?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 4.14.51
[linux/fpc-iii.git] / drivers / vhost / vhost.c
blob50e48afd88ff67f1dd644223f53da53aa226cfc0
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 vhost_vq_meta_reset(dev);
1019 vhost_del_umem_range(dev->iotlb, msg->iova,
1020 msg->iova + msg->size - 1);
1021 break;
1022 default:
1023 ret = -EINVAL;
1024 break;
1025 }
1026
1027 vhost_dev_unlock_vqs(dev);
1028 mutex_unlock(&dev->mutex);
1029
1030 return ret;
1031 }
1032 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1033 struct iov_iter *from)
1034 {
1035 struct vhost_msg_node node;
1036 unsigned size = sizeof(struct vhost_msg);
1037 size_t ret;
1038 int err;
1039
1040 if (iov_iter_count(from) < size)
1041 return 0;
1042 ret = copy_from_iter(&node.msg, size, from);
1043 if (ret != size)
1044 goto done;
1045
1046 switch (node.msg.type) {
1047 case VHOST_IOTLB_MSG:
1048 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
1049 if (err)
1050 ret = err;
1051 break;
1052 default:
1053 ret = -EINVAL;
1054 break;
1055 }
1056
1057 done:
1058 return ret;
1059 }
1060 EXPORT_SYMBOL(vhost_chr_write_iter);
1061
1062 unsigned int vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1063 poll_table *wait)
1064 {
1065 unsigned int mask = 0;
1066
1067 poll_wait(file, &dev->wait, wait);
1068
1069 if (!list_empty(&dev->read_list))
1070 mask |= POLLIN | POLLRDNORM;
1071
1072 return mask;
1073 }
1074 EXPORT_SYMBOL(vhost_chr_poll);
1075
1076 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1077 int noblock)
1078 {
1079 DEFINE_WAIT(wait);
1080 struct vhost_msg_node *node;
1081 ssize_t ret = 0;
1082 unsigned size = sizeof(struct vhost_msg);
1083
1084 if (iov_iter_count(to) < size)
1085 return 0;
1086
1087 while (1) {
1088 if (!noblock)
1089 prepare_to_wait(&dev->wait, &wait,
1090 TASK_INTERRUPTIBLE);
1091
1092 node = vhost_dequeue_msg(dev, &dev->read_list);
1093 if (node)
1094 break;
1095 if (noblock) {
1096 ret = -EAGAIN;
1097 break;
1098 }
1099 if (signal_pending(current)) {
1100 ret = -ERESTARTSYS;
1101 break;
1102 }
1103 if (!dev->iotlb) {
1104 ret = -EBADFD;
1105 break;
1106 }
1107
1108 schedule();
1109 }
1110
1111 if (!noblock)
1112 finish_wait(&dev->wait, &wait);
1113
1114 if (node) {
1115 ret = copy_to_iter(&node->msg, size, to);
1116
1117 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1118 kfree(node);
1119 return ret;
1120 }
1121
1122 vhost_enqueue_msg(dev, &dev->pending_list, node);
1123 }
1124
1125 return ret;
1126 }
1127 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1128
1129 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1130 {
1131 struct vhost_dev *dev = vq->dev;
1132 struct vhost_msg_node *node;
1133 struct vhost_iotlb_msg *msg;
1134
1135 node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1136 if (!node)
1137 return -ENOMEM;
1138
1139 msg = &node->msg.iotlb;
1140 msg->type = VHOST_IOTLB_MISS;
1141 msg->iova = iova;
1142 msg->perm = access;
1143
1144 vhost_enqueue_msg(dev, &dev->read_list, node);
1145
1146 return 0;
1147 }
1148
1149 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1150 struct vring_desc __user *desc,
1151 struct vring_avail __user *avail,
1152 struct vring_used __user *used)
1153
1154 {
1155 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1156
1157 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1158 access_ok(VERIFY_READ, avail,
1159 sizeof *avail + num * sizeof *avail->ring + s) &&
1160 access_ok(VERIFY_WRITE, used,
1161 sizeof *used + num * sizeof *used->ring + s);
1162 }
1163
1164 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1165 const struct vhost_umem_node *node,
1166 int type)
1167 {
1168 int access = (type == VHOST_ADDR_USED) ?
1169 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1170
1171 if (likely(node->perm & access))
1172 vq->meta_iotlb[type] = node;
1173 }
1174
1175 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1176 int access, u64 addr, u64 len, int type)
1177 {
1178 const struct vhost_umem_node *node;
1179 struct vhost_umem *umem = vq->iotlb;
1180 u64 s = 0, size, orig_addr = addr;
1181
1182 if (vhost_vq_meta_fetch(vq, addr, len, type))
1183 return true;
1184
1185 while (len > s) {
1186 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1187 addr,
1188 addr + len - 1);
1189 if (node == NULL || node->start > addr) {
1190 vhost_iotlb_miss(vq, addr, access);
1191 return false;
1192 } else if (!(node->perm & access)) {
1193 /* Report the possible access violation by
1194 * request another translation from userspace.
