Loading...
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/module.h>
26#include <linux/sort.h>
27#include <linux/sched/mm.h>
28#include <linux/sched/signal.h>
29#include <linux/sched/vhost_task.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 struct vhost_virtqueue *vq)
192{
193 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
194 init_poll_funcptr(&poll->table, vhost_poll_func);
195 poll->mask = mask;
196 poll->dev = dev;
197 poll->wqh = NULL;
198 poll->vq = vq;
199
200 vhost_work_init(&poll->work, fn);
201}
202EXPORT_SYMBOL_GPL(vhost_poll_init);
203
204/* Start polling a file. We add ourselves to file's wait queue. The caller must
205 * keep a reference to a file until after vhost_poll_stop is called. */
206int vhost_poll_start(struct vhost_poll *poll, struct file *file)
207{
208 __poll_t mask;
209
210 if (poll->wqh)
211 return 0;
212
213 mask = vfs_poll(file, &poll->table);
214 if (mask)
215 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
216 if (mask & EPOLLERR) {
217 vhost_poll_stop(poll);
218 return -EINVAL;
219 }
220
221 return 0;
222}
223EXPORT_SYMBOL_GPL(vhost_poll_start);
224
225/* Stop polling a file. After this function returns, it becomes safe to drop the
226 * file reference. You must also flush afterwards. */
227void vhost_poll_stop(struct vhost_poll *poll)
228{
229 if (poll->wqh) {
230 remove_wait_queue(poll->wqh, &poll->wait);
231 poll->wqh = NULL;
232 }
233}
234EXPORT_SYMBOL_GPL(vhost_poll_stop);
235
236static void vhost_worker_queue(struct vhost_worker *worker,
237 struct vhost_work *work)
238{
239 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
240 /* We can only add the work to the list after we're
241 * sure it was not in the list.
242 * test_and_set_bit() implies a memory barrier.
243 */
244 llist_add(&work->node, &worker->work_list);
245 vhost_task_wake(worker->vtsk);
246 }
247}
248
249bool vhost_vq_work_queue(struct vhost_virtqueue *vq, struct vhost_work *work)
250{
251 struct vhost_worker *worker;
252 bool queued = false;
253
254 rcu_read_lock();
255 worker = rcu_dereference(vq->worker);
256 if (worker) {
257 queued = true;
258 vhost_worker_queue(worker, work);
259 }
260 rcu_read_unlock();
261
262 return queued;
263}
264EXPORT_SYMBOL_GPL(vhost_vq_work_queue);
265
266void vhost_vq_flush(struct vhost_virtqueue *vq)
267{
268 struct vhost_flush_struct flush;
269
270 init_completion(&flush.wait_event);
271 vhost_work_init(&flush.work, vhost_flush_work);
272
273 if (vhost_vq_work_queue(vq, &flush.work))
274 wait_for_completion(&flush.wait_event);
275}
276EXPORT_SYMBOL_GPL(vhost_vq_flush);
277
278/**
279 * vhost_worker_flush - flush a worker
280 * @worker: worker to flush
281 *
282 * This does not use RCU to protect the worker, so the device or worker
283 * mutex must be held.
284 */
285static void vhost_worker_flush(struct vhost_worker *worker)
286{
287 struct vhost_flush_struct flush;
288
289 init_completion(&flush.wait_event);
290 vhost_work_init(&flush.work, vhost_flush_work);
291
292 vhost_worker_queue(worker, &flush.work);
293 wait_for_completion(&flush.wait_event);
294}
295
296void vhost_dev_flush(struct vhost_dev *dev)
297{
298 struct vhost_worker *worker;
299 unsigned long i;
300
301 xa_for_each(&dev->worker_xa, i, worker) {
302 mutex_lock(&worker->mutex);
303 if (!worker->attachment_cnt) {
304 mutex_unlock(&worker->mutex);
305 continue;
306 }
307 vhost_worker_flush(worker);
308 mutex_unlock(&worker->mutex);
309 }
310}
311EXPORT_SYMBOL_GPL(vhost_dev_flush);
312
313/* A lockless hint for busy polling code to exit the loop */
314bool vhost_vq_has_work(struct vhost_virtqueue *vq)
315{
316 struct vhost_worker *worker;
317 bool has_work = false;
318
319 rcu_read_lock();
320 worker = rcu_dereference(vq->worker);
321 if (worker && !llist_empty(&worker->work_list))
322 has_work = true;
323 rcu_read_unlock();
324
325 return has_work;
326}
327EXPORT_SYMBOL_GPL(vhost_vq_has_work);
328
329void vhost_poll_queue(struct vhost_poll *poll)
330{
331 vhost_vq_work_queue(poll->vq, &poll->work);
332}
333EXPORT_SYMBOL_GPL(vhost_poll_queue);
334
335static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
336{
337 int j;
338
339 for (j = 0; j < VHOST_NUM_ADDRS; j++)
340 vq->meta_iotlb[j] = NULL;
341}
342
343static void vhost_vq_meta_reset(struct vhost_dev *d)
344{
345 int i;
346
347 for (i = 0; i < d->nvqs; ++i)
348 __vhost_vq_meta_reset(d->vqs[i]);
349}
350
351static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
352{
353 call_ctx->ctx = NULL;
354 memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
355}
356
357bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
358{
359 return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
360}
361EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
362
363static void vhost_vq_reset(struct vhost_dev *dev,
364 struct vhost_virtqueue *vq)
365{
366 vq->num = 1;
367 vq->desc = NULL;
368 vq->avail = NULL;
369 vq->used = NULL;
370 vq->last_avail_idx = 0;
371 vq->avail_idx = 0;
372 vq->last_used_idx = 0;
373 vq->signalled_used = 0;
374 vq->signalled_used_valid = false;
375 vq->used_flags = 0;
376 vq->log_used = false;
377 vq->log_addr = -1ull;
378 vq->private_data = NULL;
379 vq->acked_features = 0;
380 vq->acked_backend_features = 0;
381 vq->log_base = NULL;
382 vq->error_ctx = NULL;
383 vq->kick = NULL;
384 vq->log_ctx = NULL;
385 vhost_disable_cross_endian(vq);
386 vhost_reset_is_le(vq);
387 vq->busyloop_timeout = 0;
388 vq->umem = NULL;
389 vq->iotlb = NULL;
390 rcu_assign_pointer(vq->worker, NULL);
391 vhost_vring_call_reset(&vq->call_ctx);
392 __vhost_vq_meta_reset(vq);
393}
394
395static bool vhost_worker(void *data)
396{
397 struct vhost_worker *worker = data;
398 struct vhost_work *work, *work_next;
399 struct llist_node *node;
400
401 node = llist_del_all(&worker->work_list);
402 if (node) {
403 __set_current_state(TASK_RUNNING);
404
405 node = llist_reverse_order(node);
406 /* make sure flag is seen after deletion */
407 smp_wmb();
408 llist_for_each_entry_safe(work, work_next, node, node) {
409 clear_bit(VHOST_WORK_QUEUED, &work->flags);
410 kcov_remote_start_common(worker->kcov_handle);
411 work->fn(work);
412 kcov_remote_stop();
413 cond_resched();
414 }
415 }
416
417 return !!node;
418}
419
420static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
421{
422 kfree(vq->indirect);
423 vq->indirect = NULL;
424 kfree(vq->log);
425 vq->log = NULL;
426 kfree(vq->heads);
427 vq->heads = NULL;
428}
429
430/* Helper to allocate iovec buffers for all vqs. */
431static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
432{
433 struct vhost_virtqueue *vq;
434 int i;
435
436 for (i = 0; i < dev->nvqs; ++i) {
437 vq = dev->vqs[i];
438 vq->indirect = kmalloc_array(UIO_MAXIOV,
439 sizeof(*vq->indirect),
440 GFP_KERNEL);
441 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
442 GFP_KERNEL);
443 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
444 GFP_KERNEL);
445 if (!vq->indirect || !vq->log || !vq->heads)
446 goto err_nomem;
447 }
448 return 0;
449
450err_nomem:
451 for (; i >= 0; --i)
452 vhost_vq_free_iovecs(dev->vqs[i]);
453 return -ENOMEM;
454}
455
456static void vhost_dev_free_iovecs(struct vhost_dev *dev)
457{
458 int i;
459
460 for (i = 0; i < dev->nvqs; ++i)
461 vhost_vq_free_iovecs(dev->vqs[i]);
462}
463
464bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
465 int pkts, int total_len)
466{
467 struct vhost_dev *dev = vq->dev;
468
469 if ((dev->byte_weight && total_len >= dev->byte_weight) ||
470 pkts >= dev->weight) {
471 vhost_poll_queue(&vq->poll);
472 return true;
473 }
474
475 return false;
476}
477EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
478
479static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
480 unsigned int num)
481{
482 size_t event __maybe_unused =
483 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
484
485 return size_add(struct_size(vq->avail, ring, num), event);
486}
487
488static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
489 unsigned int num)
490{
491 size_t event __maybe_unused =
492 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
493
494 return size_add(struct_size(vq->used, ring, num), event);
495}
496
497static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
498 unsigned int num)
499{
500 return sizeof(*vq->desc) * num;
501}
502
503void vhost_dev_init(struct vhost_dev *dev,
504 struct vhost_virtqueue **vqs, int nvqs,
505 int iov_limit, int weight, int byte_weight,
506 bool use_worker,
507 int (*msg_handler)(struct vhost_dev *dev, u32 asid,
508 struct vhost_iotlb_msg *msg))
509{
510 struct vhost_virtqueue *vq;
511 int i;
512
513 dev->vqs = vqs;
514 dev->nvqs = nvqs;
515 mutex_init(&dev->mutex);
516 dev->log_ctx = NULL;
517 dev->umem = NULL;
518 dev->iotlb = NULL;
519 dev->mm = NULL;
520 dev->iov_limit = iov_limit;
521 dev->weight = weight;
522 dev->byte_weight = byte_weight;
523 dev->use_worker = use_worker;
524 dev->msg_handler = msg_handler;
525 init_waitqueue_head(&dev->wait);
526 INIT_LIST_HEAD(&dev->read_list);
527 INIT_LIST_HEAD(&dev->pending_list);
528 spin_lock_init(&dev->iotlb_lock);
529 xa_init_flags(&dev->worker_xa, XA_FLAGS_ALLOC);
530
531 for (i = 0; i < dev->nvqs; ++i) {
532 vq = dev->vqs[i];
533 vq->log = NULL;
534 vq->indirect = NULL;
535 vq->heads = NULL;
536 vq->dev = dev;
537 mutex_init(&vq->mutex);
538 vhost_vq_reset(dev, vq);
539 if (vq->handle_kick)
540 vhost_poll_init(&vq->poll, vq->handle_kick,
541 EPOLLIN, dev, vq);
542 }
543}
544EXPORT_SYMBOL_GPL(vhost_dev_init);
545
546/* Caller should have device mutex */
547long vhost_dev_check_owner(struct vhost_dev *dev)
548{
549 /* Are you the owner? If not, I don't think you mean to do that */
550 return dev->mm == current->mm ? 0 : -EPERM;
551}
552EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
553
554/* Caller should have device mutex */
555bool vhost_dev_has_owner(struct vhost_dev *dev)
556{
557 return dev->mm;
558}
559EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
560
561static void vhost_attach_mm(struct vhost_dev *dev)
562{
563 /* No owner, become one */
564 if (dev->use_worker) {
565 dev->mm = get_task_mm(current);
566 } else {
567 /* vDPA device does not use worker thead, so there's
568 * no need to hold the address space for mm. This help
569 * to avoid deadlock in the case of mmap() which may
570 * held the refcnt of the file and depends on release
571 * method to remove vma.
572 */
573 dev->mm = current->mm;
574 mmgrab(dev->mm);
575 }
576}
577
578static void vhost_detach_mm(struct vhost_dev *dev)
579{
580 if (!dev->mm)
581 return;
582
583 if (dev->use_worker)
584 mmput(dev->mm);
585 else
586 mmdrop(dev->mm);
587
588 dev->mm = NULL;
589}
590
591static void vhost_worker_destroy(struct vhost_dev *dev,
592 struct vhost_worker *worker)
593{
594 if (!worker)
595 return;
596
597 WARN_ON(!llist_empty(&worker->work_list));
598 xa_erase(&dev->worker_xa, worker->id);
599 vhost_task_stop(worker->vtsk);
600 kfree(worker);
601}
602
603static void vhost_workers_free(struct vhost_dev *dev)
604{
605 struct vhost_worker *worker;
606 unsigned long i;
607
608 if (!dev->use_worker)
609 return;
610
611 for (i = 0; i < dev->nvqs; i++)
612 rcu_assign_pointer(dev->vqs[i]->worker, NULL);
613 /*
614 * Free the default worker we created and cleanup workers userspace
615 * created but couldn't clean up (it forgot or crashed).
616 */
617 xa_for_each(&dev->worker_xa, i, worker)
618 vhost_worker_destroy(dev, worker);
619 xa_destroy(&dev->worker_xa);
620}
621
622static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev)
623{
624 struct vhost_worker *worker;
625 struct vhost_task *vtsk;
626 char name[TASK_COMM_LEN];
627 int ret;
628 u32 id;
629
630 worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
631 if (!worker)
632 return NULL;
633
634 snprintf(name, sizeof(name), "vhost-%d", current->pid);
635
636 vtsk = vhost_task_create(vhost_worker, worker, name);
637 if (!vtsk)
638 goto free_worker;
639
640 mutex_init(&worker->mutex);
641 init_llist_head(&worker->work_list);
642 worker->kcov_handle = kcov_common_handle();
643 worker->vtsk = vtsk;
644
645 vhost_task_start(vtsk);
646
647 ret = xa_alloc(&dev->worker_xa, &id, worker, xa_limit_32b, GFP_KERNEL);
648 if (ret < 0)
649 goto stop_worker;
650 worker->id = id;
651
652 return worker;
653
654stop_worker:
655 vhost_task_stop(vtsk);
656free_worker:
657 kfree(worker);
658 return NULL;
659}
660
661/* Caller must have device mutex */
662static void __vhost_vq_attach_worker(struct vhost_virtqueue *vq,
663 struct vhost_worker *worker)
664{
665 struct vhost_worker *old_worker;
666
667 old_worker = rcu_dereference_check(vq->worker,
668 lockdep_is_held(&vq->dev->mutex));
669
670 mutex_lock(&worker->mutex);
671 worker->attachment_cnt++;
672 mutex_unlock(&worker->mutex);
673 rcu_assign_pointer(vq->worker, worker);
674
675 if (!old_worker)
676 return;
677 /*
678 * Take the worker mutex to make sure we see the work queued from
679 * device wide flushes which doesn't use RCU for execution.
680 */
681 mutex_lock(&old_worker->mutex);
682 old_worker->attachment_cnt--;
683 /*
684 * We don't want to call synchronize_rcu for every vq during setup
685 * because it will slow down VM startup. If we haven't done
686 * VHOST_SET_VRING_KICK and not done the driver specific
687 * SET_ENDPOINT/RUNNUNG then we can skip the sync since there will
688 * not be any works queued for scsi and net.
689 */
690 mutex_lock(&vq->mutex);
691 if (!vhost_vq_get_backend(vq) && !vq->kick) {
692 mutex_unlock(&vq->mutex);
693 mutex_unlock(&old_worker->mutex);
694 /*
695 * vsock can queue anytime after VHOST_VSOCK_SET_GUEST_CID.
696 * Warn if it adds support for multiple workers but forgets to
697 * handle the early queueing case.
698 */
699 WARN_ON(!old_worker->attachment_cnt &&
700 !llist_empty(&old_worker->work_list));
701 return;
702 }
703 mutex_unlock(&vq->mutex);
704
705 /* Make sure new vq queue/flush/poll calls see the new worker */
706 synchronize_rcu();
707 /* Make sure whatever was queued gets run */
708 vhost_worker_flush(old_worker);
709 mutex_unlock(&old_worker->mutex);
710}
711
712 /* Caller must have device mutex */
713static int vhost_vq_attach_worker(struct vhost_virtqueue *vq,
714 struct vhost_vring_worker *info)
715{
716 unsigned long index = info->worker_id;
717 struct vhost_dev *dev = vq->dev;
718 struct vhost_worker *worker;
719
720 if (!dev->use_worker)
721 return -EINVAL;
722
723 worker = xa_find(&dev->worker_xa, &index, UINT_MAX, XA_PRESENT);
724 if (!worker || worker->id != info->worker_id)
725 return -ENODEV;
726
727 __vhost_vq_attach_worker(vq, worker);
728 return 0;
729}
730
731/* Caller must have device mutex */
732static int vhost_new_worker(struct vhost_dev *dev,
733 struct vhost_worker_state *info)
734{
735 struct vhost_worker *worker;
736
737 worker = vhost_worker_create(dev);
738 if (!worker)
739 return -ENOMEM;
740
741 info->worker_id = worker->id;
742 return 0;
743}
744
745/* Caller must have device mutex */
746static int vhost_free_worker(struct vhost_dev *dev,
747 struct vhost_worker_state *info)
748{
749 unsigned long index = info->worker_id;
750 struct vhost_worker *worker;
751
752 worker = xa_find(&dev->worker_xa, &index, UINT_MAX, XA_PRESENT);
753 if (!worker || worker->id != info->worker_id)
754 return -ENODEV;
755
756 mutex_lock(&worker->mutex);
757 if (worker->attachment_cnt) {
758 mutex_unlock(&worker->mutex);
759 return -EBUSY;
760 }
761 mutex_unlock(&worker->mutex);
762
763 vhost_worker_destroy(dev, worker);
764 return 0;
765}
766
767static int vhost_get_vq_from_user(struct vhost_dev *dev, void __user *argp,
768 struct vhost_virtqueue **vq, u32 *id)
769{
770 u32 __user *idxp = argp;
771 u32 idx;
772 long r;
773
774 r = get_user(idx, idxp);
775 if (r < 0)
776 return r;
777
778 if (idx >= dev->nvqs)
779 return -ENOBUFS;
780
781 idx = array_index_nospec(idx, dev->nvqs);
782
783 *vq = dev->vqs[idx];
784 *id = idx;
785 return 0;
786}
787
788/* Caller must have device mutex */
789long vhost_worker_ioctl(struct vhost_dev *dev, unsigned int ioctl,
790 void __user *argp)
791{
792 struct vhost_vring_worker ring_worker;
793 struct vhost_worker_state state;
794 struct vhost_worker *worker;
795 struct vhost_virtqueue *vq;
796 long ret;
797 u32 idx;
798
799 if (!dev->use_worker)
800 return -EINVAL;
801
802 if (!vhost_dev_has_owner(dev))
803 return -EINVAL;
804
805 ret = vhost_dev_check_owner(dev);
806 if (ret)
807 return ret;
808
809 switch (ioctl) {
810 /* dev worker ioctls */
811 case VHOST_NEW_WORKER:
812 ret = vhost_new_worker(dev, &state);
813 if (!ret && copy_to_user(argp, &state, sizeof(state)))
814 ret = -EFAULT;
815 return ret;
816 case VHOST_FREE_WORKER:
817 if (copy_from_user(&state, argp, sizeof(state)))
818 return -EFAULT;
819 return vhost_free_worker(dev, &state);
820 /* vring worker ioctls */
821 case VHOST_ATTACH_VRING_WORKER:
822 case VHOST_GET_VRING_WORKER:
823 break;
824 default:
825 return -ENOIOCTLCMD;
826 }
827
828 ret = vhost_get_vq_from_user(dev, argp, &vq, &idx);
829 if (ret)
830 return ret;
831
832 switch (ioctl) {
833 case VHOST_ATTACH_VRING_WORKER:
834 if (copy_from_user(&ring_worker, argp, sizeof(ring_worker))) {
835 ret = -EFAULT;
836 break;
837 }
838
839 ret = vhost_vq_attach_worker(vq, &ring_worker);
840 break;
841 case VHOST_GET_VRING_WORKER:
842 worker = rcu_dereference_check(vq->worker,
843 lockdep_is_held(&dev->mutex));
844 if (!worker) {
845 ret = -EINVAL;
846 break;
847 }
848
849 ring_worker.index = idx;
850 ring_worker.worker_id = worker->id;
851
852 if (copy_to_user(argp, &ring_worker, sizeof(ring_worker)))
853 ret = -EFAULT;
854 break;
855 default:
856 ret = -ENOIOCTLCMD;
857 break;
858 }
859
860 return ret;
861}
862EXPORT_SYMBOL_GPL(vhost_worker_ioctl);
863
864/* Caller should have device mutex */
865long vhost_dev_set_owner(struct vhost_dev *dev)
866{
867 struct vhost_worker *worker;
868 int err, i;
869
870 /* Is there an owner already? */
871 if (vhost_dev_has_owner(dev)) {
872 err = -EBUSY;
873 goto err_mm;
874 }
875
876 vhost_attach_mm(dev);
877
878 err = vhost_dev_alloc_iovecs(dev);
879 if (err)
880 goto err_iovecs;
881
882 if (dev->use_worker) {
883 /*
884 * This should be done last, because vsock can queue work
885 * before VHOST_SET_OWNER so it simplifies the failure path
886 * below since we don't have to worry about vsock queueing
887 * while we free the worker.
