vhost: apply cgroup to vhost workers
[linux-2.6.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 /* Init poll structure */
64 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
65                      unsigned long mask, struct vhost_dev *dev)
66 {
67         struct vhost_work *work = &poll->work;
68
69         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
70         init_poll_funcptr(&poll->table, vhost_poll_func);
71         poll->mask = mask;
72         poll->dev = dev;
73
74         INIT_LIST_HEAD(&work->node);
75         work->fn = fn;
76         init_waitqueue_head(&work->done);
77         work->flushing = 0;
78         work->queue_seq = work->done_seq = 0;
79 }
80
81 /* Start polling a file. We add ourselves to file's wait queue. The caller must
82  * keep a reference to a file until after vhost_poll_stop is called. */
83 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
84 {
85         unsigned long mask;
86         mask = file->f_op->poll(file, &poll->table);
87         if (mask)
88                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
89 }
90
91 /* Stop polling a file. After this function returns, it becomes safe to drop the
92  * file reference. You must also flush afterwards. */
93 void vhost_poll_stop(struct vhost_poll *poll)
94 {
95         remove_wait_queue(poll->wqh, &poll->wait);
96 }
97
98 /* Flush any work that has been scheduled. When calling this, don't hold any
99  * locks that are also used by the callback. */
100 void vhost_poll_flush(struct vhost_poll *poll)
101 {
102         struct vhost_work *work = &poll->work;
103         unsigned seq;
104         int left;
105         int flushing;
106
107         spin_lock_irq(&poll->dev->work_lock);
108         seq = work->queue_seq;
109         work->flushing++;
110         spin_unlock_irq(&poll->dev->work_lock);
111         wait_event(work->done, ({
112                    spin_lock_irq(&poll->dev->work_lock);
113                    left = seq - work->done_seq <= 0;
114                    spin_unlock_irq(&poll->dev->work_lock);
115                    left;
116         }));
117         spin_lock_irq(&poll->dev->work_lock);
118         flushing = --work->flushing;
119         spin_unlock_irq(&poll->dev->work_lock);
120         BUG_ON(flushing < 0);
121 }
122
123 void vhost_poll_queue(struct vhost_poll *poll)
124 {
125         struct vhost_dev *dev = poll->dev;
126         struct vhost_work *work = &poll->work;
127         unsigned long flags;
128
129         spin_lock_irqsave(&dev->work_lock, flags);
130         if (list_empty(&work->node)) {
131                 list_add_tail(&work->node, &dev->work_list);
132                 work->queue_seq++;
133                 wake_up_process(dev->worker);
134         }
135         spin_unlock_irqrestore(&dev->work_lock, flags);
136 }
137
138 static void vhost_vq_reset(struct vhost_dev *dev,
139                            struct vhost_virtqueue *vq)
140 {
141         vq->num = 1;
142         vq->desc = NULL;
143         vq->avail = NULL;
144         vq->used = NULL;
145         vq->last_avail_idx = 0;
146         vq->avail_idx = 0;
147         vq->last_used_idx = 0;
148         vq->used_flags = 0;
149         vq->used_flags = 0;
150         vq->log_used = false;
151         vq->log_addr = -1ull;
152         vq->hdr_size = 0;
153         vq->private_data = NULL;
154         vq->log_base = NULL;
155         vq->error_ctx = NULL;
156         vq->error = NULL;
157         vq->kick = NULL;
158         vq->call_ctx = NULL;
159         vq->call = NULL;
160         vq->log_ctx = NULL;
161 }
162
163 static int vhost_worker(void *data)
164 {
165         struct vhost_dev *dev = data;
166         struct vhost_work *work = NULL;
167         unsigned uninitialized_var(seq);
168
169         for (;;) {
170                 /* mb paired w/ kthread_stop */
171                 set_current_state(TASK_INTERRUPTIBLE);
172
173                 spin_lock_irq(&dev->work_lock);
174                 if (work) {
175                         work->done_seq = seq;
176                         if (work->flushing)
177                                 wake_up_all(&work->done);
178                 }
179
180                 if (kthread_should_stop()) {
181                         spin_unlock_irq(&dev->work_lock);
182                         __set_current_state(TASK_RUNNING);
183                         return 0;
184                 }
185                 if (!list_empty(&dev->work_list)) {
186                         work = list_first_entry(&dev->work_list,
187                                                 struct vhost_work, node);
188                         list_del_init(&work->node);
189                         seq = work->queue_seq;
190                 } else
191                         work = NULL;
192                 spin_unlock_irq(&dev->work_lock);
193
194                 if (work) {
195                         __set_current_state(TASK_RUNNING);
196                         work->fn(work);
197                 } else
198                         schedule();
199
200         }
201 }
202
203 long vhost_dev_init(struct vhost_dev *dev,
204                     struct vhost_virtqueue *vqs, int nvqs)
205 {
206         int i;
207
208         dev->vqs = vqs;
209         dev->nvqs = nvqs;
210         mutex_init(&dev->mutex);
211         dev->log_ctx = NULL;
212         dev->log_file = NULL;
