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 36enum { 37 VHOST_MEMORY_MAX_NREGIONS = 64, 38 VHOST_MEMORY_F_LOG = 0x1, 39}; 40 41static 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 51static 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 63static 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 */ 73void 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. */ 86void 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. */ 96void vhost_poll_stop(struct vhost_poll *poll) 97{ 98 remove_wait_queue(poll->wqh, &poll->wait); 99} 100 101static 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. */ 125void vhost_poll_flush(struct vhost_poll *poll) 126{ 127 vhost_work_flush(poll->dev, &poll->work); 128} 129 130static 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 144void vhost_poll_queue(struct vhost_poll *poll) 145{ 146 vhost_work_queue(poll->dev, &poll->work); 147} 148 149static 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 175static 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 215long 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 */ 244long 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 250struct vhost_attach_cgroups_struct { 251 struct vhost_work work; 252 struct task_struct *owner; 253 int ret; 254}; 255 256static 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 263static 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 */ 274static 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; 299err_cgroup: 300 kthread_stop(worker); 301 dev->worker = NULL; 302err_worker: 303 if (dev->mm) 304 mmput(dev->mm); 305 dev->mm = NULL; 306err_mm: 307 return err; 308} 309 310/* Caller should have device mutex */ 311long 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 dev->memory = memory; 324 return 0; 325} 326 327/* Caller should have device mutex */ 328void 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(dev->memory); 356 dev->memory = NULL; 357 if (dev->mm) 358 mmput(dev->mm); 359 dev->mm = NULL; 360 361 WARN_ON(!list_empty(&dev->work_list)); 362 if (dev->worker) { 363 kthread_stop(dev->worker); 364 dev->worker = NULL; 365 } 366} 367 368static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz) 369{ 370 u64 a = addr / VHOST_PAGE_SIZE / 8; 371 /* Make sure 64 bit math will not overflow. */ 372 if (a > ULONG_MAX - (unsigned long)log_base || 373 a + (unsigned long)log_base > ULONG_MAX) 374 return -EFAULT; 375 376 return access_ok(VERIFY_WRITE, log_base + a, 377 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8); 378} 379 380/* Caller should have vq mutex and device mutex. */ 381static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem, 382 int log_all) 383{ 384 int i; 385 386 if (!mem) 387 return 0; 388 389 for (i = 0; i < mem->nregions; ++i) { 390 struct vhost_memory_region *m = mem->regions + i; 391 unsigned long a = m->userspace_addr; 392 if (m->memory_size > ULONG_MAX) 393 return 0; 394 else if (!access_ok(VERIFY_WRITE, (void __user *)a, 395 m->memory_size)) 396 return 0; 397 else if (log_all && !log_access_ok(log_base, 398 m->guest_phys_addr, 399 m->memory_size)) 400 return 0; 401 } 402 return 1; 403} 404 405/* Can we switch to this memory table? */ 406/* Caller should have device mutex but not vq mutex */ 407static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem, 408 int log_all) 409{ 410 int i; 411 for (i = 0; i < d->nvqs; ++i) { 412 int ok; 413 mutex_lock(&d->vqs[i].mutex); 414 /* If ring is inactive, will check when it's enabled. */ 415 if (d->vqs[i].private_data) 416 ok = vq_memory_access_ok(d->vqs[i].log_base, mem, 417 log_all); 418 else 419 ok = 1; 420 mutex_unlock(&d->vqs[i].mutex); 421 if (!ok) 422 return 0; 423 } 424 return 1; 425} 426 