xref: /kernel/linux/linux-5.10/drivers/vhost/vhost.c (revision 8c2ecf20)
1// SPDX-License-Identifier: GPL-2.0-only
2/* Copyright (C) 2009 Red Hat, Inc.
3 * Copyright (C) 2006 Rusty Russell IBM Corporation
4 *
5 * Author: Michael S. Tsirkin <mst@redhat.com>
6 *
7 * Inspiration, some code, and most witty comments come from
8 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
9 *
10 * Generic code for virtio server in host kernel.
11 */
12
13#include <linux/eventfd.h>
14#include <linux/vhost.h>
15#include <linux/uio.h>
16#include <linux/mm.h>
17#include <linux/miscdevice.h>
18#include <linux/mutex.h>
19#include <linux/poll.h>
20#include <linux/file.h>
21#include <linux/highmem.h>
22#include <linux/slab.h>
23#include <linux/vmalloc.h>
24#include <linux/kthread.h>
25#include <linux/cgroup.h>
26#include <linux/module.h>
27#include <linux/sort.h>
28#include <linux/sched/mm.h>
29#include <linux/sched/signal.h>
30#include <linux/interval_tree_generic.h>
31#include <linux/nospec.h>
32#include <linux/kcov.h>
33
34#include "vhost.h"
35
36static ushort max_mem_regions = 64;
37module_param(max_mem_regions, ushort, 0444);
38MODULE_PARM_DESC(max_mem_regions,
39	"Maximum number of memory regions in memory map. (default: 64)");
40static int max_iotlb_entries = 2048;
41module_param(max_iotlb_entries, int, 0444);
42MODULE_PARM_DESC(max_iotlb_entries,
43	"Maximum number of iotlb entries. (default: 2048)");
44
45enum {
46	VHOST_MEMORY_F_LOG = 0x1,
47};
48
49#define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
50#define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
51
52#ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
53static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
54{
55	vq->user_be = !virtio_legacy_is_little_endian();
56}
57
58static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
59{
60	vq->user_be = true;
61}
62
63static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
64{
65	vq->user_be = false;
66}
67
68static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
69{
70	struct vhost_vring_state s;
71
72	if (vq->private_data)
73		return -EBUSY;
74
75	if (copy_from_user(&s, argp, sizeof(s)))
76		return -EFAULT;
77
78	if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
79	    s.num != VHOST_VRING_BIG_ENDIAN)
80		return -EINVAL;
81
82	if (s.num == VHOST_VRING_BIG_ENDIAN)
83		vhost_enable_cross_endian_big(vq);
84	else
85		vhost_enable_cross_endian_little(vq);
86
87	return 0;
88}
89
90static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
91				   int __user *argp)
92{
93	struct vhost_vring_state s = {
94		.index = idx,
95		.num = vq->user_be
96	};
97
98	if (copy_to_user(argp, &s, sizeof(s)))
99		return -EFAULT;
100
101	return 0;
102}
103
104static void vhost_init_is_le(struct vhost_virtqueue *vq)
105{
106	/* Note for legacy virtio: user_be is initialized at reset time
107	 * according to the host endianness. If userspace does not set an
108	 * explicit endianness, the default behavior is native endian, as
109	 * expected by legacy virtio.
110	 */
111	vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
112}
113#else
114static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
115{
116}
117
118static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
119{
120	return -ENOIOCTLCMD;
121}
122
123static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
124				   int __user *argp)
125{
126	return -ENOIOCTLCMD;
127}
128
129static void vhost_init_is_le(struct vhost_virtqueue *vq)
130{
131	vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
132		|| virtio_legacy_is_little_endian();
133}
134#endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
135
136static void vhost_reset_is_le(struct vhost_virtqueue *vq)
137{
138	vhost_init_is_le(vq);
139}
140
141struct vhost_flush_struct {
142	struct vhost_work work;
143	struct completion wait_event;
144};
145
146static void vhost_flush_work(struct vhost_work *work)
147{
148	struct vhost_flush_struct *s;
149
150	s = container_of(work, struct vhost_flush_struct, work);
151	complete(&s->wait_event);
152}
153
154static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
155			    poll_table *pt)
156{
157	struct vhost_poll *poll;
158
159	poll = container_of(pt, struct vhost_poll, table);
160	poll->wqh = wqh;
161	add_wait_queue(wqh, &poll->wait);
162}
163
164static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
165			     void *key)
166{
167	struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
168	struct vhost_work *work = &poll->work;
169
170	if (!(key_to_poll(key) & poll->mask))
171		return 0;
172
173	if (!poll->dev->use_worker)
174		work->fn(work);
175	else
176		vhost_poll_queue(poll);
177
178	return 0;
179}
180
181void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
182{
183	clear_bit(VHOST_WORK_QUEUED, &work->flags);
184	work->fn = fn;
185}
186EXPORT_SYMBOL_GPL(vhost_work_init);
187
188/* Init poll structure */
189void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
190		     __poll_t mask, struct vhost_dev *dev)
191{
192	init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
193	init_poll_funcptr(&poll->table, vhost_poll_func);
194	poll->mask = mask;
195	poll->dev = dev;
196	poll->wqh = NULL;
197
198	vhost_work_init(&poll->work, fn);
199}
200EXPORT_SYMBOL_GPL(vhost_poll_init);
201
202/* Start polling a file. We add ourselves to file's wait queue. The caller must
203 * keep a reference to a file until after vhost_poll_stop is called. */
204int vhost_poll_start(struct vhost_poll *poll, struct file *file)
205{
206	__poll_t mask;
207
208	if (poll->wqh)
209		return 0;
210
211	mask = vfs_poll(file, &poll->table);
212	if (mask)
213		vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
214	if (mask & EPOLLERR) {
215		vhost_poll_stop(poll);
216		return -EINVAL;
217	}
218
219	return 0;
220}
221EXPORT_SYMBOL_GPL(vhost_poll_start);
222
223/* Stop polling a file. After this function returns, it becomes safe to drop the
224 * file reference. You must also flush afterwards. */
225void vhost_poll_stop(struct vhost_poll *poll)
226{
227	if (poll->wqh) {
228		remove_wait_queue(poll->wqh, &poll->wait);
229		poll->wqh = NULL;
230	}
231}
232EXPORT_SYMBOL_GPL(vhost_poll_stop);
233
234void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
235{
236	struct vhost_flush_struct flush;
237
238	if (dev->worker) {
239		init_completion(&flush.wait_event);
240		vhost_work_init(&flush.work, vhost_flush_work);
241
242		vhost_work_queue(dev, &flush.work);
243		wait_for_completion(&flush.wait_event);
244	}
245}
246EXPORT_SYMBOL_GPL(vhost_work_flush);
247
248/* Flush any work that has been scheduled. When calling this, don't hold any
249 * locks that are also used by the callback. */
250void vhost_poll_flush(struct vhost_poll *poll)
251{
252	vhost_work_flush(poll->dev, &poll->work);
253}
254EXPORT_SYMBOL_GPL(vhost_poll_flush);
255
256void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
257{
258	if (!dev->worker)
259		return;
260
261	if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
262		/* We can only add the work to the list after we're
263		 * sure it was not in the list.
264		 * test_and_set_bit() implies a memory barrier.