213         dev->memory = NULL;
214         dev->mm = NULL;
215         spin_lock_init(&dev->work_lock);
216         INIT_LIST_HEAD(&dev->work_list);
217         dev->worker = NULL;
218
219         for (i = 0; i < dev->nvqs; ++i) {
220                 dev->vqs[i].dev = dev;
221                 mutex_init(&dev->vqs[i].mutex);
222                 vhost_vq_reset(dev, dev->vqs + i);
223                 if (dev->vqs[i].handle_kick)
224                         vhost_poll_init(&dev->vqs[i].poll,
225                                         dev->vqs[i].handle_kick, POLLIN, dev);
226         }
227
228         return 0;
229 }
230
231 /* Caller should have device mutex */
232 long vhost_dev_check_owner(struct vhost_dev *dev)
233 {
234         /* Are you the owner? If not, I don't think you mean to do that */
235         return dev->mm == current->mm ? 0 : -EPERM;
236 }
237
238 /* Caller should have device mutex */
239 static long vhost_dev_set_owner(struct vhost_dev *dev)
240 {
241         struct task_struct *worker;
242         int err;
243         /* Is there an owner already? */
244         if (dev->mm) {
245                 err = -EBUSY;
246                 goto err_mm;
247         }
248         /* No owner, become one */
249         dev->mm = get_task_mm(current);
250         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
251         if (IS_ERR(worker)) {
252                 err = PTR_ERR(worker);
253                 goto err_worker;
254         }
255
256         dev->worker = worker;
257         err = cgroup_attach_task_current_cg(worker);
258         if (err)
259                 goto err_cgroup;
260         wake_up_process(worker);        /* avoid contributing to loadavg */
261
262         return 0;
263 err_cgroup:
264         kthread_stop(worker);
265 err_worker:
266         if (dev->mm)
267                 mmput(dev->mm);
268         dev->mm = NULL;
269 err_mm:
270         return err;
271 }
272
273 /* Caller should have device mutex */
274 long vhost_dev_reset_owner(struct vhost_dev *dev)
275 {
276         struct vhost_memory *memory;
277
278         /* Restore memory to default empty mapping. */
279         memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
280         if (!memory)
281                 return -ENOMEM;
282
283         vhost_dev_cleanup(dev);
284
285         memory->nregions = 0;
286         dev->memory = memory;
287         return 0;
288 }
289
290 /* Caller should have device mutex */
291 void vhost_dev_cleanup(struct vhost_dev *dev)
292 {
293         int i;
294         for (i = 0; i < dev->nvqs; ++i) {
295                 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
296                         vhost_poll_stop(&dev->vqs[i].poll);
297                         vhost_poll_flush(&dev->vqs[i].poll);
298                 }
299                 if (dev->vqs[i].error_ctx)
300                         eventfd_ctx_put(dev->vqs[i].error_ctx);
301                 if (dev->vqs[i].error)
302                         fput(dev->vqs[i].error);
303                 if (dev->vqs[i].kick)
304                         fput(dev->vqs[i].kick);
305                 if (dev->vqs[i].call_ctx)
306                         eventfd_ctx_put(dev->vqs[i].call_ctx);
307                 if (dev->vqs[i].call)
308                         fput(dev->vqs[i].call);
309                 vhost_vq_reset(dev, dev->vqs + i);
310         }
311         if (dev->log_ctx)
312                 eventfd_ctx_put(dev->log_ctx);
313         dev->log_ctx = NULL;
314         if (dev->log_file)
315                 fput(dev->log_file);
316         dev->log_file = NULL;
317         /* No one will access memory at this point */
318         kfree(dev->memory);
319         dev->memory = NULL;
320         if (dev->mm)
321                 mmput(dev->mm);
322         dev->mm = NULL;
323
324         WARN_ON(!list_empty(&dev->work_list));
325         kthread_stop(dev->worker);
326 }
327
328 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
329 {
330         u64 a = addr / VHOST_PAGE_SIZE / 8;
331         /* Make sure 64 bit math will not overflow. */
332         if (a > ULONG_MAX - (unsigned long)log_base ||
333             a + (unsigned long)log_base > ULONG_MAX)
334                 return -EFAULT;
335
336         return access_ok(VERIFY_WRITE, log_base + a,
337                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
338 }
339
340 /* Caller should have vq mutex and device mutex. */
341 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
342                                int log_all)
343 {
344         int i;
345
346         if (!mem)
347                 return 0;
348
349         for (i = 0; i < mem->nregions; ++i) {
350                 struct vhost_memory_region *m = mem->regions + i;
351                 unsigned long a = m->userspace_addr;
352                 if (m->memory_size > ULONG_MAX)
353                         return 0;
354                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
355                                     m->memory_size))
356                         return 0;
357                 else if (log_all && !log_access_ok(log_base,
358                                                    m->guest_phys_addr,
359                                                    m->memory_size))
360                         return 0;
361         }
362         return 1;
363 }
364
365 /* Can we switch to this memory table? */
