1 /* 2 * Helpers for getting linearized buffers from iov / filling buffers into iovs 3 * 4 * Copyright IBM, Corp. 2007, 2008 5 * Copyright (C) 2010 Red Hat, Inc. 6 * 7 * Author(s): 8 * Anthony Liguori <aliguori@us.ibm.com> 9 * Amit Shah <amit.shah@redhat.com> 10 * Michael Tokarev <mjt@tls.msk.ru> 11 * 12 * This work is licensed under the terms of the GNU GPL, version 2. See 13 * the COPYING file in the top-level directory. 14 * 15 * Contributions after 2012-01-13 are licensed under the terms of the 16 * GNU GPL, version 2 or (at your option) any later version. 17 */ 18 19 #include "qemu/osdep.h" 20 #include "qemu/iov.h" 21 #include "qemu/sockets.h" 22 #include "qemu/cutils.h" 23 24 size_t iov_from_buf_full(const struct iovec *iov, unsigned int iov_cnt, 25 size_t offset, const void *buf, size_t bytes) 26 { 27 size_t done; 28 unsigned int i; 29 for (i = 0, done = 0; (offset || done < bytes) && i < iov_cnt; i++) { 30 if (offset < iov[i].iov_len) { 31 size_t len = MIN(iov[i].iov_len - offset, bytes - done); 32 memcpy(iov[i].iov_base + offset, buf + done, len); 33 done += len; 34 offset = 0; 35 } else { 36 offset -= iov[i].iov_len; 37 } 38 } 39 assert(offset == 0); 40 return done; 41 } 42 43 size_t iov_to_buf_full(const struct iovec *iov, const unsigned int iov_cnt, 44 size_t offset, void *buf, size_t bytes) 45 { 46 size_t done; 47 unsigned int i; 48 for (i = 0, done = 0; (offset || done < bytes) && i < iov_cnt; i++) { 49 if (offset < iov[i].iov_len) { 50 size_t len = MIN(iov[i].iov_len - offset, bytes - done); 51 memcpy(buf + done, iov[i].iov_base + offset, len); 52 done += len; 53 offset = 0; 54 } else { 55 offset -= iov[i].iov_len; 56 } 57 } 58 assert(offset == 0); 59 return done; 60 } 61 62 size_t iov_memset(const struct iovec *iov, const unsigned int iov_cnt, 63 size_t offset, int fillc, size_t bytes) 64 { 65 size_t done; 66 unsigned int i; 67 for (i = 0, done = 0; (offset || done < bytes) && i < iov_cnt; i++) { 68 if (offset < iov[i].iov_len) { 69 size_t len = MIN(iov[i].iov_len - offset, bytes - done); 70 memset(iov[i].iov_base + offset, fillc, len); 71 done += len; 72 offset = 0; 73 } else { 74 offset -= iov[i].iov_len; 75 } 76 } 77 assert(offset == 0); 78 return done; 79 } 80 81 size_t iov_size(const struct iovec *iov, const unsigned int iov_cnt) 82 { 83 size_t len; 84 unsigned int i; 85 86 len = 0; 87 for (i = 0; i < iov_cnt; i++) { 88 len += iov[i].iov_len; 89 } 90 return len; 91 } 92 93 /* helper function for iov_send_recv() */ 94 static ssize_t 95 do_send_recv(int sockfd, struct iovec *iov, unsigned iov_cnt, bool do_send) 96 { 97 #ifdef CONFIG_POSIX 98 ssize_t ret; 99 struct msghdr msg; 100 memset(&msg, 0, sizeof(msg)); 101 msg.msg_iov = iov; 102 msg.msg_iovlen = iov_cnt; 103 do { 104 ret = do_send 105 ? sendmsg(sockfd, &msg, 0) 106 : recvmsg(sockfd, &msg, 0); 107 } while (ret < 0 && errno == EINTR); 108 return ret; 109 #else 110 /* else send piece-by-piece */ 111 /*XXX Note: windows has WSASend() and WSARecv() */ 112 unsigned i = 0; 113 ssize_t ret = 0; 114 ssize_t off = 0; 115 while (i < iov_cnt) { 116 ssize_t r = do_send 117 ? send(sockfd, iov[i].iov_base + off, iov[i].iov_len - off, 0) 118 : recv(sockfd, iov[i].iov_base + off, iov[i].iov_len - off, 0); 119 if (r > 0) { 120 ret += r; 121 off += r; 