Merge branch 'jc/diff-sane-truncate-no-more' into next
[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->lock_file = lock_file;
84         state->newfd = -1;
85         state->apply = 1;
86         state->line_termination = '\n';
87         state->p_value = 1;
88         state->p_context = UINT_MAX;
89         state->squelch_whitespace_errors = 5;
90         state->ws_error_action = warn_on_ws_error;
91         state->ws_ignore_action = ignore_ws_none;
92         state->linenr = 1;
93         string_list_init(&state->fn_table, 0);
94         string_list_init(&state->limit_by_name, 0);
95         string_list_init(&state->symlink_changes, 0);
96         strbuf_init(&state->root, 0);
97
98         git_apply_config();
99         if (apply_default_whitespace && parse_whitespace_option(state, apply_default_whitespace))
100                 return -1;
101         if (apply_default_ignorewhitespace && parse_ignorewhitespace_option(state, apply_default_ignorewhitespace))
102                 return -1;
103         return 0;
104 }
105
106 void clear_apply_state(struct apply_state *state)
107 {
108         string_list_clear(&state->limit_by_name, 0);
109         string_list_clear(&state->symlink_changes, 0);
110         strbuf_release(&state->root);
111
112         /* &state->fn_table is cleared at the end of apply_patch() */
113 }
114
115 static void mute_routine(const char *msg, va_list params)
116 {
117         /* do nothing */
118 }
119
120 int check_apply_state(struct apply_state *state, int force_apply)
121 {
122         int is_not_gitdir = !startup_info->have_repository;
123
124         if (state->apply_with_reject && state->threeway)
125                 return error(_("--reject and --3way cannot be used together."));
126         if (state->cached && state->threeway)
127                 return error(_("--cached and --3way cannot be used together."));
128         if (state->threeway) {
129                 if (is_not_gitdir)
130                         return error(_("--3way outside a repository"));
131                 state->check_index = 1;
132         }
133         if (state->apply_with_reject) {
134                 state->apply = 1;
135                 if (state->apply_verbosity == verbosity_normal)
136                         state->apply_verbosity = verbosity_verbose;
137         }
138         if (!force_apply && (state->diffstat || state->numstat || state->summary || state->check || state->fake_ancestor))
139                 state->apply = 0;
140         if (state->check_index && is_not_gitdir)
141                 return error(_("--index outside a repository"));
142         if (state->cached) {
143                 if (is_not_gitdir)
144                         return error(_("--cached outside a repository"));
145                 state->check_index = 1;
146         }
147         if (state->check_index)
148                 state->unsafe_paths = 0;
149         if (!state->lock_file)
150                 return error("BUG: state->lock_file should not be NULL");
151
152         if (state->apply_verbosity <= verbosity_silent) {
153                 state->saved_error_routine = get_error_routine();
154                 state->saved_warn_routine = get_warn_routine();
155                 set_error_routine(mute_routine);
156                 set_warn_routine(mute_routine);
157         }
158
159         return 0;
160 }
161
162 static void set_default_whitespace_mode(struct apply_state *state)
163 {
164         if (!state->whitespace_option && !apply_default_whitespace)
165                 state->ws_error_action = (state->apply ? warn_on_ws_error : nowarn_ws_error);
166 }
167
168 /*
169  * This represents one "hunk" from a patch, starting with
170  * "@@ -oldpos,oldlines +newpos,newlines @@" marker.  The
171  * patch text is pointed at by patch, and its byte length
172  * is stored in size.  leading and trailing are the number
173  * of context lines.
174  */
175 struct fragment {
176         unsigned long leading, trailing;
177         unsigned long oldpos, oldlines;
178         unsigned long newpos, newlines;
179         /*
180          * 'patch' is usually borrowed from buf in apply_patch(),
181          * but some codepaths store an allocated buffer.
182          */
183         const char *patch;
184         unsigned free_patch:1,
185                 rejected:1;
186         int size;
187         int linenr;
188         struct fragment *next;
189 };
190
191 /*
192  * When dealing with a binary patch, we reuse "leading" field
193  * to store the type of the binary hunk, either deflated "delta"
194  * or deflated "literal".
195  */
196 #define binary_patch_method leading
197 #define BINARY_DELTA_DEFLATED   1
198 #define BINARY_LITERAL_DEFLATED 2
199
200 /*
201  * This represents a "patch" to a file, both metainfo changes
202  * such as creation/deletion, filemode and content changes represented
203  * as a series of fragments.
204  */
205 struct patch {
206         char *new_name, *old_name, *def_name;
207         unsigned int old_mode, new_mode;
208         int is_new, is_delete;  /* -1 = unknown, 0 = false, 1 = true */
209         int rejected;
210         unsigned ws_rule;
211         int lines_added, lines_deleted;
212         int score;
213         int extension_linenr; /* first line specifying delete/new/rename/copy */
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 (starts_with(name, state->prefix))
789                         val = count_slashes(state->prefix);
790                 else {
791                         cp++;
792                         if (starts_with(cp, state->prefix))
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                 char *another;
965                 if (isnull)
966                         return error(_("git apply: bad git-diff - expected /dev/null, got %s on line %d"),
967                                      *name, state->linenr);
968                 another = find_name(state, line, NULL, state->p_value, TERM_TAB);
969                 if (!another || strcmp(another, *name)) {
970                         free(another);
971                         return error((side == DIFF_NEW_NAME) ?
972                             _("git apply: bad git-diff - inconsistent new filename on line %d") :
973                             _("git apply: bad git-diff - inconsistent old filename on line %d"), state->linenr);
974                 }
975                 free(another);
976         } else {
977                 if (!starts_with(line, "/dev/null\n"))
978                         return error(_("git apply: bad git-diff - expected /dev/null on line %d"), state->linenr);
979         }
980
981         return 0;
982 }
983
984 static int gitdiff_oldname(struct apply_state *state,
985                            const char *line,
986                            struct patch *patch)
987 {
988         return gitdiff_verify_name(state, line,
989                                    patch->is_new, &patch->old_name,
990                                    DIFF_OLD_NAME);
991 }
992
993 static int gitdiff_newname(struct apply_state *state,
994                            const char *line,
995                            struct patch *patch)
996 {
997         return gitdiff_verify_name(state, line,
998                                    patch->is_delete, &patch->new_name,
999                                    DIFF_NEW_NAME);
1000 }
1001
1002 static int parse_mode_line(const char *line, int linenr, unsigned int *mode)
1003 {
1004         char *end;
1005         *mode = strtoul(line, &end, 8);
1006         if (end == line || !isspace(*end))
1007                 return error(_("invalid mode on line %d: %s"), linenr, line);
1008         return 0;
1009 }
1010
1011 static int gitdiff_oldmode(struct apply_state *state,
1012                            const char *line,
1013                            struct patch *patch)
1014 {
1015         return parse_mode_line(line, state->linenr, &patch->old_mode);
1016 }
1017
1018 static int gitdiff_newmode(struct apply_state *state,
1019                            const char *line,
1020                            struct patch *patch)
1021 {
1022         return parse_mode_line(line, state->linenr, &patch->new_mode);
1023 }
1024
1025 static int gitdiff_delete(struct apply_state *state,
1026                           const char *line,
1027                           struct patch *patch)
1028 {
1029         patch->is_delete = 1;
1030         free(patch->old_name);
1031         patch->old_name = xstrdup_or_null(patch->def_name);
1032         return gitdiff_oldmode(state, line, patch);
1033 }
1034
1035 static int gitdiff_newfile(struct apply_state *state,
1036                            const char *line,
1037                            struct patch *patch)
1038 {
1039         patch->is_new = 1;
1040         free(patch->new_name);
1041         patch->new_name = xstrdup_or_null(patch->def_name);
1042         return gitdiff_newmode(state, line, patch);
1043 }
1044
1045 static int gitdiff_copysrc(struct apply_state *state,
1046                            const char *line,
1047                            struct patch *patch)
1048 {
1049         patch->is_copy = 1;
1050         free(patch->old_name);
1051         patch->old_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1052         return 0;
1053 }
1054
1055 static int gitdiff_copydst(struct apply_state *state,
1056                            const char *line,
1057                            struct patch *patch)
1058 {
1059         patch->is_copy = 1;
1060         free(patch->new_name);
1061         patch->new_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1062         return 0;
1063 }
1064
1065 static int gitdiff_renamesrc(struct apply_state *state,
1066                              const char *line,
1067                              struct patch *patch)
1068 {
1069         patch->is_rename = 1;
1070         free(patch->old_name);
1071         patch->old_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1072         return 0;
1073 }
1074
1075 static int gitdiff_renamedst(struct apply_state *state,
1076                              const char *line,
1077                              struct patch *patch)
1078 {
1079         patch->is_rename = 1;
1080         free(patch->new_name);
1081         patch->new_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1082         return 0;
1083 }
1084
1085 static int gitdiff_similarity(struct apply_state *state,
1086                               const char *line,
1087                               struct patch *patch)
1088 {
1089         unsigned long val = strtoul(line, NULL, 10);
1090         if (val <= 100)
1091                 patch->score = val;
1092         return 0;
1093 }
1094
1095 static int gitdiff_dissimilarity(struct apply_state *state,
1096                                  const char *line,
1097                                  struct patch *patch)
1098 {
1099         unsigned long val = strtoul(line, NULL, 10);
1100         if (val <= 100)
1101                 patch->score = val;
1102         return 0;
1103 }
1104
1105 static int gitdiff_index(struct apply_state *state,
1106                          const char *line,
1107                          struct patch *patch)
1108 {
1109         /*
1110          * index line is N hexadecimal, "..", N hexadecimal,
1111          * and optional space with octal mode.
1112          */
1113         const char *ptr, *eol;
1114         int len;
1115
1116         ptr = strchr(line, '.');
1117         if (!ptr || ptr[1] != '.' || 40 < ptr - line)
1118                 return 0;
1119         len = ptr - line;
1120         memcpy(patch->old_sha1_prefix, line, len);
1121         patch->old_sha1_prefix[len] = 0;
1122
1123         line = ptr + 2;
1124         ptr = strchr(line, ' ');
1125         eol = strchrnul(line, '\n');
1126
1127         if (!ptr || eol < ptr)
1128                 ptr = eol;
1129         len = ptr - line;
1130
1131         if (40 < len)
1132                 return 0;
1133         memcpy(patch->new_sha1_prefix, line, len);
1134         patch->new_sha1_prefix[len] = 0;
1135         if (*ptr == ' ')
1136                 return gitdiff_oldmode(state, ptr + 1, patch);
1137         return 0;
1138 }
1139
1140 /*
1141  * This is normal for a diff that doesn't change anything: we'll fall through
1142  * into the next diff. Tell the parser to break out.
1143  */
1144 static int gitdiff_unrecognized(struct apply_state *state,
1145                                 const char *line,
1146                                 struct patch *patch)
1147 {
1148         return 1;
1149 }
1150
1151 /*
1152  * Skip p_value leading components from "line"; as we do not accept
1153  * absolute paths, return NULL in that case.
1154  */
1155 static const char *skip_tree_prefix(struct apply_state *state,
1156                                     const char *line,
1157                                     int llen)
1158 {
1159         int nslash;
1160         int i;
1161
1162         if (!state->p_value)
1163                 return (llen && line[0] == '/') ? NULL : line;
1164
1165         nslash = state->p_value;
1166         for (i = 0; i < llen; i++) {
1167                 int ch = line[i];
1168                 if (ch == '/' && --nslash <= 0)
1169                         return (i == 0) ? NULL : &line[i + 1];
1170         }
1171         return NULL;
1172 }
1173
1174 /*
1175  * This is to extract the same name that appears on "diff --git"
1176  * line.  We do not find and return anything if it is a rename
1177  * patch, and it is OK because we will find the name elsewhere.
1178  * We need to reliably find name only when it is mode-change only,
1179  * creation or deletion of an empty file.  In any of these cases,
1180  * both sides are the same name under a/ and b/ respectively.
1181  */
1182 static char *git_header_name(struct apply_state *state,
1183                              const char *line,
1184                              int llen)
1185 {
1186         const char *name;
1187         const char *second = NULL;
1188         size_t len, line_len;
1189
1190         line += strlen("diff --git ");
1191         llen -= strlen("diff --git ");
1192
1193         if (*line == '"') {
1194                 const char *cp;
1195                 struct strbuf first = STRBUF_INIT;
1196                 struct strbuf sp = STRBUF_INIT;
1197
1198                 if (unquote_c_style(&first, line, &second))
1199                         goto free_and_fail1;
1200
1201                 /* strip the a/b prefix including trailing slash */
1202                 cp = skip_tree_prefix(state, first.buf, first.len);
1203                 if (!cp)
1204                         goto free_and_fail1;
1205                 strbuf_remove(&first, 0, cp - first.buf);
1206
1207                 /*
1208                  * second points at one past closing dq of name.
1209                  * find the second name.
1210                  */
1211                 while ((second < line + llen) && isspace(*second))
1212                         second++;
1213
1214                 if (line + llen <= second)
1215                         goto free_and_fail1;
1216                 if (*second == '"') {
1217                         if (unquote_c_style(&sp, second, NULL))
1218                                 goto free_and_fail1;
1219                         cp = skip_tree_prefix(state, sp.buf, sp.len);
1220                         if (!cp)
1221                                 goto free_and_fail1;
1222                         /* They must match, otherwise ignore */
1223                         if (strcmp(cp, first.buf))
1224                                 goto free_and_fail1;
1225                         strbuf_release(&sp);
1226                         return strbuf_detach(&first, NULL);
1227                 }
1228
1229                 /* unquoted second */
1230                 cp = skip_tree_prefix(state, second, line + llen - second);
1231                 if (!cp)
1232                         goto free_and_fail1;
1233                 if (line + llen - cp != first.len ||
1234                     memcmp(first.buf, cp, first.len))
1235                         goto free_and_fail1;
1236                 return strbuf_detach(&first, NULL);
1237
1238         free_and_fail1:
1239                 strbuf_release(&first);
1240                 strbuf_release(&sp);
1241                 return NULL;
1242         }
1243
1244         /* unquoted first name */
1245         name = skip_tree_prefix(state, line, llen);
1246         if (!name)
1247                 return NULL;
1248
1249         /*
1250          * since the first name is unquoted, a dq if exists must be
1251          * the beginning of the second name.
1252          */
1253         for (second = name; second < line + llen; second++) {
1254                 if (*second == '"') {
1255                         struct strbuf sp = STRBUF_INIT;
1256                         const char *np;
1257
1258                         if (unquote_c_style(&sp, second, NULL))
1259                                 goto free_and_fail2;
1260
1261                         np = skip_tree_prefix(state, sp.buf, sp.len);
1262                         if (!np)
1263                                 goto free_and_fail2;
1264
1265                         len = sp.buf + sp.len - np;
1266                         if (len < second - name &&
1267                             !strncmp(np, name, len) &&
1268                             isspace(name[len])) {
1269                                 /* Good */
1270                                 strbuf_remove(&sp, 0, np - sp.buf);
1271                                 return strbuf_detach(&sp, NULL);
1272                         }
1273
1274                 free_and_fail2:
1275                         strbuf_release(&sp);
1276                         return NULL;
1277                 }
1278         }
1279
1280         /*
1281          * Accept a name only if it shows up twice, exactly the same
1282          * form.
1283          */
1284         second = strchr(name, '\n');
1285         if (!second)
1286                 return NULL;
1287         line_len = second - name;
1288         for (len = 0 ; ; len++) {
1289                 switch (name[len]) {
1290                 default:
1291                         continue;
1292                 case '\n':
1293                         return NULL;
1294                 case '\t': case ' ':
1295                         /*
1296                          * Is this the separator between the preimage
1297                          * and the postimage pathname?  Again, we are
1298                          * only interested in the case where there is
1299                          * no rename, as this is only to set def_name
1300                          * and a rename patch has the names elsewhere
1301                          * in an unambiguous form.
1302                          */
1303                         if (!name[len + 1])
1304                                 return NULL; /* no postimage name */
1305                         second = skip_tree_prefix(state, name + len + 1,
1306                                                   line_len - (len + 1));
1307                         if (!second)
1308                                 return NULL;
1309                         /*
1310                          * Does len bytes starting at "name" and "second"
1311                          * (that are separated by one HT or SP we just
1312                          * found) exactly match?
1313                          */
1314                         if (second[len] == '\n' && !strncmp(name, second, len))
1315                                 return xmemdupz(name, len);
1316                 }
1317         }
1318 }
1319
1320 static int check_header_line(struct apply_state *state, struct patch *patch)
1321 {
1322         int extensions = (patch->is_delete == 1) + (patch->is_new == 1) +
1323                          (patch->is_rename == 1) + (patch->is_copy == 1);
1324         if (extensions > 1)
1325                 return error(_("inconsistent header lines %d and %d"),
1326                              patch->extension_linenr, state->linenr);
1327         if (extensions && !patch->extension_linenr)
1328                 patch->extension_linenr = state->linenr;
1329         return 0;
1330 }
1331
1332 /* Verify that we recognize the lines following a git header */
1333 static int parse_git_header(struct apply_state *state,
1334                             const char *line,
1335                             int len,
1336                             unsigned int size,
1337                             struct patch *patch)
1338 {
1339         unsigned long offset;
1340
1341         /* A git diff has explicit new/delete information, so we don't guess */
1342         patch->is_new = 0;
1343         patch->is_delete = 0;
1344
1345         /*
1346          * Some things may not have the old name in the
1347          * rest of the headers anywhere (pure mode changes,
1348          * or removing or adding empty files), so we get
1349          * the default name from the header.