366 /* Caller should have device mutex but not vq mutex */
367 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
368                             int log_all)
369 {
370         int i;
371         for (i = 0; i < d->nvqs; ++i) {
372                 int ok;
373                 mutex_lock(&d->vqs[i].mutex);
374                 /* If ring is inactive, will check when it's enabled. */
375                 if (d->vqs[i].private_data)
376                         ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
377                                                  log_all);
378                 else
379                         ok = 1;
380                 mutex_unlock(&d->vqs[i].mutex);
381                 if (!ok)
382                         return 0;
383         }
384         return 1;
385 }
386
387 static int vq_access_ok(unsigned int num,
388                         struct vring_desc __user *desc,
389                         struct vring_avail __user *avail,
390                         struct vring_used __user *used)
391 {
392         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
393                access_ok(VERIFY_READ, avail,
394                          sizeof *avail + num * sizeof *avail->ring) &&
395                access_ok(VERIFY_WRITE, used,
396                         sizeof *used + num * sizeof *used->ring);
397 }
398
399 /* Can we log writes? */
400 /* Caller should have device mutex but not vq mutex */
401 int vhost_log_access_ok(struct vhost_dev *dev)
402 {
403         return memory_access_ok(dev, dev->memory, 1);
404 }
405
406 /* Verify access for write logging. */
407 /* Caller should have vq mutex and device mutex */
408 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base)
409 {
410         return vq_memory_access_ok(log_base, vq->dev->memory,
411                             vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
412                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
413                                         sizeof *vq->used +
414                                         vq->num * sizeof *vq->used->ring));
415 }
416
417 /* Can we start vq? */
418 /* Caller should have vq mutex and device mutex */
419 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
420 {
421         return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) &&
422                 vq_log_access_ok(vq, vq->log_base);
423 }
424
425 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
426 {
427         struct vhost_memory mem, *newmem, *oldmem;
428         unsigned long size = offsetof(struct vhost_memory, regions);
429         if (copy_from_user(&mem, m, size))
430                 return -EFAULT;
431         if (mem.padding)
432                 return -EOPNOTSUPP;
433         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
434                 return -E2BIG;
435         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
436         if (!newmem)
437                 return -ENOMEM;
438
439         memcpy(newmem, &mem, size);
440         if (copy_from_user(newmem->regions, m->regions,
441                            mem.nregions * sizeof *m->regions)) {
442                 kfree(newmem);
443                 return -EFAULT;
444         }
445
446         if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL))) {
447                 kfree(newmem);
448                 return -EFAULT;
449         }
450         oldmem = d->memory;
451         rcu_assign_pointer(d->memory, newmem);
452         synchronize_rcu();
453         kfree(oldmem);
454         return 0;
455 }
456
457 static int init_used(struct vhost_virtqueue *vq,
458                      struct vring_used __user *used)
459 {
460         int r = put_user(vq->used_flags, &used->flags);
461         if (r)
462                 return r;
463         return get_user(vq->last_used_idx, &used->idx);
464 }
465
466 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
467 {
468         struct file *eventfp, *filep = NULL,
469                     *pollstart = NULL, *pollstop = NULL;
470         struct eventfd_ctx *ctx = NULL;
471         u32 __user *idxp = argp;
472         struct vhost_virtqueue *vq;
473         struct vhost_vring_state s;
474         struct vhost_vring_file f;
475         struct vhost_vring_addr a;
476         u32 idx;
477         long r;
478
479         r = get_user(idx, idxp);
480         if (r < 0)
481                 return r;
482         if (idx >= d->nvqs)
483                 return -ENOBUFS;
484
485         vq = d->vqs + idx;
486
487         mutex_lock(&vq->mutex);
488
489         switch (ioctl) {
490         case VHOST_SET_VRING_NUM:
491                 /* Resizing ring with an active backend?
492                  * You don't want to do that. */
493                 if (vq->private_data) {
494                         r = -EBUSY;
495                         break;
496                 }
497                 if (copy_from_user(&s, argp, sizeof s)) {
498                         r = -EFAULT;
499                         break;
500                 }
501                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
502                         r = -EINVAL;
503                         break;
504                 }
505                 vq->num = s.num;
506                 break;
507         case VHOST_SET_VRING_BASE:
508                 /* Moving base with an active backend?
