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// SPDX-License-Identifier: GPL-2.0-only
2/* Copyright (C) 2009 Red Hat, Inc.
3 * Copyright (C) 2006 Rusty Russell IBM Corporation
4 *
5 * Author: Michael S. Tsirkin <mst@redhat.com>
6 *
7 * Inspiration, some code, and most witty comments come from
8 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
9 *
10 * Generic code for virtio server in host kernel.
11 */
12
13#include <linux/eventfd.h>
14#include <linux/vhost.h>
15#include <linux/uio.h>
16#include <linux/mm.h>
17#include <linux/miscdevice.h>
18#include <linux/mutex.h>
19#include <linux/poll.h>
20#include <linux/file.h>
21#include <linux/highmem.h>
22#include <linux/slab.h>
23#include <linux/vmalloc.h>
24#include <linux/kthread.h>
25#include <linux/cgroup.h>
26#include <linux/module.h>
27#include <linux/sort.h>
28#include <linux/sched/mm.h>
29#include <linux/sched/signal.h>
30#include <linux/interval_tree_generic.h>
31#include <linux/nospec.h>
32#include <linux/kcov.h>
33
34#include "vhost.h"
35
36static ushort max_mem_regions = 64;
37module_param(max_mem_regions, ushort, 0444);
38MODULE_PARM_DESC(max_mem_regions,
39 "Maximum number of memory regions in memory map. (default: 64)");
40static int max_iotlb_entries = 2048;
41module_param(max_iotlb_entries, int, 0444);
42MODULE_PARM_DESC(max_iotlb_entries,
43 "Maximum number of iotlb entries. (default: 2048)");
44
45enum {
46 VHOST_MEMORY_F_LOG = 0x1,
47};
48
49#define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
50#define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
51
52#ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
53static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
54{
55 vq->user_be = !virtio_legacy_is_little_endian();
56}
57
58static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
59{
60 vq->user_be = true;
61}
62
63static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
64{
65 vq->user_be = false;
66}
67
68static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
69{
70 struct vhost_vring_state s;
71
72 if (vq->private_data)
73 return -EBUSY;
74
75 if (copy_from_user(&s, argp, sizeof(s)))
76 return -EFAULT;
77
78 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
79 s.num != VHOST_VRING_BIG_ENDIAN)
80 return -EINVAL;
81
82 if (s.num == VHOST_VRING_BIG_ENDIAN)
83 vhost_enable_cross_endian_big(vq);
84 else
85 vhost_enable_cross_endian_little(vq);
86
87 return 0;
88}
89
90static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
91 int __user *argp)
92{
93 struct vhost_vring_state s = {
94 .index = idx,
95 .num = vq->user_be
96 };
97
98 if (copy_to_user(argp, &s, sizeof(s)))
99 return -EFAULT;
100
101 return 0;
102}
103
104static void vhost_init_is_le(struct vhost_virtqueue *vq)
105{
106 /* Note for legacy virtio: user_be is initialized at reset time
107 * according to the host endianness. If userspace does not set an
108 * explicit endianness, the default behavior is native endian, as
109 * expected by legacy virtio.
110 */
111 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
112}
113#else
114static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
115{
116}
117
118static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
119{
120 return -ENOIOCTLCMD;
121}
122
123static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
124 int __user *argp)
125{
126 return -ENOIOCTLCMD;
127}
128
129static void vhost_init_is_le(struct vhost_virtqueue *vq)
130{
131 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
132 || virtio_legacy_is_little_endian();
133}
134#endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
135
136static void vhost_reset_is_le(struct vhost_virtqueue *vq)
137{
138 vhost_init_is_le(vq);
139}
140
141struct vhost_flush_struct {
142 struct vhost_work work;
143 struct completion wait_event;
144};
145
146static void vhost_flush_work(struct vhost_work *work)
147{
148 struct vhost_flush_struct *s;
149
150 s = container_of(work, struct vhost_flush_struct, work);
151 complete(&s->wait_event);
152}
153
154static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
155 poll_table *pt)
156{
157 struct vhost_poll *poll;
158
159 poll = container_of(pt, struct vhost_poll, table);
160 poll->wqh = wqh;
161 add_wait_queue(wqh, &poll->wait);
162}
163
164static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
165 void *key)
166{
167 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
168 struct vhost_work *work = &poll->work;
169
170 if (!(key_to_poll(key) & poll->mask))
171 return 0;
172
173 if (!poll->dev->use_worker)
174 work->fn(work);
175 else
176 vhost_poll_queue(poll);
177
178 return 0;
179}
180
181void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
182{
183 clear_bit(VHOST_WORK_QUEUED, &work->flags);
184 work->fn = fn;
185}
186EXPORT_SYMBOL_GPL(vhost_work_init);
187
188/* Init poll structure */
189void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
190 __poll_t mask, struct vhost_dev *dev)
191{
192 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
193 init_poll_funcptr(&poll->table, vhost_poll_func);
194 poll->mask = mask;
195 poll->dev = dev;
196 poll->wqh = NULL;
197
198 vhost_work_init(&poll->work, fn);
199}
200EXPORT_SYMBOL_GPL(vhost_poll_init);
201
202/* Start polling a file. We add ourselves to file's wait queue. The caller must
203 * keep a reference to a file until after vhost_poll_stop is called. */
204int vhost_poll_start(struct vhost_poll *poll, struct file *file)
205{
206 __poll_t mask;
207
208 if (poll->wqh)
209 return 0;
210
211 mask = vfs_poll(file, &poll->table);
212 if (mask)
213 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
214 if (mask & EPOLLERR) {
215 vhost_poll_stop(poll);
216 return -EINVAL;
217 }
218
219 return 0;
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_dev_flush(struct vhost_dev *dev)
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_dev_flush);
247
248void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
249{
250 if (!dev->worker)
251 return;
252
253 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
254 /* We can only add the work to the list after we're
255 * sure it was not in the list.
256 * test_and_set_bit() implies a memory barrier.
257 */
258 llist_add(&work->node, &dev->work_list);
259 wake_up_process(dev->worker);
260 }
261}
262EXPORT_SYMBOL_GPL(vhost_work_queue);
263
264/* A lockless hint for busy polling code to exit the loop */
265bool vhost_has_work(struct vhost_dev *dev)
266{
267 return !llist_empty(&dev->work_list);
268}
269EXPORT_SYMBOL_GPL(vhost_has_work);
270
271void vhost_poll_queue(struct vhost_poll *poll)
272{
273 vhost_work_queue(poll->dev, &poll->work);
274}
275EXPORT_SYMBOL_GPL(vhost_poll_queue);
276
277static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
278{
279 int j;
280
281 for (j = 0; j < VHOST_NUM_ADDRS; j++)
282 vq->meta_iotlb[j] = NULL;
283}
284
285static void vhost_vq_meta_reset(struct vhost_dev *d)
286{
287 int i;
288
289 for (i = 0; i < d->nvqs; ++i)
290 __vhost_vq_meta_reset(d->vqs[i]);
291}
292
293static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
294{
295 call_ctx->ctx = NULL;
296 memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
297}
298
299bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
300{
301 return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
302}
303EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
304
305static void vhost_vq_reset(struct vhost_dev *dev,
306 struct vhost_virtqueue *vq)
307{
308 vq->num = 1;
309 vq->desc = NULL;
310 vq->avail = NULL;
311 vq->used = NULL;
312 vq->last_avail_idx = 0;
313 vq->avail_idx = 0;
314 vq->last_used_idx = 0;
315 vq->signalled_used = 0;
316 vq->signalled_used_valid = false;
317 vq->used_flags = 0;
318 vq->log_used = false;
319 vq->log_addr = -1ull;
320 vq->private_data = NULL;
321 vq->acked_features = 0;
322 vq->acked_backend_features = 0;
323 vq->log_base = NULL;
324 vq->error_ctx = NULL;
325 vq->kick = NULL;
