xref: /openbmc/linux/drivers/vhost/vhost.c (revision e23feb16)
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/rcupdate.h>
22 #include <linux/poll.h>
23 #include <linux/file.h>
24 #include <linux/highmem.h>
25 #include <linux/slab.h>
26 #include <linux/kthread.h>
27 #include <linux/cgroup.h>
28 #include <linux/module.h>
29 
30 #include "vhost.h"
31 
32 enum {
33 	VHOST_MEMORY_MAX_NREGIONS = 64,
34 	VHOST_MEMORY_F_LOG = 0x1,
35 };
36 
37 #define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num])
38 #define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num])
39 
40 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
41 			    poll_table *pt)
42 {
43 	struct vhost_poll *poll;
44 
45 	poll = container_of(pt, struct vhost_poll, table);
46 	poll->wqh = wqh;
47 	add_wait_queue(wqh, &poll->wait);
48 }
49 
50 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
51 			     void *key)
52 {
53 	struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
54 
55 	if (!((unsigned long)key & poll->mask))
56 		return 0;
57 
58 	vhost_poll_queue(poll);
59 	return 0;
60 }
61 
62 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
63 {
64 	INIT_LIST_HEAD(&work->node);
65 	work->fn = fn;
66 	init_waitqueue_head(&work->done);
67 	work->flushing = 0;
68 	work->queue_seq = work->done_seq = 0;
69 }
70 EXPORT_SYMBOL_GPL(vhost_work_init);
71 
72 /* Init poll structure */
73 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
74 		     unsigned long mask, struct vhost_dev *dev)
75 {
76 	init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
77 	init_poll_funcptr(&poll->table, vhost_poll_func);
78 	poll->mask = mask;
79 	poll->dev = dev;
80 	poll->wqh = NULL;
81 
82 	vhost_work_init(&poll->work, fn);
83 }
84 EXPORT_SYMBOL_GPL(vhost_poll_init);
85 
86 /* Start polling a file. We add ourselves to file's wait queue. The caller must
87  * keep a reference to a file until after vhost_poll_stop is called. */
88 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
89 {
90 	unsigned long mask;
91 	int ret = 0;
92 
93 	if (poll->wqh)
94 		return 0;
95 
96 	mask = file->f_op->poll(file, &poll->table);
97 	if (mask)
98 		vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
99 	if (mask & POLLERR) {
100 		if (poll->wqh)
101 			remove_wait_queue(poll->wqh, &poll->wait);
102 		ret = -EINVAL;
103 	}
104 
105 	return ret;
106 }
107 EXPORT_SYMBOL_GPL(vhost_poll_start);
108 
109 /* Stop polling a file. After this function returns, it becomes safe to drop the
110  * file reference. You must also flush afterwards. */
111 void vhost_poll_stop(struct vhost_poll *poll)
112 {
113 	if (poll->wqh) {
114 		remove_wait_queue(poll->wqh, &poll->wait);
115 		poll->wqh = NULL;
116 	}
117 }
118 EXPORT_SYMBOL_GPL(vhost_poll_stop);
119 
120 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
121 				unsigned seq)
122 {
123 	int left;
124 
125 	spin_lock_irq(&dev->work_lock);
126 	left = seq - work->done_seq;
127 	spin_unlock_irq(&dev->work_lock);
128 	return left <= 0;
129 }
130 
131 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
132 {
133 	unsigned seq;
134 	int flushing;
135 
136 	spin_lock_irq(&dev->work_lock);
137 	seq = work->queue_seq;
138 	work->flushing++;
139 	spin_unlock_irq(&dev->work_lock);
140 	wait_event(work->done, vhost_work_seq_done(dev, work, seq));
141 	spin_lock_irq(&dev->work_lock);
142 	flushing = --work->flushing;
143 	spin_unlock_irq(&dev->work_lock);
144 	BUG_ON(flushing < 0);
145 }
146 EXPORT_SYMBOL_GPL(vhost_work_flush);
147 
148 /* Flush any work that has been scheduled. When calling this, don't hold any
149  * locks that are also used by the callback. */
150 void vhost_poll_flush(struct vhost_poll *poll)
151 {
152 	vhost_work_flush(poll->dev, &poll->work);
153 }
154 EXPORT_SYMBOL_GPL(vhost_poll_flush);
155 
156 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
157 {
158 	unsigned long flags;
159 
160 	spin_lock_irqsave(&dev->work_lock, flags);
161 	if (list_empty(&work->node)) {
162 		list_add_tail(&work->node, &dev->work_list);
163 		work->queue_seq++;
164 		spin_unlock_irqrestore(&dev->work_lock, flags);
165 		wake_up_process(dev->worker);
166 	} else {
167 		spin_unlock_irqrestore(&dev->work_lock, flags);
168 	}
169 }
170 EXPORT_SYMBOL_GPL(vhost_work_queue);
171 
172 void vhost_poll_queue(struct vhost_poll *poll)
173 {
174 	vhost_work_queue(poll->dev, &poll->work);
175 }
176 EXPORT_SYMBOL_GPL(vhost_poll_queue);
177 
178 static void vhost_vq_reset(struct vhost_dev *dev,
179 			   struct vhost_virtqueue *vq)
180 {
181 	vq->num = 1;
182 	vq->desc = NULL;
183 	vq->avail = NULL;
184 	vq->used = NULL;
185 	vq->last_avail_idx = 0;
186 	vq->avail_idx = 0;
187 	vq->last_used_idx = 0;
188 	vq->signalled_used = 0;
189 	vq->signalled_used_valid = false;
190 	vq->used_flags = 0;
191 	vq->log_used = false;
192 	vq->log_addr = -1ull;
193 	vq->private_data = NULL;
194 	vq->log_base = NULL;
195 	vq->error_ctx = NULL;
196 	vq->error = NULL;
197 	vq->kick = NULL;
198 	vq->call_ctx = NULL;
199 	vq->call = NULL;
200 	vq->log_ctx = NULL;
201 }
202 
203 static int vhost_worker(void *data)
204 {
205 	struct vhost_dev *dev = data;
206 	struct vhost_work *work = NULL;
207 	unsigned uninitialized_var(seq);
208 	mm_segment_t oldfs = get_fs();
209 
210 	set_fs(USER_DS);
211 	use_mm(dev->mm);
212 
213 	for (;;) {
214 		/* mb paired w/ kthread_stop */
215 		set_current_state(TASK_INTERRUPTIBLE);
216 
217 		spin_lock_irq(&dev->work_lock);
218 		if (work) {
219 			work->done_seq = seq;
220 			if (work->flushing)
221 				wake_up_all(&work->done);
222 		}
223 
224 		if (kthread_should_stop()) {
225 			spin_unlock_irq(&dev->work_lock);
226 			__set_current_state(TASK_RUNNING);
227 			break;
228 		}
229 		if (!list_empty(&dev->work_list)) {
230 			work = list_first_entry(&dev->work_list,
231 						struct vhost_work, node);
232 			list_del_init(&work->node);
233 			seq = work->queue_seq;
234 		} else