509                  * You don't want to do that. */
510                 if (vq->private_data) {
511                         r = -EBUSY;
512                         break;
513                 }
514                 if (copy_from_user(&s, argp, sizeof s)) {
515                         r = -EFAULT;
516                         break;
517                 }
518                 if (s.num > 0xffff) {
519                         r = -EINVAL;
520                         break;
521                 }
522                 vq->last_avail_idx = s.num;
523                 /* Forget the cached index value. */
524                 vq->avail_idx = vq->last_avail_idx;
525                 break;
526         case VHOST_GET_VRING_BASE:
527                 s.index = idx;
528                 s.num = vq->last_avail_idx;
529                 if (copy_to_user(argp, &s, sizeof s))
530                         r = -EFAULT;
531                 break;
532         case VHOST_SET_VRING_ADDR:
533                 if (copy_from_user(&a, argp, sizeof a)) {
534                         r = -EFAULT;
535                         break;
536                 }
537                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
538                         r = -EOPNOTSUPP;
539                         break;
540                 }
541                 /* For 32bit, verify that the top 32bits of the user
542                    data are set to zero. */
543                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
544                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
545                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
546                         r = -EFAULT;
547                         break;
548                 }
549                 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
550                     (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
551                     (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
552                         r = -EINVAL;
553                         break;
554                 }
555
556                 /* We only verify access here if backend is configured.
557                  * If it is not, we don't as size might not have been setup.
558                  * We will verify when backend is configured. */
559                 if (vq->private_data) {
560                         if (!vq_access_ok(vq->num,
561                                 (void __user *)(unsigned long)a.desc_user_addr,
562                                 (void __user *)(unsigned long)a.avail_user_addr,
563                                 (void __user *)(unsigned long)a.used_user_addr)) {
564                                 r = -EINVAL;
565                                 break;
566                         }
567
568                         /* Also validate log access for used ring if enabled. */
569                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
570                             !log_access_ok(vq->log_base, a.log_guest_addr,
571                                            sizeof *vq->used +
572                                            vq->num * sizeof *vq->used->ring)) {
573                                 r = -EINVAL;
574                                 break;
575                         }
576                 }
577
578                 r = init_used(vq, (struct vring_used __user *)(unsigned long)
579                               a.used_user_addr);
580                 if (r)
581                         break;
582                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
583                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
584                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
585                 vq->log_addr = a.log_guest_addr;
586                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
587                 break;
588         case VHOST_SET_VRING_KICK:
589                 if (copy_from_user(&f, argp, sizeof f)) {
590                         r = -EFAULT;
591                         break;
592                 }
593                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
594                 if (IS_ERR(eventfp)) {
595                         r = PTR_ERR(eventfp);
596                         break;
597                 }
598                 if (eventfp != vq->kick) {
599                         pollstop = filep = vq->kick;
600                         pollstart = vq->kick = eventfp;
601                 } else
602                         filep = eventfp;
603                 break;
604         case VHOST_SET_VRING_CALL:
605                 if (copy_from_user(&f, argp, sizeof f)) {
606                         r = -EFAULT;
607                         break;
608                 }
609                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
610                 if (IS_ERR(eventfp)) {
611                         r = PTR_ERR(eventfp);
612                         break;
613                 }
614                 if (eventfp != vq->call) {
615                         filep = vq->call;
616                         ctx = vq->call_ctx;
617                         vq->call = eventfp;
618                         vq->call_ctx = eventfp ?
619                                 eventfd_ctx_fileget(eventfp) : NULL;
620                 } else
621                         filep = eventfp;
622                 break;
623         case VHOST_SET_VRING_ERR:
624                 if (copy_from_user(&f, argp, sizeof f)) {
625                         r = -EFAULT;
626                         break;
627                 }
628                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
629                 if (IS_ERR(eventfp)) {
630                         r = PTR_ERR(eventfp);
631                         break;
632                 }
633                 if (eventfp != vq->error) {
634                         filep = vq->error;
635                         vq->error = eventfp;
636                         ctx = vq->error_ctx;
637                         vq->error_ctx = eventfp ?
