xref: /linux/drivers/vhost/vhost.c (revision a115bc070b1fc57ab23f3972401425927b5b465c)
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/lguest/lguest.c, by Rusty Russell
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.
10  *
11  * Generic code for virtio server in host kernel.
12  */
13 
14 #include <linux/eventfd.h>
15 #include <linux/vhost.h>
16 #include <linux/virtio_net.h>
17 #include <linux/mm.h>
18 #include <linux/miscdevice.h>
19 #include <linux/mutex.h>
20 #include <linux/workqueue.h>
21 #include <linux/rcupdate.h>
22 #include <linux/poll.h>
23 #include <linux/file.h>
24 #include <linux/highmem.h>
25 
26 #include <linux/net.h>
27 #include <linux/if_packet.h>
28 #include <linux/if_arp.h>
29 
30 #include <net/sock.h>
31 
32 #include "vhost.h"
33 
34 enum {
35 	VHOST_MEMORY_MAX_NREGIONS = 64,
36 	VHOST_MEMORY_F_LOG = 0x1,
37 };
38 
39 static struct workqueue_struct *vhost_workqueue;
40 
41 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
42 			    poll_table *pt)
43 {
44 	struct vhost_poll *poll;
45 	poll = container_of(pt, struct vhost_poll, table);
46 
47 	poll->wqh = wqh;
48 	add_wait_queue(wqh, &poll->wait);
49 }
50 
51 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
52 			     void *key)
53 {
54 	struct vhost_poll *poll;
55 	poll = container_of(wait, struct vhost_poll, wait);
56 	if (!((unsigned long)key & poll->mask))
57 		return 0;
58 
59 	queue_work(vhost_workqueue, &poll->work);
60 	return 0;
61 }
62 
63 /* Init poll structure */
64 void vhost_poll_init(struct vhost_poll *poll, work_func_t func,
65 		     unsigned long mask)
66 {
67 	INIT_WORK(&poll->work, func);
68 	init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
69 	init_poll_funcptr(&poll->table, vhost_poll_func);
70 	poll->mask = mask;
71 }
72 
73 /* Start polling a file. We add ourselves to file's wait queue. The caller must
74  * keep a reference to a file until after vhost_poll_stop is called. */
75 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
76 {
77 	unsigned long mask;
78 	mask = file->f_op->poll(file, &poll->table);
79 	if (mask)
80 		vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
81 }
82 
83 /* Stop polling a file. After this function returns, it becomes safe to drop the
84  * file reference. You must also flush afterwards. */
85 void vhost_poll_stop(struct vhost_poll *poll)
86 {
87 	remove_wait_queue(poll->wqh, &poll->wait);
88 }
89 
90 /* Flush any work that has been scheduled. When calling this, don't hold any
91  * locks that are also used by the callback. */
92 void vhost_poll_flush(struct vhost_poll *poll)
93 {
94 	flush_work(&poll->work);
95 }
96 
97 void vhost_poll_queue(struct vhost_poll *poll)
98 {
99 	queue_work(vhost_workqueue, &poll->work);
100 }
101 
102 static void vhost_vq_reset(struct vhost_dev *dev,
103 			   struct vhost_virtqueue *vq)
104 {
105 	vq->num = 1;
106 	vq->desc = NULL;
107 	vq->avail = NULL;
108 	vq->used = NULL;
109 	vq->last_avail_idx = 0;
110 	vq->avail_idx = 0;
111 	vq->last_used_idx = 0;
112 	vq->used_flags = 0;
113 	vq->used_flags = 0;
114 	vq->log_used = false;
115 	vq->log_addr = -1ull;
116 	vq->hdr_size = 0;
117 	vq->private_data = NULL;
118 	vq->log_base = NULL;
119 	vq->error_ctx = NULL;
120 	vq->error = NULL;
121 	vq->kick = NULL;
122 	vq->call_ctx = NULL;
123 	vq->call = NULL;
124 	vq->log_ctx = NULL;
125 }
126 
127 long vhost_dev_init(struct vhost_dev *dev,
128 		    struct vhost_virtqueue *vqs, int nvqs)
129 {
130 	int i;
131 	dev->vqs = vqs;
132 	dev->nvqs = nvqs;
133 	mutex_init(&dev->mutex);
134 	dev->log_ctx = NULL;
135 	dev->log_file = NULL;
136 	dev->memory = NULL;
137 	dev->mm = NULL;
138 
139 	for (i = 0; i < dev->nvqs; ++i) {
140 		dev->vqs[i].dev = dev;
141 		mutex_init(&dev->vqs[i].mutex);
142 		vhost_vq_reset(dev, dev->vqs + i);
143 		if (dev->vqs[i].handle_kick)
144 			vhost_poll_init(&dev->vqs[i].poll,
145 					dev->vqs[i].handle_kick,
146 					POLLIN);
147 	}
148 	return 0;
149 }
150 
151 /* Caller should have device mutex */
152 long vhost_dev_check_owner(struct vhost_dev *dev)
153 {
154 	/* Are you the owner? If not, I don't think you mean to do that */
155 	return dev->mm == current->mm ? 0 : -EPERM;
