xref: /linux/drivers/vhost/vhost.c (revision 93df8a1ed6231727c5db94a80b1a6bd5ee67cec3)
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 		if (!n)
689 			return ERR_PTR(-ENOMEM);
690 	}
691 	return n;
692 }
693 
694 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
695 {
696 	struct vhost_memory mem, *newmem, *oldmem;
697 	unsigned long size = offsetof(struct vhost_memory, regions);
698 	int i;
699 
700 	if (copy_from_user(&mem, m, size))
701 		return -EFAULT;
702 	if (mem.padding)
703 		return -EOPNOTSUPP;
704 	if (mem.nregions > max_mem_regions)
705 		return -E2BIG;
706 	newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
707 	if (!newmem)
708 		return -ENOMEM;
709 
710 	memcpy(newmem, &mem, size);
711 	if (copy_from_user(newmem->regions, m->regions,
712 			   mem.nregions * sizeof *m->regions)) {
713 		kvfree(newmem);
714 		return -EFAULT;
715 	}
716 	sort(newmem->regions, newmem->nregions, sizeof(*newmem->regions),
717 		vhost_memory_reg_sort_cmp, NULL);
718 
719 	if (!memory_access_ok(d, newmem, 0)) {
720 		kvfree(newmem);
721 		return -EFAULT;
722 	}
723 	oldmem = d->memory;
724 	d->memory = newmem;
725 
726 	/* All memory accesses are done under some VQ mutex. */
727 	for (i = 0; i < d->nvqs; ++i) {
728 		mutex_lock(&d->vqs[i]->mutex);
729 		d->vqs[i]->memory = newmem;
730 		mutex_unlock(&d->vqs[i]->mutex);
731 	}
732 	kvfree(oldmem);
733 	return 0;
734 }
735 
736 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
737 {
738 	struct file *eventfp, *filep = NULL;
739 	bool pollstart = false, pollstop = false;
740 	struct eventfd_ctx *ctx = NULL;
741 	u32 __user *idxp = argp;
742 	struct vhost_virtqueue *vq;
743 	struct vhost_vring_state s;
744 	struct vhost_vring_file f;
745 	struct vhost_vring_addr a;
746 	u32 idx;
747 	long r;
748 
749 	r = get_user(idx, idxp);
750 	if (r < 0)
751 		return r;
752 	if (idx >= d->nvqs)
753 		return -ENOBUFS;
754 
755 	vq = d->vqs[idx];
756 
757 	mutex_lock(&vq->mutex);
758 
759 	switch (ioctl) {
760 	case VHOST_SET_VRING_NUM:
761 		/* Resizing ring with an active backend?
762 		 * You don't want to do that. */
763 		if (vq->private_data) {
764 			r = -EBUSY;
765 			break;
766 		}
767 		if (copy_from_user(&s, argp, sizeof s)) {
768 			r = -EFAULT;
769 			break;
770 		}
771 		if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
772 			r = -EINVAL;
773 			break;
774 		}
775 		vq->num = s.num;
776 		break;
777 	case VHOST_SET_VRING_BASE:
778 		/* Moving base with an active backend?
779 		 * You don't want to do that. */
780 		if (vq->private_data) {
781 			r = -EBUSY;
782 			break;
783 		}
784 		if (copy_from_user(&s, argp, sizeof s)) {
785 			r = -EFAULT;
786 			break;
787 		}
788 		if (s.num > 0xffff) {
789 			r = -EINVAL;
790 			break;
791 		}
792 		vq->last_avail_idx = s.num;
793 		/* Forget the cached index value. */
794 		vq->avail_idx = vq->last_avail_idx;
795 		break;
796 	case VHOST_GET_VRING_BASE:
797 		s.index = idx;
798 		s.num = vq->last_avail_idx;
799 		if (copy_to_user(argp, &s, sizeof s))
800 			r = -EFAULT;
801 		break;
802 	case VHOST_SET_VRING_ADDR:
803 		if (copy_from_user(&a, argp, sizeof a)) {
804 			r = -EFAULT;
805 			break;
806 		}
807 		if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
808 			r = -EOPNOTSUPP;
809 			break;
810 		}
811 		/* For 32bit, verify that the top 32bits of the user
812 		   data are set to zero. */
813 		if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
814 		    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
815 		    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
816 			r = -EFAULT;
817 			break;
818 		}
819 
820 		/* Make sure it's safe to cast pointers to vring types. */
821 		BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
822 		BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
823 		if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
824 		    (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
825 		    (a.log_guest_addr & (sizeof(u64) - 1))) {
826 			r = -EINVAL;
827 			break;
828 		}
829 
830 		/* We only verify access here if backend is configured.