638                                 eventfd_ctx_fileget(eventfp) : NULL;
639                 } else
640                         filep = eventfp;
641                 break;
642         default:
643                 r = -ENOIOCTLCMD;
644         }
645
646         if (pollstop && vq->handle_kick)
647                 vhost_poll_stop(&vq->poll);
648
649         if (ctx)
650                 eventfd_ctx_put(ctx);
651         if (filep)
652                 fput(filep);
653
654         if (pollstart && vq->handle_kick)
655                 vhost_poll_start(&vq->poll, vq->kick);
656
657         mutex_unlock(&vq->mutex);
658
659         if (pollstop && vq->handle_kick)
660                 vhost_poll_flush(&vq->poll);
661         return r;
662 }
663
664 /* Caller must have device mutex */
665 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
666 {
667         void __user *argp = (void __user *)arg;
668         struct file *eventfp, *filep = NULL;
669         struct eventfd_ctx *ctx = NULL;
670         u64 p;
671         long r;
672         int i, fd;
673
674         /* If you are not the owner, you can become one */
675         if (ioctl == VHOST_SET_OWNER) {
676                 r = vhost_dev_set_owner(d);
677                 goto done;
678         }
679
680         /* You must be the owner to do anything else */
681         r = vhost_dev_check_owner(d);
682         if (r)
683                 goto done;
684
685         switch (ioctl) {
686         case VHOST_SET_MEM_TABLE:
687                 r = vhost_set_memory(d, argp);
688                 break;
689         case VHOST_SET_LOG_BASE:
690                 if (copy_from_user(&p, argp, sizeof p)) {
691                         r = -EFAULT;
692                         break;
693                 }
694                 if ((u64)(unsigned long)p != p) {
695                         r = -EFAULT;
696                         break;
697                 }
698                 for (i = 0; i < d->nvqs; ++i) {
699                         struct vhost_virtqueue *vq;
700                         void __user *base = (void __user *)(unsigned long)p;
701                         vq = d->vqs + i;
702                         mutex_lock(&vq->mutex);
703                         /* If ring is inactive, will check when it's enabled. */
704                         if (vq->private_data && !vq_log_access_ok(vq, base))
705                                 r = -EFAULT;
706                         else
707                                 vq->log_base = base;
708                         mutex_unlock(&vq->mutex);
709                 }
710                 break;
711         case VHOST_SET_LOG_FD:
712                 r = get_user(fd, (int __user *)argp);
713                 if (r < 0)
714                         break;
715                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
716                 if (IS_ERR(eventfp)) {
717                         r = PTR_ERR(eventfp);
718                         break;
719                 }
720                 if (eventfp != d->log_file) {
721                         filep = d->log_file;
722                         ctx = d->log_ctx;
723                         d->log_ctx = eventfp ?
724                                 eventfd_ctx_fileget(eventfp) : NULL;
725                 } else
726                         filep = eventfp;
727                 for (i = 0; i < d->nvqs; ++i) {
728                         mutex_lock(&d->vqs[i].mutex);
729                         d->vqs[i].log_ctx = d->log_ctx;
730                         mutex_unlock(&d->vqs[i].mutex);
731                 }
732                 if (ctx)
733                         eventfd_ctx_put(ctx);
734                 if (filep)
735                         fput(filep);
736                 break;
737         default:
738                 r = vhost_set_vring(d, ioctl, argp);
739                 break;
740         }
741 done:
742         return r;
743 }
744
745 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
746                                                      __u64 addr, __u32 len)
747 {
748         struct vhost_memory_region *reg;
749         int i;
750         /* linear search is not brilliant, but we really have on the order of 6
751          * regions in practice */
752         for (i = 0; i < mem->nregions; ++i) {
753                 reg = mem->regions + i;
754                 if (reg->guest_phys_addr <= addr &&
755                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
756                         return reg;
757         }
758         return NULL;
759 }
760
761 /* TODO: This is really inefficient.  We need something like get_user()
762  * (instruction directly accesses the data, with an exception table entry
763  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
764  */
765 static int set_bit_to_user(int nr, void __user *addr)
766 {
767         unsigned long log = (unsigned long)addr;
768         struct page *page;
769         void *base;
770         int bit = nr + (log % PAGE_SIZE) * 8;
771         int r;
772         r = get_user_pages_fast(log, 1, 1, &page);
773         if (r < 0)
774                 return r;
775         BUG_ON(r != 1);
776         base = kmap_atomic(page, KM_USER0);
777         set_bit(bit, base);
778         kunmap_atomic(base, KM_USER0);
779         set_page_dirty_lock(page);
780         put_page(page);
781         return 0;
782 }
783
784 static int log_write(void __user *log_base,
785                      u64 write_address, u64 write_length)
786 {
787         int r;
788         if (!write_length)
789                 return 0;
790         write_address /= VHOST_PAGE_SIZE;
791         for (;;) {
792                 u64 base = (u64)(unsigned long)log_base;
793                 u64 log = base + write_address / 8;
794                 int bit = write_address % 8;
795                 if ((u64)(unsigned long)log != log)
796                         return -EFAULT;
797                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
798                 if (r < 0)
799                         return r;
800                 if (write_length <= VHOST_PAGE_SIZE)
801                         break;
802                 write_length -= VHOST_PAGE_SIZE;
803                 write_address += VHOST_PAGE_SIZE;
804         }
805         return r;
806 }
807
808 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
809                     unsigned int log_num, u64 len)
810 {
811         int i, r;
812
813         /* Make sure data written is seen before log. */
814         smp_wmb();
815         for (i = 0; i < log_num; ++i) {
816                 u64 l = min(log[i].len, len);
817                 r = log_write(vq->log_base, log[i].addr, l);
818                 if (r < 0)
819                         return r;
820                 len -= l;
821                 if (!len)
