shallow: migrate shallow information into the object parser
[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 "object-store.h"
13 #include "blob.h"
14 #include "delta.h"
15 #include "diff.h"
16 #include "dir.h"
17 #include "xdiff-interface.h"
18 #include "ll-merge.h"
19 #include "lockfile.h"
20 #include "parse-options.h"
21 #include "quote.h"
22 #include "rerere.h"
23 #include "apply.h"
24
25 static void git_apply_config(void)
26 {
27         git_config_get_string_const("apply.whitespace", &apply_default_whitespace);
28         git_config_get_string_const("apply.ignorewhitespace", &apply_default_ignorewhitespace);
29         git_config(git_default_config, NULL);
30 }
31
32 static int parse_whitespace_option(struct apply_state *state, const char *option)
33 {
34         if (!option) {
35                 state->ws_error_action = warn_on_ws_error;
36                 return 0;
37         }
38         if (!strcmp(option, "warn")) {
39                 state->ws_error_action = warn_on_ws_error;
40                 return 0;
41         }
42         if (!strcmp(option, "nowarn")) {
43                 state->ws_error_action = nowarn_ws_error;
44                 return 0;
45         }
46         if (!strcmp(option, "error")) {
47                 state->ws_error_action = die_on_ws_error;
48                 return 0;
49         }
50         if (!strcmp(option, "error-all")) {
51                 state->ws_error_action = die_on_ws_error;
52                 state->squelch_whitespace_errors = 0;
53                 return 0;
54         }
55         if (!strcmp(option, "strip") || !strcmp(option, "fix")) {
56                 state->ws_error_action = correct_ws_error;
57                 return 0;
58         }
59         return error(_("unrecognized whitespace option '%s'"), option);
60 }
61
62 static int parse_ignorewhitespace_option(struct apply_state *state,
63                                                  const char *option)
64 {
65         if (!option || !strcmp(option, "no") ||
66             !strcmp(option, "false") || !strcmp(option, "never") ||
67             !strcmp(option, "none")) {
68                 state->ws_ignore_action = ignore_ws_none;
69                 return 0;
70         }
71         if (!strcmp(option, "change")) {
72                 state->ws_ignore_action = ignore_ws_change;
73                 return 0;
74         }
75         return error(_("unrecognized whitespace ignore option '%s'"), option);
76 }
77
78 int init_apply_state(struct apply_state *state,
79                      const char *prefix)
80 {
81         memset(state, 0, sizeof(*state));
82         state->prefix = prefix;
83         state->apply = 1;
84         state->line_termination = '\n';
85         state->p_value = 1;
86         state->p_context = UINT_MAX;
87         state->squelch_whitespace_errors = 5;
88         state->ws_error_action = warn_on_ws_error;
89         state->ws_ignore_action = ignore_ws_none;
90         state->linenr = 1;
91         string_list_init(&state->fn_table, 0);
92         string_list_init(&state->limit_by_name, 0);
93         string_list_init(&state->symlink_changes, 0);
94         strbuf_init(&state->root, 0);
95
96         git_apply_config();
97         if (apply_default_whitespace && parse_whitespace_option(state, apply_default_whitespace))
98                 return -1;
99         if (apply_default_ignorewhitespace && parse_ignorewhitespace_option(state, apply_default_ignorewhitespace))
100                 return -1;
101         return 0;
102 }
103
104 void clear_apply_state(struct apply_state *state)
105 {
106         string_list_clear(&state->limit_by_name, 0);
107         string_list_clear(&state->symlink_changes, 0);
108         strbuf_release(&state->root);
109
110         /* &state->fn_table is cleared at the end of apply_patch() */
111 }
112
113 static void mute_routine(const char *msg, va_list params)
114 {
115         /* do nothing */
116 }
117
118 int check_apply_state(struct apply_state *state, int force_apply)
119 {
120         int is_not_gitdir = !startup_info->have_repository;
121
122         if (state->apply_with_reject && state->threeway)
123                 return error(_("--reject and --3way cannot be used together."));
124         if (state->cached && state->threeway)
125                 return error(_("--cached and --3way cannot be used together."));
126         if (state->threeway) {
127                 if (is_not_gitdir)
128                         return error(_("--3way outside a repository"));
129                 state->check_index = 1;
130         }
131         if (state->apply_with_reject) {
132                 state->apply = 1;
133                 if (state->apply_verbosity == verbosity_normal)
134                         state->apply_verbosity = verbosity_verbose;
135         }
136         if (!force_apply && (state->diffstat || state->numstat || state->summary || state->check || state->fake_ancestor))
137                 state->apply = 0;
138         if (state->check_index && is_not_gitdir)
139                 return error(_("--index outside a repository"));
140         if (state->cached) {
141                 if (is_not_gitdir)
142                         return error(_("--cached outside a repository"));
143                 state->check_index = 1;
144         }
145         if (state->check_index)
146                 state->unsafe_paths = 0;
147
148         if (state->apply_verbosity <= verbosity_silent) {
149                 state->saved_error_routine = get_error_routine();
150                 state->saved_warn_routine = get_warn_routine();
151                 set_error_routine(mute_routine);
152                 set_warn_routine(mute_routine);
153         }
154
155         return 0;
156 }
157
158 static void set_default_whitespace_mode(struct apply_state *state)
159 {
160         if (!state->whitespace_option && !apply_default_whitespace)
161                 state->ws_error_action = (state->apply ? warn_on_ws_error : nowarn_ws_error);
162 }
163
164 /*
165  * This represents one "hunk" from a patch, starting with
166  * "@@ -oldpos,oldlines +newpos,newlines @@" marker.  The
167  * patch text is pointed at by patch, and its byte length
168  * is stored in size.  leading and trailing are the number
169  * of context lines.
170  */
171 struct fragment {
172         unsigned long leading, trailing;
173         unsigned long oldpos, oldlines;
174         unsigned long newpos, newlines;
175         /*
176          * 'patch' is usually borrowed from buf in apply_patch(),
177          * but some codepaths store an allocated buffer.
178          */
179         const char *patch;
180         unsigned free_patch:1,
181                 rejected:1;
182         int size;
183         int linenr;
184         struct fragment *next;
185 };
186
187 /*
188  * When dealing with a binary patch, we reuse "leading" field
189  * to store the type of the binary hunk, either deflated "delta"
190  * or deflated "literal".
191  */
192 #define binary_patch_method leading
193 #define BINARY_DELTA_DEFLATED   1
194 #define BINARY_LITERAL_DEFLATED 2
195
196 /*
197  * This represents a "patch" to a file, both metainfo changes
198  * such as creation/deletion, filemode and content changes represented
199  * as a series of fragments.
200  */
201 struct patch {
202         char *new_name, *old_name, *def_name;
203         unsigned int old_mode, new_mode;
204         int is_new, is_delete;  /* -1 = unknown, 0 = false, 1 = true */
205         int rejected;
206         unsigned ws_rule;
207         int lines_added, lines_deleted;
208         int score;
209         int extension_linenr; /* first line specifying delete/new/rename/copy */
210         unsigned int is_toplevel_relative:1;
211         unsigned int inaccurate_eof:1;
212         unsigned int is_binary:1;
213         unsigned int is_copy:1;
214         unsigned int is_rename:1;
215         unsigned int recount:1;
216         unsigned int conflicted_threeway:1;
217         unsigned int direct_to_threeway:1;
218         unsigned int crlf_in_old:1;
219         struct fragment *fragments;
220         char *result;
221         size_t resultsize;
222         char old_sha1_prefix[41];
223         char new_sha1_prefix[41];
224         struct patch *next;
225
226         /* three-way fallback result */
227         struct object_id threeway_stage[3];
228 };
229
230 static void free_fragment_list(struct fragment *list)
231 {
232         while (list) {
233                 struct fragment *next = list->next;
234                 if (list->free_patch)
235                         free((char *)list->patch);
236                 free(list);
237                 list = next;
238         }
239 }
240
241 static void free_patch(struct patch *patch)
242 {
243         free_fragment_list(patch->fragments);
244         free(patch->def_name);
245         free(patch->old_name);
246         free(patch->new_name);
247         free(patch->result);
248         free(patch);
249 }
250
251 static void free_patch_list(struct patch *list)
252 {
253         while (list) {
254                 struct patch *next = list->next;
255                 free_patch(list);
256                 list = next;
257         }
258 }
259
260 /*
261  * A line in a file, len-bytes long (includes the terminating LF,
262  * except for an incomplete line at the end if the file ends with
263  * one), and its contents hashes to 'hash'.
264  */
265 struct line {
266         size_t len;
267         unsigned hash : 24;
268         unsigned flag : 8;
269 #define LINE_COMMON     1
270 #define LINE_PATCHED    2
271 };
272
273 /*
274  * This represents a "file", which is an array of "lines".
275  */
276 struct image {
277         char *buf;
278         size_t len;
279         size_t nr;
280         size_t alloc;
281         struct line *line_allocated;
282         struct line *line;
283 };
284
285 static uint32_t hash_line(const char *cp, size_t len)
286 {
287         size_t i;
288         uint32_t h;
289         for (i = 0, h = 0; i < len; i++) {
290                 if (!isspace(cp[i])) {
291                         h = h * 3 + (cp[i] & 0xff);
292                 }
293         }
294         return h;
295 }
296
297 /*
298  * Compare lines s1 of length n1 and s2 of length n2, ignoring
299  * whitespace difference. Returns 1 if they match, 0 otherwise
300  */
301 static int fuzzy_matchlines(const char *s1, size_t n1,
302                             const char *s2, size_t n2)
303 {
304         const char *end1 = s1 + n1;
305         const char *end2 = s2 + n2;
306
307         /* ignore line endings */
308         while (s1 < end1 && (end1[-1] == '\r' || end1[-1] == '\n'))
309                 end1--;
310         while (s2 < end2 && (end2[-1] == '\r' || end2[-1] == '\n'))
311                 end2--;
312
313         while (s1 < end1 && s2 < end2) {
314                 if (isspace(*s1)) {
315                         /*
316                          * Skip whitespace. We check on both buffers
317                          * because we don't want "a b" to match "ab".
318                          */
319                         if (!isspace(*s2))
320                                 return 0;
321                         while (s1 < end1 && isspace(*s1))
322                                 s1++;
323                         while (s2 < end2 && isspace(*s2))
324                                 s2++;
325                 } else if (*s1++ != *s2++)
326                         return 0;
327         }
328
329         /* If we reached the end on one side only, lines don't match. */
330         return s1 == end1 && s2 == end2;
331 }
332
333 static void add_line_info(struct image *img, const char *bol, size_t len, unsigned flag)
334 {
335         ALLOC_GROW(img->line_allocated, img->nr + 1, img->alloc);
336         img->line_allocated[img->nr].len = len;
337         img->line_allocated[img->nr].hash = hash_line(bol, len);
338         img->line_allocated[img->nr].flag = flag;
339         img->nr++;
340 }
341
342 /*
343  * "buf" has the file contents to be patched (read from various sources).
344  * attach it to "image" and add line-based index to it.
345  * "image" now owns the "buf".
346  */
347 static void prepare_image(struct image *image, char *buf, size_t len,
348                           int prepare_linetable)
349 {
350         const char *cp, *ep;
351
352         memset(image, 0, sizeof(*image));
353         image->buf = buf;
354         image->len = len;
355
356         if (!prepare_linetable)
357                 return;
358
359         ep = image->buf + image->len;
360         cp = image->buf;
361         while (cp < ep) {
362                 const char *next;
363                 for (next = cp; next < ep && *next != '\n'; next++)
364                         ;
365                 if (next < ep)
366                         next++;
367                 add_line_info(image, cp, next - cp, 0);
368                 cp = next;
369         }
370         image->line = image->line_allocated;
371 }
372
373 static void clear_image(struct image *image)
374 {
375         free(image->buf);
376         free(image->line_allocated);
377         memset(image, 0, sizeof(*image));
378 }
379
380 /* fmt must contain _one_ %s and no other substitution */
381 static void say_patch_name(FILE *output, const char *fmt, struct patch *patch)
382 {
383         struct strbuf sb = STRBUF_INIT;
384
385         if (patch->old_name && patch->new_name &&
386             strcmp(patch->old_name, patch->new_name)) {
387                 quote_c_style(patch->old_name, &sb, NULL, 0);
388                 strbuf_addstr(&sb, " => ");
389                 quote_c_style(patch->new_name, &sb, NULL, 0);
390         } else {
391                 const char *n = patch->new_name;
392                 if (!n)
393                         n = patch->old_name;
394                 quote_c_style(n, &sb, NULL, 0);
395         }
396         fprintf(output, fmt, sb.buf);
397         fputc('\n', output);
398         strbuf_release(&sb);
399 }
400
401 #define SLOP (16)
402
403 static int read_patch_file(struct strbuf *sb, int fd)
404 {
405         if (strbuf_read(sb, fd, 0) < 0)
406                 return error_errno("git apply: failed to read");
407
408         /*
409          * Make sure that we have some slop in the buffer
410          * so that we can do speculative "memcmp" etc, and
411          * see to it that it is NUL-filled.
