Loading...
1/* Copyright (C) 2009 Red Hat, Inc.
2 * Copyright (C) 2006 Rusty Russell IBM Corporation
3 *
4 * Author: Michael S. Tsirkin <mst@redhat.com>
5 *
6 * Inspiration, some code, and most witty comments come from
7 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
8 *
9 * This work is licensed under the terms of the GNU GPL, version 2.
10 *
11 * Generic code for virtio server in host kernel.
12 */
13
14#include <linux/eventfd.h>
15#include <linux/vhost.h>
16#include <linux/virtio_net.h>
17#include <linux/mm.h>
18#include <linux/mmu_context.h>
19#include <linux/miscdevice.h>
20#include <linux/mutex.h>
21#include <linux/rcupdate.h>
22#include <linux/poll.h>
23#include <linux/file.h>
24#include <linux/highmem.h>
25#include <linux/slab.h>
26#include <linux/kthread.h>
27#include <linux/cgroup.h>
28
29#include <linux/net.h>
30#include <linux/if_packet.h>
31#include <linux/if_arp.h>
32
33#include "vhost.h"
34
35enum {
36 VHOST_MEMORY_MAX_NREGIONS = 64,
37 VHOST_MEMORY_F_LOG = 0x1,
38};
39
40static unsigned vhost_zcopy_mask __read_mostly;
41
42#define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num])
43#define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num])
44
45static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
46 poll_table *pt)
47{
48 struct vhost_poll *poll;
49
50 poll = container_of(pt, struct vhost_poll, table);
51 poll->wqh = wqh;
52 add_wait_queue(wqh, &poll->wait);
53}
54
55static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
56 void *key)
57{
58 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
59
60 if (!((unsigned long)key & poll->mask))
61 return 0;
62
63 vhost_poll_queue(poll);
64 return 0;
65}
66
67static void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
68{
69 INIT_LIST_HEAD(&work->node);
70 work->fn = fn;
71 init_waitqueue_head(&work->done);
72 work->flushing = 0;
73 work->queue_seq = work->done_seq = 0;
74}
75
76/* Init poll structure */
77void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
78 unsigned long mask, struct vhost_dev *dev)
79{
80 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
81 init_poll_funcptr(&poll->table, vhost_poll_func);
82 poll->mask = mask;
83 poll->dev = dev;
84
85 vhost_work_init(&poll->work, fn);
86}
87
88/* Start polling a file. We add ourselves to file's wait queue. The caller must
89 * keep a reference to a file until after vhost_poll_stop is called. */
90void vhost_poll_start(struct vhost_poll *poll, struct file *file)
91{
92 unsigned long mask;
93
94 mask = file->f_op->poll(file, &poll->table);
95 if (mask)
96 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
97}
98
99/* Stop polling a file. After this function returns, it becomes safe to drop the
100 * file reference. You must also flush afterwards. */
101void vhost_poll_stop(struct vhost_poll *poll)
102{
103 remove_wait_queue(poll->wqh, &poll->wait);
104}
105
106static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
107 unsigned seq)
108{
109 int left;
110
111 spin_lock_irq(&dev->work_lock);
112 left = seq - work->done_seq;
113 spin_unlock_irq(&dev->work_lock);
114 return left <= 0;
115}
116
117static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
118{
119 unsigned seq;
120 int flushing;
121
122 spin_lock_irq(&dev->work_lock);
123 seq = work->queue_seq;
124 work->flushing++;
125 spin_unlock_irq(&dev->work_lock);
126 wait_event(work->done, vhost_work_seq_done(dev, work, seq));
127 spin_lock_irq(&dev->work_lock);
128 flushing = --work->flushing;
129 spin_unlock_irq(&dev->work_lock);
130 BUG_ON(flushing < 0);
131}
132
133/* Flush any work that has been scheduled. When calling this, don't hold any
134 * locks that are also used by the callback. */
135void vhost_poll_flush(struct vhost_poll *poll)
136{
137 vhost_work_flush(poll->dev, &poll->work);
138}
139
140static inline void vhost_work_queue(struct vhost_dev *dev,
141 struct vhost_work *work)
142{
143 unsigned long flags;
144
145 spin_lock_irqsave(&dev->work_lock, flags);
146 if (list_empty(&work->node)) {
147 list_add_tail(&work->node, &dev->work_list);
148 work->queue_seq++;
149 wake_up_process(dev->worker);
150 }
151 spin_unlock_irqrestore(&dev->work_lock, flags);
152}
153
154void vhost_poll_queue(struct vhost_poll *poll)
155{
156 vhost_work_queue(poll->dev, &poll->work);
157}
158
159static void vhost_vq_reset(struct vhost_dev *dev,
160 struct vhost_virtqueue *vq)
161{
162 vq->num = 1;
163 vq->desc = NULL;
164 vq->avail = NULL;
165 vq->used = NULL;
166 vq->last_avail_idx = 0;
167 vq->avail_idx = 0;
168 vq->last_used_idx = 0;
169 vq->signalled_used = 0;
170 vq->signalled_used_valid = false;
171 vq->used_flags = 0;
172 vq->log_used = false;
173 vq->log_addr = -1ull;
174 vq->vhost_hlen = 0;
175 vq->sock_hlen = 0;
176 vq->private_data = NULL;
177 vq->log_base = NULL;
178 vq->error_ctx = NULL;
179 vq->error = NULL;
180 vq->kick = NULL;
181 vq->call_ctx = NULL;
182 vq->call = NULL;
183 vq->log_ctx = NULL;
184 vq->upend_idx = 0;
185 vq->done_idx = 0;
186 vq->ubufs = NULL;
187}
188
189static int vhost_worker(void *data)
190{
191 struct vhost_dev *dev = data;
192 struct vhost_work *work = NULL;
193 unsigned uninitialized_var(seq);
194
195 use_mm(dev->mm);
196
197 for (;;) {
198 /* mb paired w/ kthread_stop */
199 set_current_state(TASK_INTERRUPTIBLE);
200
201 spin_lock_irq(&dev->work_lock);
202 if (work) {
203 work->done_seq = seq;
204 if (work->flushing)
205 wake_up_all(&work->done);
206 }
207
208 if (kthread_should_stop()) {
209 spin_unlock_irq(&dev->work_lock);
210 __set_current_state(TASK_RUNNING);
211 break;
212 }
213 if (!list_empty(&dev->work_list)) {
214 work = list_first_entry(&dev->work_list,
215 struct vhost_work, node);
216 list_del_init(&work->node);
217 seq = work->queue_seq;
218 } else
219 work = NULL;
220 spin_unlock_irq(&dev->work_lock);
221
222 if (work) {
223 __set_current_state(TASK_RUNNING);
224 work->fn(work);
225 } else
226 schedule();
227
228 }
229 unuse_mm(dev->mm);
230 return 0;
231}
232
233static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
234{
235 kfree(vq->indirect);
236 vq->indirect = NULL;
237 kfree(vq->log);
238 vq->log = NULL;
239 kfree(vq->heads);
240 vq->heads = NULL;
241 kfree(vq->ubuf_info);
242 vq->ubuf_info = NULL;
243}
244
245void vhost_enable_zcopy(int vq)
246{
247 vhost_zcopy_mask |= 0x1 << vq;
248}
249
250/* Helper to allocate iovec buffers for all vqs. */
251static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
252{
253 int i;
254 bool zcopy;
255
256 for (i = 0; i < dev->nvqs; ++i) {
257 dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect *
258 UIO_MAXIOV, GFP_KERNEL);
259 dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV,
260 GFP_KERNEL);
261 dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads *
262 UIO_MAXIOV, GFP_KERNEL);
263 zcopy = vhost_zcopy_mask & (0x1 << i);
264 if (zcopy)
265 dev->vqs[i].ubuf_info =
266 kmalloc(sizeof *dev->vqs[i].ubuf_info *
267 UIO_MAXIOV, GFP_KERNEL);
268 if (!dev->vqs[i].indirect || !dev->vqs[i].log ||
269 !dev->vqs[i].heads ||
270 (zcopy && !dev->vqs[i].ubuf_info))
271 goto err_nomem;
272 }
273 return 0;
274
275err_nomem:
276 for (; i >= 0; --i)
277 vhost_vq_free_iovecs(&dev->vqs[i]);
278 return -ENOMEM;
279}
280
281static void vhost_dev_free_iovecs(struct vhost_dev *dev)
282{
283 int i;
284
285 for (i = 0; i < dev->nvqs; ++i)
286 vhost_vq_free_iovecs(&dev->vqs[i]);
287}
288
289long vhost_dev_init(struct vhost_dev *dev,
290 struct vhost_virtqueue *vqs, int nvqs)
291{
292 int i;
293
294 dev->vqs = vqs;
295 dev->nvqs = nvqs;
296 mutex_init(&dev->mutex);
297 dev->log_ctx = NULL;
298 dev->log_file = NULL;
299 dev->memory = NULL;
300 dev->mm = NULL;
301 spin_lock_init(&dev->work_lock);
302 INIT_LIST_HEAD(&dev->work_list);
303 dev->worker = NULL;
304
305 for (i = 0; i < dev->nvqs; ++i) {
306 dev->vqs[i].log = NULL;
307 dev->vqs[i].indirect = NULL;
308 dev->vqs[i].heads = NULL;
309 dev->vqs[i].ubuf_info = NULL;
310 dev->vqs[i].dev = dev;
311 mutex_init(&dev->vqs[i].mutex);
312 vhost_vq_reset(dev, dev->vqs + i);
313 if (dev->vqs[i].handle_kick)
314 vhost_poll_init(&dev->vqs[i].poll,
315 dev->vqs[i].handle_kick, POLLIN, dev);
316 }
317
318 return 0;
319}
320
321/* Caller should have device mutex */
322long vhost_dev_check_owner(struct vhost_dev *dev)
323{
324 /* Are you the owner? If not, I don't think you mean to do that */
325 return dev->mm == current->mm ? 0 : -EPERM;
326}
327
328struct vhost_attach_cgroups_struct {
329 struct vhost_work work;
330 struct task_struct *owner;
331 int ret;
332};
333
334static void vhost_attach_cgroups_work(struct vhost_work *work)
335{
336 struct vhost_attach_cgroups_struct *s;
337
338 s = container_of(work, struct vhost_attach_cgroups_struct, work);
339 s->ret = cgroup_attach_task_all(s->owner, current);
340}
341
342static int vhost_attach_cgroups(struct vhost_dev *dev)
343{
344 struct vhost_attach_cgroups_struct attach;
345
346 attach.owner = current;
347 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
348 vhost_work_queue(dev, &attach.work);
349 vhost_work_flush(dev, &attach.work);
350 return attach.ret;
351}
352
353/* Caller should have device mutex */
354static long vhost_dev_set_owner(struct vhost_dev *dev)
355{
356 struct task_struct *worker;
357 int err;
358
359 /* Is there an owner already? */
360 if (dev->mm) {
361 err = -EBUSY;
362 goto err_mm;
363 }
364
365 /* No owner, become one */
366 dev->mm = get_task_mm(current);
367 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
368 if (IS_ERR(worker)) {
369 err = PTR_ERR(worker);
370 goto err_worker;
371 }
372
373 dev->worker = worker;
374 wake_up_process(worker); /* avoid contributing to loadavg */
375
376 err = vhost_attach_cgroups(dev);
377 if (err)
378 goto err_cgroup;
379
380 err = vhost_dev_alloc_iovecs(dev);
381 if (err)
382 goto err_cgroup;
383
384 return 0;
385err_cgroup:
386 kthread_stop(worker);
387 dev->worker = NULL;
388err_worker:
389 if (dev->mm)
390 mmput(dev->mm);
391 dev->mm = NULL;
392err_mm:
393 return err;
394}
395
396/* Caller should have device mutex */
397long vhost_dev_reset_owner(struct vhost_dev *dev)
398{
399 struct vhost_memory *memory;
400
401 /* Restore memory to default empty mapping. */
402 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
403 if (!memory)
404 return -ENOMEM;
405
406 vhost_dev_cleanup(dev);
407
408 memory->nregions = 0;
409 RCU_INIT_POINTER(dev->memory, memory);
410 return 0;
411}
412
413/* In case of DMA done not in order in lower device driver for some reason.