326 vq->log_ctx = NULL;
327 vhost_disable_cross_endian(vq);
328 vhost_reset_is_le(vq);
329 vq->busyloop_timeout = 0;
330 vq->umem = NULL;
331 vq->iotlb = NULL;
332 vhost_vring_call_reset(&vq->call_ctx);
333 __vhost_vq_meta_reset(vq);
334}
335
336static int vhost_worker(void *data)
337{
338 struct vhost_dev *dev = data;
339 struct vhost_work *work, *work_next;
340 struct llist_node *node;
341
342 kthread_use_mm(dev->mm);
343
344 for (;;) {
345 /* mb paired w/ kthread_stop */
346 set_current_state(TASK_INTERRUPTIBLE);
347
348 if (kthread_should_stop()) {
349 __set_current_state(TASK_RUNNING);
350 break;
351 }
352
353 node = llist_del_all(&dev->work_list);
354 if (!node)
355 schedule();
356
357 node = llist_reverse_order(node);
358 /* make sure flag is seen after deletion */
359 smp_wmb();
360 llist_for_each_entry_safe(work, work_next, node, node) {
361 clear_bit(VHOST_WORK_QUEUED, &work->flags);
362 __set_current_state(TASK_RUNNING);
363 kcov_remote_start_common(dev->kcov_handle);
364 work->fn(work);
365 kcov_remote_stop();
366 if (need_resched())
367 schedule();
368 }
369 }
370 kthread_unuse_mm(dev->mm);
371 return 0;
372}
373
374static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
375{
376 kfree(vq->indirect);
377 vq->indirect = NULL;
378 kfree(vq->log);
379 vq->log = NULL;
380 kfree(vq->heads);
381 vq->heads = NULL;
382}
383
384/* Helper to allocate iovec buffers for all vqs. */
385static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
386{
387 struct vhost_virtqueue *vq;
388 int i;
389
390 for (i = 0; i < dev->nvqs; ++i) {
391 vq = dev->vqs[i];
392 vq->indirect = kmalloc_array(UIO_MAXIOV,
393 sizeof(*vq->indirect),
394 GFP_KERNEL);
395 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
396 GFP_KERNEL);
397 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
398 GFP_KERNEL);
399 if (!vq->indirect || !vq->log || !vq->heads)
400 goto err_nomem;
401 }
402 return 0;
403
404err_nomem:
405 for (; i >= 0; --i)
406 vhost_vq_free_iovecs(dev->vqs[i]);
407 return -ENOMEM;
408}
409
410static void vhost_dev_free_iovecs(struct vhost_dev *dev)
411{
412 int i;
413
414 for (i = 0; i < dev->nvqs; ++i)
415 vhost_vq_free_iovecs(dev->vqs[i]);
416}
417
418bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
419 int pkts, int total_len)
420{
421 struct vhost_dev *dev = vq->dev;
422
423 if ((dev->byte_weight && total_len >= dev->byte_weight) ||
424 pkts >= dev->weight) {
425 vhost_poll_queue(&vq->poll);
426 return true;
427 }
428
429 return false;
430}
431EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
432
433static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
434 unsigned int num)
435{
436 size_t event __maybe_unused =
437 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
438
439 return sizeof(*vq->avail) +
440 sizeof(*vq->avail->ring) * num + event;
441}
442
443static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
444 unsigned int num)
445{
446 size_t event __maybe_unused =
447 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
448
449 return sizeof(*vq->used) +
450 sizeof(*vq->used->ring) * num + event;
451}
452
453static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
454 unsigned int num)
455{
456 return sizeof(*vq->desc) * num;
457}
458
459void vhost_dev_init(struct vhost_dev *dev,
460 struct vhost_virtqueue **vqs, int nvqs,
461 int iov_limit, int weight, int byte_weight,
462 bool use_worker,
463 int (*msg_handler)(struct vhost_dev *dev, u32 asid,
464 struct vhost_iotlb_msg *msg))
465{
466 struct vhost_virtqueue *vq;
467 int i;
468
469 dev->vqs = vqs;
470 dev->nvqs = nvqs;
471 mutex_init(&dev->mutex);
472 dev->log_ctx = NULL;
473 dev->umem = NULL;
474 dev->iotlb = NULL;
475 dev->mm = NULL;
476 dev->worker = NULL;
477 dev->iov_limit = iov_limit;
478 dev->weight = weight;
479 dev->byte_weight = byte_weight;
480 dev->use_worker = use_worker;
481 dev->msg_handler = msg_handler;
482 init_llist_head(&dev->work_list);
483 init_waitqueue_head(&dev->wait);
484 INIT_LIST_HEAD(&dev->read_list);
485 INIT_LIST_HEAD(&dev->pending_list);
486 spin_lock_init(&dev->iotlb_lock);
487
488
489 for (i = 0; i < dev->nvqs; ++i) {
490 vq = dev->vqs[i];
491 vq->log = NULL;
492 vq->indirect = NULL;
493 vq->heads = NULL;
494 vq->dev = dev;
495 mutex_init(&vq->mutex);
496 vhost_vq_reset(dev, vq);
497 if (vq->handle_kick)
498 vhost_poll_init(&vq->poll, vq->handle_kick,
499 EPOLLIN, dev);
500 }
501}
502EXPORT_SYMBOL_GPL(vhost_dev_init);
503
504/* Caller should have device mutex */
505long vhost_dev_check_owner(struct vhost_dev *dev)
506{
507 /* Are you the owner? If not, I don't think you mean to do that */
508 return dev->mm == current->mm ? 0 : -EPERM;
509}
510EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
511
512struct vhost_attach_cgroups_struct {
513 struct vhost_work work;
514 struct task_struct *owner;
515 int ret;
516};
517
518static void vhost_attach_cgroups_work(struct vhost_work *work)
519{
520 struct vhost_attach_cgroups_struct *s;
521
522 s = container_of(work, struct vhost_attach_cgroups_struct, work);
523 s->ret = cgroup_attach_task_all(s->owner, current);
524}
525
526static int vhost_attach_cgroups(struct vhost_dev *dev)
527{
528 struct vhost_attach_cgroups_struct attach;
529
530 attach.owner = current;
531 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
532 vhost_work_queue(dev, &attach.work);
533 vhost_dev_flush(dev);
534 return attach.ret;
535}
536
537/* Caller should have device mutex */
538bool vhost_dev_has_owner(struct vhost_dev *dev)
539{
540 return dev->mm;
541}
542EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
543
544static void vhost_attach_mm(struct vhost_dev *dev)
545{
546 /* No owner, become one */
547 if (dev->use_worker) {
548 dev->mm = get_task_mm(current);
549 } else {
550 /* vDPA device does not use worker thead, so there's
551 * no need to hold the address space for mm. This help
552 * to avoid deadlock in the case of mmap() which may
553 * held the refcnt of the file and depends on release
554 * method to remove vma.
555 */
556 dev->mm = current->mm;
557 mmgrab(dev->mm);
558 }
559}
560
561static void vhost_detach_mm(struct vhost_dev *dev)
562{
563 if (!dev->mm)
564 return;
565
566 if (dev->use_worker)
567 mmput(dev->mm);
568 else
569 mmdrop(dev->mm);
570
571 dev->mm = NULL;
572}
573
574/* Caller should have device mutex */
575long vhost_dev_set_owner(struct vhost_dev *dev)
576{
577 struct task_struct *worker;
578 int err;
579
580 /* Is there an owner already? */
581 if (vhost_dev_has_owner(dev)) {
582 err = -EBUSY;
583 goto err_mm;
584 }
585
586 vhost_attach_mm(dev);
587
588 dev->kcov_handle = kcov_common_handle();
589 if (dev->use_worker) {
590 worker = kthread_create(vhost_worker, dev,
591 "vhost-%d", current->pid);
592 if (IS_ERR(worker)) {
593 err = PTR_ERR(worker);
594 goto err_worker;
595 }
596
597 dev->worker = worker;
598 wake_up_process(worker); /* avoid contributing to loadavg */
599
600 err = vhost_attach_cgroups(dev);
601 if (err)
602 goto err_cgroup;
603 }
604
605 err = vhost_dev_alloc_iovecs(dev);
606 if (err)
607 goto err_cgroup;
608
609 return 0;
610err_cgroup:
611 if (dev->worker) {
612 kthread_stop(dev->worker);
613 dev->worker = NULL;
614 }
615err_worker:
616 vhost_detach_mm(dev);
617 dev->kcov_handle = 0;
618err_mm:
619 return err;
620}
621EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
622
623static struct vhost_iotlb *iotlb_alloc(void)
624{
625 return vhost_iotlb_alloc(max_iotlb_entries,
626 VHOST_IOTLB_FLAG_RETIRE);
627}
628
629struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
630{
631 return iotlb_alloc();
632}
633EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
634
635/* Caller should have device mutex */
636void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
637{
638 int i;
639
640 vhost_dev_cleanup(dev);
641
642 dev->umem = umem;
643 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
644 * VQs aren't running.
