xref: /openbmc/linux/drivers/vhost/vhost.c (revision 95e9fd10)
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/virtio_net.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 
29 #include <linux/net.h>
30 #include <linux/if_packet.h>
31 #include <linux/if_arp.h>
32 
33 #include "vhost.h"
34 
35 enum {
36 	VHOST_MEMORY_MAX_NREGIONS = 64,
37 	VHOST_MEMORY_F_LOG = 0x1,
38 };
39 
40 static unsigned vhost_zcopy_mask __read_mostly;
41 
42 #define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num])
43 #define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num])
44 
45 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
46 			    poll_table *pt)
47 {
48 	struct vhost_poll *poll;
49 
50 	poll = container_of(pt, struct vhost_poll, table);
51 	poll->wqh = wqh;
52 	add_wait_queue(wqh, &poll->wait);
53 }
54 
55 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
56 			     void *key)
57 {
58 	struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
59 
60 	if (!((unsigned long)key & poll->mask))
61 		return 0;
62 
63 	vhost_poll_queue(poll);
64 	return 0;
65 }
66 
67 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
68 {
69 	INIT_LIST_HEAD(&work->node);
70 	work->fn = fn;
71 	init_waitqueue_head(&work->done);
72 	work->flushing = 0;
73 	work->queue_seq = work->done_seq = 0;
74 }
75 
76 /* Init poll structure */
77 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
78 		     unsigned long mask, struct vhost_dev *dev)
79 {
80 	init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
81 	init_poll_funcptr(&poll->table, vhost_poll_func);
82 	poll->mask = mask;
83 	poll->dev = dev;
84 
85 	vhost_work_init(&poll->work, fn);
86 }
87 
88 /* Start polling a file. We add ourselves to file's wait queue. The caller must
89  * keep a reference to a file until after vhost_poll_stop is called. */
90 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
91 {
92 	unsigned long mask;
93 
94 	mask = file->f_op->poll(file, &poll->table);
95 	if (mask)
96 		vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
97 }
98 
99 /* Stop polling a file. After this function returns, it becomes safe to drop the
100  * file reference. You must also flush afterwards. */
101 void vhost_poll_stop(struct vhost_poll *poll)
102 {
103 	remove_wait_queue(poll->wqh, &poll->wait);
104 }
105 
106 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
107 				unsigned seq)
108 {
109 	int left;
110 
111 	spin_lock_irq(&dev->work_lock);
112 	left = seq - work->done_seq;
113 	spin_unlock_irq(&dev->work_lock);
114 	return left <= 0;
115 }
116 
117 static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
118 {
119 	unsigned seq;
120 	int flushing;
121 
122 	spin_lock_irq(&dev->work_lock);
123 	seq = work->queue_seq;
124 	work->flushing++;
125 	spin_unlock_irq(&dev->work_lock);
126 	wait_event(work->done, vhost_work_seq_done(dev, work, seq));
127 	spin_lock_irq(&dev->work_lock);
128 	flushing = --work->flushing;
129 	spin_unlock_irq(&dev->work_lock);
130 	BUG_ON(flushing < 0);
131 }
132 
133 /* Flush any work that has been scheduled. When calling this, don't hold any
134  * locks that are also used by the callback. */
135 void vhost_poll_flush(struct vhost_poll *poll)
136 {
137 	vhost_work_flush(poll->dev, &poll->work);
138 }
139 
140 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
141 {
142 	unsigned long flags;
143 
144 	spin_lock_irqsave(&dev->work_lock, flags);
145 	if (list_empty(&work->node)) {
146 		list_add_tail(&work->node, &dev->work_list);
147 		work->queue_seq++;
148 		wake_up_process(dev->worker);
149 	}
150 	spin_unlock_irqrestore(&dev->work_lock, flags);
151 }
152 
153 void vhost_poll_queue(struct vhost_poll *poll)
154 {
155 	vhost_work_queue(poll->dev, &poll->work);
156 }
157 
158 static void vhost_vq_reset(struct vhost_dev *dev,
159 			   struct vhost_virtqueue *vq)
160 {
161 	vq->num = 1;
162 	vq->desc = NULL;
163 	vq->avail = NULL;
164 	vq->used = NULL;
165 	vq->last_avail_idx = 0;
166 	vq->avail_idx = 0;
167 	vq->last_used_idx = 0;
168 	vq->signalled_used = 0;
169 	vq->signalled_used_valid = false;
170 	vq->used_flags = 0;
171 	vq->log_used = false;
172 	vq->log_addr = -1ull;
173 	vq->vhost_hlen = 0;
174 	vq->sock_hlen = 0;
175 	vq->private_data = NULL;
176 	vq->log_base = NULL;
177 	vq->error_ctx = NULL;
178 	vq->error = NULL;
179 	vq->kick = NULL;
180 	vq->call_ctx = NULL;
181 	vq->call = NULL;
182 	vq->log_ctx = NULL;
183 	vq->upend_idx = 0;
184 	vq->done_idx = 0;
185 	vq->ubufs = NULL;
186 }
187 
188 static int vhost_worker(void *data)
189 {
190 	struct vhost_dev *dev = data;
191 	struct vhost_work *work = NULL;
192 	unsigned uninitialized_var(seq);
193 	mm_segment_t oldfs = get_fs();
194 
195 	set_fs(USER_DS);
196 	use_mm(dev->mm);
197 
198 	for (;;) {
199 		/* mb paired w/ kthread_stop */
200 		set_current_state(TASK_INTERRUPTIBLE);
201 
202 		spin_lock_irq(&dev->work_lock);
203 		if (work) {
204 			work->done_seq = seq;
205 			if (work->flushing)
206 				wake_up_all(&work->done);
207 		}
208 
209 		if (kthread_should_stop()) {
210 			spin_unlock_irq(&dev->work_lock);
211 			__set_current_state(TASK_RUNNING);
