1 /* Copyright (C) 2009 Red Hat, Inc. 2 * Copyright (C) 2006 Rusty Russell IBM Corporation 3 * 4 * Author: Michael S. Tsirkin <mst@redhat.com> 5 * 6 * Inspiration, some code, and most witty comments come from 7 * Documentation/lguest/lguest.c, by Rusty Russell 8 * 9 * This work is licensed under the terms of the GNU GPL, version 2. 10 * 11 * Generic code for virtio server in host kernel. 12 */ 13 14 #include <linux/eventfd.h> 15 #include <linux/vhost.h> 16 #include <linux/virtio_net.h> 17 #include <linux/mm.h> 18 #include <linux/miscdevice.h> 19 #include <linux/mutex.h> 20 #include <linux/rcupdate.h> 21 #include <linux/poll.h> 22 #include <linux/file.h> 23 #include <linux/highmem.h> 24 #include <linux/slab.h> 25 #include <linux/kthread.h> 26 #include <linux/cgroup.h> 27 28 #include <linux/net.h> 29 #include <linux/if_packet.h> 30 #include <linux/if_arp.h> 31 32 #include <net/sock.h> 33 34 #include "vhost.h" 35 36 enum { 37 VHOST_MEMORY_MAX_NREGIONS = 64, 38 VHOST_MEMORY_F_LOG = 0x1, 39 }; 40 41 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh, 42 poll_table *pt) 43 { 44 struct vhost_poll *poll; 45 poll = container_of(pt, struct vhost_poll, table); 46 47 poll->wqh = wqh; 48 add_wait_queue(wqh, &poll->wait); 49 } 50 51 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync, 52 void *key) 53 { 54 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait); 55 56 if (!((unsigned long)key & poll->mask)) 57 return 0; 58 59 vhost_poll_queue(poll); 60 return 0; 61 } 62 63 static void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn) 64 { 65 INIT_LIST_HEAD(&work->node); 66 work->fn = fn; 67 init_waitqueue_head(&work->done); 68 work->flushing = 0; 69 work->queue_seq = work->done_seq = 0; 70 } 71 72 /* Init poll structure */ 73 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn, 74 unsigned long mask, struct vhost_dev *dev) 75 { 76 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup); 77 init_poll_funcptr(&poll->table, vhost_poll_func); 78 poll->mask = mask; 79 poll->dev = dev; 80 81 vhost_work_init(&poll->work, fn); 82 } 83 84 /* Start polling a file. We add ourselves to file's wait queue. The caller must 85 * keep a reference to a file until after vhost_poll_stop is called. */ 86 void vhost_poll_start(struct vhost_poll *poll, struct file *file) 87 { 88 unsigned long mask; 89 mask = file->f_op->poll(file, &poll->table); 90 if (mask) 91 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask); 92 } 93 94 /* Stop polling a file. After this function returns, it becomes safe to drop the 95 * file reference. You must also flush afterwards. */ 96 void vhost_poll_stop(struct vhost_poll *poll) 97 { 98 remove_wait_queue(poll->wqh, &poll->wait); 99 } 100 101 static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work) 102 { 103 unsigned seq; 104 int left; 105 int flushing; 106 107 spin_lock_irq(&dev->work_lock); 108 seq = work->queue_seq; 109 work->flushing++; 110 spin_unlock_irq(&dev->work_lock); 111 wait_event(work->done, ({ 112 spin_lock_irq(&dev->work_lock); 113 left = seq - work->done_seq <= 0; 114 spin_unlock_irq(&dev->work_lock); 115 left; 116 })); 117 spin_lock_irq(&dev->work_lock); 118 flushing = --work->flushing; 119 spin_unlock_irq(&dev->work_lock); 120 BUG_ON(flushing < 0); 121 } 122 123 /* Flush any work that has been scheduled. When calling this, don't hold any 124 * locks that are also used by the callback. */ 125 void vhost_poll_flush(struct vhost_poll *poll) 126 { 127 vhost_work_flush(poll->dev, &poll->work); 128 } 129 130 static inline void vhost_work_queue(struct vhost_dev *dev, 131 struct vhost_work *work) 132 { 133 unsigned long flags; 134 135 spin_lock_irqsave(&dev->work_lock, flags); 136 if (list_empty(&work->node)) { 137 list_add_tail(&work->node, &dev->work_list); 138 work->queue_seq++; 139 wake_up_process(dev->worker); 140 } 141 spin_unlock_irqrestore(&dev->work_lock, flags); 142 } 143 144 void vhost_poll_queue(struct vhost_poll *poll) 145 { 146 vhost_work_queue(poll->dev, &poll->work); 147 } 148 149 static void vhost_vq_reset(struct vhost_dev *dev, 150 struct vhost_virtqueue *vq) 151 { 152 vq->num = 1; 153 vq->desc = NULL; 154 vq->avail = NULL; 155 vq->used = NULL; 156 vq->last_avail_idx = 0; 157 vq->avail_idx = 0; 158 vq->last_used_idx = 0; 159 vq->used_flags = 0; 160 vq->used_flags = 0; 161 vq->log_used = false; 162 vq->log_addr = -1ull; 163 vq->vhost_hlen = 0; 164 vq->sock_hlen = 0; 165 vq->private_data = NULL; 166 vq->log_base = NULL; 167 vq->error_ctx = NULL; 168 vq->error = NULL; 169 vq->kick = NULL; 170 vq->call_ctx = NULL; 171 vq->call = NULL; 172 vq->log_ctx = NULL; 173 } 174 175 static int vhost_worker(void *data) 176 { 177 struct vhost_dev *dev = data; 178 struct vhost_work *work = NULL; 179 unsigned uninitialized_var(seq); 180 181 for (;;) { 182 /* mb paired w/ kthread_stop */ 183 set_current_state(TASK_INTERRUPTIBLE); 184 185 spin_lock_irq(&dev->work_lock); 186 if (work) { 187 work->done_seq = seq; 188 if (work->flushing) 189 wake_up_all(&work->done); 190 } 191 192 if (kthread_should_stop()) { 193 spin_unlock_irq(&dev->work_lock); 194 __set_current_state(TASK_RUNNING); 195 return 0; 196 } 197 if (!list_empty(&dev->work_list)) { 198 work = list_first_entry(&dev->work_list, 199 struct vhost_work, node); 200 list_del_init(&work->node); 201 seq = work->queue_seq; 202 } else 203 work = NULL; 204 spin_unlock_irq(&dev->work_lock); 205 206 if (work) { 207 __set_current_state(TASK_RUNNING); 208 work->fn(work); 209 } else 210 schedule(); 211 212 } 213 } 214 215 /* Helper to allocate iovec buffers for all vqs. */ 216 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev) 217 { 218 int i; 219 for (i = 0; i < dev->nvqs; ++i) { 220 dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect * 221 UIO_MAXIOV, GFP_KERNEL); 222 dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV, 223 GFP_KERNEL); 224 dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads * 225 UIO_MAXIOV, GFP_KERNEL); 226 227 if (!dev->vqs[i].indirect || !dev->vqs[i].log || 228 !dev->vqs[i].heads) 229 goto err_nomem; 230 } 231 return 0; 232 err_nomem: 233 for (; i >= 0; --i) { 234 kfree(dev->vqs[i].indirect); 235 kfree(dev->vqs[i].log); 236 kfree(dev->vqs[i].heads); 237 } 238 return -ENOMEM; 239 } 240 241 static void vhost_dev_free_iovecs(struct vhost_dev *dev) 242 { 243 int i; 244 for (i = 0; i < dev->nvqs; ++i) { 245 kfree(dev->vqs[i].indirect); 246 dev->vqs[i].indirect = NULL; 247 kfree(dev->vqs[i].log); 248 dev->vqs[i].log = NULL; 249 kfree(dev->vqs[i].heads); 250 dev->vqs[i].heads = NULL; 251 } 252 } 253 254 long vhost_dev_init(struct vhost_dev *dev, 255 struct vhost_virtqueue *vqs, int nvqs) 256 { 257 int i; 258 259 dev->vqs = vqs; 260 dev->nvqs = nvqs; 261 mutex_init(&dev->mutex); 262 dev->log_ctx = NULL; 263 dev->log_file = NULL; 264 dev->memory = NULL; 265 dev->mm = NULL; 266 spin_lock_init(&dev->work_lock); 267 INIT_LIST_HEAD(&dev->work_list); 268 dev->worker = NULL; 269 270 for (i = 0; i < dev->nvqs; ++i) { 271 dev->vqs[i].log = NULL; 272 dev->vqs[i].indirect = NULL; 273 dev->vqs[i].heads = NULL; 274 dev->vqs[i].dev = dev; 275 mutex_init(&dev->vqs[i].mutex); 276 vhost_vq_reset(dev, dev->vqs + i); 277 if (dev->vqs[i].handle_kick) 278 vhost_poll_init(&dev->vqs[i].poll, 279 dev->vqs[i].handle_kick, POLLIN, dev); 280 } 281 282 return 0; 283 } 284 285 /* Caller should have device mutex */ 286 long vhost_dev_check_owner(struct vhost_dev *dev) 287 { 288 /* Are you the owner? If not, I don't think you mean to do that */ 289 return dev->mm == current->mm ? 0 : -EPERM; 290 } 291 292 struct vhost_attach_cgroups_struct { 293 struct vhost_work work; 294 struct task_struct *owner; 295 int ret; 296 }; 297 298 static void vhost_attach_cgroups_work(struct vhost_work *work) 299 { 300 struct vhost_attach_cgroups_struct *s; 301 s = container_of(work, struct vhost_attach_cgroups_struct, work); 302 s->ret = cgroup_attach_task_all(s->owner, current); 303 } 304 305 static int vhost_attach_cgroups(struct vhost_dev *dev) 306 { 307 struct vhost_attach_cgroups_struct attach; 308 attach.owner = current; 309 vhost_work_init(&attach.work, vhost_attach_cgroups_work); 310 vhost_work_queue(dev, &attach.work); 311 vhost_work_flush(dev, &attach.work); 312 return attach.ret; 313 } 314 315 /* Caller should have device mutex */ 316 static long vhost_dev_set_owner(struct vhost_dev *dev) 317 { 318 struct task_struct *worker; 319 int err; 320 /* Is there an owner already? */ 321 if (dev->mm) { 322 err = -EBUSY; 323 goto err_mm; 324 } 325 /* No owner, become one */ 326 dev->mm = get_task_mm(current); 327 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid); 328 if (IS_ERR(worker)) { 329 err = PTR_ERR(worker); 330 goto err_worker; 331 } 332 333 dev->worker = worker; 334 wake_up_process(worker); /* avoid contributing to loadavg */ 335 336 err = vhost_attach_cgroups(dev); 337 if (err) 338 goto err_cgroup; 339 340 err = vhost_dev_alloc_iovecs(dev); 341 if (err) 342 goto err_cgroup; 343 344 return 0; 345 err_cgroup: 346 kthread_stop(worker); 347 dev->worker = NULL; 348 err_worker: 349 if (dev->mm) 350 mmput(dev->mm); 351 dev->mm = NULL; 352 err_mm: 353 return err; 354 } 355 356 /* Caller should have device mutex */ 357 long vhost_dev_reset_owner(struct vhost_dev *dev) 358 { 359 struct vhost_memory *memory; 360 361 /* Restore memory to default empty mapping. */ 362 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL); 363 if (!memory) 364 return -ENOMEM; 365 366 vhost_dev_cleanup(dev); 367 368 memory->nregions = 0; 369 RCU_INIT_POINTER(dev->memory, memory); 370 return 0; 371 } 372 373 /* Caller should