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#include <linux/interval_tree_generic.h>
31
32#include "vhost.h"
33
34static ushort max_mem_regions = 64;
35module_param(max_mem_regions, ushort, 0444);
36MODULE_PARM_DESC(max_mem_regions,
37 "Maximum number of memory regions in memory map. (default: 64)");
38static int max_iotlb_entries = 2048;
39module_param(max_iotlb_entries, int, 0444);
40MODULE_PARM_DESC(max_iotlb_entries,
41 "Maximum number of iotlb entries. (default: 2048)");
42
43enum {
44 VHOST_MEMORY_F_LOG = 0x1,
45};
46
47#define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
48#define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
49
50INTERVAL_TREE_DEFINE(struct vhost_umem_node,
51 rb, __u64, __subtree_last,
52 START, LAST, static inline, vhost_umem_interval_tree);
53
54#ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
55static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
56{
57 vq->user_be = !virtio_legacy_is_little_endian();
58}
59
60static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
61{
62 vq->user_be = true;
63}
64
65static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
66{
67 vq->user_be = false;
68}
69
70static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
71{
72 struct vhost_vring_state s;
73
74 if (vq->private_data)
75 return -EBUSY;
76
77 if (copy_from_user(&s, argp, sizeof(s)))
78 return -EFAULT;
79
80 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
81 s.num != VHOST_VRING_BIG_ENDIAN)
82 return -EINVAL;
83
84 if (s.num == VHOST_VRING_BIG_ENDIAN)
85 vhost_enable_cross_endian_big(vq);
86 else
87 vhost_enable_cross_endian_little(vq);
88
89 return 0;
90}
91
92static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
93 int __user *argp)
94{
95 struct vhost_vring_state s = {
96 .index = idx,
97 .num = vq->user_be
98 };
99
100 if (copy_to_user(argp, &s, sizeof(s)))
101 return -EFAULT;
102
103 return 0;
104}
105
106static void vhost_init_is_le(struct vhost_virtqueue *vq)
107{
108 /* Note for legacy virtio: user_be is initialized at reset time
109 * according to the host endianness. If userspace does not set an
110 * explicit endianness, the default behavior is native endian, as
111 * expected by legacy virtio.
112 */
113 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
114}
115#else
116static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
117{
118}
119
120static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
121{
122 return -ENOIOCTLCMD;
123}
124
125static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
126 int __user *argp)
127{
128 return -ENOIOCTLCMD;
129}
130
131static void vhost_init_is_le(struct vhost_virtqueue *vq)
132{
133 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
134 || virtio_legacy_is_little_endian();
135}
136#endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
137
138static void vhost_reset_is_le(struct vhost_virtqueue *vq)
139{
140 vhost_init_is_le(vq);
141}
142
143struct vhost_flush_struct {
144 struct vhost_work work;
145 struct completion wait_event;
146};
147
148static void vhost_flush_work(struct vhost_work *work)
149{
150 struct vhost_flush_struct *s;
151
152 s = container_of(work, struct vhost_flush_struct, work);
153 complete(&s->wait_event);
154}
155
156static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
157 poll_table *pt)
158{
159 struct vhost_poll *poll;
160
161 poll = container_of(pt, struct vhost_poll, table);
162 poll->wqh = wqh;
163 add_wait_queue(wqh, &poll->wait);
164}
165
166static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
167 void *key)
168{
169 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
170
171 if (!((unsigned long)key & poll->mask))
172 return 0;
173
174 vhost_poll_queue(poll);
175 return 0;
176}
177
178void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
179{
180 clear_bit(VHOST_WORK_QUEUED, &work->flags);
181 work->fn = fn;
182 init_waitqueue_head(&work->done);
183}
184EXPORT_SYMBOL_GPL(vhost_work_init);
185
186/* Init poll structure */
187void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
188 unsigned long mask, struct vhost_dev *dev)
189{
190 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
191 init_poll_funcptr(&poll->table, vhost_poll_func);
192 poll->mask = mask;
193 poll->dev = dev;
194 poll->wqh = NULL;
195
196 vhost_work_init(&poll->work, fn);
197}
198EXPORT_SYMBOL_GPL(vhost_poll_init);
199
200/* Start polling a file. We add ourselves to file's wait queue. The caller must
201 * keep a reference to a file until after vhost_poll_stop is called. */
202int vhost_poll_start(struct vhost_poll *poll, struct file *file)
203{
204 unsigned long mask;
205 int ret = 0;
206
207 if (poll->wqh)
208 return 0;
209
210 mask = file->f_op->poll(file, &poll->table);
211 if (mask)
212 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
213 if (mask & POLLERR) {
214 if (poll->wqh)
215 remove_wait_queue(poll->wqh, &poll->wait);
216 ret = -EINVAL;
217 }
218
219 return ret;
220}
221EXPORT_SYMBOL_GPL(vhost_poll_start);
222
223/* Stop polling a file. After this function returns, it becomes safe to drop the
224 * file reference. You must also flush afterwards. */
225void vhost_poll_stop(struct vhost_poll *poll)
226{
227 if (poll->wqh) {
228 remove_wait_queue(poll->wqh, &poll->wait);
229 poll->wqh = NULL;
230 }
231}
232EXPORT_SYMBOL_GPL(vhost_poll_stop);
233
234void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
235{
236 struct vhost_flush_struct flush;
237
238 if (dev->worker) {
239 init_completion(&flush.wait_event);
240 vhost_work_init(&flush.work, vhost_flush_work);
241
242 vhost_work_queue(dev, &flush.work);
243 wait_for_completion(&flush.wait_event);
244 }
245}
246EXPORT_SYMBOL_GPL(vhost_work_flush);
247
248/* Flush any work that has been scheduled. When calling this, don't hold any
249 * locks that are also used by the callback. */
250void vhost_poll_flush(struct vhost_poll *poll)
251{
252 vhost_work_flush(poll->dev, &poll->work);
253}
254EXPORT_SYMBOL_GPL(vhost_poll_flush);
255
256void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
257{
258 if (!dev->worker)
259 return;
260
261 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
262 /* We can only add the work to the list after we're
263 * sure it was not in the list.
264 * test_and_set_bit() implies a memory barrier.
