aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/vhost/vhost.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/vhost/vhost.c')
-rw-r--r--drivers/vhost/vhost.c1098
1 files changed, 1098 insertions, 0 deletions
diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
new file mode 100644
index 000000000000..c8c25dbc5857
--- /dev/null
+++ b/drivers/vhost/vhost.c
@@ -0,0 +1,1098 @@
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
34enum {
35 VHOST_MEMORY_MAX_NREGIONS = 64,
36 VHOST_MEMORY_F_LOG = 0x1,
37};
38
39static struct workqueue_struct *vhost_workqueue;
40
41static 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
51static 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 */
64void 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. */
75void 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. */
85void 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. */
92void vhost_poll_flush(struct vhost_poll *poll)
93{
94 flush_work(&poll->work);
95}
96
97void vhost_poll_queue(struct vhost_poll *poll)
98{
99 queue_work(vhost_workqueue, &poll->work);
100}
101
102static 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}
125
126long vhost_dev_init(struct vhost_dev *dev,
127 struct vhost_virtqueue *vqs, int nvqs)
128{
129 int i;
130 dev->vqs = vqs;
131 dev->nvqs = nvqs;
132 mutex_init(&dev->mutex);
133 dev->log_ctx = NULL;
134 dev->log_file = NULL;
135 dev->memory = NULL;
136 dev->mm = NULL;
137
138 for (i = 0; i < dev->nvqs; ++i) {
139 dev->vqs[i].dev = dev;
140 mutex_init(&dev->vqs[i].mutex);
141 vhost_vq_reset(dev, dev->vqs + i);
142 if (dev->vqs[i].handle_kick)
143 vhost_poll_init(&dev->vqs[i].poll,
144 dev->vqs[i].handle_kick,
145 POLLIN);
146 }
147 return 0;
148}
149
150/* Caller should have device mutex */
151long vhost_dev_check_owner(struct vhost_dev *dev)
152{
153 /* Are you the owner? If not, I don't think you mean to do that */
154 return dev->mm == current->mm ? 0 : -EPERM;
155}
156
157/* Caller should have device mutex */
158static long vhost_dev_set_owner(struct vhost_dev *dev)
159{
160 /* Is there an owner already? */
161 if (dev->mm)
162 return -EBUSY;
163 /* No owner, become one */
164 dev->mm = get_task_mm(current);
165 return 0;
166}
167
168/* Caller should have device mutex */
169long vhost_dev_reset_owner(struct vhost_dev *dev)
170{
171 struct vhost_memory *memory;
172
173 /* Restore memory to default empty mapping. */
174 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
175 if (!memory)
176 return -ENOMEM;
177
178 vhost_dev_cleanup(dev);
179
180 memory->nregions = 0;
181 dev->memory = memory;
182 return 0;
183}
184
185/* Caller should have device mutex */
186void vhost_dev_cleanup(struct vhost_dev *dev)
187{
188 int i;
189 for (i = 0; i < dev->nvqs; ++i) {
190 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
191 vhost_poll_stop(&dev->vqs[i].poll);
192 vhost_poll_flush(&dev->vqs[i].poll);
193 }
194 if (dev->vqs[i].error_ctx)
195 eventfd_ctx_put(dev->vqs[i].error_ctx);
196 if (dev->vqs[i].error)
197 fput(dev->vqs[i].error);
198 if (dev->vqs[i].kick)
199 fput(dev->vqs[i].kick);
200 if (dev->vqs[i].call_ctx)
201 eventfd_ctx_put(dev->vqs[i].call_ctx);
202 if (dev->vqs[i].call)
203 fput(dev->vqs[i].call);
204 vhost_vq_reset(dev, dev->vqs + i);
205 }
206 if (dev->log_ctx)
207 eventfd_ctx_put(dev->log_ctx);
208 dev->log_ctx = NULL;
209 if (dev->log_file)
210 fput(dev->log_file);
211 dev->log_file = NULL;
212 /* No one will access memory at this point */
213 kfree(dev->memory);
214 dev->memory = NULL;
215 if (dev->mm)
216 mmput(dev->mm);
217 dev->mm = NULL;
218}
219
220static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
221{
222 u64 a = addr / VHOST_PAGE_SIZE / 8;
223 /* Make sure 64 bit math will not overflow. */
224 if (a > ULONG_MAX - (unsigned long)log_base ||
225 a + (unsigned long)log_base > ULONG_MAX)
226 return -EFAULT;
227
228 return access_ok(VERIFY_WRITE, log_base + a,
229 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
230}
231
232/* Caller should have vq mutex and device mutex. */