265		 */
266		llist_add(&work->node, &dev->work_list);
267		wake_up_process(dev->worker);
268	}
269}
270EXPORT_SYMBOL_GPL(vhost_work_queue);
271
272/* A lockless hint for busy polling code to exit the loop */
273bool vhost_has_work(struct vhost_dev *dev)
274{
275	return !llist_empty(&dev->work_list);
276}
277EXPORT_SYMBOL_GPL(vhost_has_work);
278
279void vhost_poll_queue(struct vhost_poll *poll)
280{
281	vhost_work_queue(poll->dev, &poll->work);
282}
283EXPORT_SYMBOL_GPL(vhost_poll_queue);
284
285static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
286{
287	int j;
288
289	for (j = 0; j < VHOST_NUM_ADDRS; j++)
290		vq->meta_iotlb[j] = NULL;
291}
292
293static void vhost_vq_meta_reset(struct vhost_dev *d)
294{
295	int i;
296
297	for (i = 0; i < d->nvqs; ++i)
298		__vhost_vq_meta_reset(d->vqs[i]);
299}
300
301static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
302{
303	call_ctx->ctx = NULL;
304	memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
305}
306
307bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
308{
309	return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
310}
311EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
312
313static void vhost_vq_reset(struct vhost_dev *dev,
314			   struct vhost_virtqueue *vq)
315{
316	vq->num = 1;
317	vq->desc = NULL;
318	vq->avail = NULL;
319	vq->used = NULL;
320	vq->last_avail_idx = 0;
321	vq->avail_idx = 0;
322	vq->last_used_idx = 0;
323	vq->signalled_used = 0;
324	vq->signalled_used_valid = false;
325	vq->used_flags = 0;
326	vq->log_used = false;
327	vq->log_addr = -1ull;
328	vq->private_data = NULL;
329	vq->acked_features = 0;
330	vq->acked_backend_features = 0;
331	vq->log_base = NULL;
332	vq->error_ctx = NULL;
333	vq->kick = NULL;
334	vq->log_ctx = NULL;
335	vhost_disable_cross_endian(vq);
336	vhost_reset_is_le(vq);
337	vq->busyloop_timeout = 0;
338	vq->umem = NULL;
339	vq->iotlb = NULL;
340	vhost_vring_call_reset(&vq->call_ctx);
341	__vhost_vq_meta_reset(vq);
342}
343
344static int vhost_worker(void *data)
345{
346	struct vhost_dev *dev = data;
347	struct vhost_work *work, *work_next;
348	struct llist_node *node;
349
350	kthread_use_mm(dev->mm);
351
352	for (;;) {
353		/* mb paired w/ kthread_stop */
354		set_current_state(TASK_INTERRUPTIBLE);
355
356		if (kthread_should_stop()) {
357			__set_current_state(TASK_RUNNING);
358			break;
359		}
360
361		node = llist_del_all(&dev->work_list);
362		if (!node)
363			schedule();
364
365		node = llist_reverse_order(node);
366		/* make sure flag is seen after deletion */
367		smp_wmb();
368		llist_for_each_entry_safe(work, work_next, node, node) {
369			clear_bit(VHOST_WORK_QUEUED, &work->flags);
370			__set_current_state(TASK_RUNNING);
371			kcov_remote_start_common(dev->kcov_handle);
372			work->fn(work);
373			kcov_remote_stop();
374			if (need_resched())
375				schedule();
376		}
377	}
378	kthread_unuse_mm(dev->mm);
379	return 0;
380}
381
382static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
383{
384	kfree(vq->indirect);
385	vq->indirect = NULL;
386	kfree(vq->log);
387	vq->log = NULL;
388	kfree(vq->heads);
389	vq->heads = NULL;
390}
391
392/* Helper to allocate iovec buffers for all vqs. */
393static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
394{
395	struct vhost_virtqueue *vq;
396	int i;
397
398	for (i = 0; i < dev->nvqs; ++i) {
399		vq = dev->vqs[i];
400		vq->indirect = kmalloc_array(UIO_MAXIOV,
401					     sizeof(*vq->indirect),
402					     GFP_KERNEL);
403		vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
404					GFP_KERNEL);
405		vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
406					  GFP_KERNEL);
407		if (!vq->indirect || !vq->log || !vq->heads)
408			goto err_nomem;
409	}
410	return 0;
411
412err_nomem:
413	for (; i >= 0; --i)
414		vhost_vq_free_iovecs(dev->vqs[i]);
415	return -ENOMEM;
416}
417
418static void vhost_dev_free_iovecs(struct vhost_dev *dev)
419{
420	int i;
421
422	for (i = 0; i < dev->nvqs; ++i)
423		vhost_vq_free_iovecs(dev->vqs[i]);
424}
425
426bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
427			  int pkts, int total_len)
428{
429	struct vhost_dev *dev = vq->dev;
430
431	if ((dev->byte_weight && total_len >= dev->byte_weight) ||
432	    pkts >= dev->weight) {
433		vhost_poll_queue(&vq->poll);
434		return true;
435	}
436
437	return false;
438}
439EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
440
441static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
442				   unsigned int num)
443{
444	size_t event __maybe_unused =
445	       vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
446
447	return sizeof(*vq->avail) +
448	       sizeof(*vq->avail->ring) * num + event;
449}
450
451static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
452				  unsigned int num)
453{
454	size_t event __maybe_unused =
455	       vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
456
457	return sizeof(*vq->used) +
458	       sizeof(*vq->used->ring) * num + event;
459}
460
461static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
462				  unsigned int num)
463{
464	return sizeof(*vq->desc) * num;
465}
466
467void vhost_dev_init(struct vhost_dev *dev,
468		    struct vhost_virtqueue **vqs, int nvqs,
469		    int iov_limit, int weight, int byte_weight,
470		    bool use_worker,
471		    int (*msg_handler)(struct vhost_dev *dev,
472				       struct vhost_iotlb_msg *msg))
473{
474	struct vhost_virtqueue *vq;
475	int i;
476
477	dev->vqs = vqs;
478	dev->nvqs = nvqs;
479	mutex_init(&dev->mutex);
480	dev->log_ctx = NULL;
481	dev->umem = NULL;
482	dev->iotlb = NULL;
483	dev->mm = NULL;
484	dev->worker = NULL;
485	dev->iov_limit = iov_limit;
486	dev->weight = weight;
487	dev->byte_weight = byte_weight;
488	dev->use_worker = use_worker;
489	dev->msg_handler = msg_handler;
490	init_llist_head(&dev->work_list);
491	init_waitqueue_head(&dev->wait);
492	INIT_LIST_HEAD(&dev->read_list);
493	INIT_LIST_HEAD(&dev->pending_list);
494	spin_lock_init(&dev->iotlb_lock);
495
496
497	for (i = 0; i < dev->nvqs; ++i) {
498		vq = dev->vqs[i];
499		vq->log = NULL;
500		vq->indirect = NULL;
501		vq->heads = NULL;
502		vq->dev = dev;
503		mutex_init(&vq->mutex);
504		vhost_vq_reset(dev, vq);
505		if (vq->handle_kick)
506			vhost_poll_init(&vq->poll, vq->handle_kick,
507					EPOLLIN, dev);
508	}
509}
510EXPORT_SYMBOL_GPL(vhost_dev_init);
511
512/* Caller should have device mutex */
513long vhost_dev_check_owner(struct vhost_dev *dev)
514{
515	/* Are you the owner? If not, I don't think you mean to do that */
516	return dev->mm == current->mm ? 0 : -EPERM;
517}
518EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
519
520struct vhost_attach_cgroups_struct {
521	struct vhost_work work;
522	struct task_struct *owner;
523	int ret;
524};
525
526static void vhost_attach_cgroups_work(struct vhost_work *work)
527{
528	struct vhost_attach_cgroups_struct *s;
529
530	s = container_of(work, struct vhost_attach_cgroups_struct, work);
531	s->ret = cgroup_attach_task_all(s->owner, current);
532}
533
534static int vhost_attach_cgroups(struct vhost_dev *dev)
535{
536	struct vhost_attach_cgroups_struct attach;
537
538	attach.owner = current;
539	vhost_work_init(&attach.work, vhost_attach_cgroups_work);
540	vhost_work_queue(dev, &attach.work);
541	vhost_work_flush(dev, &attach.work);
542	return attach.ret;
543}
544
545/* Caller should have device mutex */
546bool vhost_dev_has_owner(struct vhost_dev *dev)
547{
548	return dev->mm;
549}
550EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
551
552static void vhost_attach_mm(struct vhost_dev *dev)
553{
554	/* No owner, become one */
555	if (dev->use_worker) {
556		dev->mm = get_task_mm(current);
557	} else {
558		/* vDPA device does not use worker thead, so there's
559		 * no need to hold the address space for mm. This help
560		 * to avoid deadlock in the case of mmap() which may
561		 * held the refcnt of the file and depends on release
562		 * method to remove vma.
563		 */
564		dev->mm = current->mm;
565		mmgrab(dev->mm);
566	}
567}
568
569static void vhost_detach_mm(struct vhost_dev *dev)
570{
571	if (!dev->mm)
572		return;
573
574	if (dev->use_worker)
575		mmput(dev->mm);
576	else
577		mmdrop(dev->mm);
578
579	dev->mm = NULL;
580}
581
582/* Caller should have device mutex */
583long vhost_dev_set_owner(struct vhost_dev *dev)
584{
585	struct task_struct *worker;
586	int err;
587
588	/* Is there an owner already? */
589	if (vhost_dev_has_owner(dev)) {
590		err = -EBUSY;
591		goto err_mm;
592	}
593
594	vhost_attach_mm(dev);
595
596	dev->kcov_handle = kcov_common_handle();
597	if (dev->use_worker) {
598		worker = kthread_create(vhost_worker, dev,
599					"vhost-%d", current->pid);
600		if (IS_ERR(worker)) {
601			err = PTR_ERR(worker);
602			goto err_worker;
603		}
604
605		dev->worker = worker;
606		wake_up_process(worker); /* avoid contributing to loadavg */
607
608		err = vhost_attach_cgroups(dev);
609		if (err)
610			goto err_cgroup;
611	}
612
613	err = vhost_dev_alloc_iovecs(dev);
614	if (err)
615		goto err_cgroup;
616
617	return 0;
618err_cgroup:
619	if (dev->worker) {
620		kthread_stop(dev->worker);
621		dev->worker = NULL;
622	}
623err_worker:
624	vhost_detach_mm(dev);
625	dev->kcov_handle = 0;
626err_mm:
627	return err;
628}
629EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
630
631static struct vhost_iotlb *iotlb_alloc(void)
632{
633	return vhost_iotlb_alloc(max_iotlb_entries,
634				 VHOST_IOTLB_FLAG_RETIRE);
635}
636
637struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
638{
639	return iotlb_alloc();
640}
641EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
642
643/* Caller should have device mutex */
644void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
645{
646	int i;
647
648	vhost_dev_cleanup(dev);
649
650	dev->umem = umem;
651	/* We don't need VQ locks below since vhost_dev_cleanup makes sure
652	 * VQs aren't running.