412          */
413         strbuf_grow(sb, SLOP);
414         memset(sb->buf + sb->len, 0, SLOP);
415         return 0;
416 }
417
418 static unsigned long linelen(const char *buffer, unsigned long size)
419 {
420         unsigned long len = 0;
421         while (size--) {
422                 len++;
423                 if (*buffer++ == '\n')
424                         break;
425         }
426         return len;
427 }
428
429 static int is_dev_null(const char *str)
430 {
431         return skip_prefix(str, "/dev/null", &str) && isspace(*str);
432 }
433
434 #define TERM_SPACE      1
435 #define TERM_TAB        2
436
437 static int name_terminate(int c, int terminate)
438 {
439         if (c == ' ' && !(terminate & TERM_SPACE))
440                 return 0;
441         if (c == '\t' && !(terminate & TERM_TAB))
442                 return 0;
443
444         return 1;
445 }
446
447 /* remove double slashes to make --index work with such filenames */
448 static char *squash_slash(char *name)
449 {
450         int i = 0, j = 0;
451
452         if (!name)
453                 return NULL;
454
455         while (name[i]) {
456                 if ((name[j++] = name[i++]) == '/')
457                         while (name[i] == '/')
458                                 i++;
459         }
460         name[j] = '\0';
461         return name;
462 }
463
464 static char *find_name_gnu(struct apply_state *state,
465                            const char *line,
466                            const char *def,
467                            int p_value)
468 {
469         struct strbuf name = STRBUF_INIT;
470         char *cp;
471
472         /*
473          * Proposed "new-style" GNU patch/diff format; see
474          * http://marc.info/?l=git&m=112927316408690&w=2
475          */
476         if (unquote_c_style(&name, line, NULL)) {
477                 strbuf_release(&name);
478                 return NULL;
479         }
480
481         for (cp = name.buf; p_value; p_value--) {
482                 cp = strchr(cp, '/');
483                 if (!cp) {
484                         strbuf_release(&name);
485                         return NULL;
486                 }
487                 cp++;
488         }
489
490         strbuf_remove(&name, 0, cp - name.buf);
491         if (state->root.len)
492                 strbuf_insert(&name, 0, state->root.buf, state->root.len);
493         return squash_slash(strbuf_detach(&name, NULL));
494 }
495
496 static size_t sane_tz_len(const char *line, size_t len)
497 {
498         const char *tz, *p;
499
500         if (len < strlen(" +0500") || line[len-strlen(" +0500")] != ' ')
501                 return 0;
502         tz = line + len - strlen(" +0500");
503
504         if (tz[1] != '+' && tz[1] != '-')
505                 return 0;
506
507         for (p = tz + 2; p != line + len; p++)
508                 if (!isdigit(*p))
509                         return 0;
510
511         return line + len - tz;
512 }
513
514 static size_t tz_with_colon_len(const char *line, size_t len)
515 {
516         const char *tz, *p;
517
518         if (len < strlen(" +08:00") || line[len - strlen(":00")] != ':')
519                 return 0;
520         tz = line + len - strlen(" +08:00");
521
522         if (tz[0] != ' ' || (tz[1] != '+' && tz[1] != '-'))
523                 return 0;
524         p = tz + 2;
525         if (!isdigit(*p++) || !isdigit(*p++) || *p++ != ':' ||
526             !isdigit(*p++) || !isdigit(*p++))
527                 return 0;
528
529         return line + len - tz;
530 }
531
532 static size_t date_len(const char *line, size_t len)
533 {
534         const char *date, *p;
535
536         if (len < strlen("72-02-05") || line[len-strlen("-05")] != '-')
537                 return 0;
538         p = date = line + len - strlen("72-02-05");
539
540         if (!isdigit(*p++) || !isdigit(*p++) || *p++ != '-' ||
541             !isdigit(*p++) || !isdigit(*p++) || *p++ != '-' ||
542             !isdigit(*p++) || !isdigit(*p++))   /* Not a date. */
543                 return 0;
544
545         if (date - line >= strlen("19") &&
546             isdigit(date[-1]) && isdigit(date[-2]))     /* 4-digit year */
547                 date -= strlen("19");
548
549         return line + len - date;
550 }
551
552 static size_t short_time_len(const char *line, size_t len)
553 {
554         const char *time, *p;
555
556         if (len < strlen(" 07:01:32") || line[len-strlen(":32")] != ':')
557                 return 0;
558         p = time = line + len - strlen(" 07:01:32");
559
560         /* Permit 1-digit hours? */
561         if (*p++ != ' ' ||
562             !isdigit(*p++) || !isdigit(*p++) || *p++ != ':' ||
563             !isdigit(*p++) || !isdigit(*p++) || *p++ != ':' ||
564             !isdigit(*p++) || !isdigit(*p++))   /* Not a time. */
565                 return 0;
566
567         return line + len - time;
568 }
569
570 static size_t fractional_time_len(const char *line, size_t len)
571 {
572         const char *p;
573         size_t n;
574
575         /* Expected format: 19:41:17.620000023 */
576         if (!len || !isdigit(line[len - 1]))
577                 return 0;
578         p = line + len - 1;
579
580         /* Fractional seconds. */
581         while (p > line && isdigit(*p))
582                 p--;
583         if (*p != '.')
584                 return 0;
585
586         /* Hours, minutes, and whole seconds. */
587         n = short_time_len(line, p - line);
588         if (!n)
589                 return 0;
590
591         return line + len - p + n;
592 }
593
594 static size_t trailing_spaces_len(const char *line, size_t len)
595 {
596         const char *p;
597
598         /* Expected format: ' ' x (1 or more)  */
599         if (!len || line[len - 1] != ' ')
600                 return 0;
601
602         p = line + len;
603         while (p != line) {
604                 p--;
605                 if (*p != ' ')
606                         return line + len - (p + 1);
607         }
608
609         /* All spaces! */
610         return len;
611 }
612
613 static size_t diff_timestamp_len(const char *line, size_t len)
614 {
615         const char *end = line + len;
616         size_t n;
617
618         /*
619          * Posix: 2010-07-05 19:41:17
620          * GNU: 2010-07-05 19:41:17.620000023 -0500
621          */
622
623         if (!isdigit(end[-1]))
624                 return 0;
625
626         n = sane_tz_len(line, end - line);
627         if (!n)
628                 n = tz_with_colon_len(line, end - line);
629         end -= n;
630
631         n = short_time_len(line, end - line);
632         if (!n)
633                 n = fractional_time_len(line, end - line);
634         end -= n;
635
636         n = date_len(line, end - line);
637         if (!n) /* No date.  Too bad. */
638                 return 0;
639         end -= n;
640
641         if (end == line)        /* No space before date. */
642                 return 0;
643         if (end[-1] == '\t') {  /* Success! */
644                 end--;
645                 return line + len - end;
646         }
647         if (end[-1] != ' ')     /* No space before date. */
648                 return 0;
649
650         /* Whitespace damage. */
651         end -= trailing_spaces_len(line, end - line);
652         return line + len - end;
653 }
654
655 static char *find_name_common(struct apply_state *state,
656                               const char *line,
657                               const char *def,
658                               int p_value,
659                               const char *end,
660                               int terminate)
661 {
662         int len;
663         const char *start = NULL;
664
665         if (p_value == 0)
666                 start = line;
667         while (line != end) {
668                 char c = *line;
669
670                 if (!end && isspace(c)) {
671                         if (c == '\n')
672                                 break;
673                         if (name_terminate(c, terminate))
674                                 break;
675                 }
676                 line++;
677                 if (c == '/' && !--p_value)
678                         start = line;
679         }
680         if (!start)
681                 return squash_slash(xstrdup_or_null(def));
682         len = line - start;
683         if (!len)
684                 return squash_slash(xstrdup_or_null(def));
685
686         /*
687          * Generally we prefer the shorter name, especially
688          * if the other one is just a variation of that with
689          * something else tacked on to the end (ie "file.orig"
690          * or "file~").
691          */
692         if (def) {
693                 int deflen = strlen(def);
694                 if (deflen < len && !strncmp(start, def, deflen))
695                         return squash_slash(xstrdup(def));
696         }
697
698         if (state->root.len) {
699                 char *ret = xstrfmt("%s%.*s", state->root.buf, len, start);
700                 return squash_slash(ret);
701         }
702
703         return squash_slash(xmemdupz(start, len));
704 }
705
706 static char *find_name(struct apply_state *state,
707                        const char *line,
708                        char *def,
709                        int p_value,
710                        int terminate)
711 {
712         if (*line == '"') {
713                 char *name = find_name_gnu(state, line, def, p_value);
714                 if (name)
715                         return name;
716         }
717
718         return find_name_common(state, line, def, p_value, NULL, terminate);
719 }
720
721 static char *find_name_traditional(struct apply_state *state,
722                                    const char *line,
723                                    char *def,
724                                    int p_value)
725 {
726         size_t len;
727         size_t date_len;
728
729         if (*line == '"') {
730                 char *name = find_name_gnu(state, line, def, p_value);
731                 if (name)
732                         return name;
733         }
734
735         len = strchrnul(line, '\n') - line;
736         date_len = diff_timestamp_len(line, len);
737         if (!date_len)
738                 return find_name_common(state, line, def, p_value, NULL, TERM_TAB);
739         len -= date_len;
740
741         return find_name_common(state, line, def, p_value, line + len, 0);
742 }
743
744 /*
745  * Given the string after "--- " or "+++ ", guess the appropriate
746  * p_value for the given patch.
747  */
748 static int guess_p_value(struct apply_state *state, const char *nameline)
749 {
750         char *name, *cp;
751         int val = -1;
752
753         if (is_dev_null(nameline))
754                 return -1;
755         name = find_name_traditional(state, nameline, NULL, 0);
756         if (!name)
757                 return -1;
758         cp = strchr(name, '/');
759         if (!cp)
760                 val = 0;
761         else if (state->prefix) {
762                 /*
763                  * Does it begin with "a/$our-prefix" and such?  Then this is
764                  * very likely to apply to our directory.
765                  */
766                 if (starts_with(name, state->prefix))
767                         val = count_slashes(state->prefix);
768                 else {
769                         cp++;
770                         if (starts_with(cp, state->prefix))
771                                 val = count_slashes(state->prefix) + 1;
772                 }
773         }
774         free(name);
775         return val;
776 }
777
778 /*
779  * Does the ---/+++ line have the POSIX timestamp after the last HT?
780  * GNU diff puts epoch there to signal a creation/deletion event.  Is
781  * this such a timestamp?
782  */
783 static int has_epoch_timestamp(const char *nameline)
784 {
785         /*
786          * We are only interested in epoch timestamp; any non-zero
787          * fraction cannot be one, hence "(\.0+)?" in the regexp below.
788          * For the same reason, the date must be either 1969-12-31 or
789          * 1970-01-01, and the seconds part must be "00".
790          */
791         const char stamp_regexp[] =
792                 "^[0-2][0-9]:([0-5][0-9]):00(\\.0+)?"
793                 " "
794                 "([-+][0-2][0-9]:?[0-5][0-9])\n";
795         const char *timestamp = NULL, *cp, *colon;
796         static regex_t *stamp;
797         regmatch_t m[10];
798         int zoneoffset, epoch_hour, hour, minute;
799         int status;
800
801         for (cp = nameline; *cp != '\n'; cp++) {
802                 if (*cp == '\t')
803                         timestamp = cp + 1;
804         }
805         if (!timestamp)
806                 return 0;
807
808         /*
809          * YYYY-MM-DD hh:mm:ss must be from either 1969-12-31
810          * (west of GMT) or 1970-01-01 (east of GMT)
811          */
812         if (skip_prefix(timestamp, "1969-12-31 ", &timestamp))
813                 epoch_hour = 24;
814         else if (skip_prefix(timestamp, "1970-01-01 ", &timestamp))
815                 epoch_hour = 0;
816         else
817                 return 0;
818
819         if (!stamp) {
820                 stamp = xmalloc(sizeof(*stamp));
821                 if (regcomp(stamp, stamp_regexp, REG_EXTENDED)) {
822                         warning(_("Cannot prepare timestamp regexp %s"),
823                                 stamp_regexp);
824                         return 0;
825                 }
826         }
827
828         status = regexec(stamp, timestamp, ARRAY_SIZE(m), m, 0);
829         if (status) {
830                 if (status != REG_NOMATCH)
831                         warning(_("regexec returned %d for input: %s"),
832                                 status, timestamp);
833                 return 0;
834         }
835
836         hour = strtol(timestamp, NULL, 10);
837         minute = strtol(timestamp + m[1].rm_so, NULL, 10);
838
839         zoneoffset = strtol(timestamp + m[3].rm_so + 1, (char **) &colon, 10);
840         if (*colon == ':')
841                 zoneoffset = zoneoffset * 60 + strtol(colon + 1, NULL, 10);
842         else
843                 zoneoffset = (zoneoffset / 100) * 60 + (zoneoffset % 100);
844         if (timestamp[m[3].rm_so] == '-')
845                 zoneoffset = -zoneoffset;
846
847         return hour * 60 + minute - zoneoffset == epoch_hour * 60;
848 }
849
850 /*
851  * Get the name etc info from the ---/+++ lines of a traditional patch header
852  *
853  * FIXME! The end-of-filename heuristics are kind of screwy. For existing
854  * files, we can happily check the index for a match, but for creating a
855  * new file we should try to match whatever "patch" does. I have no idea.
856  */
857 static int parse_traditional_patch(struct apply_state *state,
858                                    const char *first,
859                                    const char *second,
860                                    struct patch *patch)
861 {
862         char *name;
863
864         first += 4;     /* skip "--- " */
865         second += 4;    /* skip "+++ " */
866         if (!state->p_value_known) {
867                 int p, q;
868                 p = guess_p_value(state, first);
869                 q = guess_p_value(state, second);
870                 if (p < 0) p = q;
871                 if (0 <= p && p == q) {
872                         state->p_value = p;
873                         state->p_value_known = 1;
874                 }
875         }
876         if (is_dev_null(first)) {
877                 patch->is_new = 1;
878                 patch->is_delete = 0;
879                 name = find_name_traditional(state, second, NULL, state->p_value);
880                 patch->new_name = name;
881         } else if (is_dev_null(second)) {
882                 patch->is_new = 0;
883                 patch->is_delete = 1;
884                 name = find_name_traditional(state, first, NULL, state->p_value);
885                 patch->old_name = name;
886         } else {
887                 char *first_name;
888                 first_name = find_name_traditional(state, first, NULL, state->p_value);
889                 name = find_name_traditional(state, second, first_name, state->p_value);
890                 free(first_name);
891                 if (has_epoch_timestamp(first)) {
892                         patch->is_new = 1;
893                         patch->is_delete = 0;
894                         patch->new_name = name;
895                 } else if (has_epoch_timestamp(second)) {
896                         patch->is_new = 0;
897                         patch->is_delete = 1;
898                         patch->old_name = name;
899                 } else {
900                         patch->old_name = name;
901                         patch->new_name = xstrdup_or_null(name);
902                 }
903         }
904         if (!name)
905                 return error(_("unable to find filename in patch at line %d"), state->linenr);
906
907         return 0;
908 }
909
910 static int gitdiff_hdrend(struct apply_state *state,
911                           const char *line,
912                           struct patch *patch)
913 {
914         return 1;
915 }
916
917 /*
918  * We're anal about diff header consistency, to make
919  * sure that we don't end up having strange ambiguous
920  * patches floating around.
921  *
922  * As a result, gitdiff_{old|new}name() will check
923  * their names against any previous information, just
924  * to make sure..
925  */
926 #define DIFF_OLD_NAME 0
927 #define DIFF_NEW_NAME 1
928
929 static int gitdiff_verify_name(struct apply_state *state,
930                                const char *line,
931                                int isnull,
932                                char **name,
933                                int side)
934 {
935         if (!*name && !isnull) {
936                 *name = find_name(state, line, NULL, state->p_value, TERM_TAB);
937                 return 0;
938         }
939
940         if (*name) {
941                 char *another;
942                 if (isnull)
943                         return error(_("git apply: bad git-diff - expected /dev/null, got %s on line %d"),
944                                      *name, state->linenr);
945                 another = find_name(state, line, NULL, state->p_value, TERM_TAB);
946                 if (!another || strcmp(another, *name)) {
947                         free(another);
948                         return error((side == DIFF_NEW_NAME) ?