156 }
157 
158 /* Caller should have device mutex */
159 static long vhost_dev_set_owner(struct vhost_dev *dev)
160 {
161 	/* Is there an owner already? */
162 	if (dev->mm)
163 		return -EBUSY;
164 	/* No owner, become one */
165 	dev->mm = get_task_mm(current);
166 	return 0;
167 }
168 
169 /* Caller should have device mutex */
170 long vhost_dev_reset_owner(struct vhost_dev *dev)
171 {
172 	struct vhost_memory *memory;
173 
174 	/* Restore memory to default empty mapping. */
175 	memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
176 	if (!memory)
177 		return -ENOMEM;
178 
179 	vhost_dev_cleanup(dev);
180 
181 	memory->nregions = 0;
182 	dev->memory = memory;
183 	return 0;
184 }
185 
186 /* Caller should have device mutex */
187 void vhost_dev_cleanup(struct vhost_dev *dev)
188 {
189 	int i;
190 	for (i = 0; i < dev->nvqs; ++i) {
191 		if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
192 			vhost_poll_stop(&dev->vqs[i].poll);
193 			vhost_poll_flush(&dev->vqs[i].poll);
194 		}
195 		if (dev->vqs[i].error_ctx)
196 			eventfd_ctx_put(dev->vqs[i].error_ctx);
197 		if (dev->vqs[i].error)
198 			fput(dev->vqs[i].error);
199 		if (dev->vqs[i].kick)
200 			fput(dev->vqs[i].kick);
201 		if (dev->vqs[i].call_ctx)
202 			eventfd_ctx_put(dev->vqs[i].call_ctx);
203 		if (dev->vqs[i].call)
204 			fput(dev->vqs[i].call);
205 		vhost_vq_reset(dev, dev->vqs + i);
206 	}
207 	if (dev->log_ctx)
208 		eventfd_ctx_put(dev->log_ctx);
209 	dev->log_ctx = NULL;
210 	if (dev->log_file)
211 		fput(dev->log_file);
212 	dev->log_file = NULL;
213 	/* No one will access memory at this point */
214 	kfree(dev->memory);
215 	dev->memory = NULL;
216 	if (dev->mm)
217 		mmput(dev->mm);
218 	dev->mm = NULL;
219 }
220 
221 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
222 {
223 	u64 a = addr / VHOST_PAGE_SIZE / 8;
224 	/* Make sure 64 bit math will not overflow. */
225 	if (a > ULONG_MAX - (unsigned long)log_base ||
226 	    a + (unsigned long)log_base > ULONG_MAX)
227 		return -EFAULT;
228 
229 	return access_ok(VERIFY_WRITE, log_base + a,
230 			 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
231 }
232 
233 /* Caller should have vq mutex and device mutex. */
234 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
235 			       int log_all)
236 {
237 	int i;
238 	for (i = 0; i < mem->nregions; ++i) {
239 		struct vhost_memory_region *m = mem->regions + i;
240 		unsigned long a = m->userspace_addr;
241 		if (m->memory_size > ULONG_MAX)
242 			return 0;
243 		else if (!access_ok(VERIFY_WRITE, (void __user *)a,
244 				    m->memory_size))
245 			return 0;
246 		else if (log_all && !log_access_ok(log_base,
247 						   m->guest_phys_addr,
248 						   m->memory_size))
249 			return 0;
250 	}
251 	return 1;
252 }
253 
254 /* Can we switch to this memory table? */
255 /* Caller should have device mutex but not vq mutex */
256 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
257 			    int log_all)
258 {
259 	int i;
260 	for (i = 0; i < d->nvqs; ++i) {
261 		int ok;
262 		mutex_lock(&d->vqs[i].mutex);
263 		/* If ring is inactive, will check when it's enabled. */
264 		if (d->vqs[i].private_data)
265 			ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
266 						 log_all);
267 		else
268 			ok = 1;
269 		mutex_unlock(&d->vqs[i].mutex);
270 		if (!ok)
271 			return 0;
272 	}
273 	return 1;
274 }
275 
276 static int vq_access_ok(unsigned int num,
277 			struct vring_desc __user *desc,
278 			struct vring_avail __user *avail,
279 			struct vring_used __user *used)
280 {
281 	return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
282 	       access_ok(VERIFY_READ, avail,
283 			 sizeof *avail + num * sizeof *avail->ring) &&
284 	       access_ok(VERIFY_WRITE, used,
285 			sizeof *used + num * sizeof *used->ring);
286 }
287 
288 /* Can we log writes? */
289 /* Caller should have device mutex but not vq mutex */
290 int vhost_log_access_ok(struct vhost_dev *dev)
291 {
292 	return memory_access_ok(dev, dev->memory, 1);
293 }
294 
295 /* Verify access for write logging. */
296 /* Caller should have vq mutex and device mutex */
297 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base)
