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/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, 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
 145     while (bytes > 0) {
 146         /* Find the start position, skipping `offset' bytes:
 147          * first, skip all full-sized vector elements, */
 148         for (niov = 0; niov < iov_cnt && offset >= iov[niov].iov_len; ++niov) {
 149             offset -= iov[niov].iov_len;
 150         }
 151
 152         /* niov == iov_cnt would only be valid if bytes == 0, which
 153          * we already ruled out in the loop condition.  */
 154         assert(niov < iov_cnt);
 155         iov += niov;
 156         iov_cnt -= niov;
 157
 158         if (offset) {
 159             /* second, skip `offset' bytes from the (now) first element,
 160              * undo it on exit */
 161             iov[0].iov_base += offset;
 162             iov[0].iov_len -= offset;
 163         }
 164         /* Find the end position skipping `bytes' bytes: */
 165         /* first, skip all full-sized elements */
 166         tail = bytes;
 167         for (niov = 0; niov < iov_cnt && iov[niov].iov_len <= tail; ++niov) {
 168             tail -= iov[niov].iov_len;
 169         }
 170         if (tail) {
 171             /* second, fixup the last element, and remember the original
 172              * length */
 173             assert(niov < iov_cnt);
 174             assert(iov[niov].iov_len > tail);
 175             orig_len = iov[niov].iov_len;
 176             iov[niov++].iov_len = tail;
 177             ret = do_send_recv(sockfd, iov, niov, do_send);
 178             /* Undo the changes above before checking for errors */
 179             iov[niov-1].iov_len = orig_len;
 180         } else {
 181             ret = do_send_recv(sockfd, iov, niov, do_send);
 182         }
 183         if (offset) {
 184             iov[0].iov_base -= offset;
 185             iov[0].iov_len += offset;
 186         }
 187
 188         if (ret < 0) {
 189             assert(errno != EINTR);
 190             if (errno == EAGAIN && total > 0) {
 191                 return total;
 192             }
 193             return -1;
 194         }
 195
 196         if (ret == 0 && !do_send) {
 197             /* recv returns 0 when the peer has performed an orderly
 198              * shutdown. */
 199             break;
 200         }
 201
 202         /* Prepare for the next iteration */
 203         offset += ret;
 204         total += ret;
 205         bytes -= ret;
 206     }
 207
 208     return total;
 209 }
 210
 211
 212 void iov_hexdump(const struct iovec *iov, const unsigned int iov_cnt,
 213                  FILE *fp, const char *prefix, size_t limit)
 214 {
 215     int v;
 216     size_t size = 0;
 217     char *buf;
 218
 219     for (v = 0; v < iov_cnt; v++) {
 220         size += iov[v].iov_len;
 221     }
 222     size = size > limit ? limit : size;
 223     buf = g_malloc(size);
 224     iov_to_buf(iov, iov_cnt, 0, buf, size);
 225     qemu_hexdump(buf, fp, prefix, size);
 226     g_free(buf);
 227 }
 228
 229 unsigned iov_copy(struct iovec *dst_iov, unsigned int dst_iov_cnt,
 230                  const struct iovec *iov, unsigned int iov_cnt,
 231                  size_t offset, size_t bytes)
 232 {
 233     size_t len;
 234     unsigned int i, j;
 235     for (i = 0, j = 0; i < iov_cnt && j < dst_iov_cnt && bytes; i++) {
 236         if (offset >= iov[i].iov_len) {
 237             offset -= iov[i].iov_len;
 238             continue;
 239         }
 240         len = MIN(bytes, iov[i].iov_len - offset);
 241
 242         dst_iov[j].iov_base = iov[i].iov_base + offset;
 243         dst_iov[j].iov_len = len;
 244         j++;
 245         bytes -= len;
 246         offset = 0;
 247     }
 248     assert(offset == 0);
 249     return j;
 250 }
 251
 252 /* io vectors */
 253
 254 void qemu_iovec_init(QEMUIOVector *qiov, int alloc_hint)
 255 {
 256     qiov->iov = g_malloc(alloc_hint * sizeof(struct iovec));
 257     qiov->niov = 0;
 258     qiov->nalloc = alloc_hint;
 259     qiov->size = 0;
 260 }
 261
 262 void qemu_iovec_init_external(QEMUIOVector *qiov, struct iovec *iov, int niov)
 263 {
 264     int i;
 265
 266     qiov->iov = iov;
 267     qiov->niov = niov;
 268     qiov->nalloc = -1;
 269     qiov->size = 0;
 270     for (i = 0; i < niov; i++)
 271         qiov->size += iov[i].iov_len;
 272 }
 273
 274 void qemu_iovec_add(QEMUIOVector *qiov, void *base, size_t len)
 275 {
 276     assert(qiov->nalloc != -1);
 277
 278     if (qiov->niov == qiov->nalloc) {
 279         qiov->nalloc = 2 * qiov->nalloc + 1;
 280         qiov->iov = g_realloc(qiov->iov, qiov->nalloc * sizeof(struct iovec));
 281     }
 282     qiov->iov[qiov->niov].iov_base = base;
 283     qiov->iov[qiov->niov].iov_len = len;
 284     qiov->size += len;
 285     ++qiov->niov;
 286 }
 287
 288 /*
 289  * Concatenates (partial) iovecs from src_iov to the end of dst.
