xref: /openbmc/qemu/util/iov.c (revision 8692aa29)
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; i < iov_cnt && j < dst_iov_cnt && bytes; i++) {
251         if (offset >= iov[i].iov_len) {
252             offset -= iov[i].iov_len;
253             continue;
254         }
255         len = MIN(bytes, iov[i].iov_len - offset);
256 
257         dst_iov[j].iov_base = iov[i].iov_base + offset;
258         dst_iov[j].iov_len = len;
259         j++;
260         bytes -= len;
261         offset = 0;
262     }
263     assert(offset == 0);
264     return j;
265 }
266 
267 /* io vectors */
268 
269 void qemu_iovec_init(QEMUIOVector *qiov, int alloc_hint)
270 {
271     qiov->iov = g_new(struct iovec, alloc_hint);
272     qiov->niov = 0;
273     qiov->nalloc = alloc_hint;
274     qiov->size = 0;
275 }
276 
277 void qemu_iovec_init_external(QEMUIOVector *qiov, struct iovec *iov, int niov)
278 {
279     int i;
280 
281     qiov->iov = iov;
282     qiov->niov = niov;
283     qiov->nalloc = -1;
284     qiov->size = 0;
285     for (i = 0; i < niov; i++)
286         qiov->size += iov[i].iov_len;
287 }
288 
289 void qemu_iovec_add(QEMUIOVector *qiov, void *base, size_t len)
290 {
291     assert(qiov->nalloc != -1);
292 
293     if (qiov->niov == qiov->nalloc) {
294         qiov->nalloc = 2 * qiov->nalloc + 1;
295         qiov->iov = g_renew(struct iovec, qiov->iov, qiov->nalloc);
296     }
297     qiov->iov[qiov->niov].iov_base = base;
298     qiov->iov[qiov->niov].iov_len = len;
299     qiov->size += len;
300     ++qiov->niov;
301 }
302 
303 /*
304  * Concatenates (partial) iovecs from src_iov to the end of dst.
305  * It starts copying after skipping `soffset' bytes at the
306  * beginning of src and adds individual vectors from src to
307  * dst copies up to `sbytes' bytes total, or up to the end
308  * of src_iov if it comes first.  This way, it is okay to specify
309  * very large value for `sbytes' to indicate "up to the end
310  * of src".
311  * Only vector pointers are processed, not the actual data buffers.
312  */
313 size_t qemu_iovec_concat_iov(QEMUIOVector *dst,
314                              struct iovec *src_iov, unsigned int src_cnt,
315                              size_t soffset, size_t sbytes)
316 {
317     int i;
318     size_t done;
319 
320     if (!sbytes) {
321         return 0;
322     }
323     assert(dst->nalloc != -1);
324     for (i = 0, done = 0; done < sbytes && i < src_cnt; i++) {
325         if (soffset < src_iov[i].iov_len) {
326             size_t len = MIN(src_iov[i].iov_len - soffset, sbytes - done);
327             qemu_iovec_add(dst, src_iov[i].iov_base + soffset, len);
328             done += len;
329             soffset = 0;
330         } else {
331             soffset -= src_iov[i].iov_len;
332         }
333     }
334     assert(soffset == 0); /* offset beyond end of src */
335 
336     return done;
337 }
338 
339 /*
340  * Concatenates (partial) iovecs from src to the end of dst.
341  * It starts copying after skipping `soffset' bytes at the
342  * beginning of src and adds individual vectors from src to
343  * dst copies up to `sbytes' bytes total, or up to the end
344  * of src if it comes first.  This way, it is okay to specify
345  * very large value for `sbytes' to indicate "up to the end
346  * of src".
347  * Only vector pointers are processed, not the actual data buffers.
348  */
349 void qemu_iovec_concat(QEMUIOVector *dst,
350                        QEMUIOVector *src, size_t soffset, size_t sbytes)
351 {
352     qemu_iovec_concat_iov(dst, src->iov, src->niov, soffset, sbytes);
353 }
354 
355 /*
356  * Check if the contents of the iovecs are all zero
357  */
358 bool qemu_iovec_is_zero(QEMUIOVector *qiov)
359 {
360     int i;
361     for (i = 0; i < qiov->niov; i++) {
362         size_t offs = QEMU_ALIGN_DOWN(qiov->iov[i].iov_len, 4 * sizeof(long));
363         uint8_t *ptr = qiov->iov[i].iov_base;
364         if (offs && !buffer_is_zero(qiov->iov[i].iov_base, offs)) {
365             return false;
366         }
367         for (; offs < qiov->iov[i].iov_len; offs++) {
368             if (ptr[offs]) {
369                 return false;
370             }
371         }
372     }
373     return true;
374 }
375 
376 void qemu_iovec_destroy(QEMUIOVector *qiov)
377 {
378     assert(qiov->nalloc != -1);
379 
380     qemu_iovec_reset(qiov);
381     g_free(qiov->iov);
382     qiov->nalloc = 0;
383     qiov->iov = NULL;
384 }
385 
386 void qemu_iovec_reset(QEMUIOVector *qiov)
387 {
388     assert(qiov->nalloc != -1);
389 
390     qiov->niov = 0;
391     qiov->size = 0;
392 }
393 
394 size_t qemu_iovec_to_buf(QEMUIOVector *qiov, size_t offset,
395                          void *buf, size_t bytes)
396 {
397     return iov_to_buf(qiov->iov, qiov->niov, offset, buf, bytes);
398 }
399 
400 size_t qemu_iovec_from_buf(QEMUIOVector *qiov, size_t offset,
401                            const void *buf, size_t bytes)
402 {
403     return iov_from_buf(qiov->iov, qiov->niov, offset, buf, bytes);
404 }
405 
406 size_t qemu_iovec_memset(QEMUIOVector *qiov, size_t offset,
407                          int fillc, size_t bytes)
408 {
409     return iov_memset(qiov->iov, qiov->niov, offset, fillc, bytes);
410 }
411 
412 /**
413  * Check that I/O vector contents are identical
414  *
415  * The IO vectors must have the same structure (same length of all parts).
