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/poll.h>
22#include <linux/file.h>
23#include <linux/highmem.h>
24#include <linux/slab.h>
25#include <linux/vmalloc.h>
26#include <linux/kthread.h>
27#include <linux/cgroup.h>
28#include <linux/module.h>
29#include <linux/sort.h>
30
31#include "vhost.h"
32
33static ushort max_mem_regions = 64;
34module_param(max_mem_regions, ushort, 0444);
35MODULE_PARM_DESC(max_mem_regions,
36 "Maximum number of memory regions in memory map. (default: 64)");
37
38enum {
39 VHOST_MEMORY_F_LOG = 0x1,
40};
41
42#define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
43#define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
44
45#ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
46static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
47{
48 vq->user_be = !virtio_legacy_is_little_endian();
49}
50
51static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
52{
53 vq->user_be = true;
54}
55
56static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
57{
58 vq->user_be = false;
59}
60
61static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
62{
63 struct vhost_vring_state s;
64
65 if (vq->private_data)
66 return -EBUSY;
67
68 if (copy_from_user(&s, argp, sizeof(s)))
69 return -EFAULT;
70
71 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
72 s.num != VHOST_VRING_BIG_ENDIAN)
73 return -EINVAL;
74
75 if (s.num == VHOST_VRING_BIG_ENDIAN)
76 vhost_enable_cross_endian_big(vq);
77 else
78 vhost_enable_cross_endian_little(vq);
79
80 return 0;
81}
82
83static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
84 int __user *argp)
85{
86 struct vhost_vring_state s = {
87 .index = idx,
88 .num = vq->user_be
89 };
90
91 if (copy_to_user(argp, &s, sizeof(s)))
92 return -EFAULT;
93
94 return 0;
95}
96
97static void vhost_init_is_le(struct vhost_virtqueue *vq)
98{
99 /* Note for legacy virtio: user_be is initialized at reset time
100 * according to the host endianness. If userspace does not set an
101 * explicit endianness, the default behavior is native endian, as
102 * expected by legacy virtio.
103 */
104 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
105}
106#else
107static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
108{
109}
110
111static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
112{
113 return -ENOIOCTLCMD;
114}
115
116static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
117 int __user *argp)
118{
119 return -ENOIOCTLCMD;
120}
121
122static void vhost_init_is_le(struct vhost_virtqueue *vq)
123{
124 if (vhost_has_feature(vq, VIRTIO_F_VERSION_1))
125 vq->is_le = true;
126}
127#endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
128
129static void vhost_reset_is_le(struct vhost_virtqueue *vq)
130{
131 vq->is_le = virtio_legacy_is_little_endian();
132}
133
134static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
135 poll_table *pt)
136{
137 struct vhost_poll *poll;
138
139 poll = container_of(pt, struct vhost_poll, table);
140 poll->wqh = wqh;
141 add_wait_queue(wqh, &poll->wait);
142}
143
144static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
145 void *key)
146{
147 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
148
149 if (!((unsigned long)key & poll->mask))
150 return 0;
151
152 vhost_poll_queue(poll);
153 return 0;
154}
155
156void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
157{
158 INIT_LIST_HEAD(&work->node);
159 work->fn = fn;
160 init_waitqueue_head(&work->done);
161 work->flushing = 0;
162 work->queue_seq = work->done_seq = 0;
163}
164EXPORT_SYMBOL_GPL(vhost_work_init);
165
166/* Init poll structure */
167void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
168 unsigned long mask, struct vhost_dev *dev)
169{
170 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
171 init_poll_funcptr(&poll->table, vhost_poll_func);
172 poll->mask = mask;
173 poll->dev = dev;
174 poll->wqh = NULL;
175
176 vhost_work_init(&poll->work, fn);
177}
178EXPORT_SYMBOL_GPL(vhost_poll_init);
179
180/* Start polling a file. We add ourselves to file's wait queue. The caller must
181 * keep a reference to a file until after vhost_poll_stop is called. */
182int vhost_poll_start(struct vhost_poll *poll, struct file *file)
183{
184 unsigned long mask;
185 int ret = 0;
186
187 if (poll->wqh)
188 return 0;
189
190 mask = file->f_op->poll(file, &poll->table);
191 if (mask)
192 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
193 if (mask & POLLERR) {
194 if (poll->wqh)
195 remove_wait_queue(poll->wqh, &poll->wait);
196 ret = -EINVAL;
197 }
198
199 return ret;
200}
201EXPORT_SYMBOL_GPL(vhost_poll_start);
202
203/* Stop polling a file. After this function returns, it becomes safe to drop the
204 * file reference. You must also flush afterwards. */
205void vhost_poll_stop(struct vhost_poll *poll)
206{
207 if (poll->wqh) {
208 remove_wait_queue(poll->wqh, &poll->wait);
209 poll->wqh = NULL;
210 }
211}
212EXPORT_SYMBOL_GPL(vhost_poll_stop);
213
214static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
215 unsigned seq)
216{
217 int left;
218
219 spin_lock_irq(&dev->work_lock);
220 left = seq - work->done_seq;
221 spin_unlock_irq(&dev->work_lock);
222 return left <= 0;
223}
224
225void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
226{
227 unsigned seq;
228 int flushing;
229
230 spin_lock_irq(&dev->work_lock);
231 seq = work->queue_seq;
232 work->flushing++;
233 spin_unlock_irq(&dev->work_lock);
234 wait_event(work->done, vhost_work_seq_done(dev, work, seq));
235 spin_lock_irq(&dev->work_lock);
236 flushing = --work->flushing;
237 spin_unlock_irq(&dev->work_lock);
238 BUG_ON(flushing < 0);
239}
240EXPORT_SYMBOL_GPL(vhost_work_flush);
241
242/* Flush any work that has been scheduled. When calling this, don't hold any
243 * locks that are also used by the callback. */
244void vhost_poll_flush(struct vhost_poll *poll)
245{
246 vhost_work_flush(poll->dev, &poll->work);
247}
248EXPORT_SYMBOL_GPL(vhost_poll_flush);
249
250void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
251{
252 unsigned long flags;
253
254 spin_lock_irqsave(&dev->work_lock, flags);
255 if (list_empty(&work->node)) {
256 list_add_tail(&work->node, &dev->work_list);
257 work->queue_seq++;
258 spin_unlock_irqrestore(&dev->work_lock, flags);
259 wake_up_process(dev->worker);
260 } else {
261 spin_unlock_irqrestore(&dev->work_lock, flags);
262 }
263}
264EXPORT_SYMBOL_GPL(vhost_work_queue);
265
266/* A lockless hint for busy polling code to exit the loop */
267bool vhost_has_work(struct vhost_dev *dev)