414 * upend_idx is used to track end of used idx, done_idx is used to track head
415 * of used idx. Once lower device DMA done contiguously, we will signal KVM
416 * guest used idx.
417 */
418int vhost_zerocopy_signal_used(struct vhost_virtqueue *vq)
419{
420 int i;
421 int j = 0;
422
423 for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
424 if ((vq->heads[i].len == VHOST_DMA_DONE_LEN)) {
425 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
426 vhost_add_used_and_signal(vq->dev, vq,
427 vq->heads[i].id, 0);
428 ++j;
429 } else
430 break;
431 }
432 if (j)
433 vq->done_idx = i;
434 return j;
435}
436
437/* Caller should have device mutex */
438void vhost_dev_cleanup(struct vhost_dev *dev)
439{
440 int i;
441
442 for (i = 0; i < dev->nvqs; ++i) {
443 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
444 vhost_poll_stop(&dev->vqs[i].poll);
445 vhost_poll_flush(&dev->vqs[i].poll);
446 }
447 /* Wait for all lower device DMAs done. */
448 if (dev->vqs[i].ubufs)
449 vhost_ubuf_put_and_wait(dev->vqs[i].ubufs);
450
451 /* Signal guest as appropriate. */
452 vhost_zerocopy_signal_used(&dev->vqs[i]);
453
454 if (dev->vqs[i].error_ctx)
455 eventfd_ctx_put(dev->vqs[i].error_ctx);
456 if (dev->vqs[i].error)
457 fput(dev->vqs[i].error);
458 if (dev->vqs[i].kick)
459 fput(dev->vqs[i].kick);
460 if (dev->vqs[i].call_ctx)
461 eventfd_ctx_put(dev->vqs[i].call_ctx);
462 if (dev->vqs[i].call)
463 fput(dev->vqs[i].call);
464 vhost_vq_reset(dev, dev->vqs + i);
465 }
466 vhost_dev_free_iovecs(dev);
467 if (dev->log_ctx)
468 eventfd_ctx_put(dev->log_ctx);
469 dev->log_ctx = NULL;
470 if (dev->log_file)
471 fput(dev->log_file);
472 dev->log_file = NULL;
473 /* No one will access memory at this point */
474 kfree(rcu_dereference_protected(dev->memory,
475 lockdep_is_held(&dev->mutex)));
476 RCU_INIT_POINTER(dev->memory, NULL);
477 WARN_ON(!list_empty(&dev->work_list));
478 if (dev->worker) {
479 kthread_stop(dev->worker);
480 dev->worker = NULL;
481 }
482 if (dev->mm)
483 mmput(dev->mm);
484 dev->mm = NULL;
485}
486
487static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
488{
489 u64 a = addr / VHOST_PAGE_SIZE / 8;
490
491 /* Make sure 64 bit math will not overflow. */
492 if (a > ULONG_MAX - (unsigned long)log_base ||
493 a + (unsigned long)log_base > ULONG_MAX)
494 return 0;
495
496 return access_ok(VERIFY_WRITE, log_base + a,
497 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
498}
499
500/* Caller should have vq mutex and device mutex. */
501static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
502 int log_all)
503{
504 int i;
505
506 if (!mem)
507 return 0;
508
509 for (i = 0; i < mem->nregions; ++i) {
510 struct vhost_memory_region *m = mem->regions + i;
511 unsigned long a = m->userspace_addr;
512 if (m->memory_size > ULONG_MAX)
513 return 0;
514 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
515 m->memory_size))
516 return 0;
517 else if (log_all && !log_access_ok(log_base,
518 m->guest_phys_addr,
519 m->memory_size))
520 return 0;
521 }
522 return 1;
523}
524
525/* Can we switch to this memory table? */
526/* Caller should have device mutex but not vq mutex */
527static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
528 int log_all)
529{
530 int i;
531
532 for (i = 0; i < d->nvqs; ++i) {
533 int ok;
534 mutex_lock(&d->vqs[i].mutex);
535 /* If ring is inactive, will check when it's enabled. */
536 if (d->vqs[i].private_data)
537 ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
538 log_all);
539 else
540 ok = 1;
541 mutex_unlock(&d->vqs[i].mutex);
542 if (!ok)
543 return 0;
544 }
545 return 1;
546}
547
548static int vq_access_ok(struct vhost_dev *d, unsigned int num,
549 struct vring_desc __user *desc,
550 struct vring_avail __user *avail,
551 struct vring_used __user *used)
552{
553 size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
554 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
555 access_ok(VERIFY_READ, avail,
556 sizeof *avail + num * sizeof *avail->ring + s) &&
557 access_ok(VERIFY_WRITE, used,
558 sizeof *used + num * sizeof *used->ring + s);
559}
560
561/* Can we log writes? */
562/* Caller should have device mutex but not vq mutex */
563int vhost_log_access_ok(struct vhost_dev *dev)
564{
565 struct vhost_memory *mp;
566
567 mp = rcu_dereference_protected(dev->memory,
568 lockdep_is_held(&dev->mutex));
569 return memory_access_ok(dev, mp, 1);
570}
571
572/* Verify access for write logging. */
573/* Caller should have vq mutex and device mutex */
574static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
575 void __user *log_base)
576{
577 struct vhost_memory *mp;
578 size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
579
580 mp = rcu_dereference_protected(vq->dev->memory,
581 lockdep_is_held(&vq->mutex));
582 return vq_memory_access_ok(log_base, mp,
583 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
584 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
585 sizeof *vq->used +
586 vq->num * sizeof *vq->used->ring + s));
587}
588
589/* Can we start vq? */
590/* Caller should have vq mutex and device mutex */
591int vhost_vq_access_ok(struct vhost_virtqueue *vq)
592{
593 return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
594 vq_log_access_ok(vq->dev, vq, vq->log_base);
595}
596
597static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
598{
599 struct vhost_memory mem, *newmem, *oldmem;
600 unsigned long size = offsetof(struct vhost_memory, regions);
601
602 if (copy_from_user(&mem, m, size))
603 return -EFAULT;
604 if (mem.padding)
605 return -EOPNOTSUPP;
606 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
607 return -E2BIG;
608 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
609 if (!newmem)
610 return -ENOMEM;
611
612 memcpy(newmem, &mem, size);
613 if (copy_from_user(newmem->regions, m->regions,
614 mem.nregions * sizeof *m->regions)) {
615 kfree(newmem);
616 return -EFAULT;
617 }
618
619 if (!memory_access_ok(d, newmem,
620 vhost_has_feature(d, VHOST_F_LOG_ALL))) {
621 kfree(newmem);
622 return -EFAULT;
623 }
624 oldmem = rcu_dereference_protected(d->memory,
625 lockdep_is_held(&d->mutex));
626 rcu_assign_pointer(d->memory, newmem);
627 synchronize_rcu();
628 kfree(oldmem);
629 return 0;
630}
631
632static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
633{
634 struct file *eventfp, *filep = NULL,
635 *pollstart = NULL, *pollstop = NULL;
636 struct eventfd_ctx *ctx = NULL;
637 u32 __user *idxp = argp;
638 struct vhost_virtqueue *vq;
639 struct vhost_vring_state s;
640 struct vhost_vring_file f;
641 struct vhost_vring_addr a;
642 u32 idx;
643 long r;
644
645 r = get_user(idx, idxp);
646 if (r < 0)
647 return r;
648 if (idx >= d->nvqs)
649 return -ENOBUFS;
650
651 vq = d->vqs + idx;
652
653 mutex_lock(&vq->mutex);
654
655 switch (ioctl) {
656 case VHOST_SET_VRING_NUM:
657 /* Resizing ring with an active backend?
658 * You don't want to do that. */
659 if (vq->private_data) {
660 r = -EBUSY;
661 break;
662 }
663 if (copy_from_user(&s, argp, sizeof s)) {
664 r = -EFAULT;
665 break;
666 }
667 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
668 r = -EINVAL;
669 break;
670 }
671 vq->num = s.num;
672 break;
673 case VHOST_SET_VRING_BASE:
674 /* Moving base with an active backend?
675 * You don't want to do that. */
676 if (vq->private_data) {
677 r = -EBUSY;
678 break;
679 }
680 if (copy_from_user(&s, argp, sizeof s)) {
681 r = -EFAULT;
682 break;
683 }
684 if (s.num > 0xffff) {
685 r = -EINVAL;
686 break;
687 }
688 vq->last_avail_idx = s.num;
689 /* Forget the cached index value. */
690 vq->avail_idx = vq->last_avail_idx;
691 break;
692 case VHOST_GET_VRING_BASE:
693 s.index = idx;
694 s.num = vq->last_avail_idx;
695 if (copy_to_user(argp, &s, sizeof s))
696 r = -EFAULT;
697 break;
698 case VHOST_SET_VRING_ADDR:
699 if (copy_from_user(&a, argp, sizeof a)) {
700 r = -EFAULT;
701 break;
702 }
703 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
704 r = -EOPNOTSUPP;
705 break;
706 }
707 /* For 32bit, verify that the top 32bits of the user
708 data are set to zero. */
709 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
710 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
711 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
712 r = -EFAULT;
713 break;
714 }
715 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
716 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
717 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
718 r = -EINVAL;
719 break;
720 }
721
722 /* We only verify access here if backend is configured.
723 * If it is not, we don't as size might not have been setup.
724 * We will verify when backend is configured. */
725 if (vq->private_data) {
726 if (!vq_access_ok(d, vq->num,
727 (void __user *)(unsigned long)a.desc_user_addr,
728 (void __user *)(unsigned long)a.avail_user_addr,
729 (void __user *)(unsigned long)a.used_user_addr)) {
730 r = -EINVAL;
731 break;
732 }
733
734 /* Also validate log access for used ring if enabled. */
735 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
736 !log_access_ok(vq->log_base, a.log_guest_addr,
737 sizeof *vq->used +
738 vq->num * sizeof *vq->used->ring)) {
739 r = -EINVAL;
740 break;
741 }
742 }
743
744 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
745 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
746 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
747 vq->log_addr = a.log_guest_addr;
748 vq->used = (void __user *)(unsigned long)a.used_user_addr;
749 break;
750 case VHOST_SET_VRING_KICK:
751 if (copy_from_user(&f, argp, sizeof f)) {
752 r = -EFAULT;
753 break;
754 }
755 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
756 if (IS_ERR(eventfp)) {
757 r = PTR_ERR(eventfp);
758 break;
759 }
760 if (eventfp != vq->kick) {
761 pollstop = filep = vq->kick;
762 pollstart = vq->kick = eventfp;
763 } else
764 filep = eventfp;
765 break;
766 case VHOST_SET_VRING_CALL:
767 if (copy_from_user(&f, argp, sizeof f)) {
768 r = -EFAULT;
769 break;
770 }
771 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
772 if (IS_ERR(eventfp)) {
773 r = PTR_ERR(eventfp);
774 break;
775 }
776 if (eventfp != vq->call) {
777 filep = vq->call;
778 ctx = vq->call_ctx;
779 vq->call = eventfp;
780 vq->call_ctx = eventfp ?
781 eventfd_ctx_fileget(eventfp) : NULL;
782 } else
783 filep = eventfp;
784 break;
785 case VHOST_SET_VRING_ERR:
786 if (copy_from_user(&f, argp, sizeof f)) {
787 r = -EFAULT;
788 break;
789 }
790 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
791 if (IS_ERR(eventfp)) {
792 r = PTR_ERR(eventfp);
793 break;
794 }
795 if (eventfp != vq->error) {
796 filep = vq->error;
797 vq->error = eventfp;
798 ctx = vq->error_ctx;
799 vq->error_ctx = eventfp ?