265 */
266 llist_add(&work->node, &dev->work_list);
267 wake_up_process(dev->worker);
268 }
269}
270EXPORT_SYMBOL_GPL(vhost_work_queue);
271
272/* A lockless hint for busy polling code to exit the loop */
273bool vhost_has_work(struct vhost_dev *dev)
274{
275 return !llist_empty(&dev->work_list);
276}
277EXPORT_SYMBOL_GPL(vhost_has_work);
278
279void vhost_poll_queue(struct vhost_poll *poll)
280{
281 vhost_work_queue(poll->dev, &poll->work);
282}
283EXPORT_SYMBOL_GPL(vhost_poll_queue);
284
285static void vhost_vq_reset(struct vhost_dev *dev,
286 struct vhost_virtqueue *vq)
287{
288 vq->num = 1;
289 vq->desc = NULL;
290 vq->avail = NULL;
291 vq->used = NULL;
292 vq->last_avail_idx = 0;
293 vq->last_used_event = 0;
294 vq->avail_idx = 0;
295 vq->last_used_idx = 0;
296 vq->signalled_used = 0;
297 vq->signalled_used_valid = false;
298 vq->used_flags = 0;
299 vq->log_used = false;
300 vq->log_addr = -1ull;
301 vq->private_data = NULL;
302 vq->acked_features = 0;
303 vq->log_base = NULL;
304 vq->error_ctx = NULL;
305 vq->error = NULL;
306 vq->kick = NULL;
307 vq->call_ctx = NULL;
308 vq->call = NULL;
309 vq->log_ctx = NULL;
310 vhost_reset_is_le(vq);
311 vhost_disable_cross_endian(vq);
312 vq->busyloop_timeout = 0;
313 vq->umem = NULL;
314 vq->iotlb = NULL;
315}
316
317static int vhost_worker(void *data)
318{
319 struct vhost_dev *dev = data;
320 struct vhost_work *work, *work_next;
321 struct llist_node *node;
322 mm_segment_t oldfs = get_fs();
323
324 set_fs(USER_DS);
325 use_mm(dev->mm);
326
327 for (;;) {
328 /* mb paired w/ kthread_stop */
329 set_current_state(TASK_INTERRUPTIBLE);
330
331 if (kthread_should_stop()) {
332 __set_current_state(TASK_RUNNING);
333 break;
334 }
335
336 node = llist_del_all(&dev->work_list);
337 if (!node)
338 schedule();
339
340 node = llist_reverse_order(node);
341 /* make sure flag is seen after deletion */
342 smp_wmb();
343 llist_for_each_entry_safe(work, work_next, node, node) {
344 clear_bit(VHOST_WORK_QUEUED, &work->flags);
345 __set_current_state(TASK_RUNNING);
346 work->fn(work);
347 if (need_resched())
348 schedule();
349 }
350 }
351 unuse_mm(dev->mm);
352 set_fs(oldfs);
353 return 0;
354}
355
356static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
357{
358 kfree(vq->indirect);
359 vq->indirect = NULL;
360 kfree(vq->log);
361 vq->log = NULL;
362 kfree(vq->heads);
363 vq->heads = NULL;
364}
365
366/* Helper to allocate iovec buffers for all vqs. */
367static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
368{
369 struct vhost_virtqueue *vq;
370 int i;
371
372 for (i = 0; i < dev->nvqs; ++i) {
373 vq = dev->vqs[i];
374 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
375 GFP_KERNEL);
376 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
377 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
378 if (!vq->indirect || !vq->log || !vq->heads)
379 goto err_nomem;
380 }
381 return 0;
382
383err_nomem:
384 for (; i >= 0; --i)
385 vhost_vq_free_iovecs(dev->vqs[i]);
386 return -ENOMEM;
387}
388
389static void vhost_dev_free_iovecs(struct vhost_dev *dev)
390{
391 int i;
392
393 for (i = 0; i < dev->nvqs; ++i)
394 vhost_vq_free_iovecs(dev->vqs[i]);
395}
396
397void vhost_dev_init(struct vhost_dev *dev,
398 struct vhost_virtqueue **vqs, int nvqs)
399{
400 struct vhost_virtqueue *vq;
401 int i;
402
403 dev->vqs = vqs;
404 dev->nvqs = nvqs;
405 mutex_init(&dev->mutex);
406 dev->log_ctx = NULL;
407 dev->log_file = NULL;
408 dev->umem = NULL;
409 dev->iotlb = NULL;
410 dev->mm = NULL;
411 dev->worker = NULL;
412 init_llist_head(&dev->work_list);
413 init_waitqueue_head(&dev->wait);
414 INIT_LIST_HEAD(&dev->read_list);
415 INIT_LIST_HEAD(&dev->pending_list);
416 spin_lock_init(&dev->iotlb_lock);
417
418
419 for (i = 0; i < dev->nvqs; ++i) {
420 vq = dev->vqs[i];
421 vq->log = NULL;
422 vq->indirect = NULL;
423 vq->heads = NULL;
424 vq->dev = dev;
425 mutex_init(&vq->mutex);
426 vhost_vq_reset(dev, vq);
427 if (vq->handle_kick)
428 vhost_poll_init(&vq->poll, vq->handle_kick,
429 POLLIN, dev);
430 }
431}
432EXPORT_SYMBOL_GPL(vhost_dev_init);
433
434/* Caller should have device mutex */
435long vhost_dev_check_owner(struct vhost_dev *dev)
436{
437 /* Are you the owner? If not, I don't think you mean to do that */
438 return dev->mm == current->mm ? 0 : -EPERM;
439}
440EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
441
442struct vhost_attach_cgroups_struct {
443 struct vhost_work work;
444 struct task_struct *owner;
445 int ret;
446};
447
448static void vhost_attach_cgroups_work(struct vhost_work *work)
449{
450 struct vhost_attach_cgroups_struct *s;
451
452 s = container_of(work, struct vhost_attach_cgroups_struct, work);
453 s->ret = cgroup_attach_task_all(s->owner, current);
454}
455
456static int vhost_attach_cgroups(struct vhost_dev *dev)
457{
458 struct vhost_attach_cgroups_struct attach;
459
460 attach.owner = current;
461 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
462 vhost_work_queue(dev, &attach.work);
463 vhost_work_flush(dev, &attach.work);
464 return attach.ret;
465}
466
467/* Caller should have device mutex */
468bool vhost_dev_has_owner(struct vhost_dev *dev)
469{
470 return dev->mm;
471}
472EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
473
474/* Caller should have device mutex */
475long vhost_dev_set_owner(struct vhost_dev *dev)
476{
477 struct task_struct *worker;
478 int err;
479
480 /* Is there an owner already? */
481 if (vhost_dev_has_owner(dev)) {
482 err = -EBUSY;
483 goto err_mm;
484 }
485
486 /* No owner, become one */
487 dev->mm = get_task_mm(current);
488 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
489 if (IS_ERR(worker)) {
490 err = PTR_ERR(worker);
491 goto err_worker;
492 }
493
494 dev->worker = worker;
495 wake_up_process(worker); /* avoid contributing to loadavg */
496
497 err = vhost_attach_cgroups(dev);
498 if (err)
499 goto err_cgroup;
500
501 err = vhost_dev_alloc_iovecs(dev);
502 if (err)
503 goto err_cgroup;
504
505 return 0;
506err_cgroup:
507 kthread_stop(worker);
508 dev->worker = NULL;
509err_worker:
510 if (dev->mm)
511 mmput(dev->mm);
512 dev->mm = NULL;
513err_mm:
514 return err;
515}
516EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
517
518static void *vhost_kvzalloc(unsigned long size)
519{
520 void *n = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
521
522 if (!n)
523 n = vzalloc(size);
524 return n;
525}
526
527struct vhost_umem *vhost_dev_reset_owner_prepare(void)
528{
529 return vhost_kvzalloc(sizeof(struct vhost_umem));
530}
531EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
532
533/* Caller should have device mutex */
534void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
535{
536 int i;
537
538 vhost_dev_cleanup(dev, true);
539
540 /* Restore memory to default empty mapping. */
541 INIT_LIST_HEAD(&umem->umem_list);
542 dev->umem = umem;
543 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
544 * VQs aren't running.
545 */
546 for (i = 0; i < dev->nvqs; ++i)
547 dev->vqs[i]->umem = umem;
548}
549EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
550
551void vhost_dev_stop(struct vhost_dev *dev)
552{
553 int i;
554
555 for (i = 0; i < dev->nvqs; ++i) {
556 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
557 vhost_poll_stop(&dev->vqs[i]->poll);
558 vhost_poll_flush(&dev->vqs[i]->poll);
559 }
560 }
561}
562EXPORT_SYMBOL_GPL(vhost_dev_stop);
563
564static void vhost_umem_free(struct vhost_umem *umem,
565 struct vhost_umem_node *node)
566{
567 vhost_umem_interval_tree_remove(node, &umem->umem_tree);
568 list_del(&node->link);
569 kfree(node);
570 umem->numem--;
571}
572
573static void vhost_umem_clean(struct vhost_umem *umem)
574{
575 struct vhost_umem_node *node, *tmp;
576
577 if (!umem)
578 return;
579
580 list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
581 vhost_umem_free(umem, node);
582
583 kvfree(umem);
584}
585
586static void vhost_clear_msg(struct vhost_dev *dev)
587{
588 struct vhost_msg_node *node, *n;
589
590 spin_lock(&dev->iotlb_lock);
591
592 list_for_each_entry_safe(node, n, &dev->read_list, node) {
593 list_del(&node->node);
594 kfree(node);
595 }
596
597 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
598 list_del(&node->node);
599 kfree(node);
600 }
601
602 spin_unlock(&dev->iotlb_lock);
603}
604
605/* Caller should have device mutex if and only if locked is set */
606void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
607{
608 int i;
609
610 for (i = 0; i < dev->nvqs; ++i) {
611 if (dev->vqs[i]->error_ctx)
612 eventfd_ctx_put(dev->vqs[i]->error_ctx);
613 if (dev->vqs[i]->error)
614 fput(dev->vqs[i]->error);
615 if (dev->vqs[i]->kick)
616 fput(dev->vqs[i]->kick);
617 if (dev->vqs[i]->call_ctx)
618 eventfd_ctx_put(dev->vqs[i]->call_ctx);
619 if (dev->vqs[i]->call)
620 fput(dev->vqs[i]->call);
621 vhost_vq_reset(dev, dev->vqs[i]);
622 }
623 vhost_dev_free_iovecs(dev);
624 if (dev->log_ctx)
625 eventfd_ctx_put(dev->log_ctx);
626 dev->log_ctx = NULL;
627 if (dev->log_file)
628 fput(dev->log_file);
629 dev->log_file = NULL;
630 /* No one will access memory at this point */
631 vhost_umem_clean(dev->umem);
632 dev->umem = NULL;
633 vhost_umem_clean(dev->iotlb);
634 dev->iotlb = NULL;
635 vhost_clear_msg(dev);
636 wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
637 WARN_ON(!llist_empty(&dev->work_list));
638 if (dev->worker) {
639 kthread_stop(dev->worker);
640 dev->worker = NULL;
641 }
642 if (dev->mm)
643 mmput(dev->mm);
644 dev->mm = NULL;
645}
646EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
647
648static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
649{
650 u64 a = addr / VHOST_PAGE_SIZE / 8;
651
652 /* Make sure 64 bit math will not overflow. */
653 if (a > ULONG_MAX - (unsigned long)log_base ||
654 a + (unsigned long)log_base > ULONG_MAX)
655 return 0;
656
657 return access_ok(VERIFY_WRITE, log_base + a,
658 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
659}
660
661static bool vhost_overflow(u64 uaddr, u64 size)
662{
663 /* Make sure 64 bit math will not overflow. */
664 return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
665}
666
667/* Caller should have vq mutex and device mutex. */
668static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
669 int log_all)
670{
671 struct vhost_umem_node *node;
672
673 if (!umem)
674 return 0;
675
676 list_for_each_entry(node, &umem->umem_list, link) {
677 unsigned long a = node->userspace_addr;
678
679 if (vhost_overflow(node->userspace_addr, node->size))
680 return 0;
681
682
683 if (!access_ok(VERIFY_WRITE, (void __user *)a,
684 node->size))
685 return 0;
686 else if (log_all && !log_access_ok(log_base,
687 node->start,
688 node->size))
689 return 0;
690 }
691 return 1;
692}
693
694/* Can we switch to this memory table? */
695/* Caller should have device mutex but not vq mutex */
696static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
697 int log_all)
698{
699 int i;
700
701 for (i = 0; i < d->nvqs; ++i) {
702 int ok;
703 bool log;
704
705 mutex_lock(&d->vqs[i]->mutex);
706 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
707 /* If ring is inactive, will check when it's enabled. */
708 if (d->vqs[i]->private_data)
709 ok = vq_memory_access_ok(d->vqs[i]->log_base,
710 umem, log);
711 else
712 ok = 1;
713 mutex_unlock(&d->vqs[i]->mutex);
714 if (!ok)
715 return 0;
716 }
717 return 1;
718}
719
720static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
721 struct iovec iov[], int iov_size, int access);
722
723static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
724 const void *from, unsigned size)
725{
726 int ret;
727
728 if (!vq->iotlb)
729 return __copy_to_user(to, from, size);
730 else {
731 /* This function should be called after iotlb
732 * prefetch, which means we're sure that all vq
733 * could be access through iotlb. So -EAGAIN should
734 * not happen in this case.
