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/iov.h" 20 #include "qemu/sockets.h" 21 22 size_t iov_from_buf(const struct iovec *iov, unsigned int iov_cnt, 23 size_t offset, const void *buf, size_t bytes) 24 { 25 size_t done; 26 unsigned int i; 27 for (i = 0, done = 0; (offset || done < bytes) && i < iov_cnt; i++) { 28 if (offset < iov[i].iov_len) { 29 size_t len = MIN(iov[i].iov_len - offset, bytes - done); 30 memcpy(iov[i].iov_base + offset, buf + done, len); 31 done += len; 32 offset = 0; 33 } else { 34 offset -= iov[i].iov_len; 35 } 36 } 37 assert(offset == 0); 38 return done; 39 } 40 41 size_t iov_to_buf(const struct iovec *iov, const unsigned int iov_cnt, 42 size_t offset, void *buf, size_t bytes) 43 { 44 size_t done; 45 unsigned int i; 46 for (i = 0, done = 0; (offset || done < bytes) && i < iov_cnt; i++) { 47 if (offset < iov[i].iov_len) { 48 size_t len = MIN(iov[i].iov_len - offset, bytes - done); 49 memcpy(buf + done, iov[i].iov_base + offset, len); 50 done += len; 51 offset = 0; 52 } else { 53 offset -= iov[i].iov_len; 54 } 55 } 56 assert(offset == 0); 57 return done; 58 } 59 60 size_t iov_memset(const struct iovec *iov, const unsigned int iov_cnt, 61 size_t offset, int fillc, size_t bytes) 62 { 63 size_t done; 64 unsigned int i; 65 for (i = 0, done = 0; (offset || done < bytes) && i < iov_cnt; i++) { 66 if (offset < iov[i].iov_len) { 67 size_t len = MIN(iov[i].iov_len - offset, bytes - done); 68 memset(iov[i].iov_base + offset, fillc, len); 69 done += len; 70 offset = 0; 71 } else { 72 offset -= iov[i].iov_len; 73 } 74 } 75 assert(offset == 0); 76 return done; 77 } 78 79 size_t iov_size(const struct iovec *iov, const unsigned int iov_cnt) 80 { 81 size_t len; 82 unsigned int i; 83 84 len = 0; 85 for (i = 0; i < iov_cnt; i++) { 86 len += iov[i].iov_len; 87 } 88 return len; 89 } 90 91 /* helper function for iov_send_recv() */ 92 static ssize_t 93 do_send_recv(int sockfd, struct iovec *iov, unsigned iov_cnt, bool do_send) 94 { 95 #ifdef CONFIG_POSIX 96 ssize_t ret; 97 struct msghdr msg; 98 memset(&msg, 0, sizeof(msg)); 99 msg.msg_iov = iov; 100 msg.msg_iovlen = iov_cnt; 101 do { 102 ret = do_send 103 ? sendmsg(sockfd, &msg, 0) 104 : recvmsg(sockfd, &msg, 0); 105 } while (ret < 0 && errno == EINTR); 106 return ret; 107 #else 108 /* else send piece-by-piece */ 109 /*XXX Note: windows has WSASend() and WSARecv() */ 110 unsigned i = 0; 111 ssize_t ret = 0; 112 while (i < iov_cnt) { 113 ssize_t r = do_send 114 ? send(sockfd, iov[i].iov_base, iov[i].iov_len, 0) 115 : recv(sockfd, iov[i].iov_base, iov[i].iov_len, 0); 116 if (r > 0) { 117 ret += r; 118 } else if (!r) { 119 break; 120 } else if (errno == EINTR) { 121 continue; 122 } else { 123 /* else it is some "other" error, 124 * only return if there was no data processed. */ 125 if (ret == 0) { 126 ret = -1; 127 } 128 break; 129 } 130 i++; 131 } 132 return ret; 133 #endif 134 } 135 136 ssize_t iov_send_recv(int sockfd, const struct iovec *_iov, unsigned iov_cnt, 137 size_t offset, size_t bytes, 138 bool do_send) 139 { 140 ssize_t total = 0; 141 ssize_t ret; 142 size_t orig_len, tail; 143 unsigned niov; 144 struct iovec *local_iov, *iov; 145 146 if (bytes <= 0) { 147 return 0; 148 } 149 150 local_iov = g_new0(struct iovec, iov_cnt); 151 iov_copy(local_iov, iov_cnt, _iov, iov_cnt, offset, bytes); 152 offset = 0; 153 iov = local_iov; 154 155 while (bytes > 0) { 156 /* Find the start position, skipping `offset' bytes: 157 * first, skip all full-sized vector elements, */ 158 for (niov = 0; niov < iov_cnt && offset >= iov[niov].iov_len; ++niov) { 159 offset -= iov[niov].iov_len; 160 } 161 162 /* niov == iov_cnt would only be valid if bytes == 0, which 163 * we already ruled out in the loop condition. */ 164 assert(niov < iov_cnt); 165 iov += niov; 166 iov_cnt -= niov; 167 168 if (offset) { 169 /* second, skip `offset' bytes from the (now) first element, 170 * undo it on exit */ 171 iov[0].iov_base += offset; 172 iov[0].iov_len -= offset; 173 } 174 /* Find the end position skipping `bytes' bytes: */ 175 /* first, skip all full-sized elements */ 176 tail = bytes; 177 for (niov = 0; niov < iov_cnt && iov[niov].iov_len <= tail; ++niov) { 178 tail -= iov[niov].iov_len; 179 } 180 if (tail) { 181 /* second, fixup the last element, and remember the original 182 * length */ 183 assert(niov < iov_cnt); 184 assert(iov[niov].iov_len > tail); 185 orig_len = iov[niov].iov_len; 186 iov[niov++].iov_len = tail; 187 ret = do_send_recv(sockfd, iov, niov, do_send); 188 /* Undo the changes above before checking for errors */ 189 iov[niov-1].iov_len = orig_len; 190 } else { 191 ret = do_send_recv(sockfd, iov, niov, do_send); 192 } 193 if (offset) { 194 iov[0].iov_base -= offset; 195 iov[0].iov_len += offset; 196 } 197 198 if (ret < 0) { 199 assert(errno != EINTR); 200 g_free(local_iov); 201 if (errno == EAGAIN && total > 0) { 202 return total; 203 } 204 return -1; 205 } 206 207 if (ret == 0 && !do_send) { 208 /* recv returns 0 when the peer has performed an orderly 209 * shutdown. */ 210 break; 211 } 212 213 /* Prepare for the next iteration */ 214 offset += ret; 215 total += ret; 216 bytes -= ret; 217 } 218 219 g_free(local_iov); 220 return total; 221 } 222 223 224 void iov_hexdump(const struct iovec *iov, const unsigned int iov_cnt, 225 FILE *fp, const char *prefix, size_t limit) 226 { 227 int v; 228 size_t size = 0; 229 char *buf; 230 231 for (v = 0; v < iov_cnt; v++) { 232 size += iov[v].iov_len; 233 } 234 size = size > limit ? limit : size; 235 buf = g_malloc(size); 236 iov_to_buf(iov, iov_cnt, 0, buf, size); 237 qemu_hexdump(buf, fp, prefix, size); 238 g_free(buf); 239 } 240 241 unsigned iov_copy(struct iovec *dst_iov, unsigned int dst_iov_cnt, 242 const struct iovec *iov, unsigned int iov_cnt, 243 size_t offset, size_t bytes) 244 { 245 size_t len; 246 unsigned int i, j; 247 for (i = 0, j = 0; i < iov_cnt && j < dst_iov_cnt && bytes; i++) { 248 if (offset >= iov[i].iov_len) { 249 offset -= iov[i].iov_len; 250 continue; 251 } 252 len = MIN(bytes, iov[i].iov_len - offset); 253 254 dst_iov[j].iov_base = iov[i].iov_base + offset; 255 dst_iov[j].iov_len = len; 256 j++; 257 bytes -= len; 258 offset = 0; 259 } 260 assert(offset == 0); 261 return j; 262 } 263 264 /* io vectors */ 265 266 void qemu_iovec_init(QEMUIOVector *qiov, int alloc_hint) 267 { 268 qiov->iov = g_new(struct iovec, alloc_hint); 269 qiov->niov = 0; 270 qiov->nalloc = alloc_hint; 271 qiov->size = 0; 272 } 273 274 void qemu_iovec_init_external(QEMUIOVector *qiov, struct iovec *iov, int niov) 275 { 276 int i; 277 278 qiov->iov = iov; 279 qiov->niov = niov; 280 qiov->nalloc = -1; 281 qiov->size = 0; 282 for (i = 0; i < niov; i++) 283 qiov->size += iov[i].iov_len; 284 } 285 286 void qemu_iovec_add(QEMUIOVector *qiov, void *base, size_t len) 287 { 288 assert(qiov->nalloc != -1); 289 290 if (qiov->niov == qiov->nalloc) { 291 qiov->nalloc = 2 * qiov->nalloc + 1; 292 qiov->iov = g_renew(struct iovec, qiov->iov, qiov->nalloc); 293 } 294 qiov->iov[qiov->niov].iov_base = base; 295 qiov->iov[qiov->niov].iov_len = len; 296 qiov->size += len; 297 ++qiov->niov; 298 } 299 300 /* 301 * Concatenates (partial) iovecs from src_iov to the end of dst. 302 * It starts copying after skipping `soffset' bytes at the 303 * beginning of src and adds individual vectors from src to 304 * dst copies up to `sbytes' bytes total, or up to the end 305 * of src_iov if it comes first. This way, it is okay to specify 306 * very large value for `sbytes' to indicate "up to the end 307 * of src". 