427static int vq_access_ok(unsigned int num, 428 struct vring_desc __user *desc, 429 struct vring_avail __user *avail, 430 struct vring_used __user *used) 431{ 432 return access_ok(VERIFY_READ, desc, num * sizeof *desc) && 433 access_ok(VERIFY_READ, avail, 434 sizeof *avail + num * sizeof *avail->ring) && 435 access_ok(VERIFY_WRITE, used, 436 sizeof *used + num * sizeof *used->ring); 437} 438 439/* Can we log writes? */ 440/* Caller should have device mutex but not vq mutex */ 441int vhost_log_access_ok(struct vhost_dev *dev) 442{ 443 return memory_access_ok(dev, dev->memory, 1); 444} 445 446/* Verify access for write logging. */ 447/* Caller should have vq mutex and device mutex */ 448static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base) 449{ 450 return vq_memory_access_ok(log_base, vq->dev->memory, 451 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) && 452 (!vq->log_used || log_access_ok(log_base, vq->log_addr, 453 sizeof *vq->used + 454 vq->num * sizeof *vq->used->ring)); 455} 456 457/* Can we start vq? */ 458/* Caller should have vq mutex and device mutex */ 459int vhost_vq_access_ok(struct vhost_virtqueue *vq) 460{ 461 return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) && 462 vq_log_access_ok(vq, vq->log_base); 463} 464 465static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m) 466{ 467 struct vhost_memory mem, *newmem, *oldmem; 468 unsigned long size = offsetof(struct vhost_memory, regions); 469 if (copy_from_user(&mem, m, size)) 470 return -EFAULT; 471 if (mem.padding) 472 return -EOPNOTSUPP; 473 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS) 474 return -E2BIG; 475 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL); 476 if (!newmem) 477 return -ENOMEM; 478 479 memcpy(newmem, &mem, size); 480 if (copy_from_user(newmem->regions, m->regions, 481 mem.nregions * sizeof *m->regions)) { 482 kfree(newmem); 483 return -EFAULT; 484 } 485 486 if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL))) { 487 kfree(newmem); 488 return -EFAULT; 489 } 490 oldmem = d->memory; 491 rcu_assign_pointer(d->memory, newmem); 492 synchronize_rcu(); 493 kfree(oldmem); 494 return 0; 495} 496 497static int init_used(struct vhost_virtqueue *vq, 498 struct vring_used __user *used) 499{ 500 int r = put_user(vq->used_flags, &used->flags); 501 if (r) 502 return r; 503 return get_user(vq->last_used_idx, &used->idx); 504} 505 506static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp) 507{ 508 struct file *eventfp, *filep = NULL, 509 *pollstart = NULL, *pollstop = NULL; 510 struct eventfd_ctx *ctx = NULL; 511 u32 __user *idxp = argp; 512 struct vhost_virtqueue *vq; 513 struct vhost_vring_state s; 514 struct vhost_vring_file f; 515 struct vhost_vring_addr a; 516 u32 idx; 517 long r; 518 519 r = get_user(idx, idxp); 520 if (r < 0) 521 return r; 522 if (idx >= d->nvqs) 523 return -ENOBUFS; 524 525 vq = d->vqs + idx; 526 527 mutex_lock(&vq->mutex); 528 529 switch (ioctl) { 530 case VHOST_SET_VRING_NUM: 531 /* Resizing ring with an active backend? 532 * You don't want to do that. */ 533 if (vq->private_data) { 534 r = -EBUSY; 535 break; 536 } 537 if (copy_from_user(&s, argp, sizeof s)) { 538 r = -EFAULT; 539 break; 540 } 541 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) { 542 r = -EINVAL; 543 break; 544 } 545 vq->num = s.num; 546 break; 547 case VHOST_SET_VRING_BASE: 548 /* Moving base with an active backend? 