Root/
1 | /* Copyright (C) 2009 Red Hat, Inc. |
2 | * Copyright (C) 2006 Rusty Russell IBM Corporation |
3 | * |
4 | * Author: Michael S. Tsirkin <mst@redhat.com> |
5 | * |
6 | * Inspiration, some code, and most witty comments come from |
7 | * Documentation/virtual/lguest/lguest.c, by Rusty Russell |
8 | * |
9 | * This work is licensed under the terms of the GNU GPL, version 2. |
10 | * |
11 | * Generic code for virtio server in host kernel. |
12 | */ |
13 | |
14 | #include <linux/eventfd.h> |
15 | #include <linux/vhost.h> |
16 | #include <linux/virtio_net.h> |
17 | #include <linux/mm.h> |
18 | #include <linux/mmu_context.h> |
19 | #include <linux/miscdevice.h> |
20 | #include <linux/mutex.h> |
21 | #include <linux/rcupdate.h> |
22 | #include <linux/poll.h> |
23 | #include <linux/file.h> |
24 | #include <linux/highmem.h> |
25 | #include <linux/slab.h> |
26 | #include <linux/kthread.h> |
27 | #include <linux/cgroup.h> |
28 | |
29 | #include <linux/net.h> |
30 | #include <linux/if_packet.h> |
31 | #include <linux/if_arp.h> |
32 | |
33 | #include "vhost.h" |
34 | |
35 | enum { |
36 | VHOST_MEMORY_MAX_NREGIONS = 64, |
37 | VHOST_MEMORY_F_LOG = 0x1, |
38 | }; |
39 | |
40 | static unsigned vhost_zcopy_mask __read_mostly; |
41 | |
42 | #define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num]) |
43 | #define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num]) |
44 | |
45 | static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh, |
46 | poll_table *pt) |
47 | { |
48 | struct vhost_poll *poll; |
49 | |
50 | poll = container_of(pt, struct vhost_poll, table); |
51 | poll->wqh = wqh; |
52 | add_wait_queue(wqh, &poll->wait); |
53 | } |
54 | |
55 | static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync, |
56 | void *key) |
57 | { |
58 | struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait); |
59 | |
60 | if (!((unsigned long)key & poll->mask)) |
61 | return 0; |
62 | |
63 | vhost_poll_queue(poll); |
64 | return 0; |
65 | } |
66 | |
67 | static void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn) |
68 | { |
69 | INIT_LIST_HEAD(&work->node); |
70 | work->fn = fn; |
71 | init_waitqueue_head(&work->done); |
72 | work->flushing = 0; |
73 | work->queue_seq = work->done_seq = 0; |
74 | } |
75 | |
76 | /* Init poll structure */ |
77 | void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn, |
78 | unsigned long mask, struct vhost_dev *dev) |
79 | { |
80 | init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup); |
81 | init_poll_funcptr(&poll->table, vhost_poll_func); |
82 | poll->mask = mask; |
83 | poll->dev = dev; |
84 | |
85 | vhost_work_init(&poll->work, fn); |
86 | } |
87 | |
88 | /* Start polling a file. We add ourselves to file's wait queue. The caller must |
89 | * keep a reference to a file until after vhost_poll_stop is called. */ |
90 | void vhost_poll_start(struct vhost_poll *poll, struct file *file) |
91 | { |
92 | unsigned long mask; |
93 | |
94 | mask = file->f_op->poll(file, &poll->table); |
95 | if (mask) |
96 | vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask); |
97 | } |
98 | |
99 | /* Stop polling a file. After this function returns, it becomes safe to drop the |
100 | * file reference. You must also flush afterwards. */ |
101 | void vhost_poll_stop(struct vhost_poll *poll) |
102 | { |
103 | remove_wait_queue(poll->wqh, &poll->wait); |
104 | } |
105 | |
106 | static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work, |
107 | unsigned seq) |
108 | { |
109 | int left; |
110 | |
111 | spin_lock_irq(&dev->work_lock); |
112 | left = seq - work->done_seq; |
113 | spin_unlock_irq(&dev->work_lock); |
114 | return left <= 0; |
115 | } |
116 | |
117 | static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work) |
118 | { |
119 | unsigned seq; |
120 | int flushing; |
121 | |
122 | spin_lock_irq(&dev->work_lock); |
123 | seq = work->queue_seq; |
124 | work->flushing++; |
125 | spin_unlock_irq(&dev->work_lock); |
126 | wait_event(work->done, vhost_work_seq_done(dev, work, seq)); |
127 | spin_lock_irq(&dev->work_lock); |
128 | flushing = --work->flushing; |
129 | spin_unlock_irq(&dev->work_lock); |
130 | BUG_ON(flushing < 0); |
131 | } |
132 | |
133 | /* Flush any work that has been scheduled. When calling this, don't hold any |
134 | * locks that are also used by the callback. */ |
135 | void vhost_poll_flush(struct vhost_poll *poll) |
136 | { |
137 | vhost_work_flush(poll->dev, &poll->work); |
138 | } |
139 | |
140 | static inline void vhost_work_queue(struct vhost_dev *dev, |
141 | struct vhost_work *work) |
142 | { |
143 | unsigned long flags; |
144 | |
145 | spin_lock_irqsave(&dev->work_lock, flags); |
146 | if (list_empty(&work->node)) { |
147 | list_add_tail(&work->node, &dev->work_list); |
148 | work->queue_seq++; |
149 | wake_up_process(dev->worker); |
150 | } |
151 | spin_unlock_irqrestore(&dev->work_lock, flags); |
152 | } |
153 | |
154 | void vhost_poll_queue(struct vhost_poll *poll) |
155 | { |
156 | vhost_work_queue(poll->dev, &poll->work); |
157 | } |
158 | |
159 | static void vhost_vq_reset(struct vhost_dev *dev, |
160 | struct vhost_virtqueue *vq) |
161 | { |
162 | vq->num = 1; |
163 | vq->desc = NULL; |
164 | vq->avail = NULL; |
165 | vq->used = NULL; |
166 | vq->last_avail_idx = 0; |
167 | vq->avail_idx = 0; |
168 | vq->last_used_idx = 0; |
169 | vq->signalled_used = 0; |
170 | vq->signalled_used_valid = false; |
171 | vq->used_flags = 0; |
172 | vq->log_used = false; |
173 | vq->log_addr = -1ull; |
174 | vq->vhost_hlen = 0; |
175 | vq->sock_hlen = 0; |
176 | vq->private_data = NULL; |
177 | vq->log_base = NULL; |
178 | vq->error_ctx = NULL; |
179 | vq->error = NULL; |
180 | vq->kick = NULL; |
181 | vq->call_ctx = NULL; |
182 | vq->call = NULL; |
183 | vq->log_ctx = NULL; |
184 | vq->upend_idx = 0; |
185 | vq->done_idx = 0; |
186 | vq->ubufs = NULL; |
187 | } |
188 | |
189 | static int vhost_worker(void *data) |
190 | { |
191 | struct vhost_dev *dev = data; |
192 | struct vhost_work *work = NULL; |
193 | unsigned uninitialized_var(seq); |
194 | |
195 | use_mm(dev->mm); |
196 | |
197 | for (;;) { |
198 | /* mb paired w/ kthread_stop */ |
199 | set_current_state(TASK_INTERRUPTIBLE); |
200 | |
201 | spin_lock_irq(&dev->work_lock); |
202 | if (work) { |
203 | work->done_seq = seq; |
204 | if (work->flushing) |
205 | wake_up_all(&work->done); |
206 | } |
207 | |
208 | if (kthread_should_stop()) { |
209 | spin_unlock_irq(&dev->work_lock); |
210 | __set_current_state(TASK_RUNNING); |