831 		 * If it is not, we don't as size might not have been setup.
832 		 * We will verify when backend is configured. */
833 		if (vq->private_data) {
834 			if (!vq_access_ok(vq, vq->num,
835 				(void __user *)(unsigned long)a.desc_user_addr,
836 				(void __user *)(unsigned long)a.avail_user_addr,
837 				(void __user *)(unsigned long)a.used_user_addr)) {
838 				r = -EINVAL;
839 				break;
840 			}
841 
842 			/* Also validate log access for used ring if enabled. */
843 			if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
844 			    !log_access_ok(vq->log_base, a.log_guest_addr,
845 					   sizeof *vq->used +
846 					   vq->num * sizeof *vq->used->ring)) {
847 				r = -EINVAL;
848 				break;
849 			}
850 		}
851 
852 		vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
853 		vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
854 		vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
855 		vq->log_addr = a.log_guest_addr;
856 		vq->used = (void __user *)(unsigned long)a.used_user_addr;
857 		break;
858 	case VHOST_SET_VRING_KICK:
859 		if (copy_from_user(&f, argp, sizeof f)) {
860 			r = -EFAULT;
861 			break;
862 		}
863 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
864 		if (IS_ERR(eventfp)) {
865 			r = PTR_ERR(eventfp);
866 			break;
867 		}
868 		if (eventfp != vq->kick) {
869 			pollstop = (filep = vq->kick) != NULL;
870 			pollstart = (vq->kick = eventfp) != NULL;
871 		} else
872 			filep = eventfp;
873 		break;
874 	case VHOST_SET_VRING_CALL:
875 		if (copy_from_user(&f, argp, sizeof f)) {
876 			r = -EFAULT;
877 			break;
878 		}
879 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
880 		if (IS_ERR(eventfp)) {
881 			r = PTR_ERR(eventfp);
882 			break;
883 		}
884 		if (eventfp != vq->call) {
885 			filep = vq->call;
886 			ctx = vq->call_ctx;
887 			vq->call = eventfp;
888 			vq->call_ctx = eventfp ?
889 				eventfd_ctx_fileget(eventfp) : NULL;
890 		} else
891 			filep = eventfp;
892 		break;
893 	case VHOST_SET_VRING_ERR:
894 		if (copy_from_user(&f, argp, sizeof f)) {
895 			r = -EFAULT;
896 			break;
897 		}
898 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
899 		if (IS_ERR(eventfp)) {
900 			r = PTR_ERR(eventfp);
901 			break;
902 		}
903 		if (eventfp != vq->error) {
904 			filep = vq->error;
905 			vq->error = eventfp;
906 			ctx = vq->error_ctx;
907 			vq->error_ctx = eventfp ?
908 				eventfd_ctx_fileget(eventfp) : NULL;
909 		} else
910 			filep = eventfp;
911 		break;
912 	case VHOST_SET_VRING_ENDIAN:
913 		r = vhost_set_vring_endian(vq, argp);
914 		break;
915 	case VHOST_GET_VRING_ENDIAN:
916 		r = vhost_get_vring_endian(vq, idx, argp);
917 		break;
918 	default:
919 		r = -ENOIOCTLCMD;
920 	}
921 
922 	if (pollstop && vq->handle_kick)
923 		vhost_poll_stop(&vq->poll);
924 
925 	if (ctx)
926 		eventfd_ctx_put(ctx);
927 	if (filep)
928 		fput(filep);
929 
930 	if (pollstart && vq->handle_kick)
931 		r = vhost_poll_start(&vq->poll, vq->kick);
932 
933 	mutex_unlock(&vq->mutex);
934 
935 	if (pollstop && vq->handle_kick)
936 		vhost_poll_flush(&vq->poll);
937 	return r;
938 }
939 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
940 
941 /* Caller must have device mutex */
942 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
943 {
944 	struct file *eventfp, *filep = NULL;
945 	struct eventfd_ctx *ctx = NULL;
946 	u64 p;
947 	long r;
948 	int i, fd;
949 
950 	/* If you are not the owner, you can become one */
951 	if (ioctl == VHOST_SET_OWNER) {
952 		r = vhost_dev_set_owner(d);
953 		goto done;
954 	}
955 
956 	/* You must be the owner to do anything else */
957 	r = vhost_dev_check_owner(d);
958 	if (r)
959 		goto done;
960 
961 	switch (ioctl) {
962 	case VHOST_SET_MEM_TABLE:
963 		r = vhost_set_memory(d, argp);
964 		break;