416  * A typical usage is to compare vectors created with qemu_iovec_clone().
417  *
418  * @a:          I/O vector
419  * @b:          I/O vector
420  * @ret:        Offset to first mismatching byte or -1 if match
421  */
422 ssize_t qemu_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
423 {
424     int i;
425     ssize_t offset = 0;
426 
427     assert(a->niov == b->niov);
428     for (i = 0; i < a->niov; i++) {
429         size_t len = 0;
430         uint8_t *p = (uint8_t *)a->iov[i].iov_base;
431         uint8_t *q = (uint8_t *)b->iov[i].iov_base;
432 
433         assert(a->iov[i].iov_len == b->iov[i].iov_len);
434         while (len < a->iov[i].iov_len && *p++ == *q++) {
435             len++;
436         }
437 
438         offset += len;
439 
440         if (len != a->iov[i].iov_len) {
441             return offset;
442         }
443     }
444     return -1;
445 }
446 
447 typedef struct {
448     int src_index;
449     struct iovec *src_iov;
450     void *dest_base;
451 } IOVectorSortElem;
452 
453 static int sortelem_cmp_src_base(const void *a, const void *b)
454 {
455     const IOVectorSortElem *elem_a = a;
456     const IOVectorSortElem *elem_b = b;
457 
458     /* Don't overflow */
459     if (elem_a->src_iov->iov_base < elem_b->src_iov->iov_base) {
460         return -1;
461     } else if (elem_a->src_iov->iov_base > elem_b->src_iov->iov_base) {
462         return 1;
463     } else {
464         return 0;
465     }
466 }
467 
468 static int sortelem_cmp_src_index(const void *a, const void *b)
469 {
470     const IOVectorSortElem *elem_a = a;
471     const IOVectorSortElem *elem_b = b;
472 
473     return elem_a->src_index - elem_b->src_index;
474 }
475 
476 /**
477  * Copy contents of I/O vector
478  *
479  * The relative relationships of overlapping iovecs are preserved.  This is
480  * necessary to ensure identical semantics in the cloned I/O vector.
481  */
482 void qemu_iovec_clone(QEMUIOVector *dest, const QEMUIOVector *src, void *buf)
483 {
484     IOVectorSortElem sortelems[src->niov];
485     void *last_end;
486     int i;
487 
488     /* Sort by source iovecs by base address */
489     for (i = 0; i < src->niov; i++) {
490         sortelems[i].src_index = i;
491         sortelems[i].src_iov = &src->iov[i];
492     }
493     qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_base);
494 
495     /* Allocate buffer space taking into account overlapping iovecs */
496     last_end = NULL;
497     for (i = 0; i < src->niov; i++) {
498         struct iovec *cur = sortelems[i].src_iov;
499         ptrdiff_t rewind = 0;
500 
501         /* Detect overlap */
502         if (last_end && last_end > cur->iov_base) {
503             rewind = last_end - cur->iov_base;
504         }
505 
506         sortelems[i].dest_base = buf - rewind;
507         buf += cur->iov_len - MIN(rewind, cur->iov_len);
508         last_end = MAX(cur->iov_base + cur->iov_len, last_end);
509     }
510 
511     /* Sort by source iovec index and build destination iovec */
512     qsort(sortelems, src->niov, sizeof(sortelems[0]), sortelem_cmp_src_index);
513     for (i = 0; i < src->niov; i++) {
514         qemu_iovec_add(dest, sortelems[i].dest_base, src->iov[i].iov_len);
515     }
516 }
517 
518 size_t iov_discard_front(struct iovec **iov, unsigned int *iov_cnt,
519                          size_t bytes)
520 {
521     size_t total = 0;
522     struct iovec *cur;
523 
524     for (cur = *iov; *iov_cnt > 0; cur++) {
525         if (cur->iov_len > bytes) {
526             cur->iov_base += bytes;
527             cur->iov_len -= bytes;
528             total += bytes;
529             break;
530         }
531 
532         bytes -= cur->iov_len;
533         total += cur->iov_len;
534         *iov_cnt -= 1;
535     }
536 
537     *iov = cur;
538     return total;
539 }
540 
541 size_t iov_discard_back(struct iovec *iov, unsigned int *iov_cnt,
542                         size_t bytes)
543 {
544     size_t total = 0;
545     struct iovec *cur;
546 
547     if (*iov_cnt == 0) {
548         return 0;
549     }
550 
551     cur = iov + (*iov_cnt - 1);
552 
553     while (*iov_cnt > 0) {
554         if (cur->iov_len > bytes) {
555             cur->iov_len -= bytes;
556             total += bytes;
557             break;
558         }
559 
560         bytes -= cur->iov_len;
561         total += cur->iov_len;
562         cur--;
563         *iov_cnt -= 1;
564     }
565 
566     return total;
567 }
568 
569 void qemu_iovec_discard_back(QEMUIOVector *qiov, size_t bytes)
570 {
571     size_t total;
572     unsigned int niov = qiov->niov;
573 
574     assert(qiov->size >= bytes);
575     total = iov_discard_back(qiov->iov, &niov, bytes);
576     assert(total == bytes);
577 
578     qiov->niov = niov;
579     qiov->size -= bytes;
580 }
581