211 | break; |
212 | } |
213 | if (!list_empty(&dev->work_list)) { |
214 | work = list_first_entry(&dev->work_list, |
215 | struct vhost_work, node); |
216 | list_del_init(&work->node); |
217 | seq = work->queue_seq; |
218 | } else |
219 | work = NULL; |
220 | spin_unlock_irq(&dev->work_lock); |
221 | |
222 | if (work) { |
223 | __set_current_state(TASK_RUNNING); |
224 | work->fn(work); |
225 | if (need_resched()) |
226 | schedule(); |
227 | } else |
228 | schedule(); |
229 | |
230 | } |
231 | unuse_mm(dev->mm); |
232 | return 0; |
233 | } |
234 | |
235 | static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq) |
236 | { |
237 | kfree(vq->indirect); |
238 | vq->indirect = NULL; |
239 | kfree(vq->log); |
240 | vq->log = NULL; |
241 | kfree(vq->heads); |
242 | vq->heads = NULL; |
243 | kfree(vq->ubuf_info); |
244 | vq->ubuf_info = NULL; |
245 | } |
246 | |
247 | void vhost_enable_zcopy(int vq) |
248 | { |
249 | vhost_zcopy_mask |= 0x1 << vq; |
250 | } |
251 | |
252 | /* Helper to allocate iovec buffers for all vqs. */ |
253 | static long vhost_dev_alloc_iovecs(struct vhost_dev *dev) |
254 | { |
255 | int i; |
256 | bool zcopy; |
257 | |
258 | for (i = 0; i < dev->nvqs; ++i) { |
259 | dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect * |
260 | UIO_MAXIOV, GFP_KERNEL); |
261 | dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV, |
262 | GFP_KERNEL); |
263 | dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads * |
264 | UIO_MAXIOV, GFP_KERNEL); |
265 | zcopy = vhost_zcopy_mask & (0x1 << i); |
266 | if (zcopy) |
267 | dev->vqs[i].ubuf_info = |
268 | kmalloc(sizeof *dev->vqs[i].ubuf_info * |
269 | UIO_MAXIOV, GFP_KERNEL); |
270 | if (!dev->vqs[i].indirect || !dev->vqs[i].log || |
271 | !dev->vqs[i].heads || |
272 | (zcopy && !dev->vqs[i].ubuf_info)) |
273 | goto err_nomem; |
274 | } |
275 | return 0; |
276 | |
277 | err_nomem: |
278 | for (; i >= 0; --i) |
279 | vhost_vq_free_iovecs(&dev->vqs[i]); |
280 | return -ENOMEM; |
281 | } |
282 | |
283 | static void vhost_dev_free_iovecs(struct vhost_dev *dev) |
284 | { |
285 | int i; |
286 | |
287 | for (i = 0; i < dev->nvqs; ++i) |
288 | vhost_vq_free_iovecs(&dev->vqs[i]); |
289 | } |
290 | |
291 | long vhost_dev_init(struct vhost_dev *dev, |
292 | struct vhost_virtqueue *vqs, int nvqs) |
293 | { |
294 | int i; |
295 | |
296 | dev->vqs = vqs; |
297 | dev->nvqs = nvqs; |
298 | mutex_init(&dev->mutex); |
299 | dev->log_ctx = NULL; |
300 | dev->log_file = NULL; |
301 | dev->memory = NULL; |
302 | dev->mm = NULL; |
303 | spin_lock_init(&dev->work_lock); |
304 | INIT_LIST_HEAD(&dev->work_list); |
305 | dev->worker = NULL; |
306 | |
307 | for (i = 0; i < dev->nvqs; ++i) { |
308 | dev->vqs[i].log = NULL; |
309 | dev->vqs[i].indirect = NULL; |
310 | dev->vqs[i].heads = NULL; |
311 | dev->vqs[i].ubuf_info = NULL; |
312 | dev->vqs[i].dev = dev; |
313 | mutex_init(&dev->vqs[i].mutex); |
314 | vhost_vq_reset(dev, dev->vqs + i); |
315 | if (dev->vqs[i].handle_kick) |
316 | vhost_poll_init(&dev->vqs[i].poll, |
317 | dev->vqs[i].handle_kick, POLLIN, dev); |
318 | } |
319 | |
320 | return 0; |
321 | } |
322 | |
323 | /* Caller should have device mutex */ |
324 | long vhost_dev_check_owner(struct vhost_dev *dev) |
325 | { |
326 | /* Are you the owner? If not, I don't think you mean to do that */ |
327 | return dev->mm == current->mm ? 0 : -EPERM; |
328 | } |
329 | |
330 | struct vhost_attach_cgroups_struct { |
331 | struct vhost_work work; |
332 | struct task_struct *owner; |
333 | int ret; |
334 | }; |
335 | |
336 | static void vhost_attach_cgroups_work(struct vhost_work *work) |
337 | { |
338 | struct vhost_attach_cgroups_struct *s; |
339 | |
340 | s = container_of(work, struct vhost_attach_cgroups_struct, work); |
341 | s->ret = cgroup_attach_task_all(s->owner, current); |
342 | } |
343 | |
344 | static int vhost_attach_cgroups(struct vhost_dev *dev) |
345 | { |
346 | struct vhost_attach_cgroups_struct attach; |
347 | |
348 | attach.owner = current; |
349 | vhost_work_init(&attach.work, vhost_attach_cgroups_work); |
350 | vhost_work_queue(dev, &attach.work); |
351 | vhost_work_flush(dev, &attach.work); |
352 | return attach.ret; |
353 | } |
354 | |
355 | /* Caller should have device mutex */ |
356 | static long vhost_dev_set_owner(struct vhost_dev *dev) |
357 | { |
358 | struct task_struct *worker; |
359 | int err; |
360 | |
361 | /* Is there an owner already? */ |
362 | if (dev->mm) { |
363 | err = -EBUSY; |
364 | goto err_mm; |
365 | } |
366 | |
367 | /* No owner, become one */ |
368 | dev->mm = get_task_mm(current); |
369 | worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid); |
370 | if (IS_ERR(worker)) { |
371 | err = PTR_ERR(worker); |
372 | goto err_worker; |
373 | } |
374 | |
375 | dev->worker = worker; |
376 | wake_up_process(worker); /* avoid contributing to loadavg */ |
377 | |
378 | err = vhost_attach_cgroups(dev); |
379 | if (err) |
380 | goto err_cgroup; |
381 | |
382 | err = vhost_dev_alloc_iovecs(dev); |
383 | if (err) |
384 | goto err_cgroup; |
385 | |
386 | return 0; |
387 | err_cgroup: |
388 | kthread_stop(worker); |
389 | dev->worker = NULL; |
390 | err_worker: |
391 | if (dev->mm) |
392 | mmput(dev->mm); |
393 | dev->mm = NULL; |
394 | err_mm: |
395 | return err; |
396 | } |
397 | |
398 | /* Caller should have device mutex */ |
399 | long vhost_dev_reset_owner(struct vhost_dev *dev) |
400 | { |
401 | struct vhost_memory *memory; |
402 | |
403 | /* Restore memory to default empty mapping. */ |
404 | memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL); |
405 | if (!memory) |
406 | return -ENOMEM; |
407 | |
408 | vhost_dev_cleanup(dev, true); |
409 | |
410 | memory->nregions = 0; |
411 | RCU_INIT_POINTER(dev->memory, memory); |
412 | return 0; |
413 | } |
414 | |
415 | /* In case of DMA done not in order in lower device driver for some reason. |
416 | * upend_idx is used to track end of used idx, done_idx is used to track head |
417 | * of used idx. Once lower device DMA done contiguously, we will signal KVM |
418 | * guest used idx. |
419 | */ |
420 | int vhost_zerocopy_signal_used(struct vhost_virtqueue *vq) |
421 | { |
422 | int i; |
423 | int j = 0; |
424 | |
425 | for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) { |
426 | if ((vq->heads[i].len == VHOST_DMA_DONE_LEN)) { |
427 | vq->heads[i].len = VHOST_DMA_CLEAR_LEN; |
428 | vhost_add_used_and_signal(vq->dev, vq, |
429 | vq->heads[i].id, 0); |
430 | ++j; |
431 | } else |
432 | break; |
433 | } |
434 | if (j) |