549 * You don't want to do that. */ 550 if (vq->private_data) { 551 r = -EBUSY; 552 break; 553 } 554 if (copy_from_user(&s, argp, sizeof s)) { 555 r = -EFAULT; 556 break; 557 } 558 if (s.num > 0xffff) { 559 r = -EINVAL; 560 break; 561 } 562 vq->last_avail_idx = s.num; 563 /* Forget the cached index value. */ 564 vq->avail_idx = vq->last_avail_idx; 565 break; 566 case VHOST_GET_VRING_BASE: 567 s.index = idx; 568 s.num = vq->last_avail_idx; 569 if (copy_to_user(argp, &s, sizeof s)) 570 r = -EFAULT; 571 break; 572 case VHOST_SET_VRING_ADDR: 573 if (copy_from_user(&a, argp, sizeof a)) { 574 r = -EFAULT; 575 break; 576 } 577 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) { 578 r = -EOPNOTSUPP; 579 break; 580 } 581 /* For 32bit, verify that the top 32bits of the user 582 data are set to zero. */ 583 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr || 584 (u64)(unsigned long)a.used_user_addr != a.used_user_addr || 585 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) { 586 r = -EFAULT; 587 break; 588 } 589 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) || 590 (a.used_user_addr & (sizeof *vq->used->ring - 1)) || 591 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) { 592 r = -EINVAL; 593 break; 594 } 595 596 /* We only verify access here if backend is configured. 597 * If it is not, we don't as size might not have been setup. 598 * We will verify when backend is configured. */ 599 if (vq->private_data) { 600 if (!vq_access_ok(vq->num, 601 (void __user *)(unsigned long)a.desc_user_addr, 602 (void __user *)(unsigned long)a.avail_user_addr, 603 (void __user *)(unsigned long)a.used_user_addr)) { 604 r = -EINVAL; 605 break; 606 } 607 608 /* Also validate log access for used ring if enabled. */ 609 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) && 610 !log_access_ok(vq->log_base, a.log_guest_addr, 611 sizeof *vq->used + 612 vq->num * sizeof *vq->used->ring)) { 613 r = -EINVAL; 614 break; 615 } 616 } 617 618 r = init_used(vq, (struct vring_used __user *)(unsigned long) 619 a.used_user_addr); 620 if (r) 621 break; 622 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG)); 623 vq->desc = (void __user *)(unsigned long)a.desc_user_addr; 624 vq->avail = (void __user *)(unsigned long)a.avail_user_addr; 625 vq->log_addr = a.log_guest_addr; 626 vq->used = (void __user *)(unsigned long)a.used_user_addr; 627 break; 628 case VHOST_SET_VRING_KICK: 629 if (copy_from_user(&f, argp, sizeof f)) { 630 r = -EFAULT; 631 break; 632 } 633 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); 634 if (IS_ERR(eventfp)) { 635 r = PTR_ERR(eventfp); 636 break; 637 } 638 if (eventfp != vq->kick) { 639 pollstop = filep = vq->kick; 640 pollstart = vq->kick = eventfp; 641 } else 642 filep = eventfp; 643 break; 644 case VHOST_SET_VRING_CALL: 645 if (copy_from_user(&f, argp, sizeof f)) { 646 r = -EFAULT; 647 break; 648 } 649 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); 650 if (IS_ERR(eventfp)) { 651 r = PTR_ERR(eventfp); 652 break; 653 } 654 if (eventfp != vq->call) { 655 filep = vq->call; 656 ctx = vq->call_ctx; 657 vq->call = eventfp; 658 vq->call_ctx = eventfp ? 659 eventfd_ctx_fileget(eventfp) : NULL; 660 } else 661 filep = eventfp; 662 break; 663 case VHOST_SET_VRING_ERR: 664 if (copy_from_user(&f, argp, sizeof f)) { 665 r = -EFAULT; 666 break; 667 } 668 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); 669 if (IS_ERR(eventfp)) { 670 r = PTR_ERR(eventfp); 671 break; 672 } 673 if (eventfp != vq->error) { 674 filep = vq->error; 675 vq->error = eventfp; 676 ctx = vq->error_ctx; 677 vq->error_ctx = eventfp ? 