122 if (off < iov[i].iov_len) { 123 continue; 124 } 125 } else if (!r) { 126 break; 127 } else if (errno == EINTR) { 128 continue; 129 } else { 130 /* else it is some "other" error, 131 * only return if there was no data processed. */ 132 if (ret == 0) { 133 ret = -1; 134 } 135 break; 136 } 137 off = 0; 138 i++; 139 } 140 return ret; 141 #endif 142 } 143 144 ssize_t iov_send_recv(int sockfd, const struct iovec *_iov, unsigned iov_cnt, 145 size_t offset, size_t bytes, 146 bool do_send) 147 { 148 ssize_t total = 0; 149 ssize_t ret; 150 size_t orig_len, tail; 151 unsigned niov; 152 struct iovec *local_iov, *iov; 153 154 if (bytes <= 0) { 155 return 0; 156 } 157 158 local_iov = g_new0(struct iovec, iov_cnt); 159 iov_copy(local_iov, iov_cnt, _iov, iov_cnt, offset, bytes); 160 offset = 0; 161 iov = local_iov; 162 163 while (bytes > 0) { 164 /* Find the start position, skipping `offset' bytes: 165 * first, skip all full-sized vector elements, */ 166 for (niov = 0; niov < iov_cnt && offset >= iov[niov].iov_len; ++niov) { 167 offset -= iov[niov].iov_len; 168 } 169 170 /* niov == iov_cnt would only be valid if bytes == 0, which 171 * we already ruled out in the loop condition. */ 172 assert(niov < iov_cnt); 173 iov += niov; 174 iov_cnt -= niov; 175 176 if (offset) { 177 /* second, skip `offset' bytes from the (now) first element, 178 * undo it on exit */ 179 iov[0].iov_base += offset; 180 iov[0].iov_len -= offset; 181 } 182 /* Find the end position skipping `bytes' bytes: */ 183 /* first, skip all full-sized elements */ 184 tail = bytes; 185 for (niov = 0; niov < iov_cnt && iov[niov].iov_len <= tail; ++niov) { 186 tail -= iov[niov].iov_len; 187 } 188 if (tail) { 189 /* second, fixup the last element, and remember the original 190 * length */ 191 assert(niov < iov_cnt); 192 assert(iov[niov].iov_len > tail); 193 orig_len = iov[niov].iov_len; 194 iov[niov++].iov_len = tail; 195 ret = do_send_recv(sockfd, iov, niov, do_send); 196 /* Undo the changes above before checking for errors */ 197 iov[niov-1].iov_len = orig_len; 198 } else { 199 ret = do_send_recv(sockfd, iov, niov, do_send); 200 } 201 if (offset) { 202 iov[0].iov_base -= offset; 203 iov[0].iov_len += offset; 204 } 205 206 if (ret < 0) { 207 assert(errno != EINTR); 208 g_free(local_iov); 209 if (errno == EAGAIN && total > 0) { 210 return total; 211 } 212 return -1; 213 } 214 215 if (ret == 0 && !do_send) { 216 /* recv returns 0 when the peer has performed an orderly 217 * shutdown. */ 218 break; 219 } 220 221 /* Prepare for the next iteration */ 222 offset += ret; 223 total += ret; 224 bytes -= ret; 225 } 226 227 g_free(local_iov); 228 return total; 229 } 230 231 232 void iov_hexdump(const struct iovec *iov, const unsigned int iov_cnt, 233 FILE *fp, const char *prefix, size_t limit) 234 { 235 int v; 236 size_t size = 0; 237 char *buf; 238 239 for (v = 0; v < iov_cnt; v++) { 240 size += iov[v].iov_len; 241 } 242 size = size > limit ? limit : size; 243 buf = g_malloc(size); 244 iov_to_buf(iov, iov_cnt, 0, buf, size); 245 qemu_hexdump(fp, prefix, buf, size); 246 g_free(buf); 247 } 248 249 unsigned iov_copy(struct iovec *dst_iov, unsigned int dst_iov_cnt, 250 const struct iovec *iov, unsigned int iov_cnt, 251 size_t offset, size_t bytes) 252 { 253 size_t len; 254 