653	 */
654	for (i = 0; i < dev->nvqs; ++i)
655		dev->vqs[i]->umem = umem;
656}
657EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
658
659void vhost_dev_stop(struct vhost_dev *dev)
660{
661	int i;
662
663	for (i = 0; i < dev->nvqs; ++i) {
664		if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
665			vhost_poll_stop(&dev->vqs[i]->poll);
666			vhost_poll_flush(&dev->vqs[i]->poll);
667		}
668	}
669}
670EXPORT_SYMBOL_GPL(vhost_dev_stop);
671
672void vhost_clear_msg(struct vhost_dev *dev)
673{
674	struct vhost_msg_node *node, *n;
675
676	spin_lock(&dev->iotlb_lock);
677
678	list_for_each_entry_safe(node, n, &dev->read_list, node) {
679		list_del(&node->node);
680		kfree(node);
681	}
682
683	list_for_each_entry_safe(node, n, &dev->pending_list, node) {
684		list_del(&node->node);
685		kfree(node);
686	}
687
688	spin_unlock(&dev->iotlb_lock);
689}
690EXPORT_SYMBOL_GPL(vhost_clear_msg);
691
692void vhost_dev_cleanup(struct vhost_dev *dev)
693{
694	int i;
695
696	for (i = 0; i < dev->nvqs; ++i) {
697		if (dev->vqs[i]->error_ctx)
698			eventfd_ctx_put(dev->vqs[i]->error_ctx);
699		if (dev->vqs[i]->kick)
700			fput(dev->vqs[i]->kick);
701		if (dev->vqs[i]->call_ctx.ctx)
702			eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
703		vhost_vq_reset(dev, dev->vqs[i]);
704	}
705	vhost_dev_free_iovecs(dev);
706	if (dev->log_ctx)
707		eventfd_ctx_put(dev->log_ctx);
708	dev->log_ctx = NULL;
709	/* No one will access memory at this point */
710	vhost_iotlb_free(dev->umem);
711	dev->umem = NULL;
712	vhost_iotlb_free(dev->iotlb);
713	dev->iotlb = NULL;
714	vhost_clear_msg(dev);
715	wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
716	WARN_ON(!llist_empty(&dev->work_list));
717	if (dev->worker) {
718		kthread_stop(dev->worker);
719		dev->worker = NULL;
720		dev->kcov_handle = 0;
721	}
722	vhost_detach_mm(dev);
723}
724EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
725
726static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
727{
728	u64 a = addr / VHOST_PAGE_SIZE / 8;
729
730	/* Make sure 64 bit math will not overflow. */
731	if (a > ULONG_MAX - (unsigned long)log_base ||
732	    a + (unsigned long)log_base > ULONG_MAX)
733		return false;
734
735	return access_ok(log_base + a,
736			 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
737}
738
739/* Make sure 64 bit math will not overflow. */
740static bool vhost_overflow(u64 uaddr, u64 size)
741{
742	if (uaddr > ULONG_MAX || size > ULONG_MAX)
743		return true;
744
745	if (!size)
746		return false;
747
748	return uaddr > ULONG_MAX - size + 1;
749}
750
751/* Caller should have vq mutex and device mutex. */
752static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
753				int log_all)
754{
755	struct vhost_iotlb_map *map;
756
757	if (!umem)
758		return false;
759
760	list_for_each_entry(map, &umem->list, link) {
761		unsigned long a = map->addr;
762
763		if (vhost_overflow(map->addr, map->size))
764			return false;
765
766
767		if (!access_ok((void __user *)a, map->size))
768			return false;
769		else if (log_all && !log_access_ok(log_base,
770						   map->start,
771						   map->size))
772			return false;
773	}
774	return true;
775}
776
777static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
778					       u64 addr, unsigned int size,
779					       int type)
780{
781	const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
782
783	if (!map)
784		return NULL;
785
786	return (void __user *)(uintptr_t)(map->addr + addr - map->start);
787}
788
789/* Can we switch to this memory table? */
790/* Caller should have device mutex but not vq mutex */
791static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
792			     int log_all)
793{
794	int i;
795
796	for (i = 0; i < d->nvqs; ++i) {
797		bool ok;
798		bool log;
799
800		mutex_lock(&d->vqs[i]->mutex);
801		log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
802		/* If ring is inactive, will check when it's enabled. */
803		if (d->vqs[i]->private_data)
804			ok = vq_memory_access_ok(d->vqs[i]->log_base,
805						 umem, log);
806		else
807			ok = true;
808		mutex_unlock(&d->vqs[i]->mutex);
809		if (!ok)
810			return false;
811	}
812	return true;
813}
814
815static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
816			  struct iovec iov[], int iov_size, int access);
817
818static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
819			      const void *from, unsigned size)
820{
821	int ret;
822
823	if (!vq->iotlb)
824		return __copy_to_user(to, from, size);
825	else {
826		/* This function should be called after iotlb
827		 * prefetch, which means we're sure that all vq
828		 * could be access through iotlb. So -EAGAIN should
829		 * not happen in this case.
830		 */
831		struct iov_iter t;
832		void __user *uaddr = vhost_vq_meta_fetch(vq,
833				     (u64)(uintptr_t)to, size,
834				     VHOST_ADDR_USED);
835
836		if (uaddr)
837			return __copy_to_user(uaddr, from, size);
838
839		ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
840				     ARRAY_SIZE(vq->iotlb_iov),
841				     VHOST_ACCESS_WO);
842		if (ret < 0)
843			goto out;
844		iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
845		ret = copy_to_iter(from, size, &t);
846		if (ret == size)
847			ret = 0;
848	}
849out:
850	return ret;
851}
852
853static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
854				void __user *from, unsigned size)
855{
856	int ret;
857
858	if (!vq->iotlb)
859		return __copy_from_user(to, from, size);
860	else {
861		/* This function should be called after iotlb
862		 * prefetch, which means we're sure that vq
863		 * could be access through iotlb. So -EAGAIN should
864		 * not happen in this case.
865		 */
866		void __user *uaddr = vhost_vq_meta_fetch(vq,
867				     (u64)(uintptr_t)from, size,
868				     VHOST_ADDR_DESC);
869		struct iov_iter f;
870
871		if (uaddr)
872			return __copy_from_user(to, uaddr, size);
873
874		ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
875				     ARRAY_SIZE(vq->iotlb_iov),
876				     VHOST_ACCESS_RO);
877		if (ret < 0) {
878			vq_err(vq, "IOTLB translation failure: uaddr "
879			       "%p size 0x%llx\n", from,
880			       (unsigned long long) size);
881			goto out;
882		}
883		iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
884		ret = copy_from_iter(to, size, &f);
885		if (ret == size)
886			ret = 0;
887	}
888
889out:
890	return ret;
891}
892
893static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
894					  void __user *addr, unsigned int size,
895					  int type)
896{
897	int ret;
898
899	ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
900			     ARRAY_SIZE(vq->iotlb_iov),
901			     VHOST_ACCESS_RO);
902	if (ret < 0) {
903		vq_err(vq, "IOTLB translation failure: uaddr "
904			"%p size 0x%llx\n", addr,
905			(unsigned long long) size);
906		return NULL;
907	}
908
909	if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
910		vq_err(vq, "Non atomic userspace memory access: uaddr "
911			"%p size 0x%llx\n", addr,
912			(unsigned long long) size);
913		return NULL;
914	}
915
916	return vq->iotlb_iov[0].iov_base;
917}
918
919/* This function should be called after iotlb
920 * prefetch, which means we're sure that vq
921 * could be access through iotlb. So -EAGAIN should
922 * not happen in this case.
923 */
924static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
925					    void __user *addr, unsigned int size,
926					    int type)
927{
928	void __user *uaddr = vhost_vq_meta_fetch(vq,
929			     (u64)(uintptr_t)addr, size, type);
930	if (uaddr)
931		return uaddr;
932
933	return __vhost_get_user_slow(vq, addr, size, type);
934}
935
936#define vhost_put_user(vq, x, ptr)		\
937({ \
938	int ret; \
939	if (!vq->iotlb) { \
940		ret = __put_user(x, ptr); \
941	} else { \
942		__typeof__(ptr) to = \
943			(__typeof__(ptr)) __vhost_get_user(vq, ptr,	\
944					  sizeof(*ptr), VHOST_ADDR_USED); \
945		if (to != NULL) \
946			ret = __put_user(x, to); \
947		else \
948			ret = -EFAULT;	\
949	} \
950	ret; \
951})
952
953static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
954{
955	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
956			      vhost_avail_event(vq));
957}
958
959static inline int vhost_put_used(struct vhost_virtqueue *vq,
960				 struct vring_used_elem *head, int idx,
961				 int count)
962{
963	return vhost_copy_to_user(vq, vq->used->ring + idx, head,
964				  count * sizeof(*head));
965}
966
967static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
968
969{
970	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
971			      &vq->used->flags);
972}
973
974static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
975
976{
977	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
978			      &vq->used->idx);
979}
980
981#define vhost_get_user(vq, x, ptr, type)		\
982({ \
983	int ret; \
984	if (!vq->iotlb) { \
985		ret = __get_user(x, ptr); \
986	} else { \
987		__typeof__(ptr) from = \
988			(__typeof__(ptr)) __vhost_get_user(vq, ptr, \
989							   sizeof(*ptr), \
990							   type); \
991		if (from != NULL) \
992			ret = __get_user(x, from); \
993		else \
994			ret = -EFAULT; \