735 */
736 /* TODO: more fast path */
737 struct iov_iter t;
738 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
739 ARRAY_SIZE(vq->iotlb_iov),
740 VHOST_ACCESS_WO);
741 if (ret < 0)
742 goto out;
743 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
744 ret = copy_to_iter(from, size, &t);
745 if (ret == size)
746 ret = 0;
747 }
748out:
749 return ret;
750}
751
752static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
753 void __user *from, unsigned size)
754{
755 int ret;
756
757 if (!vq->iotlb)
758 return __copy_from_user(to, from, size);
759 else {
760 /* This function should be called after iotlb
761 * prefetch, which means we're sure that vq
762 * could be access through iotlb. So -EAGAIN should
763 * not happen in this case.
764 */
765 /* TODO: more fast path */
766 struct iov_iter f;
767 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
768 ARRAY_SIZE(vq->iotlb_iov),
769 VHOST_ACCESS_RO);
770 if (ret < 0) {
771 vq_err(vq, "IOTLB translation failure: uaddr "
772 "%p size 0x%llx\n", from,
773 (unsigned long long) size);
774 goto out;
775 }
776 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
777 ret = copy_from_iter(to, size, &f);
778 if (ret == size)
779 ret = 0;
780 }
781
782out:
783 return ret;
784}
785
786static void __user *__vhost_get_user(struct vhost_virtqueue *vq,
787 void __user *addr, unsigned size)
788{
789 int ret;
790
791 /* This function should be called after iotlb
792 * prefetch, which means we're sure that vq
793 * could be access through iotlb. So -EAGAIN should
794 * not happen in this case.
795 */
796 /* TODO: more fast path */
797 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
798 ARRAY_SIZE(vq->iotlb_iov),
799 VHOST_ACCESS_RO);
800 if (ret < 0) {
801 vq_err(vq, "IOTLB translation failure: uaddr "
802 "%p size 0x%llx\n", addr,
803 (unsigned long long) size);
804 return NULL;
805 }
806
807 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
808 vq_err(vq, "Non atomic userspace memory access: uaddr "
809 "%p size 0x%llx\n", addr,
810 (unsigned long long) size);
811 return NULL;
812 }
813
814 return vq->iotlb_iov[0].iov_base;
815}
816
817#define vhost_put_user(vq, x, ptr) \
818({ \
819 int ret = -EFAULT; \
820 if (!vq->iotlb) { \
821 ret = __put_user(x, ptr); \
822 } else { \
823 __typeof__(ptr) to = \
824 (__typeof__(ptr)) __vhost_get_user(vq, ptr, sizeof(*ptr)); \
825 if (to != NULL) \
826 ret = __put_user(x, to); \
827 else \
828 ret = -EFAULT; \
829 } \
830 ret; \
831})
832
833#define vhost_get_user(vq, x, ptr) \
834({ \
835 int ret; \
836 if (!vq->iotlb) { \
837 ret = __get_user(x, ptr); \
838 } else { \
839 __typeof__(ptr) from = \
840 (__typeof__(ptr)) __vhost_get_user(vq, ptr, sizeof(*ptr)); \
841 if (from != NULL) \
842 ret = __get_user(x, from); \
843 else \
844 ret = -EFAULT; \
845 } \
846 ret; \
847})
848
849static void vhost_dev_lock_vqs(struct vhost_dev *d)
850{
851 int i = 0;
852 for (i = 0; i < d->nvqs; ++i)
853 mutex_lock(&d->vqs[i]->mutex);
854}
855
856static void vhost_dev_unlock_vqs(struct vhost_dev *d)
857{
858 int i = 0;
859 for (i = 0; i < d->nvqs; ++i)
860 mutex_unlock(&d->vqs[i]->mutex);
861}
862
863static int vhost_new_umem_range(struct vhost_umem *umem,
864 u64 start, u64 size, u64 end,
865 u64 userspace_addr, int perm)
866{
867 struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
868
869 if (!node)
870 return -ENOMEM;
871
872 if (umem->numem == max_iotlb_entries) {
873 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
874 vhost_umem_free(umem, tmp);
875 }
876
877 node->start = start;
878 node->size = size;
879 node->last = end;
880 node->userspace_addr = userspace_addr;
881 node->perm = perm;
882 INIT_LIST_HEAD(&node->link);
883 list_add_tail(&node->link, &umem->umem_list);
884 vhost_umem_interval_tree_insert(node, &umem->umem_tree);
885 umem->numem++;
886
887 return 0;
888}
889
890static void vhost_del_umem_range(struct vhost_umem *umem,
891 u64 start, u64 end)
892{
893 struct vhost_umem_node *node;
894
895 while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
896 start, end)))
897 vhost_umem_free(umem, node);
898}
899
900static void vhost_iotlb_notify_vq(struct vhost_dev *d,
901 struct vhost_iotlb_msg *msg)
902{
903 struct vhost_msg_node *node, *n;
904
905 spin_lock(&d->iotlb_lock);
906
907 list_for_each_entry_safe(node, n, &d->pending_list, node) {
908 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
909 if (msg->iova <= vq_msg->iova &&
910 msg->iova + msg->size - 1 > vq_msg->iova &&
911 vq_msg->type == VHOST_IOTLB_MISS) {
912 vhost_poll_queue(&node->vq->poll);
913 list_del(&node->node);
914 kfree(node);
915 }
916 }
917
918 spin_unlock(&d->iotlb_lock);
919}
920
921static int umem_access_ok(u64 uaddr, u64 size, int access)
922{
923 unsigned long a = uaddr;
924
925 /* Make sure 64 bit math will not overflow. */
926 if (vhost_overflow(uaddr, size))
927 return -EFAULT;
928
929 if ((access & VHOST_ACCESS_RO) &&
930 !access_ok(VERIFY_READ, (void __user *)a, size))
931 return -EFAULT;
932 if ((access & VHOST_ACCESS_WO) &&
933 !access_ok(VERIFY_WRITE, (void __user *)a, size))
934 return -EFAULT;
935 return 0;
936}
937
938static int vhost_process_iotlb_msg(struct vhost_dev *dev,
939 struct vhost_iotlb_msg *msg)
940{
941 int ret = 0;
942
943 vhost_dev_lock_vqs(dev);
944 switch (msg->type) {
945 case VHOST_IOTLB_UPDATE:
946 if (!dev->iotlb) {
947 ret = -EFAULT;
948 break;
949 }
950 if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
951 ret = -EFAULT;
952 break;
953 }
954 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
955 msg->iova + msg->size - 1,
956 msg->uaddr, msg->perm)) {
957 ret = -ENOMEM;
958 break;
959 }
960 vhost_iotlb_notify_vq(dev, msg);
961 break;
962 case VHOST_IOTLB_INVALIDATE:
963 vhost_del_umem_range(dev->iotlb, msg->iova,
964 msg->iova + msg->size - 1);
965 break;
966 default:
967 ret = -EINVAL;
968 break;
969 }
970
971 vhost_dev_unlock_vqs(dev);
972 return ret;
973}
974ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
975 struct iov_iter *from)
976{
977 struct vhost_msg_node node;
978 unsigned size = sizeof(struct vhost_msg);
979 size_t ret;
980 int err;
981
982 if (iov_iter_count(from) < size)
983 return 0;
984 ret = copy_from_iter(&node.msg, size, from);
985 if (ret != size)
986 goto done;
987
988 switch (node.msg.type) {
989 case VHOST_IOTLB_MSG:
990 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
991 if (err)
992 ret = err;
993 break;
994 default:
995 ret = -EINVAL;
996 break;
997 }
998
999done:
1000 return ret;
1001}
1002EXPORT_SYMBOL(vhost_chr_write_iter);
1003
1004unsigned int vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1005 poll_table *wait)
1006{
1007 unsigned int mask = 0;
1008
1009 poll_wait(file, &dev->wait, wait);
1010
1011 if (!list_empty(&dev->read_list))
1012 mask |= POLLIN | POLLRDNORM;
1013
1014 return mask;
1015}
1016EXPORT_SYMBOL(vhost_chr_poll);
1017
1018ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1019 int noblock)
1020{
1021 DEFINE_WAIT(wait);
1022 struct vhost_msg_node *node;
1023 ssize_t ret = 0;
1024 unsigned size = sizeof(struct vhost_msg);
1025