645 */
646 for (i = 0; i < dev->nvqs; ++i)
647 dev->vqs[i]->umem = umem;
648}
649EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
650
651void vhost_dev_stop(struct vhost_dev *dev)
652{
653 int i;
654
655 for (i = 0; i < dev->nvqs; ++i) {
656 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
657 vhost_poll_stop(&dev->vqs[i]->poll);
658 }
659
660 vhost_dev_flush(dev);
661}
662EXPORT_SYMBOL_GPL(vhost_dev_stop);
663
664void vhost_clear_msg(struct vhost_dev *dev)
665{
666 struct vhost_msg_node *node, *n;
667
668 spin_lock(&dev->iotlb_lock);
669
670 list_for_each_entry_safe(node, n, &dev->read_list, node) {
671 list_del(&node->node);
672 kfree(node);
673 }
674
675 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
676 list_del(&node->node);
677 kfree(node);
678 }
679
680 spin_unlock(&dev->iotlb_lock);
681}
682EXPORT_SYMBOL_GPL(vhost_clear_msg);
683
684void vhost_dev_cleanup(struct vhost_dev *dev)
685{
686 int i;
687
688 for (i = 0; i < dev->nvqs; ++i) {
689 if (dev->vqs[i]->error_ctx)
690 eventfd_ctx_put(dev->vqs[i]->error_ctx);
691 if (dev->vqs[i]->kick)
692 fput(dev->vqs[i]->kick);
693 if (dev->vqs[i]->call_ctx.ctx)
694 eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
695 vhost_vq_reset(dev, dev->vqs[i]);
696 }
697 vhost_dev_free_iovecs(dev);
698 if (dev->log_ctx)
699 eventfd_ctx_put(dev->log_ctx);
700 dev->log_ctx = NULL;
701 /* No one will access memory at this point */
702 vhost_iotlb_free(dev->umem);
703 dev->umem = NULL;
704 vhost_iotlb_free(dev->iotlb);
705 dev->iotlb = NULL;
706 vhost_clear_msg(dev);
707 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
708 WARN_ON(!llist_empty(&dev->work_list));
709 if (dev->worker) {
710 kthread_stop(dev->worker);
711 dev->worker = NULL;
712 dev->kcov_handle = 0;
713 }
714 vhost_detach_mm(dev);
715}
716EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
717
718static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
719{
720 u64 a = addr / VHOST_PAGE_SIZE / 8;
721
722 /* Make sure 64 bit math will not overflow. */
723 if (a > ULONG_MAX - (unsigned long)log_base ||
724 a + (unsigned long)log_base > ULONG_MAX)
725 return false;
726
727 return access_ok(log_base + a,
728 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
729}
730
731/* Make sure 64 bit math will not overflow. */
732static bool vhost_overflow(u64 uaddr, u64 size)
733{
734 if (uaddr > ULONG_MAX || size > ULONG_MAX)
735 return true;
736
737 if (!size)
738 return false;
739
740 return uaddr > ULONG_MAX - size + 1;
741}
742
743/* Caller should have vq mutex and device mutex. */
744static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
745 int log_all)
746{
747 struct vhost_iotlb_map *map;
748
749 if (!umem)
750 return false;
751
752 list_for_each_entry(map, &umem->list, link) {
753 unsigned long a = map->addr;
754
755 if (vhost_overflow(map->addr, map->size))
756 return false;
757
758
759 if (!access_ok((void __user *)a, map->size))
760 return false;
761 else if (log_all && !log_access_ok(log_base,
762 map->start,
763 map->size))
764 return false;
765 }
766 return true;
767}
768
769static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
770 u64 addr, unsigned int size,
771 int type)
772{
773 const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
774
775 if (!map)
776 return NULL;
777
778 return (void __user *)(uintptr_t)(map->addr + addr - map->start);
779}
780
781/* Can we switch to this memory table? */
782/* Caller should have device mutex but not vq mutex */
783static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
784 int log_all)
785{
786 int i;
787
788 for (i = 0; i < d->nvqs; ++i) {
789 bool ok;
790 bool log;
791
792 mutex_lock(&d->vqs[i]->mutex);
793 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
794 /* If ring is inactive, will check when it's enabled. */
795 if (d->vqs[i]->private_data)
796 ok = vq_memory_access_ok(d->vqs[i]->log_base,
797 umem, log);
798 else
799 ok = true;
800 mutex_unlock(&d->vqs[i]->mutex);
801 if (!ok)
802 return false;
803 }
804 return true;
805}
806
807static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
808 struct iovec iov[], int iov_size, int access);
809
810static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
811 const void *from, unsigned size)
812{
813 int ret;
814
815 if (!vq->iotlb)
816 return __copy_to_user(to, from, size);
817 else {
818 /* This function should be called after iotlb
819 * prefetch, which means we're sure that all vq
820 * could be access through iotlb. So -EAGAIN should
821 * not happen in this case.
822 */
823 struct iov_iter t;
824 void __user *uaddr = vhost_vq_meta_fetch(vq,
825 (u64)(uintptr_t)to, size,
826 VHOST_ADDR_USED);
827
828 if (uaddr)
829 return __copy_to_user(uaddr, from, size);
830
831 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
832 ARRAY_SIZE(vq->iotlb_iov),
833 VHOST_ACCESS_WO);
834 if (ret < 0)
835 goto out;
836 iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
837 ret = copy_to_iter(from, size, &t);
838 if (ret == size)
839 ret = 0;
840 }
841out:
842 return ret;
843}
844
845static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
846 void __user *from, unsigned size)
847{
848 int ret;
849
850 if (!vq->iotlb)
851 return __copy_from_user(to, from, size);
852 else {
853 /* This function should be called after iotlb
854 * prefetch, which means we're sure that vq
855 * could be access through iotlb. So -EAGAIN should
856 * not happen in this case.
857 */
858 void __user *uaddr = vhost_vq_meta_fetch(vq,
859 (u64)(uintptr_t)from, size,
860 VHOST_ADDR_DESC);
861 struct iov_iter f;
862
863 if (uaddr)
864 return __copy_from_user(to, uaddr, size);
865
866 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
867 ARRAY_SIZE(vq->iotlb_iov),
868 VHOST_ACCESS_RO);
869 if (ret < 0) {
870 vq_err(vq, "IOTLB translation failure: uaddr "
871 "%p size 0x%llx\n", from,
872 (unsigned long long) size);
873 goto out;
874 }
875 iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
876 ret = copy_from_iter(to, size, &f);
877 if (ret == size)
878 ret = 0;
879 }
880
881out:
882 return ret;
883}
884
885static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
886 void __user *addr, unsigned int size,
887 int type)
888{
889 int ret;
890
891 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
892 ARRAY_SIZE(vq->iotlb_iov),
893 VHOST_ACCESS_RO);
894 if (ret < 0) {
895 vq_err(vq, "IOTLB translation failure: uaddr "
896 "%p size 0x%llx\n", addr,
897 (unsigned long long) size);
898 return NULL;
899 }
900
901 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
902 vq_err(vq, "Non atomic userspace memory access: uaddr "
903 "%p size 0x%llx\n", addr,
904 (unsigned long long) size);
905 return NULL;
906 }
907
908 return vq->iotlb_iov[0].iov_base;
909}
910
911/* This function should be called after iotlb
912 * prefetch, which means we're sure that vq
913 * could be access through iotlb. So -EAGAIN should
914 * not happen in this case.
915 */
916static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
917 void __user *addr, unsigned int size,
918 int type)
919{
920 void __user *uaddr = vhost_vq_meta_fetch(vq,
921 (u64)(uintptr_t)addr, size, type);
922 if (uaddr)
923 return uaddr;
924
925 return __vhost_get_user_slow(vq, addr, size, type);
926}
927
928#define vhost_put_user(vq, x, ptr) \
929({ \
930 int ret; \
931 if (!vq->iotlb) { \
932 ret = __put_user(x, ptr); \
933 } else { \
934 __typeof__(ptr) to = \
935 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
936 sizeof(*ptr), VHOST_ADDR_USED); \
937 if (to != NULL) \
938 ret = __put_user(x, to); \
939 else \
940 ret = -EFAULT; \
941 } \
942 ret; \
943})
944
945static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
946{
947 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
948 vhost_avail_event(vq));
949}
950
951static inline int vhost_put_used(struct vhost_virtqueue *vq,
952 struct vring_used_elem *head, int idx,
953 int count)
954{
955 return vhost_copy_to_user(vq, vq->used->ring + idx, head,
956 count * sizeof(*head));
957}
958
959static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
960
961{
962 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
963 &vq->used->flags);
964}
965
966static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
967
968{
969 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
970 &vq->used->idx);
971}
972
973#define vhost_get_user(vq, x, ptr, type) \
974({ \
975 int ret; \
976 if (!vq->iotlb) { \
977 ret = __get_user(x, ptr); \
978 } else { \
979 __typeof__(ptr) from = \
980 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
981 sizeof(*ptr), \
982 type); \
983 if (from != NULL) \
984 ret = __get_user(x, from); \
985 else \
986 ret = -EFAULT; \
987 } \
988 ret; \
989})
990
991#define vhost_get_avail(vq, x, ptr) \
992 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
993
994#define vhost_get_used(vq, x, ptr) \
995 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
996
997static void vhost_dev_lock_vqs(struct vhost_dev *d)
998{
999 int i = 0;
1000 for (i = 0; i < d->nvqs; ++i)
1001 mutex_lock_nested(&d->vqs[i]->mutex, i);
1002}
1003
1004static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1005{
1006 int i = 0;
1007 for (i = 0; i < d->nvqs; ++i)
1008 mutex_unlock(&d->vqs[i]->mutex);
1009}
1010
1011static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1012 __virtio16 *idx)
1013{
1014 return vhost_get_avail(vq, *idx, &vq->avail->idx);
1015}
1016
1017static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1018 __virtio16 *head, int idx)
1019{
1020 return vhost_get_avail(vq, *head,
1021 &vq->avail->ring[idx & (vq->num - 1)]);
1022}
1023
1024static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1025 __virtio16 *flags)
1026{
1027 return vhost_get_avail(vq, *flags, &vq->avail->flags);
1028}
1029
1030static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1031 __virtio16 *event)
1032{
1033 return vhost_get_avail(vq, *event, vhost_used_event(vq));
1034}
1035
1036static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1037 __virtio16 *idx)
1038{
1039 return vhost_get_used(vq, *idx, &vq->used->idx);
1040}
1041
1042static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1043 struct vring_desc *desc, int idx)
1044{
1045 return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1046}
1047
1048static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1049 struct vhost_iotlb_msg *msg)
1050{
1051 struct vhost_msg_node *node, *n;
1052
1053 spin_lock(&d->iotlb_lock);
1054
1055 list_for_each_entry_safe(node, n, &d->pending_list, node) {
1056 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1057 if (msg->iova <= vq_msg->iova &&
1058 msg->iova + msg->size - 1 >= vq_msg->iova &&
1059 vq_msg->type == VHOST_IOTLB_MISS) {
1060 vhost_poll_queue(&node->vq->poll);
1061 list_del(&node->node);
1062 kfree(node);
1063 }
1064 }
1065
1066 spin_unlock(&d->iotlb_lock);
1067}
1068
1069static bool umem_access_ok(u64 uaddr, u64 size, int access)
1070{
1071 unsigned long a = uaddr;
1072
1073 /* Make sure 64 bit math will not overflow. */
1074 if (vhost_overflow(uaddr, size))
1075 return false;
1076
1077 if ((access & VHOST_ACCESS_RO) &&
1078 !access_ok((void __user *)a, size))
1079 return false;
1080 if ((access & VHOST_ACCESS_WO) &&
1081 !access_ok((void __user *)a, size))
1082 return false;
1083 return true;
1084}
1085
1086static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1087 struct vhost_iotlb_msg *msg)
1088{
1089 int ret = 0;
1090
1091 if (asid != 0)
1092 return -EINVAL;
1093
1094 mutex_lock(&dev->mutex);
1095 vhost_dev_lock_vqs(dev);
1096 switch (msg->type) {
1097 case VHOST_IOTLB_UPDATE:
1098 if (!dev->iotlb) {
1099 ret = -EFAULT;
1100 break;
1101 }
1102 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1103 ret = -EFAULT;
1104 break;
1105 }
1106 vhost_vq_meta_reset(dev);
1107 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1108 msg->iova + msg->size - 1,
1109 msg->uaddr, msg->perm)) {
1110 ret = -ENOMEM;
1111 break;
1112 }
1113 vhost_iotlb_notify_vq(dev, msg);
1114 break;
1115 case VHOST_IOTLB_INVALIDATE:
1116 if (!dev->iotlb) {
1117 ret = -EFAULT;
1118 break;
1119 }
1120 vhost_vq_meta_reset(dev);
1121 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1122 msg->iova + msg->size - 1);
1123 break;
1124 default:
1125 ret = -EINVAL;
1126 break;
1127 }
1128
1129 vhost_dev_unlock_vqs(dev);
1130 mutex_unlock(&dev->mutex);
1131
1132 return ret;
1133}
1134ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1135 struct iov_iter *from)
1136{
1137 struct vhost_iotlb_msg msg;
1138 size_t offset;
1139 int type, ret;
1140 u32 asid = 0;
1141
1142 ret = copy_from_iter(&type, sizeof(type), from);
1143 if (ret != sizeof(type)) {
1144 ret = -EINVAL;
1145 goto done;
1146 }
1147
1148 switch (type) {
1149 case VHOST_IOTLB_MSG:
1150 /* There maybe a hole after type for V1 message type,
1151 * so skip it here.
