]>
Commit | Line | Data |
---|---|---|
3a4d5c94 MT |
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> | |
64e1c807 | 18 | #include <linux/mmu_context.h> |
3a4d5c94 MT |
19 | #include <linux/miscdevice.h> |
20 | #include <linux/mutex.h> | |
3a4d5c94 MT |
21 | #include <linux/rcupdate.h> |
22 | #include <linux/poll.h> | |
23 | #include <linux/file.h> | |
24 | #include <linux/highmem.h> | |
5a0e3ad6 | 25 | #include <linux/slab.h> |
c23f3445 | 26 | #include <linux/kthread.h> |
9e3d1957 | 27 | #include <linux/cgroup.h> |
3a4d5c94 MT |
28 | |
29 | #include <linux/net.h> | |
30 | #include <linux/if_packet.h> | |
31 | #include <linux/if_arp.h> | |
32 | ||
3a4d5c94 MT |
33 | #include "vhost.h" |
34 | ||
35 | enum { | |
36 | VHOST_MEMORY_MAX_NREGIONS = 64, | |
37 | VHOST_MEMORY_F_LOG = 0x1, | |
38 | }; | |
39 | ||
3a4d5c94 MT |
40 | static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh, |
41 | poll_table *pt) | |
42 | { | |
43 | struct vhost_poll *poll; | |
3a4d5c94 | 44 | |
d47effe1 | 45 | poll = container_of(pt, struct vhost_poll, table); |
3a4d5c94 MT |
46 | poll->wqh = wqh; |
47 | add_wait_queue(wqh, &poll->wait); | |
48 | } | |
49 | ||
50 | static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync, | |
51 | void *key) | |
52 | { | |
c23f3445 TH |
53 | struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait); |
54 | ||
3a4d5c94 MT |
55 | if (!((unsigned long)key & poll->mask)) |
56 | return 0; | |
57 | ||
c23f3445 | 58 | vhost_poll_queue(poll); |
3a4d5c94 MT |
59 | return 0; |
60 | } | |
61 | ||
87d6a412 MT |
62 | static void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn) |
63 | { | |
64 | INIT_LIST_HEAD(&work->node); | |
65 | work->fn = fn; | |
66 | init_waitqueue_head(&work->done); | |
67 | work->flushing = 0; | |
68 | work->queue_seq = work->done_seq = 0; | |
69 | } | |
70 | ||
3a4d5c94 | 71 | /* Init poll structure */ |
c23f3445 TH |
72 | void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn, |
73 | unsigned long mask, struct vhost_dev *dev) | |
3a4d5c94 | 74 | { |
3a4d5c94 MT |
75 | init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup); |
76 | init_poll_funcptr(&poll->table, vhost_poll_func); | |
77 | poll->mask = mask; | |
c23f3445 TH |
78 | poll->dev = dev; |
79 | ||
87d6a412 | 80 | vhost_work_init(&poll->work, fn); |
3a4d5c94 MT |
81 | } |
82 | ||
83 | /* Start polling a file. We add ourselves to file's wait queue. The caller must | |
84 | * keep a reference to a file until after vhost_poll_stop is called. */ | |
85 | void vhost_poll_start(struct vhost_poll *poll, struct file *file) | |
86 | { | |
87 | unsigned long mask; | |
d47effe1 | 88 | |
3a4d5c94 MT |
89 | mask = file->f_op->poll(file, &poll->table); |
90 | if (mask) | |
91 | vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask); | |
92 | } | |
93 | ||
94 | /* Stop polling a file. After this function returns, it becomes safe to drop the | |
95 | * file reference. You must also flush afterwards. */ | |
96 | void vhost_poll_stop(struct vhost_poll *poll) | |
97 | { | |
98 | remove_wait_queue(poll->wqh, &poll->wait); | |
99 | } | |
100 | ||
0174b0c3 MT |
101 | static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work, |
102 | unsigned seq) | |
103 | { | |
104 | int left; | |
d47effe1 | 105 | |
0174b0c3 MT |
106 | spin_lock_irq(&dev->work_lock); |
107 | left = seq - work->done_seq; | |
108 | spin_unlock_irq(&dev->work_lock); | |
109 | return left <= 0; | |
110 | } | |
111 | ||
87d6a412 | 112 | static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work) |
3a4d5c94 | 113 | { |
c23f3445 | 114 | unsigned seq; |
c23f3445 TH |
115 | int flushing; |
116 | ||
87d6a412 | 117 | spin_lock_irq(&dev->work_lock); |
c23f3445 TH |
118 | seq = work->queue_seq; |
119 | work->flushing++; | |
87d6a412 | 120 | spin_unlock_irq(&dev->work_lock); |
0174b0c3 | 121 | wait_event(work->done, vhost_work_seq_done(dev, work, seq)); |
87d6a412 | 122 | spin_lock_irq(&dev->work_lock); |
c23f3445 | 123 | flushing = --work->flushing; |
87d6a412 | 124 | spin_unlock_irq(&dev->work_lock); |
c23f3445 | 125 | BUG_ON(flushing < 0); |
3a4d5c94 MT |
126 | } |
127 | ||
87d6a412 MT |
128 | /* Flush any work that has been scheduled. When calling this, don't hold any |
129 | * locks that are also used by the callback. */ | |
130 | void vhost_poll_flush(struct vhost_poll *poll) | |
131 | { | |
132 | vhost_work_flush(poll->dev, &poll->work); | |
133 | } | |
134 | ||
135 | static inline void vhost_work_queue(struct vhost_dev *dev, | |
136 | struct vhost_work *work) | |
3a4d5c94 | 137 | { |
c23f3445 TH |
138 | unsigned long flags; |
139 | ||
140 | spin_lock_irqsave(&dev->work_lock, flags); | |
141 | if (list_empty(&work->node)) { | |
142 | list_add_tail(&work->node, &dev->work_list); | |
143 | work->queue_seq++; | |
144 | wake_up_process(dev->worker); | |
145 | } | |
146 | spin_unlock_irqrestore(&dev->work_lock, flags); | |
3a4d5c94 MT |
147 | } |
148 | ||
87d6a412 MT |
149 | void vhost_poll_queue(struct vhost_poll *poll) |
150 | { | |
151 | vhost_work_queue(poll->dev, &poll->work); | |
152 | } | |
153 | ||
3a4d5c94 MT |
154 | static void vhost_vq_reset(struct vhost_dev *dev, |
155 | struct vhost_virtqueue *vq) | |
156 | { | |
157 | vq->num = 1; | |
158 | vq->desc = NULL; | |
159 | vq->avail = NULL; | |
160 | vq->used = NULL; | |
161 | vq->last_avail_idx = 0; | |
162 | vq->avail_idx = 0; | |
163 | vq->last_used_idx = 0; | |
164 | vq->used_flags = 0; | |
3a4d5c94 MT |
165 | vq->log_used = false; |
166 | vq->log_addr = -1ull; | |
8dd014ad DS |
167 | vq->vhost_hlen = 0; |
168 | vq->sock_hlen = 0; | |
3a4d5c94 MT |
169 | vq->private_data = NULL; |
170 | vq->log_base = NULL; | |
171 | vq->error_ctx = NULL; | |
172 | vq->error = NULL; | |
173 | vq->kick = NULL; | |
174 | vq->call_ctx = NULL; | |
175 | vq->call = NULL; | |
73a99f08 | 176 | vq->log_ctx = NULL; |
3a4d5c94 MT |
177 | } |
178 | ||
c23f3445 TH |
179 | static int vhost_worker(void *data) |
180 | { | |
181 | struct vhost_dev *dev = data; | |
182 | struct vhost_work *work = NULL; | |
183 | unsigned uninitialized_var(seq); | |
184 | ||
64e1c807 MT |
185 | use_mm(dev->mm); |
186 | ||
c23f3445 TH |
187 | for (;;) { |
188 | /* mb paired w/ kthread_stop */ | |
189 | set_current_state(TASK_INTERRUPTIBLE); | |
190 | ||
191 | spin_lock_irq(&dev->work_lock); | |