1350          */
1351         patch->def_name = git_header_name(state, line, len);
1352         if (patch->def_name && state->root.len) {
1353                 char *s = xstrfmt("%s%s", state->root.buf, patch->def_name);
1354                 free(patch->def_name);
1355                 patch->def_name = s;
1356         }
1357
1358         line += len;
1359         size -= len;
1360         state->linenr++;
1361         for (offset = len ; size > 0 ; offset += len, size -= len, line += len, state->linenr++) {
1362                 static const struct opentry {
1363                         const char *str;
1364                         int (*fn)(struct apply_state *, const char *, struct patch *);
1365                 } optable[] = {
1366                         { "@@ -", gitdiff_hdrend },
1367                         { "--- ", gitdiff_oldname },
1368                         { "+++ ", gitdiff_newname },
1369                         { "old mode ", gitdiff_oldmode },
1370                         { "new mode ", gitdiff_newmode },
1371                         { "deleted file mode ", gitdiff_delete },
1372                         { "new file mode ", gitdiff_newfile },
1373                         { "copy from ", gitdiff_copysrc },
1374                         { "copy to ", gitdiff_copydst },
1375                         { "rename old ", gitdiff_renamesrc },
1376                         { "rename new ", gitdiff_renamedst },
1377                         { "rename from ", gitdiff_renamesrc },
1378                         { "rename to ", gitdiff_renamedst },
1379                         { "similarity index ", gitdiff_similarity },
1380                         { "dissimilarity index ", gitdiff_dissimilarity },
1381                         { "index ", gitdiff_index },
1382                         { "", gitdiff_unrecognized },
1383                 };
1384                 int i;
1385
1386                 len = linelen(line, size);
1387                 if (!len || line[len-1] != '\n')
1388                         break;
1389                 for (i = 0; i < ARRAY_SIZE(optable); i++) {
1390                         const struct opentry *p = optable + i;
1391                         int oplen = strlen(p->str);
1392                         int res;
1393                         if (len < oplen || memcmp(p->str, line, oplen))
1394                                 continue;
1395                         res = p->fn(state, line + oplen, patch);
1396                         if (res < 0)
1397                                 return -1;
1398                         if (check_header_line(state, patch))
1399                                 return -1;
1400                         if (res > 0)
1401                                 return offset;
1402                         break;
1403                 }
1404         }
1405
1406         return offset;
1407 }
1408
1409 static int parse_num(const char *line, unsigned long *p)
1410 {
1411         char *ptr;
1412
1413         if (!isdigit(*line))
1414                 return 0;
1415         *p = strtoul(line, &ptr, 10);
1416         return ptr - line;
1417 }
1418
1419 static int parse_range(const char *line, int len, int offset, const char *expect,
1420                        unsigned long *p1, unsigned long *p2)
1421 {
1422         int digits, ex;
1423
1424         if (offset < 0 || offset >= len)
1425                 return -1;
1426         line += offset;
1427         len -= offset;
1428
1429         digits = parse_num(line, p1);
1430         if (!digits)
1431                 return -1;
1432
1433         offset += digits;
1434         line += digits;
1435         len -= digits;
1436
1437         *p2 = 1;
1438         if (*line == ',') {
1439                 digits = parse_num(line+1, p2);
1440                 if (!digits)
1441                         return -1;
1442
1443                 offset += digits+1;
1444                 line += digits+1;
1445                 len -= digits+1;
1446         }
1447
1448         ex = strlen(expect);
1449         if (ex > len)
1450                 return -1;
1451         if (memcmp(line, expect, ex))
1452                 return -1;
1453
1454         return offset + ex;
1455 }
1456
1457 static void recount_diff(const char *line, int size, struct fragment *fragment)
1458 {
1459         int oldlines = 0, newlines = 0, ret = 0;
1460
1461         if (size < 1) {
1462                 warning("recount: ignore empty hunk");
1463                 return;
1464         }
1465
1466         for (;;) {
1467                 int len = linelen(line, size);
1468                 size -= len;
1469                 line += len;
1470
1471                 if (size < 1)
1472                         break;
1473
1474                 switch (*line) {
1475                 case ' ': case '\n':
1476                         newlines++;
1477                         /* fall through */
1478                 case '-':
1479                         oldlines++;
1480                         continue;
1481                 case '+':
1482                         newlines++;
1483                         continue;
1484                 case '\\':
1485                         continue;
1486                 case '@':
1487                         ret = size < 3 || !starts_with(line, "@@ ");
1488                         break;
1489                 case 'd':
1490                         ret = size < 5 || !starts_with(line, "diff ");
1491                         break;
1492                 default:
1493                         ret = -1;
1494                         break;
1495                 }
1496                 if (ret) {
1497                         warning(_("recount: unexpected line: %.*s"),
1498                                 (int)linelen(line, size), line);
1499                         return;
1500                 }
1501                 break;
1502         }
1503         fragment->oldlines = oldlines;
1504         fragment->newlines = newlines;
1505 }
1506
1507 /*
1508  * Parse a unified diff fragment header of the
1509  * form "@@ -a,b +c,d @@"
1510  */
1511 static int parse_fragment_header(const char *line, int len, struct fragment *fragment)
1512 {
1513         int offset;
1514
1515         if (!len || line[len-1] != '\n')
1516                 return -1;
1517
1518         /* Figure out the number of lines in a fragment */
1519         offset = parse_range(line, len, 4, " +", &fragment->oldpos, &fragment->oldlines);
1520         offset = parse_range(line, len, offset, " @@", &fragment->newpos, &fragment->newlines);
1521
1522         return offset;
1523 }
1524
1525 /*
1526  * Find file diff header
1527  *
1528  * Returns:
1529  *  -1 if no header was found
1530  *  -128 in case of error
1531  *   the size of the header in bytes (called "offset") otherwise
1532  */
1533 static int find_header(struct apply_state *state,
1534                        const char *line,
1535                        unsigned long size,
1536                        int *hdrsize,
1537                        struct patch *patch)
1538 {
1539         unsigned long offset, len;
1540
1541         patch->is_toplevel_relative = 0;
1542         patch->is_rename = patch->is_copy = 0;
1543         patch->is_new = patch->is_delete = -1;
1544         patch->old_mode = patch->new_mode = 0;
1545         patch->old_name = patch->new_name = NULL;
1546         for (offset = 0; size > 0; offset += len, size -= len, line += len, state->linenr++) {
1547                 unsigned long nextlen;
1548
1549                 len = linelen(line, size);
1550                 if (!len)
1551                         break;
1552
1553                 /* Testing this early allows us to take a few shortcuts.. */
1554                 if (len < 6)
1555                         continue;
1556
1557                 /*
1558                  * Make sure we don't find any unconnected patch fragments.
1559                  * That's a sign that we didn't find a header, and that a
1560                  * patch has become corrupted/broken up.
1561                  */
1562                 if (!memcmp("@@ -", line, 4)) {
1563                         struct fragment dummy;
1564                         if (parse_fragment_header(line, len, &dummy) < 0)
1565                                 continue;
1566                         error(_("patch fragment without header at line %d: %.*s"),
1567                                      state->linenr, (int)len-1, line);
1568                         return -128;
1569                 }
1570
1571                 if (size < len + 6)
1572                         break;
1573
1574                 /*
1575                  * Git patch? It might not have a real patch, just a rename
1576                  * or mode change, so we handle that specially
1577                  */
1578                 if (!memcmp("diff --git ", line, 11)) {
1579                         int git_hdr_len = parse_git_header(state, line, len, size, patch);
1580                         if (git_hdr_len < 0)
1581                                 return -128;
1582                         if (git_hdr_len <= len)
1583                                 continue;
1584                         if (!patch->old_name && !patch->new_name) {
1585                                 if (!patch->def_name) {
1586                                         error(Q_("git diff header lacks filename information when removing "
1587                                                         "%d leading pathname component (line %d)",
1588                                                         "git diff header lacks filename information when removing "
1589                                                         "%d leading pathname components (line %d)",
1590                                                         state->p_value),
1591                                                      state->p_value, state->linenr);
1592                                         return -128;
1593                                 }
1594                                 patch->old_name = xstrdup(patch->def_name);
1595                                 patch->new_name = xstrdup(patch->def_name);
1596                         }
1597                         if ((!patch->new_name && !patch->is_delete) ||
1598                             (!patch->old_name && !patch->is_new)) {
1599                                 error(_("git diff header lacks filename information "
1600                                              "(line %d)"), state->linenr);
1601                                 return -128;
1602                         }
1603                         patch->is_toplevel_relative = 1;
1604                         *hdrsize = git_hdr_len;
1605                         return offset;
1606                 }
1607
1608                 /* --- followed by +++ ? */
1609                 if (memcmp("--- ", line,  4) || memcmp("+++ ", line + len, 4))
1610                         continue;
1611
1612                 /*
1613                  * We only accept unified patches, so we want it to
1614                  * at least have "@@ -a,b +c,d @@\n", which is 14 chars
1615                  * minimum ("@@ -0,0 +1 @@\n" is the shortest).
1616                  */
1617                 nextlen = linelen(line + len, size - len);
1618                 if (size < nextlen + 14 || memcmp("@@ -", line + len + nextlen, 4))
1619                         continue;
1620
1621                 /* Ok, we'll consider it a patch */
1622                 if (parse_traditional_patch(state, line, line+len, patch))
1623                         return -128;
1624                 *hdrsize = len + nextlen;
1625                 state->linenr += 2;
1626                 return offset;
1627         }
1628         return -1;
1629 }
1630
1631 static void record_ws_error(struct apply_state *state,
1632                             unsigned result,
1633                             const char *line,
1634                             int len,
1635                             int linenr)
1636 {
1637         char *err;
1638
1639         if (!result)
1640                 return;
1641
1642         state->whitespace_error++;
1643         if (state->squelch_whitespace_errors &&
1644             state->squelch_whitespace_errors < state->whitespace_error)
1645                 return;
1646
1647         err = whitespace_error_string(result);
1648         if (state->apply_verbosity > verbosity_silent)
1649                 fprintf(stderr, "%s:%d: %s.\n%.*s\n",
1650                         state->patch_input_file, linenr, err, len, line);
1651         free(err);
1652 }
1653
1654 static void check_whitespace(struct apply_state *state,
1655                              const char *line,
1656                              int len,
1657                              unsigned ws_rule)
1658 {
1659         unsigned result = ws_check(line + 1, len - 1, ws_rule);
1660
1661         record_ws_error(state, result, line + 1, len - 2, state->linenr);
1662 }
1663
1664 /*
1665  * Parse a unified diff. Note that this really needs to parse each
1666  * fragment separately, since the only way to know the difference
1667  * between a "---" that is part of a patch, and a "---" that starts
1668  * the next patch is to look at the line counts..
1669  */
1670 static int parse_fragment(struct apply_state *state,
1671                           const char *line,
1672                           unsigned long size,
1673                           struct patch *patch,
1674                           struct fragment *fragment)
1675 {
1676         int added, deleted;
1677         int len = linelen(line, size), offset;
1678         unsigned long oldlines, newlines;
1679         unsigned long leading, trailing;
1680
1681         offset = parse_fragment_header(line, len, fragment);
1682         if (offset < 0)
1683                 return -1;
1684         if (offset > 0 && patch->recount)
1685                 recount_diff(line + offset, size - offset, fragment);
1686         oldlines = fragment->oldlines;
1687         newlines = fragment->newlines;
1688         leading = 0;
1689         trailing = 0;
1690
1691         /* Parse the thing.. */
1692         line += len;
1693         size -= len;
1694         state->linenr++;
1695         added = deleted = 0;
1696         for (offset = len;
1697              0 < size;
1698              offset += len, size -= len, line += len, state->linenr++) {
1699                 if (!oldlines && !newlines)
1700                         break;
1701                 len = linelen(line, size);
1702                 if (!len || line[len-1] != '\n')
1703                         return -1;
1704                 switch (*line) {
1705                 default:
1706                         return -1;
1707                 case '\n': /* newer GNU diff, an empty context line */
1708                 case ' ':
1709                         oldlines--;
1710                         newlines--;
1711                         if (!deleted && !added)
1712                                 leading++;
1713                         trailing++;
1714                         if (!state->apply_in_reverse &&
1715                             state->ws_error_action == correct_ws_error)
1716                                 check_whitespace(state, line, len, patch->ws_rule);
1717                         break;
1718                 case '-':
1719                         if (state->apply_in_reverse &&
1720                             state->ws_error_action != nowarn_ws_error)
1721                                 check_whitespace(state, line, len, patch->ws_rule);
1722                         deleted++;
1723                         oldlines--;
1724                         trailing = 0;
1725                         break;
1726                 case '+':
1727                         if (!state->apply_in_reverse &&
1728                             state->ws_error_action != nowarn_ws_error)
1729                                 check_whitespace(state, line, len, patch->ws_rule);
1730                         added++;
1731                         newlines--;
1732                         trailing = 0;
1733                         break;
1734
1735                 /*
1736                  * We allow "\ No newline at end of file". Depending
1737                  * on locale settings when the patch was produced we
1738                  * don't know what this line looks like. The only
1739                  * thing we do know is that it begins with "\ ".
1740                  * Checking for 12 is just for sanity check -- any
1741                  * l10n of "\ No newline..." is at least that long.
1742                  */
1743                 case '\\':
1744                         if (len < 12 || memcmp(line, "\\ ", 2))
1745                                 return -1;
1746                         break;
1747                 }
1748         }
1749         if (oldlines || newlines)
1750                 return -1;
1751         if (!deleted && !added)
1752                 return -1;
1753
1754         fragment->leading = leading;
1755         fragment->trailing = trailing;
1756
1757         /*
1758          * If a fragment ends with an incomplete line, we failed to include
1759          * it in the above loop because we hit oldlines == newlines == 0
1760          * before seeing it.
1761          */
1762         if (12 < size && !memcmp(line, "\\ ", 2))
1763                 offset += linelen(line, size);
1764
1765         patch->lines_added += added;
1766         patch->lines_deleted += deleted;
1767
1768         if (0 < patch->is_new && oldlines)
1769                 return error(_("new file depends on old contents"));
1770         if (0 < patch->is_delete && newlines)
1771                 return error(_("deleted file still has contents"));
1772         return offset;
1773 }
1774
1775 /*
1776  * We have seen "diff --git a/... b/..." header (or a traditional patch
1777  * header).  Read hunks that belong to this patch into fragments and hang
1778  * them to the given patch structure.
1779  *
1780  * The (fragment->patch, fragment->size) pair points into the memory given
1781  * by the caller, not a copy, when we return.
1782  *
1783  * Returns:
1784  *   -1 in case of error,
1785  *   the number of bytes in the patch otherwise.
1786  */
1787 static int parse_single_patch(struct apply_state *state,
1788                               const char *line,
1789                               unsigned long size,
1790                               struct patch *patch)
1791 {
1792         unsigned long offset = 0;
1793         unsigned long oldlines = 0, newlines = 0, context = 0;
1794         struct fragment **fragp = &patch->fragments;
1795
1796         while (size > 4 && !memcmp(line, "@@ -", 4)) {
1797                 struct fragment *fragment;
1798                 int len;
1799
1800                 fragment = xcalloc(1, sizeof(*fragment));
1801                 fragment->linenr = state->linenr;
1802                 len = parse_fragment(state, line, size, patch, fragment);
1803                 if (len <= 0) {
1804                         free(fragment);
1805                         return error(_("corrupt patch at line %d"), state->linenr);
1806                 }
1807                 fragment->patch = line;
1808                 fragment->size = len;
1809                 oldlines += fragment->oldlines;
1810                 newlines += fragment->newlines;
1811                 context += fragment->leading + fragment->trailing;
1812
1813                 *fragp = fragment;
1814                 fragp = &fragment->next;
1815
1816                 offset += len;
1817                 line += len;
1818                 size -= len;
1819         }
1820
1821         /*
1822          * If something was removed (i.e. we have old-lines) it cannot
1823          * be creation, and if something was added it cannot be
1824          * deletion.  However, the reverse is not true; --unified=0
1825          * patches that only add are not necessarily creation even
1826          * though they do not have any old lines, and ones that only
1827          * delete are not necessarily deletion.
1828          *
1829          * Unfortunately, a real creation/deletion patch do _not_ have
1830          * any context line by definition, so we cannot safely tell it
1831          * apart with --unified=0 insanity.  At least if the patch has
1832          * more than one hunk it is not creation or deletion.
1833          */
1834         if (patch->is_new < 0 &&
1835             (oldlines || (patch->fragments && patch->fragments->next)))
1836                 patch->is_new = 0;
1837         if (patch->is_delete < 0 &&
1838             (newlines || (patch->fragments && patch->fragments->next)))
1839                 patch->is_delete = 0;
1840
1841         if (0 < patch->is_new && oldlines)
1842                 return error(_("new file %s depends on old contents"), patch->new_name);
1843         if (0 < patch->is_delete && newlines)
1844                 return error(_("deleted file %s still has contents"), patch->old_name);
1845         if (!patch->is_delete && !newlines && context && state->apply_verbosity > verbosity_silent)
1846                 fprintf_ln(stderr,
1847                            _("** warning: "
1848                              "file %s becomes empty but is not deleted"),
1849                            patch->new_name);
1850
1851         return offset;
1852 }
1853
1854 static inline int metadata_changes(struct patch *patch)
1855 {
1856         return  patch->is_rename > 0 ||
1857                 patch->is_copy > 0 ||
1858                 patch->is_new > 0 ||
1859                 patch->is_delete ||
1860                 (patch->old_mode && patch->new_mode &&
1861                  patch->old_mode != patch->new_mode);
1862 }
1863
1864 static char *inflate_it(const void *data, unsigned long size,
1865                         unsigned long inflated_size)
1866 {
1867         git_zstream stream;
1868         void *out;
1869         int st;
1870
1871         memset(&stream, 0, sizeof(stream));
1872
1873         stream.next_in = (unsigned char *)data;
1874         stream.avail_in = size;
1875         stream.next_out = out = xmalloc(inflated_size);
1876         stream.avail_out = inflated_size;
1877         git_inflate_init(&stream);
1878         st = git_inflate(&stream, Z_FINISH);
1879         git_inflate_end(&stream);
1880         if ((st != Z_STREAM_END) || stream.total_out != inflated_size) {
1881                 free(out);
1882                 return NULL;
1883         }
1884         return out;
1885 }
1886
1887 /*
1888  * Read a binary hunk and return a new fragment; fragment->patch
1889  * points at an allocated memory that the caller must free, so
1890  * it is marked as "->free_patch = 1".
1891  */
1892 static struct fragment *parse_binary_hunk(struct apply_state *state,
1893                                           char **buf_p,
1894                                           unsigned long *sz_p,
1895                                           int *status_p,
1896                                           int *used_p)
1897 {
1898         /*
1899          * Expect a line that begins with binary patch method ("literal"
1900          * or "delta"), followed by the length of data before deflating.
1901          * a sequence of 'length-byte' followed by base-85 encoded data
1902          * should follow, terminated by a newline.
1903          *
1904          * Each 5-byte sequence of base-85 encodes up to 4 bytes,
1905          * and we would limit the patch line to 66 characters,
1906          * so one line can fit up to 13 groups that would decode
1907          * to 52 bytes max.  The length byte 'A'-'Z' corresponds
1908          * to 1-26 bytes, and 'a'-'z' corresponds to 27-52 bytes.