800 eventfd_ctx_fileget(eventfp) : NULL;
801 } else
802 filep = eventfp;
803 break;
804 default:
805 r = -ENOIOCTLCMD;
806 }
807
808 if (pollstop && vq->handle_kick)
809 vhost_poll_stop(&vq->poll);
810
811 if (ctx)
812 eventfd_ctx_put(ctx);
813 if (filep)
814 fput(filep);
815
816 if (pollstart && vq->handle_kick)
817 vhost_poll_start(&vq->poll, vq->kick);
818
819 mutex_unlock(&vq->mutex);
820
821 if (pollstop && vq->handle_kick)
822 vhost_poll_flush(&vq->poll);
823 return r;
824}
825
826/* Caller must have device mutex */
827long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
828{
829 void __user *argp = (void __user *)arg;
830 struct file *eventfp, *filep = NULL;
831 struct eventfd_ctx *ctx = NULL;
832 u64 p;
833 long r;
834 int i, fd;
835
836 /* If you are not the owner, you can become one */
837 if (ioctl == VHOST_SET_OWNER) {
838 r = vhost_dev_set_owner(d);
839 goto done;
840 }
841
842 /* You must be the owner to do anything else */
843 r = vhost_dev_check_owner(d);
844 if (r)
845 goto done;
846
847 switch (ioctl) {
848 case VHOST_SET_MEM_TABLE:
849 r = vhost_set_memory(d, argp);
850 break;
851 case VHOST_SET_LOG_BASE:
852 if (copy_from_user(&p, argp, sizeof p)) {
853 r = -EFAULT;
854 break;
855 }
856 if ((u64)(unsigned long)p != p) {
857 r = -EFAULT;
858 break;
859 }
860 for (i = 0; i < d->nvqs; ++i) {
861 struct vhost_virtqueue *vq;
862 void __user *base = (void __user *)(unsigned long)p;
863 vq = d->vqs + i;
864 mutex_lock(&vq->mutex);
865 /* If ring is inactive, will check when it's enabled. */
866 if (vq->private_data && !vq_log_access_ok(d, vq, base))
867 r = -EFAULT;
868 else
869 vq->log_base = base;
870 mutex_unlock(&vq->mutex);
871 }
872 break;
873 case VHOST_SET_LOG_FD:
874 r = get_user(fd, (int __user *)argp);
875 if (r < 0)
876 break;
877 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
878 if (IS_ERR(eventfp)) {
879 r = PTR_ERR(eventfp);
880 break;
881 }
882 if (eventfp != d->log_file) {
883 filep = d->log_file;
884 ctx = d->log_ctx;
885 d->log_ctx = eventfp ?
886 eventfd_ctx_fileget(eventfp) : NULL;
887 } else
888 filep = eventfp;
889 for (i = 0; i < d->nvqs; ++i) {
890 mutex_lock(&d->vqs[i].mutex);
891 d->vqs[i].log_ctx = d->log_ctx;
892 mutex_unlock(&d->vqs[i].mutex);
893 }
894 if (ctx)
895 eventfd_ctx_put(ctx);
896 if (filep)
897 fput(filep);
898 break;
899 default:
900 r = vhost_set_vring(d, ioctl, argp);
901 break;
902 }
903done:
904 return r;
905}
906
907static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
908 __u64 addr, __u32 len)
909{
910 struct vhost_memory_region *reg;
911 int i;
912
913 /* linear search is not brilliant, but we really have on the order of 6
914 * regions in practice */
915 for (i = 0; i < mem->nregions; ++i) {
916 reg = mem->regions + i;
917 if (reg->guest_phys_addr <= addr &&
918 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
919 return reg;
920 }
921 return NULL;
922}
923
924/* TODO: This is really inefficient. We need something like get_user()
925 * (instruction directly accesses the data, with an exception table entry
926 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
927 */
928static int set_bit_to_user(int nr, void __user *addr)
929{
930 unsigned long log = (unsigned long)addr;
931 struct page *page;
932 void *base;
933 int bit = nr + (log % PAGE_SIZE) * 8;
934 int r;
935
936 r = get_user_pages_fast(log, 1, 1, &page);
937 if (r < 0)
938 return r;
939 BUG_ON(r != 1);
940 base = kmap_atomic(page, KM_USER0);
941 set_bit(bit, base);
942 kunmap_atomic(base, KM_USER0);
943 set_page_dirty_lock(page);
944 put_page(page);
945 return 0;
946}
947
948static int log_write(void __user *log_base,
949 u64 write_address, u64 write_length)
950{
951 u64 write_page = write_address / VHOST_PAGE_SIZE;
952 int r;
953
954 if (!write_length)
955 return 0;
956 write_length += write_address % VHOST_PAGE_SIZE;
957 for (;;) {
958 u64 base = (u64)(unsigned long)log_base;
959 u64 log = base + write_page / 8;
960 int bit = write_page % 8;
961 if ((u64)(unsigned long)log != log)
962 return -EFAULT;
963 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
964 if (r < 0)
965 return r;
966 if (write_length <= VHOST_PAGE_SIZE)
967 break;
968 write_length -= VHOST_PAGE_SIZE;
969 write_page += 1;
970 }
971 return r;
972}
973
974int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
975 unsigned int log_num, u64 len)
976{
977 int i, r;
978
979 /* Make sure data written is seen before log. */
980 smp_wmb();
981 for (i = 0; i < log_num; ++i) {
982 u64 l = min(log[i].len, len);
983 r = log_write(vq->log_base, log[i].addr, l);
984 if (r < 0)
985 return r;
986 len -= l;
987 if (!len) {
988 if (vq->log_ctx)
989 eventfd_signal(vq->log_ctx, 1);
990 return 0;
991 }
992 }
993 /* Length written exceeds what we have stored. This is a bug. */
994 BUG();
995 return 0;
996}
997
998static int vhost_update_used_flags(struct vhost_virtqueue *vq)
999{
1000 void __user *used;
1001 if (__put_user(vq->used_flags, &vq->used->flags) < 0)
1002 return -EFAULT;
1003 if (unlikely(vq->log_used)) {
1004 /* Make sure the flag is seen before log. */
1005 smp_wmb();
1006 /* Log used flag write. */
1007 used = &vq->used->flags;
1008 log_write(vq->log_base, vq->log_addr +
1009 (used - (void __user *)vq->used),
1010 sizeof vq->used->flags);
1011 if (vq->log_ctx)
1012 eventfd_signal(vq->log_ctx, 1);
1013 }
1014 return 0;
1015}
1016
1017static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1018{
1019 if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1020 return -EFAULT;
1021 if (unlikely(vq->log_used)) {
1022 void __user *used;
1023 /* Make sure the event is seen before log. */
1024 smp_wmb();
1025 /* Log avail event write */
1026 used = vhost_avail_event(vq);
1027 log_write(vq->log_base, vq->log_addr +
1028 (used - (void __user *)vq->used),
1029 sizeof *vhost_avail_event(vq));
1030 if (vq->log_ctx)
1031 eventfd_signal(vq->log_ctx, 1);
1032 }
1033 return 0;
1034}
1035
1036int vhost_init_used(struct vhost_virtqueue *vq)
1037{
1038 int r;
1039 if (!vq->private_data)
1040 return 0;
1041
1042 r = vhost_update_used_flags(vq);
1043 if (r)
1044 return r;
1045 vq->signalled_used_valid = false;
1046 return get_user(vq->last_used_idx, &vq->used->idx);
1047}
1048
1049static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1050 struct iovec iov[], int iov_size)
1051{
1052 const struct vhost_memory_region *reg;
1053 struct vhost_memory *mem;
1054 struct iovec *_iov;
1055 u64 s = 0;
1056 int ret = 0;
1057
1058 rcu_read_lock();
1059
1060 mem = rcu_dereference(dev->memory);
1061 while ((u64)len > s) {
1062 u64 size;
1063 if (unlikely(ret >= iov_size)) {
1064 ret = -ENOBUFS;
1065 break;
1066 }
1067 reg = find_region(mem, addr, len);
1068 if (unlikely(!reg)) {
1069 ret = -EFAULT;
1070 break;
1071 }
1072 _iov = iov + ret;
1073 size = reg->memory_size - addr + reg->guest_phys_addr;
1074 _iov->iov_len = min((u64)len, size);
1075 _iov->iov_base = (void __user *)(unsigned long)
1076 (reg->userspace_addr + addr - reg->guest_phys_addr);
1077 s += size;
1078 addr += size;
1079 ++ret;
1080 }
1081
1082 rcu_read_unlock();
1083 return ret;
1084}
1085
1086/* Each buffer in the virtqueues is actually a chain of descriptors. This
1087 * function returns the next descriptor in the chain,
1088 * or -1U if we're at the end. */
1089static unsigned next_desc(struct vring_desc *desc)
1090{
1091 unsigned int next;
1092
1093 /* If this descriptor says it doesn't chain, we're done. */
1094 if (!(desc->flags & VRING_DESC_F_NEXT))
1095 return -1U;
1096
1097 /* Check they're not leading us off end of descriptors. */
1098 next = desc->next;
1099 /* Make sure compiler knows to grab that: we don't want it changing! */
1100 /* We will use the result as an index in an array, so most
1101 * architectures only need a compiler barrier here. */
1102 read_barrier_depends();
1103
1104 return next;
1105}
1106
1107static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1108 struct iovec iov[], unsigned int iov_size,
1109 unsigned int *out_num, unsigned int *in_num,
1110 struct vhost_log *log, unsigned int *log_num,
1111 struct vring_desc *indirect)
1112{
1113 struct vring_desc desc;
1114 unsigned int i = 0, count, found = 0;
1115 int ret;
1116
1117 /* Sanity check */
1118 if (unlikely(indirect->len % sizeof desc)) {
1119 vq_err(vq, "Invalid length in indirect descriptor: "
1120 "len 0x%llx not multiple of 0x%zx\n",
1121 (unsigned long long)indirect->len,
1122 sizeof desc);
1123 return -EINVAL;
1124 }
1125
1126 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1127 UIO_MAXIOV);
1128 if (unlikely(ret < 0)) {
1129 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1130 return ret;
1131 }
1132
1133 /* We will use the result as an address to read from, so most
1134 * architectures only need a compiler barrier here. */
1135 read_barrier_depends();
1136
1137 count = indirect->len / sizeof desc;
1138 /* Buffers are chained via a 16 bit next field, so
1139 * we can have at most 2^16 of these. */
1140 if (unlikely(count > USHRT_MAX + 1)) {
1141 vq_err(vq, "Indirect buffer length too big: %d\n",
1142 indirect->len);
1143 return -E2BIG;
1144 }
1145
1146 do {
1147 unsigned iov_count = *in_num + *out_num;
1148 if (unlikely(++found > count)) {
1149 vq_err(vq, "Loop detected: last one at %u "
1150 "indirect size %u\n",
1151 i, count);
1152 return -EINVAL;
1153 }
1154 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1155 vq->indirect, sizeof desc))) {
1156 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1157 i, (size_t)indirect->addr + i * sizeof desc);
1158 return -EINVAL;
1159 }
1160 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1161 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1162 i, (size_t)indirect->addr + i * sizeof desc);
1163 return -EINVAL;
1164 }
1165
1166 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1167 iov_size - iov_count);
1168 if (unlikely(ret < 0)) {
1169 vq_err(vq, "Translation failure %d indirect idx %d\n",
1170 ret, i);
1171 return ret;
1172 }
1173 /* If this is an input descriptor, increment that count. */
1174 if (desc.flags & VRING_DESC_F_WRITE) {
1175 *in_num += ret;
1176 if (unlikely(log)) {
1177 log[*log_num].addr = desc.addr;
1178 log[*log_num].len = desc.len;
1179 ++*log_num;
1180 }
1181 } else {
1182 /* If it's an output descriptor, they're all supposed
1183 * to come before any input descriptors. */
1184 if (unlikely(*in_num)) {
1185 vq_err(vq, "Indirect descriptor "
1186 "has out after in: idx %d\n", i);
1187 return -EINVAL;
1188 }
1189 *out_num += ret;
1190 }
1191 } while ((i = next_desc(&desc)) != -1);
1192 return 0;
1193}
1194
1195/* This looks in the virtqueue and for the first available buffer, and converts
1196 * it to an iovec for convenient access. Since descriptors consist of some
1197 * number of output then some number of input descriptors, it's actually two
1198 * iovecs, but we pack them into one and note how many of each there were.