233static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
234 int log_all)
235{
236 int i;
237 for (i = 0; i < mem->nregions; ++i) {
238 struct vhost_memory_region *m = mem->regions + i;
239 unsigned long a = m->userspace_addr;
240 if (m->memory_size > ULONG_MAX)
241 return 0;
242 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
243 m->memory_size))
244 return 0;
245 else if (log_all && !log_access_ok(log_base,
246 m->guest_phys_addr,
247 m->memory_size))
248 return 0;
249 }
250 return 1;
251}
252
253/* Can we switch to this memory table? */
254/* Caller should have device mutex but not vq mutex */
255static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
256 int log_all)
257{
258 int i;
259 for (i = 0; i < d->nvqs; ++i) {
260 int ok;
261 mutex_lock(&d->vqs[i].mutex);
262 /* If ring is inactive, will check when it's enabled. */
263 if (d->vqs[i].private_data)
264 ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
265 log_all);
266 else
267 ok = 1;
268 mutex_unlock(&d->vqs[i].mutex);
269 if (!ok)
270 return 0;
271 }
272 return 1;
273}
274
275static int vq_access_ok(unsigned int num,
276 struct vring_desc __user *desc,
277 struct vring_avail __user *avail,
278 struct vring_used __user *used)
279{
280 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
281 access_ok(VERIFY_READ, avail,
282 sizeof *avail + num * sizeof *avail->ring) &&
283 access_ok(VERIFY_WRITE, used,
284 sizeof *used + num * sizeof *used->ring);
285}
286
287/* Can we log writes? */
288/* Caller should have device mutex but not vq mutex */
289int vhost_log_access_ok(struct vhost_dev *dev)
290{
291 return memory_access_ok(dev, dev->memory, 1);
292}
293
294/* Verify access for write logging. */
295/* Caller should have vq mutex and device mutex */
296static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base)
297{
298 return vq_memory_access_ok(log_base, vq->dev->memory,
299 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
300 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
301 sizeof *vq->used +
302 vq->num * sizeof *vq->used->ring));
303}
304
305/* Can we start vq? */
306/* Caller should have vq mutex and device mutex */
307int vhost_vq_access_ok(struct vhost_virtqueue *vq)
308{
309 return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) &&
310 vq_log_access_ok(vq, vq->log_base);
311}
312
313static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
314{
315 struct vhost_memory mem, *newmem, *oldmem;
316 unsigned long size = offsetof(struct vhost_memory, regions);
317 long r;
318 r = copy_from_user(&mem, m, size);
319 if (r)
320 return r;
321 if (mem.padding)
322 return -EOPNOTSUPP;
323 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
324 return -E2BIG;
325 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
326 if (!newmem)
327 return -ENOMEM;
328
329 memcpy(newmem, &mem, size);
330 r = copy_from_user(newmem->regions, m->regions,
331 mem.nregions * sizeof *m->regions);
332 if (r) {
333 kfree(newmem);
334 return r;
335 }
336
337 if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL)))
338 return -EFAULT;
339 oldmem = d->memory;
340 rcu_assign_pointer(d->memory, newmem);
341 synchronize_rcu();
342 kfree(oldmem);
343 return 0;
344}
345
346static int init_used(struct vhost_virtqueue *vq,
347 struct vring_used __user *used)
348{
349 int r = put_user(vq->used_flags, &used->flags);
350 if (r)
351 return r;
352 return get_user(vq->last_used_idx, &used->idx);
353}
354
355static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
356{
357 struct file *eventfp, *filep = NULL,
358 *pollstart = NULL, *pollstop = NULL;
359 struct eventfd_ctx *ctx = NULL;
360 u32 __user *idxp = argp;
361 struct vhost_virtqueue *vq;
362 struct vhost_vring_state s;
363 struct vhost_vring_file f;
364 struct vhost_vring_addr a;
365 u32 idx;
366 long r;
367
368 r = get_user(idx, idxp);
369 if (r < 0)
370 return r;
371 if (idx > d->nvqs)
372 return -ENOBUFS;
373
374 vq = d->vqs + idx;
375
376 mutex_lock(&vq->mutex);
377
378 switch (ioctl) {
379 case VHOST_SET_VRING_NUM:
380 /* Resizing ring with an active backend?