212 			break;
213 		}
214 		if (!list_empty(&dev->work_list)) {
215 			work = list_first_entry(&dev->work_list,
216 						struct vhost_work, node);
217 			list_del_init(&work->node);
218 			seq = work->queue_seq;
219 		} else
220 			work = NULL;
221 		spin_unlock_irq(&dev->work_lock);
222 
223 		if (work) {
224 			__set_current_state(TASK_RUNNING);
225 			work->fn(work);
226 			if (need_resched())
227 				schedule();
228 		} else
229 			schedule();
230 
231 	}
232 	unuse_mm(dev->mm);
233 	set_fs(oldfs);
234 	return 0;
235 }
236 
237 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
238 {
239 	kfree(vq->indirect);
240 	vq->indirect = NULL;
241 	kfree(vq->log);
242 	vq->log = NULL;
243 	kfree(vq->heads);
244 	vq->heads = NULL;
245 	kfree(vq->ubuf_info);
246 	vq->ubuf_info = NULL;
247 }
248 
249 void vhost_enable_zcopy(int vq)
250 {
251 	vhost_zcopy_mask |= 0x1 << vq;
252 }
253 
254 /* Helper to allocate iovec buffers for all vqs. */
255 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
256 {
257 	int i;
258 	bool zcopy;
259 
260 	for (i = 0; i < dev->nvqs; ++i) {
261 		dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect *
262 					       UIO_MAXIOV, GFP_KERNEL);
263 		dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV,
264 					  GFP_KERNEL);
265 		dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads *
266 					    UIO_MAXIOV, GFP_KERNEL);
267 		zcopy = vhost_zcopy_mask & (0x1 << i);
268 		if (zcopy)
269 			dev->vqs[i].ubuf_info =
270 				kmalloc(sizeof *dev->vqs[i].ubuf_info *
271 					UIO_MAXIOV, GFP_KERNEL);
272 		if (!dev->vqs[i].indirect || !dev->vqs[i].log ||
273 			!dev->vqs[i].heads ||
274 			(zcopy && !dev->vqs[i].ubuf_info))
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 	int i;
297 
298 	dev->vqs = vqs;
299 	dev->nvqs = nvqs;
300 	mutex_init(&dev->mutex);
301 	dev->log_ctx = NULL;
302 	dev->log_file = NULL;
303 	dev->memory = NULL;
304 	dev->mm = NULL;
305 	spin_lock_init(&dev->work_lock);
306 	INIT_LIST_HEAD(&dev->work_list);
307 	dev->worker = NULL;
308 
309 	for (i = 0; i < dev->nvqs; ++i) {
310 		dev->vqs[i].log = NULL;
311 		dev->vqs[i].indirect = NULL;
312 		dev->vqs[i].heads = NULL;
313 		dev->vqs[i].ubuf_info = NULL;
314 		dev->vqs[i].dev = dev;
315 		mutex_init(&dev->vqs[i].mutex);
316 		vhost_vq_reset(dev, dev->vqs + i);
317 		if (dev->vqs[i].handle_kick)
318 			vhost_poll_init(&dev->vqs[i].poll,
319 					dev->vqs[i].handle_kick, POLLIN, dev);
320 	}
321 
322 	return 0;
323 }
324 
325 /* Caller should have device mutex */
326 long vhost_dev_check_owner(struct vhost_dev *dev)
327 {
328 	/* Are you the owner? If not, I don't think you mean to do that */
329 	return dev->mm == current->mm ? 0 : -EPERM;
330 }
331 
332 struct vhost_attach_cgroups_struct {
333 	struct vhost_work work;
334 	struct task_struct *owner;
335 	int ret;
336 };
337 
338 static void vhost_attach_cgroups_work(struct vhost_work *work)
339 {
340 	struct vhost_attach_cgroups_struct *s;
341 
342 	s = container_of(work, struct vhost_attach_cgroups_struct, work);
343 	s->ret = cgroup_attach_task_all(s->owner, current);
344 }
345 
346 static int vhost_attach_cgroups(struct vhost_dev *dev)
347 {
348 	struct vhost_attach_cgroups_struct attach;
349 
350 	attach.owner = current;
351 	vhost_work_init(&attach.work, vhost_attach_cgroups_work);
352 	vhost_work_queue(dev, &attach.work);
353 	vhost_work_flush(dev, &attach.work);
354 	return attach.ret;
355 }
356 
357 /* Caller should have device mutex */
358 static long vhost_dev_set_owner(struct vhost_dev *dev)
359 {
360 	struct task_struct *worker;
361 	int err;
362 
363 	/* Is there an owner already? */
364 	if (dev->mm) {
365 		err = -EBUSY;
366 		goto err_mm;
367 	}
368 
369 	/* No owner, become one */
370 	dev->mm = get_task_mm(current);
371 	worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
372 	if (IS_ERR(worker)) {
373 		err = PTR_ERR(worker);
374 		goto err_worker;
375 	}
376 
377 	dev->worker = worker;
378 	wake_up_process(worker);	/* avoid contributing to loadavg */
379 
380 	err = vhost_attach_cgroups(dev);
381 	if (err)
382 		goto err_cgroup;
383 
384 	err = vhost_dev_alloc_iovecs(dev);
385 	if (err)
386 		goto err_cgroup;
387 
388 	return 0;
389 err_cgroup:
390 	kthread_stop(worker);
391 	dev->worker = NULL;
392 err_worker:
393 	if (dev->mm)
394 		mmput(dev->mm);
395 	dev->mm = NULL;
396 err_mm:
397 	return err;
398 }
399 
400 /* Caller should have device mutex */
401 long vhost_dev_reset_owner(struct vhost_dev *dev)
402 {
403 	struct vhost_memory *memory;
404 
405 	/* Restore memory to default empty mapping. */
406 	memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
407 	if (!memory)
408 		return -ENOMEM;
409 
410 	vhost_dev_cleanup(dev, true);
411 
412 	memory->nregions = 0;
413 	RCU_INIT_POINTER(dev->memory, memory);
414 	return 0;
415 }
416 
417 /* In case of DMA done not in order in lower device driver for some reason.
418  * upend_idx is used to track end of used idx, done_idx is used to track head
419  * of used idx. Once lower device DMA done contiguously, we will signal KVM
420  * guest used idx.
