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/workqueue.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 27 #include <linux/net.h> 28 #include <linux/if_packet.h> 29 #include <linux/if_arp.h> 30 31 #include <net/sock.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 struct workqueue_struct *vhost_workqueue; 41 42 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh, 43 poll_table *pt) 44 { 45 struct vhost_poll *poll; 46 poll = container_of(pt, struct vhost_poll, table); 47 48 poll->wqh = wqh; 49 add_wait_queue(wqh, &poll->wait); 50 } 51 52 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync, 53 void *key) 54 { 55 struct vhost_poll *poll; 56 poll = container_of(wait, struct vhost_poll, wait); 57 if (!((unsigned long)key & poll->mask)) 58 return 0; 59 60 queue_work(vhost_workqueue, &poll->work); 61 return 0; 62 } 63 64 /* Init poll structure */ 65 void vhost_poll_init(struct vhost_poll *poll, work_func_t func, 66 unsigned long mask) 67 { 68 INIT_WORK(&poll->work, func); 69 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup); 70 init_poll_funcptr(&poll->table, vhost_poll_func); 71 poll->mask = mask; 72 } 73 74 /* Start polling a file. We add ourselves to file's wait queue. The caller must 75 * keep a reference to a file until after vhost_poll_stop is called. */ 76 void vhost_poll_start(struct vhost_poll *poll, struct file *file) 77 { 78 unsigned long mask; 79 mask = file->f_op->poll(file, &poll->table); 80 if (mask) 81 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask); 82 } 83 84 /* Stop polling a file. After this function returns, it becomes safe to drop the 85 * file reference. You must also flush afterwards. */ 86 void vhost_poll_stop(struct vhost_poll *poll) 87 { 88 remove_wait_queue(poll->wqh, &poll->wait); 89 } 90 91 /* Flush any work that has been scheduled. When calling this, don't hold any 92 * locks that are also used by the callback. */ 93 void vhost_poll_flush(struct vhost_poll *poll) 94 { 95 flush_work(&poll->work); 96 } 97 98 void vhost_poll_queue(struct vhost_poll *poll) 99 { 100 queue_work(vhost_workqueue, &poll->work); 101 } 102 103 static void vhost_vq_reset(struct vhost_dev *dev, 104 struct vhost_virtqueue *vq) 105 { 106 vq->num = 1; 107 vq->desc = NULL; 108 vq->avail = NULL; 109 vq->used = NULL; 110 vq->last_avail_idx = 0; 111 vq->avail_idx = 0; 112 vq->last_used_idx = 0; 113 vq->used_flags = 0; 114 vq->used_flags = 0; 115 vq->log_used = false; 116 vq->log_addr = -1ull; 117 vq->hdr_size = 0; 118 vq->private_data = NULL; 119 vq->log_base = NULL; 120 vq->error_ctx = NULL; 121 vq->error = NULL; 122 vq->kick = NULL; 123 vq->call_ctx = NULL; 124 vq->call = NULL; 125 vq->log_ctx = NULL; 126 } 127 128 long vhost_dev_init(struct vhost_dev *dev, 129 struct vhost_virtqueue *vqs, int nvqs) 130 { 131 int i; 132 dev->vqs = vqs; 133 dev->nvqs = nvqs; 134 mutex_init(&dev->mutex); 135 dev->log_ctx = NULL; 136 dev->log_file = NULL; 137 dev->memory = NULL; 138 dev->mm = NULL; 139 140 for (i = 0; i < dev->nvqs; ++i) { 141 dev->vqs[i].dev = dev; 142 mutex_init(&dev->vqs[i].mutex); 143 vhost_vq_reset(dev, dev->vqs + i); 144 if (dev->vqs[i].handle_kick) 145 vhost_poll_init(&dev->vqs[i].poll, 146 dev->vqs[i].handle_kick, 147 POLLIN); 148 } 149 return 0; 150 } 151 152 /* Caller should have device mutex */ 153 long vhost_dev_check_owner(struct vhost_dev *dev) 154 { 155 /* Are you the owner? If not, I don't think you mean to do that */ 156 return dev->mm == current->mm ? 