995	} \
996	ret; \
997})
998
999#define vhost_get_avail(vq, x, ptr) \
1000	vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
1001
1002#define vhost_get_used(vq, x, ptr) \
1003	vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
1004
1005static void vhost_dev_lock_vqs(struct vhost_dev *d)
1006{
1007	int i = 0;
1008	for (i = 0; i < d->nvqs; ++i)
1009		mutex_lock_nested(&d->vqs[i]->mutex, i);
1010}
1011
1012static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1013{
1014	int i = 0;
1015	for (i = 0; i < d->nvqs; ++i)
1016		mutex_unlock(&d->vqs[i]->mutex);
1017}
1018
1019static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1020				      __virtio16 *idx)
1021{
1022	return vhost_get_avail(vq, *idx, &vq->avail->idx);
1023}
1024
1025static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1026				       __virtio16 *head, int idx)
1027{
1028	return vhost_get_avail(vq, *head,
1029			       &vq->avail->ring[idx & (vq->num - 1)]);
1030}
1031
1032static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1033					__virtio16 *flags)
1034{
1035	return vhost_get_avail(vq, *flags, &vq->avail->flags);
1036}
1037
1038static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1039				       __virtio16 *event)
1040{
1041	return vhost_get_avail(vq, *event, vhost_used_event(vq));
1042}
1043
1044static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1045				     __virtio16 *idx)
1046{
1047	return vhost_get_used(vq, *idx, &vq->used->idx);
1048}
1049
1050static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1051				 struct vring_desc *desc, int idx)
1052{
1053	return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1054}
1055
1056static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1057				  struct vhost_iotlb_msg *msg)
1058{
1059	struct vhost_msg_node *node, *n;
1060
1061	spin_lock(&d->iotlb_lock);
1062
1063	list_for_each_entry_safe(node, n, &d->pending_list, node) {
1064		struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1065		if (msg->iova <= vq_msg->iova &&
1066		    msg->iova + msg->size - 1 >= vq_msg->iova &&
1067		    vq_msg->type == VHOST_IOTLB_MISS) {
1068			vhost_poll_queue(&node->vq->poll);
1069			list_del(&node->node);
1070			kfree(node);
1071		}
1072	}
1073
1074	spin_unlock(&d->iotlb_lock);
1075}
1076
1077static bool umem_access_ok(u64 uaddr, u64 size, int access)
1078{
1079	unsigned long a = uaddr;
1080
1081	/* Make sure 64 bit math will not overflow. */
1082	if (vhost_overflow(uaddr, size))
1083		return false;
1084
1085	if ((access & VHOST_ACCESS_RO) &&
1086	    !access_ok((void __user *)a, size))
1087		return false;
1088	if ((access & VHOST_ACCESS_WO) &&
1089	    !access_ok((void __user *)a, size))
1090		return false;
1091	return true;
1092}
1093
1094static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1095				   struct vhost_iotlb_msg *msg)
1096{
1097	int ret = 0;
1098
1099	mutex_lock(&dev->mutex);
1100	vhost_dev_lock_vqs(dev);
1101	switch (msg->type) {
1102	case VHOST_IOTLB_UPDATE:
1103		if (!dev->iotlb) {
1104			ret = -EFAULT;
1105			break;
1106		}
1107		if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1108			ret = -EFAULT;
1109			break;
1110		}
1111		vhost_vq_meta_reset(dev);
1112		if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1113					  msg->iova + msg->size - 1,
1114					  msg->uaddr, msg->perm)) {
1115			ret = -ENOMEM;
1116			break;
1117		}
1118		vhost_iotlb_notify_vq(dev, msg);
1119		break;
1120	case VHOST_IOTLB_INVALIDATE:
1121		if (!dev->iotlb) {
1122			ret = -EFAULT;
1123			break;
1124		}
1125		vhost_vq_meta_reset(dev);
1126		vhost_iotlb_del_range(dev->iotlb, msg->iova,
1127				      msg->iova + msg->size - 1);
1128		break;
1129	default:
1130		ret = -EINVAL;
1131		break;
1132	}
1133
1134	vhost_dev_unlock_vqs(dev);
1135	mutex_unlock(&dev->mutex);
1136
1137	return ret;
1138}
1139ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1140			     struct iov_iter *from)
1141{
1142	struct vhost_iotlb_msg msg;
1143	size_t offset;
1144	int type, ret;
1145
1146	ret = copy_from_iter(&type, sizeof(type), from);
1147	if (ret != sizeof(type)) {
1148		ret = -EINVAL;
1149		goto done;
1150	}
1151
1152	switch (type) {
1153	case VHOST_IOTLB_MSG:
1154		/* There maybe a hole after type for V1 message type,
1155		 * so skip it here.
1156		 */
1157		offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1158		break;
1159	case VHOST_IOTLB_MSG_V2:
1160		offset = sizeof(__u32);
1161		break;
1162	default:
1163		ret = -EINVAL;
1164		goto done;
1165	}
1166
1167	iov_iter_advance(from, offset);
1168	ret = copy_from_iter(&msg, sizeof(msg), from);
1169	if (ret != sizeof(msg)) {
1170		ret = -EINVAL;
1171		goto done;
1172	}
1173
1174	if (dev->msg_handler)
1175		ret = dev->msg_handler(dev, &msg);
1176	else
1177		ret = vhost_process_iotlb_msg(dev, &msg);
1178	if (ret) {
1179		ret = -EFAULT;
1180		goto done;
1181	}
1182
1183	ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1184	      sizeof(struct vhost_msg_v2);
1185done:
1186	return ret;
1187}
1188EXPORT_SYMBOL(vhost_chr_write_iter);
1189
1190__poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1191			    poll_table *wait)
1192{
1193	__poll_t mask = 0;
1194
1195	poll_wait(file, &dev->wait, wait);
1196
1197	if (!list_empty(&dev->read_list))
1198		mask |= EPOLLIN | EPOLLRDNORM;
1199
1200	return mask;
1201}
1202EXPORT_SYMBOL(vhost_chr_poll);
1203
1204ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1205			    int noblock)
1206{
1207	DEFINE_WAIT(wait);
1208	struct vhost_msg_node *node;
1209	ssize_t ret = 0;
1210	unsigned size = sizeof(struct vhost_msg);
1211
1212	if (iov_iter_count(to) < size)
1213		return 0;
1214
1215	while (1) {
1216		if (!noblock)
1217			prepare_to_wait(&dev->wait, &wait,
1218					TASK_INTERRUPTIBLE);
1219
1220		node = vhost_dequeue_msg(dev, &dev->read_list);
1221		if (node)
1222			break;
1223		if (noblock) {
1224			ret = -EAGAIN;
1225			break;
1226		}
1227		if (signal_pending(current)) {
1228			ret = -ERESTARTSYS;
1229			break;
1230		}
1231		if (!dev->iotlb) {
1232			ret = -EBADFD;
1233			break;
1234		}
1235
1236		schedule();
1237	}
1238
1239	if (!noblock)
1240		finish_wait(&dev->wait, &wait);
1241
1242	if (node) {
1243		struct vhost_iotlb_msg *msg;
1244		void *start = &node->msg;
1245
1246		switch (node->msg.type) {
1247		case VHOST_IOTLB_MSG:
1248			size = sizeof(node->msg);
1249			msg = &node->msg.iotlb;
1250			break;
1251		case VHOST_IOTLB_MSG_V2:
1252			size = sizeof(node->msg_v2);
1253			msg = &node->msg_v2.iotlb;
1254			break;
1255		default:
1256			BUG();
1257			break;
1258		}
1259
1260		ret = copy_to_iter(start, size, to);
1261		if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1262			kfree(node);
1263			return ret;
1264		}
1265		vhost_enqueue_msg(dev, &dev->pending_list, node);
1266	}
1267
1268	return ret;
1269}
1270EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1271
1272static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1273{
1274	struct vhost_dev *dev = vq->dev;
1275	struct vhost_msg_node *node;
1276	struct vhost_iotlb_msg *msg;
1277	bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1278
1279	node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1280	if (!node)
1281		return -ENOMEM;
1282
1283	if (v2) {
1284		node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1285		msg = &node->msg_v2.iotlb;
1286	} else {
1287		msg = &node->msg.iotlb;
1288	}
1289
1290	msg->type = VHOST_IOTLB_MISS;
1291	msg->iova = iova;
1292	msg->perm = access;
1293
1294	vhost_enqueue_msg(dev, &dev->read_list, node);
1295
1296	return 0;
1297}
1298
1299static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1300			 vring_desc_t __user *desc,
1301			 vring_avail_t __user *avail,
1302			 vring_used_t __user *used)
1303
1304{
1305	/* If an IOTLB device is present, the vring addresses are
1306	 * GIOVAs. Access validation occurs at prefetch time. */
1307	if (vq->iotlb)
1308		return true;
1309
1310	return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1311	       access_ok(avail, vhost_get_avail_size(vq, num)) &&
1312	       access_ok(used, vhost_get_used_size(vq, num));
1313}
1314
1315static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1316				 const struct vhost_iotlb_map *map,
1317				 int type)
1318{
1319	int access = (type == VHOST_ADDR_USED) ?
1320		     VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1321
1322	if (likely(map->perm & access))
1323		vq->meta_iotlb[type] = map;
1324}
1325
1326static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1327			    int access, u64 addr, u64 len, int type)
1328{
1329	const struct vhost_iotlb_map *map;
1330	struct vhost_iotlb *umem = vq->iotlb;
1331	u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1332
1333	if (vhost_vq_meta_fetch(vq, addr, len, type))
1334		return true;
1335
1336	while (len > s) {
1337		map = vhost_iotlb_itree_first(umem, addr, last);
1338		if (map == NULL || map->start > addr) {
1339			vhost_iotlb_miss(vq, addr, access);
1340			return false;
1341		} else if (!(map->perm & access)) {
1342			/* Report the possible access violation by
1343			 * request another translation from userspace.