421  */
422 int vhost_zerocopy_signal_used(struct vhost_virtqueue *vq)
423 {
424 	int i;
425 	int j = 0;
426 
427 	for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
428 		if ((vq->heads[i].len == VHOST_DMA_DONE_LEN)) {
429 			vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
430 			vhost_add_used_and_signal(vq->dev, vq,
431 						  vq->heads[i].id, 0);
432 			++j;
433 		} else
434 			break;
435 	}
436 	if (j)
437 		vq->done_idx = i;
438 	return j;
439 }
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].kick && dev->vqs[i].handle_kick) {
448 			vhost_poll_stop(&dev->vqs[i].poll);
449 			vhost_poll_flush(&dev->vqs[i].poll);
450 		}
451 		/* Wait for all lower device DMAs done. */
452 		if (dev->vqs[i].ubufs)
453 			vhost_ubuf_put_and_wait(dev->vqs[i].ubufs);
454 
455 		/* Signal guest as appropriate. */
456 		vhost_zerocopy_signal_used(&dev->vqs[i]);
457 
458 		if (dev->vqs[i].error_ctx)
459 			eventfd_ctx_put(dev->vqs[i].error_ctx);
460 		if (dev->vqs[i].error)
461 			fput(dev->vqs[i].error);
462 		if (dev->vqs[i].kick)
463 			fput(dev->vqs[i].kick);
464 		if (dev->vqs[i].call_ctx)
465 			eventfd_ctx_put(dev->vqs[i].call_ctx);
466 		if (dev->vqs[i].call)
467 			fput(dev->vqs[i].call);
468 		vhost_vq_reset(dev, dev->vqs + i);
469 	}
470 	vhost_dev_free_iovecs(dev);
471 	if (dev->log_ctx)
472 		eventfd_ctx_put(dev->log_ctx);
473 	dev->log_ctx = NULL;
474 	if (dev->log_file)
475 		fput(dev->log_file);
476 	dev->log_file = NULL;
477 	/* No one will access memory at this point */
478 	kfree(rcu_dereference_protected(dev->memory,
479 					locked ==
480 						lockdep_is_held(&dev->mutex)));
481 	RCU_INIT_POINTER(dev->memory, NULL);
482 	WARN_ON(!list_empty(&dev->work_list));
483 	if (dev->worker) {
484 		kthread_stop(dev->worker);
485 		dev->worker = NULL;
486 	}
487 	if (dev->mm)
488 		mmput(dev->mm);
489 	dev->mm = NULL;
490 }
491 
492 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
493 {
494 	u64 a = addr / VHOST_PAGE_SIZE / 8;
495 
496 	/* Make sure 64 bit math will not overflow. */
497 	if (a > ULONG_MAX - (unsigned long)log_base ||
498 	    a + (unsigned long)log_base > ULONG_MAX)
499 		return 0;
500 
501 	return access_ok(VERIFY_WRITE, log_base + a,
502 			 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
503 }
504 
505 /* Caller should have vq mutex and device mutex. */
506 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
507 			       int log_all)
508 {
509 	int i;
510 
511 	if (!mem)
512 		return 0;
513 
514 	for (i = 0; i < mem->nregions; ++i) {
515 		struct vhost_memory_region *m = mem->regions + i;
516 		unsigned long a = m->userspace_addr;
517 		if (m->memory_size > ULONG_MAX)
518 			return 0;
519 		else if (!access_ok(VERIFY_WRITE, (void __user *)a,
520 				    m->memory_size))
521 			return 0;
522 		else if (log_all && !log_access_ok(log_base,
523 						   m->guest_phys_addr,
524 						   m->memory_size))
525 			return 0;
526 	}
527 	return 1;
528 }
529 
530 /* Can we switch to this memory table? */
531 /* Caller should have device mutex but not vq mutex */
532 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
533 			    int log_all)
534 {
535 	int i;
536 
537 	for (i = 0; i < d->nvqs; ++i) {
538 		int ok;
539 		mutex_lock(&d->vqs[i].mutex);
540 		/* If ring is inactive, will check when it's enabled. */
541 		if (d->vqs[i].private_data)
542 			ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
543 						 log_all);
544 		else
545 			ok = 1;
546 		mutex_unlock(&d->vqs[i].mutex);
547 		if (!ok)
548 			return 0;
549 	}
550 	return 1;
551 }
552 
553 static int vq_access_ok(struct vhost_dev *d, unsigned int num,
554 			struct vring_desc __user *desc,
555 			struct vring_avail __user *avail,
556 			struct vring_used __user *used)
557 {
558 	size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
559 	return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
560 	       access_ok(VERIFY_READ, avail,
561 			 sizeof *avail + num * sizeof *avail->ring + s) &&
562 	       access_ok(VERIFY_WRITE, used,
563 			sizeof *used + num * sizeof *used->ring + s);
564 }
565 
566 /* Can we log writes? */
567 /* Caller should have device mutex but not vq mutex */
568 int vhost_log_access_ok(struct vhost_dev *dev)
569 {
570 	struct vhost_memory *mp;
571 
572 	mp = rcu_dereference_protected(dev->memory,
573 				       lockdep_is_held(&dev->mutex));
574 	return memory_access_ok(dev, mp, 1);
575 }
576 
577 /* Verify access for write logging. */
578 /* Caller should have vq mutex and device mutex */
579 static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
580 			    void __user *log_base)
581 {
582 	struct vhost_memory *mp;
583 	size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
584 
585 	mp = rcu_dereference_protected(vq->dev->memory,
586 				       lockdep_is_held(&vq->mutex));
587 	return vq_memory_access_ok(log_base, mp,
588 			    vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
589 		(!vq->log_used || log_access_ok(log_base, vq->log_addr,
590 					sizeof *vq->used +
591 					vq->num * sizeof *vq->used->ring + s));
592 }
593 
594 /* Can we start vq? */
595 /* Caller should have vq mutex and device mutex */
596 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
597 {
598 	return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
599 		vq_log_access_ok(vq->dev, vq, vq->log_base);
600 }
601 
602 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
603 {
604 	struct vhost_memory mem, *newmem, *oldmem;
605 	unsigned long size = offsetof(struct vhost_memory, regions);
606 
607 	if (copy_from_user(&mem, m, size))
608 		return -EFAULT;
609 	if (mem.padding)
610 		return -EOPNOTSUPP;
611 	if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
612 		return -E2BIG;
613 	newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
614 	if (!newmem)
615 		return -ENOMEM;
616 
617 	memcpy(newmem, &mem, size);
618 	if (copy_from_user(newmem->regions, m->regions,
619 			   mem.nregions * sizeof *m->regions)) {
620 		kfree(newmem);
621 		return -EFAULT;
622 	}
623 
624 	if (!memory_access_ok(d, newmem,
625 			      vhost_has_feature(d, VHOST_F_LOG_ALL))) {
626 		kfree(newmem);
627 		return -EFAULT;
628 	}
629 	oldmem = rcu_dereference_protected(d->memory,
630 					   lockdep_is_held(&d->mutex));
631 	rcu_assign_pointer(d->memory, newmem);
632 	synchronize_rcu();
633 	kfree(oldmem);
634 	return 0;
635 }
636 
637 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
638 {
639 	struct file *eventfp, *filep = NULL,
640 		    *pollstart = NULL, *pollstop = NULL;
641 	struct eventfd_ctx *ctx = NULL;
642 	u32 __user *idxp = argp;
643 	struct vhost_virtqueue *vq;
644 	struct vhost_vring_state s;
645 	struct vhost_vring_file f;
646 	struct vhost_vring_addr a;
647 	u32 idx;
648 	long r;
649 
650 	r = get_user(idx, idxp);
651 	if (r < 0)
652 		return r;
653 	if (idx >= d->nvqs)
654 		return -ENOBUFS;
655 
656 	vq = d->vqs + idx;
657 
658 	mutex_lock(&vq->mutex);
659 
660 	switch (ioctl) {
661 	case VHOST_SET_VRING_NUM:
662 		/* Resizing ring with an active backend?