308 * Only vector pointers are processed, not the actual data buffers. 309 */ 310 size_t qemu_iovec_concat_iov(QEMUIOVector *dst, 311 struct iovec *src_iov, unsigned int src_cnt, 312 size_t soffset, size_t sbytes) 313 { 314 int i; 315 size_t done; 316 317 if (!sbytes) { 318 return 0; 319 } 320 assert(dst->nalloc != -1); 321 for (i = 0, done = 0; done < sbytes && i < src_cnt; i++) { 322 if (soffset < src_iov[i].iov_len) { 323 size_t len = MIN(src_iov[i].iov_len - soffset, sbytes - done); 324 qemu_iovec_add(dst, src_iov[i].iov_base + soffset, len); 325 done += len; 326 soffset = 0; 327 } else { 328 soffset -= src_iov[i].iov_len; 329 } 330 } 331 assert(soffset == 0); /* offset beyond end of src */ 332 333 return done; 334 } 335 336 /* 337 * Concatenates (partial) iovecs from src to the end of dst. 338 * It starts copying after skipping `soffset' bytes at the 339 * beginning of src and adds individual vectors from src to 340 * dst copies up to `sbytes' bytes total, or up to the end 341 * of src if it comes first. This way, it is okay to specify 342 * very large value for `sbytes' to indicate "up to the end 343 * of src". 344 * Only vector pointers are processed, not the actual data buffers. 345 */ 346 void qemu_iovec_concat(QEMUIOVector *dst, 347 QEMUIOVector *src, size_t soffset, size_t sbytes) 348 { 349 qemu_iovec_concat_iov(dst, src->iov, src->niov, soffset, sbytes); 350 } 351 352 /* 353 * Check if the contents of the iovecs are all zero 354 */ 355 bool qemu_iovec_is_zero(QEMUIOVector *qiov) 356 { 357 int i; 358 for (i = 0; i < qiov->niov; i++) { 359 size_t offs = QEMU_ALIGN_DOWN(qiov->iov[i].iov_len, 4 * sizeof(long)); 360 uint8_t *ptr = qiov->iov[i].iov_base; 361 if (offs && !buffer_is_zero(qiov->iov[i].iov_base, offs)) { 362 return false; 363 } 364 for (; offs < qiov->iov[i].iov_len; offs++) { 365 if (ptr[offs]) { 366 return false; 367 } 368 } 369 } 370 return true; 371 } 372 373 void qemu_iovec_destroy(QEMUIOVector *qiov) 374 { 375 assert(qiov->nalloc != -1); 376 377 qemu_iovec_reset(qiov); 378 g_free(qiov->iov); 379 qiov->nalloc = 0; 380 qiov->iov = NULL; 381 } 382 383 void qemu_iovec_reset(QEMUIOVector *qiov) 384 { 385 assert(qiov->nalloc != -1); 386 387 qiov->niov = 0; 388 qiov->size = 0; 389 } 390 391 size_t qemu_iovec_to_buf(QEMUIOVector *qiov, size_t offset, 392 void *buf, size_t bytes) 393 { 394 return iov_to_buf(qiov->iov, qiov->niov, offset, buf, bytes); 395 } 396 397 size_t qemu_iovec_from_buf(QEMUIOVector *qiov, size_t offset, 398 const void *buf, size_t bytes) 399 { 400 return iov_from_buf(qiov->iov, qiov->niov, offset, buf, bytes); 401 } 402 403 size_t qemu_iovec_memset(QEMUIOVector *qiov, size_t offset, 404 int fillc, size_t bytes) 405 { 406 return iov_memset(qiov->iov, qiov->niov, offset, fillc, bytes); 407 } 408 409 /** 410 * Check that I/O vector contents are identical 411 * 412 * The IO vectors must have the same structure (same length of all parts). 413 * A typical usage is to compare vectors created with qemu_iovec_clone(). 