1344			 */
1345			return false;
1346		}
1347
1348		size = map->size - addr + map->start;
1349
1350		if (orig_addr == addr && size >= len)
1351			vhost_vq_meta_update(vq, map, type);
1352
1353		s += size;
1354		addr += size;
1355	}
1356
1357	return true;
1358}
1359
1360int vq_meta_prefetch(struct vhost_virtqueue *vq)
1361{
1362	unsigned int num = vq->num;
1363
1364	if (!vq->iotlb)
1365		return 1;
1366
1367	return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1368			       vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1369	       iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1370			       vhost_get_avail_size(vq, num),
1371			       VHOST_ADDR_AVAIL) &&
1372	       iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1373			       vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1374}
1375EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1376
1377/* Can we log writes? */
1378/* Caller should have device mutex but not vq mutex */
1379bool vhost_log_access_ok(struct vhost_dev *dev)
1380{
1381	return memory_access_ok(dev, dev->umem, 1);
1382}
1383EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1384
1385static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1386				  void __user *log_base,
1387				  bool log_used,
1388				  u64 log_addr)
1389{
1390	/* If an IOTLB device is present, log_addr is a GIOVA that
1391	 * will never be logged by log_used(). */
1392	if (vq->iotlb)
1393		return true;
1394
1395	return !log_used || log_access_ok(log_base, log_addr,
1396					  vhost_get_used_size(vq, vq->num));
1397}
1398
1399/* Verify access for write logging. */
1400/* Caller should have vq mutex and device mutex */
1401static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1402			     void __user *log_base)
1403{
1404	return vq_memory_access_ok(log_base, vq->umem,
1405				   vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1406		vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1407}
1408
1409/* Can we start vq? */
1410/* Caller should have vq mutex and device mutex */
1411bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1412{
1413	if (!vq_log_access_ok(vq, vq->log_base))
1414		return false;
1415
1416	return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1417}
1418EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1419
1420static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1421{
1422	struct vhost_memory mem, *newmem;
1423	struct vhost_memory_region *region;
1424	struct vhost_iotlb *newumem, *oldumem;
1425	unsigned long size = offsetof(struct vhost_memory, regions);
1426	int i;
1427
1428	if (copy_from_user(&mem, m, size))
1429		return -EFAULT;
1430	if (mem.padding)
1431		return -EOPNOTSUPP;
1432	if (mem.nregions > max_mem_regions)
1433		return -E2BIG;
1434	newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1435			GFP_KERNEL);
1436	if (!newmem)
1437		return -ENOMEM;
1438
1439	memcpy(newmem, &mem, size);
1440	if (copy_from_user(newmem->regions, m->regions,
1441			   flex_array_size(newmem, regions, mem.nregions))) {
1442		kvfree(newmem);
1443		return -EFAULT;
1444	}
1445
1446	newumem = iotlb_alloc();
1447	if (!newumem) {
1448		kvfree(newmem);
1449		return -ENOMEM;
1450	}
1451
1452	for (region = newmem->regions;
1453	     region < newmem->regions + mem.nregions;
1454	     region++) {
1455		if (vhost_iotlb_add_range(newumem,
1456					  region->guest_phys_addr,
1457					  region->guest_phys_addr +
1458					  region->memory_size - 1,
1459					  region->userspace_addr,
1460					  VHOST_MAP_RW))
1461			goto err;
1462	}
1463
1464	if (!memory_access_ok(d, newumem, 0))
1465		goto err;
1466
1467	oldumem = d->umem;
1468	d->umem = newumem;
1469
1470	/* All memory accesses are done under some VQ mutex. */
1471	for (i = 0; i < d->nvqs; ++i) {
1472		mutex_lock(&d->vqs[i]->mutex);
1473		d->vqs[i]->umem = newumem;
1474		mutex_unlock(&d->vqs[i]->mutex);
1475	}
1476
1477	kvfree(newmem);
1478	vhost_iotlb_free(oldumem);
1479	return 0;
1480
1481err:
1482	vhost_iotlb_free(newumem);
1483	kvfree(newmem);
1484	return -EFAULT;
1485}
1486
1487static long vhost_vring_set_num(struct vhost_dev *d,
1488				struct vhost_virtqueue *vq,
1489				void __user *argp)
1490{
1491	struct vhost_vring_state s;
1492
1493	/* Resizing ring with an active backend?
1494	 * You don't want to do that. */
1495	if (vq->private_data)
1496		return -EBUSY;
1497
1498	if (copy_from_user(&s, argp, sizeof s))
1499		return -EFAULT;
1500
1501	if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1502		return -EINVAL;
1503	vq->num = s.num;
1504
1505	return 0;
1506}
1507
1508static long vhost_vring_set_addr(struct vhost_dev *d,
1509				 struct vhost_virtqueue *vq,
1510				 void __user *argp)
1511{
1512	struct vhost_vring_addr a;
1513
1514	if (copy_from_user(&a, argp, sizeof a))
1515		return -EFAULT;
1516	if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1517		return -EOPNOTSUPP;
1518
1519	/* For 32bit, verify that the top 32bits of the user
1520	   data are set to zero. */
1521	if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1522	    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1523	    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1524		return -EFAULT;
1525
1526	/* Make sure it's safe to cast pointers to vring types. */
1527	BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1528	BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1529	if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1530	    (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1531	    (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1532		return -EINVAL;
1533
1534	/* We only verify access here if backend is configured.
1535	 * If it is not, we don't as size might not have been setup.
1536	 * We will verify when backend is configured. */
1537	if (vq->private_data) {
1538		if (!vq_access_ok(vq, vq->num,
1539			(void __user *)(unsigned long)a.desc_user_addr,
1540			(void __user *)(unsigned long)a.avail_user_addr,
1541			(void __user *)(unsigned long)a.used_user_addr))
1542			return -EINVAL;
1543
1544		/* Also validate log access for used ring if enabled. */
1545		if (!vq_log_used_access_ok(vq, vq->log_base,
1546				a.flags & (0x1 << VHOST_VRING_F_LOG),
1547				a.log_guest_addr))
1548			return -EINVAL;
1549	}
1550
1551	vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1552	vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1553	vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1554	vq->log_addr = a.log_guest_addr;
1555	vq->used = (void __user *)(unsigned long)a.used_user_addr;
1556
1557	return 0;
1558}
1559
1560static long vhost_vring_set_num_addr(struct vhost_dev *d,
1561				     struct vhost_virtqueue *vq,
1562				     unsigned int ioctl,
1563				     void __user *argp)
1564{
1565	long r;
1566
1567	mutex_lock(&vq->mutex);
1568
1569	switch (ioctl) {
1570	case VHOST_SET_VRING_NUM:
1571		r = vhost_vring_set_num(d, vq, argp);
1572		break;
1573	case VHOST_SET_VRING_ADDR:
1574		r = vhost_vring_set_addr(d, vq, argp);
1575		break;
1576	default:
1577		BUG();
1578	}
1579
1580	mutex_unlock(&vq->mutex);
1581
1582	return r;
1583}
1584long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1585{
1586	struct file *eventfp, *filep = NULL;
1587	bool pollstart = false, pollstop = false;
1588	struct eventfd_ctx *ctx = NULL;
1589	u32 __user *idxp = argp;
1590	struct vhost_virtqueue *vq;
1591	struct vhost_vring_state s;
1592	struct vhost_vring_file f;
1593	u32 idx;
1594	long r;
1595
1596	r = get_user(idx, idxp);
1597	if (r < 0)
1598		return r;
1599	if (idx >= d->nvqs)
1600		return -ENOBUFS;
1601
1602	idx = array_index_nospec(idx, d->nvqs);
1603	vq = d->vqs[idx];
1604
1605	if (ioctl == VHOST_SET_VRING_NUM ||
1606	    ioctl == VHOST_SET_VRING_ADDR) {
1607		return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1608	}
1609
1610	mutex_lock(&vq->mutex);
1611
1612	switch (ioctl) {
1613	case VHOST_SET_VRING_BASE:
1614		/* Moving base with an active backend?
1615		 * You don't want to do that. */
1616		if (vq->private_data) {
1617			r = -EBUSY;
1618			break;
1619		}
1620		if (copy_from_user(&s, argp, sizeof s)) {
1621			r = -EFAULT;
1622			break;
1623		}
1624		if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
1625			vq->last_avail_idx = s.num & 0xffff;
1626			vq->last_used_idx = (s.num >> 16) & 0xffff;
1627		} else {
1628			if (s.num > 0xffff) {
1629				r = -EINVAL;
1630				break;
1631			}
1632			vq->last_avail_idx = s.num;
1633		}
1634		/* Forget the cached index value. */
1635		vq->avail_idx = vq->last_avail_idx;
1636		break;
1637	case VHOST_GET_VRING_BASE:
1638		s.index = idx;
