util/iov.c

   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(buf, fp, prefix, 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 void 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     if (mid_len) {
 429         mid_iov = qiov_slice(mid_qiov, mid_offset, mid_len,
 430                              &mid_head, &mid_tail, &mid_niov);
 431     }
 432
 433     total_niov = !!head_len + mid_niov + !!tail_len;
 434     if (total_niov == 1) {
 435         qemu_iovec_init_buf(qiov, NULL, 0);
 436         p = &qiov->local_iov;
 437     } else {
 438         qiov->niov = qiov->nalloc = total_niov;
 439         qiov->size = head_len + mid_len + tail_len;
 440         p = qiov->iov = g_new(struct iovec, qiov->niov);
 441     }
 442
 443     if (head_len) {
 444         p->iov_base = head_buf;
 445         p->iov_len = head_len;
 446         p++;
 447     }
 448
 449     assert(!mid_niov == !mid_len);
 450     if (mid_niov) {
 451         memcpy(p, mid_iov, mid_niov * sizeof(*p));
 452         p[0].iov_base = (uint8_t *)p[0].iov_base + mid_head;
 453         p[0].iov_len -= mid_head;
 454         p[mid_niov - 1].iov_len -= mid_tail;
 455         p += mid_niov;
 456     }
 457
 458     if (tail_len) {
 459         p->iov_base = tail_buf;
 460         p->iov_len = tail_len;
 461     }
 462 }
 463
 464 /*
 465  * Check if the contents of subrange of qiov data is all zeroes.
 466  */
 467 bool qemu_iovec_is_zero(QEMUIOVector *qiov, size_t offset, size_t bytes)
 468 {
 469     struct iovec *iov;
 470     size_t current_offset;
 471
 472     assert(offset + bytes <= qiov->size);
 473
 474     iov = iov_skip_offset(qiov->iov, offset, &current_offset);
 475
 476     while (bytes) {
 477         uint8_t *base = (uint8_t *)iov->iov_base + current_offset;
 478         size_t len = MIN(iov->iov_len - current_offset, bytes);
 479
 480         if (!buffer_is_zero(base, len)) {
 481             return false;
 482         }
 483
 484         current_offset = 0;
 485         bytes -= len;
 486         iov++;
 487     }
 488
 489     return true;
 490 }
 491
 492 void qemu_iovec_init_slice(QEMUIOVector *qiov, QEMUIOVector *source,
 493                            size_t offset, size_t len)
 494 {
 495     qemu_iovec_init_extended(qiov, NULL, 0, source, offset, len, NULL, 0);
 496 }
 497
 498 void qemu_iovec_destroy(QEMUIOVector *qiov)
 499 {
 500     if (qiov->nalloc != -1) {
 501         g_free(qiov->iov);
 502     }
 503
 504     memset(qiov, 0, sizeof(*qiov));
 505 }
 506
 507 void qemu_iovec_reset(QEMUIOVector *qiov)
 508 {
 509     assert(qiov->nalloc != -1);
 510
 511     qiov->niov = 0;
 512     qiov->size = 0;
 513 }
 514
 515 size_t qemu_iovec_to_buf(QEMUIOVector *qiov, size_t offset,
 516                          void *buf, size_t bytes)
 517 {
 518     return iov_to_buf(qiov->iov, qiov->niov, offset, buf, bytes);
 519 }
 520
 521 size_t qemu_iovec_from_buf(QEMUIOVector *qiov, size_t offset,
 522                            const void *buf, size_t bytes)
 523 {
 524     return iov_from_buf(qiov->iov, qiov->niov, offset, buf, bytes);
 525 }
 526
 527 size_t qemu_iovec_memset(QEMUIOVector *qiov, size_t offset,
 528                          int fillc, size_t bytes)
 529 {
 530     return iov_memset(qiov->iov, qiov->niov, offset, fillc, bytes);
 531 }
 532
 533 /**
 534  * Check that I/O vector contents are identical
 535  *
 536  * The IO vectors must have the same structure (same length of all parts).
 537  * A typical usage is to compare vectors created with qemu_iovec_clone().