0 : -EPERM; 157 } 158 159 /* Caller should have device mutex */ 160 static long vhost_dev_set_owner(struct vhost_dev *dev) 161 { 162 /* Is there an owner already? */ 163 if (dev->mm) 164 return -EBUSY; 165 /* No owner, become one */ 166 dev->mm = get_task_mm(current); 167 return 0; 168 } 169 170 /* Caller should have device mutex */ 171 long vhost_dev_reset_owner(struct vhost_dev *dev) 172 { 173 struct vhost_memory *memory; 174 175 /* Restore memory to default empty mapping. */ 176 memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL); 177 if (!memory) 178 return -ENOMEM; 179 180 vhost_dev_cleanup(dev); 181 182 memory->nregions = 0; 183 dev->memory = memory; 184 return 0; 185 } 186 187 /* Caller should have device mutex */ 188 void vhost_dev_cleanup(struct vhost_dev *dev) 189 { 190 int i; 191 for (i = 0; i < dev->nvqs; ++i) { 192 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) { 193 vhost_poll_stop(&dev->vqs[i].poll); 194 vhost_poll_flush(&dev->vqs[i].poll); 195 } 196 if (dev->vqs[i].error_ctx) 197 eventfd_ctx_put(dev->vqs[i].error_ctx); 198 if (dev->vqs[i].error) 199 fput(dev->vqs[i].error); 200 if (dev->vqs[i].kick) 201 fput(dev->vqs[i].kick); 202 if (dev->vqs[i].call_ctx) 203 eventfd_ctx_put(dev->vqs[i].call_ctx); 204 if (dev->vqs[i].call) 205 fput(dev->vqs[i].call); 206 vhost_vq_reset(dev, dev->vqs + i); 207 } 208 if (dev->log_ctx) 209 eventfd_ctx_put(dev->log_ctx); 210 dev->log_ctx = NULL; 211 if (dev->log_file) 212 fput(dev->log_file); 213 dev->log_file = NULL; 214 /* No one will access memory at this point */ 215 kfree(dev->memory); 216 dev->memory = NULL; 217 if (dev->mm) 218 mmput(dev->mm); 219 dev->mm = NULL; 220 } 221 222 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz) 223 { 224 u64 a = addr / VHOST_PAGE_SIZE / 8; 225 /* Make sure 64 bit math will not overflow. */ 226 if (a > ULONG_MAX - (unsigned long)log_base || 227 a + (unsigned long)log_base > ULONG_MAX) 228 return -EFAULT; 229 230 return access_ok(VERIFY_WRITE, log_base + a, 231 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8); 232 } 233 234 /* Caller should have vq mutex and device mutex. */ 235 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem, 236 int log_all) 237 { 238 int i; 239 for (i = 0; i < mem->nregions; ++i) { 240 struct vhost_memory_region *m = mem->regions + i; 241 unsigned long a = m->userspace_addr; 242 if (m->memory_size > ULONG_MAX) 243 return 0; 244 else if (!access_ok(VERIFY_WRITE, (void __user *)a, 245 m->memory_size)) 246 return 0; 247 else if (log_all && !log_access_ok(log_base, 248 m->guest_phys_addr, 249 m->memory_size)) 250 return 0; 251 } 252 return 1; 253 } 254 255 /* Can we switch to this memory table? */ 256 /* Caller should have device mutex but not vq mutex */ 257 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem, 258 int log_all) 259 { 260 int i; 261 for (i = 0; i < d->nvqs; ++i) { 262 int ok; 263 mutex_lock(&d->vqs[i].mutex); 264 /* If ring is inactive, will check when it's enabled. */ 265 if (d->vqs[i].private_data) 266 ok = vq_memory_access_ok(d->vqs[i].log_base, mem, 267 log_all); 268 else 269 ok = 1; 270 mutex_unlock(&d->vqs[i].mutex); 271 if (!ok) 272 return 0; 273 } 274 return 1; 275 } 276 277 static int vq_access_ok(unsigned int num, 278 struct vring_desc __user *desc, 279 struct vring_avail __user *avail, 280 struct vring_used __user *used) 281 { 282 return access_ok(VERIFY_READ, desc, num * sizeof *desc) && 283 access_ok(VERIFY_READ, avail, 284 sizeof *avail + num * sizeof *avail->ring) && 285 access_ok(VERIFY_WRITE, used, 286 sizeof *used + num * sizeof *used->ring); 287 } 288 289 /* Can we log writes? */ 290 /* Caller should have device mutex but not vq mutex */ 291 int vhost_log_access_ok(struct vhost_dev *dev) 292 { 293 return memory_access_ok(dev, dev->memory, 1); 294 } 295 296 /* Verify access for write logging. */ 297 /* Caller should have vq mutex and device mutex */ 298 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base) 299 { 300 return vq_memory_access_ok(log_base, vq->dev->memory, 301 vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) && 302 (!vq->log_used || log_access_ok(log_base, vq->log_addr, 303 sizeof *vq->used + 304 vq->num * sizeof *vq->used->ring)); 305 } 306 307 /* Can we start vq? */ 308 /* Caller should have vq mutex and device mutex */ 309 int vhost_vq_access_ok(struct vhost_virtqueue *vq) 310 { 311 return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) && 312 vq_log_access_ok(vq, vq->log_base); 313 } 314 315 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m) 316 { 317 struct vhost_memory mem, *newmem, *oldmem; 318 unsigned long size = offsetof(struct vhost_memory, regions); 319 long r; 320 r = copy_from_user(&mem, m, size); 321 if (r) 322 return r; 323 if (mem.padding) 324 return -EOPNOTSUPP; 325 if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS) 326 return -E2BIG; 327 newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL); 328 if (!newmem) 329 return -ENOMEM; 330 331 memcpy(newmem, &mem, size); 332 r = copy_from_user(newmem->regions, m->regions, 333 mem.nregions * sizeof *m->regions); 334 if (r) { 335 kfree(newmem); 336 return r; 337 } 338 339 if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL))) 340 return -EFAULT; 341 oldmem = d->memory; 342 rcu_assign_pointer(d->memory, newmem); 343 synchronize_rcu(); 344 kfree(oldmem); 345 return 0; 346 } 347 348 static int init_used(struct vhost_virtqueue *vq, 349 struct vring_used __user *used) 350 { 351 int r = put_user(vq->used_flags, &used->flags); 352 if (r) 353 return r; 354 return get_user(vq->last_used_idx, &used->idx); 355 } 356 357 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp) 358 { 359 struct file *eventfp, *filep = NULL, 360 *pollstart = NULL, *pollstop = NULL; 361 struct eventfd_ctx *ctx = NULL; 362 u32 __user *idxp = argp; 363 struct vhost_virtqueue *vq; 364 struct vhost_vring_state s; 365 struct vhost_vring_file f; 366 struct vhost_vring_addr a; 367 u32 idx; 368 long r; 369 370 r = get_user(idx, idxp); 371 if (r < 0) 372 return r; 373 if (idx > d->nvqs) 374 return -ENOBUFS; 375 376 vq = d->vqs + idx; 377 378 mutex_lock(&vq->mutex); 379 380 switch (ioctl) { 381 case VHOST_SET_VRING_NUM: 382 /* Resizing ring with an active backend? 383 * You don't want to do that. */ 384 if (vq->private_data) { 385 r = -EBUSY; 386 break; 387 } 388 r = copy_from_user(&s, argp, sizeof s); 389 if (r < 0) 390 break; 391 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) { 392 r = -EINVAL; 393 break; 394 } 395 vq->num = s.num; 396 break; 397 case VHOST_SET_VRING_BASE: 398 /* Moving base with an active backend? 399 * You don't want to do that. */ 400 if (vq->private_data) { 401 r = -EBUSY; 402 break; 403 } 404 r = copy_from_user(&s, argp, sizeof s); 405 if (r < 0) 406 break; 407 if (s.num > 0xffff) { 408 r = -EINVAL; 409 break; 410 } 411 vq->last_avail_idx = s.num; 412 /* Forget the cached index value. */ 413 vq->avail_idx = vq->last_avail_idx; 414 break; 415 case VHOST_GET_VRING_BASE: 416 s.index = idx; 417 s.num = vq->last_avail_idx; 418 r = copy_to_user(argp, &s, sizeof s); 419 break; 420 case VHOST_SET_VRING_ADDR: 421 r = copy_from_user(&a, argp, sizeof a); 422 if (r < 0) 423 break; 424 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) { 425 r = -EOPNOTSUPP; 426 break; 427 } 428 /* For 32bit, verify that the top 32bits of the user 429 data are set to zero. */ 430 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr || 431 (u64)(unsigned long)a.used_user_addr != a.used_user_addr || 432 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) { 433 r = -EFAULT; 434 break; 435 } 436 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) || 437 (a.used_user_addr & (sizeof *vq->used->ring - 1)) || 438 (a.log_guest_addr & (sizeof *vq->used->ring - 1))) { 439 r = -EINVAL; 440 break; 441 } 442 443 /* We only verify access here if backend is configured. 