1152 */
1153 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1154 break;
1155 case VHOST_IOTLB_MSG_V2:
1156 if (vhost_backend_has_feature(dev->vqs[0],
1157 VHOST_BACKEND_F_IOTLB_ASID)) {
1158 ret = copy_from_iter(&asid, sizeof(asid), from);
1159 if (ret != sizeof(asid)) {
1160 ret = -EINVAL;
1161 goto done;
1162 }
1163 offset = 0;
1164 } else
1165 offset = sizeof(__u32);
1166 break;
1167 default:
1168 ret = -EINVAL;
1169 goto done;
1170 }
1171
1172 iov_iter_advance(from, offset);
1173 ret = copy_from_iter(&msg, sizeof(msg), from);
1174 if (ret != sizeof(msg)) {
1175 ret = -EINVAL;
1176 goto done;
1177 }
1178
1179 if ((msg.type == VHOST_IOTLB_UPDATE ||
1180 msg.type == VHOST_IOTLB_INVALIDATE) &&
1181 msg.size == 0) {
1182 ret = -EINVAL;
1183 goto done;
1184 }
1185
1186 if (dev->msg_handler)
1187 ret = dev->msg_handler(dev, asid, &msg);
1188 else
1189 ret = vhost_process_iotlb_msg(dev, asid, &msg);
1190 if (ret) {
1191 ret = -EFAULT;
1192 goto done;
1193 }
1194
1195 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1196 sizeof(struct vhost_msg_v2);
1197done:
1198 return ret;
1199}
1200EXPORT_SYMBOL(vhost_chr_write_iter);
1201
1202__poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1203 poll_table *wait)
1204{
1205 __poll_t mask = 0;
1206
1207 poll_wait(file, &dev->wait, wait);
1208
1209 if (!list_empty(&dev->read_list))
1210 mask |= EPOLLIN | EPOLLRDNORM;
1211
1212 return mask;
1213}
1214EXPORT_SYMBOL(vhost_chr_poll);
1215
1216ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1217 int noblock)
1218{
1219 DEFINE_WAIT(wait);
1220 struct vhost_msg_node *node;
1221 ssize_t ret = 0;
1222 unsigned size = sizeof(struct vhost_msg);
1223
1224 if (iov_iter_count(to) < size)
1225 return 0;
1226
1227 while (1) {
1228 if (!noblock)
1229 prepare_to_wait(&dev->wait, &wait,
1230 TASK_INTERRUPTIBLE);
1231
1232 node = vhost_dequeue_msg(dev, &dev->read_list);
1233 if (node)
1234 break;
1235 if (noblock) {
1236 ret = -EAGAIN;
1237 break;
1238 }
1239 if (signal_pending(current)) {
1240 ret = -ERESTARTSYS;
1241 break;
1242 }
1243 if (!dev->iotlb) {
1244 ret = -EBADFD;
1245 break;
1246 }
1247
1248 schedule();
1249 }
1250
1251 if (!noblock)
1252 finish_wait(&dev->wait, &wait);
1253
1254 if (node) {
1255 struct vhost_iotlb_msg *msg;
1256 void *start = &node->msg;
1257
1258 switch (node->msg.type) {
1259 case VHOST_IOTLB_MSG:
1260 size = sizeof(node->msg);
1261 msg = &node->msg.iotlb;
1262 break;
1263 case VHOST_IOTLB_MSG_V2:
1264 size = sizeof(node->msg_v2);
1265 msg = &node->msg_v2.iotlb;
1266 break;
1267 default:
1268 BUG();
1269 break;
1270 }
1271
1272 ret = copy_to_iter(start, size, to);
1273 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1274 kfree(node);
1275 return ret;
1276 }
1277 vhost_enqueue_msg(dev, &dev->pending_list, node);
1278 }
1279
1280 return ret;
1281}
1282EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1283
1284static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1285{
1286 struct vhost_dev *dev = vq->dev;
1287 struct vhost_msg_node *node;
1288 struct vhost_iotlb_msg *msg;
1289 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1290
1291 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1292 if (!node)
1293 return -ENOMEM;
1294
1295 if (v2) {
1296 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1297 msg = &node->msg_v2.iotlb;
1298 } else {
1299 msg = &node->msg.iotlb;
1300 }
1301
1302 msg->type = VHOST_IOTLB_MISS;
1303 msg->iova = iova;
1304 msg->perm = access;
1305
1306 vhost_enqueue_msg(dev, &dev->read_list, node);
1307
1308 return 0;
1309}
1310
1311static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1312 vring_desc_t __user *desc,
1313 vring_avail_t __user *avail,
1314 vring_used_t __user *used)
1315
1316{
1317 /* If an IOTLB device is present, the vring addresses are
1318 * GIOVAs. Access validation occurs at prefetch time. */
1319 if (vq->iotlb)
1320 return true;
1321
1322 return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1323 access_ok(avail, vhost_get_avail_size(vq, num)) &&
1324 access_ok(used, vhost_get_used_size(vq, num));
1325}
1326
1327static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1328 const struct vhost_iotlb_map *map,
1329 int type)
1330{
1331 int access = (type == VHOST_ADDR_USED) ?
1332 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1333
1334 if (likely(map->perm & access))
1335 vq->meta_iotlb[type] = map;
1336}
1337
1338static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1339 int access, u64 addr, u64 len, int type)
1340{
1341 const struct vhost_iotlb_map *map;
1342 struct vhost_iotlb *umem = vq->iotlb;
1343 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1344
1345 if (vhost_vq_meta_fetch(vq, addr, len, type))
1346 return true;
1347
1348 while (len > s) {
1349 map = vhost_iotlb_itree_first(umem, addr, last);
1350 if (map == NULL || map->start > addr) {
1351 vhost_iotlb_miss(vq, addr, access);
1352 return false;
1353 } else if (!(map->perm & access)) {
1354 /* Report the possible access violation by
1355 * request another translation from userspace.