1026 if (iov_iter_count(to) < size)
1027 return 0;
1028
1029 while (1) {
1030 if (!noblock)
1031 prepare_to_wait(&dev->wait, &wait,
1032 TASK_INTERRUPTIBLE);
1033
1034 node = vhost_dequeue_msg(dev, &dev->read_list);
1035 if (node)
1036 break;
1037 if (noblock) {
1038 ret = -EAGAIN;
1039 break;
1040 }
1041 if (signal_pending(current)) {
1042 ret = -ERESTARTSYS;
1043 break;
1044 }
1045 if (!dev->iotlb) {
1046 ret = -EBADFD;
1047 break;
1048 }
1049
1050 schedule();
1051 }
1052
1053 if (!noblock)
1054 finish_wait(&dev->wait, &wait);
1055
1056 if (node) {
1057 ret = copy_to_iter(&node->msg, size, to);
1058
1059 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1060 kfree(node);
1061 return ret;
1062 }
1063
1064 vhost_enqueue_msg(dev, &dev->pending_list, node);
1065 }
1066
1067 return ret;
1068}
1069EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1070
1071static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1072{
1073 struct vhost_dev *dev = vq->dev;
1074 struct vhost_msg_node *node;
1075 struct vhost_iotlb_msg *msg;
1076
1077 node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1078 if (!node)
1079 return -ENOMEM;
1080
1081 msg = &node->msg.iotlb;
1082 msg->type = VHOST_IOTLB_MISS;
1083 msg->iova = iova;
1084 msg->perm = access;
1085
1086 vhost_enqueue_msg(dev, &dev->read_list, node);
1087
1088 return 0;
1089}
1090
1091static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1092 struct vring_desc __user *desc,
1093 struct vring_avail __user *avail,
1094 struct vring_used __user *used)
1095
1096{
1097 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1098
1099 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1100 access_ok(VERIFY_READ, avail,
1101 sizeof *avail + num * sizeof *avail->ring + s) &&
1102 access_ok(VERIFY_WRITE, used,
1103 sizeof *used + num * sizeof *used->ring + s);
1104}
1105
1106static int iotlb_access_ok(struct vhost_virtqueue *vq,
1107 int access, u64 addr, u64 len)
1108{
1109 const struct vhost_umem_node *node;
1110 struct vhost_umem *umem = vq->iotlb;
1111 u64 s = 0, size;
1112
1113 while (len > s) {
1114 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1115 addr,
1116 addr + len - 1);
1117 if (node == NULL || node->start > addr) {
1118 vhost_iotlb_miss(vq, addr, access);
1119 return false;
1120 } else if (!(node->perm & access)) {
1121 /* Report the possible access violation by
1122 * request another translation from userspace.
1123 */
1124 return false;
1125 }
1126
1127 size = node->size - addr + node->start;
1128 s += size;
1129 addr += size;
1130 }
1131
1132 return true;
1133}
1134
1135int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1136{
1137 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1138 unsigned int num = vq->num;
1139
1140 if (!vq->iotlb)
1141 return 1;
1142
1143 return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1144 num * sizeof *vq->desc) &&
1145 iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1146 sizeof *vq->avail +
1147 num * sizeof *vq->avail->ring + s) &&
1148 iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1149 sizeof *vq->used +
1150 num * sizeof *vq->used->ring + s);
1151}
1152EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1153
1154/* Can we log writes? */
1155/* Caller should have device mutex but not vq mutex */
1156int vhost_log_access_ok(struct vhost_dev *dev)
1157{
1158 return memory_access_ok(dev, dev->umem, 1);
1159}
1160EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1161
1162/* Verify access for write logging. */
1163/* Caller should have vq mutex and device mutex */
1164static int vq_log_access_ok(struct vhost_virtqueue *vq,
1165 void __user *log_base)
1166{
1167 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1168
1169 return vq_memory_access_ok(log_base, vq->umem,
1170 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1171 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1172 sizeof *vq->used +
1173 vq->num * sizeof *vq->used->ring + s));
1174}
1175
1176/* Can we start vq? */
1177/* Caller should have vq mutex and device mutex */
1178int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1179{
1180 if (vq->iotlb) {
1181 /* When device IOTLB was used, the access validation
1182 * will be validated during prefetching.
1183 */
1184 return 1;
1185 }
1186 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
1187 vq_log_access_ok(vq, vq->log_base);
1188}
1189EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1190
1191static struct vhost_umem *vhost_umem_alloc(void)
1192{
1193 struct vhost_umem *umem = vhost_kvzalloc(sizeof(*umem));
1194
1195 if (!umem)
1196 return NULL;
1197
1198 umem->umem_tree = RB_ROOT;
1199 umem->numem = 0;
1200 INIT_LIST_HEAD(&umem->umem_list);
1201
1202 return umem;
1203}
1204
1205static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1206{
1207 struct vhost_memory mem, *newmem;
1208 struct vhost_memory_region *region;
1209 struct vhost_umem *newumem, *oldumem;
1210 unsigned long size = offsetof(struct vhost_memory, regions);
1211 int i;
1212
1213 if (copy_from_user(&mem, m, size))
1214 return -EFAULT;
1215 if (mem.padding)
1216 return -EOPNOTSUPP;
1217 if (mem.nregions > max_mem_regions)
1218 return -E2BIG;
1219 newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
1220 if (!newmem)
1221 return -ENOMEM;
1222
1223 memcpy(newmem, &mem, size);
1224 if (copy_from_user(newmem->regions, m->regions,
1225 mem.nregions * sizeof *m->regions)) {
1226 kvfree(newmem);
1227 return -EFAULT;
1228 }
1229
1230 newumem = vhost_umem_alloc();
1231 if (!newumem) {
1232 kvfree(newmem);
1233 return -ENOMEM;
1234 }
1235
1236 for (region = newmem->regions;
1237 region < newmem->regions + mem.nregions;
1238 region++) {
1239 if (vhost_new_umem_range(newumem,
1240 region->guest_phys_addr,
1241 region->memory_size,
1242 region->guest_phys_addr +
1243 region->memory_size - 1,
1244 region->userspace_addr,
1245 VHOST_ACCESS_RW))
1246 goto err;
1247 }
1248
1249 if (!memory_access_ok(d, newumem, 0))
1250 goto err;
1251
1252 oldumem = d->umem;
1253 d->umem = newumem;
1254
1255 /* All memory accesses are done under some VQ mutex. */
1256 for (i = 0; i < d->nvqs; ++i) {
1257 mutex_lock(&d->vqs[i]->mutex);
1258 d->vqs[i]->umem = newumem;
1259 mutex_unlock(&d->vqs[i]->mutex);
1260 }
1261
1262 kvfree(newmem);
1263 vhost_umem_clean(oldumem);
1264 return 0;
1265
1266err:
1267 vhost_umem_clean(newumem);
1268 kvfree(newmem);
1269 return -EFAULT;
1270}
1271
1272long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1273{
1274 struct file *eventfp, *filep = NULL;
1275 bool pollstart = false, pollstop = false;
1276 struct eventfd_ctx *ctx = NULL;
1277 u32 __user *idxp = argp;
1278 struct vhost_virtqueue *vq;
1279 struct vhost_vring_state s;
1280 struct vhost_vring_file f;
1281 struct vhost_vring_addr a;
1282 u32 idx;
1283 long r;
1284
1285 r = get_user(idx, idxp);
1286 if (r < 0)
1287 return r;
1288 if (idx >= d->nvqs)
1289 return -ENOBUFS;
1290
1291 vq = d->vqs[idx];
1292
1293 mutex_lock(&vq->mutex);
1294
1295 switch (ioctl) {
1296 case VHOST_SET_VRING_NUM:
1297 /* Resizing ring with an active backend?