435 | vq->done_idx = i; |
436 | return j; |
437 | } |
438 | |
439 | /* Caller should have device mutex if and only if locked is set */ |
440 | void vhost_dev_cleanup(struct vhost_dev *dev, bool locked) |
441 | { |
442 | int i; |
443 | |
444 | for (i = 0; i < dev->nvqs; ++i) { |
445 | if (dev->vqs[i].kick && dev->vqs[i].handle_kick) { |
446 | vhost_poll_stop(&dev->vqs[i].poll); |
447 | vhost_poll_flush(&dev->vqs[i].poll); |
448 | } |
449 | /* Wait for all lower device DMAs done. */ |
450 | if (dev->vqs[i].ubufs) |
451 | vhost_ubuf_put_and_wait(dev->vqs[i].ubufs); |
452 | |
453 | /* Signal guest as appropriate. */ |
454 | vhost_zerocopy_signal_used(&dev->vqs[i]); |
455 | |
456 | if (dev->vqs[i].error_ctx) |
457 | eventfd_ctx_put(dev->vqs[i].error_ctx); |
458 | if (dev->vqs[i].error) |
459 | fput(dev->vqs[i].error); |
460 | if (dev->vqs[i].kick) |
461 | fput(dev->vqs[i].kick); |
462 | if (dev->vqs[i].call_ctx) |
463 | eventfd_ctx_put(dev->vqs[i].call_ctx); |
464 | if (dev->vqs[i].call) |
465 | fput(dev->vqs[i].call); |
466 | vhost_vq_reset(dev, dev->vqs + i); |
467 | } |
468 | vhost_dev_free_iovecs(dev); |
469 | if (dev->log_ctx) |
470 | eventfd_ctx_put(dev->log_ctx); |
471 | dev->log_ctx = NULL; |
472 | if (dev->log_file) |
473 | fput(dev->log_file); |
474 | dev->log_file = NULL; |
475 | /* No one will access memory at this point */ |
476 | kfree(rcu_dereference_protected(dev->memory, |
477 | locked == |
478 | lockdep_is_held(&dev->mutex))); |
479 | RCU_INIT_POINTER(dev->memory, NULL); |
480 | WARN_ON(!list_empty(&dev->work_list)); |
481 | if (dev->worker) { |
482 | kthread_stop(dev->worker); |
483 | dev->worker = NULL; |
484 | } |
485 | if (dev->mm) |
486 | mmput(dev->mm); |
487 | dev->mm = NULL; |
488 | } |
489 | |
490 | static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz) |
491 | { |
492 | u64 a = addr / VHOST_PAGE_SIZE / 8; |
493 | |
494 | /* Make sure 64 bit math will not overflow. */ |
495 | if (a > ULONG_MAX - (unsigned long)log_base || |
496 | a + (unsigned long)log_base > ULONG_MAX) |
497 | return 0; |
498 | |
499 | return access_ok(VERIFY_WRITE, log_base + a, |
500 | (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8); |
501 | } |
502 | |
503 | /* Caller should have vq mutex and device mutex. */ |
504 | static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem, |
505 | int log_all) |
506 | { |
507 | int i; |
508 | |
509 | if (!mem) |
510 | return 0; |
511 | |
512 | for (i = 0; i < mem->nregions; ++i) { |
513 | struct vhost_memory_region *m = mem->regions + i; |
514 | unsigned long a = m->userspace_addr; |
515 | if (m->memory_size > ULONG_MAX) |
516 | return 0; |
517 | else if (!access_ok(VERIFY_WRITE, (void __user *)a, |
518 | m->memory_size)) |
519 | return 0; |
520 | else if (log_all && !log_access_ok(log_base, |
521 | m->guest_phys_addr, |
522 | m->memory_size)) |
523 | return 0; |
524 | } |
525 | return 1; |
526 | } |
527 | |
528 | /* Can we switch to this memory table? */ |
529 | /* Caller should have device mutex but not vq mutex */ |
530 | static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem, |
531 | int log_all) |
532 | { |
533 | int i; |
534 | |
535 | for (i = 0; i < d->nvqs; ++i) { |
536 | int ok; |
537 | mutex_lock(&d->vqs[i].mutex); |
538 | /* If ring is inactive, will check when it's enabled. */ |
539 | if (d->vqs[i].private_data) |
540 | ok = vq_memory_access_ok(d->vqs[i].log_base, mem, |
541 | log_all); |
542 | else |
543 | ok = 1; |
544 | mutex_unlock(&d->vqs[i].mutex); |
545 | if (!ok) |
546 | return 0; |
547 | } |
548 | return 1; |
549 | } |
550 | |
551 | static int vq_access_ok(struct vhost_dev *d, unsigned int num, |
552 | struct vring_desc __user *desc, |
553 | struct vring_avail __user *avail, |
554 | struct vring_used __user *used) |
555 | { |
556 | size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0; |
557 | return access_ok(VERIFY_READ, desc, num * sizeof *desc) && |
558 | access_ok(VERIFY_READ, avail, |
559 | sizeof *avail + num * sizeof *avail->ring + s) && |
560 | access_ok(VERIFY_WRITE, used, |
561 | sizeof *used + num * sizeof *used->ring + s); |
562 | } |
563 | |
564 | /* Can we log writes? */ |
565 | /* Caller should have device mutex but not vq mutex */ |
566 | int vhost_log_access_ok(struct vhost_dev *dev) |
567 | { |
568 | struct vhost_memory *mp; |
569 | |
570 | mp = rcu_dereference_protected(dev->memory, |
571 | lockdep_is_held(&dev->mutex)); |
572 | return memory_access_ok(dev, mp, 1); |
573 | } |
574 | |
575 | /* Verify access for write logging. */ |
576 | /* Caller should have vq mutex and device mutex */ |
577 | static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq, |
578 | void __user *log_base) |
579 | { |
580 | struct vhost_memory *mp; |
581 | size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0; |
582 | |
583 | mp = rcu_dereference_protected(vq->dev->memory, |
584 | lockdep_is_held(&vq->mutex)); |
585 | return vq_memory_access_ok(log_base, mp, |
586 | vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) && |
587 | (!vq->log_used || log_access_ok(log_base, vq->log_addr, |
588 | sizeof *vq->used + |
589 | vq->num * sizeof *vq->used->ring + s)); |
590 | } |
591 | |
592 | /* Can we start vq? */ |
593 | /* Caller should have vq mutex and device mutex */ |
594 | int vhost_vq_access_ok(struct vhost_virtqueue *vq) |
595 | { |
596 | return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) && |
597 | vq_log_access_ok(vq->dev, vq, vq->log_base); |
598 | } |
599 | |
600 | static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m) |
601 | { |
602 | struct vhost_memory mem, *newmem, *oldmem; |
603 | unsigned long size = offsetof(struct vhost_memory, regions); |
604 | |
605 | if (copy_from_user(&mem, m, size)) |
606 | return -EFAULT; |
607 | if (mem.padding) |
608 | return -EOPNOTSUPP; |
609 | if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS) |
610 | return -E2BIG; |
611 | newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL); |
612 | if (!newmem) |
613 | return -ENOMEM; |
614 | |
615 | memcpy(newmem, &mem, size); |
616 | if (copy_from_user(newmem->regions, m->regions, |
617 | mem.nregions * sizeof *m->regions)) { |
618 | kfree(newmem); |
619 | return -EFAULT; |
620 | } |
621 | |
622 | if (!memory_access_ok(d, newmem, |
623 | vhost_has_feature(d, VHOST_F_LOG_ALL))) { |
624 | kfree(newmem); |
625 | return -EFAULT; |
626 | } |
627 | oldmem = rcu_dereference_protected(d->memory, |
628 | lockdep_is_held(&d->mutex)); |
629 | rcu_assign_pointer(d->memory, newmem); |