268{
269 return !list_empty(&dev->work_list);
270}
271EXPORT_SYMBOL_GPL(vhost_has_work);
272
273void vhost_poll_queue(struct vhost_poll *poll)
274{
275 vhost_work_queue(poll->dev, &poll->work);
276}
277EXPORT_SYMBOL_GPL(vhost_poll_queue);
278
279static void vhost_vq_reset(struct vhost_dev *dev,
280 struct vhost_virtqueue *vq)
281{
282 vq->num = 1;
283 vq->desc = NULL;
284 vq->avail = NULL;
285 vq->used = NULL;
286 vq->last_avail_idx = 0;
287 vq->avail_idx = 0;
288 vq->last_used_idx = 0;
289 vq->signalled_used = 0;
290 vq->signalled_used_valid = false;
291 vq->used_flags = 0;
292 vq->log_used = false;
293 vq->log_addr = -1ull;
294 vq->private_data = NULL;
295 vq->acked_features = 0;
296 vq->log_base = NULL;
297 vq->error_ctx = NULL;
298 vq->error = NULL;
299 vq->kick = NULL;
300 vq->call_ctx = NULL;
301 vq->call = NULL;
302 vq->log_ctx = NULL;
303 vq->memory = NULL;
304 vhost_reset_is_le(vq);
305 vhost_disable_cross_endian(vq);
306 vq->busyloop_timeout = 0;
307}
308
309static int vhost_worker(void *data)
310{
311 struct vhost_dev *dev = data;
312 struct vhost_work *work = NULL;
313 unsigned uninitialized_var(seq);
314 mm_segment_t oldfs = get_fs();
315
316 set_fs(USER_DS);
317 use_mm(dev->mm);
318
319 for (;;) {
320 /* mb paired w/ kthread_stop */
321 set_current_state(TASK_INTERRUPTIBLE);
322
323 spin_lock_irq(&dev->work_lock);
324 if (work) {
325 work->done_seq = seq;
326 if (work->flushing)
327 wake_up_all(&work->done);
328 }
329
330 if (kthread_should_stop()) {
331 spin_unlock_irq(&dev->work_lock);
332 __set_current_state(TASK_RUNNING);
333 break;
334 }
335 if (!list_empty(&dev->work_list)) {
336 work = list_first_entry(&dev->work_list,
337 struct vhost_work, node);
338 list_del_init(&work->node);
339 seq = work->queue_seq;
340 } else
341 work = NULL;
342 spin_unlock_irq(&dev->work_lock);
343
344 if (work) {
345 __set_current_state(TASK_RUNNING);
346 work->fn(work);
347 if (need_resched())
348 schedule();
349 } else
350 schedule();
351
352 }
353 unuse_mm(dev->mm);
354 set_fs(oldfs);
355 return 0;
356}
357
358static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
359{
360 kfree(vq->indirect);
361 vq->indirect = NULL;
362 kfree(vq->log);
363 vq->log = NULL;
364 kfree(vq->heads);
365 vq->heads = NULL;
366}
367
368/* Helper to allocate iovec buffers for all vqs. */
369static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
370{
371 struct vhost_virtqueue *vq;
372 int i;
373
374 for (i = 0; i < dev->nvqs; ++i) {
375 vq = dev->vqs[i];
376 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
377 GFP_KERNEL);
378 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
379 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
380 if (!vq->indirect || !vq->log || !vq->heads)
381 goto err_nomem;
382 }
383 return 0;
384
385err_nomem:
386 for (; i >= 0; --i)
387 vhost_vq_free_iovecs(dev->vqs[i]);
388 return -ENOMEM;
389}
390
391static void vhost_dev_free_iovecs(struct vhost_dev *dev)
392{
393 int i;
394
395 for (i = 0; i < dev->nvqs; ++i)
396 vhost_vq_free_iovecs(dev->vqs[i]);
397}
398
399void vhost_dev_init(struct vhost_dev *dev,
400 struct vhost_virtqueue **vqs, int nvqs)
401{
402 struct vhost_virtqueue *vq;
403 int i;
404
405 dev->vqs = vqs;
406 dev->nvqs = nvqs;
407 mutex_init(&dev->mutex);
408 dev->log_ctx = NULL;
409 dev->log_file = NULL;
410 dev->memory = NULL;
411 dev->mm = NULL;
412 spin_lock_init(&dev->work_lock);
413 INIT_LIST_HEAD(&dev->work_list);
414 dev->worker = NULL;
415
416 for (i = 0; i < dev->nvqs; ++i) {
417 vq = dev->vqs[i];
418 vq->log = NULL;
419 vq->indirect = NULL;
420 vq->heads = NULL;
421 vq->dev = dev;
422 mutex_init(&vq->mutex);
423 vhost_vq_reset(dev, vq);
424 if (vq->handle_kick)
425 vhost_poll_init(&vq->poll, vq->handle_kick,
426 POLLIN, dev);
427 }
428}
429EXPORT_SYMBOL_GPL(vhost_dev_init);
430
431/* Caller should have device mutex */
432long vhost_dev_check_owner(struct vhost_dev *dev)
433{
434 /* Are you the owner? If not, I don't think you mean to do that */
435 return dev->mm == current->mm ? 0 : -EPERM;
436}
437EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
438
439struct vhost_attach_cgroups_struct {
440 struct vhost_work work;
441 struct task_struct *owner;
442 int ret;
443};
444
445static void vhost_attach_cgroups_work(struct vhost_work *work)
446{
447 struct vhost_attach_cgroups_struct *s;
448
449 s = container_of(work, struct vhost_attach_cgroups_struct, work);
450 s->ret = cgroup_attach_task_all(s->owner, current);
451}
452
453static int vhost_attach_cgroups(struct vhost_dev *dev)
454{
455 struct vhost_attach_cgroups_struct attach;
456
457 attach.owner = current;
458 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
459 vhost_work_queue(dev, &attach.work);
460 vhost_work_flush(dev, &attach.work);
461 return attach.ret;
462}
463
464/* Caller should have device mutex */
465bool vhost_dev_has_owner(struct vhost_dev *dev)
466{
467 return dev->mm;
468}
469EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
470
471/* Caller should have device mutex */
472long vhost_dev_set_owner(struct vhost_dev *dev)
473{
474 struct task_struct *worker;
475 int err;
476
477 /* Is there an owner already? */
478 if (vhost_dev_has_owner(dev)) {
479 err = -EBUSY;
480 goto err_mm;
481 }
482
483 /* No owner, become one */
484 dev->mm = get_task_mm(current);
485 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
486 if (IS_ERR(worker)) {
487 err = PTR_ERR(worker);
488 goto err_worker;
489 }
490
491 dev->worker = worker;
492 wake_up_process(worker); /* avoid contributing to loadavg */
493
494 err = vhost_attach_cgroups(dev);
495 if (err)
496 goto err_cgroup;
497
498 err = vhost_dev_alloc_iovecs(dev);
499 if (err)
500 goto err_cgroup;
501
502 return 0;
503err_cgroup:
504 kthread_stop(worker);
505 dev->worker = NULL;
506err_worker:
507 if (dev->mm)
508 mmput(dev->mm);
509 dev->mm = NULL;
510err_mm:
511 return err;
512}
513EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
514
515struct vhost_memory *vhost_dev_reset_owner_prepare(void)
516{
517 return kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
518}
519EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
520
521/* Caller should have device mutex */
522void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_memory *memory)
523{
524 int i;
525
526 vhost_dev_cleanup(dev, true);
527
528 /* Restore memory to default empty mapping. */
529 memory->nregions = 0;
530 dev->memory = memory;
531 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
532 * VQs aren't running.