1298 * You don't want to do that. */
1299 if (vq->private_data) {
1300 r = -EBUSY;
1301 break;
1302 }
1303 if (copy_from_user(&s, argp, sizeof s)) {
1304 r = -EFAULT;
1305 break;
1306 }
1307 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1308 r = -EINVAL;
1309 break;
1310 }
1311 vq->num = s.num;
1312 break;
1313 case VHOST_SET_VRING_BASE:
1314 /* Moving base with an active backend?
1315 * You don't want to do that. */
1316 if (vq->private_data) {
1317 r = -EBUSY;
1318 break;
1319 }
1320 if (copy_from_user(&s, argp, sizeof s)) {
1321 r = -EFAULT;
1322 break;
1323 }
1324 if (s.num > 0xffff) {
1325 r = -EINVAL;
1326 break;
1327 }
1328 vq->last_avail_idx = vq->last_used_event = s.num;
1329 /* Forget the cached index value. */
1330 vq->avail_idx = vq->last_avail_idx;
1331 break;
1332 case VHOST_GET_VRING_BASE:
1333 s.index = idx;
1334 s.num = vq->last_avail_idx;
1335 if (copy_to_user(argp, &s, sizeof s))
1336 r = -EFAULT;
1337 break;
1338 case VHOST_SET_VRING_ADDR:
1339 if (copy_from_user(&a, argp, sizeof a)) {
1340 r = -EFAULT;
1341 break;
1342 }
1343 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1344 r = -EOPNOTSUPP;
1345 break;
1346 }
1347 /* For 32bit, verify that the top 32bits of the user
1348 data are set to zero. */
1349 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1350 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1351 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1352 r = -EFAULT;
1353 break;
1354 }
1355
1356 /* Make sure it's safe to cast pointers to vring types. */
1357 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1358 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1359 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1360 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1361 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1362 r = -EINVAL;
1363 break;
1364 }
1365
1366 /* We only verify access here if backend is configured.
1367 * If it is not, we don't as size might not have been setup.
1368 * We will verify when backend is configured. */
1369 if (vq->private_data) {
1370 if (!vq_access_ok(vq, vq->num,
1371 (void __user *)(unsigned long)a.desc_user_addr,
1372 (void __user *)(unsigned long)a.avail_user_addr,
1373 (void __user *)(unsigned long)a.used_user_addr)) {
1374 r = -EINVAL;
1375 break;
1376 }
1377
1378 /* Also validate log access for used ring if enabled. */
1379 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1380 !log_access_ok(vq->log_base, a.log_guest_addr,
1381 sizeof *vq->used +
1382 vq->num * sizeof *vq->used->ring)) {
1383 r = -EINVAL;
1384 break;
1385 }
1386 }
1387
1388 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1389 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1390 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1391 vq->log_addr = a.log_guest_addr;
1392 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1393 break;
1394 case VHOST_SET_VRING_KICK:
1395 if (copy_from_user(&f, argp, sizeof f)) {
1396 r = -EFAULT;
1397 break;
1398 }
1399 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1400 if (IS_ERR(eventfp)) {
1401 r = PTR_ERR(eventfp);
1402 break;
1403 }
1404 if (eventfp != vq->kick) {
1405 pollstop = (filep = vq->kick) != NULL;
1406 pollstart = (vq->kick = eventfp) != NULL;
1407 } else
1408 filep = eventfp;
1409 break;
1410 case VHOST_SET_VRING_CALL:
1411 if (copy_from_user(&f, argp, sizeof f)) {
1412 r = -EFAULT;
1413 break;
1414 }
1415 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1416 if (IS_ERR(eventfp)) {
1417 r = PTR_ERR(eventfp);
1418 break;
1419 }
1420 if (eventfp != vq->call) {
1421 filep = vq->call;
1422 ctx = vq->call_ctx;
1423 vq->call = eventfp;
1424 vq->call_ctx = eventfp ?
1425 eventfd_ctx_fileget(eventfp) : NULL;
1426 } else
1427 filep = eventfp;
1428 break;
1429 case VHOST_SET_VRING_ERR:
1430 if (copy_from_user(&f, argp, sizeof f)) {
1431 r = -EFAULT;
1432 break;
1433 }
1434 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1435 if (IS_ERR(eventfp)) {
1436 r = PTR_ERR(eventfp);
1437 break;
1438 }
1439 if (eventfp != vq->error) {
1440 filep = vq->error;
1441 vq->error = eventfp;
1442 ctx = vq->error_ctx;
1443 vq->error_ctx = eventfp ?
1444 eventfd_ctx_fileget(eventfp) : NULL;
1445 } else
1446 filep = eventfp;
1447 break;
1448 case VHOST_SET_VRING_ENDIAN:
1449 r = vhost_set_vring_endian(vq, argp);
1450 break;
1451 case VHOST_GET_VRING_ENDIAN:
1452 r = vhost_get_vring_endian(vq, idx, argp);
1453 break;
1454 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1455 if (copy_from_user(&s, argp, sizeof(s))) {
1456 r = -EFAULT;
1457 break;
1458 }
1459 vq->busyloop_timeout = s.num;
1460 break;
1461 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1462 s.index = idx;
1463 s.num = vq->busyloop_timeout;
1464 if (copy_to_user(argp, &s, sizeof(s)))
1465 r = -EFAULT;
1466 break;
1467 default:
1468 r = -ENOIOCTLCMD;
1469 }
1470
1471 if (pollstop && vq->handle_kick)
1472 vhost_poll_stop(&vq->poll);
1473
1474 if (ctx)
1475 eventfd_ctx_put(ctx);
1476 if (filep)
1477 fput(filep);
1478
1479 if (pollstart && vq->handle_kick)
1480 r = vhost_poll_start(&vq->poll, vq->kick);
1481
1482 mutex_unlock(&vq->mutex);
1483
1484 if (pollstop && vq->handle_kick)
1485 vhost_poll_flush(&vq->poll);
1486 return r;
1487}
1488EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1489
1490int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1491{
1492 struct vhost_umem *niotlb, *oiotlb;
1493 int i;
1494
1495 niotlb = vhost_umem_alloc();
1496 if (!niotlb)
1497 return -ENOMEM;
1498
1499 oiotlb = d->iotlb;
1500 d->iotlb = niotlb;
1501
1502 for (i = 0; i < d->nvqs; ++i) {
1503 mutex_lock(&d->vqs[i]->mutex);
1504 d->vqs[i]->iotlb = niotlb;
1505 mutex_unlock(&d->vqs[i]->mutex);
1506 }
1507
1508 vhost_umem_clean(oiotlb);
1509
1510 return 0;
1511}
1512EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1513
1514/* Caller must have device mutex */
1515long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1516{
1517 struct file *eventfp, *filep = NULL;
1518 struct eventfd_ctx *ctx = NULL;
1519 u64 p;
1520 long r;
1521 int i, fd;
1522
1523 /* If you are not the owner, you can become one */
1524 if (ioctl == VHOST_SET_OWNER) {
1525 r = vhost_dev_set_owner(d);
1526 goto done;
1527 }
1528
1529 /* You must be the owner to do anything else */
1530 r = vhost_dev_check_owner(d);
1531 if (r)
1532 goto done;
1533
1534 switch (ioctl) {
1535 case VHOST_SET_MEM_TABLE:
1536 r = vhost_set_memory(d, argp);
1537 break;
1538 case VHOST_SET_LOG_BASE:
1539 if (copy_from_user(&p, argp, sizeof p)) {
1540 r = -EFAULT;
1541 break;
1542 }
1543 if ((u64)(unsigned long)p != p) {
1544 r = -EFAULT;
1545 break;
1546 }
1547 for (i = 0; i < d->nvqs; ++i) {
1548 struct vhost_virtqueue *vq;
1549 void __user *base = (void __user *)(unsigned long)p;
1550 vq = d->vqs[i];
1551 mutex_lock(&vq->mutex);
1552 /* If ring is inactive, will check when it's enabled. */
1553 if (vq->private_data && !vq_log_access_ok(vq, base))
1554 r = -EFAULT;
1555 else
1556 vq->log_base = base;
1557 mutex_unlock(&vq->mutex);
1558 }
1559 break;
1560 case VHOST_SET_LOG_FD:
1561 r = get_user(fd, (int __user *)argp);
1562 if (r < 0)
1563 break;
1564 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1565 if (IS_ERR(eventfp)) {
1566 r = PTR_ERR(eventfp);
1567 break;
1568 }
1569 if (eventfp != d->log_file) {
1570 filep = d->log_file;
1571 d->log_file = eventfp;
1572 ctx = d->log_ctx;
1573 d->log_ctx = eventfp ?