965 	case VHOST_SET_LOG_BASE:
966 		if (copy_from_user(&p, argp, sizeof p)) {
967 			r = -EFAULT;
968 			break;
969 		}
970 		if ((u64)(unsigned long)p != p) {
971 			r = -EFAULT;
972 			break;
973 		}
974 		for (i = 0; i < d->nvqs; ++i) {
975 			struct vhost_virtqueue *vq;
976 			void __user *base = (void __user *)(unsigned long)p;
977 			vq = d->vqs[i];
978 			mutex_lock(&vq->mutex);
979 			/* If ring is inactive, will check when it's enabled. */
980 			if (vq->private_data && !vq_log_access_ok(vq, base))
981 				r = -EFAULT;
982 			else
983 				vq->log_base = base;
984 			mutex_unlock(&vq->mutex);
985 		}
986 		break;
987 	case VHOST_SET_LOG_FD:
988 		r = get_user(fd, (int __user *)argp);
989 		if (r < 0)
990 			break;
991 		eventfp = fd == -1 ? NULL : eventfd_fget(fd);
992 		if (IS_ERR(eventfp)) {
993 			r = PTR_ERR(eventfp);
994 			break;
995 		}
996 		if (eventfp != d->log_file) {
997 			filep = d->log_file;
998 			ctx = d->log_ctx;
999 			d->log_ctx = eventfp ?
1000 				eventfd_ctx_fileget(eventfp) : NULL;
1001 		} else
1002 			filep = eventfp;
1003 		for (i = 0; i < d->nvqs; ++i) {
1004 			mutex_lock(&d->vqs[i]->mutex);
1005 			d->vqs[i]->log_ctx = d->log_ctx;
1006 			mutex_unlock(&d->vqs[i]->mutex);
1007 		}
1008 		if (ctx)
1009 			eventfd_ctx_put(ctx);
1010 		if (filep)
1011 			fput(filep);
1012 		break;
1013 	default:
1014 		r = -ENOIOCTLCMD;
1015 		break;
1016 	}
1017 done:
1018 	return r;
1019 }
1020 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1021 
1022 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
1023 						     __u64 addr, __u32 len)
1024 {
1025 	const struct vhost_memory_region *reg;
1026 	int start = 0, end = mem->nregions;
1027 
1028 	while (start < end) {
1029 		int slot = start + (end - start) / 2;
1030 		reg = mem->regions + slot;
1031 		if (addr >= reg->guest_phys_addr)
1032 			end = slot;
1033 		else
1034 			start = slot + 1;
1035 	}
1036 
1037 	reg = mem->regions + start;
1038 	if (addr >= reg->guest_phys_addr &&
1039 		reg->guest_phys_addr + reg->memory_size > addr)
1040 		return reg;
1041 	return NULL;
1042 }
1043 
1044 /* TODO: This is really inefficient.  We need something like get_user()
1045  * (instruction directly accesses the data, with an exception table entry
1046  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1047  */
1048 static int set_bit_to_user(int nr, void __user *addr)
1049 {
1050 	unsigned long log = (unsigned long)addr;
1051 	struct page *page;
1052 	void *base;
1053 	int bit = nr + (log % PAGE_SIZE) * 8;
1054 	int r;
1055 
1056 	r = get_user_pages_fast(log, 1, 1, &page);
1057 	if (r < 0)
1058 		return r;
1059 	BUG_ON(r != 1);
1060 	base = kmap_atomic(page);
1061 	set_bit(bit, base);
1062 	kunmap_atomic(base);
1063 	set_page_dirty_lock(page);
1064 	put_page(page);
1065 	return 0;
1066 }
1067 
1068 static int log_write(void __user *log_base,
1069 		     u64 write_address, u64 write_length)
1070 {
1071 	u64 write_page = write_address / VHOST_PAGE_SIZE;
1072 	int r;
1073 
1074 	if (!write_length)
1075 		return 0;
1076 	write_length += write_address % VHOST_PAGE_SIZE;
1077 	for (;;) {
1078 		u64 base = (u64)(unsigned long)log_base;
1079 		u64 log = base + write_page / 8;
1080 		int bit = write_page % 8;
1081 		if ((u64)(unsigned long)log != log)
1082 			return -EFAULT;
1083 		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1084 		if (r < 0)
1085 			return r;
1086 		if (write_length <= VHOST_PAGE_SIZE)
1087 			break;
1088 		write_length -= VHOST_PAGE_SIZE;
1089 		write_page += 1;
1090 	}
1091 	return r;
1092 }
1093 
1094 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1095 		    unsigned int log_num, u64 len)
1096 {
1097 	int i, r;
1098 
1099 	/* Make sure data written is seen before log. */
1100 	smp_wmb();