1909          */
1910         int llen, used;
1911         unsigned long size = *sz_p;
1912         char *buffer = *buf_p;
1913         int patch_method;
1914         unsigned long origlen;
1915         char *data = NULL;
1916         int hunk_size = 0;
1917         struct fragment *frag;
1918
1919         llen = linelen(buffer, size);
1920         used = llen;
1921
1922         *status_p = 0;
1923
1924         if (starts_with(buffer, "delta ")) {
1925                 patch_method = BINARY_DELTA_DEFLATED;
1926                 origlen = strtoul(buffer + 6, NULL, 10);
1927         }
1928         else if (starts_with(buffer, "literal ")) {
1929                 patch_method = BINARY_LITERAL_DEFLATED;
1930                 origlen = strtoul(buffer + 8, NULL, 10);
1931         }
1932         else
1933                 return NULL;
1934
1935         state->linenr++;
1936         buffer += llen;
1937         while (1) {
1938                 int byte_length, max_byte_length, newsize;
1939                 llen = linelen(buffer, size);
1940                 used += llen;
1941                 state->linenr++;
1942                 if (llen == 1) {
1943                         /* consume the blank line */
1944                         buffer++;
1945                         size--;
1946                         break;
1947                 }
1948                 /*
1949                  * Minimum line is "A00000\n" which is 7-byte long,
1950                  * and the line length must be multiple of 5 plus 2.
1951                  */
1952                 if ((llen < 7) || (llen-2) % 5)
1953                         goto corrupt;
1954                 max_byte_length = (llen - 2) / 5 * 4;
1955                 byte_length = *buffer;
1956                 if ('A' <= byte_length && byte_length <= 'Z')
1957                         byte_length = byte_length - 'A' + 1;
1958                 else if ('a' <= byte_length && byte_length <= 'z')
1959                         byte_length = byte_length - 'a' + 27;
1960                 else
1961                         goto corrupt;
1962                 /* if the input length was not multiple of 4, we would
1963                  * have filler at the end but the filler should never
1964                  * exceed 3 bytes
1965                  */
1966                 if (max_byte_length < byte_length ||
1967                     byte_length <= max_byte_length - 4)
1968                         goto corrupt;
1969                 newsize = hunk_size + byte_length;
1970                 data = xrealloc(data, newsize);
1971                 if (decode_85(data + hunk_size, buffer + 1, byte_length))
1972                         goto corrupt;
1973                 hunk_size = newsize;
1974                 buffer += llen;
1975                 size -= llen;
1976         }
1977
1978         frag = xcalloc(1, sizeof(*frag));
1979         frag->patch = inflate_it(data, hunk_size, origlen);
1980         frag->free_patch = 1;
1981         if (!frag->patch)
1982                 goto corrupt;
1983         free(data);
1984         frag->size = origlen;
1985         *buf_p = buffer;
1986         *sz_p = size;
1987         *used_p = used;
1988         frag->binary_patch_method = patch_method;
1989         return frag;
1990
1991  corrupt:
1992         free(data);
1993         *status_p = -1;
1994         error(_("corrupt binary patch at line %d: %.*s"),
1995               state->linenr-1, llen-1, buffer);
1996         return NULL;
1997 }
1998
1999 /*
2000  * Returns:
2001  *   -1 in case of error,
2002  *   the length of the parsed binary patch otherwise
2003  */
2004 static int parse_binary(struct apply_state *state,
2005                         char *buffer,
2006                         unsigned long size,
2007                         struct patch *patch)
2008 {
2009         /*
2010          * We have read "GIT binary patch\n"; what follows is a line
2011          * that says the patch method (currently, either "literal" or
2012          * "delta") and the length of data before deflating; a
2013          * sequence of 'length-byte' followed by base-85 encoded data
2014          * follows.
2015          *
2016          * When a binary patch is reversible, there is another binary
2017          * hunk in the same format, starting with patch method (either
2018          * "literal" or "delta") with the length of data, and a sequence
2019          * of length-byte + base-85 encoded data, terminated with another
2020          * empty line.  This data, when applied to the postimage, produces
2021          * the preimage.
2022          */
2023         struct fragment *forward;
2024         struct fragment *reverse;
2025         int status;
2026         int used, used_1;
2027
2028         forward = parse_binary_hunk(state, &buffer, &size, &status, &used);
2029         if (!forward && !status)
2030                 /* there has to be one hunk (forward hunk) */
2031                 return error(_("unrecognized binary patch at line %d"), state->linenr-1);
2032         if (status)
2033                 /* otherwise we already gave an error message */
2034                 return status;
2035
2036         reverse = parse_binary_hunk(state, &buffer, &size, &status, &used_1);
2037         if (reverse)
2038                 used += used_1;
2039         else if (status) {
2040                 /*
2041                  * Not having reverse hunk is not an error, but having
2042                  * a corrupt reverse hunk is.
2043                  */
2044                 free((void*) forward->patch);
2045                 free(forward);
2046                 return status;
2047         }
2048         forward->next = reverse;
2049         patch->fragments = forward;
2050         patch->is_binary = 1;
2051         return used;
2052 }
2053
2054 static void prefix_one(struct apply_state *state, char **name)
2055 {
2056         char *old_name = *name;
2057         if (!old_name)
2058                 return;
2059         *name = prefix_filename(state->prefix, *name);
2060         free(old_name);
2061 }
2062
2063 static void prefix_patch(struct apply_state *state, struct patch *p)
2064 {
2065         if (!state->prefix || p->is_toplevel_relative)
2066                 return;
2067         prefix_one(state, &p->new_name);
2068         prefix_one(state, &p->old_name);
2069 }
2070
2071 /*
2072  * include/exclude
2073  */
2074
2075 static void add_name_limit(struct apply_state *state,
2076                            const char *name,
2077                            int exclude)
2078 {
2079         struct string_list_item *it;
2080
2081         it = string_list_append(&state->limit_by_name, name);
2082         it->util = exclude ? NULL : (void *) 1;
2083 }
2084
2085 static int use_patch(struct apply_state *state, struct patch *p)
2086 {
2087         const char *pathname = p->new_name ? p->new_name : p->old_name;
2088         int i;
2089
2090         /* Paths outside are not touched regardless of "--include" */
2091         if (state->prefix && *state->prefix) {
2092                 const char *rest;
2093                 if (!skip_prefix(pathname, state->prefix, &rest) || !*rest)
2094                         return 0;
2095         }
2096
2097         /* See if it matches any of exclude/include rule */
2098         for (i = 0; i < state->limit_by_name.nr; i++) {
2099                 struct string_list_item *it = &state->limit_by_name.items[i];
2100                 if (!wildmatch(it->string, pathname, 0))
2101                         return (it->util != NULL);
2102         }
2103
2104         /*
2105          * If we had any include, a path that does not match any rule is
2106          * not used.  Otherwise, we saw bunch of exclude rules (or none)
2107          * and such a path is used.
2108          */
2109         return !state->has_include;
2110 }
2111
2112 /*
2113  * Read the patch text in "buffer" that extends for "size" bytes; stop
2114  * reading after seeing a single patch (i.e. changes to a single file).
2115  * Create fragments (i.e. patch hunks) and hang them to the given patch.
2116  *
2117  * Returns:
2118  *   -1 if no header was found or parse_binary() failed,
2119  *   -128 on another error,
2120  *   the number of bytes consumed otherwise,
2121  *     so that the caller can call us again for the next patch.
2122  */
2123 static int parse_chunk(struct apply_state *state, char *buffer, unsigned long size, struct patch *patch)
2124 {
2125         int hdrsize, patchsize;
2126         int offset = find_header(state, buffer, size, &hdrsize, patch);
2127
2128         if (offset < 0)
2129                 return offset;
2130
2131         prefix_patch(state, patch);
2132
2133         if (!use_patch(state, patch))
2134                 patch->ws_rule = 0;
2135         else
2136                 patch->ws_rule = whitespace_rule(patch->new_name
2137                                                  ? patch->new_name
2138                                                  : patch->old_name);
2139
2140         patchsize = parse_single_patch(state,
2141                                        buffer + offset + hdrsize,
2142                                        size - offset - hdrsize,
2143                                        patch);
2144
2145         if (patchsize < 0)
2146                 return -128;
2147
2148         if (!patchsize) {
2149                 static const char git_binary[] = "GIT binary patch\n";
2150                 int hd = hdrsize + offset;
2151                 unsigned long llen = linelen(buffer + hd, size - hd);
2152
2153                 if (llen == sizeof(git_binary) - 1 &&
2154                     !memcmp(git_binary, buffer + hd, llen)) {
2155                         int used;
2156                         state->linenr++;
2157                         used = parse_binary(state, buffer + hd + llen,
2158                                             size - hd - llen, patch);
2159                         if (used < 0)
2160                                 return -1;
2161                         if (used)
2162                                 patchsize = used + llen;
2163                         else
2164                                 patchsize = 0;
2165                 }
2166                 else if (!memcmp(" differ\n", buffer + hd + llen - 8, 8)) {
2167                         static const char *binhdr[] = {
2168                                 "Binary files ",
2169                                 "Files ",
2170                                 NULL,
2171                         };
2172                         int i;
2173                         for (i = 0; binhdr[i]; i++) {
2174                                 int len = strlen(binhdr[i]);
2175                                 if (len < size - hd &&
2176                                     !memcmp(binhdr[i], buffer + hd, len)) {
2177                                         state->linenr++;
2178                                         patch->is_binary = 1;
2179                                         patchsize = llen;
2180                                         break;
2181                                 }
2182                         }
2183                 }
2184
2185                 /* Empty patch cannot be applied if it is a text patch
2186                  * without metadata change.  A binary patch appears
2187                  * empty to us here.
2188                  */
2189                 if ((state->apply || state->check) &&
2190                     (!patch->is_binary && !metadata_changes(patch))) {
2191                         error(_("patch with only garbage at line %d"), state->linenr);
2192                         return -128;
2193                 }
2194         }
2195
2196         return offset + hdrsize + patchsize;
2197 }
2198
2199 static void reverse_patches(struct patch *p)
2200 {
2201         for (; p; p = p->next) {
2202                 struct fragment *frag = p->fragments;
2203
2204                 SWAP(p->new_name, p->old_name);
2205                 SWAP(p->new_mode, p->old_mode);
2206                 SWAP(p->is_new, p->is_delete);
2207                 SWAP(p->lines_added, p->lines_deleted);
2208                 SWAP(p->old_sha1_prefix, p->new_sha1_prefix);
2209
2210                 for (; frag; frag = frag->next) {
2211                         SWAP(frag->newpos, frag->oldpos);
2212                         SWAP(frag->newlines, frag->oldlines);
2213                 }
2214         }
2215 }
2216
2217 static const char pluses[] =
2218 "++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++";
2219 static const char minuses[]=
2220 "----------------------------------------------------------------------";
2221
2222 static void show_stats(struct apply_state *state, struct patch *patch)
2223 {
2224         struct strbuf qname = STRBUF_INIT;
2225         char *cp = patch->new_name ? patch->new_name : patch->old_name;
2226         int max, add, del;
2227
2228         quote_c_style(cp, &qname, NULL, 0);
2229
2230         /*
2231          * "scale" the filename
2232          */
2233         max = state->max_len;
2234         if (max > 50)
2235                 max = 50;
2236
2237         if (qname.len > max) {
2238                 cp = strchr(qname.buf + qname.len + 3 - max, '/');
2239                 if (!cp)
2240                         cp = qname.buf + qname.len + 3 - max;
2241                 strbuf_splice(&qname, 0, cp - qname.buf, "...", 3);
2242         }
2243
2244         if (patch->is_binary) {
2245                 printf(" %-*s |  Bin\n", max, qname.buf);
2246                 strbuf_release(&qname);
2247                 return;
2248         }
2249
2250         printf(" %-*s |", max, qname.buf);
2251         strbuf_release(&qname);
2252
2253         /*
2254          * scale the add/delete
2255          */
2256         max = max + state->max_change > 70 ? 70 - max : state->max_change;
2257         add = patch->lines_added;
2258         del = patch->lines_deleted;
2259
2260         if (state->max_change > 0) {
2261                 int total = ((add + del) * max + state->max_change / 2) / state->max_change;
2262                 add = (add * max + state->max_change / 2) / state->max_change;
2263                 del = total - add;
2264         }
2265         printf("%5d %.*s%.*s\n", patch->lines_added + patch->lines_deleted,
2266                 add, pluses, del, minuses);
2267 }
2268
2269 static int read_old_data(struct stat *st, const char *path, struct strbuf *buf)
2270 {
2271         switch (st->st_mode & S_IFMT) {
2272         case S_IFLNK:
2273                 if (strbuf_readlink(buf, path, st->st_size) < 0)
2274                         return error(_("unable to read symlink %s"), path);
2275                 return 0;
2276         case S_IFREG:
2277                 if (strbuf_read_file(buf, path, st->st_size) != st->st_size)
2278                         return error(_("unable to open or read %s"), path);
2279                 convert_to_git(&the_index, path, buf->buf, buf->len, buf, 0);
2280                 return 0;
2281         default:
2282                 return -1;
2283         }
2284 }
2285
2286 /*
2287  * Update the preimage, and the common lines in postimage,
2288  * from buffer buf of length len. If postlen is 0 the postimage
2289  * is updated in place, otherwise it's updated on a new buffer
2290  * of length postlen
2291  */
2292
2293 static void update_pre_post_images(struct image *preimage,
2294                                    struct image *postimage,
2295                                    char *buf,
2296                                    size_t len, size_t postlen)
2297 {
2298         int i, ctx, reduced;
2299         char *new, *old, *fixed;
2300         struct image fixed_preimage;
2301
2302         /*
2303          * Update the preimage with whitespace fixes.  Note that we
2304          * are not losing preimage->buf -- apply_one_fragment() will
2305          * free "oldlines".
2306          */
2307         prepare_image(&fixed_preimage, buf, len, 1);
2308         assert(postlen
2309                ? fixed_preimage.nr == preimage->nr
2310                : fixed_preimage.nr <= preimage->nr);
2311         for (i = 0; i < fixed_preimage.nr; i++)
2312                 fixed_preimage.line[i].flag = preimage->line[i].flag;
2313         free(preimage->line_allocated);
2314         *preimage = fixed_preimage;
2315
2316         /*
2317          * Adjust the common context lines in postimage. This can be
2318          * done in-place when we are shrinking it with whitespace
2319          * fixing, but needs a new buffer when ignoring whitespace or
2320          * expanding leading tabs to spaces.
2321          *
2322          * We trust the caller to tell us if the update can be done
2323          * in place (postlen==0) or not.
2324          */
2325         old = postimage->buf;
2326         if (postlen)
2327                 new = postimage->buf = xmalloc(postlen);
2328         else
2329                 new = old;
2330         fixed = preimage->buf;
2331
2332         for (i = reduced = ctx = 0; i < postimage->nr; i++) {
2333                 size_t l_len = postimage->line[i].len;
2334                 if (!(postimage->line[i].flag & LINE_COMMON)) {
2335                         /* an added line -- no counterparts in preimage */
2336                         memmove(new, old, l_len);
2337                         old += l_len;
2338                         new += l_len;
2339                         continue;
2340                 }
2341
2342                 /* a common context -- skip it in the original postimage */
2343                 old += l_len;
2344
2345                 /* and find the corresponding one in the fixed preimage */
2346                 while (ctx < preimage->nr &&
2347                        !(preimage->line[ctx].flag & LINE_COMMON)) {
2348                         fixed += preimage->line[ctx].len;
2349                         ctx++;
2350                 }
2351
2352                 /*
2353                  * preimage is expected to run out, if the caller
2354                  * fixed addition of trailing blank lines.
2355                  */
2356                 if (preimage->nr <= ctx) {
2357                         reduced++;
2358                         continue;
2359                 }
2360
2361                 /* and copy it in, while fixing the line length */
2362                 l_len = preimage->line[ctx].len;
2363                 memcpy(new, fixed, l_len);
2364                 new += l_len;
2365                 fixed += l_len;
2366                 postimage->line[i].len = l_len;
2367                 ctx++;
2368         }
2369
2370         if (postlen
2371             ? postlen < new - postimage->buf
2372             : postimage->len < new - postimage->buf)
2373                 die("BUG: caller miscounted postlen: asked %d, orig = %d, used = %d",
2374                     (int)postlen, (int) postimage->len, (int)(new - postimage->buf));
2375
2376         /* Fix the length of the whole thing */
2377         postimage->len = new - postimage->buf;
2378         postimage->nr -= reduced;
2379 }
2380
2381 static int line_by_line_fuzzy_match(struct image *img,
2382                                     struct image *preimage,
2383                                     struct image *postimage,
2384                                     unsigned long try,
2385                                     int try_lno,
2386                                     int preimage_limit)
2387 {
2388         int i;
2389         size_t imgoff = 0;
2390         size_t preoff = 0;
2391         size_t postlen = postimage->len;
2392         size_t extra_chars;
2393         char *buf;
2394         char *preimage_eof;
2395         char *preimage_end;
2396         struct strbuf fixed;
2397         char *fixed_buf;
2398         size_t fixed_len;
2399
2400         for (i = 0; i < preimage_limit; i++) {
2401                 size_t prelen = preimage->line[i].len;
2402                 size_t imglen = img->line[try_lno+i].len;
2403
2404                 if (!fuzzy_matchlines(img->buf + try + imgoff, imglen,
2405                                       preimage->buf + preoff, prelen))
2406                         return 0;
2407                 if (preimage->line[i].flag & LINE_COMMON)
2408                         postlen += imglen - prelen;
2409                 imgoff += imglen;
2410                 preoff += prelen;
2411         }
2412
2413         /*
2414          * Ok, the preimage matches with whitespace fuzz.
2415          *
2416          * imgoff now holds the true length of the target that
2417          * matches the preimage before the end of the file.
2418          *
2419          * Count the number of characters in the preimage that fall
2420          * beyond the end of the file and make sure that all of them
2421          * are whitespace characters. (This can only happen if
2422          * we are removing blank lines at the end of the file.)
2423          */
2424         buf = preimage_eof = preimage->buf + preoff;
2425         for ( ; i < preimage->nr; i++)
2426                 preoff += preimage->line[i].len;
2427         preimage_end = preimage->buf + preoff;
2428         for ( ; buf < preimage_end; buf++)
2429                 if (!isspace(*buf))
2430                         return 0;
2431
2432         /*
2433          * Update the preimage and the common postimage context
2434          * lines to use the same whitespace as the target.
2435          * If whitespace is missing in the target (i.e.
2436          * if the preimage extends beyond the end of the file),
2437          * use the whitespace from the preimage.