630 | synchronize_rcu(); |
631 | kfree(oldmem); |
632 | return 0; |
633 | } |
634 | |
635 | static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp) |
636 | { |
637 | struct file *eventfp, *filep = NULL, |
638 | *pollstart = NULL, *pollstop = NULL; |
639 | struct eventfd_ctx *ctx = NULL; |
640 | u32 __user *idxp = argp; |
641 | struct vhost_virtqueue *vq; |
642 | struct vhost_vring_state s; |
643 | struct vhost_vring_file f; |
644 | struct vhost_vring_addr a; |
645 | u32 idx; |
646 | long r; |
647 | |
648 | r = get_user(idx, idxp); |
649 | if (r < 0) |
650 | return r; |
651 | if (idx >= d->nvqs) |
652 | return -ENOBUFS; |
653 | |
654 | vq = d->vqs + idx; |
655 | |
656 | mutex_lock(&vq->mutex); |
657 | |
658 | switch (ioctl) { |
659 | case VHOST_SET_VRING_NUM: |
660 | /* Resizing ring with an active backend? |
661 | * You don't want to do that. */ |
662 | if (vq->private_data) { |
663 | r = -EBUSY; |
664 | break; |
665 | } |
666 | if (copy_from_user(&s, argp, sizeof s)) { |
667 | r = -EFAULT; |
668 | break; |
669 | } |
670 | if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) { |
671 | r = -EINVAL; |
672 | break; |
673 | } |
674 | vq->num = s.num; |
675 | break; |
676 | case VHOST_SET_VRING_BASE: |
677 | /* Moving base with an active backend? |
678 | * You don't want to do that. */ |
679 | if (vq->private_data) { |
680 | r = -EBUSY; |
681 | break; |
682 | } |
683 | if (copy_from_user(&s, argp, sizeof s)) { |
684 | r = -EFAULT; |
685 | break; |
686 | } |
687 | if (s.num > 0xffff) { |
688 | r = -EINVAL; |
689 | break; |
690 | } |
691 | vq->last_avail_idx = s.num; |
692 | /* Forget the cached index value. */ |
693 | vq->avail_idx = vq->last_avail_idx; |
694 | break; |
695 | case VHOST_GET_VRING_BASE: |
696 | s.index = idx; |
697 | s.num = vq->last_avail_idx; |
698 | if (copy_to_user(argp, &s, sizeof s)) |
699 | r = -EFAULT; |
700 | break; |
701 | case VHOST_SET_VRING_ADDR: |
702 | if (copy_from_user(&a, argp, sizeof a)) { |
703 | r = -EFAULT; |
704 | break; |
705 | } |
706 | if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) { |
707 | r = -EOPNOTSUPP; |
708 | break; |
709 | } |
710 | /* For 32bit, verify that the top 32bits of the user |
711 | data are set to zero. */ |
712 | if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr || |
713 | (u64)(unsigned long)a.used_user_addr != a.used_user_addr || |
714 | (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) { |
715 | r = -EFAULT; |
716 | break; |
717 | } |
718 | if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) || |
719 | (a.used_user_addr & (sizeof *vq->used->ring - 1)) || |
720 | (a.log_guest_addr & (sizeof *vq->used->ring - 1))) { |
721 | r = -EINVAL; |
722 | break; |
723 | } |
724 | |
725 | /* We only verify access here if backend is configured. |
726 | * If it is not, we don't as size might not have been setup. |
727 | * We will verify when backend is configured. */ |
728 | if (vq->private_data) { |
729 | if (!vq_access_ok(d, vq->num, |
730 | (void __user *)(unsigned long)a.desc_user_addr, |
731 | (void __user *)(unsigned long)a.avail_user_addr, |
732 | (void __user *)(unsigned long)a.used_user_addr)) { |
733 | r = -EINVAL; |
734 | break; |
735 | } |
736 | |
737 | /* Also validate log access for used ring if enabled. */ |
738 | if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) && |
739 | !log_access_ok(vq->log_base, a.log_guest_addr, |
740 | sizeof *vq->used + |
741 | vq->num * sizeof *vq->used->ring)) { |
742 | r = -EINVAL; |
743 | break; |
744 | } |
745 | } |
746 | |
747 | vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG)); |
748 | vq->desc = (void __user *)(unsigned long)a.desc_user_addr; |
749 | vq->avail = (void __user *)(unsigned long)a.avail_user_addr; |
750 | vq->log_addr = a.log_guest_addr; |
751 | vq->used = (void __user *)(unsigned long)a.used_user_addr; |
752 | break; |
753 | case VHOST_SET_VRING_KICK: |
754 | if (copy_from_user(&f, argp, sizeof f)) { |
755 | r = -EFAULT; |
756 | break; |
757 | } |
758 | eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); |
759 | if (IS_ERR(eventfp)) { |
760 | r = PTR_ERR(eventfp); |
761 | break; |
762 | } |
763 | if (eventfp != vq->kick) { |
764 | pollstop = filep = vq->kick; |
765 | pollstart = vq->kick = eventfp; |
766 | } else |
767 | filep = eventfp; |
768 | break; |
769 | case VHOST_SET_VRING_CALL: |
770 | if (copy_from_user(&f, argp, sizeof f)) { |
771 | r = -EFAULT; |
772 | break; |
773 | } |
774 | eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); |
775 | if (IS_ERR(eventfp)) { |
776 | r = PTR_ERR(eventfp); |
777 | break; |
778 | } |
779 | if (eventfp != vq->call) { |
780 | filep = vq->call; |
781 | ctx = vq->call_ctx; |
782 | vq->call = eventfp; |
783 | vq->call_ctx = eventfp ? |
784 | eventfd_ctx_fileget(eventfp) : NULL; |
785 | } else |
786 | filep = eventfp; |
787 | break; |
788 | case VHOST_SET_VRING_ERR: |
789 | if (copy_from_user(&f, argp, sizeof f)) { |
790 | r = -EFAULT; |
791 | break; |
792 | } |
793 | eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); |
794 | if (IS_ERR(eventfp)) { |
795 | r = PTR_ERR(eventfp); |
796 | break; |
797 | } |
798 | if (eventfp != vq->error) { |
799 | filep = vq->error; |
800 | vq->error = eventfp; |
801 | ctx = vq->error_ctx; |
802 | vq->error_ctx = eventfp ? |
803 | eventfd_ctx_fileget(eventfp) : NULL; |
804 | } else |
805 | filep = eventfp; |
806 | break; |
807 | default: |
808 | r = -ENOIOCTLCMD; |
809 | } |
810 | |
811 | if (pollstop && vq->handle_kick) |
812 | vhost_poll_stop(&vq->poll); |
813 | |
814 | if (ctx) |
815 | eventfd_ctx_put(ctx); |
816 | if (filep) |
817 | fput(filep); |
818 | |
819 | if (pollstart && vq->handle_kick) |
820 | vhost_poll_start(&vq->poll, vq->kick); |
821 | |
822 | mutex_unlock(&vq->mutex); |
823 | |
824 | if (pollstop && vq->handle_kick) |
825 | vhost_poll_flush(&vq->poll); |
826 | return r; |
827 | } |
828 | |
829 | /* Caller must have device mutex */ |
830 | long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg) |
831 | { |
832 | void __user *argp = (void __user *)arg; |
833 | struct file *eventfp, *filep = NULL; |
834 | struct eventfd_ctx *ctx = NULL; |
835 | u64 p; |
836 | long r; |
837 | int i, fd; |
838 | |
839 | /* If you are not the owner, you can become one */ |
840 | if (ioctl == VHOST_SET_OWNER) { |
841 | r = vhost_dev_set_owner(d); |
842 | goto done; |
843 | } |
844 | |
845 | /* You must be the owner to do anything else */ |
846 | r = vhost_dev_check_owner(d); |
847 | if (r) |
