projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
vhost: put mm after thread stop
[deliverable/linux.git] / drivers / vhost / vhost.c
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/rcupdate.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/kthread.h>
26 #include <linux/cgroup.h>
27
28 #include <linux/net.h>
29 #include <linux/if_packet.h>
30 #include <linux/if_arp.h>
31
32 #include <net/sock.h>
33
34 #include "vhost.h"
35
36 enum {
37 VHOST_MEMORY_MAX_NREGIONS = 64,
38 VHOST_MEMORY_F_LOG = 0x1,
39 };
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 = container_of(wait, struct vhost_poll, wait);
55
56 if (!((unsigned long)key & poll->mask))
57 return 0;
58
59 vhost_poll_queue(poll);
60 return 0;
61 }
62
63 static void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
64 {
65 INIT_LIST_HEAD(&work->node);
66 work->fn = fn;
67 init_waitqueue_head(&work->done);
68 work->flushing = 0;
69 work->queue_seq = work->done_seq = 0;
70 }
71
72 /* Init poll structure */
73 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
74 unsigned long mask, struct vhost_dev *dev)
75 {
76 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
77 init_poll_funcptr(&poll->table, vhost_poll_func);
78 poll->mask = mask;
79 poll->dev = dev;
80
81 vhost_work_init(&poll->work, fn);
82 }
83
84 /* Start polling a file. We add ourselves to file's wait queue. The caller must
85 * keep a reference to a file until after vhost_poll_stop is called. */
86 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
87 {
88 unsigned long mask;
89 mask = file->f_op->poll(file, &poll->table);
90 if (mask)
91 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
92 }
93
94 /* Stop polling a file. After this function returns, it becomes safe to drop the
95 * file reference. You must also flush afterwards. */
96 void vhost_poll_stop(struct vhost_poll *poll)
97 {
98 remove_wait_queue(poll->wqh, &poll->wait);
99 }
100
101 static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
102 {
103 unsigned seq;
104 int left;
105 int flushing;
106
107 spin_lock_irq(&dev->work_lock);
108 seq = work->queue_seq;
109 work->flushing++;
110 spin_unlock_irq(&dev->work_lock);
111 wait_event(work->done, ({
112 spin_lock_irq(&dev->work_lock);
113 left = seq - work->done_seq <= 0;
114 spin_unlock_irq(&dev->work_lock);
115 left;
116 }));
117 spin_lock_irq(&dev->work_lock);
118 flushing = --work->flushing;
119 spin_unlock_irq(&dev->work_lock);
120 BUG_ON(flushing < 0);
121 }
122
123 /* Flush any work that has been scheduled. When calling this, don't hold any
124 * locks that are also used by the callback. */
125 void vhost_poll_flush(struct vhost_poll *poll)
126 {
127 vhost_work_flush(poll->dev, &poll->work);
128 }
129
130 static inline void vhost_work_queue(struct vhost_dev *dev,
131 struct vhost_work *work)
132 {
133 unsigned long flags;
134
135 spin_lock_irqsave(&dev->work_lock, flags);
136 if (list_empty(&work->node)) {
137 list_add_tail(&work->node, &dev->work_list);
138 work->queue_seq++;
139 wake_up_process(dev->worker);
140 }
141 spin_unlock_irqrestore(&dev->work_lock, flags);
142 }
143
144 void vhost_poll_queue(struct vhost_poll *poll)
145 {
146 vhost_work_queue(poll->dev, &poll->work);
147 }
148
149 static void vhost_vq_reset(struct vhost_dev *dev,
150 struct vhost_virtqueue *vq)
151 {
152 vq->num = 1;
153 vq->desc = NULL;
154 vq->avail = NULL;
155 vq->used = NULL;
156 vq->last_avail_idx = 0;
157 vq->avail_idx = 0;
158 vq->last_used_idx = 0;
159 vq->used_flags = 0;
160 vq->log_used = false;
161 vq->log_addr = -1ull;
162 vq->vhost_hlen = 0;
163 vq->sock_hlen = 0;
164 vq->private_data = NULL;
165 vq->log_base = NULL;
166 vq->error_ctx = NULL;
167 vq->error = NULL;
168 vq->kick = NULL;
169 vq->call_ctx = NULL;
170 vq->call = NULL;
171 vq->log_ctx = NULL;
172 }
173
174 static int vhost_worker(void *data)
175 {
176 struct vhost_dev *dev = data;
177 struct vhost_work *work = NULL;
178 unsigned uninitialized_var(seq);
179
180 for (;;) {
181 /* mb paired w/ kthread_stop */
182 set_current_state(TASK_INTERRUPTIBLE);
183
184 spin_lock_irq(&dev->work_lock);
185 if (work) {
186 work->done_seq = seq;
187 if (work->flushing)
188 wake_up_all(&work->done);
189 }
190
191 if (kthread_should_stop()) {
192 spin_unlock_irq(&dev->work_lock);
193 __set_current_state(TASK_RUNNING);
194 return 0;
195 }
196 if (!list_empty(&dev->work_list)) {
197 work = list_first_entry(&dev->work_list,
198 struct vhost_work, node);
199 list_del_init(&work->node);
200 seq = work->queue_seq;
201 } else
202 work = NULL;
203 spin_unlock_irq(&dev->work_lock);
204
205 if (work) {
206 __set_current_state(TASK_RUNNING);
207 work->fn(work);
208 } else
209 schedule();
210
211 }
212 }
213
214 /* Helper to allocate iovec buffers for all vqs. */
215 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
216 {
217 int i;
218 for (i = 0; i < dev->nvqs; ++i) {
219 dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect *
220 UIO_MAXIOV, GFP_KERNEL);
221 dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV,
222 GFP_KERNEL);
223 dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads *
224 UIO_MAXIOV, GFP_KERNEL);
225
226 if (!dev->vqs[i].indirect || !dev->vqs[i].log ||
227 !dev->vqs[i].heads)
228 goto err_nomem;
229 }
230 return 0;
231 err_nomem:
232 for (; i >= 0; --i) {
233 kfree(dev->vqs[i].indirect);
234 kfree(dev->vqs[i].log);
235 kfree(dev->vqs[i].heads);
236 }
237 return -ENOMEM;
238 }
239
240 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
241 {
242 int i;
243 for (i = 0; i < dev->nvqs; ++i) {
244 kfree(dev->vqs[i].indirect);
245 dev->vqs[i].indirect = NULL;
246 kfree(dev->vqs[i].log);
247 dev->vqs[i].log = NULL;
248 kfree(dev->vqs[i].heads);
249 dev->vqs[i].heads = NULL;
250 }
251 }
252
253 long vhost_dev_init(struct vhost_dev *dev,
254 struct vhost_virtqueue *vqs, int nvqs)
255 {
256 int i;
257
258 dev->vqs = vqs;
259 dev->nvqs = nvqs;
260 mutex_init(&dev->mutex);
261 dev->log_ctx = NULL;
262 dev->log_file = NULL;
263 dev->memory = NULL;
264 dev->mm = NULL;
265 spin_lock_init(&dev->work_lock);
266 INIT_LIST_HEAD(&dev->work_list);
267 dev->worker = NULL;
268
269 for (i = 0; i < dev->nvqs; ++i) {
270 dev->vqs[i].log = NULL;
271 dev->vqs[i].indirect = NULL;
272 dev->vqs[i].heads = NULL;
273 dev->vqs[i].dev = dev;
274 mutex_init(&dev->vqs[i].mutex);
275 vhost_vq_reset(dev, dev->vqs + i);
276 if (dev->vqs[i].handle_kick)
277 vhost_poll_init(&dev->vqs[i].poll,
278 dev->vqs[i].handle_kick, POLLIN, dev);
279 }
280
281 return 0;
282 }
283
284 /* Caller should have device mutex */
285 long vhost_dev_check_owner(struct vhost_dev *dev)
286 {
287 /* Are you the owner? If not, I don't think you mean to do that */
288 return dev->mm == current->mm ? 0 : -EPERM;
289 }
290
291 struct vhost_attach_cgroups_struct {
292 struct vhost_work work;
293 struct task_struct *owner;
294 int ret;
295 };
296
297 static void vhost_attach_cgroups_work(struct vhost_work *work)
298 {
299 struct vhost_attach_cgroups_struct *s;
300 s = container_of(work, struct vhost_attach_cgroups_struct, work);
301 s->ret = cgroup_attach_task_all(s->owner, current);
302 }
303
304 static int vhost_attach_cgroups(struct vhost_dev *dev)
305 {
306 struct vhost_attach_cgroups_struct attach;
307 attach.owner = current;
308 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
309 vhost_work_queue(dev, &attach.work);
310 vhost_work_flush(dev, &attach.work);
311 return attach.ret;
312 }
313
314 /* Caller should have device mutex */
315 static long vhost_dev_set_owner(struct vhost_dev *dev)
316 {
317 struct task_struct *worker;
318 int err;
319 /* Is there an owner already? */
320 if (dev->mm) {
321 err = -EBUSY;
322 goto err_mm;
323 }
324 /* No owner, become one */
325 dev->mm = get_task_mm(current);
326 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
327 if (IS_ERR(worker)) {
328 err = PTR_ERR(worker);
329 goto err_worker;
330 }
331
332 dev->worker = worker;
333 wake_up_process(worker); /* avoid contributing to loadavg */
334
335 err = vhost_attach_cgroups(dev);
336 if (err)
337 goto err_cgroup;
338
339 err = vhost_dev_alloc_iovecs(dev);
340 if (err)
341 goto err_cgroup;
342
343 return 0;
344 err_cgroup:
345 kthread_stop(worker);
346 dev->worker = NULL;
347 err_worker:
348 if (dev->mm)
349 mmput(dev->mm);
350 dev->mm = NULL;
351 err_mm:
352 return err;
353 }
354
355 /* Caller should have device mutex */
356 long vhost_dev_reset_owner(struct vhost_dev *dev)
357 {
358 struct vhost_memory *memory;
359
360 /* Restore memory to default empty mapping. */
361 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
362 if (!memory)
363 return -ENOMEM;
364
365 vhost_dev_cleanup(dev);
366
367 memory->nregions = 0;
368 RCU_INIT_POINTER(dev->memory, memory);
369 return 0;
370 }
371
372 /* Caller should have device mutex */
373 void vhost_dev_cleanup(struct vhost_dev *dev)
374 {
375 int i;
376 for (i = 0; i < dev->nvqs; ++i) {
377 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
378 vhost_poll_stop(&dev->vqs[i].poll);
379 vhost_poll_flush(&dev->vqs[i].poll);
380 }
381 if (dev->vqs[i].error_ctx)
382 eventfd_ctx_put(dev->vqs[i].error_ctx);
383 if (dev->vqs[i].error)
384 fput(dev->vqs[i].error);
385 if (dev->vqs[i].kick)
386 fput(dev->vqs[i].kick);
387 if (dev->vqs[i].call_ctx)
388 eventfd_ctx_put(dev->vqs[i].call_ctx);
389 if (dev->vqs[i].call)
390 fput(dev->vqs[i].call);
391 vhost_vq_reset(dev, dev->vqs + i);
392 }
393 vhost_dev_free_iovecs(dev);
394 if (dev->log_ctx)
395 eventfd_ctx_put(dev->log_ctx);
396 dev->log_ctx = NULL;
397 if (dev->log_file)
398 fput(dev->log_file);
399 dev->log_file = NULL;
400 /* No one will access memory at this point */
401 kfree(rcu_dereference_protected(dev->memory,
402 lockdep_is_held(&dev->mutex)));
403 RCU_INIT_POINTER(dev->memory, NULL);
404 WARN_ON(!list_empty(&dev->work_list));
405 if (dev->worker) {
406 kthread_stop(dev->worker);
407 dev->worker = NULL;