949                             _("git apply: bad git-diff - inconsistent new filename on line %d") :
950                             _("git apply: bad git-diff - inconsistent old filename on line %d"), state->linenr);
951                 }
952                 free(another);
953         } else {
954                 if (!is_dev_null(line))
955                         return error(_("git apply: bad git-diff - expected /dev/null on line %d"), state->linenr);
956         }
957
958         return 0;
959 }
960
961 static int gitdiff_oldname(struct apply_state *state,
962                            const char *line,
963                            struct patch *patch)
964 {
965         return gitdiff_verify_name(state, line,
966                                    patch->is_new, &patch->old_name,
967                                    DIFF_OLD_NAME);
968 }
969
970 static int gitdiff_newname(struct apply_state *state,
971                            const char *line,
972                            struct patch *patch)
973 {
974         return gitdiff_verify_name(state, line,
975                                    patch->is_delete, &patch->new_name,
976                                    DIFF_NEW_NAME);
977 }
978
979 static int parse_mode_line(const char *line, int linenr, unsigned int *mode)
980 {
981         char *end;
982         *mode = strtoul(line, &end, 8);
983         if (end == line || !isspace(*end))
984                 return error(_("invalid mode on line %d: %s"), linenr, line);
985         return 0;
986 }
987
988 static int gitdiff_oldmode(struct apply_state *state,
989                            const char *line,
990                            struct patch *patch)
991 {
992         return parse_mode_line(line, state->linenr, &patch->old_mode);
993 }
994
995 static int gitdiff_newmode(struct apply_state *state,
996                            const char *line,
997                            struct patch *patch)
998 {
999         return parse_mode_line(line, state->linenr, &patch->new_mode);
1000 }
1001
1002 static int gitdiff_delete(struct apply_state *state,
1003                           const char *line,
1004                           struct patch *patch)
1005 {
1006         patch->is_delete = 1;
1007         free(patch->old_name);
1008         patch->old_name = xstrdup_or_null(patch->def_name);
1009         return gitdiff_oldmode(state, line, patch);
1010 }
1011
1012 static int gitdiff_newfile(struct apply_state *state,
1013                            const char *line,
1014                            struct patch *patch)
1015 {
1016         patch->is_new = 1;
1017         free(patch->new_name);
1018         patch->new_name = xstrdup_or_null(patch->def_name);
1019         return gitdiff_newmode(state, line, patch);
1020 }
1021
1022 static int gitdiff_copysrc(struct apply_state *state,
1023                            const char *line,
1024                            struct patch *patch)
1025 {
1026         patch->is_copy = 1;
1027         free(patch->old_name);
1028         patch->old_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1029         return 0;
1030 }
1031
1032 static int gitdiff_copydst(struct apply_state *state,
1033                            const char *line,
1034                            struct patch *patch)
1035 {
1036         patch->is_copy = 1;
1037         free(patch->new_name);
1038         patch->new_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1039         return 0;
1040 }
1041
1042 static int gitdiff_renamesrc(struct apply_state *state,
1043                              const char *line,
1044                              struct patch *patch)
1045 {
1046         patch->is_rename = 1;
1047         free(patch->old_name);
1048         patch->old_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1049         return 0;
1050 }
1051
1052 static int gitdiff_renamedst(struct apply_state *state,
1053                              const char *line,
1054                              struct patch *patch)
1055 {
1056         patch->is_rename = 1;
1057         free(patch->new_name);
1058         patch->new_name = find_name(state, line, NULL, state->p_value ? state->p_value - 1 : 0, 0);
1059         return 0;
1060 }
1061
1062 static int gitdiff_similarity(struct apply_state *state,
1063                               const char *line,
1064                               struct patch *patch)
1065 {
1066         unsigned long val = strtoul(line, NULL, 10);
1067         if (val <= 100)
1068                 patch->score = val;
1069         return 0;
1070 }
1071
1072 static int gitdiff_dissimilarity(struct apply_state *state,
1073                                  const char *line,
1074                                  struct patch *patch)
1075 {
1076         unsigned long val = strtoul(line, NULL, 10);
1077         if (val <= 100)
1078                 patch->score = val;
1079         return 0;
1080 }
1081
1082 static int gitdiff_index(struct apply_state *state,
1083                          const char *line,
1084                          struct patch *patch)
1085 {
1086         /*
1087          * index line is N hexadecimal, "..", N hexadecimal,
1088          * and optional space with octal mode.
1089          */
1090         const char *ptr, *eol;
1091         int len;
1092
1093         ptr = strchr(line, '.');
1094         if (!ptr || ptr[1] != '.' || 40 < ptr - line)
1095                 return 0;
1096         len = ptr - line;
1097         memcpy(patch->old_sha1_prefix, line, len);
1098         patch->old_sha1_prefix[len] = 0;
1099
1100         line = ptr + 2;
1101         ptr = strchr(line, ' ');
1102         eol = strchrnul(line, '\n');
1103
1104         if (!ptr || eol < ptr)
1105                 ptr = eol;
1106         len = ptr - line;
1107
1108         if (40 < len)
1109                 return 0;
1110         memcpy(patch->new_sha1_prefix, line, len);
1111         patch->new_sha1_prefix[len] = 0;
1112         if (*ptr == ' ')
1113                 return gitdiff_oldmode(state, ptr + 1, patch);
1114         return 0;
1115 }
1116
1117 /*
1118  * This is normal for a diff that doesn't change anything: we'll fall through
1119  * into the next diff. Tell the parser to break out.
1120  */
1121 static int gitdiff_unrecognized(struct apply_state *state,
1122                                 const char *line,
1123                                 struct patch *patch)
1124 {
1125         return 1;
1126 }
1127
1128 /*
1129  * Skip p_value leading components from "line"; as we do not accept
1130  * absolute paths, return NULL in that case.
1131  */
1132 static const char *skip_tree_prefix(struct apply_state *state,
1133                                     const char *line,
1134                                     int llen)
1135 {
1136         int nslash;
1137         int i;
1138
1139         if (!state->p_value)
1140                 return (llen && line[0] == '/') ? NULL : line;
1141
1142         nslash = state->p_value;
1143         for (i = 0; i < llen; i++) {
1144                 int ch = line[i];
1145                 if (ch == '/' && --nslash <= 0)
1146                         return (i == 0) ? NULL : &line[i + 1];
1147         }
1148         return NULL;
1149 }
1150
1151 /*
1152  * This is to extract the same name that appears on "diff --git"
1153  * line.  We do not find and return anything if it is a rename
1154  * patch, and it is OK because we will find the name elsewhere.
1155  * We need to reliably find name only when it is mode-change only,
1156  * creation or deletion of an empty file.  In any of these cases,
1157  * both sides are the same name under a/ and b/ respectively.
1158  */
1159 static char *git_header_name(struct apply_state *state,
1160                              const char *line,
1161                              int llen)
1162 {
1163         const char *name;
1164         const char *second = NULL;
1165         size_t len, line_len;
1166
1167         line += strlen("diff --git ");
1168         llen -= strlen("diff --git ");
1169
1170         if (*line == '"') {
1171                 const char *cp;
1172                 struct strbuf first = STRBUF_INIT;
1173                 struct strbuf sp = STRBUF_INIT;
1174
1175                 if (unquote_c_style(&first, line, &second))
1176                         goto free_and_fail1;
1177
1178                 /* strip the a/b prefix including trailing slash */
1179                 cp = skip_tree_prefix(state, first.buf, first.len);
1180                 if (!cp)
1181                         goto free_and_fail1;
1182                 strbuf_remove(&first, 0, cp - first.buf);
1183
1184                 /*
1185                  * second points at one past closing dq of name.
1186                  * find the second name.
1187                  */
1188                 while ((second < line + llen) && isspace(*second))
1189                         second++;
1190
1191                 if (line + llen <= second)
1192                         goto free_and_fail1;
1193                 if (*second == '"') {
1194                         if (unquote_c_style(&sp, second, NULL))
1195                                 goto free_and_fail1;
1196                         cp = skip_tree_prefix(state, sp.buf, sp.len);
1197                         if (!cp)
1198                                 goto free_and_fail1;
1199                         /* They must match, otherwise ignore */
1200                         if (strcmp(cp, first.buf))
1201                                 goto free_and_fail1;
1202                         strbuf_release(&sp);
1203                         return strbuf_detach(&first, NULL);
1204                 }
1205
1206                 /* unquoted second */
1207                 cp = skip_tree_prefix(state, second, line + llen - second);
1208                 if (!cp)
1209                         goto free_and_fail1;
1210                 if (line + llen - cp != first.len ||
1211                     memcmp(first.buf, cp, first.len))
1212                         goto free_and_fail1;
1213                 return strbuf_detach(&first, NULL);
1214
1215         free_and_fail1:
1216                 strbuf_release(&first);
1217                 strbuf_release(&sp);
1218                 return NULL;
1219         }
1220
1221         /* unquoted first name */
1222         name = skip_tree_prefix(state, line, llen);
1223         if (!name)
1224                 return NULL;
1225
1226         /*
1227          * since the first name is unquoted, a dq if exists must be
1228          * the beginning of the second name.
1229          */
1230         for (second = name; second < line + llen; second++) {
1231                 if (*second == '"') {
1232                         struct strbuf sp = STRBUF_INIT;
1233                         const char *np;
1234
1235                         if (unquote_c_style(&sp, second, NULL))
1236                                 goto free_and_fail2;
1237
1238                         np = skip_tree_prefix(state, sp.buf, sp.len);
1239                         if (!np)
1240                                 goto free_and_fail2;
1241
1242                         len = sp.buf + sp.len - np;
1243                         if (len < second - name &&
1244                             !strncmp(np, name, len) &&
1245                             isspace(name[len])) {
1246                                 /* Good */
1247                                 strbuf_remove(&sp, 0, np - sp.buf);
1248                                 return strbuf_detach(&sp, NULL);
1249                         }
1250
1251                 free_and_fail2:
1252                         strbuf_release(&sp);
1253                         return NULL;
1254                 }
1255         }
1256
1257         /*
1258          * Accept a name only if it shows up twice, exactly the same
1259          * form.
1260          */
1261         second = strchr(name, '\n');
1262         if (!second)
1263                 return NULL;
1264         line_len = second - name;
1265         for (len = 0 ; ; len++) {
1266                 switch (name[len]) {
1267                 default:
1268                         continue;
1269                 case '\n':
1270                         return NULL;
1271                 case '\t': case ' ':
1272                         /*
1273                          * Is this the separator between the preimage
1274                          * and the postimage pathname?  Again, we are
1275                          * only interested in the case where there is
1276                          * no rename, as this is only to set def_name
1277                          * and a rename patch has the names elsewhere
1278                          * in an unambiguous form.
1279                          */
1280                         if (!name[len + 1])
1281                                 return NULL; /* no postimage name */
1282                         second = skip_tree_prefix(state, name + len + 1,
1283                                                   line_len - (len + 1));
1284                         if (!second)
1285                                 return NULL;
1286                         /*
1287                          * Does len bytes starting at "name" and "second"
1288                          * (that are separated by one HT or SP we just
1289                          * found) exactly match?
1290                          */
1291                         if (second[len] == '\n' && !strncmp(name, second, len))
1292                                 return xmemdupz(name, len);
1293                 }
1294         }
1295 }
1296
1297 static int check_header_line(struct apply_state *state, struct patch *patch)
1298 {
1299         int extensions = (patch->is_delete == 1) + (patch->is_new == 1) +
1300                          (patch->is_rename == 1) + (patch->is_copy == 1);
1301         if (extensions > 1)
1302                 return error(_("inconsistent header lines %d and %d"),
1303                              patch->extension_linenr, state->linenr);
1304         if (extensions && !patch->extension_linenr)
1305                 patch->extension_linenr = state->linenr;
1306         return 0;
1307 }
1308
1309 /* Verify that we recognize the lines following a git header */
1310 static int parse_git_header(struct apply_state *state,
1311                             const char *line,
1312                             int len,
1313                             unsigned int size,
1314                             struct patch *patch)
1315 {
1316         unsigned long offset;
1317
1318         /* A git diff has explicit new/delete information, so we don't guess */
1319         patch->is_new = 0;
1320         patch->is_delete = 0;
1321
1322         /*
1323          * Some things may not have the old name in the
1324          * rest of the headers anywhere (pure mode changes,
1325          * or removing or adding empty files), so we get
1326          * the default name from the header.
