xref: /openbmc/linux/drivers/vhost/vhost.c (revision 2572f00d)
1 /* Copyright (C) 2009 Red Hat, Inc.
2  * Copyright (C) 2006 Rusty Russell IBM Corporation
3  *
4  * Author: Michael S. Tsirkin <mst@redhat.com>
5  *
6  * Inspiration, some code, and most witty comments come from
7  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.
10  *
11  * Generic code for virtio server in host kernel.
12  */
13 
14 #include <linux/eventfd.h>
15 #include <linux/vhost.h>
16 #include <linux/uio.h>
17 #include <linux/mm.h>
18 #include <linux/mmu_context.h>
19 #include <linux/miscdevice.h>
20 #include <linux/mutex.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/vmalloc.h>
26 #include <linux/kthread.h>
27 #include <linux/cgroup.h>
28 #include <linux/module.h>
29 #include <linux/sort.h>
30 
31 #include "vhost.h"
32 
33 static ushort max_mem_regions = 64;
34 module_param(max_mem_regions, ushort, 0444);
35 MODULE_PARM_DESC(max_mem_regions,
36 	"Maximum number of memory regions in memory map. (default: 64)");
37 
38 enum {
39 	VHOST_MEMORY_F_LOG = 0x1,
40 };
41 
42 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
43 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
44 
45 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
46 static void vhost_vq_reset_user_be(struct vhost_virtqueue *vq)
47 {
48 	vq->user_be = !virtio_legacy_is_little_endian();
49 }
50 
51 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
52 {
53 	struct vhost_vring_state s;
54 
55 	if (vq->private_data)
56 		return -EBUSY;
57 
58 	if (copy_from_user(&s, argp, sizeof(s)))
59 		return -EFAULT;
60 
61 	if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
62 	    s.num != VHOST_VRING_BIG_ENDIAN)
63 		return -EINVAL;
64 
65 	vq->user_be = s.num;
66 
67 	return 0;
68 }
69 
70 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
71 				   int __user *argp)
72 {
73 	struct vhost_vring_state s = {
74 		.index = idx,
75 		.num = vq->user_be
76 	};
77 
78 	if (copy_to_user(argp, &s, sizeof(s)))
79 		return -EFAULT;
80 
81 	return 0;
82 }
83 
84 static void vhost_init_is_le(struct vhost_virtqueue *vq)
85 {
86 	/* Note for legacy virtio: user_be is initialized at reset time
87 	 * according to the host endianness. If userspace does not set an
88 	 * explicit endianness, the default behavior is native endian, as
89 	 * expected by legacy virtio.
90 	 */
91 	vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
92 }
93 #else
94 static void vhost_vq_reset_user_be(struct vhost_virtqueue *vq)
95 {
96 }
97 
98 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
99 {
100 	return -ENOIOCTLCMD;
101 }
102 
103 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
104 				   int __user *argp)
105 {
106 	return -ENOIOCTLCMD;
107 }
108 
109 static void vhost_init_is_le(struct vhost_virtqueue *vq)
110 {
111 	if (vhost_has_feature(vq, VIRTIO_F_VERSION_1))
112 		vq->is_le = true;
113 }
114 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
115 
116 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
117 			    poll_table *pt)
118 {
119 	struct vhost_poll *poll;
120 
121 	poll = container_of(pt, struct vhost_poll, table);
122 	poll->wqh = wqh;
123 	add_wait_queue(wqh, &poll->wait);
124 }
125 
126 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
127 			     void *key)
128 {
129 	struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
130 
131 	if (!((unsigned long)key & poll->mask))
132 		return 0;
133 
134 	vhost_poll_queue(poll);
135 	return 0;
136 }
137 
138 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
139 {
140 	INIT_LIST_HEAD(&work->node);
141 	work->fn = fn;
142 	init_waitqueue_head(&work->done);
143 	work->flushing = 0;
144 	work->queue_seq = work->done_seq = 0;
145 }
146 EXPORT_SYMBOL_GPL(vhost_work_init);
147 
148 /* Init poll structure */
149 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
150 		     unsigned long mask, struct vhost_dev *dev)
151 {
152 	init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
153 	init_poll_funcptr(&poll->table, vhost_poll_func);
154 	poll->mask = mask;
155 	poll->dev = dev;
156 	poll->wqh = NULL;
157 
158 	vhost_work_init(&poll->work, fn);
159 }
160 EXPORT_SYMBOL_GPL(vhost_poll_init);
161 
162 /* Start polling a file. We add ourselves to file's wait queue. The caller must
163  * keep a reference to a file until after vhost_poll_stop is called. */
164 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
165 {
166 	unsigned long mask;
167 	int ret = 0;
168 
169 	if (poll->wqh)
170 		return 0;
171 
172 	mask = file->f_op->poll(file, &poll->table);
173 	if (mask)
174 		vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
175 	if (mask & POLLERR) {
176 		if (poll->wqh)
177 			remove_wait_queue(poll->wqh, &poll->wait);
178 		ret = -EINVAL;
179 	}
180 
181 	return ret;
182 }
183 EXPORT_SYMBOL_GPL(vhost_poll_start);
184 
185 /* Stop polling a file. After this function returns, it becomes safe to drop the
186  * file reference. You must also flush afterwards. */
187 void vhost_poll_stop(struct vhost_poll *poll)
188 {
189 	if (poll->wqh) {
190 		remove_wait_queue(poll->wqh, &poll->wait);
191 		poll->wqh = NULL;
192 	}
193 }
194 EXPORT_SYMBOL_GPL(vhost_poll_stop);
195 
196 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
197 				unsigned seq)
198 {
199 	int left;
200 
201 	spin_lock_irq(&dev->work_lock);
202 	left = seq - work->done_seq;
203 	spin_unlock_irq(&dev->work_lock);
204 	return left <= 0;
205 }
206 
207 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
208 {
209 	unsigned seq;
210 	int flushing;
211 
212 	spin_lock_irq(&dev->work_lock);
213 	seq = work->queue_seq;
214 	work->flushing++;
215 	spin_unlock_irq(&dev->work_lock);
216 	wait_event(work->done, vhost_work_seq_done(dev, work, seq));
217 	spin_lock_irq(&dev->work_lock);
218 	flushing = --work->flushing;
219 	spin_unlock_irq(&dev->work_lock);
220 	BUG_ON(flushing < 0);
221 }
222 EXPORT_SYMBOL_GPL(vhost_work_flush);
223 
224 /* Flush any work that has been scheduled. When calling this, don't hold any
225  * locks that are also used by the callback. */
226 void vhost_poll_flush(struct vhost_poll *poll)
227 {
228 	vhost_work_flush(poll->dev, &poll->work);
229 }
230 EXPORT_SYMBOL_GPL(vhost_poll_flush);
231 
232 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
233 {
234 	unsigned long flags;
235 
236 	spin_lock_irqsave(&dev->work_lock, flags);
237 	if (list_empty(&work->node)) {
238 		list_add_tail(&work->node, &dev->work_list);
239 		work->queue_seq++;
240 		spin_unlock_irqrestore(&dev->work_lock, flags);
241 		wake_up_process(dev->worker);
242 	} else {
243 		spin_unlock_irqrestore(&dev->work_lock, flags);
244 	}
245 }
246 EXPORT_SYMBOL_GPL(vhost_work_queue);
247 
248 void vhost_poll_queue(struct vhost_poll *poll)