408 }
409 if (dev->mm)
410 mmput(dev->mm);
411 dev->mm = NULL;
412 }
413
414 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
415 {
416 u64 a = addr / VHOST_PAGE_SIZE / 8;
417 /* Make sure 64 bit math will not overflow. */
418 if (a > ULONG_MAX - (unsigned long)log_base ||
419 a + (unsigned long)log_base > ULONG_MAX)
420 return 0;
421
422 return access_ok(VERIFY_WRITE, log_base + a,
423 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
424 }
425
426 /* Caller should have vq mutex and device mutex. */
427 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
428 int log_all)
429 {
430 int i;
431
432 if (!mem)
433 return 0;
434
435 for (i = 0; i < mem->nregions; ++i) {
436 struct vhost_memory_region *m = mem->regions + i;
437 unsigned long a = m->userspace_addr;
438 if (m->memory_size > ULONG_MAX)
439 return 0;
440 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
441 m->memory_size))
442 return 0;
443 else if (log_all && !log_access_ok(log_base,
444 m->guest_phys_addr,
445 m->memory_size))
446 return 0;
447 }
448 return 1;
449 }
450
451 /* Can we switch to this memory table? */
452 /* Caller should have device mutex but not vq mutex */
453 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
454 int log_all)
455 {
456 int i;
457 for (i = 0; i < d->nvqs; ++i) {
458 int ok;
459 mutex_lock(&d->vqs[i].mutex);
460 /* If ring is inactive, will check when it's enabled. */
461 if (d->vqs[i].private_data)
462 ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
463 log_all);
464 else
465 ok = 1;
466 mutex_unlock(&d->vqs[i].mutex);
467 if (!ok)
468 return 0;
469 }
470 return 1;
471 }
472
473 static int vq_access_ok(unsigned int num,
474 struct vring_desc __user *desc,
475 struct vring_avail __user *avail,
476 struct vring_used __user *used)
477 {
478 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
479 access_ok(VERIFY_READ, avail,
480 sizeof *avail + num * sizeof *avail->ring) &&
481 access_ok(VERIFY_WRITE, used,
482 sizeof *used + num * sizeof *used->ring);
483 }
484
485 /* Can we log writes? */
486 /* Caller should have device mutex but not vq mutex */
487 int vhost_log_access_ok(struct vhost_dev *dev)
488 {
489 struct vhost_memory *mp;
490
491 mp = rcu_dereference_protected(dev->memory,
492 lockdep_is_held(&dev->mutex));
493 return memory_access_ok(dev, mp, 1);
494 }
495
496 /* Verify access for write logging. */
497 /* Caller should have vq mutex and device mutex */
498 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base)
499 {
500 struct vhost_memory *mp;
501
502 mp = rcu_dereference_protected(vq->dev->memory,
503 lockdep_is_held(&vq->mutex));
504 return vq_memory_access_ok(log_base, mp,
505 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
506 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
507 sizeof *vq->used +
508 vq->num * sizeof *vq->used->ring));
509 }
510
511 /* Can we start vq? */
512 /* Caller should have vq mutex and device mutex */
513 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
514 {
515 return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) &&
516 vq_log_access_ok(vq, vq->log_base);
517 }
518
519 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
520 {
521 struct vhost_memory mem, *newmem, *oldmem;
522 unsigned long size = offsetof(struct vhost_memory, regions);
523 if (copy_from_user(&mem, m, size))
524 return -EFAULT;
525 if (mem.padding)
526 return -EOPNOTSUPP;
527 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
528 return -E2BIG;
529 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
530 if (!newmem)
531 return -ENOMEM;
532
533 memcpy(newmem, &mem, size);
534 if (copy_from_user(newmem->regions, m->regions,
535 mem.nregions * sizeof *m->regions)) {
536 kfree(newmem);
537 return -EFAULT;
538 }
539
540 if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL))) {
541 kfree(newmem);
542 return -EFAULT;
543 }
544 oldmem = rcu_dereference_protected(d->memory,
545 lockdep_is_held(&d->mutex));
546 rcu_assign_pointer(d->memory, newmem);
547 synchronize_rcu();
548 kfree(oldmem);
549 return 0;
550 }
551
552 static int init_used(struct vhost_virtqueue *vq,
553 struct vring_used __user *used)
554 {
555 int r = put_user(vq->used_flags, &used->flags);
556 if (r)
557 return r;
558 return get_user(vq->last_used_idx, &used->idx);
559 }
560
561 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
562 {
563 struct file *eventfp, *filep = NULL,
564 *pollstart = NULL, *pollstop = NULL;
565 struct eventfd_ctx *ctx = NULL;
566 u32 __user *idxp = argp;
567 struct vhost_virtqueue *vq;
568 struct vhost_vring_state s;
569 struct vhost_vring_file f;
570 struct vhost_vring_addr a;
571 u32 idx;
572 long r;
573
574 r = get_user(idx, idxp);
575 if (r < 0)
576 return r;
577 if (idx >= d->nvqs)
578 return -ENOBUFS;
579
580 vq = d->vqs + idx;
581
582 mutex_lock(&vq->mutex);
583
584 switch (ioctl) {
585 case VHOST_SET_VRING_NUM:
586 /* Resizing ring with an active backend?