533 */
534 for (i = 0; i < dev->nvqs; ++i)
535 dev->vqs[i]->memory = memory;
536}
537EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
538
539void vhost_dev_stop(struct vhost_dev *dev)
540{
541 int i;
542
543 for (i = 0; i < dev->nvqs; ++i) {
544 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
545 vhost_poll_stop(&dev->vqs[i]->poll);
546 vhost_poll_flush(&dev->vqs[i]->poll);
547 }
548 }
549}
550EXPORT_SYMBOL_GPL(vhost_dev_stop);
551
552/* Caller should have device mutex if and only if locked is set */
553void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
554{
555 int i;
556
557 for (i = 0; i < dev->nvqs; ++i) {
558 if (dev->vqs[i]->error_ctx)
559 eventfd_ctx_put(dev->vqs[i]->error_ctx);
560 if (dev->vqs[i]->error)
561 fput(dev->vqs[i]->error);
562 if (dev->vqs[i]->kick)
563 fput(dev->vqs[i]->kick);
564 if (dev->vqs[i]->call_ctx)
565 eventfd_ctx_put(dev->vqs[i]->call_ctx);
566 if (dev->vqs[i]->call)
567 fput(dev->vqs[i]->call);
568 vhost_vq_reset(dev, dev->vqs[i]);
569 }
570 vhost_dev_free_iovecs(dev);
571 if (dev->log_ctx)
572 eventfd_ctx_put(dev->log_ctx);
573 dev->log_ctx = NULL;
574 if (dev->log_file)
575 fput(dev->log_file);
576 dev->log_file = NULL;
577 /* No one will access memory at this point */
578 kvfree(dev->memory);
579 dev->memory = NULL;
580 WARN_ON(!list_empty(&dev->work_list));
581 if (dev->worker) {
582 kthread_stop(dev->worker);
583 dev->worker = NULL;
584 }
585 if (dev->mm)
586 mmput(dev->mm);
587 dev->mm = NULL;
588}
589EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
590
591static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
592{
593 u64 a = addr / VHOST_PAGE_SIZE / 8;
594
595 /* Make sure 64 bit math will not overflow. */
596 if (a > ULONG_MAX - (unsigned long)log_base ||
597 a + (unsigned long)log_base > ULONG_MAX)
598 return 0;
599
600 return access_ok(VERIFY_WRITE, log_base + a,
601 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
602}
603
604/* Caller should have vq mutex and device mutex. */
605static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
606 int log_all)
607{
608 int i;
609
610 if (!mem)
611 return 0;
612
613 for (i = 0; i < mem->nregions; ++i) {
614 struct vhost_memory_region *m = mem->regions + i;
615 unsigned long a = m->userspace_addr;
616 if (m->memory_size > ULONG_MAX)
617 return 0;
618 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
619 m->memory_size))
620 return 0;
621 else if (log_all && !log_access_ok(log_base,
622 m->guest_phys_addr,
623 m->memory_size))
624 return 0;
625 }
626 return 1;
627}
628
629/* Can we switch to this memory table? */
630/* Caller should have device mutex but not vq mutex */
631static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
632 int log_all)
633{
634 int i;
635
636 for (i = 0; i < d->nvqs; ++i) {
637 int ok;
638 bool log;
639
640 mutex_lock(&d->vqs[i]->mutex);
641 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
642 /* If ring is inactive, will check when it's enabled. */
643 if (d->vqs[i]->private_data)
644 ok = vq_memory_access_ok(d->vqs[i]->log_base, mem, log);
645 else
646 ok = 1;
647 mutex_unlock(&d->vqs[i]->mutex);
648 if (!ok)
649 return 0;
650 }
651 return 1;
652}
653
654static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
655 struct vring_desc __user *desc,
656 struct vring_avail __user *avail,
657 struct vring_used __user *used)
658{
659 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
660 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
661 access_ok(VERIFY_READ, avail,
662 sizeof *avail + num * sizeof *avail->ring + s) &&
663 access_ok(VERIFY_WRITE, used,
664 sizeof *used + num * sizeof *used->ring + s);
665}
666
667/* Can we log writes? */
668/* Caller should have device mutex but not vq mutex */
669int vhost_log_access_ok(struct vhost_dev *dev)
670{
671 return memory_access_ok(dev, dev->memory, 1);
672}
673EXPORT_SYMBOL_GPL(vhost_log_access_ok);
674
675/* Verify access for write logging. */
676/* Caller should have vq mutex and device mutex */
677static int vq_log_access_ok(struct vhost_virtqueue *vq,
678 void __user *log_base)
679{
680 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
681
682 return vq_memory_access_ok(log_base, vq->memory,
683 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
684 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
685 sizeof *vq->used +
686 vq->num * sizeof *vq->used->ring + s));
687}
688
689/* Can we start vq? */
690/* Caller should have vq mutex and device mutex */
691int vhost_vq_access_ok(struct vhost_virtqueue *vq)
692{
693 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
694 vq_log_access_ok(vq, vq->log_base);
695}
696EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
697
698static int vhost_memory_reg_sort_cmp(const void *p1, const void *p2)
699{
700 const struct vhost_memory_region *r1 = p1, *r2 = p2;
701 if (r1->guest_phys_addr < r2->guest_phys_addr)
702 return 1;
703 if (r1->guest_phys_addr > r2->guest_phys_addr)
704 return -1;
705 return 0;
706}
707
708static void *vhost_kvzalloc(unsigned long size)
709{
710 void *n = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
711
712 if (!n)
713 n = vzalloc(size);
714 return n;
715}
716
717static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
718{
719 struct vhost_memory mem, *newmem, *oldmem;
720 unsigned long size = offsetof(struct vhost_memory, regions);
721 int i;
722
723 if (copy_from_user(&mem, m, size))
724 return -EFAULT;
725 if (mem.padding)
726 return -EOPNOTSUPP;