1356 */
1357 return false;
1358 }
1359
1360 size = map->size - addr + map->start;
1361
1362 if (orig_addr == addr && size >= len)
1363 vhost_vq_meta_update(vq, map, type);
1364
1365 s += size;
1366 addr += size;
1367 }
1368
1369 return true;
1370}
1371
1372int vq_meta_prefetch(struct vhost_virtqueue *vq)
1373{
1374 unsigned int num = vq->num;
1375
1376 if (!vq->iotlb)
1377 return 1;
1378
1379 return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1380 vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1381 iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1382 vhost_get_avail_size(vq, num),
1383 VHOST_ADDR_AVAIL) &&
1384 iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1385 vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1386}
1387EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1388
1389/* Can we log writes? */
1390/* Caller should have device mutex but not vq mutex */
1391bool vhost_log_access_ok(struct vhost_dev *dev)
1392{
1393 return memory_access_ok(dev, dev->umem, 1);
1394}
1395EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1396
1397static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1398 void __user *log_base,
1399 bool log_used,
1400 u64 log_addr)
1401{
1402 /* If an IOTLB device is present, log_addr is a GIOVA that
1403 * will never be logged by log_used(). */
1404 if (vq->iotlb)
1405 return true;
1406
1407 return !log_used || log_access_ok(log_base, log_addr,
1408 vhost_get_used_size(vq, vq->num));
1409}
1410
1411/* Verify access for write logging. */
1412/* Caller should have vq mutex and device mutex */
1413static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1414 void __user *log_base)
1415{
1416 return vq_memory_access_ok(log_base, vq->umem,
1417 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1418 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1419}
1420
1421/* Can we start vq? */
1422/* Caller should have vq mutex and device mutex */
1423bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1424{
1425 if (!vq_log_access_ok(vq, vq->log_base))
1426 return false;
1427
1428 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1429}
1430EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1431
1432static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1433{
1434 struct vhost_memory mem, *newmem;
1435 struct vhost_memory_region *region;
1436 struct vhost_iotlb *newumem, *oldumem;
1437 unsigned long size = offsetof(struct vhost_memory, regions);
1438 int i;
1439
1440 if (copy_from_user(&mem, m, size))
1441 return -EFAULT;
1442 if (mem.padding)
1443 return -EOPNOTSUPP;
1444 if (mem.nregions > max_mem_regions)
1445 return -E2BIG;
1446 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1447 GFP_KERNEL);
1448 if (!newmem)
1449 return -ENOMEM;
1450
1451 memcpy(newmem, &mem, size);
1452 if (copy_from_user(newmem->regions, m->regions,
1453 flex_array_size(newmem, regions, mem.nregions))) {
1454 kvfree(newmem);
1455 return -EFAULT;
1456 }
1457
1458 newumem = iotlb_alloc();
1459 if (!newumem) {
1460 kvfree(newmem);
1461 return -ENOMEM;
1462 }
1463
1464 for (region = newmem->regions;
1465 region < newmem->regions + mem.nregions;
1466 region++) {
1467 if (vhost_iotlb_add_range(newumem,
1468 region->guest_phys_addr,
1469 region->guest_phys_addr +
1470 region->memory_size - 1,
1471 region->userspace_addr,
1472 VHOST_MAP_RW))
1473 goto err;
1474 }
1475
1476 if (!memory_access_ok(d, newumem, 0))
1477 goto err;
1478
1479 oldumem = d->umem;
1480 d->umem = newumem;
1481
1482 /* All memory accesses are done under some VQ mutex. */
1483 for (i = 0; i < d->nvqs; ++i) {
1484 mutex_lock(&d->vqs[i]->mutex);
1485 d->vqs[i]->umem = newumem;
1486 mutex_unlock(&d->vqs[i]->mutex);
1487 }
1488
1489 kvfree(newmem);
1490 vhost_iotlb_free(oldumem);
1491 return 0;
1492
1493err:
1494 vhost_iotlb_free(newumem);
1495 kvfree(newmem);
1496 return -EFAULT;
1497}
1498
1499static long vhost_vring_set_num(struct vhost_dev *d,
1500 struct vhost_virtqueue *vq,
1501 void __user *argp)
1502{
1503 struct vhost_vring_state s;
1504
1505 /* Resizing ring with an active backend?
1506 * You don't want to do that. */
1507 if (vq->private_data)
1508 return -EBUSY;
1509
1510 if (copy_from_user(&s, argp, sizeof s))
1511 return -EFAULT;
1512
1513 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1514 return -EINVAL;
1515 vq->num = s.num;
1516
1517 return 0;
1518}
1519
1520static long vhost_vring_set_addr(struct vhost_dev *d,
1521 struct vhost_virtqueue *vq,
1522 void __user *argp)
1523{
1524 struct vhost_vring_addr a;
1525
1526 if (copy_from_user(&a, argp, sizeof a))
1527 return -EFAULT;
1528 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1529 return -EOPNOTSUPP;
1530
1531 /* For 32bit, verify that the top 32bits of the user
1532 data are set to zero. */
1533 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1534 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1535 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1536 return -EFAULT;
1537
1538 /* Make sure it's safe to cast pointers to vring types. */
1539 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1540 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1541 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1542 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1543 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1544 return -EINVAL;
1545
1546 /* We only verify access here if backend is configured.
1547 * If it is not, we don't as size might not have been setup.
1548 * We will verify when backend is configured. */
1549 if (vq->private_data) {
1550 if (!vq_access_ok(vq, vq->num,
1551 (void __user *)(unsigned long)a.desc_user_addr,
1552 (void __user *)(unsigned long)a.avail_user_addr,
1553 (void __user *)(unsigned long)a.used_user_addr))
1554 return -EINVAL;
1555
1556 /* Also validate log access for used ring if enabled. */
1557 if (!vq_log_used_access_ok(vq, vq->log_base,
1558 a.flags & (0x1 << VHOST_VRING_F_LOG),
1559 a.log_guest_addr))
1560 return -EINVAL;
1561 }
1562
1563 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1564 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1565 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1566 vq->log_addr = a.log_guest_addr;
1567 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1568
1569 return 0;
1570}
1571
1572static long vhost_vring_set_num_addr(struct vhost_dev *d,
1573 struct vhost_virtqueue *vq,
1574 unsigned int ioctl,
1575 void __user *argp)
1576{
1577 long r;
1578
1579 mutex_lock(&vq->mutex);
1580
1581 switch (ioctl) {
1582 case VHOST_SET_VRING_NUM:
1583 r = vhost_vring_set_num(d, vq, argp);
1584 break;
1585 case VHOST_SET_VRING_ADDR:
1586 r = vhost_vring_set_addr(d, vq, argp);
1587 break;
1588 default:
1589 BUG();
1590 }
1591
1592 mutex_unlock(&vq->mutex);
1593
1594 return r;
1595}
1596long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1597{
1598 struct file *eventfp, *filep = NULL;
1599 bool pollstart = false, pollstop = false;
1600 struct eventfd_ctx *ctx = NULL;
1601 u32 __user *idxp = argp;
1602 struct vhost_virtqueue *vq;
1603 struct vhost_vring_state s;
1604 struct vhost_vring_file f;
1605 u32 idx;
1606 long r;
1607
1608 r = get_user(idx, idxp);
1609 if (r < 0)
1610 return r;
1611 if (idx >= d->nvqs)
1612 return -ENOBUFS;
1613
1614 idx = array_index_nospec(idx, d->nvqs);
1615 vq = d->vqs[idx];
1616
1617 if (ioctl == VHOST_SET_VRING_NUM ||
1618 ioctl == VHOST_SET_VRING_ADDR) {
1619 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1620 }
1621
1622 mutex_lock(&vq->mutex);
1623
1624 switch (ioctl) {
1625 case VHOST_SET_VRING_BASE:
1626 /* Moving base with an active backend?
1627 * You don't want to do that. */
1628 if (vq->private_data) {
1629 r = -EBUSY;
1630 break;
1631 }
1632 if (copy_from_user(&s, argp, sizeof s)) {
1633 r = -EFAULT;
1634 break;
1635 }
1636 if (s.num > 0xffff) {
1637 r = -EINVAL;
1638 break;
1639 }
1640 vq->last_avail_idx = s.num;
1641 /* Forget the cached index value. */
1642 vq->avail_idx = vq->last_avail_idx;
1643 break;
1644 case VHOST_GET_VRING_BASE:
1645 s.index = idx;
1646 s.num = vq->last_avail_idx;
1647 if (copy_to_user(argp, &s, sizeof s))