587 * You don't want to do that. */
588 if (vq->private_data) {
589 r = -EBUSY;
590 break;
591 }
592 if (copy_from_user(&s, argp, sizeof s)) {
593 r = -EFAULT;
594 break;
595 }
596 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
597 r = -EINVAL;
598 break;
599 }
600 vq->num = s.num;
601 break;
602 case VHOST_SET_VRING_BASE:
603 /* Moving base with an active backend?
604 * You don't want to do that. */
605 if (vq->private_data) {
606 r = -EBUSY;
607 break;
608 }
609 if (copy_from_user(&s, argp, sizeof s)) {
610 r = -EFAULT;
611 break;
612 }
613 if (s.num > 0xffff) {
614 r = -EINVAL;
615 break;
616 }
617 vq->last_avail_idx = s.num;
618 /* Forget the cached index value. */
619 vq->avail_idx = vq->last_avail_idx;
620 break;
621 case VHOST_GET_VRING_BASE:
622 s.index = idx;
623 s.num = vq->last_avail_idx;
624 if (copy_to_user(argp, &s, sizeof s))
625 r = -EFAULT;
626 break;
627 case VHOST_SET_VRING_ADDR:
628 if (copy_from_user(&a, argp, sizeof a)) {
629 r = -EFAULT;
630 break;
631 }
632 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
633 r = -EOPNOTSUPP;
634 break;
635 }
636 /* For 32bit, verify that the top 32bits of the user
637 data are set to zero. */
638 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
639 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
640 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
641 r = -EFAULT;
642 break;
643 }
644 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
645 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
646 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
647 r = -EINVAL;
648 break;
649 }
650
651 /* We only verify access here if backend is configured.
652 * If it is not, we don't as size might not have been setup.
653 * We will verify when backend is configured. */
654 if (vq->private_data) {
655 if (!vq_access_ok(vq->num,
656 (void __user *)(unsigned long)a.desc_user_addr,
657 (void __user *)(unsigned long)a.avail_user_addr,
658 (void __user *)(unsigned long)a.used_user_addr)) {
659 r = -EINVAL;
660 break;
661 }
662
663 /* Also validate log access for used ring if enabled. */
664 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
665 !log_access_ok(vq->log_base, a.log_guest_addr,
666 sizeof *vq->used +
667 vq->num * sizeof *vq->used->ring)) {
668 r = -EINVAL;
669 break;
670 }
671 }
672
673 r = init_used(vq, (struct vring_used __user *)(unsigned long)
674 a.used_user_addr);
675 if (r)
676 break;
677 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
678 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
679 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
680 vq->log_addr = a.log_guest_addr;
681 vq->used = (void __user *)(unsigned long)a.used_user_addr;
682 break;
683 case VHOST_SET_VRING_KICK:
684 if (copy_from_user(&f, argp, sizeof f)) {
685 r = -EFAULT;
686 break;
687 }
688 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
689 if (IS_ERR(eventfp)) {
690 r = PTR_ERR(eventfp);
691 break;
692 }
693 if (eventfp != vq->kick) {
694 pollstop = filep = vq->kick;
695 pollstart = vq->kick = eventfp;
696 } else
697 filep = eventfp;
698 break;
699 case VHOST_SET_VRING_CALL:
700 if (copy_from_user(&f, argp, sizeof f)) {
701 r = -EFAULT;
702 break;
703 }
704 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
705 if (IS_ERR(eventfp)) {
706 r = PTR_ERR(eventfp);
707 break;
708 }
709 if (eventfp != vq->call) {
710 filep = vq->call;
711 ctx = vq->call_ctx;
712 vq->call = eventfp;
713 vq->call_ctx = eventfp ?
714 eventfd_ctx_fileget(eventfp) : NULL;
715 } else
716 filep = eventfp;
717 break;
718 case VHOST_SET_VRING_ERR:
719 if (copy_from_user(&f, argp, sizeof f)) {
720 r = -EFAULT;
721 break;
722 }
723 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
724 if (IS_ERR(eventfp)) {
725 r = PTR_ERR(eventfp);
726 break;
727 }
728 if (eventfp != vq->error) {
729 filep = vq->error;
730 vq->error = eventfp;
731 ctx = vq->error_ctx;
732 vq->error_ctx = eventfp ?
