access_ok(dev, dev->memory, 1);
493 /* Verify access for write logging. */
494 /* Caller should have vq mutex and device mutex */
495 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base)
497 return vq_memory_access_ok(log_base, vq->dev->memory,
498 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
499 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
501 vq->num * sizeof *vq->used->ring));
504 /* Can we start vq? */
505 /* Caller should have vq mutex and device mutex */
506 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
508 return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) &&
509 vq_log_access_ok(vq, vq->log_base);
512 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
514 struct vhost_memory mem, *newmem, *oldmem;
515 unsigned long size = offsetof(struct vhost_memory, regions);
516 if (copy_from_user(&mem, m, size))
520 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
522 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
526 memcpy(newmem, &mem, size);
527 if (copy_from_user(newmem->regions, m->regions,
528 mem.nregions * sizeof *m->regions)) {
533 if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL))) {
538 rcu_assign_pointer(d->memory, newmem);
544 static int init_used(struct vhost_virtqueue *vq,
545 struct vring_used __user *used)
547 int r = put_user(vq->used_flags, &used->flags);
550 return get_user(vq->last_used_idx, &used->idx);
553 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
555 struct file *eventfp, *filep = NULL,
556 *pollstart = NULL, *pollstop = NULL;
557 struct eventfd_ctx *ctx = NULL;
558 u32 __user *idxp = argp;
559 struct vhost_virtqueue *vq;
560 struct vhost_vring_state s;
561 struct vhost_vring_file f;
562 struct vhost_vring_addr a;
566 r = get_user(idx, idxp);
574 mutex_lock(&vq->mutex);
577 case VHOST_SET_VRING_NUM:
578 /* Resizing ring with an active backend?
579 * You don't want to do that. */
580 if (vq->private_data) {
584 if (copy_from_user(&s, argp, sizeof s)) {
588 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
594 case VHOST_SET_VRING_BASE:
595 /* Moving base with an active backend?
596 * You don't want to do that. */
597 if (vq->private_data) {
601 if (copy_from_user(&s, argp, sizeof s)) {
605 if (s.num > 0xffff) {
609 vq->last_avail_idx = s.num;
610 /* Forget the cached index value. */
611 vq->avail_idx = vq->last_avail_idx;
613 case VHOST_GET_VRING_BASE:
615 s.num = vq->last_avail_idx;
616 if (copy_to_user(argp, &s, sizeof s))
619 case VHOST_SET_VRING_ADDR:
620 if (copy_from_user(&a, argp, sizeof a)) {
624 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
628 /* For 32bit, verify that the top 32bits of the user
629 data are set to zero. */
630 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
631 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
632 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
636 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
637 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
638 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
643 /* We only verify access here if backend is configured.
644 * If it is not, we don't as size might not have been setup.
645 * We will verify when backend is configured. */
646 if (vq->private_data) {
647 if (!vq_access_ok(vq->num,
648 (void __user *)(unsigned long)a.desc_user_addr,
649 (void __user *)(unsigned long)a.avail_user_addr,
650 (void __user *)(unsigned long)a.used_user_addr)) {
655 /* Also validate log access for used ring if enabled. */
656 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
657 !log_access_ok(vq->log_base, a.log_guest_addr,
659 vq->num * sizeof *vq->used->ring)) {
665 r = init_used(vq, (struct vring_used __user *)(unsigned long)
669 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
670 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
671 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
672 vq->log_addr = a.log_guest_addr;
673 vq->used = (void __user *)(unsigned long)a.used_user_addr;
675 case VHOST_SET_VRING_KICK:
676 if (copy_from_user(&f, argp, sizeof f)) {
680 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
681 if (IS_ERR(eventfp)) {
682 r = PTR_ERR(eventfp);
685 if (eventfp != vq->kick) {
686 pollstop = filep = vq->kick;
687 pollstart = vq->kick = eventfp;
691 case VHOST_SET_VRING_CALL:
692 if (copy_from_user(&f, argp, sizeof f)) {
696 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
697 if (IS_ERR(eventfp)) {
698 r = PTR_ERR(eventfp);
701 if (eventfp != vq->call) {
705 vq->call_ctx = eventfp ?