444 * If it is not, we don't as size might not have been setup. 445 * We will verify when backend is configured. */ 446 if (vq->private_data) { 447 if (!vq_access_ok(vq->num, 448 (void __user *)(unsigned long)a.desc_user_addr, 449 (void __user *)(unsigned long)a.avail_user_addr, 450 (void __user *)(unsigned long)a.used_user_addr)) { 451 r = -EINVAL; 452 break; 453 } 454 455 /* Also validate log access for used ring if enabled. */ 456 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) && 457 !log_access_ok(vq->log_base, a.log_guest_addr, 458 sizeof *vq->used + 459 vq->num * sizeof *vq->used->ring)) { 460 r = -EINVAL; 461 break; 462 } 463 } 464 465 r = init_used(vq, (struct vring_used __user *)(unsigned long) 466 a.used_user_addr); 467 if (r) 468 break; 469 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG)); 470 vq->desc = (void __user *)(unsigned long)a.desc_user_addr; 471 vq->avail = (void __user *)(unsigned long)a.avail_user_addr; 472 vq->log_addr = a.log_guest_addr; 473 vq->used = (void __user *)(unsigned long)a.used_user_addr; 474 break; 475 case VHOST_SET_VRING_KICK: 476 r = copy_from_user(&f, argp, sizeof f); 477 if (r < 0) 478 break; 479 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); 480 if (IS_ERR(eventfp)) { 481 r = PTR_ERR(eventfp); 482 break; 483 } 484 if (eventfp != vq->kick) { 485 pollstop = filep = vq->kick; 486 pollstart = vq->kick = eventfp; 487 } else 488 filep = eventfp; 489 break; 490 case VHOST_SET_VRING_CALL: 491 r = copy_from_user(&f, argp, sizeof f); 492 if (r < 0) 493 break; 494 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); 495 if (IS_ERR(eventfp)) { 496 r = PTR_ERR(eventfp); 497 break; 498 } 499 if (eventfp != vq->call) { 500 filep = vq->call; 501 ctx = vq->call_ctx; 502 vq->call = eventfp; 503 vq->call_ctx = eventfp ? 504 eventfd_ctx_fileget(eventfp) : NULL; 505 } else 506 filep = eventfp; 507 break; 508 case VHOST_SET_VRING_ERR: 509 r = copy_from_user(&f, argp, sizeof f); 510 if (r < 0) 511 break; 512 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd); 513 if (IS_ERR(eventfp)) { 514 r = PTR_ERR(eventfp); 515 break; 516 } 517 if (eventfp != vq->error) { 518 filep = vq->error; 519 vq->error = eventfp; 520 ctx = vq->error_ctx; 521 vq->error_ctx = eventfp ? 522 eventfd_ctx_fileget(eventfp) : NULL; 523 } else 524 filep = eventfp; 525 break; 526 default: 527 r = -ENOIOCTLCMD; 528 } 529 530 if (pollstop && vq->handle_kick) 531 vhost_poll_stop(&vq->poll); 532 533 if (ctx) 534 eventfd_ctx_put(ctx); 535 if (filep) 536 fput(filep); 537 538 if (pollstart && vq->handle_kick) 539 vhost_poll_start(&vq->poll, vq->kick); 540 541 mutex_unlock(&vq->mutex); 542 543 if (pollstop && vq->handle_kick) 544 vhost_poll_flush(&vq->poll); 545 return r; 546 } 547 548 /* Caller must have device mutex */ 549 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg) 550 { 551 void __user *argp = (void __user *)arg; 552 struct file *eventfp, *filep = NULL; 553 struct eventfd_ctx *ctx = NULL; 554 u64 p; 555 long r; 556 int i, fd; 557 558 /* If you are not the owner, you can become one */ 559 if (ioctl == VHOST_SET_OWNER) { 560 r = vhost_dev_set_owner(d); 561 goto done; 562 } 563 564 /* You must be the owner to do anything else */ 565 r = vhost_dev_check_owner(d); 566 if (r) 567 goto done; 568 569 switch (ioctl) { 570 case VHOST_SET_MEM_TABLE: 571 r = vhost_set_memory(d, argp); 572 break; 573 case VHOST_SET_LOG_BASE: 574 r = copy_from_user(&p, argp, sizeof p); 575 if (r < 0) 576 break; 577 if ((u64)(unsigned long)p != p) { 578 r = -EFAULT; 579 break; 580 } 581 for (i = 0; i < d->nvqs; ++i) { 582 struct vhost_virtqueue *vq; 583 void __user *base = (void __user *)(unsigned long)p; 584 vq = d->vqs + i; 585 mutex_lock(&vq->mutex); 586 /* If ring is inactive, will check when it's enabled. */ 587 if (vq->private_data && !vq_log_access_ok(vq, base)) 588 r = -EFAULT; 589 else 590 vq->log_base = base; 591 mutex_unlock(&vq->mutex); 592 } 593 break; 594 case VHOST_SET_LOG_FD: 595 r = get_user(fd, (int __user *)argp); 596 if (r < 0) 597 break; 598 eventfp = fd == -1 ? NULL : eventfd_fget(fd); 599 if (IS_ERR(eventfp)) { 600 r = PTR_ERR(eventfp); 601 break; 602 } 603 if (eventfp != d->log_file) { 604 filep = d->log_file; 605 ctx = d->log_ctx; 606 d->log_ctx = eventfp ? 