Merge branch 'mh/packed-ref-store-prep'
[git] / apply.c
1 /*
2  * apply.c
3  *
4  * Copyright (C) Linus Torvalds, 2005
5  *
6  * This applies patches on top of some (arbitrary) version of the SCM.
7  *
8  */
9
10 #include "cache.h"
11 #include "config.h"
12 #include "blob.h"
13 #include "delta.h"
14 #include "diff.h"
15 #include "dir.h"
16 #include "xdiff-interface.h"
17 #include "ll-merge.h"
18 #include "lockfile.h"
19 #include "parse-options.h"
20 #include "quote.h"
21 #include "rerere.h"
22 #include "apply.h"
23
24 static void git_apply_config(void)
25 {
26         git_config_get_string_const("apply.whitespace", &apply_default_whitespace);
27         git_config_get_string_const("apply.ignorewhitespace", &apply_default_ignorewhitespace);
28         git_config(git_default_config, NULL);
29 }
30
31 static int parse_whitespace_option(struct apply_state *state, const char *option)
32 {
33         if (!option) {
34                 state->ws_error_action = warn_on_ws_error;
35                 return 0;
36         }
37         if (!strcmp(option, "warn")) {
38                 state->ws_error_action = warn_on_ws_error;
39                 return 0;
40         }
41         if (!strcmp(option, "nowarn")) {
42                 state->ws_error_action = nowarn_ws_error;
43                 return 0;
44         }
45         if (!strcmp(option, "error")) {
46                 state->ws_error_action = die_on_ws_error;
47                 return 0;
48         }
49         if (!strcmp(option, "error-all")) {
50                 state->ws_error_action = die_on_ws_error;
51                 state->squelch_whitespace_errors = 0;
52                 return 0;
53         }
54         if (!strcmp(option, "strip") || !strcmp(option, "fix")) {
55                 state->ws_error_action = correct_ws_error;
56                 return 0;
57         }
58         return error(_("unrecognized whitespace option '%s'"), option);
59 }
60
61 static int parse_ignorewhitespace_option(struct apply_state *state,
62                                                  const char *option)
63 {
64         if (!option || !strcmp(option, "no") ||
65             !strcmp(option, "false") || !strcmp(option, "never") ||
66             !strcmp(option, "none")) {
67                 state->ws_ignore_action = ignore_ws_none;
68                 return 0;
69         }
70         if (!strcmp(option, "change")) {
71                 state->ws_ignore_action = ignore_ws_change;
72                 return 0;
73         }
74         return error(_("unrecognized whitespace ignore option '%s'"), option);
75 }
76
77 int init_apply_state(struct apply_state *state,
78                      const char *prefix,
79                      struct lock_file *lock_file)
80 {
81         memset(state, 0, sizeof(*state));
82         state->prefix = prefix;
83         state->prefix_length = state->prefix ? strlen(state->prefix) : 0;
84         state->lock_file = lock_file;
85         state->newfd = -1;
86         state->apply = 1;
87         state->line_termination = '\n';
88         state->p_value = 1;
89         state->p_context = UINT_MAX;
90         state->squelch_whitespace_errors = 5;
91         state->ws_error_action = warn_on_ws_error;
92         state->ws_ignore_action = ignore_ws_none;
93         state->linenr = 1;
94         string_list_init(&state->fn_table, 0);
95         string_list_init(&state->limit_by_name, 0);
96         string_list_init(&state->symlink_changes, 0);
97         strbuf_init(&state->root, 0);
98
99         git_apply_config();
100         if (apply_default_whitespace && parse_whitespace_option(state, apply_default_whitespace))
101                 return -1;
102         if (apply_default_ignorewhitespace && parse_ignorewhitespace_option(state, apply_default_ignorewhitespace))
103                 return -1;
104         return 0;
105 }
106
107 void clear_apply_state(struct apply_state *state)
108 {
109         string_list_clear(&state->limit_by_name, 0);
110         string_list_clear(&state->symlink_changes, 0);
111         strbuf_release(&state->root);
112
113         /* &state->fn_table is cleared at the end of apply_patch() */
114 }
115
116 static void mute_routine(const char *msg, va_list params)
117 {
118         /* do nothing */
119 }
120
121 int check_apply_state(struct apply_state *state, int force_apply)
122 {
123         int is_not_gitdir = !startup_info->have_repository;
124
125         if (state->apply_with_reject && state->threeway)
126                 return error(_("--reject and --3way cannot be used together."));
127         if (state->cached && state->threeway)
128                 return error(_("--cached and --3way cannot be used together."));
129         if (state->threeway) {
130                 if (is_not_gitdir)
131                         return error(_("--3way outside a repository"));
132                 state->check_index = 1;
133         }
134         if (state->apply_with_reject) {
135                 state->apply = 1;
136                 if (state->apply_verbosity == verbosity_normal)
137                         state->apply_verbosity = verbosity_verbose;
138         }
139         if (!force_apply && (state->diffstat || state->numstat || state->summary || state->check || state->fake_ancestor))
140                 state->apply = 0;
141         if (state->check_index && is_not_gitdir)
142                 return error(_("--index outside a repository"));
143         if (state->cached) {
144                 if (is_not_gitdir)
145                         return error(_("--cached outside a repository"));
146                 state->check_index = 1;
147         }
148         if (state->check_index)
149                 state->unsafe_paths = 0;
150         if (!state->lock_file)
151                 return error("BUG: state->lock_file should not be NULL");
152
153         if (state->apply_verbosity <= verbosity_silent) {
154                 state->saved_error_routine = get_error_routine();
155                 state->saved_warn_routine = get_warn_routine();
156                 set_error_routine(mute_routine);
157                 set_warn_routine(mute_routine);
158         }
159
160         return 0;
161 }
162
163 static void set_default_whitespace_mode(struct apply_state *state)
164 {
165         if (!state->whitespace_option && !apply_default_whitespace)
166                 state->ws_error_action = (state->apply ? warn_on_ws_error : nowarn_ws_error);
167 }
168
169 /*
170  * This represents one "hunk" from a patch, starting with
171  * "@@ -oldpos,oldlines +newpos,newlines @@" marker.  The
172  * patch text is pointed at by patch, and its byte length
173  * is stored in size.  leading and trailing are the number
174  * of context lines.
175  */
176 struct fragment {
177         unsigned long leading, trailing;
178         unsigned long oldpos, oldlines;
179         unsigned long newpos, newlines;
180         /*
181          * 'patch' is usually borrowed from buf in apply_patch(),
182          * but some codepaths store an allocated buffer.
183          */
184         const char *patch;
185         unsigned free_patch:1,
186                 rejected:1;
187         int size;
188         int linenr;
189         struct fragment *next;
190 };
191
192 /*
193  * When dealing with a binary patch, we reuse "leading" field
194  * to store the type of the binary hunk, either deflated "delta"
195  * or deflated "literal".
196  */
197 #define binary_patch_method leading
198 #define BINARY_DELTA_DEFLATED   1
199 #define BINARY_LITERAL_DEFLATED 2
200
201 /*
202  * This represents a "patch" to a file, both metainfo changes
203  * such as creation/deletion, filemode and content changes represented
204  * as a series of fragments.
205  */
206 struct patch {
207         char *new_name, *old_name, *def_name;
208         unsigned int old_mode, new_mode;
209         int is_new, is_delete;  /* -1 = unknown, 0 = false, 1 = true */
210         int rejected;
211         unsigned ws_rule;
212         int lines_added, lines_deleted;
213         int score;
214         unsigned int is_toplevel_relative:1;
215         unsigned int inaccurate_eof:1;
216         unsigned int is_binary:1;
217         unsigned int is_copy:1;
218         unsigned int is_rename:1;
219         unsigned int recount:1;
220         unsigned int conflicted_threeway:1;
221         unsigned int direct_to_threeway:1;
222         struct fragment *fragments;
223         char *result;
224         size_t resultsize;
225         char old_sha1_prefix[41];
226         char new_sha1_prefix[41];
227         struct patch *next;
228
229         /* three-way fallback result */
230         struct object_id threeway_stage[3];
231 };
232
233 static void free_fragment_list(struct fragment *list)
234 {
235         while (list) {
236                 struct fragment *next = list->next;
237                 if (list->free_patch)
238                         free((char *)list->patch);
239                 free(list);
240                 list = next;
241         }
242 }
243
244 static void free_patch(struct patch *patch)
245 {
246         free_fragment_list(patch->fragments);
247         free(patch->def_name);
248         free(patch->old_name);
249         free(patch->new_name);
250         free(patch->result);
251         free(patch);
252 }
253
254 static void free_patch_list(struct patch *list)
255 {
256         while (list) {
257                 struct patch *next = list->next;
258                 free_patch(list);
259                 list = next;
260         }
261 }
262
263 /*
264  * A line in a file, len-bytes long (includes the terminating LF,
265  * except for an incomplete line at the end if the file ends with
266  * one), and its contents hashes to 'hash'.
267  */
268 struct line {
269         size_t len;
270         unsigned hash : 24;
271         unsigned flag : 8;
272 #define LINE_COMMON     1
273 #define LINE_PATCHED    2
274 };
275
276 /*
277  * This represents a "file", which is an array of "lines".
278  */
279 struct image {
280         char *buf;
281         size_t len;
282         size_t nr;
283         size_t alloc;
284         struct line *line_allocated;
285         struct line *line;
286 };
287
288 static uint32_t hash_line(const char *cp, size_t len)
289 {
290         size_t i;
291         uint32_t h;
292         for (i = 0, h = 0; i < len; i++) {
293                 if (!isspace(cp[i])) {
294                         h = h * 3 + (cp[i] & 0xff);
295                 }
296         }
297         return h;
298 }
299
300 /*
301  * Compare lines s1 of length n1 and s2 of length n2, ignoring
302  * whitespace difference. Returns 1 if they match, 0 otherwise
303  */
304 static int fuzzy_matchlines(const char *s1, size_t n1,
305                             const char *s2, size_t n2)
306 {
307         const char *last1 = s1 + n1 - 1;
308         const char *last2 = s2 + n2 - 1;
309         int result = 0;
310
311         /* ignore line endings */
312         while ((*last1 == '\r') || (*last1 == '\n'))
313                 last1--;
314         while ((*last2 == '\r') || (*last2 == '\n'))
315                 last2--;
316
317         /* skip leading whitespaces, if both begin with whitespace */
318         if (s1 <= last1 && s2 <= last2 && isspace(*s1) && isspace(*s2)) {
319                 while (isspace(*s1) && (s1 <= last1))
320                         s1++;
321                 while (isspace(*s2) && (s2 <= last2))
322                         s2++;
323         }
324         /* early return if both lines are empty */
325         if ((s1 > last1) && (s2 > last2))
326                 return 1;
327         while (!result) {
328                 result = *s1++ - *s2++;
329                 /*
330                  * Skip whitespace inside. We check for whitespace on
331                  * both buffers because we don't want "a b" to match
332                  * "ab"
333                  */
334                 if (isspace(*s1) && isspace(*s2)) {
335                         while (isspace(*s1) && s1 <= last1)
336                                 s1++;
337                         while (isspace(*s2) && s2 <= last2)
338                                 s2++;
339                 }
340                 /*
341                  * If we reached the end on one side only,
342                  * lines don't match
343                  */
344                 if (
345                     ((s2 > last2) && (s1 <= last1)) ||
346                     ((s1 > last1) && (s2 <= last2)))
347                         return 0;
348                 if ((s1 > last1) && (s2 > last2))
349                         break;
350         }
351
352         return !result;
353 }
354
355 static void add_line_info(struct image *img, const char *bol, size_t len, unsigned flag)
356 {
357         ALLOC_GROW(img->line_allocated, img->nr + 1, img->alloc);
358         img->line_allocated[img->nr].len = len;
359         img->line_allocated[img->nr].hash = hash_line(bol, len);
360         img->line_allocated[img->nr].flag = flag;
361         img->nr++;
362 }
363
364 /*
365  * "buf" has the file contents to be patched (read from various sources).
366  * attach it to "image" and add line-based index to it.
367  * "image" now owns the "buf".
368  */
369 static void prepare_image(struct image *image, char *buf, size_t len,
370                           int prepare_linetable)
371 {
372         const char *cp, *ep;
373
374         memset(image, 0, sizeof(*image));
375         image->buf = buf;
376         image->len = len;
377
378         if (!prepare_linetable)
379                 return;
380
381         ep = image->buf + image->len;
382         cp = image->buf;
383         while (cp < ep) {
384                 const char *next;
385                 for (next = cp; next < ep && *next != '\n'; next++)
386                         ;
387                 if (next < ep)
388                         next++;
389                 add_line_info(image, cp, next - cp, 0);
390                 cp = next;
391         }
392         image->line = image->line_allocated;
393 }
394
395 static void clear_image(struct image *image)
396 {
397         free(image->buf);
398         free(image->line_allocated);
399         memset(image, 0, sizeof(*image));
400 }
401
402 /* fmt must contain _one_ %s and no other substitution */
403 static void say_patch_name(FILE *output, const char *fmt, struct patch *patch)
404 {
405         struct strbuf sb = STRBUF_INIT;
406
407         if (patch->old_name && patch->new_name &&
408             strcmp(patch->old_name, patch->new_name)) {
409                 quote_c_style(patch->old_name, &sb, NULL, 0);
410                 strbuf_addstr(&sb, " => ");
411                 quote_c_style(patch->new_name, &sb, NULL, 0);
412         } else {
413                 const char *n = patch->new_name;
414                 if (!n)
415                         n = patch->old_name;
416                 quote_c_style(n, &sb, NULL, 0);
417         }
418         fprintf(output, fmt, sb.buf);
419         fputc('\n', output);
420         strbuf_release(&sb);
421 }
422
423 #define SLOP (16)
424
425 static int read_patch_file(struct strbuf *sb, int fd)
426 {
427         if (strbuf_read(sb, fd, 0) < 0)
428                 return error_errno("git apply: failed to read");
429
430         /*
431          * Make sure that we have some slop in the buffer
432          * so that we can do speculative "memcmp" etc, and
433          * see to it that it is NUL-filled.
434          */
435         strbuf_grow(sb, SLOP);
436         memset(sb->buf + sb->len, 0, SLOP);
437         return 0;
438 }
439
440 static unsigned long linelen(const char *buffer, unsigned long size)
441 {
442         unsigned long len = 0;
443         while (size--) {
444                 len++;
445                 if (*buffer++ == '\n')
446                         break;
447         }
448         return len;
449 }
450
451 static int is_dev_null(const char *str)
452 {
453         return skip_prefix(str, "/dev/null", &str) && isspace(*str);
454 }
455
456 #define TERM_SPACE      1
457 #define TERM_TAB        2
458
459 static int name_terminate(int c, int terminate)
460 {
461         if (c == ' ' && !(terminate & TERM_SPACE))
462                 return 0;
463         if (c == '\t' && !(terminate & TERM_TAB))
464                 return 0;
465
466         return 1;
467 }
468
469 /* remove double slashes to make --index work with such filenames */
470 static char *squash_slash(char *name)
471 {
472         int i = 0, j = 0;
473
474         if (!name)
475                 return NULL;
476
477         while (name[i]) {
478                 if ((name[j++] = name[i++]) == '/')
479                         while (name[i] == '/')
480                                 i++;
481         }
482         name[j] = '\0';
483         return name;
484 }
485
486 static char *find_name_gnu(struct apply_state *state,
487                            const char *line,
488                            const char *def,
489                            int p_value)
490 {
491         struct strbuf name = STRBUF_INIT;
492         char *cp;
493
494         /*
495          * Proposed "new-style" GNU patch/diff format; see
496          * http://marc.info/?l=git&m=112927316408690&w=2
497          */
498         if (unquote_c_style(&name, line, NULL)) {
499                 strbuf_release(&name);
500                 return NULL;
501         }
502
503         for (cp = name.buf; p_value; p_value--) {
504                 cp = strchr(cp, '/');
505                 if (!cp) {
506                         strbuf_release(&name);
507                         return NULL;
508                 }
509                 cp++;
510         }
511
512         strbuf_remove(&name, 0, cp - name.buf);
513         if (state->root.len)
514                 strbuf_insert(&name, 0, state->root.buf, state->root.len);
515         return squash_slash(strbuf_detach(&name, NULL));
516 }
517
518 static size_t sane_tz_len(const char *line, size_t len)
519 {
520         const char *tz, *p;
521
522         if (len < strlen(" +0500") || line[len-strlen(" +0500")] != ' ')
523                 return 0;
524         tz = line + len - strlen(" +0500");
525
526         if (tz[1] != '+' && tz[1] != '-')
527                 return 0;
528
529         for (p = tz + 2; p != line + len; p++)
530                 if (!isdigit(*p))
531                         return 0;
532
533         return line + len - tz;
534 }
535
536 static size_t tz_with_colon_len(const char *line, size_t len)
537 {
538         const char *tz, *p;
539
540         if (len < strlen(" +08:00") || line[len - strlen(":00")] != ':')
541                 return 0;
542         tz = line + len - strlen(" +08:00");
543
544         if (tz[0] != ' ' || (tz[1] != '+' && tz[1] != '-'))
545                 return 0;
546         p = tz + 2;
547         if (!isdigit(*p++) || !isdigit(*p++) || *p++ != ':' ||
548             !isdigit(*p++) || !isdigit(*p++))
549                 return 0;
550
551         return line + len - tz;
552 }
553
554 static size_t date_len(const char *line, size_t len)
555 {
556         const char *date, *p;
557
558         if (len < strlen("72-02-05") || line[len-strlen("-05")] != '-')
559                 return 0;
560         p = date = line + len - strlen("72-02-05");
561
562         if (!isdigit(*p++) || !isdigit(*p++) || *p++ != '-' ||
563             !isdigit(*p++) || !isdigit(*p++) || *p++ != '-' ||
564             !isdigit(*p++) || !isdigit(*p++))   /* Not a date. */
565                 return 0;
566
567         if (date - line >= strlen("19") &&
568             isdigit(date[-1]) && isdigit(date[-2]))     /* 4-digit year */
569                 date -= strlen("19");
570
571         return line + len - date;
572 }
573
574 static size_t short_time_len(const char *line, size_t len)
575 {
576         const char *time, *p;
577
578         if (len < strlen(" 07:01:32") || line[len-strlen(":32")] != ':')
579                 return 0;
580         p = time = line + len - strlen(" 07:01:32");
581
582         /* Permit 1-digit hours? */
583         if (*p++ != ' ' ||
584             !isdigit(*p++) || !isdigit(*p++) || *p++ != ':' ||
585             !isdigit(*p++) || !isdigit(*p++) || *p++ != ':' ||
586             !isdigit(*p++) || !isdigit(*p++))   /* Not a time. */
587                 return 0;
588
589         return line + len - time;
590 }
591
592 static size_t fractional_time_len(const char *line, size_t len)
593 {
594         const char *p;
595         size_t n;
596
597         /* Expected format: 19:41:17.620000023 */
598         if (!len || !isdigit(line[len - 1]))
599                 return 0;
600         p = line + len - 1;
601
602         /* Fractional seconds. */
603         while (p > line && isdigit(*p))
604                 p--;
605         if (*p != '.')
606                 return 0;
607
608         /* Hours, minutes, and whole seconds. */
609         n = short_time_len(line, p - line);
610         if (!n)
611                 return 0;
612
613         return line + len - p + n;
614 }
615
616 static size_t trailing_spaces_len(const char *line, size_t len)
617 {
618         const char *p;
619
620         /* Expected format: ' ' x (1 or more)  */
621         if (!len || line[len - 1] != ' ')
622                 return 0;
623
624         p = line + len;
625         while (p != line) {
626                 p--;
627                 if (*p != ' ')
628                         return line + len - (p + 1);
629         }
630
631         /* All spaces! */
632         return len;
633 }
634
635 static size_t diff_timestamp_len(const char *line, size_t len)
636 {
637         const char *end = line + len;
638         size_t n;
639
640         /*
641          * Posix: 2010-07-05 19:41:17
642          * GNU: 2010-07-05 19:41:17.620000023 -0500
643          */
644
645         if (!isdigit(end[-1]))
646                 return 0;
647
648         n = sane_tz_len(line, end - line);
649         if (!n)
650                 n = tz_with_colon_len(line, end - line);
651         end -= n;
652
653         n = short_time_len(line, end - line);
654         if (!n)
655                 n = fractional_time_len(line, end - line);
656         end -= n;
657
658         n = date_len(line, end - line);
659         if (!n) /* No date.  Too bad. */
660                 return 0;
661         end -= n;
662
663         if (end == line)        /* No space before date. */
664                 return 0;
665         if (end[-1] == '\t') {  /* Success! */
666                 end--;
667                 return line + len - end;
668         }
669         if (end[-1] != ' ')     /* No space before date. */
670                 return 0;
671
672         /* Whitespace damage. */
673         end -= trailing_spaces_len(line, end - line);
674         return line + len - end;
675 }
676
677 static char *find_name_common(struct apply_state *state,
678                               const char *line,
679                               const char *def,
680                               int p_value,
681                               const char *end,
682                               int terminate)
683 {
684         int len;
685         const char *start = NULL;
686
687         if (p_value == 0)
688                 start = line;
689         while (line != end) {
690                 char c = *line;
691
692                 if (!end && isspace(c)) {
693                         if (c == '\n')
694                                 break;
695                         if (name_terminate(c, terminate))
696                                 break;
697                 }
698                 line++;
699                 if (c == '/' && !--p_value)
700                         start = line;
701         }
702         if (!start)
703                 return squash_slash(xstrdup_or_null(def));
704         len = line - start;
705         if (!len)
706                 return squash_slash(xstrdup_or_null(def));
707
708         /*
709          * Generally we prefer the shorter name, especially
710          * if the other one is just a variation of that with
711          * something else tacked on to the end (ie "file.orig"
712          * or "file~").