727 if (mem.nregions > max_mem_regions)
728 return -E2BIG;
729 newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
730 if (!newmem)
731 return -ENOMEM;
732
733 memcpy(newmem, &mem, size);
734 if (copy_from_user(newmem->regions, m->regions,
735 mem.nregions * sizeof *m->regions)) {
736 kvfree(newmem);
737 return -EFAULT;
738 }
739 sort(newmem->regions, newmem->nregions, sizeof(*newmem->regions),
740 vhost_memory_reg_sort_cmp, NULL);
741
742 if (!memory_access_ok(d, newmem, 0)) {
743 kvfree(newmem);
744 return -EFAULT;
745 }
746 oldmem = d->memory;
747 d->memory = newmem;
748
749 /* All memory accesses are done under some VQ mutex. */
750 for (i = 0; i < d->nvqs; ++i) {
751 mutex_lock(&d->vqs[i]->mutex);
752 d->vqs[i]->memory = newmem;
753 mutex_unlock(&d->vqs[i]->mutex);
754 }
755 kvfree(oldmem);
756 return 0;
757}
758
759long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
760{
761 struct file *eventfp, *filep = NULL;
762 bool pollstart = false, pollstop = false;
763 struct eventfd_ctx *ctx = NULL;
764 u32 __user *idxp = argp;
765 struct vhost_virtqueue *vq;
766 struct vhost_vring_state s;
767 struct vhost_vring_file f;
768 struct vhost_vring_addr a;
769 u32 idx;
770 long r;
771
772 r = get_user(idx, idxp);
773 if (r < 0)
774 return r;
775 if (idx >= d->nvqs)
776 return -ENOBUFS;
777
778 vq = d->vqs[idx];
779
780 mutex_lock(&vq->mutex);
781
782 switch (ioctl) {
783 case VHOST_SET_VRING_NUM:
784 /* Resizing ring with an active backend?
785 * You don't want to do that. */
786 if (vq->private_data) {
787 r = -EBUSY;
788 break;
789 }
790 if (copy_from_user(&s, argp, sizeof s)) {
791 r = -EFAULT;
792 break;
793 }
794 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
795 r = -EINVAL;
796 break;
797 }
798 vq->num = s.num;
799 break;
800 case VHOST_SET_VRING_BASE:
801 /* Moving base with an active backend?
802 * You don't want to do that. */
803 if (vq->private_data) {
804 r = -EBUSY;
805 break;
806 }
807 if (copy_from_user(&s, argp, sizeof s)) {
808 r = -EFAULT;
809 break;
810 }
811 if (s.num > 0xffff) {
812 r = -EINVAL;
813 break;
814 }
815 vq->last_avail_idx = s.num;
816 /* Forget the cached index value. */
817 vq->avail_idx = vq->last_avail_idx;
818 break;
819 case VHOST_GET_VRING_BASE:
820 s.index = idx;
821 s.num = vq->last_avail_idx;
822 if (copy_to_user(argp, &s, sizeof s))
823 r = -EFAULT;
824 break;
825 case VHOST_SET_VRING_ADDR:
826 if (copy_from_user(&a, argp, sizeof a)) {
827 r = -EFAULT;
828 break;
829 }
830 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
831 r = -EOPNOTSUPP;
832 break;
833 }
834 /* For 32bit, verify that the top 32bits of the user
835 data are set to zero. */
836 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
837 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
838 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
839 r = -EFAULT;
840 break;
841 }
842
843 /* Make sure it's safe to cast pointers to vring types. */
844 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
845 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
846 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
847 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
848 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
849 r = -EINVAL;
850 break;
851 }
852
853 /* We only verify access here if backend is configured.
854 * If it is not, we don't as size might not have been setup.
855 * We will verify when backend is configured. */
856 if (vq->private_data) {
857 if (!vq_access_ok(vq, vq->num,
858 (void __user *)(unsigned long)a.desc_user_addr,
859 (void __user *)(unsigned long)a.avail_user_addr,
860 (void __user *)(unsigned long)a.used_user_addr)) {
861 r = -EINVAL;
862 break;
863 }
864
865 /* Also validate log access for used ring if enabled. */
866 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
867 !log_access_ok(vq->log_base, a.log_guest_addr,
868 sizeof *vq->used +
869 vq->num * sizeof *vq->used->ring)) {
870 r = -EINVAL;
871 break;
872 }
873 }
874
875 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
876 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
877 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
878 vq->log_addr = a.log_guest_addr;
879 vq->used = (void __user *)(unsigned long)a.used_user_addr;
880 break;
881 case VHOST_SET_VRING_KICK:
882 if (copy_from_user(&f, argp, sizeof f)) {
883 r = -EFAULT;
884 break;
885 }
886 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
887 if (IS_ERR(eventfp)) {
888 r = PTR_ERR(eventfp);
889 break;
890 }
891 if (eventfp != vq->kick) {
892 pollstop = (filep = vq->kick) != NULL;
893 pollstart = (vq->kick = eventfp) != NULL;
894 } else
895 filep = eventfp;
896 break;
897 case VHOST_SET_VRING_CALL:
898 if (copy_from_user(&f, argp, sizeof f)) {
899 r = -EFAULT;
900 break;
901 }
902 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
903 if (IS_ERR(eventfp)) {
904 r = PTR_ERR(eventfp);
905 break;
906 }
907 if (eventfp != vq->call) {
908 filep = vq->call;
909 ctx = vq->call_ctx;
910 vq->call = eventfp;
911 vq->call_ctx = eventfp ?
912 eventfd_ctx_fileget(eventfp) : NULL;
913 } else
914 filep = eventfp;
915 break;
916 case VHOST_SET_VRING_ERR:
917 if (copy_from_user(&f, argp, sizeof f)) {
918 r = -EFAULT;
919 break;
920 }
921 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
922 if (IS_ERR(eventfp)) {
923 r = PTR_ERR(eventfp);
924 break;
925 }
926 if (eventfp != vq->error) {
927 filep = vq->error;
928 vq->error = eventfp;
929 ctx = vq->error_ctx;
930 vq->error_ctx = eventfp ?