 290  * It starts copying after skipping `soffset' bytes at the
 291  * beginning of src and adds individual vectors from src to
 292  * dst copies up to `sbytes' bytes total, or up to the end
 293  * of src_iov if it comes first.  This way, it is okay to specify
 294  * very large value for `sbytes' to indicate "up to the end
 295  * of src".
 296  * Only vector pointers are processed, not the actual data buffers.
 297  */
 298 size_t qemu_iovec_concat_iov(QEMUIOVector *dst,
 299                              struct iovec *src_iov, unsigned int src_cnt,
 300                              size_t soffset, size_t sbytes)
 301 {
 302     int i;
 303     size_t done;
 304
 305     if (!sbytes) {
 306         return 0;
 307     }
 308     assert(dst->nalloc != -1);
 309     for (i = 0, done = 0; done < sbytes && i < src_cnt; i++) {
 310         if (soffset < src_iov[i].iov_len) {
 311             size_t len = MIN(src_iov[i].iov_len - soffset, sbytes - done);
 312             qemu_iovec_add(dst, src_iov[i].iov_base + soffset, len);
 313             done += len;
 314             soffset = 0;
 315         } else {
 316             soffset -= src_iov[i].iov_len;
 317         }
 318     }
 319     assert(soffset == 0); /* offset beyond end of src */
 320
 321     return done;
 322 }
 323
 324 /*
 325  * Concatenates (partial) iovecs from src to the end of dst.
 326  * It starts copying after skipping `soffset' bytes at the
 327  * beginning of src and adds individual vectors from src to
 328  * dst copies up to `sbytes' bytes total, or up to the end
 329  * of src if it comes first.  This way, it is okay to specify
 330  * very large value for `sbytes' to indicate "up to the end
 331  * of src".
 332  * Only vector pointers are processed, not the actual data buffers.
 333  */
 334 void qemu_iovec_concat(QEMUIOVector *dst,
 335                        QEMUIOVector *src, size_t soffset, size_t sbytes)
 336 {
 337     qemu_iovec_concat_iov(dst, src->iov, src->niov, soffset, sbytes);
 338 }
 339
 340 /*
 341  * Check if the contents of the iovecs are all zero
 342  */
 343 bool qemu_iovec_is_zero(QEMUIOVector *qiov)
 344 {
 345     int i;
 346     for (i = 0; i < qiov->niov; i++) {
 347         size_t offs = QEMU_ALIGN_DOWN(qiov->iov[i].iov_len, 4 * sizeof(long));
 348         uint8_t *ptr = qiov->iov[i].iov_base;
 349         if (offs && !buffer_is_zero(qiov->iov[i].iov_base, offs)) {
 350             return false;
 351         }
 352         for (; offs < qiov->iov[i].iov_len; offs++) {
 353             if (ptr[offs]) {
 354                 return false;
 355             }
 356         }
 357     }
 358     return true;
 359 }
 360
 361 void qemu_iovec_destroy(QEMUIOVector *qiov)
 362 {
 363     assert(qiov->nalloc != -1);
 364
 365     qemu_iovec_reset(qiov);
 366     g_free(qiov->iov);
 367     qiov->nalloc = 0;
 368     qiov->iov = NULL;
 369 }
 370
 371 void qemu_iovec_reset(QEMUIOVector *qiov)
 372 {
 373     assert(qiov->nalloc != -1);
 374
 375     qiov->niov = 0;
 376     qiov->size = 0;
 377 }
 378
 379 size_t qemu_iovec_to_buf(QEMUIOVector *qiov, size_t offset,
 380                          void *buf, size_t bytes)
 381 {
 382     return iov_to_buf(qiov->iov, qiov->niov, offset, buf, bytes);
 383 }
 384
 385 size_t qemu_iovec_from_buf(QEMUIOVector *qiov, size_t offset,
 386                            const void *buf, size_t bytes)
 387 {
 388     return iov_from_buf(qiov->iov, qiov->niov, offset, buf, bytes);
 389 }
 390
 391 size_t qemu_iovec_memset(QEMUIOVector *qiov, size_t offset,
 392                          int fillc, size_t bytes)
 393 {
 394     return iov_memset(qiov->iov, qiov->niov, offset, fillc, bytes);
 395 }
 396
 397 /**
 398  * Check that I/O vector contents are identical
 399  *
 400  * The IO vectors must have the same structure (same length of all parts).
 401  * A typical usage is to compare vectors created with qemu_iovec_clone().
