]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - drivers/vhost/vhost.c
projects / mirror_ubuntu-bionic-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
vhost: synchronize IOTLB message with dev cleanup
[mirror_ubuntu-bionic-kernel.git] / drivers / vhost / vhost.c
1 /* Copyright (C) 2009 Red Hat, Inc.
2 * Copyright (C) 2006 Rusty Russell IBM Corporation
3 *
4 * Author: Michael S. Tsirkin <mst@redhat.com>
5 *
6 * Inspiration, some code, and most witty comments come from
7 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
8 *
9 * This work is licensed under the terms of the GNU GPL, version 2.
10 *
11 * Generic code for virtio server in host kernel.
12 */
13
14 #include <linux/eventfd.h>
15 #include <linux/vhost.h>
16 #include <linux/uio.h>
17 #include <linux/mm.h>
18 #include <linux/mmu_context.h>
19 #include <linux/miscdevice.h>
20 #include <linux/mutex.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/vmalloc.h>
26 #include <linux/kthread.h>
27 #include <linux/cgroup.h>
28 #include <linux/module.h>
29 #include <linux/sort.h>
30 #include <linux/sched/mm.h>
31 #include <linux/sched/signal.h>
32 #include <linux/interval_tree_generic.h>
33
34 #include "vhost.h"
35
36 static ushort max_mem_regions = 64;
37 module_param(max_mem_regions, ushort, 0444);
38 MODULE_PARM_DESC(max_mem_regions,
39 "Maximum number of memory regions in memory map. (default: 64)");
40 static int max_iotlb_entries = 2048;
41 module_param(max_iotlb_entries, int, 0444);
42 MODULE_PARM_DESC(max_iotlb_entries,
43 "Maximum number of iotlb entries. (default: 2048)");
44
45 enum {
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 INTERVAL_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
57 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
58 {
59 vq->user_be = !virtio_legacy_is_little_endian();
60 }
61
62 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
63 {
64 vq->user_be = true;
65 }
66
67 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
68 {
69 vq->user_be = false;
70 }
71
72 static 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
94 static 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
108 static 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
118 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
119 {
120 }
121
122 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
123 {
124 return -ENOIOCTLCMD;
125 }
126
127 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
128 int __user *argp)
129 {
130 return -ENOIOCTLCMD;
131 }
132
133 static 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
140 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
141 {
142 vhost_init_is_le(vq);
143 }
144
145 struct vhost_flush_struct {
146 struct vhost_work work;
147 struct completion wait_event;
148 };
149
150 static 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
158 static 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
168 static 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 (!((unsigned long)key & poll->mask))
174 return 0;
175
176 vhost_poll_queue(poll);
177 return 0;
178 }
179
180 void 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 init_waitqueue_head(&work->done);
185 }
186 EXPORT_SYMBOL_GPL(vhost_work_init);
187
188 /* Init poll structure */
189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
190 unsigned long mask, struct vhost_dev *dev)
191 {
192 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
193 init_poll_funcptr(&poll->table, vhost_poll_func);
194 poll->mask = mask;
195 poll->dev = dev;
196 poll->wqh = NULL;
197
198 vhost_work_init(&poll->work, fn);
199 }
200 EXPORT_SYMBOL_GPL(vhost_poll_init);
201
202 /* Start polling a file. We add ourselves to file's wait queue. The caller must
203 * keep a reference to a file until after vhost_poll_stop is called. */
204 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
205 {
206 unsigned long mask;
207 int ret = 0;
208
209 if (poll->wqh)
210 return 0;
211
212 mask = file->f_op->poll(file, &poll->table);
213 if (mask)
214 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
215 if (mask & POLLERR) {
216 vhost_poll_stop(poll);
217 ret = -EINVAL;
218 }
219
220 return ret;
221 }
222 EXPORT_SYMBOL_GPL(vhost_poll_start);
223
224 /* Stop polling a file. After this function returns, it becomes safe to drop the
225 * file reference. You must also flush afterwards. */
226 void vhost_poll_stop(struct vhost_poll *poll)
227 {
228 if (poll->wqh) {
229 remove_wait_queue(poll->wqh, &poll->wait);
230 poll->wqh = NULL;
231 }
232 }
233 EXPORT_SYMBOL_GPL(vhost_poll_stop);
234
235 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
236 {
237 struct vhost_flush_struct flush;
238
239 if (dev->worker) {
240 init_completion(&flush.wait_event);
241 vhost_work_init(&flush.work, vhost_flush_work);
242
243 vhost_work_queue(dev, &flush.work);
244 wait_for_completion(&flush.wait_event);
245 }
246 }
247 EXPORT_SYMBOL_GPL(vhost_work_flush);
248
249 /* Flush any work that has been scheduled. When calling this, don't hold any
250 * locks that are also used by the callback. */
251 void vhost_poll_flush(struct vhost_poll *poll)
252 {
253 vhost_work_flush(poll->dev, &poll->work);
254 }
255 EXPORT_SYMBOL_GPL(vhost_poll_flush);
256
257 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
258 {
259 if (!dev->worker)
260 return;
261
262 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
263 /* We can only add the work to the list after we're
264 * sure it was not in the list.
265 * test_and_set_bit() implies a memory barrier.
266 */
267 llist_add(&work->node, &dev->work_list);
268 wake_up_process(dev->worker);
269 }
270 }
271 EXPORT_SYMBOL_GPL(vhost_work_queue);
272
273 /* A lockless hint for busy polling code to exit the loop */
274 bool vhost_has_work(struct vhost_dev *dev)
275 {
276 return !llist_empty(&dev->work_list);
277 }
278 EXPORT_SYMBOL_GPL(vhost_has_work);
279
280 void vhost_poll_queue(struct vhost_poll *poll)
281 {
282 vhost_work_queue(poll->dev, &poll->work);
283 }
284 EXPORT_SYMBOL_GPL(vhost_poll_queue);
285
286 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
287 {
288 int j;
289
290 for (j = 0; j < VHOST_NUM_ADDRS; j++)
291 vq->meta_iotlb[j] = NULL;
292 }
293
294 static void vhost_vq_meta_reset(struct vhost_dev *d)
295 {
296 int i;
297
298 for (i = 0; i < d->nvqs; ++i)
299 __vhost_vq_meta_reset(d->vqs[i]);
300 }
301
302 static void vhost_vq_reset(struct vhost_dev *dev,
303 struct vhost_virtqueue *vq)
304 {
305 vq->num = 1;
306 vq->desc = NULL;
307 vq->avail = NULL;
308 vq->used = NULL;
309 vq->last_avail_idx = 0;
310 vq->avail_idx = 0;
311 vq->last_used_idx = 0;
312 vq->signalled_used = 0;
313 vq->signalled_used_valid = false;
314 vq->used_flags = 0;
315 vq->log_used = false;
316 vq->log_addr = -1ull;
317 vq->private_data = NULL;
318 vq->acked_features = 0;
319 vq->log_base = NULL;
320 vq->error_ctx = NULL;
321 vq->error = NULL;
322 vq->kick = NULL;
323 vq->call_ctx = NULL;
324 vq->call = NULL;
325 vq->log_ctx = NULL;
326 vhost_reset_is_le(vq);
327 vhost_disable_cross_endian(vq);
328 vq->busyloop_timeout = 0;
329 vq->umem = NULL;
330 vq->iotlb = NULL;
331 __vhost_vq_meta_reset(vq);
332 }
333
334 static int vhost_worker(void *data)
335 {
336 struct vhost_dev *dev = data;
337 struct vhost_work *work, *work_next;
338 struct llist_node *node;
339 mm_segment_t oldfs = get_fs();
340
341 set_fs(USER_DS);
342 use_mm(dev->mm);
343
344 for (;;) {
345 /* mb paired w/ kthread_stop */
346 set_current_state(TASK_INTERRUPTIBLE);
347
348 if (kthread_should_stop()) {
349 __set_current_state(TASK_RUNNING);
350 break;
351 }
352
353 node = llist_del_all(&dev->work_list);
354 if (!node)
355 schedule();
356
357 node = llist_reverse_order(node);
358 /* make sure flag is seen after deletion */
359 smp_wmb();
360 llist_for_each_entry_safe(work, work_next, node, node) {
361 clear_bit(VHOST_WORK_QUEUED, &work->flags);
362 __set_current_state(TASK_RUNNING);
363 work->fn(work);
364 if (need_resched())
365 schedule();
366 }
367 }
368 unuse_mm(dev->mm);
369 set_fs(oldfs);
370 return 0;
371 }
372
373 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
374 {
375 kfree(vq->indirect);
376 vq->indirect = NULL;
377 kfree(vq->log);
378 vq->log = NULL;
379 kfree(vq->heads);
380 vq->heads = NULL;
381 }
382
383 /* Helper to allocate iovec buffers for all vqs. */
384 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
385 {
386 struct vhost_virtqueue *vq;
387 int i;
388
389 for (i = 0; i < dev->nvqs; ++i) {
390 vq = dev->vqs[i];
391 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
392 GFP_KERNEL);
393 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
394 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
395 if (!vq->indirect || !vq->log || !vq->heads)
396 goto err_nomem;
397 }
398 return 0;
399
400 err_nomem:
401 for (; i >= 0; --i)
402 vhost_vq_free_iovecs(dev->vqs[i]);
403 return -ENOMEM;
404 }
405