713          */
714         if (def) {
715                 int deflen = strlen(def);
716                 if (deflen < len && !strncmp(start, def, deflen))
717                         return squash_slash(xstrdup(def));
718         }
719
720         if (state->root.len) {
721                 char *ret = xstrfmt("%s%.*s", state->root.buf, len, start);
722                 return squash_slash(ret);
723         }
724
725         return squash_slash(xmemdupz(start, len));
726 }
727
728 static char *find_name(struct apply_state *state,
729                        const char *line,
730                        char *def,
731                        int p_value,
732                        int terminate)
733 {
734         if (*line == '"') {
735                 char *name = find_name_gnu(state, line, def, p_value);
736                 if (name)
737                         return name;
738         }
739
740         return find_name_common(state, line, def, p_value, NULL, terminate);
741 }
742
743 static char *find_name_traditional(struct apply_state *state,
744                                    const char *line,
745                                    char *def,
746                                    int p_value)
747 {
748         size_t len;
749         size_t date_len;
750
751         if (*line == '"') {
752                 char *name = find_name_gnu(state, line, def, p_value);
753                 if (name)
754                         return name;
755         }
756
757         len = strchrnul(line, '\n') - line;
758         date_len = diff_timestamp_len(line, len);
759         if (!date_len)
760                 return find_name_common(state, line, def, p_value, NULL, TERM_TAB);
761         len -= date_len;
762
763         return find_name_common(state, line, def, p_value, line + len, 0);
764 }
765
766 /*
767  * Given the string after "--- " or "+++ ", guess the appropriate
768  * p_value for the given patch.
769  */
770 static int guess_p_value(struct apply_state *state, const char *nameline)
771 {
772         char *name, *cp;
773         int val = -1;
774
775         if (is_dev_null(nameline))
776                 return -1;
777         name = find_name_traditional(state, nameline, NULL, 0);
778         if (!name)
779                 return -1;
780         cp = strchr(name, '/');
781         if (!cp)
782                 val = 0;
783         else if (state->prefix) {
784                 /*
785                  * Does it begin with "a/$our-prefix" and such?  Then this is
786                  * very likely to apply to our directory.
787                  */
788                 if (!strncmp(name, state->prefix, state->prefix_length))
789                         val = count_slashes(state->prefix);
790                 else {
791                         cp++;
792                         if (!strncmp(cp, state->prefix, state->prefix_length))
793                                 val = count_slashes(state->prefix) + 1;
794                 }
795         }
796         free(name);
797         return val;
798 }
799
800 /*
801  * Does the ---/+++ line have the POSIX timestamp after the last HT?
802  * GNU diff puts epoch there to signal a creation/deletion event.  Is
803  * this such a timestamp?
804  */
805 static int has_epoch_timestamp(const char *nameline)
806 {
807         /*
808          * We are only interested in epoch timestamp; any non-zero
809          * fraction cannot be one, hence "(\.0+)?" in the regexp below.
810          * For the same reason, the date must be either 1969-12-31 or
811          * 1970-01-01, and the seconds part must be "00".
812          */
813         const char stamp_regexp[] =
814                 "^(1969-12-31|1970-01-01)"
815                 " "
816                 "[0-2][0-9]:[0-5][0-9]:00(\\.0+)?"
817                 " "
818                 "([-+][0-2][0-9]:?[0-5][0-9])\n";
819         const char *timestamp = NULL, *cp, *colon;
820         static regex_t *stamp;
821         regmatch_t m[10];
822         int zoneoffset;
823         int hourminute;
824         int status;
825
826         for (cp = nameline; *cp != '\n'; cp++) {
827                 if (*cp == '\t')
828                         timestamp = cp + 1;
829         }
830         if (!timestamp)
831                 return 0;
832         if (!stamp) {
833                 stamp = xmalloc(sizeof(*stamp));
834                 if (regcomp(stamp, stamp_regexp, REG_EXTENDED)) {
835                         warning(_("Cannot prepare timestamp regexp %s"),
836                                 stamp_regexp);
837                         return 0;
838                 }
839         }
840
841         status = regexec(stamp, timestamp, ARRAY_SIZE(m), m, 0);
842         if (status) {
843                 if (status != REG_NOMATCH)
844                         warning(_("regexec returned %d for input: %s"),
845                                 status, timestamp);
846                 return 0;
847         }
848
849         zoneoffset = strtol(timestamp + m[3].rm_so + 1, (char **) &colon, 10);
850         if (*colon == ':')
851                 zoneoffset = zoneoffset * 60 + strtol(colon + 1, NULL, 10);
852         else
853                 zoneoffset = (zoneoffset / 100) * 60 + (zoneoffset % 100);
854         if (timestamp[m[3].rm_so] == '-')
855                 zoneoffset = -zoneoffset;
856
857         /*
858          * YYYY-MM-DD hh:mm:ss must be from either 1969-12-31
859          * (west of GMT) or 1970-01-01 (east of GMT)
860          */
861         if ((zoneoffset < 0 && memcmp(timestamp, "1969-12-31", 10)) ||
862             (0 <= zoneoffset && memcmp(timestamp, "1970-01-01", 10)))
863                 return 0;
864
865         hourminute = (strtol(timestamp + 11, NULL, 10) * 60 +
866                       strtol(timestamp + 14, NULL, 10) -
867                       zoneoffset);
868
869         return ((zoneoffset < 0 && hourminute == 1440) ||
870                 (0 <= zoneoffset && !hourminute));
871 }
872
873 /*
874  * Get the name etc info from the ---/+++ lines of a traditional patch header
875  *
876  * FIXME! The end-of-filename heuristics are kind of screwy. For existing
877  * files, we can happily check the index for a match, but for creating a
878  * new file we should try to match whatever "patch" does. I have no idea.
879  */
880 static int parse_traditional_patch(struct apply_state *state,
881                                    const char *first,
882                                    const char *second,
883                                    struct patch *patch)
884 {
885         char *name;
886
887         first += 4;     /* skip "--- " */
888         second += 4;    /* skip "+++ " */
889         if (!state->p_value_known) {
890                 int p, q;
891                 p = guess_p_value(state, first);
892                 q = guess_p_value(state, second);
893                 if (p < 0) p = q;
894                 if (0 <= p && p == q) {
895                         state->p_value = p;
896                         state->p_value_known = 1;
897                 }
898         }
899         if (is_dev_null(first)) {
900                 patch->is_new = 1;
901                 patch->is_delete = 0;
902                 name = find_name_traditional(state, second, NULL, state->p_value);
903                 patch->new_name = name;
904         } else if (is_dev_null(second)) {
905                 patch->is_new = 0;
906                 patch->is_delete = 1;
907                 name = find_name_traditional(state, first, NULL, state->p_value);
908                 patch->old_name = name;
909         } else {
910                 char *first_name;
911                 first_name = find_name_traditional(state, first, NULL, state->p_value);
912                 name = find_name_traditional(state, second, first_name, state->p_value);
913                 free(first_name);
914                 if (has_epoch_timestamp(first)) {
915                         patch->is_new = 1;
916                         patch->is_delete = 0;
917                         patch->new_name = name;
918                 } else if (has_epoch_timestamp(second)) {
919                         patch->is_new = 0;
920                         patch->is_delete = 1;
921                         patch->old_name = name;
922                 } else {
923                         patch->old_name = name;
924                         patch->new_name = xstrdup_or_null(name);
925                 }
926         }
927         if (!name)
928                 return error(_("unable to find filename in patch at line %d"), state->linenr);
929
930         return 0;
931 }
932
933 static int gitdiff_hdrend(struct apply_state *state,
934                           const char *line,
935                           struct patch *patch)
936 {
937         return 1;
938 }
939
940 /*
941  * We're anal about diff header consistency, to make
942  * sure that we don't end up having strange ambiguous
943  * patches floating around.
944  *
945  * As a result, gitdiff_{old|new}name() will check
946  * their names against any previous information, just
947  * to make sure..
948  */
949 #define DIFF_OLD_NAME 0
950 #define DIFF_NEW_NAME 1
951
952 static int gitdiff_verify_name(struct apply_state *state,
953                                const char *line,
954                                int isnull,
955                                char **name,
956                                int side)
957 {
958         if (!*name && !isnull) {
959                 *name = find_name(state, line, NULL, state->p_value, TERM_TAB);
960                 return 0;
961         }
962
963         if (*name) {
964                 int len = strlen(*name);
965                 char *another;
966                 if (isnull)
967                         return error(_("git apply: bad git-diff - expected /dev/null, got %s on line %d"),
968                                      *name, state->linenr);
969                 another = find_name(state, line, NULL, state->p_value, TERM_TAB);
970                 if (!another || memcmp(another, *name, len + 1)) {
971                         free(another);
972                         return error((side == DIFF_NEW_NAME) ?
973                             _("git apply: bad git-diff - inconsistent new filename on line %d") :
974                             _("git apply: bad git-diff - inconsistent old filename on line %d"), state->linenr);
975                 }
976                 free(another);
977         } else {
978                 /* expect "/dev/null" */
979                 if (memcmp("/dev/null", line, 9) || line[9] != '\n')
980                         return error(_("git apply: bad git-diff - expected /dev/null on line %d"), state->linenr);
981         }
982
983         return 0;
984 }
985
986 static int gitdiff_oldname(struct apply_state *state,
987                            const char *line,
988                            struct patch *patch)
989 {
990         return gitdiff_verify_name(state, line,
991                                    patch->is_new, &patch->old_name,
992                                    DIFF_OLD_NAME);
993 }
994
995 static int gitdiff_newname(struct apply_state *state,
996                            const char *line,
997                            struct patch *patch)
998 {
999         return gitdiff_verify_name(state, line,
1000                                    patch->is_delete, &patch->new_name,
1001                                    DIFF_NEW_NAME);
1002 }
1003
1004 static int gitdiff_oldmode(struct apply_state *state,
1005                            const char *line,
1006                            struct patch *patch)
1007 {
1008         patch->old_mode = strtoul(line, NULL, 8);
1009         return 0;
1010 }
1011
1012 static int gitdiff_newmode(struct apply_state *state,
1013                            const char *line,
1014                            struct patch *patch)
1015 {
1016         patch->new_mode = strtoul(line, NULL, 8);
1017         return 0;
1018 }
1019
1020 static int gitdiff_delete(struct apply_state *state,
1021                           const char *line,
1022                           struct patch *patch)
1023 {
1024         patch->is_delete = 1;
1025         free(patch->old_name);
1026         patch->old_name = xstrdup_or_null(patch->def_name);
1027         return gitdiff_oldmode(state, line, patch);
1028 }
1029
1030 static int gitdiff_newfile(struct apply_state *state,
1031                            const char *line,
1032                            struct patch *patch)
1033 {
1034         patch->is_new = 1;
1035         free(patch->new_name);
1036         patch->new_name = xstrdup_or_null(patch->def_name);
1037         return gitdiff_newmode(state, line, patch);
1038 }
1039
1040 static int gitdiff_copysrc(struct apply_state *state,
1041                            const char *line,
1042                            struct patch *patch)
1043 {
1044         patch->is_copy = 1;
1045         free(patch->old_name);
1046         patch->old_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1047         return 0;
1048 }
1049
1050 static int gitdiff_copydst(struct apply_state *state,
1051                            const char *line,
1052                            struct patch *patch)
1053 {
1054         patch->is_copy = 1;
1055         free(patch->new_name);
1056         patch->new_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1057         return 0;
1058 }
1059
1060 static int gitdiff_renamesrc(struct apply_state *state,
1061                              const char *line,
1062                              struct patch *patch)
1063 {
1064         patch->is_rename = 1;
1065         free(patch->old_name);
1066         patch->old_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1067         return 0;
1068 }
1069
1070 static int gitdiff_renamedst(struct apply_state *state,
1071                              const char *line,
1072                              struct patch *patch)
1073 {
1074         patch->is_rename = 1;
1075         free(patch->new_name);
1076         patch->new_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1077         return 0;
1078 }
1079
1080 static int gitdiff_similarity(struct apply_state *state,
1081                               const char *line,
1082                               struct patch *patch)
1083 {
1084         unsigned long val = strtoul(line, NULL, 10);
1085         if (val <= 100)
1086                 patch->score = val;
1087         return 0;
1088 }
1089
1090 static int gitdiff_dissimilarity(struct apply_state *state,
1091                                  const char *line,
1092                                  struct patch *patch)
1093 {
1094         unsigned long val = strtoul(line, NULL, 10);
1095         if (val <= 100)
1096                 patch->score = val;
1097         return 0;
1098 }
1099
1100 static int gitdiff_index(struct apply_state *state,
1101                          const char *line,
1102                          struct patch *patch)
1103 {
1104         /*
1105          * index line is N hexadecimal, "..", N hexadecimal,
1106          * and optional space with octal mode.
1107          */
1108         const char *ptr, *eol;
1109         int len;
1110
1111         ptr = strchr(line, '.');
1112         if (!ptr || ptr[1] != '.' || 40 < ptr - line)
1113                 return 0;
1114         len = ptr - line;
1115         memcpy(patch->old_sha1_prefix, line, len);
1116         patch->old_sha1_prefix[len] = 0;
1117
1118         line = ptr + 2;
1119         ptr = strchr(line, ' ');
1120         eol = strchrnul(line, '\n');
1121
1122         if (!ptr || eol < ptr)
1123                 ptr = eol;
1124         len = ptr - line;
1125
1126         if (40 < len)
1127                 return 0;
1128         memcpy(patch->new_sha1_prefix, line, len);
1129         patch->new_sha1_prefix[len] = 0;
1130         if (*ptr == ' ')
1131                 patch->old_mode = strtoul(ptr+1, NULL, 8);
1132         return 0;
1133 }
1134
1135 /*
1136  * This is normal for a diff that doesn't change anything: we'll fall through
1137  * into the next diff. Tell the parser to break out.
1138  */
1139 static int gitdiff_unrecognized(struct apply_state *state,
1140                                 const char *line,
1141                                 struct patch *patch)
1142 {
1143         return 1;
1144 }
1145
1146 /*
1147  * Skip p_value leading components from "line"; as we do not accept
1148  * absolute paths, return NULL in that case.
1149  */
1150 static const char *skip_tree_prefix(struct apply_state *state,
1151                                     const char *line,
1152                                     int llen)
1153 {
1154         int nslash;
1155         int i;
1156
1157         if (!state->p_value)
1158                 return (llen && line[0] == '/') ? NULL : line;
1159
1160         nslash = state->p_value;
1161         for (i = 0; i < llen; i++) {
1162                 int ch = line[i];
1163                 if (ch == '/' && --nslash <= 0)
1164                         return (i == 0) ? NULL : &line[i + 1];
1165         }
1166         return NULL;
1167 }
1168
1169 /*
1170  * This is to extract the same name that appears on "diff --git"
1171  * line.  We do not find and return anything if it is a rename
1172  * patch, and it is OK because we will find the name elsewhere.
1173  * We need to reliably find name only when it is mode-change only,
1174  * creation or deletion of an empty file.  In any of these cases,
1175  * both sides are the same name under a/ and b/ respectively.
1176  */
1177 static char *git_header_name(struct apply_state *state,
1178                              const char *line,
1179                              int llen)
1180 {
1181         const char *name;
1182         const char *second = NULL;
1183         size_t len, line_len;
1184
1185         line += strlen("diff --git ");
1186         llen -= strlen("diff --git ");
1187
1188         if (*line == '"') {
1189                 const char *cp;
1190                 struct strbuf first = STRBUF_INIT;
1191                 struct strbuf sp = STRBUF_INIT;
1192
1193                 if (unquote_c_style(&first, line, &second))
1194                         goto free_and_fail1;
1195
1196                 /* strip the a/b prefix including trailing slash */
1197                 cp = skip_tree_prefix(state, first.buf, first.len);
1198                 if (!cp)
1199                         goto free_and_fail1;
1200                 strbuf_remove(&first, 0, cp - first.buf);
1201
1202                 /*
1203                  * second points at one past closing dq of name.
1204                  * find the second name.
1205                  */
1206                 while ((second < line + llen) && isspace(*second))
1207                         second++;
1208
1209                 if (line + llen <= second)
1210                         goto free_and_fail1;
1211                 if (*second == '"') {
1212                         if (unquote_c_style(&sp, second, NULL))
1213                                 goto free_and_fail1;
1214                         cp = skip_tree_prefix(state, sp.buf, sp.len);
1215                         if (!cp)
1216                                 goto free_and_fail1;
1217                         /* They must match, otherwise ignore */
1218                         if (strcmp(cp, first.buf))
1219                                 goto free_and_fail1;
1220                         strbuf_release(&sp);
1221                         return strbuf_detach(&first, NULL);
1222                 }
1223
1224                 /* unquoted second */
1225                 cp = skip_tree_prefix(state, second, line + llen - second);
1226                 if (!cp)
1227                         goto free_and_fail1;
1228                 if (line + llen - cp != first.len ||
1229                     memcmp(first.buf, cp, first.len))
1230                         goto free_and_fail1;
1231                 return strbuf_detach(&first, NULL);
1232
1233         free_and_fail1:
1234                 strbuf_release(&first);
1235                 strbuf_release(&sp);
1236                 return NULL;
1237         }
1238
1239         /* unquoted first name */
1240         name = skip_tree_prefix(state, line, llen);
1241         if (!name)
1242                 return NULL;
1243
1244         /*
1245          * since the first name is unquoted, a dq if exists must be
1246          * the beginning of the second name.
1247          */
1248         for (second = name; second < line + llen; second++) {
1249                 if (*second == '"') {
1250                         struct strbuf sp = STRBUF_INIT;
1251                         const char *np;
1252
1253                         if (unquote_c_style(&sp, second, NULL))
1254                                 goto free_and_fail2;
1255
1256                         np = skip_tree_prefix(state, sp.buf, sp.len);
1257                         if (!np)
1258                                 goto free_and_fail2;
1259
1260                         len = sp.buf + sp.len - np;
1261                         if (len < second - name &&
1262                             !strncmp(np, name, len) &&
1263                             isspace(name[len])) {
1264                                 /* Good */
1265                                 strbuf_remove(&sp, 0, np - sp.buf);
1266                                 return strbuf_detach(&sp, NULL);
1267                         }
1268
1269                 free_and_fail2:
1270                         strbuf_release(&sp);
1271                         return NULL;
1272                 }
1273         }
1274
1275         /*
1276          * Accept a name only if it shows up twice, exactly the same
1277          * form.
1278          */
1279         second = strchr(name, '\n');
1280         if (!second)
1281                 return NULL;
1282         line_len = second - name;
1283         for (len = 0 ; ; len++) {
1284                 switch (name[len]) {
1285                 default:
1286                         continue;
1287                 case '\n':
1288                         return NULL;
1289                 case '\t': case ' ':
1290                         /*
1291                          * Is this the separator between the preimage
1292                          * and the postimage pathname?  Again, we are
1293                          * only interested in the case where there is
1294                          * no rename, as this is only to set def_name
1295                          * and a rename patch has the names elsewhere
1296                          * in an unambiguous form.
1297                          */
1298                         if (!name[len + 1])
1299                                 return NULL; /* no postimage name */
1300                         second = skip_tree_prefix(state, name + len + 1,
1301                                                   line_len - (len + 1));
1302                         if (!second)
1303                                 return NULL;
1304                         /*
1305                          * Does len bytes starting at "name" and "second"
1306                          * (that are separated by one HT or SP we just
1307                          * found) exactly match?
1308                          */
1309                         if (second[len] == '\n' && !strncmp(name, second, len))
1310                                 return xmemdupz(name, len);
1311                 }
1312         }
1313 }
1314
1315 /* Verify that we recognize the lines following a git header */
1316 static int parse_git_header(struct apply_state *state,
1317                             const char *line,
1318                             int len,
1319                             unsigned int size,
1320                             struct patch *patch)
1321 {
1322         unsigned long offset;
1323
1324         /* A git diff has explicit new/delete information, so we don't guess */
1325         patch->is_new = 0;
1326         patch->is_delete = 0;
1327
1328         /*
1329          * Some things may not have the old name in the
1330          * rest of the headers anywhere (pure mode changes,
1331          * or removing or adding empty files), so we get
1332          * the default name from the header.