1195 */
1196 return false;
1197 }
1198
1199 size = node->size - addr + node->start;
1200
1201 if (orig_addr == addr && size >= len)
1202 vhost_vq_meta_update(vq, node, type);
1203
1204 s += size;
1205 addr += size;
1206 }
1207
1208 return true;
1209 }
1210
1211 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1212 {
1213 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1214 unsigned int num = vq->num;
1215
1216 if (!vq->iotlb)
1217 return 1;
1218
1219 return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1220 num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1221 iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1222 sizeof *vq->avail +
1223 num * sizeof(*vq->avail->ring) + s,
1224 VHOST_ADDR_AVAIL) &&
1225 iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1226 sizeof *vq->used +
1227 num * sizeof(*vq->used->ring) + s,
1228 VHOST_ADDR_USED);
1229 }
1230 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1231
1232 /* Can we log writes? */
1233 /* Caller should have device mutex but not vq mutex */
1234 int vhost_log_access_ok(struct vhost_dev *dev)
1235 {
1236 return memory_access_ok(dev, dev->umem, 1);
1237 }
1238 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1239
1240 /* Verify access for write logging. */
1241 /* Caller should have vq mutex and device mutex */
1242 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1243 void __user *log_base)
1244 {
1245 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1246
1247 return vq_memory_access_ok(log_base, vq->umem,
1248 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1249 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1250 sizeof *vq->used +
1251 vq->num * sizeof *vq->used->ring + s));
1252 }
1253
1254 /* Can we start vq? */
1255 /* Caller should have vq mutex and device mutex */
1256 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1257 {
1258 if (!vq_log_access_ok(vq, vq->log_base))
1259 return 0;
1260
1261 /* Access validation occurs at prefetch time with IOTLB */
1262 if (vq->iotlb)
1263 return 1;
1264
1265 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1266 }
1267 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1268
1269 static struct vhost_umem *vhost_umem_alloc(void)
1270 {
1271 struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1272
1273 if (!umem)
1274 return NULL;
1275
1276 umem->umem_tree = RB_ROOT_CACHED;
1277 umem->numem = 0;
1278 INIT_LIST_HEAD(&umem->umem_list);
1279
1280 return umem;
1281 }
1282
1283 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1284 {
1285 struct vhost_memory mem, *newmem;
1286 struct vhost_memory_region *region;
1287 struct vhost_umem *newumem, *oldumem;
1288 unsigned long size = offsetof(struct vhost_memory, regions);
1289 int i;
1290
1291 if (copy_from_user(&mem, m, size))
1292 return -EFAULT;
1293 if (mem.padding)
1294 return -EOPNOTSUPP;
1295 if (mem.nregions > max_mem_regions)
1296 return -E2BIG;
1297 newmem = kvzalloc(size + mem.nregions * sizeof(*m->regions), GFP_KERNEL);
1298 if (!newmem)
1299 return -ENOMEM;
1300
1301 memcpy(newmem, &mem, size);
1302 if (copy_from_user(newmem->regions, m->regions,
1303 mem.nregions * sizeof *m->regions)) {
1304 kvfree(newmem);
1305 return -EFAULT;
1306 }
1307
1308 newumem = vhost_umem_alloc();
1309 if (!newumem) {
1310 kvfree(newmem);
1311 return -ENOMEM;
1312 }
1313
1314 for (region = newmem->regions;
1315 region < newmem->regions + mem.nregions;
1316 region++) {
1317 if (vhost_new_umem_range(newumem,
1318 region->guest_phys_addr,
1319 region->memory_size,
1320 region->guest_phys_addr +
1321 region->memory_size - 1,
1322 region->userspace_addr,
1323 VHOST_ACCESS_RW))
1324 goto err;
1325 }
1326
1327 if (!memory_access_ok(d, newumem, 0))
1328 goto err;
1329
1330 oldumem = d->umem;
1331 d->umem = newumem;
1332
1333 /* All memory accesses are done under some VQ mutex. */
1334 for (i = 0; i < d->nvqs; ++i) {
1335 mutex_lock(&d->vqs[i]->mutex);
1336 d->vqs[i]->umem = newumem;
1337 mutex_unlock(&d->vqs[i]->mutex);
1338 }
1339
1340 kvfree(newmem);
1341 vhost_umem_clean(oldumem);
1342 return 0;
1343
1344 err:
1345 vhost_umem_clean(newumem);
1346 kvfree(newmem);
1347 return -EFAULT;
1348 }
1349
1350 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1351 {
1352 struct file *eventfp, *filep = NULL;
1353 bool pollstart = false, pollstop = false;
1354 struct eventfd_ctx *ctx = NULL;
1355 u32 __user *idxp = argp;
1356 struct vhost_virtqueue *vq;
1357 struct vhost_vring_state s;
1358 struct vhost_vring_file f;
1359 struct vhost_vring_addr a;
1360 u32 idx;
1361 long r;
1362
1363 r = get_user(idx, idxp);
1364 if (r < 0)
1365 return r;
1366 if (idx >= d->nvqs)
1367 return -ENOBUFS;
1368
1369 vq = d->vqs[idx];
1370
1371 mutex_lock(&vq->mutex);
1372
1373 switch (ioctl) {
1374 case VHOST_SET_VRING_NUM:
1375 /* Resizing ring with an active backend?
1376 * You don't want to do that. */
1377 if (vq->private_data) {
1378 r = -EBUSY;
1379 break;
1380 }
1381 if (copy_from_user(&s, argp, sizeof s)) {
1382 r = -EFAULT;
1383 break;
1384 }
1385 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1386 r = -EINVAL;
1387 break;
1388 }
1389 vq->num = s.num;
1390 break;
1391 case VHOST_SET_VRING_BASE:
1392 /* Moving base with an active backend?