192 | if (work) { | |
193 | work->done_seq = seq; | |
194 | if (work->flushing) | |
195 | wake_up_all(&work->done); | |
196 | } | |
197 | ||
198 | if (kthread_should_stop()) { | |
199 | spin_unlock_irq(&dev->work_lock); | |
200 | __set_current_state(TASK_RUNNING); | |
64e1c807 | 201 | break; |
c23f3445 TH |
202 | } |
203 | if (!list_empty(&dev->work_list)) { | |
204 | work = list_first_entry(&dev->work_list, | |
205 | struct vhost_work, node); | |
206 | list_del_init(&work->node); | |
207 | seq = work->queue_seq; | |
208 | } else | |
209 | work = NULL; | |
210 | spin_unlock_irq(&dev->work_lock); | |
211 | ||
212 | if (work) { | |
213 | __set_current_state(TASK_RUNNING); | |
214 | work->fn(work); | |
215 | } else | |
216 | schedule(); | |
217 | ||
218 | } | |
64e1c807 MT |
219 | unuse_mm(dev->mm); |
220 | return 0; | |
c23f3445 TH |
221 | } |
222 | ||
e0e9b406 JW |
223 | /* Helper to allocate iovec buffers for all vqs. */ |
224 | static long vhost_dev_alloc_iovecs(struct vhost_dev *dev) | |
225 | { | |
226 | int i; | |
d47effe1 | 227 | |
e0e9b406 JW |
228 | for (i = 0; i < dev->nvqs; ++i) { |
229 | dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect * | |
230 | UIO_MAXIOV, GFP_KERNEL); | |
231 | dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV, | |
232 | GFP_KERNEL); | |
233 | dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads * | |
234 | UIO_MAXIOV, GFP_KERNEL); | |
235 | ||
236 | if (!dev->vqs[i].indirect || !dev->vqs[i].log || | |
237 | !dev->vqs[i].heads) | |
238 | goto err_nomem; | |
239 | } | |
240 | return 0; | |
d47effe1 | 241 | |
e0e9b406 JW |
242 | err_nomem: |
243 | for (; i >= 0; --i) { | |
244 | kfree(dev->vqs[i].indirect); | |
245 | kfree(dev->vqs[i].log); | |
246 | kfree(dev->vqs[i].heads); | |
247 | } | |
248 | return -ENOMEM; | |
249 | } | |
250 | ||
251 | static void vhost_dev_free_iovecs(struct vhost_dev *dev) | |
252 | { | |
253 | int i; | |
d47effe1 | 254 | |
e0e9b406 JW |
255 | for (i = 0; i < dev->nvqs; ++i) { |
256 | kfree(dev->vqs[i].indirect); | |
257 | dev->vqs[i].indirect = NULL; | |
258 | kfree(dev->vqs[i].log); | |
259 | dev->vqs[i].log = NULL; | |
260 | kfree(dev->vqs[i].heads); | |
261 | dev->vqs[i].heads = NULL; | |
262 | } | |
263 | } | |
264 | ||
3a4d5c94 MT |
265 | long vhost_dev_init(struct vhost_dev *dev, |
266 | struct vhost_virtqueue *vqs, int nvqs) | |
267 | { | |
268 | int i; | |
c23f3445 | 269 | |
3a4d5c94 MT |
270 | dev->vqs = vqs; |
271 | dev->nvqs = nvqs; | |
272 | mutex_init(&dev->mutex); | |
273 | dev->log_ctx = NULL; | |
274 | dev->log_file = NULL; | |
275 | dev->memory = NULL; | |
276 | dev->mm = NULL; | |
c23f3445 TH |
277 | spin_lock_init(&dev->work_lock); |
278 | INIT_LIST_HEAD(&dev->work_list); | |
279 | dev->worker = NULL; | |
3a4d5c94 MT |
280 | |
281 | for (i = 0; i < dev->nvqs; ++i) { | |
e0e9b406 JW |
282 | dev->vqs[i].log = NULL; |
283 | dev->vqs[i].indirect = NULL; | |
284 | dev->vqs[i].heads = NULL; | |
3a4d5c94 MT |
285 | dev->vqs[i].dev = dev; |
286 | mutex_init(&dev->vqs[i].mutex); | |
287 | vhost_vq_reset(dev, dev->vqs + i); | |
288 | if (dev->vqs[i].handle_kick) | |
289 | vhost_poll_init(&dev->vqs[i].poll, | |
c23f3445 | 290 | dev->vqs[i].handle_kick, POLLIN, dev); |
3a4d5c94 | 291 | } |
c23f3445 | 292 | |
3a4d5c94 MT |
293 | return 0; |
294 | } | |
295 | ||
296 | /* Caller should have device mutex */ | |
297 | long vhost_dev_check_owner(struct vhost_dev *dev) | |
298 | { | |
299 | /* Are you the owner? If not, I don't think you mean to do that */ | |
300 | return dev->mm == current->mm ? 0 : -EPERM; | |
301 | } | |
302 | ||
87d6a412 | 303 | struct vhost_attach_cgroups_struct { |
d47effe1 KK |
304 | struct vhost_work work; |
305 | struct task_struct *owner; | |
306 | int ret; | |
87d6a412 MT |
307 | }; |
308 | ||
309 | static void vhost_attach_cgroups_work(struct vhost_work *work) | |
310 | { | |
d47effe1 KK |
311 | struct vhost_attach_cgroups_struct *s; |
312 | ||
313 | s = container_of(work, struct vhost_attach_cgroups_struct, work); | |
314 | s->ret = cgroup_attach_task_all(s->owner, current); | |
87d6a412 MT |
315 | } |
316 | ||
317 | static int vhost_attach_cgroups(struct vhost_dev *dev) | |
318 | { | |
d47effe1 KK |
319 | struct vhost_attach_cgroups_struct attach; |
320 | ||
321 | attach.owner = current; | |
322 | vhost_work_init(&attach.work, vhost_attach_cgroups_work); | |
323 | vhost_work_queue(dev, &attach.work); | |
324 | vhost_work_flush(dev, &attach.work); | |
325 | return attach.ret; | |
87d6a412 MT |
326 | } |
327 | ||
3a4d5c94 MT |
328 | /* Caller should have device mutex */ |
329 | static long vhost_dev_set_owner(struct vhost_dev *dev) | |
330 | { | |
c23f3445 TH |
331 | struct task_struct *worker; |
332 | int err; | |
d47effe1 | 333 | |
3a4d5c94 | 334 | /* Is there an owner already? */ |
c23f3445 TH |
335 | if (dev->mm) { |
336 | err = -EBUSY; | |
337 | goto err_mm; | |
338 | } | |
d47effe1 | 339 | |
3a4d5c94 MT |
340 | /* No owner, become one */ |
341 | dev->mm = get_task_mm(current); | |
c23f3445 TH |
342 | worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid); |
343 | if (IS_ERR(worker)) { | |
344 | err = PTR_ERR(worker); | |
345 | goto err_worker; | |
346 | } | |
347 | ||
348 | dev->worker = worker; | |
87d6a412 MT |
349 | wake_up_process(worker); /* avoid contributing to loadavg */ |
350 | ||
351 | err = vhost_attach_cgroups(dev); | |
9e3d1957 MT |
352 | if (err) |
353 | goto err_cgroup; | |
c23f3445 | 354 | |
e0e9b406 JW |
355 | err = vhost_dev_alloc_iovecs(dev); |
356 | if (err) | |
357 | goto err_cgroup; | |
358 | ||
3a4d5c94 | 359 | return 0; |
9e3d1957 MT |
360 | err_cgroup: |
361 | kthread_stop(worker); | |
615cc221 | 362 | dev->worker = NULL; |
c23f3445 TH |
363 | err_worker: |
364 | if (dev->mm) | |
365 | mmput(dev->mm); | |
366 | dev->mm = NULL; | |
367 | err_mm: | |
368 | return err; | |
3a4d5c94 MT |
369 | } |
370 | ||
371 | /* Caller should have device mutex */ | |
372 | long vhost_dev_reset_owner(struct vhost_dev *dev) | |
373 | { | |
374 | struct vhost_memory *memory; | |
375 | ||
376 | /* Restore memory to default empty mapping. */ | |
377 | memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL); | |
378 | if (!memory) | |
379 | return -ENOMEM; | |
380 | ||
381 | vhost_dev_cleanup(dev); | |
382 | ||
383 | memory->nregions = 0; | |