888 */
889 worker = vhost_worker_create(dev);
890 if (!worker) {
891 err = -ENOMEM;
892 goto err_worker;
893 }
894
895 for (i = 0; i < dev->nvqs; i++)
896 __vhost_vq_attach_worker(dev->vqs[i], worker);
897 }
898
899 return 0;
900
901err_worker:
902 vhost_dev_free_iovecs(dev);
903err_iovecs:
904 vhost_detach_mm(dev);
905err_mm:
906 return err;
907}
908EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
909
910static struct vhost_iotlb *iotlb_alloc(void)
911{
912 return vhost_iotlb_alloc(max_iotlb_entries,
913 VHOST_IOTLB_FLAG_RETIRE);
914}
915
916struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
917{
918 return iotlb_alloc();
919}
920EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
921
922/* Caller should have device mutex */
923void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
924{
925 int i;
926
927 vhost_dev_cleanup(dev);
928
929 dev->umem = umem;
930 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
931 * VQs aren't running.
932 */
933 for (i = 0; i < dev->nvqs; ++i)
934 dev->vqs[i]->umem = umem;
935}
936EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
937
938void vhost_dev_stop(struct vhost_dev *dev)
939{
940 int i;
941
942 for (i = 0; i < dev->nvqs; ++i) {
943 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
944 vhost_poll_stop(&dev->vqs[i]->poll);
945 }
946
947 vhost_dev_flush(dev);
948}
949EXPORT_SYMBOL_GPL(vhost_dev_stop);
950
951void vhost_clear_msg(struct vhost_dev *dev)
952{
953 struct vhost_msg_node *node, *n;
954
955 spin_lock(&dev->iotlb_lock);
956
957 list_for_each_entry_safe(node, n, &dev->read_list, node) {
958 list_del(&node->node);
959 kfree(node);
960 }
961
962 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
963 list_del(&node->node);
964 kfree(node);
965 }
966
967 spin_unlock(&dev->iotlb_lock);
968}
969EXPORT_SYMBOL_GPL(vhost_clear_msg);
970
971void vhost_dev_cleanup(struct vhost_dev *dev)
972{
973 int i;
974
975 for (i = 0; i < dev->nvqs; ++i) {
976 if (dev->vqs[i]->error_ctx)
977 eventfd_ctx_put(dev->vqs[i]->error_ctx);
978 if (dev->vqs[i]->kick)
979 fput(dev->vqs[i]->kick);
980 if (dev->vqs[i]->call_ctx.ctx)
981 eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
982 vhost_vq_reset(dev, dev->vqs[i]);
983 }
984 vhost_dev_free_iovecs(dev);
985 if (dev->log_ctx)
986 eventfd_ctx_put(dev->log_ctx);
987 dev->log_ctx = NULL;
988 /* No one will access memory at this point */
989 vhost_iotlb_free(dev->umem);
990 dev->umem = NULL;
991 vhost_iotlb_free(dev->iotlb);
992 dev->iotlb = NULL;
993 vhost_clear_msg(dev);
994 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
995 vhost_workers_free(dev);
996 vhost_detach_mm(dev);
997}
998EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
999
1000static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
1001{
1002 u64 a = addr / VHOST_PAGE_SIZE / 8;
1003
1004 /* Make sure 64 bit math will not overflow. */
1005 if (a > ULONG_MAX - (unsigned long)log_base ||
1006 a + (unsigned long)log_base > ULONG_MAX)
1007 return false;
1008
1009 return access_ok(log_base + a,
1010 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
1011}
1012
1013/* Make sure 64 bit math will not overflow. */
1014static bool vhost_overflow(u64 uaddr, u64 size)
1015{
1016 if (uaddr > ULONG_MAX || size > ULONG_MAX)
1017 return true;
1018
1019 if (!size)
1020 return false;
1021
1022 return uaddr > ULONG_MAX - size + 1;
1023}
1024
1025/* Caller should have vq mutex and device mutex. */
1026static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
1027 int log_all)
1028{
1029 struct vhost_iotlb_map *map;
1030
1031 if (!umem)
1032 return false;
1033
1034 list_for_each_entry(map, &umem->list, link) {
1035 unsigned long a = map->addr;
1036
1037 if (vhost_overflow(map->addr, map->size))
1038 return false;
1039
1040
1041 if (!access_ok((void __user *)a, map->size))
1042 return false;
1043 else if (log_all && !log_access_ok(log_base,
1044 map->start,
1045 map->size))
1046 return false;
1047 }
1048 return true;
1049}
1050
1051static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
1052 u64 addr, unsigned int size,
1053 int type)
1054{
1055 const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
1056
1057 if (!map)
1058 return NULL;
1059
1060 return (void __user *)(uintptr_t)(map->addr + addr - map->start);
1061}
1062
1063/* Can we switch to this memory table? */
1064/* Caller should have device mutex but not vq mutex */
1065static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
1066 int log_all)
1067{
1068 int i;
1069
1070 for (i = 0; i < d->nvqs; ++i) {
1071 bool ok;
1072 bool log;
1073
1074 mutex_lock(&d->vqs[i]->mutex);
1075 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
1076 /* If ring is inactive, will check when it's enabled. */
1077 if (d->vqs[i]->private_data)
1078 ok = vq_memory_access_ok(d->vqs[i]->log_base,
1079 umem, log);
1080 else
1081 ok = true;
1082 mutex_unlock(&d->vqs[i]->mutex);
1083 if (!ok)
1084 return false;
1085 }
1086 return true;
1087}
1088
1089static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1090 struct iovec iov[], int iov_size, int access);
1091
1092static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
1093 const void *from, unsigned size)
1094{
1095 int ret;
1096
1097 if (!vq->iotlb)
1098 return __copy_to_user(to, from, size);
1099 else {
1100 /* This function should be called after iotlb
1101 * prefetch, which means we're sure that all vq
1102 * could be access through iotlb. So -EAGAIN should
1103 * not happen in this case.
1104 */
1105 struct iov_iter t;
1106 void __user *uaddr = vhost_vq_meta_fetch(vq,
1107 (u64)(uintptr_t)to, size,
1108 VHOST_ADDR_USED);
1109
1110 if (uaddr)
1111 return __copy_to_user(uaddr, from, size);
1112
1113 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
1114 ARRAY_SIZE(vq->iotlb_iov),
1115 VHOST_ACCESS_WO);
1116 if (ret < 0)
1117 goto out;
1118 iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
1119 ret = copy_to_iter(from, size, &t);
1120 if (ret == size)
1121 ret = 0;
1122 }
1123out:
1124 return ret;
1125}
1126
1127static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
1128 void __user *from, unsigned size)
1129{
1130 int ret;
1131
1132 if (!vq->iotlb)
1133 return __copy_from_user(to, from, size);
1134 else {
1135 /* This function should be called after iotlb
1136 * prefetch, which means we're sure that vq
1137 * could be access through iotlb. So -EAGAIN should
1138 * not happen in this case.
1139 */
1140 void __user *uaddr = vhost_vq_meta_fetch(vq,
1141 (u64)(uintptr_t)from, size,
1142 VHOST_ADDR_DESC);
1143 struct iov_iter f;
1144
1145 if (uaddr)
1146 return __copy_from_user(to, uaddr, size);
1147
1148 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
1149 ARRAY_SIZE(vq->iotlb_iov),
1150 VHOST_ACCESS_RO);
1151 if (ret < 0) {
1152 vq_err(vq, "IOTLB translation failure: uaddr "
1153 "%p size 0x%llx\n", from,
1154 (unsigned long long) size);
1155 goto out;
1156 }
1157 iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
1158 ret = copy_from_iter(to, size, &f);
1159 if (ret == size)
1160 ret = 0;
1161 }
1162
1163out:
1164 return ret;
1165}
1166
1167static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
1168 void __user *addr, unsigned int size,
1169 int type)
1170{
1171 int ret;
1172
1173 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
1174 ARRAY_SIZE(vq->iotlb_iov),
1175 VHOST_ACCESS_RO);
1176 if (ret < 0) {
1177 vq_err(vq, "IOTLB translation failure: uaddr "
1178 "%p size 0x%llx\n", addr,
1179 (unsigned long long) size);
1180 return NULL;
1181 }
1182
1183 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
1184 vq_err(vq, "Non atomic userspace memory access: uaddr "
1185 "%p size 0x%llx\n", addr,
1186 (unsigned long long) size);
1187 return NULL;
1188 }
1189
1190 return vq->iotlb_iov[0].iov_base;
1191}
1192
1193/* This function should be called after iotlb
1194 * prefetch, which means we're sure that vq
1195 * could be access through iotlb. So -EAGAIN should
1196 * not happen in this case.
1197 */
1198static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
1199 void __user *addr, unsigned int size,
1200 int type)
1201{
1202 void __user *uaddr = vhost_vq_meta_fetch(vq,
1203 (u64)(uintptr_t)addr, size, type);
1204 if (uaddr)
1205 return uaddr;
1206
1207 return __vhost_get_user_slow(vq, addr, size, type);
1208}
1209
1210#define vhost_put_user(vq, x, ptr) \
1211({ \
1212 int ret; \
1213 if (!vq->iotlb) { \
1214 ret = __put_user(x, ptr); \
1215 } else { \
1216 __typeof__(ptr) to = \
1217 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
1218 sizeof(*ptr), VHOST_ADDR_USED); \
1219 if (to != NULL) \
1220 ret = __put_user(x, to); \
1221 else \
1222 ret = -EFAULT; \
1223 } \
1224 ret; \
1225})
1226
1227static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
1228{
1229 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1230 vhost_avail_event(vq));
1231}
1232
1233static inline int vhost_put_used(struct vhost_virtqueue *vq,
1234 struct vring_used_elem *head, int idx,
1235 int count)
1236{
1237 return vhost_copy_to_user(vq, vq->used->ring + idx, head,
1238 count * sizeof(*head));
1239}
1240
1241static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
1242
1243{
1244 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1245 &vq->used->flags);
1246}
1247
1248static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
1249
1250{
1251 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
1252 &vq->used->idx);
1253}
1254
1255#define vhost_get_user(vq, x, ptr, type) \
1256({ \
1257 int ret; \
1258 if (!vq->iotlb) { \
1259 ret = __get_user(x, ptr); \
1260 } else { \
1261 __typeof__(ptr) from = \
1262 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
1263 sizeof(*ptr), \
1264 type); \
1265 if (from != NULL) \
1266 ret = __get_user(x, from); \
1267 else \
1268 ret = -EFAULT; \
1269 } \
1270 ret; \
1271})
1272
1273#define vhost_get_avail(vq, x, ptr) \
1274 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
1275
1276#define vhost_get_used(vq, x, ptr) \
1277 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
1278
1279static void vhost_dev_lock_vqs(struct vhost_dev *d)
1280{
1281 int i = 0;
1282 for (i = 0; i < d->nvqs; ++i)
1283 mutex_lock_nested(&d->vqs[i]->mutex, i);
1284}
1285
1286static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1287{
1288 int i = 0;
1289 for (i = 0; i < d->nvqs; ++i)
1290 mutex_unlock(&d->vqs[i]->mutex);
1291}
1292
1293static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1294 __virtio16 *idx)
1295{
1296 return vhost_get_avail(vq, *idx, &vq->avail->idx);
1297}
1298
1299static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1300 __virtio16 *head, int idx)
1301{
1302 return vhost_get_avail(vq, *head,
1303 &vq->avail->ring[idx & (vq->num - 1)]);
1304}
1305
1306static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1307 __virtio16 *flags)
1308{
1309 return vhost_get_avail(vq, *flags, &vq->avail->flags);
1310}
1311
1312static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1313 __virtio16 *event)
1314{
1315 return vhost_get_avail(vq, *event, vhost_used_event(vq));
1316}
1317
1318static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1319 __virtio16 *idx)
1320{
1321 return vhost_get_used(vq, *idx, &vq->used->idx);
1322}
1323
1324static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1325 struct vring_desc *desc, int idx)
1326{
1327 return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1328}
1329
1330static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1331 struct vhost_iotlb_msg *msg)
1332{
1333 struct vhost_msg_node *node, *n;
1334
1335 spin_lock(&d->iotlb_lock);
1336
1337 list_for_each_entry_safe(node, n, &d->pending_list, node) {
1338 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1339 if (msg->iova <= vq_msg->iova &&
1340 msg->iova + msg->size - 1 >= vq_msg->iova &&
1341 vq_msg->type == VHOST_IOTLB_MISS) {
1342 vhost_poll_queue(&node->vq->poll);
1343 list_del(&node->node);
1344 kfree(node);
1345 }
1346 }
1347
1348 spin_unlock(&d->iotlb_lock);
1349}
1350
1351static bool umem_access_ok(u64 uaddr, u64 size, int access)
1352{
1353 unsigned long a = uaddr;
1354
1355 /* Make sure 64 bit math will not overflow. */
1356 if (vhost_overflow(uaddr, size))
1357 return false;
1358
1359 if ((access & VHOST_ACCESS_RO) &&
1360 !access_ok((void __user *)a, size))
1361 return false;
1362 if ((access & VHOST_ACCESS_WO) &&
1363 !access_ok((void __user *)a, size))
1364 return false;
1365 return true;
1366}
1367
1368static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1369 struct vhost_iotlb_msg *msg)
1370{
1371 int ret = 0;
1372
1373 if (asid != 0)
1374 return -EINVAL;
1375
1376 mutex_lock(&dev->mutex);
1377 vhost_dev_lock_vqs(dev);
1378 switch (msg->type) {
1379 case VHOST_IOTLB_UPDATE:
1380 if (!dev->iotlb) {
1381 ret = -EFAULT;
1382 break;
1383 }
1384 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1385 ret = -EFAULT;
1386 break;
1387 }
1388 vhost_vq_meta_reset(dev);
1389 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1390 msg->iova + msg->size - 1,
1391 msg->uaddr, msg->perm)) {
1392 ret = -ENOMEM;
1393 break;
1394 }
1395 vhost_iotlb_notify_vq(dev, msg);
1396 break;
1397 case VHOST_IOTLB_INVALIDATE:
1398 if (!dev->iotlb) {
1399 ret = -EFAULT;
1400 break;
1401 }
1402 vhost_vq_meta_reset(dev);
1403 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1404 msg->iova + msg->size - 1);
1405 break;
1406 default:
1407 ret = -EINVAL;
1408 break;
1409 }
1410
1411 vhost_dev_unlock_vqs(dev);
1412 mutex_unlock(&dev->mutex);
1413
1414 return ret;
1415}
1416ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1417 struct iov_iter *from)
1418{
1419 struct vhost_iotlb_msg msg;
1420 size_t offset;
1421 int type, ret;
1422 u32 asid = 0;
1423
1424 ret = copy_from_iter(&type, sizeof(type), from);
1425 if (ret != sizeof(type)) {
1426 ret = -EINVAL;
1427 goto done;
1428 }
1429
1430 switch (type) {
1431 case VHOST_IOTLB_MSG:
1432 /* There maybe a hole after type for V1 message type,
1433 * so skip it here.