have device mutex */ 374 void vhost_dev_cleanup(struct vhost_dev *dev) 375 { 376 int i; 377 for (i = 0; i < dev->nvqs; ++i) { 378 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) { 379 vhost_poll_stop(&dev->vqs[i].poll); 380 vhost_poll_flush(&dev->vqs[i].poll); 381 } 382 if (dev->vqs[i].error_ctx) 383 eventfd_ctx_put(dev->vqs[i].error_ctx); 384 if (dev->vqs[i].error) 385 fput(dev->vqs[i].error); 386 if (dev->vqs[i].kick) 387 fput(dev->vqs[i].kick); 388 if (dev->vqs[i].call_ctx) 389 eventfd_ctx_put(dev->vqs[i].call_ctx); 390 if (dev->vqs[i].call) 391 fput(dev->vqs[i].call); 392 vhost_vq_reset(dev, dev->vqs + i); 393 } 394 vhost_dev_free_iovecs(dev); 395 if (dev->log_ctx) 396 eventfd_ctx_put(dev->log_ctx); 397 dev->log_ctx = NULL; 398 if (dev->log_file) 399 fput(dev->log_file); 400 dev->log_file = NULL; 401 /* No one will access memory at this point */ 402 kfree(rcu_dereference_protected(dev->memory, 403 lockdep_is_held(&dev->mutex))); 404 RCU_INIT_POINTER(dev->memory, NULL); 405 if (dev->mm) 406 mmput(dev->mm); 407 dev->mm = NULL; 408 409 WARN_ON(!list_empty(&dev->work_list)); 410 if (dev->worker) { 411 kthread_stop(dev->worker); 412 dev->worker = NULL; 413 } 414 } 415 416 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz) 417 { 418 u64 a = addr / VHOST_PAGE_SIZE / 8; 419 /* Make sure 64 bit math will not overflow. */ 420 if (a > ULONG_MAX - (unsigned long)log_base || 421 a + (unsigned long)log_base > ULONG_MAX) 422 return 0; 423 424 return access_ok(VERIFY_WRITE, log_base + a, 425 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8); 426 } 427 428 /* Caller should have vq mutex and device mutex. */ 429 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem, 430 int log_all) 431 { 432 int i; 433 434 if (!mem) 435 return 0; 436 437 for (i = 0; i < mem->nregions; ++i) { 438 struct vhost_memory_region *m = mem->regions + i; 439 unsigned long a = m->userspace_addr; 440 if (m->memory_size > ULONG_MAX) 441 return 0; 442 else if (!access_ok(VERIFY_WRITE, (void __user *)a, 443 m->memory_size)) 444 return 0; 445 else if (log_all && !log_access_ok(log_base, 446 m->guest_phys_addr, 447 m->memory_size)) 448 return 0; 449 } 450 return 1; 451 } 452 453 /* Can we switch to this memory table? */ 454 /* Caller should have device mutex but not vq mutex */ 455 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem, 456 int log_all) 457 { 458 int i; 459 for (i = 0; i < d->nvqs; ++i) { 460 int ok; 461 mutex_lock(&d->vqs[i].mutex); 462 /* If ring is inactive, will check when it's enabled. */ 463 if (d->vqs[i].private_data) 464 ok = vq_memory_access_ok(d->vqs[i].log_base, mem, 465 log_all); 466 else 467 ok = 1; 468 mutex_unlock(&d->vqs[i].mutex); 469 if (!ok) 470 return 0; 471 } 472 return 1; 473 } 474 475 static int vq_access_ok(unsigned int num, 476 struct vring_desc __user *desc, 477 struct vring_avail __user *avail, 478 struct vring_used __user *used) 479 { 480 return access_ok(VERIFY_READ, desc, num * sizeof *desc) && 481 access_ok(VERIFY_READ, avail, 482 sizeof *avail + num * sizeof *avail->ring) && 483 access_ok(VERIFY_WRITE, used, 484 sizeof *used + num * sizeof *used->ring); 485 } 486 487 /* Can we log writes? */ 488 /* Caller should have device mutex but not vq mutex */ 489 int vhost_log_access_ok(struct vhost_dev *dev) 490 { 491 struct vhost_memory *mp; 492 493 mp = rcu_dereference_protected(dev->memory, 494 lockdep_is_held(&dev->mutex)); 495 return memory_access_ok(dev, mp, 1); 496 } 497 498 /* Verify access for write logging. */ 499 /* Caller should have vq mutex and device mutex */ 500 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base) 501 { 502 struct vhost_memory *mp; 503 504 mp = rcu_dereference_protected(vq->dev->memory, 505 lockdep_is_held(&vq->mutex)); 506 return vq_memory_access_ok(log_base, mp, 507 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) && 508 (!vq->log_used || log_access_ok(log_base, vq->log_addr, 509 sizeof *vq->used + 510 vq->num * sizeof *vq->used->ring)); 511 } 512 513 /* Can we start vq? */ 514 /* Caller should have vq mutex and device mutex */ 515 int vhost_vq_access_ok(struct vhost_virtqueue *vq) 516 { 517 return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) && 518 vq_log_access_ok(vq, vq->log_base); 519 } 520 521 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m) 522 { 523 struct vhost_memory mem, *newmem, *oldmem; 524 unsigned long size = offsetof(struct vhost_memory, regions); 525 if (copy_from_user(&mem, m, size)) 526 return -EFAULT; 527 if (mem.padding) 528 return -EOPNOTSUPP; 529 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS) 530 return -E2BIG; 531 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL); 532 if (!newmem) 533 return -ENOMEM; 534 535 memcpy(newmem, &mem, size); 536 if (copy_from_user(newmem->regions, m->regions, 