28457ee6 | 384 | RCU_INIT_POINTER(dev->memory, memory); |
3a4d5c94 MT |
385 | return 0; |
386 | } | |
387 | ||
388 | /* Caller should have device mutex */ | |
389 | void vhost_dev_cleanup(struct vhost_dev *dev) | |
390 | { | |
391 | int i; | |
d47effe1 | 392 | |
3a4d5c94 MT |
393 | for (i = 0; i < dev->nvqs; ++i) { |
394 | if (dev->vqs[i].kick && dev->vqs[i].handle_kick) { | |
395 | vhost_poll_stop(&dev->vqs[i].poll); | |
396 | vhost_poll_flush(&dev->vqs[i].poll); | |
397 | } | |
398 | if (dev->vqs[i].error_ctx) | |
399 | eventfd_ctx_put(dev->vqs[i].error_ctx); | |
400 | if (dev->vqs[i].error) | |
401 | fput(dev->vqs[i].error); | |
402 | if (dev->vqs[i].kick) | |
403 | fput(dev->vqs[i].kick); | |
404 | if (dev->vqs[i].call_ctx) | |
405 | eventfd_ctx_put(dev->vqs[i].call_ctx); | |
406 | if (dev->vqs[i].call) | |
407 | fput(dev->vqs[i].call); | |
408 | vhost_vq_reset(dev, dev->vqs + i); | |
409 | } | |
e0e9b406 | 410 | vhost_dev_free_iovecs(dev); |
3a4d5c94 MT |
411 | if (dev->log_ctx) |
412 | eventfd_ctx_put(dev->log_ctx); | |
413 | dev->log_ctx = NULL; | |
414 | if (dev->log_file) | |
415 | fput(dev->log_file); | |
416 | dev->log_file = NULL; | |
417 | /* No one will access memory at this point */ | |
28457ee6 AB |
418 | kfree(rcu_dereference_protected(dev->memory, |
419 | lockdep_is_held(&dev->mutex))); | |
420 | RCU_INIT_POINTER(dev->memory, NULL); | |
c23f3445 | 421 | WARN_ON(!list_empty(&dev->work_list)); |
78b620ce ED |
422 | if (dev->worker) { |
423 | kthread_stop(dev->worker); | |
424 | dev->worker = NULL; | |
425 | } | |
533a19b4 MT |
426 | if (dev->mm) |
427 | mmput(dev->mm); | |
428 | dev->mm = NULL; | |
3a4d5c94 MT |
429 | } |
430 | ||
431 | static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz) | |
432 | { | |
433 | u64 a = addr / VHOST_PAGE_SIZE / 8; | |
d47effe1 | 434 | |
3a4d5c94 MT |
435 | /* Make sure 64 bit math will not overflow. */ |
436 | if (a > ULONG_MAX - (unsigned long)log_base || | |
437 | a + (unsigned long)log_base > ULONG_MAX) | |
6d97e55f | 438 | return 0; |
3a4d5c94 MT |
439 | |
440 | return access_ok(VERIFY_WRITE, log_base + a, | |
441 | (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8); | |
442 | } | |
443 | ||
444 | /* Caller should have vq mutex and device mutex. */ | |
445 | static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem, | |
446 | int log_all) | |
447 | { | |
448 | int i; | |
179b284e | 449 | |
f8322fbe MT |
450 | if (!mem) |
451 | return 0; | |
179b284e | 452 | |
3a4d5c94 MT |
453 | for (i = 0; i < mem->nregions; ++i) { |
454 | struct vhost_memory_region *m = mem->regions + i; | |
455 | unsigned long a = m->userspace_addr; | |
456 | if (m->memory_size > ULONG_MAX) | |
457 | return 0; | |
458 | else if (!access_ok(VERIFY_WRITE, (void __user *)a, | |
459 | m->memory_size)) | |
460 | return 0; | |
461 | else if (log_all && !log_access_ok(log_base, | |
462 | m->guest_phys_addr, | |
463 | m->memory_size)) | |
464 | return 0; | |
465 | } | |
466 | return 1; | |
467 | } | |
468 | ||
469 | /* Can we switch to this memory table? */ | |
470 | /* Caller should have device mutex but not vq mutex */ | |
471 | static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem, | |
472 | int log_all) | |
473 | { | |
474 | int i; | |
d47effe1 | 475 | |
3a4d5c94 MT |
476 | for (i = 0; i < d->nvqs; ++i) { |
477 | int ok; | |
478 | mutex_lock(&d->vqs[i].mutex); | |
479 | /* If ring is inactive, will check when it's enabled. */ | |
480 | if (d->vqs[i].private_data) | |
481 | ok = vq_memory_access_ok(d->vqs[i].log_base, mem, | |
482 | log_all); | |
483 | else | |
484 | ok = 1; | |
485 | mutex_unlock(&d->vqs[i].mutex); | |
486 | if (!ok) | |
487 | return 0; | |
488 | } | |
489 | return 1; | |
490 | } | |
491 | ||
492 | static int vq_access_ok(unsigned int num, | |
493 | struct vring_desc __user *desc, | |
494 | struct vring_avail __user *avail, | |
495 | struct vring_used __user *used) | |
496 | { | |
497 | return access_ok(VERIFY_READ, desc, num * sizeof *desc) && | |
498 | access_ok(VERIFY_READ, avail, | |
499 | sizeof *avail + num * sizeof *avail->ring) && | |
500 | access_ok(VERIFY_WRITE, used, | |
501 | sizeof *used + num * sizeof *used->ring); | |
502 | } | |
503 | ||
504 | /* Can we log writes? */ | |
505 | /* Caller should have device mutex but not vq mutex */ | |
506 | int vhost_log_access_ok(struct vhost_dev *dev) | |
507 | { | |
28457ee6 AB |
508 | struct vhost_memory *mp; |
509 | ||
510 | mp = rcu_dereference_protected(dev->memory, | |
511 | lockdep_is_held(&dev->mutex)); | |
512 | return memory_access_ok(dev, mp, 1); | |
3a4d5c94 MT |
513 | } |
514 | ||
515 | /* Verify access for write logging. */ | |
516 | /* Caller should have vq mutex and device mutex */ | |
517 | static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base) | |
518 | { | |
28457ee6 AB |
519 | struct vhost_memory *mp; |
520 | ||
521 | mp = rcu_dereference_protected(vq->dev->memory, | |
522 | lockdep_is_held(&vq->mutex)); | |
523 | return vq_memory_access_ok(log_base, mp, | |
3a4d5c94 MT |
524 | vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) && |
525 | (!vq->log_used || log_access_ok(log_base, vq->log_addr, | |
526 | sizeof *vq->used + | |
527 | vq->num * sizeof *vq->used->ring)); | |
528 | } | |
529 | ||
530 | /* Can we start vq? */ | |
531 | /* Caller should have vq mutex and device mutex */ | |
532 | int vhost_vq_access_ok(struct vhost_virtqueue *vq) | |
533 | { | |
534 | return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) && | |
535 | vq_log_access_ok(vq, vq->log_base); | |
536 | } | |
537 | ||
538 | static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m) | |
539 | { | |
540 | struct vhost_memory mem, *newmem, *oldmem; | |
541 | unsigned long size = offsetof(struct vhost_memory, regions); | |
d47effe1 | 542 | |
7ad9c9d2 TY |
543 | if (copy_from_user(&mem, m, size)) |
544 | return -EFAULT; | |
3a4d5c94 MT |
545 | if (mem.padding) |
546 | return -EOPNOTSUPP; | |
547 | if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS) | |
548 | return -E2BIG; | |
549 | newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL); | |
550 | if (!newmem) | |
551 | return -ENOMEM; | |
552 | ||
553 | memcpy(newmem, &mem, size); | |
7ad9c9d2 TY |
554 | if (copy_from_user(newmem->regions, m->regions, |
555 | mem.nregions * sizeof *m->regions)) { | |