1199 *
1200 * This function returns the descriptor number found, or vq->num (which is
1201 * never a valid descriptor number) if none was found. A negative code is
1202 * returned on error. */
1203int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1204 struct iovec iov[], unsigned int iov_size,
1205 unsigned int *out_num, unsigned int *in_num,
1206 struct vhost_log *log, unsigned int *log_num)
1207{
1208 struct vring_desc desc;
1209 unsigned int i, head, found = 0;
1210 u16 last_avail_idx;
1211 int ret;
1212
1213 /* Check it isn't doing very strange things with descriptor numbers. */
1214 last_avail_idx = vq->last_avail_idx;
1215 if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1216 vq_err(vq, "Failed to access avail idx at %p\n",
1217 &vq->avail->idx);
1218 return -EFAULT;
1219 }
1220
1221 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1222 vq_err(vq, "Guest moved used index from %u to %u",
1223 last_avail_idx, vq->avail_idx);
1224 return -EFAULT;
1225 }
1226
1227 /* If there's nothing new since last we looked, return invalid. */
1228 if (vq->avail_idx == last_avail_idx)
1229 return vq->num;
1230
1231 /* Only get avail ring entries after they have been exposed by guest. */
1232 smp_rmb();
1233
1234 /* Grab the next descriptor number they're advertising, and increment
1235 * the index we've seen. */
1236 if (unlikely(__get_user(head,
1237 &vq->avail->ring[last_avail_idx % vq->num]))) {
1238 vq_err(vq, "Failed to read head: idx %d address %p\n",
1239 last_avail_idx,
1240 &vq->avail->ring[last_avail_idx % vq->num]);
1241 return -EFAULT;
1242 }
1243
1244 /* If their number is silly, that's an error. */
1245 if (unlikely(head >= vq->num)) {
1246 vq_err(vq, "Guest says index %u > %u is available",
1247 head, vq->num);
1248 return -EINVAL;
1249 }
1250
1251 /* When we start there are none of either input nor output. */
1252 *out_num = *in_num = 0;
1253 if (unlikely(log))
1254 *log_num = 0;
1255
1256 i = head;
1257 do {
1258 unsigned iov_count = *in_num + *out_num;
1259 if (unlikely(i >= vq->num)) {
1260 vq_err(vq, "Desc index is %u > %u, head = %u",
1261 i, vq->num, head);
1262 return -EINVAL;
1263 }
1264 if (unlikely(++found > vq->num)) {
1265 vq_err(vq, "Loop detected: last one at %u "
1266 "vq size %u head %u\n",
1267 i, vq->num, head);
1268 return -EINVAL;
1269 }
1270 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1271 if (unlikely(ret)) {
1272 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1273 i, vq->desc + i);
1274 return -EFAULT;
1275 }
1276 if (desc.flags & VRING_DESC_F_INDIRECT) {
1277 ret = get_indirect(dev, vq, iov, iov_size,
1278 out_num, in_num,
1279 log, log_num, &desc);
1280 if (unlikely(ret < 0)) {
1281 vq_err(vq, "Failure detected "
1282 "in indirect descriptor at idx %d\n", i);
1283 return ret;
1284 }
1285 continue;
1286 }
1287
1288 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1289 iov_size - iov_count);
1290 if (unlikely(ret < 0)) {
1291 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1292 ret, i);
1293 return ret;
1294 }
1295 if (desc.flags & VRING_DESC_F_WRITE) {
1296 /* If this is an input descriptor,
1297 * increment that count. */
1298 *in_num += ret;
1299 if (unlikely(log)) {
1300 log[*log_num].addr = desc.addr;
1301 log[*log_num].len = desc.len;
1302 ++*log_num;
1303 }
1304 } else {
1305 /* If it's an output descriptor, they're all supposed
1306 * to come before any input descriptors. */
1307 if (unlikely(*in_num)) {
1308 vq_err(vq, "Descriptor has out after in: "
1309 "idx %d\n", i);
1310 return -EINVAL;
1311 }
1312 *out_num += ret;
1313 }
1314 } while ((i = next_desc(&desc)) != -1);
1315
1316 /* On success, increment avail index. */
1317 vq->last_avail_idx++;
1318
1319 /* Assume notifications from guest are disabled at this point,
1320 * if they aren't we would need to update avail_event index. */
1321 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1322 return head;
1323}
1324
1325/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1326void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1327{
1328 vq->last_avail_idx -= n;
1329}
1330
1331/* After we've used one of their buffers, we tell them about it. We'll then
1332 * want to notify the guest, using eventfd. */
1333int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1334{
1335 struct vring_used_elem __user *used;
1336
1337 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1338 * next entry in that used ring. */
1339 used = &vq->used->ring[vq->last_used_idx % vq->num];
1340 if (__put_user(head, &used->id)) {
1341 vq_err(vq, "Failed to write used id");
1342 return -EFAULT;
1343 }
1344 if (__put_user(len, &used->len)) {
1345 vq_err(vq, "Failed to write used len");
1346 return -EFAULT;
1347 }
1348 /* Make sure buffer is written before we update index. */
1349 smp_wmb();
1350 if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1351 vq_err(vq, "Failed to increment used idx");
1352 return -EFAULT;
1353 }
1354 if (unlikely(vq->log_used)) {
1355 /* Make sure data is seen before log. */
1356 smp_wmb();
1357 /* Log used ring entry write. */
1358 log_write(vq->log_base,
1359 vq->log_addr +
1360 ((void __user *)used - (void __user *)vq->used),
1361 sizeof *used);
1362 /* Log used index update. */
1363 log_write(vq->log_base,
1364 vq->log_addr + offsetof(struct vring_used, idx),
1365 sizeof vq->used->idx);
1366 if (vq->log_ctx)
1367 eventfd_signal(vq->log_ctx, 1);
1368 }
1369 vq->last_used_idx++;
1370 /* If the driver never bothers to signal in a very long while,
1371 * used index might wrap around. If that happens, invalidate
1372 * signalled_used index we stored. TODO: make sure driver
1373 * signals at least once in 2^16 and remove this. */
1374 if (unlikely(vq->last_used_idx == vq->signalled_used))
1375 vq->signalled_used_valid = false;
1376 return 0;
1377}
1378
1379static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1380 struct vring_used_elem *heads,
1381 unsigned count)
1382{
1383 struct vring_used_elem __user *used;
1384 u16 old, new;
1385 int start;
1386
1387 start = vq->last_used_idx % vq->num;
1388 used = vq->used->ring + start;
1389 if (__copy_to_user(used, heads, count * sizeof *used)) {
1390 vq_err(vq, "Failed to write used");
1391 return -EFAULT;
1392 }
1393 if (unlikely(vq->log_used)) {
1394 /* Make sure data is seen before log. */
1395 smp_wmb();
1396 /* Log used ring entry write. */
1397 log_write(vq->log_base,
1398 vq->log_addr +
1399 ((void __user *)used - (void __user *)vq->used),
1400 count * sizeof *used);
1401 }
1402 old = vq->last_used_idx;
1403 new = (vq->last_used_idx += count);
1404 /* If the driver never bothers to signal in a very long while,
1405 * used index might wrap around. If that happens, invalidate
1406 * signalled_used index we stored. TODO: make sure driver
1407 * signals at least once in 2^16 and remove this. */
1408 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1409 vq->signalled_used_valid = false;
1410 return 0;
1411}
1412
1413/* After we've used one of their buffers, we tell them about it. We'll then
1414 * want to notify the guest, using eventfd. */
1415int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1416 unsigned count)
1417{
1418 int start, n, r;
1419
1420 start = vq->last_used_idx % vq->num;
1421 n = vq->num - start;
1422 if (n < count) {
1423 r = __vhost_add_used_n(vq, heads, n);
1424 if (r < 0)
1425 return r;
1426 heads += n;
1427 count -= n;
1428 }
1429 r = __vhost_add_used_n(vq, heads, count);
1430
1431 /* Make sure buffer is written before we update index. */
1432 smp_wmb();
1433 if (put_user(vq->last_used_idx, &vq->used->idx)) {
1434 vq_err(vq, "Failed to increment used idx");
1435 return -EFAULT;
1436 }
1437 if (unlikely(vq->log_used)) {
1438 /* Log used index update. */
1439 log_write(vq->log_base,
1440 vq->log_addr + offsetof(struct vring_used, idx),
1441 sizeof vq->used->idx);
1442 if (vq->log_ctx)
1443 eventfd_signal(vq->log_ctx, 1);
1444 }
1445 return r;
1446}
1447
1448static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1449{
1450 __u16 old, new, event;
1451 bool v;
1452 /* Flush out used index updates. This is paired
1453 * with the barrier that the Guest executes when enabling
1454 * interrupts. */
1455 smp_mb();
1456
1457 if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1458 unlikely(vq->avail_idx == vq->last_avail_idx))
1459 return true;
1460
1461 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1462 __u16 flags;
1463 if (__get_user(flags, &vq->avail->flags)) {
1464 vq_err(vq, "Failed to get flags");
1465 return true;
1466 }
1467 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1468 }
1469 old = vq->signalled_used;
1470 v = vq->signalled_used_valid;
1471 new = vq->signalled_used = vq->last_used_idx;
1472 vq->signalled_used_valid = true;
1473
1474 if (unlikely(!v))
1475 return true;
1476
1477 if (get_user(event, vhost_used_event(vq))) {
1478 vq_err(vq, "Failed to get used event idx");
1479 return true;
1480 }
1481 return vring_need_event(event, new, old);
1482}
1483
1484/* This actually signals the guest, using eventfd. */
1485void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1486{
1487 /* Signal the Guest tell them we used something up. */
1488 if (vq->call_ctx && vhost_notify(dev, vq))
1489 eventfd_signal(vq->call_ctx, 1);
1490}
1491
1492/* And here's the combo meal deal. Supersize me! */
1493void vhost_add_used_and_signal(struct vhost_dev *dev,
1494 struct vhost_virtqueue *vq,
1495 unsigned int head, int len)
1496{
1497 vhost_add_used(vq, head, len);
1498 vhost_signal(dev, vq);
1499}
1500
1501/* multi-buffer version of vhost_add_used_and_signal */
1502void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1503 struct vhost_virtqueue *vq,
1504 struct vring_used_elem *heads, unsigned count)
1505{
1506 vhost_add_used_n(vq, heads, count);
1507 vhost_signal(dev, vq);
1508}
1509
1510/* OK, now we need to know about added descriptors. */
1511bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1512{
1513 u16 avail_idx;
1514 int r;
1515
1516 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1517 return false;
1518 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1519 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1520 r = vhost_update_used_flags(vq);
1521 if (r) {
1522 vq_err(vq, "Failed to enable notification at %p: %d\n",
1523 &vq->used->flags, r);
1524 return false;
1525 }
1526 } else {
1527 r = vhost_update_avail_event(vq, vq->avail_idx);
1528 if (r) {
1529 vq_err(vq, "Failed to update avail event index at %p: %d\n",
1530 vhost_avail_event(vq), r);
1531 return false;
1532 }
1533 }
1534 /* They could have slipped one in as we were doing that: make
1535 * sure it's written, then check again. */
1536 smp_mb();
1537 r = __get_user(avail_idx, &vq->avail->idx);
1538 if (r) {
1539 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1540 &vq->avail->idx, r);
1541 return false;
1542 }
1543
1544 return avail_idx != vq->avail_idx;
1545}
1546
1547/* We don't need to be notified again. */
1548void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1549{
1550 int r;
1551
1552 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1553 return;
1554 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1555 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1556 r = vhost_update_used_flags(vq);
1557 if (r)
1558 vq_err(vq, "Failed to enable notification at %p: %d\n",
1559 &vq->used->flags, r);
1560 }
1561}
1562
1563static void vhost_zerocopy_done_signal(struct kref *kref)
1564{
1565 struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
1566 kref);
1567 wake_up(&ubufs->wait);
1568}
1569
1570struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
1571 bool zcopy)
1572{
1573 struct vhost_ubuf_ref *ubufs;
1574 /* No zero copy backend? Nothing to count. */
1575 if (!zcopy)
1576 return NULL;
1577 ubufs = kmalloc(sizeof *ubufs, GFP_KERNEL);
1578 if (!ubufs)
1579 return ERR_PTR(-ENOMEM);
1580 kref_init(&ubufs->kref);
1581 init_waitqueue_head(&ubufs->wait);
1582 ubufs->vq = vq;
1583 return ubufs;
1584}
1585
1586void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
1587{
1588 kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1589}
1590
1591void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
1592{
1593 kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1594 wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
1595 kfree(ubufs);
1596}
1597
1598void vhost_zerocopy_callback(void *arg)
1599{
1600 struct ubuf_info *ubuf = arg;
1601 struct vhost_ubuf_ref *ubufs = ubuf->arg;
1602 struct vhost_virtqueue *vq = ubufs->vq;
1603
1604 /* set len = 1 to mark this desc buffers done DMA */
1605 vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
1606 kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1607}
1/* Copyright (C) 2009 Red Hat, Inc.