848 | goto done; |
849 | |
850 | switch (ioctl) { |
851 | case VHOST_SET_MEM_TABLE: |
852 | r = vhost_set_memory(d, argp); |
853 | break; |
854 | case VHOST_SET_LOG_BASE: |
855 | if (copy_from_user(&p, argp, sizeof p)) { |
856 | r = -EFAULT; |
857 | break; |
858 | } |
859 | if ((u64)(unsigned long)p != p) { |
860 | r = -EFAULT; |
861 | break; |
862 | } |
863 | for (i = 0; i < d->nvqs; ++i) { |
864 | struct vhost_virtqueue *vq; |
865 | void __user *base = (void __user *)(unsigned long)p; |
866 | vq = d->vqs + i; |
867 | mutex_lock(&vq->mutex); |
868 | /* If ring is inactive, will check when it's enabled. */ |
869 | if (vq->private_data && !vq_log_access_ok(d, vq, base)) |
870 | r = -EFAULT; |
871 | else |
872 | vq->log_base = base; |
873 | mutex_unlock(&vq->mutex); |
874 | } |
875 | break; |
876 | case VHOST_SET_LOG_FD: |
877 | r = get_user(fd, (int __user *)argp); |
878 | if (r < 0) |
879 | break; |
880 | eventfp = fd == -1 ? NULL : eventfd_fget(fd); |
881 | if (IS_ERR(eventfp)) { |
882 | r = PTR_ERR(eventfp); |
883 | break; |
884 | } |
885 | if (eventfp != d->log_file) { |
886 | filep = d->log_file; |
887 | ctx = d->log_ctx; |
888 | d->log_ctx = eventfp ? |
889 | eventfd_ctx_fileget(eventfp) : NULL; |
890 | } else |
891 | filep = eventfp; |
892 | for (i = 0; i < d->nvqs; ++i) { |
893 | mutex_lock(&d->vqs[i].mutex); |
894 | d->vqs[i].log_ctx = d->log_ctx; |
895 | mutex_unlock(&d->vqs[i].mutex); |
896 | } |
897 | if (ctx) |
898 | eventfd_ctx_put(ctx); |
899 | if (filep) |
900 | fput(filep); |
901 | break; |
902 | default: |
903 | r = vhost_set_vring(d, ioctl, argp); |
904 | break; |
905 | } |
906 | done: |
907 | return r; |
908 | } |
909 | |
910 | static const struct vhost_memory_region *find_region(struct vhost_memory *mem, |
911 | __u64 addr, __u32 len) |
912 | { |
913 | struct vhost_memory_region *reg; |
914 | int i; |
915 | |
916 | /* linear search is not brilliant, but we really have on the order of 6 |
917 | * regions in practice */ |
918 | for (i = 0; i < mem->nregions; ++i) { |
919 | reg = mem->regions + i; |
920 | if (reg->guest_phys_addr <= addr && |
921 | reg->guest_phys_addr + reg->memory_size - 1 >= addr) |
922 | return reg; |
923 | } |
924 | return NULL; |
925 | } |
926 | |
927 | /* TODO: This is really inefficient. We need something like get_user() |
928 | * (instruction directly accesses the data, with an exception table entry |
929 | * returning -EFAULT). See Documentation/x86/exception-tables.txt. |
930 | */ |
931 | static int set_bit_to_user(int nr, void __user *addr) |
932 | { |
933 | unsigned long log = (unsigned long)addr; |
934 | struct page *page; |
935 | void *base; |
936 | int bit = nr + (log % PAGE_SIZE) * 8; |
937 | int r; |
938 | |
939 | r = get_user_pages_fast(log, 1, 1, &page); |
940 | if (r < 0) |
941 | return r; |
942 | BUG_ON(r != 1); |
943 | base = kmap_atomic(page); |
944 | set_bit(bit, base); |
945 | kunmap_atomic(base); |
946 | set_page_dirty_lock(page); |
947 | put_page(page); |
948 | return 0; |
949 | } |
950 | |
951 | static int log_write(void __user *log_base, |
952 | u64 write_address, u64 write_length) |
953 | { |
954 | u64 write_page = write_address / VHOST_PAGE_SIZE; |
955 | int r; |
956 | |
957 | if (!write_length) |
958 | return 0; |
959 | write_length += write_address % VHOST_PAGE_SIZE; |
960 | for (;;) { |
961 | u64 base = (u64)(unsigned long)log_base; |
962 | u64 log = base + write_page / 8; |
963 | int bit = write_page % 8; |
964 | if ((u64)(unsigned long)log != log) |
965 | return -EFAULT; |
966 | r = set_bit_to_user(bit, (void __user *)(unsigned long)log); |
967 | if (r < 0) |
968 | return r; |
969 | if (write_length <= VHOST_PAGE_SIZE) |
970 | break; |
971 | write_length -= VHOST_PAGE_SIZE; |
972 | write_page += 1; |
973 | } |
974 | return r; |
975 | } |
976 | |
977 | int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log, |
978 | unsigned int log_num, u64 len) |
979 | { |
980 | int i, r; |
981 | |
982 | /* Make sure data written is seen before log. */ |
983 | smp_wmb(); |
984 | for (i = 0; i < log_num; ++i) { |
985 | u64 l = min(log[i].len, len); |
986 | r = log_write(vq->log_base, log[i].addr, l); |
987 | if (r < 0) |
988 | return r; |
989 | len -= l; |
990 | if (!len) { |
991 | if (vq->log_ctx) |
992 | eventfd_signal(vq->log_ctx, 1); |
993 | return 0; |
994 | } |
995 | } |
996 | /* Length written exceeds what we have stored. This is a bug. */ |
997 | BUG(); |
998 | return 0; |
999 | } |
1000 | |
1001 | static int vhost_update_used_flags(struct vhost_virtqueue *vq) |
1002 | { |
1003 | void __user *used; |
1004 | if (__put_user(vq->used_flags, &vq->used->flags) < 0) |
1005 | return -EFAULT; |
1006 | if (unlikely(vq->log_used)) { |
1007 | /* Make sure the flag is seen before log. */ |
1008 | smp_wmb(); |
1009 | /* Log used flag write. */ |
1010 | used = &vq->used->flags; |
1011 | log_write(vq->log_base, vq->log_addr + |
1012 | (used - (void __user *)vq->used), |
1013 | sizeof vq->used->flags); |
1014 | if (vq->log_ctx) |
1015 | eventfd_signal(vq->log_ctx, 1); |
1016 | } |
1017 | return 0; |
1018 | } |
1019 | |
1020 | static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event) |
1021 | { |
1022 | if (__put_user(vq->avail_idx, vhost_avail_event(vq))) |
1023 | return -EFAULT; |
1024 | if (unlikely(vq->log_used)) { |
1025 | void __user *used; |
1026 | /* Make sure the event is seen before log. */ |
1027 | smp_wmb(); |
1028 | /* Log avail event write */ |
1029 | used = vhost_avail_event(vq); |
1030 | log_write(vq->log_base, vq->log_addr + |
1031 | (used - (void __user *)vq->used), |
1032 | sizeof *vhost_avail_event(vq)); |
1033 | if (vq->log_ctx) |
1034 | eventfd_signal(vq->log_ctx, 1); |
1035 | } |
1036 | return 0; |
1037 | } |
1038 | |
1039 | int vhost_init_used(struct vhost_virtqueue *vq) |
1040 | { |
1041 | int r; |
1042 | if (!vq->private_data) |
1043 | return 0; |
1044 | |
1045 | r = vhost_update_used_flags(vq); |
1046 | if (r) |
1047 | return r; |
1048 | vq->signalled_used_valid = false; |
1049 | return get_user(vq->last_used_idx, &vq->used->idx); |
1050 | } |
1051 | |
1052 | static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len, |
1053 | struct iovec iov[], int iov_size) |
1054 | { |
1055 | const struct vhost_memory_region *reg; |
1056 | struct vhost_memory *mem; |
1057 | struct iovec *_iov; |
1058 | u64 s = 0; |
1059 | int ret = 0; |
1060 | |
1061 | rcu_read_lock(); |
1062 | |
1063 | mem = rcu_dereference(dev->memory); |
1064 | while ((u64)len > s) { |
1065 | u64 size; |