2438          */
2439         extra_chars = preimage_end - preimage_eof;
2440         strbuf_init(&fixed, imgoff + extra_chars);
2441         strbuf_add(&fixed, img->buf + try, imgoff);
2442         strbuf_add(&fixed, preimage_eof, extra_chars);
2443         fixed_buf = strbuf_detach(&fixed, &fixed_len);
2444         update_pre_post_images(preimage, postimage,
2445                                fixed_buf, fixed_len, postlen);
2446         return 1;
2447 }
2448
2449 static int match_fragment(struct apply_state *state,
2450                           struct image *img,
2451                           struct image *preimage,
2452                           struct image *postimage,
2453                           unsigned long try,
2454                           int try_lno,
2455                           unsigned ws_rule,
2456                           int match_beginning, int match_end)
2457 {
2458         int i;
2459         char *fixed_buf, *buf, *orig, *target;
2460         struct strbuf fixed;
2461         size_t fixed_len, postlen;
2462         int preimage_limit;
2463
2464         if (preimage->nr + try_lno <= img->nr) {
2465                 /*
2466                  * The hunk falls within the boundaries of img.
2467                  */
2468                 preimage_limit = preimage->nr;
2469                 if (match_end && (preimage->nr + try_lno != img->nr))
2470                         return 0;
2471         } else if (state->ws_error_action == correct_ws_error &&
2472                    (ws_rule & WS_BLANK_AT_EOF)) {
2473                 /*
2474                  * This hunk extends beyond the end of img, and we are
2475                  * removing blank lines at the end of the file.  This
2476                  * many lines from the beginning of the preimage must
2477                  * match with img, and the remainder of the preimage
2478                  * must be blank.
2479                  */
2480                 preimage_limit = img->nr - try_lno;
2481         } else {
2482                 /*
2483                  * The hunk extends beyond the end of the img and
2484                  * we are not removing blanks at the end, so we
2485                  * should reject the hunk at this position.
2486                  */
2487                 return 0;
2488         }
2489
2490         if (match_beginning && try_lno)
2491                 return 0;
2492
2493         /* Quick hash check */
2494         for (i = 0; i < preimage_limit; i++)
2495                 if ((img->line[try_lno + i].flag & LINE_PATCHED) ||
2496                     (preimage->line[i].hash != img->line[try_lno + i].hash))
2497                         return 0;
2498
2499         if (preimage_limit == preimage->nr) {
2500                 /*
2501                  * Do we have an exact match?  If we were told to match
2502                  * at the end, size must be exactly at try+fragsize,
2503                  * otherwise try+fragsize must be still within the preimage,
2504                  * and either case, the old piece should match the preimage
2505                  * exactly.
2506                  */
2507                 if ((match_end
2508                      ? (try + preimage->len == img->len)
2509                      : (try + preimage->len <= img->len)) &&
2510                     !memcmp(img->buf + try, preimage->buf, preimage->len))
2511                         return 1;
2512         } else {
2513                 /*
2514                  * The preimage extends beyond the end of img, so
2515                  * there cannot be an exact match.
2516                  *
2517                  * There must be one non-blank context line that match
2518                  * a line before the end of img.
2519                  */
2520                 char *buf_end;
2521
2522                 buf = preimage->buf;
2523                 buf_end = buf;
2524                 for (i = 0; i < preimage_limit; i++)
2525                         buf_end += preimage->line[i].len;
2526
2527                 for ( ; buf < buf_end; buf++)
2528                         if (!isspace(*buf))
2529                                 break;
2530                 if (buf == buf_end)
2531                         return 0;
2532         }
2533
2534         /*
2535          * No exact match. If we are ignoring whitespace, run a line-by-line
2536          * fuzzy matching. We collect all the line length information because
2537          * we need it to adjust whitespace if we match.
2538          */
2539         if (state->ws_ignore_action == ignore_ws_change)
2540                 return line_by_line_fuzzy_match(img, preimage, postimage,
2541                                                 try, try_lno, preimage_limit);
2542
2543         if (state->ws_error_action != correct_ws_error)
2544                 return 0;
2545
2546         /*
2547          * The hunk does not apply byte-by-byte, but the hash says
2548          * it might with whitespace fuzz. We weren't asked to
2549          * ignore whitespace, we were asked to correct whitespace
2550          * errors, so let's try matching after whitespace correction.
2551          *
2552          * While checking the preimage against the target, whitespace
2553          * errors in both fixed, we count how large the corresponding
2554          * postimage needs to be.  The postimage prepared by
2555          * apply_one_fragment() has whitespace errors fixed on added
2556          * lines already, but the common lines were propagated as-is,
2557          * which may become longer when their whitespace errors are
2558          * fixed.
2559          */
2560
2561         /* First count added lines in postimage */
2562         postlen = 0;
2563         for (i = 0; i < postimage->nr; i++) {
2564                 if (!(postimage->line[i].flag & LINE_COMMON))
2565                         postlen += postimage->line[i].len;
2566         }
2567
2568         /*
2569          * The preimage may extend beyond the end of the file,
2570          * but in this loop we will only handle the part of the
2571          * preimage that falls within the file.
2572          */
2573         strbuf_init(&fixed, preimage->len + 1);
2574         orig = preimage->buf;
2575         target = img->buf + try;
2576         for (i = 0; i < preimage_limit; i++) {
2577                 size_t oldlen = preimage->line[i].len;
2578                 size_t tgtlen = img->line[try_lno + i].len;
2579                 size_t fixstart = fixed.len;
2580                 struct strbuf tgtfix;
2581                 int match;
2582
2583                 /* Try fixing the line in the preimage */
2584                 ws_fix_copy(&fixed, orig, oldlen, ws_rule, NULL);
2585
2586                 /* Try fixing the line in the target */
2587                 strbuf_init(&tgtfix, tgtlen);
2588                 ws_fix_copy(&tgtfix, target, tgtlen, ws_rule, NULL);
2589
2590                 /*
2591                  * If they match, either the preimage was based on
2592                  * a version before our tree fixed whitespace breakage,
2593                  * or we are lacking a whitespace-fix patch the tree
2594                  * the preimage was based on already had (i.e. target
2595                  * has whitespace breakage, the preimage doesn't).
2596                  * In either case, we are fixing the whitespace breakages
2597                  * so we might as well take the fix together with their
2598                  * real change.
2599                  */
2600                 match = (tgtfix.len == fixed.len - fixstart &&
2601                          !memcmp(tgtfix.buf, fixed.buf + fixstart,
2602                                              fixed.len - fixstart));
2603
2604                 /* Add the length if this is common with the postimage */
2605                 if (preimage->line[i].flag & LINE_COMMON)
2606                         postlen += tgtfix.len;
2607
2608                 strbuf_release(&tgtfix);
2609                 if (!match)
2610                         goto unmatch_exit;
2611
2612                 orig += oldlen;
2613                 target += tgtlen;
2614         }
2615
2616
2617         /*
2618          * Now handle the lines in the preimage that falls beyond the
2619          * end of the file (if any). They will only match if they are
2620          * empty or only contain whitespace (if WS_BLANK_AT_EOL is
2621          * false).
2622          */
2623         for ( ; i < preimage->nr; i++) {
2624                 size_t fixstart = fixed.len; /* start of the fixed preimage */
2625                 size_t oldlen = preimage->line[i].len;
2626                 int j;
2627
2628                 /* Try fixing the line in the preimage */
2629                 ws_fix_copy(&fixed, orig, oldlen, ws_rule, NULL);
2630
2631                 for (j = fixstart; j < fixed.len; j++)
2632                         if (!isspace(fixed.buf[j]))
2633                                 goto unmatch_exit;
2634
2635                 orig += oldlen;
2636         }
2637
2638         /*
2639          * Yes, the preimage is based on an older version that still
2640          * has whitespace breakages unfixed, and fixing them makes the
2641          * hunk match.  Update the context lines in the postimage.
2642          */
2643         fixed_buf = strbuf_detach(&fixed, &fixed_len);
2644         if (postlen < postimage->len)
2645                 postlen = 0;
2646         update_pre_post_images(preimage, postimage,
2647                                fixed_buf, fixed_len, postlen);
2648         return 1;
2649
2650  unmatch_exit:
2651         strbuf_release(&fixed);
2652         return 0;
2653 }
2654
2655 static int find_pos(struct apply_state *state,
2656                     struct image *img,
2657                     struct image *preimage,
2658                     struct image *postimage,
2659                     int line,
2660                     unsigned ws_rule,
2661                     int match_beginning, int match_end)
2662 {
2663         int i;
2664         unsigned long backwards, forwards, try;
2665         int backwards_lno, forwards_lno, try_lno;
2666
2667         /*
2668          * If match_beginning or match_end is specified, there is no
2669          * point starting from a wrong line that will never match and
2670          * wander around and wait for a match at the specified end.
2671          */
2672         if (match_beginning)
2673                 line = 0;
2674         else if (match_end)
2675                 line = img->nr - preimage->nr;
2676
2677         /*
2678          * Because the comparison is unsigned, the following test
2679          * will also take care of a negative line number that can
2680          * result when match_end and preimage is larger than the target.
2681          */
2682         if ((size_t) line > img->nr)
2683                 line = img->nr;
2684
2685         try = 0;
2686         for (i = 0; i < line; i++)
2687                 try += img->line[i].len;
2688
2689         /*
2690          * There's probably some smart way to do this, but I'll leave
2691          * that to the smart and beautiful people. I'm simple and stupid.
2692          */
2693         backwards = try;
2694         backwards_lno = line;
2695         forwards = try;
2696         forwards_lno = line;
2697         try_lno = line;
2698
2699         for (i = 0; ; i++) {
2700                 if (match_fragment(state, img, preimage, postimage,
2701                                    try, try_lno, ws_rule,
2702                                    match_beginning, match_end))
2703                         return try_lno;
2704
2705         again:
2706                 if (backwards_lno == 0 && forwards_lno == img->nr)
2707                         break;
2708
2709                 if (i & 1) {
2710                         if (backwards_lno == 0) {
2711                                 i++;
2712                                 goto again;
2713                         }
2714                         backwards_lno--;
2715                         backwards -= img->line[backwards_lno].len;
2716                         try = backwards;
2717                         try_lno = backwards_lno;
2718                 } else {
2719                         if (forwards_lno == img->nr) {
2720                                 i++;
2721                                 goto again;
2722                         }
2723                         forwards += img->line[forwards_lno].len;
2724                         forwards_lno++;
2725                         try = forwards;
2726                         try_lno = forwards_lno;
2727                 }
2728
2729         }
2730         return -1;
2731 }
2732
2733 static void remove_first_line(struct image *img)
2734 {
2735         img->buf += img->line[0].len;
2736         img->len -= img->line[0].len;
2737         img->line++;
2738         img->nr--;
2739 }
2740
2741 static void remove_last_line(struct image *img)
2742 {
2743         img->len -= img->line[--img->nr].len;
2744 }
2745
2746 /*
2747  * The change from "preimage" and "postimage" has been found to
2748  * apply at applied_pos (counts in line numbers) in "img".
2749  * Update "img" to remove "preimage" and replace it with "postimage".
2750  */
2751 static void update_image(struct apply_state *state,
2752                          struct image *img,
2753                          int applied_pos,
2754                          struct image *preimage,
2755                          struct image *postimage)
2756 {
2757         /*
2758          * remove the copy of preimage at offset in img
2759          * and replace it with postimage
2760          */
2761         int i, nr;
2762         size_t remove_count, insert_count, applied_at = 0;
2763         char *result;
2764         int preimage_limit;
2765
2766         /*
2767          * If we are removing blank lines at the end of img,
2768          * the preimage may extend beyond the end.
2769          * If that is the case, we must be careful only to
2770          * remove the part of the preimage that falls within
2771          * the boundaries of img. Initialize preimage_limit
2772          * to the number of lines in the preimage that falls
2773          * within the boundaries.
2774          */
2775         preimage_limit = preimage->nr;
2776         if (preimage_limit > img->nr - applied_pos)
2777                 preimage_limit = img->nr - applied_pos;
2778
2779         for (i = 0; i < applied_pos; i++)
2780                 applied_at += img->line[i].len;
2781
2782         remove_count = 0;
2783         for (i = 0; i < preimage_limit; i++)
2784                 remove_count += img->line[applied_pos + i].len;
2785         insert_count = postimage->len;
2786
2787         /* Adjust the contents */
2788         result = xmalloc(st_add3(st_sub(img->len, remove_count), insert_count, 1));
2789         memcpy(result, img->buf, applied_at);
2790         memcpy(result + applied_at, postimage->buf, postimage->len);
2791         memcpy(result + applied_at + postimage->len,
2792                img->buf + (applied_at + remove_count),
2793                img->len - (applied_at + remove_count));
2794         free(img->buf);
2795         img->buf = result;
2796         img->len += insert_count - remove_count;
2797         result[img->len] = '\0';
2798
2799         /* Adjust the line table */
2800         nr = img->nr + postimage->nr - preimage_limit;
2801         if (preimage_limit < postimage->nr) {
2802                 /*
2803                  * NOTE: this knows that we never call remove_first_line()
2804                  * on anything other than pre/post image.
2805                  */
2806                 REALLOC_ARRAY(img->line, nr);
2807                 img->line_allocated = img->line;
2808         }
2809         if (preimage_limit != postimage->nr)
2810                 MOVE_ARRAY(img->line + applied_pos + postimage->nr,
2811                            img->line + applied_pos + preimage_limit,
2812                            img->nr - (applied_pos + preimage_limit));
2813         COPY_ARRAY(img->line + applied_pos, postimage->line, postimage->nr);
2814         if (!state->allow_overlap)
2815                 for (i = 0; i < postimage->nr; i++)
2816                         img->line[applied_pos + i].flag |= LINE_PATCHED;
2817         img->nr = nr;
2818 }
2819
2820 /*
2821  * Use the patch-hunk text in "frag" to prepare two images (preimage and
2822  * postimage) for the hunk.  Find lines that match "preimage" in "img" and
2823  * replace the part of "img" with "postimage" text.
2824  */
2825 static int apply_one_fragment(struct apply_state *state,
2826                               struct image *img, struct fragment *frag,
2827                               int inaccurate_eof, unsigned ws_rule,
2828                               int nth_fragment)
2829 {
2830         int match_beginning, match_end;
2831         const char *patch = frag->patch;
2832         int size = frag->size;
2833         char *old, *oldlines;
2834         struct strbuf newlines;
2835         int new_blank_lines_at_end = 0;
2836         int found_new_blank_lines_at_end = 0;
2837         int hunk_linenr = frag->linenr;
2838         unsigned long leading, trailing;
2839         int pos, applied_pos;
2840         struct image preimage;
2841         struct image postimage;
2842
2843         memset(&preimage, 0, sizeof(preimage));
2844         memset(&postimage, 0, sizeof(postimage));
2845         oldlines = xmalloc(size);
2846         strbuf_init(&newlines, size);
2847
2848         old = oldlines;
2849         while (size > 0) {
2850                 char first;
2851                 int len = linelen(patch, size);
2852                 int plen;
2853                 int added_blank_line = 0;
2854                 int is_blank_context = 0;
2855                 size_t start;
2856
2857                 if (!len)
2858                         break;
2859
2860                 /*
2861                  * "plen" is how much of the line we should use for
2862                  * the actual patch data. Normally we just remove the
2863                  * first character on the line, but if the line is
2864                  * followed by "\ No newline", then we also remove the
2865                  * last one (which is the newline, of course).
2866                  */
2867                 plen = len - 1;
2868                 if (len < size && patch[len] == '\\')
2869                         plen--;
2870                 first = *patch;
2871                 if (state->apply_in_reverse) {
2872                         if (first == '-')
2873                                 first = '+';
2874                         else if (first == '+')
2875                                 first = '-';
2876                 }
2877
2878                 switch (first) {
2879                 case '\n':
2880                         /* Newer GNU diff, empty context line */
2881                         if (plen < 0)
2882                                 /* ... followed by '\No newline'; nothing */
2883                                 break;
2884                         *old++ = '\n';
2885                         strbuf_addch(&newlines, '\n');
2886                         add_line_info(&preimage, "\n", 1, LINE_COMMON);
2887                         add_line_info(&postimage, "\n", 1, LINE_COMMON);
2888                         is_blank_context = 1;
2889                         break;
2890                 case ' ':
2891                         if (plen && (ws_rule & WS_BLANK_AT_EOF) &&
2892                             ws_blank_line(patch + 1, plen, ws_rule))
2893                                 is_blank_context = 1;
2894                 case '-':
2895                         memcpy(old, patch + 1, plen);
2896                         add_line_info(&preimage, old, plen,
2897                                       (first == ' ' ? LINE_COMMON : 0));
2898                         old += plen;
2899                         if (first == '-')
2900                                 break;
2901                 /* Fall-through for ' ' */
2902                 case '+':
2903                         /* --no-add does not add new lines */
2904                         if (first == '+' && state->no_add)
2905                                 break;
2906
2907                         start = newlines.len;
2908                         if (first != '+' ||
2909                             !state->whitespace_error ||
2910                             state->ws_error_action != correct_ws_error) {
2911                                 strbuf_add(&newlines, patch + 1, plen);
2912                         }
2913                         else {
2914                                 ws_fix_copy(&newlines, patch + 1, plen, ws_rule, &state->applied_after_fixing_ws);
2915                         }
2916                         add_line_info(&postimage, newlines.buf + start, newlines.len - start,
2917                                       (first == '+' ? 0 : LINE_COMMON));
2918                         if (first == '+' &&
2919                             (ws_rule & WS_BLANK_AT_EOF) &&
2920                             ws_blank_line(patch + 1, plen, ws_rule))
2921                                 added_blank_line = 1;
2922                         break;
2923                 case '@': case '\\':
2924                         /* Ignore it, we already handled it */
2925                         break;
2926                 default:
2927                         if (state->apply_verbosity > verbosity_normal)
2928                                 error(_("invalid start of line: '%c'"), first);
2929                         applied_pos = -1;
2930                         goto out;
2931                 }
2932                 if (added_blank_line) {
2933                         if (!new_blank_lines_at_end)
2934                                 found_new_blank_lines_at_end = hunk_linenr;
2935                         new_blank_lines_at_end++;
2936                 }
2937                 else if (is_blank_context)
2938                         ;
2939                 else
2940                         new_blank_lines_at_end = 0;
2941                 patch += len;
2942                 size -= len;
2943                 hunk_linenr++;
2944         }
2945         if (inaccurate_eof &&
2946             old > oldlines && old[-1] == '\n' &&
2947             newlines.len > 0 && newlines.buf[newlines.len - 1] == '\n') {
2948                 old--;
2949                 strbuf_setlen(&newlines, newlines.len - 1);
2950         }
2951
2952         leading = frag->leading;
2953         trailing = frag->trailing;
2954
2955         /*
2956          * A hunk to change lines at the beginning would begin with
2957          * @@ -1,L +N,M @@
2958          * but we need to be careful.  -U0 that inserts before the second
2959          * line also has this pattern.