249 {
250 	vhost_work_queue(poll->dev, &poll->work);
251 }
252 EXPORT_SYMBOL_GPL(vhost_poll_queue);
253 
254 static void vhost_vq_reset(struct vhost_dev *dev,
255 			   struct vhost_virtqueue *vq)
256 {
257 	vq->num = 1;
258 	vq->desc = NULL;
259 	vq->avail = NULL;
260 	vq->used = NULL;
261 	vq->last_avail_idx = 0;
262 	vq->avail_idx = 0;
263 	vq->last_used_idx = 0;
264 	vq->signalled_used = 0;
265 	vq->signalled_used_valid = false;
266 	vq->used_flags = 0;
267 	vq->log_used = false;
268 	vq->log_addr = -1ull;
269 	vq->private_data = NULL;
270 	vq->acked_features = 0;
271 	vq->log_base = NULL;
272 	vq->error_ctx = NULL;
273 	vq->error = NULL;
274 	vq->kick = NULL;
275 	vq->call_ctx = NULL;
276 	vq->call = NULL;
277 	vq->log_ctx = NULL;
278 	vq->memory = NULL;
279 	vq->is_le = virtio_legacy_is_little_endian();
280 	vhost_vq_reset_user_be(vq);
281 }
282 
283 static int vhost_worker(void *data)
284 {
285 	struct vhost_dev *dev = data;
286 	struct vhost_work *work = NULL;
287 	unsigned uninitialized_var(seq);
288 	mm_segment_t oldfs = get_fs();
289 
290 	set_fs(USER_DS);
291 	use_mm(dev->mm);
292 
293 	for (;;) {
294 		/* mb paired w/ kthread_stop */
295 		set_current_state(TASK_INTERRUPTIBLE);
296 
297 		spin_lock_irq(&dev->work_lock);
298 		if (work) {
299 			work->done_seq = seq;
300 			if (work->flushing)
301 				wake_up_all(&work->done);
302 		}
303 
304 		if (kthread_should_stop()) {
305 			spin_unlock_irq(&dev->work_lock);
306 			__set_current_state(TASK_RUNNING);
307 			break;
308 		}
309 		if (!list_empty(&dev->work_list)) {
310 			work = list_first_entry(&dev->work_list,
311 						struct vhost_work, node);
312 			list_del_init(&work->node);
313 			seq = work->queue_seq;
314 		} else
315 			work = NULL;
316 		spin_unlock_irq(&dev->work_lock);
317 
318 		if (work) {
319 			__set_current_state(TASK_RUNNING);
320 			work->fn(work);
321 			if (need_resched())
322 				schedule();
323 		} else
324 			schedule();
325 
326 	}
327 	unuse_mm(dev->mm);
328 	set_fs(oldfs);
329 	return 0;
330 }
331 
332 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
333 {
334 	kfree(vq->indirect);
335 	vq->indirect = NULL;
336 	kfree(vq->log);
337 	vq->log = NULL;
338 	kfree(vq->heads);
339 	vq->heads = NULL;
340 }
341 
342 /* Helper to allocate iovec buffers for all vqs. */
343 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
344 {
345 	struct vhost_virtqueue *vq;
346 	int i;
347 
348 	for (i = 0; i < dev->nvqs; ++i) {
349 		vq = dev->vqs[i];
350 		vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
351 				       GFP_KERNEL);
352 		vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
353 		vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
354 		if (!vq->indirect || !vq->log || !vq->heads)
355 			goto err_nomem;
356 	}
357 	return 0;
358 
359 err_nomem:
360 	for (; i >= 0; --i)
361 		vhost_vq_free_iovecs(dev->vqs[i]);
362 	return -ENOMEM;
363 }
364 
365 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
366 {
367 	int i;
368 
369 	for (i = 0; i < dev->nvqs; ++i)
370 		vhost_vq_free_iovecs(dev->vqs[i]);
371 }
372 
373 void vhost_dev_init(struct vhost_dev *dev,
374 		    struct vhost_virtqueue **vqs, int nvqs)
375 {
376 	struct vhost_virtqueue *vq;
377 	int i;
378 
379 	dev->vqs = vqs;
380 	dev->nvqs = nvqs;
381 	mutex_init(&dev->mutex);
382 	dev->log_ctx = NULL;
383 	dev->log_file = NULL;
384 	dev->memory = NULL;
385 	dev->mm = NULL;
386 	spin_lock_init(&dev->work_lock);
387 	INIT_LIST_HEAD(&dev->work_list);
388 	dev->worker = NULL;
389 
390 	for (i = 0; i < dev->nvqs; ++i) {
391 		vq = dev->vqs[i];
392 		vq->log = NULL;
393 		vq->indirect = NULL;
394 		vq->heads = NULL;
395 		vq->dev = dev;
396 		mutex_init(&vq->mutex);
397 		vhost_vq_reset(dev, vq);
398 		if (vq->handle_kick)
399 			vhost_poll_init(&vq->poll, vq->handle_kick,
400 					POLLIN, dev);
401 	}
402 }
403 EXPORT_SYMBOL_GPL(vhost_dev_init);
404 
405 /* Caller should have device mutex */
406 long vhost_dev_check_owner(struct vhost_dev *dev)
407 {
408 	/* Are you the owner? If not, I don't think you mean to do that */
409 	return dev->mm == current->mm ? 0 : -EPERM;
410 }
411 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
412 
413 struct vhost_attach_cgroups_struct {
414 	struct vhost_work work;
415 	struct task_struct *owner;
416 	int ret;
417 };
418 
419 static void vhost_attach_cgroups_work(struct vhost_work *work)
420 {
421 	struct vhost_attach_cgroups_struct *s;
422 
423 	s = container_of(work, struct vhost_attach_cgroups_struct, work);
424 	s->ret = cgroup_attach_task_all(s->owner, current);
425 }
426 
427 static int vhost_attach_cgroups(struct vhost_dev *dev)
428 {
429 	struct vhost_attach_cgroups_struct attach;
430 
431 	attach.owner = current;
432 	vhost_work_init(&attach.work, vhost_attach_cgroups_work);
433 	vhost_work_queue(dev, &attach.work);
434 	vhost_work_flush(dev, &attach.work);
435 	return attach.ret;
436 }
437 
438 /* Caller should have device mutex */
439 bool vhost_dev_has_owner(struct vhost_dev *dev)
440 {
441 	return dev->mm;
442 }
443 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
444 
445 /* Caller should have device mutex */
446 long vhost_dev_set_owner(struct vhost_dev *dev)
447 {
448 	struct task_struct *worker;
449 	int err;
450 
451 	/* Is there an owner already? */
452 	if (vhost_dev_has_owner(dev)) {
453 		err = -EBUSY;
454 		goto err_mm;
455 	}
456 
457 	/* No owner, become one */
458 	dev->mm = get_task_mm(current);
459 	worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
460 	if (IS_ERR(worker)) {
461 		err = PTR_ERR(worker);
462 		goto err_worker;
463 	}
464 
465 	dev->worker = worker;
466 	wake_up_process(worker);	/* avoid contributing to loadavg */
467 
468 	err = vhost_attach_cgroups(dev);
469 	if (err)
470 		goto err_cgroup;
471 
472 	err = vhost_dev_alloc_iovecs(dev);
473 	if (err)
474 		goto err_cgroup;
475 
476 	return 0;
477 err_cgroup:
478 	kthread_stop(worker);
479 	dev->worker = NULL;
480 err_worker:
481 	if (dev->mm)
482 		mmput(dev->mm);
483 	dev->mm = NULL;
484 err_mm:
485 	return err;
486 }
487 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
488 
489 struct vhost_memory *vhost_dev_reset_owner_prepare(void)
490 {
491 	return kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
492 }
493 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
494 
495 /* Caller should have device mutex */
496 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_memory *memory)
497 {
498 	int i;
499 
500 	vhost_dev_cleanup(dev, true);
501 
502 	/* Restore memory to default empty mapping. */
503 	memory->nregions = 0;
504 	dev->memory = memory;
505 	/* We don't need VQ locks below since vhost_dev_cleanup makes sure
506 	 * VQs aren't running.