1101 	for (i = 0; i < log_num; ++i) {
1102 		u64 l = min(log[i].len, len);
1103 		r = log_write(vq->log_base, log[i].addr, l);
1104 		if (r < 0)
1105 			return r;
1106 		len -= l;
1107 		if (!len) {
1108 			if (vq->log_ctx)
1109 				eventfd_signal(vq->log_ctx, 1);
1110 			return 0;
1111 		}
1112 	}
1113 	/* Length written exceeds what we have stored. This is a bug. */
1114 	BUG();
1115 	return 0;
1116 }
1117 EXPORT_SYMBOL_GPL(vhost_log_write);
1118 
1119 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1120 {
1121 	void __user *used;
1122 	if (__put_user(cpu_to_vhost16(vq, vq->used_flags), &vq->used->flags) < 0)
1123 		return -EFAULT;
1124 	if (unlikely(vq->log_used)) {
1125 		/* Make sure the flag is seen before log. */
1126 		smp_wmb();
1127 		/* Log used flag write. */
1128 		used = &vq->used->flags;
1129 		log_write(vq->log_base, vq->log_addr +
1130 			  (used - (void __user *)vq->used),
1131 			  sizeof vq->used->flags);
1132 		if (vq->log_ctx)
1133 			eventfd_signal(vq->log_ctx, 1);
1134 	}
1135 	return 0;
1136 }
1137 
1138 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1139 {
1140 	if (__put_user(cpu_to_vhost16(vq, vq->avail_idx), vhost_avail_event(vq)))
1141 		return -EFAULT;
1142 	if (unlikely(vq->log_used)) {
1143 		void __user *used;
1144 		/* Make sure the event is seen before log. */
1145 		smp_wmb();
1146 		/* Log avail event write */
1147 		used = vhost_avail_event(vq);
1148 		log_write(vq->log_base, vq->log_addr +
1149 			  (used - (void __user *)vq->used),
1150 			  sizeof *vhost_avail_event(vq));
1151 		if (vq->log_ctx)
1152 			eventfd_signal(vq->log_ctx, 1);
1153 	}
1154 	return 0;
1155 }
1156 
1157 int vhost_init_used(struct vhost_virtqueue *vq)
1158 {
1159 	__virtio16 last_used_idx;
1160 	int r;
1161 	if (!vq->private_data) {
1162 		vq->is_le = virtio_legacy_is_little_endian();
1163 		return 0;
1164 	}
1165 
1166 	vhost_init_is_le(vq);
1167 
1168 	r = vhost_update_used_flags(vq);
1169 	if (r)
1170 		return r;
1171 	vq->signalled_used_valid = false;
1172 	if (!access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx))
1173 		return -EFAULT;
1174 	r = __get_user(last_used_idx, &vq->used->idx);
1175 	if (r)
1176 		return r;
1177 	vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1178 	return 0;
1179 }
1180 EXPORT_SYMBOL_GPL(vhost_init_used);
1181 
1182 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1183 			  struct iovec iov[], int iov_size)
1184 {
1185 	const struct vhost_memory_region *reg;
1186 	struct vhost_memory *mem;
1187 	struct iovec *_iov;
1188 	u64 s = 0;
1189 	int ret = 0;
1190 
1191 	mem = vq->memory;
1192 	while ((u64)len > s) {
1193 		u64 size;
1194 		if (unlikely(ret >= iov_size)) {
1195 			ret = -ENOBUFS;
1196 			break;
1197 		}
1198 		reg = find_region(mem, addr, len);
1199 		if (unlikely(!reg)) {
1200 			ret = -EFAULT;
1201 			break;
1202 		}
1203 		_iov = iov + ret;
1204 		size = reg->memory_size - addr + reg->guest_phys_addr;
1205 		_iov->iov_len = min((u64)len - s, size);
1206 		_iov->iov_base = (void __user *)(unsigned long)
1207 			(reg->userspace_addr + addr - reg->guest_phys_addr);
1208 		s += size;
1209 		addr += size;
1210 		++ret;
1211 	}
1212 
1213 	return ret;
1214 }
1215 
1216 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1217  * function returns the next descriptor in the chain,
1218  * or -1U if we're at the end. */
1219 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1220 {
1221 	unsigned int next;
1222 
1223 	/* If this descriptor says it doesn't chain, we're done. */
1224 	if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1225 		return -1U;
1226 
1227 	/* Check they're not leading us off end of descriptors. */
1228 	next = vhost16_to_cpu(vq, desc->next);
1229 	/* Make sure compiler knows to grab that: we don't want it changing! */