2 * Copyright (C) 2006 Rusty Russell IBM Corporation
3 *
4 * Author: Michael S. Tsirkin <mst@redhat.com>
5 *
6 * Inspiration, some code, and most witty comments come from
7 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
8 *
9 * This work is licensed under the terms of the GNU GPL, version 2.
10 *
11 * Generic code for virtio server in host kernel.
12 */
13
14#include <linux/eventfd.h>
15#include <linux/vhost.h>
16#include <linux/uio.h>
17#include <linux/mm.h>
18#include <linux/mmu_context.h>
19#include <linux/miscdevice.h>
20#include <linux/mutex.h>
21#include <linux/rcupdate.h>
22#include <linux/poll.h>
23#include <linux/file.h>
24#include <linux/highmem.h>
25#include <linux/slab.h>
26#include <linux/kthread.h>
27#include <linux/cgroup.h>
28#include <linux/module.h>
29
30#include "vhost.h"
31
32enum {
33 VHOST_MEMORY_MAX_NREGIONS = 64,
34 VHOST_MEMORY_F_LOG = 0x1,
35};
36
37#define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num])
38#define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num])
39
40static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
41 poll_table *pt)
42{
43 struct vhost_poll *poll;
44
45 poll = container_of(pt, struct vhost_poll, table);
46 poll->wqh = wqh;
47 add_wait_queue(wqh, &poll->wait);
48}
49
50static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
51 void *key)
52{
53 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
54
55 if (!((unsigned long)key & poll->mask))
56 return 0;
57
58 vhost_poll_queue(poll);
59 return 0;
60}
61
62void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
63{
64 INIT_LIST_HEAD(&work->node);
65 work->fn = fn;
66 init_waitqueue_head(&work->done);
67 work->flushing = 0;
68 work->queue_seq = work->done_seq = 0;
69}
70EXPORT_SYMBOL_GPL(vhost_work_init);
71
72/* Init poll structure */
73void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
74 unsigned long mask, struct vhost_dev *dev)
75{
76 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
77 init_poll_funcptr(&poll->table, vhost_poll_func);
78 poll->mask = mask;
79 poll->dev = dev;
80 poll->wqh = NULL;
81
82 vhost_work_init(&poll->work, fn);
83}
84EXPORT_SYMBOL_GPL(vhost_poll_init);
85
86/* Start polling a file. We add ourselves to file's wait queue. The caller must
87 * keep a reference to a file until after vhost_poll_stop is called. */
88int vhost_poll_start(struct vhost_poll *poll, struct file *file)
89{
90 unsigned long mask;
91 int ret = 0;
92
93 if (poll->wqh)
94 return 0;
95
96 mask = file->f_op->poll(file, &poll->table);
97 if (mask)
98 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
99 if (mask & POLLERR) {
100 if (poll->wqh)
101 remove_wait_queue(poll->wqh, &poll->wait);
102 ret = -EINVAL;
103 }
104
105 return ret;
106}
107EXPORT_SYMBOL_GPL(vhost_poll_start);
108
109/* Stop polling a file. After this function returns, it becomes safe to drop the
110 * file reference. You must also flush afterwards. */
111void vhost_poll_stop(struct vhost_poll *poll)
112{
113 if (poll->wqh) {
114 remove_wait_queue(poll->wqh, &poll->wait);
115 poll->wqh = NULL;
116 }
117}
118EXPORT_SYMBOL_GPL(vhost_poll_stop);
119
120static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
121 unsigned seq)
122{
123 int left;
124
125 spin_lock_irq(&dev->work_lock);
126 left = seq - work->done_seq;
127 spin_unlock_irq(&dev->work_lock);
128 return left <= 0;
129}
130
131void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
132{
133 unsigned seq;
134 int flushing;
135
136 spin_lock_irq(&dev->work_lock);
137 seq = work->queue_seq;
138 work->flushing++;
139 spin_unlock_irq(&dev->work_lock);
140 wait_event(work->done, vhost_work_seq_done(dev, work, seq));
141 spin_lock_irq(&dev->work_lock);
142 flushing = --work->flushing;
143 spin_unlock_irq(&dev->work_lock);
144 BUG_ON(flushing < 0);
145}
146EXPORT_SYMBOL_GPL(vhost_work_flush);
147
148/* Flush any work that has been scheduled. When calling this, don't hold any
149 * locks that are also used by the callback. */
150void vhost_poll_flush(struct vhost_poll *poll)
151{
152 vhost_work_flush(poll->dev, &poll->work);
153}
154EXPORT_SYMBOL_GPL(vhost_poll_flush);
155
156void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
157{
158 unsigned long flags;
159
160 spin_lock_irqsave(&dev->work_lock, flags);
161 if (list_empty(&work->node)) {
162 list_add_tail(&work->node, &dev->work_list);
163 work->queue_seq++;
164 spin_unlock_irqrestore(&dev->work_lock, flags);
165 wake_up_process(dev->worker);
166 } else {
167 spin_unlock_irqrestore(&dev->work_lock, flags);
168 }
169}
170EXPORT_SYMBOL_GPL(vhost_work_queue);
171
172void vhost_poll_queue(struct vhost_poll *poll)
173{
174 vhost_work_queue(poll->dev, &poll->work);
175}
176EXPORT_SYMBOL_GPL(vhost_poll_queue);
177
178static void vhost_vq_reset(struct vhost_dev *dev,
179 struct vhost_virtqueue *vq)
180{
181 vq->num = 1;
182 vq->desc = NULL;
183 vq->avail = NULL;
184 vq->used = NULL;
185 vq->last_avail_idx = 0;
186 vq->avail_idx = 0;
187 vq->last_used_idx = 0;
188 vq->signalled_used = 0;
189 vq->signalled_used_valid = false;
190 vq->used_flags = 0;
191 vq->log_used = false;
192 vq->log_addr = -1ull;
193 vq->private_data = NULL;
194 vq->log_base = NULL;
195 vq->error_ctx = NULL;
196 vq->error = NULL;
197 vq->kick = NULL;
198 vq->call_ctx = NULL;
199 vq->call = NULL;
200 vq->log_ctx = NULL;
201}
202
203static int vhost_worker(void *data)
204{
205 struct vhost_dev *dev = data;
206 struct vhost_work *work = NULL;
207 unsigned uninitialized_var(seq);
208 mm_segment_t oldfs = get_fs();
209
210 set_fs(USER_DS);
211 use_mm(dev->mm);
212
213 for (;;) {
214 /* mb paired w/ kthread_stop */
215 set_current_state(TASK_INTERRUPTIBLE);
216
217 spin_lock_irq(&dev->work_lock);
218 if (work) {
219 work->done_seq = seq;
220 if (work->flushing)
221 wake_up_all(&work->done);
222 }
223
224 if (kthread_should_stop()) {
225 spin_unlock_irq(&dev->work_lock);
226 __set_current_state(TASK_RUNNING);
227 break;
228 }
229 if (!list_empty(&dev->work_list)) {
230 work = list_first_entry(&dev->work_list,
231 struct vhost_work, node);
232 list_del_init(&work->node);
233 seq = work->queue_seq;
234 } else
235 work = NULL;
236 spin_unlock_irq(&dev->work_lock);
237
238 if (work) {
239 __set_current_state(TASK_RUNNING);
240 work->fn(work);
241 if (need_resched())
242 schedule();
243 } else
244 schedule();
245
246 }
247 unuse_mm(dev->mm);
248 set_fs(oldfs);
249 return 0;
250}
251
252static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
253{
254 kfree(vq->indirect);
255 vq->indirect = NULL;
256 kfree(vq->log);
257 vq->log = NULL;
258 kfree(vq->heads);
259 vq->heads = NULL;
260}
261
262/* Helper to allocate iovec buffers for all vqs. */
263static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
264{
265 struct vhost_virtqueue *vq;
266 int i;
267
268 for (i = 0; i < dev->nvqs; ++i) {
269 vq = dev->vqs[i];
270 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
271 GFP_KERNEL);
272 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
273 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
274 if (!vq->indirect || !vq->log || !vq->heads)
275 goto err_nomem;
276 }
277 return 0;
278
279err_nomem:
280 for (; i >= 0; --i)
281 vhost_vq_free_iovecs(dev->vqs[i]);
282 return -ENOMEM;
283}
284
285static void vhost_dev_free_iovecs(struct vhost_dev *dev)
286{
287 int i;
288
289 for (i = 0; i < dev->nvqs; ++i)
290 vhost_vq_free_iovecs(dev->vqs[i]);
291}
292
293void vhost_dev_init(struct vhost_dev *dev,
294 struct vhost_virtqueue **vqs, int nvqs)
295{
296 struct vhost_virtqueue *vq;
297 int i;
298
299 dev->vqs = vqs;
300 dev->nvqs = nvqs;
301 mutex_init(&dev->mutex);
302 dev->log_ctx = NULL;
303 dev->log_file = NULL;
304 dev->memory = NULL;
305 dev->mm = NULL;
306 spin_lock_init(&dev->work_lock);
307 INIT_LIST_HEAD(&dev->work_list);
308 dev->worker = NULL;
309
310 for (i = 0; i < dev->nvqs; ++i) {
311 vq = dev->vqs[i];
312 vq->log = NULL;
313 vq->indirect = NULL;
314 vq->heads = NULL;
315 vq->dev = dev;
316 mutex_init(&vq->mutex);
317 vhost_vq_reset(dev, vq);
318 if (vq->handle_kick)
319 vhost_poll_init(&vq->poll, vq->handle_kick,
320 POLLIN, dev);
321 }
322}
323EXPORT_SYMBOL_GPL(vhost_dev_init);
324
325/* Caller should have device mutex */
326long vhost_dev_check_owner(struct vhost_dev *dev)
327{
328 /* Are you the owner? If not, I don't think you mean to do that */
329 return dev->mm == current->mm ? 0 : -EPERM;
330}
331EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
332
333struct vhost_attach_cgroups_struct {
334 struct vhost_work work;
335 struct task_struct *owner;
336 int ret;
337};
338
339static void vhost_attach_cgroups_work(struct vhost_work *work)
340{
341 struct vhost_attach_cgroups_struct *s;
342
343 s = container_of(work, struct vhost_attach_cgroups_struct, work);
344 s->ret = cgroup_attach_task_all(s->owner, current);
345}
346
347static int vhost_attach_cgroups(struct vhost_dev *dev)
348{
349 struct vhost_attach_cgroups_struct attach;
350
351 attach.owner = current;
352 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
353 vhost_work_queue(dev, &attach.work);
354 vhost_work_flush(dev, &attach.work);
355 return attach.ret;
356}
357
358/* Caller should have device mutex */
359bool vhost_dev_has_owner(struct vhost_dev *dev)
360{
361 return dev->mm;
362}
363EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
364
365/* Caller should have device mutex */
366long vhost_dev_set_owner(struct vhost_dev *dev)
367{
368 struct task_struct *worker;
369 int err;
370
371 /* Is there an owner already? */
372 if (vhost_dev_has_owner(dev)) {
373 err = -EBUSY;
374 goto err_mm;
375 }
376
377 /* No owner, become one */
378 dev->mm = get_task_mm(current);
379 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
380 if (IS_ERR(worker)) {
381 err = PTR_ERR(worker);
382 goto err_worker;
383 }
384
385 dev->worker = worker;
386 wake_up_process(worker); /* avoid contributing to loadavg */
387
388 err = vhost_attach_cgroups(dev);
389 if (err)
390 goto err_cgroup;
391
392 err = vhost_dev_alloc_iovecs(dev);
393 if (err)
394 goto err_cgroup;
395
396 return 0;
397err_cgroup:
398 kthread_stop(worker);
399 dev->worker = NULL;
400err_worker:
401 if (dev->mm)
402 mmput(dev->mm);
403 dev->mm = NULL;
404err_mm:
405 return err;
406}
407EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
408
409struct vhost_memory *vhost_dev_reset_owner_prepare(void)
410{
411 return kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
412}
413EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
414
415/* Caller should have device mutex */