1574 eventfd_ctx_fileget(eventfp) : NULL;
1575 } else
1576 filep = eventfp;
1577 for (i = 0; i < d->nvqs; ++i) {
1578 mutex_lock(&d->vqs[i]->mutex);
1579 d->vqs[i]->log_ctx = d->log_ctx;
1580 mutex_unlock(&d->vqs[i]->mutex);
1581 }
1582 if (ctx)
1583 eventfd_ctx_put(ctx);
1584 if (filep)
1585 fput(filep);
1586 break;
1587 default:
1588 r = -ENOIOCTLCMD;
1589 break;
1590 }
1591done:
1592 return r;
1593}
1594EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1595
1596/* TODO: This is really inefficient. We need something like get_user()
1597 * (instruction directly accesses the data, with an exception table entry
1598 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1599 */
1600static int set_bit_to_user(int nr, void __user *addr)
1601{
1602 unsigned long log = (unsigned long)addr;
1603 struct page *page;
1604 void *base;
1605 int bit = nr + (log % PAGE_SIZE) * 8;
1606 int r;
1607
1608 r = get_user_pages_fast(log, 1, 1, &page);
1609 if (r < 0)
1610 return r;
1611 BUG_ON(r != 1);
1612 base = kmap_atomic(page);
1613 set_bit(bit, base);
1614 kunmap_atomic(base);
1615 set_page_dirty_lock(page);
1616 put_page(page);
1617 return 0;
1618}
1619
1620static int log_write(void __user *log_base,
1621 u64 write_address, u64 write_length)
1622{
1623 u64 write_page = write_address / VHOST_PAGE_SIZE;
1624 int r;
1625
1626 if (!write_length)
1627 return 0;
1628 write_length += write_address % VHOST_PAGE_SIZE;
1629 for (;;) {
1630 u64 base = (u64)(unsigned long)log_base;
1631 u64 log = base + write_page / 8;
1632 int bit = write_page % 8;
1633 if ((u64)(unsigned long)log != log)
1634 return -EFAULT;
1635 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1636 if (r < 0)
1637 return r;
1638 if (write_length <= VHOST_PAGE_SIZE)
1639 break;
1640 write_length -= VHOST_PAGE_SIZE;
1641 write_page += 1;
1642 }
1643 return r;
1644}
1645
1646int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1647 unsigned int log_num, u64 len)
1648{
1649 int i, r;
1650
1651 /* Make sure data written is seen before log. */
1652 smp_wmb();
1653 for (i = 0; i < log_num; ++i) {
1654 u64 l = min(log[i].len, len);
1655 r = log_write(vq->log_base, log[i].addr, l);
1656 if (r < 0)
1657 return r;
1658 len -= l;
1659 if (!len) {
1660 if (vq->log_ctx)
1661 eventfd_signal(vq->log_ctx, 1);
1662 return 0;
1663 }
1664 }
1665 /* Length written exceeds what we have stored. This is a bug. */
1666 BUG();
1667 return 0;
1668}
1669EXPORT_SYMBOL_GPL(vhost_log_write);
1670
1671static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1672{
1673 void __user *used;
1674 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1675 &vq->used->flags) < 0)
1676 return -EFAULT;
1677 if (unlikely(vq->log_used)) {
1678 /* Make sure the flag is seen before log. */
1679 smp_wmb();
1680 /* Log used flag write. */
1681 used = &vq->used->flags;
1682 log_write(vq->log_base, vq->log_addr +
1683 (used - (void __user *)vq->used),
1684 sizeof vq->used->flags);
1685 if (vq->log_ctx)
1686 eventfd_signal(vq->log_ctx, 1);
1687 }
1688 return 0;
1689}
1690
1691static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1692{
1693 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1694 vhost_avail_event(vq)))
1695 return -EFAULT;
1696 if (unlikely(vq->log_used)) {
1697 void __user *used;
1698 /* Make sure the event is seen before log. */
1699 smp_wmb();
1700 /* Log avail event write */
1701 used = vhost_avail_event(vq);
1702 log_write(vq->log_base, vq->log_addr +
1703 (used - (void __user *)vq->used),
1704 sizeof *vhost_avail_event(vq));
1705 if (vq->log_ctx)
1706 eventfd_signal(vq->log_ctx, 1);
1707 }
1708 return 0;
1709}
1710
1711int vhost_vq_init_access(struct vhost_virtqueue *vq)
1712{
1713 __virtio16 last_used_idx;
1714 int r;
1715 bool is_le = vq->is_le;
1716
1717 if (!vq->private_data)
1718 return 0;
1719
1720 vhost_init_is_le(vq);
1721
1722 r = vhost_update_used_flags(vq);
1723 if (r)
1724 goto err;
1725 vq->signalled_used_valid = false;
1726 if (!vq->iotlb &&
1727 !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1728 r = -EFAULT;
1729 goto err;
1730 }
1731 r = vhost_get_user(vq, last_used_idx, &vq->used->idx);
1732 if (r) {
1733 vq_err(vq, "Can't access used idx at %p\n",
1734 &vq->used->idx);
1735 goto err;
1736 }
1737 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1738 return 0;
1739
1740err:
1741 vq->is_le = is_le;
1742 return r;
1743}
1744EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1745
1746static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1747 struct iovec iov[], int iov_size, int access)
1748{
1749 const struct vhost_umem_node *node;
1750 struct vhost_dev *dev = vq->dev;
1751 struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1752 struct iovec *_iov;
1753 u64 s = 0;
1754 int ret = 0;
1755
1756 while ((u64)len > s) {
1757 u64 size;
1758 if (unlikely(ret >= iov_size)) {
1759 ret = -ENOBUFS;
1760 break;
1761 }
1762
1763 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1764 addr, addr + len - 1);
1765 if (node == NULL || node->start > addr) {
1766 if (umem != dev->iotlb) {
1767 ret = -EFAULT;
1768 break;
1769 }
1770 ret = -EAGAIN;
1771 break;
1772 } else if (!(node->perm & access)) {
1773 ret = -EPERM;
1774 break;
1775 }
1776
1777 _iov = iov + ret;
1778 size = node->size - addr + node->start;
1779 _iov->iov_len = min((u64)len - s, size);
1780 _iov->iov_base = (void __user *)(unsigned long)
1781 (node->userspace_addr + addr - node->start);
1782 s += size;
1783 addr += size;
1784 ++ret;
1785 }
1786
1787 if (ret == -EAGAIN)
1788 vhost_iotlb_miss(vq, addr, access);
1789 return ret;
1790}
1791
1792/* Each buffer in the virtqueues is actually a chain of descriptors. This
1793 * function returns the next descriptor in the chain,
1794 * or -1U if we're at the end. */
1795static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1796{
1797 unsigned int next;
1798
1799 /* If this descriptor says it doesn't chain, we're done. */
1800 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1801 return -1U;
1802
1803 /* Check they're not leading us off end of descriptors. */
1804 next = vhost16_to_cpu(vq, desc->next);
1805 /* Make sure compiler knows to grab that: we don't want it changing! */
1806 /* We will use the result as an index in an array, so most
1807 * architectures only need a compiler barrier here. */
1808 read_barrier_depends();
1809
1810 return next;
1811}
1812
1813static int get_indirect(struct vhost_virtqueue *vq,
1814 struct iovec iov[], unsigned int iov_size,
1815 unsigned int *out_num, unsigned int *in_num,
1816 struct vhost_log *log, unsigned int *log_num,
1817 struct vring_desc *indirect)
1818{
1819 struct vring_desc desc;
1820 unsigned int i = 0, count, found = 0;
1821 u32 len = vhost32_to_cpu(vq, indirect->len);
1822 struct iov_iter from;
1823 int ret, access;
1824
1825 /* Sanity check */
1826 if (unlikely(len % sizeof desc)) {
1827 vq_err(vq, "Invalid length in indirect descriptor: "
1828 "len 0x%llx not multiple of 0x%zx\n",
1829 (unsigned long long)len,
1830 sizeof desc);
1831 return -EINVAL;
1832 }
1833
1834 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1835 UIO_MAXIOV, VHOST_ACCESS_RO);
1836 if (unlikely(ret < 0)) {
1837 if (ret != -EAGAIN)
1838 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1839 return ret;
1840 }
1841 iov_iter_init(&from, READ, vq->indirect, ret, len);
1842
1843 /* We will use the result as an address to read from, so most
1844 * architectures only need a compiler barrier here. */
1845 read_barrier_depends();
1846
1847 count = len / sizeof desc;
1848 /* Buffers are chained via a 16 bit next field, so
1849 * we can have at most 2^16 of these. */
1850 if (unlikely(count > USHRT_MAX + 1)) {
1851 vq_err(vq, "Indirect buffer length too big: %d\n",
1852 indirect->len);
1853 return -E2BIG;
1854 }
1855
1856 do {
1857 unsigned iov_count = *in_num + *out_num;
1858 if (unlikely(++found > count)) {