298 {
299 	return vq_memory_access_ok(log_base, vq->dev->memory,
300 			    vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
301 		(!vq->log_used || log_access_ok(log_base, vq->log_addr,
302 					sizeof *vq->used +
303 					vq->num * sizeof *vq->used->ring));
304 }
305 
306 /* Can we start vq? */
307 /* Caller should have vq mutex and device mutex */
308 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
309 {
310 	return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) &&
311 		vq_log_access_ok(vq, vq->log_base);
312 }
313 
314 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
315 {
316 	struct vhost_memory mem, *newmem, *oldmem;
317 	unsigned long size = offsetof(struct vhost_memory, regions);
318 	long r;
319 	r = copy_from_user(&mem, m, size);
320 	if (r)
321 		return r;
322 	if (mem.padding)
323 		return -EOPNOTSUPP;
324 	if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
325 		return -E2BIG;
326 	newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
327 	if (!newmem)
328 		return -ENOMEM;
329 
330 	memcpy(newmem, &mem, size);
331 	r = copy_from_user(newmem->regions, m->regions,
332 			   mem.nregions * sizeof *m->regions);
333 	if (r) {
334 		kfree(newmem);
335 		return r;
336 	}
337 
338 	if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL)))
339 		return -EFAULT;
340 	oldmem = d->memory;
341 	rcu_assign_pointer(d->memory, newmem);
342 	synchronize_rcu();
343 	kfree(oldmem);
344 	return 0;
345 }
346 
347 static int init_used(struct vhost_virtqueue *vq,
348 		     struct vring_used __user *used)
349 {
350 	int r = put_user(vq->used_flags, &used->flags);
351 	if (r)
352 		return r;
353 	return get_user(vq->last_used_idx, &used->idx);
354 }
355 
356 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
357 {
358 	struct file *eventfp, *filep = NULL,
359 		    *pollstart = NULL, *pollstop = NULL;
360 	struct eventfd_ctx *ctx = NULL;
361 	u32 __user *idxp = argp;
362 	struct vhost_virtqueue *vq;
363 	struct vhost_vring_state s;
364 	struct vhost_vring_file f;
365 	struct vhost_vring_addr a;
366 	u32 idx;
367 	long r;
368 
369 	r = get_user(idx, idxp);
370 	if (r < 0)
371 		return r;
372 	if (idx > d->nvqs)
373 		return -ENOBUFS;
374 
375 	vq = d->vqs + idx;
376 
377 	mutex_lock(&vq->mutex);
378 
379 	switch (ioctl) {
380 	case VHOST_SET_VRING_NUM:
381 		/* Resizing ring with an active backend?
382 		 * You don't want to do that. */
383 		if (vq->private_data) {
384 			r = -EBUSY;
385 			break;
386 		}
387 		r = copy_from_user(&s, argp, sizeof s);
388 		if (r < 0)
389 			break;
390 		if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
391 			r = -EINVAL;
392 			break;
393 		}
394 		vq->num = s.num;
395 		break;
396 	case VHOST_SET_VRING_BASE:
397 		/* Moving base with an active backend?
398 		 * You don't want to do that. */
399 		if (vq->private_data) {
400 			r = -EBUSY;
401 			break;
402 		}
403 		r = copy_from_user(&s, argp, sizeof s);
404 		if (r < 0)
405 			break;
406 		if (s.num > 0xffff) {
407 			r = -EINVAL;
408 			break;
409 		}
410 		vq->last_avail_idx = s.num;
411 		/* Forget the cached index value. */
412 		vq->avail_idx = vq->last_avail_idx;
413 		break;
414 	case VHOST_GET_VRING_BASE:
415 		s.index = idx;
416 		s.num = vq->last_avail_idx;
417 		r = copy_to_user(argp, &s, sizeof s);
418 		break;
419 	case VHOST_SET_VRING_ADDR:
420 		r = copy_from_user(&a, argp, sizeof a);
421 		if (r < 0)
422 			break;
423 		if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
424 			r = -EOPNOTSUPP;
425 			break;
426 		}
427 		/* For 32bit, verify that the top 32bits of the user
428 		   data are set to zero. */
429 		if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
430 		    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
431 		    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
432 			r = -EFAULT;
433 			break;
434 		}
435 		if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
436 		    (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
437 		    (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
438 			r = -EINVAL;
439 			break;
440 		}
441 
442 		/* We only verify access here if backend is configured.