1648 r = -EFAULT;
1649 break;
1650 case VHOST_SET_VRING_KICK:
1651 if (copy_from_user(&f, argp, sizeof f)) {
1652 r = -EFAULT;
1653 break;
1654 }
1655 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1656 if (IS_ERR(eventfp)) {
1657 r = PTR_ERR(eventfp);
1658 break;
1659 }
1660 if (eventfp != vq->kick) {
1661 pollstop = (filep = vq->kick) != NULL;
1662 pollstart = (vq->kick = eventfp) != NULL;
1663 } else
1664 filep = eventfp;
1665 break;
1666 case VHOST_SET_VRING_CALL:
1667 if (copy_from_user(&f, argp, sizeof f)) {
1668 r = -EFAULT;
1669 break;
1670 }
1671 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1672 if (IS_ERR(ctx)) {
1673 r = PTR_ERR(ctx);
1674 break;
1675 }
1676
1677 swap(ctx, vq->call_ctx.ctx);
1678 break;
1679 case VHOST_SET_VRING_ERR:
1680 if (copy_from_user(&f, argp, sizeof f)) {
1681 r = -EFAULT;
1682 break;
1683 }
1684 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1685 if (IS_ERR(ctx)) {
1686 r = PTR_ERR(ctx);
1687 break;
1688 }
1689 swap(ctx, vq->error_ctx);
1690 break;
1691 case VHOST_SET_VRING_ENDIAN:
1692 r = vhost_set_vring_endian(vq, argp);
1693 break;
1694 case VHOST_GET_VRING_ENDIAN:
1695 r = vhost_get_vring_endian(vq, idx, argp);
1696 break;
1697 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1698 if (copy_from_user(&s, argp, sizeof(s))) {
1699 r = -EFAULT;
1700 break;
1701 }
1702 vq->busyloop_timeout = s.num;
1703 break;
1704 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1705 s.index = idx;
1706 s.num = vq->busyloop_timeout;
1707 if (copy_to_user(argp, &s, sizeof(s)))
1708 r = -EFAULT;
1709 break;
1710 default:
1711 r = -ENOIOCTLCMD;
1712 }
1713
1714 if (pollstop && vq->handle_kick)
1715 vhost_poll_stop(&vq->poll);
1716
1717 if (!IS_ERR_OR_NULL(ctx))
1718 eventfd_ctx_put(ctx);
1719 if (filep)
1720 fput(filep);
1721
1722 if (pollstart && vq->handle_kick)
1723 r = vhost_poll_start(&vq->poll, vq->kick);
1724
1725 mutex_unlock(&vq->mutex);
1726
1727 if (pollstop && vq->handle_kick)
1728 vhost_dev_flush(vq->poll.dev);
1729 return r;
1730}
1731EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1732
1733int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1734{
1735 struct vhost_iotlb *niotlb, *oiotlb;
1736 int i;
1737
1738 niotlb = iotlb_alloc();
1739 if (!niotlb)
1740 return -ENOMEM;
1741
1742 oiotlb = d->iotlb;
1743 d->iotlb = niotlb;
1744
1745 for (i = 0; i < d->nvqs; ++i) {
1746 struct vhost_virtqueue *vq = d->vqs[i];
1747
1748 mutex_lock(&vq->mutex);
1749 vq->iotlb = niotlb;
1750 __vhost_vq_meta_reset(vq);
1751 mutex_unlock(&vq->mutex);
1752 }
1753
1754 vhost_iotlb_free(oiotlb);
1755
1756 return 0;
1757}
1758EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1759
1760/* Caller must have device mutex */
1761long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1762{
1763 struct eventfd_ctx *ctx;
1764 u64 p;
1765 long r;
1766 int i, fd;
1767
1768 /* If you are not the owner, you can become one */
1769 if (ioctl == VHOST_SET_OWNER) {
1770 r = vhost_dev_set_owner(d);
1771 goto done;
1772 }
1773
1774 /* You must be the owner to do anything else */
1775 r = vhost_dev_check_owner(d);
1776 if (r)
1777 goto done;
1778
1779 switch (ioctl) {
1780 case VHOST_SET_MEM_TABLE:
1781 r = vhost_set_memory(d, argp);
1782 break;
1783 case VHOST_SET_LOG_BASE:
1784 if (copy_from_user(&p, argp, sizeof p)) {
1785 r = -EFAULT;
1786 break;
1787 }
1788 if ((u64)(unsigned long)p != p) {
1789 r = -EFAULT;
1790 break;
1791 }
1792 for (i = 0; i < d->nvqs; ++i) {
1793 struct vhost_virtqueue *vq;
1794 void __user *base = (void __user *)(unsigned long)p;
1795 vq = d->vqs[i];
1796 mutex_lock(&vq->mutex);
1797 /* If ring is inactive, will check when it's enabled. */
1798 if (vq->private_data && !vq_log_access_ok(vq, base))
1799 r = -EFAULT;
1800 else
1801 vq->log_base = base;
1802 mutex_unlock(&vq->mutex);
1803 }
1804 break;
1805 case VHOST_SET_LOG_FD:
1806 r = get_user(fd, (int __user *)argp);
1807 if (r < 0)
1808 break;
1809 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1810 if (IS_ERR(ctx)) {
1811 r = PTR_ERR(ctx);
1812 break;
1813 }
1814 swap(ctx, d->log_ctx);
1815 for (i = 0; i < d->nvqs; ++i) {
1816 mutex_lock(&d->vqs[i]->mutex);
1817 d->vqs[i]->log_ctx = d->log_ctx;
1818 mutex_unlock(&d->vqs[i]->mutex);
1819 }
1820 if (ctx)
1821 eventfd_ctx_put(ctx);
1822 break;
1823 default:
1824 r = -ENOIOCTLCMD;
1825 break;
1826 }
1827done:
1828 return r;
1829}
1830EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1831
1832/* TODO: This is really inefficient. We need something like get_user()
1833 * (instruction directly accesses the data, with an exception table entry
1834 * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1835 */
1836static int set_bit_to_user(int nr, void __user *addr)
1837{
1838 unsigned long log = (unsigned long)addr;
1839 struct page *page;
1840 void *base;
1841 int bit = nr + (log % PAGE_SIZE) * 8;
1842 int r;
1843
1844 r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1845 if (r < 0)
1846 return r;
1847 BUG_ON(r != 1);
1848 base = kmap_atomic(page);
1849 set_bit(bit, base);
1850 kunmap_atomic(base);
1851 unpin_user_pages_dirty_lock(&page, 1, true);
1852 return 0;
1853}
1854
1855static int log_write(void __user *log_base,
1856 u64 write_address, u64 write_length)
1857{
1858 u64 write_page = write_address / VHOST_PAGE_SIZE;
1859 int r;
1860
1861 if (!write_length)
1862 return 0;
1863 write_length += write_address % VHOST_PAGE_SIZE;
1864 for (;;) {
1865 u64 base = (u64)(unsigned long)log_base;
1866 u64 log = base + write_page / 8;
1867 int bit = write_page % 8;
1868 if ((u64)(unsigned long)log != log)
1869 return -EFAULT;
1870 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1871 if (r < 0)
1872 return r;
1873 if (write_length <= VHOST_PAGE_SIZE)
1874 break;
1875 write_length -= VHOST_PAGE_SIZE;
1876 write_page += 1;
1877 }
1878 return r;
1879}
1880
1881static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1882{
1883 struct vhost_iotlb *umem = vq->umem;
1884 struct vhost_iotlb_map *u;
1885 u64 start, end, l, min;
1886 int r;
1887 bool hit = false;
1888
1889 while (len) {
1890 min = len;
1891 /* More than one GPAs can be mapped into a single HVA. So
1892 * iterate all possible umems here to be safe.
1893 */
1894 list_for_each_entry(u, &umem->list, link) {
1895 if (u->addr > hva - 1 + len ||
1896 u->addr - 1 + u->size < hva)
1897 continue;
1898 start = max(u->addr, hva);
1899 end = min(u->addr - 1 + u->size, hva - 1 + len);
1900 l = end - start + 1;
1901 r = log_write(vq->log_base,
1902 u->start + start - u->addr,
1903 l);
1904 if (r < 0)
1905 return r;
1906 hit = true;
1907 min = min(l, min);
1908 }
1909
1910 if (!hit)
1911 return -EFAULT;
1912
1913 len -= min;
1914 hva += min;
1915 }
1916
1917 return 0;
1918}
1919
1920static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1921{
1922 struct iovec *iov = vq->log_iov;
1923 int i, ret;
1924
1925 if (!vq->iotlb)
1926 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1927
1928 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1929 len, iov, 64, VHOST_ACCESS_WO);
1930 if (ret < 0)
1931 return ret;
1932
1933 for (i = 0; i < ret; i++) {
1934 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1935 iov[i].iov_len);
1936 if (ret)
1937 return ret;
1938 }
1939
1940 return 0;
1941}
1942
1943int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1944 unsigned int log_num, u64 len, struct iovec *iov, int count)
1945{
1946 int i, r;
1947
1948 /* Make sure data written is seen before log. */
1949 smp_wmb();
1950
1951 if (vq->iotlb) {
1952 for (i = 0; i < count; i++) {
1953 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1954 iov[i].iov_len);
1955 if (r < 0)
1956 return r;
1957 }
1958 return 0;
1959 }
1960
1961 for (i = 0; i < log_num; ++i) {
1962 u64 l = min(log[i].len, len);
1963 r = log_write(vq->log_base, log[i].addr, l);
1964 if (r < 0)
1965 return r;
1966 len -= l;
1967 if (!len) {
1968 if (vq->log_ctx)
1969 eventfd_signal(vq->log_ctx, 1);
1970 return 0;
1971 }
1972 }
1973 /* Length written exceeds what we have stored. This is a bug. */
1974 BUG();
1975 return 0;
1976}
1977EXPORT_SYMBOL_GPL(vhost_log_write);
1978