706 eventfd_ctx_fileget(eventfp) : NULL;
710 case VHOST_SET_VRING_ERR:
711 if (copy_from_user(&f, argp, sizeof f)) {
715 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
716 if (IS_ERR(eventfp)) {
717 r = PTR_ERR(eventfp);
720 if (eventfp != vq->error) {
724 vq->error_ctx = eventfp ?
725 eventfd_ctx_fileget(eventfp) : NULL;
733 if (pollstop && vq->handle_kick)
734 vhost_poll_stop(&vq->poll);
737 eventfd_ctx_put(ctx);
741 if (pollstart && vq->handle_kick)
742 vhost_poll_start(&vq->poll, vq->kick);
744 mutex_unlock(&vq->mutex);
746 if (pollstop && vq->handle_kick)
747 vhost_poll_flush(&vq->poll);
751 /* Caller must have device mutex */
752 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
754 void __user *argp = (void __user *)arg;
755 struct file *eventfp, *filep = NULL;
756 struct eventfd_ctx *ctx = NULL;
761 /* If you are not the owner, you can become one */
762 if (ioctl == VHOST_SET_OWNER) {
763 r = vhost_dev_set_owner(d);
767 /* You must be the owner to do anything else */
768 r = vhost_dev_check_owner(d);
773 case VHOST_SET_MEM_TABLE:
774 r = vhost_set_memory(d, argp);
776 case VHOST_SET_LOG_BASE:
777 if (copy_from_user(&p, argp, sizeof p)) {
781 if ((u64)(unsigned long)p != p) {
785 for (i = 0; i < d->nvqs; ++i) {
786 struct vhost_virtqueue *vq;
787 void __user *base = (void __user *)(unsigned long)p;
789 mutex_lock(&vq->mutex);
790 /* If ring is inactive, will check when it's enabled. */
791 if (vq->private_data && !vq_log_access_ok(vq, base))
795 mutex_unlock(&vq->mutex);
798 case VHOST_SET_LOG_FD:
799 r = get_user(fd, (int __user *)argp);
802 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
803 if (IS_ERR(eventfp)) {
804 r = PTR_ERR(eventfp);
807 if (eventfp != d->log_file) {
810 d->log_ctx = eventfp ?
811 eventfd_ctx_fileget(eventfp) : NULL;
814 for (i = 0; i < d->nvqs; ++i) {
815 mutex_lock(&d->vqs[i].mutex);
816 d->vqs[i].log_ctx = d->log_ctx;
817 mutex_unlock(&d->vqs[i].mutex);
820 eventfd_ctx_put(ctx);
825 r = vhost_set_vring(d, ioctl, argp);
832 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
833 __u64 addr, __u32 len)
835 struct vhost_memory_region *reg;
837 /* linear search is not brilliant, but we really have on the order of 6
838 * regions in practice */
839 for (i = 0; i < mem->nregions; ++i) {
840 reg = mem->regions + i;
841 if (reg->guest_phys_addr <= addr &&
842 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
848 /* TODO: This is really inefficient. We need something like get_user()
849 * (instruction directly accesses the data, with an exception table entry
850 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
852 static int set_bit_to_user(int nr, void __user *addr)
854 unsigned long log = (unsigned long)addr;
857 int bit = nr + (log % PAGE_SIZE) * 8;
859 r = get_user_pages_fast(log, 1, 1, &page);
863 base = kmap_atomic(page, KM_USER0);
865 kunmap_atomic(base, KM_USER0);
866 set_page_dirty_lock(page);
871 static int log_write(void __user *log_base,
872 u64 write_address, u64 write_length)
877 write_address /= VHOST_PAGE_SIZE;
879 u64 base = (u64)(unsigned long)log_base;
880 u64 log = base + write_address / 8;
881 int bit = write_address % 8;
882 if ((u64)(unsigned long)log != log)
884 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
887 if (write_length <= VHOST_PAGE_SIZE)
889 write_length -= VHOST_PAGE_SIZE;
890 write_address += VHOST_PAGE_SIZE;
895 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
896 unsigned int log_num, u64 len)
900 /* Make sure data written is seen before log. */
902 for (i = 0; i < log_num; ++i) {
903 u64 l = min(log[i].len, len);
904 r = log_write(vq->log_base, log[i].addr, l);
910 eventfd_signal(vq->log_ctx, 1);
914 /* Length written exceeds what we have stored. This is a bug. */