406 static 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
414 void vhost_dev_init(struct vhost_dev *dev,
415 struct vhost_virtqueue **vqs, int nvqs)
416 {
417 struct vhost_virtqueue *vq;
418 int i;
419
420 dev->vqs = vqs;
421 dev->nvqs = nvqs;
422 mutex_init(&dev->mutex);
423 dev->log_ctx = NULL;
424 dev->log_file = NULL;
425 dev->umem = NULL;
426 dev->iotlb = NULL;
427 dev->mm = NULL;
428 dev->worker = NULL;
429 init_llist_head(&dev->work_list);
430 init_waitqueue_head(&dev->wait);
431 INIT_LIST_HEAD(&dev->read_list);
432 INIT_LIST_HEAD(&dev->pending_list);
433 spin_lock_init(&dev->iotlb_lock);
434
435
436 for (i = 0; i < dev->nvqs; ++i) {
437 vq = dev->vqs[i];
438 vq->log = NULL;
439 vq->indirect = NULL;
440 vq->heads = NULL;
441 vq->dev = dev;
442 mutex_init(&vq->mutex);
443 vhost_vq_reset(dev, vq);
444 if (vq->handle_kick)
445 vhost_poll_init(&vq->poll, vq->handle_kick,
446 POLLIN, dev);
447 }
448 }
449 EXPORT_SYMBOL_GPL(vhost_dev_init);
450
451 /* Caller should have device mutex */
452 long vhost_dev_check_owner(struct vhost_dev *dev)
453 {
454 /* Are you the owner? If not, I don't think you mean to do that */
455 return dev->mm == current->mm ? 0 : -EPERM;
456 }
457 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
458
459 struct vhost_attach_cgroups_struct {
460 struct vhost_work work;
461 struct task_struct *owner;
462 int ret;
463 };
464
465 static void vhost_attach_cgroups_work(struct vhost_work *work)
466 {
467 struct vhost_attach_cgroups_struct *s;
468
469 s = container_of(work, struct vhost_attach_cgroups_struct, work);
470 s->ret = cgroup_attach_task_all(s->owner, current);
471 }
472
473 static int vhost_attach_cgroups(struct vhost_dev *dev)
474 {
475 struct vhost_attach_cgroups_struct attach;
476
477 attach.owner = current;
478 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
479 vhost_work_queue(dev, &attach.work);
480 vhost_work_flush(dev, &attach.work);
481 return attach.ret;
482 }
483
484 /* Caller should have device mutex */
485 bool vhost_dev_has_owner(struct vhost_dev *dev)
486 {
487 return dev->mm;
488 }
489 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
490
491 /* Caller should have device mutex */
492 long vhost_dev_set_owner(struct vhost_dev *dev)
493 {
494 struct task_struct *worker;
495 int err;
496
497 /* Is there an owner already? */
498 if (vhost_dev_has_owner(dev)) {
499 err = -EBUSY;
500 goto err_mm;
501 }
502
503 /* No owner, become one */
504 dev->mm = get_task_mm(current);
505 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
506 if (IS_ERR(worker)) {
507 err = PTR_ERR(worker);
508 goto err_worker;
509 }
510
511 dev->worker = worker;
512 wake_up_process(worker); /* avoid contributing to loadavg */
513
514 err = vhost_attach_cgroups(dev);
515 if (err)
516 goto err_cgroup;
517
518 err = vhost_dev_alloc_iovecs(dev);
519 if (err)
520 goto err_cgroup;
521
522 return 0;
523 err_cgroup:
524 kthread_stop(worker);
525 dev->worker = NULL;
526 err_worker:
527 if (dev->mm)
528 mmput(dev->mm);
529 dev->mm = NULL;
530 err_mm:
531 return err;
532 }
533 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
534
535 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
536 {
537 return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
538 }
539 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
540
541 /* Caller should have device mutex */
542 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
543 {
544 int i;
545
546 vhost_dev_cleanup(dev, true);
547
548 /* Restore memory to default empty mapping. */
549 INIT_LIST_HEAD(&umem->umem_list);
550 dev->umem = umem;
551 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
552 * VQs aren't running.
553 */
554 for (i = 0; i < dev->nvqs; ++i)
555 dev->vqs[i]->umem = umem;
556 }
557 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
558
559 void vhost_dev_stop(struct vhost_dev *dev)
560 {
561 int i;
562
563 for (i = 0; i < dev->nvqs; ++i) {
564 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
565 vhost_poll_stop(&dev->vqs[i]->poll);
566 vhost_poll_flush(&dev->vqs[i]->poll);
567 }
568 }
569 }
570 EXPORT_SYMBOL_GPL(vhost_dev_stop);
571
572 static void vhost_umem_free(struct vhost_umem *umem,
573 struct vhost_umem_node *node)
574 {
575 vhost_umem_interval_tree_remove(node, &umem->umem_tree);
576 list_del(&node->link);
577 kfree(node);
578 umem->numem--;
579 }
580
581 static void vhost_umem_clean(struct vhost_umem *umem)
582 {
583 struct vhost_umem_node *node, *tmp;
584
585 if (!umem)
586 return;
587
588 list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
589 vhost_umem_free(umem, node);
590
591 kvfree(umem);
592 }
593
594 static void vhost_clear_msg(struct vhost_dev *dev)
595 {
596 struct vhost_msg_node *node, *n;
597
598 spin_lock(&dev->iotlb_lock);
599
600 list_for_each_entry_safe(node, n, &dev->read_list, node) {
601 list_del(&node->node);
602 kfree(node);
603 }
604
605 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
606 list_del(&node->node);
607 kfree(node);
608 }
609
610 spin_unlock(&dev->iotlb_lock);
611 }
612
613 /* Caller should have device mutex if and only if locked is set */
614 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
615 {
616 int i;
617
618 for (i = 0; i < dev->nvqs; ++i) {
619 if (dev->vqs[i]->error_ctx)
620 eventfd_ctx_put(dev->vqs[i]->error_ctx);
621 if (dev->vqs[i]->error)
622 fput(dev->vqs[i]->error);
623 if (dev->vqs[i]->kick)
624 fput(dev->vqs[i]->kick);
625 if (dev->vqs[i]->call_ctx)
626 eventfd_ctx_put(dev->vqs[i]->call_ctx);
627 if (dev->vqs[i]->call)
628 fput(dev->vqs[i]->call);
629 vhost_vq_reset(dev, dev->vqs[i]);
630 }
631 vhost_dev_free_iovecs(dev);
632 if (dev->log_ctx)
633 eventfd_ctx_put(dev->log_ctx);
634 dev->log_ctx = NULL;
635 if (dev->log_file)
636 fput(dev->log_file);
637 dev->log_file = NULL;
638 /* No one will access memory at this point */
639 vhost_umem_clean(dev->umem);
640 dev->umem = NULL;
641 vhost_umem_clean(dev->iotlb);
642 dev->iotlb = NULL;
643 vhost_clear_msg(dev);
644 wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
645 WARN_ON(!llist_empty(&dev->work_list));
646 if (dev->worker) {
647 kthread_stop(dev->worker);
648 dev->worker = NULL;
649 }
650 if (dev->mm)
651 mmput(dev->mm);
652 dev->mm = NULL;
653 }
654 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
655
656 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
657 {
658 u64 a = addr / VHOST_PAGE_SIZE / 8;
659
660 /* Make sure 64 bit math will not overflow. */
661 if (a > ULONG_MAX - (unsigned long)log_base ||
662 a + (unsigned long)log_base > ULONG_MAX)
663 return 0;
664
665 return access_ok(VERIFY_WRITE, log_base + a,
666 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
667 }
668
669 static bool vhost_overflow(u64 uaddr, u64 size)
670 {
671 /* Make sure 64 bit math will not overflow. */
672 return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
673 }
674
675 /* Caller should have vq mutex and device mutex. */
676 static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
677 int log_all)
678 {
679 struct vhost_umem_node *node;
680
681 if (!umem)
682 return 0;
683
684 list_for_each_entry(node, &umem->umem_list, link) {
685 unsigned long a = node->userspace_addr;
686
687 if (vhost_overflow(node->userspace_addr, node->size))
688 return 0;
689
690
691 if (!access_ok(VERIFY_WRITE, (void __user *)a,
692 node->size))
693 return 0;
694 else if (log_all && !log_access_ok(log_base,
695 node->start,
696 node->size))
697 return 0;
698 }
699 return 1;
700 }
701
702 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
703 u64 addr, unsigned int size,
704 int type)
705 {
706 const struct vhost_umem_node *node = vq->meta_iotlb[type];
707
708 if (!node)
709 return NULL;
710
711 return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
712 }
713
714 /* Can we switch to this memory table? */
715 /* Caller should have device mutex but not vq mutex */
716 static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
717 int log_all)
718 {
719 int i;
720
721 for (i = 0; i < d->nvqs; ++i) {
722 int ok;
723 bool log;
724
725 mutex_lock(&d->vqs[i]->mutex);
726 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
727 /* If ring is inactive, will check when it's enabled. */
728 if (d->vqs[i]->private_data)
729 ok = vq_memory_access_ok(d->vqs[i]->log_base,
730 umem, log);
731 else
732 ok = 1;
733 mutex_unlock(&d->vqs[i]->mutex);
734 if (!ok)
735 return 0;
736 }
737 return 1;
738 }
739
740 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
741 struct iovec iov[], int iov_size, int access);
742
743 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
744 const void *from, unsigned size)
745 {
746 int ret;
747
748 if (!vq->iotlb)
749 return __copy_to_user(to, from, size);
750 else {
751 /* This function should be called after iotlb
752 * prefetch, which means we're sure that all vq
753 * could be access through iotlb. So -EAGAIN should
754 * not happen in this case.