1230 	/* We will use the result as an index in an array, so most
1231 	 * architectures only need a compiler barrier here. */
1232 	read_barrier_depends();
1233 
1234 	return next;
1235 }
1236 
1237 static int get_indirect(struct vhost_virtqueue *vq,
1238 			struct iovec iov[], unsigned int iov_size,
1239 			unsigned int *out_num, unsigned int *in_num,
1240 			struct vhost_log *log, unsigned int *log_num,
1241 			struct vring_desc *indirect)
1242 {
1243 	struct vring_desc desc;
1244 	unsigned int i = 0, count, found = 0;
1245 	u32 len = vhost32_to_cpu(vq, indirect->len);
1246 	struct iov_iter from;
1247 	int ret;
1248 
1249 	/* Sanity check */
1250 	if (unlikely(len % sizeof desc)) {
1251 		vq_err(vq, "Invalid length in indirect descriptor: "
1252 		       "len 0x%llx not multiple of 0x%zx\n",
1253 		       (unsigned long long)len,
1254 		       sizeof desc);
1255 		return -EINVAL;
1256 	}
1257 
1258 	ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1259 			     UIO_MAXIOV);
1260 	if (unlikely(ret < 0)) {
1261 		vq_err(vq, "Translation failure %d in indirect.\n", ret);
1262 		return ret;
1263 	}
1264 	iov_iter_init(&from, READ, vq->indirect, ret, len);
1265 
1266 	/* We will use the result as an address to read from, so most
1267 	 * architectures only need a compiler barrier here. */
1268 	read_barrier_depends();
1269 
1270 	count = len / sizeof desc;
1271 	/* Buffers are chained via a 16 bit next field, so
1272 	 * we can have at most 2^16 of these. */
1273 	if (unlikely(count > USHRT_MAX + 1)) {
1274 		vq_err(vq, "Indirect buffer length too big: %d\n",
1275 		       indirect->len);
1276 		return -E2BIG;
1277 	}
1278 
1279 	do {
1280 		unsigned iov_count = *in_num + *out_num;
1281 		if (unlikely(++found > count)) {
1282 			vq_err(vq, "Loop detected: last one at %u "
1283 			       "indirect size %u\n",
1284 			       i, count);
1285 			return -EINVAL;
1286 		}
1287 		if (unlikely(copy_from_iter(&desc, sizeof(desc), &from) !=
1288 			     sizeof(desc))) {
1289 			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1290 			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1291 			return -EINVAL;
1292 		}
1293 		if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1294 			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1295 			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1296 			return -EINVAL;
1297 		}
1298 
1299 		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1300 				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
1301 				     iov_size - iov_count);
1302 		if (unlikely(ret < 0)) {
1303 			vq_err(vq, "Translation failure %d indirect idx %d\n",
1304 			       ret, i);
1305 			return ret;
1306 		}
1307 		/* If this is an input descriptor, increment that count. */
1308 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1309 			*in_num += ret;
1310 			if (unlikely(log)) {
1311 				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1312 				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1313 				++*log_num;
1314 			}
1315 		} else {
1316 			/* If it's an output descriptor, they're all supposed
1317 			 * to come before any input descriptors. */
1318 			if (unlikely(*in_num)) {
1319 				vq_err(vq, "Indirect descriptor "
1320 				       "has out after in: idx %d\n", i);
1321 				return -EINVAL;
1322 			}
1323 			*out_num += ret;
1324 		}
1325 	} while ((i = next_desc(vq, &desc)) != -1);
1326 	return 0;
1327 }
1328 
1329 /* This looks in the virtqueue and for the first available buffer, and converts
1330  * it to an iovec for convenient access.  Since descriptors consist of some
1331  * number of output then some number of input descriptors, it's actually two
1332  * iovecs, but we pack them into one and note how many of each there were.