1434 */
1435 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1436 break;
1437 case VHOST_IOTLB_MSG_V2:
1438 if (vhost_backend_has_feature(dev->vqs[0],
1439 VHOST_BACKEND_F_IOTLB_ASID)) {
1440 ret = copy_from_iter(&asid, sizeof(asid), from);
1441 if (ret != sizeof(asid)) {
1442 ret = -EINVAL;
1443 goto done;
1444 }
1445 offset = 0;
1446 } else
1447 offset = sizeof(__u32);
1448 break;
1449 default:
1450 ret = -EINVAL;
1451 goto done;
1452 }
1453
1454 iov_iter_advance(from, offset);
1455 ret = copy_from_iter(&msg, sizeof(msg), from);
1456 if (ret != sizeof(msg)) {
1457 ret = -EINVAL;
1458 goto done;
1459 }
1460
1461 if (msg.type == VHOST_IOTLB_UPDATE && msg.size == 0) {
1462 ret = -EINVAL;
1463 goto done;
1464 }
1465
1466 if (dev->msg_handler)
1467 ret = dev->msg_handler(dev, asid, &msg);
1468 else
1469 ret = vhost_process_iotlb_msg(dev, asid, &msg);
1470 if (ret) {
1471 ret = -EFAULT;
1472 goto done;
1473 }
1474
1475 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1476 sizeof(struct vhost_msg_v2);
1477done:
1478 return ret;
1479}
1480EXPORT_SYMBOL(vhost_chr_write_iter);
1481
1482__poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1483 poll_table *wait)
1484{
1485 __poll_t mask = 0;
1486
1487 poll_wait(file, &dev->wait, wait);
1488
1489 if (!list_empty(&dev->read_list))
1490 mask |= EPOLLIN | EPOLLRDNORM;
1491
1492 return mask;
1493}
1494EXPORT_SYMBOL(vhost_chr_poll);
1495
1496ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1497 int noblock)
1498{
1499 DEFINE_WAIT(wait);
1500 struct vhost_msg_node *node;
1501 ssize_t ret = 0;
1502 unsigned size = sizeof(struct vhost_msg);
1503
1504 if (iov_iter_count(to) < size)
1505 return 0;
1506
1507 while (1) {
1508 if (!noblock)
1509 prepare_to_wait(&dev->wait, &wait,
1510 TASK_INTERRUPTIBLE);
1511
1512 node = vhost_dequeue_msg(dev, &dev->read_list);
1513 if (node)
1514 break;
1515 if (noblock) {
1516 ret = -EAGAIN;
1517 break;
1518 }
1519 if (signal_pending(current)) {
1520 ret = -ERESTARTSYS;
1521 break;
1522 }
1523 if (!dev->iotlb) {
1524 ret = -EBADFD;
1525 break;
1526 }
1527
1528 schedule();
1529 }
1530
1531 if (!noblock)
1532 finish_wait(&dev->wait, &wait);
1533
1534 if (node) {
1535 struct vhost_iotlb_msg *msg;
1536 void *start = &node->msg;
1537
1538 switch (node->msg.type) {
1539 case VHOST_IOTLB_MSG:
1540 size = sizeof(node->msg);
1541 msg = &node->msg.iotlb;
1542 break;
1543 case VHOST_IOTLB_MSG_V2:
1544 size = sizeof(node->msg_v2);
1545 msg = &node->msg_v2.iotlb;
1546 break;
1547 default:
1548 BUG();
1549 break;
1550 }
1551
1552 ret = copy_to_iter(start, size, to);
1553 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1554 kfree(node);
1555 return ret;
1556 }
1557 vhost_enqueue_msg(dev, &dev->pending_list, node);
1558 }
1559
1560 return ret;
1561}
1562EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1563
1564static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1565{
1566 struct vhost_dev *dev = vq->dev;
1567 struct vhost_msg_node *node;
1568 struct vhost_iotlb_msg *msg;
1569 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1570
1571 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1572 if (!node)
1573 return -ENOMEM;
1574
1575 if (v2) {
1576 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1577 msg = &node->msg_v2.iotlb;
1578 } else {
1579 msg = &node->msg.iotlb;
1580 }
1581
1582 msg->type = VHOST_IOTLB_MISS;
1583 msg->iova = iova;
1584 msg->perm = access;
1585
1586 vhost_enqueue_msg(dev, &dev->read_list, node);
1587
1588 return 0;
1589}
1590
1591static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1592 vring_desc_t __user *desc,
1593 vring_avail_t __user *avail,
1594 vring_used_t __user *used)
1595
1596{
1597 /* If an IOTLB device is present, the vring addresses are
1598 * GIOVAs. Access validation occurs at prefetch time. */
1599 if (vq->iotlb)
1600 return true;
1601
1602 return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1603 access_ok(avail, vhost_get_avail_size(vq, num)) &&
1604 access_ok(used, vhost_get_used_size(vq, num));
1605}
1606
1607static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1608 const struct vhost_iotlb_map *map,
1609 int type)
1610{
1611 int access = (type == VHOST_ADDR_USED) ?
1612 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1613
1614 if (likely(map->perm & access))
1615 vq->meta_iotlb[type] = map;
1616}
1617
1618static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1619 int access, u64 addr, u64 len, int type)
1620{
1621 const struct vhost_iotlb_map *map;
1622 struct vhost_iotlb *umem = vq->iotlb;
1623 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1624
1625 if (vhost_vq_meta_fetch(vq, addr, len, type))
1626 return true;
1627
1628 while (len > s) {
1629 map = vhost_iotlb_itree_first(umem, addr, last);
1630 if (map == NULL || map->start > addr) {
1631 vhost_iotlb_miss(vq, addr, access);
1632 return false;
1633 } else if (!(map->perm & access)) {
1634 /* Report the possible access violation by
1635 * request another translation from userspace.
1636 */
1637 return false;
1638 }
1639
1640 size = map->size - addr + map->start;
1641
1642 if (orig_addr == addr && size >= len)
1643 vhost_vq_meta_update(vq, map, type);
1644
1645 s += size;
1646 addr += size;
1647 }
1648
1649 return true;
1650}
1651
1652int vq_meta_prefetch(struct vhost_virtqueue *vq)
1653{
1654 unsigned int num = vq->num;
1655
1656 if (!vq->iotlb)
1657 return 1;
1658
1659 return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1660 vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1661 iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1662 vhost_get_avail_size(vq, num),
1663 VHOST_ADDR_AVAIL) &&
1664 iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1665 vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1666}
1667EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1668
1669/* Can we log writes? */
1670/* Caller should have device mutex but not vq mutex */
1671bool vhost_log_access_ok(struct vhost_dev *dev)
1672{
1673 return memory_access_ok(dev, dev->umem, 1);
1674}
1675EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1676
1677static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1678 void __user *log_base,
1679 bool log_used,
1680 u64 log_addr)
1681{
1682 /* If an IOTLB device is present, log_addr is a GIOVA that
1683 * will never be logged by log_used(). */
1684 if (vq->iotlb)
1685 return true;
1686
1687 return !log_used || log_access_ok(log_base, log_addr,
1688 vhost_get_used_size(vq, vq->num));
1689}
1690
1691/* Verify access for write logging. */
1692/* Caller should have vq mutex and device mutex */
1693static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1694 void __user *log_base)
1695{
1696 return vq_memory_access_ok(log_base, vq->umem,
1697 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1698 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1699}
1700
1701/* Can we start vq? */
1702/* Caller should have vq mutex and device mutex */
1703bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1704{
1705 if (!vq_log_access_ok(vq, vq->log_base))
1706 return false;
1707
1708 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1709}
1710EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1711
1712static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1713{
1714 struct vhost_memory mem, *newmem;
1715 struct vhost_memory_region *region;
1716 struct vhost_iotlb *newumem, *oldumem;
1717 unsigned long size = offsetof(struct vhost_memory, regions);
1718 int i;
1719
1720 if (copy_from_user(&mem, m, size))
1721 return -EFAULT;
1722 if (mem.padding)
1723 return -EOPNOTSUPP;
1724 if (mem.nregions > max_mem_regions)
1725 return -E2BIG;
1726 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1727 GFP_KERNEL);
1728 if (!newmem)
1729 return -ENOMEM;
1730
1731 memcpy(newmem, &mem, size);
1732 if (copy_from_user(newmem->regions, m->regions,
1733 flex_array_size(newmem, regions, mem.nregions))) {
1734 kvfree(newmem);
1735 return -EFAULT;
1736 }
1737
1738 newumem = iotlb_alloc();
1739 if (!newumem) {
1740 kvfree(newmem);
1741 return -ENOMEM;
1742 }
1743
1744 for (region = newmem->regions;
1745 region < newmem->regions + mem.nregions;
1746 region++) {
1747 if (vhost_iotlb_add_range(newumem,
1748 region->guest_phys_addr,
1749 region->guest_phys_addr +
1750 region->memory_size - 1,
1751 region->userspace_addr,
1752 VHOST_MAP_RW))
1753 goto err;
1754 }
1755
1756 if (!memory_access_ok(d, newumem, 0))
1757 goto err;
1758
1759 oldumem = d->umem;
1760 d->umem = newumem;
1761
1762 /* All memory accesses are done under some VQ mutex. */
1763 for (i = 0; i < d->nvqs; ++i) {
1764 mutex_lock(&d->vqs[i]->mutex);
1765 d->vqs[i]->umem = newumem;
1766 mutex_unlock(&d->vqs[i]->mutex);
1767 }
1768
1769 kvfree(newmem);
1770 vhost_iotlb_free(oldumem);
1771 return 0;
1772
1773err:
1774 vhost_iotlb_free(newumem);
1775 kvfree(newmem);
1776 return -EFAULT;
1777}
1778
1779static long vhost_vring_set_num(struct vhost_dev *d,
1780 struct vhost_virtqueue *vq,
1781 void __user *argp)
1782{
1783 struct vhost_vring_state s;
1784
1785 /* Resizing ring with an active backend?
1786 * You don't want to do that. */
1787 if (vq->private_data)
1788 return -EBUSY;
1789
1790 if (copy_from_user(&s, argp, sizeof s))
1791 return -EFAULT;
1792
1793 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1794 return -EINVAL;
1795 vq->num = s.num;
1796
1797 return 0;
1798}
1799
1800static long vhost_vring_set_addr(struct vhost_dev *d,
1801 struct vhost_virtqueue *vq,
1802 void __user *argp)
1803{
1804 struct vhost_vring_addr a;
1805
1806 if (copy_from_user(&a, argp, sizeof a))
1807 return -EFAULT;
1808 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1809 return -EOPNOTSUPP;
1810
1811 /* For 32bit, verify that the top 32bits of the user
1812 data are set to zero. */
1813 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1814 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1815 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1816 return -EFAULT;
1817
1818 /* Make sure it's safe to cast pointers to vring types. */
1819 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1820 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1821 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1822 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1823 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1824 return -EINVAL;
1825
1826 /* We only verify access here if backend is configured.
1827 * If it is not, we don't as size might not have been setup.
1828 * We will verify when backend is configured. */
1829 if (vq->private_data) {
1830 if (!vq_access_ok(vq, vq->num,
1831 (void __user *)(unsigned long)a.desc_user_addr,
1832 (void __user *)(unsigned long)a.avail_user_addr,
1833 (void __user *)(unsigned long)a.used_user_addr))
1834 return -EINVAL;
1835
1836 /* Also validate log access for used ring if enabled. */
1837 if (!vq_log_used_access_ok(vq, vq->log_base,
1838 a.flags & (0x1 << VHOST_VRING_F_LOG),
1839 a.log_guest_addr))
1840 return -EINVAL;
1841 }
1842
1843 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1844 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1845 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1846 vq->log_addr = a.log_guest_addr;
1847 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1848
1849 return 0;
1850}
1851
1852static long vhost_vring_set_num_addr(struct vhost_dev *d,
1853 struct vhost_virtqueue *vq,
1854 unsigned int ioctl,
1855 void __user *argp)
1856{
1857 long r;
1858
1859 mutex_lock(&vq->mutex);
1860
1861 switch (ioctl) {
1862 case VHOST_SET_VRING_NUM:
1863 r = vhost_vring_set_num(d, vq, argp);
1864 break;
1865 case VHOST_SET_VRING_ADDR:
1866 r = vhost_vring_set_addr(d, vq, argp);
1867 break;
1868 default:
1869 BUG();
1870 }
1871
1872 mutex_unlock(&vq->mutex);
1873
1874 return r;
1875}
1876long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1877{
1878 struct file *eventfp, *filep = NULL;
1879 bool pollstart = false, pollstop = false;
1880 struct eventfd_ctx *ctx = NULL;
1881 struct vhost_virtqueue *vq;
1882 struct vhost_vring_state s;
1883 struct vhost_vring_file f;
1884 u32 idx;
1885 long r;
1886
1887 r = vhost_get_vq_from_user(d, argp, &vq, &idx);
1888 if (r < 0)
1889 return r;
1890
1891 if (ioctl == VHOST_SET_VRING_NUM ||
1892 ioctl == VHOST_SET_VRING_ADDR) {
1893 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1894 }
1895
1896 mutex_lock(&vq->mutex);
1897
1898 switch (ioctl) {
1899 case VHOST_SET_VRING_BASE:
1900 /* Moving base with an active backend?
1901 * You don't want to do that. */
1902 if (vq->private_data) {
1903 r = -EBUSY;
1904 break;
1905 }
1906 if (copy_from_user(&s, argp, sizeof s)) {
1907 r = -EFAULT;
1908 break;
1909 }
1910 if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
1911 vq->last_avail_idx = s.num & 0xffff;
1912 vq->last_used_idx = (s.num >> 16) & 0xffff;
1913 } else {
1914 if (s.num > 0xffff) {
1915 r = -EINVAL;
1916 break;
1917 }
1918 vq->last_avail_idx = s.num;
1919 }
1920 /* Forget the cached index value. */
1921 vq->avail_idx = vq->last_avail_idx;
1922 break;
1923 case VHOST_GET_VRING_BASE:
1924 s.index = idx;
1925 if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED))
1926 s.num = (u32)vq->last_avail_idx | ((u32)vq->last_used_idx << 16);
1927 else
1928 s.num = vq->last_avail_idx;
1929 if (copy_to_user(argp, &s, sizeof s))
1930 r = -EFAULT;
1931 break;
1932 case VHOST_SET_VRING_KICK:
1933 if (copy_from_user(&f, argp, sizeof f)) {
1934 r = -EFAULT;
1935 break;
1936 }
1937 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1938 if (IS_ERR(eventfp)) {
1939 r = PTR_ERR(eventfp);
1940 break;
1941 }
1942 if (eventfp != vq->kick) {
1943 pollstop = (filep = vq->kick) != NULL;
1944 pollstart = (vq->kick = eventfp) != NULL;
1945 } else
1946 filep = eventfp;
1947 break;
1948 case VHOST_SET_VRING_CALL:
1949 if (copy_from_user(&f, argp, sizeof f)) {
1950 r = -EFAULT;
1951 break;
1952 }
1953 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1954 if (IS_ERR(ctx)) {
1955 r = PTR_ERR(ctx);
1956 break;
1957 }
1958
1959 swap(ctx, vq->call_ctx.ctx);
1960 break;
1961 case VHOST_SET_VRING_ERR:
1962 if (copy_from_user(&f, argp, sizeof f)) {
1963 r = -EFAULT;
1964 break;
1965 }
1966 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1967 if (IS_ERR(ctx)) {
1968 r = PTR_ERR(ctx);
1969 break;
1970 }
1971 swap(ctx, vq->error_ctx);
1972 break;
1973 case VHOST_SET_VRING_ENDIAN:
1974 r = vhost_set_vring_endian(vq, argp);
1975 break;
1976 case VHOST_GET_VRING_ENDIAN:
1977 r = vhost_get_vring_endian(vq, idx, argp);
1978 break;
1979 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1980 if (copy_from_user(&s, argp, sizeof(s))) {
1981 r = -EFAULT;
1982 break;
1983 }
1984 vq->busyloop_timeout = s.num;
1985 break;
1986 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1987 s.index = idx;
1988 s.num = vq->busyloop_timeout;
1989 if (copy_to_user(argp, &s, sizeof(s)))
1990 r = -EFAULT;
1991 break;
1992 default:
1993 r = -ENOIOCTLCMD;
1994 }
1995
1996 if (pollstop && vq->handle_kick)
1997 vhost_poll_stop(&vq->poll);
1998
1999 if (!IS_ERR_OR_NULL(ctx))
2000 eventfd_ctx_put(ctx);
2001 if (filep)
2002 fput(filep);
2003
2004 if (pollstart && vq->handle_kick)
2005 r = vhost_poll_start(&vq->poll, vq->kick);
2006
2007 mutex_unlock(&vq->mutex);
2008
2009 if (pollstop && vq->handle_kick)
2010 vhost_dev_flush(vq->poll.dev);
2011 return r;
2012}
2013EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
2014
2015int vhost_init_device_iotlb(struct vhost_dev *d)
2016{
2017 struct vhost_iotlb *niotlb, *oiotlb;
2018 int i;
2019
2020 niotlb = iotlb_alloc();
2021 if (!niotlb)
2022 return -ENOMEM;
2023
2024 oiotlb = d->iotlb;
2025 d->iotlb = niotlb;
2026
2027 for (i = 0; i < d->nvqs; ++i) {
2028 struct vhost_virtqueue *vq = d->vqs[i];
2029
2030 mutex_lock(&vq->mutex);
2031 vq->iotlb = niotlb;
2032 __vhost_vq_meta_reset(vq);
2033 mutex_unlock(&vq->mutex);
2034 }
2035
2036 vhost_iotlb_free(oiotlb);
2037
2038 return 0;
2039}
2040EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
2041
2042/* Caller must have device mutex */
2043long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
2044{
2045 struct eventfd_ctx *ctx;
2046 u64 p;
2047 long r;
2048 int i, fd;
2049
2050 /* If you are not the owner, you can become one */
2051 if (ioctl == VHOST_SET_OWNER) {
2052 r = vhost_dev_set_owner(d);
2053 goto done;
2054 }
2055
2056 /* You must be the owner to do anything else */
2057 r = vhost_dev_check_owner(d);
2058 if (r)
2059 goto done;
2060
2061 switch (ioctl) {
2062 case VHOST_SET_MEM_TABLE:
2063 r = vhost_set_memory(d, argp);
2064 break;
2065 case VHOST_SET_LOG_BASE:
2066 if (copy_from_user(&p, argp, sizeof p)) {
2067 r = -EFAULT;
2068 break;
2069 }
2070 if ((u64)(unsigned long)p != p) {
2071 r = -EFAULT;
2072 break;
2073 }
2074 for (i = 0; i < d->nvqs; ++i) {
2075 struct vhost_virtqueue *vq;
2076 void __user *base = (void __user *)(unsigned long)p;
2077 vq = d->vqs[i];
2078 mutex_lock(&vq->mutex);
2079 /* If ring is inactive, will check when it's enabled. */
2080 if (vq->private_data && !vq_log_access_ok(vq, base))
2081 r = -EFAULT;
2082 else
2083 vq->log_base = base;
2084 mutex_unlock(&vq->mutex);
2085 }
2086 break;
2087 case VHOST_SET_LOG_FD:
2088 r = get_user(fd, (int __user *)argp);
2089 if (r < 0)
2090 break;
2091 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
2092 if (IS_ERR(ctx)) {
2093 r = PTR_ERR(ctx);
2094 break;
2095 }
2096 swap(ctx, d->log_ctx);
2097 for (i = 0; i < d->nvqs; ++i) {
2098 mutex_lock(&d->vqs[i]->mutex);
2099 d->vqs[i]->log_ctx = d->log_ctx;
2100 mutex_unlock(&d->vqs[i]->mutex);
2101 }
2102 if (ctx)
2103 eventfd_ctx_put(ctx);
2104 break;
2105 default:
2106 r = -ENOIOCTLCMD;
2107 break;
2108 }
2109done:
2110 return r;
2111}
2112EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
2113
2114/* TODO: This is really inefficient. We need something like get_user()
2115 * (instruction directly accesses the data, with an exception table entry
2116 * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
2117 */
2118static int set_bit_to_user(int nr, void __user *addr)
2119{
2120 unsigned long log = (unsigned long)addr;
2121 struct page *page;
2122 void *base;
2123 int bit = nr + (log % PAGE_SIZE) * 8;
2124 int r;
2125
2126 r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
2127 if (r < 0)
2128 return r;
2129 BUG_ON(r != 1);
2130 base = kmap_atomic(page);
2131 set_bit(bit, base);
2132 kunmap_atomic(base);
2133 unpin_user_pages_dirty_lock(&page, 1, true);
2134 return 0;
2135}
2136
2137static int log_write(void __user *log_base,
2138 u64 write_address, u64 write_length)
2139{
2140 u64 write_page = write_address / VHOST_PAGE_SIZE;
2141 int r;
2142
2143 if (!write_length)
2144 return 0;
2145 write_length += write_address % VHOST_PAGE_SIZE;
2146 for (;;) {
2147 u64 base = (u64)(unsigned long)log_base;
2148 u64 log = base + write_page / 8;
2149 int bit = write_page % 8;
2150 if ((u64)(unsigned long)log != log)
2151 return -EFAULT;
2152 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
2153 if (r < 0)
2154 return r;
2155 if (write_length <= VHOST_PAGE_SIZE)
2156 break;
2157 write_length -= VHOST_PAGE_SIZE;
2158 write_page += 1;
2159 }
2160 return r;
2161}
2162
2163static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
2164{
2165 struct vhost_iotlb *umem = vq->umem;
2166 struct vhost_iotlb_map *u;
2167 u64 start, end, l, min;
2168 int r;
2169 bool hit = false;
2170
2171 while (len) {
2172 min = len;
2173 /* More than one GPAs can be mapped into a single HVA. So
2174 * iterate all possible umems here to be safe.