235 			work = NULL;
236 		spin_unlock_irq(&dev->work_lock);
237 
238 		if (work) {
239 			__set_current_state(TASK_RUNNING);
240 			work->fn(work);
241 			if (need_resched())
242 				schedule();
243 		} else
244 			schedule();
245 
246 	}
247 	unuse_mm(dev->mm);
248 	set_fs(oldfs);
249 	return 0;
250 }
251 
252 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
253 {
254 	kfree(vq->indirect);
255 	vq->indirect = NULL;
256 	kfree(vq->log);
257 	vq->log = NULL;
258 	kfree(vq->heads);
259 	vq->heads = NULL;
260 }
261 
262 /* Helper to allocate iovec buffers for all vqs. */
263 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
264 {
265 	struct vhost_virtqueue *vq;
266 	int i;
267 
268 	for (i = 0; i < dev->nvqs; ++i) {
269 		vq = dev->vqs[i];
270 		vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
271 				       GFP_KERNEL);
272 		vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
273 		vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
274 		if (!vq->indirect || !vq->log || !vq->heads)
275 			goto err_nomem;
276 	}
277 	return 0;
278 
279 err_nomem:
280 	for (; i >= 0; --i)
281 		vhost_vq_free_iovecs(dev->vqs[i]);
282 	return -ENOMEM;
283 }
284 
285 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
286 {
287 	int i;
288 
289 	for (i = 0; i < dev->nvqs; ++i)
290 		vhost_vq_free_iovecs(dev->vqs[i]);
291 }
292 
293 long vhost_dev_init(struct vhost_dev *dev,
294 		    struct vhost_virtqueue **vqs, int nvqs)
295 {
296 	struct vhost_virtqueue *vq;
297 	int i;
298 
299 	dev->vqs = vqs;
300 	dev->nvqs = nvqs;
301 	mutex_init(&dev->mutex);
302 	dev->log_ctx = NULL;
303 	dev->log_file = NULL;
304 	dev->memory = NULL;
305 	dev->mm = NULL;
306 	spin_lock_init(&dev->work_lock);
307 	INIT_LIST_HEAD(&dev->work_list);
308 	dev->worker = NULL;
309 
310 	for (i = 0; i < dev->nvqs; ++i) {
311 		vq = dev->vqs[i];
312 		vq->log = NULL;
313 		vq->indirect = NULL;
314 		vq->heads = NULL;
315 		vq->dev = dev;
316 		mutex_init(&vq->mutex);
317 		vhost_vq_reset(dev, vq);
318 		if (vq->handle_kick)
319 			vhost_poll_init(&vq->poll, vq->handle_kick,
320 					POLLIN, dev);
321 	}
322 
323 	return 0;
324 }
325 EXPORT_SYMBOL_GPL(vhost_dev_init);
326 
327 /* Caller should have device mutex */
328 long vhost_dev_check_owner(struct vhost_dev *dev)
329 {
330 	/* Are you the owner? If not, I don't think you mean to do that */
331 	return dev->mm == current->mm ? 0 : -EPERM;
332 }
333 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
334 
335 struct vhost_attach_cgroups_struct {
336 	struct vhost_work work;
337 	struct task_struct *owner;
338 	int ret;
339 };
340 
341 static void vhost_attach_cgroups_work(struct vhost_work *work)
342 {
343 	struct vhost_attach_cgroups_struct *s;
344 
345 	s = container_of(work, struct vhost_attach_cgroups_struct, work);
346 	s->ret = cgroup_attach_task_all(s->owner, current);
347 }
348 
349 static int vhost_attach_cgroups(struct vhost_dev *dev)
350 {
351 	struct vhost_attach_cgroups_struct attach;
352 
353 	attach.owner = current;
354 	vhost_work_init(&attach.work, vhost_attach_cgroups_work);
355 	vhost_work_queue(dev, &attach.work);
356 	vhost_work_flush(dev, &attach.work);
357 	return attach.ret;
358 }
359 
360 /* Caller should have device mutex */
361 bool vhost_dev_has_owner(struct vhost_dev *dev)
362 {
363 	return dev->mm;
364 }
365 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
366 
367 /* Caller should have device mutex */
368 long vhost_dev_set_owner(struct vhost_dev *dev)
369 {
370 	struct task_struct *worker;
371 	int err;
372 
373 	/* Is there an owner already? */
374 	if (vhost_dev_has_owner(dev)) {
375 		err = -EBUSY;
376 		goto err_mm;
377 	}
378 
379 	/* No owner, become one */
380 	dev->mm = get_task_mm(current);
381 	worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
382 	if (IS_ERR(worker)) {
383 		err = PTR_ERR(worker);
384 		goto err_worker;
385 	}
386 
387 	dev->worker = worker;
388 	wake_up_process(worker);	/* avoid contributing to loadavg */
389 
390 	err = vhost_attach_cgroups(dev);
391 	if (err)
392 		goto err_cgroup;
393 
394 	err = vhost_dev_alloc_iovecs(dev);
395 	if (err)
396 		goto err_cgroup;
397 
398 	return 0;
399 err_cgroup:
400 	kthread_stop(worker);
401 	dev->worker = NULL;
402 err_worker:
403 	if (dev->mm)
404 		mmput(dev->mm);
405 	dev->mm = NULL;
406 err_mm:
407 	return err;
408 }
409 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
410 
411 struct vhost_memory *vhost_dev_reset_owner_prepare(void)
412 {
413 	return kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
414 }
415 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
416 
417 /* Caller should have device mutex */
418 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_memory *memory)
419 {
420 	vhost_dev_cleanup(dev, true);
421 
422 	/* Restore memory to default empty mapping. */
423 	memory->nregions = 0;
424 	RCU_INIT_POINTER(dev->memory, memory);
425 }
426 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
427 
428 void vhost_dev_stop(struct vhost_dev *dev)
429 {
430 	int i;
431 
432 	for (i = 0; i < dev->nvqs; ++i) {
433 		if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
434 			vhost_poll_stop(&dev->vqs[i]->poll);
435 			vhost_poll_flush(&dev->vqs[i]->poll);
436 		}
437 	}
438 }
439 EXPORT_SYMBOL_GPL(vhost_dev_stop);
440 
441 /* Caller should have device mutex if and only if locked is set */
442 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
443 {
444 	int i;
445 
446 	for (i = 0; i < dev->nvqs; ++i) {
447 		if (dev->vqs[i]->error_ctx)
448 			eventfd_ctx_put(dev->vqs[i]->error_ctx);
449 		if (dev->vqs[i]->error)
450 			fput(dev->vqs[i]->error);
451 		if (dev->vqs[i]->kick)
452 			fput(dev->vqs[i]->kick);
453 		if (dev->vqs[i]->call_ctx)
454 			eventfd_ctx_put(dev->vqs[i]->call_ctx);
455 		if (dev->vqs[i]->call)
456 			fput(dev->vqs[i]->call);
457 		vhost_vq_reset(dev, dev->vqs[i]);
458 	}
459 	vhost_dev_free_iovecs(dev);