416void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_memory *memory)
417{
418 vhost_dev_cleanup(dev, true);
419
420 /* Restore memory to default empty mapping. */
421 memory->nregions = 0;
422 RCU_INIT_POINTER(dev->memory, memory);
423}
424EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
425
426void vhost_dev_stop(struct vhost_dev *dev)
427{
428 int i;
429
430 for (i = 0; i < dev->nvqs; ++i) {
431 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
432 vhost_poll_stop(&dev->vqs[i]->poll);
433 vhost_poll_flush(&dev->vqs[i]->poll);
434 }
435 }
436}
437EXPORT_SYMBOL_GPL(vhost_dev_stop);
438
439/* Caller should have device mutex if and only if locked is set */
440void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
441{
442 int i;
443
444 for (i = 0; i < dev->nvqs; ++i) {
445 if (dev->vqs[i]->error_ctx)
446 eventfd_ctx_put(dev->vqs[i]->error_ctx);
447 if (dev->vqs[i]->error)
448 fput(dev->vqs[i]->error);
449 if (dev->vqs[i]->kick)
450 fput(dev->vqs[i]->kick);
451 if (dev->vqs[i]->call_ctx)
452 eventfd_ctx_put(dev->vqs[i]->call_ctx);
453 if (dev->vqs[i]->call)
454 fput(dev->vqs[i]->call);
455 vhost_vq_reset(dev, dev->vqs[i]);
456 }
457 vhost_dev_free_iovecs(dev);
458 if (dev->log_ctx)
459 eventfd_ctx_put(dev->log_ctx);
460 dev->log_ctx = NULL;
461 if (dev->log_file)
462 fput(dev->log_file);
463 dev->log_file = NULL;
464 /* No one will access memory at this point */
465 kfree(rcu_dereference_protected(dev->memory,
466 locked ==
467 lockdep_is_held(&dev->mutex)));
468 RCU_INIT_POINTER(dev->memory, NULL);
469 WARN_ON(!list_empty(&dev->work_list));
470 if (dev->worker) {
471 kthread_stop(dev->worker);
472 dev->worker = NULL;
473 }
474 if (dev->mm)
475 mmput(dev->mm);
476 dev->mm = NULL;
477}
478EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
479
480static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
481{
482 u64 a = addr / VHOST_PAGE_SIZE / 8;
483
484 /* Make sure 64 bit math will not overflow. */
485 if (a > ULONG_MAX - (unsigned long)log_base ||
486 a + (unsigned long)log_base > ULONG_MAX)
487 return 0;
488
489 return access_ok(VERIFY_WRITE, log_base + a,
490 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
491}
492
493/* Caller should have vq mutex and device mutex. */
494static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
495 int log_all)
496{
497 int i;
498
499 if (!mem)
500 return 0;
501
502 for (i = 0; i < mem->nregions; ++i) {
503 struct vhost_memory_region *m = mem->regions + i;
504 unsigned long a = m->userspace_addr;
505 if (m->memory_size > ULONG_MAX)
506 return 0;
507 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
508 m->memory_size))
509 return 0;
510 else if (log_all && !log_access_ok(log_base,
511 m->guest_phys_addr,
512 m->memory_size))
513 return 0;
514 }
515 return 1;
516}
517
518/* Can we switch to this memory table? */
519/* Caller should have device mutex but not vq mutex */
520static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
521 int log_all)
522{
523 int i;
524
525 for (i = 0; i < d->nvqs; ++i) {
526 int ok;
527 mutex_lock(&d->vqs[i]->mutex);
528 /* If ring is inactive, will check when it's enabled. */
529 if (d->vqs[i]->private_data)
530 ok = vq_memory_access_ok(d->vqs[i]->log_base, mem,
531 log_all);
532 else
533 ok = 1;
534 mutex_unlock(&d->vqs[i]->mutex);
535 if (!ok)
536 return 0;
537 }
538 return 1;
539}
540
541static int vq_access_ok(struct vhost_dev *d, unsigned int num,
542 struct vring_desc __user *desc,
543 struct vring_avail __user *avail,
544 struct vring_used __user *used)
545{
546 size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
547 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
548 access_ok(VERIFY_READ, avail,
549 sizeof *avail + num * sizeof *avail->ring + s) &&
550 access_ok(VERIFY_WRITE, used,
551 sizeof *used + num * sizeof *used->ring + s);
552}
553
554/* Can we log writes? */
555/* Caller should have device mutex but not vq mutex */
556int vhost_log_access_ok(struct vhost_dev *dev)
557{
558 struct vhost_memory *mp;
559
560 mp = rcu_dereference_protected(dev->memory,
561 lockdep_is_held(&dev->mutex));
562 return memory_access_ok(dev, mp, 1);
563}
564EXPORT_SYMBOL_GPL(vhost_log_access_ok);
565
566/* Verify access for write logging. */
567/* Caller should have vq mutex and device mutex */
568static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
569 void __user *log_base)
570{
571 struct vhost_memory *mp;
572 size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
573
574 mp = rcu_dereference_protected(vq->dev->memory,
575 lockdep_is_held(&vq->mutex));
576 return vq_memory_access_ok(log_base, mp,
577 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
578 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
579 sizeof *vq->used +
580 vq->num * sizeof *vq->used->ring + s));
581}
582
583/* Can we start vq? */
584/* Caller should have vq mutex and device mutex */
585int vhost_vq_access_ok(struct vhost_virtqueue *vq)
586{
587 return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
588 vq_log_access_ok(vq->dev, vq, vq->log_base);
589}
590EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
591
592static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
593{
594 struct vhost_memory mem, *newmem, *oldmem;
595 unsigned long size = offsetof(struct vhost_memory, regions);
596
597 if (copy_from_user(&mem, m, size))
598 return -EFAULT;
599 if (mem.padding)
600 return -EOPNOTSUPP;
601 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
602 return -E2BIG;
603 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
604 if (!newmem)
605 return -ENOMEM;
606
607 memcpy(newmem, &mem, size);
608 if (copy_from_user(newmem->regions, m->regions,
609 mem.nregions * sizeof *m->regions)) {
610 kfree(newmem);
611 return -EFAULT;
612 }
613
614 if (!memory_access_ok(d, newmem,
615 vhost_has_feature(d, VHOST_F_LOG_ALL))) {
616 kfree(newmem);
617 return -EFAULT;
618 }
619 oldmem = rcu_dereference_protected(d->memory,
620 lockdep_is_held(&d->mutex));
621 rcu_assign_pointer(d->memory, newmem);
622 synchronize_rcu();
623 kfree(oldmem);
624 return 0;
625}
626
627long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
628{
629 struct file *eventfp, *filep = NULL;
630 bool pollstart = false, pollstop = false;
631 struct eventfd_ctx *ctx = NULL;
632 u32 __user *idxp = argp;
633 struct vhost_virtqueue *vq;
634 struct vhost_vring_state s;
635 struct vhost_vring_file f;
636 struct vhost_vring_addr a;
637 u32 idx;
638 long r;
639
640 r = get_user(idx, idxp);
641 if (r < 0)
642 return r;
643 if (idx >= d->nvqs)
644 return -ENOBUFS;
645
646 vq = d->vqs[idx];
647
648 mutex_lock(&vq->mutex);
649
650 switch (ioctl) {
651 case VHOST_SET_VRING_NUM:
652 /* Resizing ring with an active backend?
653 * You don't want to do that. */
654 if (vq->private_data) {
655 r = -EBUSY;
656 break;
657 }
658 if (copy_from_user(&s, argp, sizeof s)) {
659 r = -EFAULT;
660 break;
661 }
662 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
663 r = -EINVAL;
664 break;
665 }
666 vq->num = s.num;
667 break;
668 case VHOST_SET_VRING_BASE:
669 /* Moving base with an active backend?
670 * You don't want to do that. */
671 if (vq->private_data) {
672 r = -EBUSY;
673 break;
674 }
675 if (copy_from_user(&s, argp, sizeof s)) {
676 r = -EFAULT;
677 break;
678 }
679 if (s.num > 0xffff) {
680 r = -EINVAL;
681 break;
682 }
683 vq->last_avail_idx = s.num;
684 /* Forget the cached index value. */
685 vq->avail_idx = vq->last_avail_idx;
686 break;
687 case VHOST_GET_VRING_BASE:
688 s.index = idx;
689 s.num = vq->last_avail_idx;
690 if (copy_to_user(argp, &s, sizeof s))
691 r = -EFAULT;
692 break;
693 case VHOST_SET_VRING_ADDR:
694 if (copy_from_user(&a, argp, sizeof a)) {
695 r = -EFAULT;
696 break;
697 }
698 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
699 r = -EOPNOTSUPP;
700 break;
701 }
702 /* For 32bit, verify that the top 32bits of the user
703 data are set to zero. */
704 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
705 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
706 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
707 r = -EFAULT;
708 break;
709 }
710 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
711 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
712 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
713 r = -EINVAL;
714 break;
715 }
716
717 /* We only verify access here if backend is configured.
718 * If it is not, we don't as size might not have been setup.
719 * We will verify when backend is configured. */
720 if (vq->private_data) {
721 if (!vq_access_ok(d, vq->num,
722 (void __user *)(unsigned long)a.desc_user_addr,
723 (void __user *)(unsigned long)a.avail_user_addr,
724 (void __user *)(unsigned long)a.used_user_addr)) {
725 r = -EINVAL;
726 break;
727 }
728
729 /* Also validate log access for used ring if enabled. */
730 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
731 !log_access_ok(vq->log_base, a.log_guest_addr,
732 sizeof *vq->used +
733 vq->num * sizeof *vq->used->ring)) {
734 r = -EINVAL;
735 break;
736 }
737 }
738
739 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
740 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
741 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
742 vq->log_addr = a.log_guest_addr;
743 vq->used = (void __user *)(unsigned long)a.used_user_addr;
744 break;
745 case VHOST_SET_VRING_KICK:
746 if (copy_from_user(&f, argp, sizeof f)) {
747 r = -EFAULT;
748 break;
749 }
750 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
751 if (IS_ERR(eventfp)) {
752 r = PTR_ERR(eventfp);
753 break;
754 }
755 if (eventfp != vq->kick) {
756 pollstop = (filep = vq->kick) != NULL;
757 pollstart = (vq->kick = eventfp) != NULL;
758 } else
759 filep = eventfp;
760 break;
761 case VHOST_SET_VRING_CALL:
762 if (copy_from_user(&f, argp, sizeof f)) {
763 r = -EFAULT;
764 break;
765 }
766 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
767 if (IS_ERR(eventfp)) {
768 r = PTR_ERR(eventfp);
769 break;
770 }
771 if (eventfp != vq->call) {
772 filep = vq->call;
773 ctx = vq->call_ctx;
774 vq->call = eventfp;
775 vq->call_ctx = eventfp ?
776 eventfd_ctx_fileget(eventfp) : NULL;
777 } else
778 filep = eventfp;
779 break;
780 case VHOST_SET_VRING_ERR:
781 if (copy_from_user(&f, argp, sizeof f)) {
782 r = -EFAULT;
783 break;
784 }
785 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
786 if (IS_ERR(eventfp)) {
787 r = PTR_ERR(eventfp);
788 break;
789 }
790 if (eventfp != vq->error) {
791 filep = vq->error;
792 vq->error = eventfp;
793 ctx = vq->error_ctx;
794 vq->error_ctx = eventfp ?