1333          */
1334         patch->def_name = git_header_name(state, line, len);
1335         if (patch->def_name && state->root.len) {
1336                 char *s = xstrfmt("%s%s", state->root.buf, patch->def_name);
1337                 free(patch->def_name);
1338                 patch->def_name = s;
1339         }
1340
1341         line += len;
1342         size -= len;
1343         state->linenr++;
1344         for (offset = len ; size > 0 ; offset += len, size -= len, line += len, state->linenr++) {
1345                 static const struct opentry {
1346                         const char *str;
1347                         int (*fn)(struct apply_state *, const char *, struct patch *);
1348                 } optable[] = {
1349                         { "@@ -", gitdiff_hdrend },
1350                         { "--- ", gitdiff_oldname },
1351                         { "+++ ", gitdiff_newname },
1352                         { "old mode ", gitdiff_oldmode },
1353                         { "new mode ", gitdiff_newmode },
1354                         { "deleted file mode ", gitdiff_delete },
1355                         { "new file mode ", gitdiff_newfile },
1356                         { "copy from ", gitdiff_copysrc },
1357                         { "copy to ", gitdiff_copydst },
1358                         { "rename old ", gitdiff_renamesrc },
1359                         { "rename new ", gitdiff_renamedst },
1360                         { "rename from ", gitdiff_renamesrc },
1361                         { "rename to ", gitdiff_renamedst },
1362                         { "similarity index ", gitdiff_similarity },
1363                         { "dissimilarity index ", gitdiff_dissimilarity },
1364                         { "index ", gitdiff_index },
1365                         { "", gitdiff_unrecognized },
1366                 };
1367                 int i;
1368
1369                 len = linelen(line, size);
1370                 if (!len || line[len-1] != '\n')
1371                         break;
1372                 for (i = 0; i < ARRAY_SIZE(optable); i++) {
1373                         const struct opentry *p = optable + i;
1374                         int oplen = strlen(p->str);
1375                         int res;
1376                         if (len < oplen || memcmp(p->str, line, oplen))
1377                                 continue;
1378                         res = p->fn(state, line + oplen, patch);
1379                         if (res < 0)
1380                                 return -1;
1381                         if (res > 0)
1382                                 return offset;
1383                         break;
1384                 }
1385         }
1386
1387         return offset;
1388 }
1389
1390 static int parse_num(const char *line, unsigned long *p)
1391 {
1392         char *ptr;
1393
1394         if (!isdigit(*line))
1395                 return 0;
1396         *p = strtoul(line, &ptr, 10);
1397         return ptr - line;
1398 }
1399
1400 static int parse_range(const char *line, int len, int offset, const char *expect,
1401                        unsigned long *p1, unsigned long *p2)
1402 {
1403         int digits, ex;
1404
1405         if (offset < 0 || offset >= len)
1406                 return -1;
1407         line += offset;
1408         len -= offset;
1409
1410         digits = parse_num(line, p1);
1411         if (!digits)
1412                 return -1;
1413
1414         offset += digits;
1415         line += digits;
1416         len -= digits;
1417
1418         *p2 = 1;
1419         if (*line == ',') {
1420                 digits = parse_num(line+1, p2);
1421                 if (!digits)
1422                         return -1;
1423
1424                 offset += digits+1;
1425                 line += digits+1;
1426                 len -= digits+1;
1427         }
1428
1429         ex = strlen(expect);
1430         if (ex > len)
1431                 return -1;
1432         if (memcmp(line, expect, ex))
1433                 return -1;
1434
1435         return offset + ex;
1436 }
1437
1438 static void recount_diff(const char *line, int size, struct fragment *fragment)
1439 {
1440         int oldlines = 0, newlines = 0, ret = 0;
1441
1442         if (size < 1) {
1443                 warning("recount: ignore empty hunk");
1444                 return;
1445         }
1446
1447         for (;;) {
1448                 int len = linelen(line, size);
1449                 size -= len;
1450                 line += len;
1451
1452                 if (size < 1)
1453                         break;
1454
1455                 switch (*line) {
1456                 case ' ': case '\n':
1457                         newlines++;
1458                         /* fall through */
1459                 case '-':
1460                         oldlines++;
1461                         continue;
1462                 case '+':
1463                         newlines++;
1464                         continue;
1465                 case '\\':
1466                         continue;
1467                 case '@':
1468                         ret = size < 3 || !starts_with(line, "@@ ");
1469                         break;
1470                 case 'd':
1471                         ret = size < 5 || !starts_with(line, "diff ");
1472                         break;
1473                 default:
1474                         ret = -1;
1475                         break;
1476                 }
1477                 if (ret) {
1478                         warning(_("recount: unexpected line: %.*s"),
1479                                 (int)linelen(line, size), line);
1480                         return;
1481                 }
1482                 break;
1483         }
1484         fragment->oldlines = oldlines;
1485         fragment->newlines = newlines;
1486 }
1487
1488 /*
1489  * Parse a unified diff fragment header of the
1490  * form "@@ -a,b +c,d @@"
1491  */
1492 static int parse_fragment_header(const char *line, int len, struct fragment *fragment)
1493 {
1494         int offset;
1495
1496         if (!len || line[len-1] != '\n')
1497                 return -1;
1498
1499         /* Figure out the number of lines in a fragment */
1500         offset = parse_range(line, len, 4, " +", &fragment->oldpos, &fragment->oldlines);
1501         offset = parse_range(line, len, offset, " @@", &fragment->newpos, &fragment->newlines);
1502
1503         return offset;
1504 }
1505
1506 /*
1507  * Find file diff header
1508  *
1509  * Returns:
1510  *  -1 if no header was found
1511  *  -128 in case of error
1512  *   the size of the header in bytes (called "offset") otherwise
1513  */
1514 static int find_header(struct apply_state *state,
1515                        const char *line,
1516                        unsigned long size,
1517                        int *hdrsize,
1518                        struct patch *patch)
1519 {
1520         unsigned long offset, len;
1521
1522         patch->is_toplevel_relative = 0;
1523         patch->is_rename = patch->is_copy = 0;
1524         patch->is_new = patch->is_delete = -1;
1525         patch->old_mode = patch->new_mode = 0;
1526         patch->old_name = patch->new_name = NULL;
1527         for (offset = 0; size > 0; offset += len, size -= len, line += len, state->linenr++) {
1528                 unsigned long nextlen;
1529
1530                 len = linelen(line, size);
1531                 if (!len)
1532                         break;
1533
1534                 /* Testing this early allows us to take a few shortcuts.. */
1535                 if (len < 6)
1536                         continue;
1537
1538                 /*
1539                  * Make sure we don't find any unconnected patch fragments.
1540                  * That's a sign that we didn't find a header, and that a
1541                  * patch has become corrupted/broken up.
1542                  */
1543                 if (!memcmp("@@ -", line, 4)) {
1544                         struct fragment dummy;
1545                         if (parse_fragment_header(line, len, &dummy) < 0)
1546                                 continue;
1547                         error(_("patch fragment without header at line %d: %.*s"),
1548                                      state->linenr, (int)len-1, line);
1549                         return -128;
1550                 }
1551
1552                 if (size < len + 6)
1553                         break;
1554
1555                 /*
1556                  * Git patch? It might not have a real patch, just a rename
1557                  * or mode change, so we handle that specially
1558                  */
1559                 if (!memcmp("diff --git ", line, 11)) {
1560                         int git_hdr_len = parse_git_header(state, line, len, size, patch);
1561                         if (git_hdr_len < 0)
1562                                 return -128;
1563                         if (git_hdr_len <= len)
1564                                 continue;
1565                         if (!patch->old_name && !patch->new_name) {
1566                                 if (!patch->def_name) {
1567                                         error(Q_("git diff header lacks filename information when removing "
1568                                                         "%d leading pathname component (line %d)",
1569                                                         "git diff header lacks filename information when removing "
1570                                                         "%d leading pathname components (line %d)",
1571                                                         state->p_value),
1572                                                      state->p_value, state->linenr);
1573                                         return -128;
1574                                 }
1575                                 patch->old_name = xstrdup(patch->def_name);
1576                                 patch->new_name = xstrdup(patch->def_name);
1577                         }
1578                         if (!patch->is_delete && !patch->new_name) {
1579                                 error(_("git diff header lacks filename information "
1580                                              "(line %d)"), state->linenr);
1581                                 return -128;
1582                         }
1583                         patch->is_toplevel_relative = 1;
1584                         *hdrsize = git_hdr_len;
1585                         return offset;
1586                 }
1587
1588                 /* --- followed by +++ ? */
1589                 if (memcmp("--- ", line,  4) || memcmp("+++ ", line + len, 4))
1590                         continue;
1591
1592                 /*
1593                  * We only accept unified patches, so we want it to
1594                  * at least have "@@ -a,b +c,d @@\n", which is 14 chars
1595                  * minimum ("@@ -0,0 +1 @@\n" is the shortest).
1596                  */
1597                 nextlen = linelen(line + len, size - len);
1598                 if (size < nextlen + 14 || memcmp("@@ -", line + len + nextlen, 4))
1599                         continue;
1600
1601                 /* Ok, we'll consider it a patch */
1602                 if (parse_traditional_patch(state, line, line+len, patch))
1603                         return -128;
1604                 *hdrsize = len + nextlen;
1605                 state->linenr += 2;
1606                 return offset;
1607         }
1608         return -1;
1609 }
1610
1611 static void record_ws_error(struct apply_state *state,
1612                             unsigned result,
1613                             const char *line,
1614                             int len,
1615                             int linenr)
1616 {
1617         char *err;
1618
1619         if (!result)
1620                 return;
1621
1622         state->whitespace_error++;
1623         if (state->squelch_whitespace_errors &&
1624             state->squelch_whitespace_errors < state->whitespace_error)
1625                 return;
1626
1627         err = whitespace_error_string(result);
1628         if (state->apply_verbosity > verbosity_silent)
1629                 fprintf(stderr, "%s:%d: %s.\n%.*s\n",
1630                         state->patch_input_file, linenr, err, len, line);
1631         free(err);
1632 }
1633
1634 static void check_whitespace(struct apply_state *state,
1635                              const char *line,
1636                              int len,
1637                              unsigned ws_rule)
1638 {
1639         unsigned result = ws_check(line + 1, len - 1, ws_rule);
1640
1641         record_ws_error(state, result, line + 1, len - 2, state->linenr);
1642 }
1643
1644 /*
1645  * Parse a unified diff. Note that this really needs to parse each
1646  * fragment separately, since the only way to know the difference
1647  * between a "---" that is part of a patch, and a "---" that starts
1648  * the next patch is to look at the line counts..
1649  */
1650 static int parse_fragment(struct apply_state *state,
1651                           const char *line,
1652                           unsigned long size,
1653                           struct patch *patch,
1654                           struct fragment *fragment)
1655 {
1656         int added, deleted;
1657         int len = linelen(line, size), offset;
1658         unsigned long oldlines, newlines;
1659         unsigned long leading, trailing;
1660
1661         offset = parse_fragment_header(line, len, fragment);
1662         if (offset < 0)
1663                 return -1;
1664         if (offset > 0 && patch->recount)
1665                 recount_diff(line + offset, size - offset, fragment);
1666         oldlines = fragment->oldlines;
1667         newlines = fragment->newlines;
1668         leading = 0;
1669         trailing = 0;
1670
1671         /* Parse the thing.. */
1672         line += len;
1673         size -= len;
1674         state->linenr++;
1675         added = deleted = 0;
1676         for (offset = len;
1677              0 < size;
1678              offset += len, size -= len, line += len, state->linenr++) {
1679                 if (!oldlines && !newlines)
1680                         break;
1681                 len = linelen(line, size);
1682                 if (!len || line[len-1] != '\n')
1683                         return -1;
1684                 switch (*line) {
1685                 default:
1686                         return -1;
1687                 case '\n': /* newer GNU diff, an empty context line */
1688                 case ' ':
1689                         oldlines--;
1690                         newlines--;
1691                         if (!deleted && !added)
1692                                 leading++;
1693                         trailing++;
1694                         if (!state->apply_in_reverse &&
1695                             state->ws_error_action == correct_ws_error)
1696                                 check_whitespace(state, line, len, patch->ws_rule);
1697                         break;
1698                 case '-':
1699                         if (state->apply_in_reverse &&
1700                             state->ws_error_action != nowarn_ws_error)
1701                                 check_whitespace(state, line, len, patch->ws_rule);
1702                         deleted++;
1703                         oldlines--;
1704                         trailing = 0;
1705                         break;
1706                 case '+':
1707                         if (!state->apply_in_reverse &&
1708                             state->ws_error_action != nowarn_ws_error)
1709                                 check_whitespace(state, line, len, patch->ws_rule);
1710                         added++;
1711                         newlines--;
1712                         trailing = 0;
1713                         break;
1714
1715                 /*
1716                  * We allow "\ No newline at end of file". Depending
1717                  * on locale settings when the patch was produced we
1718                  * don't know what this line looks like. The only
1719                  * thing we do know is that it begins with "\ ".
1720                  * Checking for 12 is just for sanity check -- any
1721                  * l10n of "\ No newline..." is at least that long.
1722                  */
1723                 case '\\':
1724                         if (len < 12 || memcmp(line, "\\ ", 2))
1725                                 return -1;
1726                         break;
1727                 }
1728         }
1729         if (oldlines || newlines)
1730                 return -1;
1731         if (!deleted && !added)
1732                 return -1;
1733
1734         fragment->leading = leading;
1735         fragment->trailing = trailing;
1736
1737         /*
1738          * If a fragment ends with an incomplete line, we failed to include
1739          * it in the above loop because we hit oldlines == newlines == 0
1740          * before seeing it.
1741          */
1742         if (12 < size && !memcmp(line, "\\ ", 2))
1743                 offset += linelen(line, size);
1744
1745         patch->lines_added += added;
1746         patch->lines_deleted += deleted;
1747
1748         if (0 < patch->is_new && oldlines)
1749                 return error(_("new file depends on old contents"));
1750         if (0 < patch->is_delete && newlines)
1751                 return error(_("deleted file still has contents"));
1752         return offset;
1753 }
1754
1755 /*
1756  * We have seen "diff --git a/... b/..." header (or a traditional patch
1757  * header).  Read hunks that belong to this patch into fragments and hang
1758  * them to the given patch structure.
1759  *
1760  * The (fragment->patch, fragment->size) pair points into the memory given
1761  * by the caller, not a copy, when we return.
1762  *
1763  * Returns:
1764  *   -1 in case of error,
1765  *   the number of bytes in the patch otherwise.
1766  */
1767 static int parse_single_patch(struct apply_state *state,
1768                               const char *line,
1769                               unsigned long size,
1770                               struct patch *patch)
1771 {
1772         unsigned long offset = 0;
1773         unsigned long oldlines = 0, newlines = 0, context = 0;
1774         struct fragment **fragp = &patch->fragments;
1775
1776         while (size > 4 && !memcmp(line, "@@ -", 4)) {
1777                 struct fragment *fragment;
1778                 int len;
1779
1780                 fragment = xcalloc(1, sizeof(*fragment));
1781                 fragment->linenr = state->linenr;
1782                 len = parse_fragment(state, line, size, patch, fragment);
1783                 if (len <= 0) {
1784                         free(fragment);
1785                         return error(_("corrupt patch at line %d"), state->linenr);
1786                 }
1787                 fragment->patch = line;
1788                 fragment->size = len;
1789                 oldlines += fragment->oldlines;
1790                 newlines += fragment->newlines;
1791                 context += fragment->leading + fragment->trailing;
1792
1793                 *fragp = fragment;
1794                 fragp = &fragment->next;
1795
1796                 offset += len;
1797                 line += len;
1798                 size -= len;
1799         }
1800
1801         /*
1802          * If something was removed (i.e. we have old-lines) it cannot
1803          * be creation, and if something was added it cannot be
1804          * deletion.  However, the reverse is not true; --unified=0
1805          * patches that only add are not necessarily creation even
1806          * though they do not have any old lines, and ones that only
1807          * delete are not necessarily deletion.
1808          *
1809          * Unfortunately, a real creation/deletion patch do _not_ have
1810          * any context line by definition, so we cannot safely tell it
1811          * apart with --unified=0 insanity.  At least if the patch has
1812          * more than one hunk it is not creation or deletion.
1813          */
1814         if (patch->is_new < 0 &&
1815             (oldlines || (patch->fragments && patch->fragments->next)))
1816                 patch->is_new = 0;
1817         if (patch->is_delete < 0 &&
1818             (newlines || (patch->fragments && patch->fragments->next)))
1819                 patch->is_delete = 0;
1820
1821         if (0 < patch->is_new && oldlines)
1822                 return error(_("new file %s depends on old contents"), patch->new_name);
1823         if (0 < patch->is_delete && newlines)
1824                 return error(_("deleted file %s still has contents"), patch->old_name);
1825         if (!patch->is_delete && !newlines && context && state->apply_verbosity > verbosity_silent)
1826                 fprintf_ln(stderr,
1827                            _("** warning: "
1828                              "file %s becomes empty but is not deleted"),
1829                            patch->new_name);
1830
1831         return offset;
1832 }
1833
1834 static inline int metadata_changes(struct patch *patch)
1835 {
1836         return  patch->is_rename > 0 ||
1837                 patch->is_copy > 0 ||
1838                 patch->is_new > 0 ||
1839                 patch->is_delete ||
1840                 (patch->old_mode && patch->new_mode &&
1841                  patch->old_mode != patch->new_mode);
1842 }
1843
1844 static char *inflate_it(const void *data, unsigned long size,
1845                         unsigned long inflated_size)
1846 {
1847         git_zstream stream;
1848         void *out;
1849         int st;
1850
1851         memset(&stream, 0, sizeof(stream));
1852
1853         stream.next_in = (unsigned char *)data;
1854         stream.avail_in = size;
1855         stream.next_out = out = xmalloc(inflated_size);
1856         stream.avail_out = inflated_size;
1857         git_inflate_init(&stream);
1858         st = git_inflate(&stream, Z_FINISH);
1859         git_inflate_end(&stream);
1860         if ((st != Z_STREAM_END) || stream.total_out != inflated_size) {
1861                 free(out);
1862                 return NULL;
1863         }
1864         return out;
1865 }
1866
1867 /*
1868  * Read a binary hunk and return a new fragment; fragment->patch
1869  * points at an allocated memory that the caller must free, so
1870  * it is marked as "->free_patch = 1".
1871  */
1872 static struct fragment *parse_binary_hunk(struct apply_state *state,
1873                                           char **buf_p,
1874                                           unsigned long *sz_p,
1875                                           int *status_p,
1876                                           int *used_p)
1877 {
1878         /*
1879          * Expect a line that begins with binary patch method ("literal"
1880          * or "delta"), followed by the length of data before deflating.
1881          * a sequence of 'length-byte' followed by base-85 encoded data
1882          * should follow, terminated by a newline.
1883          *
1884          * Each 5-byte sequence of base-85 encodes up to 4 bytes,
1885          * and we would limit the patch line to 66 characters,
1886          * so one line can fit up to 13 groups that would decode
1887          * to 52 bytes max.  The length byte 'A'-'Z' corresponds
1888          * to 1-26 bytes, and 'a'-'z' corresponds to 27-52 bytes.
1889          */
1890         int llen, used;
1891         unsigned long size = *sz_p;
1892         char *buffer = *buf_p;
1893         int patch_method;
1894         unsigned long origlen;
1895         char *data = NULL;
1896         int hunk_size = 0;
1897         struct fragment *frag;
1898
1899         llen = linelen(buffer, size);
1900         used = llen;
1901
1902         *status_p = 0;
1903
1904         if (starts_with(buffer, "delta ")) {
1905                 patch_method = BINARY_DELTA_DEFLATED;
1906                 origlen = strtoul(buffer + 6, NULL, 10);
1907         }
1908         else if (starts_with(buffer, "literal ")) {
1909                 patch_method = BINARY_LITERAL_DEFLATED;
1910                 origlen = strtoul(buffer + 8, NULL, 10);
1911         }
1912         else
1913                 return NULL;
1914
1915         state->linenr++;
1916         buffer += llen;
1917         while (1) {
1918                 int byte_length, max_byte_length, newsize;
1919                 llen = linelen(buffer, size);
1920                 used += llen;
1921                 state->linenr++;
1922                 if (llen == 1) {
1923                         /* consume the blank line */
1924                         buffer++;
1925                         size--;
1926                         break;
1927                 }
1928                 /*
1929                  * Minimum line is "A00000\n" which is 7-byte long,
1930                  * and the line length must be multiple of 5 plus 2.
1931                  */
1932                 if ((llen < 7) || (llen-2) % 5)
1933                         goto corrupt;
1934                 max_byte_length = (llen - 2) / 5 * 4;
1935                 byte_length = *buffer;
1936                 if ('A' <= byte_length && byte_length <= 'Z')
1937                         byte_length = byte_length - 'A' + 1;
1938                 else if ('a' <= byte_length && byte_length <= 'z')
1939                         byte_length = byte_length - 'a' + 27;
1940                 else
1941                         goto corrupt;
1942                 /* if the input length was not multiple of 4, we would
1943                  * have filler at the end but the filler should never
1944                  * exceed 3 bytes
1945                  */
1946                 if (max_byte_length < byte_length ||
1947                     byte_length <= max_byte_length - 4)
1948                         goto corrupt;
1949                 newsize = hunk_size + byte_length;
1950                 data = xrealloc(data, newsize);
1951                 if (decode_85(data + hunk_size, buffer + 1, byte_length))
1952                         goto corrupt;
1953                 hunk_size = newsize;
1954                 buffer += llen;
1955                 size -= llen;
1956         }
1957
1958         frag = xcalloc(1, sizeof(*frag));
1959         frag->patch = inflate_it(data, hunk_size, origlen);
1960         frag->free_patch = 1;
1961         if (!frag->patch)
1962                 goto corrupt;
1963         free(data);
1964         frag->size = origlen;
1965         *buf_p = buffer;
1966         *sz_p = size;
1967         *used_p = used;
1968         frag->binary_patch_method = patch_method;
1969         return frag;
1970
1971  corrupt:
1972         free(data);
1973         *status_p = -1;
1974         error(_("corrupt binary patch at line %d: %.*s"),
1975               state->linenr-1, llen-1, buffer);
1976         return NULL;
1977 }
1978
1979 /*
1980  * Returns:
1981  *   -1 in case of error,
1982  *   the length of the parsed binary patch otherwise
1983  */
1984 static int parse_binary(struct apply_state *state,
1985                         char *buffer,
1986                         unsigned long size,
1987                         struct patch *patch)
1988 {
1989         /*
1990          * We have read "GIT binary patch\n"; what follows is a line
1991          * that says the patch method (currently, either "literal" or
1992          * "delta") and the length of data before deflating; a
1993          * sequence of 'length-byte' followed by base-85 encoded data
1994          * follows.