1639		if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED))
1640			s.num = (u32)vq->last_avail_idx | ((u32)vq->last_used_idx << 16);
1641		else
1642			s.num = vq->last_avail_idx;
1643		if (copy_to_user(argp, &s, sizeof s))
1644			r = -EFAULT;
1645		break;
1646	case VHOST_SET_VRING_KICK:
1647		if (copy_from_user(&f, argp, sizeof f)) {
1648			r = -EFAULT;
1649			break;
1650		}
1651		eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1652		if (IS_ERR(eventfp)) {
1653			r = PTR_ERR(eventfp);
1654			break;
1655		}
1656		if (eventfp != vq->kick) {
1657			pollstop = (filep = vq->kick) != NULL;
1658			pollstart = (vq->kick = eventfp) != NULL;
1659		} else
1660			filep = eventfp;
1661		break;
1662	case VHOST_SET_VRING_CALL:
1663		if (copy_from_user(&f, argp, sizeof f)) {
1664			r = -EFAULT;
1665			break;
1666		}
1667		ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1668		if (IS_ERR(ctx)) {
1669			r = PTR_ERR(ctx);
1670			break;
1671		}
1672
1673		swap(ctx, vq->call_ctx.ctx);
1674		break;
1675	case VHOST_SET_VRING_ERR:
1676		if (copy_from_user(&f, argp, sizeof f)) {
1677			r = -EFAULT;
1678			break;
1679		}
1680		ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1681		if (IS_ERR(ctx)) {
1682			r = PTR_ERR(ctx);
1683			break;
1684		}
1685		swap(ctx, vq->error_ctx);
1686		break;
1687	case VHOST_SET_VRING_ENDIAN:
1688		r = vhost_set_vring_endian(vq, argp);
1689		break;
1690	case VHOST_GET_VRING_ENDIAN:
1691		r = vhost_get_vring_endian(vq, idx, argp);
1692		break;
1693	case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1694		if (copy_from_user(&s, argp, sizeof(s))) {
1695			r = -EFAULT;
1696			break;
1697		}
1698		vq->busyloop_timeout = s.num;
1699		break;
1700	case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1701		s.index = idx;
1702		s.num = vq->busyloop_timeout;
1703		if (copy_to_user(argp, &s, sizeof(s)))
1704			r = -EFAULT;
1705		break;
1706	default:
1707		r = -ENOIOCTLCMD;
1708	}
1709
1710	if (pollstop && vq->handle_kick)
1711		vhost_poll_stop(&vq->poll);
1712
1713	if (!IS_ERR_OR_NULL(ctx))
1714		eventfd_ctx_put(ctx);
1715	if (filep)
1716		fput(filep);
1717
1718	if (pollstart && vq->handle_kick)
1719		r = vhost_poll_start(&vq->poll, vq->kick);
1720
1721	mutex_unlock(&vq->mutex);
1722
1723	if (pollstop && vq->handle_kick)
1724		vhost_poll_flush(&vq->poll);
1725	return r;
1726}
1727EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1728
1729int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1730{
1731	struct vhost_iotlb *niotlb, *oiotlb;
1732	int i;
1733
1734	niotlb = iotlb_alloc();
1735	if (!niotlb)
1736		return -ENOMEM;
1737
1738	oiotlb = d->iotlb;
1739	d->iotlb = niotlb;
1740
1741	for (i = 0; i < d->nvqs; ++i) {
1742		struct vhost_virtqueue *vq = d->vqs[i];
1743
1744		mutex_lock(&vq->mutex);
1745		vq->iotlb = niotlb;
1746		__vhost_vq_meta_reset(vq);
1747		mutex_unlock(&vq->mutex);
1748	}
1749
1750	vhost_iotlb_free(oiotlb);
1751
1752	return 0;
1753}
1754EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1755
1756/* Caller must have device mutex */
1757long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1758{
1759	struct eventfd_ctx *ctx;
1760	u64 p;
1761	long r;
1762	int i, fd;
1763
1764	/* If you are not the owner, you can become one */
1765	if (ioctl == VHOST_SET_OWNER) {
1766		r = vhost_dev_set_owner(d);
1767		goto done;
1768	}
1769
1770	/* You must be the owner to do anything else */
1771	r = vhost_dev_check_owner(d);
1772	if (r)
1773		goto done;
1774
1775	switch (ioctl) {
1776	case VHOST_SET_MEM_TABLE:
1777		r = vhost_set_memory(d, argp);
1778		break;
1779	case VHOST_SET_LOG_BASE:
1780		if (copy_from_user(&p, argp, sizeof p)) {
1781			r = -EFAULT;
1782			break;
1783		}
1784		if ((u64)(unsigned long)p != p) {
1785			r = -EFAULT;
1786			break;
1787		}
1788		for (i = 0; i < d->nvqs; ++i) {
1789			struct vhost_virtqueue *vq;
1790			void __user *base = (void __user *)(unsigned long)p;
1791			vq = d->vqs[i];
1792			mutex_lock(&vq->mutex);
1793			/* If ring is inactive, will check when it's enabled. */
1794			if (vq->private_data && !vq_log_access_ok(vq, base))
1795				r = -EFAULT;
1796			else
1797				vq->log_base = base;
1798			mutex_unlock(&vq->mutex);
1799		}
1800		break;
1801	case VHOST_SET_LOG_FD:
1802		r = get_user(fd, (int __user *)argp);
1803		if (r < 0)
1804			break;
1805		ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1806		if (IS_ERR(ctx)) {
1807			r = PTR_ERR(ctx);
1808			break;
1809		}
1810		swap(ctx, d->log_ctx);
1811		for (i = 0; i < d->nvqs; ++i) {
1812			mutex_lock(&d->vqs[i]->mutex);
1813			d->vqs[i]->log_ctx = d->log_ctx;
1814			mutex_unlock(&d->vqs[i]->mutex);
1815		}
1816		if (ctx)
1817			eventfd_ctx_put(ctx);
1818		break;
1819	default:
1820		r = -ENOIOCTLCMD;
1821		break;
1822	}
1823done:
1824	return r;
1825}
1826EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1827
1828/* TODO: This is really inefficient.  We need something like get_user()
1829 * (instruction directly accesses the data, with an exception table entry
1830 * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1831 */
1832static int set_bit_to_user(int nr, void __user *addr)
1833{
1834	unsigned long log = (unsigned long)addr;
1835	struct page *page;
1836	void *base;
1837	int bit = nr + (log % PAGE_SIZE) * 8;
1838	int r;
1839
1840	r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1841	if (r < 0)
1842		return r;
1843	BUG_ON(r != 1);
1844	base = kmap_atomic(page);
1845	set_bit(bit, base);
1846	kunmap_atomic(base);
1847	unpin_user_pages_dirty_lock(&page, 1, true);
1848	return 0;
1849}
1850
1851static int log_write(void __user *log_base,
1852		     u64 write_address, u64 write_length)
1853{
1854	u64 write_page = write_address / VHOST_PAGE_SIZE;
1855	int r;
1856
1857	if (!write_length)
1858		return 0;
1859	write_length += write_address % VHOST_PAGE_SIZE;
1860	for (;;) {
1861		u64 base = (u64)(unsigned long)log_base;
1862		u64 log = base + write_page / 8;
1863		int bit = write_page % 8;
1864		if ((u64)(unsigned long)log != log)
1865			return -EFAULT;
1866		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1867		if (r < 0)
1868			return r;
1869		if (write_length <= VHOST_PAGE_SIZE)
1870			break;
1871		write_length -= VHOST_PAGE_SIZE;
1872		write_page += 1;
1873	}
1874	return r;
1875}
1876
1877static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1878{
1879	struct vhost_iotlb *umem = vq->umem;
1880	struct vhost_iotlb_map *u;
1881	u64 start, end, l, min;
1882	int r;
1883	bool hit = false;
1884
1885	while (len) {
1886		min = len;
1887		/* More than one GPAs can be mapped into a single HVA. So
1888		 * iterate all possible umems here to be safe.
1889		 */
1890		list_for_each_entry(u, &umem->list, link) {
1891			if (u->addr > hva - 1 + len ||
1892			    u->addr - 1 + u->size < hva)
1893				continue;
1894			start = max(u->addr, hva);
1895			end = min(u->addr - 1 + u->size, hva - 1 + len);
1896			l = end - start + 1;
1897			r = log_write(vq->log_base,
1898				      u->start + start - u->addr,
1899				      l);
1900			if (r < 0)
1901				return r;
1902			hit = true;
1903			min = min(l, min);
1904		}
1905
1906		if (!hit)
1907			return -EFAULT;
1908
1909		len -= min;
1910		hva += min;
1911	}
1912
1913	return 0;
1914}
1915
1916static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1917{
1918	struct iovec *iov = vq->log_iov;
1919	int i, ret;
1920
1921	if (!vq->iotlb)
1922		return log_write(vq->log_base, vq->log_addr + used_offset, len);
1923
1924	ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1925			     len, iov, 64, VHOST_ACCESS_WO);
1926	if (ret < 0)
1927		return ret;
1928
1929	for (i = 0; i < ret; i++) {
1930		ret = log_write_hva(vq,	(uintptr_t)iov[i].iov_base,
1931				    iov[i].iov_len);
1932		if (ret)
1933			return ret;
1934	}
1935
1936	return 0;
1937}
1938
1939int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1940		    unsigned int log_num, u64 len, struct iovec *iov, int count)
1941{
1942	int i, r;
1943
1944	/* Make sure data written is seen before log. */
1945	smp_wmb();
1946
1947	if (vq->iotlb) {
1948		for (i = 0; i < count; i++) {
1949			r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1950					  iov[i].iov_len);
1951			if (r < 0)
1952				return r;
1953		}
1954		return 0;
1955	}
1956
1957	for (i = 0; i < log_num; ++i) {
1958		u64 l = min(log[i].len, len);
1959		r = log_write(vq->log_base, log[i].addr, l);
1960		if (r < 0)
1961			return r;
1962		len -= l;
1963		if (!len) {
1964			if (vq->log_ctx)
1965				eventfd_signal(vq->log_ctx, 1);
1966			return 0;
1967		}
1968	}
1969	/* Length written exceeds what we have stored. This is a bug. */
1970	BUG();