919 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
920 struct iovec iov[], int iov_size)
922 const struct vhost_memory_region *reg;
923 struct vhost_memory *mem;
930 mem = rcu_dereference(dev->memory);
931 while ((u64)len > s) {
933 if (unlikely(ret >= iov_size)) {
937 reg = find_region(mem, addr, len);
938 if (unlikely(!reg)) {
943 size = reg->memory_size - addr + reg->guest_phys_addr;
944 _iov->iov_len = min((u64)len, size);
945 _iov->iov_base = (void __user *)(unsigned long)
946 (reg->userspace_addr + addr - reg->guest_phys_addr);
956 /* Each buffer in the virtqueues is actually a chain of descriptors. This
957 * function returns the next descriptor in the chain,
958 * or -1U if we're at the end. */
959 static unsigned next_desc(struct vring_desc *desc)
963 /* If this descriptor says it doesn't chain, we're done. */
964 if (!(desc->flags & VRING_DESC_F_NEXT))
967 /* Check they're not leading us off end of descriptors. */
969 /* Make sure compiler knows to grab that: we don't want it changing! */
970 /* We will use the result as an index in an array, so most
971 * architectures only need a compiler barrier here. */
972 read_barrier_depends();
977 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
978 struct iovec iov[], unsigned int iov_size,
979 unsigned int *out_num, unsigned int *in_num,
980 struct vhost_log *log, unsigned int *log_num,
981 struct vring_desc *indirect)
983 struct vring_desc desc;
984 unsigned int i = 0, count, found = 0;
988 if (unlikely(indirect->len % sizeof desc)) {
989 vq_err(vq, "Invalid length in indirect descriptor: "
990 "len 0x%llx not multiple of 0x%zx\n",
991 (unsigned long long)indirect->len,
996 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
998 if (unlikely(ret < 0)) {
999 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1003 /* We will use the result as an address to read from, so most
1004 * architectures only need a compiler barrier here. */
1005 read_barrier_depends();
1007 count = indirect->len / sizeof desc;
1008 /* Buffers are chained via a 16 bit next field, so
1009 * we can have at most 2^16 of these. */
1010 if (unlikely(count > USHRT_MAX + 1)) {
1011 vq_err(vq, "Indirect buffer length too big: %d\n",
1017 unsigned iov_count = *in_num + *out_num;
1018 if (unlikely(++found > count)) {
1019 vq_err(vq, "Loop detected: last one at %u "
1020 "indirect size %u\n",
1024 if (unlikely(memcpy_fromiovec((unsigned char *)&desc, vq->indirect,
1026 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1027 i, (size_t)indirect->addr + i * sizeof desc);
1030 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1031 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1032 i, (size_t)indirect->addr + i * sizeof desc);
1036 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1037 iov_size - iov_count);
1038 if (unlikely(ret < 0)) {
1039 vq_err(vq, "Translation failure %d indirect idx %d\n",
1043 /* If this is an input descriptor, increment that count. */
1044 if (desc.flags & VRING_DESC_F_WRITE) {
1046 if (unlikely(log)) {
1047 log[*log_num].addr = desc.addr;
1048 log[*log_num].len = desc.len;
1052 /* If it's an output descriptor, they're all supposed
1053 * to come before any input descriptors. */
1054 if (unlikely(*in_num)) {
1055 vq_err(vq, "Indirect descriptor "
1056 "has out after in: idx %d\n", i);
1061 } while ((i = next_desc(&desc)) != -1);
1065 /* This looks in the virtqueue and for the first available buffer, and converts
1066 * it to an iovec for convenient access. Since descriptors consist of some
1067 * number of output then some number of input descriptors, it's actually two
1068 * iovecs, but we pack them into one and note how many of each there were.