443 		 * If it is not, we don't as size might not have been setup.
444 		 * We will verify when backend is configured. */
445 		if (vq->private_data) {
446 			if (!vq_access_ok(vq->num,
447 				(void __user *)(unsigned long)a.desc_user_addr,
448 				(void __user *)(unsigned long)a.avail_user_addr,
449 				(void __user *)(unsigned long)a.used_user_addr)) {
450 				r = -EINVAL;
451 				break;
452 			}
453 
454 			/* Also validate log access for used ring if enabled. */
455 			if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
456 			    !log_access_ok(vq->log_base, a.log_guest_addr,
457 					   sizeof *vq->used +
458 					   vq->num * sizeof *vq->used->ring)) {
459 				r = -EINVAL;
460 				break;
461 			}
462 		}
463 
464 		r = init_used(vq, (struct vring_used __user *)(unsigned long)
465 			      a.used_user_addr);
466 		if (r)
467 			break;
468 		vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
469 		vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
470 		vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
471 		vq->log_addr = a.log_guest_addr;
472 		vq->used = (void __user *)(unsigned long)a.used_user_addr;
473 		break;
474 	case VHOST_SET_VRING_KICK:
475 		r = copy_from_user(&f, argp, sizeof f);
476 		if (r < 0)
477 			break;
478 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
479 		if (IS_ERR(eventfp))
480 			return PTR_ERR(eventfp);
481 		if (eventfp != vq->kick) {
482 			pollstop = filep = vq->kick;
483 			pollstart = vq->kick = eventfp;
484 		} else
485 			filep = eventfp;
486 		break;
487 	case VHOST_SET_VRING_CALL:
488 		r = copy_from_user(&f, argp, sizeof f);
489 		if (r < 0)
490 			break;
491 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
492 		if (IS_ERR(eventfp))
493 			return PTR_ERR(eventfp);
494 		if (eventfp != vq->call) {
495 			filep = vq->call;
496 			ctx = vq->call_ctx;
497 			vq->call = eventfp;
498 			vq->call_ctx = eventfp ?
499 				eventfd_ctx_fileget(eventfp) : NULL;
500 		} else
501 			filep = eventfp;
502 		break;
503 	case VHOST_SET_VRING_ERR:
504 		r = copy_from_user(&f, argp, sizeof f);
505 		if (r < 0)
506 			break;
507 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
508 		if (IS_ERR(eventfp))
509 			return PTR_ERR(eventfp);
510 		if (eventfp != vq->error) {
511 			filep = vq->error;
512 			vq->error = eventfp;
513 			ctx = vq->error_ctx;
514 			vq->error_ctx = eventfp ?
515 				eventfd_ctx_fileget(eventfp) : NULL;
516 		} else
517 			filep = eventfp;
518 		break;
519 	default:
520 		r = -ENOIOCTLCMD;
521 	}
522 
523 	if (pollstop && vq->handle_kick)
524 		vhost_poll_stop(&vq->poll);
525 
526 	if (ctx)
527 		eventfd_ctx_put(ctx);
528 	if (filep)
529 		fput(filep);
530 
531 	if (pollstart && vq->handle_kick)
532 		vhost_poll_start(&vq->poll, vq->kick);
533 
534 	mutex_unlock(&vq->mutex);
535 
536 	if (pollstop && vq->handle_kick)
537 		vhost_poll_flush(&vq->poll);
538 	return r;
539 }
540 
541 /* Caller must have device mutex */
542 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
543 {
544 	void __user *argp = (void __user *)arg;
545 	struct file *eventfp, *filep = NULL;
546 	struct eventfd_ctx *ctx = NULL;
547 	u64 p;
548 	long r;
549 	int i, fd;
550 
551 	/* If you are not the owner, you can become one */
552 	if (ioctl == VHOST_SET_OWNER) {
553 		r = vhost_dev_set_owner(d);
554 		goto done;
555 	}
556 
557 	/* You must be the owner to do anything else */
558 	r = vhost_dev_check_owner(d);
559 	if (r)
560 		goto done;
561 
562 	switch (ioctl) {
563 	case VHOST_SET_MEM_TABLE:
564 		r = vhost_set_memory(d, argp);
565 		break;
566 	case VHOST_SET_LOG_BASE:
567 		r = copy_from_user(&p, argp, sizeof p);
568 		if (r < 0)
569 			break;
570 		if ((u64)(unsigned long)p != p) {
571 			r = -EFAULT;
572 			break;
573 		}
574 		for (i = 0; i < d->nvqs; ++i) {
575 			struct vhost_virtqueue *vq;
576 			void __user *base = (void __user *)(unsigned long)p;
577 			vq = d->vqs + i;
578 			mutex_lock(&vq->mutex);
579 			/* If ring is inactive, will check when it's enabled. */
580 			if (vq->private_data && !vq_log_access_ok(vq, base))
581 				r = -EFAULT;
582 			else
583 				vq->log_base = base;
584 			mutex_unlock(&vq->mutex);
585 		}
586 		break;
587 	case VHOST_SET_LOG_FD:
588 		r = get_user(fd, (int __user *)argp);
589 		if (r < 0)
590 			break;
591 		eventfp = fd == -1 ? NULL : eventfd_fget(fd);
592 		if (IS_ERR(eventfp)) {
593 			r = PTR_ERR(eventfp);
594 			break;
595 		}
596 		if (eventfp != d->log_file) {
597 			filep = d->log_file;
598 			ctx = d->log_ctx;
599 			d->log_ctx = eventfp ?