663 		 * You don't want to do that. */
664 		if (vq->private_data) {
665 			r = -EBUSY;
666 			break;
667 		}
668 		if (copy_from_user(&s, argp, sizeof s)) {
669 			r = -EFAULT;
670 			break;
671 		}
672 		if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
673 			r = -EINVAL;
674 			break;
675 		}
676 		vq->num = s.num;
677 		break;
678 	case VHOST_SET_VRING_BASE:
679 		/* Moving base with an active backend?
680 		 * You don't want to do that. */
681 		if (vq->private_data) {
682 			r = -EBUSY;
683 			break;
684 		}
685 		if (copy_from_user(&s, argp, sizeof s)) {
686 			r = -EFAULT;
687 			break;
688 		}
689 		if (s.num > 0xffff) {
690 			r = -EINVAL;
691 			break;
692 		}
693 		vq->last_avail_idx = s.num;
694 		/* Forget the cached index value. */
695 		vq->avail_idx = vq->last_avail_idx;
696 		break;
697 	case VHOST_GET_VRING_BASE:
698 		s.index = idx;
699 		s.num = vq->last_avail_idx;
700 		if (copy_to_user(argp, &s, sizeof s))
701 			r = -EFAULT;
702 		break;
703 	case VHOST_SET_VRING_ADDR:
704 		if (copy_from_user(&a, argp, sizeof a)) {
705 			r = -EFAULT;
706 			break;
707 		}
708 		if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
709 			r = -EOPNOTSUPP;
710 			break;
711 		}
712 		/* For 32bit, verify that the top 32bits of the user
713 		   data are set to zero. */
714 		if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
715 		    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
716 		    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
717 			r = -EFAULT;
718 			break;
719 		}
720 		if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
721 		    (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
722 		    (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
723 			r = -EINVAL;
724 			break;
725 		}
726 
727 		/* We only verify access here if backend is configured.
728 		 * If it is not, we don't as size might not have been setup.
729 		 * We will verify when backend is configured. */
730 		if (vq->private_data) {
731 			if (!vq_access_ok(d, vq->num,
732 				(void __user *)(unsigned long)a.desc_user_addr,
733 				(void __user *)(unsigned long)a.avail_user_addr,
734 				(void __user *)(unsigned long)a.used_user_addr)) {
735 				r = -EINVAL;
736 				break;
737 			}
738 
739 			/* Also validate log access for used ring if enabled. */
740 			if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
741 			    !log_access_ok(vq->log_base, a.log_guest_addr,
742 					   sizeof *vq->used +
743 					   vq->num * sizeof *vq->used->ring)) {
744 				r = -EINVAL;
745 				break;
746 			}
747 		}
748 
749 		vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
750 		vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
751 		vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
752 		vq->log_addr = a.log_guest_addr;
753 		vq->used = (void __user *)(unsigned long)a.used_user_addr;
754 		break;
755 	case VHOST_SET_VRING_KICK:
756 		if (copy_from_user(&f, argp, sizeof f)) {
757 			r = -EFAULT;
758 			break;
759 		}
760 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
761 		if (IS_ERR(eventfp)) {
762 			r = PTR_ERR(eventfp);
763 			break;
764 		}
765 		if (eventfp != vq->kick) {
766 			pollstop = filep = vq->kick;
767 			pollstart = vq->kick = eventfp;
768 		} else
769 			filep = eventfp;
770 		break;
771 	case VHOST_SET_VRING_CALL:
772 		if (copy_from_user(&f, argp, sizeof f)) {
773 			r = -EFAULT;
774 			break;
775 		}
776 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
777 		if (IS_ERR(eventfp)) {
778 			r = PTR_ERR(eventfp);
779 			break;
780 		}
781 		if (eventfp != vq->call) {
782 			filep = vq->call;
783 			ctx = vq->call_ctx;
784 			vq->call = eventfp;
785 			vq->call_ctx = eventfp ?
786 				eventfd_ctx_fileget(eventfp) : NULL;
787 		} else
788 			filep = eventfp;
789 		break;
790 	case VHOST_SET_VRING_ERR:
791 		if (copy_from_user(&f, argp, sizeof f)) {
792 			r = -EFAULT;
793 			break;
794 		}
795 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
796 		if (IS_ERR(eventfp)) {
797 			r = PTR_ERR(eventfp);
798 			break;
799 		}
800 		if (eventfp != vq->error) {
801 			filep = vq->error;
802 			vq->error = eventfp;
803 			ctx = vq->error_ctx;
804 			vq->error_ctx = eventfp ?
805 				eventfd_ctx_fileget(eventfp) : NULL;
806 		} else
807 			filep = eventfp;
808 		break;
809 	default:
810 		r = -ENOIOCTLCMD;
811 	}
812 
813 	if (pollstop && vq->handle_kick)
814 		vhost_poll_stop(&vq->poll);
815 
816 	if (ctx)
817 		eventfd_ctx_put(ctx);
818 	if (filep)
819 		fput(filep);
820 
821 	if (pollstart && vq->handle_kick)
822 		vhost_poll_start(&vq->poll, vq->kick);
823 
824 	mutex_unlock(&vq->mutex);
825 
826 	if (pollstop && vq->handle_kick)
827 		vhost_poll_flush(&vq->poll);
828 	return r;
829 }
830 
831 /* Caller must have device mutex */
832 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
833 {
834 	void __user *argp = (void __user *)arg;
835 	struct file *eventfp, *filep = NULL;
836 	struct eventfd_ctx *ctx = NULL;
837 	u64 p;
838 	long r;
839 	int i, fd;
840 
841 	/* If you are not the owner, you can become one */
842 	if (ioctl == VHOST_SET_OWNER) {
843 		r = vhost_dev_set_owner(d);
844 		goto done;
845 	}
846 
847 	/* You must be the owner to do anything else */
848 	r = vhost_dev_check_owner(d);
849 	if (r)
850 		goto done;
851 
852 	switch (ioctl) {
853 	case VHOST_SET_MEM_TABLE:
854 		r = vhost_set_memory(d, argp);
855 		break;
856 	case VHOST_SET_LOG_BASE:
857 		if (copy_from_user(&p, argp, sizeof p)) {
858 			r = -EFAULT;
859 			break;
860 		}
861 		if ((u64)(unsigned long)p != p) {
862 			r = -EFAULT;
863 			break;
864 		}
865 		for (i = 0; i < d->nvqs; ++i) {
866 			struct vhost_virtqueue *vq;
867 			void __user *base = (void __user *)(unsigned long)p;
868 			vq = d->vqs + i;
869 			mutex_lock(&vq->mutex);
870 			/* If ring is inactive, will check when it's enabled. */
871 			if (vq->private_data && !vq_log_access_ok(d, vq, base))
872 				r = -EFAULT;
873 			else
874 				vq->log_base = base;
875 			mutex_unlock(&vq->mutex);
876 		}
877 		break;
878 	case VHOST_SET_LOG_FD:
879 		r = get_user(fd, (int __user *)argp);
880 		if (r < 0)
881 			break;
882 		eventfp = fd == -1 ? NULL : eventfd_fget(fd);
883 		if (IS_ERR(eventfp)) {
884 			r = PTR_ERR(eventfp);
885 			break;
886 		}
887 		if (eventfp != d->log_file) {
888 			filep = d->log_file;
889 			ctx = d->log_ctx;
890 			d->log_ctx = eventfp ?