460 	if (dev->log_ctx)
461 		eventfd_ctx_put(dev->log_ctx);
462 	dev->log_ctx = NULL;
463 	if (dev->log_file)
464 		fput(dev->log_file);
465 	dev->log_file = NULL;
466 	/* No one will access memory at this point */
467 	kfree(rcu_dereference_protected(dev->memory,
468 					locked ==
469 						lockdep_is_held(&dev->mutex)));
470 	RCU_INIT_POINTER(dev->memory, NULL);
471 	WARN_ON(!list_empty(&dev->work_list));
472 	if (dev->worker) {
473 		kthread_stop(dev->worker);
474 		dev->worker = NULL;
475 	}
476 	if (dev->mm)
477 		mmput(dev->mm);
478 	dev->mm = NULL;
479 }
480 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
481 
482 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
483 {
484 	u64 a = addr / VHOST_PAGE_SIZE / 8;
485 
486 	/* Make sure 64 bit math will not overflow. */
487 	if (a > ULONG_MAX - (unsigned long)log_base ||
488 	    a + (unsigned long)log_base > ULONG_MAX)
489 		return 0;
490 
491 	return access_ok(VERIFY_WRITE, log_base + a,
492 			 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
493 }
494 
495 /* Caller should have vq mutex and device mutex. */
496 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
497 			       int log_all)
498 {
499 	int i;
500 
501 	if (!mem)
502 		return 0;
503 
504 	for (i = 0; i < mem->nregions; ++i) {
505 		struct vhost_memory_region *m = mem->regions + i;
506 		unsigned long a = m->userspace_addr;
507 		if (m->memory_size > ULONG_MAX)
508 			return 0;
509 		else if (!access_ok(VERIFY_WRITE, (void __user *)a,
510 				    m->memory_size))
511 			return 0;
512 		else if (log_all && !log_access_ok(log_base,
513 						   m->guest_phys_addr,
514 						   m->memory_size))
515 			return 0;
516 	}
517 	return 1;
518 }
519 
520 /* Can we switch to this memory table? */
521 /* Caller should have device mutex but not vq mutex */
522 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
523 			    int log_all)
524 {
525 	int i;
526 
527 	for (i = 0; i < d->nvqs; ++i) {
528 		int ok;
529 		mutex_lock(&d->vqs[i]->mutex);
530 		/* If ring is inactive, will check when it's enabled. */
531 		if (d->vqs[i]->private_data)
532 			ok = vq_memory_access_ok(d->vqs[i]->log_base, mem,
533 						 log_all);
534 		else
535 			ok = 1;
536 		mutex_unlock(&d->vqs[i]->mutex);
537 		if (!ok)
538 			return 0;
539 	}
540 	return 1;
541 }
542 
543 static int vq_access_ok(struct vhost_dev *d, unsigned int num,
544 			struct vring_desc __user *desc,
545 			struct vring_avail __user *avail,
546 			struct vring_used __user *used)
547 {
548 	size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
549 	return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
550 	       access_ok(VERIFY_READ, avail,
551 			 sizeof *avail + num * sizeof *avail->ring + s) &&
552 	       access_ok(VERIFY_WRITE, used,
553 			sizeof *used + num * sizeof *used->ring + s);
554 }
555 
556 /* Can we log writes? */
557 /* Caller should have device mutex but not vq mutex */
558 int vhost_log_access_ok(struct vhost_dev *dev)
559 {
560 	struct vhost_memory *mp;
561 
562 	mp = rcu_dereference_protected(dev->memory,
563 				       lockdep_is_held(&dev->mutex));
564 	return memory_access_ok(dev, mp, 1);
565 }
566 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
567 
568 /* Verify access for write logging. */
569 /* Caller should have vq mutex and device mutex */
570 static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
571 			    void __user *log_base)
572 {
573 	struct vhost_memory *mp;
574 	size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
575 
576 	mp = rcu_dereference_protected(vq->dev->memory,
577 				       lockdep_is_held(&vq->mutex));
578 	return vq_memory_access_ok(log_base, mp,
579 			    vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
580 		(!vq->log_used || log_access_ok(log_base, vq->log_addr,
581 					sizeof *vq->used +
582 					vq->num * sizeof *vq->used->ring + s));
583 }
584 
585 /* Can we start vq? */
586 /* Caller should have vq mutex and device mutex */
587 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
588 {
589 	return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
590 		vq_log_access_ok(vq->dev, vq, vq->log_base);
591 }
592 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
593 
594 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
595 {
596 	struct vhost_memory mem, *newmem, *oldmem;
597 	unsigned long size = offsetof(struct vhost_memory, regions);
598 
599 	if (copy_from_user(&mem, m, size))
600 		return -EFAULT;
601 	if (mem.padding)
602 		return -EOPNOTSUPP;
603 	if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
604 		return -E2BIG;
605 	newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
606 	if (!newmem)
607 		return -ENOMEM;
608 
609 	memcpy(newmem, &mem, size);
610 	if (copy_from_user(newmem->regions, m->regions,
611 			   mem.nregions * sizeof *m->regions)) {
612 		kfree(newmem);
613 		return -EFAULT;
614 	}
615 
616 	if (!memory_access_ok(d, newmem,
617 			      vhost_has_feature(d, VHOST_F_LOG_ALL))) {
618 		kfree(newmem);
619 		return -EFAULT;
620 	}
621 	oldmem = rcu_dereference_protected(d->memory,
622 					   lockdep_is_held(&d->mutex));
623 	rcu_assign_pointer(d->memory, newmem);
624 	synchronize_rcu();
625 	kfree(oldmem);
626 	return 0;
627 }
628 
629 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
630 {
631 	struct file *eventfp, *filep = NULL;
632 	bool pollstart = false, pollstop = false;
633 	struct eventfd_ctx *ctx = NULL;
634 	u32 __user *idxp = argp;
635 	struct vhost_virtqueue *vq;
636 	struct vhost_vring_state s;
637 	struct vhost_vring_file f;
638 	struct vhost_vring_addr a;
639 	u32 idx;
640 	long r;
641 
642 	r = get_user(idx, idxp);
643 	if (r < 0)
644 		return r;
645 	if (idx >= d->nvqs)
646 		return -ENOBUFS;
647 
648 	vq = d->vqs[idx];
649 
650 	mutex_lock(&vq->mutex);
651 
652 	switch (ioctl) {
653 	case VHOST_SET_VRING_NUM:
654 		/* Resizing ring with an active backend?