 538  *
 539  * @a:          I/O vector
 540  * @b:          I/O vector
 541  * @ret:        Offset to first mismatching byte or -1 if match
 542  */
 543 ssize_t qemu_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
 544 {
 545     int i;
 546     ssize_t offset = 0;
 547
 548     assert(a->niov == b->niov);
 549     for (i = 0; i < a->niov; i++) {
 550         size_t len = 0;
 551         uint8_t *p = (uint8_t *)a->iov[i].iov_base;
 552         uint8_t *q = (uint8_t *)b->iov[i].iov_base;
 553
 554         assert(a->iov[i].iov_len == b->iov[i].iov_len);
 555         while (len < a->iov[i].iov_len && *p++ == *q++) {
 556             len++;
 557         }
 558
 559         offset += len;
 560
 561         if (len != a->iov[i].iov_len) {
 562             return offset;
 563         }
 564     }
 565     return -1;
 566 }
 567
 568 typedef struct {
 569     int src_index;
 570     struct iovec *src_iov;
 571     void *dest_base;
 572 } IOVectorSortElem;
 573
 574 static int sortelem_cmp_src_base(const void *a, const void *b)
 575 {
 576     const IOVectorSortElem *elem_a = a;
 577     const IOVectorSortElem *elem_b = b;
 578
 579     /* Don't overflow */
 580     if (elem_a->src_iov->iov_base < elem_b->src_iov->iov_base) {
 581         return -1;
 582     } else if (elem_a->src_iov->iov_base > elem_b->src_iov->iov_base) {
 583         return 1;
 584     } else {
 585         return 0;
 586     }
 587 }
 588
 589 static int sortelem_cmp_src_index(const void *a, const void *b)
 590 {
 591     const IOVectorSortElem *elem_a = a;
 592     const IOVectorSortElem *elem_b = b;
 593
 594     return elem_a->src_index - elem_b->src_index;
 595 }
 596
 597 /**
 598  * Copy contents of I/O vector
 599  *
 600  * The relative relationships of overlapping iovecs are preserved.  This is
 601  * necessary to ensure identical semantics in the cloned I/O vector.
 602  */
 603 void qemu_iovec_clone(QEMUIOVector *dest, const QEMUIOVector *src, void *buf)
 604 {
 605     IOVectorSortElem sortelems[src->niov];
 606     void *last_end;
 607     int i;
 608
 609     /* Sort by source iovecs by base address */
 610     for (i = 0; i < src->niov; i++) {
 611         sortelems[i].src_index = i;
 612         sortelems[i].src_iov = &src->iov[i];
 613     }
 614     qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_base);
 615
 616     /* Allocate buffer space taking into account overlapping iovecs */
 617     last_end = NULL;
 618     for (i = 0; i < src->niov; i++) {
 619         struct iovec *cur = sortelems[i].src_iov;
 620         ptrdiff_t rewind = 0;
 621
 622         /* Detect overlap */
 623         if (last_end && last_end > cur->iov_base) {
 624             rewind = last_end - cur->iov_base;
 625         }
 626
 627         sortelems[i].dest_base = buf - rewind;
 628         buf += cur->iov_len - MIN(rewind, cur->iov_len);
 629         last_end = MAX(cur->iov_base + cur->iov_len, last_end);
 630     }
 631
 632     /* Sort by source iovec index and build destination iovec */
 633     qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_index);
 634     for (i = 0; i < src->niov; i++) {
 635         qemu_iovec_add(dest, sortelems[i].dest_base, src->iov[i].iov_len);
 636     }
 637 }
 638
 639 size_t iov_discard_front(struct iovec **iov, unsigned int *iov_cnt,
 640                          size_t bytes)
 641 {
 642     size_t total = 0;
 643     struct iovec *cur;
 644
 645     for (cur = *iov; *iov_cnt > 0; cur++) {
 646         if (cur->iov_len > bytes) {
 647             cur->iov_base += bytes;
 648             cur->iov_len -= bytes;
 649             total += bytes;
 650             break;
 651         }
 652
 653         bytes -= cur->iov_len;
 654         total += cur->iov_len;
 655         *iov_cnt -= 1;
 656     }
 657
 658     *iov = cur;
 659     return total;
 660 }
 661
 662 size_t iov_discard_back(struct iovec *iov, unsigned int *iov_cnt,
 663                         size_t bytes)
 664 {
 665     size_t total = 0;
 666     struct iovec *cur;
 667
 668     if (*iov_cnt == 0) {
 669         return 0;
 670     }
 671
 672     cur = iov + (*iov_cnt - 1);
 673
 674     while (*iov_cnt > 0) {
 675         if (cur->iov_len > bytes) {
 676             cur->iov_len -= bytes;
 677             total += bytes;
 678             break;
 679         }
 680
 681         bytes -= cur->iov_len;
 682         total += cur->iov_len;
 683         cur--;
 684         *iov_cnt -= 1;
 685     }
 686
 687     return total;
 688 }
 689
 690 void qemu_iovec_discard_back(QEMUIOVector *qiov, size_t bytes)
 691 {
 692     size_t total;
 693     unsigned int niov = qiov->niov;
 694
 695     assert(qiov->size >= bytes);
 696     total = iov_discard_back(qiov->iov, &niov, bytes);
 697     assert(total == bytes);
 698
 699     qiov->niov = niov;
 700     qiov->size -= bytes;
 701 }