600 				eventfd_ctx_fileget(eventfp) : NULL;
601 		} else
602 			filep = eventfp;
603 		for (i = 0; i < d->nvqs; ++i) {
604 			mutex_lock(&d->vqs[i].mutex);
605 			d->vqs[i].log_ctx = d->log_ctx;
606 			mutex_unlock(&d->vqs[i].mutex);
607 		}
608 		if (ctx)
609 			eventfd_ctx_put(ctx);
610 		if (filep)
611 			fput(filep);
612 		break;
613 	default:
614 		r = vhost_set_vring(d, ioctl, argp);
615 		break;
616 	}
617 done:
618 	return r;
619 }
620 
621 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
622 						     __u64 addr, __u32 len)
623 {
624 	struct vhost_memory_region *reg;
625 	int i;
626 	/* linear search is not brilliant, but we really have on the order of 6
627 	 * regions in practice */
628 	for (i = 0; i < mem->nregions; ++i) {
629 		reg = mem->regions + i;
630 		if (reg->guest_phys_addr <= addr &&
631 		    reg->guest_phys_addr + reg->memory_size - 1 >= addr)
632 			return reg;
633 	}
634 	return NULL;
635 }
636 
637 /* TODO: This is really inefficient.  We need something like get_user()
638  * (instruction directly accesses the data, with an exception table entry
639  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
640  */
641 static int set_bit_to_user(int nr, void __user *addr)
642 {
643 	unsigned long log = (unsigned long)addr;
644 	struct page *page;
645 	void *base;
646 	int bit = nr + (log % PAGE_SIZE) * 8;
647 	int r;
648 	r = get_user_pages_fast(log, 1, 1, &page);
649 	if (r < 0)
650 		return r;
651 	BUG_ON(r != 1);
652 	base = kmap_atomic(page, KM_USER0);
653 	set_bit(bit, base);
654 	kunmap_atomic(base, KM_USER0);
655 	set_page_dirty_lock(page);
656 	put_page(page);
657 	return 0;
658 }
659 
660 static int log_write(void __user *log_base,
661 		     u64 write_address, u64 write_length)
662 {
663 	int r;
664 	if (!write_length)
665 		return 0;
666 	write_address /= VHOST_PAGE_SIZE;
667 	for (;;) {
668 		u64 base = (u64)(unsigned long)log_base;
669 		u64 log = base + write_address / 8;
670 		int bit = write_address % 8;
671 		if ((u64)(unsigned long)log != log)
672 			return -EFAULT;
673 		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
674 		if (r < 0)
675 			return r;
676 		if (write_length <= VHOST_PAGE_SIZE)
677 			break;
678 		write_length -= VHOST_PAGE_SIZE;
679 		write_address += VHOST_PAGE_SIZE;
680 	}
681 	return r;
682 }
683 
684 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
685 		    unsigned int log_num, u64 len)
686 {
687 	int i, r;
688 
689 	/* Make sure data written is seen before log. */
690 	smp_wmb();
691 	for (i = 0; i < log_num; ++i) {
692 		u64 l = min(log[i].len, len);
693 		r = log_write(vq->log_base, log[i].addr, l);
694 		if (r < 0)
695 			return r;
696 		len -= l;
697 		if (!len)
698 			return 0;
699 	}
700 	if (vq->log_ctx)
701 		eventfd_signal(vq->log_ctx, 1);
702 	/* Length written exceeds what we have stored. This is a bug. */
703 	BUG();
704 	return 0;
705 }
706 
707 int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
708 		   struct iovec iov[], int iov_size)
709 {
710 	const struct vhost_memory_region *reg;
711 	struct vhost_memory *mem;
712 	struct iovec *_iov;
713 	u64 s = 0;
714 	int ret = 0;
715 
716 	rcu_read_lock();
717 
718 	mem = rcu_dereference(dev->memory);
719 	while ((u64)len > s) {
720 		u64 size;
721 		if (ret >= iov_size) {
722 			ret = -ENOBUFS;
723 			break;
724 		}
725 		reg = find_region(mem, addr, len);
726 		if (!reg) {
727 			ret = -EFAULT;
728 			break;
729 		}
730 		_iov = iov + ret;
731 		size = reg->memory_size - addr + reg->guest_phys_addr;
732 		_iov->iov_len = min((u64)len, size);
733 		_iov->iov_base = (void *)(unsigned long)
734 			(reg->userspace_addr + addr - reg->guest_phys_addr);
735 		s += size;
736 		addr += size;
737 		++ret;
738 	}
739 
740 	rcu_read_unlock();
741 	return ret;
742 }
743 
744 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