822                         return 0;
823         }
824         if (vq->log_ctx)
825                 eventfd_signal(vq->log_ctx, 1);
826         /* Length written exceeds what we have stored. This is a bug. */
827         BUG();
828         return 0;
829 }
830
831 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
832                           struct iovec iov[], int iov_size)
833 {
834         const struct vhost_memory_region *reg;
835         struct vhost_memory *mem;
836         struct iovec *_iov;
837         u64 s = 0;
838         int ret = 0;
839
840         rcu_read_lock();
841
842         mem = rcu_dereference(dev->memory);
843         while ((u64)len > s) {
844                 u64 size;
845                 if (unlikely(ret >= iov_size)) {
846                         ret = -ENOBUFS;
847                         break;
848                 }
849                 reg = find_region(mem, addr, len);
850                 if (unlikely(!reg)) {
851                         ret = -EFAULT;
852                         break;
853                 }
854                 _iov = iov + ret;
855                 size = reg->memory_size - addr + reg->guest_phys_addr;
856                 _iov->iov_len = min((u64)len, size);
857                 _iov->iov_base = (void __user *)(unsigned long)
858                         (reg->userspace_addr + addr - reg->guest_phys_addr);
859                 s += size;
860                 addr += size;
861                 ++ret;
862         }
863
864         rcu_read_unlock();
865         return ret;
866 }
867
868 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
869  * function returns the next descriptor in the chain,
870  * or -1U if we're at the end. */
871 static unsigned next_desc(struct vring_desc *desc)
872 {
873         unsigned int next;
874
875         /* If this descriptor says it doesn't chain, we're done. */
876         if (!(desc->flags & VRING_DESC_F_NEXT))
877                 return -1U;
878
879         /* Check they're not leading us off end of descriptors. */
880         next = desc->next;
881         /* Make sure compiler knows to grab that: we don't want it changing! */
882         /* We will use the result as an index in an array, so most
883          * architectures only need a compiler barrier here. */
884         read_barrier_depends();
885
886         return next;
887 }
888
889 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
890                         struct iovec iov[], unsigned int iov_size,
891                         unsigned int *out_num, unsigned int *in_num,
892                         struct vhost_log *log, unsigned int *log_num,
893                         struct vring_desc *indirect)
894 {
895         struct vring_desc desc;
896         unsigned int i = 0, count, found = 0;
897         int ret;
898
899         /* Sanity check */
900         if (unlikely(indirect->len % sizeof desc)) {
901                 vq_err(vq, "Invalid length in indirect descriptor: "
902                        "len 0x%llx not multiple of 0x%zx\n",
903                        (unsigned long long)indirect->len,
904                        sizeof desc);
905                 return -EINVAL;
906         }
907
908         ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
909                              ARRAY_SIZE(vq->indirect));
910         if (unlikely(ret < 0)) {
911                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
912                 return ret;
913         }
914
915         /* We will use the result as an address to read from, so most
916          * architectures only need a compiler barrier here. */
917         read_barrier_depends();
918
919         count = indirect->len / sizeof desc;
920         /* Buffers are chained via a 16 bit next field, so
921          * we can have at most 2^16 of these. */
922         if (unlikely(count > USHRT_MAX + 1)) {
923                 vq_err(vq, "Indirect buffer length too big: %d\n",
924                        indirect->len);
925                 return -E2BIG;
926         }
927
928         do {
929                 unsigned iov_count = *in_num + *out_num;
930                 if (unlikely(++found > count)) {
931                         vq_err(vq, "Loop detected: last one at %u "
932                                "indirect size %u\n",
933                                i, count);
934                         return -EINVAL;
935                 }
936                 if (unlikely(memcpy_fromiovec((unsigned char *)&desc, vq->indirect,
937                                               sizeof desc))) {
938                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
939                                i, (size_t)indirect->addr + i * sizeof desc);
940                         return -EINVAL;
941                 }
942                 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
943                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
944                                i, (size_t)indirect->addr + i * sizeof desc);
945                         return -EINVAL;
946                 }
947
948                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
949                                      iov_size - iov_count);
950                 if (unlikely(ret < 0)) {
951                         vq_err(vq, "Translation failure %d indirect idx %d\n",
952                                ret, i);
953                         return ret;
954                 }
955                 /* If this is an input descriptor, increment that count. */
956                 if (desc.flags & VRING_DESC_F_WRITE) {
957                         *in_num += ret;
958                         if (unlikely(log)) {
959                                 log[*log_num].addr = desc.addr;
960                                 log[*log_num].len = desc.len;
961                                 ++*log_num;
962                         }
963                 } else {
964                         /* If it's an output descriptor, they're all supposed
965                          * to come before any input descriptors. */
966                         if (unlikely(*in_num)) {
967                                 vq_err(vq, "Indirect descriptor "
968                                        "has out after in: idx %d\n", i);
969                                 return -EINVAL;
970                         }
971                         *out_num += ret;
972                 }
973         } while ((i = next_desc(&desc)) != -1);
974         return 0;
975 }
976
977 /* This looks in the virtqueue and for the first available buffer, and converts
978  * it to an iovec for convenient access.  Since descriptors consist of some
979  * number of output then some number of input descriptors, it's actually two
980  * iovecs, but we pack them into one and note how many of each there were.