1995          *
1996          * When a binary patch is reversible, there is another binary
1997          * hunk in the same format, starting with patch method (either
1998          * "literal" or "delta") with the length of data, and a sequence
1999          * of length-byte + base-85 encoded data, terminated with another
2000          * empty line.  This data, when applied to the postimage, produces
2001          * the preimage.
2002          */
2003         struct fragment *forward;
2004         struct fragment *reverse;
2005         int status;
2006         int used, used_1;
2007
2008         forward = parse_binary_hunk(state, &buffer, &size, &status, &used);
2009         if (!forward && !status)
2010                 /* there has to be one hunk (forward hunk) */
2011                 return error(_("unrecognized binary patch at line %d"), state->linenr-1);
2012         if (status)
2013                 /* otherwise we already gave an error message */
2014                 return status;
2015
2016         reverse = parse_binary_hunk(state, &buffer, &size, &status, &used_1);
2017         if (reverse)
2018                 used += used_1;
2019         else if (status) {
2020                 /*
2021                  * Not having reverse hunk is not an error, but having
2022                  * a corrupt reverse hunk is.
2023                  */
2024                 free((void*) forward->patch);
2025                 free(forward);
2026                 return status;
2027         }
2028         forward->next = reverse;
2029         patch->fragments = forward;
2030         patch->is_binary = 1;
2031         return used;
2032 }
2033
2034 static void prefix_one(struct apply_state *state, char **name)
2035 {
2036         char *old_name = *name;
2037         if (!old_name)
2038                 return;
2039         *name = prefix_filename(state->prefix, *name);
2040         free(old_name);
2041 }
2042
2043 static void prefix_patch(struct apply_state *state, struct patch *p)
2044 {
2045         if (!state->prefix || p->is_toplevel_relative)
2046                 return;
2047         prefix_one(state, &p->new_name);
2048         prefix_one(state, &p->old_name);
2049 }
2050
2051 /*
2052  * include/exclude
2053  */
2054
2055 static void add_name_limit(struct apply_state *state,
2056                            const char *name,
2057                            int exclude)
2058 {
2059         struct string_list_item *it;
2060
2061         it = string_list_append(&state->limit_by_name, name);
2062         it->util = exclude ? NULL : (void *) 1;
2063 }
2064
2065 static int use_patch(struct apply_state *state, struct patch *p)
2066 {
2067         const char *pathname = p->new_name ? p->new_name : p->old_name;
2068         int i;
2069
2070         /* Paths outside are not touched regardless of "--include" */
2071         if (0 < state->prefix_length) {
2072                 int pathlen = strlen(pathname);
2073                 if (pathlen <= state->prefix_length ||
2074                     memcmp(state->prefix, pathname, state->prefix_length))
2075                         return 0;
2076         }
2077
2078         /* See if it matches any of exclude/include rule */
2079         for (i = 0; i < state->limit_by_name.nr; i++) {
2080                 struct string_list_item *it = &state->limit_by_name.items[i];
2081                 if (!wildmatch(it->string, pathname, 0, NULL))
2082                         return (it->util != NULL);
2083         }
2084
2085         /*
2086          * If we had any include, a path that does not match any rule is
2087          * not used.  Otherwise, we saw bunch of exclude rules (or none)
2088          * and such a path is used.
2089          */
2090         return !state->has_include;
2091 }
2092
2093 /*
2094  * Read the patch text in "buffer" that extends for "size" bytes; stop
2095  * reading after seeing a single patch (i.e. changes to a single file).
2096  * Create fragments (i.e. patch hunks) and hang them to the given patch.
2097  *
2098  * Returns:
2099  *   -1 if no header was found or parse_binary() failed,
2100  *   -128 on another error,
2101  *   the number of bytes consumed otherwise,
2102  *     so that the caller can call us again for the next patch.
2103  */
2104 static int parse_chunk(struct apply_state *state, char *buffer, unsigned long size, struct patch *patch)
2105 {
2106         int hdrsize, patchsize;
2107         int offset = find_header(state, buffer, size, &hdrsize, patch);
2108
2109         if (offset < 0)
2110                 return offset;
2111
2112         prefix_patch(state, patch);
2113
2114         if (!use_patch(state, patch))
2115                 patch->ws_rule = 0;
2116         else
2117                 patch->ws_rule = whitespace_rule(patch->new_name
2118                                                  ? patch->new_name
2119                                                  : patch->old_name);
2120
2121         patchsize = parse_single_patch(state,
2122                                        buffer + offset + hdrsize,
2123                                        size - offset - hdrsize,
2124                                        patch);
2125
2126         if (patchsize < 0)
2127                 return -128;
2128
2129         if (!patchsize) {
2130                 static const char git_binary[] = "GIT binary patch\n";
2131                 int hd = hdrsize + offset;
2132                 unsigned long llen = linelen(buffer + hd, size - hd);
2133
2134                 if (llen == sizeof(git_binary) - 1 &&
2135                     !memcmp(git_binary, buffer + hd, llen)) {
2136                         int used;
2137                         state->linenr++;
2138                         used = parse_binary(state, buffer + hd + llen,
2139                                             size - hd - llen, patch);
2140                         if (used < 0)
2141                                 return -1;
2142                         if (used)
2143                                 patchsize = used + llen;
2144                         else
2145                                 patchsize = 0;
2146                 }
2147                 else if (!memcmp(" differ\n", buffer + hd + llen - 8, 8)) {
2148                         static const char *binhdr[] = {
2149                                 "Binary files ",
2150                                 "Files ",
2151                                 NULL,
2152                         };
2153                         int i;
2154                         for (i = 0; binhdr[i]; i++) {
2155                                 int len = strlen(binhdr[i]);
2156                                 if (len < size - hd &&
2157                                     !memcmp(binhdr[i], buffer + hd, len)) {
2158                                         state->linenr++;
2159                                         patch->is_binary = 1;
2160                                         patchsize = llen;
2161                                         break;
2162                                 }
2163                         }
2164                 }
2165
2166                 /* Empty patch cannot be applied if it is a text patch
2167                  * without metadata change.  A binary patch appears
2168                  * empty to us here.
2169                  */
2170                 if ((state->apply || state->check) &&
2171                     (!patch->is_binary && !metadata_changes(patch))) {
2172                         error(_("patch with only garbage at line %d"), state->linenr);
2173                         return -128;
2174                 }
2175         }
2176
2177         return offset + hdrsize + patchsize;
2178 }
2179
2180 static void reverse_patches(struct patch *p)
2181 {
2182         for (; p; p = p->next) {
2183                 struct fragment *frag = p->fragments;
2184
2185                 SWAP(p->new_name, p->old_name);
2186                 SWAP(p->new_mode, p->old_mode);
2187                 SWAP(p->is_new, p->is_delete);
2188                 SWAP(p->lines_added, p->lines_deleted);
2189                 SWAP(p->old_sha1_prefix, p->new_sha1_prefix);
2190
2191                 for (; frag; frag = frag->next) {
2192                         SWAP(frag->newpos, frag->oldpos);
2193                         SWAP(frag->newlines, frag->oldlines);
2194                 }
2195         }
2196 }
2197
2198 static const char pluses[] =
2199 "++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++";
2200 static const char minuses[]=
2201 "----------------------------------------------------------------------";
2202
2203 static void show_stats(struct apply_state *state, struct patch *patch)
2204 {
2205         struct strbuf qname = STRBUF_INIT;
2206         char *cp = patch->new_name ? patch->new_name : patch->old_name;
2207         int max, add, del;
2208
2209         quote_c_style(cp, &qname, NULL, 0);
2210
2211         /*
2212          * "scale" the filename
2213          */
2214         max = state->max_len;
2215         if (max > 50)
2216                 max = 50;
2217
2218         if (qname.len > max) {
2219                 cp = strchr(qname.buf + qname.len + 3 - max, '/');
2220                 if (!cp)
2221                         cp = qname.buf + qname.len + 3 - max;
2222                 strbuf_splice(&qname, 0, cp - qname.buf, "...", 3);
2223         }
2224
2225         if (patch->is_binary) {
2226                 printf(" %-*s |  Bin\n", max, qname.buf);
2227                 strbuf_release(&qname);
2228                 return;
2229         }
2230
2231         printf(" %-*s |", max, qname.buf);
2232         strbuf_release(&qname);
2233
2234         /*
2235          * scale the add/delete
2236          */
2237         max = max + state->max_change > 70 ? 70 - max : state->max_change;
2238         add = patch->lines_added;
2239         del = patch->lines_deleted;
2240
2241         if (state->max_change > 0) {
2242                 int total = ((add + del) * max + state->max_change / 2) / state->max_change;
2243                 add = (add * max + state->max_change / 2) / state->max_change;
2244                 del = total - add;
2245         }
2246         printf("%5d %.*s%.*s\n", patch->lines_added + patch->lines_deleted,
2247                 add, pluses, del, minuses);
2248 }
2249
2250 static int read_old_data(struct stat *st, const char *path, struct strbuf *buf)
2251 {
2252         switch (st->st_mode & S_IFMT) {
2253         case S_IFLNK:
2254                 if (strbuf_readlink(buf, path, st->st_size) < 0)
2255                         return error(_("unable to read symlink %s"), path);
2256                 return 0;
2257         case S_IFREG:
2258                 if (strbuf_read_file(buf, path, st->st_size) != st->st_size)
2259                         return error(_("unable to open or read %s"), path);
2260                 convert_to_git(&the_index, path, buf->buf, buf->len, buf, 0);
2261                 return 0;
2262         default:
2263                 return -1;
2264         }
2265 }
2266
2267 /*
2268  * Update the preimage, and the common lines in postimage,
2269  * from buffer buf of length len. If postlen is 0 the postimage
2270  * is updated in place, otherwise it's updated on a new buffer
2271  * of length postlen
2272  */
2273
2274 static void update_pre_post_images(struct image *preimage,
2275                                    struct image *postimage,
2276                                    char *buf,
2277                                    size_t len, size_t postlen)
2278 {
2279         int i, ctx, reduced;
2280         char *new, *old, *fixed;
2281         struct image fixed_preimage;
2282
2283         /*
2284          * Update the preimage with whitespace fixes.  Note that we
2285          * are not losing preimage->buf -- apply_one_fragment() will
2286          * free "oldlines".
2287          */
2288         prepare_image(&fixed_preimage, buf, len, 1);
2289         assert(postlen
2290                ? fixed_preimage.nr == preimage->nr
2291                : fixed_preimage.nr <= preimage->nr);
2292         for (i = 0; i < fixed_preimage.nr; i++)
2293                 fixed_preimage.line[i].flag = preimage->line[i].flag;
2294         free(preimage->line_allocated);
2295         *preimage = fixed_preimage;
2296
2297         /*
2298          * Adjust the common context lines in postimage. This can be
2299          * done in-place when we are shrinking it with whitespace
2300          * fixing, but needs a new buffer when ignoring whitespace or
2301          * expanding leading tabs to spaces.
2302          *
2303          * We trust the caller to tell us if the update can be done
2304          * in place (postlen==0) or not.
2305          */
2306         old = postimage->buf;
2307         if (postlen)
2308                 new = postimage->buf = xmalloc(postlen);
2309         else
2310                 new = old;
2311         fixed = preimage->buf;
2312
2313         for (i = reduced = ctx = 0; i < postimage->nr; i++) {
2314                 size_t l_len = postimage->line[i].len;
2315                 if (!(postimage->line[i].flag & LINE_COMMON)) {
2316                         /* an added line -- no counterparts in preimage */
2317                         memmove(new, old, l_len);
2318                         old += l_len;
2319                         new += l_len;
2320                         continue;
2321                 }
2322
2323                 /* a common context -- skip it in the original postimage */
2324                 old += l_len;
2325
2326                 /* and find the corresponding one in the fixed preimage */
2327                 while (ctx < preimage->nr &&
2328                        !(preimage->line[ctx].flag & LINE_COMMON)) {
2329                         fixed += preimage->line[ctx].len;
2330                         ctx++;
2331                 }
2332
2333                 /*
2334                  * preimage is expected to run out, if the caller
2335                  * fixed addition of trailing blank lines.
2336                  */
2337                 if (preimage->nr <= ctx) {
2338                         reduced++;
2339                         continue;
2340                 }
2341
2342                 /* and copy it in, while fixing the line length */
2343                 l_len = preimage->line[ctx].len;
2344                 memcpy(new, fixed, l_len);
2345                 new += l_len;
2346                 fixed += l_len;
2347                 postimage->line[i].len = l_len;
2348                 ctx++;
2349         }
2350
2351         if (postlen
2352             ? postlen < new - postimage->buf
2353             : postimage->len < new - postimage->buf)
2354                 die("BUG: caller miscounted postlen: asked %d, orig = %d, used = %d",
2355                     (int)postlen, (int) postimage->len, (int)(new - postimage->buf));
2356
2357         /* Fix the length of the whole thing */
2358         postimage->len = new - postimage->buf;
2359         postimage->nr -= reduced;
2360 }
2361
2362 static int line_by_line_fuzzy_match(struct image *img,
2363                                     struct image *preimage,
2364                                     struct image *postimage,
2365                                     unsigned long try,
2366                                     int try_lno,
2367                                     int preimage_limit)
2368 {
2369         int i;
2370         size_t imgoff = 0;
2371         size_t preoff = 0;
2372         size_t postlen = postimage->len;
2373         size_t extra_chars;
2374         char *buf;
2375         char *preimage_eof;
2376         char *preimage_end;
2377         struct strbuf fixed;
2378         char *fixed_buf;
2379         size_t fixed_len;
2380
2381         for (i = 0; i < preimage_limit; i++) {
2382                 size_t prelen = preimage->line[i].len;
2383                 size_t imglen = img->line[try_lno+i].len;
2384
2385                 if (!fuzzy_matchlines(img->buf + try + imgoff, imglen,
2386                                       preimage->buf + preoff, prelen))
2387                         return 0;
2388                 if (preimage->line[i].flag & LINE_COMMON)
2389                         postlen += imglen - prelen;
2390                 imgoff += imglen;
2391                 preoff += prelen;
2392         }
2393
2394         /*
2395          * Ok, the preimage matches with whitespace fuzz.
2396          *
2397          * imgoff now holds the true length of the target that
2398          * matches the preimage before the end of the file.
2399          *
2400          * Count the number of characters in the preimage that fall
2401          * beyond the end of the file and make sure that all of them
2402          * are whitespace characters. (This can only happen if
2403          * we are removing blank lines at the end of the file.)
2404          */
2405         buf = preimage_eof = preimage->buf + preoff;
2406         for ( ; i < preimage->nr; i++)
2407                 preoff += preimage->line[i].len;
2408         preimage_end = preimage->buf + preoff;
2409         for ( ; buf < preimage_end; buf++)
2410                 if (!isspace(*buf))
2411                         return 0;
2412
2413         /*
2414          * Update the preimage and the common postimage context
2415          * lines to use the same whitespace as the target.
2416          * If whitespace is missing in the target (i.e.
2417          * if the preimage extends beyond the end of the file),
2418          * use the whitespace from the preimage.
2419          */
2420         extra_chars = preimage_end - preimage_eof;
2421         strbuf_init(&fixed, imgoff + extra_chars);
2422         strbuf_add(&fixed, img->buf + try, imgoff);
2423         strbuf_add(&fixed, preimage_eof, extra_chars);
2424         fixed_buf = strbuf_detach(&fixed, &fixed_len);
2425         update_pre_post_images(preimage, postimage,
2426                                fixed_buf, fixed_len, postlen);
2427         return 1;
2428 }
2429
2430 static int match_fragment(struct apply_state *state,
2431                           struct image *img,
2432                           struct image *preimage,
2433                           struct image *postimage,
2434                           unsigned long try,
2435                           int try_lno,
2436                           unsigned ws_rule,
2437                           int match_beginning, int match_end)
2438 {
2439         int i;
2440         char *fixed_buf, *buf, *orig, *target;
2441         struct strbuf fixed;
2442         size_t fixed_len, postlen;
2443         int preimage_limit;
2444
2445         if (preimage->nr + try_lno <= img->nr) {
2446                 /*
2447                  * The hunk falls within the boundaries of img.
2448                  */
2449                 preimage_limit = preimage->nr;
2450                 if (match_end && (preimage->nr + try_lno != img->nr))
2451                         return 0;
2452         } else if (state->ws_error_action == correct_ws_error &&
2453                    (ws_rule & WS_BLANK_AT_EOF)) {
2454                 /*
2455                  * This hunk extends beyond the end of img, and we are
2456                  * removing blank lines at the end of the file.  This
2457                  * many lines from the beginning of the preimage must
2458                  * match with img, and the remainder of the preimage
2459                  * must be blank.
2460                  */
2461                 preimage_limit = img->nr - try_lno;
2462         } else {
2463                 /*
2464                  * The hunk extends beyond the end of the img and
2465                  * we are not removing blanks at the end, so we
2466                  * should reject the hunk at this position.
2467                  */
2468                 return 0;
2469         }
2470
2471         if (match_beginning && try_lno)
2472                 return 0;
2473
2474         /* Quick hash check */
2475         for (i = 0; i < preimage_limit; i++)
2476                 if ((img->line[try_lno + i].flag & LINE_PATCHED) ||
2477                     (preimage->line[i].hash != img->line[try_lno + i].hash))
2478                         return 0;
2479
2480         if (preimage_limit == preimage->nr) {
2481                 /*
2482                  * Do we have an exact match?  If we were told to match
2483                  * at the end, size must be exactly at try+fragsize,
2484                  * otherwise try+fragsize must be still within the preimage,
2485                  * and either case, the old piece should match the preimage
2486                  * exactly.
2487                  */
2488                 if ((match_end
2489                      ? (try + preimage->len == img->len)
2490                      : (try + preimage->len <= img->len)) &&
2491                     !memcmp(img->buf + try, preimage->buf, preimage->len))
2492                         return 1;
2493         } else {
2494                 /*
2495                  * The preimage extends beyond the end of img, so
2496                  * there cannot be an exact match.
2497                  *
2498                  * There must be one non-blank context line that match
2499                  * a line before the end of img.
2500                  */
2501                 char *buf_end;
2502
2503                 buf = preimage->buf;
2504                 buf_end = buf;
2505                 for (i = 0; i < preimage_limit; i++)
2506                         buf_end += preimage->line[i].len;
2507
2508                 for ( ; buf < buf_end; buf++)
2509                         if (!isspace(*buf))
2510                                 break;
2511                 if (buf == buf_end)
2512                         return 0;
2513         }
2514
2515         /*
2516          * No exact match. If we are ignoring whitespace, run a line-by-line
2517          * fuzzy matching. We collect all the line length information because
2518          * we need it to adjust whitespace if we match.
2519          */
2520         if (state->ws_ignore_action == ignore_ws_change)
2521                 return line_by_line_fuzzy_match(img, preimage, postimage,
2522                                                 try, try_lno, preimage_limit);
2523
2524         if (state->ws_error_action != correct_ws_error)
2525                 return 0;
2526
2527         /*
2528          * The hunk does not apply byte-by-byte, but the hash says
2529          * it might with whitespace fuzz. We weren't asked to
2530          * ignore whitespace, we were asked to correct whitespace
2531          * errors, so let's try matching after whitespace correction.
2532          *
2533          * While checking the preimage against the target, whitespace
2534          * errors in both fixed, we count how large the corresponding
2535          * postimage needs to be.  The postimage prepared by
2536          * apply_one_fragment() has whitespace errors fixed on added
2537          * lines already, but the common lines were propagated as-is,
2538          * which may become longer when their whitespace errors are
2539          * fixed.
2540          */
2541
2542         /* First count added lines in postimage */
2543         postlen = 0;
2544         for (i = 0; i < postimage->nr; i++) {
2545                 if (!(postimage->line[i].flag & LINE_COMMON))
2546                         postlen += postimage->line[i].len;
2547         }
2548
2549         /*
2550          * The preimage may extend beyond the end of the file,
2551          * but in this loop we will only handle the part of the
2552          * preimage that falls within the file.
2553          */
2554         strbuf_init(&fixed, preimage->len + 1);
2555         orig = preimage->buf;
2556         target = img->buf + try;
2557         for (i = 0; i < preimage_limit; i++) {
2558                 size_t oldlen = preimage->line[i].len;
2559                 size_t tgtlen = img->line[try_lno + i].len;
2560                 size_t fixstart = fixed.len;
2561                 struct strbuf tgtfix;
2562                 int match;
2563
2564                 /* Try fixing the line in the preimage */
2565                 ws_fix_copy(&fixed, orig, oldlen, ws_rule, NULL);
2566
2567                 /* Try fixing the line in the target */
2568                 strbuf_init(&tgtfix, tgtlen);
2569                 ws_fix_copy(&tgtfix, target, tgtlen, ws_rule, NULL);
2570
2571                 /*
2572                  * If they match, either the preimage was based on
2573                  * a version before our tree fixed whitespace breakage,
2574                  * or we are lacking a whitespace-fix patch the tree
2575                  * the preimage was based on already had (i.e. target
2576                  * has whitespace breakage, the preimage doesn't).
2577                  * In either case, we are fixing the whitespace breakages
2578                  * so we might as well take the fix together with their
2579                  * real change.