414 * 415 * @a: I/O vector 416 * @b: I/O vector 417 * @ret: Offset to first mismatching byte or -1 if match 418 */ 419 ssize_t qemu_iovec_compare(QEMUIOVector *a, QEMUIOVector *b) 420 { 421 int i; 422 ssize_t offset = 0; 423 424 assert(a->niov == b->niov); 425 for (i = 0; i < a->niov; i++) { 426 size_t len = 0; 427 uint8_t *p = (uint8_t *)a->iov[i].iov_base; 428 uint8_t *q = (uint8_t *)b->iov[i].iov_base; 429 430 assert(a->iov[i].iov_len == b->iov[i].iov_len); 431 while (len < a->iov[i].iov_len && *p++ == *q++) { 432 len++; 433 } 434 435 offset += len; 436 437 if (len != a->iov[i].iov_len) { 438 return offset; 439 } 440 } 441 return -1; 442 } 443 444 typedef struct { 445 int src_index; 446 struct iovec *src_iov; 447 void *dest_base; 448 } IOVectorSortElem; 449 450 static int sortelem_cmp_src_base(const void *a, const void *b) 451 { 452 const IOVectorSortElem *elem_a = a; 453 const IOVectorSortElem *elem_b = b; 454 455 /* Don't overflow */ 456 if (elem_a->src_iov->iov_base < elem_b->src_iov->iov_base) { 457 return -1; 458 } else if (elem_a->src_iov->iov_base > elem_b->src_iov->iov_base) { 459 return 1; 460 } else { 461 return 0; 462 } 463 } 464 465 static int sortelem_cmp_src_index(const void *a, const void *b) 466 { 467 const IOVectorSortElem *elem_a = a; 468 const IOVectorSortElem *elem_b = b; 469 470 return elem_a->src_index - elem_b->src_index; 471 } 472 473 /** 474 * Copy contents of I/O vector 475 * 476 * The relative relationships of overlapping iovecs are preserved. This is 477 * necessary to ensure identical semantics in the cloned I/O vector. 478 */ 479 void qemu_iovec_clone(QEMUIOVector *dest, const QEMUIOVector *src, void *buf) 480 { 481 IOVectorSortElem sortelems[src->niov]; 482 void *last_end; 483 int i; 484 485 /* Sort by source iovecs by base address */ 486 for (i = 0; i < src->niov; i++) { 487 sortelems[i].src_index = i; 488 sortelems[i].src_iov = &src->iov[i]; 489 } 490 qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_base); 491 492 /* Allocate buffer space taking into account overlapping iovecs */ 493 last_end = NULL; 494 for (i = 0; i < src->niov; i++) { 495 struct iovec *cur = sortelems[i].src_iov; 496 ptrdiff_t rewind = 0; 497 498 /* Detect overlap */ 499 if (last_end && last_end > cur->iov_base) { 500 rewind = last_end - cur->iov_base; 501 } 502 503 sortelems[i].dest_base = buf - rewind; 504 buf += cur->iov_len - MIN(rewind, cur->iov_len); 505 last_end = MAX(cur->iov_base + cur->iov_len, last_end); 506 } 507 508 /* Sort by source iovec index and build destination iovec */ 509 qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_index); 510 for (i = 0; i < src->niov; i++) { 511 qemu_iovec_add(dest, sortelems[i].dest_base, src->iov[i].iov_len); 512 } 513 } 514 515 size_t iov_discard_front(struct iovec **iov, unsigned int *iov_cnt, 516 size_t bytes) 517 { 518 size_t total = 0; 519 struct iovec *cur; 520 521 for (cur = *iov; *iov_cnt > 0; cur++) { 522 if (cur->iov_len > bytes) { 523 cur->iov_base += bytes; 524 cur->iov_len -= bytes; 525 total += bytes; 526 break; 527 } 528 529 bytes -= cur->iov_len; 530 total += cur->iov_len; 531 *iov_cnt -= 1; 532 } 533 534 *iov = cur; 535 return total; 536 } 537 538 size_t iov_discard_back(struct iovec *iov, unsigned int *iov_cnt, 539 size_t bytes) 540 { 541 size_t total = 0; 542 struct iovec *cur; 543 544 if (*iov_cnt == 0) { 545 return 0; 546 } 547 548 cur = iov + (*iov_cnt - 1); 549 550 while (*iov_cnt > 0) { 551 if (cur->iov_len > bytes) { 552 cur->iov_len -= bytes; 553 total += bytes; 554 break; 555 } 556 557 bytes -= cur->iov_len; 558 total += cur->iov_len; 559 cur--; 560 *iov_cnt -= 1; 561 } 562 563 return total; 564 } 565 566 void qemu_iovec_discard_back(QEMUIOVector *qiov, size_t bytes) 567 { 568 size_t total; 569 unsigned int niov = qiov->niov; 570 571 assert(qiov->size >= bytes); 572 total = iov_discard_back(qiov->iov, &niov, bytes); 573 assert(total == bytes); 574 575 qiov->niov = niov; 576 qiov->size -= bytes; 577 } 578