381 * You don't want to do that. */
382 if (vq->private_data) {
383 r = -EBUSY;
384 break;
385 }
386 r = copy_from_user(&s, argp, sizeof s);
387 if (r < 0)
388 break;
389 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
390 r = -EINVAL;
391 break;
392 }
393 vq->num = s.num;
394 break;
395 case VHOST_SET_VRING_BASE:
396 /* Moving base with an active backend?
397 * You don't want to do that. */
398 if (vq->private_data) {
399 r = -EBUSY;
400 break;
401 }
402 r = copy_from_user(&s, argp, sizeof s);
403 if (r < 0)
404 break;
405 if (s.num > 0xffff) {
406 r = -EINVAL;
407 break;
408 }
409 vq->last_avail_idx = s.num;
410 /* Forget the cached index value. */
411 vq->avail_idx = vq->last_avail_idx;
412 break;
413 case VHOST_GET_VRING_BASE:
414 s.index = idx;
415 s.num = vq->last_avail_idx;
416 r = copy_to_user(argp, &s, sizeof s);
417 break;
418 case VHOST_SET_VRING_ADDR:
419 r = copy_from_user(&a, argp, sizeof a);
420 if (r < 0)
421 break;
422 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
423 r = -EOPNOTSUPP;
424 break;
425 }
426 /* For 32bit, verify that the top 32bits of the user
427 data are set to zero. */
428 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
429 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
430 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
431 r = -EFAULT;
432 break;
433 }
434 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
435 (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
436 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
437 r = -EINVAL;
438 break;
439 }
440
441 /* We only verify access here if backend is configured.
442 * If it is not, we don't as size might not have been setup.
443 * We will verify when backend is configured. */
444 if (vq->private_data) {
445 if (!vq_access_ok(vq->num,
446 (void __user *)(unsigned long)a.desc_user_addr,
447 (void __user *)(unsigned long)a.avail_user_addr,
448 (void __user *)(unsigned long)a.used_user_addr)) {
449 r = -EINVAL;
450 break;
451 }
452
453 /* Also validate log access for used ring if enabled. */
454 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
455 !log_access_ok(vq->log_base, a.log_guest_addr,
456 sizeof *vq->used +
457 vq->num * sizeof *vq->used->ring)) {
458 r = -EINVAL;
459 break;
460 }
461 }
462
463 r = init_used(vq, (struct vring_used __user *)(unsigned long)
464 a.used_user_addr);
465 if (r)
466 break;
467 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
468 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
469 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
470 vq->log_addr = a.log_guest_addr;
471 vq->used = (void __user *)(unsigned long)a.used_user_addr;
472 break;
473 case VHOST_SET_VRING_KICK:
474 r = copy_from_user(&f, argp, sizeof f);
475 if (r < 0)
476 break;
477 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
478 if (IS_ERR(eventfp))
479 return PTR_ERR(eventfp);
480 if (eventfp != vq->kick) {
481 pollstop = filep = vq->kick;
482 pollstart = vq->kick = eventfp;
483 } else
484 filep = eventfp;
485 break;
486 case VHOST_SET_VRING_CALL:
487 r = copy_from_user(&f, argp, sizeof f);
488 if (r < 0)
489 break;
490 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
491 if (IS_ERR(eventfp))
492 return PTR_ERR(eventfp);
493 if (eventfp != vq->call) {
494 filep = vq->call;
495 ctx = vq->call_ctx;
496 vq->call = eventfp;
497 vq->call_ctx = eventfp ?
