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 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(buf, fp, prefix, 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; i < iov_cnt && j < dst_iov_cnt && bytes; i++) { 250 if (offset >= iov[i].iov_len) { 251 offset -= iov[i].iov_len; 252 continue; 253 } 254 len = MIN(bytes, iov[i].iov_len - offset); 255 256 dst_iov[j].iov_base = iov[i].iov_base + offset; 257 dst_iov[j].iov_len = len; 258 j++; 259 bytes -= len; 260 offset = 0; 261 } 262 assert(offset == 0); 263 return j; 264 } 265 266 /* io vectors */ 267 268 void qemu_iovec_init(QEMUIOVector *qiov, int alloc_hint) 269 { 270 qiov->iov = g_new(struct iovec, alloc_hint); 271 qiov->niov = 0; 272 qiov->nalloc = alloc_hint; 273 qiov->size = 0; 274 } 275 276 void qemu_iovec_init_external(QEMUIOVector *qiov, struct iovec *iov, int niov) 277 { 278 int i; 279 280 qiov->iov = iov; 281 qiov->niov = niov; 282 qiov->nalloc = -1; 283 qiov->size = 0; 284 for (i = 0; i < niov; i++) 285 qiov->size += iov[i].iov_len; 286 } 287 288 void qemu_iovec_add(QEMUIOVector *qiov, void *base, size_t len) 289 { 290 assert(qiov->nalloc != -1); 291 292 if (qiov->niov == qiov->nalloc) { 293 qiov->nalloc = 2 * qiov->nalloc + 1; 294 qiov->iov = g_renew(struct iovec, qiov->iov, qiov->nalloc); 295 } 296 qiov->iov[qiov->niov].iov_base = base; 297 qiov->iov[qiov->niov].iov_len = len; 298 qiov->size += len; 299 ++qiov->niov; 300 } 301 302 /* 303 * Concatenates (partial) iovecs from src_iov to the end of dst. 304 * It starts copying after skipping `soffset' bytes at the 305 * beginning of src and adds individual vectors from src to 306 * dst copies up to `sbytes' bytes total, or up to the end 307 * of src_iov if it comes first. This way, it is okay to specify 308 * very large value for `sbytes' to indicate "up to the end 309 * of src". 310 * Only vector pointers are processed, not the actual data buffers. 311 */ 312 size_t qemu_iovec_concat_iov(QEMUIOVector *dst, 313 struct iovec *src_iov, unsigned int src_cnt, 314 size_t soffset, size_t sbytes) 315 { 316 int i; 317 size_t done; 318 319 if (!sbytes) { 320 return 0; 321 } 322 assert(dst->nalloc != -1); 323 for (i = 0, done = 0; done < sbytes && i < src_cnt; i++) { 324 if (soffset < src_iov[i].iov_len) { 325 size_t len = MIN(src_iov[i].iov_len - soffset, sbytes - done); 326 qemu_iovec_add(dst, src_iov[i].iov_base + soffset, len); 327 done += len; 328 soffset = 0; 329 } else { 330 soffset -= src_iov[i].iov_len; 331 } 332 } 333 assert(soffset == 0); /* offset beyond end of src */ 334 335 return done; 336 } 337 338 /* 339 * Concatenates (partial) iovecs from src to the end of dst. 340 * It starts copying after skipping `soffset' bytes at the 341 * beginning of src and adds individual vectors from src to 342 * dst copies up to `sbytes' bytes total, or up to the end 343 * of src if it comes first. This way, it is okay to specify 344 * very large value for `sbytes' to indicate "up to the end 345 * of src". 346 * Only vector pointers are processed, not the actual data buffers. 347 */ 348 void qemu_iovec_concat(QEMUIOVector *dst, 349 QEMUIOVector *src, size_t soffset, size_t sbytes) 350 { 351 qemu_iovec_concat_iov(dst, src->iov, src->niov, soffset, sbytes); 352 } 353 354 /* 355 * Check if the contents of the iovecs are all zero 356 */ 357 bool qemu_iovec_is_zero(QEMUIOVector *qiov) 358 { 359 int i; 360 for (i = 0; i < qiov->niov; i++) { 361 size_t offs = QEMU_ALIGN_DOWN(qiov->iov[i].iov_len, 4 * sizeof(long)); 362 uint8_t *ptr = qiov->iov[i].iov_base; 363 if (offs && !buffer_is_zero(qiov->iov[i].iov_base, offs)) { 364 return false; 365 } 366 for (; offs < qiov->iov[i].iov_len; offs++) { 367 if (ptr[offs]) { 368 return false; 369 } 370 } 371 } 372 return true; 373 } 374 375 void qemu_iovec_destroy(QEMUIOVector *qiov) 376 { 377 assert(qiov->nalloc != -1); 378 379 qemu_iovec_reset(qiov); 380 g_free(qiov->iov); 381 qiov->nalloc = 0; 382 qiov->iov = NULL; 383 } 384 385 void qemu_iovec_reset(QEMUIOVector *qiov) 386 { 387 assert(qiov->nalloc != -1); 388 389 qiov->niov = 0; 390 qiov->size = 0; 391 } 392 393 size_t qemu_iovec_to_buf(QEMUIOVector *qiov, size_t offset, 394 void *buf, size_t bytes) 395 { 396 return iov_to_buf(qiov->iov, qiov->niov, offset, buf, bytes); 397 } 398 399 size_t qemu_iovec_from_buf(QEMUIOVector *qiov, size_t offset, 400 const void *buf, size_t bytes) 401 { 402 return iov_from_buf(qiov->iov, qiov->niov, offset, buf, bytes); 403 } 404 405 size_t qemu_iovec_memset(QEMUIOVector *qiov, size_t offset, 406 int fillc, size_t bytes) 407 { 408 return iov_memset(qiov->iov, qiov->niov, offset, fillc, bytes); 409 } 410 411 /** 412 * Check that I/O vector contents are identical 413 * 414 * The IO vectors must have the same structure (same length of all parts). 