unsigned int i, j; 255 for (i = 0, j = 0; 256 i < iov_cnt && j < dst_iov_cnt && (offset || bytes); i++) { 257 if (offset >= iov[i].iov_len) { 258 offset -= iov[i].iov_len; 259 continue; 260 } 261 len = MIN(bytes, iov[i].iov_len - offset); 262 263 dst_iov[j].iov_base = iov[i].iov_base + offset; 264 dst_iov[j].iov_len = len; 265 j++; 266 bytes -= len; 267 offset = 0; 268 } 269 assert(offset == 0); 270 return j; 271 } 272 273 /* io vectors */ 274 275 void qemu_iovec_init(QEMUIOVector *qiov, int alloc_hint) 276 { 277 qiov->iov = g_new(struct iovec, alloc_hint); 278 qiov->niov = 0; 279 qiov->nalloc = alloc_hint; 280 qiov->size = 0; 281 } 282 283 void qemu_iovec_init_external(QEMUIOVector *qiov, struct iovec *iov, int niov) 284 { 285 int i; 286 287 qiov->iov = iov; 288 qiov->niov = niov; 289 qiov->nalloc = -1; 290 qiov->size = 0; 291 for (i = 0; i < niov; i++) 292 qiov->size += iov[i].iov_len; 293 } 294 295 void qemu_iovec_add(QEMUIOVector *qiov, void *base, size_t len) 296 { 297 assert(qiov->nalloc != -1); 298 299 if (qiov->niov == qiov->nalloc) { 300 qiov->nalloc = 2 * qiov->nalloc + 1; 301 qiov->iov = g_renew(struct iovec, qiov->iov, qiov->nalloc); 302 } 303 qiov->iov[qiov->niov].iov_base = base; 304 qiov->iov[qiov->niov].iov_len = len; 305 qiov->size += len; 306 ++qiov->niov; 307 } 308 309 /* 310 * Concatenates (partial) iovecs from src_iov to the end of dst. 311 * It starts copying after skipping `soffset' bytes at the 312 * beginning of src and adds individual vectors from src to 313 * dst copies up to `sbytes' bytes total, or up to the end 314 * of src_iov if it comes first. This way, it is okay to specify 315 * very large value for `sbytes' to indicate "up to the end 316 * of src". 317 * Only vector pointers are processed, not the actual data buffers. 318 */ 319 size_t qemu_iovec_concat_iov(QEMUIOVector *dst, 320 struct iovec *src_iov, unsigned int src_cnt, 321 size_t soffset, size_t sbytes) 322 { 323 int i; 324 size_t done; 325 326 if (!sbytes) { 327 return 0; 328 } 329 assert(dst->nalloc != -1); 330 for (i = 0, done = 0; done < sbytes && i < src_cnt; i++) { 331 if (soffset < src_iov[i].iov_len) { 332 size_t len = MIN(src_iov[i].iov_len - soffset, sbytes - done); 333 qemu_iovec_add(dst, src_iov[i].iov_base + soffset, len); 334 done += len; 335 soffset = 0; 336 } else { 337 soffset -= src_iov[i].iov_len; 338 } 339 } 340 assert(soffset == 0); /* offset beyond end of src */ 341 342 return done; 343 } 344 345 /* 346 * Concatenates (partial) iovecs from src to the end of dst. 347 * It starts copying after skipping `soffset' bytes at the 348 * beginning of src and adds individual vectors from src to 349 * dst copies up to `sbytes' bytes total, or up to the end 350 * of src if it comes first. This way, it is okay to specify 351 * very large value for `sbytes' to indicate "up to the end 352 * of src". 353 * Only vector pointers are processed, not the actual data buffers. 