1393 * You don't want to do that. */
1394 if (vq->private_data) {
1395 r = -EBUSY;
1396 break;
1397 }
1398 if (copy_from_user(&s, argp, sizeof s)) {
1399 r = -EFAULT;
1400 break;
1401 }
1402 if (s.num > 0xffff) {
1403 r = -EINVAL;
1404 break;
1405 }
1406 vq->last_avail_idx = s.num;
1407 /* Forget the cached index value. */
1408 vq->avail_idx = vq->last_avail_idx;
1409 break;
1410 case VHOST_GET_VRING_BASE:
1411 s.index = idx;
1412 s.num = vq->last_avail_idx;
1413 if (copy_to_user(argp, &s, sizeof s))
1414 r = -EFAULT;
1415 break;
1416 case VHOST_SET_VRING_ADDR:
1417 if (copy_from_user(&a, argp, sizeof a)) {
1418 r = -EFAULT;
1419 break;
1420 }
1421 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1422 r = -EOPNOTSUPP;
1423 break;
1424 }
1425 /* For 32bit, verify that the top 32bits of the user
1426 data are set to zero. */
1427 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1428 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1429 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1430 r = -EFAULT;
1431 break;
1432 }
1433
1434 /* Make sure it's safe to cast pointers to vring types. */
1435 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1436 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1437 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1438 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1439 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1440 r = -EINVAL;
1441 break;
1442 }
1443
1444 /* We only verify access here if backend is configured.
1445 * If it is not, we don't as size might not have been setup.
1446 * We will verify when backend is configured. */
1447 if (vq->private_data) {
1448 if (!vq_access_ok(vq, vq->num,
1449 (void __user *)(unsigned long)a.desc_user_addr,
1450 (void __user *)(unsigned long)a.avail_user_addr,
1451 (void __user *)(unsigned long)a.used_user_addr)) {
1452 r = -EINVAL;
1453 break;
1454 }
1455
1456 /* Also validate log access for used ring if enabled. */
1457 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1458 !log_access_ok(vq->log_base, a.log_guest_addr,
1459 sizeof *vq->used +
1460 vq->num * sizeof *vq->used->ring)) {
1461 r = -EINVAL;
1462 break;
1463 }
1464 }
1465
1466 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1467 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1468 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1469 vq->log_addr = a.log_guest_addr;
1470 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1471 break;
1472 case VHOST_SET_VRING_KICK:
1473 if (copy_from_user(&f, argp, sizeof f)) {
1474 r = -EFAULT;
1475 break;
1476 }
1477 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1478 if (IS_ERR(eventfp)) {
1479 r = PTR_ERR(eventfp);
1480 break;
1481 }
1482 if (eventfp != vq->kick) {
1483 pollstop = (filep = vq->kick) != NULL;
1484 pollstart = (vq->kick = eventfp) != NULL;
1485 } else
1486 filep = eventfp;
1487 break;
1488 case VHOST_SET_VRING_CALL:
1489 if (copy_from_user(&f, argp, sizeof f)) {
1490 r = -EFAULT;
1491 break;
1492 }
1493 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1494 if (IS_ERR(eventfp)) {
1495 r = PTR_ERR(eventfp);
1496 break;
1497 }
1498 if (eventfp != vq->call) {
1499 filep = vq->call;
1500 ctx = vq->call_ctx;
1501 vq->call = eventfp;
1502 vq->call_ctx = eventfp ?
1503 eventfd_ctx_fileget(eventfp) : NULL;
1504 } else
1505 filep = eventfp;
1506 break;
1507 case VHOST_SET_VRING_ERR:
1508 if (copy_from_user(&f, argp, sizeof f)) {
1509 r = -EFAULT;
1510 break;
1511 }
1512 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1513 if (IS_ERR(eventfp)) {
1514 r = PTR_ERR(eventfp);
1515 break;
1516 }
1517 if (eventfp != vq->error) {
1518 filep = vq->error;
1519 vq->error = eventfp;
1520 ctx = vq->error_ctx;
1521 vq->error_ctx = eventfp ?