981  *
982  * This function returns the descriptor number found, or vq->num (which is
983  * never a valid descriptor number) if none was found.  A negative code is
984  * returned on error. */
985 int vhost_get_vq_desc(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 {
990         struct vring_desc desc;
991         unsigned int i, head, found = 0;
992         u16 last_avail_idx;
993         int ret;
994
995         /* Check it isn't doing very strange things with descriptor numbers. */
996         last_avail_idx = vq->last_avail_idx;
997         if (unlikely(get_user(vq->avail_idx, &vq->avail->idx))) {
998                 vq_err(vq, "Failed to access avail idx at %p\n",
999                        &vq->avail->idx);
1000                 return -EFAULT;
1001         }
1002
1003         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1004                 vq_err(vq, "Guest moved used index from %u to %u",
1005                        last_avail_idx, vq->avail_idx);
1006                 return -EFAULT;
1007         }
1008
1009         /* If there's nothing new since last we looked, return invalid. */
1010         if (vq->avail_idx == last_avail_idx)
1011                 return vq->num;
1012
1013         /* Only get avail ring entries after they have been exposed by guest. */
1014         smp_rmb();
1015
1016         /* Grab the next descriptor number they're advertising, and increment
1017          * the index we've seen. */
1018         if (unlikely(get_user(head,
1019                               &vq->avail->ring[last_avail_idx % vq->num]))) {
1020                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1021                        last_avail_idx,
1022                        &vq->avail->ring[last_avail_idx % vq->num]);
1023                 return -EFAULT;
1024         }
1025
1026         /* If their number is silly, that's an error. */
1027         if (unlikely(head >= vq->num)) {
1028                 vq_err(vq, "Guest says index %u > %u is available",
1029                        head, vq->num);
1030                 return -EINVAL;
1031         }
1032
1033         /* When we start there are none of either input nor output. */
1034         *out_num = *in_num = 0;
1035         if (unlikely(log))
1036                 *log_num = 0;
1037
1038         i = head;
1039         do {
1040                 unsigned iov_count = *in_num + *out_num;
1041                 if (unlikely(i >= vq->num)) {
1042                         vq_err(vq, "Desc index is %u > %u, head = %u",
1043                                i, vq->num, head);
1044                         return -EINVAL;
1045                 }
1046                 if (unlikely(++found > vq->num)) {
1047                         vq_err(vq, "Loop detected: last one at %u "
1048                                "vq size %u head %u\n",
1049                                i, vq->num, head);
1050                         return -EINVAL;
1051                 }
1052                 ret = copy_from_user(&desc, vq->desc + i, sizeof desc);
1053                 if (unlikely(ret)) {
1054                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1055                                i, vq->desc + i);
1056                         return -EFAULT;
1057                 }
1058                 if (desc.flags & VRING_DESC_F_INDIRECT) {
1059                         ret = get_indirect(dev, vq, iov, iov_size,
1060                                            out_num, in_num,
1061                                            log, log_num, &desc);
1062                         if (unlikely(ret < 0)) {
1063                                 vq_err(vq, "Failure detected "
1064                                        "in indirect descriptor at idx %d\n", i);
1065                                 return ret;
1066                         }
1067                         continue;
1068                 }
1069
1070                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1071                                      iov_size - iov_count);
1072                 if (unlikely(ret < 0)) {
1073                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1074                                ret, i);
1075                         return ret;
1076                 }
1077                 if (desc.flags & VRING_DESC_F_WRITE) {
1078                         /* If this is an input descriptor,
1079                          * increment that count. */
1080                         *in_num += ret;
1081                         if (unlikely(log)) {
1082                                 log[*log_num].addr = desc.addr;
1083                                 log[*log_num].len = desc.len;
1084                                 ++*log_num;
1085                         }
1086                 } else {
1087                         /* If it's an output descriptor, they're all supposed
1088                          * to come before any input descriptors. */
1089                         if (unlikely(*in_num)) {
1090                                 vq_err(vq, "Descriptor has out after in: "
1091                                        "idx %d\n", i);
1092                                 return -EINVAL;
1093                         }