 402  *
 403  * @a:          I/O vector
 404  * @b:          I/O vector
 405  * @ret:        Offset to first mismatching byte or -1 if match
 406  */
 407 ssize_t qemu_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
 408 {
 409     int i;
 410     ssize_t offset = 0;
 411
 412     assert(a->niov == b->niov);
 413     for (i = 0; i < a->niov; i++) {
 414         size_t len = 0;
 415         uint8_t *p = (uint8_t *)a->iov[i].iov_base;
 416         uint8_t *q = (uint8_t *)b->iov[i].iov_base;
 417
 418         assert(a->iov[i].iov_len == b->iov[i].iov_len);
 419         while (len < a->iov[i].iov_len && *p++ == *q++) {
 420             len++;
 421         }
 422
 423         offset += len;
 424
 425         if (len != a->iov[i].iov_len) {
 426             return offset;
 427         }
 428     }
 429     return -1;
 430 }
 431
 432 typedef struct {
 433     int src_index;
 434     struct iovec *src_iov;
 435     void *dest_base;
 436 } IOVectorSortElem;
 437
 438 static int sortelem_cmp_src_base(const void *a, const void *b)
 439 {
 440     const IOVectorSortElem *elem_a = a;
 441     const IOVectorSortElem *elem_b = b;
 442
 443     /* Don't overflow */
 444     if (elem_a->src_iov->iov_base < elem_b->src_iov->iov_base) {
 445         return -1;
 446     } else if (elem_a->src_iov->iov_base > elem_b->src_iov->iov_base) {
 447         return 1;
 448     } else {
 449         return 0;
 450     }
 451 }
 452
 453 static int sortelem_cmp_src_index(const void *a, const void *b)
 454 {
 455     const IOVectorSortElem *elem_a = a;
 456     const IOVectorSortElem *elem_b = b;
 457
 458     return elem_a->src_index - elem_b->src_index;
 459 }
 460
 461 /**
 462  * Copy contents of I/O vector
 463  *
 464  * The relative relationships of overlapping iovecs are preserved.  This is
 465  * necessary to ensure identical semantics in the cloned I/O vector.
 466  */
 467 void qemu_iovec_clone(QEMUIOVector *dest, const QEMUIOVector *src, void *buf)
 468 {
 469     IOVectorSortElem sortelems[src->niov];
 470     void *last_end;
 471     int i;
 472
 473     /* Sort by source iovecs by base address */
 474     for (i = 0; i < src->niov; i++) {
 475         sortelems[i].src_index = i;
 476         sortelems[i].src_iov = &src->iov[i];
 477     }
 478     qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_base);
 479
 480     /* Allocate buffer space taking into account overlapping iovecs */
 481     last_end = NULL;
 482     for (i = 0; i < src->niov; i++) {
 483         struct iovec *cur = sortelems[i].src_iov;
 484         ptrdiff_t rewind = 0;
 485
 486         /* Detect overlap */
 487         if (last_end && last_end > cur->iov_base) {
 488             rewind = last_end - cur->iov_base;
 489         }
 490
 491         sortelems[i].dest_base = buf - rewind;
 492         buf += cur->iov_len - MIN(rewind, cur->iov_len);
 493         last_end = MAX(cur->iov_base + cur->iov_len, last_end);
 494     }
 495
 496     /* Sort by source iovec index and build destination iovec */
 497     qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_index);
 498     for (i = 0; i < src->niov; i++) {
 499         qemu_iovec_add(dest, sortelems[i].dest_base, src->iov[i].iov_len);
 500     }
 501 }
 502
 503 size_t iov_discard_front(struct iovec **iov, unsigned int *iov_cnt,
 504                          size_t bytes)
 505 {
 506     size_t total = 0;
 507     struct iovec *cur;
 508
 509     for (cur = *iov; *iov_cnt > 0; cur++) {
 510         if (cur->iov_len > bytes) {
 511             cur->iov_base += bytes;
 512             cur->iov_len -= bytes;
 513             total += bytes;
 514             break;
 515         }
 516
 517         bytes -= cur->iov_len;
 518         total += cur->iov_len;
 519         *iov_cnt -= 1;
 520     }
 521
 522     *iov = cur;
 523     return total;
 524 }
 525
 526 size_t iov_discard_back(struct iovec *iov, unsigned int *iov_cnt,
 527                         size_t bytes)
 528 {
 529     size_t total = 0;
 530     struct iovec *cur;
 531
 532     if (*iov_cnt == 0) {
 533         return 0;
 534     }
 535
 536     cur = iov + (*iov_cnt - 1);
 537
 538     while (*iov_cnt > 0) {
 539         if (cur->iov_len > bytes) {
 540             cur->iov_len -= bytes;
 541             total += bytes;
 542             break;
 543         }
 544
 545         bytes -= cur->iov_len;
 546         total += cur->iov_len;
 547         cur--;
 548         *iov_cnt -= 1;
 549     }
 550
 551     return total;
 552 }
 553
 554 void qemu_iovec_discard_back(QEMUIOVector *qiov, size_t bytes)
 555 {
 556     size_t total;
 557     unsigned int niov = qiov->niov;
 558
 559     assert(qiov->size >= bytes);
 560     total = iov_discard_back(qiov->iov, &niov, bytes);
 561     assert(total == bytes);
 562
 563     qiov->niov = niov;
 564     qiov->size -= bytes;
 565 }