3a4d5c94 | 556 | kfree(newmem); |
7ad9c9d2 | 557 | return -EFAULT; |
3a4d5c94 MT |
558 | } |
559 | ||
d47effe1 KK |
560 | if (!memory_access_ok(d, newmem, |
561 | vhost_has_feature(d, VHOST_F_LOG_ALL))) { | |
a02c3789 | 562 | kfree(newmem); |
3a4d5c94 | 563 | return -EFAULT; |
a02c3789 | 564 | } |
28457ee6 AB |
565 | oldmem = rcu_dereference_protected(d->memory, |
566 | lockdep_is_held(&d->mutex)); | |
3a4d5c94 MT |
567 | rcu_assign_pointer(d->memory, newmem); |
568 | synchronize_rcu(); | |
569 | kfree(oldmem); | |
570 | return 0; | |
571 | } | |
572 | ||
573 | static int init_used(struct vhost_virtqueue *vq, | |
574 | struct vring_used __user *used) | |
575 | { | |
576 | int r = put_user(vq->used_flags, &used->flags); | |
d47effe1 | 577 | |
3a4d5c94 MT |
578 | if (r) |
579 | return r; | |
580 | return get_user(vq->last_used_idx, &used->idx); | |
581 | } | |
582 | ||
583 | static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp) | |
584 | { | |
585 | struct file *eventfp, *filep = NULL, | |
586 | *pollstart = NULL, *pollstop = NULL; | |
587 | struct eventfd_ctx *ctx = NULL; | |
588 | u32 __user *idxp = argp; | |
589 | struct vhost_virtqueue *vq; | |
590 | struct vhost_vring_state s; | |
591 | struct vhost_vring_file f; | |
592 | struct vhost_vring_addr a; | |
593 | u32 idx; | |
594 | long r; | |
595 | ||
596 | r = get_user(idx, idxp); | |
597 | if (r < 0) | |
598 | return r; | |
0f3d9a17 | 599 | if (idx >= d->nvqs) |
3a4d5c94 MT |
600 | return -ENOBUFS; |
601 | ||
602 | vq = d->vqs + idx; | |
603 | ||
604 | mutex_lock(&vq->mutex); | |
605 | ||
606 | switch (ioctl) { | |
607 | case VHOST_SET_VRING_NUM: | |
608 | /* Resizing ring with an active backend? | |
609 | * You don't want to do that. */ | |
610 | if (vq->private_data) { | |
611 | r = -EBUSY; | |
612 | break; | |
613 | } | |
7ad9c9d2 TY |
614 | if (copy_from_user(&s, argp, sizeof s)) { |
615 | r = -EFAULT; | |
3a4d5c94 | 616 | break; |
7ad9c9d2 | 617 | } |
3a4d5c94 MT |
618 | if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) { |
619 | r = -EINVAL; | |
620 | break; | |
621 | } | |
622 | vq->num = s.num; | |
623 | break; | |
624 | case VHOST_SET_VRING_BASE: | |
625 | /* Moving base with an active backend? | |
626 | * You don't want to do that. */ | |
627 | if (vq->private_data) { | |
628 | r = -EBUSY; | |
629 | break; | |
630 | } | |
7ad9c9d2 TY |
631 | if (copy_from_user(&s, argp, sizeof s)) { |
632 | r = -EFAULT; | |
3a4d5c94 | 633 | break; |
7ad9c9d2 | 634 | } |
3a4d5c94 MT |
635 | if (s.num > 0xffff) { |
636 | r = -EINVAL; | |
637 | break; | |
638 | } | |
639 | vq->last_avail_idx = s.num; | |
640 | /* Forget the cached index value. */ | |
641 | vq->avail_idx = vq->last_avail_idx; | |
642 | break; | |
643 | case VHOST_GET_VRING_BASE: | |
644 | s.index = idx; | |
645 | s.num = vq->last_avail_idx; | |
7ad9c9d2 TY |
646 | if (copy_to_user(argp, &s, sizeof s)) |
647 | r = -EFAULT; | |
3a4d5c94 MT |
648 | break; |
649 | case VHOST_SET_VRING_ADDR: | |
7ad9c9d2 TY |
650 | if (copy_from_user(&a, argp, sizeof a)) { |
651 | r = -EFAULT; | |
3a4d5c94 | 652 | break; |
7ad9c9d2 | 653 | } |
3a4d5c94 MT |
654 | if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) { |
655 | r = -EOPNOTSUPP; | |
656 | break; | |
657 | } | |
658 | /* For 32bit, verify that the top 32bits of the user | |
659 | data are set to zero. */ | |
660 | if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr || | |
661 | (u64)(unsigned long)a.used_user_addr != a.used_user_addr || | |
662 | (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) { | |
663 | r = -EFAULT; | |
664 | break; | |
665 | } | |
666 | if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) || | |
667 | (a.used_user_addr & (sizeof *vq->used->ring - 1)) || | |
668 | (a.log_guest_addr & (sizeof *vq->used->ring - 1))) { | |
669 | r = -EINVAL; | |
670 | break; | |
671 | } | |
672 | ||
673 | /* We only verify access here if backend is configured. | |
674 | * If it is not, we don't as size might not have been setup. | |
675 | * We will verify when backend is configured. */ | |
676 | if (vq->private_data) { | |
677 | if (!vq_access_ok(vq->num, | |
678 | (void __user *)(unsigned long)a.desc_user_addr, | |
679 | (void __user *)(unsigned long)a.avail_user_addr, | |
680 | (void __user *)(unsigned long)a.used_user_addr)) { | |
681 | r = -EINVAL; | |
682 | break; | |
683 | } | |
684 | ||
685 | /* Also validate log access for used ring if enabled. */ | |
686 | if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) && | |
687 | !log_access_ok(vq->log_base, a.log_guest_addr, | |
688 | sizeof *vq->used + | |
689 | vq->num * sizeof *vq->used->ring)) { | |
690 | r = -EINVAL; | |
691 | break; | |
692 | } | |
693 | } | |
694 | ||
695 | r = init_used(vq, (struct vring_used __user *)(unsigned long) | |
696 | a.used_user_addr); | |
697 | if (r) | |
698 | break; | |
699 | vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG)); | |
700 | vq->desc = (void __user *)(unsigned long)a.desc_user_addr; | |
701 | vq->avail = (void __user *)(unsigned long)a.avail_user_addr; | |
702 | vq->log_addr = a.log_guest_addr; | |
703 | vq->used = (void __user *)(unsigned long)a.used_user_addr; | |
704 | break; | |
705 | case VHOST_SET_VRING_KICK: | |
7ad9c9d2 TY |
706 | if (copy_from_user(&f, argp, sizeof f)) { |
707 | r = -EFAULT; | |
3a4d5c94 | 708 | break; |
7ad9c9d2 | 709 | } |
3a4d5c94 | 710 | eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); |
535297a6 MT |
711 | if (IS_ERR(eventfp)) { |
712 | r = PTR_ERR(eventfp); | |
713 | break; | |
714 | } | |
3a4d5c94 MT |
715 | if (eventfp != vq->kick) { |
716 | pollstop = filep = vq->kick; | |
717 | pollstart = vq->kick = eventfp; | |
718 | } else | |
719 | filep = eventfp; | |
720 | break; | |
721 | case VHOST_SET_VRING_CALL: | |
7ad9c9d2 TY |
722 | if (copy_from_user(&f, argp, sizeof f)) { |
723 | r = -EFAULT; | |
3a4d5c94 | 724 | break; |
7ad9c9d2 | 725 | } |
3a4d5c94 | 726 | eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); |
535297a6 MT |
727 | if (IS_ERR(eventfp)) { |
728 | r = PTR_ERR(eventfp); | |
729 | break; | |
730 | } | |
3a4d5c94 MT |
731 | if (eventfp != vq->call) { |
732 | filep = vq->call; | |
733 | ctx = vq->call_ctx; | |
734 | vq->call = eventfp; | |
735 | vq->call_ctx = eventfp ? | |
736 | eventfd_ctx_fileget(eventfp) : NULL; | |