537 mem.nregions * sizeof *m->regions)) { 538 kfree(newmem); 539 return -EFAULT; 540 } 541 542 if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL))) { 543 kfree(newmem); 544 return -EFAULT; 545 } 546 oldmem = rcu_dereference_protected(d->memory, 547 lockdep_is_held(&d->mutex)); 548 rcu_assign_pointer(d->memory, newmem); 549 synchronize_rcu(); 550 kfree(oldmem); 551 return 0; 552 } 553 554 static int init_used(struct vhost_virtqueue *vq, 555 struct vring_used __user *used) 556 { 557 int r = put_user(vq->used_flags, &used->flags); 558 if (r) 559 return r; 560 return get_user(vq->last_used_idx, &used->idx); 561 } 562 563 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp) 564 { 565 struct file *eventfp, *filep = NULL, 566 *pollstart = NULL, *pollstop = NULL; 567 struct eventfd_ctx *ctx = NULL; 568 u32 __user *idxp = argp; 569 struct vhost_virtqueue *vq; 570 struct vhost_vring_state s; 571 struct vhost_vring_file f; 572 struct vhost_vring_addr a; 573 u32 idx; 574 long r; 575 576 r = get_user(idx, idxp); 577 if (r < 0) 578 return r; 579 if (idx >= d->nvqs) 580 return -ENOBUFS; 581 582 vq = d->vqs + idx; 583 584 mutex_lock(&vq->mutex); 585 586 switch (ioctl) { 587 case VHOST_SET_VRING_NUM: 588 /* Resizing ring with an active backend? 589 * You don't want to do that. */ 590 if (vq->private_data) { 591 r = -EBUSY; 592 break; 593 } 594 if (copy_from_user(&s, argp, sizeof s)) { 595 r = -EFAULT; 596 break; 597 } 598 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) { 599 r = -EINVAL; 600 break; 601 } 602 vq->num = s.num; 603 break; 604 case VHOST_SET_VRING_BASE: 605 /* Moving base with an active backend? 606 * You don't want to do that. */ 607 if (vq->private_data) { 608 r = -EBUSY; 609 break; 610 } 611 if (copy_from_user(&s, argp, sizeof s)) { 612 r = -EFAULT; 613 break; 614 } 615 if (s.num > 0xffff) { 616 r = -EINVAL; 617 break; 618 } 619 vq->last_avail_idx = s.num; 620 /* Forget the cached index value. */ 621 vq->avail_idx = vq->last_avail_idx; 622 break; 623 case VHOST_GET_VRING_BASE: 624 s.index = idx; 625 s.num = vq->last_avail_idx; 626 if (copy_to_user(argp, &s, sizeof s)) 627 r = -EFAULT; 628 break; 629 case VHOST_SET_VRING_ADDR: 630 if (copy_from_user(&a, argp, sizeof a)) { 631 r = -EFAULT; 632 break; 633 } 634 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) { 635 r = -EOPNOTSUPP; 636 break; 637 } 638 /* For 32bit, verify that the top 32bits of the user 639 data are set to zero. */ 640 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr || 641 (u64)(unsigned long)a.used_user_addr != a.used_user_addr || 642 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) { 643 r = -EFAULT; 644 break; 645 } 646 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) || 647 (a.used_user_addr & (sizeof *vq->used->ring - 1)) || 648 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) { 649 r = -EINVAL; 650 break; 651 } 652 653 /* We only verify access here if backend is configured. 654 * If it is not, we don't as size might not have been setup. 655 * We will verify when backend is configured. */ 656 if (vq->private_data) { 657 if (!vq_access_ok(vq->num, 658 (void __user *)(unsigned long)a.desc_user_addr, 659 (void __user *)(unsigned long)a.avail_user_addr, 660 (void __user *)(unsigned long)a.used_user_addr)) { 661 r = -EINVAL; 662 break; 663 } 664 665 /* Also validate log access for used ring if enabled. */ 666 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) && 667 !log_access_ok(vq->log_base, a.log_guest_addr, 668 sizeof *vq->used + 669 vq->num * sizeof *vq->used->ring)) { 670 r = -EINVAL; 671 break; 672 } 673 } 674 675 r = init_used(vq, (struct vring_used __user *)(unsigned long) 676 a.used_user_addr); 677 if (r) 678 break; 679 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG)); 680 vq->desc = (void __user *)(unsigned long)a.desc_user_addr; 681 vq->avail = (void __user *)(unsigned long)a.avail_user_addr; 682 vq->log_addr = a.log_guest_addr; 683 vq->used = (void __user *)(unsigned long)a.used_user_addr; 684 break; 685 case VHOST_SET_VRING_KICK: 686 if (copy_from_user(&f, argp, sizeof f)) { 687 r = -EFAULT; 688 break; 689 } 690 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); 691 if (IS_ERR(eventfp)) { 692 r = PTR_ERR(eventfp); 693 break; 694 } 695 if (eventfp != vq->kick) { 696 pollstop = filep = vq->kick; 697 pollstart = vq->kick = eventfp; 698 } else 699 filep = eventfp; 700 break; 701 case VHOST_SET_VRING_CALL: 702 if (copy_from_user(&f, argp, sizeof f)) { 703 r = -EFAULT; 704 break; 705 } 706 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); 707 if (IS_ERR(eventfp)) { 708 r = PTR_ERR(eventfp); 709 break; 710 } 711 if (eventfp != vq->call) { 712 filep = vq->call; 713 ctx = vq->call_ctx; 714 vq->call = eventfp; 715 vq->call_ctx = eventfp ? 