415 * A typical usage is to compare vectors created with qemu_iovec_clone(). 416 * 417 * @a: I/O vector 418 * @b: I/O vector 419 * @ret: Offset to first mismatching byte or -1 if match 420 */ 421 ssize_t qemu_iovec_compare(QEMUIOVector *a, QEMUIOVector *b) 422 { 423 int i; 424 ssize_t offset = 0; 425 426 assert(a->niov == b->niov); 427 for (i = 0; i < a->niov; i++) { 428 size_t len = 0; 429 uint8_t *p = (uint8_t *)a->iov[i].iov_base; 430 uint8_t *q = (uint8_t *)b->iov[i].iov_base; 431 432 assert(a->iov[i].iov_len == b->iov[i].iov_len); 433 while (len < a->iov[i].iov_len && *p++ == *q++) { 434 len++; 435 } 436 437 offset += len; 438 439 if (len != a->iov[i].iov_len) { 440 return offset; 441 } 442 } 443 return -1; 444 } 445 446 typedef struct { 447 int src_index; 448 struct iovec *src_iov; 449 void *dest_base; 450 } IOVectorSortElem; 451 452 static int sortelem_cmp_src_base(const void *a, const void *b) 453 { 454 const IOVectorSortElem *elem_a = a; 455 const IOVectorSortElem *elem_b = b; 456 457 /* Don't overflow */ 458 if (elem_a->src_iov->iov_base < elem_b->src_iov->iov_base) { 459 return -1; 460 } else if (elem_a->src_iov->iov_base > elem_b->src_iov->iov_base) { 461 return 1; 462 } else { 463 return 0; 464 } 465 } 466 467 static int sortelem_cmp_src_index(const void *a, const void *b) 468 { 469 const IOVectorSortElem *elem_a = a; 470 const IOVectorSortElem *elem_b = b; 471 472 return elem_a->src_index - elem_b->src_index; 473 } 474 475 /** 476 * Copy contents of I/O vector 477 * 478 * The relative relationships of overlapping iovecs are preserved. This is 479 * necessary to ensure identical semantics in the cloned I/O vector. 480 */ 481 void qemu_iovec_clone(QEMUIOVector *dest, const QEMUIOVector *src, void *buf) 482 { 483 IOVectorSortElem sortelems[src->niov]; 484 void *last_end; 485 int i; 486 487 /* Sort by source iovecs by base address */ 488 for (i = 0; i < src->niov; i++) { 489 sortelems[i].src_index = i; 490 sortelems[i].src_iov = &src->iov[i]; 491 } 492 qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_base); 493 494 /* Allocate buffer space taking into account overlapping iovecs */ 495 last_end = NULL; 496 for (i = 0; i < src->niov; i++) { 497 struct iovec *cur = sortelems[i].src_iov; 498 ptrdiff_t rewind = 0; 499 500 /* Detect overlap */ 501 if (last_end && last_end > cur->iov_base) { 502 rewind = last_end - cur->iov_base; 503 } 504 505 sortelems[i].dest_base = buf - rewind; 506 buf += cur->iov_len - MIN(rewind, cur->iov_len); 507 last_end = MAX(cur->iov_base + cur->iov_len, last_end); 508 } 509 510 /* Sort by source iovec index and build destination iovec */ 511 qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_index); 512 for (i = 0; i < src->niov; i++) { 513 qemu_iovec_add(dest, sortelems[i].dest_base, src->iov[i].iov_len); 514 } 515 } 516 517 size_t iov_discard_front(struct iovec **iov, unsigned int *iov_cnt, 518 size_t bytes) 519 { 520 size_t total = 0; 521 struct iovec *cur; 522 523 for (cur = *iov; *iov_cnt > 0; cur++) { 524 if (cur->iov_len > bytes) { 525 cur->iov_base += bytes; 526 cur->iov_len -= bytes; 527 total += bytes; 528 break; 529 } 530 531 bytes -= cur->iov_len; 532 total += cur->iov_len; 533 *iov_cnt -= 1; 534 } 535 536 *iov = cur; 537 return total; 538 } 539 540 size_t iov_discard_back(struct iovec *iov, unsigned int *iov_cnt, 541 size_t bytes) 542 { 543 size_t total = 0; 544 struct iovec *cur; 545 546 if (*iov_cnt == 0) { 547 return 0; 548 } 549 550 cur = iov + (*iov_cnt - 1); 551 552 while (*iov_cnt > 0) { 553 if (cur->iov_len > bytes) { 554 cur->iov_len -= bytes; 555 total += bytes; 556 break; 557 } 558 559 bytes -= cur->iov_len; 560 total += cur->iov_len; 561 cur--; 562 *iov_cnt -= 1; 563 } 564 565 return total; 566 } 567 568 void qemu_iovec_discard_back(QEMUIOVector *qiov, size_t bytes) 569 { 570 size_t total; 571 unsigned int niov = qiov->niov; 572 573 assert(qiov->size >= bytes); 574 total = iov_discard_back(qiov->iov, &niov, bytes); 575 assert(total == bytes); 576 577 qiov->niov = niov; 578 qiov->size -= bytes; 579 } 580