1327          */
1328         patch->def_name = git_header_name(state, line, len);
1329         if (patch->def_name && state->root.len) {
1330                 char *s = xstrfmt("%s%s", state->root.buf, patch->def_name);
1331                 free(patch->def_name);
1332                 patch->def_name = s;
1333         }
1334
1335         line += len;
1336         size -= len;
1337         state->linenr++;
1338         for (offset = len ; size > 0 ; offset += len, size -= len, line += len, state->linenr++) {
1339                 static const struct opentry {
1340                         const char *str;
1341                         int (*fn)(struct apply_state *, const char *, struct patch *);
1342                 } optable[] = {
1343                         { "@@ -", gitdiff_hdrend },
1344                         { "--- ", gitdiff_oldname },
1345                         { "+++ ", gitdiff_newname },
1346                         { "old mode ", gitdiff_oldmode },
1347                         { "new mode ", gitdiff_newmode },
1348                         { "deleted file mode ", gitdiff_delete },
1349                         { "new file mode ", gitdiff_newfile },
1350                         { "copy from ", gitdiff_copysrc },
1351                         { "copy to ", gitdiff_copydst },
1352                         { "rename old ", gitdiff_renamesrc },
1353                         { "rename new ", gitdiff_renamedst },
1354                         { "rename from ", gitdiff_renamesrc },
1355                         { "rename to ", gitdiff_renamedst },
1356                         { "similarity index ", gitdiff_similarity },
1357                         { "dissimilarity index ", gitdiff_dissimilarity },
1358                         { "index ", gitdiff_index },
1359                         { "", gitdiff_unrecognized },
1360                 };
1361                 int i;
1362
1363                 len = linelen(line, size);
1364                 if (!len || line[len-1] != '\n')
1365                         break;
1366                 for (i = 0; i < ARRAY_SIZE(optable); i++) {
1367                         const struct opentry *p = optable + i;
1368                         int oplen = strlen(p->str);
1369                         int res;
1370                         if (len < oplen || memcmp(p->str, line, oplen))
1371                                 continue;
1372                         res = p->fn(state, line + oplen, patch);
1373                         if (res < 0)
1374                                 return -1;
1375                         if (check_header_line(state, patch))
1376                                 return -1;
1377                         if (res > 0)
1378                                 return offset;
1379                         break;
1380                 }
1381         }
1382
1383         return offset;
1384 }
1385
1386 static int parse_num(const char *line, unsigned long *p)
1387 {
1388         char *ptr;
1389
1390         if (!isdigit(*line))
1391                 return 0;
1392         *p = strtoul(line, &ptr, 10);
1393         return ptr - line;
1394 }
1395
1396 static int parse_range(const char *line, int len, int offset, const char *expect,
1397                        unsigned long *p1, unsigned long *p2)
1398 {
1399         int digits, ex;
1400
1401         if (offset < 0 || offset >= len)
1402                 return -1;
1403         line += offset;
1404         len -= offset;
1405
1406         digits = parse_num(line, p1);
1407         if (!digits)
1408                 return -1;
1409
1410         offset += digits;
1411         line += digits;
1412         len -= digits;
1413
1414         *p2 = 1;
1415         if (*line == ',') {
1416                 digits = parse_num(line+1, p2);
1417                 if (!digits)
1418                         return -1;
1419
1420                 offset += digits+1;
1421                 line += digits+1;
1422                 len -= digits+1;
1423         }
1424
1425         ex = strlen(expect);
1426         if (ex > len)
1427                 return -1;
1428         if (memcmp(line, expect, ex))
1429                 return -1;
1430
1431         return offset + ex;
1432 }
1433
1434 static void recount_diff(const char *line, int size, struct fragment *fragment)
1435 {
1436         int oldlines = 0, newlines = 0, ret = 0;
1437
1438         if (size < 1) {
1439                 warning("recount: ignore empty hunk");
1440                 return;
1441         }
1442
1443         for (;;) {
1444                 int len = linelen(line, size);
1445                 size -= len;
1446                 line += len;
1447
1448                 if (size < 1)
1449                         break;
1450
1451                 switch (*line) {
1452                 case ' ': case '\n':
1453                         newlines++;
1454                         /* fall through */
1455                 case '-':
1456                         oldlines++;
1457                         continue;
1458                 case '+':
1459                         newlines++;
1460                         continue;
1461                 case '\\':
1462                         continue;
1463                 case '@':
1464                         ret = size < 3 || !starts_with(line, "@@ ");
1465                         break;
1466                 case 'd':
1467                         ret = size < 5 || !starts_with(line, "diff ");
1468                         break;
1469                 default:
1470                         ret = -1;
1471                         break;
1472                 }
1473                 if (ret) {
1474                         warning(_("recount: unexpected line: %.*s"),
1475                                 (int)linelen(line, size), line);
1476                         return;
1477                 }
1478                 break;
1479         }
1480         fragment->oldlines = oldlines;
1481         fragment->newlines = newlines;
1482 }
1483
1484 /*
1485  * Parse a unified diff fragment header of the
1486  * form "@@ -a,b +c,d @@"
1487  */
1488 static int parse_fragment_header(const char *line, int len, struct fragment *fragment)
1489 {
1490         int offset;
1491
1492         if (!len || line[len-1] != '\n')
1493                 return -1;
1494
1495         /* Figure out the number of lines in a fragment */
1496         offset = parse_range(line, len, 4, " +", &fragment->oldpos, &fragment->oldlines);
1497         offset = parse_range(line, len, offset, " @@", &fragment->newpos, &fragment->newlines);
1498
1499         return offset;
1500 }
1501
1502 /*
1503  * Find file diff header
1504  *
1505  * Returns:
1506  *  -1 if no header was found
1507  *  -128 in case of error
1508  *   the size of the header in bytes (called "offset") otherwise
1509  */
1510 static int find_header(struct apply_state *state,
1511                        const char *line,
1512                        unsigned long size,
1513                        int *hdrsize,
1514                        struct patch *patch)
1515 {
1516         unsigned long offset, len;
1517
1518         patch->is_toplevel_relative = 0;
1519         patch->is_rename = patch->is_copy = 0;
1520         patch->is_new = patch->is_delete = -1;
1521         patch->old_mode = patch->new_mode = 0;
1522         patch->old_name = patch->new_name = NULL;
1523         for (offset = 0; size > 0; offset += len, size -= len, line += len, state->linenr++) {
1524                 unsigned long nextlen;
1525
1526                 len = linelen(line, size);
1527                 if (!len)
1528                         break;
1529
1530                 /* Testing this early allows us to take a few shortcuts.. */
1531                 if (len < 6)
1532                         continue;
1533
1534                 /*
1535                  * Make sure we don't find any unconnected patch fragments.
1536                  * That's a sign that we didn't find a header, and that a
1537                  * patch has become corrupted/broken up.
1538                  */
1539                 if (!memcmp("@@ -", line, 4)) {
1540                         struct fragment dummy;
1541                         if (parse_fragment_header(line, len, &dummy) < 0)
1542                                 continue;
1543                         error(_("patch fragment without header at line %d: %.*s"),
1544                                      state->linenr, (int)len-1, line);
1545                         return -128;
1546                 }
1547
1548                 if (size < len + 6)
1549                         break;
1550
1551                 /*
1552                  * Git patch? It might not have a real patch, just a rename
1553                  * or mode change, so we handle that specially
1554                  */
1555                 if (!memcmp("diff --git ", line, 11)) {
1556                         int git_hdr_len = parse_git_header(state, line, len, size, patch);
1557                         if (git_hdr_len < 0)
1558                                 return -128;
1559                         if (git_hdr_len <= len)
1560                                 continue;
1561                         if (!patch->old_name && !patch->new_name) {
1562                                 if (!patch->def_name) {
1563                                         error(Q_("git diff header lacks filename information when removing "
1564                                                         "%d leading pathname component (line %d)",
1565                                                         "git diff header lacks filename information when removing "
1566                                                         "%d leading pathname components (line %d)",
1567                                                         state->p_value),
1568                                                      state->p_value, state->linenr);
1569                                         return -128;
1570                                 }
1571                                 patch->old_name = xstrdup(patch->def_name);
1572                                 patch->new_name = xstrdup(patch->def_name);
1573                         }
1574                         if ((!patch->new_name && !patch->is_delete) ||
1575                             (!patch->old_name && !patch->is_new)) {
1576                                 error(_("git diff header lacks filename information "
1577                                              "(line %d)"), state->linenr);
1578                                 return -128;
1579                         }
1580                         patch->is_toplevel_relative = 1;
1581                         *hdrsize = git_hdr_len;
1582                         return offset;
1583                 }
1584
1585                 /* --- followed by +++ ? */
1586                 if (memcmp("--- ", line,  4) || memcmp("+++ ", line + len, 4))
1587                         continue;
1588
1589                 /*
1590                  * We only accept unified patches, so we want it to
1591                  * at least have "@@ -a,b +c,d @@\n", which is 14 chars
1592                  * minimum ("@@ -0,0 +1 @@\n" is the shortest).
1593                  */
1594                 nextlen = linelen(line + len, size - len);
1595                 if (size < nextlen + 14 || memcmp("@@ -", line + len + nextlen, 4))
1596                         continue;
1597
1598                 /* Ok, we'll consider it a patch */
1599                 if (parse_traditional_patch(state, line, line+len, patch))
1600                         return -128;
1601                 *hdrsize = len + nextlen;
1602                 state->linenr += 2;
1603                 return offset;
1604         }
1605         return -1;
1606 }
1607
1608 static void record_ws_error(struct apply_state *state,
1609                             unsigned result,
1610                             const char *line,
1611                             int len,
1612                             int linenr)
1613 {
1614         char *err;
1615
1616         if (!result)
1617                 return;
1618
1619         state->whitespace_error++;
1620         if (state->squelch_whitespace_errors &&
1621             state->squelch_whitespace_errors < state->whitespace_error)
1622                 return;
1623
1624         err = whitespace_error_string(result);
1625         if (state->apply_verbosity > verbosity_silent)
1626                 fprintf(stderr, "%s:%d: %s.\n%.*s\n",
1627                         state->patch_input_file, linenr, err, len, line);
1628         free(err);
1629 }
1630
1631 static void check_whitespace(struct apply_state *state,
1632                              const char *line,
1633                              int len,
1634                              unsigned ws_rule)
1635 {
1636         unsigned result = ws_check(line + 1, len - 1, ws_rule);
1637
1638         record_ws_error(state, result, line + 1, len - 2, state->linenr);
1639 }
1640
1641 /*
1642  * Check if the patch has context lines with CRLF or
1643  * the patch wants to remove lines with CRLF.
1644  */
1645 static void check_old_for_crlf(struct patch *patch, const char *line, int len)
1646 {
1647         if (len >= 2 && line[len-1] == '\n' && line[len-2] == '\r') {
1648                 patch->ws_rule |= WS_CR_AT_EOL;
1649                 patch->crlf_in_old = 1;
1650         }
1651 }
1652
1653
1654 /*
1655  * Parse a unified diff. Note that this really needs to parse each
1656  * fragment separately, since the only way to know the difference
1657  * between a "---" that is part of a patch, and a "---" that starts
1658  * the next patch is to look at the line counts..
1659  */
1660 static int parse_fragment(struct apply_state *state,
1661                           const char *line,
1662                           unsigned long size,
1663                           struct patch *patch,
1664                           struct fragment *fragment)
1665 {
1666         int added, deleted;
1667         int len = linelen(line, size), offset;
1668         unsigned long oldlines, newlines;
1669         unsigned long leading, trailing;
1670
1671         offset = parse_fragment_header(line, len, fragment);
1672         if (offset < 0)
1673                 return -1;
1674         if (offset > 0 && patch->recount)
1675                 recount_diff(line + offset, size - offset, fragment);
1676         oldlines = fragment->oldlines;
1677         newlines = fragment->newlines;
1678         leading = 0;
1679         trailing = 0;
1680
1681         /* Parse the thing.. */
1682         line += len;
1683         size -= len;
1684         state->linenr++;
1685         added = deleted = 0;
1686         for (offset = len;
1687              0 < size;
1688              offset += len, size -= len, line += len, state->linenr++) {
1689                 if (!oldlines && !newlines)
1690                         break;
1691                 len = linelen(line, size);
1692                 if (!len || line[len-1] != '\n')
1693                         return -1;
1694                 switch (*line) {
1695                 default:
1696                         return -1;
1697                 case '\n': /* newer GNU diff, an empty context line */
1698                 case ' ':
1699                         oldlines--;
1700                         newlines--;
1701                         if (!deleted && !added)
1702                                 leading++;
1703                         trailing++;
1704                         check_old_for_crlf(patch, line, len);
1705                         if (!state->apply_in_reverse &&
1706                             state->ws_error_action == correct_ws_error)
1707                                 check_whitespace(state, line, len, patch->ws_rule);
1708                         break;
1709                 case '-':
1710                         if (!state->apply_in_reverse)
1711                                 check_old_for_crlf(patch, line, len);
1712                         if (state->apply_in_reverse &&
1713                             state->ws_error_action != nowarn_ws_error)
1714                                 check_whitespace(state, line, len, patch->ws_rule);
1715                         deleted++;
1716                         oldlines--;
1717                         trailing = 0;
1718                         break;
1719                 case '+':
1720                         if (state->apply_in_reverse)
1721                                 check_old_for_crlf(patch, line, len);
1722                         if (!state->apply_in_reverse &&
1723                             state->ws_error_action != nowarn_ws_error)
1724                                 check_whitespace(state, line, len, patch->ws_rule);
1725                         added++;
1726                         newlines--;
1727                         trailing = 0;
1728                         break;
1729
1730                 /*
1731                  * We allow "\ No newline at end of file". Depending
1732                  * on locale settings when the patch was produced we
1733                  * don't know what this line looks like. The only
1734                  * thing we do know is that it begins with "\ ".
1735                  * Checking for 12 is just for sanity check -- any
1736                  * l10n of "\ No newline..." is at least that long.
1737                  */
1738                 case '\\':
1739                         if (len < 12 || memcmp(line, "\\ ", 2))
1740                                 return -1;
1741                         break;
1742                 }
1743         }
1744         if (oldlines || newlines)
1745                 return -1;
1746         if (!deleted && !added)
1747                 return -1;
1748
1749         fragment->leading = leading;
1750         fragment->trailing = trailing;
1751
1752         /*
1753          * If a fragment ends with an incomplete line, we failed to include
1754          * it in the above loop because we hit oldlines == newlines == 0
1755          * before seeing it.
1756          */
1757         if (12 < size && !memcmp(line, "\\ ", 2))
1758                 offset += linelen(line, size);
1759
1760         patch->lines_added += added;
1761         patch->lines_deleted += deleted;
1762
1763         if (0 < patch->is_new && oldlines)
1764                 return error(_("new file depends on old contents"));
1765         if (0 < patch->is_delete && newlines)
1766                 return error(_("deleted file still has contents"));
1767         return offset;
1768 }
1769
1770 /*
1771  * We have seen "diff --git a/... b/..." header (or a traditional patch
1772  * header).  Read hunks that belong to this patch into fragments and hang
1773  * them to the given patch structure.
1774  *
1775  * The (fragment->patch, fragment->size) pair points into the memory given
1776  * by the caller, not a copy, when we return.
1777  *
1778  * Returns:
1779  *   -1 in case of error,
1780  *   the number of bytes in the patch otherwise.
1781  */
1782 static int parse_single_patch(struct apply_state *state,
1783                               const char *line,
1784                               unsigned long size,
1785                               struct patch *patch)
1786 {
1787         unsigned long offset = 0;
1788         unsigned long oldlines = 0, newlines = 0, context = 0;
1789         struct fragment **fragp = &patch->fragments;
1790
1791         while (size > 4 && !memcmp(line, "@@ -", 4)) {
1792                 struct fragment *fragment;
1793                 int len;
1794
1795                 fragment = xcalloc(1, sizeof(*fragment));
1796                 fragment->linenr = state->linenr;
1797                 len = parse_fragment(state, line, size, patch, fragment);
1798                 if (len <= 0) {
1799                         free(fragment);
1800                         return error(_("corrupt patch at line %d"), state->linenr);
1801                 }
1802                 fragment->patch = line;
1803                 fragment->size = len;
1804                 oldlines += fragment->oldlines;
1805                 newlines += fragment->newlines;
1806                 context += fragment->leading + fragment->trailing;
1807
1808                 *fragp = fragment;
1809                 fragp = &fragment->next;
1810
1811                 offset += len;
1812                 line += len;
1813                 size -= len;
1814         }
1815
1816         /*
1817          * If something was removed (i.e. we have old-lines) it cannot
1818          * be creation, and if something was added it cannot be
1819          * deletion.  However, the reverse is not true; --unified=0
1820          * patches that only add are not necessarily creation even
1821          * though they do not have any old lines, and ones that only
1822          * delete are not necessarily deletion.