733 eventfd_ctx_fileget(eventfp) : NULL;
734 } else
735 filep = eventfp;
736 break;
737 default:
738 r = -ENOIOCTLCMD;
739 }
740
741 if (pollstop && vq->handle_kick)
742 vhost_poll_stop(&vq->poll);
743
744 if (ctx)
745 eventfd_ctx_put(ctx);
746 if (filep)
747 fput(filep);
748
749 if (pollstart && vq->handle_kick)
750 vhost_poll_start(&vq->poll, vq->kick);
751
752 mutex_unlock(&vq->mutex);
753
754 if (pollstop && vq->handle_kick)
755 vhost_poll_flush(&vq->poll);
756 return r;
757 }
758
759 /* Caller must have device mutex */
760 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
761 {
762 void __user *argp = (void __user *)arg;
763 struct file *eventfp, *filep = NULL;
764 struct eventfd_ctx *ctx = NULL;
765 u64 p;
766 long r;
767 int i, fd;
768
769 /* If you are not the owner, you can become one */
770 if (ioctl == VHOST_SET_OWNER) {
771 r = vhost_dev_set_owner(d);
772 goto done;
773 }
774
775 /* You must be the owner to do anything else */
776 r = vhost_dev_check_owner(d);
777 if (r)
778 goto done;
779
780 switch (ioctl) {
781 case VHOST_SET_MEM_TABLE:
782 r = vhost_set_memory(d, argp);
783 break;
784 case VHOST_SET_LOG_BASE:
785 if (copy_from_user(&p, argp, sizeof p)) {
786 r = -EFAULT;
787 break;
788 }
789 if ((u64)(unsigned long)p != p) {
790 r = -EFAULT;
791 break;
792 }
793 for (i = 0; i < d->nvqs; ++i) {
794 struct vhost_virtqueue *vq;
795 void __user *base = (void __user *)(unsigned long)p;
796 vq = d->vqs + i;
797 mutex_lock(&vq->mutex);
798 /* If ring is inactive, will check when it's enabled. */
799 if (vq->private_data && !vq_log_access_ok(vq, base))
800 r = -EFAULT;
801 else
802 vq->log_base = base;
803 mutex_unlock(&vq->mutex);
804 }
805 break;
806 case VHOST_SET_LOG_FD:
807 r = get_user(fd, (int __user *)argp);
808 if (r < 0)
809 break;
810 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
811 if (IS_ERR(eventfp)) {
812 r = PTR_ERR(eventfp);
813 break;
814 }
815 if (eventfp != d->log_file) {
816 filep = d->log_file;
817 ctx = d->log_ctx;
818 d->log_ctx = eventfp ?
819 eventfd_ctx_fileget(eventfp) : NULL;
820 } else
821 filep = eventfp;
822 for (i = 0; i < d->nvqs; ++i) {
823 mutex_lock(&d->vqs[i].mutex);
824 d->vqs[i].log_ctx = d->log_ctx;
825 mutex_unlock(&d->vqs[i].mutex);
826 }
827 if (ctx)
828 eventfd_ctx_put(ctx);
829 if (filep)
830 fput(filep);
831 break;
832 default:
833 r = vhost_set_vring(d, ioctl, argp);
834 break;
835 }
836 done:
837 return r;
838 }
839
840 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
841 __u64 addr, __u32 len)
842 {
843 struct vhost_memory_region *reg;
844 int i;
845 /* linear search is not brilliant, but we really have on the order of 6
846 * regions in practice */
847 for (i = 0; i < mem->nregions; ++i) {
848 reg = mem->regions + i;
849 if (reg->guest_phys_addr <= addr &&
850 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
851 return reg;
852 }
853 return NULL;
854 }
855
856 /* TODO: This is really inefficient. We need something like get_user()
857 * (instruction directly accesses the data, with an exception table entry
858 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
859 */
860 static int set_bit_to_user(int nr, void __user *addr)
861 {
862 unsigned long log = (unsigned long)addr;
863 struct page *page;
864 void *base;
865 int bit = nr + (log % PAGE_SIZE) * 8;
866 int r;
867 r = get_user_pages_fast(log, 1, 1, &page);
868 if (r < 0)
869 return r;
870 BUG_ON(r != 1);
871 base = kmap_atomic(page, KM_USER0);
872 set_bit(bit, base);
873 kunmap_atomic(base, KM_USER0);
874 set_page_dirty_lock(page);
875 put_page(page);
876 return 0;
877 }
878
879 static int log_write(void __user *log_base,
880 u64 write_address, u64 write_length)
881 {
882 int r;
883 if (!write_length)
884 return 0;
885 write_address /= VHOST_PAGE_SIZE;
886 for (;;) {
887 u64 base = (u64)(unsigned long)log_base;
888 u64 log = base + write_address / 8;
889 int bit = write_address % 8;
890 if ((u64)(unsigned long)log != log)
891 return -EFAULT;
892 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
893 if (r < 0)
894 return r;
895 if (write_length <= VHOST_PAGE_SIZE)
896 break;
897 write_length -= VHOST_PAGE_SIZE;
898 write_address += VHOST_PAGE_SIZE;
899 }
900 return r;
901 }
902
903 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
904 unsigned int log_num, u64 len)
905 {
906 int i, r;
907
908 /* Make sure data written is seen before log. */
909 smp_wmb();
910 for (i = 0; i < log_num; ++i) {
911 u64 l = min(log[i].len, len);
912 r = log_write(vq->log_base, log[i].addr, l);
913 if (r < 0)
914 return r;
915 len -= l;
916 if (!len) {
917 if (vq->log_ctx)
918 eventfd_signal(vq->log_ctx, 1);
919 return 0;
920 }
921 }
922 /* Length written exceeds what we have stored. This is a bug. */
