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