354 */ 355 void qemu_iovec_concat(QEMUIOVector *dst, 356 QEMUIOVector *src, size_t soffset, size_t sbytes) 357 { 358 qemu_iovec_concat_iov(dst, src->iov, src->niov, soffset, sbytes); 359 } 360 361 /* 362 * qiov_find_iov 363 * 364 * Return pointer to iovec structure, where byte at @offset in original vector 365 * @iov exactly is. 366 * Set @remaining_offset to be offset inside that iovec to the same byte. 367 */ 368 static struct iovec *iov_skip_offset(struct iovec *iov, size_t offset, 369 size_t *remaining_offset) 370 { 371 while (offset > 0 && offset >= iov->iov_len) { 372 offset -= iov->iov_len; 373 iov++; 374 } 375 *remaining_offset = offset; 376 377 return iov; 378 } 379 380 /* 381 * qiov_slice 382 * 383 * Find subarray of iovec's, containing requested range. @head would 384 * be offset in first iov (returned by the function), @tail would be 385 * count of extra bytes in last iovec (returned iov + @niov - 1). 386 */ 387 static struct iovec *qiov_slice(QEMUIOVector *qiov, 388 size_t offset, size_t len, 389 size_t *head, size_t *tail, int *niov) 390 { 391 struct iovec *iov, *end_iov; 392 393 assert(offset + len <= qiov->size); 394 395 iov = iov_skip_offset(qiov->iov, offset, head); 396 end_iov = iov_skip_offset(iov, *head + len, tail); 397 398 if (*tail > 0) { 399 assert(*tail < end_iov->iov_len); 400 *tail = end_iov->iov_len - *tail; 401 end_iov++; 402 } 403 404 *niov = end_iov - iov; 405 406 return iov; 407 } 408 409 int qemu_iovec_subvec_niov(QEMUIOVector *qiov, size_t offset, size_t len) 410 { 411 size_t head, tail; 412 int niov; 413 414 qiov_slice(qiov, offset, len, &head, &tail, &niov); 415 416 return niov; 417 } 418 419 /* 420 * Compile new iovec, combining @head_buf buffer, sub-qiov of @mid_qiov, 421 * and @tail_buf buffer into new qiov. 422 */ 423 int qemu_iovec_init_extended( 424 QEMUIOVector *qiov, 425 void *head_buf, size_t head_len, 426 QEMUIOVector *mid_qiov, size_t mid_offset, size_t mid_len, 427 void *tail_buf, size_t tail_len) 428 { 429 size_t mid_head, mid_tail; 430 int total_niov, mid_niov = 0; 431 struct iovec *p, *mid_iov = NULL; 432 433 assert(mid_qiov->niov <= IOV_MAX); 434 435 if (SIZE_MAX - head_len < mid_len || 436 SIZE_MAX - head_len - mid_len < tail_len) 437 { 438 return -EINVAL; 439 } 440 441 if (mid_len) { 442 mid_iov = qiov_slice(mid_qiov, mid_offset, mid_len, 443 &mid_head, &mid_tail, &mid_niov); 444 } 445 446 total_niov = !!head_len + mid_niov + !!tail_len; 447 if (total_niov > IOV_MAX) { 448 return -EINVAL; 449 } 450 451 if (total_niov == 1) { 452 qemu_iovec_init_buf(qiov, NULL, 0); 453 p = &qiov->local_iov; 454 } else { 455 qiov->niov = qiov->nalloc = total_niov; 456 qiov->size = head_len + mid_len + tail_len; 457 p = qiov->iov = g_new(struct iovec, qiov->niov); 458 } 459 460 if (head_len) { 461 p->iov_base = head_buf; 462 p->iov_len = head_len; 463 p++; 464 } 465 466 assert(!mid_niov == !mid_len); 467 if (mid_niov) { 468 memcpy(p, mid_iov, mid_niov * sizeof(*p)); 469 p[0].iov_base = (uint8_t *)p[0].iov_base + mid_head; 470 p[0].iov_len -= mid_head; 471 p[mid_niov - 1].iov_len -= mid_tail; 472 p += mid_niov; 473 } 474 475 if (tail_len) { 476 p->iov_base = tail_buf; 477 p->iov_len = tail_len; 478 } 479 480 return 0; 481 } 482 483 /* 484 * Check if the contents of subrange of qiov data is all zeroes. 