498 eventfd_ctx_fileget(eventfp) : NULL;
499 } else
500 filep = eventfp;
501 break;
502 case VHOST_SET_VRING_ERR:
503 r = copy_from_user(&f, argp, sizeof f);
504 if (r < 0)
505 break;
506 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
507 if (IS_ERR(eventfp))
508 return PTR_ERR(eventfp);
509 if (eventfp != vq->error) {
510 filep = vq->error;
511 vq->error = eventfp;
512 ctx = vq->error_ctx;
513 vq->error_ctx = eventfp ?
514 eventfd_ctx_fileget(eventfp) : NULL;
515 } else
516 filep = eventfp;
517 break;
518 default:
519 r = -ENOIOCTLCMD;
520 }
521
522 if (pollstop && vq->handle_kick)
523 vhost_poll_stop(&vq->poll);
524
525 if (ctx)
526 eventfd_ctx_put(ctx);
527 if (filep)
528 fput(filep);
529
530 if (pollstart && vq->handle_kick)
531 vhost_poll_start(&vq->poll, vq->kick);
532
533 mutex_unlock(&vq->mutex);
534
535 if (pollstop && vq->handle_kick)
536 vhost_poll_flush(&vq->poll);
537 return r;
538}
539
540/* Caller must have device mutex */
541long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
542{
543 void __user *argp = (void __user *)arg;
544 struct file *eventfp, *filep = NULL;
545 struct eventfd_ctx *ctx = NULL;
546 u64 p;
547 long r;
548 int i, fd;
549
550 /* If you are not the owner, you can become one */
551 if (ioctl == VHOST_SET_OWNER) {
552 r = vhost_dev_set_owner(d);
553 goto done;
554 }
555
556 /* You must be the owner to do anything else */
557 r = vhost_dev_check_owner(d);
558 if (r)
559 goto done;
560
561 switch (ioctl) {
562 case VHOST_SET_MEM_TABLE:
563 r = vhost_set_memory(d, argp);
564 break;
565 case VHOST_SET_LOG_BASE:
566 r = copy_from_user(&p, argp, sizeof p);
567 if (r < 0)
568 break;
569 if ((u64)(unsigned long)p != p) {
570 r = -EFAULT;
571 break;
572 }
573 for (i = 0; i < d->nvqs; ++i) {
574 struct vhost_virtqueue *vq;
575 void __user *base = (void __user *)(unsigned long)p;
576 vq = d->vqs + i;
577 mutex_lock(&vq->mutex);
578 /* If ring is inactive, will check when it's enabled. */
579 if (vq->private_data && !vq_log_access_ok(vq, base))
580 r = -EFAULT;
581 else
582 vq->log_base = base;
583 mutex_unlock(&vq->mutex);
584 }
585 break;
586 case VHOST_SET_LOG_FD:
587 r = get_user(fd, (int __user *)argp);
588 if (r < 0)
589 break;
590 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
591 if (IS_ERR(eventfp)) {
592 r = PTR_ERR(eventfp);
593 break;
594 }
595 if (eventfp != d->log_file) {
596 filep = d->log_file;
597 ctx = d->log_ctx;
598 d->log_ctx = eventfp ?