607 eventfd_ctx_fileget(eventfp) : NULL; 608 } else 609 filep = eventfp; 610 for (i = 0; i < d->nvqs; ++i) { 611 mutex_lock(&d->vqs[i].mutex); 612 d->vqs[i].log_ctx = d->log_ctx; 613 mutex_unlock(&d->vqs[i].mutex); 614 } 615 if (ctx) 616 eventfd_ctx_put(ctx); 617 if (filep) 618 fput(filep); 619 break; 620 default: 621 r = vhost_set_vring(d, ioctl, argp); 622 break; 623 } 624 done: 625 return r; 626 } 627 628 static const struct vhost_memory_region *find_region(struct vhost_memory *mem, 629 __u64 addr, __u32 len) 630 { 631 struct vhost_memory_region *reg; 632 int i; 633 /* linear search is not brilliant, but we really have on the order of 6 634 * regions in practice */ 635 for (i = 0; i < mem->nregions; ++i) { 636 reg = mem->regions + i; 637 if (reg->guest_phys_addr <= addr && 638 reg->guest_phys_addr + reg->memory_size - 1 >= addr) 639 return reg; 640 } 641 return NULL; 642 } 643 644 /* TODO: This is really inefficient. We need something like get_user() 645 * (instruction directly accesses the data, with an exception table entry 646 * returning -EFAULT). See Documentation/x86/exception-tables.txt. 647 */ 648 static int set_bit_to_user(int nr, void __user *addr) 649 { 650 unsigned long log = (unsigned long)addr; 651 struct page *page; 652 void *base; 653 int bit = nr + (log % PAGE_SIZE) * 8; 654 int r; 655 r = get_user_pages_fast(log, 1, 1, &page); 656 if (r < 0) 657 return r; 658 BUG_ON(r != 1); 659 base = kmap_atomic(page, KM_USER0); 660 set_bit(bit, base); 661 kunmap_atomic(base, KM_USER0); 662 set_page_dirty_lock(page); 663 put_page(page); 664 return 0; 665 } 666 667 static int log_write(void __user *log_base, 668 u64 write_address, u64 write_length) 669 { 670 int r; 671 if (!write_length) 672 return 0; 673 write_address /= VHOST_PAGE_SIZE; 674 for (;;) { 675 u64 base = (u64)(unsigned long)log_base; 676 u64 log = base + write_address / 8; 677 int bit = write_address % 8; 678 if ((u64)(unsigned long)log != log) 679 return -EFAULT; 680 r = set_bit_to_user(bit, (void __user *)(unsigned long)log); 681 if (r < 0) 682 return r; 683 if (write_length <= VHOST_PAGE_SIZE) 684 break; 685 write_length -= VHOST_PAGE_SIZE; 686 write_address += VHOST_PAGE_SIZE; 687 } 688 return r; 689 } 690 691 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log, 692 unsigned int log_num, u64 len) 693 { 694 int i, r; 695 696 /* Make sure data written is seen before log. */ 697 smp_wmb(); 698 for (i = 0; i < log_num; ++i) { 699 u64 l = min(log[i].len, len); 700 r = log_write(vq->log_base, log[i].addr, l); 701 if (r < 0) 702 return r; 703 len -= l; 704 if (!len) 705 return 0; 706 } 707 if (vq->log_ctx) 708 eventfd_signal(vq->log_ctx, 1); 709 /* Length written exceeds what we have stored. This is a bug. */ 710 BUG(); 711 return 0; 712 } 713 714 int translate_desc(struct vhost_dev *dev, u64 addr, u32 len, 715 struct iovec iov[], int iov_size) 716 { 717 const struct vhost_memory_region *reg; 718 struct vhost_memory *mem; 719 struct iovec *_iov; 720 u64 s = 0; 721 int ret = 0; 722 723 rcu_read_lock(); 724 725 mem = rcu_dereference(dev->memory); 726 while ((u64)len > s) { 727 u64 size; 728 if (ret >= iov_size) { 729 ret = -ENOBUFS; 730 break; 731 } 732 reg = find_region(mem, addr, len); 733 if (!reg) { 734 ret = -EFAULT; 735 break; 736 } 737 _iov = iov + ret; 738 size = reg->memory_size - addr + reg->guest_phys_addr; 739 _iov->iov_len = min((u64)len, size); 740 _iov->iov_base = (void *)(unsigned long) 741 (reg->userspace_addr + addr - reg->guest_phys_addr); 742 s += size; 743 addr += size; 744 ++ret; 745 } 746 747 rcu_read_unlock(); 748 return ret; 749 } 750 751 /* Each buffer in the virtqueues is actually a chain of descriptors. This 752 * function returns the next descriptor in the chain, 753 * or -1U if we're at the end. */ 754 static unsigned next_desc(struct vring_desc *desc) 755 { 756 unsigned int next; 757 758 /* If this descriptor says it doesn't chain, we're done. */ 759 if (!(desc->flags & VRING_DESC_F_NEXT)) 760 return -1U; 761 762 /* Check they're not leading us off end of descriptors. */ 763 next = desc->next; 764 /* Make sure compiler knows to grab that: we don't want it changing! */ 765 /* We will use the result as an index in an array, so most 766 * architectures only need a compiler barrier here. */ 767 read_barrier_depends(); 768 769 return next; 770 } 771 772 static unsigned get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq, 773 struct iovec iov[], unsigned int iov_size, 774 unsigned int *out_num, unsigned int *in_num, 775 struct vhost_log *log, unsigned int *log_num, 776 struct vring_desc *indirect) 777 { 778 struct vring_desc desc; 779 unsigned int i = 0, count, found = 0; 780 int ret; 781 782 /* Sanity check */ 783 if (indirect->len % sizeof desc) { 784 vq_err(vq, "Invalid length in indirect descriptor: " 785 "len 0x%llx not multiple of 0x%zx\n", 786 (unsigned long long)indirect->len, 787 sizeof desc); 788 return -EINVAL; 789 } 790 791 ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect, 792 ARRAY_SIZE(vq->indirect)); 793 if (ret < 0) { 794 vq_err(vq, "Translation failure %d in indirect.\n", ret); 795 return ret; 796 } 797 798 /* We will use the result as an address to read from, so most 799 * architectures only need a compiler barrier here. */ 800 read_barrier_depends(); 801 802 count = indirect->len / sizeof desc; 803 /* Buffers are chained via a 16 bit next field, so 804 * we can have at most 2^16 of these. */ 805 if (count > USHORT_MAX + 1) { 806 vq_err(vq, "Indirect buffer length too big: %d\n", 807 indirect->len); 808 return -E2BIG; 809 } 810 811 do { 812 unsigned iov_count = *in_num + *out_num; 813 if (++found > count) { 814 vq_err(vq, "Loop detected: last one at %u " 815 "indirect size %u\n", 816 i, count); 817 return -EINVAL; 818 } 819 if (memcpy_fromiovec((unsigned char *)&desc, vq->indirect, 820 sizeof desc)) { 821 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n", 822 i, (size_t)indirect->addr + i * sizeof desc); 823 return -EINVAL; 824 } 825 if (desc.flags & VRING_DESC_F_INDIRECT) { 826 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n", 827 i, (size_t)indirect->addr + i * sizeof desc); 828 return -EINVAL; 829 } 830 831 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count, 832 iov_size - iov_count); 833 if (ret < 0) { 834 vq_err(vq, "Translation failure %d indirect idx %d\n", 835 ret, i); 836 return ret; 837 } 838 /* If this is an input descriptor, increment that count. */ 839 if (desc.flags & VRING_DESC_F_WRITE) { 840 *in_num += ret; 841 if (unlikely(log)) { 842 log[*log_num].addr = desc.addr; 843 log[*log_num].len = desc.len; 844 ++*log_num; 845 } 846 } else { 847 /* If it's an output descriptor, they're all supposed 848 * to come before any input descriptors. */ 849 if (*in_num) { 850 vq_err(vq, "Indirect descriptor " 851 "has out after in: idx %d\n", i); 852 return -EINVAL; 853 } 854 *out_num += ret; 855 } 856 } while ((i = next_desc(&desc)) != -1); 857 return 0; 858 } 859 860 /* This looks in the virtqueue and for the first available buffer, and converts 861 * it to an iovec for convenient access. Since descriptors consist of some 862 * number