931 eventfd_ctx_fileget(eventfp) : NULL;
932 } else
933 filep = eventfp;
934 break;
935 case VHOST_SET_VRING_ENDIAN:
936 r = vhost_set_vring_endian(vq, argp);
937 break;
938 case VHOST_GET_VRING_ENDIAN:
939 r = vhost_get_vring_endian(vq, idx, argp);
940 break;
941 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
942 if (copy_from_user(&s, argp, sizeof(s))) {
943 r = -EFAULT;
944 break;
945 }
946 vq->busyloop_timeout = s.num;
947 break;
948 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
949 s.index = idx;
950 s.num = vq->busyloop_timeout;
951 if (copy_to_user(argp, &s, sizeof(s)))
952 r = -EFAULT;
953 break;
954 default:
955 r = -ENOIOCTLCMD;
956 }
957
958 if (pollstop && vq->handle_kick)
959 vhost_poll_stop(&vq->poll);
960
961 if (ctx)
962 eventfd_ctx_put(ctx);
963 if (filep)
964 fput(filep);
965
966 if (pollstart && vq->handle_kick)
967 r = vhost_poll_start(&vq->poll, vq->kick);
968
969 mutex_unlock(&vq->mutex);
970
971 if (pollstop && vq->handle_kick)
972 vhost_poll_flush(&vq->poll);
973 return r;
974}
975EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
976
977/* Caller must have device mutex */
978long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
979{
980 struct file *eventfp, *filep = NULL;
981 struct eventfd_ctx *ctx = NULL;
982 u64 p;
983 long r;
984 int i, fd;
985
986 /* If you are not the owner, you can become one */
987 if (ioctl == VHOST_SET_OWNER) {
988 r = vhost_dev_set_owner(d);
989 goto done;
990 }
991
992 /* You must be the owner to do anything else */
993 r = vhost_dev_check_owner(d);
994 if (r)
995 goto done;
996
997 switch (ioctl) {
998 case VHOST_SET_MEM_TABLE:
999 r = vhost_set_memory(d, argp);
1000 break;
1001 case VHOST_SET_LOG_BASE:
1002 if (copy_from_user(&p, argp, sizeof p)) {
1003 r = -EFAULT;
1004 break;
1005 }
1006 if ((u64)(unsigned long)p != p) {
1007 r = -EFAULT;
1008 break;
1009 }
1010 for (i = 0; i < d->nvqs; ++i) {
1011 struct vhost_virtqueue *vq;
1012 void __user *base = (void __user *)(unsigned long)p;
1013 vq = d->vqs[i];
1014 mutex_lock(&vq->mutex);
1015 /* If ring is inactive, will check when it's enabled. */
1016 if (vq->private_data && !vq_log_access_ok(vq, base))
1017 r = -EFAULT;
1018 else
1019 vq->log_base = base;
1020 mutex_unlock(&vq->mutex);
1021 }
1022 break;
1023 case VHOST_SET_LOG_FD:
1024 r = get_user(fd, (int __user *)argp);
1025 if (r < 0)
1026 break;
1027 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1028 if (IS_ERR(eventfp)) {
1029 r = PTR_ERR(eventfp);
1030 break;
1031 }
1032 if (eventfp != d->log_file) {
1033 filep = d->log_file;
1034 d->log_file = eventfp;
1035 ctx = d->log_ctx;
1036 d->log_ctx = eventfp ?
1037 eventfd_ctx_fileget(eventfp) : NULL;
1038 } else
1039 filep = eventfp;
1040 for (i = 0; i < d->nvqs; ++i) {
1041 mutex_lock(&d->vqs[i]->mutex);
1042 d->vqs[i]->log_ctx = d->log_ctx;
1043 mutex_unlock(&d->vqs[i]->mutex);
1044 }
1045 if (ctx)
1046 eventfd_ctx_put(ctx);
1047 if (filep)
1048 fput(filep);
1049 break;
1050 default:
1051 r = -ENOIOCTLCMD;
1052 break;
1053 }
1054done:
1055 return r;
1056}
1057EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1058
1059static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
1060 __u64 addr, __u32 len)
1061{
1062 const struct vhost_memory_region *reg;
1063 int start = 0, end = mem->nregions;
1064
1065 while (start < end) {
1066 int slot = start + (end - start) / 2;
1067 reg = mem->regions + slot;
1068 if (addr >= reg->guest_phys_addr)
1069 end = slot;
1070 else
1071 start = slot + 1;
1072 }
1073
1074 reg = mem->regions + start;
1075 if (addr >= reg->guest_phys_addr &&
1076 reg->guest_phys_addr + reg->memory_size > addr)
1077 return reg;
1078 return NULL;
1079}
1080
1081/* TODO: This is really inefficient. We need something like get_user()
1082 * (instruction directly accesses the data, with an exception table entry
1083 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1084 */
1085static int set_bit_to_user(int nr, void __user *addr)
1086{
1087 unsigned long log = (unsigned long)addr;
1088 struct page *page;
1089 void *base;
1090 int bit = nr + (log % PAGE_SIZE) * 8;
1091 int r;
1092
1093 r = get_user_pages_fast(log, 1, 1, &page);
1094 if (r < 0)
1095 return r;
1096 BUG_ON(r != 1);
1097 base = kmap_atomic(page);
1098 set_bit(bit, base);
1099 kunmap_atomic(base);
1100 set_page_dirty_lock(page);
1101 put_page(page);
1102 return 0;
1103}
1104
1105static int log_write(void __user *log_base,
1106 u64 write_address, u64 write_length)
1107{
1108 u64 write_page = write_address / VHOST_PAGE_SIZE;
1109 int r;
1110
1111 if (!write_length)
1112 return 0;
1113 write_length += write_address % VHOST_PAGE_SIZE;
1114 for (;;) {
1115 u64 base = (u64)(unsigned long)log_base;
1116 u64 log = base + write_page / 8;
1117 int bit = write_page % 8;
1118 if ((u64)(unsigned long)log != log)
1119 return -EFAULT;
1120 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1121 if (r < 0)
1122 return r;
1123 if (write_length <= VHOST_PAGE_SIZE)
1124 break;
1125 write_length -= VHOST_PAGE_SIZE;
1126 write_page += 1;
1127 }
1128 return r;
1129}
1130
1131int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1132 unsigned int log_num, u64 len)
1133{
1134 int i, r;
1135
1136 /* Make sure data written is seen before log. */
1137 smp_wmb();
1138 for (i = 0; i < log_num; ++i) {
1139 u64 l = min(log[i].len, len);
1140 r = log_write(vq->log_base, log[i].addr, l);
1141 if (r < 0)
1142 return r;
1143 len -= l;
1144 if (!len) {
1145 if (vq->log_ctx)
1146 eventfd_signal(vq->log_ctx, 1);
1147 return 0;
1148 }
1149 }
1150 /* Length written exceeds what we have stored. This is a bug. */
1151 BUG();
1152 return 0;
1153}
1154EXPORT_SYMBOL_GPL(vhost_log_write);
1155
1156static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1157{
1158 void __user *used;
1159 if (__put_user(cpu_to_vhost16(vq, vq->used_flags), &vq->used->flags) < 0)
1160 return -EFAULT;
1161 if (unlikely(vq->log_used)) {
1162 /* Make sure the flag is seen before log. */
1163 smp_wmb();
1164 /* Log used flag write. */
1165 used = &vq->used->flags;
1166 log_write(vq->log_base, vq->log_addr +
1167 (used - (void __user *)vq->used),
1168 sizeof vq->used->flags);
1169 if (vq->log_ctx)
1170 eventfd_signal(vq->log_ctx, 1);
1171 }