1333  *
1334  * This function returns the descriptor number found, or vq->num (which is
1335  * never a valid descriptor number) if none was found.  A negative code is
1336  * returned on error. */
1337 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1338 		      struct iovec iov[], unsigned int iov_size,
1339 		      unsigned int *out_num, unsigned int *in_num,
1340 		      struct vhost_log *log, unsigned int *log_num)
1341 {
1342 	struct vring_desc desc;
1343 	unsigned int i, head, found = 0;
1344 	u16 last_avail_idx;
1345 	__virtio16 avail_idx;
1346 	__virtio16 ring_head;
1347 	int ret;
1348 
1349 	/* Check it isn't doing very strange things with descriptor numbers. */
1350 	last_avail_idx = vq->last_avail_idx;
1351 	if (unlikely(__get_user(avail_idx, &vq->avail->idx))) {
1352 		vq_err(vq, "Failed to access avail idx at %p\n",
1353 		       &vq->avail->idx);
1354 		return -EFAULT;
1355 	}
1356 	vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1357 
1358 	if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1359 		vq_err(vq, "Guest moved used index from %u to %u",
1360 		       last_avail_idx, vq->avail_idx);
1361 		return -EFAULT;
1362 	}
1363 
1364 	/* If there's nothing new since last we looked, return invalid. */
1365 	if (vq->avail_idx == last_avail_idx)
1366 		return vq->num;
1367 
1368 	/* Only get avail ring entries after they have been exposed by guest. */
1369 	smp_rmb();
1370 
1371 	/* Grab the next descriptor number they're advertising, and increment
1372 	 * the index we've seen. */
1373 	if (unlikely(__get_user(ring_head,
1374 				&vq->avail->ring[last_avail_idx % vq->num]))) {
1375 		vq_err(vq, "Failed to read head: idx %d address %p\n",
1376 		       last_avail_idx,
1377 		       &vq->avail->ring[last_avail_idx % vq->num]);
1378 		return -EFAULT;
1379 	}
1380 
1381 	head = vhost16_to_cpu(vq, ring_head);
1382 
1383 	/* If their number is silly, that's an error. */
1384 	if (unlikely(head >= vq->num)) {
1385 		vq_err(vq, "Guest says index %u > %u is available",
1386 		       head, vq->num);
1387 		return -EINVAL;
1388 	}
1389 
1390 	/* When we start there are none of either input nor output. */
1391 	*out_num = *in_num = 0;
1392 	if (unlikely(log))
1393 		*log_num = 0;
1394 
1395 	i = head;
1396 	do {
1397 		unsigned iov_count = *in_num + *out_num;
1398 		if (unlikely(i >= vq->num)) {
1399 			vq_err(vq, "Desc index is %u > %u, head = %u",
1400 			       i, vq->num, head);
1401 			return -EINVAL;
1402 		}
1403 		if (unlikely(++found > vq->num)) {
1404 			vq_err(vq, "Loop detected: last one at %u "
1405 			       "vq size %u head %u\n",
1406 			       i, vq->num, head);
1407 			return -EINVAL;
1408 		}
1409 		ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1410 		if (unlikely(ret)) {
1411 			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1412 			       i, vq->desc + i);
1413 			return -EFAULT;
1414 		}
1415 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
1416 			ret = get_indirect(vq, iov, iov_size,
1417 					   out_num, in_num,
1418 					   log, log_num, &desc);
1419 			if (unlikely(ret < 0)) {
1420 				vq_err(vq, "Failure detected "
1421 				       "in indirect descriptor at idx %d\n", i);
1422 				return ret;
1423 			}
1424 			continue;
1425 		}
1426 
1427 		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1428 				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
1429 				     iov_size - iov_count);
1430 		if (unlikely(ret < 0)) {
1431 			vq_err(vq, "Translation failure %d descriptor idx %d\n",
1432 			       ret, i);
1433 			return ret;
1434 		}
1435 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1436 			/* If this is an input descriptor,
1437 			 * increment that count. */
1438 			*in_num += ret;
1439 			if (unlikely(log)) {
1440 				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1441 				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1442 				++*log_num;
1443 			}
1444 		} else {
1445 			/* If it's an output descriptor, they're all supposed
1446 			 * to come before any input descriptors. */
1447 			if (unlikely(*in_num)) {