891 				eventfd_ctx_fileget(eventfp) : NULL;
892 		} else
893 			filep = eventfp;
894 		for (i = 0; i < d->nvqs; ++i) {
895 			mutex_lock(&d->vqs[i].mutex);
896 			d->vqs[i].log_ctx = d->log_ctx;
897 			mutex_unlock(&d->vqs[i].mutex);
898 		}
899 		if (ctx)
900 			eventfd_ctx_put(ctx);
901 		if (filep)
902 			fput(filep);
903 		break;
904 	default:
905 		r = vhost_set_vring(d, ioctl, argp);
906 		break;
907 	}
908 done:
909 	return r;
910 }
911 
912 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
913 						     __u64 addr, __u32 len)
914 {
915 	struct vhost_memory_region *reg;
916 	int i;
917 
918 	/* linear search is not brilliant, but we really have on the order of 6
919 	 * regions in practice */
920 	for (i = 0; i < mem->nregions; ++i) {
921 		reg = mem->regions + i;
922 		if (reg->guest_phys_addr <= addr &&
923 		    reg->guest_phys_addr + reg->memory_size - 1 >= addr)
924 			return reg;
925 	}
926 	return NULL;
927 }
928 
929 /* TODO: This is really inefficient.  We need something like get_user()
930  * (instruction directly accesses the data, with an exception table entry
931  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
932  */
933 static int set_bit_to_user(int nr, void __user *addr)
934 {
935 	unsigned long log = (unsigned long)addr;
936 	struct page *page;
937 	void *base;
938 	int bit = nr + (log % PAGE_SIZE) * 8;
939 	int r;
940 
941 	r = get_user_pages_fast(log, 1, 1, &page);
942 	if (r < 0)
943 		return r;
944 	BUG_ON(r != 1);
945 	base = kmap_atomic(page);
946 	set_bit(bit, base);
947 	kunmap_atomic(base);
948 	set_page_dirty_lock(page);
949 	put_page(page);
950 	return 0;
951 }
952 
953 static int log_write(void __user *log_base,
954 		     u64 write_address, u64 write_length)
955 {
956 	u64 write_page = write_address / VHOST_PAGE_SIZE;
957 	int r;
958 
959 	if (!write_length)
960 		return 0;
961 	write_length += write_address % VHOST_PAGE_SIZE;
962 	for (;;) {
963 		u64 base = (u64)(unsigned long)log_base;
964 		u64 log = base + write_page / 8;
965 		int bit = write_page % 8;
966 		if ((u64)(unsigned long)log != log)
967 			return -EFAULT;
968 		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
969 		if (r < 0)
970 			return r;
971 		if (write_length <= VHOST_PAGE_SIZE)
972 			break;
973 		write_length -= VHOST_PAGE_SIZE;
974 		write_page += 1;
975 	}
976 	return r;
977 }
978 
979 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
980 		    unsigned int log_num, u64 len)
981 {
982 	int i, r;
983 
984 	/* Make sure data written is seen before log. */
985 	smp_wmb();
986 	for (i = 0; i < log_num; ++i) {
987 		u64 l = min(log[i].len, len);
988 		r = log_write(vq->log_base, log[i].addr, l);
989 		if (r < 0)
990 			return r;
991 		len -= l;
992 		if (!len) {
993 			if (vq->log_ctx)
994 				eventfd_signal(vq->log_ctx, 1);
995 			return 0;
996 		}
997 	}
998 	/* Length written exceeds what we have stored. This is a bug. */
999 	BUG();
1000 	return 0;
1001 }
1002 
1003 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1004 {
1005 	void __user *used;
1006 	if (__put_user(vq->used_flags, &vq->used->flags) < 0)
1007 		return -EFAULT;
1008 	if (unlikely(vq->log_used)) {
1009 		/* Make sure the flag is seen before log. */
1010 		smp_wmb();
1011 		/* Log used flag write. */
1012 		used = &vq->used->flags;
1013 		log_write(vq->log_base, vq->log_addr +
1014 			  (used - (void __user *)vq->used),
1015 			  sizeof vq->used->flags);
1016 		if (vq->log_ctx)
1017 			eventfd_signal(vq->log_ctx, 1);
1018 	}
1019 	return 0;
1020 }
1021 
1022 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1023 {
1024 	if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1025 		return -EFAULT;
1026 	if (unlikely(vq->log_used)) {
1027 		void __user *used;
1028 		/* Make sure the event is seen before log. */
1029 		smp_wmb();
1030 		/* Log avail event write */
1031 		used = vhost_avail_event(vq);
1032 		log_write(vq->log_base, vq->log_addr +
1033 			  (used - (void __user *)vq->used),
1034 			  sizeof *vhost_avail_event(vq));
1035 		if (vq->log_ctx)
1036 			eventfd_signal(vq->log_ctx, 1);
1037 	}
1038 	return 0;
1039 }
1040 
1041 int vhost_init_used(struct vhost_virtqueue *vq)
1042 {
1043 	int r;
1044 	if (!vq->private_data)
1045 		return 0;
1046 
1047 	r = vhost_update_used_flags(vq);
1048 	if (r)
1049 		return r;
1050 	vq->signalled_used_valid = false;
1051 	return get_user(vq->last_used_idx, &vq->used->idx);
1052 }
1053 
1054 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1055 			  struct iovec iov[], int iov_size)
1056 {
1057 	const struct vhost_memory_region *reg;
1058 	struct vhost_memory *mem;
1059 	struct iovec *_iov;
1060 	u64 s = 0;
1061 	int ret = 0;
1062 
1063 	rcu_read_lock();
1064 
1065 	mem = rcu_dereference(dev->memory);