2175 */
2176 list_for_each_entry(u, &umem->list, link) {
2177 if (u->addr > hva - 1 + len ||
2178 u->addr - 1 + u->size < hva)
2179 continue;
2180 start = max(u->addr, hva);
2181 end = min(u->addr - 1 + u->size, hva - 1 + len);
2182 l = end - start + 1;
2183 r = log_write(vq->log_base,
2184 u->start + start - u->addr,
2185 l);
2186 if (r < 0)
2187 return r;
2188 hit = true;
2189 min = min(l, min);
2190 }
2191
2192 if (!hit)
2193 return -EFAULT;
2194
2195 len -= min;
2196 hva += min;
2197 }
2198
2199 return 0;
2200}
2201
2202static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
2203{
2204 struct iovec *iov = vq->log_iov;
2205 int i, ret;
2206
2207 if (!vq->iotlb)
2208 return log_write(vq->log_base, vq->log_addr + used_offset, len);
2209
2210 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
2211 len, iov, 64, VHOST_ACCESS_WO);
2212 if (ret < 0)
2213 return ret;
2214
2215 for (i = 0; i < ret; i++) {
2216 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
2217 iov[i].iov_len);
2218 if (ret)
2219 return ret;
2220 }
2221
2222 return 0;
2223}
2224
2225int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
2226 unsigned int log_num, u64 len, struct iovec *iov, int count)
2227{
2228 int i, r;
2229
2230 /* Make sure data written is seen before log. */
2231 smp_wmb();
2232
2233 if (vq->iotlb) {
2234 for (i = 0; i < count; i++) {
2235 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
2236 iov[i].iov_len);
2237 if (r < 0)
2238 return r;
2239 }
2240 return 0;
2241 }
2242
2243 for (i = 0; i < log_num; ++i) {
2244 u64 l = min(log[i].len, len);
2245 r = log_write(vq->log_base, log[i].addr, l);
2246 if (r < 0)
2247 return r;
2248 len -= l;
2249 if (!len) {
2250 if (vq->log_ctx)
2251 eventfd_signal(vq->log_ctx);
2252 return 0;
2253 }
2254 }
2255 /* Length written exceeds what we have stored. This is a bug. */
2256 BUG();
2257 return 0;
2258}
2259EXPORT_SYMBOL_GPL(vhost_log_write);
2260
2261static int vhost_update_used_flags(struct vhost_virtqueue *vq)
2262{
2263 void __user *used;
2264 if (vhost_put_used_flags(vq))
2265 return -EFAULT;
2266 if (unlikely(vq->log_used)) {
2267 /* Make sure the flag is seen before log. */
2268 smp_wmb();
2269 /* Log used flag write. */
2270 used = &vq->used->flags;
2271 log_used(vq, (used - (void __user *)vq->used),
2272 sizeof vq->used->flags);
2273 if (vq->log_ctx)
2274 eventfd_signal(vq->log_ctx);
2275 }
2276 return 0;
2277}
2278
2279static int vhost_update_avail_event(struct vhost_virtqueue *vq)
2280{
2281 if (vhost_put_avail_event(vq))
2282 return -EFAULT;
2283 if (unlikely(vq->log_used)) {
2284 void __user *used;
2285 /* Make sure the event is seen before log. */
2286 smp_wmb();
2287 /* Log avail event write */
2288 used = vhost_avail_event(vq);
2289 log_used(vq, (used - (void __user *)vq->used),
2290 sizeof *vhost_avail_event(vq));
2291 if (vq->log_ctx)
2292 eventfd_signal(vq->log_ctx);
2293 }
2294 return 0;
2295}
2296
2297int vhost_vq_init_access(struct vhost_virtqueue *vq)
2298{
2299 __virtio16 last_used_idx;
2300 int r;
2301 bool is_le = vq->is_le;
2302
2303 if (!vq->private_data)
2304 return 0;
2305
2306 vhost_init_is_le(vq);
2307
2308 r = vhost_update_used_flags(vq);
2309 if (r)
2310 goto err;
2311 vq->signalled_used_valid = false;
2312 if (!vq->iotlb &&
2313 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2314 r = -EFAULT;
2315 goto err;
2316 }
2317 r = vhost_get_used_idx(vq, &last_used_idx);
2318 if (r) {
2319 vq_err(vq, "Can't access used idx at %p\n",
2320 &vq->used->idx);
2321 goto err;
2322 }
2323 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2324 return 0;
2325
2326err:
2327 vq->is_le = is_le;
2328 return r;
2329}
2330EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2331
2332static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2333 struct iovec iov[], int iov_size, int access)
2334{
2335 const struct vhost_iotlb_map *map;
2336 struct vhost_dev *dev = vq->dev;
2337 struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2338 struct iovec *_iov;
2339 u64 s = 0, last = addr + len - 1;
2340 int ret = 0;
2341
2342 while ((u64)len > s) {
2343 u64 size;
2344 if (unlikely(ret >= iov_size)) {
2345 ret = -ENOBUFS;
2346 break;
2347 }
2348
2349 map = vhost_iotlb_itree_first(umem, addr, last);
2350 if (map == NULL || map->start > addr) {
2351 if (umem != dev->iotlb) {
2352 ret = -EFAULT;
2353 break;
2354 }
2355 ret = -EAGAIN;
2356 break;
2357 } else if (!(map->perm & access)) {
2358 ret = -EPERM;
2359 break;
2360 }
2361
2362 _iov = iov + ret;
2363 size = map->size - addr + map->start;
2364 _iov->iov_len = min((u64)len - s, size);
2365 _iov->iov_base = (void __user *)(unsigned long)
2366 (map->addr + addr - map->start);
2367 s += size;
2368 addr += size;
2369 ++ret;
2370 }
2371
2372 if (ret == -EAGAIN)
2373 vhost_iotlb_miss(vq, addr, access);
2374 return ret;
2375}
2376
2377/* Each buffer in the virtqueues is actually a chain of descriptors. This
2378 * function returns the next descriptor in the chain,
2379 * or -1U if we're at the end. */
2380static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2381{
2382 unsigned int next;
2383
2384 /* If this descriptor says it doesn't chain, we're done. */
2385 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2386 return -1U;
2387
2388 /* Check they're not leading us off end of descriptors. */
2389 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2390 return next;
2391}
2392
2393static int get_indirect(struct vhost_virtqueue *vq,
2394 struct iovec iov[], unsigned int iov_size,
2395 unsigned int *out_num, unsigned int *in_num,
2396 struct vhost_log *log, unsigned int *log_num,
2397 struct vring_desc *indirect)
2398{
2399 struct vring_desc desc;
2400 unsigned int i = 0, count, found = 0;
2401 u32 len = vhost32_to_cpu(vq, indirect->len);
2402 struct iov_iter from;
2403 int ret, access;
2404
2405 /* Sanity check */
2406 if (unlikely(len % sizeof desc)) {
2407 vq_err(vq, "Invalid length in indirect descriptor: "
2408 "len 0x%llx not multiple of 0x%zx\n",
2409 (unsigned long long)len,
2410 sizeof desc);
2411 return -EINVAL;
2412 }
2413
2414 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2415 UIO_MAXIOV, VHOST_ACCESS_RO);
2416 if (unlikely(ret < 0)) {
2417 if (ret != -EAGAIN)
2418 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2419 return ret;
2420 }
2421 iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2422 count = len / sizeof desc;
2423 /* Buffers are chained via a 16 bit next field, so
2424 * we can have at most 2^16 of these. */
2425 if (unlikely(count > USHRT_MAX + 1)) {
2426 vq_err(vq, "Indirect buffer length too big: %d\n",
2427 indirect->len);
2428 return -E2BIG;
2429 }
2430
2431 do {
2432 unsigned iov_count = *in_num + *out_num;
2433 if (unlikely(++found > count)) {
2434 vq_err(vq, "Loop detected: last one at %u "
2435 "indirect size %u\n",
2436 i, count);
2437 return -EINVAL;
2438 }
2439 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2440 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2441 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2442 return -EINVAL;
2443 }
2444 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2445 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2446 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2447 return -EINVAL;
2448 }
2449
2450 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2451 access = VHOST_ACCESS_WO;
2452 else
2453 access = VHOST_ACCESS_RO;
2454
2455 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2456 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2457 iov_size - iov_count, access);
2458 if (unlikely(ret < 0)) {
2459 if (ret != -EAGAIN)
2460 vq_err(vq, "Translation failure %d indirect idx %d\n",
2461 ret, i);
2462 return ret;
2463 }
2464 /* If this is an input descriptor, increment that count. */
2465 if (access == VHOST_ACCESS_WO) {
2466 *in_num += ret;
2467 if (unlikely(log && ret)) {
2468 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2469 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2470 ++*log_num;
2471 }
2472 } else {
2473 /* If it's an output descriptor, they're all supposed
2474 * to come before any input descriptors. */
2475 if (unlikely(*in_num)) {
2476 vq_err(vq, "Indirect descriptor "
2477 "has out after in: idx %d\n", i);
2478 return -EINVAL;
2479 }
2480 *out_num += ret;
2481 }
2482 } while ((i = next_desc(vq, &desc)) != -1);
2483 return 0;
2484}
2485
2486/* This looks in the virtqueue and for the first available buffer, and converts
2487 * it to an iovec for convenient access. Since descriptors consist of some
2488 * number of output then some number of input descriptors, it's actually two
2489 * iovecs, but we pack them into one and note how many of each there were.
2490 *
2491 * This function returns the descriptor number found, or vq->num (which is
2492 * never a valid descriptor number) if none was found. A negative code is
2493 * returned on error. */
2494int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2495 struct iovec iov[], unsigned int iov_size,
2496 unsigned int *out_num, unsigned int *in_num,
2497 struct vhost_log *log, unsigned int *log_num)
2498{
2499 struct vring_desc desc;
2500 unsigned int i, head, found = 0;
2501 u16 last_avail_idx;
2502 __virtio16 avail_idx;
2503 __virtio16 ring_head;
2504 int ret, access;
2505
2506 /* Check it isn't doing very strange things with descriptor numbers. */
2507 last_avail_idx = vq->last_avail_idx;
2508
2509 if (vq->avail_idx == vq->last_avail_idx) {
2510 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2511 vq_err(vq, "Failed to access avail idx at %p\n",
2512 &vq->avail->idx);
2513 return -EFAULT;
2514 }
2515 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2516
2517 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2518 vq_err(vq, "Guest moved used index from %u to %u",
2519 last_avail_idx, vq->avail_idx);
2520 return -EFAULT;
2521 }
2522
2523 /* If there's nothing new since last we looked, return
2524 * invalid.
2525 */
2526 if (vq->avail_idx == last_avail_idx)
2527 return vq->num;
2528
2529 /* Only get avail ring entries after they have been
2530 * exposed by guest.
2531 */
2532 smp_rmb();
2533 }
2534
2535 /* Grab the next descriptor number they're advertising, and increment
2536 * the index we've seen. */
2537 if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2538 vq_err(vq, "Failed to read head: idx %d address %p\n",
2539 last_avail_idx,
2540 &vq->avail->ring[last_avail_idx % vq->num]);
2541 return -EFAULT;
2542 }
2543
2544 head = vhost16_to_cpu(vq, ring_head);
2545
2546 /* If their number is silly, that's an error. */
2547 if (unlikely(head >= vq->num)) {
2548 vq_err(vq, "Guest says index %u > %u is available",
2549 head, vq->num);
2550 return -EINVAL;
2551 }
2552
2553 /* When we start there are none of either input nor output. */
2554 *out_num = *in_num = 0;
2555 if (unlikely(log))
2556 *log_num = 0;
2557
2558 i = head;
2559 do {
2560 unsigned iov_count = *in_num + *out_num;
2561 if (unlikely(i >= vq->num)) {
2562 vq_err(vq, "Desc index is %u > %u, head = %u",
2563 i, vq->num, head);
2564 return -EINVAL;
2565 }
2566 if (unlikely(++found > vq->num)) {
2567 vq_err(vq, "Loop detected: last one at %u "
2568 "vq size %u head %u\n",
2569 i, vq->num, head);
2570 return -EINVAL;
2571 }
2572 ret = vhost_get_desc(vq, &desc, i);
2573 if (unlikely(ret)) {
2574 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2575 i, vq->desc + i);
2576 return -EFAULT;
2577 }
2578 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2579 ret = get_indirect(vq, iov, iov_size,
2580 out_num, in_num,
2581 log, log_num, &desc);
2582 if (unlikely(ret < 0)) {
2583 if (ret != -EAGAIN)
2584 vq_err(vq, "Failure detected "
2585 "in indirect descriptor at idx %d\n", i);
2586 return ret;
2587 }
2588 continue;
2589 }
2590
2591 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2592 access = VHOST_ACCESS_WO;
2593 else
2594 access = VHOST_ACCESS_RO;
2595 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2596 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2597 iov_size - iov_count, access);
2598 if (unlikely(ret < 0)) {
2599 if (ret != -EAGAIN)
2600 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2601 ret, i);
2602 return ret;
2603 }
2604 if (access == VHOST_ACCESS_WO) {
2605 /* If this is an input descriptor,
2606 * increment that count. */
2607 *in_num += ret;
2608 if (unlikely(log && ret)) {
2609 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2610 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2611 ++*log_num;
2612 }
2613 } else {
2614 /* If it's an output descriptor, they're all supposed
2615 * to come before any input descriptors. */
2616 if (unlikely(*in_num)) {
2617 vq_err(vq, "Descriptor has out after in: "
2618 "idx %d\n", i);
2619 return -EINVAL;
2620 }
2621 *out_num += ret;
2622 }
2623 } while ((i = next_desc(vq, &desc)) != -1);
2624
2625 /* On success, increment avail index. */
2626 vq->last_avail_idx++;
2627
2628 /* Assume notifications from guest are disabled at this point,
2629 * if they aren't we would need to update avail_event index. */
2630 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2631 return head;
2632}
2633EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2634
2635/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2636void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2637{
2638 vq->last_avail_idx -= n;
2639}
2640EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2641
2642/* After we've used one of their buffers, we tell them about it. We'll then
2643 * want to notify the guest, using eventfd. */
2644int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2645{
2646 struct vring_used_elem heads = {
2647 cpu_to_vhost32(vq, head),
2648 cpu_to_vhost32(vq, len)
2649 };
2650
2651 return vhost_add_used_n(vq, &heads, 1);
2652}
2653EXPORT_SYMBOL_GPL(vhost_add_used);
2654
2655static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2656 struct vring_used_elem *heads,
2657 unsigned count)
2658{
2659 vring_used_elem_t __user *used;
2660 u16 old, new;
2661 int start;
2662
2663 start = vq->last_used_idx & (vq->num - 1);
2664 used = vq->used->ring + start;
2665 if (vhost_put_used(vq, heads, start, count)) {
2666 vq_err(vq, "Failed to write used");
2667 return -EFAULT;
2668 }
2669 if (unlikely(vq->log_used)) {
2670 /* Make sure data is seen before log. */
2671 smp_wmb();
2672 /* Log used ring entry write. */
2673 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2674 count * sizeof *used);
2675 }
2676 old = vq->last_used_idx;
2677 new = (vq->last_used_idx += count);
2678 /* If the driver never bothers to signal in a very long while,
2679 * used index might wrap around. If that happens, invalidate
2680 * signalled_used index we stored. TODO: make sure driver
2681 * signals at least once in 2^16 and remove this. */
2682 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2683 vq->signalled_used_valid = false;
2684 return 0;
2685}
2686
2687/* After we've used one of their buffers, we tell them about it. We'll then
2688 * want to notify the guest, using eventfd. */
2689int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2690 unsigned count)
2691{
2692 int start, n, r;
2693
2694 start = vq->last_used_idx & (vq->num - 1);
2695 n = vq->num - start;
2696 if (n < count) {
2697 r = __vhost_add_used_n(vq, heads, n);
2698 if (r < 0)
2699 return r;
2700 heads += n;
2701 count -= n;
2702 }
2703 r = __vhost_add_used_n(vq, heads, count);
2704
2705 /* Make sure buffer is written before we update index. */
2706 smp_wmb();
2707 if (vhost_put_used_idx(vq)) {
2708 vq_err(vq, "Failed to increment used idx");
2709 return -EFAULT;
2710 }
2711 if (unlikely(vq->log_used)) {
2712 /* Make sure used idx is seen before log. */
2713 smp_wmb();
2714 /* Log used index update. */
2715 log_used(vq, offsetof(struct vring_used, idx),
2716 sizeof vq->used->idx);
2717 if (vq->log_ctx)
2718 eventfd_signal(vq->log_ctx);
2719 }
2720 return r;
2721}
2722EXPORT_SYMBOL_GPL(vhost_add_used_n);
2723
2724static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2725{
2726 __u16 old, new;
2727 __virtio16 event;
2728 bool v;
2729 /* Flush out used index updates. This is paired
2730 * with the barrier that the Guest executes when enabling
2731 * interrupts. */
2732 smp_mb();
2733
2734 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2735 unlikely(vq->avail_idx == vq->last_avail_idx))
2736 return true;
2737
2738 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2739 __virtio16 flags;
2740 if (vhost_get_avail_flags(vq, &flags)) {
2741 vq_err(vq, "Failed to get flags");
2742 return true;
2743 }
2744 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2745 }
2746 old = vq->signalled_used;
2747 v = vq->signalled_used_valid;
2748 new = vq->signalled_used = vq->last_used_idx;
2749 vq->signalled_used_valid = true;
2750
2751 if (unlikely(!v))
2752 return true;
2753
2754 if (vhost_get_used_event(vq, &event)) {
2755 vq_err(vq, "Failed to get used event idx");
2756 return true;
2757 }
2758 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2759}
2760
2761/* This actually signals the guest, using eventfd. */
2762void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2763{
2764 /* Signal the Guest tell them we used something up. */
2765 if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2766 eventfd_signal(vq->call_ctx.ctx);
2767}
2768EXPORT_SYMBOL_GPL(vhost_signal);
2769
2770/* And here's the combo meal deal. Supersize me! */
2771void vhost_add_used_and_signal(struct vhost_dev *dev,
2772 struct vhost_virtqueue *vq,
2773 unsigned int head, int len)
2774{
2775 vhost_add_used(vq, head, len);
2776 vhost_signal(dev, vq);
2777}
2778EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2779
2780/* multi-buffer version of vhost_add_used_and_signal */
2781void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2782 struct vhost_virtqueue *vq,
2783 struct vring_used_elem *heads, unsigned count)
2784{
2785 vhost_add_used_n(vq, heads, count);
2786 vhost_signal(dev, vq);
2787}
2788EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2789
2790/* return true if we're sure that avaiable ring is empty */
2791bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2792{
2793 __virtio16 avail_idx;
2794 int r;
2795
2796 if (vq->avail_idx != vq->last_avail_idx)
2797 return false;
2798
2799 r = vhost_get_avail_idx(vq, &avail_idx);
2800 if (unlikely(r))
2801 return false;
2802 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2803
2804 return vq->avail_idx == vq->last_avail_idx;
2805}
2806EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2807
2808/* OK, now we need to know about added descriptors. */
2809bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2810{
2811 __virtio16 avail_idx;
2812 int r;
2813
2814 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2815 return false;
2816 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2817 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2818 r = vhost_update_used_flags(vq);
2819 if (r) {
2820 vq_err(vq, "Failed to enable notification at %p: %d\n",
2821 &vq->used->flags, r);
2822 return false;
2823 }
2824 } else {
2825 r = vhost_update_avail_event(vq);
2826 if (r) {
2827 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2828 vhost_avail_event(vq), r);
2829 return false;
2830 }
2831 }
2832 /* They could have slipped one in as we were doing that: make
2833 * sure it's written, then check again. */
2834 smp_mb();
2835 r = vhost_get_avail_idx(vq, &avail_idx);
2836 if (r) {
2837 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2838 &vq->avail->idx, r);
2839 return false;
2840 }
2841 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2842
2843 return vq->avail_idx != vq->last_avail_idx;
2844}
2845EXPORT_SYMBOL_GPL(vhost_enable_notify);
2846
2847/* We don't need to be notified again. */
2848void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2849{
2850 int r;
2851
2852 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2853 return;
2854 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2855 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2856 r = vhost_update_used_flags(vq);
2857 if (r)
2858 vq_err(vq, "Failed to disable notification at %p: %d\n",
2859 &vq->used->flags, r);
2860 }
2861}
2862EXPORT_SYMBOL_GPL(vhost_disable_notify);
2863
2864/* Create a new message. */
2865struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2866{
2867 /* Make sure all padding within the structure is initialized. */
2868 struct vhost_msg_node *node = kzalloc(sizeof(*node), GFP_KERNEL);
2869 if (!node)
2870 return NULL;
2871
2872 node->vq = vq;
2873 node->msg.type = type;
2874 return node;
2875}
2876EXPORT_SYMBOL_GPL(vhost_new_msg);
2877
2878void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2879 struct vhost_msg_node *node)
2880{
2881 spin_lock(&dev->iotlb_lock);
2882 list_add_tail(&node->node, head);
2883 spin_unlock(&dev->iotlb_lock);
2884
2885 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2886}
2887EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2888
2889struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2890 struct list_head *head)
2891{
2892 struct vhost_msg_node *node = NULL;
2893
2894 spin_lock(&dev->iotlb_lock);
2895 if (!list_empty(head)) {
2896 node = list_first_entry(head, struct vhost_msg_node,
2897 node);
2898 list_del(&node->node);
2899 }
2900 spin_unlock(&dev->iotlb_lock);
2901
2902 return node;
2903}
2904EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2905
2906void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2907{
2908 struct vhost_virtqueue *vq;
2909 int i;
2910
2911 mutex_lock(&dev->mutex);
2912 for (i = 0; i < dev->nvqs; ++i) {
2913 vq = dev->vqs[i];
2914 mutex_lock(&vq->mutex);
2915 vq->acked_backend_features = features;
2916 mutex_unlock(&vq->mutex);
2917 }
2918 mutex_unlock(&dev->mutex);
2919}
2920EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2921
2922static int __init vhost_init(void)
2923{
2924 return 0;
2925}
2926
2927static void __exit vhost_exit(void)
2928{
2929}
2930
2931module_init(vhost_init);
2932module_exit(vhost_exit);
2933
2934MODULE_VERSION("0.0.1");
2935MODULE_LICENSE("GPL v2");
2936MODULE_AUTHOR("Michael S. Tsirkin");
2937MODULE_DESCRIPTION("Host kernel accelerator for virtio");
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/mmu_context.h>
18#include <linux/miscdevice.h>
19#include <linux/mutex.h>
20#include <linux/poll.h>
21#include <linux/file.h>
22#include <linux/highmem.h>
23#include <linux/slab.h>
24#include <linux/vmalloc.h>
25#include <linux/kthread.h>
26#include <linux/cgroup.h>
27#include <linux/module.h>
28#include <linux/sort.h>
29#include <linux/sched/mm.h>
30#include <linux/sched/signal.h>
31#include <linux/interval_tree_generic.h>
32#include <linux/nospec.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
52INTERVAL_TREE_DEFINE(struct vhost_umem_node,
53 rb, __u64, __subtree_last,
54 START, LAST, static inline, vhost_umem_interval_tree);
55
56#ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
57static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
58{
59 vq->user_be = !virtio_legacy_is_little_endian();
60}
61
62static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
63{
64 vq->user_be = true;
65}
66
67static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
68{
69 vq->user_be = false;
70}
71
72static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
73{
74 struct vhost_vring_state s;
75
76 if (vq->private_data)
77 return -EBUSY;
78
79 if (copy_from_user(&s, argp, sizeof(s)))
80 return -EFAULT;
81
82 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
83 s.num != VHOST_VRING_BIG_ENDIAN)
84 return -EINVAL;
85
86 if (s.num == VHOST_VRING_BIG_ENDIAN)
87 vhost_enable_cross_endian_big(vq);
88 else
89 vhost_enable_cross_endian_little(vq);
90
91 return 0;
92}
93
94static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
95 int __user *argp)
96{
97 struct vhost_vring_state s = {
98 .index = idx,
99 .num = vq->user_be
100 };
101
102 if (copy_to_user(argp, &s, sizeof(s)))
103 return -EFAULT;
104
105 return 0;
106}
107
108static void vhost_init_is_le(struct vhost_virtqueue *vq)
109{
110 /* Note for legacy virtio: user_be is initialized at reset time
111 * according to the host endianness. If userspace does not set an
112 * explicit endianness, the default behavior is native endian, as
113 * expected by legacy virtio.