745  * function returns the next descriptor in the chain,
746  * or -1U if we're at the end. */
747 static unsigned next_desc(struct vring_desc *desc)
748 {
749 	unsigned int next;
750 
751 	/* If this descriptor says it doesn't chain, we're done. */
752 	if (!(desc->flags & VRING_DESC_F_NEXT))
753 		return -1U;
754 
755 	/* Check they're not leading us off end of descriptors. */
756 	next = desc->next;
757 	/* Make sure compiler knows to grab that: we don't want it changing! */
758 	/* We will use the result as an index in an array, so most
759 	 * architectures only need a compiler barrier here. */
760 	read_barrier_depends();
761 
762 	return next;
763 }
764 
765 static unsigned get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
766 			     struct iovec iov[], unsigned int iov_size,
767 			     unsigned int *out_num, unsigned int *in_num,
768 			     struct vhost_log *log, unsigned int *log_num,
769 			     struct vring_desc *indirect)
770 {
771 	struct vring_desc desc;
772 	unsigned int i = 0, count, found = 0;
773 	int ret;
774 
775 	/* Sanity check */
776 	if (indirect->len % sizeof desc) {
777 		vq_err(vq, "Invalid length in indirect descriptor: "
778 		       "len 0x%llx not multiple of 0x%zx\n",
779 		       (unsigned long long)indirect->len,
780 		       sizeof desc);
781 		return -EINVAL;
782 	}
783 
784 	ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
785 			     ARRAY_SIZE(vq->indirect));
786 	if (ret < 0) {
787 		vq_err(vq, "Translation failure %d in indirect.\n", ret);
788 		return ret;
789 	}
790 
791 	/* We will use the result as an address to read from, so most
792 	 * architectures only need a compiler barrier here. */
793 	read_barrier_depends();
794 
795 	count = indirect->len / sizeof desc;
796 	/* Buffers are chained via a 16 bit next field, so
797 	 * we can have at most 2^16 of these. */
798 	if (count > USHORT_MAX + 1) {
799 		vq_err(vq, "Indirect buffer length too big: %d\n",
800 		       indirect->len);
801 		return -E2BIG;
802 	}
803 
804 	do {
805 		unsigned iov_count = *in_num + *out_num;
806 		if (++found > count) {
807 			vq_err(vq, "Loop detected: last one at %u "
808 			       "indirect size %u\n",
809 			       i, count);
810 			return -EINVAL;
811 		}
812 		if (memcpy_fromiovec((unsigned char *)&desc, vq->indirect,
813 				     sizeof desc)) {
814 			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
815 			       i, (size_t)indirect->addr + i * sizeof desc);
816 			return -EINVAL;
817 		}
818 		if (desc.flags & VRING_DESC_F_INDIRECT) {
819 			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
820 			       i, (size_t)indirect->addr + i * sizeof desc);
821 			return -EINVAL;
822 		}
823 
824 		ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
825 				     iov_size - iov_count);
826 		if (ret < 0) {
827 			vq_err(vq, "Translation failure %d indirect idx %d\n",
828 			       ret, i);
829 			return ret;
830 		}
831 		/* If this is an input descriptor, increment that count. */
832 		if (desc.flags & VRING_DESC_F_WRITE) {
833 			*in_num += ret;
834 			if (unlikely(log)) {
835 				log[*log_num].addr = desc.addr;
836 				log[*log_num].len = desc.len;
837 				++*log_num;
838 			}
839 		} else {
840 			/* If it's an output descriptor, they're all supposed
841 			 * to come before any input descriptors. */
842 			if (*in_num) {
843 				vq_err(vq, "Indirect descriptor "
844 				       "has out after in: idx %d\n", i);
845 				return -EINVAL;
846 			}
847 			*out_num += ret;
848 		}
849 	} while ((i = next_desc(&desc)) != -1);
850 	return 0;
851 }
852 
853 /* This looks in the virtqueue and for the first available buffer, and converts
854  * it to an iovec for convenient access.  Since descriptors consist of some
855  * number of output then some number of input descriptors, it's actually two
856  * iovecs, but we pack them into one and note how many of each there were.