737 | } else | |
738 | filep = eventfp; | |
739 | break; | |
740 | case VHOST_SET_VRING_ERR: | |
7ad9c9d2 TY |
741 | if (copy_from_user(&f, argp, sizeof f)) { |
742 | r = -EFAULT; | |
3a4d5c94 | 743 | break; |
7ad9c9d2 | 744 | } |
3a4d5c94 | 745 | eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); |
535297a6 MT |
746 | if (IS_ERR(eventfp)) { |
747 | r = PTR_ERR(eventfp); | |
748 | break; | |
749 | } | |
3a4d5c94 MT |
750 | if (eventfp != vq->error) { |
751 | filep = vq->error; | |
752 | vq->error = eventfp; | |
753 | ctx = vq->error_ctx; | |
754 | vq->error_ctx = eventfp ? | |
755 | eventfd_ctx_fileget(eventfp) : NULL; | |
756 | } else | |
757 | filep = eventfp; | |
758 | break; | |
759 | default: | |
760 | r = -ENOIOCTLCMD; | |
761 | } | |
762 | ||
763 | if (pollstop && vq->handle_kick) | |
764 | vhost_poll_stop(&vq->poll); | |
765 | ||
766 | if (ctx) | |
767 | eventfd_ctx_put(ctx); | |
768 | if (filep) | |
769 | fput(filep); | |
770 | ||
771 | if (pollstart && vq->handle_kick) | |
772 | vhost_poll_start(&vq->poll, vq->kick); | |
773 | ||
774 | mutex_unlock(&vq->mutex); | |
775 | ||
776 | if (pollstop && vq->handle_kick) | |
777 | vhost_poll_flush(&vq->poll); | |
778 | return r; | |
779 | } | |
780 | ||
781 | /* Caller must have device mutex */ | |
782 | long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg) | |
783 | { | |
784 | void __user *argp = (void __user *)arg; | |
785 | struct file *eventfp, *filep = NULL; | |
786 | struct eventfd_ctx *ctx = NULL; | |
787 | u64 p; | |
788 | long r; | |
789 | int i, fd; | |
790 | ||
791 | /* If you are not the owner, you can become one */ | |
792 | if (ioctl == VHOST_SET_OWNER) { | |
793 | r = vhost_dev_set_owner(d); | |
794 | goto done; | |
795 | } | |
796 | ||
797 | /* You must be the owner to do anything else */ | |
798 | r = vhost_dev_check_owner(d); | |
799 | if (r) | |
800 | goto done; | |
801 | ||
802 | switch (ioctl) { | |
803 | case VHOST_SET_MEM_TABLE: | |
804 | r = vhost_set_memory(d, argp); | |
805 | break; | |
806 | case VHOST_SET_LOG_BASE: | |
7ad9c9d2 TY |
807 | if (copy_from_user(&p, argp, sizeof p)) { |
808 | r = -EFAULT; | |
3a4d5c94 | 809 | break; |
7ad9c9d2 | 810 | } |
3a4d5c94 MT |
811 | if ((u64)(unsigned long)p != p) { |
812 | r = -EFAULT; | |
813 | break; | |
814 | } | |
815 | for (i = 0; i < d->nvqs; ++i) { | |
816 | struct vhost_virtqueue *vq; | |
817 | void __user *base = (void __user *)(unsigned long)p; | |
818 | vq = d->vqs + i; | |
819 | mutex_lock(&vq->mutex); | |
820 | /* If ring is inactive, will check when it's enabled. */ | |
821 | if (vq->private_data && !vq_log_access_ok(vq, base)) | |
822 | r = -EFAULT; | |
823 | else | |
824 | vq->log_base = base; | |
825 | mutex_unlock(&vq->mutex); | |
826 | } | |
827 | break; | |
828 | case VHOST_SET_LOG_FD: | |
829 | r = get_user(fd, (int __user *)argp); | |
830 | if (r < 0) | |
831 | break; | |
832 | eventfp = fd == -1 ? NULL : eventfd_fget(fd); | |
833 | if (IS_ERR(eventfp)) { | |
834 | r = PTR_ERR(eventfp); | |
835 | break; | |
836 | } | |
837 | if (eventfp != d->log_file) { | |
838 | filep = d->log_file; | |
839 | ctx = d->log_ctx; | |
840 | d->log_ctx = eventfp ? | |
841 | eventfd_ctx_fileget(eventfp) : NULL; | |
842 | } else | |
843 | filep = eventfp; | |
844 | for (i = 0; i < d->nvqs; ++i) { | |
845 | mutex_lock(&d->vqs[i].mutex); | |
846 | d->vqs[i].log_ctx = d->log_ctx; | |
847 | mutex_unlock(&d->vqs[i].mutex); | |
848 | } | |
849 | if (ctx) | |
850 | eventfd_ctx_put(ctx); | |
851 | if (filep) | |
852 | fput(filep); | |
853 | break; | |
854 | default: | |
855 | r = vhost_set_vring(d, ioctl, argp); | |
856 | break; | |
857 | } | |
858 | done: | |
859 | return r; | |
860 | } | |
861 | ||
862 | static const struct vhost_memory_region *find_region(struct vhost_memory *mem, | |
863 | __u64 addr, __u32 len) | |
864 | { | |
865 | struct vhost_memory_region *reg; | |
866 | int i; | |
d47effe1 | 867 | |
3a4d5c94 MT |
868 | /* linear search is not brilliant, but we really have on the order of 6 |
869 | * regions in practice */ | |
870 | for (i = 0; i < mem->nregions; ++i) { | |
871 | reg = mem->regions + i; | |
872 | if (reg->guest_phys_addr <= addr && | |
873 | reg->guest_phys_addr + reg->memory_size - 1 >= addr) | |
874 | return reg; | |
875 | } | |
876 | return NULL; | |
877 | } | |
878 | ||
879 | /* TODO: This is really inefficient. We need something like get_user() | |
880 | * (instruction directly accesses the data, with an exception table entry | |
881 | * returning -EFAULT). See Documentation/x86/exception-tables.txt. | |
882 | */ | |
883 | static int set_bit_to_user(int nr, void __user *addr) | |
884 | { | |
885 | unsigned long log = (unsigned long)addr; | |
886 | struct page *page; | |
887 | void *base; | |
888 | int bit = nr + (log % PAGE_SIZE) * 8; | |
889 | int r; | |
d47effe1 | 890 | |
3a4d5c94 | 891 | r = get_user_pages_fast(log, 1, 1, &page); |
d6db3f5c | 892 | if (r < 0) |
3a4d5c94 | 893 | return r; |
d6db3f5c | 894 | BUG_ON(r != 1); |
3a4d5c94 MT |
895 | base = kmap_atomic(page, KM_USER0); |
896 | set_bit(bit, base); | |
897 | kunmap_atomic(base, KM_USER0); | |
898 | set_page_dirty_lock(page); | |
899 | put_page(page); | |
900 | return 0; | |
901 | } | |
902 | ||
903 | static int log_write(void __user *log_base, | |
904 | u64 write_address, u64 write_length) | |
905 | { | |
28831ee6 | 906 | u64 write_page = write_address / VHOST_PAGE_SIZE; |
3a4d5c94 | 907 | int r; |
d47effe1 | 908 | |
3a4d5c94 MT |
909 | if (!write_length) |
910 | return 0; | |
3bf9be40 | 911 | write_length += write_address % VHOST_PAGE_SIZE; |
3a4d5c94 MT |
912 | for (;;) { |
913 | u64 base = (u64)(unsigned long)log_base; | |
28831ee6 MT |
914 | u64 log = base + write_page / 8; |
915 | int bit = write_page % 8; | |
3a4d5c94 MT |
916 | if ((u64)(unsigned long)log != log) |
917 | return -EFAULT; | |
918 | r = set_bit_to_user(bit, (void __user *)(unsigned long)log); | |
919 | if (r < 0) | |
920 | return r; | |
921 | if (write_length <= VHOST_PAGE_SIZE) | |
922 | break; | |
923 | write_length -= VHOST_PAGE_SIZE; | |
28831ee6 | 924 | write_page += 1; |
3a4d5c94 MT |
925 | } |
926 | return r; | |
927 | } | |
928 | ||
929 | int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log, | |
930 | unsigned int log_num, u64 len) | |
931 | { | |
932 | int i, r; | |
933 | ||