755 */
756 struct iov_iter t;
757 void __user *uaddr = vhost_vq_meta_fetch(vq,
758 (u64)(uintptr_t)to, size,
759 VHOST_ADDR_USED);
760
761 if (uaddr)
762 return __copy_to_user(uaddr, from, size);
763
764 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
765 ARRAY_SIZE(vq->iotlb_iov),
766 VHOST_ACCESS_WO);
767 if (ret < 0)
768 goto out;
769 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
770 ret = copy_to_iter(from, size, &t);
771 if (ret == size)
772 ret = 0;
773 }
774 out:
775 return ret;
776 }
777
778 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
779 void __user *from, unsigned size)
780 {
781 int ret;
782
783 if (!vq->iotlb)
784 return __copy_from_user(to, from, size);
785 else {
786 /* This function should be called after iotlb
787 * prefetch, which means we're sure that vq
788 * could be access through iotlb. So -EAGAIN should
789 * not happen in this case.
790 */
791 void __user *uaddr = vhost_vq_meta_fetch(vq,
792 (u64)(uintptr_t)from, size,
793 VHOST_ADDR_DESC);
794 struct iov_iter f;
795
796 if (uaddr)
797 return __copy_from_user(to, uaddr, size);
798
799 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
800 ARRAY_SIZE(vq->iotlb_iov),
801 VHOST_ACCESS_RO);
802 if (ret < 0) {
803 vq_err(vq, "IOTLB translation failure: uaddr "
804 "%p size 0x%llx\n", from,
805 (unsigned long long) size);
806 goto out;
807 }
808 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
809 ret = copy_from_iter(to, size, &f);
810 if (ret == size)
811 ret = 0;
812 }
813
814 out:
815 return ret;
816 }
817
818 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
819 void __user *addr, unsigned int size,
820 int type)
821 {
822 int ret;
823
824 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
825 ARRAY_SIZE(vq->iotlb_iov),
826 VHOST_ACCESS_RO);
827 if (ret < 0) {
828 vq_err(vq, "IOTLB translation failure: uaddr "
829 "%p size 0x%llx\n", addr,
830 (unsigned long long) size);
831 return NULL;
832 }
833
834 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
835 vq_err(vq, "Non atomic userspace memory access: uaddr "
836 "%p size 0x%llx\n", addr,
837 (unsigned long long) size);
838 return NULL;
839 }
840
841 return vq->iotlb_iov[0].iov_base;
842 }
843
844 /* This function should be called after iotlb
845 * prefetch, which means we're sure that vq
846 * could be access through iotlb. So -EAGAIN should
847 * not happen in this case.
848 */
849 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
850 void *addr, unsigned int size,
851 int type)
852 {
853 void __user *uaddr = vhost_vq_meta_fetch(vq,
854 (u64)(uintptr_t)addr, size, type);
855 if (uaddr)
856 return uaddr;
857
858 return __vhost_get_user_slow(vq, addr, size, type);
859 }
860
861 #define vhost_put_user(vq, x, ptr) \
862 ({ \
863 int ret = -EFAULT; \
864 if (!vq->iotlb) { \
865 ret = __put_user(x, ptr); \
866 } else { \
867 __typeof__(ptr) to = \
868 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
869 sizeof(*ptr), VHOST_ADDR_USED); \
870 if (to != NULL) \
871 ret = __put_user(x, to); \
872 else \
873 ret = -EFAULT; \
874 } \
875 ret; \
876 })
877
878 #define vhost_get_user(vq, x, ptr, type) \
879 ({ \
880 int ret; \
881 if (!vq->iotlb) { \
882 ret = __get_user(x, ptr); \
883 } else { \
884 __typeof__(ptr) from = \
885 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
886 sizeof(*ptr), \
887 type); \
888 if (from != NULL) \
889 ret = __get_user(x, from); \
890 else \
891 ret = -EFAULT; \
892 } \
893 ret; \
894 })
895
896 #define vhost_get_avail(vq, x, ptr) \
897 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
898
899 #define vhost_get_used(vq, x, ptr) \
900 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
901
902 static void vhost_dev_lock_vqs(struct vhost_dev *d)
903 {
904 int i = 0;
905 for (i = 0; i < d->nvqs; ++i)
906 mutex_lock_nested(&d->vqs[i]->mutex, i);
907 }
908
909 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
910 {
911 int i = 0;
912 for (i = 0; i < d->nvqs; ++i)
913 mutex_unlock(&d->vqs[i]->mutex);
914 }
915
916 static int vhost_new_umem_range(struct vhost_umem *umem,
917 u64 start, u64 size, u64 end,
918 u64 userspace_addr, int perm)
919 {
920 struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
921
922 if (!node)
923 return -ENOMEM;
924
925 if (umem->numem == max_iotlb_entries) {
926 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
927 vhost_umem_free(umem, tmp);
928 }
929
930 node->start = start;
931 node->size = size;
932 node->last = end;
933 node->userspace_addr = userspace_addr;
934 node->perm = perm;
935 INIT_LIST_HEAD(&node->link);
936 list_add_tail(&node->link, &umem->umem_list);
937 vhost_umem_interval_tree_insert(node, &umem->umem_tree);
938 umem->numem++;
939
940 return 0;
941 }
942
943 static void vhost_del_umem_range(struct vhost_umem *umem,
944 u64 start, u64 end)
945 {
946 struct vhost_umem_node *node;
947
948 while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
949 start, end)))
950 vhost_umem_free(umem, node);
951 }
952
953 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
954 struct vhost_iotlb_msg *msg)
955 {
956 struct vhost_msg_node *node, *n;
957
958 spin_lock(&d->iotlb_lock);
959
960 list_for_each_entry_safe(node, n, &d->pending_list, node) {
961 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
962 if (msg->iova <= vq_msg->iova &&
963 msg->iova + msg->size - 1 > vq_msg->iova &&
964 vq_msg->type == VHOST_IOTLB_MISS) {
965 vhost_poll_queue(&node->vq->poll);
966 list_del(&node->node);
967 kfree(node);
968 }
969 }
970
971 spin_unlock(&d->iotlb_lock);
972 }
973
974 static int umem_access_ok(u64 uaddr, u64 size, int access)
975 {
976 unsigned long a = uaddr;
977
978 /* Make sure 64 bit math will not overflow. */
979 if (vhost_overflow(uaddr, size))
980 return -EFAULT;
981
982 if ((access & VHOST_ACCESS_RO) &&
983 !access_ok(VERIFY_READ, (void __user *)a, size))
984 return -EFAULT;
985 if ((access & VHOST_ACCESS_WO) &&
986 !access_ok(VERIFY_WRITE, (void __user *)a, size))
987 return -EFAULT;
988 return 0;
989 }
990
991 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
992 struct vhost_iotlb_msg *msg)
993 {
994 int ret = 0;
995
996 mutex_lock(&dev->mutex);
997 vhost_dev_lock_vqs(dev);
998 switch (msg->type) {
999 case VHOST_IOTLB_UPDATE:
1000 if (!dev->iotlb) {
1001 ret = -EFAULT;
1002 break;
1003 }
1004 if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1005 ret = -EFAULT;
1006 break;
1007 }
1008 vhost_vq_meta_reset(dev);
1009 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1010 msg->iova + msg->size - 1,
1011 msg->uaddr, msg->perm)) {
1012 ret = -ENOMEM;
1013 break;
1014 }
1015 vhost_iotlb_notify_vq(dev, msg);
1016 break;
1017 case VHOST_IOTLB_INVALIDATE:
1018 if (!dev->iotlb) {
1019 ret = -EFAULT;
1020 break;
1021 }
1022 vhost_vq_meta_reset(dev);
1023 vhost_del_umem_range(dev->iotlb, msg->iova,
1024 msg->iova + msg->size - 1);
1025 break;
1026 default:
1027 ret = -EINVAL;
1028 break;
1029 }
1030
1031 vhost_dev_unlock_vqs(dev);
1032 mutex_unlock(&dev->mutex);
1033
1034 return ret;
1035 }
1036 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1037 struct iov_iter *from)
1038 {
1039 struct vhost_msg_node node;
1040 unsigned size = sizeof(struct vhost_msg);
1041 size_t ret;
1042 int err;
1043
1044 if (iov_iter_count(from) < size)
1045 return 0;
1046 ret = copy_from_iter(&node.msg, size, from);
1047 if (ret != size)
1048 goto done;
1049
1050 switch (node.msg.type) {
1051 case VHOST_IOTLB_MSG:
1052 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
1053 if (err)
1054 ret = err;
1055 break;
1056 default:
1057 ret = -EINVAL;
1058 break;
1059 }
1060
1061 done:
1062 return ret;
1063 }
1064 EXPORT_SYMBOL(vhost_chr_write_iter);
1065
1066 unsigned int vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1067 poll_table *wait)
1068 {
1069 unsigned int mask = 0;
1070
1071 poll_wait(file, &dev->wait, wait);
1072
1073 if (!list_empty(&dev->read_list))
1074 mask |= POLLIN | POLLRDNORM;
1075
1076 return mask;
1077 }