2960          *
2961          * And a hunk to add to an empty file would begin with
2962          * @@ -0,0 +N,M @@
2963          *
2964          * In other words, a hunk that is (frag->oldpos <= 1) with or
2965          * without leading context must match at the beginning.
2966          */
2967         match_beginning = (!frag->oldpos ||
2968                            (frag->oldpos == 1 && !state->unidiff_zero));
2969
2970         /*
2971          * A hunk without trailing lines must match at the end.
2972          * However, we simply cannot tell if a hunk must match end
2973          * from the lack of trailing lines if the patch was generated
2974          * with unidiff without any context.
2975          */
2976         match_end = !state->unidiff_zero && !trailing;
2977
2978         pos = frag->newpos ? (frag->newpos - 1) : 0;
2979         preimage.buf = oldlines;
2980         preimage.len = old - oldlines;
2981         postimage.buf = newlines.buf;
2982         postimage.len = newlines.len;
2983         preimage.line = preimage.line_allocated;
2984         postimage.line = postimage.line_allocated;
2985
2986         for (;;) {
2987
2988                 applied_pos = find_pos(state, img, &preimage, &postimage, pos,
2989                                        ws_rule, match_beginning, match_end);
2990
2991                 if (applied_pos >= 0)
2992                         break;
2993
2994                 /* Am I at my context limits? */
2995                 if ((leading <= state->p_context) && (trailing <= state->p_context))
2996                         break;
2997                 if (match_beginning || match_end) {
2998                         match_beginning = match_end = 0;
2999                         continue;
3000                 }
3001
3002                 /*
3003                  * Reduce the number of context lines; reduce both
3004                  * leading and trailing if they are equal otherwise
3005                  * just reduce the larger context.
3006                  */
3007                 if (leading >= trailing) {
3008                         remove_first_line(&preimage);
3009                         remove_first_line(&postimage);
3010                         pos--;
3011                         leading--;
3012                 }
3013                 if (trailing > leading) {
3014                         remove_last_line(&preimage);
3015                         remove_last_line(&postimage);
3016                         trailing--;
3017                 }
3018         }
3019
3020         if (applied_pos >= 0) {
3021                 if (new_blank_lines_at_end &&
3022                     preimage.nr + applied_pos >= img->nr &&
3023                     (ws_rule & WS_BLANK_AT_EOF) &&
3024                     state->ws_error_action != nowarn_ws_error) {
3025                         record_ws_error(state, WS_BLANK_AT_EOF, "+", 1,
3026                                         found_new_blank_lines_at_end);
3027                         if (state->ws_error_action == correct_ws_error) {
3028                                 while (new_blank_lines_at_end--)
3029                                         remove_last_line(&postimage);
3030                         }
3031                         /*
3032                          * We would want to prevent write_out_results()
3033                          * from taking place in apply_patch() that follows
3034                          * the callchain led us here, which is:
3035                          * apply_patch->check_patch_list->check_patch->
3036                          * apply_data->apply_fragments->apply_one_fragment
3037                          */
3038                         if (state->ws_error_action == die_on_ws_error)
3039                                 state->apply = 0;
3040                 }
3041
3042                 if (state->apply_verbosity > verbosity_normal && applied_pos != pos) {
3043                         int offset = applied_pos - pos;
3044                         if (state->apply_in_reverse)
3045                                 offset = 0 - offset;
3046                         fprintf_ln(stderr,
3047                                    Q_("Hunk #%d succeeded at %d (offset %d line).",
3048                                       "Hunk #%d succeeded at %d (offset %d lines).",
3049                                       offset),
3050                                    nth_fragment, applied_pos + 1, offset);
3051                 }
3052
3053                 /*
3054                  * Warn if it was necessary to reduce the number
3055                  * of context lines.
3056                  */
3057                 if ((leading != frag->leading ||
3058                      trailing != frag->trailing) && state->apply_verbosity > verbosity_silent)
3059                         fprintf_ln(stderr, _("Context reduced to (%ld/%ld)"
3060                                              " to apply fragment at %d"),
3061                                    leading, trailing, applied_pos+1);
3062                 update_image(state, img, applied_pos, &preimage, &postimage);
3063         } else {
3064                 if (state->apply_verbosity > verbosity_normal)
3065                         error(_("while searching for:\n%.*s"),
3066                               (int)(old - oldlines), oldlines);
3067         }
3068
3069 out:
3070         free(oldlines);
3071         strbuf_release(&newlines);
3072         free(preimage.line_allocated);
3073         free(postimage.line_allocated);
3074
3075         return (applied_pos < 0);
3076 }
3077
3078 static int apply_binary_fragment(struct apply_state *state,
3079                                  struct image *img,
3080                                  struct patch *patch)
3081 {
3082         struct fragment *fragment = patch->fragments;
3083         unsigned long len;
3084         void *dst;
3085
3086         if (!fragment)
3087                 return error(_("missing binary patch data for '%s'"),
3088                              patch->new_name ?
3089                              patch->new_name :
3090                              patch->old_name);
3091
3092         /* Binary patch is irreversible without the optional second hunk */
3093         if (state->apply_in_reverse) {
3094                 if (!fragment->next)
3095                         return error(_("cannot reverse-apply a binary patch "
3096                                        "without the reverse hunk to '%s'"),
3097                                      patch->new_name
3098                                      ? patch->new_name : patch->old_name);
3099                 fragment = fragment->next;
3100         }
3101         switch (fragment->binary_patch_method) {
3102         case BINARY_DELTA_DEFLATED:
3103                 dst = patch_delta(img->buf, img->len, fragment->patch,
3104                                   fragment->size, &len);
3105                 if (!dst)
3106                         return -1;
3107                 clear_image(img);
3108                 img->buf = dst;
3109                 img->len = len;
3110                 return 0;
3111         case BINARY_LITERAL_DEFLATED:
3112                 clear_image(img);
3113                 img->len = fragment->size;
3114                 img->buf = xmemdupz(fragment->patch, img->len);
3115                 return 0;
3116         }
3117         return -1;
3118 }
3119
3120 /*
3121  * Replace "img" with the result of applying the binary patch.
3122  * The binary patch data itself in patch->fragment is still kept
3123  * but the preimage prepared by the caller in "img" is freed here
3124  * or in the helper function apply_binary_fragment() this calls.
3125  */
3126 static int apply_binary(struct apply_state *state,
3127                         struct image *img,
3128                         struct patch *patch)
3129 {
3130         const char *name = patch->old_name ? patch->old_name : patch->new_name;
3131         struct object_id oid;
3132
3133         /*
3134          * For safety, we require patch index line to contain
3135          * full 40-byte textual SHA1 for old and new, at least for now.
3136          */
3137         if (strlen(patch->old_sha1_prefix) != 40 ||
3138             strlen(patch->new_sha1_prefix) != 40 ||
3139             get_oid_hex(patch->old_sha1_prefix, &oid) ||
3140             get_oid_hex(patch->new_sha1_prefix, &oid))
3141                 return error(_("cannot apply binary patch to '%s' "
3142                                "without full index line"), name);
3143
3144         if (patch->old_name) {
3145                 /*
3146                  * See if the old one matches what the patch
3147                  * applies to.
3148                  */
3149                 hash_sha1_file(img->buf, img->len, blob_type, oid.hash);
3150                 if (strcmp(oid_to_hex(&oid), patch->old_sha1_prefix))
3151                         return error(_("the patch applies to '%s' (%s), "
3152                                        "which does not match the "
3153                                        "current contents."),
3154                                      name, oid_to_hex(&oid));
3155         }
3156         else {
3157                 /* Otherwise, the old one must be empty. */
3158                 if (img->len)
3159                         return error(_("the patch applies to an empty "
3160                                        "'%s' but it is not empty"), name);
3161         }
3162
3163         get_oid_hex(patch->new_sha1_prefix, &oid);
3164         if (is_null_oid(&oid)) {
3165                 clear_image(img);
3166                 return 0; /* deletion patch */
3167         }
3168
3169         if (has_sha1_file(oid.hash)) {
3170                 /* We already have the postimage */
3171                 enum object_type type;
3172                 unsigned long size;
3173                 char *result;
3174
3175                 result = read_sha1_file(oid.hash, &type, &size);
3176                 if (!result)
3177                         return error(_("the necessary postimage %s for "
3178                                        "'%s' cannot be read"),
3179                                      patch->new_sha1_prefix, name);
3180                 clear_image(img);
3181                 img->buf = result;
3182                 img->len = size;
3183         } else {
3184                 /*
3185                  * We have verified buf matches the preimage;
3186                  * apply the patch data to it, which is stored
3187                  * in the patch->fragments->{patch,size}.
3188                  */
3189                 if (apply_binary_fragment(state, img, patch))
3190                         return error(_("binary patch does not apply to '%s'"),
3191                                      name);
3192
3193                 /* verify that the result matches */
3194                 hash_sha1_file(img->buf, img->len, blob_type, oid.hash);
3195                 if (strcmp(oid_to_hex(&oid), patch->new_sha1_prefix))
3196                         return error(_("binary patch to '%s' creates incorrect result (expecting %s, got %s)"),
3197                                 name, patch->new_sha1_prefix, oid_to_hex(&oid));
3198         }
3199
3200         return 0;
3201 }
3202
3203 static int apply_fragments(struct apply_state *state, struct image *img, struct patch *patch)
3204 {
3205         struct fragment *frag = patch->fragments;
3206         const char *name = patch->old_name ? patch->old_name : patch->new_name;
3207         unsigned ws_rule = patch->ws_rule;
3208         unsigned inaccurate_eof = patch->inaccurate_eof;
3209         int nth = 0;
3210
3211         if (patch->is_binary)
3212                 return apply_binary(state, img, patch);
3213
3214         while (frag) {
3215                 nth++;
3216                 if (apply_one_fragment(state, img, frag, inaccurate_eof, ws_rule, nth)) {
3217                         error(_("patch failed: %s:%ld"), name, frag->oldpos);
3218                         if (!state->apply_with_reject)
3219                                 return -1;
3220                         frag->rejected = 1;
3221                 }
3222                 frag = frag->next;
3223         }
3224         return 0;
3225 }
3226
3227 static int read_blob_object(struct strbuf *buf, const struct object_id *oid, unsigned mode)
3228 {
3229         if (S_ISGITLINK(mode)) {
3230                 strbuf_grow(buf, 100);
3231                 strbuf_addf(buf, "Subproject commit %s\n", oid_to_hex(oid));
3232         } else {
3233                 enum object_type type;
3234                 unsigned long sz;
3235                 char *result;
3236
3237                 result = read_sha1_file(oid->hash, &type, &sz);
3238                 if (!result)
3239                         return -1;
3240                 /* XXX read_sha1_file NUL-terminates */
3241                 strbuf_attach(buf, result, sz, sz + 1);
3242         }
3243         return 0;
3244 }
3245
3246 static int read_file_or_gitlink(const struct cache_entry *ce, struct strbuf *buf)
3247 {
3248         if (!ce)
3249                 return 0;
3250         return read_blob_object(buf, &ce->oid, ce->ce_mode);
3251 }
3252
3253 static struct patch *in_fn_table(struct apply_state *state, const char *name)
3254 {
3255         struct string_list_item *item;
3256
3257         if (name == NULL)
3258                 return NULL;
3259
3260         item = string_list_lookup(&state->fn_table, name);
3261         if (item != NULL)
3262                 return (struct patch *)item->util;
3263
3264         return NULL;
3265 }
3266
3267 /*
3268  * item->util in the filename table records the status of the path.
3269  * Usually it points at a patch (whose result records the contents
3270  * of it after applying it), but it could be PATH_WAS_DELETED for a
3271  * path that a previously applied patch has already removed, or
3272  * PATH_TO_BE_DELETED for a path that a later patch would remove.
3273  *
3274  * The latter is needed to deal with a case where two paths A and B
3275  * are swapped by first renaming A to B and then renaming B to A;
3276  * moving A to B should not be prevented due to presence of B as we
3277  * will remove it in a later patch.
3278  */
3279 #define PATH_TO_BE_DELETED ((struct patch *) -2)
3280 #define PATH_WAS_DELETED ((struct patch *) -1)
3281
3282 static int to_be_deleted(struct patch *patch)
3283 {
3284         return patch == PATH_TO_BE_DELETED;
3285 }
3286
3287 static int was_deleted(struct patch *patch)
3288 {
3289         return patch == PATH_WAS_DELETED;
3290 }
3291
3292 static void add_to_fn_table(struct apply_state *state, struct patch *patch)
3293 {
3294         struct string_list_item *item;
3295
3296         /*
3297          * Always add new_name unless patch is a deletion
3298          * This should cover the cases for normal diffs,
3299          * file creations and copies
3300          */
3301         if (patch->new_name != NULL) {
3302                 item = string_list_insert(&state->fn_table, patch->new_name);
3303                 item->util = patch;
3304         }
3305
3306         /*
3307          * store a failure on rename/deletion cases because
3308          * later chunks shouldn't patch old names
3309          */
3310         if ((patch->new_name == NULL) || (patch->is_rename)) {
3311                 item = string_list_insert(&state->fn_table, patch->old_name);
3312                 item->util = PATH_WAS_DELETED;
3313         }
3314 }
3315
3316 static void prepare_fn_table(struct apply_state *state, struct patch *patch)
3317 {
3318         /*
3319          * store information about incoming file deletion
3320          */
3321         while (patch) {
3322                 if ((patch->new_name == NULL) || (patch->is_rename)) {
3323                         struct string_list_item *item;
3324                         item = string_list_insert(&state->fn_table, patch->old_name);
3325                         item->util = PATH_TO_BE_DELETED;
3326                 }
3327                 patch = patch->next;
3328         }
3329 }
3330
3331 static int checkout_target(struct index_state *istate,
3332                            struct cache_entry *ce, struct stat *st)
3333 {
3334         struct checkout costate = CHECKOUT_INIT;
3335
3336         costate.refresh_cache = 1;
3337         costate.istate = istate;
3338         if (checkout_entry(ce, &costate, NULL) || lstat(ce->name, st))
3339                 return error(_("cannot checkout %s"), ce->name);
3340         return 0;
3341 }
3342
3343 static struct patch *previous_patch(struct apply_state *state,
3344                                     struct patch *patch,
3345                                     int *gone)
3346 {
3347         struct patch *previous;
3348
3349         *gone = 0;
3350         if (patch->is_copy || patch->is_rename)
3351                 return NULL; /* "git" patches do not depend on the order */
3352
3353         previous = in_fn_table(state, patch->old_name);
3354         if (!previous)
3355                 return NULL;
3356
3357         if (to_be_deleted(previous))
3358                 return NULL; /* the deletion hasn't happened yet */
3359
3360         if (was_deleted(previous))
3361                 *gone = 1;
3362
3363         return previous;
3364 }
3365
3366 static int verify_index_match(const struct cache_entry *ce, struct stat *st)
3367 {
3368         if (S_ISGITLINK(ce->ce_mode)) {
3369                 if (!S_ISDIR(st->st_mode))
3370                         return -1;
3371                 return 0;
3372         }
3373         return ce_match_stat(ce, st, CE_MATCH_IGNORE_VALID|CE_MATCH_IGNORE_SKIP_WORKTREE);
3374 }
3375
3376 #define SUBMODULE_PATCH_WITHOUT_INDEX 1
3377
3378 static int load_patch_target(struct apply_state *state,
3379                              struct strbuf *buf,
3380                              const struct cache_entry *ce,
3381                              struct stat *st,
3382                              const char *name,
3383                              unsigned expected_mode)
3384 {
3385         if (state->cached || state->check_index) {
3386                 if (read_file_or_gitlink(ce, buf))
3387                         return error(_("failed to read %s"), name);
3388         } else if (name) {
3389                 if (S_ISGITLINK(expected_mode)) {
3390                         if (ce)
3391                                 return read_file_or_gitlink(ce, buf);
3392                         else
3393                                 return SUBMODULE_PATCH_WITHOUT_INDEX;
3394                 } else if (has_symlink_leading_path(name, strlen(name))) {
3395                         return error(_("reading from '%s' beyond a symbolic link"), name);
3396                 } else {
3397                         if (read_old_data(st, name, buf))
3398                                 return error(_("failed to read %s"), name);
3399                 }
3400         }
3401         return 0;
3402 }
3403
3404 /*
3405  * We are about to apply "patch"; populate the "image" with the
3406  * current version we have, from the working tree or from the index,
3407  * depending on the situation e.g. --cached/--index.  If we are
3408  * applying a non-git patch that incrementally updates the tree,
3409  * we read from the result of a previous diff.
3410  */
3411 static int load_preimage(struct apply_state *state,
3412                          struct image *image,
3413                          struct patch *patch, struct stat *st,
3414                          const struct cache_entry *ce)
3415 {
3416         struct strbuf buf = STRBUF_INIT;
3417         size_t len;
3418         char *img;
3419         struct patch *previous;
3420         int status;
3421
3422         previous = previous_patch(state, patch, &status);
3423         if (status)
3424                 return error(_("path %s has been renamed/deleted"),
3425                              patch->old_name);
3426         if (previous) {
3427                 /* We have a patched copy in memory; use that. */
3428                 strbuf_add(&buf, previous->result, previous->resultsize);
3429         } else {
3430                 status = load_patch_target(state, &buf, ce, st,
3431                                            patch->old_name, patch->old_mode);
3432                 if (status < 0)
3433                         return status;
3434                 else if (status == SUBMODULE_PATCH_WITHOUT_INDEX) {
3435                         /*
3436                          * There is no way to apply subproject
3437                          * patch without looking at the index.
3438                          * NEEDSWORK: shouldn't this be flagged
3439                          * as an error???