1823          *
1824          * Unfortunately, a real creation/deletion patch do _not_ have
1825          * any context line by definition, so we cannot safely tell it
1826          * apart with --unified=0 insanity.  At least if the patch has
1827          * more than one hunk it is not creation or deletion.
1828          */
1829         if (patch->is_new < 0 &&
1830             (oldlines || (patch->fragments && patch->fragments->next)))
1831                 patch->is_new = 0;
1832         if (patch->is_delete < 0 &&
1833             (newlines || (patch->fragments && patch->fragments->next)))
1834                 patch->is_delete = 0;
1835
1836         if (0 < patch->is_new && oldlines)
1837                 return error(_("new file %s depends on old contents"), patch->new_name);
1838         if (0 < patch->is_delete && newlines)
1839                 return error(_("deleted file %s still has contents"), patch->old_name);
1840         if (!patch->is_delete && !newlines && context && state->apply_verbosity > verbosity_silent)
1841                 fprintf_ln(stderr,
1842                            _("** warning: "
1843                              "file %s becomes empty but is not deleted"),
1844                            patch->new_name);
1845
1846         return offset;
1847 }
1848
1849 static inline int metadata_changes(struct patch *patch)
1850 {
1851         return  patch->is_rename > 0 ||
1852                 patch->is_copy > 0 ||
1853                 patch->is_new > 0 ||
1854                 patch->is_delete ||
1855                 (patch->old_mode && patch->new_mode &&
1856                  patch->old_mode != patch->new_mode);
1857 }
1858
1859 static char *inflate_it(const void *data, unsigned long size,
1860                         unsigned long inflated_size)
1861 {
1862         git_zstream stream;
1863         void *out;
1864         int st;
1865
1866         memset(&stream, 0, sizeof(stream));
1867
1868         stream.next_in = (unsigned char *)data;
1869         stream.avail_in = size;
1870         stream.next_out = out = xmalloc(inflated_size);
1871         stream.avail_out = inflated_size;
1872         git_inflate_init(&stream);
1873         st = git_inflate(&stream, Z_FINISH);
1874         git_inflate_end(&stream);
1875         if ((st != Z_STREAM_END) || stream.total_out != inflated_size) {
1876                 free(out);
1877                 return NULL;
1878         }
1879         return out;
1880 }
1881
1882 /*
1883  * Read a binary hunk and return a new fragment; fragment->patch
1884  * points at an allocated memory that the caller must free, so
1885  * it is marked as "->free_patch = 1".
1886  */
1887 static struct fragment *parse_binary_hunk(struct apply_state *state,
1888                                           char **buf_p,
1889                                           unsigned long *sz_p,
1890                                           int *status_p,
1891                                           int *used_p)
1892 {
1893         /*
1894          * Expect a line that begins with binary patch method ("literal"
1895          * or "delta"), followed by the length of data before deflating.
1896          * a sequence of 'length-byte' followed by base-85 encoded data
1897          * should follow, terminated by a newline.
1898          *
1899          * Each 5-byte sequence of base-85 encodes up to 4 bytes,
1900          * and we would limit the patch line to 66 characters,
1901          * so one line can fit up to 13 groups that would decode
1902          * to 52 bytes max.  The length byte 'A'-'Z' corresponds
1903          * to 1-26 bytes, and 'a'-'z' corresponds to 27-52 bytes.
1904          */
1905         int llen, used;
1906         unsigned long size = *sz_p;
1907         char *buffer = *buf_p;
1908         int patch_method;
1909         unsigned long origlen;
1910         char *data = NULL;
1911         int hunk_size = 0;
1912         struct fragment *frag;
1913
1914         llen = linelen(buffer, size);
1915         used = llen;
1916
1917         *status_p = 0;
1918
1919         if (starts_with(buffer, "delta ")) {
1920                 patch_method = BINARY_DELTA_DEFLATED;
1921                 origlen = strtoul(buffer + 6, NULL, 10);
1922         }
1923         else if (starts_with(buffer, "literal ")) {
1924                 patch_method = BINARY_LITERAL_DEFLATED;
1925                 origlen = strtoul(buffer + 8, NULL, 10);
1926         }
1927         else
1928                 return NULL;
1929
1930         state->linenr++;
1931         buffer += llen;
1932         while (1) {
1933                 int byte_length, max_byte_length, newsize;
1934                 llen = linelen(buffer, size);
1935                 used += llen;
1936                 state->linenr++;
1937                 if (llen == 1) {
1938                         /* consume the blank line */
1939                         buffer++;
1940                         size--;
1941                         break;
1942                 }
1943                 /*
1944                  * Minimum line is "A00000\n" which is 7-byte long,
1945                  * and the line length must be multiple of 5 plus 2.
1946                  */
1947                 if ((llen < 7) || (llen-2) % 5)
1948                         goto corrupt;
1949                 max_byte_length = (llen - 2) / 5 * 4;
1950                 byte_length = *buffer;
1951                 if ('A' <= byte_length && byte_length <= 'Z')
1952                         byte_length = byte_length - 'A' + 1;
1953                 else if ('a' <= byte_length && byte_length <= 'z')
1954                         byte_length = byte_length - 'a' + 27;
1955                 else
1956                         goto corrupt;
1957                 /* if the input length was not multiple of 4, we would
1958                  * have filler at the end but the filler should never
1959                  * exceed 3 bytes
1960                  */
1961                 if (max_byte_length < byte_length ||
1962                     byte_length <= max_byte_length - 4)
1963                         goto corrupt;
1964                 newsize = hunk_size + byte_length;
1965                 data = xrealloc(data, newsize);
1966                 if (decode_85(data + hunk_size, buffer + 1, byte_length))
1967                         goto corrupt;
1968                 hunk_size = newsize;
1969                 buffer += llen;
1970                 size -= llen;
1971         }
1972
1973         frag = xcalloc(1, sizeof(*frag));
1974         frag->patch = inflate_it(data, hunk_size, origlen);
1975         frag->free_patch = 1;
1976         if (!frag->patch)
1977                 goto corrupt;
1978         free(data);
1979         frag->size = origlen;
1980         *buf_p = buffer;
1981         *sz_p = size;
1982         *used_p = used;
1983         frag->binary_patch_method = patch_method;
1984         return frag;
1985
1986  corrupt:
1987         free(data);
1988         *status_p = -1;
1989         error(_("corrupt binary patch at line %d: %.*s"),
1990               state->linenr-1, llen-1, buffer);
1991         return NULL;
1992 }
1993
1994 /*
1995  * Returns:
1996  *   -1 in case of error,
1997  *   the length of the parsed binary patch otherwise
1998  */
1999 static int parse_binary(struct apply_state *state,
2000                         char *buffer,
2001                         unsigned long size,
2002                         struct patch *patch)
2003 {
2004         /*
2005          * We have read "GIT binary patch\n"; what follows is a line
2006          * that says the patch method (currently, either "literal" or
2007          * "delta") and the length of data before deflating; a
2008          * sequence of 'length-byte' followed by base-85 encoded data
2009          * follows.
2010          *
2011          * When a binary patch is reversible, there is another binary
2012          * hunk in the same format, starting with patch method (either
2013          * "literal" or "delta") with the length of data, and a sequence
2014          * of length-byte + base-85 encoded data, terminated with another
2015          * empty line.  This data, when applied to the postimage, produces
2016          * the preimage.
2017          */
2018         struct fragment *forward;
2019         struct fragment *reverse;
2020         int status;
2021         int used, used_1;
2022
2023         forward = parse_binary_hunk(state, &buffer, &size, &status, &used);
2024         if (!forward && !status)
2025                 /* there has to be one hunk (forward hunk) */
2026                 return error(_("unrecognized binary patch at line %d"), state->linenr-1);
2027         if (status)
2028                 /* otherwise we already gave an error message */
2029                 return status;
2030
2031         reverse = parse_binary_hunk(state, &buffer, &size, &status, &used_1);
2032         if (reverse)
2033                 used += used_1;
2034         else if (status) {
2035                 /*
2036                  * Not having reverse hunk is not an error, but having
2037                  * a corrupt reverse hunk is.
2038                  */
2039                 free((void*) forward->patch);
2040                 free(forward);
2041                 return status;
2042         }
2043         forward->next = reverse;
2044         patch->fragments = forward;
2045         patch->is_binary = 1;
2046         return used;
2047 }
2048
2049 static void prefix_one(struct apply_state *state, char **name)
2050 {
2051         char *old_name = *name;
2052         if (!old_name)
2053                 return;
2054         *name = prefix_filename(state->prefix, *name);
2055         free(old_name);
2056 }
2057
2058 static void prefix_patch(struct apply_state *state, struct patch *p)
2059 {
2060         if (!state->prefix || p->is_toplevel_relative)
2061                 return;
2062         prefix_one(state, &p->new_name);
2063         prefix_one(state, &p->old_name);
2064 }
2065
2066 /*
2067  * include/exclude
2068  */
2069
2070 static void add_name_limit(struct apply_state *state,
2071                            const char *name,
2072                            int exclude)
2073 {
2074         struct string_list_item *it;
2075
2076         it = string_list_append(&state->limit_by_name, name);
2077         it->util = exclude ? NULL : (void *) 1;
2078 }
2079
2080 static int use_patch(struct apply_state *state, struct patch *p)
2081 {
2082         const char *pathname = p->new_name ? p->new_name : p->old_name;
2083         int i;
2084
2085         /* Paths outside are not touched regardless of "--include" */
2086         if (state->prefix && *state->prefix) {
2087                 const char *rest;
2088                 if (!skip_prefix(pathname, state->prefix, &rest) || !*rest)
2089                         return 0;
2090         }
2091
2092         /* See if it matches any of exclude/include rule */
2093         for (i = 0; i < state->limit_by_name.nr; i++) {
2094                 struct string_list_item *it = &state->limit_by_name.items[i];
2095                 if (!wildmatch(it->string, pathname, 0))
2096                         return (it->util != NULL);
2097         }
2098
2099         /*
2100          * If we had any include, a path that does not match any rule is
2101          * not used.  Otherwise, we saw bunch of exclude rules (or none)
2102          * and such a path is used.
2103          */
2104         return !state->has_include;
2105 }
2106
2107 /*
2108  * Read the patch text in "buffer" that extends for "size" bytes; stop
2109  * reading after seeing a single patch (i.e. changes to a single file).
2110  * Create fragments (i.e. patch hunks) and hang them to the given patch.
2111  *
2112  * Returns:
2113  *   -1 if no header was found or parse_binary() failed,
2114  *   -128 on another error,
2115  *   the number of bytes consumed otherwise,
2116  *     so that the caller can call us again for the next patch.
2117  */
2118 static int parse_chunk(struct apply_state *state, char *buffer, unsigned long size, struct patch *patch)
2119 {
2120         int hdrsize, patchsize;
2121         int offset = find_header(state, buffer, size, &hdrsize, patch);
2122
2123         if (offset < 0)
2124                 return offset;
2125
2126         prefix_patch(state, patch);
2127
2128         if (!use_patch(state, patch))
2129                 patch->ws_rule = 0;
2130         else
2131                 patch->ws_rule = whitespace_rule(patch->new_name
2132                                                  ? patch->new_name
2133                                                  : patch->old_name);
2134
2135         patchsize = parse_single_patch(state,
2136                                        buffer + offset + hdrsize,
2137                                        size - offset - hdrsize,
2138                                        patch);
2139
2140         if (patchsize < 0)
2141                 return -128;
2142
2143         if (!patchsize) {
2144                 static const char git_binary[] = "GIT binary patch\n";
2145                 int hd = hdrsize + offset;
2146                 unsigned long llen = linelen(buffer + hd, size - hd);
2147
2148                 if (llen == sizeof(git_binary) - 1 &&
2149                     !memcmp(git_binary, buffer + hd, llen)) {
2150                         int used;
2151                         state->linenr++;
2152                         used = parse_binary(state, buffer + hd + llen,
2153                                             size - hd - llen, patch);
2154                         if (used < 0)
2155                                 return -1;
2156                         if (used)
2157                                 patchsize = used + llen;
2158                         else
2159                                 patchsize = 0;
2160                 }
2161                 else if (!memcmp(" differ\n", buffer + hd + llen - 8, 8)) {
2162                         static const char *binhdr[] = {
2163                                 "Binary files ",
2164                                 "Files ",
2165                                 NULL,
2166                         };
2167                         int i;
2168                         for (i = 0; binhdr[i]; i++) {
2169                                 int len = strlen(binhdr[i]);
2170                                 if (len < size - hd &&
2171                                     !memcmp(binhdr[i], buffer + hd, len)) {
2172                                         state->linenr++;
2173                                         patch->is_binary = 1;
2174                                         patchsize = llen;
2175                                         break;
2176                                 }
2177                         }
2178                 }
2179
2180                 /* Empty patch cannot be applied if it is a text patch
2181                  * without metadata change.  A binary patch appears
2182                  * empty to us here.
2183                  */
2184                 if ((state->apply || state->check) &&
2185                     (!patch->is_binary && !metadata_changes(patch))) {
2186                         error(_("patch with only garbage at line %d"), state->linenr);
2187                         return -128;
2188                 }
2189         }
2190
2191         return offset + hdrsize + patchsize;
2192 }
2193
2194 static void reverse_patches(struct patch *p)
2195 {
2196         for (; p; p = p->next) {
2197                 struct fragment *frag = p->fragments;
2198
2199                 SWAP(p->new_name, p->old_name);
2200                 SWAP(p->new_mode, p->old_mode);
2201                 SWAP(p->is_new, p->is_delete);
2202                 SWAP(p->lines_added, p->lines_deleted);
2203                 SWAP(p->old_sha1_prefix, p->new_sha1_prefix);
2204
2205                 for (; frag; frag = frag->next) {
2206                         SWAP(frag->newpos, frag->oldpos);
2207                         SWAP(frag->newlines, frag->oldlines);
2208                 }
2209         }
2210 }
2211
2212 static const char pluses[] =
2213 "++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++";
2214 static const char minuses[]=
2215 "----------------------------------------------------------------------";
2216
2217 static void show_stats(struct apply_state *state, struct patch *patch)
2218 {
2219         struct strbuf qname = STRBUF_INIT;
2220         char *cp = patch->new_name ? patch->new_name : patch->old_name;
2221         int max, add, del;
2222
2223         quote_c_style(cp, &qname, NULL, 0);
2224
2225         /*
2226          * "scale" the filename
2227          */
2228         max = state->max_len;
2229         if (max > 50)
2230                 max = 50;
2231
2232         if (qname.len > max) {
2233                 cp = strchr(qname.buf + qname.len + 3 - max, '/');
2234                 if (!cp)
2235                         cp = qname.buf + qname.len + 3 - max;
2236                 strbuf_splice(&qname, 0, cp - qname.buf, "...", 3);
2237         }
2238
2239         if (patch->is_binary) {
2240                 printf(" %-*s |  Bin\n", max, qname.buf);
2241                 strbuf_release(&qname);
2242                 return;
2243         }
2244
2245         printf(" %-*s |", max, qname.buf);
2246         strbuf_release(&qname);
2247
2248         /*
2249          * scale the add/delete
2250          */
2251         max = max + state->max_change > 70 ? 70 - max : state->max_change;
2252         add = patch->lines_added;
2253         del = patch->lines_deleted;
2254
2255         if (state->max_change > 0) {
2256                 int total = ((add + del) * max + state->max_change / 2) / state->max_change;
2257                 add = (add * max + state->max_change / 2) / state->max_change;
2258                 del = total - add;
2259         }
2260         printf("%5d %.*s%.*s\n", patch->lines_added + patch->lines_deleted,
2261                 add, pluses, del, minuses);
2262 }
2263
2264 static int read_old_data(struct stat *st, struct patch *patch,
2265                          const char *path, struct strbuf *buf)
2266 {
2267         int conv_flags = patch->crlf_in_old ?