1859 vq_err(vq, "Loop detected: last one at %u "
1860 "indirect size %u\n",
1861 i, count);
1862 return -EINVAL;
1863 }
1864 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1865 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1866 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1867 return -EINVAL;
1868 }
1869 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1870 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1871 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1872 return -EINVAL;
1873 }
1874
1875 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1876 access = VHOST_ACCESS_WO;
1877 else
1878 access = VHOST_ACCESS_RO;
1879
1880 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1881 vhost32_to_cpu(vq, desc.len), iov + iov_count,
1882 iov_size - iov_count, access);
1883 if (unlikely(ret < 0)) {
1884 if (ret != -EAGAIN)
1885 vq_err(vq, "Translation failure %d indirect idx %d\n",
1886 ret, i);
1887 return ret;
1888 }
1889 /* If this is an input descriptor, increment that count. */
1890 if (access == VHOST_ACCESS_WO) {
1891 *in_num += ret;
1892 if (unlikely(log)) {
1893 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1894 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1895 ++*log_num;
1896 }
1897 } else {
1898 /* If it's an output descriptor, they're all supposed
1899 * to come before any input descriptors. */
1900 if (unlikely(*in_num)) {
1901 vq_err(vq, "Indirect descriptor "
1902 "has out after in: idx %d\n", i);
1903 return -EINVAL;
1904 }
1905 *out_num += ret;
1906 }
1907 } while ((i = next_desc(vq, &desc)) != -1);
1908 return 0;
1909}
1910
1911/* This looks in the virtqueue and for the first available buffer, and converts
1912 * it to an iovec for convenient access. Since descriptors consist of some
1913 * number of output then some number of input descriptors, it's actually two
1914 * iovecs, but we pack them into one and note how many of each there were.
1915 *
1916 * This function returns the descriptor number found, or vq->num (which is
1917 * never a valid descriptor number) if none was found. A negative code is
1918 * returned on error. */
1919int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1920 struct iovec iov[], unsigned int iov_size,
1921 unsigned int *out_num, unsigned int *in_num,
1922 struct vhost_log *log, unsigned int *log_num)
1923{
1924 struct vring_desc desc;
1925 unsigned int i, head, found = 0;
1926 u16 last_avail_idx;
1927 __virtio16 avail_idx;
1928 __virtio16 ring_head;
1929 int ret, access;
1930
1931 /* Check it isn't doing very strange things with descriptor numbers. */
1932 last_avail_idx = vq->last_avail_idx;
1933 if (unlikely(vhost_get_user(vq, avail_idx, &vq->avail->idx))) {
1934 vq_err(vq, "Failed to access avail idx at %p\n",
1935 &vq->avail->idx);
1936 return -EFAULT;
1937 }
1938 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1939
1940 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1941 vq_err(vq, "Guest moved used index from %u to %u",
1942 last_avail_idx, vq->avail_idx);
1943 return -EFAULT;
1944 }
1945
1946 /* If there's nothing new since last we looked, return invalid. */
1947 if (vq->avail_idx == last_avail_idx)
1948 return vq->num;
1949
1950 /* Only get avail ring entries after they have been exposed by guest. */
1951 smp_rmb();
1952
1953 /* Grab the next descriptor number they're advertising, and increment
1954 * the index we've seen. */
1955 if (unlikely(vhost_get_user(vq, ring_head,
1956 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
1957 vq_err(vq, "Failed to read head: idx %d address %p\n",
1958 last_avail_idx,
1959 &vq->avail->ring[last_avail_idx % vq->num]);
1960 return -EFAULT;
1961 }
1962
1963 head = vhost16_to_cpu(vq, ring_head);
1964
1965 /* If their number is silly, that's an error. */
1966 if (unlikely(head >= vq->num)) {
1967 vq_err(vq, "Guest says index %u > %u is available",
1968 head, vq->num);
1969 return -EINVAL;
1970 }
1971
1972 /* When we start there are none of either input nor output. */
1973 *out_num = *in_num = 0;
1974 if (unlikely(log))
1975 *log_num = 0;
1976
1977 i = head;
1978 do {
1979 unsigned iov_count = *in_num + *out_num;
1980 if (unlikely(i >= vq->num)) {
1981 vq_err(vq, "Desc index is %u > %u, head = %u",
1982 i, vq->num, head);
1983 return -EINVAL;
1984 }
1985 if (unlikely(++found > vq->num)) {
1986 vq_err(vq, "Loop detected: last one at %u "
1987 "vq size %u head %u\n",
1988 i, vq->num, head);
1989 return -EINVAL;
1990 }
1991 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
1992 sizeof desc);
1993 if (unlikely(ret)) {
1994 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1995 i, vq->desc + i);
1996 return -EFAULT;
1997 }
1998 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
1999 ret = get_indirect(vq, iov, iov_size,
2000 out_num, in_num,
2001 log, log_num, &desc);
2002 if (unlikely(ret < 0)) {
2003 if (ret != -EAGAIN)
2004 vq_err(vq, "Failure detected "
2005 "in indirect descriptor at idx %d\n", i);
2006 return ret;
2007 }
2008 continue;
2009 }
2010
2011 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2012 access = VHOST_ACCESS_WO;
2013 else
2014 access = VHOST_ACCESS_RO;
2015 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2016 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2017 iov_size - iov_count, access);
2018 if (unlikely(ret < 0)) {
2019 if (ret != -EAGAIN)
2020 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2021 ret, i);
2022 return ret;
2023 }
2024 if (access == VHOST_ACCESS_WO) {
2025 /* If this is an input descriptor,
2026 * increment that count. */
2027 *in_num += ret;
2028 if (unlikely(log)) {
2029 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2030 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2031 ++*log_num;
2032 }
2033 } else {
2034 /* If it's an output descriptor, they're all supposed
2035 * to come before any input descriptors. */
2036 if (unlikely(*in_num)) {
2037 vq_err(vq, "Descriptor has out after in: "
2038 "idx %d\n", i);
2039 return -EINVAL;
2040 }
2041 *out_num += ret;
2042 }
2043 } while ((i = next_desc(vq, &desc)) != -1);
2044
2045 /* On success, increment avail index. */
2046 vq->last_avail_idx++;
2047
2048 /* Assume notifications from guest are disabled at this point,
2049 * if they aren't we would need to update avail_event index. */
2050 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2051 return head;
2052}
2053EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2054
2055/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2056void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2057{
2058 vq->last_avail_idx -= n;
2059}
2060EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2061
2062/* After we've used one of their buffers, we tell them about it. We'll then
2063 * want to notify the guest, using eventfd. */
2064int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2065{
2066 struct vring_used_elem heads = {
2067 cpu_to_vhost32(vq, head),
2068 cpu_to_vhost32(vq, len)
2069 };
2070
2071 return vhost_add_used_n(vq, &heads, 1);
2072}
2073EXPORT_SYMBOL_GPL(vhost_add_used);
2074
2075static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2076 struct vring_used_elem *heads,
2077 unsigned count)
2078{
2079 struct vring_used_elem __user *used;
2080 u16 old, new;
2081 int start;
2082
2083 start = vq->last_used_idx & (vq->num - 1);
2084 used = vq->used->ring + start;
2085 if (count == 1) {
2086 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2087 vq_err(vq, "Failed to write used id");
2088 return -EFAULT;
2089 }
2090 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2091 vq_err(vq, "Failed to write used len");
2092 return -EFAULT;
2093 }
2094 } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2095 vq_err(vq, "Failed to write used");
2096 return -EFAULT;
2097 }
2098 if (unlikely(vq->log_used)) {
2099 /* Make sure data is seen before log. */
2100 smp_wmb();