795 eventfd_ctx_fileget(eventfp) : NULL;
796 } else
797 filep = eventfp;
798 break;
799 default:
800 r = -ENOIOCTLCMD;
801 }
802
803 if (pollstop && vq->handle_kick)
804 vhost_poll_stop(&vq->poll);
805
806 if (ctx)
807 eventfd_ctx_put(ctx);
808 if (filep)
809 fput(filep);
810
811 if (pollstart && vq->handle_kick)
812 r = vhost_poll_start(&vq->poll, vq->kick);
813
814 mutex_unlock(&vq->mutex);
815
816 if (pollstop && vq->handle_kick)
817 vhost_poll_flush(&vq->poll);
818 return r;
819}
820EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
821
822/* Caller must have device mutex */
823long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
824{
825 struct file *eventfp, *filep = NULL;
826 struct eventfd_ctx *ctx = NULL;
827 u64 p;
828 long r;
829 int i, fd;
830
831 /* If you are not the owner, you can become one */
832 if (ioctl == VHOST_SET_OWNER) {
833 r = vhost_dev_set_owner(d);
834 goto done;
835 }
836
837 /* You must be the owner to do anything else */
838 r = vhost_dev_check_owner(d);
839 if (r)
840 goto done;
841
842 switch (ioctl) {
843 case VHOST_SET_MEM_TABLE:
844 r = vhost_set_memory(d, argp);
845 break;
846 case VHOST_SET_LOG_BASE:
847 if (copy_from_user(&p, argp, sizeof p)) {
848 r = -EFAULT;
849 break;
850 }
851 if ((u64)(unsigned long)p != p) {
852 r = -EFAULT;
853 break;
854 }
855 for (i = 0; i < d->nvqs; ++i) {
856 struct vhost_virtqueue *vq;
857 void __user *base = (void __user *)(unsigned long)p;
858 vq = d->vqs[i];
859 mutex_lock(&vq->mutex);
860 /* If ring is inactive, will check when it's enabled. */
861 if (vq->private_data && !vq_log_access_ok(d, vq, base))
862 r = -EFAULT;
863 else
864 vq->log_base = base;
865 mutex_unlock(&vq->mutex);
866 }
867 break;
868 case VHOST_SET_LOG_FD:
869 r = get_user(fd, (int __user *)argp);
870 if (r < 0)
871 break;
872 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
873 if (IS_ERR(eventfp)) {
874 r = PTR_ERR(eventfp);
875 break;
876 }
877 if (eventfp != d->log_file) {
878 filep = d->log_file;
879 ctx = d->log_ctx;
880 d->log_ctx = eventfp ?
881 eventfd_ctx_fileget(eventfp) : NULL;
882 } else
883 filep = eventfp;
884 for (i = 0; i < d->nvqs; ++i) {
885 mutex_lock(&d->vqs[i]->mutex);
886 d->vqs[i]->log_ctx = d->log_ctx;
887 mutex_unlock(&d->vqs[i]->mutex);
888 }
889 if (ctx)
890 eventfd_ctx_put(ctx);
891 if (filep)
892 fput(filep);
893 break;
894 default:
895 r = -ENOIOCTLCMD;
896 break;
897 }
898done:
899 return r;
900}
901EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
902
903static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
904 __u64 addr, __u32 len)
905{
906 struct vhost_memory_region *reg;
907 int i;
908
909 /* linear search is not brilliant, but we really have on the order of 6
910 * regions in practice */
911 for (i = 0; i < mem->nregions; ++i) {
912 reg = mem->regions + i;
913 if (reg->guest_phys_addr <= addr &&
914 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
915 return reg;
916 }
917 return NULL;
918}
919
920/* TODO: This is really inefficient. We need something like get_user()
921 * (instruction directly accesses the data, with an exception table entry
922 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
923 */
924static int set_bit_to_user(int nr, void __user *addr)
925{
926 unsigned long log = (unsigned long)addr;
927 struct page *page;
928 void *base;
929 int bit = nr + (log % PAGE_SIZE) * 8;
930 int r;
931
932 r = get_user_pages_fast(log, 1, 1, &page);
933 if (r < 0)
934 return r;
935 BUG_ON(r != 1);
936 base = kmap_atomic(page);
937 set_bit(bit, base);
938 kunmap_atomic(base);
939 set_page_dirty_lock(page);
940 put_page(page);
941 return 0;
942}
943
944static int log_write(void __user *log_base,
945 u64 write_address, u64 write_length)
946{
947 u64 write_page = write_address / VHOST_PAGE_SIZE;
948 int r;
949
950 if (!write_length)
951 return 0;
952 write_length += write_address % VHOST_PAGE_SIZE;
953 for (;;) {
954 u64 base = (u64)(unsigned long)log_base;
955 u64 log = base + write_page / 8;
956 int bit = write_page % 8;
957 if ((u64)(unsigned long)log != log)
958 return -EFAULT;
959 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
960 if (r < 0)
961 return r;
962 if (write_length <= VHOST_PAGE_SIZE)
963 break;
964 write_length -= VHOST_PAGE_SIZE;
965 write_page += 1;
966 }
967 return r;
968}
969
970int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
971 unsigned int log_num, u64 len)
972{
973 int i, r;
974
975 /* Make sure data written is seen before log. */
976 smp_wmb();
977 for (i = 0; i < log_num; ++i) {
978 u64 l = min(log[i].len, len);
979 r = log_write(vq->log_base, log[i].addr, l);
980 if (r < 0)
981 return r;
982 len -= l;
983 if (!len) {
984 if (vq->log_ctx)
985 eventfd_signal(vq->log_ctx, 1);
986 return 0;
987 }
988 }
989 /* Length written exceeds what we have stored. This is a bug. */
990 BUG();
991 return 0;
992}
993EXPORT_SYMBOL_GPL(vhost_log_write);
994
995static int vhost_update_used_flags(struct vhost_virtqueue *vq)
996{
997 void __user *used;
998 if (__put_user(vq->used_flags, &vq->used->flags) < 0)
999 return -EFAULT;
1000 if (unlikely(vq->log_used)) {
1001 /* Make sure the flag is seen before log. */
1002 smp_wmb();
1003 /* Log used flag write. */
1004 used = &vq->used->flags;
1005 log_write(vq->log_base, vq->log_addr +
1006 (used - (void __user *)vq->used),
1007 sizeof vq->used->flags);
1008 if (vq->log_ctx)
1009 eventfd_signal(vq->log_ctx, 1);
1010 }
1011 return 0;
1012}
1013
1014static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1015{
1016 if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1017 return -EFAULT;
1018 if (unlikely(vq->log_used)) {
1019 void __user *used;
1020 /* Make sure the event is seen before log. */
1021 smp_wmb();
1022 /* Log avail event write */
1023 used = vhost_avail_event(vq);
1024 log_write(vq->log_base, vq->log_addr +
1025 (used - (void __user *)vq->used),
1026 sizeof *vhost_avail_event(vq));
1027 if (vq->log_ctx)
1028 eventfd_signal(vq->log_ctx, 1);
1029 }
1030 return 0;
1031}
1032
1033int vhost_init_used(struct vhost_virtqueue *vq)
1034{
1035 int r;
1036 if (!vq->private_data)
1037 return 0;
1038
1039 r = vhost_update_used_flags(vq);
1040 if (r)
1041 return r;
1042 vq->signalled_used_valid = false;
1043 return get_user(vq->last_used_idx, &vq->used->idx);
1044}
1045EXPORT_SYMBOL_GPL(vhost_init_used);
1046
1047static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1048 struct iovec iov[], int iov_size)
1049{
1050 const struct vhost_memory_region *reg;
1051 struct vhost_memory *mem;
1052 struct iovec *_iov;
1053 u64 s = 0;
1054 int ret = 0;
1055
1056 rcu_read_lock();
1057
1058 mem = rcu_dereference(dev->memory);
1059 while ((u64)len > s) {
1060 u64 size;
1061 if (unlikely(ret >= iov_size)) {
1062 ret = -ENOBUFS;
1063 break;
1064 }
1065 reg = find_region(mem, addr, len);
1066 if (unlikely(!reg)) {
1067 ret = -EFAULT;
1068 break;
1069 }
1070 _iov = iov + ret;
1071 size = reg->memory_size - addr + reg->guest_phys_addr;
1072 _iov->iov_len = min((u64)len - s, size);
1073 _iov->iov_base = (void __user *)(unsigned long)
1074 (reg->userspace_addr + addr - reg->guest_phys_addr);
1075 s += size;
1076 addr += size;
1077 ++ret;
1078 }
1079
1080 rcu_read_unlock();
1081 return ret;
1082}
1083
1084/* Each buffer in the virtqueues is actually a chain of descriptors. This
1085 * function returns the next descriptor in the chain,
1086 * or -1U if we're at the end. */
1087static unsigned next_desc(struct vring_desc *desc)
1088{
1089 unsigned int next;
1090
1091 /* If this descriptor says it doesn't chain, we're done. */
1092 if (!(desc->flags & VRING_DESC_F_NEXT))
1093 return -1U;
1094
1095 /* Check they're not leading us off end of descriptors. */
1096 next = desc->next;
1097 /* Make sure compiler knows to grab that: we don't want it changing! */
1098 /* We will use the result as an index in an array, so most
1099 * architectures only need a compiler barrier here. */
1100 read_barrier_depends();
1101
1102 return next;
1103}
1104
1105static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1106 struct iovec iov[], unsigned int iov_size,
1107 unsigned int *out_num, unsigned int *in_num,
1108 struct vhost_log *log, unsigned int *log_num,
1109 struct vring_desc *indirect)
1110{
1111 struct vring_desc desc;
1112 unsigned int i = 0, count, found = 0;
1113 int ret;
1114
1115 /* Sanity check */
1116 if (unlikely(indirect->len % sizeof desc)) {
1117 vq_err(vq, "Invalid length in indirect descriptor: "
1118 "len 0x%llx not multiple of 0x%zx\n",
1119 (unsigned long long)indirect->len,
1120 sizeof desc);
1121 return -EINVAL;
1122 }
1123
1124 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1125 UIO_MAXIOV);
1126 if (unlikely(ret < 0)) {
1127 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1128 return ret;
1129 }
1130
1131 /* We will use the result as an address to read from, so most
1132 * architectures only need a compiler barrier here. */
1133 read_barrier_depends();
1134
1135 count = indirect->len / sizeof desc;
1136 /* Buffers are chained via a 16 bit next field, so
1137 * we can have at most 2^16 of these. */
1138 if (unlikely(count > USHRT_MAX + 1)) {
1139 vq_err(vq, "Indirect buffer length too big: %d\n",
1140 indirect->len);
1141 return -E2BIG;
1142 }
1143
1144 do {
1145 unsigned iov_count = *in_num + *out_num;
1146 if (unlikely(++found > count)) {
1147 vq_err(vq, "Loop detected: last one at %u "
1148 "indirect size %u\n",
1149 i, count);
1150 return -EINVAL;
1151 }
1152 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1153 vq->indirect, sizeof desc))) {
1154 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1155 i, (size_t)indirect->addr + i * sizeof desc);
1156 return -EINVAL;
1157 }
1158 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1159 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1160 i, (size_t)indirect->addr + i * sizeof desc);
1161 return -EINVAL;
1162 }
1163
1164 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1165 iov_size - iov_count);
1166 if (unlikely(ret < 0)) {
1167 vq_err(vq, "Translation failure %d indirect idx %d\n",
1168 ret, i);
1169 return ret;
1170 }
1171 /* If this is an input descriptor, increment that count. */
1172 if (desc.flags & VRING_DESC_F_WRITE) {
1173 *in_num += ret;
1174 if (unlikely(log)) {
1175 log[*log_num].addr = desc.addr;
1176 log[*log_num].len = desc.len;
1177 ++*log_num;
1178 }
1179 } else {
1180 /* If it's an output descriptor, they're all supposed
1181 * to come before any input descriptors. */
1182 if (unlikely(*in_num)) {
1183 vq_err(vq, "Indirect descriptor "
1184 "has out after in: idx %d\n", i);
1185 return -EINVAL;
1186 }
1187 *out_num += ret;
1188 }
1189 } while ((i = next_desc(&desc)) != -1);
1190 return 0;
1191}
1192
1193/* This looks in the virtqueue and for the first available buffer, and converts
1194 * it to an iovec for convenient access. Since descriptors consist of some
1195 * number of output then some number of input descriptors, it's actually two
1196 * iovecs, but we pack them into one and note how many of each there were.