3440                          */
3441                         free_fragment_list(patch->fragments);
3442                         patch->fragments = NULL;
3443                 } else if (status) {
3444                         return error(_("failed to read %s"), patch->old_name);
3445                 }
3446         }
3447
3448         img = strbuf_detach(&buf, &len);
3449         prepare_image(image, img, len, !patch->is_binary);
3450         return 0;
3451 }
3452
3453 static int three_way_merge(struct image *image,
3454                            char *path,
3455                            const struct object_id *base,
3456                            const struct object_id *ours,
3457                            const struct object_id *theirs)
3458 {
3459         mmfile_t base_file, our_file, their_file;
3460         mmbuffer_t result = { NULL };
3461         int status;
3462
3463         read_mmblob(&base_file, base);
3464         read_mmblob(&our_file, ours);
3465         read_mmblob(&their_file, theirs);
3466         status = ll_merge(&result, path,
3467                           &base_file, "base",
3468                           &our_file, "ours",
3469                           &their_file, "theirs", NULL);
3470         free(base_file.ptr);
3471         free(our_file.ptr);
3472         free(their_file.ptr);
3473         if (status < 0 || !result.ptr) {
3474                 free(result.ptr);
3475                 return -1;
3476         }
3477         clear_image(image);
3478         image->buf = result.ptr;
3479         image->len = result.size;
3480
3481         return status;
3482 }
3483
3484 /*
3485  * When directly falling back to add/add three-way merge, we read from
3486  * the current contents of the new_name.  In no cases other than that
3487  * this function will be called.
3488  */
3489 static int load_current(struct apply_state *state,
3490                         struct image *image,
3491                         struct patch *patch)
3492 {
3493         struct strbuf buf = STRBUF_INIT;
3494         int status, pos;
3495         size_t len;
3496         char *img;
3497         struct stat st;
3498         struct cache_entry *ce;
3499         char *name = patch->new_name;
3500         unsigned mode = patch->new_mode;
3501
3502         if (!patch->is_new)
3503                 die("BUG: patch to %s is not a creation", patch->old_name);
3504
3505         pos = cache_name_pos(name, strlen(name));
3506         if (pos < 0)
3507                 return error(_("%s: does not exist in index"), name);
3508         ce = active_cache[pos];
3509         if (lstat(name, &st)) {
3510                 if (errno != ENOENT)
3511                         return error_errno("%s", name);
3512                 if (checkout_target(&the_index, ce, &st))
3513                         return -1;
3514         }
3515         if (verify_index_match(ce, &st))
3516                 return error(_("%s: does not match index"), name);
3517
3518         status = load_patch_target(state, &buf, ce, &st, name, mode);
3519         if (status < 0)
3520                 return status;
3521         else if (status)
3522                 return -1;
3523         img = strbuf_detach(&buf, &len);
3524         prepare_image(image, img, len, !patch->is_binary);
3525         return 0;
3526 }
3527
3528 static int try_threeway(struct apply_state *state,
3529                         struct image *image,
3530                         struct patch *patch,
3531                         struct stat *st,
3532                         const struct cache_entry *ce)
3533 {
3534         struct object_id pre_oid, post_oid, our_oid;
3535         struct strbuf buf = STRBUF_INIT;
3536         size_t len;
3537         int status;
3538         char *img;
3539         struct image tmp_image;
3540
3541         /* No point falling back to 3-way merge in these cases */
3542         if (patch->is_delete ||
3543             S_ISGITLINK(patch->old_mode) || S_ISGITLINK(patch->new_mode))
3544                 return -1;
3545
3546         /* Preimage the patch was prepared for */
3547         if (patch->is_new)
3548                 write_sha1_file("", 0, blob_type, pre_oid.hash);
3549         else if (get_oid(patch->old_sha1_prefix, &pre_oid) ||
3550                  read_blob_object(&buf, &pre_oid, patch->old_mode))
3551                 return error(_("repository lacks the necessary blob to fall back on 3-way merge."));
3552
3553         if (state->apply_verbosity > verbosity_silent)
3554                 fprintf(stderr, _("Falling back to three-way merge...\n"));
3555
3556         img = strbuf_detach(&buf, &len);
3557         prepare_image(&tmp_image, img, len, 1);
3558         /* Apply the patch to get the post image */
3559         if (apply_fragments(state, &tmp_image, patch) < 0) {
3560                 clear_image(&tmp_image);
3561                 return -1;
3562         }
3563         /* post_oid is theirs */
3564         write_sha1_file(tmp_image.buf, tmp_image.len, blob_type, post_oid.hash);
3565         clear_image(&tmp_image);
3566
3567         /* our_oid is ours */
3568         if (patch->is_new) {
3569                 if (load_current(state, &tmp_image, patch))
3570                         return error(_("cannot read the current contents of '%s'"),
3571                                      patch->new_name);
3572         } else {
3573                 if (load_preimage(state, &tmp_image, patch, st, ce))
3574                         return error(_("cannot read the current contents of '%s'"),
3575                                      patch->old_name);
3576         }
3577         write_sha1_file(tmp_image.buf, tmp_image.len, blob_type, our_oid.hash);
3578         clear_image(&tmp_image);
3579
3580         /* in-core three-way merge between post and our using pre as base */
3581         status = three_way_merge(image, patch->new_name,
3582                                  &pre_oid, &our_oid, &post_oid);
3583         if (status < 0) {
3584                 if (state->apply_verbosity > verbosity_silent)
3585                         fprintf(stderr,
3586                                 _("Failed to fall back on three-way merge...\n"));
3587                 return status;
3588         }
3589
3590         if (status) {
3591                 patch->conflicted_threeway = 1;
3592                 if (patch->is_new)
3593                         oidclr(&patch->threeway_stage[0]);
3594                 else
3595                         oidcpy(&patch->threeway_stage[0], &pre_oid);
3596                 oidcpy(&patch->threeway_stage[1], &our_oid);
3597                 oidcpy(&patch->threeway_stage[2], &post_oid);
3598                 if (state->apply_verbosity > verbosity_silent)
3599                         fprintf(stderr,
3600                                 _("Applied patch to '%s' with conflicts.\n"),
3601                                 patch->new_name);
3602         } else {
3603                 if (state->apply_verbosity > verbosity_silent)
3604                         fprintf(stderr,
3605                                 _("Applied patch to '%s' cleanly.\n"),
3606                                 patch->new_name);
3607         }
3608         return 0;
3609 }
3610
3611 static int apply_data(struct apply_state *state, struct patch *patch,
3612                       struct stat *st, const struct cache_entry *ce)
3613 {
3614         struct image image;
3615
3616         if (load_preimage(state, &image, patch, st, ce) < 0)
3617                 return -1;
3618
3619         if (patch->direct_to_threeway ||
3620             apply_fragments(state, &image, patch) < 0) {
3621                 /* Note: with --reject, apply_fragments() returns 0 */
3622                 if (!state->threeway || try_threeway(state, &image, patch, st, ce) < 0)
3623                         return -1;
3624         }
3625         patch->result = image.buf;
3626         patch->resultsize = image.len;
3627         add_to_fn_table(state, patch);
3628         free(image.line_allocated);
3629
3630         if (0 < patch->is_delete && patch->resultsize)
3631                 return error(_("removal patch leaves file contents"));
3632
3633         return 0;
3634 }
3635
3636 /*
3637  * If "patch" that we are looking at modifies or deletes what we have,
3638  * we would want it not to lose any local modification we have, either
3639  * in the working tree or in the index.
3640  *
3641  * This also decides if a non-git patch is a creation patch or a
3642  * modification to an existing empty file.  We do not check the state
3643  * of the current tree for a creation patch in this function; the caller
3644  * check_patch() separately makes sure (and errors out otherwise) that
3645  * the path the patch creates does not exist in the current tree.
3646  */
3647 static int check_preimage(struct apply_state *state,
3648                           struct patch *patch,
3649                           struct cache_entry **ce,
3650                           struct stat *st)
3651 {
3652         const char *old_name = patch->old_name;
3653         struct patch *previous = NULL;
3654         int stat_ret = 0, status;
3655         unsigned st_mode = 0;
3656
3657         if (!old_name)
3658                 return 0;
3659
3660         assert(patch->is_new <= 0);
3661         previous = previous_patch(state, patch, &status);
3662
3663         if (status)
3664                 return error(_("path %s has been renamed/deleted"), old_name);
3665         if (previous) {
3666                 st_mode = previous->new_mode;
3667         } else if (!state->cached) {
3668                 stat_ret = lstat(old_name, st);
3669                 if (stat_ret && errno != ENOENT)
3670                         return error_errno("%s", old_name);
3671         }
3672
3673         if (state->check_index && !previous) {
3674                 int pos = cache_name_pos(old_name, strlen(old_name));
3675                 if (pos < 0) {
3676                         if (patch->is_new < 0)
3677                                 goto is_new;
3678                         return error(_("%s: does not exist in index"), old_name);
3679                 }
3680                 *ce = active_cache[pos];
3681                 if (stat_ret < 0) {
3682                         if (checkout_target(&the_index, *ce, st))
3683                                 return -1;
3684                 }
3685                 if (!state->cached && verify_index_match(*ce, st))
3686                         return error(_("%s: does not match index"), old_name);
3687                 if (state->cached)
3688                         st_mode = (*ce)->ce_mode;
3689         } else if (stat_ret < 0) {
3690                 if (patch->is_new < 0)
3691                         goto is_new;
3692                 return error_errno("%s", old_name);
3693         }
3694
3695         if (!state->cached && !previous)
3696                 st_mode = ce_mode_from_stat(*ce, st->st_mode);
3697
3698         if (patch->is_new < 0)
3699                 patch->is_new = 0;
3700         if (!patch->old_mode)
3701                 patch->old_mode = st_mode;
3702         if ((st_mode ^ patch->old_mode) & S_IFMT)
3703                 return error(_("%s: wrong type"), old_name);
3704         if (st_mode != patch->old_mode)
3705                 warning(_("%s has type %o, expected %o"),
3706                         old_name, st_mode, patch->old_mode);
3707         if (!patch->new_mode && !patch->is_delete)
3708                 patch->new_mode = st_mode;
3709         return 0;
3710
3711  is_new:
3712         patch->is_new = 1;
3713         patch->is_delete = 0;
3714         FREE_AND_NULL(patch->old_name);
3715         return 0;
3716 }
3717
3718
3719 #define EXISTS_IN_INDEX 1
3720 #define EXISTS_IN_WORKTREE 2
3721
3722 static int check_to_create(struct apply_state *state,
3723                            const char *new_name,
3724                            int ok_if_exists)
3725 {
3726         struct stat nst;
3727
3728         if (state->check_index &&
3729             cache_name_pos(new_name, strlen(new_name)) >= 0 &&
3730             !ok_if_exists)
3731                 return EXISTS_IN_INDEX;
3732         if (state->cached)
3733                 return 0;
3734
3735         if (!lstat(new_name, &nst)) {
3736                 if (S_ISDIR(nst.st_mode) || ok_if_exists)
3737                         return 0;
3738                 /*
3739                  * A leading component of new_name might be a symlink
3740                  * that is going to be removed with this patch, but
3741                  * still pointing at somewhere that has the path.
3742                  * In such a case, path "new_name" does not exist as
3743                  * far as git is concerned.
3744                  */
3745                 if (has_symlink_leading_path(new_name, strlen(new_name)))
3746                         return 0;
3747
3748                 return EXISTS_IN_WORKTREE;
3749         } else if (!is_missing_file_error(errno)) {
3750                 return error_errno("%s", new_name);
3751         }
3752         return 0;
3753 }
3754
3755 static uintptr_t register_symlink_changes(struct apply_state *state,
3756                                           const char *path,
3757                                           uintptr_t what)
3758 {
3759         struct string_list_item *ent;
3760
3761         ent = string_list_lookup(&state->symlink_changes, path);
3762         if (!ent) {
3763                 ent = string_list_insert(&state->symlink_changes, path);
3764                 ent->util = (void *)0;
3765         }
3766         ent->util = (void *)(what | ((uintptr_t)ent->util));
3767         return (uintptr_t)ent->util;
3768 }
3769
3770 static uintptr_t check_symlink_changes(struct apply_state *state, const char *path)
3771 {
3772         struct string_list_item *ent;
3773
3774         ent = string_list_lookup(&state->symlink_changes, path);
3775         if (!ent)
3776                 return 0;
3777         return (uintptr_t)ent->util;
3778 }
3779
3780 static void prepare_symlink_changes(struct apply_state *state, struct patch *patch)
3781 {
3782         for ( ; patch; patch = patch->next) {
3783                 if ((patch->old_name && S_ISLNK(patch->old_mode)) &&
3784                     (patch->is_rename || patch->is_delete))
3785                         /* the symlink at patch->old_name is removed */
3786                         register_symlink_changes(state, patch->old_name, APPLY_SYMLINK_GOES_AWAY);
3787
3788                 if (patch->new_name && S_ISLNK(patch->new_mode))
3789                         /* the symlink at patch->new_name is created or remains */
3790                         register_symlink_changes(state, patch->new_name, APPLY_SYMLINK_IN_RESULT);
3791         }
3792 }
3793
3794 static int path_is_beyond_symlink_1(struct apply_state *state, struct strbuf *name)
3795 {
3796         do {
3797                 unsigned int change;
3798
3799                 while (--name->len && name->buf[name->len] != '/')
3800                         ; /* scan backwards */
3801                 if (!name->len)
3802                         break;
3803                 name->buf[name->len] = '\0';
3804                 change = check_symlink_changes(state, name->buf);
3805                 if (change & APPLY_SYMLINK_IN_RESULT)
3806                         return 1;
3807                 if (change & APPLY_SYMLINK_GOES_AWAY)
3808                         /*
3809                          * This cannot be "return 0", because we may
3810                          * see a new one created at a higher level.
3811                          */
3812                         continue;
3813
3814                 /* otherwise, check the preimage */
3815                 if (state->check_index) {
3816                         struct cache_entry *ce;
3817
3818                         ce = cache_file_exists(name->buf, name->len, ignore_case);
3819                         if (ce && S_ISLNK(ce->ce_mode))
3820                                 return 1;
3821                 } else {
3822                         struct stat st;
3823                         if (!lstat(name->buf, &st) && S_ISLNK(st.st_mode))
3824                                 return 1;
3825                 }
3826         } while (1);
3827         return 0;
3828 }
3829
3830 static int path_is_beyond_symlink(struct apply_state *state, const char *name_)
3831 {
3832         int ret;
3833         struct strbuf name = STRBUF_INIT;
3834
3835         assert(*name_ != '\0');
3836         strbuf_addstr(&name, name_);
3837         ret = path_is_beyond_symlink_1(state, &name);
3838         strbuf_release(&name);
3839
3840         return ret;
3841 }
3842
3843 static int check_unsafe_path(struct patch *patch)
3844 {
3845         const char *old_name = NULL;
3846         const char *new_name = NULL;
3847         if (patch->is_delete)
3848                 old_name = patch->old_name;
3849         else if (!patch->is_new && !patch->is_copy)
3850                 old_name = patch->old_name;
3851         if (!patch->is_delete)
3852                 new_name = patch->new_name;
3853
3854         if (old_name && !verify_path(old_name))
3855                 return error(_("invalid path '%s'"), old_name);
3856         if (new_name && !verify_path(new_name))
3857                 return error(_("invalid path '%s'"), new_name);
3858         return 0;
3859 }
3860
3861 /*
3862  * Check and apply the patch in-core; leave the result in patch->result
3863  * for the caller to write it out to the final destination.
3864  */
3865 static int check_patch(struct apply_state *state, struct patch *patch)
3866 {
3867         struct stat st;
3868         const char *old_name = patch->old_name;
3869         const char *new_name = patch->new_name;
3870         const char *name = old_name ? old_name : new_name;
3871         struct cache_entry *ce = NULL;
3872         struct patch *tpatch;
3873         int ok_if_exists;
3874         int status;
3875
3876         patch->rejected = 1; /* we will drop this after we succeed */
3877
3878         status = check_preimage(state, patch, &ce, &st);
3879         if (status)
3880                 return status;
3881         old_name = patch->old_name;
3882
3883         /*
3884          * A type-change diff is always split into a patch to delete
3885          * old, immediately followed by a patch to create new (see
3886          * diff.c::run_diff()); in such a case it is Ok that the entry
3887          * to be deleted by the previous patch is still in the working
3888          * tree and in the index.
3889          *
3890          * A patch to swap-rename between A and B would first rename A
3891          * to B and then rename B to A.  While applying the first one,
3892          * the presence of B should not stop A from getting renamed to
3893          * B; ask to_be_deleted() about the later rename.  Removal of
3894          * B and rename from A to B is handled the same way by asking
3895          * was_deleted().
3896          */
3897         if ((tpatch = in_fn_table(state, new_name)) &&
3898             (was_deleted(tpatch) || to_be_deleted(tpatch)))
3899                 ok_if_exists = 1;
3900         else
3901                 ok_if_exists = 0;
3902
3903         if (new_name &&
3904             ((0 < patch->is_new) || patch->is_rename || patch->is_copy)) {
3905                 int err = check_to_create(state, new_name, ok_if_exists);
3906
3907                 if (err && state->threeway) {
3908                         patch->direct_to_threeway = 1;
3909                 } else switch (err) {
3910                 case 0:
3911                         break; /* happy */
3912                 case EXISTS_IN_INDEX:
3913                         return error(_("%s: already exists in index"), new_name);
3914                         break;
3915                 case EXISTS_IN_WORKTREE:
3916                         return error(_("%s: already exists in working directory"),
3917                                      new_name);
3918                 default:
3919                         return err;
3920                 }
3921
3922                 if (!patch->new_mode) {
3923                         if (0 < patch->is_new)
3924                                 patch->new_mode = S_IFREG | 0644;
3925                         else
3926                                 patch->new_mode = patch->old_mode;
3927                 }
3928         }
3929
3930         if (new_name && old_name) {
3931                 int same = !strcmp(old_name, new_name);
3932                 if (!patch->new_mode)
3933                         patch->new_mode = patch->old_mode;
3934                 if ((patch->old_mode ^ patch->new_mode) & S_IFMT) {
3935                         if (same)
3936                                 return error(_("new mode (%o) of %s does not "
3937                                                "match old mode (%o)"),
3938                                         patch->new_mode, new_name,
3939                                         patch->old_mode);
3940                         else
3941                                 return error(_("new mode (%o) of %s does not "
3942                                                "match old mode (%o) of %s"),
3943                                         patch->new_mode, new_name,
3944                                         patch->old_mode, old_name);
3945                 }
3946         }
3947
3948         if (!state->unsafe_paths && check_unsafe_path(patch))
3949                 return -128;
3950
3951         /*
3952          * An attempt to read from or delete a path that is beyond a
3953          * symbolic link will be prevented by load_patch_target() that
3954          * is called at the beginning of apply_data() so we do not
3955          * have to worry about a patch marked with "is_delete" bit
3956          * here.  We however need to make sure that the patch result
3957          * is not deposited to a path that is beyond a symbolic link
3958          * here.