of output then some number of input descriptors, it's actually two 863 * iovecs, but we pack them into one and note how many of each there were. 864 * 865 * This function returns the descriptor number found, or vq->num (which 866 * is never a valid descriptor number) if none was found. */ 867 unsigned vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq, 868 struct iovec iov[], unsigned int iov_size, 869 unsigned int *out_num, unsigned int *in_num, 870 struct vhost_log *log, unsigned int *log_num) 871 { 872 struct vring_desc desc; 873 unsigned int i, head, found = 0; 874 u16 last_avail_idx; 875 int ret; 876 877 /* Check it isn't doing very strange things with descriptor numbers. */ 878 last_avail_idx = vq->last_avail_idx; 879 if (get_user(vq->avail_idx, &vq->avail->idx)) { 880 vq_err(vq, "Failed to access avail idx at %p\n", 881 &vq->avail->idx); 882 return vq->num; 883 } 884 885 if ((u16)(vq->avail_idx - last_avail_idx) > vq->num) { 886 vq_err(vq, "Guest moved used index from %u to %u", 887 last_avail_idx, vq->avail_idx); 888 return vq->num; 889 } 890 891 /* If there's nothing new since last we looked, return invalid. */ 892 if (vq->avail_idx == last_avail_idx) 893 return vq->num; 894 895 /* Only get avail ring entries after they have been exposed by guest. */ 896 smp_rmb(); 897 898 /* Grab the next descriptor number they're advertising, and increment 899 * the index we've seen. */ 900 if (get_user(head, &vq->avail->ring[last_avail_idx % vq->num])) { 901 vq_err(vq, "Failed to read head: idx %d address %p\n", 902 last_avail_idx, 903 &vq->avail->ring[last_avail_idx % vq->num]); 904 return vq->num; 905 } 906 907 /* If their number is silly, that's an error. */ 908 if (head >= vq->num) { 909 vq_err(vq, "Guest says index %u > %u is available", 910 head, vq->num); 911 return vq->num; 912 } 913 914 /* When we start there are none of either input nor output. */ 915 *out_num = *in_num = 0; 916 if (unlikely(log)) 917 *log_num = 0; 918 919 i = head; 920 do { 921 unsigned iov_count = *in_num + *out_num; 922 if (i >= vq->num) { 923 vq_err(vq, "Desc index is %u > %u, head = %u", 924 i, vq->num, head); 925 return vq->num; 926 } 927 if (++found > vq->num) { 928 vq_err(vq, "Loop detected: last one at %u " 929 "vq size %u head %u\n", 930 i, vq->num, head); 931 return vq->num; 932 } 933 ret = copy_from_user(&desc, vq->desc + i, sizeof desc); 934 if (ret) { 935 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n", 936 i, vq->desc + i); 937 return vq->num; 938 } 939 if (desc.flags & VRING_DESC_F_INDIRECT) { 940 ret = get_indirect(dev, vq, iov, iov_size, 941 out_num, in_num, 942 log, log_num, &desc); 943 if (ret < 0) { 944 vq_err(vq, "Failure detected " 945 "in indirect descriptor at idx %d\n", i); 946 return vq->num; 947 } 948 continue; 949 } 950 951 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count, 952 iov_size - iov_count); 953 if (ret < 0) { 954 vq_err(vq, "Translation failure %d descriptor idx %d\n", 955 ret, i); 956 return vq->num; 957 } 958 if (desc.flags & VRING_DESC_F_WRITE) { 959 /* If this is an input descriptor, 960 * increment that count. */ 961 *in_num += ret; 962 if (unlikely(log)) { 963 log[*log_num].addr = desc.addr; 964 log[*log_num].len = desc.len; 965 ++*log_num; 966 } 967 } else { 968 /* If it's an output descriptor, they're all supposed 969 * to come before any input descriptors. */ 970 if (*in_num) { 971 vq_err(vq, "Descriptor has out after in: " 972 "idx %d\n", i); 973 return vq->num; 974 } 975 *out_num += ret; 976 } 977 } while ((i = next_desc(&desc)) != -1); 978 979 /* On success, increment avail index. */ 980 vq->last_avail_idx++; 981 return head; 982 } 983 984 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */ 985 void vhost_discard_vq_desc(struct vhost_virtqueue *vq) 986 { 987 