1066 	while ((u64)len > s) {
1067 		u64 size;
1068 		if (unlikely(ret >= iov_size)) {
1069 			ret = -ENOBUFS;
1070 			break;
1071 		}
1072 		reg = find_region(mem, addr, len);
1073 		if (unlikely(!reg)) {
1074 			ret = -EFAULT;
1075 			break;
1076 		}
1077 		_iov = iov + ret;
1078 		size = reg->memory_size - addr + reg->guest_phys_addr;
1079 		_iov->iov_len = min((u64)len, size);
1080 		_iov->iov_base = (void __user *)(unsigned long)
1081 			(reg->userspace_addr + addr - reg->guest_phys_addr);
1082 		s += size;
1083 		addr += size;
1084 		++ret;
1085 	}
1086 
1087 	rcu_read_unlock();
1088 	return ret;
1089 }
1090 
1091 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1092  * function returns the next descriptor in the chain,
1093  * or -1U if we're at the end. */
1094 static unsigned next_desc(struct vring_desc *desc)
1095 {
1096 	unsigned int next;
1097 
1098 	/* If this descriptor says it doesn't chain, we're done. */
1099 	if (!(desc->flags & VRING_DESC_F_NEXT))
1100 		return -1U;
1101 
1102 	/* Check they're not leading us off end of descriptors. */
1103 	next = desc->next;
1104 	/* Make sure compiler knows to grab that: we don't want it changing! */
1105 	/* We will use the result as an index in an array, so most
1106 	 * architectures only need a compiler barrier here. */
1107 	read_barrier_depends();
1108 
1109 	return next;
1110 }
1111 
1112 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1113 			struct iovec iov[], unsigned int iov_size,
1114 			unsigned int *out_num, unsigned int *in_num,
1115 			struct vhost_log *log, unsigned int *log_num,
1116 			struct vring_desc *indirect)
1117 {
1118 	struct vring_desc desc;
1119 	unsigned int i = 0, count, found = 0;
1120 	int ret;
1121 
1122 	/* Sanity check */
1123 	if (unlikely(indirect->len % sizeof desc)) {
1124 		vq_err(vq, "Invalid length in indirect descriptor: "
1125 		       "len 0x%llx not multiple of 0x%zx\n",
1126 		       (unsigned long long)indirect->len,
1127 		       sizeof desc);
1128 		return -EINVAL;
1129 	}
1130 
1131 	ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1132 			     UIO_MAXIOV);
1133 	if (unlikely(ret < 0)) {
1134 		vq_err(vq, "Translation failure %d in indirect.\n", ret);
1135 		return ret;
1136 	}
1137 
1138 	/* We will use the result as an address to read from, so most
1139 	 * architectures only need a compiler barrier here. */
1140 	read_barrier_depends();
1141 
1142 	count = indirect->len / sizeof desc;
1143 	/* Buffers are chained via a 16 bit next field, so
1144 	 * we can have at most 2^16 of these. */
1145 	if (unlikely(count > USHRT_MAX + 1)) {
1146 		vq_err(vq, "Indirect buffer length too big: %d\n",
1147 		       indirect->len);
1148 		return -E2BIG;
1149 	}
1150 
1151 	do {
1152 		unsigned iov_count = *in_num + *out_num;
1153 		if (unlikely(++found > count)) {
1154 			vq_err(vq, "Loop detected: last one at %u "
1155 			       "indirect size %u\n",
1156 			       i, count);
1157 			return -EINVAL;
1158 		}
1159 		if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1160 					      vq->indirect, sizeof desc))) {
1161 			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1162 			       i, (size_t)indirect->addr + i * sizeof desc);
1163 			return -EINVAL;
1164 		}
1165 		if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1166 			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1167 			       i, (size_t)indirect->addr + i * sizeof desc);
1168 			return -EINVAL;
1169 		}
1170 
1171 		ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1172 				     iov_size - iov_count);
1173 		if (unlikely(ret < 0)) {
1174 			vq_err(vq, "Translation failure %d indirect idx %d\n",
1175 			       ret, i);
1176 			return ret;
1177 		}
1178 		/* If this is an input descriptor, increment that count. */
1179 		if (desc.flags & VRING_DESC_F_WRITE) {
1180 			*in_num += ret;
1181 			if (unlikely(log)) {
1182 				log[*log_num].addr = desc.addr;
1183 				log[*log_num].len = desc.len;
1184 				++*log_num;
1185 			}
1186 		} else {
1187 			/* If it's an output descriptor, they're all supposed
1188 			 * to come before any input descriptors. */
1189 			if (unlikely(*in_num)) {
1190 				vq_err(vq, "Indirect descriptor "
1191 				       "has out after in: idx %d\n", i);
1192 				return -EINVAL;
1193 			}
1194 			*out_num += ret;
1195 		}
1196 	} while ((i = next_desc(&desc)) != -1);
1197 	return 0;
1198 }
1199 
1200 /* This looks in the virtqueue and for the first available buffer, and converts
1201  * it to an iovec for convenient access.  Since descriptors consist of some
1202  * number of output then some number of input descriptors, it's actually two
1203  * iovecs, but we pack them into one and note how many of each there were.