485 */ 486 bool qemu_iovec_is_zero(QEMUIOVector *qiov, size_t offset, size_t bytes) 487 { 488 struct iovec *iov; 489 size_t current_offset; 490 491 assert(offset + bytes <= qiov->size); 492 493 iov = iov_skip_offset(qiov->iov, offset, ¤t_offset); 494 495 while (bytes) { 496 uint8_t *base = (uint8_t *)iov->iov_base + current_offset; 497 size_t len = MIN(iov->iov_len - current_offset, bytes); 498 499 if (!buffer_is_zero(base, len)) { 500 return false; 501 } 502 503 current_offset = 0; 504 bytes -= len; 505 iov++; 506 } 507 508 return true; 509 } 510 511 void qemu_iovec_init_slice(QEMUIOVector *qiov, QEMUIOVector *source, 512 size_t offset, size_t len) 513 { 514 int ret; 515 516 assert(source->size >= len); 517 assert(source->size - len >= offset); 518 519 /* We shrink the request, so we can't overflow neither size_t nor MAX_IOV */ 520 ret = qemu_iovec_init_extended(qiov, NULL, 0, source, offset, len, NULL, 0); 521 assert(ret == 0); 522 } 523 524 void qemu_iovec_destroy(QEMUIOVector *qiov) 525 { 526 if (qiov->nalloc != -1) { 527 g_free(qiov->iov); 528 } 529 530 memset(qiov, 0, sizeof(*qiov)); 531 } 532 533 void qemu_iovec_reset(QEMUIOVector *qiov) 534 { 535 assert(qiov->nalloc != -1); 536 537 qiov->niov = 0; 538 qiov->size = 0; 539 } 540 541 size_t qemu_iovec_to_buf(QEMUIOVector *qiov, size_t offset, 542 void *buf, size_t bytes) 543 { 544 return iov_to_buf(qiov->iov, qiov->niov, offset, buf, bytes); 545 } 546 547 size_t qemu_iovec_from_buf(QEMUIOVector *qiov, size_t offset, 548 const void *buf, size_t bytes) 549 { 550 return iov_from_buf(qiov->iov, qiov->niov, offset, buf, bytes); 551 } 552 553 size_t qemu_iovec_memset(QEMUIOVector *qiov, size_t offset, 554 int fillc, size_t bytes) 555 { 556 return iov_memset(qiov->iov, qiov->niov, offset, fillc, bytes); 557 } 558 559 /** 560 * Check that I/O vector contents are identical 561 * 562 * The IO vectors must have the same structure (same length of all parts). 563 * A typical usage is to compare vectors created with qemu_iovec_clone(). 564 * 565 * @a: I/O vector 566 * @b: I/O vector 567 * @ret: Offset to first mismatching byte or -1 if match 568 */ 569 ssize_t qemu_iovec_compare(QEMUIOVector *a, QEMUIOVector *b) 570 { 571 int i; 572 ssize_t offset = 0; 573 574 assert(a->niov == b->niov); 575 for (i = 0; i < a->niov; i++) { 576 size_t len = 0; 577 uint8_t *p = (uint8_t *)a->iov[i].iov_base; 578 uint8_t *q = (uint8_t *)b->iov[i].iov_base; 579 580 assert(a->iov[i].iov_len == b->iov[i].iov_len); 581 while (len < a->iov[i].iov_len && *p++ == *q++) { 582 len++; 583 } 584 585 offset += len; 586 587 if (len != a->iov[i].iov_len) { 588 return offset; 589 } 590 } 591 return -1; 592 } 593 594 typedef struct { 595 int src_index; 596 struct iovec *src_iov; 597 void *dest_base; 598 } IOVectorSortElem; 599 600 static int sortelem_cmp_src_base(const void *a, const void *b) 601 { 602 const IOVectorSortElem *elem_a = a; 603 const IOVectorSortElem *elem_b = b; 604 605 /* Don't overflow */ 606 if (elem_a->src_iov->iov_base < elem_b->src_iov->iov_base) { 607 return -1; 608 } else if (elem_a->src_iov->iov_base > elem_b->src_iov->iov_base) { 609 return 1; 610 } else { 611 return 0; 612 } 613 } 614 615 static int sortelem_cmp_src_index(const void *a, const void *b) 616 { 617 const IOVectorSortElem *elem_a = a; 618 const IOVectorSortElem *elem_b = b; 619 620 return elem_a->src_index - elem_b->src_index; 621 } 622 623 /** 624 * Copy contents of I/O vector 625 * 626 * The relative relationships of overlapping iovecs are preserved. This is 627 * necessary to ensure identical semantics in the cloned I/O vector. 