1172 return 0;
1173}
1174
1175static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1176{
1177 if (__put_user(cpu_to_vhost16(vq, vq->avail_idx), vhost_avail_event(vq)))
1178 return -EFAULT;
1179 if (unlikely(vq->log_used)) {
1180 void __user *used;
1181 /* Make sure the event is seen before log. */
1182 smp_wmb();
1183 /* Log avail event write */
1184 used = vhost_avail_event(vq);
1185 log_write(vq->log_base, vq->log_addr +
1186 (used - (void __user *)vq->used),
1187 sizeof *vhost_avail_event(vq));
1188 if (vq->log_ctx)
1189 eventfd_signal(vq->log_ctx, 1);
1190 }
1191 return 0;
1192}
1193
1194int vhost_vq_init_access(struct vhost_virtqueue *vq)
1195{
1196 __virtio16 last_used_idx;
1197 int r;
1198 bool is_le = vq->is_le;
1199
1200 if (!vq->private_data) {
1201 vhost_reset_is_le(vq);
1202 return 0;
1203 }
1204
1205 vhost_init_is_le(vq);
1206
1207 r = vhost_update_used_flags(vq);
1208 if (r)
1209 goto err;
1210 vq->signalled_used_valid = false;
1211 if (!access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1212 r = -EFAULT;
1213 goto err;
1214 }
1215 r = __get_user(last_used_idx, &vq->used->idx);
1216 if (r)
1217 goto err;
1218 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1219 return 0;
1220err:
1221 vq->is_le = is_le;
1222 return r;
1223}
1224EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1225
1226static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1227 struct iovec iov[], int iov_size)
1228{
1229 const struct vhost_memory_region *reg;
1230 struct vhost_memory *mem;
1231 struct iovec *_iov;
1232 u64 s = 0;
1233 int ret = 0;
1234
1235 mem = vq->memory;
1236 while ((u64)len > s) {
1237 u64 size;
1238 if (unlikely(ret >= iov_size)) {
1239 ret = -ENOBUFS;
1240 break;
1241 }
1242 reg = find_region(mem, addr, len);
1243 if (unlikely(!reg)) {
1244 ret = -EFAULT;
1245 break;
1246 }
1247 _iov = iov + ret;
1248 size = reg->memory_size - addr + reg->guest_phys_addr;
1249 _iov->iov_len = min((u64)len - s, size);
1250 _iov->iov_base = (void __user *)(unsigned long)
1251 (reg->userspace_addr + addr - reg->guest_phys_addr);
1252 s += size;
1253 addr += size;
1254 ++ret;
1255 }
1256
1257 return ret;
1258}
1259
1260/* Each buffer in the virtqueues is actually a chain of descriptors. This
1261 * function returns the next descriptor in the chain,
1262 * or -1U if we're at the end. */
1263static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1264{
1265 unsigned int next;
1266
1267 /* If this descriptor says it doesn't chain, we're done. */
1268 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1269 return -1U;
1270
1271 /* Check they're not leading us off end of descriptors. */
1272 next = vhost16_to_cpu(vq, desc->next);
1273 /* Make sure compiler knows to grab that: we don't want it changing! */
1274 /* We will use the result as an index in an array, so most
1275 * architectures only need a compiler barrier here. */
1276 read_barrier_depends();
1277
1278 return next;
1279}
1280
1281static int get_indirect(struct vhost_virtqueue *vq,
1282 struct iovec iov[], unsigned int iov_size,
1283 unsigned int *out_num, unsigned int *in_num,
1284 struct vhost_log *log, unsigned int *log_num,
1285 struct vring_desc *indirect)
1286{
1287 struct vring_desc desc;
1288 unsigned int i = 0, count, found = 0;
1289 u32 len = vhost32_to_cpu(vq, indirect->len);
1290 struct iov_iter from;
1291 int ret;
1292
1293 /* Sanity check */
1294 if (unlikely(len % sizeof desc)) {
1295 vq_err(vq, "Invalid length in indirect descriptor: "
1296 "len 0x%llx not multiple of 0x%zx\n",
1297 (unsigned long long)len,
1298 sizeof desc);
1299 return -EINVAL;
1300 }
1301
1302 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1303 UIO_MAXIOV);
1304 if (unlikely(ret < 0)) {
1305 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1306 return ret;
1307 }
1308 iov_iter_init(&from, READ, vq->indirect, ret, len);
1309
1310 /* We will use the result as an address to read from, so most
1311 * architectures only need a compiler barrier here. */
1312 read_barrier_depends();
1313
1314 count = len / sizeof desc;
1315 /* Buffers are chained via a 16 bit next field, so
1316 * we can have at most 2^16 of these. */
1317 if (unlikely(count > USHRT_MAX + 1)) {
1318 vq_err(vq, "Indirect buffer length too big: %d\n",
1319 indirect->len);
1320 return -E2BIG;
1321 }
1322
1323 do {
1324 unsigned iov_count = *in_num + *out_num;
1325 if (unlikely(++found > count)) {
1326 vq_err(vq, "Loop detected: last one at %u "
1327 "indirect size %u\n",
1328 i, count);
1329 return -EINVAL;
1330 }
1331 if (unlikely(copy_from_iter(&desc, sizeof(desc), &from) !=
1332 sizeof(desc))) {
1333 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1334 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1335 return -EINVAL;
1336 }
1337 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1338 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1339 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1340 return -EINVAL;
1341 }
1342
1343 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1344 vhost32_to_cpu(vq, desc.len), iov + iov_count,
1345 iov_size - iov_count);
1346 if (unlikely(ret < 0)) {
1347 vq_err(vq, "Translation failure %d indirect idx %d\n",
1348 ret, i);
1349 return ret;
1350 }
1351 /* If this is an input descriptor, increment that count. */
1352 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1353 *in_num += ret;
1354 if (unlikely(log)) {
1355 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1356 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1357 ++*log_num;
1358 }
1359 } else {
1360 /* If it's an output descriptor, they're all supposed
1361 * to come before any input descriptors. */
1362 if (unlikely(*in_num)) {
1363 vq_err(vq, "Indirect descriptor "
1364 "has out after in: idx %d\n", i);
1365 return -EINVAL;
1366 }
1367 *out_num += ret;
1368 }
1369 } while ((i = next_desc(vq, &desc)) != -1);
1370 return 0;
1371}
1372
1373/* This looks in the virtqueue and for the first available buffer, and converts
1374 * it to an iovec for convenient access. Since descriptors consist of some
1375 * number of output then some number of input descriptors, it's actually two
1376 * iovecs, but we pack them into one and note how many of each there were.