1522 eventfd_ctx_fileget(eventfp) : NULL;
1523 } else
1524 filep = eventfp;
1525 break;
1526 case VHOST_SET_VRING_ENDIAN:
1527 r = vhost_set_vring_endian(vq, argp);
1528 break;
1529 case VHOST_GET_VRING_ENDIAN:
1530 r = vhost_get_vring_endian(vq, idx, argp);
1531 break;
1532 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1533 if (copy_from_user(&s, argp, sizeof(s))) {
1534 r = -EFAULT;
1535 break;
1536 }
1537 vq->busyloop_timeout = s.num;
1538 break;
1539 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1540 s.index = idx;
1541 s.num = vq->busyloop_timeout;
1542 if (copy_to_user(argp, &s, sizeof(s)))
1543 r = -EFAULT;
1544 break;
1545 default:
1546 r = -ENOIOCTLCMD;
1547 }
1548
1549 if (pollstop && vq->handle_kick)
1550 vhost_poll_stop(&vq->poll);
1551
1552 if (ctx)
1553 eventfd_ctx_put(ctx);
1554 if (filep)
1555 fput(filep);
1556
1557 if (pollstart && vq->handle_kick)
1558 r = vhost_poll_start(&vq->poll, vq->kick);
1559
1560 mutex_unlock(&vq->mutex);
1561
1562 if (pollstop && vq->handle_kick)
1563 vhost_poll_flush(&vq->poll);
1564 return r;
1565 }
1566 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1567
1568 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1569 {
1570 struct vhost_umem *niotlb, *oiotlb;
1571 int i;
1572
1573 niotlb = vhost_umem_alloc();
1574 if (!niotlb)
1575 return -ENOMEM;
1576
1577 oiotlb = d->iotlb;
1578 d->iotlb = niotlb;
1579
1580 for (i = 0; i < d->nvqs; ++i) {
1581 mutex_lock(&d->vqs[i]->mutex);
1582 d->vqs[i]->iotlb = niotlb;
1583 mutex_unlock(&d->vqs[i]->mutex);
1584 }
1585
1586 vhost_umem_clean(oiotlb);
1587
1588 return 0;
1589 }
1590 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1591
1592 /* Caller must have device mutex */
1593 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1594 {
1595 struct file *eventfp, *filep = NULL;
1596 struct eventfd_ctx *ctx = NULL;
1597 u64 p;
1598 long r;
1599 int i, fd;
1600
1601 /* If you are not the owner, you can become one */
1602 if (ioctl == VHOST_SET_OWNER) {
1603 r = vhost_dev_set_owner(d);
1604 goto done;
1605 }
1606
1607 /* You must be the owner to do anything else */
1608 r = vhost_dev_check_owner(d);
1609 if (r)
1610 goto done;
1611
1612 switch (ioctl) {
1613 case VHOST_SET_MEM_TABLE:
1614 r = vhost_set_memory(d, argp);
1615 break;
1616 case VHOST_SET_LOG_BASE:
1617 if (copy_from_user(&p, argp, sizeof p)) {
1618 r = -EFAULT;
1619 break;
1620 }
1621 if ((u64)(unsigned long)p != p) {
1622 r = -EFAULT;
1623 break;
1624 }
1625 for (i = 0; i < d->nvqs; ++i) {
1626 struct vhost_virtqueue *vq;
1627 void __user *base = (void __user *)(unsigned long)p;
1628 vq = d->vqs[i];
1629 mutex_lock(&vq->mutex);
1630 /* If ring is inactive, will check when it's enabled. */
1631 if (vq->private_data && !vq_log_access_ok(vq, base))
1632 r = -EFAULT;
1633 else
1634 vq->log_base = base;
1635 mutex_unlock(&vq->mutex);
1636 }
1637 break;
1638 case VHOST_SET_LOG_FD:
1639 r = get_user(fd, (int __user *)argp);
1640 if (r < 0)
1641 break;
1642 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1643 if (IS_ERR(eventfp)) {
1644 r = PTR_ERR(eventfp);
1645 break;
1646 }
1647 if (eventfp != d->log_file) {
1648 filep = d->log_file;
1649 d->log_file = eventfp;
1650 ctx = d->log_ctx;
1651 d->log_ctx = eventfp ?
1652 eventfd_ctx_fileget(eventfp) : NULL;
1653 } else
1654 filep = eventfp;
1655 for (i = 0; i < d->nvqs; ++i) {
1656 mutex_lock(&d->vqs[i]->mutex);
1657 d->vqs[i]->log_ctx = d->log_ctx;
1658 mutex_unlock(&d->vqs[i]->mutex);
1659 }
1660 if (ctx)
1661 eventfd_ctx_put(ctx);
1662 if (filep)
1663 fput(filep);
1664 break;
1665 default:
1666 r = -ENOIOCTLCMD;
1667 break;
1668 }
1669 done:
1670 return r;
1671 }
1672 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1673
1674 /* TODO: This is really inefficient. We need something like get_user()
1675 * (instruction directly accesses the data, with an exception table entry
1676 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1677 */
1678 static int set_bit_to_user(int nr, void __user *addr)
1679 {
1680 unsigned long log = (unsigned long)addr;
1681 struct page *page;
1682 void *base;
1683 int bit = nr + (log % PAGE_SIZE) * 8;
1684 int r;
1685
1686 r = get_user_pages_fast(log, 1, 1, &page);
1687 if (r < 0)
1688 return r;
1689 BUG_ON(r != 1);
1690 base = kmap_atomic(page);
1691 set_bit(bit, base);
1692 kunmap_atomic(base);
1693 set_page_dirty_lock(page);
1694 put_page(page);
1695 return 0;
1696 }
1697
1698 static int log_write(void __user *log_base,
1699 u64 write_address, u64 write_length)
1700 {
1701 u64 write_page = write_address / VHOST_PAGE_SIZE;
1702 int r;
1703
1704 if (!write_length)
1705 return 0;
1706 write_length += write_address % VHOST_PAGE_SIZE;
1707 for (;;) {
1708 u64 base = (u64)(unsigned long)log_base;
1709 u64 log = base + write_page / 8;
1710 int bit = write_page % 8;