678 eventfd_ctx_fileget(eventfp) : NULL; 679 } else 680 filep = eventfp; 681 break; 682 default: 683 r = -ENOIOCTLCMD; 684 } 685 686 if (pollstop && vq->handle_kick) 687 vhost_poll_stop(&vq->poll); 688 689 if (ctx) 690 eventfd_ctx_put(ctx); 691 if (filep) 692 fput(filep); 693 694 if (pollstart && vq->handle_kick) 695 vhost_poll_start(&vq->poll, vq->kick); 696 697 mutex_unlock(&vq->mutex); 698 699 if (pollstop && vq->handle_kick) 700 vhost_poll_flush(&vq->poll); 701 return r; 702} 703 704/* Caller must have device mutex */ 705long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg) 706{ 707 void __user *argp = (void __user *)arg; 708 struct file *eventfp, *filep = NULL; 709 struct eventfd_ctx *ctx = NULL; 710 u64 p; 711 long r; 712 int i, fd; 713 714 /* If you are not the owner, you can become one */ 715 if (ioctl == VHOST_SET_OWNER) { 716 r = vhost_dev_set_owner(d); 717 goto done; 718 } 719 720 /* You must be the owner to do anything else */ 721 r = vhost_dev_check_owner(d); 722 if (r) 723 goto done; 724 725 switch (ioctl) { 726 case VHOST_SET_MEM_TABLE: 727 r = vhost_set_memory(d, argp); 728 break; 729 case VHOST_SET_LOG_BASE: 730 if (copy_from_user(&p, argp, sizeof p)) { 731 r = -EFAULT; 732 break; 733 } 734 if ((u64)(unsigned long)p != p) { 735 r = -EFAULT; 736 break; 737 } 738 for (i = 0; i < d->nvqs; ++i) { 739 struct vhost_virtqueue *vq; 740 void __user *base = (void __user *)(unsigned long)p; 741 vq = d->vqs + i; 742 mutex_lock(&vq->mutex); 743 /* If ring is inactive, will check when it's enabled. */ 744 if (vq->private_data && !vq_log_access_ok(vq, base)) 745 r = -EFAULT; 746 else 747 vq->log_base = base; 748 mutex_unlock(&vq->mutex); 749 } 750 break; 751 case VHOST_SET_LOG_FD: 752 r = get_user(fd, (int __user *)argp); 753 if (r < 0) 754 break; 755 eventfp = fd == -1 ? NULL : eventfd_fget(fd); 756 if (IS_ERR(eventfp)) { 757 r = PTR_ERR(eventfp); 758 break; 759 } 760 if (eventfp != d->log_file) { 761 filep = d->log_file; 762 ctx = d->log_ctx; 763 d->log_ctx = eventfp ? 764 eventfd_ctx_fileget(eventfp) : NULL; 765 } else 766 filep = eventfp; 767 for (i = 0; i < d->nvqs; ++i) { 768 mutex_lock(&d->vqs[i].mutex); 769 d->vqs[i].log_ctx = d->log_ctx; 770 mutex_unlock(&d->vqs[i].mutex); 771 } 772 if (ctx) 773 eventfd_ctx_put(ctx); 774 if (filep) 775 fput(filep); 776 break; 777 default: 778 r = vhost_set_vring(d, ioctl, argp); 779 break; 780 } 781done: 782 return r; 783} 784 785static const struct vhost_memory_region *find_region(struct vhost_memory *mem, 786 __u64 addr, __u32 len) 787{ 788 struct vhost_memory_region *reg; 789 int i; 790 /* linear search is not brilliant, but we really have on the order of 6 791 * regions in practice */ 792 for (i = 0; i < mem->nregions; ++i) { 793 reg = mem->regions + i; 794 if (reg->guest_phys_addr <= addr && 795 reg->guest_phys_addr + reg->memory_size - 1 >= addr) 796 return reg; 797 } 798 return NULL; 799} 800 801/* TODO: This is really inefficient. We need something like get_user() 802 * (instruction directly accesses the data, with an exception table entry 803 * returning -EFAULT). See Documentation/x86/exception-tables.txt. 804 */ 805static int set_bit_to_user(int nr, void __user *addr) 806{ 807 unsigned long log = (unsigned long)addr; 808 struct page *page; 809 void *base; 810 int bit = nr + (log % PAGE_SIZE) * 8; 811 int r; 812 r = get_user_pages_fast(log, 1, 1, &page); 813 if (r < 0) 814 return r; 815 BUG_ON(r != 1); 816 base = kmap_atomic(page, KM_USER0); 817 set_bit(bit, base); 818 kunmap_atomic(base, KM_USER0); 819 set_page_dirty_lock(page); 820 put_page(page); 821 return 0; 822} 823 824static int log_write(void __user *log_base, 825 u64 write_address, u64 write_length) 826{ 827 int r; 828 if (!write_length) 829 return 0; 830 write_address /= VHOST_PAGE_SIZE; 831 for (;;) { 832 u64 base = (u64)(unsigned long)log_base; 833 u64 log = base + write_address / 8; 834 int bit = write_address % 8; 835 if ((u64)(unsigned long)log != log) 836 return -EFAULT; 837 r = set_bit_to_user(bit, (void __user *)(unsigned long)log); 838 if (r < 0) 839 return r; 840 if (write_length <= VHOST_PAGE_SIZE) 841 break; 842 write_length -= VHOST_PAGE_SIZE; 843 write_address += VHOST_PAGE_SIZE; 844 } 845 return r; 846} 847 848int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log, 849 unsigned int log_num, u64 len) 850{ 851 int i, r; 852 853 /* Make sure data written is seen before log. */ 854 smp_wmb(); 855 for (i = 0; i < log_num; ++i) { 856 u64 l = min(log[i].len, len); 857 r = log_write(vq->log_base, log[i].addr, l); 858 if (r < 0) 859 return r; 860 len -= l; 861 if (!len) { 862 if (vq->log_ctx) 863 eventfd_signal(vq->log_ctx, 1); 864 return 0; 865 } 866 } 867 /* Length written exceeds what we have stored. This is a bug. */ 868 BUG(); 869 return 0; 870} 871 872static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len, 873 struct iovec iov[], int iov_size) 874{ 875 const struct vhost_memory_region *reg; 876 struct vhost_memory *mem; 877 struct iovec *_iov; 878 u64 s = 0; 879 int ret = 0; 880 881 rcu_read_lock(); 882 883 mem = rcu_dereference(dev->memory); 884 while ((u64)len > s) { 885 u64 size; 886 if (unlikely(ret >= iov_size)) { 887 ret = -ENOBUFS; 888 break; 889 } 890 reg = find_region(mem, addr, len); 891 if (unlikely(!reg)) { 892 ret = -EFAULT; 893 break; 894 } 895 _iov = iov + ret; 896 size = reg->memory_size - addr + reg->guest_phys_addr; 897 _iov->iov_len = min((u64)len, size); 898 _iov->iov_base = (void __user *)(unsigned long) 899 (reg->userspace_addr + addr - reg->guest_phys_addr); 900 s += size; 901 addr += size; 902 ++ret; 903 } 904 905 rcu_read_unlock(); 906 return ret; 907} 908 909/* Each buffer in the virtqueues is actually a chain of descriptors. This 910 * function returns the next descriptor in the chain, 911 * or -1U if we're at the end. */ 912static unsigned next_desc(struct vring_desc *desc) 913{ 914 unsigned int next; 915 916 /* If this descriptor says it doesn't chain, we're done. */ 917 if (!(desc->flags & VRING_DESC_F_NEXT)) 918 return -1U; 919 920 /* Check they're not leading us off end of descriptors. */ 921 next = desc->next; 922 /* Make sure compiler knows to grab that: we don't want it changing! */ 923 /* We will use the result as an index in an array, so most 924 * architectures only need a compiler barrier here. */ 925 read_barrier_depends(); 926 927 return next; 928} 929 930static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq, 931 struct iovec iov[], unsigned int iov_size, 932 unsigned int *out_num, unsigned int *in_num, 933 struct vhost_log *log, unsigned int *log_num, 934 struct vring_desc *indirect) 935{ 936 struct vring_desc desc; 937 unsigned int i = 0, count, found = 0; 938 int ret; 939 940 /* Sanity check */ 941 if (unlikely(indirect->len % sizeof desc)) { 942 vq_err(vq, "Invalid length in indirect descriptor: " 943 "len 0x%llx not multiple of 0x%zx\n", 944 (unsigned long long)indirect->len, 945 sizeof desc); 946 return -EINVAL; 947 } 948 949 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect, 950 ARRAY_SIZE(vq->indirect)); 951 if (unlikely(ret < 0)) { 952 vq_err(vq, "Translation failure %d in indirect.\n", ret); 953 return ret; 954 } 955 956 /* We will use the result as an address to read from, so most 957 * architectures only need a compiler barrier here. */ 958 read_barrier_depends(); 959 960 count = indirect->len / sizeof desc; 961 /* Buffers are chained via a 16 bit next field, so 962 * we can have at most 2^16 of these. */ 963 if (unlikely(count > USHRT_MAX + 1)) { 964 vq_err(vq, "Indirect buffer length too big: %d\n", 965 indirect->len); 966 return -E2BIG; 967 } 968 969 do { 970 unsigned iov_count = *in_num + *out_num; 971 if (unlikely(++found > count)) { 972 vq_err(vq, "Loop detected: last one at %u " 973 "indirect size %u\n", 974 i, count); 975 return -EINVAL; 976 } 977 if (unlikely(memcpy_fromiovec((unsigned