2268                 CONV_EOL_KEEP_CRLF : CONV_EOL_RENORMALIZE;
2269         switch (st->st_mode & S_IFMT) {
2270         case S_IFLNK:
2271                 if (strbuf_readlink(buf, path, st->st_size) < 0)
2272                         return error(_("unable to read symlink %s"), path);
2273                 return 0;
2274         case S_IFREG:
2275                 if (strbuf_read_file(buf, path, st->st_size) != st->st_size)
2276                         return error(_("unable to open or read %s"), path);
2277                 /*
2278                  * "git apply" without "--index/--cached" should never look
2279                  * at the index; the target file may not have been added to
2280                  * the index yet, and we may not even be in any Git repository.
2281                  * Pass NULL to convert_to_git() to stress this; the function
2282                  * should never look at the index when explicit crlf option
2283                  * is given.
2284                  */
2285                 convert_to_git(NULL, path, buf->buf, buf->len, buf, conv_flags);
2286                 return 0;
2287         default:
2288                 return -1;
2289         }
2290 }
2291
2292 /*
2293  * Update the preimage, and the common lines in postimage,
2294  * from buffer buf of length len. If postlen is 0 the postimage
2295  * is updated in place, otherwise it's updated on a new buffer
2296  * of length postlen
2297  */
2298
2299 static void update_pre_post_images(struct image *preimage,
2300                                    struct image *postimage,
2301                                    char *buf,
2302                                    size_t len, size_t postlen)
2303 {
2304         int i, ctx, reduced;
2305         char *new_buf, *old_buf, *fixed;
2306         struct image fixed_preimage;
2307
2308         /*
2309          * Update the preimage with whitespace fixes.  Note that we
2310          * are not losing preimage->buf -- apply_one_fragment() will
2311          * free "oldlines".
2312          */
2313         prepare_image(&fixed_preimage, buf, len, 1);
2314         assert(postlen
2315                ? fixed_preimage.nr == preimage->nr
2316                : fixed_preimage.nr <= preimage->nr);
2317         for (i = 0; i < fixed_preimage.nr; i++)
2318                 fixed_preimage.line[i].flag = preimage->line[i].flag;
2319         free(preimage->line_allocated);
2320         *preimage = fixed_preimage;
2321
2322         /*
2323          * Adjust the common context lines in postimage. This can be
2324          * done in-place when we are shrinking it with whitespace
2325          * fixing, but needs a new buffer when ignoring whitespace or
2326          * expanding leading tabs to spaces.
2327          *
2328          * We trust the caller to tell us if the update can be done
2329          * in place (postlen==0) or not.
2330          */
2331         old_buf = postimage->buf;
2332         if (postlen)
2333                 new_buf = postimage->buf = xmalloc(postlen);
2334         else
2335                 new_buf = old_buf;
2336         fixed = preimage->buf;
2337
2338         for (i = reduced = ctx = 0; i < postimage->nr; i++) {
2339                 size_t l_len = postimage->line[i].len;
2340                 if (!(postimage->line[i].flag & LINE_COMMON)) {
2341                         /* an added line -- no counterparts in preimage */
2342                         memmove(new_buf, old_buf, l_len);
2343                         old_buf += l_len;
2344                         new_buf += l_len;
2345                         continue;
2346                 }
2347
2348                 /* a common context -- skip it in the original postimage */
2349                 old_buf += l_len;
2350
2351                 /* and find the corresponding one in the fixed preimage */
2352                 while (ctx < preimage->nr &&
2353                        !(preimage->line[ctx].flag & LINE_COMMON)) {
2354                         fixed += preimage->line[ctx].len;
2355                         ctx++;
2356                 }
2357
2358                 /*
2359                  * preimage is expected to run out, if the caller
2360                  * fixed addition of trailing blank lines.
2361                  */
2362                 if (preimage->nr <= ctx) {
2363                         reduced++;
2364                         continue;
2365                 }
2366
2367                 /* and copy it in, while fixing the line length */
2368                 l_len = preimage->line[ctx].len;
2369                 memcpy(new_buf, fixed, l_len);
2370                 new_buf += l_len;
2371                 fixed += l_len;
2372                 postimage->line[i].len = l_len;
2373                 ctx++;
2374         }
2375
2376         if (postlen
2377             ? postlen < new_buf - postimage->buf
2378             : postimage->len < new_buf - postimage->buf)
2379                 die("BUG: caller miscounted postlen: asked %d, orig = %d, used = %d",
2380                     (int)postlen, (int) postimage->len, (int)(new_buf - postimage->buf));
2381
2382         /* Fix the length of the whole thing */
2383         postimage->len = new_buf - postimage->buf;
2384         postimage->nr -= reduced;
2385 }
2386
2387 static int line_by_line_fuzzy_match(struct image *img,
2388                                     struct image *preimage,
2389                                     struct image *postimage,
2390                                     unsigned long current,
2391                                     int current_lno,
2392                                     int preimage_limit)
2393 {
2394         int i;
2395         size_t imgoff = 0;
2396         size_t preoff = 0;
2397         size_t postlen = postimage->len;
2398         size_t extra_chars;
2399         char *buf;
2400         char *preimage_eof;
2401         char *preimage_end;
2402         struct strbuf fixed;
2403         char *fixed_buf;
2404         size_t fixed_len;
2405
2406         for (i = 0; i < preimage_limit; i++) {
2407                 size_t prelen = preimage->line[i].len;
2408                 size_t imglen = img->line[current_lno+i].len;
2409
2410                 if (!fuzzy_matchlines(img->buf + current + imgoff, imglen,
2411                                       preimage->buf + preoff, prelen))
2412                         return 0;
2413                 if (preimage->line[i].flag & LINE_COMMON)
2414                         postlen += imglen - prelen;
2415                 imgoff += imglen;
2416                 preoff += prelen;
2417         }
2418
2419         /*
2420          * Ok, the preimage matches with whitespace fuzz.
2421          *
2422          * imgoff now holds the true length of the target that
2423          * matches the preimage before the end of the file.
2424          *
2425          * Count the number of characters in the preimage that fall
2426          * beyond the end of the file and make sure that all of them
2427          * are whitespace characters. (This can only happen if
2428          * we are removing blank lines at the end of the file.)
2429          */
2430         buf = preimage_eof = preimage->buf + preoff;
2431         for ( ; i < preimage->nr; i++)
2432                 preoff += preimage->line[i].len;
2433         preimage_end = preimage->buf + preoff;
2434         for ( ; buf < preimage_end; buf++)
2435                 if (!isspace(*buf))
2436                         return 0;
2437
2438         /*
2439          * Update the preimage and the common postimage context
2440          * lines to use the same whitespace as the target.
2441          * If whitespace is missing in the target (i.e.
2442          * if the preimage extends beyond the end of the file),
2443          * use the whitespace from the preimage.
2444          */
2445         extra_chars = preimage_end - preimage_eof;
2446         strbuf_init(&fixed, imgoff + extra_chars);
2447         strbuf_add(&fixed, img->buf + current, imgoff);
2448         strbuf_add(&fixed, preimage_eof, extra_chars);
2449         fixed_buf = strbuf_detach(&fixed, &fixed_len);
2450         update_pre_post_images(preimage, postimage,
2451                                fixed_buf, fixed_len, postlen);
2452         return 1;
2453 }
2454
2455 static int match_fragment(struct apply_state *state,
2456                           struct image *img,
2457                           struct image *preimage,
2458                           struct image *postimage,
2459                           unsigned long current,
2460                           int current_lno,
2461                           unsigned ws_rule,
2462                           int match_beginning, int match_end)
2463 {
2464         int i;
2465         char *fixed_buf, *buf, *orig, *target;
2466         struct strbuf fixed;
2467         size_t fixed_len, postlen;
2468         int preimage_limit;
2469
2470         if (preimage->nr + current_lno <= img->nr) {
2471                 /*
2472                  * The hunk falls within the boundaries of img.
2473                  */
2474                 preimage_limit = preimage->nr;
2475                 if (match_end && (preimage->nr + current_lno != img->nr))
2476                         return 0;
2477         } else if (state->ws_error_action == correct_ws_error &&
2478                    (ws_rule & WS_BLANK_AT_EOF)) {
2479                 /*
2480                  * This hunk extends beyond the end of img, and we are
2481                  * removing blank lines at the end of the file.  This
2482                  * many lines from the beginning of the preimage must
2483                  * match with img, and the remainder of the preimage
2484                  * must be blank.
2485                  */
2486                 preimage_limit = img->nr - current_lno;
2487         } else {
2488                 /*
2489                  * The hunk extends beyond the end of the img and
2490                  * we are not removing blanks at the end, so we
2491                  * should reject the hunk at this position.
2492                  */
2493                 return 0;
2494         }
2495
2496         if (match_beginning && current_lno)
2497                 return 0;
2498
2499         /* Quick hash check */
2500         for (i = 0; i < preimage_limit; i++)
2501                 if ((img->line[current_lno + i].flag & LINE_PATCHED) ||
2502                     (preimage->line[i].hash != img->line[current_lno + i].hash))
2503                         return 0;
2504
2505         if (preimage_limit == preimage->nr) {
2506                 /*
2507                  * Do we have an exact match?  If we were told to match
2508                  * at the end, size must be exactly at current+fragsize,
2509                  * otherwise current+fragsize must be still within the preimage,
2510                  * and either case, the old piece should match the preimage
2511                  * exactly.
2512                  */
2513                 if ((match_end
2514                      ? (current + preimage->len == img->len)
2515                      : (current + preimage->len <= img->len)) &&
2516                     !memcmp(img->buf + current, preimage->buf, preimage->len))
2517                         return 1;
2518         } else {
2519                 /*
2520                  * The preimage extends beyond the end of img, so
2521                  * there cannot be an exact match.
2522                  *
2523                  * There must be one non-blank context line that match
2524                  * a line before the end of img.
2525                  */
2526                 char *buf_end;
2527
2528                 buf = preimage->buf;
2529                 buf_end = buf;
2530                 for (i = 0; i < preimage_limit; i++)
2531                         buf_end += preimage->line[i].len;
2532
2533                 for ( ; buf < buf_end; buf++)
2534                         if (!isspace(*buf))
2535                                 break;
2536                 if (buf == buf_end)
2537                         return 0;
2538         }
2539
2540         /*
2541          * No exact match. If we are ignoring whitespace, run a line-by-line
2542          * fuzzy matching. We collect all the line length information because
2543          * we need it to adjust whitespace if we match.
2544          */
2545         if (state->ws_ignore_action == ignore_ws_change)
2546                 return line_by_line_fuzzy_match(img, preimage, postimage,
2547                                                 current, current_lno, preimage_limit);
2548
2549         if (state->ws_error_action != correct_ws_error)
2550                 return 0;
2551
2552         /*
2553          * The hunk does not apply byte-by-byte, but the hash says
2554          * it might with whitespace fuzz. We weren't asked to
2555          * ignore whitespace, we were asked to correct whitespace
2556          * errors, so let's try matching after whitespace correction.
2557          *
2558          * While checking the preimage against the target, whitespace
2559          * errors in both fixed, we count how large the corresponding
2560          * postimage needs to be.  The postimage prepared by
2561          * apply_one_fragment() has whitespace errors fixed on added
2562          * lines already, but the common lines were propagated as-is,
2563          * which may become longer when their whitespace errors are
2564          * fixed.
2565          */
2566
2567         /* First count added lines in postimage */
2568         postlen = 0;
2569         for (i = 0; i < postimage->nr; i++) {
2570                 if (!(postimage->line[i].flag & LINE_COMMON))
2571                         postlen += postimage->line[i].len;
2572         }
2573
2574         /*
2575          * The preimage may extend beyond the end of the file,
2576          * but in this loop we will only handle the part of the
2577          * preimage that falls within the file.
2578          */
2579         strbuf_init(&fixed, preimage->len + 1);
2580         orig = preimage->buf;
2581         target = img->buf + current;
2582         for (i = 0; i < preimage_limit; i++) {
2583                 size_t oldlen = preimage->line[i].len;
2584                 size_t tgtlen = img->line[current_lno + i].len;
2585                 size_t fixstart = fixed.len;
2586                 struct strbuf tgtfix;
2587                 int match;
2588
2589                 /* Try fixing the line in the preimage */
2590                 ws_fix_copy(&fixed, orig, oldlen, ws_rule, NULL);
2591
2592                 /* Try fixing the line in the target */
2593                 strbuf_init(&tgtfix, tgtlen);
2594                 ws_fix_copy(&tgtfix, target, tgtlen, ws_rule, NULL);
2595
2596                 /*
2597                  * If they match, either the preimage was based on
2598                  * a version before our tree fixed whitespace breakage,
2599                  * or we are lacking a whitespace-fix patch the tree
2600                  * the preimage was based on already had (i.e. target
2601                  * has whitespace breakage, the preimage doesn't).
2602                  * In either case, we are fixing the whitespace breakages
2603                  * so we might as well take the fix together with their
2604                  * real change.