1377 *
1378 * This function returns the descriptor number found, or vq->num (which is
1379 * never a valid descriptor number) if none was found. A negative code is
1380 * returned on error. */
1381int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1382 struct iovec iov[], unsigned int iov_size,
1383 unsigned int *out_num, unsigned int *in_num,
1384 struct vhost_log *log, unsigned int *log_num)
1385{
1386 struct vring_desc desc;
1387 unsigned int i, head, found = 0;
1388 u16 last_avail_idx;
1389 __virtio16 avail_idx;
1390 __virtio16 ring_head;
1391 int ret;
1392
1393 /* Check it isn't doing very strange things with descriptor numbers. */
1394 last_avail_idx = vq->last_avail_idx;
1395 if (unlikely(__get_user(avail_idx, &vq->avail->idx))) {
1396 vq_err(vq, "Failed to access avail idx at %p\n",
1397 &vq->avail->idx);
1398 return -EFAULT;
1399 }
1400 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1401
1402 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1403 vq_err(vq, "Guest moved used index from %u to %u",
1404 last_avail_idx, vq->avail_idx);
1405 return -EFAULT;
1406 }
1407
1408 /* If there's nothing new since last we looked, return invalid. */
1409 if (vq->avail_idx == last_avail_idx)
1410 return vq->num;
1411
1412 /* Only get avail ring entries after they have been exposed by guest. */
1413 smp_rmb();
1414
1415 /* Grab the next descriptor number they're advertising, and increment
1416 * the index we've seen. */
1417 if (unlikely(__get_user(ring_head,
1418 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
1419 vq_err(vq, "Failed to read head: idx %d address %p\n",
1420 last_avail_idx,
1421 &vq->avail->ring[last_avail_idx % vq->num]);
1422 return -EFAULT;
1423 }
1424
1425 head = vhost16_to_cpu(vq, ring_head);
1426
1427 /* If their number is silly, that's an error. */
1428 if (unlikely(head >= vq->num)) {
1429 vq_err(vq, "Guest says index %u > %u is available",
1430 head, vq->num);
1431 return -EINVAL;
1432 }
1433
1434 /* When we start there are none of either input nor output. */
1435 *out_num = *in_num = 0;
1436 if (unlikely(log))
1437 *log_num = 0;
1438
1439 i = head;
1440 do {
1441 unsigned iov_count = *in_num + *out_num;
1442 if (unlikely(i >= vq->num)) {
1443 vq_err(vq, "Desc index is %u > %u, head = %u",
1444 i, vq->num, head);
1445 return -EINVAL;
1446 }
1447 if (unlikely(++found > vq->num)) {
1448 vq_err(vq, "Loop detected: last one at %u "
1449 "vq size %u head %u\n",
1450 i, vq->num, head);
1451 return -EINVAL;
1452 }
1453 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1454 if (unlikely(ret)) {
1455 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1456 i, vq->desc + i);
1457 return -EFAULT;
1458 }
1459 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
1460 ret = get_indirect(vq, iov, iov_size,
1461 out_num, in_num,
1462 log, log_num, &desc);
1463 if (unlikely(ret < 0)) {
1464 vq_err(vq, "Failure detected "
1465 "in indirect descriptor at idx %d\n", i);
1466 return ret;
1467 }
1468 continue;
1469 }
1470
1471 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1472 vhost32_to_cpu(vq, desc.len), iov + iov_count,
1473 iov_size - iov_count);
1474 if (unlikely(ret < 0)) {
1475 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1476 ret, i);
1477 return ret;
1478 }
1479 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1480 /* If this is an input descriptor,
1481 * increment that count. */
1482 *in_num += ret;
1483 if (unlikely(log)) {
1484 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1485 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1486 ++*log_num;
1487 }
1488 } else {
1489 /* If it's an output descriptor, they're all supposed
1490 * to come before any input descriptors. */
1491 if (unlikely(*in_num)) {
1492 vq_err(vq, "Descriptor has out after in: "
1493 "idx %d\n", i);
1494 return -EINVAL;
1495 }
1496 *out_num += ret;
1497 }
1498 } while ((i = next_desc(vq, &desc)) != -1);
1499
1500 /* On success, increment avail index. */
1501 vq->last_avail_idx++;
1502
1503 /* Assume notifications from guest are disabled at this point,
1504 * if they aren't we would need to update avail_event index. */
1505 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1506 return head;
1507}
1508EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1509
1510/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1511void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1512{
1513 vq->last_avail_idx -= n;
1514}
1515EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1516
1517/* After we've used one of their buffers, we tell them about it. We'll then
1518 * want to notify the guest, using eventfd. */
1519int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1520{
1521 struct vring_used_elem heads = {
1522 cpu_to_vhost32(vq, head),
1523 cpu_to_vhost32(vq, len)
1524 };
1525
1526 return vhost_add_used_n(vq, &heads, 1);
1527}
1528EXPORT_SYMBOL_GPL(vhost_add_used);
1529
1530static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1531 struct vring_used_elem *heads,
1532 unsigned count)
1533{
1534 struct vring_used_elem __user *used;
1535 u16 old, new;
1536 int start;
1537
1538 start = vq->last_used_idx & (vq->num - 1);
1539 used = vq->used->ring + start;
1540 if (count == 1) {
1541 if (__put_user(heads[0].id, &used->id)) {
1542 vq_err(vq, "Failed to write used id");
1543 return -EFAULT;
1544 }
1545 if (__put_user(heads[0].len, &used->len)) {
1546 vq_err(vq, "Failed to write used len");
1547 return -EFAULT;
1548 }
1549 } else if (__copy_to_user(used, heads, count * sizeof *used)) {
1550 vq_err(vq, "Failed to write used");
1551 return -EFAULT;
1552 }
1553 if (unlikely(vq->log_used)) {
1554 /* Make sure data is seen before log. */
1555 smp_wmb();
1556 /* Log used ring entry write. */
1557 log_write(vq->log_base,
1558 vq->log_addr +
1559 ((void __user *)used - (void __user *)vq->used),
1560 count * sizeof *used);
1561 }
1562 old = vq->last_used_idx;
1563 new = (vq->last_used_idx += count);