2580                  */
2581                 match = (tgtfix.len == fixed.len - fixstart &&
2582                          !memcmp(tgtfix.buf, fixed.buf + fixstart,
2583                                              fixed.len - fixstart));
2584
2585                 /* Add the length if this is common with the postimage */
2586                 if (preimage->line[i].flag & LINE_COMMON)
2587                         postlen += tgtfix.len;
2588
2589                 strbuf_release(&tgtfix);
2590                 if (!match)
2591                         goto unmatch_exit;
2592
2593                 orig += oldlen;
2594                 target += tgtlen;
2595         }
2596
2597
2598         /*
2599          * Now handle the lines in the preimage that falls beyond the
2600          * end of the file (if any). They will only match if they are
2601          * empty or only contain whitespace (if WS_BLANK_AT_EOL is
2602          * false).
2603          */
2604         for ( ; i < preimage->nr; i++) {
2605                 size_t fixstart = fixed.len; /* start of the fixed preimage */
2606                 size_t oldlen = preimage->line[i].len;
2607                 int j;
2608
2609                 /* Try fixing the line in the preimage */
2610                 ws_fix_copy(&fixed, orig, oldlen, ws_rule, NULL);
2611
2612                 for (j = fixstart; j < fixed.len; j++)
2613                         if (!isspace(fixed.buf[j]))
2614                                 goto unmatch_exit;
2615
2616                 orig += oldlen;
2617         }
2618
2619         /*
2620          * Yes, the preimage is based on an older version that still
2621          * has whitespace breakages unfixed, and fixing them makes the
2622          * hunk match.  Update the context lines in the postimage.
2623          */
2624         fixed_buf = strbuf_detach(&fixed, &fixed_len);
2625         if (postlen < postimage->len)
2626                 postlen = 0;
2627         update_pre_post_images(preimage, postimage,
2628                                fixed_buf, fixed_len, postlen);
2629         return 1;
2630
2631  unmatch_exit:
2632         strbuf_release(&fixed);
2633         return 0;
2634 }
2635
2636 static int find_pos(struct apply_state *state,
2637                     struct image *img,
2638                     struct image *preimage,
2639                     struct image *postimage,
2640                     int line,
2641                     unsigned ws_rule,
2642                     int match_beginning, int match_end)
2643 {
2644         int i;
2645         unsigned long backwards, forwards, try;
2646         int backwards_lno, forwards_lno, try_lno;
2647
2648         /*
2649          * If match_beginning or match_end is specified, there is no
2650          * point starting from a wrong line that will never match and
2651          * wander around and wait for a match at the specified end.
2652          */
2653         if (match_beginning)
2654                 line = 0;
2655         else if (match_end)
2656                 line = img->nr - preimage->nr;
2657
2658         /*
2659          * Because the comparison is unsigned, the following test
2660          * will also take care of a negative line number that can
2661          * result when match_end and preimage is larger than the target.
2662          */
2663         if ((size_t) line > img->nr)
2664                 line = img->nr;
2665
2666         try = 0;
2667         for (i = 0; i < line; i++)
2668                 try += img->line[i].len;
2669
2670         /*
2671          * There's probably some smart way to do this, but I'll leave
2672          * that to the smart and beautiful people. I'm simple and stupid.
2673          */
2674         backwards = try;
2675         backwards_lno = line;
2676         forwards = try;
2677         forwards_lno = line;
2678         try_lno = line;
2679
2680         for (i = 0; ; i++) {
2681                 if (match_fragment(state, img, preimage, postimage,
2682                                    try, try_lno, ws_rule,
2683                                    match_beginning, match_end))
2684                         return try_lno;
2685
2686         again:
2687                 if (backwards_lno == 0 && forwards_lno == img->nr)
2688                         break;
2689
2690                 if (i & 1) {
2691                         if (backwards_lno == 0) {
2692                                 i++;
2693                                 goto again;
2694                         }
2695                         backwards_lno--;
2696                         backwards -= img->line[backwards_lno].len;
2697                         try = backwards;
2698                         try_lno = backwards_lno;
2699                 } else {
2700                         if (forwards_lno == img->nr) {
2701                                 i++;
2702                                 goto again;
2703                         }
2704                         forwards += img->line[forwards_lno].len;
2705                         forwards_lno++;
2706                         try = forwards;
2707                         try_lno = forwards_lno;
2708                 }
2709
2710         }
2711         return -1;
2712 }
2713
2714 static void remove_first_line(struct image *img)
2715 {
2716         img->buf += img->line[0].len;
2717         img->len -= img->line[0].len;
2718         img->line++;
2719         img->nr--;
2720 }
2721
2722 static void remove_last_line(struct image *img)
2723 {
2724         img->len -= img->line[--img->nr].len;
2725 }
2726
2727 /*
2728  * The change from "preimage" and "postimage" has been found to
2729  * apply at applied_pos (counts in line numbers) in "img".
2730  * Update "img" to remove "preimage" and replace it with "postimage".
2731  */
2732 static void update_image(struct apply_state *state,
2733                          struct image *img,
2734                          int applied_pos,
2735                          struct image *preimage,
2736                          struct image *postimage)
2737 {
2738         /*
2739          * remove the copy of preimage at offset in img
2740          * and replace it with postimage
2741          */
2742         int i, nr;
2743         size_t remove_count, insert_count, applied_at = 0;
2744         char *result;
2745         int preimage_limit;
2746
2747         /*
2748          * If we are removing blank lines at the end of img,
2749          * the preimage may extend beyond the end.
2750          * If that is the case, we must be careful only to
2751          * remove the part of the preimage that falls within
2752          * the boundaries of img. Initialize preimage_limit
2753          * to the number of lines in the preimage that falls
2754          * within the boundaries.
2755          */
2756         preimage_limit = preimage->nr;
2757         if (preimage_limit > img->nr - applied_pos)
2758                 preimage_limit = img->nr - applied_pos;
2759
2760         for (i = 0; i < applied_pos; i++)
2761                 applied_at += img->line[i].len;
2762
2763         remove_count = 0;
2764         for (i = 0; i < preimage_limit; i++)
2765                 remove_count += img->line[applied_pos + i].len;
2766         insert_count = postimage->len;
2767
2768         /* Adjust the contents */
2769         result = xmalloc(st_add3(st_sub(img->len, remove_count), insert_count, 1));
2770         memcpy(result, img->buf, applied_at);
2771         memcpy(result + applied_at, postimage->buf, postimage->len);
2772         memcpy(result + applied_at + postimage->len,
2773                img->buf + (applied_at + remove_count),
2774                img->len - (applied_at + remove_count));
2775         free(img->buf);
2776         img->buf = result;
2777         img->len += insert_count - remove_count;
2778         result[img->len] = '\0';
2779
2780         /* Adjust the line table */
2781         nr = img->nr + postimage->nr - preimage_limit;
2782         if (preimage_limit < postimage->nr) {
2783                 /*
2784                  * NOTE: this knows that we never call remove_first_line()
2785                  * on anything other than pre/post image.
2786                  */
2787                 REALLOC_ARRAY(img->line, nr);
2788                 img->line_allocated = img->line;
2789         }
2790         if (preimage_limit != postimage->nr)
2791                 memmove(img->line + applied_pos + postimage->nr,
2792                         img->line + applied_pos + preimage_limit,
2793                         (img->nr - (applied_pos + preimage_limit)) *
2794                         sizeof(*img->line));
2795         memcpy(img->line + applied_pos,
2796                postimage->line,
2797                postimage->nr * sizeof(*img->line));
2798         if (!state->allow_overlap)
2799                 for (i = 0; i < postimage->nr; i++)
2800                         img->line[applied_pos + i].flag |= LINE_PATCHED;
2801         img->nr = nr;
2802 }
2803
2804 /*
2805  * Use the patch-hunk text in "frag" to prepare two images (preimage and
2806  * postimage) for the hunk.  Find lines that match "preimage" in "img" and
2807  * replace the part of "img" with "postimage" text.
2808  */
2809 static int apply_one_fragment(struct apply_state *state,
2810                               struct image *img, struct fragment *frag,
2811                               int inaccurate_eof, unsigned ws_rule,
2812                               int nth_fragment)
2813 {
2814         int match_beginning, match_end;
2815         const char *patch = frag->patch;
2816         int size = frag->size;
2817         char *old, *oldlines;
2818         struct strbuf newlines;
2819         int new_blank_lines_at_end = 0;
2820         int found_new_blank_lines_at_end = 0;
2821         int hunk_linenr = frag->linenr;
2822         unsigned long leading, trailing;
2823         int pos, applied_pos;
2824         struct image preimage;
2825         struct image postimage;
2826
2827         memset(&preimage, 0, sizeof(preimage));
2828         memset(&postimage, 0, sizeof(postimage));
2829         oldlines = xmalloc(size);
2830         strbuf_init(&newlines, size);
2831
2832         old = oldlines;
2833         while (size > 0) {
2834                 char first;
2835                 int len = linelen(patch, size);
2836                 int plen;
2837                 int added_blank_line = 0;
2838                 int is_blank_context = 0;
2839                 size_t start;
2840
2841                 if (!len)
2842                         break;
2843
2844                 /*
2845                  * "plen" is how much of the line we should use for
2846                  * the actual patch data. Normally we just remove the
2847                  * first character on the line, but if the line is
2848                  * followed by "\ No newline", then we also remove the
2849                  * last one (which is the newline, of course).
2850                  */
2851                 plen = len - 1;
2852                 if (len < size && patch[len] == '\\')
2853                         plen--;
2854                 first = *patch;
2855                 if (state->apply_in_reverse) {
2856                         if (first == '-')
2857                                 first = '+';
2858                         else if (first == '+')
2859                                 first = '-';
2860                 }
2861
2862                 switch (first) {
2863                 case '\n':
2864                         /* Newer GNU diff, empty context line */
2865                         if (plen < 0)
2866                                 /* ... followed by '\No newline'; nothing */
2867                                 break;
2868                         *old++ = '\n';
2869                         strbuf_addch(&newlines, '\n');
2870                         add_line_info(&preimage, "\n", 1, LINE_COMMON);
2871                         add_line_info(&postimage, "\n", 1, LINE_COMMON);
2872                         is_blank_context = 1;
2873                         break;
2874                 case ' ':
2875                         if (plen && (ws_rule & WS_BLANK_AT_EOF) &&
2876                             ws_blank_line(patch + 1, plen, ws_rule))
2877                                 is_blank_context = 1;
2878                 case '-':
2879                         memcpy(old, patch + 1, plen);
2880                         add_line_info(&preimage, old, plen,
2881                                       (first == ' ' ? LINE_COMMON : 0));
2882                         old += plen;
2883                         if (first == '-')
2884                                 break;
2885                 /* Fall-through for ' ' */
2886                 case '+':
2887                         /* --no-add does not add new lines */
2888                         if (first == '+' && state->no_add)
2889                                 break;
2890
2891                         start = newlines.len;
2892                         if (first != '+' ||
2893                             !state->whitespace_error ||
2894                             state->ws_error_action != correct_ws_error) {
2895                                 strbuf_add(&newlines, patch + 1, plen);
2896                         }
2897                         else {
2898                                 ws_fix_copy(&newlines, patch + 1, plen, ws_rule, &state->applied_after_fixing_ws);
2899                         }
2900                         add_line_info(&postimage, newlines.buf + start, newlines.len - start,
2901                                       (first == '+' ? 0 : LINE_COMMON));
2902                         if (first == '+' &&
2903                             (ws_rule & WS_BLANK_AT_EOF) &&
2904                             ws_blank_line(patch + 1, plen, ws_rule))
2905                                 added_blank_line = 1;
2906                         break;
2907                 case '@': case '\\':
2908                         /* Ignore it, we already handled it */
2909                         break;
2910                 default:
2911                         if (state->apply_verbosity > verbosity_normal)
2912                                 error(_("invalid start of line: '%c'"), first);
2913                         applied_pos = -1;
2914                         goto out;
2915                 }
2916                 if (added_blank_line) {
2917                         if (!new_blank_lines_at_end)
2918                                 found_new_blank_lines_at_end = hunk_linenr;
2919                         new_blank_lines_at_end++;
2920                 }
2921                 else if (is_blank_context)
2922                         ;
2923                 else
2924                         new_blank_lines_at_end = 0;
2925                 patch += len;
2926                 size -= len;
2927                 hunk_linenr++;
2928         }
2929         if (inaccurate_eof &&
2930             old > oldlines && old[-1] == '\n' &&
2931             newlines.len > 0 && newlines.buf[newlines.len - 1] == '\n') {
2932                 old--;
2933                 strbuf_setlen(&newlines, newlines.len - 1);
2934         }
2935
2936         leading = frag->leading;
2937         trailing = frag->trailing;
2938
2939         /*
2940          * A hunk to change lines at the beginning would begin with
2941          * @@ -1,L +N,M @@
2942          * but we need to be careful.  -U0 that inserts before the second
2943          * line also has this pattern.
2944          *
2945          * And a hunk to add to an empty file would begin with
2946          * @@ -0,0 +N,M @@
2947          *
2948          * In other words, a hunk that is (frag->oldpos <= 1) with or
2949          * without leading context must match at the beginning.
2950          */
2951         match_beginning = (!frag->oldpos ||
2952                            (frag->oldpos == 1 && !state->unidiff_zero));
2953
2954         /*
2955          * A hunk without trailing lines must match at the end.
2956          * However, we simply cannot tell if a hunk must match end
2957          * from the lack of trailing lines if the patch was generated
2958          * with unidiff without any context.
2959          */
2960         match_end = !state->unidiff_zero && !trailing;
2961
2962         pos = frag->newpos ? (frag->newpos - 1) : 0;
2963         preimage.buf = oldlines;
2964         preimage.len = old - oldlines;
2965         postimage.buf = newlines.buf;
2966         postimage.len = newlines.len;
2967         preimage.line = preimage.line_allocated;
2968         postimage.line = postimage.line_allocated;
2969
2970         for (;;) {
2971
2972                 applied_pos = find_pos(state, img, &preimage, &postimage, pos,
2973                                        ws_rule, match_beginning, match_end);
2974
2975                 if (applied_pos >= 0)
2976                         break;
2977
2978                 /* Am I at my context limits? */
2979                 if ((leading <= state->p_context) && (trailing <= state->p_context))
2980                         break;
2981                 if (match_beginning || match_end) {
2982                         match_beginning = match_end = 0;
2983                         continue;
2984                 }
2985
2986                 /*
2987                  * Reduce the number of context lines; reduce both
2988                  * leading and trailing if they are equal otherwise
2989                  * just reduce the larger context.
2990                  */
2991                 if (leading >= trailing) {
2992                         remove_first_line(&preimage);
2993                         remove_first_line(&postimage);
2994                         pos--;
2995                         leading--;
2996                 }
2997                 if (trailing > leading) {
2998                         remove_last_line(&preimage);
2999                         remove_last_line(&postimage);
3000                         trailing--;
3001                 }
3002         }
3003
3004         if (applied_pos >= 0) {
3005                 if (new_blank_lines_at_end &&
3006                     preimage.nr + applied_pos >= img->nr &&
3007                     (ws_rule & WS_BLANK_AT_EOF) &&
3008                     state->ws_error_action != nowarn_ws_error) {
3009                         record_ws_error(state, WS_BLANK_AT_EOF, "+", 1,
3010                                         found_new_blank_lines_at_end);
3011                         if (state->ws_error_action == correct_ws_error) {
3012                                 while (new_blank_lines_at_end--)
3013                                         remove_last_line(&postimage);
3014                         }
3015                         /*
3016                          * We would want to prevent write_out_results()
3017                          * from taking place in apply_patch() that follows
3018                          * the callchain led us here, which is:
3019                          * apply_patch->check_patch_list->check_patch->
3020                          * apply_data->apply_fragments->apply_one_fragment
3021                          */
3022                         if (state->ws_error_action == die_on_ws_error)
3023                                 state->apply = 0;
3024                 }
3025
3026                 if (state->apply_verbosity > verbosity_normal && applied_pos != pos) {
3027                         int offset = applied_pos - pos;
3028                         if (state->apply_in_reverse)
3029                                 offset = 0 - offset;
3030                         fprintf_ln(stderr,
3031                                    Q_("Hunk #%d succeeded at %d (offset %d line).",
3032                                       "Hunk #%d succeeded at %d (offset %d lines).",
3033                                       offset),
3034                                    nth_fragment, applied_pos + 1, offset);
3035                 }
3036
3037                 /*
3038                  * Warn if it was necessary to reduce the number
3039                  * of context lines.
3040                  */
3041                 if ((leading != frag->leading ||
3042                      trailing != frag->trailing) && state->apply_verbosity > verbosity_silent)
3043                         fprintf_ln(stderr, _("Context reduced to (%ld/%ld)"
3044                                              " to apply fragment at %d"),
3045                                    leading, trailing, applied_pos+1);
3046                 update_image(state, img, applied_pos, &preimage, &postimage);
3047         } else {
3048                 if (state->apply_verbosity > verbosity_normal)
3049                         error(_("while searching for:\n%.*s"),
3050                               (int)(old - oldlines), oldlines);
3051         }
3052
3053 out:
3054         free(oldlines);
3055         strbuf_release(&newlines);
3056         free(preimage.line_allocated);
3057         free(postimage.line_allocated);
3058
3059         return (applied_pos < 0);
3060 }
3061
3062 static int apply_binary_fragment(struct apply_state *state,
3063                                  struct image *img,
3064                                  struct patch *patch)
3065 {
3066         struct fragment *fragment = patch->fragments;
3067         unsigned long len;
3068         void *dst;
3069
3070         if (!fragment)
3071                 return error(_("missing binary patch data for '%s'"),
3072                              patch->new_name ?
3073                              patch->new_name :
3074                              patch->old_name);
3075
3076         /* Binary patch is irreversible without the optional second hunk */
3077         if (state->apply_in_reverse) {
3078                 if (!fragment->next)
3079                         return error(_("cannot reverse-apply a binary patch "
3080                                        "without the reverse hunk to '%s'"),
3081                                      patch->new_name
3082                                      ? patch->new_name : patch->old_name);
3083                 fragment = fragment->next;
3084         }
3085         switch (fragment->binary_patch_method) {
3086         case BINARY_DELTA_DEFLATED:
3087                 dst = patch_delta(img->buf, img->len, fragment->patch,
3088                                   fragment->size, &len);
3089                 if (!dst)
3090                         return -1;
3091                 clear_image(img);
3092                 img->buf = dst;
3093                 img->len = len;
3094                 return 0;
3095         case BINARY_LITERAL_DEFLATED:
3096                 clear_image(img);
3097                 img->len = fragment->size;
3098                 img->buf = xmemdupz(fragment->patch, img->len);
3099                 return 0;
3100         }
3101         return -1;
3102 }
3103
3104 /*
3105  * Replace "img" with the result of applying the binary patch.
3106  * The binary patch data itself in patch->fragment is still kept
3107  * but the preimage prepared by the caller in "img" is freed here
3108  * or in the helper function apply_binary_fragment() this calls.
3109  */
3110 static int apply_binary(struct apply_state *state,
3111                         struct image *img,
3112                         struct patch *patch)
3113 {
3114         const char *name = patch->old_name ? patch->old_name : patch->new_name;
3115         struct object_id oid;
3116
3117         /*
3118          * For safety, we require patch index line to contain
3119          * full 40-byte textual SHA1 for old and new, at least for now.
3120          */
3121         if (strlen(patch->old_sha1_prefix) != 40 ||
3122             strlen(patch->new_sha1_prefix) != 40 ||
3123             get_oid_hex(patch->old_sha1_prefix, &oid) ||
3124             get_oid_hex(patch->new_sha1_prefix, &oid))
3125                 return error(_("cannot apply binary patch to '%s' "
3126                                "without full index line"), name);
3127
3128         if (patch->old_name) {
3129                 /*
3130                  * See if the old one matches what the patch
3131                  * applies to.
3132                  */
3133                 hash_sha1_file(img->buf, img->len, blob_type, oid.hash);
3134                 if (strcmp(oid_to_hex(&oid), patch->old_sha1_prefix))
3135                         return error(_("the patch applies to '%s' (%s), "
3136                                        "which does not match the "
3137                                        "current contents."),
3138                                      name, oid_to_hex(&oid));
3139         }
3140         else {
3141                 /* Otherwise, the old one must be empty. */
3142                 if (img->len)
3143                         return error(_("the patch applies to an empty "
3144                                        "'%s' but it is not empty"), name);
3145         }
3146
3147         get_oid_hex(patch->new_sha1_prefix, &oid);
3148         if (is_null_oid(&oid)) {
3149                 clear_image(img);
3150                 return 0; /* deletion patch */
3151         }
3152
3153         if (has_sha1_file(oid.hash)) {
3154                 /* We already have the postimage */
3155                 enum object_type type;
3156                 unsigned long size;
3157                 char *result;
3158
3159                 result = read_sha1_file(oid.hash, &type, &size);
3160                 if (!result)
3161                         return error(_("the necessary postimage %s for "
3162                                        "'%s' cannot be read"),
3163                                      patch->new_sha1_prefix, name);
3164                 clear_image(img);
3165                 img->buf = result;
3166                 img->len = size;
3167         } else {
3168                 /*
3169                  * We have verified buf matches the preimage;
3170                  * apply the patch data to it, which is stored
3171                  * in the patch->fragments->{patch,size}.