1204  *
1205  * This function returns the descriptor number found, or vq->num (which is
1206  * never a valid descriptor number) if none was found.  A negative code is
1207  * returned on error. */
1208 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1209 		      struct iovec iov[], unsigned int iov_size,
1210 		      unsigned int *out_num, unsigned int *in_num,
1211 		      struct vhost_log *log, unsigned int *log_num)
1212 {
1213 	struct vring_desc desc;
1214 	unsigned int i, head, found = 0;
1215 	u16 last_avail_idx;
1216 	int ret;
1217 
1218 	/* Check it isn't doing very strange things with descriptor numbers. */
1219 	last_avail_idx = vq->last_avail_idx;
1220 	if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1221 		vq_err(vq, "Failed to access avail idx at %p\n",
1222 		       &vq->avail->idx);
1223 		return -EFAULT;
1224 	}
1225 
1226 	if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1227 		vq_err(vq, "Guest moved used index from %u to %u",
1228 		       last_avail_idx, vq->avail_idx);
1229 		return -EFAULT;
1230 	}
1231 
1232 	/* If there's nothing new since last we looked, return invalid. */
1233 	if (vq->avail_idx == last_avail_idx)
1234 		return vq->num;
1235 
1236 	/* Only get avail ring entries after they have been exposed by guest. */
1237 	smp_rmb();
1238 
1239 	/* Grab the next descriptor number they're advertising, and increment
1240 	 * the index we've seen. */
1241 	if (unlikely(__get_user(head,
1242 				&vq->avail->ring[last_avail_idx % vq->num]))) {
1243 		vq_err(vq, "Failed to read head: idx %d address %p\n",
1244 		       last_avail_idx,
1245 		       &vq->avail->ring[last_avail_idx % vq->num]);
1246 		return -EFAULT;
1247 	}
1248 
1249 	/* If their number is silly, that's an error. */
1250 	if (unlikely(head >= vq->num)) {
1251 		vq_err(vq, "Guest says index %u > %u is available",
1252 		       head, vq->num);
1253 		return -EINVAL;
1254 	}
1255 
1256 	/* When we start there are none of either input nor output. */
1257 	*out_num = *in_num = 0;
1258 	if (unlikely(log))
1259 		*log_num = 0;
1260 
1261 	i = head;
1262 	do {
1263 		unsigned iov_count = *in_num + *out_num;
1264 		if (unlikely(i >= vq->num)) {
1265 			vq_err(vq, "Desc index is %u > %u, head = %u",
1266 			       i, vq->num, head);
1267 			return -EINVAL;
1268 		}
1269 		if (unlikely(++found > vq->num)) {
1270 			vq_err(vq, "Loop detected: last one at %u "
1271 			       "vq size %u head %u\n",
1272 			       i, vq->num, head);
1273 			return -EINVAL;
1274 		}
1275 		ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1276 		if (unlikely(ret)) {
1277 			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1278 			       i, vq->desc + i);
1279 			return -EFAULT;
1280 		}
1281 		if (desc.flags & VRING_DESC_F_INDIRECT) {
1282 			ret = get_indirect(dev, vq, iov, iov_size,
1283 					   out_num, in_num,
1284 					   log, log_num, &desc);
1285 			if (unlikely(ret < 0)) {
1286 				vq_err(vq, "Failure detected "
1287 				       "in indirect descriptor at idx %d\n", i);
1288 				return ret;
1289 			}
1290 			continue;
1291 		}
1292 
1293 		ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1294 				     iov_size - iov_count);
1295 		if (unlikely(ret < 0)) {
1296 			vq_err(vq, "Translation failure %d descriptor idx %d\n",
1297 			       ret, i);
1298 			return ret;
1299 		}
1300 		if (desc.flags & VRING_DESC_F_WRITE) {
1301 			/* If this is an input descriptor,
1302 			 * increment that count. */
1303 			*in_num += ret;
1304 			if (unlikely(log)) {
1305 				log[*log_num].addr = desc.addr;
1306 				log[*log_num].len = desc.len;
1307 				++*log_num;
1308 			}
1309 		} else {
1310 			/* If it's an output descriptor, they're all supposed
1311 			 * to come before any input descriptors. */
1312 			if (unlikely(*in_num)) {
1313 				vq_err(vq, "Descriptor has out after in: "
1314 				       "idx %d\n", i);
1315 				return -EINVAL;
1316 			}
1317 			*out_num += ret;
1318 		}
1319 	} while ((i = next_desc(&desc)) != -1);
1320 
1321 	/* On success, increment avail index. */
1322 	vq->last_avail_idx++;
1323 
1324 	/* Assume notifications from guest are disabled at this point,
1325 	 * if they aren't we would need to update avail_event index. */
1326 	BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1327 	return head;
1328 }
1329 
1330 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1331 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1332 {
1333 	vq->last_avail_idx -= n;
1334 }
1335 
1336 /* After we've used one of their buffers, we tell them about it.  We'll then
1337  * want to notify the guest, using eventfd. */
1338 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1339 {
1340 	struct vring_used_elem __user *used;
1341 
1342 	/* The virtqueue contains a ring of used buffers.  Get a pointer to the
1343 	 * next entry in that used ring. */
1344 	used = &vq->used->ring[vq->last_used_idx % vq->num];
1345 	if (__put_user(head, &used->id)) {
1346 		vq_err(vq, "Failed to write used id");
1347 		return -EFAULT;
1348 	}
1349 	if (__put_user(len, &used->len)) {
1350 		vq_err(vq, "Failed to write used len");
1351 		return -EFAULT;
1352 	}
1353 	/* Make sure buffer is written before we update index. */
1354 	smp_wmb();
1355 	if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1356 		vq_err(vq, "Failed to increment used idx");
1357 		return -EFAULT;
1358 	}
1359 	if (unlikely(vq->log_used)) {
1360 		/* Make sure data is seen before log. */
1361 		smp_wmb();
1362 		/* Log used ring entry write. */
1363 		log_write(vq->log_base,
1364 			  vq->log_addr +
1365 			   ((void __user *)used - (void __user *)vq->used),
1366 			  sizeof *used);
1367 		/* Log used index update. */
1368 		log_write(vq->log_base,
1369 			  vq->log_addr + offsetof(struct vring_used, idx),
1370 			  sizeof vq->used->idx);
1371 		if (vq->log_ctx)
1372 			eventfd_signal(vq->log_ctx, 1);
1373 	}
1374 	vq->last_used_idx++;
1375 	/* If the driver never bothers to signal in a very long while,
1376 	 * used index might wrap around. If that happens, invalidate
1377 	 * signalled_used index we stored. TODO: make sure driver
1378 	 * signals at least once in 2^16 and remove this. */
1379 	if (unlikely(vq->last_used_idx == vq->signalled_used))
1380 		vq->signalled_used_valid = false;
1381 	return 0;
1382 }
1383 
1384 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1385 			    struct vring_used_elem *heads,
1386 			    unsigned count)
1387 {
1388 	struct vring_used_elem __user *used;
1389 	u16 old, new;
1390 	int start;
1391 
1392 	start = vq->last_used_idx % vq->num;
1393 	used = vq->used->ring + start;
1394 	if (__copy_to_user(used, heads, count * sizeof *used)) {
1395 		vq_err(vq, "Failed to write used");
1396 		return -EFAULT;
1397 	}
1398 	if (unlikely(vq->log_used)) {
1399 		/* Make sure data is seen before log. */