vq->last_avail_idx--; 988 } 989 990 /* After we've used one of their buffers, we tell them about it. We'll then 991 * want to notify the guest, using eventfd. */ 992 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len) 993 { 994 struct vring_used_elem *used; 995 996 /* The virtqueue contains a ring of used buffers. Get a pointer to the 997 * next entry in that used ring. */ 998 used = &vq->used->ring[vq->last_used_idx % vq->num]; 999 if (put_user(head, &used->id)) { 1000 vq_err(vq, "Failed to write used id"); 1001 return -EFAULT; 1002 } 1003 if (put_user(len, &used->len)) { 1004 vq_err(vq, "Failed to write used len"); 1005 return -EFAULT; 1006 } 1007 /* Make sure buffer is written before we update index. */ 1008 smp_wmb(); 1009 if (put_user(vq->last_used_idx + 1, &vq->used->idx)) { 1010 vq_err(vq, "Failed to increment used idx"); 1011 return -EFAULT; 1012 } 1013 if (unlikely(vq->log_used)) { 1014 /* Make sure data is seen before log. */ 1015 smp_wmb(); 1016 /* Log used ring entry write. */ 1017 log_write(vq->log_base, 1018 vq->log_addr + ((void *)used - (void *)vq->used), 1019 sizeof *used); 1020 /* Log used index update. */ 1021 log_write(vq->log_base, 1022 vq->log_addr + offsetof(struct vring_used, idx), 1023 sizeof vq->used->idx); 1024 if (vq->log_ctx) 1025 eventfd_signal(vq->log_ctx, 1); 1026 } 1027 vq->last_used_idx++; 1028 return 0; 1029 } 1030 1031 /* This actually signals the guest, using eventfd. */ 1032 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq) 1033 { 1034 __u16 flags = 0; 1035 if (get_user(flags, &vq->avail->flags)) { 1036 vq_err(vq, "Failed to get flags"); 1037 return; 1038 } 1039 1040 /* If they don't want an interrupt, don't signal, unless empty. */ 1041 if ((flags & VRING_AVAIL_F_NO_INTERRUPT) && 1042 (vq->avail_idx != vq->last_avail_idx || 1043 !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY))) 1044 return; 1045 1046 /* Signal the Guest tell them we used something up. */ 1047 if (vq->call_ctx) 1048 eventfd_signal(vq->call_ctx, 1); 1049 } 1050 1051 /* And here's the combo meal deal. Supersize me! */ 1052 void vhost_add_used_and_signal(struct vhost_dev *dev, 1053 struct vhost_virtqueue *vq, 1054 unsigned int head, int len) 1055 { 1056 vhost_add_used(vq, head, len); 1057 vhost_signal(dev, vq); 1058 } 1059 1060 /* OK, now we need to know about added descriptors. */ 1061 bool vhost_enable_notify(struct vhost_virtqueue *vq) 1062 { 1063 u16 avail_idx; 1064 int r; 1065 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY)) 1066 return false; 1067 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY; 1068 r = put_user(vq->used_flags, &vq->used->flags); 1069 if (r) { 1070 vq_err(vq, "Failed to enable notification at %p: %d\n", 1071 &vq->used->flags, r); 1072 return false; 1073 } 1074 /* They could have slipped one in as we were doing that: make 1075 * sure it's written, then check again. */ 1076 smp_mb(); 1077 r = get_user(avail_idx, &vq->avail->idx); 1078 if (r) { 1079 vq_err(vq, "Failed to check avail idx at %p: %d\n", 1080 &vq->avail->idx, r); 1081 return false; 1082 } 1083 1084 return avail_idx != vq->last_avail_idx; 1085 } 1086 1087 /* We don't need to be notified again. */ 1088 void vhost_disable_notify(struct vhost_virtqueue *vq) 1089 { 1090 int r; 1091 if (vq->used_flags & VRING_USED_F_NO_NOTIFY) 1092 return; 1093 vq->used_flags |= VRING_USED_F_NO_NOTIFY; 1094 r = put_user(vq->used_flags, &vq->used->flags); 1095 if (r) 1096 vq_err(vq, "Failed to enable notification at %p: %d\n", 1097 &vq->used->flags, r); 1098 } 1099 1100 int vhost_init(void) 1101 { 1102 vhost_workqueue = create_singlethread_workqueue("vhost"); 1103 if (!vhost_workqueue) 1104 return -ENOMEM; 1105 return 0; 1106 } 1107 1108 void vhost_cleanup(void) 1109 { 1110 destroy_workqueue(vhost_workqueue); 1111 } 1112