114 */
115 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
116}
117#else
118static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
119{
120}
121
122static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
123{
124 return -ENOIOCTLCMD;
125}
126
127static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
128 int __user *argp)
129{
130 return -ENOIOCTLCMD;
131}
132
133static void vhost_init_is_le(struct vhost_virtqueue *vq)
134{
135 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
136 || virtio_legacy_is_little_endian();
137}
138#endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
139
140static void vhost_reset_is_le(struct vhost_virtqueue *vq)
141{
142 vhost_init_is_le(vq);
143}
144
145struct vhost_flush_struct {
146 struct vhost_work work;
147 struct completion wait_event;
148};
149
150static void vhost_flush_work(struct vhost_work *work)
151{
152 struct vhost_flush_struct *s;
153
154 s = container_of(work, struct vhost_flush_struct, work);
155 complete(&s->wait_event);
156}
157
158static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
159 poll_table *pt)
160{
161 struct vhost_poll *poll;
162
163 poll = container_of(pt, struct vhost_poll, table);
164 poll->wqh = wqh;
165 add_wait_queue(wqh, &poll->wait);
166}
167
168static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
169 void *key)
170{
171 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
172
173 if (!(key_to_poll(key) & poll->mask))
174 return 0;
175
176 vhost_poll_queue(poll);
177 return 0;
178}
179
180void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
181{
182 clear_bit(VHOST_WORK_QUEUED, &work->flags);
183 work->fn = fn;
184}
185EXPORT_SYMBOL_GPL(vhost_work_init);
186
187/* Init poll structure */
188void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
189 __poll_t mask, struct vhost_dev *dev)
190{
191 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
192 init_poll_funcptr(&poll->table, vhost_poll_func);
193 poll->mask = mask;
194 poll->dev = dev;
195 poll->wqh = NULL;
196
197 vhost_work_init(&poll->work, fn);
198}
199EXPORT_SYMBOL_GPL(vhost_poll_init);
200
201/* Start polling a file. We add ourselves to file's wait queue. The caller must
202 * keep a reference to a file until after vhost_poll_stop is called. */
203int vhost_poll_start(struct vhost_poll *poll, struct file *file)
204{
205 __poll_t mask;
206
207 if (poll->wqh)
208 return 0;
209
210 mask = vfs_poll(file, &poll->table);
211 if (mask)
212 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
213 if (mask & EPOLLERR) {
214 vhost_poll_stop(poll);
215 return -EINVAL;
216 }
217
218 return 0;
219}
220EXPORT_SYMBOL_GPL(vhost_poll_start);
221
222/* Stop polling a file. After this function returns, it becomes safe to drop the
223 * file reference. You must also flush afterwards. */
224void vhost_poll_stop(struct vhost_poll *poll)
225{
226 if (poll->wqh) {
227 remove_wait_queue(poll->wqh, &poll->wait);
228 poll->wqh = NULL;
229 }
230}
231EXPORT_SYMBOL_GPL(vhost_poll_stop);
232
233void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
234{
235 struct vhost_flush_struct flush;
236
237 if (dev->worker) {
238 init_completion(&flush.wait_event);
239 vhost_work_init(&flush.work, vhost_flush_work);
240
241 vhost_work_queue(dev, &flush.work);
242 wait_for_completion(&flush.wait_event);
243 }
244}
245EXPORT_SYMBOL_GPL(vhost_work_flush);
246
247/* Flush any work that has been scheduled. When calling this, don't hold any
248 * locks that are also used by the callback. */
249void vhost_poll_flush(struct vhost_poll *poll)
250{
251 vhost_work_flush(poll->dev, &poll->work);
252}
253EXPORT_SYMBOL_GPL(vhost_poll_flush);
254
255void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
256{
257 if (!dev->worker)
258 return;
259
260 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
261 /* We can only add the work to the list after we're
262 * sure it was not in the list.
263 * test_and_set_bit() implies a memory barrier.
264 */
265 llist_add(&work->node, &dev->work_list);
266 wake_up_process(dev->worker);
267 }
268}
269EXPORT_SYMBOL_GPL(vhost_work_queue);
270
271/* A lockless hint for busy polling code to exit the loop */
272bool vhost_has_work(struct vhost_dev *dev)
273{
274 return !llist_empty(&dev->work_list);
275}
276EXPORT_SYMBOL_GPL(vhost_has_work);
277
278void vhost_poll_queue(struct vhost_poll *poll)
279{
280 vhost_work_queue(poll->dev, &poll->work);
281}
282EXPORT_SYMBOL_GPL(vhost_poll_queue);
283
284static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
285{
286 int j;
287
288 for (j = 0; j < VHOST_NUM_ADDRS; j++)
289 vq->meta_iotlb[j] = NULL;
290}
291
292static void vhost_vq_meta_reset(struct vhost_dev *d)
293{
294 int i;
295
296 for (i = 0; i < d->nvqs; ++i)
297 __vhost_vq_meta_reset(d->vqs[i]);
298}
299
300static void vhost_vq_reset(struct vhost_dev *dev,
301 struct vhost_virtqueue *vq)
302{
303 vq->num = 1;
304 vq->desc = NULL;
305 vq->avail = NULL;
306 vq->used = NULL;
307 vq->last_avail_idx = 0;
308 vq->avail_idx = 0;
309 vq->last_used_idx = 0;
310 vq->signalled_used = 0;
311 vq->signalled_used_valid = false;
312 vq->used_flags = 0;
313 vq->log_used = false;
314 vq->log_addr = -1ull;
315 vq->private_data = NULL;
316 vq->acked_features = 0;
317 vq->acked_backend_features = 0;
318 vq->log_base = NULL;
319 vq->error_ctx = NULL;
320 vq->kick = NULL;
321 vq->call_ctx = NULL;
322 vq->log_ctx = NULL;
323 vhost_reset_is_le(vq);
324 vhost_disable_cross_endian(vq);
325 vq->busyloop_timeout = 0;
326 vq->umem = NULL;
327 vq->iotlb = NULL;
328 __vhost_vq_meta_reset(vq);
329}
330
331static int vhost_worker(void *data)
332{
333 struct vhost_dev *dev = data;
334 struct vhost_work *work, *work_next;
335 struct llist_node *node;
336 mm_segment_t oldfs = get_fs();
337
338 set_fs(USER_DS);
339 use_mm(dev->mm);
340
341 for (;;) {
342 /* mb paired w/ kthread_stop */
343 set_current_state(TASK_INTERRUPTIBLE);
344
345 if (kthread_should_stop()) {
346 __set_current_state(TASK_RUNNING);
347 break;
348 }
349
350 node = llist_del_all(&dev->work_list);
351 if (!node)
352 schedule();
353
354 node = llist_reverse_order(node);
355 /* make sure flag is seen after deletion */
356 smp_wmb();
357 llist_for_each_entry_safe(work, work_next, node, node) {
358 clear_bit(VHOST_WORK_QUEUED, &work->flags);
359 __set_current_state(TASK_RUNNING);
360 work->fn(work);
361 if (need_resched())
362 schedule();
363 }
364 }
365 unuse_mm(dev->mm);
366 set_fs(oldfs);
367 return 0;
368}
369
370static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
371{
372 kfree(vq->indirect);
373 vq->indirect = NULL;
374 kfree(vq->log);
375 vq->log = NULL;
376 kfree(vq->heads);
377 vq->heads = NULL;
378}
379
380/* Helper to allocate iovec buffers for all vqs. */
381static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
382{
383 struct vhost_virtqueue *vq;
384 int i;
385
386 for (i = 0; i < dev->nvqs; ++i) {
387 vq = dev->vqs[i];
388 vq->indirect = kmalloc_array(UIO_MAXIOV,
389 sizeof(*vq->indirect),
390 GFP_KERNEL);
391 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
392 GFP_KERNEL);
393 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
394 GFP_KERNEL);
395 if (!vq->indirect || !vq->log || !vq->heads)
396 goto err_nomem;
397 }
398 return 0;
399
400err_nomem:
401 for (; i >= 0; --i)
402 vhost_vq_free_iovecs(dev->vqs[i]);
403 return -ENOMEM;
404}
405
406static void vhost_dev_free_iovecs(struct vhost_dev *dev)
407{
408 int i;
409
410 for (i = 0; i < dev->nvqs; ++i)
411 vhost_vq_free_iovecs(dev->vqs[i]);
412}
413
414bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
415 int pkts, int total_len)
416{
417 struct vhost_dev *dev = vq->dev;
418
419 if ((dev->byte_weight && total_len >= dev->byte_weight) ||
420 pkts >= dev->weight) {
421 vhost_poll_queue(&vq->poll);
422 return true;
423 }
424
425 return false;
426}
427EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
428
429static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
430 unsigned int num)
431{
432 size_t event __maybe_unused =
433 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
434
435 return sizeof(*vq->avail) +
436 sizeof(*vq->avail->ring) * num + event;
437}
438
439static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
440 unsigned int num)
441{
442 size_t event __maybe_unused =
443 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
444
445 return sizeof(*vq->used) +
446 sizeof(*vq->used->ring) * num + event;
447}
448
449static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
450 unsigned int num)
451{
452 return sizeof(*vq->desc) * num;
453}
454
455void vhost_dev_init(struct vhost_dev *dev,
456 struct vhost_virtqueue **vqs, int nvqs,
457 int iov_limit, int weight, int byte_weight)
458{
459 struct vhost_virtqueue *vq;
460 int i;
461
462 dev->vqs = vqs;
463 dev->nvqs = nvqs;
464 mutex_init(&dev->mutex);
465 dev->log_ctx = NULL;
466 dev->umem = NULL;
467 dev->iotlb = NULL;
468 dev->mm = NULL;
469 dev->worker = NULL;
470 dev->iov_limit = iov_limit;
471 dev->weight = weight;
472 dev->byte_weight = byte_weight;
473 init_llist_head(&dev->work_list);
474 init_waitqueue_head(&dev->wait);
475 INIT_LIST_HEAD(&dev->read_list);
476 INIT_LIST_HEAD(&dev->pending_list);
477 spin_lock_init(&dev->iotlb_lock);
478
479
480 for (i = 0; i < dev->nvqs; ++i) {
481 vq = dev->vqs[i];
482 vq->log = NULL;
483 vq->indirect = NULL;
484 vq->heads = NULL;
485 vq->dev = dev;
486 mutex_init(&vq->mutex);
487 vhost_vq_reset(dev, vq);
488 if (vq->handle_kick)
489 vhost_poll_init(&vq->poll, vq->handle_kick,
490 EPOLLIN, dev);
491 }
492}
493EXPORT_SYMBOL_GPL(vhost_dev_init);
494
495/* Caller should have device mutex */
496long vhost_dev_check_owner(struct vhost_dev *dev)
497{
498 /* Are you the owner? If not, I don't think you mean to do that */
499 return dev->mm == current->mm ? 0 : -EPERM;
500}
501EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
502
503struct vhost_attach_cgroups_struct {
504 struct vhost_work work;
505 struct task_struct *owner;
506 int ret;
507};
508
509static void vhost_attach_cgroups_work(struct vhost_work *work)
510{
511 struct vhost_attach_cgroups_struct *s;
512
513 s = container_of(work, struct vhost_attach_cgroups_struct, work);
514 s->ret = cgroup_attach_task_all(s->owner, current);
515}
516
517static int vhost_attach_cgroups(struct vhost_dev *dev)
518{
519 struct vhost_attach_cgroups_struct attach;
520
521 attach.owner = current;
522 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
523 vhost_work_queue(dev, &attach.work);
524 vhost_work_flush(dev, &attach.work);
525 return attach.ret;
526}
527
528/* Caller should have device mutex */
529bool vhost_dev_has_owner(struct vhost_dev *dev)
530{
531 return dev->mm;
532}
533EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
534
535/* Caller should have device mutex */
536long vhost_dev_set_owner(struct vhost_dev *dev)
537{
538 struct task_struct *worker;
539 int err;
540
541 /* Is there an owner already? */
542 if (vhost_dev_has_owner(dev)) {
543 err = -EBUSY;
544 goto err_mm;
545 }
546
547 /* No owner, become one */
548 dev->mm = get_task_mm(current);
549 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
550 if (IS_ERR(worker)) {
551 err = PTR_ERR(worker);
552 goto err_worker;
553 }
554
555 dev->worker = worker;
556 wake_up_process(worker); /* avoid contributing to loadavg */
557
558 err = vhost_attach_cgroups(dev);
559 if (err)
560 goto err_cgroup;
561
562 err = vhost_dev_alloc_iovecs(dev);
563 if (err)
564 goto err_cgroup;
565
566 return 0;
567err_cgroup:
568 kthread_stop(worker);
569 dev->worker = NULL;
570err_worker:
571 if (dev->mm)
572 mmput(dev->mm);
573 dev->mm = NULL;
574err_mm:
575 return err;
576}
577EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
578
579struct vhost_umem *vhost_dev_reset_owner_prepare(void)
580{
581 return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
582}
583EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
584
585/* Caller should have device mutex */
586void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
587{
588 int i;
589
590 vhost_dev_cleanup(dev);
591
592 /* Restore memory to default empty mapping. */
593 INIT_LIST_HEAD(&umem->umem_list);
594 dev->umem = umem;
595 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
596 * VQs aren't running.