599 eventfd_ctx_fileget(eventfp) : NULL;
600 } else
601 filep = eventfp;
602 for (i = 0; i < d->nvqs; ++i) {
603 mutex_lock(&d->vqs[i].mutex);
604 d->vqs[i].log_ctx = d->log_ctx;
605 mutex_unlock(&d->vqs[i].mutex);
606 }
607 if (ctx)
608 eventfd_ctx_put(ctx);
609 if (filep)
610 fput(filep);
611 break;
612 default:
613 r = vhost_set_vring(d, ioctl, argp);
614 break;
615 }
616done:
617 return r;
618}
619
620static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
621 __u64 addr, __u32 len)
622{
623 struct vhost_memory_region *reg;
624 int i;
625 /* linear search is not brilliant, but we really have on the order of 6
626 * regions in practice */
627 for (i = 0; i < mem->nregions; ++i) {
628 reg = mem->regions + i;
629 if (reg->guest_phys_addr <= addr &&
630 reg->guest_phys_addr + reg->memory_size - 1 >= addr)
631 return reg;
632 }
633 return NULL;
634}
635
636/* TODO: This is really inefficient. We need something like get_user()
637 * (instruction directly accesses the data, with an exception table entry
638 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
639 */
640static int set_bit_to_user(int nr, void __user *addr)
641{
642 unsigned long log = (unsigned long)addr;
643 struct page *page;
644 void *base;
645 int bit = nr + (log % PAGE_SIZE) * 8;
646 int r;
647 r = get_user_pages_fast(log, 1, 1, &page);
648 if (r)
649 return r;
650 base = kmap_atomic(page, KM_USER0);
651 set_bit(bit, base);
652 kunmap_atomic(base, KM_USER0);
653 set_page_dirty_lock(page);
654 put_page(page);
655 return 0;
656}
657
658static int log_write(void __user *log_base,
659 u64 write_address, u64 write_length)
660{
661 int r;
662 if (!write_length)
663 return 0;
664 write_address /= VHOST_PAGE_SIZE;
665 for (;;) {
666 u64 base = (u64)(unsigned long)log_base;
667 u64 log = base + write_address / 8;
668 int bit = write_address % 8;
669 if ((u64)(unsigned long)log != log)
670 return -EFAULT;
671 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
672 if (r < 0)
673 return r;
674 if (write_length <= VHOST_PAGE_SIZE)
675 break;
676 write_length -= VHOST_PAGE_SIZE;
677 write_address += VHOST_PAGE_SIZE;
678 }
679 return r;
680}
681
682int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
683 unsigned int log_num, u64 len)
684{
685 int i, r;
686
687 /* Make sure data written is seen before log. */
688 wmb();
689 for (i = 0; i < log_num; ++i) {
690 u64 l = min(log[i].len, len);
691 r = log_write(vq->log_base, log[i].addr, l);
692 if (r < 0)
693 return r;
694 len -= l;
695 if (!len)
696 return 0;
697 }
698 if (vq->log_ctx)
699 eventfd_signal(vq->log_ctx, 1);
700 /* Length written exceeds what we have stored. This is a bug. */
701 BUG();
702 return 0;
703}
704
705int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
706 struct iovec iov[], int iov_size)
707{
708 const struct vhost_memory_region *reg;
709 struct vhost_memory *mem;
710 struct iovec *_iov;
711 u64 s = 0;
712 int ret = 0;
713
714 rcu_read_lock();
715
716 mem = rcu_dereference(dev->memory);
717 while ((u64)len > s) {
718 u64 size;
719 if (ret >= iov_size) {
720 ret = -ENOBUFS;
721 break;
722 }
723 reg = find_region(mem, addr, len);
724 if (!reg) {
725 ret = -EFAULT;
726 break;
727 }
728 _iov = iov + ret;
729 size = reg->memory_size - addr + reg->guest_phys_addr;
730 _iov->iov_len = min((u64)len, size);
731 _iov->iov_base = (void *)(unsigned long)
732 (reg->userspace_addr + addr - reg->guest_phys_addr);
733 s += size;
734 addr += size;
735 ++ret;
736 }
737
738 rcu_read_unlock();
739 return ret;
740}
741
742/* Each buffer in the virtqueues is actually a chain of descriptors. This