857  *
858  * This function returns the descriptor number found, or vq->num (which
859  * is never a valid descriptor number) if none was found. */
860 unsigned vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
861 			   struct iovec iov[], unsigned int iov_size,
862 			   unsigned int *out_num, unsigned int *in_num,
863 			   struct vhost_log *log, unsigned int *log_num)
864 {
865 	struct vring_desc desc;
866 	unsigned int i, head, found = 0;
867 	u16 last_avail_idx;
868 	int ret;
869 
870 	/* Check it isn't doing very strange things with descriptor numbers. */
871 	last_avail_idx = vq->last_avail_idx;
872 	if (get_user(vq->avail_idx, &vq->avail->idx)) {
873 		vq_err(vq, "Failed to access avail idx at %p\n",
874 		       &vq->avail->idx);
875 		return vq->num;
876 	}
877 
878 	if ((u16)(vq->avail_idx - last_avail_idx) > vq->num) {
879 		vq_err(vq, "Guest moved used index from %u to %u",
880 		       last_avail_idx, vq->avail_idx);
881 		return vq->num;
882 	}
883 
884 	/* If there's nothing new since last we looked, return invalid. */
885 	if (vq->avail_idx == last_avail_idx)
886 		return vq->num;
887 
888 	/* Only get avail ring entries after they have been exposed by guest. */
889 	smp_rmb();
890 
891 	/* Grab the next descriptor number they're advertising, and increment
892 	 * the index we've seen. */
893 	if (get_user(head, &vq->avail->ring[last_avail_idx % vq->num])) {
894 		vq_err(vq, "Failed to read head: idx %d address %p\n",
895 		       last_avail_idx,
896 		       &vq->avail->ring[last_avail_idx % vq->num]);
897 		return vq->num;
898 	}
899 
900 	/* If their number is silly, that's an error. */
901 	if (head >= vq->num) {
902 		vq_err(vq, "Guest says index %u > %u is available",
903 		       head, vq->num);
904 		return vq->num;
905 	}
906 
907 	/* When we start there are none of either input nor output. */
908 	*out_num = *in_num = 0;
909 	if (unlikely(log))
910 		*log_num = 0;
911 
912 	i = head;
913 	do {
914 		unsigned iov_count = *in_num + *out_num;
915 		if (i >= vq->num) {
916 			vq_err(vq, "Desc index is %u > %u, head = %u",
917 			       i, vq->num, head);
918 			return vq->num;
919 		}
920 		if (++found > vq->num) {
921 			vq_err(vq, "Loop detected: last one at %u "
922 			       "vq size %u head %u\n",
923 			       i, vq->num, head);
924 			return vq->num;
925 		}
926 		ret = copy_from_user(&desc, vq->desc + i, sizeof desc);
927 		if (ret) {
928 			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
929 			       i, vq->desc + i);
930 			return vq->num;
931 		}
932 		if (desc.flags & VRING_DESC_F_INDIRECT) {
933 			ret = get_indirect(dev, vq, iov, iov_size,
934 					   out_num, in_num,
935 					   log, log_num, &desc);
936 			if (ret < 0) {
937 				vq_err(vq, "Failure detected "
938 				       "in indirect descriptor at idx %d\n", i);
939 				return vq->num;
940 			}
941 			continue;
942 		}
943 
944 		ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
945 				     iov_size - iov_count);
946 		if (ret < 0) {
947 			vq_err(vq, "Translation failure %d descriptor idx %d\n",
948 			       ret, i);
949 			return vq->num;
950 		}
951 		if (desc.flags & VRING_DESC_F_WRITE) {
952 			/* If this is an input descriptor,
953 			 * increment that count. */
954 			*in_num += ret;
955 			if (unlikely(log)) {
956 				log[*log_num].addr = desc.addr;
957 				log[*log_num].len = desc.len;
958 				++*log_num;
959 			}
960 		} else {
961 			/* If it's an output descriptor, they're all supposed
962 			 * to come before any input descriptors. */
963 			if (*in_num) {
964 				vq_err(vq, "Descriptor has out after in: "
965 				       "idx %d\n", i);
966 				return vq->num;
967 			}
968 			*out_num += ret;
969 		}
970 	} while ((i = next_desc(&desc)) != -1);
971 
972 	/* On success, increment avail index. */
973 	vq->last_avail_idx++;
974 	return head;
975 }
976 
977 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
978 void vhost_discard_vq_desc(struct vhost_virtqueue *vq)
979 {