1070 * This function returns the descriptor number found, or vq->num (which is
1071 * never a valid descriptor number) if none was found. A negative code is
1072 * returned on error. */
1073 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1074 struct iovec iov[], unsigned int iov_size,
1075 unsigned int *out_num, unsigned int *in_num,
1076 struct vhost_log *log, unsigned int *log_num)
1078 struct vring_desc desc;
1079 unsigned int i, head, found = 0;
1083 /* Check it isn't doing very strange things with descriptor numbers. */
1084 last_avail_idx = vq->last_avail_idx;
1085 if (unlikely(get_user(vq->avail_idx, &vq->avail->idx))) {
1086 vq_err(vq, "Failed to access avail idx at %p\n",
1091 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1092 vq_err(vq, "Guest moved used index from %u to %u",
1093 last_avail_idx, vq->avail_idx);
1097 /* If there's nothing new since last we looked, return invalid. */
1098 if (vq->avail_idx == last_avail_idx)
1101 /* Only get avail ring entries after they have been exposed by guest. */
1104 /* Grab the next descriptor number they're advertising, and increment
1105 * the index we've seen. */
1106 if (unlikely(get_user(head,
1107 &vq->avail->ring[last_avail_idx % vq->num]))) {
1108 vq_err(vq, "Failed to read head: idx %d address %p\n",
1110 &vq->avail->ring[last_avail_idx % vq->num]);
1114 /* If their number is silly, that's an error. */
1115 if (unlikely(head >= vq->num)) {
1116 vq_err(vq, "Guest says index %u > %u is available",
1121 /* When we start there are none of either input nor output. */
1122 *out_num = *in_num = 0;
1128 unsigned iov_count = *in_num + *out_num;
1129 if (unlikely(i >= vq->num)) {
1130 vq_err(vq, "Desc index is %u > %u, head = %u",
1134 if (unlikely(++found > vq->num)) {
1135 vq_err(vq, "Loop detected: last one at %u "
1136 "vq size %u head %u\n",
1140 ret = copy_from_user(&desc, vq->desc + i, sizeof desc);
1141 if (unlikely(ret)) {
1142 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1146 if (desc.flags & VRING_DESC_F_INDIRECT) {
1147 ret = get_indirect(dev, vq, iov, iov_size,
1149 log, log_num, &desc);
1150 if (unlikely(ret < 0)) {
1151 vq_err(vq, "Failure detected "
1152 "in indirect descriptor at idx %d\n", i);
1158 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1159 iov_size - iov_count);
1160 if (unlikely(ret < 0)) {
1161 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1165 if (desc.flags & VRING_DESC_F_WRITE) {
1166 /* If this is an input descriptor,
1167 * increment that count. */
1169 if (unlikely(log)) {
1170 log[*log_num].addr = desc.addr;
1171 log[*log_num].len = desc.len;
1175 /* If it's an output descriptor, they're all supposed
1176 * to come before any input descriptors. */
1177 if (unlikely(*in_num)) {
1178 vq_err(vq, "Descriptor has out after in: "
1184 } while ((i = next_desc(&desc)) != -1);
1186 /* On success, increment avail index. */
1187 vq->last_avail_idx++;
1191 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1192 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1194 vq->last_avail_idx -= n;
1197 /* After we've used one of their buffers, we tell them about it. We'll then
1198 * want to notify the guest, using eventfd. */
1199 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1201 struct vring_used_elem __user *used;
1203 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1204 * next entry in that used ring. */
1205 used = &vq->used->ring[vq->last_used_idx % vq->num];
1206 if (put_user(head, &used->id)) {
1207 vq_err(vq, "Failed to write used id");
1210 if (put_user(len, &used->len)) {
1211 vq_err(vq, "Failed to write used len");
1214 /* Make sure buffer is written before we update index. */
1216 if (put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1217 vq_err(vq, "Failed to increment used idx");
1220 if (unlikely(vq->log_used)) {
1221 /* Make sure data is seen before log. */