743 * function returns the next descriptor in the chain,
744 * or -1U if we're at the end. */
745static unsigned next_desc(struct vring_desc *desc)
746{
747 unsigned int next;
748
749 /* If this descriptor says it doesn't chain, we're done. */
750 if (!(desc->flags & VRING_DESC_F_NEXT))
751 return -1U;
752
753 /* Check they're not leading us off end of descriptors. */
754 next = desc->next;
755 /* Make sure compiler knows to grab that: we don't want it changing! */
756 /* We will use the result as an index in an array, so most
757 * architectures only need a compiler barrier here. */
758 read_barrier_depends();
759
760 return next;
761}
762
763static unsigned get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
764 struct iovec iov[], unsigned int iov_size,
765 unsigned int *out_num, unsigned int *in_num,
766 struct vhost_log *log, unsigned int *log_num,
767 struct vring_desc *indirect)
768{
769 struct vring_desc desc;
770 unsigned int i = 0, count, found = 0;
771 int ret;
772
773 /* Sanity check */
774 if (indirect->len % sizeof desc) {
775 vq_err(vq, "Invalid length in indirect descriptor: "
776 "len 0x%llx not multiple of 0x%zx\n",
777 (unsigned long long)indirect->len,
778 sizeof desc);
779 return -EINVAL;
780 }
781
782 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
783 ARRAY_SIZE(vq->indirect));
784 if (ret < 0) {
785 vq_err(vq, "Translation failure %d in indirect.\n", ret);
786 return ret;
787 }
788
789 /* We will use the result as an address to read from, so most
790 * architectures only need a compiler barrier here. */
791 read_barrier_depends();
792
793 count = indirect->len / sizeof desc;
794 /* Buffers are chained via a 16 bit next field, so
795 * we can have at most 2^16 of these. */
796 if (count > USHORT_MAX + 1) {
797 vq_err(vq, "Indirect buffer length too big: %d\n",
798 indirect->len);
799 return -E2BIG;
800 }
801
802 do {
803 unsigned iov_count = *in_num + *out_num;
804 if (++found > count) {
805 vq_err(vq, "Loop detected: last one at %u "
806 "indirect size %u\n",
807 i, count);
808 return -EINVAL;
809 }
810 if (memcpy_fromiovec((unsigned char *)&desc, vq->indirect,
811 sizeof desc)) {
812 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
813 i, (size_t)indirect->addr + i * sizeof desc);
814 return -EINVAL;
815 }
816 if (desc.flags & VRING_DESC_F_INDIRECT) {
817 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
818 i, (size_t)indirect->addr + i * sizeof desc);
819 return -EINVAL;
820 }
821
822 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
823 iov_size - iov_count);
824 if (ret < 0) {
825 vq_err(vq, "Translation failure %d indirect idx %d\n",
826 ret, i);
827 return ret;
828 }
829 /* If this is an input descriptor, increment that count. */
830 if (desc.flags & VRING_DESC_F_WRITE) {
831 *in_num += ret;
832 if (unlikely(log)) {
833 log[*log_num].addr = desc.addr;
834 log[*log_num].len = desc.len;
835 ++*log_num;
836 }
837 } else {
838 /* If it's an output descriptor, they're all supposed
839 * to come before any input descriptors. */
840 if (*in_num) {
841 vq_err(vq, "Indirect descriptor "
842 "has out after in: idx %d\n", i);
843 return -EINVAL;
844 }
845 *out_num += ret;
846 }
847 } while ((i = next_desc(&desc)) != -1);
848 return 0;
849}
850
851/* This looks in the virtqueue and for the first available buffer, and converts
852 * it to an iovec for convenient access. Since descriptors consist of some
853 * number of output then some number of input descriptors, it's actually two
854 * iovecs, but we pack them into one and note how many of each there were.