507 	 */
508 	for (i = 0; i < dev->nvqs; ++i)
509 		dev->vqs[i]->memory = memory;
510 }
511 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
512 
513 void vhost_dev_stop(struct vhost_dev *dev)
514 {
515 	int i;
516 
517 	for (i = 0; i < dev->nvqs; ++i) {
518 		if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
519 			vhost_poll_stop(&dev->vqs[i]->poll);
520 			vhost_poll_flush(&dev->vqs[i]->poll);
521 		}
522 	}
523 }
524 EXPORT_SYMBOL_GPL(vhost_dev_stop);
525 
526 /* Caller should have device mutex if and only if locked is set */
527 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
528 {
529 	int i;
530 
531 	for (i = 0; i < dev->nvqs; ++i) {
532 		if (dev->vqs[i]->error_ctx)
533 			eventfd_ctx_put(dev->vqs[i]->error_ctx);
534 		if (dev->vqs[i]->error)
535 			fput(dev->vqs[i]->error);
536 		if (dev->vqs[i]->kick)
537 			fput(dev->vqs[i]->kick);
538 		if (dev->vqs[i]->call_ctx)
539 			eventfd_ctx_put(dev->vqs[i]->call_ctx);
540 		if (dev->vqs[i]->call)
541 			fput(dev->vqs[i]->call);
542 		vhost_vq_reset(dev, dev->vqs[i]);
543 	}
544 	vhost_dev_free_iovecs(dev);
545 	if (dev->log_ctx)
546 		eventfd_ctx_put(dev->log_ctx);
547 	dev->log_ctx = NULL;
548 	if (dev->log_file)
549 		fput(dev->log_file);
550 	dev->log_file = NULL;
551 	/* No one will access memory at this point */
552 	kvfree(dev->memory);
553 	dev->memory = NULL;
554 	WARN_ON(!list_empty(&dev->work_list));
555 	if (dev->worker) {
556 		kthread_stop(dev->worker);
557 		dev->worker = NULL;
558 	}
559 	if (dev->mm)
560 		mmput(dev->mm);
561 	dev->mm = NULL;
562 }
563 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
564 
565 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
566 {
567 	u64 a = addr / VHOST_PAGE_SIZE / 8;
568 
569 	/* Make sure 64 bit math will not overflow. */
570 	if (a > ULONG_MAX - (unsigned long)log_base ||
571 	    a + (unsigned long)log_base > ULONG_MAX)
572 		return 0;
573 
574 	return access_ok(VERIFY_WRITE, log_base + a,
575 			 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
576 }
577 
578 /* Caller should have vq mutex and device mutex. */
579 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
580 			       int log_all)
581 {
582 	int i;
583 
584 	if (!mem)
585 		return 0;
586 
587 	for (i = 0; i < mem->nregions; ++i) {
588 		struct vhost_memory_region *m = mem->regions + i;
589 		unsigned long a = m->userspace_addr;
590 		if (m->memory_size > ULONG_MAX)
591 			return 0;
592 		else if (!access_ok(VERIFY_WRITE, (void __user *)a,
593 				    m->memory_size))
594 			return 0;
595 		else if (log_all && !log_access_ok(log_base,
596 						   m->guest_phys_addr,
597 						   m->memory_size))
598 			return 0;
599 	}
600 	return 1;
601 }
602 
603 /* Can we switch to this memory table? */
604 /* Caller should have device mutex but not vq mutex */
605 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
606 			    int log_all)
607 {
608 	int i;
609 
610 	for (i = 0; i < d->nvqs; ++i) {
611 		int ok;
612 		bool log;
613 
614 		mutex_lock(&d->vqs[i]->mutex);
615 		log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
616 		/* If ring is inactive, will check when it's enabled. */
617 		if (d->vqs[i]->private_data)
618 			ok = vq_memory_access_ok(d->vqs[i]->log_base, mem, log);
619 		else
620 			ok = 1;
621 		mutex_unlock(&d->vqs[i]->mutex);
622 		if (!ok)
623 			return 0;
624 	}
625 	return 1;
626 }
627 
628 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
629 			struct vring_desc __user *desc,
630 			struct vring_avail __user *avail,
631 			struct vring_used __user *used)
632 {
633 	size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
634 	return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
635 	       access_ok(VERIFY_READ, avail,
636 			 sizeof *avail + num * sizeof *avail->ring + s) &&
637 	       access_ok(VERIFY_WRITE, used,
638 			sizeof *used + num * sizeof *used->ring + s);
639 }
640 
641 /* Can we log writes? */
642 /* Caller should have device mutex but not vq mutex */
643 int vhost_log_access_ok(struct vhost_dev *dev)
644 {
645 	return memory_access_ok(dev, dev->memory, 1);
646 }
647 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
648 
649 /* Verify access for write logging. */
650 /* Caller should have vq mutex and device mutex */
651 static int vq_log_access_ok(struct vhost_virtqueue *vq,
652 			    void __user *log_base)
653 {
654 	size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
655 
656 	return vq_memory_access_ok(log_base, vq->memory,
657 				   vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
658 		(!vq->log_used || log_access_ok(log_base, vq->log_addr,
659 					sizeof *vq->used +
660 					vq->num * sizeof *vq->used->ring + s));
661 }
662 
663 /* Can we start vq? */
664 /* Caller should have vq mutex and device mutex */
665 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
666 {
667 	return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
668 		vq_log_access_ok(vq, vq->log_base);
669 }
670 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
671 
672 static int vhost_memory_reg_sort_cmp(const void *p1, const void *p2)
673 {
674 	const struct vhost_memory_region *r1 = p1, *r2 = p2;
675 	if (r1->guest_phys_addr < r2->guest_phys_addr)
676 		return 1;
677 	if (r1->guest_phys_addr > r2->guest_phys_addr)
678 		return -1;
679 	return 0;
680 }
681 
682 static void *vhost_kvzalloc(unsigned long size)
683 {
684 	void *n = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
685 
686 	if (!n)
687 		n = vzalloc(size);
688 	return n;
689 }
690 
691 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
692 {
693 	struct vhost_memory mem, *newmem, *oldmem;