1400 		smp_wmb();
1401 		/* Log used ring entry write. */
1402 		log_write(vq->log_base,
1403 			  vq->log_addr +
1404 			   ((void __user *)used - (void __user *)vq->used),
1405 			  count * sizeof *used);
1406 	}
1407 	old = vq->last_used_idx;
1408 	new = (vq->last_used_idx += count);
1409 	/* If the driver never bothers to signal in a very long while,
1410 	 * used index might wrap around. If that happens, invalidate
1411 	 * signalled_used index we stored. TODO: make sure driver
1412 	 * signals at least once in 2^16 and remove this. */
1413 	if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1414 		vq->signalled_used_valid = false;
1415 	return 0;
1416 }
1417 
1418 /* After we've used one of their buffers, we tell them about it.  We'll then
1419  * want to notify the guest, using eventfd. */
1420 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1421 		     unsigned count)
1422 {
1423 	int start, n, r;
1424 
1425 	start = vq->last_used_idx % vq->num;
1426 	n = vq->num - start;
1427 	if (n < count) {
1428 		r = __vhost_add_used_n(vq, heads, n);
1429 		if (r < 0)
1430 			return r;
1431 		heads += n;
1432 		count -= n;
1433 	}
1434 	r = __vhost_add_used_n(vq, heads, count);
1435 
1436 	/* Make sure buffer is written before we update index. */
1437 	smp_wmb();
1438 	if (put_user(vq->last_used_idx, &vq->used->idx)) {
1439 		vq_err(vq, "Failed to increment used idx");
1440 		return -EFAULT;
1441 	}
1442 	if (unlikely(vq->log_used)) {
1443 		/* Log used index update. */
1444 		log_write(vq->log_base,
1445 			  vq->log_addr + offsetof(struct vring_used, idx),
1446 			  sizeof vq->used->idx);
1447 		if (vq->log_ctx)
1448 			eventfd_signal(vq->log_ctx, 1);
1449 	}
1450 	return r;
1451 }
1452 
1453 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1454 {
1455 	__u16 old, new, event;
1456 	bool v;
1457 	/* Flush out used index updates. This is paired
1458 	 * with the barrier that the Guest executes when enabling
1459 	 * interrupts. */
1460 	smp_mb();
1461 
1462 	if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1463 	    unlikely(vq->avail_idx == vq->last_avail_idx))
1464 		return true;
1465 
1466 	if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1467 		__u16 flags;
1468 		if (__get_user(flags, &vq->avail->flags)) {
1469 			vq_err(vq, "Failed to get flags");
1470 			return true;
1471 		}
1472 		return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1473 	}
1474 	old = vq->signalled_used;
1475 	v = vq->signalled_used_valid;
1476 	new = vq->signalled_used = vq->last_used_idx;
1477 	vq->signalled_used_valid = true;
1478 
1479 	if (unlikely(!v))
1480 		return true;
1481 
1482 	if (get_user(event, vhost_used_event(vq))) {
1483 		vq_err(vq, "Failed to get used event idx");
1484 		return true;
1485 	}
1486 	return vring_need_event(event, new, old);
1487 }
1488 
1489 /* This actually signals the guest, using eventfd. */
1490 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1491 {
1492 	/* Signal the Guest tell them we used something up. */
1493 	if (vq->call_ctx && vhost_notify(dev, vq))
1494 		eventfd_signal(vq->call_ctx, 1);
1495 }
1496 
1497 /* And here's the combo meal deal.  Supersize me! */
1498 void vhost_add_used_and_signal(struct vhost_dev *dev,
1499 			       struct vhost_virtqueue *vq,
1500 			       unsigned int head, int len)
1501 {
1502 	vhost_add_used(vq, head, len);
1503 	vhost_signal(dev, vq);
1504 }
1505 
1506 /* multi-buffer version of vhost_add_used_and_signal */
1507 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1508 				 struct vhost_virtqueue *vq,
1509 				 struct vring_used_elem *heads, unsigned count)
1510 {
1511 	vhost_add_used_n(vq, heads, count);
1512 	vhost_signal(dev, vq);
1513 }
1514 
1515 /* OK, now we need to know about added descriptors. */
1516 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1517 {
1518 	u16 avail_idx;
1519 	int r;
1520 
1521 	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1522 		return false;
1523 	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1524 	if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1525 		r = vhost_update_used_flags(vq);
1526 		if (r) {
1527 			vq_err(vq, "Failed to enable notification at %p: %d\n",
1528 			       &vq->used->flags, r);
1529 			return false;
1530 		}
1531 	} else {
1532 		r = vhost_update_avail_event(vq, vq->avail_idx);
1533 		if (r) {
1534 			vq_err(vq, "Failed to update avail event index at %p: %d\n",
1535 			       vhost_avail_event(vq), r);
1536 			return false;
1537 		}
1538 	}
1539 	/* They could have slipped one in as we were doing that: make
1540 	 * sure it's written, then check again. */
1541 	smp_mb();
1542 	r = __get_user(avail_idx, &vq->avail->idx);
1543 	if (r) {
1544 		vq_err(vq, "Failed to check avail idx at %p: %d\n",
1545 		       &vq->avail->idx, r);
1546 		return false;
1547 	}
1548 
1549 	return avail_idx != vq->avail_idx;
1550 }
1551 
1552 /* We don't need to be notified again. */
1553 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1554 {
1555 	int r;
1556 
1557 	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1558 		return;
1559 	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1560 	if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1561 		r = vhost_update_used_flags(vq);
1562 		if (r)
1563 			vq_err(vq, "Failed to enable notification at %p: %d\n",
1564 			       &vq->used->flags, r);
1565 	}
1566 }
1567 
1568 static void vhost_zerocopy_done_signal(struct kref *kref)
1569 {
1570 	struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
1571 						    kref);
1572 	wake_up(&ubufs->wait);
1573 }
1574 
1575 struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
1576 					bool zcopy)
1577 {
1578 	struct vhost_ubuf_ref *ubufs;
1579 	/* No zero copy backend? Nothing to count. */
1580 	if (!zcopy)
1581 		return NULL;
1582 	ubufs = kmalloc(sizeof *ubufs, GFP_KERNEL);
1583 	if (!ubufs)
1584 		return ERR_PTR(-ENOMEM);
1585 	kref_init(&ubufs->kref);
1586 	init_waitqueue_head(&ubufs->wait);
1587 	ubufs->vq = vq;
1588 	return ubufs;
1589 }
1590 
1591 void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
1592 {
1593 	kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1594 }
1595 
1596 void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
1597 {
1598 	kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1599 	wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
1600 	kfree(ubufs);
1601 }
1602 
1603 void vhost_zerocopy_callback(struct ubuf_info *ubuf)
1604 {
1605 	struct vhost_ubuf_ref *ubufs = ubuf->ctx;
1606 	struct vhost_virtqueue *vq = ubufs->vq;
1607 
1608 	vhost_poll_queue(&vq->poll);
1609 	/* set len = 1 to mark this desc buffers done DMA */
1610 	vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
1611 	kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1612 }
1613