1564 /* If the driver never bothers to signal in a very long while,
1565 * used index might wrap around. If that happens, invalidate
1566 * signalled_used index we stored. TODO: make sure driver
1567 * signals at least once in 2^16 and remove this. */
1568 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1569 vq->signalled_used_valid = false;
1570 return 0;
1571}
1572
1573/* After we've used one of their buffers, we tell them about it. We'll then
1574 * want to notify the guest, using eventfd. */
1575int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1576 unsigned count)
1577{
1578 int start, n, r;
1579
1580 start = vq->last_used_idx & (vq->num - 1);
1581 n = vq->num - start;
1582 if (n < count) {
1583 r = __vhost_add_used_n(vq, heads, n);
1584 if (r < 0)
1585 return r;
1586 heads += n;
1587 count -= n;
1588 }
1589 r = __vhost_add_used_n(vq, heads, count);
1590
1591 /* Make sure buffer is written before we update index. */
1592 smp_wmb();
1593 if (__put_user(cpu_to_vhost16(vq, vq->last_used_idx), &vq->used->idx)) {
1594 vq_err(vq, "Failed to increment used idx");
1595 return -EFAULT;
1596 }
1597 if (unlikely(vq->log_used)) {
1598 /* Log used index update. */
1599 log_write(vq->log_base,
1600 vq->log_addr + offsetof(struct vring_used, idx),
1601 sizeof vq->used->idx);
1602 if (vq->log_ctx)
1603 eventfd_signal(vq->log_ctx, 1);
1604 }
1605 return r;
1606}
1607EXPORT_SYMBOL_GPL(vhost_add_used_n);
1608
1609static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1610{
1611 __u16 old, new;
1612 __virtio16 event;
1613 bool v;
1614 /* Flush out used index updates. This is paired
1615 * with the barrier that the Guest executes when enabling
1616 * interrupts. */
1617 smp_mb();
1618
1619 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1620 unlikely(vq->avail_idx == vq->last_avail_idx))
1621 return true;
1622
1623 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1624 __virtio16 flags;
1625 if (__get_user(flags, &vq->avail->flags)) {
1626 vq_err(vq, "Failed to get flags");
1627 return true;
1628 }
1629 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
1630 }
1631 old = vq->signalled_used;
1632 v = vq->signalled_used_valid;
1633 new = vq->signalled_used = vq->last_used_idx;
1634 vq->signalled_used_valid = true;
1635
1636 if (unlikely(!v))
1637 return true;
1638
1639 if (__get_user(event, vhost_used_event(vq))) {
1640 vq_err(vq, "Failed to get used event idx");
1641 return true;
1642 }
1643 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
1644}
1645
1646/* This actually signals the guest, using eventfd. */
1647void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1648{
1649 /* Signal the Guest tell them we used something up. */
1650 if (vq->call_ctx && vhost_notify(dev, vq))
1651 eventfd_signal(vq->call_ctx, 1);
1652}
1653EXPORT_SYMBOL_GPL(vhost_signal);
1654
1655/* And here's the combo meal deal. Supersize me! */
1656void vhost_add_used_and_signal(struct vhost_dev *dev,
1657 struct vhost_virtqueue *vq,
1658 unsigned int head, int len)
1659{
1660 vhost_add_used(vq, head, len);
1661 vhost_signal(dev, vq);
1662}
1663EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1664
1665/* multi-buffer version of vhost_add_used_and_signal */
1666void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1667 struct vhost_virtqueue *vq,
1668 struct vring_used_elem *heads, unsigned count)
1669{
1670 vhost_add_used_n(vq, heads, count);
1671 vhost_signal(dev, vq);
1672}
1673EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1674
1675/* return true if we're sure that avaiable ring is empty */
1676bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1677{
1678 __virtio16 avail_idx;
1679 int r;
1680
1681 r = __get_user(avail_idx, &vq->avail->idx);
1682 if (r)
1683 return false;
1684
1685 return vhost16_to_cpu(vq, avail_idx) == vq->avail_idx;
1686}
1687EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
1688
1689/* OK, now we need to know about added descriptors. */
1690bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1691{
1692 __virtio16 avail_idx;
1693 int r;
1694
1695 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1696 return false;
1697 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1698 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1699 r = vhost_update_used_flags(vq);
1700 if (r) {
1701 vq_err(vq, "Failed to enable notification at %p: %d\n",
1702 &vq->used->flags, r);
1703 return false;
1704 }
1705 } else {
1706 r = vhost_update_avail_event(vq, vq->avail_idx);
1707 if (r) {
1708 vq_err(vq, "Failed to update avail event index at %p: %d\n",
1709 vhost_avail_event(vq), r);
1710 return false;
1711 }
1712 }
1713 /* They could have slipped one in as we were doing that: make
1714 * sure it's written, then check again. */
1715 smp_mb();
1716 r = __get_user(avail_idx, &vq->avail->idx);
1717 if (r) {
1718 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1719 &vq->avail->idx, r);
1720 return false;
1721 }
1722
1723 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
1724}
1725EXPORT_SYMBOL_GPL(vhost_enable_notify);
1726
1727/* We don't need to be notified again. */
1728void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1729{
1730 int r;
1731
1732 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1733 return;
1734 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1735 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1736 r = vhost_update_used_flags(vq);
1737 if (r)
1738 vq_err(vq, "Failed to enable notification at %p: %d\n",
1739 &vq->used->flags, r);
1740 }
1741}
1742EXPORT_SYMBOL_GPL(vhost_disable_notify);
1743
1744static int __init vhost_init(void)
1745{
1746 return 0;
1747}
1748
1749static void __exit vhost_exit(void)
1750{
1751}
1752
1753module_init(vhost_init);
1754module_exit(vhost_exit);
1755
1756MODULE_VERSION("0.0.1");
1757MODULE_LICENSE("GPL v2");
1758MODULE_AUTHOR("Michael S. Tsirkin");
1759MODULE_DESCRIPTION("Host kernel accelerator for virtio");