694 	unsigned long size = offsetof(struct vhost_memory, regions);
695 	int i;
696 
697 	if (copy_from_user(&mem, m, size))
698 		return -EFAULT;
699 	if (mem.padding)
700 		return -EOPNOTSUPP;
701 	if (mem.nregions > max_mem_regions)
702 		return -E2BIG;
703 	newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
704 	if (!newmem)
705 		return -ENOMEM;
706 
707 	memcpy(newmem, &mem, size);
708 	if (copy_from_user(newmem->regions, m->regions,
709 			   mem.nregions * sizeof *m->regions)) {
710 		kvfree(newmem);
711 		return -EFAULT;
712 	}
713 	sort(newmem->regions, newmem->nregions, sizeof(*newmem->regions),
714 		vhost_memory_reg_sort_cmp, NULL);
715 
716 	if (!memory_access_ok(d, newmem, 0)) {
717 		kvfree(newmem);
718 		return -EFAULT;
719 	}
720 	oldmem = d->memory;
721 	d->memory = newmem;
722 
723 	/* All memory accesses are done under some VQ mutex. */
724 	for (i = 0; i < d->nvqs; ++i) {
725 		mutex_lock(&d->vqs[i]->mutex);
726 		d->vqs[i]->memory = newmem;
727 		mutex_unlock(&d->vqs[i]->mutex);
728 	}
729 	kvfree(oldmem);
730 	return 0;
731 }
732 
733 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
734 {
735 	struct file *eventfp, *filep = NULL;
736 	bool pollstart = false, pollstop = false;
737 	struct eventfd_ctx *ctx = NULL;
738 	u32 __user *idxp = argp;
739 	struct vhost_virtqueue *vq;
740 	struct vhost_vring_state s;
741 	struct vhost_vring_file f;
742 	struct vhost_vring_addr a;
743 	u32 idx;
744 	long r;
745 
746 	r = get_user(idx, idxp);
747 	if (r < 0)
748 		return r;
749 	if (idx >= d->nvqs)
750 		return -ENOBUFS;
751 
752 	vq = d->vqs[idx];
753 
754 	mutex_lock(&vq->mutex);
755 
756 	switch (ioctl) {
757 	case VHOST_SET_VRING_NUM:
758 		/* Resizing ring with an active backend?
759 		 * You don't want to do that. */
760 		if (vq->private_data) {
761 			r = -EBUSY;
762 			break;
763 		}
764 		if (copy_from_user(&s, argp, sizeof s)) {
765 			r = -EFAULT;
766 			break;
767 		}
768 		if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
769 			r = -EINVAL;
770 			break;
771 		}
772 		vq->num = s.num;
773 		break;
774 	case VHOST_SET_VRING_BASE:
775 		/* Moving base with an active backend?
776 		 * You don't want to do that. */
777 		if (vq->private_data) {
778 			r = -EBUSY;
779 			break;
780 		}
781 		if (copy_from_user(&s, argp, sizeof s)) {
782 			r = -EFAULT;
783 			break;
784 		}
785 		if (s.num > 0xffff) {
786 			r = -EINVAL;
787 			break;
788 		}
789 		vq->last_avail_idx = s.num;
790 		/* Forget the cached index value. */
791 		vq->avail_idx = vq->last_avail_idx;
792 		break;
793 	case VHOST_GET_VRING_BASE:
794 		s.index = idx;
795 		s.num = vq->last_avail_idx;
796 		if (copy_to_user(argp, &s, sizeof s))
797 			r = -EFAULT;
798 		break;
799 	case VHOST_SET_VRING_ADDR:
800 		if (copy_from_user(&a, argp, sizeof a)) {
801 			r = -EFAULT;
802 			break;
803 		}
804 		if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
805 			r = -EOPNOTSUPP;
806 			break;
807 		}
808 		/* For 32bit, verify that the top 32bits of the user
809 		   data are set to zero. */
810 		if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
811 		    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
812 		    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
813 			r = -EFAULT;
814 			break;
815 		}
816 
817 		/* Make sure it's safe to cast pointers to vring types. */
818 		BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
819 		BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
820 		if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
821 		    (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
822 		    (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
823 			r = -EINVAL;
824 			break;
825 		}
826 
827 		/* We only verify access here if backend is configured.
828 		 * If it is not, we don't as size might not have been setup.
829 		 * We will verify when backend is configured. */
830 		if (vq->private_data) {
831 			if (!vq_access_ok(vq, vq->num,
832 				(void __user *)(unsigned long)a.desc_user_addr,
833 				(void __user *)(unsigned long)a.avail_user_addr,
834 				(void __user *)(unsigned long)a.used_user_addr)) {
835 				r = -EINVAL;
836 				break;
837 			}
838 
839 			/* Also validate log access for used ring if enabled. */
840 			if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
841 			    !log_access_ok(vq->log_base, a.log_guest_addr,
842 					   sizeof *vq->used +
843 					   vq->num * sizeof *vq->used->ring)) {
844 				r = -EINVAL;
845 				break;
846 			}
847 		}
848 
849 		vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
850 		vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
851 		vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
852 		vq->log_addr = a.log_guest_addr;
853 		vq->used = (void __user *)(unsigned long)a.used_user_addr;
854 		break;
855 	case VHOST_SET_VRING_KICK:
856 		if (copy_from_user(&f, argp, sizeof f)) {
857 			r = -EFAULT;
858 			break;
859 		}
860 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
861 		if (IS_ERR(eventfp)) {
862 			r = PTR_ERR(eventfp);
863 			break;
864 		}
865 		if (eventfp != vq->kick) {
866 			pollstop = (filep = vq->kick) != NULL;
867 			pollstart = (vq->kick = eventfp) != NULL;
868 		} else
869 			filep = eventfp;
870 		break;
871 	case VHOST_SET_VRING_CALL:
872 		if (copy_from_user(&f, argp, sizeof f)) {
873 			r = -EFAULT;
874 			break;
875 		}
876 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
877 		if (IS_ERR(eventfp)) {
878 			r = PTR_ERR(eventfp);
879 			break;
880 		}
881 		if (eventfp != vq->call) {
882 			filep = vq->call;
883 			ctx = vq->call_ctx;
884 			vq->call = eventfp;
885 			vq->call_ctx = eventfp ?