716 eventfd_ctx_fileget(eventfp) : NULL; 717 } else 718 filep = eventfp; 719 break; 720 case VHOST_SET_VRING_ERR: 721 if (copy_from_user(&f, argp, sizeof f)) { 722 r = -EFAULT; 723 break; 724 } 725 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); 726 if (IS_ERR(eventfp)) { 727 r = PTR_ERR(eventfp); 728 break; 729 } 730 if (eventfp != vq->error) { 731 filep = vq->error; 732 vq->error = eventfp; 733 ctx = vq->error_ctx; 734 vq->error_ctx = eventfp ? 735 eventfd_ctx_fileget(eventfp) : NULL; 736 } else 737 filep = eventfp; 738 break; 739 default: 740 r = -ENOIOCTLCMD; 741 } 742 743 if (pollstop && vq->handle_kick) 744 vhost_poll_stop(&vq->poll); 745 746 if (ctx) 747 eventfd_ctx_put(ctx); 748 if (filep) 749 fput(filep); 750 751 if (pollstart && vq->handle_kick) 752 vhost_poll_start(&vq->poll, vq->kick); 753 754 mutex_unlock(&vq->mutex); 755 756 if (pollstop && vq->handle_kick) 757 vhost_poll_flush(&vq->poll); 758 return r; 759 } 760 761 /* Caller must have device mutex */ 762 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg) 763 { 764 void __user *argp = (void __user *)arg; 765 struct file *eventfp, *filep = NULL; 766 struct eventfd_ctx *ctx = NULL; 767 u64 p; 768 long r; 769 int i, fd; 770 771 /* If you are not the owner, you can become one */ 772 if (ioctl == VHOST_SET_OWNER) { 773 r = vhost_dev_set_owner(d); 774 goto done; 775 } 776 777 /* You must be the owner to do anything else */ 778 r = vhost_dev_check_owner(d); 779 if (r) 780 goto done; 781 782 switch (ioctl) { 783 case VHOST_SET_MEM_TABLE: 784 r = vhost_set_memory(d, argp); 785 break; 786 case VHOST_SET_LOG_BASE: 787 if (copy_from_user(&p, argp, sizeof p)) { 788 r = -EFAULT; 789 break; 790 } 791 if ((u64)(unsigned long)p != p) { 792 r = -EFAULT; 793 break; 794 } 795 for (i = 0; i < d->nvqs; ++i) { 796 struct vhost_virtqueue *vq; 797 void __user *base = (void __user *)(unsigned long)p; 798 vq = d->vqs + i; 799 mutex_lock(&vq->mutex); 800 /* If ring is inactive, will check when it's enabled. */ 801 if (vq->private_data && !vq_log_access_ok(vq, base)) 802 r = -EFAULT; 803 else 804 vq->log_base = base; 805 mutex_unlock(&vq->mutex); 806 } 807 break; 808 case VHOST_SET_LOG_FD: 809 r = get_user(fd, (int __user *)argp); 810 if (r < 0) 811 break; 812 eventfp = fd == -1 ? NULL : eventfd_fget(fd); 813 if (IS_ERR(eventfp)) { 814 r = PTR_ERR(eventfp); 815 break; 816 } 817 if (eventfp != d->log_file) { 818 filep = d->log_file; 819 ctx = d->log_ctx; 820 d->log_ctx = eventfp ? 821 eventfd_ctx_fileget(eventfp) : NULL; 822 } else 823 filep = eventfp; 824 for (i = 0; i < d->nvqs; ++i) { 825 mutex_lock(&d->vqs[i].mutex); 826 d->vqs[i].log_ctx = d->log_ctx; 827 mutex_unlock(&d->vqs[i].mutex); 828 } 829 if (ctx) 830 eventfd_ctx_put(ctx); 831 if (filep) 832 fput(filep); 833 break; 834 default: 835 r = vhost_set_vring(d, ioctl, argp); 836 break; 837 } 838 done: 839 return r; 840 } 841 842 static const struct vhost_memory_region *find_region(struct vhost_memory *mem, 843 __u64 addr, __u32 len) 844 { 845 struct vhost_memory_region *reg; 846 int i; 847 /* linear search is not brilliant, but we really have on the order of 6 848 * regions in practice */ 849 for (i = 0; i < mem->nregions; ++i) { 850 reg = mem->regions + i; 851 if (reg->guest_phys_addr <= addr && 852 reg->guest_phys_addr + reg->memory_size - 1 >= addr) 853 return reg; 854 } 855 return NULL; 856 } 857 858 /* TODO: This is really inefficient. We need something like get_user() 859 * (instruction directly accesses the data, with an exception table entry 860 * returning -EFAULT). See Documentation/x86/exception-tables.txt. 861 */ 862 static int set_bit_to_user(int nr, void __user *addr) 863 { 864 unsigned long log = (unsigned long)addr; 865 struct page *page; 866 void *base; 867 int bit = nr + (log % PAGE_SIZE) * 8; 868 int r; 869 r = get_user_pages_fast(log, 1, 1, &page); 870 if (r < 0) 871 return r; 872 BUG_ON(r != 1); 873 base = kmap_atomic(page, KM_USER0); 874 set_bit(bit, base); 875 kunmap_atomic(base, KM_USER0); 876 set_page_dirty_lock(page); 877 put_page(page); 878 return 0; 879 } 880 881 static int log_write(void __user *log_base, 882 u64 write_address, u64 write_length) 883 { 884 int r; 885 if (!write_length) 886 return 0; 887 write_address /= VHOST_PAGE_SIZE; 888 for (;;) { 889 u64 base = (u64)(unsigned long)log_base; 890 u64 log = base + write_address / 8; 891 int bit = write_address % 8; 892 if ((u64)(unsigned long)log != log) 893 return -EFAULT; 894 r = set_bit_to_user(bit, (void __user *)(unsigned long)log); 895 if (r < 0) 896 return r; 897 if (write_length <= VHOST_PAGE_SIZE) 898 break; 899 write_length -= VHOST_PAGE_SIZE; 900 write_address += VHOST_PAGE_SIZE; 901 } 902 return r; 903 } 904 905 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log, 906 unsigned int