923 BUG();
924 return 0;
925 }
926
927 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
928 struct iovec iov[], int iov_size)
929 {
930 const struct vhost_memory_region *reg;
931 struct vhost_memory *mem;
932 struct iovec *_iov;
933 u64 s = 0;
934 int ret = 0;
935
936 rcu_read_lock();
937
938 mem = rcu_dereference(dev->memory);
939 while ((u64)len > s) {
940 u64 size;
941 if (unlikely(ret >= iov_size)) {
942 ret = -ENOBUFS;
943 break;
944 }
945 reg = find_region(mem, addr, len);
946 if (unlikely(!reg)) {
947 ret = -EFAULT;
948 break;
949 }
950 _iov = iov + ret;
951 size = reg->memory_size - addr + reg->guest_phys_addr;
952 _iov->iov_len = min((u64)len, size);
953 _iov->iov_base = (void __user *)(unsigned long)
954 (reg->userspace_addr + addr - reg->guest_phys_addr);
955 s += size;
956 addr += size;
957 ++ret;
958 }
959
960 rcu_read_unlock();
961 return ret;
962 }
963
964 /* Each buffer in the virtqueues is actually a chain of descriptors. This
965 * function returns the next descriptor in the chain,
966 * or -1U if we're at the end. */
967 static unsigned next_desc(struct vring_desc *desc)
968 {
969 unsigned int next;
970
971 /* If this descriptor says it doesn't chain, we're done. */
972 if (!(desc->flags & VRING_DESC_F_NEXT))
973 return -1U;
974
975 /* Check they're not leading us off end of descriptors. */
976 next = desc->next;
977 /* Make sure compiler knows to grab that: we don't want it changing! */
978 /* We will use the result as an index in an array, so most
979 * architectures only need a compiler barrier here. */
980 read_barrier_depends();
981
982 return next;
983 }
984
985 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
986 struct iovec iov[], unsigned int iov_size,
987 unsigned int *out_num, unsigned int *in_num,
988 struct vhost_log *log, unsigned int *log_num,
989 struct vring_desc *indirect)
990 {
991 struct vring_desc desc;
992 unsigned int i = 0, count, found = 0;
993 int ret;
994
995 /* Sanity check */
996 if (unlikely(indirect->len % sizeof desc)) {
997 vq_err(vq, "Invalid length in indirect descriptor: "
998 "len 0x%llx not multiple of 0x%zx\n",
999 (unsigned long long)indirect->len,
1000 sizeof desc);
1001 return -EINVAL;
1002 }
1003
1004 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1005 UIO_MAXIOV);
1006 if (unlikely(ret < 0)) {
1007 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1008 return ret;
1009 }
1010
1011 /* We will use the result as an address to read from, so most
1012 * architectures only need a compiler barrier here. */
1013 read_barrier_depends();
1014
1015 count = indirect->len / sizeof desc;
1016 /* Buffers are chained via a 16 bit next field, so
1017 * we can have at most 2^16 of these. */
1018 if (unlikely(count > USHRT_MAX + 1)) {
1019 vq_err(vq, "Indirect buffer length too big: %d\n",
1020 indirect->len);
1021 return -E2BIG;
1022 }
1023
1024 do {
1025 unsigned iov_count = *in_num + *out_num;
1026 if (unlikely(++found > count)) {
1027 vq_err(vq, "Loop detected: last one at %u "
1028 "indirect size %u\n",
1029 i, count);
1030 return -EINVAL;
1031 }
1032 if (unlikely(memcpy_fromiovec((unsigned char *)&desc, vq->indirect,
1033 sizeof desc))) {
1034 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1035 i, (size_t)indirect->addr + i * sizeof desc);
1036 return -EINVAL;
1037 }
1038 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1039 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1040 i, (size_t)indirect->addr + i * sizeof desc);
1041 return -EINVAL;
1042 }
1043
1044 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1045 iov_size - iov_count);
1046 if (unlikely(ret < 0)) {
1047 vq_err(vq, "Translation failure %d indirect idx %d\n",
1048 ret, i);
1049 return ret;
1050 }
1051 /* If this is an input descriptor, increment that count. */
1052 if (desc.flags & VRING_DESC_F_WRITE) {
1053 *in_num += ret;
1054 if (unlikely(log)) {
1055 log[*log_num].addr = desc.addr;
1056 log[*log_num].len = desc.len;
1057 ++*log_num;
1058 }
1059 } else {
1060 /* If it's an output descriptor, they're all supposed
1061 * to come before any input descriptors. */
1062 if (unlikely(*in_num)) {
1063 vq_err(vq, "Indirect descriptor "
1064 "has out after in: idx %d\n", i);
1065 return -EINVAL;
1066 }
1067 *out_num += ret;
1068 }
1069 } while ((i = next_desc(&desc)) != -1);
1070 return 0;
1071 }
1072
1073 /* This looks in the virtqueue and for the first available buffer, and converts
1074 * it to an iovec for convenient access. Since descriptors consist of some
1075 * number of output then some number of input descriptors, it's actually two
1076 * iovecs, but we pack them into one and note how many of each there were.