1711 if ((u64)(unsigned long)log != log)
1712 return -EFAULT;
1713 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1714 if (r < 0)
1715 return r;
1716 if (write_length <= VHOST_PAGE_SIZE)
1717 break;
1718 write_length -= VHOST_PAGE_SIZE;
1719 write_page += 1;
1720 }
1721 return r;
1722 }
1723
1724 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1725 unsigned int log_num, u64 len)
1726 {
1727 int i, r;
1728
1729 /* Make sure data written is seen before log. */
1730 smp_wmb();
1731 for (i = 0; i < log_num; ++i) {
1732 u64 l = min(log[i].len, len);
1733 r = log_write(vq->log_base, log[i].addr, l);
1734 if (r < 0)
1735 return r;
1736 len -= l;
1737 if (!len) {
1738 if (vq->log_ctx)
1739 eventfd_signal(vq->log_ctx, 1);
1740 return 0;
1741 }
1742 }
1743 /* Length written exceeds what we have stored. This is a bug. */
1744 BUG();
1745 return 0;
1746 }
1747 EXPORT_SYMBOL_GPL(vhost_log_write);
1748
1749 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1750 {
1751 void __user *used;
1752 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1753 &vq->used->flags) < 0)
1754 return -EFAULT;
1755 if (unlikely(vq->log_used)) {
1756 /* Make sure the flag is seen before log. */
1757 smp_wmb();
1758 /* Log used flag write. */
1759 used = &vq->used->flags;
1760 log_write(vq->log_base, vq->log_addr +
1761 (used - (void __user *)vq->used),
1762 sizeof vq->used->flags);
1763 if (vq->log_ctx)
1764 eventfd_signal(vq->log_ctx, 1);
1765 }
1766 return 0;
1767 }
1768
1769 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1770 {
1771 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1772 vhost_avail_event(vq)))
1773 return -EFAULT;
1774 if (unlikely(vq->log_used)) {
1775 void __user *used;
1776 /* Make sure the event is seen before log. */
1777 smp_wmb();
1778 /* Log avail event write */
1779 used = vhost_avail_event(vq);
1780 log_write(vq->log_base, vq->log_addr +
1781 (used - (void __user *)vq->used),
1782 sizeof *vhost_avail_event(vq));
1783 if (vq->log_ctx)
1784 eventfd_signal(vq->log_ctx, 1);
1785 }
1786 return 0;
1787 }
1788
1789 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1790 {
1791 __virtio16 last_used_idx;
1792 int r;
1793 bool is_le = vq->is_le;
1794
1795 if (!vq->private_data)
1796 return 0;
1797
1798 vhost_init_is_le(vq);
1799
1800 r = vhost_update_used_flags(vq);
1801 if (r)
1802 goto err;
1803 vq->signalled_used_valid = false;
1804 if (!vq->iotlb &&
1805 !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1806 r = -EFAULT;
1807 goto err;
1808 }
1809 r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1810 if (r) {
1811 vq_err(vq, "Can't access used idx at %p\n",
1812 &vq->used->idx);
1813 goto err;
1814 }
1815 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1816 return 0;
1817
1818 err:
1819 vq->is_le = is_le;
1820 return r;
1821 }
1822 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1823
1824 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1825 struct iovec iov[], int iov_size, int access)
1826 {
1827 const struct vhost_umem_node *node;
1828 struct vhost_dev *dev = vq->dev;
1829 struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1830 struct iovec *_iov;
1831 u64 s = 0;
1832 int ret = 0;
1833
1834 while ((u64)len > s) {
1835 u64 size;
1836 if (unlikely(ret >= iov_size)) {
1837 ret = -ENOBUFS;
1838 break;
1839 }
1840
1841 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1842 addr, addr + len - 1);
1843 if (node == NULL || node->start > addr) {
1844 if (umem != dev->iotlb) {
1845 ret = -EFAULT;
1846 break;
1847 }
1848 ret = -EAGAIN;
1849 break;
1850 } else if (!(node->perm & access)) {
1851 ret = -EPERM;
1852 break;
1853 }
1854
1855 _iov = iov + ret;
1856 size = node->size - addr + node->start;
1857 _iov->iov_len = min((u64)len - s, size);
1858 _iov->iov_base = (void __user *)(unsigned long)
1859 (node->userspace_addr + addr - node->start);
1860 s += size;
1861 addr += size;
1862 ++ret;
1863 }
1864
1865 if (ret == -EAGAIN)
1866 vhost_iotlb_miss(vq, addr, access);
1867 return ret;
1868 }
1869
1870 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1871 * function returns the next descriptor in the chain,
1872 * or -1U if we're at the end. */
1873 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1874 {
1875 unsigned int next;
1876
1877 /* If this descriptor says it doesn't chain, we're done. */
1878 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1879 return -1U;
1880
1881 /* Check they're not leading us off end of descriptors. */
1882 next = vhost16_to_cpu(vq, desc->next);
1883 /* Make sure compiler knows to grab that: we don't want it changing! */
1884 /* We will use the result as an index in an array, so most
1885 * architectures only need a compiler barrier here. */
1886 read_barrier_depends();
1887
1888 return next;
1889 }
1890
1891 static int get_indirect(struct vhost_virtqueue *vq,