1979static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1980{
1981 void __user *used;
1982 if (vhost_put_used_flags(vq))
1983 return -EFAULT;
1984 if (unlikely(vq->log_used)) {
1985 /* Make sure the flag is seen before log. */
1986 smp_wmb();
1987 /* Log used flag write. */
1988 used = &vq->used->flags;
1989 log_used(vq, (used - (void __user *)vq->used),
1990 sizeof vq->used->flags);
1991 if (vq->log_ctx)
1992 eventfd_signal(vq->log_ctx, 1);
1993 }
1994 return 0;
1995}
1996
1997static int vhost_update_avail_event(struct vhost_virtqueue *vq)
1998{
1999 if (vhost_put_avail_event(vq))
2000 return -EFAULT;
2001 if (unlikely(vq->log_used)) {
2002 void __user *used;
2003 /* Make sure the event is seen before log. */
2004 smp_wmb();
2005 /* Log avail event write */
2006 used = vhost_avail_event(vq);
2007 log_used(vq, (used - (void __user *)vq->used),
2008 sizeof *vhost_avail_event(vq));
2009 if (vq->log_ctx)
2010 eventfd_signal(vq->log_ctx, 1);
2011 }
2012 return 0;
2013}
2014
2015int vhost_vq_init_access(struct vhost_virtqueue *vq)
2016{
2017 __virtio16 last_used_idx;
2018 int r;
2019 bool is_le = vq->is_le;
2020
2021 if (!vq->private_data)
2022 return 0;
2023
2024 vhost_init_is_le(vq);
2025
2026 r = vhost_update_used_flags(vq);
2027 if (r)
2028 goto err;
2029 vq->signalled_used_valid = false;
2030 if (!vq->iotlb &&
2031 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2032 r = -EFAULT;
2033 goto err;
2034 }
2035 r = vhost_get_used_idx(vq, &last_used_idx);
2036 if (r) {
2037 vq_err(vq, "Can't access used idx at %p\n",
2038 &vq->used->idx);
2039 goto err;
2040 }
2041 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2042 return 0;
2043
2044err:
2045 vq->is_le = is_le;
2046 return r;
2047}
2048EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2049
2050static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2051 struct iovec iov[], int iov_size, int access)
2052{
2053 const struct vhost_iotlb_map *map;
2054 struct vhost_dev *dev = vq->dev;
2055 struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2056 struct iovec *_iov;
2057 u64 s = 0, last = addr + len - 1;
2058 int ret = 0;
2059
2060 while ((u64)len > s) {
2061 u64 size;
2062 if (unlikely(ret >= iov_size)) {
2063 ret = -ENOBUFS;
2064 break;
2065 }
2066
2067 map = vhost_iotlb_itree_first(umem, addr, last);
2068 if (map == NULL || map->start > addr) {
2069 if (umem != dev->iotlb) {
2070 ret = -EFAULT;
2071 break;
2072 }
2073 ret = -EAGAIN;
2074 break;
2075 } else if (!(map->perm & access)) {
2076 ret = -EPERM;
2077 break;
2078 }
2079
2080 _iov = iov + ret;
2081 size = map->size - addr + map->start;
2082 _iov->iov_len = min((u64)len - s, size);
2083 _iov->iov_base = (void __user *)(unsigned long)
2084 (map->addr + addr - map->start);
2085 s += size;
2086 addr += size;
2087 ++ret;
2088 }
2089
2090 if (ret == -EAGAIN)
2091 vhost_iotlb_miss(vq, addr, access);
2092 return ret;
2093}
2094
2095/* Each buffer in the virtqueues is actually a chain of descriptors. This
2096 * function returns the next descriptor in the chain,
2097 * or -1U if we're at the end. */
2098static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2099{
2100 unsigned int next;
2101
2102 /* If this descriptor says it doesn't chain, we're done. */
2103 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2104 return -1U;
2105
2106 /* Check they're not leading us off end of descriptors. */
2107 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2108 return next;
2109}
2110
2111static int get_indirect(struct vhost_virtqueue *vq,
2112 struct iovec iov[], unsigned int iov_size,
2113 unsigned int *out_num, unsigned int *in_num,
2114 struct vhost_log *log, unsigned int *log_num,
2115 struct vring_desc *indirect)
2116{
2117 struct vring_desc desc;
2118 unsigned int i = 0, count, found = 0;
2119 u32 len = vhost32_to_cpu(vq, indirect->len);
2120 struct iov_iter from;
2121 int ret, access;
2122
2123 /* Sanity check */
2124 if (unlikely(len % sizeof desc)) {
2125 vq_err(vq, "Invalid length in indirect descriptor: "
2126 "len 0x%llx not multiple of 0x%zx\n",
2127 (unsigned long long)len,
2128 sizeof desc);
2129 return -EINVAL;
2130 }
2131
2132 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2133 UIO_MAXIOV, VHOST_ACCESS_RO);
2134 if (unlikely(ret < 0)) {
2135 if (ret != -EAGAIN)
2136 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2137 return ret;
2138 }
2139 iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2140 count = len / sizeof desc;
2141 /* Buffers are chained via a 16 bit next field, so
2142 * we can have at most 2^16 of these. */
2143 if (unlikely(count > USHRT_MAX + 1)) {
2144 vq_err(vq, "Indirect buffer length too big: %d\n",
2145 indirect->len);
2146 return -E2BIG;
2147 }
2148
2149 do {
2150 unsigned iov_count = *in_num + *out_num;
2151 if (unlikely(++found > count)) {
2152 vq_err(vq, "Loop detected: last one at %u "
2153 "indirect size %u\n",
2154 i, count);
2155 return -EINVAL;
2156 }
2157 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2158 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2159 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2160 return -EINVAL;
2161 }
2162 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2163 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2164 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2165 return -EINVAL;
2166 }
2167
2168 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2169 access = VHOST_ACCESS_WO;
2170 else
2171 access = VHOST_ACCESS_RO;
2172
2173 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2174 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2175 iov_size - iov_count, access);
2176 if (unlikely(ret < 0)) {
2177 if (ret != -EAGAIN)
2178 vq_err(vq, "Translation failure %d indirect idx %d\n",
2179 ret, i);
2180 return ret;
2181 }
2182 /* If this is an input descriptor, increment that count. */
2183 if (access == VHOST_ACCESS_WO) {
2184 *in_num += ret;
2185 if (unlikely(log && ret)) {
2186 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2187 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2188 ++*log_num;
2189 }
2190 } else {
2191 /* If it's an output descriptor, they're all supposed
2192 * to come before any input descriptors. */
2193 if (unlikely(*in_num)) {
2194 vq_err(vq, "Indirect descriptor "
2195 "has out after in: idx %d\n", i);
2196 return -EINVAL;
2197 }
2198 *out_num += ret;
2199 }
2200 } while ((i = next_desc(vq, &desc)) != -1);
2201 return 0;
2202}
2203
2204/* This looks in the virtqueue and for the first available buffer, and converts
2205 * it to an iovec for convenient access. Since descriptors consist of some
2206 * number of output then some number of input descriptors, it's actually two
2207 * iovecs, but we pack them into one and note how many of each there were.
2208 *
2209 * This function returns the descriptor number found, or vq->num (which is
2210 * never a valid descriptor number) if none was found. A negative code is
2211 * returned on error. */
2212int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2213 struct iovec iov[], unsigned int iov_size,
2214 unsigned int *out_num, unsigned int *in_num,
2215 struct vhost_log *log, unsigned int *log_num)
2216{
2217 struct vring_desc desc;
2218 unsigned int i, head, found = 0;
2219 u16 last_avail_idx;
2220 __virtio16 avail_idx;
2221 __virtio16 ring_head;
2222 int ret, access;
2223
2224 /* Check it isn't doing very strange things with descriptor numbers. */
2225 last_avail_idx = vq->last_avail_idx;
2226
2227 if (vq->avail_idx == vq->last_avail_idx) {
2228 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2229 vq_err(vq, "Failed to access avail idx at %p\n",
2230 &vq->avail->idx);
2231 return -EFAULT;
2232 }
2233 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2234
2235 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2236 vq_err(vq, "Guest moved used index from %u to %u",
2237 last_avail_idx, vq->avail_idx);
2238 return -EFAULT;
2239 }
2240
2241 /* If there's nothing new since last we looked, return
2242 * invalid.
2243 */
2244 if (vq->avail_idx == last_avail_idx)
2245 return vq->num;
2246
2247 /* Only get avail ring entries after they have been
2248 * exposed by guest.