1448 				vq_err(vq, "Descriptor has out after in: "
1449 				       "idx %d\n", i);
1450 				return -EINVAL;
1451 			}
1452 			*out_num += ret;
1453 		}
1454 	} while ((i = next_desc(vq, &desc)) != -1);
1455 
1456 	/* On success, increment avail index. */
1457 	vq->last_avail_idx++;
1458 
1459 	/* Assume notifications from guest are disabled at this point,
1460 	 * if they aren't we would need to update avail_event index. */
1461 	BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1462 	return head;
1463 }
1464 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1465 
1466 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1467 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1468 {
1469 	vq->last_avail_idx -= n;
1470 }
1471 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1472 
1473 /* After we've used one of their buffers, we tell them about it.  We'll then
1474  * want to notify the guest, using eventfd. */
1475 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1476 {
1477 	struct vring_used_elem heads = {
1478 		cpu_to_vhost32(vq, head),
1479 		cpu_to_vhost32(vq, len)
1480 	};
1481 
1482 	return vhost_add_used_n(vq, &heads, 1);
1483 }
1484 EXPORT_SYMBOL_GPL(vhost_add_used);
1485 
1486 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1487 			    struct vring_used_elem *heads,
1488 			    unsigned count)
1489 {
1490 	struct vring_used_elem __user *used;
1491 	u16 old, new;
1492 	int start;
1493 
1494 	start = vq->last_used_idx % vq->num;
1495 	used = vq->used->ring + start;
1496 	if (count == 1) {
1497 		if (__put_user(heads[0].id, &used->id)) {
1498 			vq_err(vq, "Failed to write used id");
1499 			return -EFAULT;
1500 		}
1501 		if (__put_user(heads[0].len, &used->len)) {
1502 			vq_err(vq, "Failed to write used len");
1503 			return -EFAULT;
1504 		}
1505 	} else if (__copy_to_user(used, heads, count * sizeof *used)) {
1506 		vq_err(vq, "Failed to write used");
1507 		return -EFAULT;
1508 	}
1509 	if (unlikely(vq->log_used)) {
1510 		/* Make sure data is seen before log. */
1511 		smp_wmb();
1512 		/* Log used ring entry write. */
1513 		log_write(vq->log_base,
1514 			  vq->log_addr +
1515 			   ((void __user *)used - (void __user *)vq->used),
1516 			  count * sizeof *used);
1517 	}
1518 	old = vq->last_used_idx;
1519 	new = (vq->last_used_idx += count);
1520 	/* If the driver never bothers to signal in a very long while,
1521 	 * used index might wrap around. If that happens, invalidate
1522 	 * signalled_used index we stored. TODO: make sure driver
1523 	 * signals at least once in 2^16 and remove this. */
1524 	if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1525 		vq->signalled_used_valid = false;
1526 	return 0;
1527 }
1528 
1529 /* After we've used one of their buffers, we tell them about it.  We'll then
1530  * want to notify the guest, using eventfd. */
1531 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1532 		     unsigned count)
1533 {
1534 	int start, n, r;
1535 
1536 	start = vq->last_used_idx % vq->num;
1537 	n = vq->num - start;
1538 	if (n < count) {
1539 		r = __vhost_add_used_n(vq, heads, n);
1540 		if (r < 0)
1541 			return r;
1542 		heads += n;
1543 		count -= n;
1544 	}
1545 	r = __vhost_add_used_n(vq, heads, count);
1546 
1547 	/* Make sure buffer is written before we update index. */
1548 	smp_wmb();
1549 	if (__put_user(cpu_to_vhost16(vq, vq->last_used_idx), &vq->used->idx)) {
1550 		vq_err(vq, "Failed to increment used idx");
1551 		return -EFAULT;
1552 	}
1553 	if (unlikely(vq->log_used)) {
1554 		/* Log used index update. */
1555 		log_write(vq->log_base,
1556 			  vq->log_addr + offsetof(struct vring_used, idx),
1557 			  sizeof vq->used->idx);
1558 		if (vq->log_ctx)
1559 			eventfd_signal(vq->log_ctx, 1);
1560 	}
1561 	return r;
1562 }
1563 EXPORT_SYMBOL_GPL(vhost_add_used_n);
1564 
1565 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1566 {
1567 	__u16 old, new;
1568 	__virtio16 event;
1569 	bool v;
1570 	/* Flush out used index updates. This is paired
1571 	 * with the barrier that the Guest executes when enabling
1572 	 * interrupts. */