1892 struct iovec iov[], unsigned int iov_size,
1893 unsigned int *out_num, unsigned int *in_num,
1894 struct vhost_log *log, unsigned int *log_num,
1895 struct vring_desc *indirect)
1896 {
1897 struct vring_desc desc;
1898 unsigned int i = 0, count, found = 0;
1899 u32 len = vhost32_to_cpu(vq, indirect->len);
1900 struct iov_iter from;
1901 int ret, access;
1902
1903 /* Sanity check */
1904 if (unlikely(len % sizeof desc)) {
1905 vq_err(vq, "Invalid length in indirect descriptor: "
1906 "len 0x%llx not multiple of 0x%zx\n",
1907 (unsigned long long)len,
1908 sizeof desc);
1909 return -EINVAL;
1910 }
1911
1912 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1913 UIO_MAXIOV, VHOST_ACCESS_RO);
1914 if (unlikely(ret < 0)) {
1915 if (ret != -EAGAIN)
1916 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1917 return ret;
1918 }
1919 iov_iter_init(&from, READ, vq->indirect, ret, len);
1920
1921 /* We will use the result as an address to read from, so most
1922 * architectures only need a compiler barrier here. */
1923 read_barrier_depends();
1924
1925 count = len / sizeof desc;
1926 /* Buffers are chained via a 16 bit next field, so
1927 * we can have at most 2^16 of these. */
1928 if (unlikely(count > USHRT_MAX + 1)) {
1929 vq_err(vq, "Indirect buffer length too big: %d\n",
1930 indirect->len);
1931 return -E2BIG;
1932 }
1933
1934 do {
1935 unsigned iov_count = *in_num + *out_num;
1936 if (unlikely(++found > count)) {
1937 vq_err(vq, "Loop detected: last one at %u "
1938 "indirect size %u\n",
1939 i, count);
1940 return -EINVAL;
1941 }
1942 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1943 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1944 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1945 return -EINVAL;
1946 }
1947 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1948 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1949 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1950 return -EINVAL;
1951 }
1952
1953 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1954 access = VHOST_ACCESS_WO;
1955 else
1956 access = VHOST_ACCESS_RO;
1957
1958 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1959 vhost32_to_cpu(vq, desc.len), iov + iov_count,
1960 iov_size - iov_count, access);
1961 if (unlikely(ret < 0)) {
1962 if (ret != -EAGAIN)
1963 vq_err(vq, "Translation failure %d indirect idx %d\n",
1964 ret, i);
1965 return ret;
1966 }
1967 /* If this is an input descriptor, increment that count. */
1968 if (access == VHOST_ACCESS_WO) {
1969 *in_num += ret;
1970 if (unlikely(log)) {
1971 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1972 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1973 ++*log_num;
1974 }
1975 } else {
1976 /* If it's an output descriptor, they're all supposed
1977 * to come before any input descriptors. */
1978 if (unlikely(*in_num)) {
1979 vq_err(vq, "Indirect descriptor "
1980 "has out after in: idx %d\n", i);
1981 return -EINVAL;
1982 }
1983 *out_num += ret;
1984 }
1985 } while ((i = next_desc(vq, &desc)) != -1);
1986 return 0;
1987 }
1988
1989 /* This looks in the virtqueue and for the first available buffer, and converts
1990 * it to an iovec for convenient access. Since descriptors consist of some
1991 * number of output then some number of input descriptors, it's actually two
1992 * iovecs, but we pack them into one and note how many of each there were.
1993 *
1994 * This function returns the descriptor number found, or vq->num (which is
1995 * never a valid descriptor number) if none was found. A negative code is
1996 * returned on error. */
1997 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1998 struct iovec iov[], unsigned int iov_size,
1999 unsigned int *out_num, unsigned int *in_num,
2000 struct vhost_log *log, unsigned int *log_num)
2001 {
2002 struct vring_desc desc;
2003 unsigned int i, head, found = 0;
2004 u16 last_avail_idx;
2005 __virtio16 avail_idx;
2006 __virtio16 ring_head;
2007 int ret, access;
2008
2009 /* Check it isn't doing very strange things with descriptor numbers. */
2010 last_avail_idx = vq->last_avail_idx;
2011
2012 if (vq->avail_idx == vq->last_avail_idx) {
2013 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2014 vq_err(vq, "Failed to access avail idx at %p\n",
2015 &vq->avail->idx);
2016 return -EFAULT;
2017 }
2018 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2019
2020 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2021 vq_err(vq, "Guest moved used index from %u to %u",
2022 last_avail_idx, vq->avail_idx);
2023 return -EFAULT;
2024 }
2025
2026 /* If there's nothing new since last we looked, return
2027 * invalid.
2028 */
2029 if (vq->avail_idx == last_avail_idx)
2030 return vq->num;
2031
2032 /* Only get avail ring entries after they have been
2033 * exposed by guest.