1094                         *out_num += ret;
1095                 }
1096         } while ((i = next_desc(&desc)) != -1);
1097
1098         /* On success, increment avail index. */
1099         vq->last_avail_idx++;
1100         return head;
1101 }
1102
1103 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1104 void vhost_discard_vq_desc(struct vhost_virtqueue *vq)
1105 {
1106         vq->last_avail_idx--;
1107 }
1108
1109 /* After we've used one of their buffers, we tell them about it.  We'll then
1110  * want to notify the guest, using eventfd. */
1111 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1112 {
1113         struct vring_used_elem __user *used;
1114
1115         /* The virtqueue contains a ring of used buffers.  Get a pointer to the
1116          * next entry in that used ring. */
1117         used = &vq->used->ring[vq->last_used_idx % vq->num];
1118         if (put_user(head, &used->id)) {
1119                 vq_err(vq, "Failed to write used id");
1120                 return -EFAULT;
1121         }
1122         if (put_user(len, &used->len)) {
1123                 vq_err(vq, "Failed to write used len");
1124                 return -EFAULT;
1125         }
1126         /* Make sure buffer is written before we update index. */
1127         smp_wmb();
1128         if (put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1129                 vq_err(vq, "Failed to increment used idx");
1130                 return -EFAULT;
1131         }
1132         if (unlikely(vq->log_used)) {
1133                 /* Make sure data is seen before log. */
1134                 smp_wmb();
1135                 /* Log used ring entry write. */
1136                 log_write(vq->log_base,
1137                           vq->log_addr +
1138                            ((void __user *)used - (void __user *)vq->used),
1139                           sizeof *used);
1140                 /* Log used index update. */
1141                 log_write(vq->log_base,
1142                           vq->log_addr + offsetof(struct vring_used, idx),
1143                           sizeof vq->used->idx);
1144                 if (vq->log_ctx)
1145                         eventfd_signal(vq->log_ctx, 1);
1146         }
1147         vq->last_used_idx++;
1148         return 0;
1149 }
1150
1151 /* This actually signals the guest, using eventfd. */
1152 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1153 {
1154         __u16 flags;
1155         /* Flush out used index updates. This is paired
1156          * with the barrier that the Guest executes when enabling
1157          * interrupts. */
1158         smp_mb();
1159
1160         if (get_user(flags, &vq->avail->flags)) {
1161                 vq_err(vq, "Failed to get flags");
1162                 return;
1163         }
1164
1165         /* If they don't want an interrupt, don't signal, unless empty. */
1166         if ((flags & VRING_AVAIL_F_NO_INTERRUPT) &&
1167             (vq->avail_idx != vq->last_avail_idx ||
1168              !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY)))
1169                 return;
1170
1171         /* Signal the Guest tell them we used something up. */
1172         if (vq->call_ctx)
1173                 eventfd_signal(vq->call_ctx, 1);
1174 }
1175
1176 /* And here's the combo meal deal.  Supersize me! */
1177 void vhost_add_used_and_signal(struct vhost_dev *dev,
1178                                struct vhost_virtqueue *vq,
1179                                unsigned int head, int len)
1180 {
1181         vhost_add_used(vq, head, len);
1182         vhost_signal(dev, vq);
1183 }
1184
1185 /* OK, now we need to know about added descriptors. */
1186 bool vhost_enable_notify(struct vhost_virtqueue *vq)
1187 {
1188         u16 avail_idx;
1189         int r;
1190         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1191                 return false;
1192         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1193         r = put_user(vq->used_flags, &vq->used->flags);
1194         if (r) {
1195                 vq_err(vq, "Failed to enable notification at %p: %d\n",
1196                        &vq->used->flags, r);
1197                 return false;
1198         }
1199         /* They could have slipped one in as we were doing that: make
1200          * sure it's written, then check again. */
1201         smp_mb();
1202         r = get_user(avail_idx, &vq->avail->idx);
1203         if (r) {
1204                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1205                        &vq->avail->idx, r);
1206                 return false;
1207         }
1208
1209         return avail_idx != vq->last_avail_idx;
1210 }
1211
1212 /* We don't need to be notified again. */
1213 void vhost_disable_notify(struct vhost_virtqueue *vq)
1214 {
1215         int r;
1216         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1217                 return;
1218         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1219         r = put_user(vq->used_flags, &vq->used->flags);
1220         if (r)
1221                 vq_err(vq, "Failed to enable notification at %p: %d\n",
1222                        &vq->used->flags, r);
1223 }