1223 /* Log used ring entry write. */
1224 log_write(vq->log_base,
1226 ((void __user *)used - (void __user *)vq->used),
1228 /* Log used index update. */
1229 log_write(vq->log_base,
1230 vq->log_addr + offsetof(struct vring_used, idx),
1231 sizeof vq->used->idx);
1233 eventfd_signal(vq->log_ctx, 1);
1235 vq->last_used_idx++;
1239 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1240 struct vring_used_elem *heads,
1243 struct vring_used_elem __user *used;
1246 start = vq->last_used_idx % vq->num;
1247 used = vq->used->ring + start;
1248 if (copy_to_user(used, heads, count * sizeof *used)) {
1249 vq_err(vq, "Failed to write used");
1252 if (unlikely(vq->log_used)) {
1253 /* Make sure data is seen before log. */
1255 /* Log used ring entry write. */
1256 log_write(vq->log_base,
1258 ((void __user *)used - (void __user *)vq->used),
1259 count * sizeof *used);
1261 vq->last_used_idx += count;
1265 /* After we've used one of their buffers, we tell them about it. We'll then
1266 * want to notify the guest, using eventfd. */
1267 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1272 start = vq->last_used_idx % vq->num;
1273 n = vq->num - start;
1275 r = __vhost_add_used_n(vq, heads, n);
1281 r = __vhost_add_used_n(vq, heads, count);
1283 /* Make sure buffer is written before we update index. */
1285 if (put_user(vq->last_used_idx, &vq->used->idx)) {
1286 vq_err(vq, "Failed to increment used idx");
1289 if (unlikely(vq->log_used)) {
1290 /* Log used index update. */
1291 log_write(vq->log_base,
1292 vq->log_addr + offsetof(struct vring_used, idx),
1293 sizeof vq->used->idx);
1295 eventfd_signal(vq->log_ctx, 1);
1300 /* This actually signals the guest, using eventfd. */
1301 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1304 /* Flush out used index updates. This is paired
1305 * with the barrier that the Guest executes when enabling
1309 if (get_user(flags, &vq->avail->flags)) {
1310 vq_err(vq, "Failed to get flags");
1314 /* If they don't want an interrupt, don't signal, unless empty. */
1315 if ((flags & VRING_AVAIL_F_NO_INTERRUPT) &&
1316 (vq->avail_idx != vq->last_avail_idx ||
1317 !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY)))
1320 /* Signal the Guest tell them we used something up. */
1322 eventfd_signal(vq->call_ctx, 1);
1325 /* And here's the combo meal deal. Supersize me! */
1326 void vhost_add_used_and_signal(struct vhost_dev *dev,
1327 struct vhost_virtqueue *vq,
1328 unsigned int head, int len)
1330 vhost_add_used(vq, head, len);
1331 vhost_signal(dev, vq);
1334 /* multi-buffer version of vhost_add_used_and_signal */
1335 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1336 struct vhost_virtqueue *vq,
1337 struct vring_used_elem *heads, unsigned count)
1339 vhost_add_used_n(vq, heads, count);
1340 vhost_signal(dev, vq);
1343 /* OK, now we need to know about added descriptors. */
1344 bool vhost_enable_notify(struct vhost_virtqueue *vq)
1348 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1350 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1351 r = put_user(vq->used_flags, &vq->used->flags);
1353 vq_err(vq, "Failed to enable notification at %p: %d\n",
1354 &vq->used->flags, r);
1357 /* They could have slipped one in as we were doing that: make
1358 * sure it's written, then check again. */
1360 r = get_user(avail_idx, &vq->avail->idx);
1362 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1363 &vq->avail->idx, r);
1367 return avail_idx != vq->avail_idx;
1370 /* We don't need to be notified again. */
1371 void vhost_disable_notify(struct vhost_virtqueue *vq)
1374 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1376 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1377 r = put_user(vq->used_flags, &vq->used->flags);
1379 vq_err(vq, "Failed to enable notification at %p: %d\n",
1380 &vq->used->flags, r);