886 				eventfd_ctx_fileget(eventfp) : NULL;
887 		} else
888 			filep = eventfp;
889 		break;
890 	case VHOST_SET_VRING_ERR:
891 		if (copy_from_user(&f, argp, sizeof f)) {
892 			r = -EFAULT;
893 			break;
894 		}
895 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
896 		if (IS_ERR(eventfp)) {
897 			r = PTR_ERR(eventfp);
898 			break;
899 		}
900 		if (eventfp != vq->error) {
901 			filep = vq->error;
902 			vq->error = eventfp;
903 			ctx = vq->error_ctx;
904 			vq->error_ctx = eventfp ?
905 				eventfd_ctx_fileget(eventfp) : NULL;
906 		} else
907 			filep = eventfp;
908 		break;
909 	case VHOST_SET_VRING_ENDIAN:
910 		r = vhost_set_vring_endian(vq, argp);
911 		break;
912 	case VHOST_GET_VRING_ENDIAN:
913 		r = vhost_get_vring_endian(vq, idx, argp);
914 		break;
915 	default:
916 		r = -ENOIOCTLCMD;
917 	}
918 
919 	if (pollstop && vq->handle_kick)
920 		vhost_poll_stop(&vq->poll);
921 
922 	if (ctx)
923 		eventfd_ctx_put(ctx);
924 	if (filep)
925 		fput(filep);
926 
927 	if (pollstart && vq->handle_kick)
928 		r = vhost_poll_start(&vq->poll, vq->kick);
929 
930 	mutex_unlock(&vq->mutex);
931 
932 	if (pollstop && vq->handle_kick)
933 		vhost_poll_flush(&vq->poll);
934 	return r;
935 }
936 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
937 
938 /* Caller must have device mutex */
939 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
940 {
941 	struct file *eventfp, *filep = NULL;
942 	struct eventfd_ctx *ctx = NULL;
943 	u64 p;
944 	long r;
945 	int i, fd;
946 
947 	/* If you are not the owner, you can become one */
948 	if (ioctl == VHOST_SET_OWNER) {
949 		r = vhost_dev_set_owner(d);
950 		goto done;
951 	}
952 
953 	/* You must be the owner to do anything else */
954 	r = vhost_dev_check_owner(d);
955 	if (r)
956 		goto done;
957 
958 	switch (ioctl) {
959 	case VHOST_SET_MEM_TABLE:
960 		r = vhost_set_memory(d, argp);
961 		break;
962 	case VHOST_SET_LOG_BASE:
963 		if (copy_from_user(&p, argp, sizeof p)) {
964 			r = -EFAULT;
965 			break;
966 		}
967 		if ((u64)(unsigned long)p != p) {
968 			r = -EFAULT;
969 			break;
970 		}
971 		for (i = 0; i < d->nvqs; ++i) {
972 			struct vhost_virtqueue *vq;
973 			void __user *base = (void __user *)(unsigned long)p;
974 			vq = d->vqs[i];
975 			mutex_lock(&vq->mutex);
976 			/* If ring is inactive, will check when it's enabled. */
977 			if (vq->private_data && !vq_log_access_ok(vq, base))
978 				r = -EFAULT;
979 			else
980 				vq->log_base = base;
981 			mutex_unlock(&vq->mutex);
982 		}
983 		break;
984 	case VHOST_SET_LOG_FD:
985 		r = get_user(fd, (int __user *)argp);
986 		if (r < 0)
987 			break;
988 		eventfp = fd == -1 ? NULL : eventfd_fget(fd);
989 		if (IS_ERR(eventfp)) {
990 			r = PTR_ERR(eventfp);
991 			break;
992 		}
993 		if (eventfp != d->log_file) {
994 			filep = d->log_file;
995 			d->log_file = eventfp;
996 			ctx = d->log_ctx;
997 			d->log_ctx = eventfp ?
998 				eventfd_ctx_fileget(eventfp) : NULL;
999 		} else
1000 			filep = eventfp;
1001 		for (i = 0; i < d->nvqs; ++i) {
1002 			mutex_lock(&d->vqs[i]->mutex);
1003 			d->vqs[i]->log_ctx = d->log_ctx;
1004 			mutex_unlock(&d->vqs[i]->mutex);
1005 		}
1006 		if (ctx)
1007 			eventfd_ctx_put(ctx);
1008 		if (filep)
1009 			fput(filep);
1010 		break;
1011 	default:
1012 		r = -ENOIOCTLCMD;
1013 		break;
1014 	}
1015 done:
1016 	return r;
1017 }
1018 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1019 
1020 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
1021 						     __u64 addr, __u32 len)
1022 {
1023 	const struct vhost_memory_region *reg;
1024 	int start = 0, end = mem->nregions;
1025 
1026 	while (start < end) {
1027 		int slot = start + (end - start) / 2;
1028 		reg = mem->regions + slot;
1029 		if (addr >= reg->guest_phys_addr)
1030 			end = slot;
1031 		else
1032 			start = slot + 1;
1033 	}
1034 
1035 	reg = mem->regions + start;
1036 	if (addr >= reg->guest_phys_addr &&
1037 		reg->guest_phys_addr + reg->memory_size > addr)
1038 		return reg;
1039 	return NULL;
1040 }
1041 
1042 /* TODO: This is really inefficient.  We need something like get_user()
1043  * (instruction directly accesses the data, with an exception table entry
1044  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1045  */
1046 static int set_bit_to_user(int nr, void __user *addr)
1047 {
1048 	unsigned long log = (unsigned long)addr;
1049 	struct page *page;
1050 	void *base;
1051 	int bit = nr + (log % PAGE_SIZE) * 8;
1052 	int r;
1053 
1054 	r = get_user_pages_fast(log, 1, 1, &page);
1055 	if (r < 0)
1056 		return r;
1057 	BUG_ON(r != 1);
1058 	base = kmap_atomic(page);
1059 	set_bit(bit, base);
1060 	kunmap_atomic(base);
1061 	set_page_dirty_lock(page);
1062 	put_page(page);
1063 	return 0;
1064 }
1065 
1066 static int log_write(void __user *log_base,
1067 		     u64 write_address, u64 write_length)
1068 {
1069 	u64 write_page = write_address / VHOST_PAGE_SIZE;
1070 	int r;
1071 
1072 	if (!write_length)
1073 		return 0;
1074 	write_length += write_address % VHOST_PAGE_SIZE;
1075 	for (;;) {
1076 		u64 base = (u64)(unsigned long)log_base;
1077 		u64 log = base + write_page / 8;
1078 		int bit = write_page % 8;
1079 		if ((u64)(unsigned long)log != log)
1080 			return -EFAULT;
1081 		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1082 		if (r < 0)
1083 			return r;
1084 		if (write_length <= VHOST_PAGE_SIZE)
1085 			break;
1086 		write_length -= VHOST_PAGE_SIZE;
1087 		write_page += 1;
1088 	}
1089 	return r;
1090 }
1091 
1092 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1093 		    unsigned int log_num, u64 len)
1094 {
1095 	int i, r;
1096 
1097 	/* Make sure data written is seen before log. */
1098 	smp_wmb();
1099 	for (i = 0; i < log_num; ++i) {
1100 		u64 l = min(log[i].len, len);
1101 		r = log_write(vq->log_base, log[i].addr, l);
1102 		if (r < 0)
1103 			return r;
1104 		len -= l;
1105 		if (!len) {
1106 			if (vq->log_ctx)
1107 				eventfd_signal(vq->log_ctx, 1);
1108 			return 0;
1109 		}
1110 	}
1111 	/* Length written exceeds what we have stored. This is a bug. */
1112 	BUG();
1113 	return 0;
1114 }
1115 EXPORT_SYMBOL_GPL(vhost_log_write);
1116 
1117 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1118 {
1119 	void __user *used;
1120 	if (__put_user(cpu_to_vhost16(vq, vq->used_flags), &vq->used->flags) < 0)
1121 		return -EFAULT;
1122 	if (unlikely(vq->log_used)) {
1123 		/* Make sure the flag is seen before log. */
1124 		smp_wmb();
1125 		/* Log used flag write. */
1126 		used = &vq->used->flags;
1127 		log_write(vq->log_base, vq->log_addr +
1128 			  (used - (void __user *)vq->used),
1129 			  sizeof vq->used->flags);