628 */ 629 void qemu_iovec_clone(QEMUIOVector *dest, const QEMUIOVector *src, void *buf) 630 { 631 IOVectorSortElem sortelems[src->niov]; 632 void *last_end; 633 int i; 634 635 /* Sort by source iovecs by base address */ 636 for (i = 0; i < src->niov; i++) { 637 sortelems[i].src_index = i; 638 sortelems[i].src_iov = &src->iov[i]; 639 } 640 qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_base); 641 642 /* Allocate buffer space taking into account overlapping iovecs */ 643 last_end = NULL; 644 for (i = 0; i < src->niov; i++) { 645 struct iovec *cur = sortelems[i].src_iov; 646 ptrdiff_t rewind = 0; 647 648 /* Detect overlap */ 649 if (last_end && last_end > cur->iov_base) { 650 rewind = last_end - cur->iov_base; 651 } 652 653 sortelems[i].dest_base = buf - rewind; 654 buf += cur->iov_len - MIN(rewind, cur->iov_len); 655 last_end = MAX(cur->iov_base + cur->iov_len, last_end); 656 } 657 658 /* Sort by source iovec index and build destination iovec */ 659 qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_index); 660 for (i = 0; i < src->niov; i++) { 661 qemu_iovec_add(dest, sortelems[i].dest_base, src->iov[i].iov_len); 662 } 663 } 664 665 void iov_discard_undo(IOVDiscardUndo *undo) 666 { 667 /* Restore original iovec if it was modified */ 668 if (undo->modified_iov) { 669 *undo->modified_iov = undo->orig; 670 } 671 } 672 673 size_t iov_discard_front_undoable(struct iovec **iov, 674 unsigned int *iov_cnt, 675 size_t bytes, 676 IOVDiscardUndo *undo) 677 { 678 size_t total = 0; 679 struct iovec *cur; 680 681 if (undo) { 682 undo->modified_iov = NULL; 683 } 684 685 for (cur = *iov; *iov_cnt > 0; cur++) { 686 if (cur->iov_len > bytes) { 687 if (undo) { 688 undo->modified_iov = cur; 689 undo->orig = *cur; 690 } 691 692 cur->iov_base += bytes; 693 cur->iov_len -= bytes; 694 total += bytes; 695 break; 696 } 697 698 bytes -= cur->iov_len; 699 total += cur->iov_len; 700 *iov_cnt -= 1; 701 } 702 703 *iov = cur; 704 return total; 705 } 706 707 size_t iov_discard_front(struct iovec **iov, unsigned int *iov_cnt, 708 size_t bytes) 709 { 710 return iov_discard_front_undoable(iov, iov_cnt, bytes, NULL); 711 } 712 713 size_t iov_discard_back_undoable(struct iovec *iov, 714 unsigned int *iov_cnt, 715 size_t bytes, 716 IOVDiscardUndo *undo) 717 { 718 size_t total = 0; 719 struct iovec *cur; 720 721 if (undo) { 722 undo->modified_iov = NULL; 723 } 724 725 if (*iov_cnt == 0) { 726 return 0; 727 } 728 729 cur = iov + (*iov_cnt - 1); 730 731 while (*iov_cnt > 0) { 732 if (cur->iov_len > bytes) { 733 if (undo) { 734 undo->modified_iov = cur; 735 undo->orig = *cur; 736 } 737 738 cur->iov_len -= bytes; 739 total += bytes; 740 break; 741 } 742 743 bytes -= cur->iov_len; 744 total += cur->iov_len; 745 cur--; 746 *iov_cnt -= 1; 747 } 748 749 return total; 750 } 751 752 size_t iov_discard_back(struct iovec *iov, unsigned int *iov_cnt, 753 size_t bytes) 754 { 755 return iov_discard_back_undoable(iov, iov_cnt, bytes, NULL); 756 } 757 758 void qemu_iovec_discard_back(QEMUIOVector *qiov, size_t bytes) 759 { 760 size_t total; 761 unsigned int niov = qiov->niov; 762 763 assert(qiov->size >= bytes); 764 total = iov_discard_back(qiov->iov, &niov, bytes); 765 assert(total == bytes); 766 767 qiov->niov = niov; 768 qiov->size -= bytes; 769 } 770