1077 *
1078 * This function returns the descriptor number found, or vq->num (which is
1079 * never a valid descriptor number) if none was found. A negative code is
1080 * returned on error. */
1081 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1082 struct iovec iov[], unsigned int iov_size,
1083 unsigned int *out_num, unsigned int *in_num,
1084 struct vhost_log *log, unsigned int *log_num)
1085 {
1086 struct vring_desc desc;
1087 unsigned int i, head, found = 0;
1088 u16 last_avail_idx;
1089 int ret;
1090
1091 /* Check it isn't doing very strange things with descriptor numbers. */
1092 last_avail_idx = vq->last_avail_idx;
1093 if (unlikely(get_user(vq->avail_idx, &vq->avail->idx))) {
1094 vq_err(vq, "Failed to access avail idx at %p\n",
1095 &vq->avail->idx);
1096 return -EFAULT;
1097 }
1098
1099 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1100 vq_err(vq, "Guest moved used index from %u to %u",
1101 last_avail_idx, vq->avail_idx);
1102 return -EFAULT;
1103 }
1104
1105 /* If there's nothing new since last we looked, return invalid. */
1106 if (vq->avail_idx == last_avail_idx)
1107 return vq->num;
1108
1109 /* Only get avail ring entries after they have been exposed by guest. */
1110 smp_rmb();
1111
1112 /* Grab the next descriptor number they're advertising, and increment
1113 * the index we've seen. */
1114 if (unlikely(get_user(head,
1115 &vq->avail->ring[last_avail_idx % vq->num]))) {
1116 vq_err(vq, "Failed to read head: idx %d address %p\n",
1117 last_avail_idx,
1118 &vq->avail->ring[last_avail_idx % vq->num]);
1119 return -EFAULT;
1120 }
1121
1122 /* If their number is silly, that's an error. */
1123 if (unlikely(head >= vq->num)) {
1124 vq_err(vq, "Guest says index %u > %u is available",
1125 head, vq->num);
1126 return -EINVAL;
1127 }
1128
1129 /* When we start there are none of either input nor output. */
1130 *out_num = *in_num = 0;
1131 if (unlikely(log))
1132 *log_num = 0;
1133
1134 i = head;
1135 do {
1136 unsigned iov_count = *in_num + *out_num;
1137 if (unlikely(i >= vq->num)) {
1138 vq_err(vq, "Desc index is %u > %u, head = %u",
1139 i, vq->num, head);
1140 return -EINVAL;
1141 }
1142 if (unlikely(++found > vq->num)) {
1143 vq_err(vq, "Loop detected: last one at %u "
1144 "vq size %u head %u\n",
1145 i, vq->num, head);
1146 return -EINVAL;
1147 }
1148 ret = copy_from_user(&desc, vq->desc + i, sizeof desc);
1149 if (unlikely(ret)) {
1150 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1151 i, vq->desc + i);
1152 return -EFAULT;
1153 }
1154 if (desc.flags & VRING_DESC_F_INDIRECT) {
1155 ret = get_indirect(dev, vq, iov, iov_size,
1156 out_num, in_num,
1157 log, log_num, &desc);
1158 if (unlikely(ret < 0)) {
1159 vq_err(vq, "Failure detected "
1160 "in indirect descriptor at idx %d\n", i);
1161 return ret;
1162 }
1163 continue;
1164 }
1165
1166 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1167 iov_size - iov_count);
1168 if (unlikely(ret < 0)) {
1169 vq_err(vq, "Translation failure %d descriptor idx %d\n",
1170 ret, i);
1171 return ret;
1172 }
1173 if (desc.flags & VRING_DESC_F_WRITE) {
1174 /* If this is an input descriptor,
1175 * increment that count. */
1176 *in_num += ret;
1177 if (unlikely(log)) {
1178 log[*log_num].addr = desc.addr;
1179 log[*log_num].len = desc.len;
1180 ++*log_num;
1181 }
1182 } else {
1183 /* If it's an output descriptor, they're all supposed
1184 * to come before any input descriptors. */
1185 if (unlikely(*in_num)) {
1186 vq_err(vq, "Descriptor has out after in: "
1187 "idx %d\n", i);
1188 return -EINVAL;
1189 }
1190 *out_num += ret;
1191 }
1192 } while ((i = next_desc(&desc)) != -1);
1193
1194 /* On success, increment avail index. */
1195 vq->last_avail_idx++;
1196 return head;
1197 }
1198
1199 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1200 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1201 {
1202 vq->last_avail_idx -= n;
1203 }
1204
1205 /* After we've used one of their buffers, we tell them about it. We'll then
1206 * want to notify the guest, using eventfd. */
1207 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1208 {
1209 struct vring_used_elem __user *used;
1210
1211 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1212 * next entry in that used ring. */
1213 used = &vq->used->ring[vq->last_used_idx % vq->num];
1214 if (put_user(head, &used->id)) {
1215 vq_err(vq, "Failed to write used id");
1216 return -EFAULT;
1217 }
1218 if (put_user(len, &used->len)) {
1219 vq_err(vq, "Failed to write used len");
1220 return -EFAULT;
1221 }
1222 /* Make sure buffer is written before we update index. */
1223 smp_wmb();
1224 if (put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1225 vq_err(vq, "Failed to increment used idx");
1226 return -EFAULT;
1227 }
1228 if (unlikely(vq->log_used)) {
1229 /* Make sure data is seen before log. */
1230 smp_wmb();