2101 /* Log used ring entry write. */
2102 log_write(vq->log_base,
2103 vq->log_addr +
2104 ((void __user *)used - (void __user *)vq->used),
2105 count * sizeof *used);
2106 }
2107 old = vq->last_used_idx;
2108 new = (vq->last_used_idx += count);
2109 /* If the driver never bothers to signal in a very long while,
2110 * used index might wrap around. If that happens, invalidate
2111 * signalled_used index we stored. TODO: make sure driver
2112 * signals at least once in 2^16 and remove this. */
2113 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2114 vq->signalled_used_valid = false;
2115 return 0;
2116}
2117
2118/* After we've used one of their buffers, we tell them about it. We'll then
2119 * want to notify the guest, using eventfd. */
2120int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2121 unsigned count)
2122{
2123 int start, n, r;
2124
2125 start = vq->last_used_idx & (vq->num - 1);
2126 n = vq->num - start;
2127 if (n < count) {
2128 r = __vhost_add_used_n(vq, heads, n);
2129 if (r < 0)
2130 return r;
2131 heads += n;
2132 count -= n;
2133 }
2134 r = __vhost_add_used_n(vq, heads, count);
2135
2136 /* Make sure buffer is written before we update index. */
2137 smp_wmb();
2138 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2139 &vq->used->idx)) {
2140 vq_err(vq, "Failed to increment used idx");
2141 return -EFAULT;
2142 }
2143 if (unlikely(vq->log_used)) {
2144 /* Log used index update. */
2145 log_write(vq->log_base,
2146 vq->log_addr + offsetof(struct vring_used, idx),
2147 sizeof vq->used->idx);
2148 if (vq->log_ctx)
2149 eventfd_signal(vq->log_ctx, 1);
2150 }
2151 return r;
2152}
2153EXPORT_SYMBOL_GPL(vhost_add_used_n);
2154
2155static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2156{
2157 __u16 old, new;
2158 __virtio16 event;
2159 bool v;
2160
2161 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2162 unlikely(vq->avail_idx == vq->last_avail_idx))
2163 return true;
2164
2165 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2166 __virtio16 flags;
2167 /* Flush out used index updates. This is paired
2168 * with the barrier that the Guest executes when enabling
2169 * interrupts. */
2170 smp_mb();
2171 if (vhost_get_user(vq, flags, &vq->avail->flags)) {
2172 vq_err(vq, "Failed to get flags");
2173 return true;
2174 }
2175 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2176 }
2177 old = vq->signalled_used;
2178 v = vq->signalled_used_valid;
2179 new = vq->signalled_used = vq->last_used_idx;
2180 vq->signalled_used_valid = true;
2181
2182 if (unlikely(!v))
2183 return true;
2184
2185 /* We're sure if the following conditions are met, there's no
2186 * need to notify guest:
2187 * 1) cached used event is ahead of new
2188 * 2) old to new updating does not cross cached used event. */
2189 if (vring_need_event(vq->last_used_event, new + vq->num, new) &&
2190 !vring_need_event(vq->last_used_event, new, old))
2191 return false;
2192
2193 /* Flush out used index updates. This is paired
2194 * with the barrier that the Guest executes when enabling
2195 * interrupts. */
2196 smp_mb();
2197
2198 if (vhost_get_user(vq, event, vhost_used_event(vq))) {
2199 vq_err(vq, "Failed to get used event idx");
2200 return true;
2201 }
2202 vq->last_used_event = vhost16_to_cpu(vq, event);
2203
2204 return vring_need_event(vq->last_used_event, new, old);
2205}
2206
2207/* This actually signals the guest, using eventfd. */
2208void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2209{
2210 /* Signal the Guest tell them we used something up. */
2211 if (vq->call_ctx && vhost_notify(dev, vq))
2212 eventfd_signal(vq->call_ctx, 1);
2213}
2214EXPORT_SYMBOL_GPL(vhost_signal);
2215
2216/* And here's the combo meal deal. Supersize me! */
2217void vhost_add_used_and_signal(struct vhost_dev *dev,
2218 struct vhost_virtqueue *vq,
2219 unsigned int head, int len)
2220{
2221 vhost_add_used(vq, head, len);
2222 vhost_signal(dev, vq);
2223}
2224EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2225
2226/* multi-buffer version of vhost_add_used_and_signal */
2227void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2228 struct vhost_virtqueue *vq,
2229 struct vring_used_elem *heads, unsigned count)
2230{
2231 vhost_add_used_n(vq, heads, count);
2232 vhost_signal(dev, vq);
2233}
2234EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2235
2236/* return true if we're sure that avaiable ring is empty */
2237bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2238{
2239 __virtio16 avail_idx;
2240 int r;
2241
2242 r = vhost_get_user(vq, avail_idx, &vq->avail->idx);
2243 if (r)
2244 return false;
2245
2246 return vhost16_to_cpu(vq, avail_idx) == vq->avail_idx;
2247}
2248EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2249
2250/* OK, now we need to know about added descriptors. */
2251bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2252{
2253 __virtio16 avail_idx;
2254 int r;
2255
2256 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2257 return false;
2258 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2259 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2260 r = vhost_update_used_flags(vq);
2261 if (r) {
2262 vq_err(vq, "Failed to enable notification at %p: %d\n",
2263 &vq->used->flags, r);
2264 return false;
2265 }
2266 } else {
2267 r = vhost_update_avail_event(vq, vq->avail_idx);
2268 if (r) {
2269 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2270 vhost_avail_event(vq), r);
2271 return false;
2272 }
2273 }
2274 /* They could have slipped one in as we were doing that: make
2275 * sure it's written, then check again. */
2276 smp_mb();
2277 r = vhost_get_user(vq, avail_idx, &vq->avail->idx);
2278 if (r) {
2279 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2280 &vq->avail->idx, r);
2281 return false;
2282 }
2283
2284 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2285}
2286EXPORT_SYMBOL_GPL(vhost_enable_notify);
2287
2288/* We don't need to be notified again. */
2289void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2290{
2291 int r;
2292
2293 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2294 return;
2295 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2296 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2297 r = vhost_update_used_flags(vq);
2298 if (r)
2299 vq_err(vq, "Failed to enable notification at %p: %d\n",
2300 &vq->used->flags, r);
2301 }
2302}
2303EXPORT_SYMBOL_GPL(vhost_disable_notify);
2304
2305/* Create a new message. */
2306struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2307{
2308 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2309 if (!node)
2310 return NULL;
2311 node->vq = vq;
2312 node->msg.type = type;
2313 return node;
2314}
2315EXPORT_SYMBOL_GPL(vhost_new_msg);
2316
2317void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2318 struct vhost_msg_node *node)
2319{
2320 spin_lock(&dev->iotlb_lock);
2321 list_add_tail(&node->node, head);
2322 spin_unlock(&dev->iotlb_lock);
2323
2324 wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
2325}
2326EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2327
2328struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2329 struct list_head *head)
2330{
2331 struct vhost_msg_node *node = NULL;
2332
2333 spin_lock(&dev->iotlb_lock);
2334 if (!list_empty(head)) {
2335 node = list_first_entry(head, struct vhost_msg_node,
2336 node);
2337 list_del(&node->node);
2338 }
2339 spin_unlock(&dev->iotlb_lock);
2340
2341 return node;
2342}
2343EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2344
2345
2346static int __init vhost_init(void)
2347{
2348 return 0;
2349}
2350
2351static void __exit vhost_exit(void)
2352{
2353}
2354
2355module_init(vhost_init);
2356module_exit(vhost_exit);
2357
2358MODULE_VERSION("0.0.1");
2359MODULE_LICENSE("GPL v2");
2360MODULE_AUTHOR("Michael S. Tsirkin");
2361MODULE_DESCRIPTION("Host kernel accelerator for virtio");