log_num, u64 len) 907 { 908 int i, r; 909 910 /* Make sure data written is seen before log. */ 911 smp_wmb(); 912 for (i = 0; i < log_num; ++i) { 913 u64 l = min(log[i].len, len); 914 r = log_write(vq->log_base, log[i].addr, l); 915 if (r < 0) 916 return r; 917 len -= l; 918 if (!len) { 919 if (vq->log_ctx) 920 eventfd_signal(vq->log_ctx, 1); 921 return 0; 922 } 923 } 924 /* Length written exceeds what we have stored. This is a bug. */ 925 BUG(); 926 return 0; 927 } 928 929 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len, 930 struct iovec iov[], int iov_size) 931 { 932 const struct vhost_memory_region *reg; 933 struct vhost_memory *mem; 934 struct iovec *_iov; 935 u64 s = 0; 936 int ret = 0; 937 938 rcu_read_lock(); 939 940 mem = rcu_dereference(dev->memory); 941 while ((u64)len > s) { 942 u64 size; 943 if (unlikely(ret >= iov_size)) { 944 ret = -ENOBUFS; 945 break; 946 } 947 reg = find_region(mem, addr, len); 948 if (unlikely(!reg)) { 949 ret = -EFAULT; 950 break; 951 } 952 _iov = iov + ret; 953 size = reg->memory_size - addr + reg->guest_phys_addr; 954 _iov->iov_len = min((u64)len, size); 955 _iov->iov_base = (void __user *)(unsigned long) 956 (reg->userspace_addr + addr - reg->guest_phys_addr); 957 s += size; 958 addr += size; 959 ++ret; 960 } 961 962 rcu_read_unlock(); 963 return ret; 964 } 965 966 /* Each buffer in the virtqueues is actually a chain of descriptors. This 967 * function returns the next descriptor in the chain, 968 * or -1U if we're at the end. */ 969 static unsigned next_desc(struct vring_desc *desc) 970 { 971 unsigned int next; 972 973 /* If this descriptor says it doesn't chain, we're done. */ 974 if (!(desc->flags & VRING_DESC_F_NEXT)) 975 return -1U; 976 977 /* Check they're not leading us off end of descriptors. */ 978 next = desc->next; 979 /* Make sure compiler knows to grab that: we don't want it changing! */ 980 /* We will use the result as an index in an array, so most 981 * architectures only need a compiler barrier here. */ 982 read_barrier_depends(); 983 984 return next; 985 } 986 987 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq, 988 struct iovec iov[], unsigned int iov_size, 989 unsigned int *out_num, unsigned int *in_num, 990 struct vhost_log *log, unsigned int *log_num, 991 struct vring_desc *indirect) 992 { 993 struct vring_desc desc; 994 unsigned int i = 0, count, found = 0; 995 int ret; 996 997 /* Sanity check */ 998 if (unlikely(indirect->len % sizeof desc)) { 999 vq_err(vq, "Invalid length in indirect descriptor: " 1000 "len 0x%llx not multiple of 0x%zx\n", 1001 (unsigned long long)indirect->len, 1002 sizeof desc); 1003 return -EINVAL; 1004 } 1005 1006 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect, 1007 UIO_MAXIOV); 1008 if (unlikely(ret < 0)) { 1009 vq_err(vq, "Translation failure %d in indirect.\n", ret); 1010 return ret; 1011 } 1012 1013 /* We will use the result as an address to read from, so most 1014 * architectures only need a compiler barrier here. */ 1015 read_barrier_depends(); 1016 1017 count = indirect->len / sizeof desc; 1018 /* Buffers are chained via a 16 bit next field, so 1019 * we can have at most 2^16 of these. */ 1020 if (unlikely(count > USHRT_MAX + 1)) { 1021 vq_err(vq, "Indirect buffer length too big: %d\n", 1022 indirect->len); 1023 return -E2BIG; 1024 } 1025 1026 do { 1027 unsigned iov_count = *in_num + *out_num; 1028 if (unlikely(++found > count)) { 1029 vq_err(vq, "Loop detected: last one at %u " 1030 "indirect size %u\n", 1031 i, count); 1032 return -EINVAL; 1033 } 1034 if (unlikely(memcpy_fromiovec((unsigned char *)&desc, vq->indirect, 1035 sizeof desc))) { 1036 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n", 1037 i, (size_t)indirect->addr + i * sizeof desc); 1038 return -EINVAL; 1039 } 1040 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) { 1041 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n", 1042 i, (size_t)indirect->addr + i * sizeof desc); 1043 return -EINVAL; 1044 } 1045 1046 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count, 1047 iov_size - iov_count); 1048 if (unlikely(ret < 0)) { 1049 vq_err(vq, "Translation failure %d indirect idx %d\n", 1050 ret, i); 1051 return ret; 1052 } 1053 /* If this is an input descriptor, increment that count. */ 1054 if (desc.flags & VRING_DESC_F_WRITE) { 1055 *in_num += ret; 1056 if (unlikely(log)) { 1057 log[*log_num].addr = desc.addr; 1058 log[*log_num].len = desc.len; 1059 ++*log_num; 1060 } 1061 } else { 1062 /* If it's an output descriptor, they're all supposed 1063 * to come before any input descriptors. */ 1064 if (unlikely(*in_num)) { 1065 vq_err(vq, "Indirect descriptor " 1066 "has out after in: idx %d\n", i); 1067 return -EINVAL; 1068 } 1069 *out_num += ret; 1070 } 1071 } while ((i = next_desc(&desc)) != -1); 