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