1130 		if (vq->log_ctx)
1131 			eventfd_signal(vq->log_ctx, 1);
1132 	}
1133 	return 0;
1134 }
1135 
1136 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1137 {
1138 	if (__put_user(cpu_to_vhost16(vq, vq->avail_idx), vhost_avail_event(vq)))
1139 		return -EFAULT;
1140 	if (unlikely(vq->log_used)) {
1141 		void __user *used;
1142 		/* Make sure the event is seen before log. */
1143 		smp_wmb();
1144 		/* Log avail event write */
1145 		used = vhost_avail_event(vq);
1146 		log_write(vq->log_base, vq->log_addr +
1147 			  (used - (void __user *)vq->used),
1148 			  sizeof *vhost_avail_event(vq));
1149 		if (vq->log_ctx)
1150 			eventfd_signal(vq->log_ctx, 1);
1151 	}
1152 	return 0;
1153 }
1154 
1155 int vhost_init_used(struct vhost_virtqueue *vq)
1156 {
1157 	__virtio16 last_used_idx;
1158 	int r;
1159 	if (!vq->private_data) {
1160 		vq->is_le = virtio_legacy_is_little_endian();
1161 		return 0;
1162 	}
1163 
1164 	vhost_init_is_le(vq);
1165 
1166 	r = vhost_update_used_flags(vq);
1167 	if (r)
1168 		return r;
1169 	vq->signalled_used_valid = false;
1170 	if (!access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx))
1171 		return -EFAULT;
1172 	r = __get_user(last_used_idx, &vq->used->idx);
1173 	if (r)
1174 		return r;
1175 	vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1176 	return 0;
1177 }
1178 EXPORT_SYMBOL_GPL(vhost_init_used);
1179 
1180 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1181 			  struct iovec iov[], int iov_size)
1182 {
1183 	const struct vhost_memory_region *reg;
1184 	struct vhost_memory *mem;
1185 	struct iovec *_iov;
1186 	u64 s = 0;
1187 	int ret = 0;
1188 
1189 	mem = vq->memory;
1190 	while ((u64)len > s) {
1191 		u64 size;
1192 		if (unlikely(ret >= iov_size)) {
1193 			ret = -ENOBUFS;
1194 			break;
1195 		}
1196 		reg = find_region(mem, addr, len);
1197 		if (unlikely(!reg)) {
1198 			ret = -EFAULT;
1199 			break;
1200 		}
1201 		_iov = iov + ret;
1202 		size = reg->memory_size - addr + reg->guest_phys_addr;
1203 		_iov->iov_len = min((u64)len - s, size);
1204 		_iov->iov_base = (void __user *)(unsigned long)
1205 			(reg->userspace_addr + addr - reg->guest_phys_addr);
1206 		s += size;
1207 		addr += size;
1208 		++ret;
1209 	}
1210 
1211 	return ret;
1212 }
1213 
1214 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1215  * function returns the next descriptor in the chain,
1216  * or -1U if we're at the end. */
1217 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1218 {
1219 	unsigned int next;
1220 
1221 	/* If this descriptor says it doesn't chain, we're done. */
1222 	if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1223 		return -1U;
1224 
1225 	/* Check they're not leading us off end of descriptors. */
1226 	next = vhost16_to_cpu(vq, desc->next);
1227 	/* Make sure compiler knows to grab that: we don't want it changing! */
1228 	/* We will use the result as an index in an array, so most
1229 	 * architectures only need a compiler barrier here. */
1230 	read_barrier_depends();
1231 
1232 	return next;
1233 }
1234 
1235 static int get_indirect(struct vhost_virtqueue *vq,
1236 			struct iovec iov[], unsigned int iov_size,
1237 			unsigned int *out_num, unsigned int *in_num,
1238 			struct vhost_log *log, unsigned int *log_num,
1239 			struct vring_desc *indirect)
1240 {
1241 	struct vring_desc desc;
1242 	unsigned int i = 0, count, found = 0;
1243 	u32 len = vhost32_to_cpu(vq, indirect->len);
1244 	struct iov_iter from;
1245 	int ret;
1246 
1247 	/* Sanity check */
1248 	if (unlikely(len % sizeof desc)) {
1249 		vq_err(vq, "Invalid length in indirect descriptor: "
1250 		       "len 0x%llx not multiple of 0x%zx\n",
1251 		       (unsigned long long)len,
1252 		       sizeof desc);
1253 		return -EINVAL;
1254 	}
1255 
1256 	ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1257 			     UIO_MAXIOV);
1258 	if (unlikely(ret < 0)) {
1259 		vq_err(vq, "Translation failure %d in indirect.\n", ret);
1260 		return ret;
1261 	}
1262 	iov_iter_init(&from, READ, vq->indirect, ret, len);
1263 
1264 	/* We will use the result as an address to read from, so most
1265 	 * architectures only need a compiler barrier here. */
1266 	read_barrier_depends();
1267 
1268 	count = len / sizeof desc;
1269 	/* Buffers are chained via a 16 bit next field, so
1270 	 * we can have at most 2^16 of these. */
1271 	if (unlikely(count > USHRT_MAX + 1)) {
1272 		vq_err(vq, "Indirect buffer length too big: %d\n",
1273 		       indirect->len);
1274 		return -E2BIG;
1275 	}
1276 
1277 	do {
1278 		unsigned iov_count = *in_num + *out_num;
1279 		if (unlikely(++found > count)) {
1280 			vq_err(vq, "Loop detected: last one at %u "
1281 			       "indirect size %u\n",
1282 			       i, count);
1283 			return -EINVAL;
1284 		}
1285 		if (unlikely(copy_from_iter(&desc, sizeof(desc), &from) !=
1286 			     sizeof(desc))) {
1287 			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1288 			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1289 			return -EINVAL;
1290 		}
1291 		if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1292 			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1293 			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1294 			return -EINVAL;
1295 		}
1296 
1297 		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1298 				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
1299 				     iov_size - iov_count);
1300 		if (unlikely(ret < 0)) {
1301 			vq_err(vq, "Translation failure %d indirect idx %d\n",
1302 			       ret, i);
1303 			return ret;
1304 		}
1305 		/* If this is an input descriptor, increment that count. */
1306 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1307 			*in_num += ret;
1308 			if (unlikely(log)) {
1309 				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1310 				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1311 				++*log_num;
1312 			}
1313 		} else {
1314 			/* If it's an output descriptor, they're all supposed
1315 			 * to come before any input descriptors. */
1316 			if (unlikely(*in_num)) {
1317 				vq_err(vq, "Indirect descriptor "
1318 				       "has out after in: idx %d\n", i);
1319 				return -EINVAL;
1320 			}
1321 			*out_num += ret;
1322 		}
1323 	} while ((i = next_desc(vq, &desc)) != -1);
1324 	return 0;
1325 }
1326 
1327 /* This looks in the virtqueue and for the first available buffer, and converts
1328  * it to an iovec for convenient access.  Since descriptors consist of some
1329  * number of output then some number of input descriptors, it's actually two
1330  * iovecs, but we pack them into one and note how many of each there were.