655 		 * You don't want to do that. */
656 		if (vq->private_data) {
657 			r = -EBUSY;
658 			break;
659 		}
660 		if (copy_from_user(&s, argp, sizeof s)) {
661 			r = -EFAULT;
662 			break;
663 		}
664 		if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
665 			r = -EINVAL;
666 			break;
667 		}
668 		vq->num = s.num;
669 		break;
670 	case VHOST_SET_VRING_BASE:
671 		/* Moving base with an active backend?
672 		 * You don't want to do that. */
673 		if (vq->private_data) {
674 			r = -EBUSY;
675 			break;
676 		}
677 		if (copy_from_user(&s, argp, sizeof s)) {
678 			r = -EFAULT;
679 			break;
680 		}
681 		if (s.num > 0xffff) {
682 			r = -EINVAL;
683 			break;
684 		}
685 		vq->last_avail_idx = s.num;
686 		/* Forget the cached index value. */
687 		vq->avail_idx = vq->last_avail_idx;
688 		break;
689 	case VHOST_GET_VRING_BASE:
690 		s.index = idx;
691 		s.num = vq->last_avail_idx;
692 		if (copy_to_user(argp, &s, sizeof s))
693 			r = -EFAULT;
694 		break;
695 	case VHOST_SET_VRING_ADDR:
696 		if (copy_from_user(&a, argp, sizeof a)) {
697 			r = -EFAULT;
698 			break;
699 		}
700 		if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
701 			r = -EOPNOTSUPP;
702 			break;
703 		}
704 		/* For 32bit, verify that the top 32bits of the user
705 		   data are set to zero. */
706 		if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
707 		    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
708 		    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
709 			r = -EFAULT;
710 			break;
711 		}
712 		if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
713 		    (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
714 		    (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
715 			r = -EINVAL;
716 			break;
717 		}
718 
719 		/* We only verify access here if backend is configured.
720 		 * If it is not, we don't as size might not have been setup.
721 		 * We will verify when backend is configured. */
722 		if (vq->private_data) {
723 			if (!vq_access_ok(d, vq->num,
724 				(void __user *)(unsigned long)a.desc_user_addr,
725 				(void __user *)(unsigned long)a.avail_user_addr,
726 				(void __user *)(unsigned long)a.used_user_addr)) {
727 				r = -EINVAL;
728 				break;
729 			}
730 
731 			/* Also validate log access for used ring if enabled. */
732 			if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
733 			    !log_access_ok(vq->log_base, a.log_guest_addr,
734 					   sizeof *vq->used +
735 					   vq->num * sizeof *vq->used->ring)) {
736 				r = -EINVAL;
737 				break;
738 			}
739 		}
740 
741 		vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
742 		vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
743 		vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
744 		vq->log_addr = a.log_guest_addr;
745 		vq->used = (void __user *)(unsigned long)a.used_user_addr;
746 		break;
747 	case VHOST_SET_VRING_KICK:
748 		if (copy_from_user(&f, argp, sizeof f)) {
749 			r = -EFAULT;
750 			break;
751 		}
752 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
753 		if (IS_ERR(eventfp)) {
754 			r = PTR_ERR(eventfp);
755 			break;
756 		}
757 		if (eventfp != vq->kick) {
758 			pollstop = (filep = vq->kick) != NULL;
759 			pollstart = (vq->kick = eventfp) != NULL;
760 		} else
761 			filep = eventfp;
762 		break;
763 	case VHOST_SET_VRING_CALL:
764 		if (copy_from_user(&f, argp, sizeof f)) {
765 			r = -EFAULT;
766 			break;
767 		}
768 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
769 		if (IS_ERR(eventfp)) {
770 			r = PTR_ERR(eventfp);
771 			break;
772 		}
773 		if (eventfp != vq->call) {
774 			filep = vq->call;
775 			ctx = vq->call_ctx;
776 			vq->call = eventfp;
777 			vq->call_ctx = eventfp ?
778 				eventfd_ctx_fileget(eventfp) : NULL;
779 		} else
780 			filep = eventfp;
781 		break;
782 	case VHOST_SET_VRING_ERR:
783 		if (copy_from_user(&f, argp, sizeof f)) {
784 			r = -EFAULT;
785 			break;
786 		}
787 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
788 		if (IS_ERR(eventfp)) {
789 			r = PTR_ERR(eventfp);
790 			break;
791 		}
792 		if (eventfp != vq->error) {
793 			filep = vq->error;
794 			vq->error = eventfp;
795 			ctx = vq->error_ctx;
796 			vq->error_ctx = eventfp ?