1072 return 0; 1073 } 1074 1075 /* This looks in the virtqueue and for the first available buffer, and converts 1076 * it to an iovec for convenient access. Since descriptors consist of some 1077 * number of output then some number of input descriptors, it's actually two 1078 * iovecs, but we pack them into one and note how many of each there were. 1079 * 1080 * This function returns the descriptor number found, or vq->num (which is 1081 * never a valid descriptor number) if none was found. A negative code is 1082 * returned on error. */ 1083 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq, 1084 struct iovec iov[], unsigned int iov_size, 1085 unsigned int *out_num, unsigned int *in_num, 1086 struct vhost_log *log, unsigned int *log_num) 1087 { 1088 struct vring_desc desc; 1089 unsigned int i, head, found = 0; 1090 u16 last_avail_idx; 1091 int ret; 1092 1093 /* Check it isn't doing very strange things with descriptor numbers. */ 1094 last_avail_idx = vq->last_avail_idx; 1095 if (unlikely(get_user(vq->avail_idx, &vq->avail->idx))) { 1096 vq_err(vq, "Failed to access avail idx at %p\n", 1097 &vq->avail->idx); 1098 return -EFAULT; 1099 } 1100 1101 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) { 1102 vq_err(vq, "Guest moved used index from %u to %u", 1103 last_avail_idx, vq->avail_idx); 1104 return -EFAULT; 1105 } 1106 1107 /* If there's nothing new since last we looked, return invalid. */ 1108 if (vq->avail_idx == last_avail_idx) 1109 return vq->num; 1110 1111 /* Only get avail ring entries after they have been exposed by guest. */ 1112 smp_rmb(); 1113 1114 /* Grab the next descriptor number they're advertising, and increment 1115 * the index we've seen. */ 1116 if (unlikely(get_user(head, 1117 &vq->avail->ring[last_avail_idx % vq->num]))) { 1118 vq_err(vq, "Failed to read head: idx %d address %p\n", 1119 last_avail_idx, 1120 &vq->avail->ring[last_avail_idx % vq->num]); 1121 return -EFAULT; 1122 } 1123 1124 /* If their number is silly, that's an error. */ 1125 if (unlikely(head >= vq->num)) { 1126 vq_err(vq, "Guest says index %u > %u is available", 1127 head, vq->num); 1128 return -EINVAL; 1129 } 1130 1131 /* When we start there are none of either input nor output. */ 1132 *out_num = *in_num = 0; 1133 if (unlikely(log)) 1134 *log_num = 0; 1135 1136 i = head; 1137 do { 1138 unsigned iov_count = *in_num + *out_num; 1139 if (unlikely(i >= vq->num)) { 1140 vq_err(vq, "Desc index is %u > %u, head = %u", 1141 i, vq->num, head); 1142 return -EINVAL; 1143 } 1144 if (unlikely(++found > vq->num)) { 1145 vq_err(vq, "Loop detected: last one at %u " 1146 "vq size %u head %u\n", 1147 i, vq->num, head); 1148 return -EINVAL; 1149 } 1150 ret = copy_from_user(&desc, vq->desc + i, sizeof desc); 1151 if (unlikely(ret)) { 1152 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n", 1153 i, vq->desc + i); 1154 return -EFAULT; 1155 } 1156 if (desc.flags & VRING_DESC_F_INDIRECT) { 1157 ret = get_indirect(dev, vq, iov, iov_size, 1158 out_num, in_num, 1159 log, log_num, &desc); 1160 if (unlikely(ret < 0)) { 1161 vq_err(vq, "Failure detected " 1162 "in indirect descriptor at idx %d\n", i); 1163 return ret; 1164 } 1165 continue; 1166 } 1167 1168 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count, 1169 iov_size - iov_count); 1170 if (unlikely(ret < 0)) { 1171 vq_err(vq, "Translation failure %d descriptor idx %d\n", 1172 ret, i); 1173 return ret; 1174 } 1175 if (desc.flags & VRING_DESC_F_WRITE) { 1176 /* If this is an input descriptor, 1177 * increment that count. */ 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, "Descriptor has out after in: " 1189 "idx %d\n", i); 1190 return -EINVAL; 1191 } 1192 *out_num += ret; 1193 } 1194 } while ((i = next_desc(&desc)) != -1); 1195 1196 /* On success, increment avail index. */ 1197 vq->last_avail_idx++; 1198 return head; 1199 } 1200 1201 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */ 1202 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n) 1203 { 1204 vq->last_avail_idx -= n; 1205 } 1206 1207 /* After we've used one of their buffers, we tell them about it. We'll then 1208 * want to notify the guest, using eventfd. */ 1209 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len) 1210 { 1211 struct vring_used_elem __user *used; 1212 1213 /* The virtqueue contains a ring of used buffers. Get a pointer to the 1214 * next entry in that used ring. */ 1215 used = &vq->used->ring[vq->last_used_idx % vq->num]; 1216 if (put_user(head, &used->id)) { 1217 vq_err(vq, "Failed to write used id"); 1218 return -EFAULT; 1219 } 1220 if (put_user(len, &used->len)) { 1221 vq_err(vq, "Failed to write used len"); 1222 return -EFAULT; 1223 } 1224 /* Make sure buffer is written before we update index. */ 1225 smp_wmb(); 1226 if (put_user(vq->last_used_idx + 