1231 /* Log used ring entry write. */
1232 log_write(vq->log_base,
1233 vq->log_addr +
1234 ((void __user *)used - (void __user *)vq->used),
1235 sizeof *used);
1236 /* Log used index update. */
1237 log_write(vq->log_base,
1238 vq->log_addr + offsetof(struct vring_used, idx),
1239 sizeof vq->used->idx);
1240 if (vq->log_ctx)
1241 eventfd_signal(vq->log_ctx, 1);
1242 }
1243 vq->last_used_idx++;
1244 return 0;
1245 }
1246
1247 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1248 struct vring_used_elem *heads,
1249 unsigned count)
1250 {
1251 struct vring_used_elem __user *used;
1252 int start;
1253
1254 start = vq->last_used_idx % vq->num;
1255 used = vq->used->ring + start;
1256 if (copy_to_user(used, heads, count * sizeof *used)) {
1257 vq_err(vq, "Failed to write used");
1258 return -EFAULT;
1259 }
1260 if (unlikely(vq->log_used)) {
1261 /* Make sure data is seen before log. */
1262 smp_wmb();
1263 /* Log used ring entry write. */
1264 log_write(vq->log_base,
1265 vq->log_addr +
1266 ((void __user *)used - (void __user *)vq->used),
1267 count * sizeof *used);
1268 }
1269 vq->last_used_idx += count;
1270 return 0;
1271 }
1272
1273 /* After we've used one of their buffers, we tell them about it. We'll then
1274 * want to notify the guest, using eventfd. */
1275 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1276 unsigned count)
1277 {
1278 int start, n, r;
1279
1280 start = vq->last_used_idx % vq->num;
1281 n = vq->num - start;
1282 if (n < count) {
1283 r = __vhost_add_used_n(vq, heads, n);
1284 if (r < 0)
1285 return r;
1286 heads += n;
1287 count -= n;
1288 }
1289 r = __vhost_add_used_n(vq, heads, count);
1290
1291 /* Make sure buffer is written before we update index. */
1292 smp_wmb();
1293 if (put_user(vq->last_used_idx, &vq->used->idx)) {
1294 vq_err(vq, "Failed to increment used idx");
1295 return -EFAULT;
1296 }
1297 if (unlikely(vq->log_used)) {
1298 /* Log used index update. */
1299 log_write(vq->log_base,
1300 vq->log_addr + offsetof(struct vring_used, idx),
1301 sizeof vq->used->idx);
1302 if (vq->log_ctx)
1303 eventfd_signal(vq->log_ctx, 1);
1304 }
1305 return r;
1306 }
1307
1308 /* This actually signals the guest, using eventfd. */
1309 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1310 {
1311 __u16 flags;
1312 /* Flush out used index updates. This is paired
1313 * with the barrier that the Guest executes when enabling
1314 * interrupts. */
1315 smp_mb();
1316
1317 if (get_user(flags, &vq->avail->flags)) {
1318 vq_err(vq, "Failed to get flags");
1319 return;
1320 }
1321
1322 /* If they don't want an interrupt, don't signal, unless empty. */
1323 if ((flags & VRING_AVAIL_F_NO_INTERRUPT) &&
1324 (vq->avail_idx != vq->last_avail_idx ||
1325 !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY)))
1326 return;
1327
1328 /* Signal the Guest tell them we used something up. */
1329 if (vq->call_ctx)
1330 eventfd_signal(vq->call_ctx, 1);
1331 }
1332
1333 /* And here's the combo meal deal. Supersize me! */
1334 void vhost_add_used_and_signal(struct vhost_dev *dev,
1335 struct vhost_virtqueue *vq,
1336 unsigned int head, int len)
1337 {
1338 vhost_add_used(vq, head, len);
1339 vhost_signal(dev, vq);
1340 }
1341
1342 /* multi-buffer version of vhost_add_used_and_signal */
1343 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1344 struct vhost_virtqueue *vq,
1345 struct vring_used_elem *heads, unsigned count)
1346 {
1347 vhost_add_used_n(vq, heads, count);
1348 vhost_signal(dev, vq);
1349 }
1350
1351 /* OK, now we need to know about added descriptors. */
1352 bool vhost_enable_notify(struct vhost_virtqueue *vq)
1353 {
1354 u16 avail_idx;
1355 int r;
1356 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1357 return false;
1358 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1359 r = put_user(vq->used_flags, &vq->used->flags);
1360 if (r) {
1361 vq_err(vq, "Failed to enable notification at %p: %d\n",
1362 &vq->used->flags, r);
1363 return false;
1364 }
1365 /* They could have slipped one in as we were doing that: make
1366 * sure it's written, then check again. */
1367 smp_mb();
1368 r = get_user(avail_idx, &vq->avail->idx);
1369 if (r) {
1370 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1371 &vq->avail->idx, r);
1372 return false;
1373 }
1374
1375 return avail_idx != vq->avail_idx;
1376 }
1377
1378 /* We don't need to be notified again. */
1379 void vhost_disable_notify(struct vhost_virtqueue *vq)
1380 {
1381 int r;
1382 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1383 return;
1384 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1385 r = put_user(vq->used_flags, &vq->used->flags);
1386 if (r)
1387 vq_err(vq, "Failed to enable notification at %p: %d\n",
1388 &vq->used->flags, r);
1389 }
Linux kernel - Deliverables
This page took 0.059829 seconds and 5 git commands to generate.