797 				eventfd_ctx_fileget(eventfp) : NULL;
798 		} else
799 			filep = eventfp;
800 		break;
801 	default:
802 		r = -ENOIOCTLCMD;
803 	}
804 
805 	if (pollstop && vq->handle_kick)
806 		vhost_poll_stop(&vq->poll);
807 
808 	if (ctx)
809 		eventfd_ctx_put(ctx);
810 	if (filep)
811 		fput(filep);
812 
813 	if (pollstart && vq->handle_kick)
814 		r = vhost_poll_start(&vq->poll, vq->kick);
815 
816 	mutex_unlock(&vq->mutex);
817 
818 	if (pollstop && vq->handle_kick)
819 		vhost_poll_flush(&vq->poll);
820 	return r;
821 }
822 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
823 
824 /* Caller must have device mutex */
825 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
826 {
827 	struct file *eventfp, *filep = NULL;
828 	struct eventfd_ctx *ctx = NULL;
829 	u64 p;
830 	long r;
831 	int i, fd;
832 
833 	/* If you are not the owner, you can become one */
834 	if (ioctl == VHOST_SET_OWNER) {
835 		r = vhost_dev_set_owner(d);
836 		goto done;
837 	}
838 
839 	/* You must be the owner to do anything else */
840 	r = vhost_dev_check_owner(d);
841 	if (r)
842 		goto done;
843 
844 	switch (ioctl) {
845 	case VHOST_SET_MEM_TABLE:
846 		r = vhost_set_memory(d, argp);
847 		break;
848 	case VHOST_SET_LOG_BASE:
849 		if (copy_from_user(&p, argp, sizeof p)) {
850 			r = -EFAULT;
851 			break;
852 		}
853 		if ((u64)(unsigned long)p != p) {
854 			r = -EFAULT;
855 			break;
856 		}
857 		for (i = 0; i < d->nvqs; ++i) {
858 			struct vhost_virtqueue *vq;
859 			void __user *base = (void __user *)(unsigned long)p;
860 			vq = d->vqs[i];
861 			mutex_lock(&vq->mutex);
862 			/* If ring is inactive, will check when it's enabled. */
863 			if (vq->private_data && !vq_log_access_ok(d, vq, base))
864 				r = -EFAULT;
865 			else
866 				vq->log_base = base;
867 			mutex_unlock(&vq->mutex);
868 		}
869 		break;
870 	case VHOST_SET_LOG_FD:
871 		r = get_user(fd, (int __user *)argp);
872 		if (r < 0)
873 			break;
874 		eventfp = fd == -1 ? NULL : eventfd_fget(fd);
875 		if (IS_ERR(eventfp)) {
876 			r = PTR_ERR(eventfp);
877 			break;
878 		}
879 		if (eventfp != d->log_file) {
880 			filep = d->log_file;
881 			ctx = d->log_ctx;
882 			d->log_ctx = eventfp ?
883 				eventfd_ctx_fileget(eventfp) : NULL;
884 		} else
885 			filep = eventfp;
886 		for (i = 0; i < d->nvqs; ++i) {
887 			mutex_lock(&d->vqs[i]->mutex);
888 			d->vqs[i]->log_ctx = d->log_ctx;
889 			mutex_unlock(&d->vqs[i]->mutex);
890 		}
891 		if (ctx)
892 			eventfd_ctx_put(ctx);
893 		if (filep)
894 			fput(filep);
895 		break;
896 	default:
897 		r = -ENOIOCTLCMD;
898 		break;
899 	}
900 done:
901 	return r;
902 }
903 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
904 
905 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
906 						     __u64 addr, __u32 len)
907 {
908 	struct vhost_memory_region *reg;
909 	int i;
910 
911 	/* linear search is not brilliant, but we really have on the order of 6
912 	 * regions in practice */
913 	for (i = 0; i < mem->nregions; ++i) {
914 		reg = mem->regions + i;
915 		if (reg->guest_phys_addr <= addr &&
916 		    reg->guest_phys_addr + reg->memory_size - 1 >= addr)
917 			return reg;
918 	}
919 	return NULL;
920 }
921 
922 /* TODO: This is really inefficient.  We need something like get_user()
923  * (instruction directly accesses the data, with an exception table entry
924  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
925  */
926 static int set_bit_to_user(int nr, void __user *addr)
927 {
928 	unsigned long log = (unsigned long)addr;
929 	struct page *page;
930 	void *base;
931 	int bit = nr + (log % PAGE_SIZE) * 8;
932 	int r;
933 
934 	r = get_user_pages_fast(log, 1, 1, &page);
935 	if (r < 0)
936 		return r;
937 	BUG_ON(r != 1);
938 	base = kmap_atomic(page);
939 	set_bit(bit, base);
940 	kunmap_atomic(base);
941 	set_page_dirty_lock(page);
942 	put_page(page);
943 	return 0;
944 }
945 
946 static int log_write(void __user *log_base,
947 		     u64 write_address, u64 write_length)
948 {
949 	u64 write_page = write_address / VHOST_PAGE_SIZE;
950 	int r;
951 
952 	if (!write_length)
953 		return 0;
954 	write_length += write_address % VHOST_PAGE_SIZE;
955 	for (;;) {
956 		u64 base = (u64)(unsigned long)log_base;
957 		u64 log = base + write_page / 8;
958 		int bit = write_page % 8;
959 		if ((u64)(unsigned long)log != log)
960 			return -EFAULT;
961 		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
962 		if (r < 0)
963 			return r;
964 		if (write_length <= VHOST_PAGE_SIZE)
965 			break;
966 		write_length -= VHOST_PAGE_SIZE;
967 		write_page += 1;
968 	}
969 	return r;
970 }
971 
972 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
973 		    unsigned int log_num, u64 len)
974 {
975 	int i, r;
976 
977 	/* Make sure data written is seen before log. */
978 	smp_wmb();
979 	for (i = 0; i < log_num; ++i) {
980 		u64 l = min(log[i].len, len);
981 		r = log_write(vq->log_base, log[i].addr, l);
982 		if (r < 0)
983 			return r;
984 		len -= l;
985 		if (!len) {
986 			if (vq->log_ctx)
987 				eventfd_signal(vq->log_ctx, 1);
988 			return 0;
989 		}
990 	}
991 	/* Length written exceeds what we have stored. This is a bug. */
992 	BUG();
993 	return 0;
994 }
995 EXPORT_SYMBOL_GPL(vhost_log_write);
996 
997 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
998 {
999 	void __user *used;
1000 	if (__put_user(vq->used_flags, &vq->used->flags) < 0)
1001 		return -EFAULT;
1002 	if (unlikely(vq->log_used)) {
1003 		/* Make sure the flag is seen before log. */
1004 		smp_wmb();
1005 		/* Log used flag write. */
1006 		used = &vq->used->flags;
1007 		log_write(vq->log_base, vq->log_addr +
1008 			  (used - (void __user *)vq->used),
1009 			  sizeof vq->used->flags);
1010 		if (vq->log_ctx)
1011 			eventfd_signal(vq->log_ctx, 1);
1012 	}
1013 	return 0;
1014 }
1015 
1016 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1017 {