2605                  */
2606                 match = (tgtfix.len == fixed.len - fixstart &&
2607                          !memcmp(tgtfix.buf, fixed.buf + fixstart,
2608                                              fixed.len - fixstart));
2609
2610                 /* Add the length if this is common with the postimage */
2611                 if (preimage->line[i].flag & LINE_COMMON)
2612                         postlen += tgtfix.len;
2613
2614                 strbuf_release(&tgtfix);
2615                 if (!match)
2616                         goto unmatch_exit;
2617
2618                 orig += oldlen;
2619                 target += tgtlen;
2620         }
2621
2622
2623         /*
2624          * Now handle the lines in the preimage that falls beyond the
2625          * end of the file (if any). They will only match if they are
2626          * empty or only contain whitespace (if WS_BLANK_AT_EOL is
2627          * false).
2628          */
2629         for ( ; i < preimage->nr; i++) {
2630                 size_t fixstart = fixed.len; /* start of the fixed preimage */
2631                 size_t oldlen = preimage->line[i].len;
2632                 int j;
2633
2634                 /* Try fixing the line in the preimage */
2635                 ws_fix_copy(&fixed, orig, oldlen, ws_rule, NULL);
2636
2637                 for (j = fixstart; j < fixed.len; j++)
2638                         if (!isspace(fixed.buf[j]))
2639                                 goto unmatch_exit;
2640
2641                 orig += oldlen;
2642         }
2643
2644         /*
2645          * Yes, the preimage is based on an older version that still
2646          * has whitespace breakages unfixed, and fixing them makes the
2647          * hunk match.  Update the context lines in the postimage.
2648          */
2649         fixed_buf = strbuf_detach(&fixed, &fixed_len);
2650         if (postlen < postimage->len)
2651                 postlen = 0;
2652         update_pre_post_images(preimage, postimage,
2653                                fixed_buf, fixed_len, postlen);
2654         return 1;
2655
2656  unmatch_exit:
2657         strbuf_release(&fixed);
2658         return 0;
2659 }
2660
2661 static int find_pos(struct apply_state *state,
2662                     struct image *img,
2663                     struct image *preimage,
2664                     struct image *postimage,
2665                     int line,
2666                     unsigned ws_rule,
2667                     int match_beginning, int match_end)
2668 {
2669         int i;
2670         unsigned long backwards, forwards, current;
2671         int backwards_lno, forwards_lno, current_lno;
2672
2673         /*
2674          * If match_beginning or match_end is specified, there is no
2675          * point starting from a wrong line that will never match and
2676          * wander around and wait for a match at the specified end.
2677          */
2678         if (match_beginning)
2679                 line = 0;
2680         else if (match_end)
2681                 line = img->nr - preimage->nr;
2682
2683         /*
2684          * Because the comparison is unsigned, the following test
2685          * will also take care of a negative line number that can
2686          * result when match_end and preimage is larger than the target.
2687          */
2688         if ((size_t) line > img->nr)
2689                 line = img->nr;
2690
2691         current = 0;
2692         for (i = 0; i < line; i++)
2693                 current += img->line[i].len;
2694
2695         /*
2696          * There's probably some smart way to do this, but I'll leave
2697          * that to the smart and beautiful people. I'm simple and stupid.
2698          */
2699         backwards = current;
2700         backwards_lno = line;
2701         forwards = current;
2702         forwards_lno = line;
2703         current_lno = line;
2704
2705         for (i = 0; ; i++) {
2706                 if (match_fragment(state, img, preimage, postimage,
2707                                    current, current_lno, ws_rule,
2708                                    match_beginning, match_end))
2709                         return current_lno;
2710
2711         again:
2712                 if (backwards_lno == 0 && forwards_lno == img->nr)
2713                         break;
2714
2715                 if (i & 1) {
2716                         if (backwards_lno == 0) {
2717                                 i++;
2718                                 goto again;
2719                         }
2720                         backwards_lno--;
2721                         backwards -= img->line[backwards_lno].len;
2722                         current = backwards;
2723                         current_lno = backwards_lno;
2724                 } else {
2725                         if (forwards_lno == img->nr) {
2726                                 i++;
2727                                 goto again;
2728                         }
2729                         forwards += img->line[forwards_lno].len;
2730                         forwards_lno++;
2731                         current = forwards;
2732                         current_lno = forwards_lno;
2733                 }
2734
2735         }
2736         return -1;
2737 }
2738
2739 static void remove_first_line(struct image *img)
2740 {
2741         img->buf += img->line[0].len;
2742         img->len -= img->line[0].len;
2743         img->line++;
2744         img->nr--;
2745 }
2746
2747 static void remove_last_line(struct image *img)
2748 {
2749         img->len -= img->line[--img->nr].len;
2750 }
2751
2752 /*
2753  * The change from "preimage" and "postimage" has been found to
2754  * apply at applied_pos (counts in line numbers) in "img".
2755  * Update "img" to remove "preimage" and replace it with "postimage".
2756  */
2757 static void update_image(struct apply_state *state,
2758                          struct image *img,
2759                          int applied_pos,
2760                          struct image *preimage,
2761                          struct image *postimage)
2762 {
2763         /*
2764          * remove the copy of preimage at offset in img
2765          * and replace it with postimage
2766          */
2767         int i, nr;
2768         size_t remove_count, insert_count, applied_at = 0;
2769         char *result;
2770         int preimage_limit;
2771
2772         /*
2773          * If we are removing blank lines at the end of img,
2774          * the preimage may extend beyond the end.
2775          * If that is the case, we must be careful only to
2776          * remove the part of the preimage that falls within
2777          * the boundaries of img. Initialize preimage_limit
2778          * to the number of lines in the preimage that falls
2779          * within the boundaries.
2780          */
2781         preimage_limit = preimage->nr;
2782         if (preimage_limit > img->nr - applied_pos)
2783                 preimage_limit = img->nr - applied_pos;
2784
2785         for (i = 0; i < applied_pos; i++)
2786                 applied_at += img->line[i].len;
2787
2788         remove_count = 0;
2789         for (i = 0; i < preimage_limit; i++)
2790                 remove_count += img->line[applied_pos + i].len;
2791         insert_count = postimage->len;
2792
2793         /* Adjust the contents */
2794         result = xmalloc(st_add3(st_sub(img->len, remove_count), insert_count, 1));
2795         memcpy(result, img->buf, applied_at);
2796         memcpy(result + applied_at, postimage->buf, postimage->len);
2797         memcpy(result + applied_at + postimage->len,
2798                img->buf + (applied_at + remove_count),
2799                img->len - (applied_at + remove_count));
2800         free(img->buf);
2801         img->buf = result;
2802         img->len += insert_count - remove_count;
2803         result[img->len] = '\0';
2804
2805         /* Adjust the line table */
2806         nr = img->nr + postimage->nr - preimage_limit;
2807         if (preimage_limit < postimage->nr) {
2808                 /*
2809                  * NOTE: this knows that we never call remove_first_line()
2810                  * on anything other than pre/post image.
2811                  */
2812                 REALLOC_ARRAY(img->line, nr);
2813                 img->line_allocated = img->line;
2814         }
2815         if (preimage_limit != postimage->nr)
2816                 MOVE_ARRAY(img->line + applied_pos + postimage->nr,
2817                            img->line + applied_pos + preimage_limit,
2818                            img->nr - (applied_pos + preimage_limit));
2819         COPY_ARRAY(img->line + applied_pos, postimage->line, postimage->nr);
2820         if (!state->allow_overlap)
2821                 for (i = 0; i < postimage->nr; i++)
2822                         img->line[applied_pos + i].flag |= LINE_PATCHED;
2823         img->nr = nr;
2824 }
2825
2826 /*
2827  * Use the patch-hunk text in "frag" to prepare two images (preimage and
2828  * postimage) for the hunk.  Find lines that match "preimage" in "img" and
2829  * replace the part of "img" with "postimage" text.
2830  */
2831 static int apply_one_fragment(struct apply_state *state,
2832                               struct image *img, struct fragment *frag,
2833                               int inaccurate_eof, unsigned ws_rule,
2834                               int nth_fragment)
2835 {
2836         int match_beginning, match_end;
2837         const char *patch = frag->patch;
2838         int size = frag->size;
2839         char *old, *oldlines;
2840         struct strbuf newlines;
2841         int new_blank_lines_at_end = 0;
2842         int found_new_blank_lines_at_end = 0;
2843         int hunk_linenr = frag->linenr;
2844         unsigned long leading, trailing;
2845         int pos, applied_pos;
2846         struct image preimage;
2847         struct image postimage;
2848
2849         memset(&preimage, 0, sizeof(preimage));
2850         memset(&postimage, 0, sizeof(postimage));
2851         oldlines = xmalloc(size);
2852         strbuf_init(&newlines, size);
2853
2854         old = oldlines;
2855         while (size > 0) {
2856                 char first;
2857                 int len = linelen(patch, size);
2858                 int plen;
2859                 int added_blank_line = 0;
2860                 int is_blank_context = 0;
2861                 size_t start;
2862
2863                 if (!len)
2864                         break;
2865
2866                 /*
2867                  * "plen" is how much of the line we should use for
2868                  * the actual patch data. Normally we just remove the
2869                  * first character on the line, but if the line is
2870                  * followed by "\ No newline", then we also remove the
2871                  * last one (which is the newline, of course).
2872                  */
2873                 plen = len - 1;
2874                 if (len < size && patch[len] == '\\')
2875                         plen--;
2876                 first = *patch;
2877                 if (state->apply_in_reverse) {
2878                         if (first == '-')
2879                                 first = '+';
2880                         else if (first == '+')
2881                                 first = '-';
2882                 }
2883
2884                 switch (first) {
2885                 case '\n':
2886                         /* Newer GNU diff, empty context line */
2887                         if (plen < 0)
2888                                 /* ... followed by '\No newline'; nothing */
2889                                 break;
2890                         *old++ = '\n';
2891                         strbuf_addch(&newlines, '\n');
2892                         add_line_info(&preimage, "\n", 1, LINE_COMMON);
2893                         add_line_info(&postimage, "\n", 1, LINE_COMMON);
2894                         is_blank_context = 1;
2895                         break;
2896                 case ' ':
2897                         if (plen && (ws_rule & WS_BLANK_AT_EOF) &&
2898                             ws_blank_line(patch + 1, plen, ws_rule))
2899                                 is_blank_context = 1;
2900                         /* fallthrough */
2901                 case '-':
2902                         memcpy(old, patch + 1, plen);
2903                         add_line_info(&preimage, old, plen,
2904                                       (first == ' ' ? LINE_COMMON : 0));
2905                         old += plen;
2906                         if (first == '-')
2907                                 break;
2908                         /* fallthrough */
2909                 case '+':
2910                         /* --no-add does not add new lines */
2911                         if (first == '+' && state->no_add)
2912                                 break;
2913
2914                         start = newlines.len;
2915                         if (first != '+' ||
2916                             !state->whitespace_error ||
2917                             state->ws_error_action != correct_ws_error) {
2918                                 strbuf_add(&newlines, patch + 1, plen);
2919                         }
2920                         else {
2921                                 ws_fix_copy(&newlines, patch + 1, plen, ws_rule, &state->applied_after_fixing_ws);
2922                         }
2923                         add_line_info(&postimage, newlines.buf + start, newlines.len - start,
2924                                       (first == '+' ? 0 : LINE_COMMON));
2925                         if (first == '+' &&
2926                             (ws_rule & WS_BLANK_AT_EOF) &&
2927                             ws_blank_line(patch + 1, plen, ws_rule))
2928                                 added_blank_line = 1;
2929                         break;
2930                 case '@': case '\\':
2931                         /* Ignore it, we already handled it */
2932                         break;
2933                 default:
2934                         if (state->apply_verbosity > verbosity_normal)
2935                                 error(_("invalid start of line: '%c'"), first);
2936                         applied_pos = -1;
2937                         goto out;
2938                 }
2939                 if (added_blank_line) {
2940                         if (!new_blank_lines_at_end)
2941                                 found_new_blank_lines_at_end = hunk_linenr;
2942                         new_blank_lines_at_end++;
2943                 }
2944                 else if (is_blank_context)
2945                         ;
2946                 else
2947                         new_blank_lines_at_end = 0;
2948                 patch += len;
2949                 size -= len;
2950                 hunk_linenr++;
2951         }
2952         if (inaccurate_eof &&
2953             old > oldlines && old[-1] == '\n' &&
2954             newlines.len > 0 && newlines.buf[newlines.len - 1] == '\n') {
2955                 old--;
2956                 strbuf_setlen(&newlines, newlines.len - 1);
2957                 preimage.line_allocated[preimage.nr - 1].len--;
2958                 postimage.line_allocated[postimage.nr - 1].len--;
2959         }
2960
2961         leading = frag->leading;
2962         trailing = frag->trailing;
2963
2964         /*
2965          * A hunk to change lines at the beginning would begin with
2966          * @@ -1,L +N,M @@
2967          * but we need to be careful.  -U0 that inserts before the second
2968          * line also has this pattern.
2969          *
2970          * And a hunk to add to an empty file would begin with
2971          * @@ -0,0 +N,M @@
2972          *
2973          * In other words, a hunk that is (frag->oldpos <= 1) with or
2974          * without leading context must match at the beginning.
2975          */
2976         match_beginning = (!frag->oldpos ||
2977                            (frag->oldpos == 1 && !state->unidiff_zero));
2978
2979         /*
2980          * A hunk without trailing lines must match at the end.
2981          * However, we simply cannot tell if a hunk must match end
2982          * from the lack of trailing lines if the patch was generated
2983          * with unidiff without any context.
2984          */
2985         match_end = !state->unidiff_zero && !trailing;
2986
2987         pos = frag->newpos ? (frag->newpos - 1) : 0;
2988         preimage.buf = oldlines;
2989         preimage.len = old - oldlines;
2990         postimage.buf = newlines.buf;
2991         postimage.len = newlines.len;
2992         preimage.line = preimage.line_allocated;
2993         postimage.line = postimage.line_allocated;
2994
2995         for (;;) {
2996
2997                 applied_pos = find_pos(state, img, &preimage, &postimage, pos,
2998                                        ws_rule, match_beginning, match_end);
2999
3000                 if (applied_pos >= 0)
3001                         break;
3002
3003                 /* Am I at my context limits? */
3004                 if ((leading <= state->p_context) && (trailing <= state->p_context))
3005                         break;
3006                 if (match_beginning || match_end) {
3007                         match_beginning = match_end = 0;
3008                         continue;
3009                 }
3010
3011                 /*
3012                  * Reduce the number of context lines; reduce both
3013                  * leading and trailing if they are equal otherwise
3014                  * just reduce the larger context.