3959          */
3960         if (!patch->is_delete && path_is_beyond_symlink(state, patch->new_name))
3961                 return error(_("affected file '%s' is beyond a symbolic link"),
3962                              patch->new_name);
3963
3964         if (apply_data(state, patch, &st, ce) < 0)
3965                 return error(_("%s: patch does not apply"), name);
3966         patch->rejected = 0;
3967         return 0;
3968 }
3969
3970 static int check_patch_list(struct apply_state *state, struct patch *patch)
3971 {
3972         int err = 0;
3973
3974         prepare_symlink_changes(state, patch);
3975         prepare_fn_table(state, patch);
3976         while (patch) {
3977                 int res;
3978                 if (state->apply_verbosity > verbosity_normal)
3979                         say_patch_name(stderr,
3980                                        _("Checking patch %s..."), patch);
3981                 res = check_patch(state, patch);
3982                 if (res == -128)
3983                         return -128;
3984                 err |= res;
3985                 patch = patch->next;
3986         }
3987         return err;
3988 }
3989
3990 static int read_apply_cache(struct apply_state *state)
3991 {
3992         if (state->index_file)
3993                 return read_cache_from(state->index_file);
3994         else
3995                 return read_cache();
3996 }
3997
3998 /* This function tries to read the object name from the current index */
3999 static int get_current_oid(struct apply_state *state, const char *path,
4000                            struct object_id *oid)
4001 {
4002         int pos;
4003
4004         if (read_apply_cache(state) < 0)
4005                 return -1;
4006         pos = cache_name_pos(path, strlen(path));
4007         if (pos < 0)
4008                 return -1;
4009         oidcpy(oid, &active_cache[pos]->oid);
4010         return 0;
4011 }
4012
4013 static int preimage_oid_in_gitlink_patch(struct patch *p, struct object_id *oid)
4014 {
4015         /*
4016          * A usable gitlink patch has only one fragment (hunk) that looks like:
4017          * @@ -1 +1 @@
4018          * -Subproject commit <old sha1>
4019          * +Subproject commit <new sha1>
4020          * or
4021          * @@ -1 +0,0 @@
4022          * -Subproject commit <old sha1>
4023          * for a removal patch.
4024          */
4025         struct fragment *hunk = p->fragments;
4026         static const char heading[] = "-Subproject commit ";
4027         char *preimage;
4028
4029         if (/* does the patch have only one hunk? */
4030             hunk && !hunk->next &&
4031             /* is its preimage one line? */
4032             hunk->oldpos == 1 && hunk->oldlines == 1 &&
4033             /* does preimage begin with the heading? */
4034             (preimage = memchr(hunk->patch, '\n', hunk->size)) != NULL &&
4035             starts_with(++preimage, heading) &&
4036             /* does it record full SHA-1? */
4037             !get_oid_hex(preimage + sizeof(heading) - 1, oid) &&
4038             preimage[sizeof(heading) + GIT_SHA1_HEXSZ - 1] == '\n' &&
4039             /* does the abbreviated name on the index line agree with it? */
4040             starts_with(preimage + sizeof(heading) - 1, p->old_sha1_prefix))
4041                 return 0; /* it all looks fine */
4042
4043         /* we may have full object name on the index line */
4044         return get_oid_hex(p->old_sha1_prefix, oid);
4045 }
4046
4047 /* Build an index that contains the just the files needed for a 3way merge */
4048 static int build_fake_ancestor(struct apply_state *state, struct patch *list)
4049 {
4050         struct patch *patch;
4051         struct index_state result = { NULL };
4052         static struct lock_file lock;
4053         int res;
4054
4055         /* Once we start supporting the reverse patch, it may be
4056          * worth showing the new sha1 prefix, but until then...
4057          */
4058         for (patch = list; patch; patch = patch->next) {
4059                 struct object_id oid;
4060                 struct cache_entry *ce;
4061                 const char *name;
4062
4063                 name = patch->old_name ? patch->old_name : patch->new_name;
4064                 if (0 < patch->is_new)
4065                         continue;
4066
4067                 if (S_ISGITLINK(patch->old_mode)) {
4068                         if (!preimage_oid_in_gitlink_patch(patch, &oid))
4069                                 ; /* ok, the textual part looks sane */
4070                         else
4071                                 return error(_("sha1 information is lacking or "
4072                                                "useless for submodule %s"), name);
4073                 } else if (!get_oid_blob(patch->old_sha1_prefix, &oid)) {
4074                         ; /* ok */
4075                 } else if (!patch->lines_added && !patch->lines_deleted) {
4076                         /* mode-only change: update the current */
4077                         if (get_current_oid(state, patch->old_name, &oid))
4078                                 return error(_("mode change for %s, which is not "
4079                                                "in current HEAD"), name);
4080                 } else
4081                         return error(_("sha1 information is lacking or useless "
4082                                        "(%s)."), name);
4083
4084                 ce = make_cache_entry(patch->old_mode, oid.hash, name, 0, 0);
4085                 if (!ce)
4086                         return error(_("make_cache_entry failed for path '%s'"),
4087                                      name);
4088                 if (add_index_entry(&result, ce, ADD_CACHE_OK_TO_ADD)) {
4089                         free(ce);
4090                         return error(_("could not add %s to temporary index"),
4091                                      name);
4092                 }
4093         }
4094
4095         hold_lock_file_for_update(&lock, state->fake_ancestor, LOCK_DIE_ON_ERROR);
4096         res = write_locked_index(&result, &lock, COMMIT_LOCK);
4097         discard_index(&result);
4098
4099         if (res)
4100                 return error(_("could not write temporary index to %s"),
4101                              state->fake_ancestor);
4102
4103         return 0;
4104 }
4105
4106 static void stat_patch_list(struct apply_state *state, struct patch *patch)
4107 {
4108         int files, adds, dels;
4109
4110         for (files = adds = dels = 0 ; patch ; patch = patch->next) {
4111                 files++;
4112                 adds += patch->lines_added;
4113                 dels += patch->lines_deleted;
4114                 show_stats(state, patch);
4115         }
4116
4117         print_stat_summary(stdout, files, adds, dels);
4118 }
4119
4120 static void numstat_patch_list(struct apply_state *state,
4121                                struct patch *patch)
4122 {
4123         for ( ; patch; patch = patch->next) {
4124                 const char *name;
4125                 name = patch->new_name ? patch->new_name : patch->old_name;
4126                 if (patch->is_binary)
4127                         printf("-\t-\t");
4128                 else
4129                         printf("%d\t%d\t", patch->lines_added, patch->lines_deleted);
4130                 write_name_quoted(name, stdout, state->line_termination);
4131         }
4132 }
4133
4134 static void show_file_mode_name(const char *newdelete, unsigned int mode, const char *name)
4135 {
4136         if (mode)
4137                 printf(" %s mode %06o %s\n", newdelete, mode, name);
4138         else
4139                 printf(" %s %s\n", newdelete, name);
4140 }
4141
4142 static void show_mode_change(struct patch *p, int show_name)
4143 {
4144         if (p->old_mode && p->new_mode && p->old_mode != p->new_mode) {
4145                 if (show_name)
4146                         printf(" mode change %06o => %06o %s\n",
4147                                p->old_mode, p->new_mode, p->new_name);
4148                 else
4149                         printf(" mode change %06o => %06o\n",
4150                                p->old_mode, p->new_mode);
4151         }
4152 }
4153
4154 static void show_rename_copy(struct patch *p)
4155 {
4156         const char *renamecopy = p->is_rename ? "rename" : "copy";
4157         const char *old, *new;
4158
4159         /* Find common prefix */
4160         old = p->old_name;
4161         new = p->new_name;
4162         while (1) {
4163                 const char *slash_old, *slash_new;
4164                 slash_old = strchr(old, '/');
4165                 slash_new = strchr(new, '/');
4166                 if (!slash_old ||
4167                     !slash_new ||
4168                     slash_old - old != slash_new - new ||
4169                     memcmp(old, new, slash_new - new))
4170                         break;
4171                 old = slash_old + 1;
4172                 new = slash_new + 1;
4173         }
4174         /* p->old_name thru old is the common prefix, and old and new
4175          * through the end of names are renames
4176          */
4177         if (old != p->old_name)
4178                 printf(" %s %.*s{%s => %s} (%d%%)\n", renamecopy,
4179                        (int)(old - p->old_name), p->old_name,
4180                        old, new, p->score);
4181         else
4182                 printf(" %s %s => %s (%d%%)\n", renamecopy,
4183                        p->old_name, p->new_name, p->score);
4184         show_mode_change(p, 0);
4185 }
4186
4187 static void summary_patch_list(struct patch *patch)
4188 {
4189         struct patch *p;
4190
4191         for (p = patch; p; p = p->next) {
4192                 if (p->is_new)
4193                         show_file_mode_name("create", p->new_mode, p->new_name);
4194                 else if (p->is_delete)
4195                         show_file_mode_name("delete", p->old_mode, p->old_name);
4196                 else {
4197                         if (p->is_rename || p->is_copy)
4198                                 show_rename_copy(p);
4199                         else {
4200                                 if (p->score) {
4201                                         printf(" rewrite %s (%d%%)\n",
4202                                                p->new_name, p->score);
4203                                         show_mode_change(p, 0);
4204                                 }
4205                                 else
4206                                         show_mode_change(p, 1);
4207                         }
4208                 }
4209         }
4210 }
4211
4212 static void patch_stats(struct apply_state *state, struct patch *patch)
4213 {
4214         int lines = patch->lines_added + patch->lines_deleted;
4215
4216         if (lines > state->max_change)
4217                 state->max_change = lines;
4218         if (patch->old_name) {
4219                 int len = quote_c_style(patch->old_name, NULL, NULL, 0);
4220                 if (!len)
4221                         len = strlen(patch->old_name);
4222                 if (len > state->max_len)
4223                         state->max_len = len;
4224         }
4225         if (patch->new_name) {
4226                 int len = quote_c_style(patch->new_name, NULL, NULL, 0);
4227                 if (!len)
4228                         len = strlen(patch->new_name);
4229                 if (len > state->max_len)
4230                         state->max_len = len;
4231         }
4232 }
4233
4234 static int remove_file(struct apply_state *state, struct patch *patch, int rmdir_empty)
4235 {
4236         if (state->update_index) {
4237                 if (remove_file_from_cache(patch->old_name) < 0)
4238                         return error(_("unable to remove %s from index"), patch->old_name);
4239         }
4240         if (!state->cached) {
4241                 if (!remove_or_warn(patch->old_mode, patch->old_name) && rmdir_empty) {
4242                         remove_path(patch->old_name);
4243                 }
4244         }
4245         return 0;
4246 }
4247
4248 static int add_index_file(struct apply_state *state,
4249                           const char *path,
4250                           unsigned mode,
4251                           void *buf,
4252                           unsigned long size)
4253 {
4254         struct stat st;
4255         struct cache_entry *ce;
4256         int namelen = strlen(path);
4257         unsigned ce_size = cache_entry_size(namelen);
4258
4259         if (!state->update_index)
4260                 return 0;
4261
4262         ce = xcalloc(1, ce_size);
4263         memcpy(ce->name, path, namelen);
4264         ce->ce_mode = create_ce_mode(mode);
4265         ce->ce_flags = create_ce_flags(0);
4266         ce->ce_namelen = namelen;
4267         if (S_ISGITLINK(mode)) {
4268                 const char *s;
4269
4270                 if (!skip_prefix(buf, "Subproject commit ", &s) ||
4271                     get_oid_hex(s, &ce->oid)) {
4272                         free(ce);
4273                        return error(_("corrupt patch for submodule %s"), path);
4274                 }
4275         } else {
4276                 if (!state->cached) {
4277                         if (lstat(path, &st) < 0) {
4278                                 free(ce);
4279                                 return error_errno(_("unable to stat newly "
4280                                                      "created file '%s'"),
4281                                                    path);
4282                         }
4283                         fill_stat_cache_info(ce, &st);
4284                 }
4285                 if (write_sha1_file(buf, size, blob_type, ce->oid.hash) < 0) {
4286                         free(ce);
4287                         return error(_("unable to create backing store "
4288                                        "for newly created file %s"), path);
4289                 }
4290         }
4291         if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD) < 0) {
4292                 free(ce);
4293                 return error(_("unable to add cache entry for %s"), path);
4294         }
4295
4296         return 0;
4297 }
4298
4299 /*
4300  * Returns:
4301  *  -1 if an unrecoverable error happened
4302  *   0 if everything went well
4303  *   1 if a recoverable error happened
4304  */
4305 static int try_create_file(const char *path, unsigned int mode, const char *buf, unsigned long size)
4306 {
4307         int fd, res;
4308         struct strbuf nbuf = STRBUF_INIT;
4309
4310         if (S_ISGITLINK(mode)) {
4311                 struct stat st;
4312                 if (!lstat(path, &st) && S_ISDIR(st.st_mode))
4313                         return 0;
4314                 return !!mkdir(path, 0777);
4315         }
4316
4317         if (has_symlinks && S_ISLNK(mode))
4318                 /* Although buf:size is counted string, it also is NUL
4319                  * terminated.
4320                  */
4321                 return !!symlink(buf, path);
4322
4323         fd = open(path, O_CREAT | O_EXCL | O_WRONLY, (mode & 0100) ? 0777 : 0666);
4324         if (fd < 0)
4325                 return 1;
4326
4327         if (convert_to_working_tree(path, buf, size, &nbuf)) {
4328                 size = nbuf.len;
4329                 buf  = nbuf.buf;
4330         }
4331
4332         res = write_in_full(fd, buf, size) < 0;
4333         if (res)
4334                 error_errno(_("failed to write to '%s'"), path);
4335         strbuf_release(&nbuf);
4336
4337         if (close(fd) < 0 && !res)
4338                 return error_errno(_("closing file '%s'"), path);
4339
4340         return res ? -1 : 0;
4341 }
4342
4343 /*
4344  * We optimistically assume that the directories exist,
4345  * which is true 99% of the time anyway. If they don't,
4346  * we create them and try again.
4347  *
4348  * Returns:
4349  *   -1 on error
4350  *   0 otherwise
4351  */
4352 static int create_one_file(struct apply_state *state,
4353                            char *path,
4354                            unsigned mode,
4355                            const char *buf,
4356                            unsigned long size)
4357 {
4358         int res;
4359
4360         if (state->cached)
4361                 return 0;
4362
4363         res = try_create_file(path, mode, buf, size);
4364         if (res < 0)
4365                 return -1;
4366         if (!res)
4367                 return 0;
4368
4369         if (errno == ENOENT) {
4370                 if (safe_create_leading_directories(path))
4371                         return 0;
4372                 res = try_create_file(path, mode, buf, size);
4373                 if (res < 0)
4374                         return -1;
4375                 if (!res)
4376                         return 0;
4377         }
4378
4379         if (errno == EEXIST || errno == EACCES) {
4380                 /* We may be trying to create a file where a directory
4381                  * used to be.
4382                  */
4383                 struct stat st;
4384                 if (!lstat(path, &st) && (!S_ISDIR(st.st_mode) || !rmdir(path)))
4385                         errno = EEXIST;
4386         }
4387
4388         if (errno == EEXIST) {
4389                 unsigned int nr = getpid();
4390
4391                 for (;;) {
4392                         char newpath[PATH_MAX];
4393                         mksnpath(newpath, sizeof(newpath), "%s~%u", path, nr);
4394                         res = try_create_file(newpath, mode, buf, size);
4395                         if (res < 0)
4396                                 return -1;
4397                         if (!res) {
4398                                 if (!rename(newpath, path))
4399                                         return 0;
4400                                 unlink_or_warn(newpath);
4401                                 break;
4402                         }
4403                         if (errno != EEXIST)
4404                                 break;
4405                         ++nr;
4406                 }
4407         }
4408         return error_errno(_("unable to write file '%s' mode %o"),
4409                            path, mode);
4410 }
4411
4412 static int add_conflicted_stages_file(struct apply_state *state,
4413                                        struct patch *patch)
4414 {
4415         int stage, namelen;
4416         unsigned ce_size, mode;
4417         struct cache_entry *ce;
4418
4419         if (!state->update_index)
4420                 return 0;
4421         namelen = strlen(patch->new_name);
4422         ce_size = cache_entry_size(namelen);
4423         mode = patch->new_mode ? patch->new_mode : (S_IFREG | 0644);
4424
4425         remove_file_from_cache(patch->new_name);
4426         for (stage = 1; stage < 4; stage++) {
4427                 if (is_null_oid(&patch->threeway_stage[stage - 1]))
4428                         continue;
4429                 ce = xcalloc(1, ce_size);
4430                 memcpy(ce->name, patch->new_name, namelen);
4431                 ce->ce_mode = create_ce_mode(mode);
4432                 ce->ce_flags = create_ce_flags(stage);
4433                 ce->ce_namelen = namelen;
4434                 oidcpy(&ce->oid, &patch->threeway_stage[stage - 1]);
4435                 if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD) < 0) {
4436                         free(ce);
4437                         return error(_("unable to add cache entry for %s"),
4438                                      patch->new_name);
4439                 }
4440         }
4441
4442         return 0;
4443 }
4444
4445 static int create_file(struct apply_state *state, struct patch *patch)
4446 {
4447         char *path = patch->new_name;
4448         unsigned mode = patch->new_mode;
4449         unsigned long size = patch->resultsize;
4450         char *buf = patch->result;
4451
4452         if (!mode)
4453                 mode = S_IFREG | 0644;
4454         if (create_one_file(state, path, mode, buf, size))
4455                 return -1;
4456
4457         if (patch->conflicted_threeway)
4458                 return add_conflicted_stages_file(state, patch);
4459         else
4460                 return add_index_file(state, path, mode, buf, size);
4461 }
4462
4463 /* phase zero is to remove, phase one is to create */
4464 static int write_out_one_result(struct apply_state *state,
4465                                 struct patch *patch,
4466                                 int phase)
4467 {
4468         if (patch->is_delete > 0) {
4469                 if (phase == 0)
4470                         return remove_file(state, patch, 1);
4471                 return 0;
4472         }
4473         if (patch->is_new > 0 || patch->is_copy) {
4474                 if (phase == 1)
4475                         return create_file(state, patch);
4476                 return 0;
4477         }
4478         /*
4479          * Rename or modification boils down to the same
4480          * thing: remove the old, write the new
4481          */
4482         if (phase == 0)
4483                 return remove_file(state, patch, patch->is_rename);
4484         if (phase == 1)
4485                 return create_file(state, patch);
4486         return 0;
4487 }
4488
4489 static int write_out_one_reject(struct apply_state *state, struct patch *patch)
4490 {
4491         FILE *rej;
4492         char namebuf[PATH_MAX];
4493         struct fragment *frag;
4494         int cnt = 0;
4495         struct strbuf sb = STRBUF_INIT;
4496
4497         for (cnt = 0, frag = patch->fragments; frag; frag = frag->next) {
4498                 if (!frag->rejected)
4499                         continue;
4500                 cnt++;
4501         }
4502
4503         if (!cnt) {
4504                 if (state->apply_verbosity > verbosity_normal)
4505                         say_patch_name(stderr,
4506                                        _("Applied patch %s cleanly."), patch);
4507                 return 0;
4508         }
4509
4510         /* This should not happen, because a removal patch that leaves
4511          * contents are marked "rejected" at the patch level.