3172                  */
3173                 if (apply_binary_fragment(state, img, patch))
3174                         return error(_("binary patch does not apply to '%s'"),
3175                                      name);
3176
3177                 /* verify that the result matches */
3178                 hash_sha1_file(img->buf, img->len, blob_type, oid.hash);
3179                 if (strcmp(oid_to_hex(&oid), patch->new_sha1_prefix))
3180                         return error(_("binary patch to '%s' creates incorrect result (expecting %s, got %s)"),
3181                                 name, patch->new_sha1_prefix, oid_to_hex(&oid));
3182         }
3183
3184         return 0;
3185 }
3186
3187 static int apply_fragments(struct apply_state *state, struct image *img, struct patch *patch)
3188 {
3189         struct fragment *frag = patch->fragments;
3190         const char *name = patch->old_name ? patch->old_name : patch->new_name;
3191         unsigned ws_rule = patch->ws_rule;
3192         unsigned inaccurate_eof = patch->inaccurate_eof;
3193         int nth = 0;
3194
3195         if (patch->is_binary)
3196                 return apply_binary(state, img, patch);
3197
3198         while (frag) {
3199                 nth++;
3200                 if (apply_one_fragment(state, img, frag, inaccurate_eof, ws_rule, nth)) {
3201                         error(_("patch failed: %s:%ld"), name, frag->oldpos);
3202                         if (!state->apply_with_reject)
3203                                 return -1;
3204                         frag->rejected = 1;
3205                 }
3206                 frag = frag->next;
3207         }
3208         return 0;
3209 }
3210
3211 static int read_blob_object(struct strbuf *buf, const struct object_id *oid, unsigned mode)
3212 {
3213         if (S_ISGITLINK(mode)) {
3214                 strbuf_grow(buf, 100);
3215                 strbuf_addf(buf, "Subproject commit %s\n", oid_to_hex(oid));
3216         } else {
3217                 enum object_type type;
3218                 unsigned long sz;
3219                 char *result;
3220
3221                 result = read_sha1_file(oid->hash, &type, &sz);
3222                 if (!result)
3223                         return -1;
3224                 /* XXX read_sha1_file NUL-terminates */
3225                 strbuf_attach(buf, result, sz, sz + 1);
3226         }
3227         return 0;
3228 }
3229
3230 static int read_file_or_gitlink(const struct cache_entry *ce, struct strbuf *buf)
3231 {
3232         if (!ce)
3233                 return 0;
3234         return read_blob_object(buf, &ce->oid, ce->ce_mode);
3235 }
3236
3237 static struct patch *in_fn_table(struct apply_state *state, const char *name)
3238 {
3239         struct string_list_item *item;
3240
3241         if (name == NULL)
3242                 return NULL;
3243
3244         item = string_list_lookup(&state->fn_table, name);
3245         if (item != NULL)
3246                 return (struct patch *)item->util;
3247
3248         return NULL;
3249 }
3250
3251 /*
3252  * item->util in the filename table records the status of the path.
3253  * Usually it points at a patch (whose result records the contents
3254  * of it after applying it), but it could be PATH_WAS_DELETED for a
3255  * path that a previously applied patch has already removed, or
3256  * PATH_TO_BE_DELETED for a path that a later patch would remove.
3257  *
3258  * The latter is needed to deal with a case where two paths A and B
3259  * are swapped by first renaming A to B and then renaming B to A;
3260  * moving A to B should not be prevented due to presence of B as we
3261  * will remove it in a later patch.
3262  */
3263 #define PATH_TO_BE_DELETED ((struct patch *) -2)
3264 #define PATH_WAS_DELETED ((struct patch *) -1)
3265
3266 static int to_be_deleted(struct patch *patch)
3267 {
3268         return patch == PATH_TO_BE_DELETED;
3269 }
3270
3271 static int was_deleted(struct patch *patch)
3272 {
3273         return patch == PATH_WAS_DELETED;
3274 }
3275
3276 static void add_to_fn_table(struct apply_state *state, struct patch *patch)
3277 {
3278         struct string_list_item *item;
3279
3280         /*
3281          * Always add new_name unless patch is a deletion
3282          * This should cover the cases for normal diffs,
3283          * file creations and copies
3284          */
3285         if (patch->new_name != NULL) {
3286                 item = string_list_insert(&state->fn_table, patch->new_name);
3287                 item->util = patch;
3288         }
3289
3290         /*
3291          * store a failure on rename/deletion cases because
3292          * later chunks shouldn't patch old names
3293          */
3294         if ((patch->new_name == NULL) || (patch->is_rename)) {
3295                 item = string_list_insert(&state->fn_table, patch->old_name);
3296                 item->util = PATH_WAS_DELETED;
3297         }
3298 }
3299
3300 static void prepare_fn_table(struct apply_state *state, struct patch *patch)
3301 {
3302         /*
3303          * store information about incoming file deletion
3304          */
3305         while (patch) {
3306                 if ((patch->new_name == NULL) || (patch->is_rename)) {
3307                         struct string_list_item *item;
3308                         item = string_list_insert(&state->fn_table, patch->old_name);
3309                         item->util = PATH_TO_BE_DELETED;
3310                 }
3311                 patch = patch->next;
3312         }
3313 }
3314
3315 static int checkout_target(struct index_state *istate,
3316                            struct cache_entry *ce, struct stat *st)
3317 {
3318         struct checkout costate = CHECKOUT_INIT;
3319
3320         costate.refresh_cache = 1;
3321         costate.istate = istate;
3322         if (checkout_entry(ce, &costate, NULL) || lstat(ce->name, st))
3323                 return error(_("cannot checkout %s"), ce->name);
3324         return 0;
3325 }
3326
3327 static struct patch *previous_patch(struct apply_state *state,
3328                                     struct patch *patch,
3329                                     int *gone)
3330 {
3331         struct patch *previous;
3332
3333         *gone = 0;
3334         if (patch->is_copy || patch->is_rename)
3335                 return NULL; /* "git" patches do not depend on the order */
3336
3337         previous = in_fn_table(state, patch->old_name);
3338         if (!previous)
3339                 return NULL;
3340
3341         if (to_be_deleted(previous))
3342                 return NULL; /* the deletion hasn't happened yet */
3343
3344         if (was_deleted(previous))
3345                 *gone = 1;
3346
3347         return previous;
3348 }
3349
3350 static int verify_index_match(const struct cache_entry *ce, struct stat *st)
3351 {
3352         if (S_ISGITLINK(ce->ce_mode)) {
3353                 if (!S_ISDIR(st->st_mode))
3354                         return -1;
3355                 return 0;
3356         }
3357         return ce_match_stat(ce, st, CE_MATCH_IGNORE_VALID|CE_MATCH_IGNORE_SKIP_WORKTREE);
3358 }
3359
3360 #define SUBMODULE_PATCH_WITHOUT_INDEX 1
3361
3362 static int load_patch_target(struct apply_state *state,
3363                              struct strbuf *buf,
3364                              const struct cache_entry *ce,
3365                              struct stat *st,
3366                              const char *name,
3367                              unsigned expected_mode)
3368 {
3369         if (state->cached || state->check_index) {
3370                 if (read_file_or_gitlink(ce, buf))
3371                         return error(_("failed to read %s"), name);
3372         } else if (name) {
3373                 if (S_ISGITLINK(expected_mode)) {
3374                         if (ce)
3375                                 return read_file_or_gitlink(ce, buf);
3376                         else
3377                                 return SUBMODULE_PATCH_WITHOUT_INDEX;
3378                 } else if (has_symlink_leading_path(name, strlen(name))) {
3379                         return error(_("reading from '%s' beyond a symbolic link"), name);
3380                 } else {
3381                         if (read_old_data(st, name, buf))
3382                                 return error(_("failed to read %s"), name);
3383                 }
3384         }
3385         return 0;
3386 }
3387
3388 /*
3389  * We are about to apply "patch"; populate the "image" with the
3390  * current version we have, from the working tree or from the index,
3391  * depending on the situation e.g. --cached/--index.  If we are
3392  * applying a non-git patch that incrementally updates the tree,
3393  * we read from the result of a previous diff.
3394  */
3395 static int load_preimage(struct apply_state *state,
3396                          struct image *image,
3397                          struct patch *patch, struct stat *st,
3398                          const struct cache_entry *ce)
3399 {
3400         struct strbuf buf = STRBUF_INIT;
3401         size_t len;
3402         char *img;
3403         struct patch *previous;
3404         int status;
3405
3406         previous = previous_patch(state, patch, &status);
3407         if (status)
3408                 return error(_("path %s has been renamed/deleted"),
3409                              patch->old_name);
3410         if (previous) {
3411                 /* We have a patched copy in memory; use that. */
3412                 strbuf_add(&buf, previous->result, previous->resultsize);
3413         } else {
3414                 status = load_patch_target(state, &buf, ce, st,
3415                                            patch->old_name, patch->old_mode);
3416                 if (status < 0)
3417                         return status;
3418                 else if (status == SUBMODULE_PATCH_WITHOUT_INDEX) {
3419                         /*
3420                          * There is no way to apply subproject
3421                          * patch without looking at the index.
3422                          * NEEDSWORK: shouldn't this be flagged
3423                          * as an error???
3424                          */
3425                         free_fragment_list(patch->fragments);
3426                         patch->fragments = NULL;
3427                 } else if (status) {
3428                         return error(_("failed to read %s"), patch->old_name);
3429                 }
3430         }
3431
3432         img = strbuf_detach(&buf, &len);
3433         prepare_image(image, img, len, !patch->is_binary);
3434         return 0;
3435 }
3436
3437 static int three_way_merge(struct image *image,
3438                            char *path,
3439                            const struct object_id *base,
3440                            const struct object_id *ours,
3441                            const struct object_id *theirs)
3442 {
3443         mmfile_t base_file, our_file, their_file;
3444         mmbuffer_t result = { NULL };
3445         int status;
3446
3447         read_mmblob(&base_file, base);
3448         read_mmblob(&our_file, ours);
3449         read_mmblob(&their_file, theirs);
3450         status = ll_merge(&result, path,
3451                           &base_file, "base",
3452                           &our_file, "ours",
3453                           &their_file, "theirs", NULL);
3454         free(base_file.ptr);
3455         free(our_file.ptr);
3456         free(their_file.ptr);
3457         if (status < 0 || !result.ptr) {
3458                 free(result.ptr);
3459                 return -1;
3460         }
3461         clear_image(image);
3462         image->buf = result.ptr;
3463         image->len = result.size;
3464
3465         return status;
3466 }
3467
3468 /*
3469  * When directly falling back to add/add three-way merge, we read from
3470  * the current contents of the new_name.  In no cases other than that
3471  * this function will be called.
3472  */
3473 static int load_current(struct apply_state *state,
3474                         struct image *image,
3475                         struct patch *patch)
3476 {
3477         struct strbuf buf = STRBUF_INIT;
3478         int status, pos;
3479         size_t len;
3480         char *img;
3481         struct stat st;
3482         struct cache_entry *ce;
3483         char *name = patch->new_name;
3484         unsigned mode = patch->new_mode;
3485
3486         if (!patch->is_new)
3487                 die("BUG: patch to %s is not a creation", patch->old_name);
3488
3489         pos = cache_name_pos(name, strlen(name));
3490         if (pos < 0)
3491                 return error(_("%s: does not exist in index"), name);
3492         ce = active_cache[pos];
3493         if (lstat(name, &st)) {
3494                 if (errno != ENOENT)
3495                         return error_errno("%s", name);
3496                 if (checkout_target(&the_index, ce, &st))
3497                         return -1;
3498         }
3499         if (verify_index_match(ce, &st))
3500                 return error(_("%s: does not match index"), name);
3501
3502         status = load_patch_target(state, &buf, ce, &st, name, mode);
3503         if (status < 0)
3504                 return status;
3505         else if (status)
3506                 return -1;
3507         img = strbuf_detach(&buf, &len);
3508         prepare_image(image, img, len, !patch->is_binary);
3509         return 0;
3510 }
3511
3512 static int try_threeway(struct apply_state *state,
3513                         struct image *image,
3514                         struct patch *patch,
3515                         struct stat *st,
3516                         const struct cache_entry *ce)
3517 {
3518         struct object_id pre_oid, post_oid, our_oid;
3519         struct strbuf buf = STRBUF_INIT;
3520         size_t len;
3521         int status;
3522         char *img;
3523         struct image tmp_image;
3524
3525         /* No point falling back to 3-way merge in these cases */
3526         if (patch->is_delete ||
3527             S_ISGITLINK(patch->old_mode) || S_ISGITLINK(patch->new_mode))
3528                 return -1;
3529
3530         /* Preimage the patch was prepared for */
3531         if (patch->is_new)
3532                 write_sha1_file("", 0, blob_type, pre_oid.hash);
3533         else if (get_sha1(patch->old_sha1_prefix, pre_oid.hash) ||
3534                  read_blob_object(&buf, &pre_oid, patch->old_mode))
3535                 return error(_("repository lacks the necessary blob to fall back on 3-way merge."));
3536
3537         if (state->apply_verbosity > verbosity_silent)
3538                 fprintf(stderr, _("Falling back to three-way merge...\n"));
3539
3540         img = strbuf_detach(&buf, &len);
3541         prepare_image(&tmp_image, img, len, 1);
3542         /* Apply the patch to get the post image */
3543         if (apply_fragments(state, &tmp_image, patch) < 0) {
3544                 clear_image(&tmp_image);
3545                 return -1;
3546         }
3547         /* post_oid is theirs */
3548         write_sha1_file(tmp_image.buf, tmp_image.len, blob_type, post_oid.hash);
3549         clear_image(&tmp_image);
3550
3551         /* our_oid is ours */
3552         if (patch->is_new) {
3553                 if (load_current(state, &tmp_image, patch))
3554                         return error(_("cannot read the current contents of '%s'"),
3555                                      patch->new_name);
3556         } else {
3557                 if (load_preimage(state, &tmp_image, patch, st, ce))
3558                         return error(_("cannot read the current contents of '%s'"),
3559                                      patch->old_name);
3560         }
3561         write_sha1_file(tmp_image.buf, tmp_image.len, blob_type, our_oid.hash);
3562         clear_image(&tmp_image);
3563
3564         /* in-core three-way merge between post and our using pre as base */
3565         status = three_way_merge(image, patch->new_name,
3566                                  &pre_oid, &our_oid, &post_oid);
3567         if (status < 0) {
3568                 if (state->apply_verbosity > verbosity_silent)
3569                         fprintf(stderr,
3570                                 _("Failed to fall back on three-way merge...\n"));
3571                 return status;
3572         }
3573
3574         if (status) {
3575                 patch->conflicted_threeway = 1;
3576                 if (patch->is_new)
3577                         oidclr(&patch->threeway_stage[0]);
3578                 else
3579                         oidcpy(&patch->threeway_stage[0], &pre_oid);
3580                 oidcpy(&patch->threeway_stage[1], &our_oid);
3581                 oidcpy(&patch->threeway_stage[2], &post_oid);
3582                 if (state->apply_verbosity > verbosity_silent)
3583                         fprintf(stderr,
3584                                 _("Applied patch to '%s' with conflicts.\n"),
3585                                 patch->new_name);
3586         } else {
3587                 if (state->apply_verbosity > verbosity_silent)
3588                         fprintf(stderr,
3589                                 _("Applied patch to '%s' cleanly.\n"),
3590                                 patch->new_name);
3591         }
3592         return 0;
3593 }
3594
3595 static int apply_data(struct apply_state *state, struct patch *patch,
3596                       struct stat *st, const struct cache_entry *ce)
3597 {
3598         struct image image;
3599
3600         if (load_preimage(state, &image, patch, st, ce) < 0)
3601                 return -1;
3602
3603         if (patch->direct_to_threeway ||
3604             apply_fragments(state, &image, patch) < 0) {
3605                 /* Note: with --reject, apply_fragments() returns 0 */
3606                 if (!state->threeway || try_threeway(state, &image, patch, st, ce) < 0)
3607                         return -1;
3608         }
3609         patch->result = image.buf;
3610         patch->resultsize = image.len;
3611         add_to_fn_table(state, patch);
3612         free(image.line_allocated);
3613
3614         if (0 < patch->is_delete && patch->resultsize)
3615                 return error(_("removal patch leaves file contents"));
3616
3617         return 0;
3618 }
3619
3620 /*
3621  * If "patch" that we are looking at modifies or deletes what we have,
3622  * we would want it not to lose any local modification we have, either
3623  * in the working tree or in the index.
3624  *
3625  * This also decides if a non-git patch is a creation patch or a
3626  * modification to an existing empty file.  We do not check the state
3627  * of the current tree for a creation patch in this function; the caller
3628  * check_patch() separately makes sure (and errors out otherwise) that
3629  * the path the patch creates does not exist in the current tree.
3630  */
3631 static int check_preimage(struct apply_state *state,
3632                           struct patch *patch,
3633                           struct cache_entry **ce,
3634                           struct stat *st)
3635 {
3636         const char *old_name = patch->old_name;
3637         struct patch *previous = NULL;
3638         int stat_ret = 0, status;
3639         unsigned st_mode = 0;
3640
3641         if (!old_name)
3642                 return 0;
3643
3644         assert(patch->is_new <= 0);
3645         previous = previous_patch(state, patch, &status);
3646
3647         if (status)
3648                 return error(_("path %s has been renamed/deleted"), old_name);
3649         if (previous) {
3650                 st_mode = previous->new_mode;
3651         } else if (!state->cached) {
3652                 stat_ret = lstat(old_name, st);
3653                 if (stat_ret && errno != ENOENT)
3654                         return error_errno("%s", old_name);
3655         }
3656
3657         if (state->check_index && !previous) {
3658                 int pos = cache_name_pos(old_name, strlen(old_name));
3659                 if (pos < 0) {
3660                         if (patch->is_new < 0)
3661                                 goto is_new;
3662                         return error(_("%s: does not exist in index"), old_name);
3663                 }
3664                 *ce = active_cache[pos];
3665                 if (stat_ret < 0) {
3666                         if (checkout_target(&the_index, *ce, st))
3667                                 return -1;
3668                 }
3669                 if (!state->cached && verify_index_match(*ce, st))
3670                         return error(_("%s: does not match index"), old_name);
3671                 if (state->cached)
3672                         st_mode = (*ce)->ce_mode;
3673         } else if (stat_ret < 0) {
3674                 if (patch->is_new < 0)
3675                         goto is_new;
3676                 return error_errno("%s", old_name);
3677         }
3678
3679         if (!state->cached && !previous)
3680                 st_mode = ce_mode_from_stat(*ce, st->st_mode);
3681
3682         if (patch->is_new < 0)
3683                 patch->is_new = 0;
3684         if (!patch->old_mode)
3685                 patch->old_mode = st_mode;
3686         if ((st_mode ^ patch->old_mode) & S_IFMT)
3687                 return error(_("%s: wrong type"), old_name);
3688         if (st_mode != patch->old_mode)
3689                 warning(_("%s has type %o, expected %o"),
3690                         old_name, st_mode, patch->old_mode);
3691         if (!patch->new_mode && !patch->is_delete)
3692                 patch->new_mode = st_mode;
3693         return 0;
3694
3695  is_new:
3696         patch->is_new = 1;
3697         patch->is_delete = 0;
3698         FREE_AND_NULL(patch->old_name);
3699         return 0;
3700 }
3701
3702
3703 #define EXISTS_IN_INDEX 1
3704 #define EXISTS_IN_WORKTREE 2
3705
3706 static int check_to_create(struct apply_state *state,
3707                            const char *new_name,
3708                            int ok_if_exists)
3709 {
3710         struct stat nst;
3711
3712         if (state->check_index &&
3713             cache_name_pos(new_name, strlen(new_name)) >= 0 &&
3714             !ok_if_exists)
3715                 return EXISTS_IN_INDEX;
3716         if (state->cached)
3717                 return 0;
3718
3719         if (!lstat(new_name, &nst)) {
3720                 if (S_ISDIR(nst.st_mode) || ok_if_exists)
3721                         return 0;
3722                 /*
3723                  * A leading component of new_name might be a symlink
3724                  * that is going to be removed with this patch, but
3725                  * still pointing at somewhere that has the path.
3726                  * In such a case, path "new_name" does not exist as
3727                  * far as git is concerned.
3728                  */
3729                 if (has_symlink_leading_path(new_name, strlen(new_name)))
3730                         return 0;
3731
3732                 return EXISTS_IN_WORKTREE;
3733         } else if (!is_missing_file_error(errno)) {
3734                 return error_errno("%s", new_name);
3735         }
3736         return 0;
3737 }
3738
3739 static uintptr_t register_symlink_changes(struct apply_state *state,
3740                                           const char *path,
3741                                           uintptr_t what)
3742 {
3743         struct string_list_item *ent;
3744
3745         ent = string_list_lookup(&state->symlink_changes, path);
3746         if (!ent) {
3747                 ent = string_list_insert(&state->symlink_changes, path);
3748                 ent->util = (void *)0;
3749         }
3750         ent->util = (void *)(what | ((uintptr_t)ent->util));
3751         return (uintptr_t)ent->util;
3752 }
3753
3754 static uintptr_t check_symlink_changes(struct apply_state *state, const char *path)
3755 {
3756         struct string_list_item *ent;
3757
3758         ent = string_list_lookup(&state->symlink_changes, path);
3759         if (!ent)
3760                 return 0;
3761         return (uintptr_t)ent->util;
3762 }
3763
3764 static void prepare_symlink_changes(struct apply_state *state, struct patch *patch)
3765 {
3766         for ( ; patch; patch = patch->next) {
3767                 if ((patch->old_name && S_ISLNK(patch->old_mode)) &&
3768                     (patch->is_rename || patch->is_delete))
3769                         /* the symlink at patch->old_name is removed */
3770                         register_symlink_changes(state, patch->old_name, APPLY_SYMLINK_GOES_AWAY);
3771
3772                 if (patch->new_name && S_ISLNK(patch->new_mode))
3773                         /* the symlink at patch->new_name is created or remains */
3774                         register_symlink_changes(state, patch->new_name, APPLY_SYMLINK_IN_RESULT);
3775         }
3776 }
3777
3778 static int path_is_beyond_symlink_1(struct apply_state *state, struct strbuf *name)
3779 {
3780         do {
3781                 unsigned int change;
3782
3783                 while (--name->len && name->buf[name->len] != '/')
3784                         ; /* scan backwards */
3785                 if (!name->len)
3786                         break;
3787                 name->buf[name->len] = '\0';
3788                 change = check_symlink_changes(state, name->buf);
3789                 if (change & APPLY_SYMLINK_IN_RESULT)
3790                         return 1;
3791                 if (change & APPLY_SYMLINK_GOES_AWAY)
3792                         /*
3793                          * This cannot be "return 0", because we may
3794                          * see a new one created at a higher level.