1018 	if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1019 		return -EFAULT;
1020 	if (unlikely(vq->log_used)) {
1021 		void __user *used;
1022 		/* Make sure the event is seen before log. */
1023 		smp_wmb();
1024 		/* Log avail event write */
1025 		used = vhost_avail_event(vq);
1026 		log_write(vq->log_base, vq->log_addr +
1027 			  (used - (void __user *)vq->used),
1028 			  sizeof *vhost_avail_event(vq));
1029 		if (vq->log_ctx)
1030 			eventfd_signal(vq->log_ctx, 1);
1031 	}
1032 	return 0;
1033 }
1034 
1035 int vhost_init_used(struct vhost_virtqueue *vq)
1036 {
1037 	int r;
1038 	if (!vq->private_data)
1039 		return 0;
1040 
1041 	r = vhost_update_used_flags(vq);
1042 	if (r)
1043 		return r;
1044 	vq->signalled_used_valid = false;
1045 	return get_user(vq->last_used_idx, &vq->used->idx);
1046 }
1047 EXPORT_SYMBOL_GPL(vhost_init_used);
1048 
1049 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1050 			  struct iovec iov[], int iov_size)
1051 {
1052 	const struct vhost_memory_region *reg;
1053 	struct vhost_memory *mem;
1054 	struct iovec *_iov;
1055 	u64 s = 0;
1056 	int ret = 0;
1057 
1058 	rcu_read_lock();
1059 
1060 	mem = rcu_dereference(dev->memory);
1061 	while ((u64)len > s) {
1062 		u64 size;
1063 		if (unlikely(ret >= iov_size)) {
1064 			ret = -ENOBUFS;
1065 			break;
1066 		}
1067 		reg = find_region(mem, addr, len);
1068 		if (unlikely(!reg)) {
1069 			ret = -EFAULT;
1070 			break;
1071 		}
1072 		_iov = iov + ret;
1073 		size = reg->memory_size - addr + reg->guest_phys_addr;
1074 		_iov->iov_len = min((u64)len - s, size);
1075 		_iov->iov_base = (void __user *)(unsigned long)
1076 			(reg->userspace_addr + addr - reg->guest_phys_addr);
1077 		s += size;
1078 		addr += size;
1079 		++ret;
1080 	}
1081 
1082 	rcu_read_unlock();
1083 	return ret;
1084 }
1085 
1086 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1087  * function returns the next descriptor in the chain,
1088  * or -1U if we're at the end. */
1089 static unsigned next_desc(struct vring_desc *desc)
1090 {
1091 	unsigned int next;
1092 
1093 	/* If this descriptor says it doesn't chain, we're done. */
1094 	if (!(desc->flags & VRING_DESC_F_NEXT))
1095 		return -1U;
1096 
1097 	/* Check they're not leading us off end of descriptors. */
1098 	next = desc->next;
1099 	/* Make sure compiler knows to grab that: we don't want it changing! */
1100 	/* We will use the result as an index in an array, so most
1101 	 * architectures only need a compiler barrier here. */
1102 	read_barrier_depends();
1103 
1104 	return next;
1105 }
1106 
1107 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1108 			struct iovec iov[], unsigned int iov_size,
1109 			unsigned int *out_num, unsigned int *in_num,
1110 			struct vhost_log *log, unsigned int *log_num,
1111 			struct vring_desc *indirect)
1112 {
1113 	struct vring_desc desc;
1114 	unsigned int i = 0, count, found = 0;
1115 	int ret;
1116 
1117 	/* Sanity check */
1118 	if (unlikely(indirect->len % sizeof desc)) {
1119 		vq_err(vq, "Invalid length in indirect descriptor: "
1120 		       "len 0x%llx not multiple of 0x%zx\n",
1121 		       (unsigned long long)indirect->len,
1122 		       sizeof desc);
1123 		return -EINVAL;
1124 	}
1125 
1126 	ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1127 			     UIO_MAXIOV);
1128 	if (unlikely(ret < 0)) {
1129 		vq_err(vq, "Translation failure %d in indirect.\n", ret);
1130 		return ret;
1131 	}
1132 
1133 	/* We will use the result as an address to read from, so most
1134 	 * architectures only need a compiler barrier here. */
1135 	read_barrier_depends();
1136 
1137 	count = indirect->len / sizeof desc;
1138 	/* Buffers are chained via a 16 bit next field, so
1139 	 * we can have at most 2^16 of these. */
1140 	if (unlikely(count > USHRT_MAX + 1)) {
1141 		vq_err(vq, "Indirect buffer length too big: %d\n",
1142 		       indirect->len);
1143 		return -E2BIG;
1144 	}
1145 
1146 	do {
1147 		unsigned iov_count = *in_num + *out_num;
1148 		if (unlikely(++found > count)) {
1149 			vq_err(vq, "Loop detected: last one at %u "
1150 			       "indirect size %u\n",
1151 			       i, count);
1152 			return -EINVAL;
1153 		}
1154 		if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1155 					      vq->indirect, sizeof desc))) {
1156 			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1157 			       i, (size_t)indirect->addr + i * sizeof desc);
1158 			return -EINVAL;
1159 		}
1160 		if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1161 			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1162 			       i, (size_t)indirect->addr + i * sizeof desc);
1163 			return -EINVAL;
1164 		}
1165 
1166 		ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1167 				     iov_size - iov_count);
1168 		if (unlikely(ret < 0)) {
1169 			vq_err(vq, "Translation failure %d indirect idx %d\n",
1170 			       ret, i);
1171 			return ret;
1172 		}
1173 		/* If this is an input descriptor, increment that count. */
1174 		if (desc.flags & VRING_DESC_F_WRITE) {
1175 			*in_num += ret;
1176 			if (unlikely(log)) {
1177 				log[*log_num].addr = desc.addr;
1178 				log[*log_num].len = desc.len;
1179 				++*log_num;
1180 			}
1181 		} else {
1182 			/* If it's an output descriptor, they're all supposed
1183 			 * to come before any input descriptors. */
1184 			if (unlikely(*in_num)) {
1185 				vq_err(vq, "Indirect descriptor "
1186 				       "has out after in: idx %d\n", i);
1187 				return -EINVAL;
1188 			}
1189 			*out_num += ret;
1190 		}
1191 	} while ((i = next_desc(&desc)) != -1);
1192 	return 0;
1193 }
1194 
1195 /* This looks in the virtqueue and for the first available buffer, and converts
1196  * it to an iovec for convenient access.  Since descriptors consist of some
1197  * number of output then some number of input descriptors, it's actually two
1198  * iovecs, but we pack them into one and note how many of each there were.