2249 */
2250 smp_rmb();
2251 }
2252
2253 /* Grab the next descriptor number they're advertising, and increment
2254 * the index we've seen. */
2255 if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2256 vq_err(vq, "Failed to read head: idx %d address %p\n",
2257 last_avail_idx,
2258 &vq->avail->ring[last_avail_idx % vq->num]);
2259 return -EFAULT;
2260 }
2261
2262 head = vhost16_to_cpu(vq, ring_head);
2263
2264 /* If their number is silly, that's an error. */
2265 if (unlikely(head >= vq->num)) {
2266 vq_err(vq, "Guest says index %u > %u is available",
2267 head, vq->num);
2268 return -EINVAL;
2269 }
2270
2271 /* When we start there are none of either input nor output. */
2272 *out_num = *in_num = 0;
2273 if (unlikely(log))
2274 *log_num = 0;
2275
2276 i = head;
2277 do {
2278 unsigned iov_count = *in_num + *out_num;
2279 if (unlikely(i >= vq->num)) {
2280 vq_err(vq, "Desc index is %u > %u, head = %u",
2281 i, vq->num, head);
2282 return -EINVAL;
2283 }
2284 if (unlikely(++found > vq->num)) {
2285 vq_err(vq, "Loop detected: last one at %u "
2286 "vq size %u head %u\n",
2287 i, vq->num, head);
2288 return -EINVAL;
2289 }
2290 ret = vhost_get_desc(vq, &desc, i);
2291 if (unlikely(ret)) {
2292 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2293 i, vq->desc + i);
2294 return -EFAULT;
2295 }
2296 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2297 ret = get_indirect(vq, iov, iov_size,
2298 out_num, in_num,
2299 log, log_num, &desc);
2300 if (unlikely(ret < 0)) {
2301 if (ret != -EAGAIN)
2302 vq_err(vq, "Failure detected "
2303 "in indirect descriptor at idx %d\n", i);
2304 return ret;
2305 }
2306 continue;
2307 }
2308
2309 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2310 access = VHOST_ACCESS_WO;
2311 else
2312 access = VHOST_ACCESS_RO;
2313 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2314 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2315 iov_size - iov_count, access);
2316 if (unlikely(ret < 0)) {
2317 if (ret != -EAGAIN)
2318 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2319 ret, i);
2320 return ret;
2321 }
2322 if (access == VHOST_ACCESS_WO) {
2323 /* If this is an input descriptor,
2324 * increment that count. */
2325 *in_num += ret;
2326 if (unlikely(log && ret)) {
2327 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2328 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2329 ++*log_num;
2330 }
2331 } else {
2332 /* If it's an output descriptor, they're all supposed
2333 * to come before any input descriptors. */
2334 if (unlikely(*in_num)) {
2335 vq_err(vq, "Descriptor has out after in: "
2336 "idx %d\n", i);
2337 return -EINVAL;
2338 }
2339 *out_num += ret;
2340 }
2341 } while ((i = next_desc(vq, &desc)) != -1);
2342
2343 /* On success, increment avail index. */
2344 vq->last_avail_idx++;
2345
2346 /* Assume notifications from guest are disabled at this point,
2347 * if they aren't we would need to update avail_event index. */
2348 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2349 return head;
2350}
2351EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2352
2353/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2354void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2355{
2356 vq->last_avail_idx -= n;
2357}
2358EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2359
2360/* After we've used one of their buffers, we tell them about it. We'll then
2361 * want to notify the guest, using eventfd. */
2362int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2363{
2364 struct vring_used_elem heads = {
2365 cpu_to_vhost32(vq, head),
2366 cpu_to_vhost32(vq, len)
2367 };
2368
2369 return vhost_add_used_n(vq, &heads, 1);
2370}
2371EXPORT_SYMBOL_GPL(vhost_add_used);
2372
2373static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2374 struct vring_used_elem *heads,
2375 unsigned count)
2376{
2377 vring_used_elem_t __user *used;
2378 u16 old, new;
2379 int start;
2380
2381 start = vq->last_used_idx & (vq->num - 1);
2382 used = vq->used->ring + start;
2383 if (vhost_put_used(vq, heads, start, count)) {
2384 vq_err(vq, "Failed to write used");
2385 return -EFAULT;
2386 }
2387 if (unlikely(vq->log_used)) {
2388 /* Make sure data is seen before log. */
2389 smp_wmb();
2390 /* Log used ring entry write. */
2391 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2392 count * sizeof *used);
2393 }
2394 old = vq->last_used_idx;
2395 new = (vq->last_used_idx += count);
2396 /* If the driver never bothers to signal in a very long while,
2397 * used index might wrap around. If that happens, invalidate
2398 * signalled_used index we stored. TODO: make sure driver
2399 * signals at least once in 2^16 and remove this. */
2400 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2401 vq->signalled_used_valid = false;
2402 return 0;
2403}
2404
2405/* After we've used one of their buffers, we tell them about it. We'll then
2406 * want to notify the guest, using eventfd. */
2407int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2408 unsigned count)
2409{
2410 int start, n, r;
2411
2412 start = vq->last_used_idx & (vq->num - 1);
2413 n = vq->num - start;
2414 if (n < count) {
2415 r = __vhost_add_used_n(vq, heads, n);
2416 if (r < 0)
2417 return r;
2418 heads += n;
2419 count -= n;
2420 }
2421 r = __vhost_add_used_n(vq, heads, count);
2422
2423 /* Make sure buffer is written before we update index. */
2424 smp_wmb();
2425 if (vhost_put_used_idx(vq)) {
2426 vq_err(vq, "Failed to increment used idx");
2427 return -EFAULT;
2428 }
2429 if (unlikely(vq->log_used)) {
2430 /* Make sure used idx is seen before log. */
2431 smp_wmb();
2432 /* Log used index update. */
2433 log_used(vq, offsetof(struct vring_used, idx),
2434 sizeof vq->used->idx);
2435 if (vq->log_ctx)
2436 eventfd_signal(vq->log_ctx, 1);
2437 }
2438 return r;
2439}
2440EXPORT_SYMBOL_GPL(vhost_add_used_n);
2441
2442static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2443{
2444 __u16 old, new;
2445 __virtio16 event;
2446 bool v;
2447 /* Flush out used index updates. This is paired
2448 * with the barrier that the Guest executes when enabling
2449 * interrupts. */
2450 smp_mb();
2451
2452 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2453 unlikely(vq->avail_idx == vq->last_avail_idx))
2454 return true;
2455
2456 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2457 __virtio16 flags;
2458 if (vhost_get_avail_flags(vq, &flags)) {
2459 vq_err(vq, "Failed to get flags");
2460 return true;
2461 }
2462 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2463 }
2464 old = vq->signalled_used;
2465 v = vq->signalled_used_valid;
2466 new = vq->signalled_used = vq->last_used_idx;
2467 vq->signalled_used_valid = true;
2468
2469 if (unlikely(!v))
2470 return true;
2471
2472 if (vhost_get_used_event(vq, &event)) {
2473 vq_err(vq, "Failed to get used event idx");
2474 return true;
2475 }
2476 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2477}
2478
2479/* This actually signals the guest, using eventfd. */
2480void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2481{
2482 /* Signal the Guest tell them we used something up. */
2483 if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2484 eventfd_signal(vq->call_ctx.ctx, 1);
2485}
2486EXPORT_SYMBOL_GPL(vhost_signal);
2487
2488/* And here's the combo meal deal. Supersize me! */
2489void vhost_add_used_and_signal(struct vhost_dev *dev,
2490 struct vhost_virtqueue *vq,
2491 unsigned int head, int len)
2492{
2493 vhost_add_used(vq, head, len);
2494 vhost_signal(dev, vq);
2495}
2496EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2497
2498/* multi-buffer version of vhost_add_used_and_signal */
2499void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2500 struct vhost_virtqueue *vq,
2501 struct vring_used_elem *heads, unsigned count)
2502{
2503 vhost_add_used_n(vq, heads, count);
2504 vhost_signal(dev, vq);
2505}
2506EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2507
2508/* return true if we're sure that avaiable ring is empty */
2509bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2510{
2511 __virtio16 avail_idx;
2512 int r;
2513
2514 if (vq->avail_idx != vq->last_avail_idx)
2515 return false;
2516
2517 r = vhost_get_avail_idx(vq, &avail_idx);
2518 if (unlikely(r))
2519 return false;
2520 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2521
2522 return vq->avail_idx == vq->last_avail_idx;
2523}
2524EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2525
2526/* OK, now we need to know about added descriptors. */
2527bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2528{
2529 __virtio16 avail_idx;
2530 int r;
2531
2532 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2533 return false;
2534 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2535 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2536 r = vhost_update_used_flags(vq);
2537 if (r) {
2538 vq_err(vq, "Failed to enable notification at %p: %d\n",
2539 &vq->used->flags, r);
2540 return false;
2541 }
2542 } else {
2543 r = vhost_update_avail_event(vq);
2544 if (r) {
2545 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2546 vhost_avail_event(vq), r);
2547 return false;
2548 }
2549 }
2550 /* They could have slipped one in as we were doing that: make
2551 * sure it's written, then check again. */
2552 smp_mb();
2553 r = vhost_get_avail_idx(vq, &avail_idx);
2554 if (r) {
2555 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2556 &vq->avail->idx, r);
2557 return false;
2558 }
2559 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2560
2561 return vq->avail_idx != vq->last_avail_idx;
2562}
2563EXPORT_SYMBOL_GPL(vhost_enable_notify);
2564
2565/* We don't need to be notified again. */
2566void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2567{
2568 int r;
2569
2570 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2571 return;
2572 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2573 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2574 r = vhost_update_used_flags(vq);
2575 if (r)
2576 vq_err(vq, "Failed to disable notification at %p: %d\n",
2577 &vq->used->flags, r);
2578 }
2579}
2580EXPORT_SYMBOL_GPL(vhost_disable_notify);
2581
2582/* Create a new message. */
2583struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2584{
2585 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2586 if (!node)
2587 return NULL;
2588
2589 /* Make sure all padding within the structure is initialized. */
2590 memset(&node->msg, 0, sizeof node->msg);
2591 node->vq = vq;
2592 node->msg.type = type;
2593 return node;
2594}
2595EXPORT_SYMBOL_GPL(vhost_new_msg);
2596
2597void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2598 struct vhost_msg_node *node)
2599{
2600 spin_lock(&dev->iotlb_lock);
2601 list_add_tail(&node->node, head);
2602 spin_unlock(&dev->iotlb_lock);
2603
2604 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2605}
2606EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2607
2608struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2609 struct list_head *head)
2610{
2611 struct vhost_msg_node *node = NULL;
2612
2613 spin_lock(&dev->iotlb_lock);
2614 if (!list_empty(head)) {
2615 node = list_first_entry(head, struct vhost_msg_node,
2616 node);
2617 list_del(&node->node);
2618 }
2619 spin_unlock(&dev->iotlb_lock);
2620
2621 return node;
2622}
2623EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2624
2625void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2626{
2627 struct vhost_virtqueue *vq;
2628 int i;
2629
2630 mutex_lock(&dev->mutex);
2631 for (i = 0; i < dev->nvqs; ++i) {
2632 vq = dev->vqs[i];
2633 mutex_lock(&vq->mutex);
2634 vq->acked_backend_features = features;
2635 mutex_unlock(&vq->mutex);
2636 }
2637 mutex_unlock(&dev->mutex);
2638}
2639EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2640
2641static int __init vhost_init(void)
2642{
2643 return 0;
2644}
2645
2646static void __exit vhost_exit(void)
2647{
2648}
2649
2650module_init(vhost_init);
2651module_exit(vhost_exit);
2652
2653MODULE_VERSION("0.0.1");
2654MODULE_LICENSE("GPL v2");
2655MODULE_AUTHOR("Michael S. Tsirkin");
2656MODULE_DESCRIPTION("Host kernel accelerator for virtio");