1078 EXPORT_SYMBOL(vhost_chr_poll);
1079
1080 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1081 int noblock)
1082 {
1083 DEFINE_WAIT(wait);
1084 struct vhost_msg_node *node;
1085 ssize_t ret = 0;
1086 unsigned size = sizeof(struct vhost_msg);
1087
1088 if (iov_iter_count(to) < size)
1089 return 0;
1090
1091 while (1) {
1092 if (!noblock)
1093 prepare_to_wait(&dev->wait, &wait,
1094 TASK_INTERRUPTIBLE);
1095
1096 node = vhost_dequeue_msg(dev, &dev->read_list);
1097 if (node)
1098 break;
1099 if (noblock) {
1100 ret = -EAGAIN;
1101 break;
1102 }
1103 if (signal_pending(current)) {
1104 ret = -ERESTARTSYS;
1105 break;
1106 }
1107 if (!dev->iotlb) {
1108 ret = -EBADFD;
1109 break;
1110 }
1111
1112 schedule();
1113 }
1114
1115 if (!noblock)
1116 finish_wait(&dev->wait, &wait);
1117
1118 if (node) {
1119 ret = copy_to_iter(&node->msg, size, to);
1120
1121 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1122 kfree(node);
1123 return ret;
1124 }
1125
1126 vhost_enqueue_msg(dev, &dev->pending_list, node);
1127 }
1128
1129 return ret;
1130 }
1131 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1132
1133 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1134 {
1135 struct vhost_dev *dev = vq->dev;
1136 struct vhost_msg_node *node;
1137 struct vhost_iotlb_msg *msg;
1138
1139 node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1140 if (!node)
1141 return -ENOMEM;
1142
1143 msg = &node->msg.iotlb;
1144 msg->type = VHOST_IOTLB_MISS;
1145 msg->iova = iova;
1146 msg->perm = access;
1147
1148 vhost_enqueue_msg(dev, &dev->read_list, node);
1149
1150 return 0;
1151 }
1152
1153 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1154 struct vring_desc __user *desc,
1155 struct vring_avail __user *avail,
1156 struct vring_used __user *used)
1157
1158 {
1159 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1160
1161 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1162 access_ok(VERIFY_READ, avail,
1163 sizeof *avail + num * sizeof *avail->ring + s) &&
1164 access_ok(VERIFY_WRITE, used,
1165 sizeof *used + num * sizeof *used->ring + s);
1166 }
1167
1168 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1169 const struct vhost_umem_node *node,
1170 int type)
1171 {
1172 int access = (type == VHOST_ADDR_USED) ?
1173 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1174
1175 if (likely(node->perm & access))
1176 vq->meta_iotlb[type] = node;
1177 }
1178
1179 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1180 int access, u64 addr, u64 len, int type)
1181 {
1182 const struct vhost_umem_node *node;
1183 struct vhost_umem *umem = vq->iotlb;
1184 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1185
1186 if (vhost_vq_meta_fetch(vq, addr, len, type))
1187 return true;
1188
1189 while (len > s) {
1190 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1191 addr,
1192 last);
1193 if (node == NULL || node->start > addr) {
1194 vhost_iotlb_miss(vq, addr, access);
1195 return false;
1196 } else if (!(node->perm & access)) {
1197 /* Report the possible access violation by
1198 * request another translation from userspace.
1199 */
1200 return false;
1201 }
1202
1203 size = node->size - addr + node->start;
1204
1205 if (orig_addr == addr && size >= len)
1206 vhost_vq_meta_update(vq, node, type);
1207
1208 s += size;
1209 addr += size;
1210 }
1211
1212 return true;
1213 }
1214
1215 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1216 {
1217 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1218 unsigned int num = vq->num;
1219
1220 if (!vq->iotlb)
1221 return 1;
1222
1223 return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1224 num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1225 iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1226 sizeof *vq->avail +
1227 num * sizeof(*vq->avail->ring) + s,
1228 VHOST_ADDR_AVAIL) &&
1229 iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1230 sizeof *vq->used +
1231 num * sizeof(*vq->used->ring) + s,
1232 VHOST_ADDR_USED);
1233 }
1234 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1235
1236 /* Can we log writes? */
1237 /* Caller should have device mutex but not vq mutex */
1238 int vhost_log_access_ok(struct vhost_dev *dev)
1239 {
1240 return memory_access_ok(dev, dev->umem, 1);
1241 }
1242 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1243
1244 /* Verify access for write logging. */
1245 /* Caller should have vq mutex and device mutex */
1246 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1247 void __user *log_base)
1248 {
1249 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1250
1251 return vq_memory_access_ok(log_base, vq->umem,
1252 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1253 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1254 sizeof *vq->used +
1255 vq->num * sizeof *vq->used->ring + s));
1256 }
1257
1258 /* Can we start vq? */
1259 /* Caller should have vq mutex and device mutex */
1260 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1261 {
1262 if (!vq_log_access_ok(vq, vq->log_base))
1263 return 0;
1264
1265 /* Access validation occurs at prefetch time with IOTLB */
1266 if (vq->iotlb)
1267 return 1;
1268
1269 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1270 }
1271 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1272
1273 static struct vhost_umem *vhost_umem_alloc(void)
1274 {
1275 struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1276
1277 if (!umem)
1278 return NULL;
1279
1280 umem->umem_tree = RB_ROOT_CACHED;
1281 umem->numem = 0;
1282 INIT_LIST_HEAD(&umem->umem_list);
1283
1284 return umem;
1285 }
1286
1287 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1288 {
1289 struct vhost_memory mem, *newmem;
1290 struct vhost_memory_region *region;
1291 struct vhost_umem *newumem, *oldumem;
1292 unsigned long size = offsetof(struct vhost_memory, regions);
1293 int i;
1294
1295 if (copy_from_user(&mem, m, size))
1296 return -EFAULT;
1297 if (mem.padding)
1298 return -EOPNOTSUPP;
1299 if (mem.nregions > max_mem_regions)
1300 return -E2BIG;
1301 newmem = kvzalloc(size + mem.nregions * sizeof(*m->regions), GFP_KERNEL);
1302 if (!newmem)
1303 return -ENOMEM;
1304
1305 memcpy(newmem, &mem, size);
1306 if (copy_from_user(newmem->regions, m->regions,
1307 mem.nregions * sizeof *m->regions)) {
1308 kvfree(newmem);
1309 return -EFAULT;
1310 }
1311
1312 newumem = vhost_umem_alloc();
1313 if (!newumem) {
1314 kvfree(newmem);
1315 return -ENOMEM;
1316 }
1317
1318 for (region = newmem->regions;
1319 region < newmem->regions + mem.nregions;
1320 region++) {
1321 if (vhost_new_umem_range(newumem,
1322 region->guest_phys_addr,
1323 region->memory_size,
1324 region->guest_phys_addr +
1325 region->memory_size - 1,
1326 region->userspace_addr,
1327 VHOST_ACCESS_RW))
1328 goto err;
1329 }
1330
1331 if (!memory_access_ok(d, newumem, 0))
1332 goto err;
1333
1334 oldumem = d->umem;
1335 d->umem = newumem;
1336
1337 /* All memory accesses are done under some VQ mutex. */
1338 for (i = 0; i < d->nvqs; ++i) {
1339 mutex_lock(&d->vqs[i]->mutex);
1340 d->vqs[i]->umem = newumem;
1341 mutex_unlock(&d->vqs[i]->mutex);
1342 }
1343
1344 kvfree(newmem);
1345 vhost_umem_clean(oldumem);
1346 return 0;
1347
1348 err:
1349 vhost_umem_clean(newumem);
1350 kvfree(newmem);
1351 return -EFAULT;
1352 }
1353
1354 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1355 {
1356 struct file *eventfp, *filep = NULL;
1357 bool pollstart = false, pollstop = false;
1358 struct eventfd_ctx *ctx = NULL;
1359 u32 __user *idxp = argp;
1360 struct vhost_virtqueue *vq;
1361 struct vhost_vring_state s;
1362 struct vhost_vring_file f;
1363 struct vhost_vring_addr a;
1364 u32 idx;
1365 long r;
1366
1367 r = get_user(idx, idxp);
1368 if (r < 0)
1369 return r;
1370 if (idx >= d->nvqs)
1371 return -ENOBUFS;
1372
1373 vq = d->vqs[idx];
1374
1375 mutex_lock(&vq->mutex);
1376
1377 switch (ioctl) {
1378 case VHOST_SET_VRING_NUM:
1379 /* Resizing ring with an active backend?