597 */
598 for (i = 0; i < dev->nvqs; ++i)
599 dev->vqs[i]->umem = umem;
600}
601EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
602
603void vhost_dev_stop(struct vhost_dev *dev)
604{
605 int i;
606
607 for (i = 0; i < dev->nvqs; ++i) {
608 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
609 vhost_poll_stop(&dev->vqs[i]->poll);
610 vhost_poll_flush(&dev->vqs[i]->poll);
611 }
612 }
613}
614EXPORT_SYMBOL_GPL(vhost_dev_stop);
615
616static void vhost_umem_free(struct vhost_umem *umem,
617 struct vhost_umem_node *node)
618{
619 vhost_umem_interval_tree_remove(node, &umem->umem_tree);
620 list_del(&node->link);
621 kfree(node);
622 umem->numem--;
623}
624
625static void vhost_umem_clean(struct vhost_umem *umem)
626{
627 struct vhost_umem_node *node, *tmp;
628
629 if (!umem)
630 return;
631
632 list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
633 vhost_umem_free(umem, node);
634
635 kvfree(umem);
636}
637
638static void vhost_clear_msg(struct vhost_dev *dev)
639{
640 struct vhost_msg_node *node, *n;
641
642 spin_lock(&dev->iotlb_lock);
643
644 list_for_each_entry_safe(node, n, &dev->read_list, node) {
645 list_del(&node->node);
646 kfree(node);
647 }
648
649 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
650 list_del(&node->node);
651 kfree(node);
652 }
653
654 spin_unlock(&dev->iotlb_lock);
655}
656
657void vhost_dev_cleanup(struct vhost_dev *dev)
658{
659 int i;
660
661 for (i = 0; i < dev->nvqs; ++i) {
662 if (dev->vqs[i]->error_ctx)
663 eventfd_ctx_put(dev->vqs[i]->error_ctx);
664 if (dev->vqs[i]->kick)
665 fput(dev->vqs[i]->kick);
666 if (dev->vqs[i]->call_ctx)
667 eventfd_ctx_put(dev->vqs[i]->call_ctx);
668 vhost_vq_reset(dev, dev->vqs[i]);
669 }
670 vhost_dev_free_iovecs(dev);
671 if (dev->log_ctx)
672 eventfd_ctx_put(dev->log_ctx);
673 dev->log_ctx = NULL;
674 /* No one will access memory at this point */
675 vhost_umem_clean(dev->umem);
676 dev->umem = NULL;
677 vhost_umem_clean(dev->iotlb);
678 dev->iotlb = NULL;
679 vhost_clear_msg(dev);
680 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
681 WARN_ON(!llist_empty(&dev->work_list));
682 if (dev->worker) {
683 kthread_stop(dev->worker);
684 dev->worker = NULL;
685 }
686 if (dev->mm)
687 mmput(dev->mm);
688 dev->mm = NULL;
689}
690EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
691
692static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
693{
694 u64 a = addr / VHOST_PAGE_SIZE / 8;
695
696 /* Make sure 64 bit math will not overflow. */
697 if (a > ULONG_MAX - (unsigned long)log_base ||
698 a + (unsigned long)log_base > ULONG_MAX)
699 return false;
700
701 return access_ok(log_base + a,
702 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
703}
704
705static bool vhost_overflow(u64 uaddr, u64 size)
706{
707 /* Make sure 64 bit math will not overflow. */
708 return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
709}
710
711/* Caller should have vq mutex and device mutex. */
712static bool vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
713 int log_all)
714{
715 struct vhost_umem_node *node;
716
717 if (!umem)
718 return false;
719
720 list_for_each_entry(node, &umem->umem_list, link) {
721 unsigned long a = node->userspace_addr;
722
723 if (vhost_overflow(node->userspace_addr, node->size))
724 return false;
725
726
727 if (!access_ok((void __user *)a,
728 node->size))
729 return false;
730 else if (log_all && !log_access_ok(log_base,
731 node->start,
732 node->size))
733 return false;
734 }
735 return true;
736}
737
738static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
739 u64 addr, unsigned int size,
740 int type)
741{
742 const struct vhost_umem_node *node = vq->meta_iotlb[type];
743
744 if (!node)
745 return NULL;
746
747 return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
748}
749
750/* Can we switch to this memory table? */
751/* Caller should have device mutex but not vq mutex */
752static bool memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
753 int log_all)
754{
755 int i;
756
757 for (i = 0; i < d->nvqs; ++i) {
758 bool ok;
759 bool log;
760
761 mutex_lock(&d->vqs[i]->mutex);
762 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
763 /* If ring is inactive, will check when it's enabled. */
764 if (d->vqs[i]->private_data)
765 ok = vq_memory_access_ok(d->vqs[i]->log_base,
766 umem, log);
767 else
768 ok = true;
769 mutex_unlock(&d->vqs[i]->mutex);
770 if (!ok)
771 return false;
772 }
773 return true;
774}
775
776static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
777 struct iovec iov[], int iov_size, int access);
778
779static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
780 const void *from, unsigned size)
781{
782 int ret;
783
784 if (!vq->iotlb)
785 return __copy_to_user(to, from, size);
786 else {
787 /* This function should be called after iotlb
788 * prefetch, which means we're sure that all vq
789 * could be access through iotlb. So -EAGAIN should
790 * not happen in this case.
791 */
792 struct iov_iter t;
793 void __user *uaddr = vhost_vq_meta_fetch(vq,
794 (u64)(uintptr_t)to, size,
795 VHOST_ADDR_USED);
796
797 if (uaddr)
798 return __copy_to_user(uaddr, from, size);
799
800 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
801 ARRAY_SIZE(vq->iotlb_iov),
802 VHOST_ACCESS_WO);
803 if (ret < 0)
804 goto out;
805 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
806 ret = copy_to_iter(from, size, &t);
807 if (ret == size)
808 ret = 0;
809 }
810out:
811 return ret;
812}
813
814static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
815 void __user *from, unsigned size)
816{
817 int ret;
818
819 if (!vq->iotlb)
820 return __copy_from_user(to, from, size);
821 else {
822 /* This function should be called after iotlb
823 * prefetch, which means we're sure that vq
824 * could be access through iotlb. So -EAGAIN should
825 * not happen in this case.
826 */
827 void __user *uaddr = vhost_vq_meta_fetch(vq,
828 (u64)(uintptr_t)from, size,
829 VHOST_ADDR_DESC);
830 struct iov_iter f;
831
832 if (uaddr)
833 return __copy_from_user(to, uaddr, size);
834
835 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
836 ARRAY_SIZE(vq->iotlb_iov),
837 VHOST_ACCESS_RO);
838 if (ret < 0) {
839 vq_err(vq, "IOTLB translation failure: uaddr "
840 "%p size 0x%llx\n", from,
841 (unsigned long long) size);
842 goto out;
843 }
844 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
845 ret = copy_from_iter(to, size, &f);
846 if (ret == size)
847 ret = 0;
848 }
849
850out:
851 return ret;
852}
853
854static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
855 void __user *addr, unsigned int size,
856 int type)
857{
858 int ret;
859
860 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
861 ARRAY_SIZE(vq->iotlb_iov),
862 VHOST_ACCESS_RO);
863 if (ret < 0) {
864 vq_err(vq, "IOTLB translation failure: uaddr "
865 "%p size 0x%llx\n", addr,
866 (unsigned long long) size);
867 return NULL;
868 }
869
870 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
871 vq_err(vq, "Non atomic userspace memory access: uaddr "
872 "%p size 0x%llx\n", addr,
873 (unsigned long long) size);
874 return NULL;
875 }
876
877 return vq->iotlb_iov[0].iov_base;
878}
879
880/* This function should be called after iotlb
881 * prefetch, which means we're sure that vq
882 * could be access through iotlb. So -EAGAIN should
883 * not happen in this case.
884 */
885static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
886 void *addr, unsigned int size,
887 int type)
888{
889 void __user *uaddr = vhost_vq_meta_fetch(vq,
890 (u64)(uintptr_t)addr, size, type);
891 if (uaddr)
892 return uaddr;
893
894 return __vhost_get_user_slow(vq, addr, size, type);
895}
896
897#define vhost_put_user(vq, x, ptr) \
898({ \
899 int ret = -EFAULT; \
900 if (!vq->iotlb) { \
901 ret = __put_user(x, ptr); \
902 } else { \
903 __typeof__(ptr) to = \
904 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
905 sizeof(*ptr), VHOST_ADDR_USED); \
906 if (to != NULL) \
907 ret = __put_user(x, to); \
908 else \
909 ret = -EFAULT; \
910 } \
911 ret; \
912})
913
914static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
915{
916 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
917 vhost_avail_event(vq));
918}
919
920static inline int vhost_put_used(struct vhost_virtqueue *vq,
921 struct vring_used_elem *head, int idx,
922 int count)
923{
924 return vhost_copy_to_user(vq, vq->used->ring + idx, head,
925 count * sizeof(*head));
926}
927
928static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
929
930{
931 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
932 &vq->used->flags);
933}
934
935static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
936
937{
938 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
939 &vq->used->idx);
940}
941
942#define vhost_get_user(vq, x, ptr, type) \
943({ \
944 int ret; \
945 if (!vq->iotlb) { \
946 ret = __get_user(x, ptr); \
947 } else { \
948 __typeof__(ptr) from = \
949 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
950 sizeof(*ptr), \
951 type); \
952 if (from != NULL) \
953 ret = __get_user(x, from); \
954 else \
955 ret = -EFAULT; \
956 } \
957 ret; \
958})
959
960#define vhost_get_avail(vq, x, ptr) \
961 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
962
963#define vhost_get_used(vq, x, ptr) \
964 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
965
966static void vhost_dev_lock_vqs(struct vhost_dev *d)
967{
968 int i = 0;
969 for (i = 0; i < d->nvqs; ++i)
970 mutex_lock_nested(&d->vqs[i]->mutex, i);
971}
972
973static void vhost_dev_unlock_vqs(struct vhost_dev *d)
974{
975 int i = 0;
976 for (i = 0; i < d->nvqs; ++i)
977 mutex_unlock(&d->vqs[i]->mutex);
978}
979
980static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
981 __virtio16 *idx)
982{
983 return vhost_get_avail(vq, *idx, &vq->avail->idx);
984}
985
986static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
987 __virtio16 *head, int idx)
988{
989 return vhost_get_avail(vq, *head,
990 &vq->avail->ring[idx & (vq->num - 1)]);
991}
992
993static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
994 __virtio16 *flags)
995{
996 return vhost_get_avail(vq, *flags, &vq->avail->flags);
997}
998
999static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1000 __virtio16 *event)
1001{
1002 return vhost_get_avail(vq, *event, vhost_used_event(vq));
1003}
1004
1005static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1006 __virtio16 *idx)
1007{
1008 return vhost_get_used(vq, *idx, &vq->used->idx);
1009}
1010
1011static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1012 struct vring_desc *desc, int idx)
1013{
1014 return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1015}
1016
1017static int vhost_new_umem_range(struct vhost_umem *umem,
1018 u64 start, u64 size, u64 end,
1019 u64 userspace_addr, int perm)
1020{
1021 struct vhost_umem_node *tmp, *node;
1022
1023 if (!size)
1024 return -EFAULT;
1025
1026 node = kmalloc(sizeof(*node), GFP_ATOMIC);
1027 if (!node)
1028 return -ENOMEM;
1029
1030 if (umem->numem == max_iotlb_entries) {
1031 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
1032 vhost_umem_free(umem, tmp);
1033 }
1034
1035 node->start = start;
1036 node->size = size;
1037 node->last = end;
1038 node->userspace_addr = userspace_addr;
1039 node->perm = perm;
1040 INIT_LIST_HEAD(&node->link);
1041 list_add_tail(&node->link, &umem->umem_list);
1042 vhost_umem_interval_tree_insert(node, &umem->umem_tree);
1043 umem->numem++;
1044
1045 return 0;
1046}
1047
1048static void vhost_del_umem_range(struct vhost_umem *umem,
1049 u64 start, u64 end)
1050{
1051 struct vhost_umem_node *node;
1052
1053 while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1054 start, end)))
1055 vhost_umem_free(umem, node);
1056}
1057
1058static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1059 struct vhost_iotlb_msg *msg)
1060{
1061 struct vhost_msg_node *node, *n;
1062
1063 spin_lock(&d->iotlb_lock);
1064
1065 list_for_each_entry_safe(node, n, &d->pending_list, node) {
1066 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1067 if (msg->iova <= vq_msg->iova &&
1068 msg->iova + msg->size - 1 >= vq_msg->iova &&
1069 vq_msg->type == VHOST_IOTLB_MISS) {
1070 vhost_poll_queue(&node->vq->poll);
1071 list_del(&node->node);
1072 kfree(node);
1073 }
1074 }
1075
1076 spin_unlock(&d->iotlb_lock);
1077}
1078
1079static bool umem_access_ok(u64 uaddr, u64 size, int access)
1080{
1081 unsigned long a = uaddr;
1082
1083 /* Make sure 64 bit math will not overflow. */
1084 if (vhost_overflow(uaddr, size))
1085 return false;
1086
1087 if ((access & VHOST_ACCESS_RO) &&
1088 !access_ok((void __user *)a, size))
1089 return false;
1090 if ((access & VHOST_ACCESS_WO) &&
1091 !access_ok((void __user *)a, size))
1092 return false;
1093 return true;
1094}
1095
1096static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1097 struct vhost_iotlb_msg *msg)
1098{
1099 int ret = 0;
1100
1101 mutex_lock(&dev->mutex);
1102 vhost_dev_lock_vqs(dev);
1103 switch (msg->type) {
1104 case VHOST_IOTLB_UPDATE:
1105 if (!dev->iotlb) {
1106 ret = -EFAULT;
1107 break;
1108 }
1109 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1110 ret = -EFAULT;
1111 break;
1112 }
1113 vhost_vq_meta_reset(dev);
1114 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1115 msg->iova + msg->size - 1,
1116 msg->uaddr, msg->perm)) {
1117 ret = -ENOMEM;
1118 break;
1119 }
1120 vhost_iotlb_notify_vq(dev, msg);
1121 break;
1122 case VHOST_IOTLB_INVALIDATE:
1123 if (!dev->iotlb) {
1124 ret = -EFAULT;
1125 break;
1126 }
1127 vhost_vq_meta_reset(dev);
1128 vhost_del_umem_range(dev->iotlb, msg->iova,
1129 msg->iova + msg->size - 1);
1130 break;
1131 default:
1132 ret = -EINVAL;
1133 break;
1134 }
1135
1136 vhost_dev_unlock_vqs(dev);
1137 mutex_unlock(&dev->mutex);
1138
1139 return ret;
1140}
1141ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1142 struct iov_iter *from)
1143{
1144 struct vhost_iotlb_msg msg;
1145 size_t offset;
1146 int type, ret;
1147
1148 ret = copy_from_iter(&type, sizeof(type), from);
1149 if (ret != sizeof(type)) {
1150 ret = -EINVAL;
1151 goto done;
1152 }
1153
1154 switch (type) {
1155 case VHOST_IOTLB_MSG:
1156 /* There maybe a hole after type for V1 message type,
1157 * so skip it here.
1158 */
1159 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1160 break;
1161 case VHOST_IOTLB_MSG_V2:
1162 offset = sizeof(__u32);
1163 break;
1164 default:
1165 ret = -EINVAL;
1166 goto done;
1167 }
1168
1169 iov_iter_advance(from, offset);
1170 ret = copy_from_iter(&msg, sizeof(msg), from);
1171 if (ret != sizeof(msg)) {
1172 ret = -EINVAL;
1173 goto done;
1174 }
1175 if (vhost_process_iotlb_msg(dev, &msg)) {
1176 ret = -EFAULT;
1177 goto done;
1178 }
1179
1180 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1181 sizeof(struct vhost_msg_v2);
1182done:
1183 return ret;
1184}
1185EXPORT_SYMBOL(vhost_chr_write_iter);
1186
1187__poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1188 poll_table *wait)
1189{
1190 __poll_t mask = 0;
1191
1192 poll_wait(file, &dev->wait, wait);
1193
1194 if (!list_empty(&dev->read_list))
1195 mask |= EPOLLIN | EPOLLRDNORM;
1196
1197 return mask;
1198}
1199EXPORT_SYMBOL(vhost_chr_poll);
1200
1201ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1202 int noblock)
1203{
1204 DEFINE_WAIT(wait);
1205 struct vhost_msg_node *node;
1206 ssize_t ret = 0;
1207 unsigned size = sizeof(struct vhost_msg);
1208
1209 if (iov_iter_count(to) < size)
1210 return 0;
1211
1212 while (1) {
1213 if (!noblock)
1214 prepare_to_wait(&dev->wait, &wait,
1215 TASK_INTERRUPTIBLE);
1216
1217 node = vhost_dequeue_msg(dev, &dev->read_list);
1218 if (node)
1219 break;
1220 if (noblock) {
1221 ret = -EAGAIN;
1222 break;
1223 }
1224 if (signal_pending(current)) {
1225 ret = -ERESTARTSYS;
1226 break;
1227 }
1228 if (!dev->iotlb) {
1229 ret = -EBADFD;
1230 break;
1231 }
1232
1233 schedule();
1234 }
1235
1236 if (!noblock)
1237 finish_wait(&dev->wait, &wait);
1238
1239 if (node) {
1240 struct vhost_iotlb_msg *msg;
1241 void *start = &node->msg;
1242
1243 switch (node->msg.type) {
1244 case VHOST_IOTLB_MSG:
1245 size = sizeof(node->msg);
1246 msg = &node->msg.iotlb;
1247 break;
1248 case VHOST_IOTLB_MSG_V2:
1249 size = sizeof(node->msg_v2);
1250 msg = &node->msg_v2.iotlb;
1251 break;
1252 default:
1253 BUG();
1254 break;
1255 }
1256
1257 ret = copy_to_iter(start, size, to);
1258 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1259 kfree(node);
1260 return ret;
1261 }
1262 vhost_enqueue_msg(dev, &dev->pending_list, node);
1263 }
1264
1265 return ret;
1266}
1267EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1268
1269static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1270{
1271 struct vhost_dev *dev = vq->dev;
1272 struct vhost_msg_node *node;
1273 struct vhost_iotlb_msg *msg;
1274 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1275
1276 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1277 if (!node)
1278 return -ENOMEM;
1279
1280 if (v2) {
1281 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1282 msg = &node->msg_v2.iotlb;
1283 } else {
1284 msg = &node->msg.iotlb;
1285 }
1286
1287 msg->type = VHOST_IOTLB_MISS;
1288 msg->iova = iova;
1289 msg->perm = access;
1290
1291 vhost_enqueue_msg(dev, &dev->read_list, node);
1292
1293 return 0;
1294}
1295
1296static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1297 struct vring_desc __user *desc,
1298 struct vring_avail __user *avail,
1299 struct vring_used __user *used)
1300
1301{
1302 return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1303 access_ok(avail, vhost_get_avail_size(vq, num)) &&
1304 access_ok(used, vhost_get_used_size(vq, num));
1305}
1306
1307static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1308 const struct vhost_umem_node *node,
1309 int type)
1310{
1311 int access = (type == VHOST_ADDR_USED) ?
1312 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1313
1314 if (likely(node->perm & access))
1315 vq->meta_iotlb[type] = node;
1316}
1317
1318static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1319 int access, u64 addr, u64 len, int type)
1320{
1321 const struct vhost_umem_node *node;
1322 struct vhost_umem *umem = vq->iotlb;
1323 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1324
1325 if (vhost_vq_meta_fetch(vq, addr, len, type))
1326 return true;
1327
1328 while (len > s) {
1329 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1330 addr,
1331 last);
1332 if (node == NULL || node->start > addr) {
1333 vhost_iotlb_miss(vq, addr, access);
1334 return false;
1335 } else if (!(node->perm & access)) {
1336 /* Report the possible access violation by
1337 * request another translation from userspace.