2034 */
2035 smp_rmb();
2036 }
2037
2038 /* Grab the next descriptor number they're advertising, and increment
2039 * the index we've seen. */
2040 if (unlikely(vhost_get_avail(vq, ring_head,
2041 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2042 vq_err(vq, "Failed to read head: idx %d address %p\n",
2043 last_avail_idx,
2044 &vq->avail->ring[last_avail_idx % vq->num]);
2045 return -EFAULT;
2046 }
2047
2048 head = vhost16_to_cpu(vq, ring_head);
2049
2050 /* If their number is silly, that's an error. */
2051 if (unlikely(head >= vq->num)) {
2052 vq_err(vq, "Guest says index %u > %u is available",
2053 head, vq->num);
2054 return -EINVAL;
2055 }
2056
2057 /* When we start there are none of either input nor output. */
2058 *out_num = *in_num = 0;
2059 if (unlikely(log))
2060 *log_num = 0;
2061
2062 i = head;
2063 do {
2064 unsigned iov_count = *in_num + *out_num;
2065 if (unlikely(i >= vq->num)) {
2066 vq_err(vq, "Desc index is %u > %u, head = %u",
2067 i, vq->num, head);
2068 return -EINVAL;
2069 }
2070 if (unlikely(++found > vq->num)) {
2071 vq_err(vq, "Loop detected: last one at %u "
2072 "vq size %u head %u\n",
2073 i, vq->num, head);
2074 return -EINVAL;
2075 }
2076 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2077 sizeof desc);
2078 if (unlikely(ret)) {
2079 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2080 i, vq->desc + i);
2081 return -EFAULT;
2082 }
2083 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2084 ret = get_indirect(vq, iov, iov_size,
2085 out_num, in_num,
2086 log, log_num, &desc);
2087 if (unlikely(ret < 0)) {
2088 if (ret != -EAGAIN)
2089 vq_err(vq, "Failure detected "
2090 "in indirect descriptor at idx %d\n", i);
2091 return ret;
2092 }
2093 continue;
2094 }
2095
2096 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2097 access = VHOST_ACCESS_WO;
2098 else
2099 access = VHOST_ACCESS_RO;
2100 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2101 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2102 iov_size - iov_count, access);
2103 if (unlikely(ret < 0)) {
2104 if (ret != -EAGAIN)
2105 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2106 ret, i);
2107 return ret;
2108 }
2109 if (access == VHOST_ACCESS_WO) {
2110 /* If this is an input descriptor,
2111 * increment that count. */
2112 *in_num += ret;
2113 if (unlikely(log)) {
2114 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2115 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2116 ++*log_num;
2117 }
2118 } else {
2119 /* If it's an output descriptor, they're all supposed
2120 * to come before any input descriptors. */
2121 if (unlikely(*in_num)) {
2122 vq_err(vq, "Descriptor has out after in: "
2123 "idx %d\n", i);
2124 return -EINVAL;
2125 }
2126 *out_num += ret;
2127 }
2128 } while ((i = next_desc(vq, &desc)) != -1);
2129
2130 /* On success, increment avail index. */
2131 vq->last_avail_idx++;
2132
2133 /* Assume notifications from guest are disabled at this point,
2134 * if they aren't we would need to update avail_event index. */
2135 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2136 return head;
2137 }
2138 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2139
2140 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2141 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2142 {
2143 vq->last_avail_idx -= n;
2144 }
2145 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2146
2147 /* After we've used one of their buffers, we tell them about it. We'll then
2148 * want to notify the guest, using eventfd. */
2149 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2150 {
2151 struct vring_used_elem heads = {
2152 cpu_to_vhost32(vq, head),
2153 cpu_to_vhost32(vq, len)
2154 };
2155
2156 return vhost_add_used_n(vq, &heads, 1);
2157 }
2158 EXPORT_SYMBOL_GPL(vhost_add_used);
2159
2160 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2161 struct vring_used_elem *heads,
2162 unsigned count)
2163 {
2164 struct vring_used_elem __user *used;
2165 u16 old, new;
2166 int start;
2167
2168 start = vq->last_used_idx & (vq->num - 1);
2169 used = vq->used->ring + start;
2170 if (count == 1) {
2171 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2172 vq_err(vq, "Failed to write used id");
2173 return -EFAULT;
2174 }
2175 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2176 vq_err(vq, "Failed to write used len");
2177 return -EFAULT;
2178 }
2179 } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2180 vq_err(vq, "Failed to write used");
2181 return -EFAULT;
2182 }
2183 if (unlikely(vq->log_used)) {
2184 /* Make sure data is seen before log. */
2185 smp_wmb();
2186 /* Log used ring entry write. */
2187 log_write(vq->log_base,
2188 vq->log_addr +
2189 ((void __user *)used - (void __user *)vq->used),
2190 count * sizeof *used);
2191 }
2192 old = vq->last_used_idx;
2193 new = (vq->last_used_idx += count);
2194 /* If the driver never bothers to signal in a very long while,
2195 * used index might wrap around. If that happens, invalidate
2196 * signalled_used index we stored. TODO: make sure driver
2197 * signals at least once in 2^16 and remove this. */
2198 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2199 vq->signalled_used_valid = false;
2200 return 0;
2201 }
2202
2203 /* After we've used one of their buffers, we tell them about it. We'll then
2204 * want to notify the guest, using eventfd. */
2205 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2206 unsigned count)
2207 {
2208 int start, n, r;
2209
2210 start = vq->last_used_idx & (vq->num - 1);
2211 n = vq->num - start;
2212 if (n < count) {
2213 r = __vhost_add_used_n(vq, heads, n);
2214 if (r < 0)
2215 return r;
2216 heads += n;
2217 count -= n;
2218 }
2219 r = __vhost_add_used_n(vq, heads, count);
2220
2221 /* Make sure buffer is written before we update index. */
2222 smp_wmb();
2223 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2224 &vq->used->idx)) {
2225 vq_err(vq, "Failed to increment used idx");
2226 return -EFAULT;
2227 }
2228 if (unlikely(vq->log_used)) {
2229 /* Log used index update. */
2230 log_write(vq->log_base,
