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