1066 | if (unlikely(ret >= iov_size)) { |
1067 | ret = -ENOBUFS; |
1068 | break; |
1069 | } |
1070 | reg = find_region(mem, addr, len); |
1071 | if (unlikely(!reg)) { |
1072 | ret = -EFAULT; |
1073 | break; |
1074 | } |
1075 | _iov = iov + ret; |
1076 | size = reg->memory_size - addr + reg->guest_phys_addr; |
1077 | _iov->iov_len = min((u64)len, size); |
1078 | _iov->iov_base = (void __user *)(unsigned long) |
1079 | (reg->userspace_addr + addr - reg->guest_phys_addr); |
1080 | s += size; |
1081 | addr += size; |
1082 | ++ret; |
1083 | } |
1084 | |
1085 | rcu_read_unlock(); |
1086 | return ret; |
1087 | } |
1088 | |
1089 | /* Each buffer in the virtqueues is actually a chain of descriptors. This |
1090 | * function returns the next descriptor in the chain, |
1091 | * or -1U if we're at the end. */ |
1092 | static unsigned next_desc(struct vring_desc *desc) |
1093 | { |
1094 | unsigned int next; |
1095 | |
1096 | /* If this descriptor says it doesn't chain, we're done. */ |
1097 | if (!(desc->flags & VRING_DESC_F_NEXT)) |
1098 | return -1U; |
1099 | |
1100 | /* Check they're not leading us off end of descriptors. */ |
1101 | next = desc->next; |
1102 | /* Make sure compiler knows to grab that: we don't want it changing! */ |
1103 | /* We will use the result as an index in an array, so most |
1104 | * architectures only need a compiler barrier here. */ |
1105 | read_barrier_depends(); |
1106 | |
1107 | return next; |
1108 | } |
1109 | |
1110 | static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq, |
1111 | struct iovec iov[], unsigned int iov_size, |
1112 | unsigned int *out_num, unsigned int *in_num, |
1113 | struct vhost_log *log, unsigned int *log_num, |
1114 | struct vring_desc *indirect) |
1115 | { |
1116 | struct vring_desc desc; |
1117 | unsigned int i = 0, count, found = 0; |
1118 | int ret; |
1119 | |
1120 | /* Sanity check */ |
1121 | if (unlikely(indirect->len % sizeof desc)) { |
1122 | vq_err(vq, "Invalid length in indirect descriptor: " |
1123 | "len 0x%llx not multiple of 0x%zx\n", |
1124 | (unsigned long long)indirect->len, |
1125 | sizeof desc); |
1126 | return -EINVAL; |
1127 | } |
1128 | |
1129 | ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect, |
1130 | UIO_MAXIOV); |
1131 | if (unlikely(ret < 0)) { |
1132 | vq_err(vq, "Translation failure %d in indirect.\n", ret); |
1133 | return ret; |
1134 | } |
1135 | |
1136 | /* We will use the result as an address to read from, so most |
1137 | * architectures only need a compiler barrier here. */ |
1138 | read_barrier_depends(); |
1139 | |
1140 | count = indirect->len / sizeof desc; |
1141 | /* Buffers are chained via a 16 bit next field, so |
1142 | * we can have at most 2^16 of these. */ |
1143 | if (unlikely(count > USHRT_MAX + 1)) { |
1144 | vq_err(vq, "Indirect buffer length too big: %d\n", |
1145 | indirect->len); |
1146 | return -E2BIG; |
1147 | } |
1148 | |
1149 | do { |
1150 | unsigned iov_count = *in_num + *out_num; |
1151 | if (unlikely(++found > count)) { |
1152 | vq_err(vq, "Loop detected: last one at %u " |
1153 | "indirect size %u\n", |
1154 | i, count); |
1155 | return -EINVAL; |
1156 | } |
1157 | if (unlikely(memcpy_fromiovec((unsigned char *)&desc, |
1158 | vq->indirect, sizeof desc))) { |
1159 | vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n", |
1160 | i, (size_t)indirect->addr + i * sizeof desc); |
1161 | return -EINVAL; |
1162 | } |
1163 | if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) { |
1164 | vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n", |
1165 | i, (size_t)indirect->addr + i * sizeof desc); |
1166 | return -EINVAL; |
1167 | } |
1168 | |
1169 | ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count, |
1170 | iov_size - iov_count); |
1171 | if (unlikely(ret < 0)) { |
1172 | vq_err(vq, "Translation failure %d indirect idx %d\n", |
1173 | ret, i); |
1174 | return ret; |
1175 | } |
1176 | /* If this is an input descriptor, increment that count. */ |
1177 | if (desc.flags & VRING_DESC_F_WRITE) { |
1178 | *in_num += ret; |
1179 | if (unlikely(log)) { |
1180 | log[*log_num].addr = desc.addr; |
1181 | log[*log_num].len = desc.len; |
1182 | ++*log_num; |
1183 | } |
1184 | } else { |
1185 | /* If it's an output descriptor, they're all supposed |
1186 | * to come before any input descriptors. */ |
1187 | if (unlikely(*in_num)) { |
1188 | vq_err(vq, "Indirect descriptor " |
1189 | "has out after in: idx %d\n", i); |
1190 | return -EINVAL; |
1191 | } |
1192 | *out_num += ret; |
1193 | } |
1194 | } while ((i = next_desc(&desc)) != -1); |
1195 | return 0; |
1196 | } |
1197 | |
1198 | /* This looks in the virtqueue and for the first available buffer, and converts |
1199 | * it to an iovec for convenient access. Since descriptors consist of some |
1200 | * number of output then some number of input descriptors, it's actually two |
1201 | * iovecs, but we pack them into one and note how many of each there were. |
1202 | * |
1203 | * This function returns the descriptor number found, or vq->num (which is |
1204 | * never a valid descriptor number) if none was found. A negative code is |
1205 | * returned on error. */ |
1206 | int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq, |
1207 | struct iovec iov[], unsigned int iov_size, |
1208 | unsigned int *out_num, unsigned int *in_num, |
1209 | struct vhost_log *log, unsigned int *log_num) |
1210 | { |
1211 | struct vring_desc desc; |
1212 | unsigned int i, head, found = 0; |
1213 | u16 last_avail_idx; |
1214 | int ret; |
1215 | |
1216 | /* Check it isn't doing very strange things with descriptor numbers. */ |
1217 | last_avail_idx = vq->last_avail_idx; |
1218 | if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) { |
1219 | vq_err(vq, "Failed to access avail idx at %p\n", |
1220 | &vq->avail->idx); |
1221 | return -EFAULT; |
1222 | } |
1223 | |
1224 | if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) { |
1225 | vq_err(vq, "Guest moved used index from %u to %u", |
1226 | last_avail_idx, vq->avail_idx); |
1227 | return -EFAULT; |
1228 | } |
1229 | |
1230 | /* If there's nothing new since last we looked, return invalid. */ |
1231 | if (vq->avail_idx == last_avail_idx) |
1232 | return vq->num; |
1233 | |
1234 | /* Only get avail ring entries after they have been exposed by guest. */ |
1235 | smp_rmb(); |
1236 | |
1237 | /* Grab the next descriptor number they're advertising, and increment |
1238 | * the index we've seen. */ |
1239 | if (unlikely(__get_user(head, |