4512          */
4513         if (!patch->new_name)
4514                 die(_("internal error"));
4515
4516         /* Say this even without --verbose */
4517         strbuf_addf(&sb, Q_("Applying patch %%s with %d reject...",
4518                             "Applying patch %%s with %d rejects...",
4519                             cnt),
4520                     cnt);
4521         if (state->apply_verbosity > verbosity_silent)
4522                 say_patch_name(stderr, sb.buf, patch);
4523         strbuf_release(&sb);
4524
4525         cnt = strlen(patch->new_name);
4526         if (ARRAY_SIZE(namebuf) <= cnt + 5) {
4527                 cnt = ARRAY_SIZE(namebuf) - 5;
4528                 warning(_("truncating .rej filename to %.*s.rej"),
4529                         cnt - 1, patch->new_name);
4530         }
4531         memcpy(namebuf, patch->new_name, cnt);
4532         memcpy(namebuf + cnt, ".rej", 5);
4533
4534         rej = fopen(namebuf, "w");
4535         if (!rej)
4536                 return error_errno(_("cannot open %s"), namebuf);
4537
4538         /* Normal git tools never deal with .rej, so do not pretend
4539          * this is a git patch by saying --git or giving extended
4540          * headers.  While at it, maybe please "kompare" that wants
4541          * the trailing TAB and some garbage at the end of line ;-).
4542          */
4543         fprintf(rej, "diff a/%s b/%s\t(rejected hunks)\n",
4544                 patch->new_name, patch->new_name);
4545         for (cnt = 1, frag = patch->fragments;
4546              frag;
4547              cnt++, frag = frag->next) {
4548                 if (!frag->rejected) {
4549                         if (state->apply_verbosity > verbosity_silent)
4550                                 fprintf_ln(stderr, _("Hunk #%d applied cleanly."), cnt);
4551                         continue;
4552                 }
4553                 if (state->apply_verbosity > verbosity_silent)
4554                         fprintf_ln(stderr, _("Rejected hunk #%d."), cnt);
4555                 fprintf(rej, "%.*s", frag->size, frag->patch);
4556                 if (frag->patch[frag->size-1] != '\n')
4557                         fputc('\n', rej);
4558         }
4559         fclose(rej);
4560         return -1;
4561 }
4562
4563 /*
4564  * Returns:
4565  *  -1 if an error happened
4566  *   0 if the patch applied cleanly
4567  *   1 if the patch did not apply cleanly
4568  */
4569 static int write_out_results(struct apply_state *state, struct patch *list)
4570 {
4571         int phase;
4572         int errs = 0;
4573         struct patch *l;
4574         struct string_list cpath = STRING_LIST_INIT_DUP;
4575
4576         for (phase = 0; phase < 2; phase++) {
4577                 l = list;
4578                 while (l) {
4579                         if (l->rejected)
4580                                 errs = 1;
4581                         else {
4582                                 if (write_out_one_result(state, l, phase)) {
4583                                         string_list_clear(&cpath, 0);
4584                                         return -1;
4585                                 }
4586                                 if (phase == 1) {
4587                                         if (write_out_one_reject(state, l))
4588                                                 errs = 1;
4589                                         if (l->conflicted_threeway) {
4590                                                 string_list_append(&cpath, l->new_name);
4591                                                 errs = 1;
4592                                         }
4593                                 }
4594                         }
4595                         l = l->next;
4596                 }
4597         }
4598
4599         if (cpath.nr) {
4600                 struct string_list_item *item;
4601
4602                 string_list_sort(&cpath);
4603                 if (state->apply_verbosity > verbosity_silent) {
4604                         for_each_string_list_item(item, &cpath)
4605                                 fprintf(stderr, "U %s\n", item->string);
4606                 }
4607                 string_list_clear(&cpath, 0);
4608
4609                 rerere(0);
4610         }
4611
4612         return errs;
4613 }
4614
4615 /*
4616  * Try to apply a patch.
4617  *
4618  * Returns:
4619  *  -128 if a bad error happened (like patch unreadable)
4620  *  -1 if patch did not apply and user cannot deal with it
4621  *   0 if the patch applied
4622  *   1 if the patch did not apply but user might fix it
4623  */
4624 static int apply_patch(struct apply_state *state,
4625                        int fd,
4626                        const char *filename,
4627                        int options)
4628 {
4629         size_t offset;
4630         struct strbuf buf = STRBUF_INIT; /* owns the patch text */
4631         struct patch *list = NULL, **listp = &list;
4632         int skipped_patch = 0;
4633         int res = 0;
4634
4635         state->patch_input_file = filename;
4636         if (read_patch_file(&buf, fd) < 0)
4637                 return -128;
4638         offset = 0;
4639         while (offset < buf.len) {
4640                 struct patch *patch;
4641                 int nr;
4642
4643                 patch = xcalloc(1, sizeof(*patch));
4644                 patch->inaccurate_eof = !!(options & APPLY_OPT_INACCURATE_EOF);
4645                 patch->recount =  !!(options & APPLY_OPT_RECOUNT);
4646                 nr = parse_chunk(state, buf.buf + offset, buf.len - offset, patch);
4647                 if (nr < 0) {
4648                         free_patch(patch);
4649                         if (nr == -128) {
4650                                 res = -128;
4651                                 goto end;
4652                         }
4653                         break;
4654                 }
4655                 if (state->apply_in_reverse)
4656                         reverse_patches(patch);
4657                 if (use_patch(state, patch)) {
4658                         patch_stats(state, patch);
4659                         *listp = patch;
4660                         listp = &patch->next;
4661                 }
4662                 else {
4663                         if (state->apply_verbosity > verbosity_normal)
4664                                 say_patch_name(stderr, _("Skipped patch '%s'."), patch);
4665                         free_patch(patch);
4666                         skipped_patch++;
4667                 }
4668                 offset += nr;
4669         }
4670
4671         if (!list && !skipped_patch) {
4672                 error(_("unrecognized input"));
4673                 res = -128;
4674                 goto end;
4675         }
4676
4677         if (state->whitespace_error && (state->ws_error_action == die_on_ws_error))
4678                 state->apply = 0;
4679
4680         state->update_index = state->check_index && state->apply;
4681         if (state->update_index && state->newfd < 0) {
4682                 if (state->index_file)
4683                         state->newfd = hold_lock_file_for_update(state->lock_file,
4684                                                                  state->index_file,
4685                                                                  LOCK_DIE_ON_ERROR);
4686                 else
4687                         state->newfd = hold_locked_index(state->lock_file, LOCK_DIE_ON_ERROR);
4688         }
4689
4690         if (state->check_index && read_apply_cache(state) < 0) {
4691                 error(_("unable to read index file"));
4692                 res = -128;
4693                 goto end;
4694         }
4695
4696         if (state->check || state->apply) {
4697                 int r = check_patch_list(state, list);
4698                 if (r == -128) {
4699                         res = -128;
4700                         goto end;
4701                 }
4702                 if (r < 0 && !state->apply_with_reject) {
4703                         res = -1;
4704                         goto end;
4705                 }
4706         }
4707
4708         if (state->apply) {
4709                 int write_res = write_out_results(state, list);
4710                 if (write_res < 0) {
4711                         res = -128;
4712                         goto end;
4713                 }
4714                 if (write_res > 0) {
4715                         /* with --3way, we still need to write the index out */
4716                         res = state->apply_with_reject ? -1 : 1;
4717                         goto end;
4718                 }
4719         }
4720
4721         if (state->fake_ancestor &&
4722             build_fake_ancestor(state, list)) {
4723                 res = -128;
4724                 goto end;
4725         }
4726
4727         if (state->diffstat && state->apply_verbosity > verbosity_silent)
4728                 stat_patch_list(state, list);
4729
4730         if (state->numstat && state->apply_verbosity > verbosity_silent)
4731                 numstat_patch_list(state, list);
4732
4733         if (state->summary && state->apply_verbosity > verbosity_silent)
4734                 summary_patch_list(list);
4735
4736 end:
4737         free_patch_list(list);
4738         strbuf_release(&buf);
4739         string_list_clear(&state->fn_table, 0);
4740         return res;
4741 }
4742
4743 static int apply_option_parse_exclude(const struct option *opt,
4744                                       const char *arg, int unset)
4745 {
4746         struct apply_state *state = opt->value;
4747         add_name_limit(state, arg, 1);
4748         return 0;
4749 }
4750
4751 static int apply_option_parse_include(const struct option *opt,
4752                                       const char *arg, int unset)
4753 {
4754         struct apply_state *state = opt->value;
4755         add_name_limit(state, arg, 0);
4756         state->has_include = 1;
4757         return 0;
4758 }
4759
4760 static int apply_option_parse_p(const struct option *opt,
4761                                 const char *arg,
4762                                 int unset)
4763 {
4764         struct apply_state *state = opt->value;
4765         state->p_value = atoi(arg);
4766         state->p_value_known = 1;
4767         return 0;
4768 }
4769
4770 static int apply_option_parse_space_change(const struct option *opt,
4771                                            const char *arg, int unset)
4772 {
4773         struct apply_state *state = opt->value;
4774         if (unset)
4775                 state->ws_ignore_action = ignore_ws_none;
4776         else
4777                 state->ws_ignore_action = ignore_ws_change;
4778         return 0;
4779 }
4780
4781 static int apply_option_parse_whitespace(const struct option *opt,
4782                                          const char *arg, int unset)
4783 {
4784         struct apply_state *state = opt->value;
4785         state->whitespace_option = arg;
4786         if (parse_whitespace_option(state, arg))
4787                 exit(1);
4788         return 0;
4789 }
4790
4791 static int apply_option_parse_directory(const struct option *opt,
4792                                         const char *arg, int unset)
4793 {
4794         struct apply_state *state = opt->value;
4795         strbuf_reset(&state->root);
4796         strbuf_addstr(&state->root, arg);
4797         strbuf_complete(&state->root, '/');
4798         return 0;
4799 }
4800
4801 int apply_all_patches(struct apply_state *state,
4802                       int argc,
4803                       const char **argv,
4804                       int options)
4805 {
4806         int i;
4807         int res;
4808         int errs = 0;
4809         int read_stdin = 1;
4810
4811         for (i = 0; i < argc; i++) {
4812                 const char *arg = argv[i];
4813                 char *to_free = NULL;
4814                 int fd;
4815
4816                 if (!strcmp(arg, "-")) {
4817                         res = apply_patch(state, 0, "<stdin>", options);
4818                         if (res < 0)
4819                                 goto end;
4820                         errs |= res;
4821                         read_stdin = 0;
4822                         continue;
4823                 } else
4824                         arg = to_free = prefix_filename(state->prefix, arg);
4825
4826                 fd = open(arg, O_RDONLY);
4827                 if (fd < 0) {
4828                         error(_("can't open patch '%s': %s"), arg, strerror(errno));
4829                         res = -128;
4830                         free(to_free);
4831                         goto end;
4832                 }
4833                 read_stdin = 0;
4834                 set_default_whitespace_mode(state);
4835                 res = apply_patch(state, fd, arg, options);
4836                 close(fd);
4837                 free(to_free);
4838                 if (res < 0)
4839                         goto end;
4840                 errs |= res;
4841         }
4842         set_default_whitespace_mode(state);
4843         if (read_stdin) {
4844                 res = apply_patch(state, 0, "<stdin>", options);
4845                 if (res < 0)
4846                         goto end;
4847                 errs |= res;
4848         }
4849
4850         if (state->whitespace_error) {
4851                 if (state->squelch_whitespace_errors &&
4852                     state->squelch_whitespace_errors < state->whitespace_error) {
4853                         int squelched =
4854                                 state->whitespace_error - state->squelch_whitespace_errors;
4855                         warning(Q_("squelched %d whitespace error",
4856                                    "squelched %d whitespace errors",
4857                                    squelched),
4858                                 squelched);
4859                 }
4860                 if (state->ws_error_action == die_on_ws_error) {
4861                         error(Q_("%d line adds whitespace errors.",
4862                                  "%d lines add whitespace errors.",
4863                                  state->whitespace_error),
4864                               state->whitespace_error);
4865                         res = -128;
4866                         goto end;
4867                 }
4868                 if (state->applied_after_fixing_ws && state->apply)
4869                         warning(Q_("%d line applied after"
4870                                    " fixing whitespace errors.",
4871                                    "%d lines applied after"
4872                                    " fixing whitespace errors.",
4873                                    state->applied_after_fixing_ws),
4874                                 state->applied_after_fixing_ws);
4875                 else if (state->whitespace_error)
4876                         warning(Q_("%d line adds whitespace errors.",
4877                                    "%d lines add whitespace errors.",
4878                                    state->whitespace_error),
4879                                 state->whitespace_error);
4880         }
4881
4882         if (state->update_index) {
4883                 res = write_locked_index(&the_index, state->lock_file, COMMIT_LOCK);
4884                 if (res) {
4885                         error(_("Unable to write new index file"));
4886                         res = -128;
4887                         goto end;
4888                 }
4889                 state->newfd = -1;
4890         }
4891
4892         res = !!errs;
4893
4894 end:
4895         if (state->newfd >= 0) {
4896                 rollback_lock_file(state->lock_file);
4897                 state->newfd = -1;
4898         }
4899
4900         if (state->apply_verbosity <= verbosity_silent) {
4901                 set_error_routine(state->saved_error_routine);
4902                 set_warn_routine(state->saved_warn_routine);
4903         }
4904
4905         if (res > -1)
4906                 return res;
4907         return (res == -1 ? 1 : 128);
4908 }
4909
4910 int apply_parse_options(int argc, const char **argv,
4911                         struct apply_state *state,
4912                         int *force_apply, int *options,
4913                         const char * const *apply_usage)
4914 {
4915         struct option builtin_apply_options[] = {
4916                 { OPTION_CALLBACK, 0, "exclude", state, N_("path"),
4917                         N_("don't apply changes matching the given path"),
4918                         0, apply_option_parse_exclude },
4919                 { OPTION_CALLBACK, 0, "include", state, N_("path"),
4920                         N_("apply changes matching the given path"),
4921                         0, apply_option_parse_include },
4922                 { OPTION_CALLBACK, 'p', NULL, state, N_("num"),
4923                         N_("remove <num> leading slashes from traditional diff paths"),
4924                         0, apply_option_parse_p },
4925                 OPT_BOOL(0, "no-add", &state->no_add,
4926                         N_("ignore additions made by the patch")),
4927                 OPT_BOOL(0, "stat", &state->diffstat,
4928                         N_("instead of applying the patch, output diffstat for the input")),
4929                 OPT_NOOP_NOARG(0, "allow-binary-replacement"),
4930                 OPT_NOOP_NOARG(0, "binary"),
4931                 OPT_BOOL(0, "numstat", &state->numstat,
4932                         N_("show number of added and deleted lines in decimal notation")),
4933                 OPT_BOOL(0, "summary", &state->summary,
4934                         N_("instead of applying the patch, output a summary for the input")),
4935                 OPT_BOOL(0, "check", &state->check,
4936                         N_("instead of applying the patch, see if the patch is applicable")),
4937                 OPT_BOOL(0, "index", &state->check_index,
4938                         N_("make sure the patch is applicable to the current index")),
4939                 OPT_BOOL(0, "cached", &state->cached,
4940                         N_("apply a patch without touching the working tree")),
4941                 OPT_BOOL(0, "unsafe-paths", &state->unsafe_paths,
4942                         N_("accept a patch that touches outside the working area")),
4943                 OPT_BOOL(0, "apply", force_apply,
4944                         N_("also apply the patch (use with --stat/--summary/--check)")),
4945                 OPT_BOOL('3', "3way", &state->threeway,
4946                          N_( "attempt three-way merge if a patch does not apply")),
4947                 OPT_FILENAME(0, "build-fake-ancestor", &state->fake_ancestor,
4948                         N_("build a temporary index based on embedded index information")),
4949                 /* Think twice before adding "--nul" synonym to this */
4950                 OPT_SET_INT('z', NULL, &state->line_termination,
4951                         N_("paths are separated with NUL character"), '\0'),
4952                 OPT_INTEGER('C', NULL, &state->p_context,
4953                                 N_("ensure at least <n> lines of context match")),
4954                 { OPTION_CALLBACK, 0, "whitespace", state, N_("action"),
4955                         N_("detect new or modified lines that have whitespace errors"),
4956                         0, apply_option_parse_whitespace },
4957                 { OPTION_CALLBACK, 0, "ignore-space-change", state, NULL,
4958                         N_("ignore changes in whitespace when finding context"),
4959                         PARSE_OPT_NOARG, apply_option_parse_space_change },
4960                 { OPTION_CALLBACK, 0, "ignore-whitespace", state, NULL,
4961                         N_("ignore changes in whitespace when finding context"),
4962                         PARSE_OPT_NOARG, apply_option_parse_space_change },
4963                 OPT_BOOL('R', "reverse", &state->apply_in_reverse,
4964                         N_("apply the patch in reverse")),
4965                 OPT_BOOL(0, "unidiff-zero", &state->unidiff_zero,
4966                         N_("don't expect at least one line of context")),
4967                 OPT_BOOL(0, "reject", &state->apply_with_reject,
4968                         N_("leave the rejected hunks in corresponding *.rej files")),
4969                 OPT_BOOL(0, "allow-overlap", &state->allow_overlap,
4970                         N_("allow overlapping hunks")),
4971                 OPT__VERBOSE(&state->apply_verbosity, N_("be verbose")),
4972                 OPT_BIT(0, "inaccurate-eof", options,
4973                         N_("tolerate incorrectly detected missing new-line at the end of file"),
4974                         APPLY_OPT_INACCURATE_EOF),
4975                 OPT_BIT(0, "recount", options,
4976                         N_("do not trust the line counts in the hunk headers"),
4977                         APPLY_OPT_RECOUNT),
4978                 { OPTION_CALLBACK, 0, "directory", state, N_("root"),
4979                         N_("prepend <root> to all filenames"),
4980                         0, apply_option_parse_directory },
4981                 OPT_END()
4982         };
4983
4984         return parse_options(argc, argv, state->prefix, builtin_apply_options, apply_usage, 0);
4985 }