1338 */
1339 return false;
1340 }
1341
1342 size = node->size - addr + node->start;
1343
1344 if (orig_addr == addr && size >= len)
1345 vhost_vq_meta_update(vq, node, type);
1346
1347 s += size;
1348 addr += size;
1349 }
1350
1351 return true;
1352}
1353
1354int vq_meta_prefetch(struct vhost_virtqueue *vq)
1355{
1356 unsigned int num = vq->num;
1357
1358 if (!vq->iotlb)
1359 return 1;
1360
1361 return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1362 vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1363 iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1364 vhost_get_avail_size(vq, num),
1365 VHOST_ADDR_AVAIL) &&
1366 iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1367 vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1368}
1369EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1370
1371/* Can we log writes? */
1372/* Caller should have device mutex but not vq mutex */
1373bool vhost_log_access_ok(struct vhost_dev *dev)
1374{
1375 return memory_access_ok(dev, dev->umem, 1);
1376}
1377EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1378
1379/* Verify access for write logging. */
1380/* Caller should have vq mutex and device mutex */
1381static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1382 void __user *log_base)
1383{
1384 return vq_memory_access_ok(log_base, vq->umem,
1385 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1386 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1387 vhost_get_used_size(vq, vq->num)));
1388}
1389
1390/* Can we start vq? */
1391/* Caller should have vq mutex and device mutex */
1392bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1393{
1394 if (!vq_log_access_ok(vq, vq->log_base))
1395 return false;
1396
1397 /* Access validation occurs at prefetch time with IOTLB */
1398 if (vq->iotlb)
1399 return true;
1400
1401 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1402}
1403EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1404
1405static struct vhost_umem *vhost_umem_alloc(void)
1406{
1407 struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1408
1409 if (!umem)
1410 return NULL;
1411
1412 umem->umem_tree = RB_ROOT_CACHED;
1413 umem->numem = 0;
1414 INIT_LIST_HEAD(&umem->umem_list);
1415
1416 return umem;
1417}
1418
1419static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1420{
1421 struct vhost_memory mem, *newmem;
1422 struct vhost_memory_region *region;
1423 struct vhost_umem *newumem, *oldumem;
1424 unsigned long size = offsetof(struct vhost_memory, regions);
1425 int i;
1426
1427 if (copy_from_user(&mem, m, size))
1428 return -EFAULT;
1429 if (mem.padding)
1430 return -EOPNOTSUPP;
1431 if (mem.nregions > max_mem_regions)
1432 return -E2BIG;
1433 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1434 GFP_KERNEL);
1435 if (!newmem)
1436 return -ENOMEM;
1437
1438 memcpy(newmem, &mem, size);
1439 if (copy_from_user(newmem->regions, m->regions,
1440 mem.nregions * sizeof *m->regions)) {
1441 kvfree(newmem);
1442 return -EFAULT;
1443 }
1444
1445 newumem = vhost_umem_alloc();
1446 if (!newumem) {
1447 kvfree(newmem);
1448 return -ENOMEM;
1449 }
1450
1451 for (region = newmem->regions;
1452 region < newmem->regions + mem.nregions;
1453 region++) {
1454 if (vhost_new_umem_range(newumem,
1455 region->guest_phys_addr,
1456 region->memory_size,
1457 region->guest_phys_addr +
1458 region->memory_size - 1,
1459 region->userspace_addr,
1460 VHOST_ACCESS_RW))
1461 goto err;
1462 }
1463
1464 if (!memory_access_ok(d, newumem, 0))
1465 goto err;
1466
1467 oldumem = d->umem;
1468 d->umem = newumem;
1469
1470 /* All memory accesses are done under some VQ mutex. */
1471 for (i = 0; i < d->nvqs; ++i) {
1472 mutex_lock(&d->vqs[i]->mutex);
1473 d->vqs[i]->umem = newumem;
1474 mutex_unlock(&d->vqs[i]->mutex);
1475 }
1476
1477 kvfree(newmem);
1478 vhost_umem_clean(oldumem);
1479 return 0;
1480
1481err:
1482 vhost_umem_clean(newumem);
1483 kvfree(newmem);
1484 return -EFAULT;
1485}
1486
1487static long vhost_vring_set_num(struct vhost_dev *d,
1488 struct vhost_virtqueue *vq,
1489 void __user *argp)
1490{
1491 struct vhost_vring_state s;
1492
1493 /* Resizing ring with an active backend?
1494 * You don't want to do that. */
1495 if (vq->private_data)
1496 return -EBUSY;
1497
1498 if (copy_from_user(&s, argp, sizeof s))
1499 return -EFAULT;
1500
1501 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1502 return -EINVAL;
1503 vq->num = s.num;
1504
1505 return 0;
1506}
1507
1508static long vhost_vring_set_addr(struct vhost_dev *d,
1509 struct vhost_virtqueue *vq,
1510 void __user *argp)
1511{
1512 struct vhost_vring_addr a;
1513
1514 if (copy_from_user(&a, argp, sizeof a))
1515 return -EFAULT;
1516 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1517 return -EOPNOTSUPP;
1518
1519 /* For 32bit, verify that the top 32bits of the user
1520 data are set to zero. */
1521 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1522 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1523 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1524 return -EFAULT;
1525
1526 /* Make sure it's safe to cast pointers to vring types. */
1527 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1528 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1529 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1530 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1531 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1532 return -EINVAL;
1533
1534 /* We only verify access here if backend is configured.
1535 * If it is not, we don't as size might not have been setup.
1536 * We will verify when backend is configured. */
1537 if (vq->private_data) {
1538 if (!vq_access_ok(vq, vq->num,
1539 (void __user *)(unsigned long)a.desc_user_addr,
1540 (void __user *)(unsigned long)a.avail_user_addr,
1541 (void __user *)(unsigned long)a.used_user_addr))
1542 return -EINVAL;
1543
1544 /* Also validate log access for used ring if enabled. */
1545 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1546 !log_access_ok(vq->log_base, a.log_guest_addr,
1547 sizeof *vq->used +
1548 vq->num * sizeof *vq->used->ring))
1549 return -EINVAL;
1550 }
1551
1552 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1553 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1554 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1555 vq->log_addr = a.log_guest_addr;
1556 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1557
1558 return 0;
1559}
1560
1561static long vhost_vring_set_num_addr(struct vhost_dev *d,
1562 struct vhost_virtqueue *vq,
1563 unsigned int ioctl,
1564 void __user *argp)
1565{
1566 long r;
1567
1568 mutex_lock(&vq->mutex);
1569
1570 switch (ioctl) {
1571 case VHOST_SET_VRING_NUM:
1572 r = vhost_vring_set_num(d, vq, argp);
1573 break;
1574 case VHOST_SET_VRING_ADDR:
1575 r = vhost_vring_set_addr(d, vq, argp);
1576 break;
1577 default:
1578 BUG();
1579 }
1580
1581 mutex_unlock(&vq->mutex);
1582
1583 return r;
1584}
1585long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1586{
1587 struct file *eventfp, *filep = NULL;
1588 bool pollstart = false, pollstop = false;
1589 struct eventfd_ctx *ctx = NULL;
1590 u32 __user *idxp = argp;
1591 struct vhost_virtqueue *vq;
1592 struct vhost_vring_state s;
1593 struct vhost_vring_file f;
1594 u32 idx;
1595 long r;
1596
1597 r = get_user(idx, idxp);
1598 if (r < 0)
1599 return r;
1600 if (idx >= d->nvqs)
1601 return -ENOBUFS;
1602
1603 idx = array_index_nospec(idx, d->nvqs);
1604 vq = d->vqs[idx];
1605
1606 if (ioctl == VHOST_SET_VRING_NUM ||
1607 ioctl == VHOST_SET_VRING_ADDR) {
1608 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1609 }
1610
1611 mutex_lock(&vq->mutex);
1612
1613 switch (ioctl) {
1614 case VHOST_SET_VRING_BASE:
1615 /* Moving base with an active backend?
1616 * You don't want to do that. */
1617 if (vq->private_data) {
1618 r = -EBUSY;
1619 break;
1620 }
1621 if (copy_from_user(&s, argp, sizeof s)) {
1622 r = -EFAULT;
1623 break;
1624 }
1625 if (s.num > 0xffff) {
1626 r = -EINVAL;
1627 break;
1628 }
1629 vq->last_avail_idx = s.num;
1630 /* Forget the cached index value. */
1631 vq->avail_idx = vq->last_avail_idx;
1632 break;
1633 case VHOST_GET_VRING_BASE:
1634 s.index = idx;
1635 s.num = vq->last_avail_idx;
1636 if (copy_to_user(argp, &s, sizeof s))
1637 r = -EFAULT;
1638 break;
1639 case VHOST_SET_VRING_KICK:
1640 if (copy_from_user(&f, argp, sizeof f)) {
1641 r = -EFAULT;
1642 break;
1643 }
1644 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1645 if (IS_ERR(eventfp)) {
1646 r = PTR_ERR(eventfp);
1647 break;
1648 }
1649 if (eventfp != vq->kick) {
1650 pollstop = (filep = vq->kick) != NULL;
1651 pollstart = (vq->kick = eventfp) != NULL;
1652 } else
1653 filep = eventfp;
1654 break;
1655 case VHOST_SET_VRING_CALL:
1656 if (copy_from_user(&f, argp, sizeof f)) {
1657 r = -EFAULT;
1658 break;
1659 }
1660 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1661 if (IS_ERR(ctx)) {
1662 r = PTR_ERR(ctx);
1663 break;
1664 }
1665 swap(ctx, vq->call_ctx);
1666 break;
1667 case VHOST_SET_VRING_ERR:
1668 if (copy_from_user(&f, argp, sizeof f)) {
1669 r = -EFAULT;
1670 break;
1671 }
1672 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1673 if (IS_ERR(ctx)) {
1674 r = PTR_ERR(ctx);
1675 break;
1676 }
1677 swap(ctx, vq->error_ctx);
1678 break;
1679 case VHOST_SET_VRING_ENDIAN:
1680 r = vhost_set_vring_endian(vq, argp);
1681 break;
1682 case VHOST_GET_VRING_ENDIAN:
1683 r = vhost_get_vring_endian(vq, idx, argp);
1684 break;
1685 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1686 if (copy_from_user(&s, argp, sizeof(s))) {
1687 r = -EFAULT;
1688 break;
1689 }
1690 vq->busyloop_timeout = s.num;
1691 break;
1692 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1693 s.index = idx;
1694 s.num = vq->busyloop_timeout;
1695 if (copy_to_user(argp, &s, sizeof(s)))
1696 r = -EFAULT;
1697 break;
1698 default:
1699 r = -ENOIOCTLCMD;
1700 }
1701
1702 if (pollstop && vq->handle_kick)
1703 vhost_poll_stop(&vq->poll);
1704
1705 if (!IS_ERR_OR_NULL(ctx))
1706 eventfd_ctx_put(ctx);
1707 if (filep)
1708 fput(filep);
1709
1710 if (pollstart && vq->handle_kick)
1711 r = vhost_poll_start(&vq->poll, vq->kick);
1712
1713 mutex_unlock(&vq->mutex);
1714
1715 if (pollstop && vq->handle_kick)
1716 vhost_poll_flush(&vq->poll);
1717 return r;
1718}
1719EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1720
1721int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1722{
1723 struct vhost_umem *niotlb, *oiotlb;
1724 int i;
1725
1726 niotlb = vhost_umem_alloc();
1727 if (!niotlb)
1728 return -ENOMEM;
1729
1730 oiotlb = d->iotlb;
1731 d->iotlb = niotlb;
1732
1733 for (i = 0; i < d->nvqs; ++i) {
1734 struct vhost_virtqueue *vq = d->vqs[i];
1735
1736 mutex_lock(&vq->mutex);
1737 vq->iotlb = niotlb;
1738 __vhost_vq_meta_reset(vq);
1739 mutex_unlock(&vq->mutex);
1740 }
1741
1742 vhost_umem_clean(oiotlb);
1743
1744 return 0;
1745}
1746EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1747
1748/* Caller must have device mutex */
1749long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1750{
1751 struct eventfd_ctx *ctx;
1752 u64 p;
1753 long r;
1754 int i, fd;
1755
1756 /* If you are not the owner, you can become one */
1757 if (ioctl == VHOST_SET_OWNER) {
1758 r = vhost_dev_set_owner(d);
1759 goto done;
1760 }
1761
1762 /* You must be the owner to do anything else */
1763 r = vhost_dev_check_owner(d);
1764 if (r)
1765 goto done;
1766
1767 switch (ioctl) {
1768 case VHOST_SET_MEM_TABLE:
1769 r = vhost_set_memory(d, argp);
1770 break;
1771 case VHOST_SET_LOG_BASE:
1772 if (copy_from_user(&p, argp, sizeof p)) {
1773 r = -EFAULT;
1774 break;
1775 }
1776 if ((u64)(unsigned long)p != p) {
1777 r = -EFAULT;
1778 break;
1779 }
1780 for (i = 0; i < d->nvqs; ++i) {
1781 struct vhost_virtqueue *vq;
1782 void __user *base = (void __user *)(unsigned long)p;
1783 vq = d->vqs[i];
1784 mutex_lock(&vq->mutex);
1785 /* If ring is inactive, will check when it's enabled. */
1786 if (vq->private_data && !vq_log_access_ok(vq, base))
1787 r = -EFAULT;
1788 else
1789 vq->log_base = base;
1790 mutex_unlock(&vq->mutex);
1791 }
1792 break;
1793 case VHOST_SET_LOG_FD:
1794 r = get_user(fd, (int __user *)argp);
1795 if (r < 0)
1796 break;
1797 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1798 if (IS_ERR(ctx)) {
1799 r = PTR_ERR(ctx);
1800 break;
1801 }
1802 swap(ctx, d->log_ctx);
1803 for (i = 0; i < d->nvqs; ++i) {
1804 mutex_lock(&d->vqs[i]->mutex);
1805 d->vqs[i]->log_ctx = d->log_ctx;
1806 mutex_unlock(&d->vqs[i]->mutex);
1807 }
1808 if (ctx)
1809 eventfd_ctx_put(ctx);
1810 break;
1811 default:
1812 r = -ENOIOCTLCMD;
1813 break;
1814 }
1815done:
1816 return r;
1817}
1818EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1819
1820/* TODO: This is really inefficient. We need something like get_user()
1821 * (instruction directly accesses the data, with an exception table entry
1822 * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1823 */
1824static int set_bit_to_user(int nr, void __user *addr)
1825{
1826 unsigned long log = (unsigned long)addr;
1827 struct page *page;
1828 void *base;
1829 int bit = nr + (log % PAGE_SIZE) * 8;
1830 int r;
1831
1832 r = get_user_pages_fast(log, 1, FOLL_WRITE, &page);
1833 if (r < 0)
1834 return r;
1835 BUG_ON(r != 1);
1836 base = kmap_atomic(page);
1837 set_bit(bit, base);
1838 kunmap_atomic(base);
1839 set_page_dirty_lock(page);
1840 put_page(page);
1841 return 0;
1842}
1843
1844static int log_write(void __user *log_base,
1845 u64 write_address, u64 write_length)
1846{
1847 u64 write_page = write_address / VHOST_PAGE_SIZE;
1848 int r;
1849
1850 if (!write_length)
1851 return 0;
1852 write_length += write_address % VHOST_PAGE_SIZE;
1853 for (;;) {
1854 u64 base = (u64)(unsigned long)log_base;
1855 u64 log = base + write_page / 8;
1856 int bit = write_page % 8;
1857 if ((u64)(unsigned long)log != log)
1858 return -EFAULT;
1859 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1860 if (r < 0)
1861 return r;
1862 if (write_length <= VHOST_PAGE_SIZE)
1863 break;
1864 write_length -= VHOST_PAGE_SIZE;
1865 write_page += 1;
1866 }
1867 return r;
1868}
1869
1870static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1871{
1872 struct vhost_umem *umem = vq->umem;
1873 struct vhost_umem_node *u;
1874 u64 start, end, l, min;
1875 int r;
1876 bool hit = false;
1877
1878 while (len) {
1879 min = len;
1880 /* More than one GPAs can be mapped into a single HVA. So
1881 * iterate all possible umems here to be safe.