3795                          */
3796                         continue;
3797
3798                 /* otherwise, check the preimage */
3799                 if (state->check_index) {
3800                         struct cache_entry *ce;
3801
3802                         ce = cache_file_exists(name->buf, name->len, ignore_case);
3803                         if (ce && S_ISLNK(ce->ce_mode))
3804                                 return 1;
3805                 } else {
3806                         struct stat st;
3807                         if (!lstat(name->buf, &st) && S_ISLNK(st.st_mode))
3808                                 return 1;
3809                 }
3810         } while (1);
3811         return 0;
3812 }
3813
3814 static int path_is_beyond_symlink(struct apply_state *state, const char *name_)
3815 {
3816         int ret;
3817         struct strbuf name = STRBUF_INIT;
3818
3819         assert(*name_ != '\0');
3820         strbuf_addstr(&name, name_);
3821         ret = path_is_beyond_symlink_1(state, &name);
3822         strbuf_release(&name);
3823
3824         return ret;
3825 }
3826
3827 static int check_unsafe_path(struct patch *patch)
3828 {
3829         const char *old_name = NULL;
3830         const char *new_name = NULL;
3831         if (patch->is_delete)
3832                 old_name = patch->old_name;
3833         else if (!patch->is_new && !patch->is_copy)
3834                 old_name = patch->old_name;
3835         if (!patch->is_delete)
3836                 new_name = patch->new_name;
3837
3838         if (old_name && !verify_path(old_name))
3839                 return error(_("invalid path '%s'"), old_name);
3840         if (new_name && !verify_path(new_name))
3841                 return error(_("invalid path '%s'"), new_name);
3842         return 0;
3843 }
3844
3845 /*
3846  * Check and apply the patch in-core; leave the result in patch->result
3847  * for the caller to write it out to the final destination.
3848  */
3849 static int check_patch(struct apply_state *state, struct patch *patch)
3850 {
3851         struct stat st;
3852         const char *old_name = patch->old_name;
3853         const char *new_name = patch->new_name;
3854         const char *name = old_name ? old_name : new_name;
3855         struct cache_entry *ce = NULL;
3856         struct patch *tpatch;
3857         int ok_if_exists;
3858         int status;
3859
3860         patch->rejected = 1; /* we will drop this after we succeed */
3861
3862         status = check_preimage(state, patch, &ce, &st);
3863         if (status)
3864                 return status;
3865         old_name = patch->old_name;
3866
3867         /*
3868          * A type-change diff is always split into a patch to delete
3869          * old, immediately followed by a patch to create new (see
3870          * diff.c::run_diff()); in such a case it is Ok that the entry
3871          * to be deleted by the previous patch is still in the working
3872          * tree and in the index.
3873          *
3874          * A patch to swap-rename between A and B would first rename A
3875          * to B and then rename B to A.  While applying the first one,
3876          * the presence of B should not stop A from getting renamed to
3877          * B; ask to_be_deleted() about the later rename.  Removal of
3878          * B and rename from A to B is handled the same way by asking
3879          * was_deleted().
3880          */
3881         if ((tpatch = in_fn_table(state, new_name)) &&
3882             (was_deleted(tpatch) || to_be_deleted(tpatch)))
3883                 ok_if_exists = 1;
3884         else
3885                 ok_if_exists = 0;
3886
3887         if (new_name &&
3888             ((0 < patch->is_new) || patch->is_rename || patch->is_copy)) {
3889                 int err = check_to_create(state, new_name, ok_if_exists);
3890
3891                 if (err && state->threeway) {
3892                         patch->direct_to_threeway = 1;
3893                 } else switch (err) {
3894                 case 0:
3895                         break; /* happy */
3896                 case EXISTS_IN_INDEX:
3897                         return error(_("%s: already exists in index"), new_name);
3898                         break;
3899                 case EXISTS_IN_WORKTREE:
3900                         return error(_("%s: already exists in working directory"),
3901                                      new_name);
3902                 default:
3903                         return err;
3904                 }
3905
3906                 if (!patch->new_mode) {
3907                         if (0 < patch->is_new)
3908                                 patch->new_mode = S_IFREG | 0644;
3909                         else
3910                                 patch->new_mode = patch->old_mode;
3911                 }
3912         }
3913
3914         if (new_name && old_name) {
3915                 int same = !strcmp(old_name, new_name);
3916                 if (!patch->new_mode)
3917                         patch->new_mode = patch->old_mode;
3918                 if ((patch->old_mode ^ patch->new_mode) & S_IFMT) {
3919                         if (same)
3920                                 return error(_("new mode (%o) of %s does not "
3921                                                "match old mode (%o)"),
3922                                         patch->new_mode, new_name,
3923                                         patch->old_mode);
3924                         else
3925                                 return error(_("new mode (%o) of %s does not "
3926                                                "match old mode (%o) of %s"),
3927                                         patch->new_mode, new_name,
3928                                         patch->old_mode, old_name);
3929                 }
3930         }
3931
3932         if (!state->unsafe_paths && check_unsafe_path(patch))
3933                 return -128;
3934
3935         /*
3936          * An attempt to read from or delete a path that is beyond a
3937          * symbolic link will be prevented by load_patch_target() that
3938          * is called at the beginning of apply_data() so we do not
3939          * have to worry about a patch marked with "is_delete" bit
3940          * here.  We however need to make sure that the patch result
3941          * is not deposited to a path that is beyond a symbolic link
3942          * here.
3943          */
3944         if (!patch->is_delete && path_is_beyond_symlink(state, patch->new_name))
3945                 return error(_("affected file '%s' is beyond a symbolic link"),
3946                              patch->new_name);
3947
3948         if (apply_data(state, patch, &st, ce) < 0)
3949                 return error(_("%s: patch does not apply"), name);
3950         patch->rejected = 0;
3951         return 0;
3952 }
3953
3954 static int check_patch_list(struct apply_state *state, struct patch *patch)
3955 {
3956         int err = 0;
3957
3958         prepare_symlink_changes(state, patch);
3959         prepare_fn_table(state, patch);
3960         while (patch) {
3961                 int res;
3962                 if (state->apply_verbosity > verbosity_normal)
3963                         say_patch_name(stderr,
3964                                        _("Checking patch %s..."), patch);
3965                 res = check_patch(state, patch);
3966                 if (res == -128)
3967                         return -128;
3968                 err |= res;
3969                 patch = patch->next;
3970         }
3971         return err;
3972 }
3973
3974 static int read_apply_cache(struct apply_state *state)
3975 {
3976         if (state->index_file)
3977                 return read_cache_from(state->index_file);
3978         else
3979                 return read_cache();
3980 }
3981
3982 /* This function tries to read the object name from the current index */
3983 static int get_current_oid(struct apply_state *state, const char *path,
3984                            struct object_id *oid)
3985 {
3986         int pos;
3987
3988         if (read_apply_cache(state) < 0)
3989                 return -1;
3990         pos = cache_name_pos(path, strlen(path));
3991         if (pos < 0)
3992                 return -1;
3993         oidcpy(oid, &active_cache[pos]->oid);
3994         return 0;
3995 }
3996
3997 static int preimage_oid_in_gitlink_patch(struct patch *p, struct object_id *oid)
3998 {
3999         /*
4000          * A usable gitlink patch has only one fragment (hunk) that looks like:
4001          * @@ -1 +1 @@
4002          * -Subproject commit <old sha1>
4003          * +Subproject commit <new sha1>
4004          * or
4005          * @@ -1 +0,0 @@
4006          * -Subproject commit <old sha1>
4007          * for a removal patch.
4008          */
4009         struct fragment *hunk = p->fragments;
4010         static const char heading[] = "-Subproject commit ";
4011         char *preimage;
4012
4013         if (/* does the patch have only one hunk? */
4014             hunk && !hunk->next &&
4015             /* is its preimage one line? */
4016             hunk->oldpos == 1 && hunk->oldlines == 1 &&
4017             /* does preimage begin with the heading? */
4018             (preimage = memchr(hunk->patch, '\n', hunk->size)) != NULL &&
4019             starts_with(++preimage, heading) &&
4020             /* does it record full SHA-1? */
4021             !get_oid_hex(preimage + sizeof(heading) - 1, oid) &&
4022             preimage[sizeof(heading) + GIT_SHA1_HEXSZ - 1] == '\n' &&
4023             /* does the abbreviated name on the index line agree with it? */
4024             starts_with(preimage + sizeof(heading) - 1, p->old_sha1_prefix))
4025                 return 0; /* it all looks fine */
4026
4027         /* we may have full object name on the index line */
4028         return get_oid_hex(p->old_sha1_prefix, oid);
4029 }
4030
4031 /* Build an index that contains the just the files needed for a 3way merge */
4032 static int build_fake_ancestor(struct apply_state *state, struct patch *list)
4033 {
4034         struct patch *patch;
4035         struct index_state result = { NULL };
4036         static struct lock_file lock;
4037         int res;
4038
4039         /* Once we start supporting the reverse patch, it may be
4040          * worth showing the new sha1 prefix, but until then...
4041          */
4042         for (patch = list; patch; patch = patch->next) {
4043                 struct object_id oid;
4044                 struct cache_entry *ce;
4045                 const char *name;
4046
4047                 name = patch->old_name ? patch->old_name : patch->new_name;
4048                 if (0 < patch->is_new)
4049                         continue;
4050
4051                 if (S_ISGITLINK(patch->old_mode)) {
4052                         if (!preimage_oid_in_gitlink_patch(patch, &oid))
4053                                 ; /* ok, the textual part looks sane */
4054                         else
4055                                 return error(_("sha1 information is lacking or "
4056                                                "useless for submodule %s"), name);
4057                 } else if (!get_sha1_blob(patch->old_sha1_prefix, oid.hash)) {
4058                         ; /* ok */
4059                 } else if (!patch->lines_added && !patch->lines_deleted) {
4060                         /* mode-only change: update the current */
4061                         if (get_current_oid(state, patch->old_name, &oid))
4062                                 return error(_("mode change for %s, which is not "
4063                                                "in current HEAD"), name);
4064                 } else
4065                         return error(_("sha1 information is lacking or useless "
4066                                        "(%s)."), name);
4067
4068                 ce = make_cache_entry(patch->old_mode, oid.hash, name, 0, 0);
4069                 if (!ce)
4070                         return error(_("make_cache_entry failed for path '%s'"),
4071                                      name);
4072                 if (add_index_entry(&result, ce, ADD_CACHE_OK_TO_ADD)) {
4073                         free(ce);
4074                         return error(_("could not add %s to temporary index"),
4075                                      name);
4076                 }
4077         }
4078
4079         hold_lock_file_for_update(&lock, state->fake_ancestor, LOCK_DIE_ON_ERROR);
4080         res = write_locked_index(&result, &lock, COMMIT_LOCK);
4081         discard_index(&result);
4082
4083         if (res)
4084                 return error(_("could not write temporary index to %s"),
4085                              state->fake_ancestor);
4086
4087         return 0;
4088 }
4089
4090 static void stat_patch_list(struct apply_state *state, struct patch *patch)
4091 {
4092         int files, adds, dels;
4093
4094         for (files = adds = dels = 0 ; patch ; patch = patch->next) {
4095                 files++;
4096                 adds += patch->lines_added;
4097                 dels += patch->lines_deleted;
4098                 show_stats(state, patch);
4099         }
4100
4101         print_stat_summary(stdout, files, adds, dels);
4102 }
4103
4104 static void numstat_patch_list(struct apply_state *state,
4105                                struct patch *patch)
4106 {
4107         for ( ; patch; patch = patch->next) {
4108                 const char *name;
4109                 name = patch->new_name ? patch->new_name : patch->old_name;
4110                 if (patch->is_binary)
4111                         printf("-\t-\t");
4112                 else
4113                         printf("%d\t%d\t", patch->lines_added, patch->lines_deleted);
4114                 write_name_quoted(name, stdout, state->line_termination);
4115         }
4116 }
4117
4118 static void show_file_mode_name(const char *newdelete, unsigned int mode, const char *name)
4119 {
4120         if (mode)
4121                 printf(" %s mode %06o %s\n", newdelete, mode, name);
4122         else
4123                 printf(" %s %s\n", newdelete, name);
4124 }
4125
4126 static void show_mode_change(struct patch *p, int show_name)
4127 {
4128         if (p->old_mode && p->new_mode && p->old_mode != p->new_mode) {
4129                 if (show_name)
4130                         printf(" mode change %06o => %06o %s\n",
4131                                p->old_mode, p->new_mode, p->new_name);
4132                 else
4133                         printf(" mode change %06o => %06o\n",
4134                                p->old_mode, p->new_mode);
4135         }
4136 }
4137
4138 static void show_rename_copy(struct patch *p)
4139 {
4140         const char *renamecopy = p->is_rename ? "rename" : "copy";
4141         const char *old, *new;
4142
4143         /* Find common prefix */
4144         old = p->old_name;
4145         new = p->new_name;
4146         while (1) {
4147                 const char *slash_old, *slash_new;
4148                 slash_old = strchr(old, '/');
4149                 slash_new = strchr(new, '/');
4150                 if (!slash_old ||
4151                     !slash_new ||
4152                     slash_old - old != slash_new - new ||
4153                     memcmp(old, new, slash_new - new))
4154                         break;
4155                 old = slash_old + 1;
4156                 new = slash_new + 1;
4157         }
4158         /* p->old_name thru old is the common prefix, and old and new
4159          * through the end of names are renames
4160          */
4161         if (old != p->old_name)
4162                 printf(" %s %.*s{%s => %s} (%d%%)\n", renamecopy,
4163                        (int)(old - p->old_name), p->old_name,
4164                        old, new, p->score);
4165         else
4166                 printf(" %s %s => %s (%d%%)\n", renamecopy,
4167                        p->old_name, p->new_name, p->score);
4168         show_mode_change(p, 0);
4169 }
4170
4171 static void summary_patch_list(struct patch *patch)
4172 {
4173         struct patch *p;
4174
4175         for (p = patch; p; p = p->next) {
4176                 if (p->is_new)
4177                         show_file_mode_name("create", p->new_mode, p->new_name);
4178                 else if (p->is_delete)
4179                         show_file_mode_name("delete", p->old_mode, p->old_name);
4180                 else {
4181                         if (p->is_rename || p->is_copy)
4182                                 show_rename_copy(p);
4183                         else {
4184                                 if (p->score) {
4185                                         printf(" rewrite %s (%d%%)\n",
4186                                                p->new_name, p->score);
4187                                         show_mode_change(p, 0);
4188                                 }
4189                                 else
4190                                         show_mode_change(p, 1);
4191                         }
4192                 }
4193         }
4194 }
4195
4196 static void patch_stats(struct apply_state *state, struct patch *patch)
4197 {
4198         int lines = patch->lines_added + patch->lines_deleted;
4199
4200         if (lines > state->max_change)
4201                 state->max_change = lines;
4202         if (patch->old_name) {
4203                 int len = quote_c_style(patch->old_name, NULL, NULL, 0);
4204                 if (!len)
4205                         len = strlen(patch->old_name);
4206                 if (len > state->max_len)
4207                         state->max_len = len;
4208         }
4209         if (patch->new_name) {
4210                 int len = quote_c_style(patch->new_name, NULL, NULL, 0);
4211                 if (!len)
4212                         len = strlen(patch->new_name);
4213                 if (len > state->max_len)
4214                         state->max_len = len;
4215         }
4216 }
4217
4218 static int remove_file(struct apply_state *state, struct patch *patch, int rmdir_empty)
4219 {
4220         if (state->update_index) {
4221                 if (remove_file_from_cache(patch->old_name) < 0)
4222                         return error(_("unable to remove %s from index"), patch->old_name);
4223         }
4224         if (!state->cached) {
4225                 if (!remove_or_warn(patch->old_mode, patch->old_name) && rmdir_empty) {
4226                         remove_path(patch->old_name);
4227                 }
4228         }
4229         return 0;
4230 }
4231
4232 static int add_index_file(struct apply_state *state,
4233                           const char *path,
4234                           unsigned mode,
4235                           void *buf,
4236                           unsigned long size)
4237 {
4238         struct stat st;
4239         struct cache_entry *ce;
4240         int namelen = strlen(path);
4241         unsigned ce_size = cache_entry_size(namelen);
4242
4243         if (!state->update_index)
4244                 return 0;
4245
4246         ce = xcalloc(1, ce_size);
4247         memcpy(ce->name, path, namelen);
4248         ce->ce_mode = create_ce_mode(mode);
4249         ce->ce_flags = create_ce_flags(0);
4250         ce->ce_namelen = namelen;
4251         if (S_ISGITLINK(mode)) {
4252                 const char *s;
4253
4254                 if (!skip_prefix(buf, "Subproject commit ", &s) ||
4255                     get_oid_hex(s, &ce->oid)) {
4256                         free(ce);
4257                        return error(_("corrupt patch for submodule %s"), path);
4258                 }
4259         } else {
4260                 if (!state->cached) {
4261                         if (lstat(path, &st) < 0) {
4262                                 free(ce);
4263                                 return error_errno(_("unable to stat newly "
4264                                                      "created file '%s'"),
4265                                                    path);
4266                         }
4267                         fill_stat_cache_info(ce, &st);
4268                 }
4269                 if (write_sha1_file(buf, size, blob_type, ce->oid.hash) < 0) {
4270                         free(ce);
4271                         return error(_("unable to create backing store "
4272                                        "for newly created file %s"), path);
4273                 }
4274         }
4275         if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD) < 0) {
4276                 free(ce);
4277                 return error(_("unable to add cache entry for %s"), path);
4278         }
4279
4280         return 0;
4281 }
4282
4283 /*
4284  * Returns:
4285  *  -1 if an unrecoverable error happened
4286  *   0 if everything went well
4287  *   1 if a recoverable error happened
4288  */
4289 static int try_create_file(const char *path, unsigned int mode, const char *buf, unsigned long size)
4290 {
4291         int fd, res;
4292         struct strbuf nbuf = STRBUF_INIT;
4293
4294         if (S_ISGITLINK(mode)) {
4295                 struct stat st;
4296                 if (!lstat(path, &st) && S_ISDIR(st.st_mode))
4297                         return 0;
4298                 return !!mkdir(path, 0777);
4299         }
4300
4301         if (has_symlinks && S_ISLNK(mode))
4302                 /* Although buf:size is counted string, it also is NUL
4303                  * terminated.
4304                  */
4305                 return !!symlink(buf, path);
4306
4307         fd = open(path, O_CREAT | O_EXCL | O_WRONLY, (mode & 0100) ? 0777 : 0666);
4308         if (fd < 0)
4309                 return 1;
4310
4311         if (convert_to_working_tree(path, buf, size, &nbuf)) {
4312                 size = nbuf.len;
4313                 buf  = nbuf.buf;
4314         }
4315
4316         res = write_in_full(fd, buf, size) < 0;
4317         if (res)
4318                 error_errno(_("failed to write to '%s'"), path);
4319         strbuf_release(&nbuf);
4320
4321         if (close(fd) < 0 && !res)
4322                 return error_errno(_("closing file '%s'"), path);
4323
4324         return res ? -1 : 0;
4325 }
4326
4327 /*
4328  * We optimistically assume that the directories exist,
4329  * which is true 99% of the time anyway. If they don't,
4330  * we create them and try again.
4331  *
4332  * Returns:
4333  *   -1 on error
4334  *   0 otherwise
4335  */
4336 static int create_one_file(struct apply_state *state,
4337                            char *path,
4338                            unsigned mode,
4339                            const char *buf,
4340                            unsigned long size)
4341 {
4342         int res;
4343
4344         if (state->cached)
4345                 return 0;
4346
4347         res = try_create_file(path, mode, buf, size);
4348         if (res < 0)
4349                 return -1;
4350         if (!res)
4351                 return 0;
4352
4353         if (errno == ENOENT) {
4354                 if (safe_create_leading_directories(path))
4355                         return 0;
4356                 res = try_create_file(path, mode, buf, size);
4357                 if (res < 0)
4358                         return -1;
4359                 if (!res)
4360                         return 0;
4361         }
4362
4363         if (errno == EEXIST || errno == EACCES) {
4364                 /* We may be trying to create a file where a directory
4365                  * used to be.