855 *
856 * This function returns the descriptor number found, or vq->num (which
857 * is never a valid descriptor number) if none was found. */
858unsigned vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
859 struct iovec iov[], unsigned int iov_size,
860 unsigned int *out_num, unsigned int *in_num,
861 struct vhost_log *log, unsigned int *log_num)
862{
863 struct vring_desc desc;
864 unsigned int i, head, found = 0;
865 u16 last_avail_idx;
866 int ret;
867
868 /* Check it isn't doing very strange things with descriptor numbers. */
869 last_avail_idx = vq->last_avail_idx;
870 if (get_user(vq->avail_idx, &vq->avail->idx)) {
871 vq_err(vq, "Failed to access avail idx at %p\n",
872 &vq->avail->idx);
873 return vq->num;
874 }
875
876 if ((u16)(vq->avail_idx - last_avail_idx) > vq->num) {
877 vq_err(vq, "Guest moved used index from %u to %u",
878 last_avail_idx, vq->avail_idx);
879 return vq->num;
880 }
881
882 /* If there's nothing new since last we looked, return invalid. */
883 if (vq->avail_idx == last_avail_idx)
884 return vq->num;
885
886 /* Only get avail ring entries after they have been exposed by guest. */
887 rmb();
888
889 /* Grab the next descriptor number they're advertising, and increment
890 * the index we've seen. */
891 if (get_user(head, &vq->avail->ring[last_avail_idx % vq->num])) {
892 vq_err(vq, "Failed to read head: idx %d address %p\n",
893 last_avail_idx,
894 &vq->avail->ring[last_avail_idx % vq->num]);
895 return vq->num;
896 }
897
898 /* If their number is silly, that's an error. */
899 if (head >= vq->num) {
900 vq_err(vq, "Guest says index %u > %u is available",
901 head, vq->num);
902 return vq->num;
903 }
904
905 /* When we start there are none of either input nor output. */
906 *out_num = *in_num = 0;
907 if (unlikely(log))
908 *log_num = 0;
909
910 i = head;
911 do {
912 unsigned iov_count = *in_num + *out_num;
913 if (i >= vq->num) {
914 vq_err(vq, "Desc index is %u > %u, head = %u",
915 i, vq->num, head);
916 return vq->num;
917 }
918 if (++found > vq->num) {
919 vq_err(vq, "Loop detected: last one at %u "
920 "vq size %u head %u\n",
921 i, vq->num, head);
922 return vq->num;
923 }
924 ret = copy_from_user(&desc, vq->desc + i, sizeof desc);
925 if (ret) {
926 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
927 i, vq->desc + i);
928 return vq->num;
929 }
930 if (desc.flags & VRING_DESC_F_INDIRECT) {
931 ret = get_indirect(dev, vq, iov, iov_size,
932 out_num, in_num,
933 log, log_num, &desc);
934 if (ret < 0) {
935 vq_err(vq, "Failure detected "
936 "in indirect descriptor at idx %d\n", i);
937 return vq->num;
938 }
939 continue;
940 }
941
942 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
943 iov_size - iov_count);
944 if (ret < 0) {
945 vq_err(vq, "Translation failure %d descriptor idx %d\n",
946 ret, i);
947 return vq->num;
948 }
949 if (desc.flags & VRING_DESC_F_WRITE) {
950 /* If this is an input descriptor,
951 * increment that count. */
952 *in_num += ret;
953 if (unlikely(log)) {
954 log[*log_num].addr = desc.addr;
955 log[*log_num].len = desc.len;
956 ++*log_num;
957 }
958 } else {
959 /* If it's an output descriptor, they're all supposed
960 * to come before any input descriptors. */
961 if (*in_num) {
962 vq_err(vq, "Descriptor has out after in: "
963 "idx %d\n", i);
964 return vq->num;
965 }
966 *out_num += ret;
967 }
968 } while ((i = next_desc(&desc)) != -1);
969
970 /* On success, increment avail index. */
971 vq->last_avail_idx++;
972 return head;
973}
974
975/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
976void vhost_discard_vq_desc(struct vhost_virtqueue *vq)
977{
978 vq->last_avail_idx--;
979}
980