2231 vq->log_addr + offsetof(struct vring_used, idx),
2232 sizeof vq->used->idx);
2233 if (vq->log_ctx)
2234 eventfd_signal(vq->log_ctx, 1);
2235 }
2236 return r;
2237 }
2238 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2239
2240 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2241 {
2242 __u16 old, new;
2243 __virtio16 event;
2244 bool v;
2245 /* Flush out used index updates. This is paired
2246 * with the barrier that the Guest executes when enabling
2247 * interrupts. */
2248 smp_mb();
2249
2250 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2251 unlikely(vq->avail_idx == vq->last_avail_idx))
2252 return true;
2253
2254 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2255 __virtio16 flags;
2256 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2257 vq_err(vq, "Failed to get flags");
2258 return true;
2259 }
2260 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2261 }
2262 old = vq->signalled_used;
2263 v = vq->signalled_used_valid;
2264 new = vq->signalled_used = vq->last_used_idx;
2265 vq->signalled_used_valid = true;
2266
2267 if (unlikely(!v))
2268 return true;
2269
2270 if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2271 vq_err(vq, "Failed to get used event idx");
2272 return true;
2273 }
2274 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2275 }
2276
2277 /* This actually signals the guest, using eventfd. */
2278 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2279 {
2280 /* Signal the Guest tell them we used something up. */
2281 if (vq->call_ctx && vhost_notify(dev, vq))
2282 eventfd_signal(vq->call_ctx, 1);
2283 }
2284 EXPORT_SYMBOL_GPL(vhost_signal);
2285
2286 /* And here's the combo meal deal. Supersize me! */
2287 void vhost_add_used_and_signal(struct vhost_dev *dev,
2288 struct vhost_virtqueue *vq,
2289 unsigned int head, int len)
2290 {
2291 vhost_add_used(vq, head, len);
2292 vhost_signal(dev, vq);
2293 }
2294 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2295
2296 /* multi-buffer version of vhost_add_used_and_signal */
2297 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2298 struct vhost_virtqueue *vq,
2299 struct vring_used_elem *heads, unsigned count)
2300 {
2301 vhost_add_used_n(vq, heads, count);
2302 vhost_signal(dev, vq);
2303 }
2304 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2305
2306 /* return true if we're sure that avaiable ring is empty */
2307 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2308 {
2309 __virtio16 avail_idx;
2310 int r;
2311
2312 if (vq->avail_idx != vq->last_avail_idx)
2313 return false;
2314
2315 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2316 if (unlikely(r))
2317 return false;
2318 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2319
2320 return vq->avail_idx == vq->last_avail_idx;
2321 }
2322 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2323
2324 /* OK, now we need to know about added descriptors. */
2325 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2326 {
2327 __virtio16 avail_idx;
2328 int r;
2329
2330 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2331 return false;
2332 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2333 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2334 r = vhost_update_used_flags(vq);
2335 if (r) {
2336 vq_err(vq, "Failed to enable notification at %p: %d\n",
2337 &vq->used->flags, r);
2338 return false;
2339 }
2340 } else {
2341 r = vhost_update_avail_event(vq, vq->avail_idx);
2342 if (r) {
2343 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2344 vhost_avail_event(vq), r);
2345 return false;
2346 }
2347 }
2348 /* They could have slipped one in as we were doing that: make
2349 * sure it's written, then check again. */
2350 smp_mb();
2351 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2352 if (r) {
2353 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2354 &vq->avail->idx, r);
2355 return false;
2356 }
2357
2358 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2359 }
2360 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2361
2362 /* We don't need to be notified again. */
2363 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2364 {
2365 int r;
2366
2367 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2368 return;
2369 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2370 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2371 r = vhost_update_used_flags(vq);
2372 if (r)
2373 vq_err(vq, "Failed to enable notification at %p: %d\n",
2374 &vq->used->flags, r);
2375 }
2376 }
2377 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2378
2379 /* Create a new message. */
2380 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2381 {
2382 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2383 if (!node)
2384 return NULL;
2385 node->vq = vq;
2386 node->msg.type = type;
2387 return node;
2388 }
2389 EXPORT_SYMBOL_GPL(vhost_new_msg);
2390
2391 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2392 struct vhost_msg_node *node)
2393 {
2394 spin_lock(&dev->iotlb_lock);
2395 list_add_tail(&node->node, head);
2396 spin_unlock(&dev->iotlb_lock);
2397
2398 wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
2399 }
2400 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2401
2402 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2403 struct list_head *head)
2404 {
2405 struct vhost_msg_node *node = NULL;
2406
2407 spin_lock(&dev->iotlb_lock);
2408 if (!list_empty(head)) {
2409 node = list_first_entry(head, struct vhost_msg_node,
2410 node);
2411 list_del(&node->node);
2412 }
2413 spin_unlock(&dev->iotlb_lock);
2414
2415 return node;
2416 }
2417 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2418
2419
2420 static int __init vhost_init(void)
2421 {
2422 return 0;
2423 }
2424
2425 static void __exit vhost_exit(void)
2426 {
2427 }
2428
2429 module_init(vhost_init);
2430 module_exit(vhost_exit);
2431
2432 MODULE_VERSION("0.0.1");
2433 MODULE_LICENSE("GPL v2");
2434 MODULE_AUTHOR("Michael S. Tsirkin");
2435 MODULE_DESCRIPTION("Host kernel accelerator for virtio");
Linux with added FPC-III support