1, &vq->used->idx)) { 1227 vq_err(vq, "Failed to increment used idx"); 1228 return -EFAULT; 1229 } 1230 if (unlikely(vq->log_used)) { 1231 /* Make sure data is seen before log. */ 1232 smp_wmb(); 1233 /* Log used ring entry write. */ 1234 log_write(vq->log_base, 1235 vq->log_addr + 1236 ((void __user *)used - (void __user *)vq->used), 1237 sizeof *used); 1238 /* Log used index update. */ 1239 log_write(vq->log_base, 1240 vq->log_addr + offsetof(struct vring_used, idx), 1241 sizeof vq->used->idx); 1242 if (vq->log_ctx) 1243 eventfd_signal(vq->log_ctx, 1); 1244 } 1245 vq->last_used_idx++; 1246 return 0; 1247 } 1248 1249 static int __vhost_add_used_n(struct vhost_virtqueue *vq, 1250 struct vring_used_elem *heads, 1251 unsigned count) 1252 { 1253 struct vring_used_elem __user *used; 1254 int start; 1255 1256 start = vq->last_used_idx % vq->num; 1257 used = vq->used->ring + start; 1258 if (copy_to_user(used, heads, count * sizeof *used)) { 1259 vq_err(vq, "Failed to write used"); 1260 return -EFAULT; 1261 } 1262 if (unlikely(vq->log_used)) { 1263 /* Make sure data is seen before log. */ 1264 smp_wmb(); 1265 /* Log used ring entry write. */ 1266 log_write(vq->log_base, 1267 vq->log_addr + 1268 ((void __user *)used - (void __user *)vq->used), 1269 count * sizeof *used); 1270 } 1271 vq->last_used_idx += count; 1272 return 0; 1273 } 1274 1275 /* After we've used one of their buffers, we tell them about it. We'll then 1276 * want to notify the guest, using eventfd. */ 1277 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads, 1278 unsigned count) 1279 { 1280 int start, n, r; 1281 1282 start = vq->last_used_idx % vq->num; 1283 n = vq->num - start; 1284 if (n < count) { 1285 r = __vhost_add_used_n(vq, heads, n); 1286 if (r < 0) 1287 return r; 1288 heads += n; 1289 count -= n; 1290 } 1291 r = __vhost_add_used_n(vq, heads, count); 1292 1293 /* Make sure buffer is written before we update index. */ 1294 smp_wmb(); 1295 if (put_user(vq->last_used_idx, &vq->used->idx)) { 1296 vq_err(vq, "Failed to increment used idx"); 1297 return -EFAULT; 1298 } 1299 if (unlikely(vq->log_used)) { 1300 /* Log used index update. */ 1301 log_write(vq->log_base, 1302 vq->log_addr + offsetof(struct vring_used, idx), 1303 sizeof vq->used->idx); 1304 if (vq->log_ctx) 1305 eventfd_signal(vq->log_ctx, 1); 1306 } 1307 return r; 1308 } 1309 1310 /* This actually signals the guest, using eventfd. */ 1311 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq) 1312 { 1313 __u16 flags; 1314 /* Flush out used index updates. This is paired 1315 * with the barrier that the Guest executes when enabling 1316 * interrupts. */ 1317 smp_mb(); 1318 1319 if (get_user(flags, &vq->avail->flags)) { 1320 vq_err(vq, "Failed to get flags"); 1321 return; 1322 } 1323 1324 /* If they don't want an interrupt, don't signal, unless empty. */ 1325 if ((flags & VRING_AVAIL_F_NO_INTERRUPT) && 1326 (vq->avail_idx != vq->last_avail_idx || 1327 !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY))) 1328 return; 1329 1330 /* Signal the Guest tell them we used something up. */ 1331 if (vq->call_ctx) 1332 eventfd_signal(vq->call_ctx, 1); 1333 } 1334 1335 /* And here's the combo meal deal. Supersize me! */ 1336 void vhost_add_used_and_signal(struct vhost_dev *dev, 1337 struct vhost_virtqueue *vq, 1338 unsigned int head, int len) 1339 { 1340 vhost_add_used(vq, head, len); 1341 vhost_signal(dev, vq); 1342 } 1343 1344 /* multi-buffer version of vhost_add_used_and_signal */ 1345 void vhost_add_used_and_signal_n(struct vhost_dev *dev, 1346 struct vhost_virtqueue *vq, 1347 struct vring_used_elem *heads, unsigned count) 1348 { 1349 vhost_add_used_n(vq, heads, count); 1350 vhost_signal(dev, vq); 1351 } 1352 1353 /* OK, now we need to know about added descriptors. */ 1354 bool vhost_enable_notify(struct vhost_virtqueue *vq) 1355 { 1356 u16 avail_idx; 1357 int r; 1358 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY)) 1359 return false; 1360 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY; 1361 r = put_user(vq->used_flags, &vq->used->flags); 1362 if (r) { 1363 vq_err(vq, "Failed to enable notification at %p: %d\n", 1364 &vq->used->flags, r); 1365 return false; 1366 } 1367 /* They could have slipped one in as we were doing that: make 1368 * sure it's written, then check again. */ 1369 smp_mb(); 1370 r = get_user(avail_idx, &vq->avail->idx); 1371 if (r) { 1372 vq_err(vq, "Failed to check avail idx at %p: %d\n", 1373 &vq->avail->idx, r); 1374 return false; 1375 } 1376 1377 return avail_idx != vq->avail_idx; 1378 } 1379 1380 /* We don't need to be notified again. */ 1381 void vhost_disable_notify(struct vhost_virtqueue *vq) 1382 { 1383 int r; 1384 if (vq->used_flags & VRING_USED_F_NO_NOTIFY) 1385 return; 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 } 1392