1971	return 0;
1972}
1973EXPORT_SYMBOL_GPL(vhost_log_write);
1974
1975static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1976{
1977	void __user *used;
1978	if (vhost_put_used_flags(vq))
1979		return -EFAULT;
1980	if (unlikely(vq->log_used)) {
1981		/* Make sure the flag is seen before log. */
1982		smp_wmb();
1983		/* Log used flag write. */
1984		used = &vq->used->flags;
1985		log_used(vq, (used - (void __user *)vq->used),
1986			 sizeof vq->used->flags);
1987		if (vq->log_ctx)
1988			eventfd_signal(vq->log_ctx, 1);
1989	}
1990	return 0;
1991}
1992
1993static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1994{
1995	if (vhost_put_avail_event(vq))
1996		return -EFAULT;
1997	if (unlikely(vq->log_used)) {
1998		void __user *used;
1999		/* Make sure the event is seen before log. */
2000		smp_wmb();
2001		/* Log avail event write */
2002		used = vhost_avail_event(vq);
2003		log_used(vq, (used - (void __user *)vq->used),
2004			 sizeof *vhost_avail_event(vq));
2005		if (vq->log_ctx)
2006			eventfd_signal(vq->log_ctx, 1);
2007	}
2008	return 0;
2009}
2010
2011int vhost_vq_init_access(struct vhost_virtqueue *vq)
2012{
2013	__virtio16 last_used_idx;
2014	int r;
2015	bool is_le = vq->is_le;
2016
2017	if (!vq->private_data)
2018		return 0;
2019
2020	vhost_init_is_le(vq);
2021
2022	r = vhost_update_used_flags(vq);
2023	if (r)
2024		goto err;
2025	vq->signalled_used_valid = false;
2026	if (!vq->iotlb &&
2027	    !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2028		r = -EFAULT;
2029		goto err;
2030	}
2031	r = vhost_get_used_idx(vq, &last_used_idx);
2032	if (r) {
2033		vq_err(vq, "Can't access used idx at %p\n",
2034		       &vq->used->idx);
2035		goto err;
2036	}
2037	vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2038	return 0;
2039
2040err:
2041	vq->is_le = is_le;
2042	return r;
2043}
2044EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2045
2046static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2047			  struct iovec iov[], int iov_size, int access)
2048{
2049	const struct vhost_iotlb_map *map;
2050	struct vhost_dev *dev = vq->dev;
2051	struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2052	struct iovec *_iov;
2053	u64 s = 0, last = addr + len - 1;
2054	int ret = 0;
2055
2056	while ((u64)len > s) {
2057		u64 size;
2058		if (unlikely(ret >= iov_size)) {
2059			ret = -ENOBUFS;
2060			break;
2061		}
2062
2063		map = vhost_iotlb_itree_first(umem, addr, last);
2064		if (map == NULL || map->start > addr) {
2065			if (umem != dev->iotlb) {
2066				ret = -EFAULT;
2067				break;
2068			}
2069			ret = -EAGAIN;
2070			break;
2071		} else if (!(map->perm & access)) {
2072			ret = -EPERM;
2073			break;
2074		}
2075
2076		_iov = iov + ret;
2077		size = map->size - addr + map->start;
2078		_iov->iov_len = min((u64)len - s, size);
2079		_iov->iov_base = (void __user *)(unsigned long)
2080				 (map->addr + addr - map->start);
2081		s += size;
2082		addr += size;
2083		++ret;
2084	}
2085
2086	if (ret == -EAGAIN)
2087		vhost_iotlb_miss(vq, addr, access);
2088	return ret;
2089}
2090
2091/* Each buffer in the virtqueues is actually a chain of descriptors.  This
2092 * function returns the next descriptor in the chain,
2093 * or -1U if we're at the end. */
2094static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2095{
2096	unsigned int next;
2097
2098	/* If this descriptor says it doesn't chain, we're done. */
2099	if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2100		return -1U;
2101
2102	/* Check they're not leading us off end of descriptors. */
2103	next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2104	return next;
2105}
2106
2107static int get_indirect(struct vhost_virtqueue *vq,
2108			struct iovec iov[], unsigned int iov_size,
2109			unsigned int *out_num, unsigned int *in_num,
2110			struct vhost_log *log, unsigned int *log_num,
2111			struct vring_desc *indirect)
2112{
2113	struct vring_desc desc;
2114	unsigned int i = 0, count, found = 0;
2115	u32 len = vhost32_to_cpu(vq, indirect->len);
2116	struct iov_iter from;
2117	int ret, access;
2118
2119	/* Sanity check */
2120	if (unlikely(len % sizeof desc)) {
2121		vq_err(vq, "Invalid length in indirect descriptor: "
2122		       "len 0x%llx not multiple of 0x%zx\n",
2123		       (unsigned long long)len,
2124		       sizeof desc);
2125		return -EINVAL;
2126	}
2127
2128	ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2129			     UIO_MAXIOV, VHOST_ACCESS_RO);
2130	if (unlikely(ret < 0)) {
2131		if (ret != -EAGAIN)
2132			vq_err(vq, "Translation failure %d in indirect.\n", ret);
2133		return ret;
2134	}
2135	iov_iter_init(&from, READ, vq->indirect, ret, len);
2136	count = len / sizeof desc;
2137	/* Buffers are chained via a 16 bit next field, so
2138	 * we can have at most 2^16 of these. */
2139	if (unlikely(count > USHRT_MAX + 1)) {
2140		vq_err(vq, "Indirect buffer length too big: %d\n",
2141		       indirect->len);
2142		return -E2BIG;
2143	}
2144
2145	do {
2146		unsigned iov_count = *in_num + *out_num;
2147		if (unlikely(++found > count)) {
2148			vq_err(vq, "Loop detected: last one at %u "
2149			       "indirect size %u\n",
2150			       i, count);
2151			return -EINVAL;
2152		}
2153		if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2154			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2155			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2156			return -EINVAL;
2157		}
2158		if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2159			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2160			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2161			return -EINVAL;
2162		}
2163
2164		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2165			access = VHOST_ACCESS_WO;
2166		else
2167			access = VHOST_ACCESS_RO;
2168
2169		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2170				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
2171				     iov_size - iov_count, access);
2172		if (unlikely(ret < 0)) {
2173			if (ret != -EAGAIN)
2174				vq_err(vq, "Translation failure %d indirect idx %d\n",
2175					ret, i);
2176			return ret;
2177		}
2178		/* If this is an input descriptor, increment that count. */
2179		if (access == VHOST_ACCESS_WO) {
2180			*in_num += ret;
2181			if (unlikely(log && ret)) {
2182				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2183				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2184				++*log_num;
2185			}
2186		} else {
2187			/* If it's an output descriptor, they're all supposed
2188			 * to come before any input descriptors. */
2189			if (unlikely(*in_num)) {
2190				vq_err(vq, "Indirect descriptor "
2191				       "has out after in: idx %d\n", i);
2192				return -EINVAL;
2193			}
2194			*out_num += ret;
2195		}
2196	} while ((i = next_desc(vq, &desc)) != -1);
2197	return 0;
2198}
2199
2200/* This looks in the virtqueue and for the first available buffer, and converts
2201 * it to an iovec for convenient access.  Since descriptors consist of some
2202 * number of output then some number of input descriptors, it's actually two
2203 * iovecs, but we pack them into one and note how many of each there were.
2204 *
2205 * This function returns the descriptor number found, or vq->num (which is
2206 * never a valid descriptor number) if none was found.  A negative code is
2207 * returned on error. */
2208int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2209		      struct iovec iov[], unsigned int iov_size,
2210		      unsigned int *out_num, unsigned int *in_num,
2211		      struct vhost_log *log, unsigned int *log_num)
2212{
2213	struct vring_desc desc;
2214	unsigned int i, head, found = 0;
2215	u16 last_avail_idx;
2216	__virtio16 avail_idx;
2217	__virtio16 ring_head;
2218	int ret, access;
2219
2220	/* Check it isn't doing very strange things with descriptor numbers. */
2221	last_avail_idx = vq->last_avail_idx;
2222
2223	if (vq->avail_idx == vq->last_avail_idx) {
2224		if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2225			vq_err(vq, "Failed to access avail idx at %p\n",
2226				&vq->avail->idx);
2227			return -EFAULT;
2228		}
2229		vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2230
2231		if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2232			vq_err(vq, "Guest moved used index from %u to %u",
2233				last_avail_idx, vq->avail_idx);
2234			return -EFAULT;
2235		}
2236
2237		/* If there's nothing new since last we looked, return
2238		 * invalid.
2239		 */
2240		if (vq->avail_idx == last_avail_idx)
2241			return vq->num;
2242
2243		/* Only get avail ring entries after they have been
2244		 * exposed by guest.