1331  *
1332  * This function returns the descriptor number found, or vq->num (which is
1333  * never a valid descriptor number) if none was found.  A negative code is
1334  * returned on error. */
1335 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1336 		      struct iovec iov[], unsigned int iov_size,
1337 		      unsigned int *out_num, unsigned int *in_num,
1338 		      struct vhost_log *log, unsigned int *log_num)
1339 {
1340 	struct vring_desc desc;
1341 	unsigned int i, head, found = 0;
1342 	u16 last_avail_idx;
1343 	__virtio16 avail_idx;
1344 	__virtio16 ring_head;
1345 	int ret;
1346 
1347 	/* Check it isn't doing very strange things with descriptor numbers. */
1348 	last_avail_idx = vq->last_avail_idx;
1349 	if (unlikely(__get_user(avail_idx, &vq->avail->idx))) {
1350 		vq_err(vq, "Failed to access avail idx at %p\n",
1351 		       &vq->avail->idx);
1352 		return -EFAULT;
1353 	}
1354 	vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1355 
1356 	if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1357 		vq_err(vq, "Guest moved used index from %u to %u",
1358 		       last_avail_idx, vq->avail_idx);
1359 		return -EFAULT;
1360 	}
1361 
1362 	/* If there's nothing new since last we looked, return invalid. */
1363 	if (vq->avail_idx == last_avail_idx)
1364 		return vq->num;
1365 
1366 	/* Only get avail ring entries after they have been exposed by guest. */
1367 	smp_rmb();
1368 
1369 	/* Grab the next descriptor number they're advertising, and increment
1370 	 * the index we've seen. */
1371 	if (unlikely(__get_user(ring_head,
1372 				&vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
1373 		vq_err(vq, "Failed to read head: idx %d address %p\n",
1374 		       last_avail_idx,
1375 		       &vq->avail->ring[last_avail_idx % vq->num]);
1376 		return -EFAULT;
1377 	}
1378 
1379 	head = vhost16_to_cpu(vq, ring_head);
1380 
1381 	/* If their number is silly, that's an error. */
1382 	if (unlikely(head >= vq->num)) {
1383 		vq_err(vq, "Guest says index %u > %u is available",
1384 		       head, vq->num);
1385 		return -EINVAL;
1386 	}
1387 
1388 	/* When we start there are none of either input nor output. */
1389 	*out_num = *in_num = 0;
1390 	if (unlikely(log))
1391 		*log_num = 0;
1392 
1393 	i = head;
1394 	do {
1395 		unsigned iov_count = *in_num + *out_num;
1396 		if (unlikely(i >= vq->num)) {
1397 			vq_err(vq, "Desc index is %u > %u, head = %u",
1398 			       i, vq->num, head);
1399 			return -EINVAL;
1400 		}
1401 		if (unlikely(++found > vq->num)) {
1402 			vq_err(vq, "Loop detected: last one at %u "
1403 			       "vq size %u head %u\n",
1404 			       i, vq->num, head);
1405 			return -EINVAL;
1406 		}
1407 		ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1408 		if (unlikely(ret)) {
1409 			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1410 			       i, vq->desc + i);
1411 			return -EFAULT;
1412 		}
1413 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
1414 			ret = get_indirect(vq, iov, iov_size,
1415 					   out_num, in_num,
1416 					   log, log_num, &desc);
1417 			if (unlikely(ret < 0)) {
1418 				vq_err(vq, "Failure detected "
1419 				       "in indirect descriptor at idx %d\n", i);
1420 				return ret;
1421 			}
1422 			continue;
1423 		}
1424 
1425 		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1426 				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
1427 				     iov_size - iov_count);
1428 		if (unlikely(ret < 0)) {
1429 			vq_err(vq, "Translation failure %d descriptor idx %d\n",
1430 			       ret, i);
1431 			return ret;
1432 		}
1433 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1434 			/* If this is an input descriptor,
1435 			 * increment that count. */
1436 			*in_num += ret;
1437 			if (unlikely(log)) {
1438 				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1439 				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1440 				++*log_num;
1441 			}
1442 		} else {
1443 			/* If it's an output descriptor, they're all supposed
1444 			 * to come before any input descriptors. */
1445 			if (unlikely(*in_num)) {
1446 				vq_err(vq, "Descriptor has out after in: "
1447 				       "idx %d\n", i);
1448 				return -EINVAL;
1449 			}
1450 			*out_num += ret;
1451 		}
1452 	} while ((i = next_desc(vq, &desc)) != -1);
1453 
1454 	/* On success, increment avail index. */
1455 	vq->last_avail_idx++;
1456 
1457 	/* Assume notifications from guest are disabled at this point,
1458 	 * if they aren't we would need to update avail_event index. */
1459 	BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1460 	return head;
1461 }
1462 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1463 
1464 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1465 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1466 {
1467 	vq->last_avail_idx -= n;
1468 }
1469 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1470 
1471 /* After we've used one of their buffers, we tell them about it.  We'll then
1472  * want to notify the guest, using eventfd. */
1473 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1474 {
1475 	struct vring_used_elem heads = {
1476 		cpu_to_vhost32(vq, head),
1477 		cpu_to_vhost32(vq, len)
1478 	};
1479 
1480 	return vhost_add_used_n(vq, &heads, 1);
1481 }
1482 EXPORT_SYMBOL_GPL(vhost_add_used);
1483 
1484 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1485 			    struct vring_used_elem *heads,
1486 			    unsigned count)
1487 {
1488 	struct vring_used_elem __user *used;
1489 	u16 old, new;
1490 	int start;
1491 
1492 	start = vq->last_used_idx & (vq->num - 1);
1493 	used = vq->used->ring + start;
1494 	if (count == 1) {
1495 		if (__put_user(heads[0].id, &used->id)) {
1496 			vq_err(vq, "Failed to write used id");
1497 			return -EFAULT;
1498 		}
1499 		if (__put_user(heads[0].len, &used->len)) {
1500 			vq_err(vq, "Failed to write used len");
1501 			return -EFAULT;
1502 		}
1503 	} else if (__copy_to_user(used, heads, count * sizeof *used)) {
1504 		vq_err(vq, "Failed to write used");
1505 		return -EFAULT;
1506 	}
1507 	if (unlikely(vq->log_used)) {
1508 		/* Make sure data is seen before log. */