1240 | &vq->avail->ring[last_avail_idx % vq->num]))) { |
1241 | vq_err(vq, "Failed to read head: idx %d address %p\n", |
1242 | last_avail_idx, |
1243 | &vq->avail->ring[last_avail_idx % vq->num]); |
1244 | return -EFAULT; |
1245 | } |
1246 | |
1247 | /* If their number is silly, that's an error. */ |
1248 | if (unlikely(head >= vq->num)) { |
1249 | vq_err(vq, "Guest says index %u > %u is available", |
1250 | head, vq->num); |
1251 | return -EINVAL; |
1252 | } |
1253 | |
1254 | /* When we start there are none of either input nor output. */ |
1255 | *out_num = *in_num = 0; |
1256 | if (unlikely(log)) |
1257 | *log_num = 0; |
1258 | |
1259 | i = head; |
1260 | do { |
1261 | unsigned iov_count = *in_num + *out_num; |
1262 | if (unlikely(i >= vq->num)) { |
1263 | vq_err(vq, "Desc index is %u > %u, head = %u", |
1264 | i, vq->num, head); |
1265 | return -EINVAL; |
1266 | } |
1267 | if (unlikely(++found > vq->num)) { |
1268 | vq_err(vq, "Loop detected: last one at %u " |
1269 | "vq size %u head %u\n", |
1270 | i, vq->num, head); |
1271 | return -EINVAL; |
1272 | } |
1273 | ret = __copy_from_user(&desc, vq->desc + i, sizeof desc); |
1274 | if (unlikely(ret)) { |
1275 | vq_err(vq, "Failed to get descriptor: idx %d addr %p\n", |
1276 | i, vq->desc + i); |
1277 | return -EFAULT; |
1278 | } |
1279 | if (desc.flags & VRING_DESC_F_INDIRECT) { |
1280 | ret = get_indirect(dev, vq, iov, iov_size, |
1281 | out_num, in_num, |
1282 | log, log_num, &desc); |
1283 | if (unlikely(ret < 0)) { |
1284 | vq_err(vq, "Failure detected " |
1285 | "in indirect descriptor at idx %d\n", i); |
1286 | return ret; |
1287 | } |
1288 | continue; |
1289 | } |
1290 | |
1291 | ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count, |
1292 | iov_size - iov_count); |
1293 | if (unlikely(ret < 0)) { |
1294 | vq_err(vq, "Translation failure %d descriptor idx %d\n", |
1295 | ret, i); |
1296 | return ret; |
1297 | } |
1298 | if (desc.flags & VRING_DESC_F_WRITE) { |
1299 | /* If this is an input descriptor, |
1300 | * increment that count. */ |
1301 | *in_num += ret; |
1302 | if (unlikely(log)) { |
1303 | log[*log_num].addr = desc.addr; |
1304 | log[*log_num].len = desc.len; |
1305 | ++*log_num; |
1306 | } |
1307 | } else { |
1308 | /* If it's an output descriptor, they're all supposed |
1309 | * to come before any input descriptors. */ |
1310 | if (unlikely(*in_num)) { |
1311 | vq_err(vq, "Descriptor has out after in: " |
1312 | "idx %d\n", i); |
1313 | return -EINVAL; |
1314 | } |
1315 | *out_num += ret; |
1316 | } |
1317 | } while ((i = next_desc(&desc)) != -1); |
1318 | |
1319 | /* On success, increment avail index. */ |
1320 | vq->last_avail_idx++; |
1321 | |
1322 | /* Assume notifications from guest are disabled at this point, |
1323 | * if they aren't we would need to update avail_event index. */ |
1324 | BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY)); |
1325 | return head; |
1326 | } |
1327 | |
1328 | /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */ |
1329 | void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n) |
1330 | { |
1331 | vq->last_avail_idx -= n; |
1332 | } |
1333 | |
1334 | /* After we've used one of their buffers, we tell them about it. We'll then |
1335 | * want to notify the guest, using eventfd. */ |
1336 | int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len) |
1337 | { |
1338 | struct vring_used_elem __user *used; |
1339 | |
1340 | /* The virtqueue contains a ring of used buffers. Get a pointer to the |
1341 | * next entry in that used ring. */ |
1342 | used = &vq->used->ring[vq->last_used_idx % vq->num]; |
1343 | if (__put_user(head, &used->id)) { |
1344 | vq_err(vq, "Failed to write used id"); |
1345 | return -EFAULT; |
1346 | } |
1347 | if (__put_user(len, &used->len)) { |
1348 | vq_err(vq, "Failed to write used len"); |
1349 | return -EFAULT; |
1350 | } |
1351 | /* Make sure buffer is written before we update index. */ |
1352 | smp_wmb(); |
1353 | if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) { |
1354 | vq_err(vq, "Failed to increment used idx"); |
1355 | return -EFAULT; |
1356 | } |
1357 | if (unlikely(vq->log_used)) { |
1358 | /* Make sure data is seen before log. */ |
1359 | smp_wmb(); |
1360 | /* Log used ring entry write. */ |
1361 | log_write(vq->log_base, |
1362 | vq->log_addr + |
1363 | ((void __user *)used - (void __user *)vq->used), |
1364 | sizeof *used); |
1365 | /* Log used index update. */ |
1366 | log_write(vq->log_base, |
1367 | vq->log_addr + offsetof(struct vring_used, idx), |
1368 | sizeof vq->used->idx); |
1369 | if (vq->log_ctx) |
1370 | eventfd_signal(vq->log_ctx, 1); |
1371 | } |
1372 | vq->last_used_idx++; |
1373 | /* If the driver never bothers to signal in a very long while, |
1374 | * used index might wrap around. If that happens, invalidate |
1375 | * signalled_used index we stored. TODO: make sure driver |
1376 | * signals at least once in 2^16 and remove this. */ |
1377 | if (unlikely(vq->last_used_idx == vq->signalled_used)) |
1378 | vq->signalled_used_valid = false; |
1379 | return 0; |
1380 | } |
1381 | |
1382 | static int __vhost_add_used_n(struct vhost_virtqueue *vq, |
1383 | struct vring_used_elem *heads, |
1384 | unsigned count) |
1385 | { |
1386 | struct vring_used_elem __user *used; |
1387 | u16 old, new; |
1388 | int start; |
1389 | |
1390 | start = vq->last_used_idx % vq->num; |
1391 | used = vq->used->ring + start; |
1392 | if (__copy_to_user(used, heads, count * sizeof *used)) { |
1393 | vq_err(vq, "Failed to write used"); |
1394 | return -EFAULT; |
1395 | } |
1396 | if (unlikely(vq->log_used)) { |
1397 | /* Make sure data is seen before log. */ |
1398 | smp_wmb(); |
1399 | /* Log used ring entry write. */ |
1400 | log_write(vq->log_base, |
1401 | vq->log_addr + |
1402 | ((void __user *)used - (void __user *)vq->used), |
1403 | count * sizeof *used); |
1404 | } |
1405 | old = vq->last_used_idx; |
1406 | new = (vq->last_used_idx += count); |
1407 | /* If the driver never bothers to signal in a very long while, |
1408 | * used index might wrap around. If that happens, invalidate |
1409 | * signalled_used index we stored. TODO: make sure driver |
1410 | * signals at least once in 2^16 and remove this. */ |
1411 | if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old))) |
1412 | vq->signalled_used_valid = false; |
1413 | return 0; |
1414 | } |
1415 | |
1416 | /* After we've used one of their buffers, we tell them about it. We'll then |
1417 | * want to notify the guest, using eventfd. */ |