2245		 */
2246		smp_rmb();
2247	}
2248
2249	/* Grab the next descriptor number they're advertising, and increment
2250	 * the index we've seen. */
2251	if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2252		vq_err(vq, "Failed to read head: idx %d address %p\n",
2253		       last_avail_idx,
2254		       &vq->avail->ring[last_avail_idx % vq->num]);
2255		return -EFAULT;
2256	}
2257
2258	head = vhost16_to_cpu(vq, ring_head);
2259
2260	/* If their number is silly, that's an error. */
2261	if (unlikely(head >= vq->num)) {
2262		vq_err(vq, "Guest says index %u > %u is available",
2263		       head, vq->num);
2264		return -EINVAL;
2265	}
2266
2267	/* When we start there are none of either input nor output. */
2268	*out_num = *in_num = 0;
2269	if (unlikely(log))
2270		*log_num = 0;
2271
2272	i = head;
2273	do {
2274		unsigned iov_count = *in_num + *out_num;
2275		if (unlikely(i >= vq->num)) {
2276			vq_err(vq, "Desc index is %u > %u, head = %u",
2277			       i, vq->num, head);
2278			return -EINVAL;
2279		}
2280		if (unlikely(++found > vq->num)) {
2281			vq_err(vq, "Loop detected: last one at %u "
2282			       "vq size %u head %u\n",
2283			       i, vq->num, head);
2284			return -EINVAL;
2285		}
2286		ret = vhost_get_desc(vq, &desc, i);
2287		if (unlikely(ret)) {
2288			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2289			       i, vq->desc + i);
2290			return -EFAULT;
2291		}
2292		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2293			ret = get_indirect(vq, iov, iov_size,
2294					   out_num, in_num,
2295					   log, log_num, &desc);
2296			if (unlikely(ret < 0)) {
2297				if (ret != -EAGAIN)
2298					vq_err(vq, "Failure detected "
2299						"in indirect descriptor at idx %d\n", i);
2300				return ret;
2301			}
2302			continue;
2303		}
2304
2305		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2306			access = VHOST_ACCESS_WO;
2307		else
2308			access = VHOST_ACCESS_RO;
2309		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2310				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
2311				     iov_size - iov_count, access);
2312		if (unlikely(ret < 0)) {
2313			if (ret != -EAGAIN)
2314				vq_err(vq, "Translation failure %d descriptor idx %d\n",
2315					ret, i);
2316			return ret;
2317		}
2318		if (access == VHOST_ACCESS_WO) {
2319			/* If this is an input descriptor,
2320			 * increment that count. */
2321			*in_num += ret;
2322			if (unlikely(log && ret)) {
2323				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2324				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2325				++*log_num;
2326			}
2327		} else {
2328			/* If it's an output descriptor, they're all supposed
2329			 * to come before any input descriptors. */
2330			if (unlikely(*in_num)) {
2331				vq_err(vq, "Descriptor has out after in: "
2332				       "idx %d\n", i);
2333				return -EINVAL;
2334			}
2335			*out_num += ret;
2336		}
2337	} while ((i = next_desc(vq, &desc)) != -1);
2338
2339	/* On success, increment avail index. */
2340	vq->last_avail_idx++;
2341
2342	/* Assume notifications from guest are disabled at this point,
2343	 * if they aren't we would need to update avail_event index. */
2344	BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2345	return head;
2346}
2347EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2348
2349/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2350void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2351{
2352	vq->last_avail_idx -= n;
2353}
2354EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2355
2356/* After we've used one of their buffers, we tell them about it.  We'll then
2357 * want to notify the guest, using eventfd. */
2358int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2359{
2360	struct vring_used_elem heads = {
2361		cpu_to_vhost32(vq, head),
2362		cpu_to_vhost32(vq, len)
2363	};
2364
2365	return vhost_add_used_n(vq, &heads, 1);
2366}
2367EXPORT_SYMBOL_GPL(vhost_add_used);
2368
2369static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2370			    struct vring_used_elem *heads,
2371			    unsigned count)
2372{
2373	vring_used_elem_t __user *used;
2374	u16 old, new;
2375	int start;
2376
2377	start = vq->last_used_idx & (vq->num - 1);
2378	used = vq->used->ring + start;
2379	if (vhost_put_used(vq, heads, start, count)) {
2380		vq_err(vq, "Failed to write used");
2381		return -EFAULT;
2382	}
2383	if (unlikely(vq->log_used)) {
2384		/* Make sure data is seen before log. */
2385		smp_wmb();
2386		/* Log used ring entry write. */
2387		log_used(vq, ((void __user *)used - (void __user *)vq->used),
2388			 count * sizeof *used);
2389	}
2390	old = vq->last_used_idx;
2391	new = (vq->last_used_idx += count);
2392	/* If the driver never bothers to signal in a very long while,
2393	 * used index might wrap around. If that happens, invalidate
2394	 * signalled_used index we stored. TODO: make sure driver
2395	 * signals at least once in 2^16 and remove this. */
2396	if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2397		vq->signalled_used_valid = false;
2398	return 0;
2399}
2400
2401/* After we've used one of their buffers, we tell them about it.  We'll then
2402 * want to notify the guest, using eventfd. */
2403int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2404		     unsigned count)
2405{
2406	int start, n, r;
2407
2408	start = vq->last_used_idx & (vq->num - 1);
2409	n = vq->num - start;
2410	if (n < count) {
2411		r = __vhost_add_used_n(vq, heads, n);
2412		if (r < 0)
2413			return r;
2414		heads += n;
2415		count -= n;
2416	}
2417	r = __vhost_add_used_n(vq, heads, count);
2418
2419	/* Make sure buffer is written before we update index. */
2420	smp_wmb();
2421	if (vhost_put_used_idx(vq)) {
2422		vq_err(vq, "Failed to increment used idx");
2423		return -EFAULT;
2424	}
2425	if (unlikely(vq->log_used)) {
2426		/* Make sure used idx is seen before log. */
2427		smp_wmb();
2428		/* Log used index update. */
2429		log_used(vq, offsetof(struct vring_used, idx),
2430			 sizeof vq->used->idx);
2431		if (vq->log_ctx)
2432			eventfd_signal(vq->log_ctx, 1);
2433	}
2434	return r;
2435}
2436EXPORT_SYMBOL_GPL(vhost_add_used_n);
2437
2438static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2439{
2440	__u16 old, new;
2441	__virtio16 event;
2442	bool v;
2443	/* Flush out used index updates. This is paired
2444	 * with the barrier that the Guest executes when enabling
2445	 * interrupts. */
2446	smp_mb();
2447
2448	if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2449	    unlikely(vq->avail_idx == vq->last_avail_idx))
2450		return true;
2451
2452	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2453		__virtio16 flags;
2454		if (vhost_get_avail_flags(vq, &flags)) {
2455			vq_err(vq, "Failed to get flags");
2456			return true;
2457		}
2458		return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2459	}
2460	old = vq->signalled_used;
2461	v = vq->signalled_used_valid;
2462	new = vq->signalled_used = vq->last_used_idx;
2463	vq->signalled_used_valid = true;
2464
2465	if (unlikely(!v))
2466		return true;
2467
2468	if (vhost_get_used_event(vq, &event)) {
2469		vq_err(vq, "Failed to get used event idx");
2470		return true;
2471	}
2472	return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2473}
2474
2475/* This actually signals the guest, using eventfd. */
2476void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2477{
2478	/* Signal the Guest tell them we used something up. */
2479	if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2480		eventfd_signal(vq->call_ctx.ctx, 1);
2481}
2482EXPORT_SYMBOL_GPL(vhost_signal);
2483
2484/* And here's the combo meal deal.  Supersize me! */
2485void vhost_add_used_and_signal(struct vhost_dev *dev,
2486			       struct vhost_virtqueue *vq,
2487			       unsigned int head, int len)
2488{
2489	vhost_add_used(vq, head, len);
2490	vhost_signal(dev, vq);
2491}
2492EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2493
2494/* multi-buffer version of vhost_add_used_and_signal */
2495void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2496				 struct vhost_virtqueue *vq,
2497				 struct vring_used_elem *heads, unsigned count)
2498{
2499	vhost_add_used_n(vq, heads, count);
2500	vhost_signal(dev, vq);
2501}
2502EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2503
2504/* return true if we're sure that avaiable ring is empty */
2505bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2506{
2507	__virtio16 avail_idx;
2508	int r;
2509
2510	if (vq->avail_idx != vq->last_avail_idx)
2511		return false;
2512
2513	r = vhost_get_avail_idx(vq, &avail_idx);
2514	if (unlikely(r))
2515		return false;
2516	vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2517
2518	return vq->avail_idx == vq->last_avail_idx;
2519}
2520EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2521
2522/* OK, now we need to know about added descriptors. */
2523bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2524{
2525	__virtio16 avail_idx;
2526	int r;
2527
2528	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2529		return false;
2530	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2531	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2532		r = vhost_update_used_flags(vq);
2533		if (r) {
2534			vq_err(vq, "Failed to enable notification at %p: %d\n",
2535			       &vq->used->flags, r);
2536			return false;
2537		}
2538	} else {
2539		r = vhost_update_avail_event(vq, vq->avail_idx);
2540		if (r) {
2541			vq_err(vq, "Failed to update avail event index at %p: %d\n",
2542			       vhost_avail_event(vq), r);
2543			return false;
2544		}
2545	}
2546	/* They could have slipped one in as we were doing that: make
2547	 * sure it's written, then check again. */
2548	smp_mb();
2549	r = vhost_get_avail_idx(vq, &avail_idx);
2550	if (r) {
2551		vq_err(vq, "Failed to check avail idx at %p: %d\n",
2552		       &vq->avail->idx, r);
2553		return false;
2554	}
2555
2556	return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2557}
2558EXPORT_SYMBOL_GPL(vhost_enable_notify);
2559
2560/* We don't need to be notified again. */
2561void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2562{
2563	int r;
2564
2565	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2566		return;
2567	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2568	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2569		r = vhost_update_used_flags(vq);
2570		if (r)
2571			vq_err(vq, "Failed to disable notification at %p: %d\n",
2572			       &vq->used->flags, r);
2573	}
2574}
2575EXPORT_SYMBOL_GPL(vhost_disable_notify);
2576
2577/* Create a new message. */
2578struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2579{
2580	/* Make sure all padding within the structure is initialized. */
2581	struct vhost_msg_node *node = kzalloc(sizeof(*node), GFP_KERNEL);
2582	if (!node)
2583		return NULL;
2584
2585	node->vq = vq;
2586	node->msg.type = type;
2587	return node;
2588}
2589EXPORT_SYMBOL_GPL(vhost_new_msg);
2590
2591void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2592		       struct vhost_msg_node *node)
2593{
2594	spin_lock(&dev->iotlb_lock);
2595	list_add_tail(&node->node, head);
2596	spin_unlock(&dev->iotlb_lock);
2597
2598	wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2599}
2600EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2601
2602struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2603					 struct list_head *head)
2604{
2605	struct vhost_msg_node *node = NULL;
2606
2607	spin_lock(&dev->iotlb_lock);
2608	if (!list_empty(head)) {
2609		node = list_first_entry(head, struct vhost_msg_node,
2610					node);
2611		list_del(&node->node);
2612	}
2613	spin_unlock(&dev->iotlb_lock);
2614
2615	return node;
2616}
2617EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2618
2619void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2620{
2621	struct vhost_virtqueue *vq;
2622	int i;
2623
2624	mutex_lock(&dev->mutex);
2625	for (i = 0; i < dev->nvqs; ++i) {
2626		vq = dev->vqs[i];
2627		mutex_lock(&vq->mutex);
2628		vq->acked_backend_features = features;
2629		mutex_unlock(&vq->mutex);
2630	}
2631	mutex_unlock(&dev->mutex);
2632}
2633EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2634
2635static int __init vhost_init(void)
2636{
2637	return 0;
2638}
2639
2640static void __exit vhost_exit(void)
2641{
2642}
2643
2644module_init(vhost_init);
2645module_exit(vhost_exit);
2646
2647MODULE_VERSION("0.0.1");
2648MODULE_LICENSE("GPL v2");
2649MODULE_AUTHOR("Michael S. Tsirkin");
2650MODULE_DESCRIPTION("Host kernel accelerator for virtio");
2651