1509 		smp_wmb();
1510 		/* Log used ring entry write. */
1511 		log_write(vq->log_base,
1512 			  vq->log_addr +
1513 			   ((void __user *)used - (void __user *)vq->used),
1514 			  count * sizeof *used);
1515 	}
1516 	old = vq->last_used_idx;
1517 	new = (vq->last_used_idx += count);
1518 	/* If the driver never bothers to signal in a very long while,
1519 	 * used index might wrap around. If that happens, invalidate
1520 	 * signalled_used index we stored. TODO: make sure driver
1521 	 * signals at least once in 2^16 and remove this. */
1522 	if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1523 		vq->signalled_used_valid = false;
1524 	return 0;
1525 }
1526 
1527 /* After we've used one of their buffers, we tell them about it.  We'll then
1528  * want to notify the guest, using eventfd. */
1529 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1530 		     unsigned count)
1531 {
1532 	int start, n, r;
1533 
1534 	start = vq->last_used_idx & (vq->num - 1);
1535 	n = vq->num - start;
1536 	if (n < count) {
1537 		r = __vhost_add_used_n(vq, heads, n);
1538 		if (r < 0)
1539 			return r;
1540 		heads += n;
1541 		count -= n;
1542 	}
1543 	r = __vhost_add_used_n(vq, heads, count);
1544 
1545 	/* Make sure buffer is written before we update index. */
1546 	smp_wmb();
1547 	if (__put_user(cpu_to_vhost16(vq, vq->last_used_idx), &vq->used->idx)) {
1548 		vq_err(vq, "Failed to increment used idx");
1549 		return -EFAULT;
1550 	}
1551 	if (unlikely(vq->log_used)) {
1552 		/* Log used index update. */
1553 		log_write(vq->log_base,
1554 			  vq->log_addr + offsetof(struct vring_used, idx),
1555 			  sizeof vq->used->idx);
1556 		if (vq->log_ctx)
1557 			eventfd_signal(vq->log_ctx, 1);
1558 	}
1559 	return r;
1560 }
1561 EXPORT_SYMBOL_GPL(vhost_add_used_n);
1562 
1563 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1564 {
1565 	__u16 old, new;
1566 	__virtio16 event;
1567 	bool v;
1568 	/* Flush out used index updates. This is paired
1569 	 * with the barrier that the Guest executes when enabling
1570 	 * interrupts. */
1571 	smp_mb();
1572 
1573 	if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1574 	    unlikely(vq->avail_idx == vq->last_avail_idx))
1575 		return true;
1576 
1577 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1578 		__virtio16 flags;
1579 		if (__get_user(flags, &vq->avail->flags)) {
1580 			vq_err(vq, "Failed to get flags");
1581 			return true;
1582 		}
1583 		return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
1584 	}
1585 	old = vq->signalled_used;
1586 	v = vq->signalled_used_valid;
1587 	new = vq->signalled_used = vq->last_used_idx;
1588 	vq->signalled_used_valid = true;
1589 
1590 	if (unlikely(!v))
1591 		return true;
1592 
1593 	if (__get_user(event, vhost_used_event(vq))) {
1594 		vq_err(vq, "Failed to get used event idx");
1595 		return true;
1596 	}
1597 	return vring_need_event(vhost16_to_cpu(vq, event), new, old);
1598 }
1599 
1600 /* This actually signals the guest, using eventfd. */
1601 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1602 {
1603 	/* Signal the Guest tell them we used something up. */
1604 	if (vq->call_ctx && vhost_notify(dev, vq))
1605 		eventfd_signal(vq->call_ctx, 1);
1606 }
1607 EXPORT_SYMBOL_GPL(vhost_signal);
1608 
1609 /* And here's the combo meal deal.  Supersize me! */
1610 void vhost_add_used_and_signal(struct vhost_dev *dev,
1611 			       struct vhost_virtqueue *vq,
1612 			       unsigned int head, int len)
1613 {
1614 	vhost_add_used(vq, head, len);
1615 	vhost_signal(dev, vq);
1616 }
1617 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1618 
1619 /* multi-buffer version of vhost_add_used_and_signal */
1620 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1621 				 struct vhost_virtqueue *vq,
1622 				 struct vring_used_elem *heads, unsigned count)
1623 {
1624 	vhost_add_used_n(vq, heads, count);
1625 	vhost_signal(dev, vq);
1626 }
1627 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1628 
1629 /* OK, now we need to know about added descriptors. */
1630 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1631 {
1632 	__virtio16 avail_idx;
1633 	int r;
1634 
1635 	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1636 		return false;
1637 	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1638 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1639 		r = vhost_update_used_flags(vq);
1640 		if (r) {
1641 			vq_err(vq, "Failed to enable notification at %p: %d\n",
1642 			       &vq->used->flags, r);
1643 			return false;
1644 		}
1645 	} else {
1646 		r = vhost_update_avail_event(vq, vq->avail_idx);
1647 		if (r) {
1648 			vq_err(vq, "Failed to update avail event index at %p: %d\n",
1649 			       vhost_avail_event(vq), r);
1650 			return false;
1651 		}
1652 	}
1653 	/* They could have slipped one in as we were doing that: make
1654 	 * sure it's written, then check again. */
1655 	smp_mb();
1656 	r = __get_user(avail_idx, &vq->avail->idx);
1657 	if (r) {
1658 		vq_err(vq, "Failed to check avail idx at %p: %d\n",
1659 		       &vq->avail->idx, r);
1660 		return false;
1661 	}
1662 
1663 	return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
1664 }
1665 EXPORT_SYMBOL_GPL(vhost_enable_notify);
1666 
1667 /* We don't need to be notified again. */
1668 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1669 {
1670 	int r;
1671 
1672 	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1673 		return;
1674 	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1675 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1676 		r = vhost_update_used_flags(vq);
1677 		if (r)
1678 			vq_err(vq, "Failed to enable notification at %p: %d\n",
1679 			       &vq->used->flags, r);
1680 	}
1681 }
1682 EXPORT_SYMBOL_GPL(vhost_disable_notify);
1683 
1684 static int __init vhost_init(void)
1685 {
1686 	return 0;
1687 }
1688 
1689 static void __exit vhost_exit(void)
1690 {
1691 }
1692 
1693 module_init(vhost_init);
1694 module_exit(vhost_exit);
1695 
1696 MODULE_VERSION("0.0.1");
1697 MODULE_LICENSE("GPL v2");
1698 MODULE_AUTHOR("Michael S. Tsirkin");
1699 MODULE_DESCRIPTION("Host kernel accelerator for virtio");
1700