1380 * You don't want to do that. */
1381 if (vq->private_data) {
1382 r = -EBUSY;
1383 break;
1384 }
1385 if (copy_from_user(&s, argp, sizeof s)) {
1386 r = -EFAULT;
1387 break;
1388 }
1389 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1390 r = -EINVAL;
1391 break;
1392 }
1393 vq->num = s.num;
1394 break;
1395 case VHOST_SET_VRING_BASE:
1396 /* Moving base with an active backend?
1397 * You don't want to do that. */
1398 if (vq->private_data) {
1399 r = -EBUSY;
1400 break;
1401 }
1402 if (copy_from_user(&s, argp, sizeof s)) {
1403 r = -EFAULT;
1404 break;
1405 }
1406 if (s.num > 0xffff) {
1407 r = -EINVAL;
1408 break;
1409 }
1410 vq->last_avail_idx = s.num;
1411 /* Forget the cached index value. */
1412 vq->avail_idx = vq->last_avail_idx;
1413 break;
1414 case VHOST_GET_VRING_BASE:
1415 s.index = idx;
1416 s.num = vq->last_avail_idx;
1417 if (copy_to_user(argp, &s, sizeof s))
1418 r = -EFAULT;
1419 break;
1420 case VHOST_SET_VRING_ADDR:
1421 if (copy_from_user(&a, argp, sizeof a)) {
1422 r = -EFAULT;
1423 break;
1424 }
1425 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1426 r = -EOPNOTSUPP;
1427 break;
1428 }
1429 /* For 32bit, verify that the top 32bits of the user
1430 data are set to zero. */
1431 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1432 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1433 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1434 r = -EFAULT;
1435 break;
1436 }
1437
1438 /* Make sure it's safe to cast pointers to vring types. */
1439 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1440 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1441 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1442 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1443 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1444 r = -EINVAL;
1445 break;
1446 }
1447
1448 /* We only verify access here if backend is configured.
1449 * If it is not, we don't as size might not have been setup.
1450 * We will verify when backend is configured. */
1451 if (vq->private_data) {
1452 if (!vq_access_ok(vq, vq->num,
1453 (void __user *)(unsigned long)a.desc_user_addr,
1454 (void __user *)(unsigned long)a.avail_user_addr,
1455 (void __user *)(unsigned long)a.used_user_addr)) {
1456 r = -EINVAL;
1457 break;
1458 }
1459
1460 /* Also validate log access for used ring if enabled. */
1461 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1462 !log_access_ok(vq->log_base, a.log_guest_addr,
1463 sizeof *vq->used +
1464 vq->num * sizeof *vq->used->ring)) {
1465 r = -EINVAL;
1466 break;
1467 }
1468 }
1469
1470 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1471 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1472 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1473 vq->log_addr = a.log_guest_addr;
1474 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1475 break;
1476 case VHOST_SET_VRING_KICK:
1477 if (copy_from_user(&f, argp, sizeof f)) {
1478 r = -EFAULT;
1479 break;
1480 }
1481 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1482 if (IS_ERR(eventfp)) {
1483 r = PTR_ERR(eventfp);
1484 break;
1485 }
1486 if (eventfp != vq->kick) {
1487 pollstop = (filep = vq->kick) != NULL;
1488 pollstart = (vq->kick = eventfp) != NULL;
1489 } else
1490 filep = eventfp;
1491 break;
1492 case VHOST_SET_VRING_CALL:
1493 if (copy_from_user(&f, argp, sizeof f)) {
1494 r = -EFAULT;
1495 break;
1496 }
1497 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1498 if (IS_ERR(eventfp)) {
1499 r = PTR_ERR(eventfp);
1500 break;
1501 }
1502 if (eventfp != vq->call) {
1503 filep = vq->call;
1504 ctx = vq->call_ctx;
1505 vq->call = eventfp;
1506 vq->call_ctx = eventfp ?
1507 eventfd_ctx_fileget(eventfp) : NULL;
1508 } else
1509 filep = eventfp;
1510 break;
1511 case VHOST_SET_VRING_ERR:
1512 if (copy_from_user(&f, argp, sizeof f)) {
1513 r = -EFAULT;
1514 break;
1515 }
1516 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1517 if (IS_ERR(eventfp)) {
1518 r = PTR_ERR(eventfp);
1519 break;
1520 }
1521 if (eventfp != vq->error) {
1522 filep = vq->error;
1523 vq->error = eventfp;
1524 ctx = vq->error_ctx;
1525 vq->error_ctx = eventfp ?