980 	vq->last_avail_idx--;
981 }
982 
983 /* After we've used one of their buffers, we tell them about it.  We'll then
984  * want to notify the guest, using eventfd. */
985 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
986 {
987 	struct vring_used_elem *used;
988 
989 	/* The virtqueue contains a ring of used buffers.  Get a pointer to the
990 	 * next entry in that used ring. */
991 	used = &vq->used->ring[vq->last_used_idx % vq->num];
992 	if (put_user(head, &used->id)) {
993 		vq_err(vq, "Failed to write used id");
994 		return -EFAULT;
995 	}
996 	if (put_user(len, &used->len)) {
997 		vq_err(vq, "Failed to write used len");
998 		return -EFAULT;
999 	}
1000 	/* Make sure buffer is written before we update index. */
1001 	smp_wmb();
1002 	if (put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1003 		vq_err(vq, "Failed to increment used idx");
1004 		return -EFAULT;
1005 	}
1006 	if (unlikely(vq->log_used)) {
1007 		/* Make sure data is seen before log. */
1008 		smp_wmb();
1009 		/* Log used ring entry write. */
1010 		log_write(vq->log_base,
1011 			  vq->log_addr + ((void *)used - (void *)vq->used),
1012 			  sizeof *used);
1013 		/* Log used index update. */
1014 		log_write(vq->log_base,
1015 			  vq->log_addr + offsetof(struct vring_used, idx),
1016 			  sizeof vq->used->idx);
1017 		if (vq->log_ctx)
1018 			eventfd_signal(vq->log_ctx, 1);
1019 	}
1020 	vq->last_used_idx++;
1021 	return 0;
1022 }
1023 
1024 /* This actually signals the guest, using eventfd. */
1025 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1026 {
1027 	__u16 flags = 0;
1028 	if (get_user(flags, &vq->avail->flags)) {
1029 		vq_err(vq, "Failed to get flags");
1030 		return;
1031 	}
1032 
1033 	/* If they don't want an interrupt, don't signal, unless empty. */
1034 	if ((flags & VRING_AVAIL_F_NO_INTERRUPT) &&
1035 	    (vq->avail_idx != vq->last_avail_idx ||
1036 	     !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY)))
1037 		return;
1038 
1039 	/* Signal the Guest tell them we used something up. */
1040 	if (vq->call_ctx)
1041 		eventfd_signal(vq->call_ctx, 1);
1042 }
1043 
1044 /* And here's the combo meal deal.  Supersize me! */
1045 void vhost_add_used_and_signal(struct vhost_dev *dev,
1046 			       struct vhost_virtqueue *vq,
1047 			       unsigned int head, int len)
1048 {
1049 	vhost_add_used(vq, head, len);
1050 	vhost_signal(dev, vq);
1051 }
1052 
1053 /* OK, now we need to know about added descriptors. */
1054 bool vhost_enable_notify(struct vhost_virtqueue *vq)
1055 {
1056 	u16 avail_idx;
1057 	int r;
1058 	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1059 		return false;
1060 	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1061 	r = put_user(vq->used_flags, &vq->used->flags);
1062 	if (r) {
1063 		vq_err(vq, "Failed to enable notification at %p: %d\n",
1064 		       &vq->used->flags, r);
1065 		return false;
1066 	}
1067 	/* They could have slipped one in as we were doing that: make
1068 	 * sure it's written, then check again. */
1069 	smp_mb();
1070 	r = get_user(avail_idx, &vq->avail->idx);
1071 	if (r) {
1072 		vq_err(vq, "Failed to check avail idx at %p: %d\n",
1073 		       &vq->avail->idx, r);
1074 		return false;
1075 	}
1076 
1077 	return avail_idx != vq->last_avail_idx;
1078 }
1079 
1080 /* We don't need to be notified again. */
1081 void vhost_disable_notify(struct vhost_virtqueue *vq)
1082 {
1083 	int r;
1084 	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1085 		return;
1086 	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1087 	r = put_user(vq->used_flags, &vq->used->flags);
1088 	if (r)
1089 		vq_err(vq, "Failed to enable notification at %p: %d\n",
1090 		       &vq->used->flags, r);
1091 }
1092 
1093 int vhost_init(void)
1094 {
1095 	vhost_workqueue = create_singlethread_workqueue("vhost");
1096 	if (!vhost_workqueue)
1097 		return -ENOMEM;
1098 	return 0;
1099 }
1100 
1101 void vhost_cleanup(void)
1102 {
1103 	destroy_workqueue(vhost_workqueue);
1104 }
1105