1573 	smp_mb();
1574 
1575 	if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1576 	    unlikely(vq->avail_idx == vq->last_avail_idx))
1577 		return true;
1578 
1579 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1580 		__virtio16 flags;
1581 		if (__get_user(flags, &vq->avail->flags)) {
1582 			vq_err(vq, "Failed to get flags");
1583 			return true;
1584 		}
1585 		return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
1586 	}
1587 	old = vq->signalled_used;
1588 	v = vq->signalled_used_valid;
1589 	new = vq->signalled_used = vq->last_used_idx;
1590 	vq->signalled_used_valid = true;
1591 
1592 	if (unlikely(!v))
1593 		return true;
1594 
1595 	if (__get_user(event, vhost_used_event(vq))) {
1596 		vq_err(vq, "Failed to get used event idx");
1597 		return true;
1598 	}
1599 	return vring_need_event(vhost16_to_cpu(vq, event), new, old);
1600 }
1601 
1602 /* This actually signals the guest, using eventfd. */
1603 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1604 {
1605 	/* Signal the Guest tell them we used something up. */
1606 	if (vq->call_ctx && vhost_notify(dev, vq))
1607 		eventfd_signal(vq->call_ctx, 1);
1608 }
1609 EXPORT_SYMBOL_GPL(vhost_signal);
1610 
1611 /* And here's the combo meal deal.  Supersize me! */
1612 void vhost_add_used_and_signal(struct vhost_dev *dev,
1613 			       struct vhost_virtqueue *vq,
1614 			       unsigned int head, int len)
1615 {
1616 	vhost_add_used(vq, head, len);
1617 	vhost_signal(dev, vq);
1618 }
1619 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1620 
1621 /* multi-buffer version of vhost_add_used_and_signal */
1622 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1623 				 struct vhost_virtqueue *vq,
1624 				 struct vring_used_elem *heads, unsigned count)
1625 {
1626 	vhost_add_used_n(vq, heads, count);
1627 	vhost_signal(dev, vq);
1628 }
1629 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1630 
1631 /* OK, now we need to know about added descriptors. */
1632 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1633 {
1634 	__virtio16 avail_idx;
1635 	int r;
1636 
1637 	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1638 		return false;
1639 	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1640 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1641 		r = vhost_update_used_flags(vq);
1642 		if (r) {
1643 			vq_err(vq, "Failed to enable notification at %p: %d\n",
1644 			       &vq->used->flags, r);
1645 			return false;
1646 		}
1647 	} else {
1648 		r = vhost_update_avail_event(vq, vq->avail_idx);
1649 		if (r) {
1650 			vq_err(vq, "Failed to update avail event index at %p: %d\n",
1651 			       vhost_avail_event(vq), r);
1652 			return false;
1653 		}
1654 	}
1655 	/* They could have slipped one in as we were doing that: make
1656 	 * sure it's written, then check again. */
1657 	smp_mb();
1658 	r = __get_user(avail_idx, &vq->avail->idx);
1659 	if (r) {
1660 		vq_err(vq, "Failed to check avail idx at %p: %d\n",
1661 		       &vq->avail->idx, r);
1662 		return false;
1663 	}
1664 
1665 	return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
1666 }
1667 EXPORT_SYMBOL_GPL(vhost_enable_notify);
1668 
1669 /* We don't need to be notified again. */
1670 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1671 {
1672 	int r;
1673 
1674 	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1675 		return;
1676 	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1677 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1678 		r = vhost_update_used_flags(vq);
1679 		if (r)
1680 			vq_err(vq, "Failed to enable notification at %p: %d\n",
1681 			       &vq->used->flags, r);
1682 	}
1683 }
1684 EXPORT_SYMBOL_GPL(vhost_disable_notify);
1685 
1686 static int __init vhost_init(void)
1687 {
1688 	return 0;
1689 }
1690 
1691 static void __exit vhost_exit(void)
1692 {
1693 }
1694 
1695 module_init(vhost_init);
1696 module_exit(vhost_exit);
1697 
1698 MODULE_VERSION("0.0.1");
1699 MODULE_LICENSE("GPL v2");
1700 MODULE_AUTHOR("Michael S. Tsirkin");
1701 MODULE_DESCRIPTION("Host kernel accelerator for virtio");
1702