1526 eventfd_ctx_fileget(eventfp) : NULL;
1527 } else
1528 filep = eventfp;
1529 break;
1530 case VHOST_SET_VRING_ENDIAN:
1531 r = vhost_set_vring_endian(vq, argp);
1532 break;
1533 case VHOST_GET_VRING_ENDIAN:
1534 r = vhost_get_vring_endian(vq, idx, argp);
1535 break;
1536 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1537 if (copy_from_user(&s, argp, sizeof(s))) {
1538 r = -EFAULT;
1539 break;
1540 }
1541 vq->busyloop_timeout = s.num;
1542 break;
1543 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1544 s.index = idx;
1545 s.num = vq->busyloop_timeout;
1546 if (copy_to_user(argp, &s, sizeof(s)))
1547 r = -EFAULT;
1548 break;
1549 default:
1550 r = -ENOIOCTLCMD;
1551 }
1552
1553 if (pollstop && vq->handle_kick)
1554 vhost_poll_stop(&vq->poll);
1555
1556 if (ctx)
1557 eventfd_ctx_put(ctx);
1558 if (filep)
1559 fput(filep);
1560
1561 if (pollstart && vq->handle_kick)
1562 r = vhost_poll_start(&vq->poll, vq->kick);
1563
1564 mutex_unlock(&vq->mutex);
1565
1566 if (pollstop && vq->handle_kick)
1567 vhost_poll_flush(&vq->poll);
1568 return r;
1569 }
1570 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1571
1572 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1573 {
1574 struct vhost_umem *niotlb, *oiotlb;
1575 int i;
1576
1577 niotlb = vhost_umem_alloc();
1578 if (!niotlb)
1579 return -ENOMEM;
1580
1581 oiotlb = d->iotlb;
1582 d->iotlb = niotlb;
1583
1584 for (i = 0; i < d->nvqs; ++i) {
1585 mutex_lock(&d->vqs[i]->mutex);
1586 d->vqs[i]->iotlb = niotlb;
1587 mutex_unlock(&d->vqs[i]->mutex);
1588 }
1589
1590 vhost_umem_clean(oiotlb);
1591
1592 return 0;
1593 }
1594 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1595
1596 /* Caller must have device mutex */
1597 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1598 {
1599 struct file *eventfp, *filep = NULL;
1600 struct eventfd_ctx *ctx = NULL;
1601 u64 p;
1602 long r;
1603 int i, fd;
1604
1605 /* If you are not the owner, you can become one */
1606 if (ioctl == VHOST_SET_OWNER) {
1607 r = vhost_dev_set_owner(d);
1608 goto done;
1609 }
1610
1611 /* You must be the owner to do anything else */
1612 r = vhost_dev_check_owner(d);
1613 if (r)
1614 goto done;
1615
1616 switch (ioctl) {
1617 case VHOST_SET_MEM_TABLE:
1618 r = vhost_set_memory(d, argp);
1619 break;
1620 case VHOST_SET_LOG_BASE:
1621 if (copy_from_user(&p, argp, sizeof p)) {
1622 r = -EFAULT;
1623 break;
1624 }
1625 if ((u64)(unsigned long)p != p) {
1626 r = -EFAULT;
1627 break;
1628 }
1629 for (i = 0; i < d->nvqs; ++i) {
1630 struct vhost_virtqueue *vq;
1631 void __user *base = (void __user *)(unsigned long)p;
1632 vq = d->vqs[i];
1633 mutex_lock(&vq->mutex);
1634 /* If ring is inactive, will check when it's enabled. */
1635 if (vq->private_data && !vq_log_access_ok(vq, base))
1636 r = -EFAULT;
1637 else
1638 vq->log_base = base;
1639 mutex_unlock(&vq->mutex);
1640 }
1641 break;
1642 case VHOST_SET_LOG_FD:
1643 r = get_user(fd, (int __user *)argp);
1644 if (r < 0)
1645 break;
1646 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1647 if (IS_ERR(eventfp)) {
1648 r = PTR_ERR(eventfp);
1649 break;
1650 }
1651 if (eventfp != d->log_file) {
1652 filep = d->log_file;
1653 d->log_file = eventfp;
1654 ctx = d->log_ctx;
1655 d->log_ctx = eventfp ?
1656 eventfd_ctx_fileget(eventfp) : NULL;
1657 } else
1658 filep = eventfp;
1659 for (i = 0; i < d->nvqs; ++i) {
1660 mutex_lock(&d->vqs[i]->mutex);
1661 d->vqs[i]->log_ctx = d->log_ctx;
1662 mutex_unlock(&d->vqs[i]->mutex);
1663 }
1664 if (ctx)
1665 eventfd_ctx_put(ctx);
1666 if (filep)
1667 fput(filep);
1668 break;
1669 default:
1670 r = -ENOIOCTLCMD;
1671 break;
1672 }
1673 done:
1674 return r;
1675 }
1676 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1677
1678 /* TODO: This is really inefficient. We need something like get_user()
1679 * (instruction directly accesses the data, with an exception table entry
1680 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1681 */
1682 static int set_bit_to_user(int nr, void __user *addr)
1683 {
1684 unsigned long log = (unsigned long)addr;
1685 struct page *page;
1686 void *base;
1687 int bit = nr + (log % PAGE_SIZE) * 8;
1688 int r;
1689
1690 r = get_user_pages_fast(log, 1, 1, &page);
1691 if (r < 0)
1692 return r;
1693 BUG_ON(r != 1);
1694 base = kmap_atomic(page);
1695 set_bit(bit, base);
1696 kunmap_atomic(base);
1697 set_page_dirty_lock(page);
1698 put_page(page);
1699 return 0;
1700 }
1701
1702 static int log_write(void __user *log_base,
1703 u64 write_address, u64 write_length)
1704 {
1705 u64 write_page = write_address / VHOST_PAGE_SIZE;
1706 int r;
1707
1708 if (!write_length)
1709 return 0;
1710 write_length += write_address % VHOST_PAGE_SIZE;
1711 for (;;) {
1712 u64 base = (u64)(unsigned long)log_base;
1713 u64 log = base + write_page / 8;
1714 int bit = write_page % 8;
1715 if ((u64)(unsigned long)log != log)
1716 return -EFAULT;
1717 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1718 if (r < 0)
1719 return r;
1720 if (write_length <= VHOST_PAGE_SIZE)
1721 break;
1722 write_length -= VHOST_PAGE_SIZE;
1723 write_page += 1;
1724 }
1725 return r;
1726 }
1727
1728 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1729 unsigned int log_num, u64 len)
1730 {
1731 int i, r;
1732
1733 /* Make sure data written is seen before log. */
1734 smp_wmb();
1735 for (i = 0; i < log_num; ++i) {
1736 u64 l = min(log[i].len, len);
1737 r = log_write(vq->log_base, log[i].addr, l);
1738 if (r < 0)
1739 return r;
1740 len -= l;
1741 if (!len) {
1742 if (vq->log_ctx)
1743 eventfd_signal(vq->log_ctx, 1);
1744 return 0;
1745 }
1746 }
1747 /* Length written exceeds what we have stored. This is a bug. */
1748 BUG();
1749 return 0;
1750 }
1751 EXPORT_SYMBOL_GPL(vhost_log_write);
1752
1753 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1754 {
1755 void __user *used;
1756 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1757 &vq->used->flags) < 0)
1758 return -EFAULT;
1759 if (unlikely(vq->log_used)) {
1760 /* Make sure the flag is seen before log. */
1761 smp_wmb();
1762 /* Log used flag write. */
1763 used = &vq->used->flags;
1764 log_write(vq->log_base, vq->log_addr +
1765 (used - (void __user *)vq->used),
1766 sizeof vq->used->flags);
1767 if (vq->log_ctx)
1768 eventfd_signal(vq->log_ctx, 1);
1769 }
1770 return 0;
1771 }
1772
1773 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1774 {
1775 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1776 vhost_avail_event(vq)))
1777 return -EFAULT;
1778 if (unlikely(vq->log_used)) {
1779 void __user *used;
1780 /* Make sure the event is seen before log. */
1781 smp_wmb();
1782 /* Log avail event write */
1783 used = vhost_avail_event(vq);
1784 log_write(vq->log_base, vq->log_addr +
1785 (used - (void __user *)vq->used),
1786 sizeof *vhost_avail_event(vq));
1787 if (vq->log_ctx)
1788 eventfd_signal(vq->log_ctx, 1);
1789 }
1790 return 0;
1791 }
1792
1793 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1794 {
1795 __virtio16 last_used_idx;
1796 int r;
1797 bool is_le = vq->is_le;
1798
1799 if (!vq->private_data)
1800 return 0;
1801
1802 vhost_init_is_le(vq);
1803
1804 r = vhost_update_used_flags(vq);
1805 if (r)
1806 goto err;
1807 vq->signalled_used_valid = false;
1808 if (!vq->iotlb &&
1809 !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1810 r = -EFAULT;
1811 goto err;
1812 }
1813 r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1814 if (r) {
1815 vq_err(vq, "Can't access used idx at %p\n",
1816 &vq->used->idx);
1817 goto err;
1818 }
1819 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1820 return 0;
1821
1822 err:
1823 vq->is_le = is_le;
1824 return r;
1825 }
1826 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1827
1828 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1829 struct iovec iov[], int iov_size, int access)
1830 {
1831 const struct vhost_umem_node *node;
1832 struct vhost_dev *dev = vq->dev;
1833 struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1834 struct iovec *_iov;
1835 u64 s = 0;
1836 int ret = 0;
1837
1838 while ((u64)len > s) {
1839 u64 size;
1840 if (unlikely(ret >= iov_size)) {
1841 ret = -ENOBUFS;
1842 break;
1843 }
1844
1845 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1846 addr, addr + len - 1);
1847 if (node == NULL || node->start > addr) {
1848 if (umem != dev->iotlb) {
1849 ret = -EFAULT;
1850 break;
1851 }
1852 ret = -EAGAIN;
1853 break;
1854 } else if (!(node->perm & access)) {
1855 ret = -EPERM;
1856 break;
1857 }
1858
1859 _iov = iov + ret;
1860 size = node->size - addr + node->start;
1861 _iov->iov_len = min((u64)len - s, size);
1862 _iov->iov_base = (void __user *)(unsigned long)
1863 (node->userspace_addr + addr - node->start);
1864 s += size;
1865 addr += size;
1866 ++ret;
1867 }
1868
1869 if (ret == -EAGAIN)
1870 vhost_iotlb_miss(vq, addr, access);
1871 return ret;
1872 }
1873
1874 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1875 * function returns the next descriptor in the chain,