1199  *
1200  * This function returns the descriptor number found, or vq->num (which is
1201  * never a valid descriptor number) if none was found.  A negative code is
1202  * returned on error. */
1203 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1204 		      struct iovec iov[], unsigned int iov_size,
1205 		      unsigned int *out_num, unsigned int *in_num,
1206 		      struct vhost_log *log, unsigned int *log_num)
1207 {
1208 	struct vring_desc desc;
1209 	unsigned int i, head, found = 0;
1210 	u16 last_avail_idx;
1211 	int ret;
1212 
1213 	/* Check it isn't doing very strange things with descriptor numbers. */
1214 	last_avail_idx = vq->last_avail_idx;
1215 	if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1216 		vq_err(vq, "Failed to access avail idx at %p\n",
1217 		       &vq->avail->idx);
1218 		return -EFAULT;
1219 	}
1220 
1221 	if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1222 		vq_err(vq, "Guest moved used index from %u to %u",
1223 		       last_avail_idx, vq->avail_idx);
1224 		return -EFAULT;
1225 	}
1226 
1227 	/* If there's nothing new since last we looked, return invalid. */
1228 	if (vq->avail_idx == last_avail_idx)
1229 		return vq->num;
1230 
1231 	/* Only get avail ring entries after they have been exposed by guest. */
1232 	smp_rmb();
1233 
1234 	/* Grab the next descriptor number they're advertising, and increment
1235 	 * the index we've seen. */
1236 	if (unlikely(__get_user(head,
1237 				&vq->avail->ring[last_avail_idx % vq->num]))) {
1238 		vq_err(vq, "Failed to read head: idx %d address %p\n",
1239 		       last_avail_idx,
1240 		       &vq->avail->ring[last_avail_idx % vq->num]);
1241 		return -EFAULT;
1242 	}
1243 
1244 	/* If their number is silly, that's an error. */
1245 	if (unlikely(head >= vq->num)) {
1246 		vq_err(vq, "Guest says index %u > %u is available",
1247 		       head, vq->num);
1248 		return -EINVAL;
1249 	}
1250 
1251 	/* When we start there are none of either input nor output. */
1252 	*out_num = *in_num = 0;
1253 	if (unlikely(log))
1254 		*log_num = 0;
1255 
1256 	i = head;
1257 	do {
1258 		unsigned iov_count = *in_num + *out_num;
1259 		if (unlikely(i >= vq->num)) {
1260 			vq_err(vq, "Desc index is %u > %u, head = %u",
1261 			       i, vq->num, head);
1262 			return -EINVAL;
1263 		}
1264 		if (unlikely(++found > vq->num)) {
1265 			vq_err(vq, "Loop detected: last one at %u "
1266 			       "vq size %u head %u\n",
1267 			       i, vq->num, head);
1268 			return -EINVAL;
1269 		}
1270 		ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1271 		if (unlikely(ret)) {
1272 			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1273 			       i, vq->desc + i);
1274 			return -EFAULT;
1275 		}
1276 		if (desc.flags & VRING_DESC_F_INDIRECT) {
1277 			ret = get_indirect(dev, vq, iov, iov_size,
1278 					   out_num, in_num,
1279 					   log, log_num, &desc);
1280 			if (unlikely(ret < 0)) {
1281 				vq_err(vq, "Failure detected "
1282 				       "in indirect descriptor at idx %d\n", i);
1283 				return ret;
1284 			}
1285 			continue;
1286 		}
1287 
1288 		ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1289 				     iov_size - iov_count);
1290 		if (unlikely(ret < 0)) {
1291 			vq_err(vq, "Translation failure %d descriptor idx %d\n",
1292 			       ret, i);
1293 			return ret;
1294 		}
1295 		if (desc.flags & VRING_DESC_F_WRITE) {
1296 			/* If this is an input descriptor,
1297 			 * increment that count. */
1298 			*in_num += ret;
1299 			if (unlikely(log)) {
1300 				log[*log_num].addr = desc.addr;
1301 				log[*log_num].len = desc.len;
1302 				++*log_num;
1303 			}
1304 		} else {
1305 			/* If it's an output descriptor, they're all supposed
1306 			 * to come before any input descriptors. */
1307 			if (unlikely(*in_num)) {
1308 				vq_err(vq, "Descriptor has out after in: "
1309 				       "idx %d\n", i);
1310 				return -EINVAL;
1311 			}
1312 			*out_num += ret;
1313 		}
1314 	} while ((i = next_desc(&desc)) != -1);
1315 
1316 	/* On success, increment avail index. */
1317 	vq->last_avail_idx++;
1318 
1319 	/* Assume notifications from guest are disabled at this point,
1320 	 * if they aren't we would need to update avail_event index. */
1321 	BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1322 	return head;
1323 }
1324 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1325 
1326 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1327 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1328 {
1329 	vq->last_avail_idx -= n;
1330 }
1331 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1332 
1333 /* After we've used one of their buffers, we tell them about it.  We'll then
1334  * want to notify the guest, using eventfd. */
1335 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1336 {
1337 	struct vring_used_elem heads = { head, len };
1338 
1339 	return vhost_add_used_n(vq, &heads, 1);
1340 }
1341 EXPORT_SYMBOL_GPL(vhost_add_used);
1342 
1343 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1344 			    struct vring_used_elem *heads,
1345 			    unsigned count)
1346 {
1347 	struct vring_used_elem __user *used;
1348 	u16 old, new;
1349 	int start;
1350 
1351 	start = vq->last_used_idx % vq->num;
1352 	used = vq->used->ring + start;
1353 	if (count == 1) {
1354 		if (__put_user(heads[0].id, &used->id)) {
1355 			vq_err(vq, "Failed to write used id");
1356 			return -EFAULT;
1357 		}
1358 		if (__put_user(heads[0].len, &used->len)) {
1359 			vq_err(vq, "Failed to write used len");
1360 			return -EFAULT;
1361 		}
1362 	} else if (__copy_to_user(used, heads, count * sizeof *used)) {
1363 		vq_err(vq, "Failed to write used");
1364 		return -EFAULT;
1365 	}
1366 	if (unlikely(vq->log_used)) {
1367 		/* Make sure data is seen before log. */
1368 		smp_wmb();
1369 		/* Log used ring entry write. */
1370 		log_write(vq->log_base,
1371 			  vq->log_addr +
1372 			   ((void __user *)used - (void __user *)vq->used),
