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