1876 * or -1U if we're at the end. */
1877 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1878 {
1879 unsigned int next;
1880
1881 /* If this descriptor says it doesn't chain, we're done. */
1882 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1883 return -1U;
1884
1885 /* Check they're not leading us off end of descriptors. */
1886 next = vhost16_to_cpu(vq, desc->next);
1887 /* Make sure compiler knows to grab that: we don't want it changing! */
1888 /* We will use the result as an index in an array, so most
1889 * architectures only need a compiler barrier here. */
1890 read_barrier_depends();
1891
1892 return next;
1893 }
1894
1895 static int get_indirect(struct vhost_virtqueue *vq,
1896 struct iovec iov[], unsigned int iov_size,
1897 unsigned int *out_num, unsigned int *in_num,
1898 struct vhost_log *log, unsigned int *log_num,
1899 struct vring_desc *indirect)
1900 {
1901 struct vring_desc desc;
1902 unsigned int i = 0, count, found = 0;
1903 u32 len = vhost32_to_cpu(vq, indirect->len);
1904 struct iov_iter from;
1905 int ret, access;
1906
1907 /* Sanity check */
1908 if (unlikely(len % sizeof desc)) {
1909 vq_err(vq, "Invalid length in indirect descriptor: "
1910 "len 0x%llx not multiple of 0x%zx\n",
1911 (unsigned long long)len,
1912 sizeof desc);
1913 return -EINVAL;
1914 }
1915
1916 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1917 UIO_MAXIOV, VHOST_ACCESS_RO);
1918 if (unlikely(ret < 0)) {
1919 if (ret != -EAGAIN)
1920 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1921 return ret;
1922 }
1923 iov_iter_init(&from, READ, vq->indirect, ret, len);
1924
1925 /* We will use the result as an address to read from, so most
1926 * architectures only need a compiler barrier here. */
1927 read_barrier_depends();
1928
1929 count = len / sizeof desc;
1930 /* Buffers are chained via a 16 bit next field, so
1931 * we can have at most 2^16 of these. */
1932 if (unlikely(count > USHRT_MAX + 1)) {
1933 vq_err(vq, "Indirect buffer length too big: %d\n",
1934 indirect->len);
1935 return -E2BIG;
1936 }
1937
1938 do {
1939 unsigned iov_count = *in_num + *out_num;
1940 if (unlikely(++found > count)) {
1941 vq_err(vq, "Loop detected: last one at %u "
1942 "indirect size %u\n",
1943 i, count);
1944 return -EINVAL;
1945 }
1946 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1947 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1948 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1949 return -EINVAL;
1950 }
1951 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1952 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1953 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1954 return -EINVAL;
1955 }
1956
1957 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1958 access = VHOST_ACCESS_WO;
1959 else
1960 access = VHOST_ACCESS_RO;
1961
1962 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1963 vhost32_to_cpu(vq, desc.len), iov + iov_count,
1964 iov_size - iov_count, access);
1965 if (unlikely(ret < 0)) {
1966 if (ret != -EAGAIN)
1967 vq_err(vq, "Translation failure %d indirect idx %d\n",
1968 ret, i);
1969 return ret;
1970 }
1971 /* If this is an input descriptor, increment that count. */
1972 if (access == VHOST_ACCESS_WO) {
1973 *in_num += ret;
1974 if (unlikely(log)) {
1975 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1976 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1977 ++*log_num;
1978 }
1979 } else {
1980 /* If it's an output descriptor, they're all supposed
1981 * to come before any input descriptors. */
1982 if (unlikely(*in_num)) {
1983 vq_err(vq, "Indirect descriptor "
1984 "has out after in: idx %d\n", i);
1985 return -EINVAL;
1986 }
1987 *out_num += ret;
1988 }
1989 } while ((i = next_desc(vq, &desc)) != -1);
1990 return 0;
1991 }
1992
1993 /* This looks in the virtqueue and for the first available buffer, and converts
1994 * it to an iovec for convenient access. Since descriptors consist of some
1995 * number of output then some number of input descriptors, it's actually two
1996 * iovecs, but we pack them into one and note how many of each there were.
1997 *
1998 * This function returns the descriptor number found, or vq->num (which is
1999 * never a valid descriptor number) if none was found. A negative code is
2000 * returned on error. */
2001 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2002 struct iovec iov[], unsigned int iov_size,
2003 unsigned int *out_num, unsigned int *in_num,
2004 struct vhost_log *log, unsigned int *log_num)
2005 {
2006 struct vring_desc desc;
2007 unsigned int i, head, found = 0;
2008 u16 last_avail_idx;
2009 __virtio16 avail_idx;
2010 __virtio16 ring_head;
2011 int ret, access;
2012
2013 /* Check it isn't doing very strange things with descriptor numbers. */
2014 last_avail_idx = vq->last_avail_idx;
2015
2016 if (vq->avail_idx == vq->last_avail_idx) {
2017 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2018 vq_err(vq, "Failed to access avail idx at %p\n",
2019 &vq->avail->idx);
2020 return -EFAULT;
2021 }
2022 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2023
2024 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2025 vq_err(vq, "Guest moved used index from %u to %u",
2026 last_avail_idx, vq->avail_idx);
2027 return -EFAULT;
2028 }
2029
2030 /* If there's nothing new since last we looked, return
2031 * invalid.
2032 */
2033 if (vq->avail_idx == last_avail_idx)
2034 return vq->num;
2035
2036 /* Only get avail ring entries after they have been
2037 * exposed by guest.
2038 */
2039 smp_rmb();
2040 }
2041
2042 /* Grab the next descriptor number they're advertising, and increment
2043 * the index we've seen. */
2044 if (unlikely(vhost_get_avail(vq, ring_head,
2045 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2046 vq_err(vq, "Failed to read head: idx %d address %p\n",
2047 last_avail_idx,
2048 &vq->avail->ring[last_avail_idx % vq->num]);
2049 return -EFAULT;
2050 }
2051
2052 head = vhost16_to_cpu(vq, ring_head);
2053
2054 /* If their number is silly, that's an error. */
2055 if (unlikely(head >= vq->num)) {
2056 vq_err(vq, "Guest says index %u > %u is available",
2057 head, vq->num);
2058 return -EINVAL;
2059 }
2060
2061 /* When we start there are none of either input nor output. */
2062 *out_num = *in_num = 0;
2063 if (unlikely(log))
2064 *log_num = 0;
2065
2066 i = head;
2067 do {
2068 unsigned iov_count = *in_num + *out_num;
2069 if (unlikely(i >= vq->num)) {
2070 vq_err(vq, "Desc index is %u > %u, head = %u",
2071 i, vq->num, head);
2072 return -EINVAL;
2073 }
2074 if (unlikely(++found > vq->num)) {
2075 vq_err(vq, "Loop detected: last one at %u "
2076 "vq size %u head %u\n",
2077 i, vq->num, head);
2078 return -EINVAL;
2079 }
2080 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2081 sizeof desc);
2082 if (unlikely(ret)) {
2083 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2084 i, vq->desc + i);
2085 return -EFAULT;
2086 }
2087 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2088 ret = get_indirect(vq, iov, iov_size,
2089 out_num, in_num,
2090 log, log_num, &desc);
2091 if (unlikely(ret < 0)) {
2092 if (ret != -EAGAIN)
2093 vq_err(vq, "Failure detected "
2094 "in indirect descriptor at idx %d\n", i);
2095 return ret;
2096 }
2097 continue;
2098 }
2099
2100 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2101 access = VHOST_ACCESS_WO;
2102 else
2103 access = VHOST_ACCESS_RO;
2104 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2105 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2106 iov_size - iov_count, access);
2107 if (unlikely(ret < 0)) {
2108 if (ret != -EAGAIN)
2109 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2110 ret, i);
2111 return ret;
2112 }
2113 if (access == VHOST_ACCESS_WO) {
2114 /* If this is an input descriptor,
2115 * increment that count. */
2116 *in_num += ret;
2117 if (unlikely(log)) {
2118 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2119 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2120 ++*log_num;
2121 }
2122 } else {
2123 /* If it's an output descriptor, they're all supposed
2124 * to come before any input descriptors. */
2125 if (unlikely(*in_num)) {
2126 vq_err(vq, "Descriptor has out after in: "
2127 "idx %d\n", i);
2128 return -EINVAL;
2129 }
2130 *out_num += ret;
2131 }
2132 } while ((i = next_desc(vq, &desc)) != -1);
2133
2134 /* On success, increment avail index. */
2135 vq->last_avail_idx++;
2136
2137 /* Assume notifications from guest are disabled at this point,
2138 * if they aren't we would need to update avail_event index. */
2139 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2140 return head;
2141 }
2142 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2143
2144 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2145 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2146 {
2147 vq->last_avail_idx -= n;
2148 }