934 | /* Make sure data written is seen before log. */ | |
5659338c | 935 | smp_wmb(); |
3a4d5c94 MT |
936 | for (i = 0; i < log_num; ++i) { |
937 | u64 l = min(log[i].len, len); | |
938 | r = log_write(vq->log_base, log[i].addr, l); | |
939 | if (r < 0) | |
940 | return r; | |
941 | len -= l; | |
5786aee8 MT |
942 | if (!len) { |
943 | if (vq->log_ctx) | |
944 | eventfd_signal(vq->log_ctx, 1); | |
3a4d5c94 | 945 | return 0; |
5786aee8 | 946 | } |
3a4d5c94 | 947 | } |
3a4d5c94 MT |
948 | /* Length written exceeds what we have stored. This is a bug. */ |
949 | BUG(); | |
950 | return 0; | |
951 | } | |
952 | ||
a8d3782f CH |
953 | static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len, |
954 | struct iovec iov[], int iov_size) | |
3a4d5c94 MT |
955 | { |
956 | const struct vhost_memory_region *reg; | |
957 | struct vhost_memory *mem; | |
958 | struct iovec *_iov; | |
959 | u64 s = 0; | |
960 | int ret = 0; | |
961 | ||
962 | rcu_read_lock(); | |
963 | ||
964 | mem = rcu_dereference(dev->memory); | |
965 | while ((u64)len > s) { | |
966 | u64 size; | |
7b3384fc | 967 | if (unlikely(ret >= iov_size)) { |
3a4d5c94 MT |
968 | ret = -ENOBUFS; |
969 | break; | |
970 | } | |
971 | reg = find_region(mem, addr, len); | |
7b3384fc | 972 | if (unlikely(!reg)) { |
3a4d5c94 MT |
973 | ret = -EFAULT; |
974 | break; | |
975 | } | |
976 | _iov = iov + ret; | |
977 | size = reg->memory_size - addr + reg->guest_phys_addr; | |
978 | _iov->iov_len = min((u64)len, size); | |
a8d3782f | 979 | _iov->iov_base = (void __user *)(unsigned long) |
3a4d5c94 MT |
980 | (reg->userspace_addr + addr - reg->guest_phys_addr); |
981 | s += size; | |
982 | addr += size; | |
983 | ++ret; | |
984 | } | |
985 | ||
986 | rcu_read_unlock(); | |
987 | return ret; | |
988 | } | |
989 | ||
990 | /* Each buffer in the virtqueues is actually a chain of descriptors. This | |
991 | * function returns the next descriptor in the chain, | |
992 | * or -1U if we're at the end. */ | |
993 | static unsigned next_desc(struct vring_desc *desc) | |
994 | { | |
995 | unsigned int next; | |
996 | ||
997 | /* If this descriptor says it doesn't chain, we're done. */ | |
998 | if (!(desc->flags & VRING_DESC_F_NEXT)) | |
999 | return -1U; | |
1000 | ||
1001 | /* Check they're not leading us off end of descriptors. */ | |
1002 | next = desc->next; | |
1003 | /* Make sure compiler knows to grab that: we don't want it changing! */ | |
1004 | /* We will use the result as an index in an array, so most | |
1005 | * architectures only need a compiler barrier here. */ | |
1006 | read_barrier_depends(); | |
1007 | ||
1008 | return next; | |
1009 | } | |
1010 | ||
7b3384fc MT |
1011 | static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq, |
1012 | struct iovec iov[], unsigned int iov_size, | |
1013 | unsigned int *out_num, unsigned int *in_num, | |
1014 | struct vhost_log *log, unsigned int *log_num, | |
1015 | struct vring_desc *indirect) | |
3a4d5c94 MT |
1016 | { |
1017 | struct vring_desc desc; | |
1018 | unsigned int i = 0, count, found = 0; | |
1019 | int ret; | |
1020 | ||
1021 | /* Sanity check */ | |
7b3384fc | 1022 | if (unlikely(indirect->len % sizeof desc)) { |
3a4d5c94 MT |
1023 | vq_err(vq, "Invalid length in indirect descriptor: " |
1024 | "len 0x%llx not multiple of 0x%zx\n", | |
1025 | (unsigned long long)indirect->len, | |
1026 | sizeof desc); | |
1027 | return -EINVAL; | |
1028 | } | |
1029 | ||
1030 | ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect, | |
e0e9b406 | 1031 | UIO_MAXIOV); |
7b3384fc | 1032 | if (unlikely(ret < 0)) { |
3a4d5c94 MT |
1033 | vq_err(vq, "Translation failure %d in indirect.\n", ret); |
1034 | return ret; | |
1035 | } | |
1036 | ||
1037 | /* We will use the result as an address to read from, so most | |
1038 | * architectures only need a compiler barrier here. */ | |
1039 | read_barrier_depends(); | |
1040 | ||
1041 | count = indirect->len / sizeof desc; | |
1042 | /* Buffers are chained via a 16 bit next field, so | |
1043 | * we can have at most 2^16 of these. */ | |
7b3384fc | 1044 | if (unlikely(count > USHRT_MAX + 1)) { |
3a4d5c94 MT |
1045 | vq_err(vq, "Indirect buffer length too big: %d\n", |
1046 | indirect->len); | |
1047 | return -E2BIG; | |
1048 | } | |
1049 | ||
1050 | do { | |
1051 | unsigned iov_count = *in_num + *out_num; | |
7b3384fc | 1052 | if (unlikely(++found > count)) { |
3a4d5c94 MT |
1053 | vq_err(vq, "Loop detected: last one at %u " |
1054 | "indirect size %u\n", | |
1055 | i, count); | |
1056 | return -EINVAL; | |
1057 | } | |
d47effe1 KK |
1058 | if (unlikely(memcpy_fromiovec((unsigned char *)&desc, |
1059 | vq->indirect, sizeof desc))) { | |
3a4d5c94 MT |
1060 | vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n", |
1061 | i, (size_t)indirect->addr + i * sizeof desc); | |
1062 | return -EINVAL; | |
1063 | } | |
7b3384fc | 1064 | if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) { |
3a4d5c94 MT |
1065 | vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n", |
1066 | i, (size_t)indirect->addr + i * sizeof desc); | |
1067 | return -EINVAL; | |
1068 | } | |
1069 | ||
1070 | ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count, | |
1071 | iov_size - iov_count); | |
7b3384fc | 1072 | if (unlikely(ret < 0)) { |
3a4d5c94 MT |
1073 | vq_err(vq, "Translation failure %d indirect idx %d\n", |
1074 | ret, i); | |
1075 | return ret; | |
1076 | } | |
1077 | /* If this is an input descriptor, increment that count. */ | |
1078 | if (desc.flags & VRING_DESC_F_WRITE) { | |
1079 | *in_num += ret; | |
1080 | if (unlikely(log)) { | |
1081 | log[*log_num].addr = desc.addr; | |
1082 | log[*log_num].len = desc.len; | |
1083 | ++*log_num; | |
1084 | } | |
1085 | } else { | |
1086 | /* If it's an output descriptor, they're all supposed | |
1087 | * to come before any input descriptors. */ | |
7b3384fc | 1088 | if (unlikely(*in_num)) { |
3a4d5c94 MT |
1089 | vq_err(vq, "Indirect descriptor " |
1090 | "has out after in: idx %d\n", i); | |
1091 | return -EINVAL; | |
1092 | } | |
1093 | *out_num += ret; | |
1094 | } | |
1095 | } while ((i = next_desc(&desc)) != -1); | |
1096 | return 0; | |
1097 | } | |
1098 | ||
1099 | /* This looks in the virtqueue and for the first available buffer, and converts | |
1100 | * it to an iovec for convenient access. Since descriptors consist of some | |