981/* After we've used one of their buffers, we tell them about it. We'll then
982 * want to notify the guest, using eventfd. */
983int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
984{
985 struct vring_used_elem *used;
986
987 /* The virtqueue contains a ring of used buffers. Get a pointer to the
988 * next entry in that used ring. */
989 used = &vq->used->ring[vq->last_used_idx % vq->num];
990 if (put_user(head, &used->id)) {
991 vq_err(vq, "Failed to write used id");
992 return -EFAULT;
993 }
994 if (put_user(len, &used->len)) {
995 vq_err(vq, "Failed to write used len");
996 return -EFAULT;
997 }
998 /* Make sure buffer is written before we update index. */
999 wmb();
1000 if (put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1001 vq_err(vq, "Failed to increment used idx");
1002 return -EFAULT;
1003 }
1004 if (unlikely(vq->log_used)) {
1005 /* Make sure data is seen before log. */
1006 wmb();
1007 log_write(vq->log_base, vq->log_addr + sizeof *vq->used->ring *
1008 (vq->last_used_idx % vq->num),
1009 sizeof *vq->used->ring);
1010 log_write(vq->log_base, vq->log_addr, sizeof *vq->used->ring);
1011 if (vq->log_ctx)
1012 eventfd_signal(vq->log_ctx, 1);
1013 }
1014 vq->last_used_idx++;
1015 return 0;
1016}
1017
1018/* This actually signals the guest, using eventfd. */
1019void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1020{
1021 __u16 flags = 0;
1022 if (get_user(flags, &vq->avail->flags)) {
1023 vq_err(vq, "Failed to get flags");
1024 return;
1025 }
1026
1027 /* If they don't want an interrupt, don't signal, unless empty. */
1028 if ((flags & VRING_AVAIL_F_NO_INTERRUPT) &&
1029 (vq->avail_idx != vq->last_avail_idx ||
1030 !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY)))
1031 return;
1032
1033 /* Signal the Guest tell them we used something up. */
1034 if (vq->call_ctx)
1035 eventfd_signal(vq->call_ctx, 1);
1036}
1037
1038/* And here's the combo meal deal. Supersize me! */
1039void vhost_add_used_and_signal(struct vhost_dev *dev,
1040 struct vhost_virtqueue *vq,
1041 unsigned int head, int len)
1042{
1043 vhost_add_used(vq, head, len);
1044 vhost_signal(dev, vq);
1045}
1046
1047/* OK, now we need to know about added descriptors. */
1048bool vhost_enable_notify(struct vhost_virtqueue *vq)
1049{
1050 u16 avail_idx;
1051 int r;
1052 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1053 return false;
1054 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1055 r = put_user(vq->used_flags, &vq->used->flags);
1056 if (r) {
1057 vq_err(vq, "Failed to enable notification at %p: %d\n",
1058 &vq->used->flags, r);
1059 return false;
1060 }
1061 /* They could have slipped one in as we were doing that: make
1062 * sure it's written, then check again. */
1063 mb();
1064 r = get_user(avail_idx, &vq->avail->idx);
1065 if (r) {
1066 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1067 &vq->avail->idx, r);
1068 return false;
1069 }
1070
1071 return avail_idx != vq->last_avail_idx;
1072}
1073
1074/* We don't need to be notified again. */
1075void vhost_disable_notify(struct vhost_virtqueue *vq)
1076{
1077 int r;
1078 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1079 return;
1080 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1081 r = put_user(vq->used_flags, &vq->used->flags);
1082 if (r)
1083 vq_err(vq, "Failed to enable notification at %p: %d\n",
1084 &vq->used->flags, r);
1085}
1086
1087int vhost_init(void)
1088{
1089 vhost_workqueue = create_singlethread_workqueue("vhost");
1090 if (!vhost_workqueue)
1091 return -ENOMEM;
1092 return 0;
1093}
1094
1095void vhost_cleanup(void)
1096{
1097 destroy_workqueue(vhost_workqueue);
1098}
generated by cgit v1.2.2 (git 2.25.0) at 2025-01-04 06:06:42 -0500