1197 *
1198 * This function returns the descriptor number found, or vq->num (which is
1199 * never a valid descriptor number) if none was found. A negative code is
1200 * returned on error. */
1201int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1202 struct iovec iov[], unsigned int iov_size,
1203 unsigned int *out_num, unsigned int *in_num,
1204 struct vhost_log *log, unsigned int *log_num)
1205{
1206 struct vring_desc desc;
1207 unsigned int i, head, found = 0;
1208 u16 last_avail_idx;
1209 int ret;
1210
1211 /* Check it isn't doing very strange things with descriptor numbers. */
1212 last_avail_idx = vq->last_avail_idx;
1213 if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1214 vq_err(vq, "Failed to access avail idx at %p\n",
1215 &vq->avail->idx);
1216 return -EFAULT;
1217 }
1218
1219 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1220 vq_err(vq, "Guest moved used index from %u to %u",
1221 last_avail_idx, vq->avail_idx);
1222 return -EFAULT;
1223 }
1224
1225 /* If there's nothing new since last we looked, return invalid. */
1226 if (vq->avail_idx == last_avail_idx)
1227 return vq->num;
1228
1229 /* Only get avail ring entries after they have been exposed by guest. */
1230 smp_rmb();
1231
1232 /* Grab the next descriptor number they're advertising, and increment
1233 * the index we've seen. */
1234 if (unlikely(__get_user(head,
1235 &vq->avail->ring[last_avail_idx % vq->num]))) {
1236 vq_err(vq, "Failed to read head: idx %d address %p\n",
1237 last_avail_idx,
1238 &vq->avail->ring[last_avail_idx % vq->num]);
1239 return -EFAULT;
1240 }
1241
1242 /* If their number is silly, that's an error. */
1243 if (unlikely(head >= vq->num)) {
1244 vq_err(vq, "Guest says index %u > %u is available",
1245 head, vq->num);
1246 return -EINVAL;
1247 }
1248
1249 /* When we start there are none of either input nor output. */
1250 *out_num = *in_num = 0;
1251 if (unlikely(log))
1252 *log_num = 0;
1253
1254 i = head;
1255 do {
1256 unsigned iov_count = *in_num + *out_num;
1257 if (unlikely(i >= vq->num)) {
1258 vq_err(vq, "Desc index is %u > %u, head = %u",
1259 i, vq->num, head);
1260 return -EINVAL;
1261 }
1262 if (unlikely(++found > vq->num)) {
1263 vq_err(vq, "Loop detected: last one at %u "
1264 "vq size %u head %u\n",
1265 i, vq->num, head);
1266 return -EINVAL;
1267 }
1268 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1269 if (unlikely(ret)) {
1270 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1271 i, vq->desc + i);
1272 return -EFAULT;
1273 }
1274 if (desc.flags & VRING_DESC_F_INDIRECT) {
1275 ret = get_indirect(dev, vq, iov, iov_size,
1276 out_num, in_num,
1277 log, log_num, &desc);
1278 if (unlikely(ret < 0)) {
1279 vq_err(vq, "Failure detected "
1280 "in indirect descriptor at idx %d\n", i);
1281 return ret;
1282 }
1283 continue;
1284 }
1285
1286 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1287 iov_size - iov_count);
1288 if (unlikely(ret < 0)) {
1289 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1290 ret, i);
1291 return ret;
1292 }
1293 if (desc.flags & VRING_DESC_F_WRITE) {
1294 /* If this is an input descriptor,
1295 * increment that count. */
1296 *in_num += ret;
1297 if (unlikely(log)) {
1298 log[*log_num].addr = desc.addr;
1299 log[*log_num].len = desc.len;
1300 ++*log_num;
1301 }
1302 } else {
1303 /* If it's an output descriptor, they're all supposed
1304 * to come before any input descriptors. */
1305 if (unlikely(*in_num)) {
1306 vq_err(vq, "Descriptor has out after in: "
1307 "idx %d\n", i);
1308 return -EINVAL;
1309 }
1310 *out_num += ret;
1311 }
1312 } while ((i = next_desc(&desc)) != -1);
1313
1314 /* On success, increment avail index. */
1315 vq->last_avail_idx++;
1316
1317 /* Assume notifications from guest are disabled at this point,
1318 * if they aren't we would need to update avail_event index. */
1319 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1320 return head;
1321}
1322EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1323
1324/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1325void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1326{
1327 vq->last_avail_idx -= n;
1328}
1329EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1330
1331/* After we've used one of their buffers, we tell them about it. We'll then
1332 * want to notify the guest, using eventfd. */
1333int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1334{
1335 struct vring_used_elem heads = { head, len };
1336
1337 return vhost_add_used_n(vq, &heads, 1);
1338}
1339EXPORT_SYMBOL_GPL(vhost_add_used);
1340
1341static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1342 struct vring_used_elem *heads,
1343 unsigned count)
1344{
1345 struct vring_used_elem __user *used;
1346 u16 old, new;
1347 int start;
1348
1349 start = vq->last_used_idx % vq->num;
1350 used = vq->used->ring + start;
1351 if (count == 1) {
1352 if (__put_user(heads[0].id, &used->id)) {
1353 vq_err(vq, "Failed to write used id");
1354 return -EFAULT;
1355 }
1356 if (__put_user(heads[0].len, &used->len)) {
1357 vq_err(vq, "Failed to write used len");
1358 return -EFAULT;
1359 }
1360 } else if (__copy_to_user(used, heads, count * sizeof *used)) {
1361 vq_err(vq, "Failed to write used");
1362 return -EFAULT;
1363 }
1364 if (unlikely(vq->log_used)) {
1365 /* Make sure data is seen before log. */
1366 smp_wmb();
1367 /* Log used ring entry write. */
1368 log_write(vq->log_base,
1369 vq->log_addr +
1370 ((void __user *)used - (void __user *)vq->used),
1371 count * sizeof *used);
1372 }
1373 old = vq->last_used_idx;
1374 new = (vq->last_used_idx += count);
1375 /* If the driver never bothers to signal in a very long while,
1376 * used index might wrap around. If that happens, invalidate
1377 * signalled_used index we stored. TODO: make sure driver
1378 * signals at least once in 2^16 and remove this. */
1379 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1380 vq->signalled_used_valid = false;
1381 return 0;
1382}
1383
1384/* After we've used one of their buffers, we tell them about it. We'll then
1385 * want to notify the guest, using eventfd. */
1386int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1387 unsigned count)
1388{
1389 int start, n, r;
1390
1391 start = vq->last_used_idx % vq->num;
1392 n = vq->num - start;
1393 if (n < count) {
1394 r = __vhost_add_used_n(vq, heads, n);
1395 if (r < 0)
1396 return r;
1397 heads += n;
1398 count -= n;
1399 }
1400 r = __vhost_add_used_n(vq, heads, count);
1401
1402 /* Make sure buffer is written before we update index. */
1403 smp_wmb();
1404 if (put_user(vq->last_used_idx, &vq->used->idx)) {
1405 vq_err(vq, "Failed to increment used idx");
1406 return -EFAULT;
1407 }
1408 if (unlikely(vq->log_used)) {
1409 /* Log used index update. */
1410 log_write(vq->log_base,
1411 vq->log_addr + offsetof(struct vring_used, idx),
1412 sizeof vq->used->idx);
1413 if (vq->log_ctx)
1414 eventfd_signal(vq->log_ctx, 1);
1415 }
1416 return r;
1417}
1418EXPORT_SYMBOL_GPL(vhost_add_used_n);
1419
1420static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1421{
1422 __u16 old, new, event;
1423 bool v;
1424 /* Flush out used index updates. This is paired
1425 * with the barrier that the Guest executes when enabling
1426 * interrupts. */
1427 smp_mb();
1428
1429 if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1430 unlikely(vq->avail_idx == vq->last_avail_idx))
1431 return true;
1432
1433 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1434 __u16 flags;
1435 if (__get_user(flags, &vq->avail->flags)) {
1436 vq_err(vq, "Failed to get flags");
1437 return true;
1438 }
1439 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1440 }
1441 old = vq->signalled_used;
1442 v = vq->signalled_used_valid;
1443 new = vq->signalled_used = vq->last_used_idx;
1444 vq->signalled_used_valid = true;
1445
1446 if (unlikely(!v))
1447 return true;
1448
1449 if (get_user(event, vhost_used_event(vq))) {
1450 vq_err(vq, "Failed to get used event idx");
1451 return true;
1452 }
1453 return vring_need_event(event, new, old);
1454}
1455
1456/* This actually signals the guest, using eventfd. */
1457void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1458{
1459 /* Signal the Guest tell them we used something up. */
1460 if (vq->call_ctx && vhost_notify(dev, vq))
1461 eventfd_signal(vq->call_ctx, 1);
1462}
1463EXPORT_SYMBOL_GPL(vhost_signal);
1464
1465/* And here's the combo meal deal. Supersize me! */
1466void vhost_add_used_and_signal(struct vhost_dev *dev,
1467 struct vhost_virtqueue *vq,
1468 unsigned int head, int len)
1469{
1470 vhost_add_used(vq, head, len);
1471 vhost_signal(dev, vq);
1472}
1473EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1474
1475/* multi-buffer version of vhost_add_used_and_signal */
1476void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1477 struct vhost_virtqueue *vq,
1478 struct vring_used_elem *heads, unsigned count)
1479{
1480 vhost_add_used_n(vq, heads, count);
1481 vhost_signal(dev, vq);
1482}
1483EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1484
1485/* OK, now we need to know about added descriptors. */
1486bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1487{
1488 u16 avail_idx;
1489 int r;
1490
1491 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1492 return false;
1493 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1494 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1495 r = vhost_update_used_flags(vq);
1496 if (r) {
1497 vq_err(vq, "Failed to enable notification at %p: %d\n",
1498 &vq->used->flags, r);
1499 return false;
1500 }
1501 } else {
1502 r = vhost_update_avail_event(vq, vq->avail_idx);
1503 if (r) {
1504 vq_err(vq, "Failed to update avail event index at %p: %d\n",
1505 vhost_avail_event(vq), r);
1506 return false;
1507 }
1508 }
1509 /* They could have slipped one in as we were doing that: make
1510 * sure it's written, then check again. */
1511 smp_mb();
1512 r = __get_user(avail_idx, &vq->avail->idx);
1513 if (r) {
1514 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1515 &vq->avail->idx, r);
1516 return false;
1517 }
1518
1519 return avail_idx != vq->avail_idx;
1520}
1521EXPORT_SYMBOL_GPL(vhost_enable_notify);
1522
1523/* We don't need to be notified again. */
1524void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1525{
1526 int r;
1527
1528 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1529 return;
1530 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1531 if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1532 r = vhost_update_used_flags(vq);
1533 if (r)
1534 vq_err(vq, "Failed to enable notification at %p: %d\n",
1535 &vq->used->flags, r);
1536 }
1537}
1538EXPORT_SYMBOL_GPL(vhost_disable_notify);
1539
1540static int __init vhost_init(void)
1541{
1542 return 0;
1543}
1544
1545static void __exit vhost_exit(void)
1546{
1547}
1548
1549module_init(vhost_init);
1550module_exit(vhost_exit);
1551
1552MODULE_VERSION("0.0.1");
1553MODULE_LICENSE("GPL v2");
1554MODULE_AUTHOR("Michael S. Tsirkin");
1555MODULE_DESCRIPTION("Host kernel accelerator for virtio");