1882 */
1883 list_for_each_entry(u, &umem->umem_list, link) {
1884 if (u->userspace_addr > hva - 1 + len ||
1885 u->userspace_addr - 1 + u->size < hva)
1886 continue;
1887 start = max(u->userspace_addr, hva);
1888 end = min(u->userspace_addr - 1 + u->size,
1889 hva - 1 + len);
1890 l = end - start + 1;
1891 r = log_write(vq->log_base,
1892 u->start + start - u->userspace_addr,
1893 l);
1894 if (r < 0)
1895 return r;
1896 hit = true;
1897 min = min(l, min);
1898 }
1899
1900 if (!hit)
1901 return -EFAULT;
1902
1903 len -= min;
1904 hva += min;
1905 }
1906
1907 return 0;
1908}
1909
1910static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1911{
1912 struct iovec iov[64];
1913 int i, ret;
1914
1915 if (!vq->iotlb)
1916 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1917
1918 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1919 len, iov, 64, VHOST_ACCESS_WO);
1920 if (ret < 0)
1921 return ret;
1922
1923 for (i = 0; i < ret; i++) {
1924 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1925 iov[i].iov_len);
1926 if (ret)
1927 return ret;
1928 }
1929
1930 return 0;
1931}
1932
1933int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1934 unsigned int log_num, u64 len, struct iovec *iov, int count)
1935{
1936 int i, r;
1937
1938 /* Make sure data written is seen before log. */
1939 smp_wmb();
1940
1941 if (vq->iotlb) {
1942 for (i = 0; i < count; i++) {
1943 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1944 iov[i].iov_len);
1945 if (r < 0)
1946 return r;
1947 }
1948 return 0;
1949 }
1950
1951 for (i = 0; i < log_num; ++i) {
1952 u64 l = min(log[i].len, len);
1953 r = log_write(vq->log_base, log[i].addr, l);
1954 if (r < 0)
1955 return r;
1956 len -= l;
1957 if (!len) {
1958 if (vq->log_ctx)
1959 eventfd_signal(vq->log_ctx, 1);
1960 return 0;
1961 }
1962 }
1963 /* Length written exceeds what we have stored. This is a bug. */
1964 BUG();
1965 return 0;
1966}
1967EXPORT_SYMBOL_GPL(vhost_log_write);
1968
1969static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1970{
1971 void __user *used;
1972 if (vhost_put_used_flags(vq))
1973 return -EFAULT;
1974 if (unlikely(vq->log_used)) {
1975 /* Make sure the flag is seen before log. */
1976 smp_wmb();
1977 /* Log used flag write. */
1978 used = &vq->used->flags;
1979 log_used(vq, (used - (void __user *)vq->used),
1980 sizeof vq->used->flags);
1981 if (vq->log_ctx)
1982 eventfd_signal(vq->log_ctx, 1);
1983 }
1984 return 0;
1985}
1986
1987static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1988{
1989 if (vhost_put_avail_event(vq))
1990 return -EFAULT;
1991 if (unlikely(vq->log_used)) {
1992 void __user *used;
1993 /* Make sure the event is seen before log. */
1994 smp_wmb();
1995 /* Log avail event write */
1996 used = vhost_avail_event(vq);
1997 log_used(vq, (used - (void __user *)vq->used),
1998 sizeof *vhost_avail_event(vq));
1999 if (vq->log_ctx)
2000 eventfd_signal(vq->log_ctx, 1);
2001 }
2002 return 0;
2003}
2004
2005int vhost_vq_init_access(struct vhost_virtqueue *vq)
2006{
2007 __virtio16 last_used_idx;
2008 int r;
2009 bool is_le = vq->is_le;
2010
2011 if (!vq->private_data)
2012 return 0;
2013
2014 vhost_init_is_le(vq);
2015
2016 r = vhost_update_used_flags(vq);
2017 if (r)
2018 goto err;
2019 vq->signalled_used_valid = false;
2020 if (!vq->iotlb &&
2021 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2022 r = -EFAULT;
2023 goto err;
2024 }
2025 r = vhost_get_used_idx(vq, &last_used_idx);
2026 if (r) {
2027 vq_err(vq, "Can't access used idx at %p\n",
2028 &vq->used->idx);
2029 goto err;
2030 }
2031 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2032 return 0;
2033
2034err:
2035 vq->is_le = is_le;
2036 return r;
2037}
2038EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2039
2040static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2041 struct iovec iov[], int iov_size, int access)
2042{
2043 const struct vhost_umem_node *node;
2044 struct vhost_dev *dev = vq->dev;
2045 struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
2046 struct iovec *_iov;
2047 u64 s = 0;
2048 int ret = 0;
2049
2050 while ((u64)len > s) {
2051 u64 size;
2052 if (unlikely(ret >= iov_size)) {
2053 ret = -ENOBUFS;
2054 break;
2055 }
2056
2057 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
2058 addr, addr + len - 1);
2059 if (node == NULL || node->start > addr) {
2060 if (umem != dev->iotlb) {
2061 ret = -EFAULT;
2062 break;
2063 }
2064 ret = -EAGAIN;
2065 break;
2066 } else if (!(node->perm & access)) {
2067 ret = -EPERM;
2068 break;
2069 }
2070
2071 _iov = iov + ret;
2072 size = node->size - addr + node->start;
2073 _iov->iov_len = min((u64)len - s, size);
2074 _iov->iov_base = (void __user *)(unsigned long)
2075 (node->userspace_addr + addr - node->start);
2076 s += size;
2077 addr += size;
2078 ++ret;
2079 }
2080
2081 if (ret == -EAGAIN)
2082 vhost_iotlb_miss(vq, addr, access);
2083 return ret;
2084}
2085
2086/* Each buffer in the virtqueues is actually a chain of descriptors. This
2087 * function returns the next descriptor in the chain,
2088 * or -1U if we're at the end. */
2089static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2090{
2091 unsigned int next;
2092
2093 /* If this descriptor says it doesn't chain, we're done. */
2094 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2095 return -1U;
2096
2097 /* Check they're not leading us off end of descriptors. */
2098 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2099 return next;
2100}
2101
2102static int get_indirect(struct vhost_virtqueue *vq,
2103 struct iovec iov[], unsigned int iov_size,
2104 unsigned int *out_num, unsigned int *in_num,
2105 struct vhost_log *log, unsigned int *log_num,
2106 struct vring_desc *indirect)
2107{
2108 struct vring_desc desc;
2109 unsigned int i = 0, count, found = 0;
2110 u32 len = vhost32_to_cpu(vq, indirect->len);
2111 struct iov_iter from;
2112 int ret, access;
2113
2114 /* Sanity check */
2115 if (unlikely(len % sizeof desc)) {
2116 vq_err(vq, "Invalid length in indirect descriptor: "
2117 "len 0x%llx not multiple of 0x%zx\n",
2118 (unsigned long long)len,
2119 sizeof desc);
2120 return -EINVAL;
2121 }
2122
2123 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2124 UIO_MAXIOV, VHOST_ACCESS_RO);
2125 if (unlikely(ret < 0)) {
2126 if (ret != -EAGAIN)
2127 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2128 return ret;
2129 }
2130 iov_iter_init(&from, READ, vq->indirect, ret, len);
2131
2132 /* We will use the result as an address to read from, so most
2133 * architectures only need a compiler barrier here. */
2134 read_barrier_depends();
2135
2136 count = len / sizeof desc;
2137 /* Buffers are chained via a 16 bit next field, so
2138 * we can have at most 2^16 of these. */
2139 if (unlikely(count > USHRT_MAX + 1)) {
2140 vq_err(vq, "Indirect buffer length too big: %d\n",
2141 indirect->len);
2142 return -E2BIG;
2143 }
2144
2145 do {
2146 unsigned iov_count = *in_num + *out_num;
2147 if (unlikely(++found > count)) {
2148 vq_err(vq, "Loop detected: last one at %u "
2149 "indirect size %u\n",
2150 i, count);
2151 return -EINVAL;
2152 }
2153 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2154 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2155 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2156 return -EINVAL;
2157 }
2158 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2159 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2160 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2161 return -EINVAL;
2162 }
2163
2164 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2165 access = VHOST_ACCESS_WO;
2166 else
2167 access = VHOST_ACCESS_RO;
2168
2169 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2170 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2171 iov_size - iov_count, access);
2172 if (unlikely(ret < 0)) {
2173 if (ret != -EAGAIN)
2174 vq_err(vq, "Translation failure %d indirect idx %d\n",
2175 ret, i);
2176 return ret;
2177 }
2178 /* If this is an input descriptor, increment that count. */
2179 if (access == VHOST_ACCESS_WO) {
2180 *in_num += ret;
2181 if (unlikely(log && ret)) {
2182 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2183 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2184 ++*log_num;
2185 }
2186 } else {
2187 /* If it's an output descriptor, they're all supposed
2188 * to come before any input descriptors. */
2189 if (unlikely(*in_num)) {
2190 vq_err(vq, "Indirect descriptor "
2191 "has out after in: idx %d\n", i);
2192 return -EINVAL;
2193 }
2194 *out_num += ret;
2195 }
2196 } while ((i = next_desc(vq, &desc)) != -1);
2197 return 0;
2198}
2199
2200/* This looks in the virtqueue and for the first available buffer, and converts
2201 * it to an iovec for convenient access. Since descriptors consist of some
2202 * number of output then some number of input descriptors, it's actually two
2203 * iovecs, but we pack them into one and note how many of each there were.
2204 *
2205 * This function returns the descriptor number found, or vq->num (which is
2206 * never a valid descriptor number) if none was found. A negative code is
2207 * returned on error. */
2208int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2209 struct iovec iov[], unsigned int iov_size,
2210 unsigned int *out_num, unsigned int *in_num,
2211 struct vhost_log *log, unsigned int *log_num)
2212{
2213 struct vring_desc desc;
2214 unsigned int i, head, found = 0;
2215 u16 last_avail_idx;
2216 __virtio16 avail_idx;
2217 __virtio16 ring_head;
2218 int ret, access;
2219
2220 /* Check it isn't doing very strange things with descriptor numbers. */
2221 last_avail_idx = vq->last_avail_idx;
2222
2223 if (vq->avail_idx == vq->last_avail_idx) {
2224 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2225 vq_err(vq, "Failed to access avail idx at %p\n",
2226 &vq->avail->idx);
2227 return -EFAULT;
2228 }
2229 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2230
2231 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2232 vq_err(vq, "Guest moved used index from %u to %u",
2233 last_avail_idx, vq->avail_idx);
2234 return -EFAULT;
2235 }
2236
2237 /* If there's nothing new since last we looked, return
2238 * invalid.
2239 */
2240 if (vq->avail_idx == last_avail_idx)
2241 return vq->num;
2242
2243 /* Only get avail ring entries after they have been
2244 * exposed by guest.
2245 */
2246 smp_rmb();
2247 }
2248
2249 /* Grab the next descriptor number they're advertising, and increment
2250 * the index we've seen. */
2251 if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2252 vq_err(vq, "Failed to read head: idx %d address %p\n",
2253 last_avail_idx,
2254 &vq->avail->ring[last_avail_idx % vq->num]);
2255 return -EFAULT;
2256 }
2257
2258 head = vhost16_to_cpu(vq, ring_head);
2259
2260 /* If their number is silly, that's an error. */
2261 if (unlikely(head >= vq->num)) {
2262 vq_err(vq, "Guest says index %u > %u is available",
2263 head, vq->num);
2264 return -EINVAL;
2265 }
2266
2267 /* When we start there are none of either input nor output. */
2268 *out_num = *in_num = 0;
2269 if (unlikely(log))
2270 *log_num = 0;
2271
2272 i = head;
2273 do {
2274 unsigned iov_count = *in_num + *out_num;
2275 if (unlikely(i >= vq->num)) {
2276 vq_err(vq, "Desc index is %u > %u, head = %u",
2277 i, vq->num, head);
2278 return -EINVAL;
2279 }
2280 if (unlikely(++found > vq->num)) {
2281 vq_err(vq, "Loop detected: last one at %u "
2282 "vq size %u head %u\n",
2283 i, vq->num, head);
2284 return -EINVAL;
2285 }
2286 ret = vhost_get_desc(vq, &desc, i);
2287 if (unlikely(ret)) {
2288 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2289 i, vq->desc + i);
2290 return -EFAULT;
2291 }
2292 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2293 ret = get_indirect(vq, iov, iov_size,
2294 out_num, in_num,
2295 log, log_num, &desc);
2296 if (unlikely(ret < 0)) {
2297 if (ret != -EAGAIN)
2298 vq_err(vq, "Failure detected "
2299 "in indirect descriptor at idx %d\n", i);
2300 return ret;
2301 }
2302 continue;
2303 }
2304
2305 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2306 access = VHOST_ACCESS_WO;
2307 else
2308 access = VHOST_ACCESS_RO;
2309 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2310 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2311 iov_size - iov_count, access);
2312 if (unlikely(ret < 0)) {
2313 if (ret != -EAGAIN)
2314 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2315 ret, i);
2316 return ret;
2317 }
2318 if (access == VHOST_ACCESS_WO) {
2319 /* If this is an input descriptor,
2320 * increment that count. */
2321 *in_num += ret;
2322 if (unlikely(log && ret)) {
2323 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2324 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2325 ++*log_num;
2326 }
2327 } else {
2328 /* If it's an output descriptor, they're all supposed
2329 * to come before any input descriptors. */
2330 if (unlikely(*in_num)) {
2331 vq_err(vq, "Descriptor has out after in: "
2332 "idx %d\n", i);
2333 return -EINVAL;
2334 }
2335 *out_num += ret;
2336 }
2337 } while ((i = next_desc(vq, &desc)) != -1);
2338
2339 /* On success, increment avail index. */
2340 vq->last_avail_idx++;
2341
2342 /* Assume notifications from guest are disabled at this point,
2343 * if they aren't we would need to update avail_event index. */
2344 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2345 return head;
2346}
2347EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2348
2349/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2350void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2351{
2352 vq->last_avail_idx -= n;
2353}
2354EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2355
2356/* After we've used one of their buffers, we tell them about it. We'll then
2357 * want to notify the guest, using eventfd. */
2358int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2359{
2360 struct vring_used_elem heads = {
2361 cpu_to_vhost32(vq, head),
2362 cpu_to_vhost32(vq, len)
2363 };
2364
2365 return vhost_add_used_n(vq, &heads, 1);
2366}
2367EXPORT_SYMBOL_GPL(vhost_add_used);
2368
2369static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2370 struct vring_used_elem *heads,
2371 unsigned count)
2372{
2373 struct vring_used_elem __user *used;
2374 u16 old, new;
2375 int start;
2376
2377 start = vq->last_used_idx & (vq->num - 1);
2378 used = vq->used->ring + start;
2379 if (vhost_put_used(vq, heads, start, count)) {
2380 vq_err(vq, "Failed to write used");
2381 return -EFAULT;
2382 }
2383 if (unlikely(vq->log_used)) {
2384 /* Make sure data is seen before log. */
2385 smp_wmb();
2386 /* Log used ring entry write. */
2387 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2388 count * sizeof *used);
2389 }
2390 old = vq->last_used_idx;
2391 new = (vq->last_used_idx += count);
2392 /* If the driver never bothers to signal in a very long while,
2393 * used index might wrap around. If that happens, invalidate
2394 * signalled_used index we stored. TODO: make sure driver
2395 * signals at least once in 2^16 and remove this. */
2396 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2397 vq->signalled_used_valid = false;
2398 return 0;
2399}
2400
2401/* After we've used one of their buffers, we tell them about it. We'll then
2402 * want to notify the guest, using eventfd. */
2403int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2404 unsigned count)
2405{
2406 int start, n, r;
2407
2408 start = vq->last_used_idx & (vq->num - 1);
2409 n = vq->num - start;
2410 if (n < count) {
2411 r = __vhost_add_used_n(vq, heads, n);
2412 if (r < 0)
2413 return r;
2414 heads += n;
2415 count -= n;
2416 }
2417 r = __vhost_add_used_n(vq, heads, count);
2418
2419 /* Make sure buffer is written before we update index. */
2420 smp_wmb();
2421 if (vhost_put_used_idx(vq)) {
2422 vq_err(vq, "Failed to increment used idx");
2423 return -EFAULT;
2424 }
2425 if (unlikely(vq->log_used)) {
2426 /* Make sure used idx is seen before log. */
2427 smp_wmb();
2428 /* Log used index update. */
2429 log_used(vq, offsetof(struct vring_used, idx),
2430 sizeof vq->used->idx);
2431 if (vq->log_ctx)
2432 eventfd_signal(vq->log_ctx, 1);
2433 }
2434 return r;
2435}
2436EXPORT_SYMBOL_GPL(vhost_add_used_n);
2437
2438static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2439{
2440 __u16 old, new;
2441 __virtio16 event;
2442 bool v;
2443 /* Flush out used index updates. This is paired
2444 * with the barrier that the Guest executes when enabling
2445 * interrupts. */
2446 smp_mb();
2447
2448 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2449 unlikely(vq->avail_idx == vq->last_avail_idx))
2450 return true;
2451
2452 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2453 __virtio16 flags;
2454 if (vhost_get_avail_flags(vq, &flags)) {
2455 vq_err(vq, "Failed to get flags");
2456 return true;
2457 }
2458 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2459 }
2460 old = vq->signalled_used;
2461 v = vq->signalled_used_valid;
2462 new = vq->signalled_used = vq->last_used_idx;
2463 vq->signalled_used_valid = true;
2464
2465 if (unlikely(!v))
2466 return true;
2467
2468 if (vhost_get_used_event(vq, &event)) {
2469 vq_err(vq, "Failed to get used event idx");
2470 return true;
2471 }
2472 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2473}
2474
2475/* This actually signals the guest, using eventfd. */
2476void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2477{
2478 /* Signal the Guest tell them we used something up. */
2479 if (vq->call_ctx && vhost_notify(dev, vq))
2480 eventfd_signal(vq->call_ctx, 1);
2481}
2482EXPORT_SYMBOL_GPL(vhost_signal);
2483
2484/* And here's the combo meal deal. Supersize me! */
2485void vhost_add_used_and_signal(struct vhost_dev *dev,
2486 struct vhost_virtqueue *vq,
2487 unsigned int head, int len)
2488{
2489 vhost_add_used(vq, head, len);
2490 vhost_signal(dev, vq);
2491}
2492EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2493
2494/* multi-buffer version of vhost_add_used_and_signal */
2495void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2496 struct vhost_virtqueue *vq,
2497 struct vring_used_elem *heads, unsigned count)
2498{
2499 vhost_add_used_n(vq, heads, count);
2500 vhost_signal(dev, vq);
2501}
2502EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2503
2504/* return true if we're sure that avaiable ring is empty */
2505bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2506{
2507 __virtio16 avail_idx;
2508 int r;
2509
2510 if (vq->avail_idx != vq->last_avail_idx)
2511 return false;
2512
2513 r = vhost_get_avail_idx(vq, &avail_idx);
2514 if (unlikely(r))
2515 return false;
2516 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2517
2518 return vq->avail_idx == vq->last_avail_idx;
2519}
2520EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2521
2522/* OK, now we need to know about added descriptors. */
2523bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2524{
2525 __virtio16 avail_idx;
2526 int r;
2527
2528 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2529 return false;
2530 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2531 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2532 r = vhost_update_used_flags(vq);
2533 if (r) {
2534 vq_err(vq, "Failed to enable notification at %p: %d\n",
2535 &vq->used->flags, r);
2536 return false;
2537 }
2538 } else {
2539 r = vhost_update_avail_event(vq, vq->avail_idx);
2540 if (r) {
2541 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2542 vhost_avail_event(vq), r);
2543 return false;
2544 }
2545 }
2546 /* They could have slipped one in as we were doing that: make
2547 * sure it's written, then check again. */
2548 smp_mb();
2549 r = vhost_get_avail_idx(vq, &avail_idx);
2550 if (r) {
2551 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2552 &vq->avail->idx, r);
2553 return false;
2554 }
2555
2556 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2557}
2558EXPORT_SYMBOL_GPL(vhost_enable_notify);
2559
2560/* We don't need to be notified again. */
2561void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2562{
2563 int r;
2564
2565 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2566 return;
2567 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2568 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2569 r = vhost_update_used_flags(vq);
2570 if (r)
2571 vq_err(vq, "Failed to enable notification at %p: %d\n",
2572 &vq->used->flags, r);
2573 }
2574}
2575EXPORT_SYMBOL_GPL(vhost_disable_notify);
2576
2577/* Create a new message. */
2578struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2579{
2580 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2581 if (!node)
2582 return NULL;
2583
2584 /* Make sure all padding within the structure is initialized. */
2585 memset(&node->msg, 0, sizeof node->msg);
2586 node->vq = vq;
2587 node->msg.type = type;
2588 return node;
2589}
2590EXPORT_SYMBOL_GPL(vhost_new_msg);
2591
2592void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2593 struct vhost_msg_node *node)
2594{
2595 spin_lock(&dev->iotlb_lock);
2596 list_add_tail(&node->node, head);
2597 spin_unlock(&dev->iotlb_lock);
2598
2599 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2600}
2601EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2602
2603struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2604 struct list_head *head)
2605{
2606 struct vhost_msg_node *node = NULL;
2607
2608 spin_lock(&dev->iotlb_lock);
2609 if (!list_empty(head)) {
2610 node = list_first_entry(head, struct vhost_msg_node,
2611 node);
2612 list_del(&node->node);
2613 }
2614 spin_unlock(&dev->iotlb_lock);
2615
2616 return node;
2617}
2618EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2619
2620
2621static int __init vhost_init(void)
2622{
2623 return 0;
2624}
2625
2626static void __exit vhost_exit(void)
2627{
2628}
2629
2630module_init(vhost_init);
2631module_exit(vhost_exit);
2632
2633MODULE_VERSION("0.0.1");
2634MODULE_LICENSE("GPL v2");
2635MODULE_AUTHOR("Michael S. Tsirkin");
2636MODULE_DESCRIPTION("Host kernel accelerator for virtio");