1373 			  count * sizeof *used);
1374 	}
1375 	old = vq->last_used_idx;
1376 	new = (vq->last_used_idx += count);
1377 	/* If the driver never bothers to signal in a very long while,
1378 	 * used index might wrap around. If that happens, invalidate
1379 	 * signalled_used index we stored. TODO: make sure driver
1380 	 * signals at least once in 2^16 and remove this. */
1381 	if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1382 		vq->signalled_used_valid = false;
1383 	return 0;
1384 }
1385 
1386 /* After we've used one of their buffers, we tell them about it.  We'll then
1387  * want to notify the guest, using eventfd. */
1388 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1389 		     unsigned count)
1390 {
1391 	int start, n, r;
1392 
1393 	start = vq->last_used_idx % vq->num;
1394 	n = vq->num - start;
1395 	if (n < count) {
1396 		r = __vhost_add_used_n(vq, heads, n);
1397 		if (r < 0)
1398 			return r;
1399 		heads += n;
1400 		count -= n;
1401 	}
1402 	r = __vhost_add_used_n(vq, heads, count);
1403 
1404 	/* Make sure buffer is written before we update index. */
1405 	smp_wmb();
1406 	if (put_user(vq->last_used_idx, &vq->used->idx)) {
1407 		vq_err(vq, "Failed to increment used idx");
1408 		return -EFAULT;
1409 	}
1410 	if (unlikely(vq->log_used)) {
1411 		/* Log used index update. */
1412 		log_write(vq->log_base,
1413 			  vq->log_addr + offsetof(struct vring_used, idx),
1414 			  sizeof vq->used->idx);
1415 		if (vq->log_ctx)
1416 			eventfd_signal(vq->log_ctx, 1);
1417 	}
1418 	return r;
1419 }
1420 EXPORT_SYMBOL_GPL(vhost_add_used_n);
1421 
1422 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1423 {
1424 	__u16 old, new, event;
1425 	bool v;
1426 	/* Flush out used index updates. This is paired
1427 	 * with the barrier that the Guest executes when enabling
1428 	 * interrupts. */
1429 	smp_mb();
1430 
1431 	if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1432 	    unlikely(vq->avail_idx == vq->last_avail_idx))
1433 		return true;
1434 
1435 	if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1436 		__u16 flags;
1437 		if (__get_user(flags, &vq->avail->flags)) {
1438 			vq_err(vq, "Failed to get flags");
1439 			return true;
1440 		}
1441 		return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1442 	}
1443 	old = vq->signalled_used;
1444 	v = vq->signalled_used_valid;
1445 	new = vq->signalled_used = vq->last_used_idx;
1446 	vq->signalled_used_valid = true;
1447 
1448 	if (unlikely(!v))
1449 		return true;
1450 
1451 	if (get_user(event, vhost_used_event(vq))) {
1452 		vq_err(vq, "Failed to get used event idx");
1453 		return true;
1454 	}
1455 	return vring_need_event(event, new, old);
1456 }
1457 
1458 /* This actually signals the guest, using eventfd. */
1459 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1460 {
1461 	/* Signal the Guest tell them we used something up. */
1462 	if (vq->call_ctx && vhost_notify(dev, vq))
1463 		eventfd_signal(vq->call_ctx, 1);
1464 }
1465 EXPORT_SYMBOL_GPL(vhost_signal);
1466 
1467 /* And here's the combo meal deal.  Supersize me! */
1468 void vhost_add_used_and_signal(struct vhost_dev *dev,
1469 			       struct vhost_virtqueue *vq,
1470 			       unsigned int head, int len)
1471 {
1472 	vhost_add_used(vq, head, len);
1473 	vhost_signal(dev, vq);
1474 }
1475 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1476 
1477 /* multi-buffer version of vhost_add_used_and_signal */
1478 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1479 				 struct vhost_virtqueue *vq,
1480 				 struct vring_used_elem *heads, unsigned count)
1481 {
1482 	vhost_add_used_n(vq, heads, count);
1483 	vhost_signal(dev, vq);
1484 }
1485 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1486 
1487 /* OK, now we need to know about added descriptors. */
1488 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1489 {
1490 	u16 avail_idx;
1491 	int r;
1492 
1493 	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1494 		return false;
1495 	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1496 	if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1497 		r = vhost_update_used_flags(vq);
1498 		if (r) {
1499 			vq_err(vq, "Failed to enable notification at %p: %d\n",
1500 			       &vq->used->flags, r);
1501 			return false;
1502 		}
1503 	} else {
1504 		r = vhost_update_avail_event(vq, vq->avail_idx);
1505 		if (r) {
1506 			vq_err(vq, "Failed to update avail event index at %p: %d\n",
1507 			       vhost_avail_event(vq), r);
1508 			return false;
1509 		}
1510 	}
1511 	/* They could have slipped one in as we were doing that: make
1512 	 * sure it's written, then check again. */
1513 	smp_mb();
1514 	r = __get_user(avail_idx, &vq->avail->idx);
1515 	if (r) {
1516 		vq_err(vq, "Failed to check avail idx at %p: %d\n",
1517 		       &vq->avail->idx, r);
1518 		return false;
1519 	}
1520 
1521 	return avail_idx != vq->avail_idx;
1522 }
1523 EXPORT_SYMBOL_GPL(vhost_enable_notify);
1524 
1525 /* We don't need to be notified again. */
1526 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1527 {
1528 	int r;
1529 
1530 	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1531 		return;
1532 	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1533 	if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1534 		r = vhost_update_used_flags(vq);
1535 		if (r)
1536 			vq_err(vq, "Failed to enable notification at %p: %d\n",
1537 			       &vq->used->flags, r);
1538 	}
1539 }
1540 EXPORT_SYMBOL_GPL(vhost_disable_notify);
1541 
1542 static int __init vhost_init(void)
1543 {
1544 	return 0;
1545 }
1546 
1547 static void __exit vhost_exit(void)
1548 {
1549 }
1550 
1551 module_init(vhost_init);
1552 module_exit(vhost_exit);
1553 
1554 MODULE_VERSION("0.0.1");
1555 MODULE_LICENSE("GPL v2");
1556 MODULE_AUTHOR("Michael S. Tsirkin");
1557 MODULE_DESCRIPTION("Host kernel accelerator for virtio");
1558