char *)&desc, vq->indirect, 978 sizeof desc))) { 979 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n", 980 i, (size_t)indirect->addr + i * sizeof desc); 981 return -EINVAL; 982 } 983 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) { 984 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n", 985 i, (size_t)indirect->addr + i * sizeof desc); 986 return -EINVAL; 987 } 988 989 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count, 990 iov_size - iov_count); 991 if (unlikely(ret < 0)) { 992 vq_err(vq, "Translation failure %d indirect idx %d\n", 993 ret, i); 994 return ret; 995 } 996 /* If this is an input descriptor, increment that count. */ 997 if (desc.flags & VRING_DESC_F_WRITE) { 998 *in_num += ret; 999 if (unlikely(log)) { 1000 log[*log_num].addr = desc.addr; 1001 log[*log_num].len = desc.len; 1002 ++*log_num; 1003 } 1004 } else { 1005 /* If it's an output descriptor, they're all supposed 1006 * to come before any input descriptors. */ 1007 if (unlikely(*in_num)) { 1008 vq_err(vq, "Indirect descriptor " 1009 "has out after in: idx %d\n", i); 1010 return -EINVAL; 1011 } 1012 *out_num += ret; 1013 } 1014 } while ((i = next_desc(&desc)) != -1); 1015 return 0; 1016} 1017 1018/* This looks in the virtqueue and for the first available buffer, and converts 1019 * it to an iovec for convenient access. Since descriptors consist of some 1020 * number of output then some number of input descriptors, it's actually two 1021 * iovecs, but we pack them into one and note how many of each there were. 1022 * 1023 * This function returns the descriptor number found, or vq->num (which is 1024 * never a valid descriptor number) if none was found. A negative code is 1025 * returned on error. */ 1026int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq, 1027 struct iovec iov[], unsigned int iov_size, 1028 unsigned int *out_num, unsigned int *in_num, 1029 struct vhost_log *log, unsigned int *log_num) 1030{ 1031 struct vring_desc desc; 1032 unsigned int i, head, found = 0; 1033 u16 last_avail_idx; 1034 int ret; 1035 1036 /* Check it isn't doing very strange things with descriptor numbers. */ 1037 last_avail_idx = vq->last_avail_idx; 1038 if (unlikely(get_user(vq->avail_idx, &vq->avail->idx))) { 1039 vq_err(vq, "Failed to access avail idx at %p\n", 1040 &vq->avail->idx); 1041 return -EFAULT; 1042 } 1043 1044 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) { 1045 vq_err(vq, "Guest moved used index from %u to %u", 1046 last_avail_idx, vq->avail_idx); 1047 return -EFAULT; 1048 } 1049 1050 /* If there's nothing new since last we looked, return invalid. */ 1051 if (vq->avail_idx == last_avail_idx) 1052 return vq->num; 1053 1054 /* Only get avail ring entries after they have been exposed by guest. */ 1055 smp_rmb(); 1056 1057 /* Grab the next descriptor number they're advertising, and increment 1058 * the index we've seen. */ 1059 if (unlikely(get_user(head, 1060 &vq->avail->ring[last_avail_idx % vq->num]))) { 1061 vq_err(vq, "Failed to read head: idx %d address %p\n", 1062 last_avail_idx, 1063 &vq->avail->ring[last_avail_idx % vq->num]); 1064 return -EFAULT; 1065 } 1066 1067 /* If their number is silly, that's an error. */ 1068 if (unlikely(head >= vq->num)) { 1069 vq_err(vq, "Guest says index %u > %u is available", 1070 head, vq->num); 1071 return -EINVAL; 1072 } 1073 1074 /* When we start there are none of either input nor output. */ 1075 *out_num = *in_num = 0; 1076 if (unlikely(log)) 1077 *log_num = 0; 1078 1079 i = head; 1080 do { 1081 unsigned iov_count = *in_num + *out_num; 1082 if (unlikely(i >= vq->num)) { 1083 vq_err(vq, "Desc index is %u > %u, head = %u", 1084 i, vq->num, head); 1085 return -EINVAL; 1086 } 1087 if (unlikely(++found > vq->num)) { 1088 vq_err(vq, "Loop detected: last one at %u " 1089 "vq size %u head %u\n", 1090 i, vq->num, head); 1091 return -EINVAL; 1092 } 1093 ret = copy_from_user(&desc, vq->desc + i, sizeof desc); 1094 if (unlikely(ret)) { 1095 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n", 1096 i, vq->desc + i); 1097 return -EFAULT; 1098 } 1099 if (desc.flags & VRING_DESC_F_INDIRECT) { 1100 ret = get_indirect(dev, vq, iov, iov_size, 1101 out_num, in_num, 1102 log, log_num, &desc); 1103 if (unlikely(ret < 0)) { 1104 vq_err(vq, "Failure detected " 1105 "in indirect descriptor at idx %d\n", i); 1106 return ret; 1107 } 1108 continue; 1109 } 1110 1111 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count, 1112 iov_size - iov_count); 1113 if (unlikely(ret < 0)) { 1114 vq_err(vq, "Translation failure %d descriptor idx %d\n", 1115 ret, i); 1116 return ret; 1117 } 1118 if (desc.flags & VRING_DESC_F_WRITE) { 1119 /* If this is an input descriptor, 1120 * increment that count. */ 1121 *in_num += ret; 1122 if (unlikely(log)) { 1123 log[*log_num].addr = desc.addr; 1124 log[*log_num].len = desc.len; 1125 ++*log_num; 1126 } 1127 } else { 1128 /* If it's an output descriptor, they're all supposed 1129 * to come before any input descriptors. */ 1130 if (unlikely(*in_num)) { 1131 vq_err(vq, "Descriptor has out after in: " 1132 "idx %d\n", i); 1133 return -EINVAL; 1134 } 1135 *out_num += ret; 1136 } 1137 } while ((i = next_desc(&desc)) != -1); 1138 1139 /* On success, increment avail index. */ 1140 vq->last_avail_idx++; 1141 return head; 1142} 1143 1144/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */ 1145void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n) 1146{ 1147 vq->last_avail_idx -= n; 1148} 1149 1150/* After we've used one of their buffers, we tell them about it. We'll then 1151 * want to notify the guest, using eventfd. */ 1152int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len) 1153{ 1154 struct vring_used_elem __user *used; 1155 1156 /* The virtqueue contains a ring of used buffers. Get a pointer to the 1157 * next entry in that used ring. */ 1158 used = &vq->used->ring[vq->last_used_idx % vq->num]; 1159 if (put_user(head, &used->id)) { 1160 vq_err(vq, "Failed to write used id"); 1161 return -EFAULT; 1162 } 1163 if (put_user(len, &used->len)) { 1164 vq_err(vq, "Failed to write used len"); 1165 return -EFAULT; 1166 } 1167 /* Make sure buffer is written before we update index. */ 1168 smp_wmb(); 1169 if (put_user(vq->last_used_idx + 1, &vq->used->idx)) { 1170 vq_err(vq, "Failed to increment used idx"); 1171 return -EFAULT; 1172 } 1173 if (unlikely(vq->log_used)) { 1174 /* Make sure data is seen before log. */ 1175 smp_wmb(); 1176 /* Log used ring entry write. */ 1177 log_write(vq->log_base, 1178 vq->log_addr + 1179 ((void __user *)used - (void __user *)vq->used), 1180 sizeof *used); 1181 /* Log used index update. */ 1182 log_write(vq->log_base, 1183 vq->log_addr + offsetof(struct vring_used, idx), 1184 sizeof vq->used->idx); 1185 if (vq->log_ctx) 1186 eventfd_signal(vq->log_ctx, 1); 1187 } 1188 vq->last_used_idx++; 1189 return 0; 1190} 1191 1192static int __vhost_add_used_n(struct vhost_virtqueue *vq, 1193 struct vring_used_elem *heads, 1194 unsigned count) 1195{ 1196 struct vring_used_elem __user *used; 1197 int start; 1198 1199 start = vq->last_used_idx % vq->num; 1200 used = vq->used->ring + start; 1201 if (copy_to_user(used, heads, count * sizeof *used)) { 1202 vq_err(vq, "Failed to write used"); 1203 return -EFAULT; 1204 } 1205 if (unlikely(vq->log_used)) { 1206 /* Make sure data is seen before log. */ 1207 smp_wmb(); 1208 /* Log used ring entry write. */ 1209 log_write(vq->log_base, 1210 vq->log_addr + 1211 ((void __user *)used - (void __user *)vq->used), 1212 count * sizeof *used); 1213 } 1214 vq->last_used_idx += count; 1215 return 0; 1216} 1217 1218/* After we've used one of their buffers, we tell them about it. We'll then 1219 * want to notify the guest, using eventfd. */ 1220int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads, 1221 unsigned count) 1222{ 1223 int start, n, r; 1224 1225 start = vq->last_used_idx % vq->num; 1226 n = vq->num - start; 1227 if (n < count) { 1228 r = __vhost_add_used_n(vq, heads, n); 1229 if (r < 0) 1230 return r; 1231 heads += n; 1232 count -= n; 1233 } 1234 r = __vhost_add_used_n(vq, heads, count); 1235 1236 /* Make sure buffer is written before we update index. */ 1237 smp_wmb(); 1238 if (put_user(vq->last_used_idx, &vq->used->idx)) { 1239 vq_err(vq, "Failed to increment used idx"); 1240 return -EFAULT; 1241 } 1242 if (unlikely(vq->log_used)) { 1243 /* Log used index update. */ 1244 log_write(vq->log_base, 1245 vq->log_addr + offsetof(struct vring_used, idx), 1246 sizeof vq->used->idx); 1247 if (vq->log_ctx) 1248 eventfd_signal(vq->log_ctx, 1); 1249 } 1250 return r; 1251} 1252 1253/* This actually signals the guest, using eventfd. */ 1254void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq) 1255{ 1256 __u16 flags; 1257 /* Flush out used index updates. This is paired 1258 * with the barrier that the Guest executes when enabling 1259 * interrupts. */ 1260 smp_mb(); 1261 1262 if (get_user(flags, &vq->avail->flags)) { 1263 vq_err(vq, "Failed to get flags"); 1264 return; 1265 } 1266 1267 /* If they don't want an interrupt, don't signal, unless empty. */ 1268 if ((flags & VRING_AVAIL_F_NO_INTERRUPT) && 1269 (vq->avail_idx != vq->last_avail_idx || 1270 !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY))) 1271 return; 1272 1273 /* Signal the Guest tell them we used something up. */ 1274 if (vq->call_ctx) 1275 eventfd_signal(vq->call_ctx, 1); 1276} 1277 1278/* And here's the combo meal deal. Supersize me! */ 1279void vhost_add_used_and_signal(struct vhost_dev *dev, 1280 struct vhost_virtqueue *vq, 1281 unsigned int head, int len) 1282{ 1283 vhost_add_used(vq, head, len); 1284 vhost_signal(dev, vq); 1285} 1286 1287/* multi-buffer version of vhost_add_used_and_signal */ 1288void vhost_add_used_and_signal_n(struct vhost_dev *dev, 1289 struct vhost_virtqueue *vq, 1290 struct vring_used_elem *heads, unsigned count) 1291{ 1292 vhost_add_used_n(vq, heads, count); 1293 vhost_signal(dev, vq); 1294} 1295 1296/* OK, now we need to know about added descriptors. */ 1297bool vhost_enable_notify(struct vhost_virtqueue *vq) 1298{ 1299 u16 avail_idx; 1300 int r; 1301 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY)) 1302 return false; 1303 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY; 1304 r = put_user(vq->used_flags, &vq->used->flags); 1305 if (r) { 1306 vq_err(vq, "Failed to enable notification at %p: %d\n", 1307 &vq->used->flags, r); 1308 return false; 1309 } 1310 /* They could have slipped one in as we were doing that: make 1311 * sure it's written, then check again. */ 1312 smp_mb(); 1313 r = get_user(avail_idx, &vq->avail->idx); 1314 if (r) { 1315 vq_err(vq, "Failed to check avail idx at %p: %d\n", 1316 &vq->avail->idx, r); 1317 return false; 1318 } 1319 1320 return avail_idx != vq->avail_idx; 1321} 1322 1323/* We don't need to be notified again. */ 1324void vhost_disable_notify(struct vhost_virtqueue *vq) 1325{ 1326 int r; 1327 if (vq->used_flags & VRING_USED_F_NO_NOTIFY) 1328 return; 1329 vq->used_flags |= VRING_USED_F_NO_NOTIFY; 1330 r = put_user(vq->used_flags, &vq->used->flags); 1331 if (r) 1332 vq_err(vq, "Failed to enable notification at %p: %d\n", 1333 &vq->used->flags, r); 1334} 1335