4366                  */
4367                 struct stat st;
4368                 if (!lstat(path, &st) && (!S_ISDIR(st.st_mode) || !rmdir(path)))
4369                         errno = EEXIST;
4370         }
4371
4372         if (errno == EEXIST) {
4373                 unsigned int nr = getpid();
4374
4375                 for (;;) {
4376                         char newpath[PATH_MAX];
4377                         mksnpath(newpath, sizeof(newpath), "%s~%u", path, nr);
4378                         res = try_create_file(newpath, mode, buf, size);
4379                         if (res < 0)
4380                                 return -1;
4381                         if (!res) {
4382                                 if (!rename(newpath, path))
4383                                         return 0;
4384                                 unlink_or_warn(newpath);
4385                                 break;
4386                         }
4387                         if (errno != EEXIST)
4388                                 break;
4389                         ++nr;
4390                 }
4391         }
4392         return error_errno(_("unable to write file '%s' mode %o"),
4393                            path, mode);
4394 }
4395
4396 static int add_conflicted_stages_file(struct apply_state *state,
4397                                        struct patch *patch)
4398 {
4399         int stage, namelen;
4400         unsigned ce_size, mode;
4401         struct cache_entry *ce;
4402
4403         if (!state->update_index)
4404                 return 0;
4405         namelen = strlen(patch->new_name);
4406         ce_size = cache_entry_size(namelen);
4407         mode = patch->new_mode ? patch->new_mode : (S_IFREG | 0644);
4408
4409         remove_file_from_cache(patch->new_name);
4410         for (stage = 1; stage < 4; stage++) {
4411                 if (is_null_oid(&patch->threeway_stage[stage - 1]))
4412                         continue;
4413                 ce = xcalloc(1, ce_size);
4414                 memcpy(ce->name, patch->new_name, namelen);
4415                 ce->ce_mode = create_ce_mode(mode);
4416                 ce->ce_flags = create_ce_flags(stage);
4417                 ce->ce_namelen = namelen;
4418                 oidcpy(&ce->oid, &patch->threeway_stage[stage - 1]);
4419                 if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD) < 0) {
4420                         free(ce);
4421                         return error(_("unable to add cache entry for %s"),
4422                                      patch->new_name);
4423                 }
4424         }
4425
4426         return 0;
4427 }
4428
4429 static int create_file(struct apply_state *state, struct patch *patch)
4430 {
4431         char *path = patch->new_name;
4432         unsigned mode = patch->new_mode;
4433         unsigned long size = patch->resultsize;
4434         char *buf = patch->result;
4435
4436         if (!mode)
4437                 mode = S_IFREG | 0644;
4438         if (create_one_file(state, path, mode, buf, size))
4439                 return -1;
4440
4441         if (patch->conflicted_threeway)
4442                 return add_conflicted_stages_file(state, patch);
4443         else
4444                 return add_index_file(state, path, mode, buf, size);
4445 }
4446
4447 /* phase zero is to remove, phase one is to create */
4448 static int write_out_one_result(struct apply_state *state,
4449                                 struct patch *patch,
4450                                 int phase)
4451 {
4452         if (patch->is_delete > 0) {
4453                 if (phase == 0)
4454                         return remove_file(state, patch, 1);
4455                 return 0;
4456         }
4457         if (patch->is_new > 0 || patch->is_copy) {
4458                 if (phase == 1)
4459                         return create_file(state, patch);
4460                 return 0;
4461         }
4462         /*
4463          * Rename or modification boils down to the same
4464          * thing: remove the old, write the new
4465          */
4466         if (phase == 0)
4467                 return remove_file(state, patch, patch->is_rename);
4468         if (phase == 1)
4469                 return create_file(state, patch);
4470         return 0;
4471 }
4472
4473 static int write_out_one_reject(struct apply_state *state, struct patch *patch)
4474 {
4475         FILE *rej;
4476         char namebuf[PATH_MAX];
4477         struct fragment *frag;
4478         int cnt = 0;
4479         struct strbuf sb = STRBUF_INIT;
4480
4481         for (cnt = 0, frag = patch->fragments; frag; frag = frag->next) {
4482                 if (!frag->rejected)
4483                         continue;
4484                 cnt++;
4485         }
4486
4487         if (!cnt) {
4488                 if (state->apply_verbosity > verbosity_normal)
4489                         say_patch_name(stderr,
4490                                        _("Applied patch %s cleanly."), patch);
4491                 return 0;
4492         }
4493
4494         /* This should not happen, because a removal patch that leaves
4495          * contents are marked "rejected" at the patch level.
4496          */
4497         if (!patch->new_name)
4498                 die(_("internal error"));
4499
4500         /* Say this even without --verbose */
4501         strbuf_addf(&sb, Q_("Applying patch %%s with %d reject...",
4502                             "Applying patch %%s with %d rejects...",
4503                             cnt),
4504                     cnt);
4505         if (state->apply_verbosity > verbosity_silent)
4506                 say_patch_name(stderr, sb.buf, patch);
4507         strbuf_release(&sb);
4508
4509         cnt = strlen(patch->new_name);
4510         if (ARRAY_SIZE(namebuf) <= cnt + 5) {
4511                 cnt = ARRAY_SIZE(namebuf) - 5;
4512                 warning(_("truncating .rej filename to %.*s.rej"),
4513                         cnt - 1, patch->new_name);
4514         }
4515         memcpy(namebuf, patch->new_name, cnt);
4516         memcpy(namebuf + cnt, ".rej", 5);
4517
4518         rej = fopen(namebuf, "w");
4519         if (!rej)
4520                 return error_errno(_("cannot open %s"), namebuf);
4521
4522         /* Normal git tools never deal with .rej, so do not pretend
4523          * this is a git patch by saying --git or giving extended
4524          * headers.  While at it, maybe please "kompare" that wants
4525          * the trailing TAB and some garbage at the end of line ;-).
4526          */
4527         fprintf(rej, "diff a/%s b/%s\t(rejected hunks)\n",
4528                 patch->new_name, patch->new_name);
4529         for (cnt = 1, frag = patch->fragments;
4530              frag;
4531              cnt++, frag = frag->next) {
4532                 if (!frag->rejected) {
4533                         if (state->apply_verbosity > verbosity_silent)
4534                                 fprintf_ln(stderr, _("Hunk #%d applied cleanly."), cnt);
4535                         continue;
4536                 }
4537                 if (state->apply_verbosity > verbosity_silent)
4538                         fprintf_ln(stderr, _("Rejected hunk #%d."), cnt);
4539                 fprintf(rej, "%.*s", frag->size, frag->patch);
4540                 if (frag->patch[frag->size-1] != '\n')
4541                         fputc('\n', rej);
4542         }
4543         fclose(rej);
4544         return -1;
4545 }
4546
4547 /*
4548  * Returns:
4549  *  -1 if an error happened
4550  *   0 if the patch applied cleanly
4551  *   1 if the patch did not apply cleanly
4552  */
4553 static int write_out_results(struct apply_state *state, struct patch *list)
4554 {
4555         int phase;
4556         int errs = 0;
4557         struct patch *l;
4558         struct string_list cpath = STRING_LIST_INIT_DUP;
4559
4560         for (phase = 0; phase < 2; phase++) {
4561                 l = list;
4562                 while (l) {
4563                         if (l->rejected)
4564                                 errs = 1;
4565                         else {
4566                                 if (write_out_one_result(state, l, phase)) {
4567                                         string_list_clear(&cpath, 0);
4568                                         return -1;
4569                                 }
4570                                 if (phase == 1) {
4571                                         if (write_out_one_reject(state, l))
4572                                                 errs = 1;
4573                                         if (l->conflicted_threeway) {
4574                                                 string_list_append(&cpath, l->new_name);
4575                                                 errs = 1;
4576                                         }
4577                                 }
4578                         }
4579                         l = l->next;
4580                 }
4581         }
4582
4583         if (cpath.nr) {
4584                 struct string_list_item *item;
4585
4586                 string_list_sort(&cpath);
4587                 if (state->apply_verbosity > verbosity_silent) {
4588                         for_each_string_list_item(item, &cpath)
4589                                 fprintf(stderr, "U %s\n", item->string);
4590                 }
4591                 string_list_clear(&cpath, 0);
4592
4593                 rerere(0);
4594         }
4595
4596         return errs;
4597 }
4598
4599 /*
4600  * Try to apply a patch.
4601  *
4602  * Returns:
4603  *  -128 if a bad error happened (like patch unreadable)
4604  *  -1 if patch did not apply and user cannot deal with it
4605  *   0 if the patch applied
4606  *   1 if the patch did not apply but user might fix it
4607  */
4608 static int apply_patch(struct apply_state *state,
4609                        int fd,
4610                        const char *filename,
4611                        int options)
4612 {
4613         size_t offset;
4614         struct strbuf buf = STRBUF_INIT; /* owns the patch text */
4615         struct patch *list = NULL, **listp = &list;
4616         int skipped_patch = 0;
4617         int res = 0;
4618
4619         state->patch_input_file = filename;
4620         if (read_patch_file(&buf, fd) < 0)
4621                 return -128;
4622         offset = 0;
4623         while (offset < buf.len) {
4624                 struct patch *patch;
4625                 int nr;
4626
4627                 patch = xcalloc(1, sizeof(*patch));
4628                 patch->inaccurate_eof = !!(options & APPLY_OPT_INACCURATE_EOF);
4629                 patch->recount =  !!(options & APPLY_OPT_RECOUNT);
4630                 nr = parse_chunk(state, buf.buf + offset, buf.len - offset, patch);
4631                 if (nr < 0) {
4632                         free_patch(patch);
4633                         if (nr == -128) {
4634                                 res = -128;
4635                                 goto end;
4636                         }
4637                         break;
4638                 }
4639                 if (state->apply_in_reverse)
4640                         reverse_patches(patch);
4641                 if (use_patch(state, patch)) {
4642                         patch_stats(state, patch);
4643                         *listp = patch;
4644                         listp = &patch->next;
4645                 }
4646                 else {
4647                         if (state->apply_verbosity > verbosity_normal)
4648                                 say_patch_name(stderr, _("Skipped patch '%s'."), patch);
4649                         free_patch(patch);
4650                         skipped_patch++;
4651                 }
4652                 offset += nr;
4653         }
4654
4655         if (!list && !skipped_patch) {
4656                 error(_("unrecognized input"));
4657                 res = -128;
4658                 goto end;
4659         }
4660
4661         if (state->whitespace_error && (state->ws_error_action == die_on_ws_error))
4662                 state->apply = 0;
4663
4664         state->update_index = state->check_index && state->apply;
4665         if (state->update_index && state->newfd < 0) {
4666                 if (state->index_file)
4667                         state->newfd = hold_lock_file_for_update(state->lock_file,
4668                                                                  state->index_file,
4669                                                                  LOCK_DIE_ON_ERROR);
4670                 else
4671                         state->newfd = hold_locked_index(state->lock_file, LOCK_DIE_ON_ERROR);
4672         }
4673
4674         if (state->check_index && read_apply_cache(state) < 0) {
4675                 error(_("unable to read index file"));
4676                 res = -128;
4677                 goto end;
4678         }
4679
4680         if (state->check || state->apply) {
4681                 int r = check_patch_list(state, list);
4682                 if (r == -128) {
4683                         res = -128;
4684                         goto end;
4685                 }
4686                 if (r < 0 && !state->apply_with_reject) {
4687                         res = -1;
4688                         goto end;
4689                 }
4690         }
4691
4692         if (state->apply) {
4693                 int write_res = write_out_results(state, list);
4694                 if (write_res < 0) {
4695                         res = -128;
4696                         goto end;
4697                 }
4698                 if (write_res > 0) {
4699                         /* with --3way, we still need to write the index out */
4700                         res = state->apply_with_reject ? -1 : 1;
4701                         goto end;
4702                 }
4703         }
4704
4705         if (state->fake_ancestor &&
4706             build_fake_ancestor(state, list)) {
4707                 res = -128;
4708                 goto end;
4709         }
4710
4711         if (state->diffstat && state->apply_verbosity > verbosity_silent)
4712                 stat_patch_list(state, list);
4713
4714         if (state->numstat && state->apply_verbosity > verbosity_silent)
4715                 numstat_patch_list(state, list);
4716
4717         if (state->summary && state->apply_verbosity > verbosity_silent)
4718                 summary_patch_list(list);
4719
4720 end:
4721         free_patch_list(list);
4722         strbuf_release(&buf);
4723         string_list_clear(&state->fn_table, 0);
4724         return res;
4725 }
4726
4727 static int apply_option_parse_exclude(const struct option *opt,
4728                                       const char *arg, int unset)
4729 {
4730         struct apply_state *state = opt->value;
4731         add_name_limit(state, arg, 1);
4732         return 0;
4733 }
4734
4735 static int apply_option_parse_include(const struct option *opt,
4736                                       const char *arg, int unset)
4737 {
4738         struct apply_state *state = opt->value;
4739         add_name_limit(state, arg, 0);
4740         state->has_include = 1;
4741         return 0;
4742 }
4743
4744 static int apply_option_parse_p(const struct option *opt,
4745                                 const char *arg,
4746                                 int unset)
4747 {
4748         struct apply_state *state = opt->value;
4749         state->p_value = atoi(arg);
4750         state->p_value_known = 1;
4751         return 0;
4752 }
4753
4754 static int apply_option_parse_space_change(const struct option *opt,
4755                                            const char *arg, int unset)
4756 {
4757         struct apply_state *state = opt->value;
4758         if (unset)
4759                 state->ws_ignore_action = ignore_ws_none;
4760         else
4761                 state->ws_ignore_action = ignore_ws_change;
4762         return 0;
4763 }
4764
4765 static int apply_option_parse_whitespace(const struct option *opt,
4766                                          const char *arg, int unset)
4767 {
4768         struct apply_state *state = opt->value;
4769         state->whitespace_option = arg;
4770         if (parse_whitespace_option(state, arg))
4771                 exit(1);
4772         return 0;
4773 }
4774
4775 static int apply_option_parse_directory(const struct option *opt,
4776                                         const char *arg, int unset)
4777 {
4778         struct apply_state *state = opt->value;
4779         strbuf_reset(&state->root);
4780         strbuf_addstr(&state->root, arg);
4781         strbuf_complete(&state->root, '/');
4782         return 0;
4783 }
4784
4785 int apply_all_patches(struct apply_state *state,
4786                       int argc,
4787                       const char **argv,
4788                       int options)
4789 {
4790         int i;
4791         int res;
4792         int errs = 0;
4793         int read_stdin = 1;
4794
4795         for (i = 0; i < argc; i++) {
4796                 const char *arg = argv[i];
4797                 char *to_free = NULL;
4798                 int fd;
4799
4800                 if (!strcmp(arg, "-")) {
4801                         res = apply_patch(state, 0, "<stdin>", options);
4802                         if (res < 0)
4803                                 goto end;
4804                         errs |= res;
4805                         read_stdin = 0;
4806                         continue;
4807                 } else
4808                         arg = to_free = prefix_filename(state->prefix, arg);
4809
4810                 fd = open(arg, O_RDONLY);
4811                 if (fd < 0) {
4812                         error(_("can't open patch '%s': %s"), arg, strerror(errno));
4813                         res = -128;
4814                         free(to_free);
4815                         goto end;
4816                 }
4817                 read_stdin = 0;
4818                 set_default_whitespace_mode(state);
4819                 res = apply_patch(state, fd, arg, options);
4820                 close(fd);
4821                 free(to_free);
4822                 if (res < 0)
4823                         goto end;
4824                 errs |= res;
4825         }
4826         set_default_whitespace_mode(state);
4827         if (read_stdin) {
4828                 res = apply_patch(state, 0, "<stdin>", options);
4829                 if (res < 0)
4830                         goto end;
4831                 errs |= res;
4832         }
4833
4834         if (state->whitespace_error) {
4835                 if (state->squelch_whitespace_errors &&
4836                     state->squelch_whitespace_errors < state->whitespace_error) {
4837                         int squelched =
4838                                 state->whitespace_error - state->squelch_whitespace_errors;
4839                         warning(Q_("squelched %d whitespace error",
4840                                    "squelched %d whitespace errors",
4841                                    squelched),
4842                                 squelched);
4843                 }
4844                 if (state->ws_error_action == die_on_ws_error) {
4845                         error(Q_("%d line adds whitespace errors.",
4846                                  "%d lines add whitespace errors.",
4847                                  state->whitespace_error),
4848                               state->whitespace_error);
4849                         res = -128;
4850                         goto end;
4851                 }
4852                 if (state->applied_after_fixing_ws && state->apply)
4853                         warning(Q_("%d line applied after"
4854                                    " fixing whitespace errors.",
4855                                    "%d lines applied after"
4856                                    " fixing whitespace errors.",
4857                                    state->applied_after_fixing_ws),
4858                                 state->applied_after_fixing_ws);
4859                 else if (state->whitespace_error)
4860                         warning(Q_("%d line adds whitespace errors.",
4861                                    "%d lines add whitespace errors.",
4862                                    state->whitespace_error),
4863                                 state->whitespace_error);
4864         }
4865
4866         if (state->update_index) {
4867                 res = write_locked_index(&the_index, state->lock_file, COMMIT_LOCK);
4868                 if (res) {
4869                         error(_("Unable to write new index file"));
4870                         res = -128;
4871                         goto end;
4872                 }
4873                 state->newfd = -1;
4874         }
4875
4876         res = !!errs;
4877
4878 end:
4879         if (state->newfd >= 0) {
4880                 rollback_lock_file(state->lock_file);
4881                 state->newfd = -1;
4882         }
4883
4884         if (state->apply_verbosity <= verbosity_silent) {
4885                 set_error_routine(state->saved_error_routine);
4886                 set_warn_routine(state->saved_warn_routine);
4887         }
4888
4889         if (res > -1)
4890                 return res;
4891         return (res == -1 ? 1 : 128);
4892 }
4893
4894 int apply_parse_options(int argc, const char **argv,
4895                         struct apply_state *state,
4896                         int *force_apply, int *options,
4897                         const char * const *apply_usage)
4898 {
4899         struct option builtin_apply_options[] = {
4900                 { OPTION_CALLBACK, 0, "exclude", state, N_("path"),
4901                         N_("don't apply changes matching the given path"),
4902                         0, apply_option_parse_exclude },
4903                 { OPTION_CALLBACK, 0, "include", state, N_("path"),
4904                         N_("apply changes matching the given path"),
4905                         0, apply_option_parse_include },
4906                 { OPTION_CALLBACK, 'p', NULL, state, N_("num"),
4907                         N_("remove <num> leading slashes from traditional diff paths"),
4908                         0, apply_option_parse_p },
4909                 OPT_BOOL(0, "no-add", &state->no_add,
4910                         N_("ignore additions made by the patch")),
4911                 OPT_BOOL(0, "stat", &state->diffstat,
4912                         N_("instead of applying the patch, output diffstat for the input")),
4913                 OPT_NOOP_NOARG(0, "allow-binary-replacement"),
4914                 OPT_NOOP_NOARG(0, "binary"),
4915                 OPT_BOOL(0, "numstat", &state->numstat,
4916                         N_("show number of added and deleted lines in decimal notation")),
4917                 OPT_BOOL(0, "summary", &state->summary,
4918                         N_("instead of applying the patch, output a summary for the input")),
4919                 OPT_BOOL(0, "check", &state->check,
4920                         N_("instead of applying the patch, see if the patch is applicable")),
4921                 OPT_BOOL(0, "index", &state->check_index,
4922                         N_("make sure the patch is applicable to the current index")),
4923                 OPT_BOOL(0, "cached", &state->cached,
4924                         N_("apply a patch without touching the working tree")),
4925                 OPT_BOOL(0, "unsafe-paths", &state->unsafe_paths,
4926                         N_("accept a patch that touches outside the working area")),
4927                 OPT_BOOL(0, "apply", force_apply,
4928                         N_("also apply the patch (use with --stat/--summary/--check)")),
4929                 OPT_BOOL('3', "3way", &state->threeway,
4930                          N_( "attempt three-way merge if a patch does not apply")),
4931                 OPT_FILENAME(0, "build-fake-ancestor", &state->fake_ancestor,
4932                         N_("build a temporary index based on embedded index information")),
4933                 /* Think twice before adding "--nul" synonym to this */
4934                 OPT_SET_INT('z', NULL, &state->line_termination,
4935                         N_("paths are separated with NUL character"), '\0'),
4936                 OPT_INTEGER('C', NULL, &state->p_context,
4937                                 N_("ensure at least <n> lines of context match")),
4938                 { OPTION_CALLBACK, 0, "whitespace", state, N_("action"),
4939                         N_("detect new or modified lines that have whitespace errors"),
4940                         0, apply_option_parse_whitespace },
4941                 { OPTION_CALLBACK, 0, "ignore-space-change", state, NULL,
4942                         N_("ignore changes in whitespace when finding context"),
4943                         PARSE_OPT_NOARG, apply_option_parse_space_change },
4944                 { OPTION_CALLBACK, 0, "ignore-whitespace", state, NULL,
4945                         N_("ignore changes in whitespace when finding context"),
4946                         PARSE_OPT_NOARG, apply_option_parse_space_change },
4947                 OPT_BOOL('R', "reverse", &state->apply_in_reverse,
4948                         N_("apply the patch in reverse")),
4949                 OPT_BOOL(0, "unidiff-zero", &state->unidiff_zero,
4950                         N_("don't expect at least one line of context")),
4951                 OPT_BOOL(0, "reject", &state->apply_with_reject,
4952                         N_("leave the rejected hunks in corresponding *.rej files")),
4953                 OPT_BOOL(0, "allow-overlap", &state->allow_overlap,
4954                         N_("allow overlapping hunks")),
4955                 OPT__VERBOSE(&state->apply_verbosity, N_("be verbose")),
4956                 OPT_BIT(0, "inaccurate-eof", options,
4957                         N_("tolerate incorrectly detected missing new-line at the end of file"),
4958                         APPLY_OPT_INACCURATE_EOF),
4959                 OPT_BIT(0, "recount", options,
4960                         N_("do not trust the line counts in the hunk headers"),
4961                         APPLY_OPT_RECOUNT),
4962                 { OPTION_CALLBACK, 0, "directory", state, N_("root"),
4963                         N_("prepend <root> to all filenames"),
4964                         0, apply_option_parse_directory },
4965                 OPT_END()
4966         };
4967
4968         return parse_options(argc, argv, state->prefix, builtin_apply_options, apply_usage, 0);
4969 }