1101 | * number of output then some number of input descriptors, it's actually two | |
1102 | * iovecs, but we pack them into one and note how many of each there were. | |
1103 | * | |
d5675bd2 MT |
1104 | * This function returns the descriptor number found, or vq->num (which is |
1105 | * never a valid descriptor number) if none was found. A negative code is | |
1106 | * returned on error. */ | |
1107 | int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq, | |
1108 | struct iovec iov[], unsigned int iov_size, | |
1109 | unsigned int *out_num, unsigned int *in_num, | |
1110 | struct vhost_log *log, unsigned int *log_num) | |
3a4d5c94 MT |
1111 | { |
1112 | struct vring_desc desc; | |
1113 | unsigned int i, head, found = 0; | |
1114 | u16 last_avail_idx; | |
1115 | int ret; | |
1116 | ||
1117 | /* Check it isn't doing very strange things with descriptor numbers. */ | |
1118 | last_avail_idx = vq->last_avail_idx; | |
8b7347aa | 1119 | if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) { |
3a4d5c94 MT |
1120 | vq_err(vq, "Failed to access avail idx at %p\n", |
1121 | &vq->avail->idx); | |
d5675bd2 | 1122 | return -EFAULT; |
3a4d5c94 MT |
1123 | } |
1124 | ||
7b3384fc | 1125 | if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) { |
3a4d5c94 MT |
1126 | vq_err(vq, "Guest moved used index from %u to %u", |
1127 | last_avail_idx, vq->avail_idx); | |
d5675bd2 | 1128 | return -EFAULT; |
3a4d5c94 MT |
1129 | } |
1130 | ||
1131 | /* If there's nothing new since last we looked, return invalid. */ | |
1132 | if (vq->avail_idx == last_avail_idx) | |
1133 | return vq->num; | |
1134 | ||
1135 | /* Only get avail ring entries after they have been exposed by guest. */ | |
5659338c | 1136 | smp_rmb(); |
3a4d5c94 MT |
1137 | |
1138 | /* Grab the next descriptor number they're advertising, and increment | |
1139 | * the index we've seen. */ | |
8b7347aa MT |
1140 | if (unlikely(__get_user(head, |
1141 | &vq->avail->ring[last_avail_idx % vq->num]))) { | |
3a4d5c94 MT |
1142 | vq_err(vq, "Failed to read head: idx %d address %p\n", |
1143 | last_avail_idx, | |
1144 | &vq->avail->ring[last_avail_idx % vq->num]); | |
d5675bd2 | 1145 | return -EFAULT; |
3a4d5c94 MT |
1146 | } |
1147 | ||
1148 | /* If their number is silly, that's an error. */ | |
7b3384fc | 1149 | if (unlikely(head >= vq->num)) { |
3a4d5c94 MT |
1150 | vq_err(vq, "Guest says index %u > %u is available", |
1151 | head, vq->num); | |
d5675bd2 | 1152 | return -EINVAL; |
3a4d5c94 MT |
1153 | } |
1154 | ||
1155 | /* When we start there are none of either input nor output. */ | |
1156 | *out_num = *in_num = 0; | |
1157 | if (unlikely(log)) | |
1158 | *log_num = 0; | |
1159 | ||
1160 | i = head; | |
1161 | do { | |
1162 | unsigned iov_count = *in_num + *out_num; | |
7b3384fc | 1163 | if (unlikely(i >= vq->num)) { |
3a4d5c94 MT |
1164 | vq_err(vq, "Desc index is %u > %u, head = %u", |
1165 | i, vq->num, head); | |
d5675bd2 | 1166 | return -EINVAL; |
3a4d5c94 | 1167 | } |
7b3384fc | 1168 | if (unlikely(++found > vq->num)) { |
3a4d5c94 MT |
1169 | vq_err(vq, "Loop detected: last one at %u " |
1170 | "vq size %u head %u\n", | |
1171 | i, vq->num, head); | |
d5675bd2 | 1172 | return -EINVAL; |
3a4d5c94 | 1173 | } |
fcc042a2 | 1174 | ret = __copy_from_user(&desc, vq->desc + i, sizeof desc); |
7b3384fc | 1175 | if (unlikely(ret)) { |
3a4d5c94 MT |
1176 | vq_err(vq, "Failed to get descriptor: idx %d addr %p\n", |
1177 | i, vq->desc + i); | |
d5675bd2 | 1178 | return -EFAULT; |
3a4d5c94 MT |
1179 | } |
1180 | if (desc.flags & VRING_DESC_F_INDIRECT) { | |
1181 | ret = get_indirect(dev, vq, iov, iov_size, | |
1182 | out_num, in_num, | |
1183 | log, log_num, &desc); | |
7b3384fc | 1184 | if (unlikely(ret < 0)) { |
3a4d5c94 MT |
1185 | vq_err(vq, "Failure detected " |
1186 | "in indirect descriptor at idx %d\n", i); | |
d5675bd2 | 1187 | return ret; |
3a4d5c94 MT |
1188 | } |
1189 | continue; | |
1190 | } | |
1191 | ||
1192 | ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count, | |
1193 | iov_size - iov_count); | |
7b3384fc | 1194 | if (unlikely(ret < 0)) { |
3a4d5c94 MT |
1195 | vq_err(vq, "Translation failure %d descriptor idx %d\n", |
1196 | ret, i); | |
d5675bd2 | 1197 | return ret; |
3a4d5c94 MT |
1198 | } |
1199 | if (desc.flags & VRING_DESC_F_WRITE) { | |
1200 | /* If this is an input descriptor, | |
1201 | * increment that count. */ | |
1202 | *in_num += ret; | |
1203 | if (unlikely(log)) { | |
1204 | log[*log_num].addr = desc.addr; | |
1205 | log[*log_num].len = desc.len; | |
1206 | ++*log_num; | |
1207 | } | |
1208 | } else { | |
1209 | /* If it's an output descriptor, they're all supposed | |
1210 | * to come before any input descriptors. */ | |
7b3384fc | 1211 | if (unlikely(*in_num)) { |
3a4d5c94 MT |
1212 | vq_err(vq, "Descriptor has out after in: " |
1213 | "idx %d\n", i); | |
d5675bd2 | 1214 | return -EINVAL; |
3a4d5c94 MT |
1215 | } |
1216 | *out_num += ret; | |
1217 | } | |
1218 | } while ((i = next_desc(&desc)) != -1); | |
1219 | ||
1220 | /* On success, increment avail index. */ | |
1221 | vq->last_avail_idx++; | |
1222 | return head; | |
1223 | } | |
1224 | ||
1225 | /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */ | |
8dd014ad | 1226 | void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n) |
3a4d5c94 | 1227 | { |
8dd014ad | 1228 | vq->last_avail_idx -= n; |
3a4d5c94 MT |
1229 | } |
1230 | ||
1231 | /* After we've used one of their buffers, we tell them about it. We'll then | |
1232 | * want to notify the guest, using eventfd. */ | |
1233 | int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len) | |
1234 | { | |
a8d3782f | 1235 | struct vring_used_elem __user *used; |
3a4d5c94 MT |
1236 | |
1237 | /* The virtqueue contains a ring of used buffers. Get a pointer to the | |
1238 | * next entry in that used ring. */ | |
1239 | used = &vq->used->ring[vq->last_used_idx % vq->num]; | |
8b7347aa | 1240 | if (__put_user(head, &used->id)) { |
3a4d5c94 MT |
1241 | vq_err(vq, "Failed to write used id"); |
1242 | return -EFAULT; | |
1243 | } | |
8b7347aa | 1244 | if (__put_user(len, &used->len)) { |
3a4d5c94 MT |
1245 | vq_err(vq, "Failed to write used len"); |
1246 | return -EFAULT; | |
1247 | } | |
1248 | /* Make sure buffer is written before we update index. */ | |
5659338c | 1249 | smp_wmb(); |