1418 | int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads, |
1419 | unsigned count) |
1420 | { |
1421 | int start, n, r; |
1422 | |
1423 | start = vq->last_used_idx % vq->num; |
1424 | n = vq->num - start; |
1425 | if (n < count) { |
1426 | r = __vhost_add_used_n(vq, heads, n); |
1427 | if (r < 0) |
1428 | return r; |
1429 | heads += n; |
1430 | count -= n; |
1431 | } |
1432 | r = __vhost_add_used_n(vq, heads, count); |
1433 | |
1434 | /* Make sure buffer is written before we update index. */ |
1435 | smp_wmb(); |
1436 | if (put_user(vq->last_used_idx, &vq->used->idx)) { |
1437 | vq_err(vq, "Failed to increment used idx"); |
1438 | return -EFAULT; |
1439 | } |
1440 | if (unlikely(vq->log_used)) { |
1441 | /* Log used index update. */ |
1442 | log_write(vq->log_base, |
1443 | vq->log_addr + offsetof(struct vring_used, idx), |
1444 | sizeof vq->used->idx); |
1445 | if (vq->log_ctx) |
1446 | eventfd_signal(vq->log_ctx, 1); |
1447 | } |
1448 | return r; |
1449 | } |
1450 | |
1451 | static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq) |
1452 | { |
1453 | __u16 old, new, event; |
1454 | bool v; |
1455 | /* Flush out used index updates. This is paired |
1456 | * with the barrier that the Guest executes when enabling |
1457 | * interrupts. */ |
1458 | smp_mb(); |
1459 | |
1460 | if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) && |
1461 | unlikely(vq->avail_idx == vq->last_avail_idx)) |
1462 | return true; |
1463 | |
1464 | if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) { |
1465 | __u16 flags; |
1466 | if (__get_user(flags, &vq->avail->flags)) { |
1467 | vq_err(vq, "Failed to get flags"); |
1468 | return true; |
1469 | } |
1470 | return !(flags & VRING_AVAIL_F_NO_INTERRUPT); |
1471 | } |
1472 | old = vq->signalled_used; |
1473 | v = vq->signalled_used_valid; |
1474 | new = vq->signalled_used = vq->last_used_idx; |
1475 | vq->signalled_used_valid = true; |
1476 | |
1477 | if (unlikely(!v)) |
1478 | return true; |
1479 | |
1480 | if (get_user(event, vhost_used_event(vq))) { |
1481 | vq_err(vq, "Failed to get used event idx"); |
1482 | return true; |
1483 | } |
1484 | return vring_need_event(event, new, old); |
1485 | } |
1486 | |
1487 | /* This actually signals the guest, using eventfd. */ |
1488 | void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq) |
1489 | { |
1490 | /* Signal the Guest tell them we used something up. */ |
1491 | if (vq->call_ctx && vhost_notify(dev, vq)) |
1492 | eventfd_signal(vq->call_ctx, 1); |
1493 | } |
1494 | |
1495 | /* And here's the combo meal deal. Supersize me! */ |
1496 | void vhost_add_used_and_signal(struct vhost_dev *dev, |
1497 | struct vhost_virtqueue *vq, |
1498 | unsigned int head, int len) |
1499 | { |
1500 | vhost_add_used(vq, head, len); |
1501 | vhost_signal(dev, vq); |
1502 | } |
1503 | |
1504 | /* multi-buffer version of vhost_add_used_and_signal */ |
1505 | void vhost_add_used_and_signal_n(struct vhost_dev *dev, |
1506 | struct vhost_virtqueue *vq, |
1507 | struct vring_used_elem *heads, unsigned count) |
1508 | { |
1509 | vhost_add_used_n(vq, heads, count); |
1510 | vhost_signal(dev, vq); |
1511 | } |
1512 | |
1513 | /* OK, now we need to know about added descriptors. */ |
1514 | bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq) |
1515 | { |
1516 | u16 avail_idx; |
1517 | int r; |
1518 | |
1519 | if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY)) |
1520 | return false; |
1521 | vq->used_flags &= ~VRING_USED_F_NO_NOTIFY; |
1522 | if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) { |
1523 | r = vhost_update_used_flags(vq); |
1524 | if (r) { |
1525 | vq_err(vq, "Failed to enable notification at %p: %d\n", |
1526 | &vq->used->flags, r); |
1527 | return false; |
1528 | } |
1529 | } else { |
1530 | r = vhost_update_avail_event(vq, vq->avail_idx); |
1531 | if (r) { |
1532 | vq_err(vq, "Failed to update avail event index at %p: %d\n", |
1533 | vhost_avail_event(vq), r); |
1534 | return false; |
1535 | } |
1536 | } |
1537 | /* They could have slipped one in as we were doing that: make |
1538 | * sure it's written, then check again. */ |
1539 | smp_mb(); |
1540 | r = __get_user(avail_idx, &vq->avail->idx); |
1541 | if (r) { |
1542 | vq_err(vq, "Failed to check avail idx at %p: %d\n", |
1543 | &vq->avail->idx, r); |
1544 | return false; |
1545 | } |
1546 | |
1547 | return avail_idx != vq->avail_idx; |
1548 | } |
1549 | |
1550 | /* We don't need to be notified again. */ |
1551 | void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq) |
1552 | { |
1553 | int r; |
1554 | |
1555 | if (vq->used_flags & VRING_USED_F_NO_NOTIFY) |
1556 | return; |
1557 | vq->used_flags |= VRING_USED_F_NO_NOTIFY; |
1558 | if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) { |
1559 | r = vhost_update_used_flags(vq); |
1560 | if (r) |
1561 | vq_err(vq, "Failed to enable notification at %p: %d\n", |
1562 | &vq->used->flags, r); |
1563 | } |
1564 | } |
1565 | |
1566 | static void vhost_zerocopy_done_signal(struct kref *kref) |
1567 | { |
1568 | struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref, |
1569 | kref); |
1570 | wake_up(&ubufs->wait); |
1571 | } |
1572 | |
1573 | struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq, |
1574 | bool zcopy) |
1575 | { |
1576 | struct vhost_ubuf_ref *ubufs; |
1577 | /* No zero copy backend? Nothing to count. */ |
1578 | if (!zcopy) |
1579 | return NULL; |
1580 | ubufs = kmalloc(sizeof *ubufs, GFP_KERNEL); |
1581 | if (!ubufs) |
1582 | return ERR_PTR(-ENOMEM); |
1583 | kref_init(&ubufs->kref); |
1584 | init_waitqueue_head(&ubufs->wait); |
1585 | ubufs->vq = vq; |
1586 | return ubufs; |
1587 | } |
1588 | |
1589 | void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs) |
1590 | { |
1591 | kref_put(&ubufs->kref, vhost_zerocopy_done_signal); |
1592 | } |
1593 | |
1594 | void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs) |
1595 | { |
1596 | kref_put(&ubufs->kref, vhost_zerocopy_done_signal); |
1597 | wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount)); |
1598 | kfree(ubufs); |
1599 | } |
1600 | |
1601 | void vhost_zerocopy_callback(struct ubuf_info *ubuf) |
1602 | { |
1603 | struct vhost_ubuf_ref *ubufs = ubuf->ctx; |
1604 | struct vhost_virtqueue *vq = ubufs->vq; |
1605 | |
1606 | /* set len = 1 to mark this desc buffers done DMA */ |
1607 | vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN; |
1608 | kref_put(&ubufs->kref, vhost_zerocopy_done_signal); |
1609 | } |
1610 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9