2149 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2150
2151 /* After we've used one of their buffers, we tell them about it. We'll then
2152 * want to notify the guest, using eventfd. */
2153 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2154 {
2155 struct vring_used_elem heads = {
2156 cpu_to_vhost32(vq, head),
2157 cpu_to_vhost32(vq, len)
2158 };
2159
2160 return vhost_add_used_n(vq, &heads, 1);
2161 }
2162 EXPORT_SYMBOL_GPL(vhost_add_used);
2163
2164 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2165 struct vring_used_elem *heads,
2166 unsigned count)
2167 {
2168 struct vring_used_elem __user *used;
2169 u16 old, new;
2170 int start;
2171
2172 start = vq->last_used_idx & (vq->num - 1);
2173 used = vq->used->ring + start;
2174 if (count == 1) {
2175 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2176 vq_err(vq, "Failed to write used id");
2177 return -EFAULT;
2178 }
2179 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2180 vq_err(vq, "Failed to write used len");
2181 return -EFAULT;
2182 }
2183 } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2184 vq_err(vq, "Failed to write used");
2185 return -EFAULT;
2186 }
2187 if (unlikely(vq->log_used)) {
2188 /* Make sure data is seen before log. */
2189 smp_wmb();
2190 /* Log used ring entry write. */
2191 log_write(vq->log_base,
2192 vq->log_addr +
2193 ((void __user *)used - (void __user *)vq->used),
2194 count * sizeof *used);
2195 }
2196 old = vq->last_used_idx;
2197 new = (vq->last_used_idx += count);
2198 /* If the driver never bothers to signal in a very long while,
2199 * used index might wrap around. If that happens, invalidate
2200 * signalled_used index we stored. TODO: make sure driver
2201 * signals at least once in 2^16 and remove this. */
2202 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2203 vq->signalled_used_valid = false;
2204 return 0;
2205 }
2206
2207 /* After we've used one of their buffers, we tell them about it. We'll then
2208 * want to notify the guest, using eventfd. */
2209 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2210 unsigned count)
2211 {
2212 int start, n, r;
2213
2214 start = vq->last_used_idx & (vq->num - 1);
2215 n = vq->num - start;
2216 if (n < count) {
2217 r = __vhost_add_used_n(vq, heads, n);
2218 if (r < 0)
2219 return r;
2220 heads += n;
2221 count -= n;
2222 }
2223 r = __vhost_add_used_n(vq, heads, count);
2224
2225 /* Make sure buffer is written before we update index. */
2226 smp_wmb();
2227 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2228 &vq->used->idx)) {
2229 vq_err(vq, "Failed to increment used idx");
2230 return -EFAULT;
2231 }
2232 if (unlikely(vq->log_used)) {
2233 /* Log used index update. */
2234 log_write(vq->log_base,
2235 vq->log_addr + offsetof(struct vring_used, idx),
2236 sizeof vq->used->idx);
2237 if (vq->log_ctx)
2238 eventfd_signal(vq->log_ctx, 1);
2239 }
2240 return r;
2241 }
2242 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2243
2244 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2245 {
2246 __u16 old, new;
2247 __virtio16 event;
2248 bool v;
2249 /* Flush out used index updates. This is paired
2250 * with the barrier that the Guest executes when enabling
2251 * interrupts. */
2252 smp_mb();
2253
2254 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2255 unlikely(vq->avail_idx == vq->last_avail_idx))
2256 return true;
2257
2258 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2259 __virtio16 flags;
2260 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2261 vq_err(vq, "Failed to get flags");
2262 return true;
2263 }
2264 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2265 }
2266 old = vq->signalled_used;
2267 v = vq->signalled_used_valid;
2268 new = vq->signalled_used = vq->last_used_idx;
2269 vq->signalled_used_valid = true;
2270
2271 if (unlikely(!v))
2272 return true;
2273
2274 if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2275 vq_err(vq, "Failed to get used event idx");
2276 return true;
2277 }
2278 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2279 }
2280
2281 /* This actually signals the guest, using eventfd. */
2282 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2283 {
2284 /* Signal the Guest tell them we used something up. */
2285 if (vq->call_ctx && vhost_notify(dev, vq))
2286 eventfd_signal(vq->call_ctx, 1);
2287 }
2288 EXPORT_SYMBOL_GPL(vhost_signal);
2289
2290 /* And here's the combo meal deal. Supersize me! */
2291 void vhost_add_used_and_signal(struct vhost_dev *dev,
2292 struct vhost_virtqueue *vq,
2293 unsigned int head, int len)
2294 {
2295 vhost_add_used(vq, head, len);
2296 vhost_signal(dev, vq);
2297 }
2298 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2299
2300 /* multi-buffer version of vhost_add_used_and_signal */
2301 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2302 struct vhost_virtqueue *vq,
2303 struct vring_used_elem *heads, unsigned count)
2304 {
2305 vhost_add_used_n(vq, heads, count);
2306 vhost_signal(dev, vq);
2307 }
2308 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2309
2310 /* return true if we're sure that avaiable ring is empty */
2311 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2312 {
2313 __virtio16 avail_idx;
2314 int r;
2315
2316 if (vq->avail_idx != vq->last_avail_idx)
2317 return false;
2318
2319 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2320 if (unlikely(r))
2321 return false;
2322 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2323
2324 return vq->avail_idx == vq->last_avail_idx;
2325 }
2326 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2327
2328 /* OK, now we need to know about added descriptors. */
2329 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2330 {
2331 __virtio16 avail_idx;
2332 int r;
2333
2334 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2335 return false;
2336 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2337 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2338 r = vhost_update_used_flags(vq);
2339 if (r) {
2340 vq_err(vq, "Failed to enable notification at %p: %d\n",
2341 &vq->used->flags, r);
2342 return false;
2343 }
2344 } else {
2345 r = vhost_update_avail_event(vq, vq->avail_idx);
2346 if (r) {
2347 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2348 vhost_avail_event(vq), r);
2349 return false;
2350 }
2351 }
2352 /* They could have slipped one in as we were doing that: make
2353 * sure it's written, then check again. */
2354 smp_mb();
2355 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2356 if (r) {
2357 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2358 &vq->avail->idx, r);
2359 return false;
2360 }
2361
2362 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2363 }
2364 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2365
2366 /* We don't need to be notified again. */
2367 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2368 {
2369 int r;
2370
2371 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2372 return;
2373 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2374 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2375 r = vhost_update_used_flags(vq);
2376 if (r)
2377 vq_err(vq, "Failed to enable notification at %p: %d\n",
2378 &vq->used->flags, r);
2379 }
2380 }
2381 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2382
2383 /* Create a new message. */
2384 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2385 {
2386 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2387 if (!node)
2388 return NULL;
2389
2390 /* Make sure all padding within the structure is initialized. */
2391 memset(&node->msg, 0, sizeof node->msg);
2392 node->vq = vq;
2393 node->msg.type = type;
2394 return node;
2395 }
2396 EXPORT_SYMBOL_GPL(vhost_new_msg);
2397
2398 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2399 struct vhost_msg_node *node)
2400 {
2401 spin_lock(&dev->iotlb_lock);
2402 list_add_tail(&node->node, head);
2403 spin_unlock(&dev->iotlb_lock);
2404
2405 wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
2406 }
2407 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2408
2409 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2410 struct list_head *head)
2411 {
2412 struct vhost_msg_node *node = NULL;
2413
2414 spin_lock(&dev->iotlb_lock);
2415 if (!list_empty(head)) {
2416 node = list_first_entry(head, struct vhost_msg_node,
2417 node);
2418 list_del(&node->node);
2419 }
2420 spin_unlock(&dev->iotlb_lock);
2421
2422 return node;
2423 }
2424 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2425
2426
2427 static int __init vhost_init(void)
2428 {
2429 return 0;
2430 }
2431
2432 static void __exit vhost_exit(void)
2433 {
2434 }
2435
2436 module_init(vhost_init);
2437 module_exit(vhost_exit);
2438
2439 MODULE_VERSION("0.0.1");
2440 MODULE_LICENSE("GPL v2");
2441 MODULE_AUTHOR("Michael S. Tsirkin");
2442 MODULE_DESCRIPTION("Host kernel accelerator for virtio");
ubuntu-bionic-kernel mirror