8b7347aa | 1250 | if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) { |
3a4d5c94 MT |
1251 | vq_err(vq, "Failed to increment used idx"); |
1252 | return -EFAULT; | |
1253 | } | |
1254 | if (unlikely(vq->log_used)) { | |
1255 | /* Make sure data is seen before log. */ | |
5659338c | 1256 | smp_wmb(); |
86e9424d MT |
1257 | /* Log used ring entry write. */ |
1258 | log_write(vq->log_base, | |
a8d3782f CH |
1259 | vq->log_addr + |
1260 | ((void __user *)used - (void __user *)vq->used), | |
86e9424d MT |
1261 | sizeof *used); |
1262 | /* Log used index update. */ | |
1263 | log_write(vq->log_base, | |
1264 | vq->log_addr + offsetof(struct vring_used, idx), | |
1265 | sizeof vq->used->idx); | |
3a4d5c94 MT |
1266 | if (vq->log_ctx) |
1267 | eventfd_signal(vq->log_ctx, 1); | |
1268 | } | |
1269 | vq->last_used_idx++; | |
1270 | return 0; | |
1271 | } | |
1272 | ||
8dd014ad DS |
1273 | static int __vhost_add_used_n(struct vhost_virtqueue *vq, |
1274 | struct vring_used_elem *heads, | |
1275 | unsigned count) | |
1276 | { | |
1277 | struct vring_used_elem __user *used; | |
1278 | int start; | |
1279 | ||
1280 | start = vq->last_used_idx % vq->num; | |
1281 | used = vq->used->ring + start; | |
dfe5ac5b | 1282 | if (__copy_to_user(used, heads, count * sizeof *used)) { |
8dd014ad DS |
1283 | vq_err(vq, "Failed to write used"); |
1284 | return -EFAULT; | |
1285 | } | |
1286 | if (unlikely(vq->log_used)) { | |
1287 | /* Make sure data is seen before log. */ | |
1288 | smp_wmb(); | |
1289 | /* Log used ring entry write. */ | |
1290 | log_write(vq->log_base, | |
1291 | vq->log_addr + | |
1292 | ((void __user *)used - (void __user *)vq->used), | |
1293 | count * sizeof *used); | |
1294 | } | |
1295 | vq->last_used_idx += count; | |
1296 | return 0; | |
1297 | } | |
1298 | ||
1299 | /* After we've used one of their buffers, we tell them about it. We'll then | |
1300 | * want to notify the guest, using eventfd. */ | |
1301 | int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads, | |
1302 | unsigned count) | |
1303 | { | |
1304 | int start, n, r; | |
1305 | ||
1306 | start = vq->last_used_idx % vq->num; | |
1307 | n = vq->num - start; | |
1308 | if (n < count) { | |
1309 | r = __vhost_add_used_n(vq, heads, n); | |
1310 | if (r < 0) | |
1311 | return r; | |
1312 | heads += n; | |
1313 | count -= n; | |
1314 | } | |
1315 | r = __vhost_add_used_n(vq, heads, count); | |
1316 | ||
1317 | /* Make sure buffer is written before we update index. */ | |
1318 | smp_wmb(); | |
1319 | if (put_user(vq->last_used_idx, &vq->used->idx)) { | |
1320 | vq_err(vq, "Failed to increment used idx"); | |
1321 | return -EFAULT; | |
1322 | } | |
1323 | if (unlikely(vq->log_used)) { | |
1324 | /* Log used index update. */ | |
1325 | log_write(vq->log_base, | |
1326 | vq->log_addr + offsetof(struct vring_used, idx), | |
1327 | sizeof vq->used->idx); | |
1328 | if (vq->log_ctx) | |
1329 | eventfd_signal(vq->log_ctx, 1); | |
1330 | } | |
1331 | return r; | |
1332 | } | |
1333 | ||
3a4d5c94 MT |
1334 | /* This actually signals the guest, using eventfd. */ |
1335 | void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq) | |
1336 | { | |
0d499356 | 1337 | __u16 flags; |
d47effe1 | 1338 | |
0d499356 MT |
1339 | /* Flush out used index updates. This is paired |
1340 | * with the barrier that the Guest executes when enabling | |
1341 | * interrupts. */ | |
1342 | smp_mb(); | |
1343 | ||
8b7347aa | 1344 | if (__get_user(flags, &vq->avail->flags)) { |
3a4d5c94 MT |
1345 | vq_err(vq, "Failed to get flags"); |
1346 | return; | |
1347 | } | |
1348 | ||
1349 | /* If they don't want an interrupt, don't signal, unless empty. */ | |
1350 | if ((flags & VRING_AVAIL_F_NO_INTERRUPT) && | |
1351 | (vq->avail_idx != vq->last_avail_idx || | |
1352 | !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY))) | |
1353 | return; | |
1354 | ||
1355 | /* Signal the Guest tell them we used something up. */ | |
1356 | if (vq->call_ctx) | |
1357 | eventfd_signal(vq->call_ctx, 1); | |
1358 | } | |
1359 | ||
1360 | /* And here's the combo meal deal. Supersize me! */ | |
1361 | void vhost_add_used_and_signal(struct vhost_dev *dev, | |
1362 | struct vhost_virtqueue *vq, | |
1363 | unsigned int head, int len) | |
1364 | { | |
1365 | vhost_add_used(vq, head, len); | |
1366 | vhost_signal(dev, vq); | |
1367 | } | |
1368 | ||
8dd014ad DS |
1369 | /* multi-buffer version of vhost_add_used_and_signal */ |
1370 | void vhost_add_used_and_signal_n(struct vhost_dev *dev, | |
1371 | struct vhost_virtqueue *vq, | |
1372 | struct vring_used_elem *heads, unsigned count) | |
1373 | { | |
1374 | vhost_add_used_n(vq, heads, count); | |
1375 | vhost_signal(dev, vq); | |
1376 | } | |
1377 | ||
3a4d5c94 MT |
1378 | /* OK, now we need to know about added descriptors. */ |
1379 | bool vhost_enable_notify(struct vhost_virtqueue *vq) | |
1380 | { | |
1381 | u16 avail_idx; | |
1382 | int r; | |
d47effe1 | 1383 | |
3a4d5c94 MT |
1384 | if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY)) |
1385 | return false; | |
1386 | vq->used_flags &= ~VRING_USED_F_NO_NOTIFY; | |
1387 | r = put_user(vq->used_flags, &vq->used->flags); | |
1388 | if (r) { | |
1389 | vq_err(vq, "Failed to enable notification at %p: %d\n", | |
1390 | &vq->used->flags, r); | |
1391 | return false; | |
1392 | } | |
1393 | /* They could have slipped one in as we were doing that: make | |
1394 | * sure it's written, then check again. */ | |
5659338c | 1395 | smp_mb(); |
8b7347aa | 1396 | r = __get_user(avail_idx, &vq->avail->idx); |
3a4d5c94 MT |
1397 | if (r) { |
1398 | vq_err(vq, "Failed to check avail idx at %p: %d\n", | |
1399 | &vq->avail->idx, r); | |
1400 | return false; | |
1401 | } | |
1402 | ||
8dd014ad | 1403 | return avail_idx != vq->avail_idx; |
3a4d5c94 MT |
1404 | } |
1405 | ||
1406 | /* We don't need to be notified again. */ | |
1407 | void vhost_disable_notify(struct vhost_virtqueue *vq) | |
1408 | { | |
1409 | int r; | |
d47effe1 | 1410 | |
3a4d5c94 MT |
1411 | if (vq->used_flags & VRING_USED_F_NO_NOTIFY) |
1412 | return; | |
1413 | vq->used_flags |= VRING_USED_F_NO_NOTIFY; | |
1414 | r = put_user(vq->used_flags, &vq->used->flags); | |
1415 | if (r) | |
1416 | vq_err(vq, "Failed to enable notification at %p: %d\n", | |
1417 | &vq->used->flags, r); | |
1418 | } |