3015                  */
3016                 if (leading >= trailing) {
3017                         remove_first_line(&preimage);
3018                         remove_first_line(&postimage);
3019                         pos--;
3020                         leading--;
3021                 }
3022                 if (trailing > leading) {
3023                         remove_last_line(&preimage);
3024                         remove_last_line(&postimage);
3025                         trailing--;
3026                 }
3027         }
3028
3029         if (applied_pos >= 0) {
3030                 if (new_blank_lines_at_end &&
3031                     preimage.nr + applied_pos >= img->nr &&
3032                     (ws_rule & WS_BLANK_AT_EOF) &&
3033                     state->ws_error_action != nowarn_ws_error) {
3034                         record_ws_error(state, WS_BLANK_AT_EOF, "+", 1,
3035                                         found_new_blank_lines_at_end);
3036                         if (state->ws_error_action == correct_ws_error) {
3037                                 while (new_blank_lines_at_end--)
3038                                         remove_last_line(&postimage);
3039                         }
3040                         /*
3041                          * We would want to prevent write_out_results()
3042                          * from taking place in apply_patch() that follows
3043                          * the callchain led us here, which is:
3044                          * apply_patch->check_patch_list->check_patch->
3045                          * apply_data->apply_fragments->apply_one_fragment
3046                          */
3047                         if (state->ws_error_action == die_on_ws_error)
3048                                 state->apply = 0;
3049                 }
3050
3051                 if (state->apply_verbosity > verbosity_normal && applied_pos != pos) {
3052                         int offset = applied_pos - pos;
3053                         if (state->apply_in_reverse)
3054                                 offset = 0 - offset;
3055                         fprintf_ln(stderr,
3056                                    Q_("Hunk #%d succeeded at %d (offset %d line).",
3057                                       "Hunk #%d succeeded at %d (offset %d lines).",
3058                                       offset),
3059                                    nth_fragment, applied_pos + 1, offset);
3060                 }
3061
3062                 /*
3063                  * Warn if it was necessary to reduce the number
3064                  * of context lines.
3065                  */
3066                 if ((leading != frag->leading ||
3067                      trailing != frag->trailing) && state->apply_verbosity > verbosity_silent)
3068                         fprintf_ln(stderr, _("Context reduced to (%ld/%ld)"
3069                                              " to apply fragment at %d"),
3070                                    leading, trailing, applied_pos+1);
3071                 update_image(state, img, applied_pos, &preimage, &postimage);
3072         } else {
3073                 if (state->apply_verbosity > verbosity_normal)
3074                         error(_("while searching for:\n%.*s"),
3075                               (int)(old - oldlines), oldlines);
3076         }
3077
3078 out:
3079         free(oldlines);
3080         strbuf_release(&newlines);
3081         free(preimage.line_allocated);
3082         free(postimage.line_allocated);
3083
3084         return (applied_pos < 0);
3085 }
3086
3087 static int apply_binary_fragment(struct apply_state *state,
3088                                  struct image *img,
3089                                  struct patch *patch)
3090 {
3091         struct fragment *fragment = patch->fragments;
3092         unsigned long len;
3093         void *dst;
3094
3095         if (!fragment)
3096                 return error(_("missing binary patch data for '%s'"),
3097                              patch->new_name ?
3098                              patch->new_name :
3099                              patch->old_name);
3100
3101         /* Binary patch is irreversible without the optional second hunk */
3102         if (state->apply_in_reverse) {
3103                 if (!fragment->next)
3104                         return error(_("cannot reverse-apply a binary patch "
3105                                        "without the reverse hunk to '%s'"),
3106                                      patch->new_name
3107                                      ? patch->new_name : patch->old_name);
3108                 fragment = fragment->next;
3109         }
3110         switch (fragment->binary_patch_method) {
3111         case BINARY_DELTA_DEFLATED:
3112                 dst = patch_delta(img->buf, img->len, fragment->patch,
3113                                   fragment->size, &len);
3114                 if (!dst)
3115                         return -1;
3116                 clear_image(img);
3117                 img->buf = dst;
3118                 img->len = len;
3119                 return 0;
3120         case BINARY_LITERAL_DEFLATED:
3121                 clear_image(img);
3122                 img->len = fragment->size;
3123                 img->buf = xmemdupz(fragment->patch, img->len);
3124                 return 0;
3125         }
3126         return -1;
3127 }
3128
3129 /*
3130  * Replace "img" with the result of applying the binary patch.
3131  * The binary patch data itself in patch->fragment is still kept
3132  * but the preimage prepared by the caller in "img" is freed here
3133  * or in the helper function apply_binary_fragment() this calls.
3134  */
3135 static int apply_binary(struct apply_state *state,
3136                         struct image *img,
3137                         struct patch *patch)
3138 {
3139         const char *name = patch->old_name ? patch->old_name : patch->new_name;
3140         struct object_id oid;
3141
3142         /*
3143          * For safety, we require patch index line to contain
3144          * full 40-byte textual SHA1 for old and new, at least for now.
3145          */
3146         if (strlen(patch->old_sha1_prefix) != 40 ||
3147             strlen(patch->new_sha1_prefix) != 40 ||
3148             get_oid_hex(patch->old_sha1_prefix, &oid) ||
3149             get_oid_hex(patch->new_sha1_prefix, &oid))
3150                 return error(_("cannot apply binary patch to '%s' "
3151                                "without full index line"), name);
3152
3153         if (patch->old_name) {
3154                 /*
3155                  * See if the old one matches what the patch
3156                  * applies to.
3157                  */
3158                 hash_object_file(img->buf, img->len, blob_type, &oid);
3159                 if (strcmp(oid_to_hex(&oid), patch->old_sha1_prefix))
3160                         return error(_("the patch applies to '%s' (%s), "
3161                                        "which does not match the "
3162                                        "current contents."),
3163                                      name, oid_to_hex(&oid));
3164         }
3165         else {
3166                 /* Otherwise, the old one must be empty. */
3167                 if (img->len)
3168                         return error(_("the patch applies to an empty "
3169                                        "'%s' but it is not empty"), name);
3170         }
3171
3172         get_oid_hex(patch->new_sha1_prefix, &oid);
3173         if (is_null_oid(&oid)) {
3174                 clear_image(img);
3175                 return 0; /* deletion patch */
3176         }
3177
3178         if (has_sha1_file(oid.hash)) {
3179                 /* We already have the postimage */
3180                 enum object_type type;
3181                 unsigned long size;
3182                 char *result;
3183
3184                 result = read_object_file(&oid, &type, &size);
3185                 if (!result)
3186                         return error(_("the necessary postimage %s for "
3187                                        "'%s' cannot be read"),
3188                                      patch->new_sha1_prefix, name);
3189                 clear_image(img);
3190                 img->buf = result;
3191                 img->len = size;
3192         } else {
3193                 /*
3194                  * We have verified buf matches the preimage;
3195                  * apply the patch data to it, which is stored
3196                  * in the patch->fragments->{patch,size}.
3197                  */
3198                 if (apply_binary_fragment(state, img, patch))
3199                         return error(_("binary patch does not apply to '%s'"),
3200                                      name);
3201
3202                 /* verify that the result matches */
3203                 hash_object_file(img->buf, img->len, blob_type, &oid);
3204                 if (strcmp(oid_to_hex(&oid), patch->new_sha1_prefix))
3205                         return error(_("binary patch to '%s' creates incorrect result (expecting %s, got %s)"),
3206                                 name, patch->new_sha1_prefix, oid_to_hex(&oid));
3207         }
3208
3209         return 0;
3210 }
3211
3212 static int apply_fragments(struct apply_state *state, struct image *img, struct patch *patch)
3213 {
3214         struct fragment *frag = patch->fragments;
3215         const char *name = patch->old_name ? patch->old_name : patch->new_name;
3216         unsigned ws_rule = patch->ws_rule;
3217         unsigned inaccurate_eof = patch->inaccurate_eof;
3218         int nth = 0;
3219
3220         if (patch->is_binary)
3221                 return apply_binary(state, img, patch);
3222
3223         while (frag) {
3224                 nth++;
3225                 if (apply_one_fragment(state, img, frag, inaccurate_eof, ws_rule, nth)) {
3226                         error(_("patch failed: %s:%ld"), name, frag->oldpos);
3227                         if (!state->apply_with_reject)
3228                                 return -1;
3229                         frag->rejected = 1;
3230                 }
3231                 frag = frag->next;
3232         }
3233         return 0;
3234 }
3235
3236 static int read_blob_object(struct strbuf *buf, const struct object_id *oid, unsigned mode)
3237 {
3238         if (S_ISGITLINK(mode)) {
3239                 strbuf_grow(buf, 100);
3240                 strbuf_addf(buf, "Subproject commit %s\n", oid_to_hex(oid));
3241         } else {
3242                 enum object_type type;
3243                 unsigned long sz;
3244                 char *result;
3245
3246                 result = read_object_file(oid, &type, &sz);
3247                 if (!result)
3248                         return -1;
3249                 /* XXX read_sha1_file NUL-terminates */
3250                 strbuf_attach(buf, result, sz, sz + 1);
3251         }
3252         return 0;
3253 }
3254
3255 static int read_file_or_gitlink(const struct cache_entry *ce, struct strbuf *buf)
3256 {
3257         if (!ce)
3258                 return 0;
3259         return read_blob_object(buf, &ce->oid, ce->ce_mode);
3260 }
3261
3262 static struct patch *in_fn_table(struct apply_state *state, const char *name)
3263 {
3264         struct string_list_item *item;
3265
3266         if (name == NULL)
3267                 return NULL;
3268
3269         item = string_list_lookup(&state->fn_table, name);
3270         if (item != NULL)
3271                 return (struct patch *)item->util;
3272
3273         return NULL;
3274 }
3275
3276 /*
3277  * item->util in the filename table records the status of the path.
3278  * Usually it points at a patch (whose result records the contents
3279  * of it after applying it), but it could be PATH_WAS_DELETED for a
3280  * path that a previously applied patch has already removed, or
3281  * PATH_TO_BE_DELETED for a path that a later patch would remove.
3282  *
3283  * The latter is needed to deal with a case where two paths A and B
3284  * are swapped by first renaming A to B and then renaming B to A;
3285  * moving A to B should not be prevented due to presence of B as we
3286  * will remove it in a later patch.
3287  */
3288 #define PATH_TO_BE_DELETED ((struct patch *) -2)
3289 #define PATH_WAS_DELETED ((struct patch *) -1)
3290
3291 static int to_be_deleted(struct patch *patch)
3292 {
3293         return patch == PATH_TO_BE_DELETED;
3294 }
3295
3296 static int was_deleted(struct patch *patch)
3297 {
3298         return patch == PATH_WAS_DELETED;
3299 }
3300
3301 static void add_to_fn_table(struct apply_state *state, struct patch *patch)
3302 {
3303         struct string_list_item *item;
3304
3305         /*
3306          * Always add new_name unless patch is a deletion
3307          * This should cover the cases for normal diffs,
3308          * file creations and copies
3309          */
3310         if (patch->new_name != NULL) {
3311                 item = string_list_insert(&state->fn_table, patch->new_name);
3312                 item->util = patch;
3313         }
3314
3315         /*
3316          * store a failure on rename/deletion cases because
3317          * later chunks shouldn't patch old names
3318          */
3319         if ((patch->new_name == NULL) || (patch->is_rename)) {
3320                 item = string_list_insert(&state->fn_table, patch->old_name);
3321                 item->util = PATH_WAS_DELETED;
3322         }
3323 }
3324
3325 static void prepare_fn_table(struct apply_state *state, struct patch *patch)
3326 {
3327         /*
3328          * store information about incoming file deletion
3329          */
3330         while (patch) {
3331                 if ((patch->new_name == NULL) || (patch->is_rename)) {
3332                         struct string_list_item *item;
3333                         item = string_list_insert(&state->fn_table, patch->old_name);
3334                         item->util = PATH_TO_BE_DELETED;
3335                 }
3336                 patch = patch->next;
3337         }
3338 }
3339
3340 static int checkout_target(struct index_state *istate,
3341                            struct cache_entry *ce, struct stat *st)
3342 {
3343         struct checkout costate = CHECKOUT_INIT;
3344
3345         costate.refresh_cache = 1;
3346         costate.istate = istate;
3347         if (checkout_entry(ce, &costate, NULL) || lstat(ce->name, st))
3348                 return error(_("cannot checkout %s"), ce->name);
3349         return 0;
3350 }
3351
3352 static struct patch *previous_patch(struct apply_state *state,
3353                                     struct patch *patch,
3354                                     int *gone)
3355 {
3356         struct patch *previous;
3357
3358         *gone = 0;
3359         if (patch->is_copy || patch->is_rename)
3360                 return NULL; /* "git" patches do not depend on the order */
3361
3362         previous = in_fn_table(state, patch->old_name);
3363         if (!previous)
3364                 return NULL;
3365
3366         if (to_be_deleted(previous))
3367                 return NULL; /* the deletion hasn't happened yet */
3368
3369         if (was_deleted(previous))
3370                 *gone = 1;
3371
3372         return previous;
3373 }
3374
3375 static int verify_index_match(const struct cache_entry *ce, struct stat *st)
3376 {
3377         if (S_ISGITLINK(ce->ce_mode)) {
3378                 if (!S_ISDIR(st->st_mode))
3379                         return -1;
3380                 return 0;
3381         }
3382         return ce_match_stat(ce, st, CE_MATCH_IGNORE_VALID|CE_MATCH_IGNORE_SKIP_WORKTREE);
3383 }
3384
3385 #define SUBMODULE_PATCH_WITHOUT_INDEX 1
3386
3387 static int load_patch_target(struct apply_state *state,
3388                              struct strbuf *buf,
3389                              const struct cache_entry *ce,
3390                              struct stat *st,
3391                              struct patch *patch,
3392                              const char *name,
3393                              unsigned expected_mode)
3394 {
3395         if (state->cached || state->check_index) {
3396                 if (read_file_or_gitlink(ce, buf))
3397                         return error(_("failed to read %s"), name);
3398         } else if (name) {
3399                 if (S_ISGITLINK(expected_mode)) {
3400                         if (ce)
3401                                 return read_file_or_gitlink(ce, buf);
3402                         else
3403                                 return SUBMODULE_PATCH_WITHOUT_INDEX;
3404                 } else if (has_symlink_leading_path(name, strlen(name))) {
3405                         return error(_("reading from '%s' beyond a symbolic link"), name);
3406                 } else {
3407                         if (read_old_data(st, patch, name, buf))
3408                                 return error(_("failed to read %s"), name);
3409                 }
3410         }
3411         return 0;
3412 }
3413
3414 /*
3415  * We are about to apply "patch"; populate the "image" with the
3416  * current version we have, from the working tree or from the index,
3417  * depending on the situation e.g. --cached/--index.  If we are
3418  * applying a non-git patch that incrementally updates the tree,
3419  * we read from the result of a previous diff.
3420  */
3421 static int load_preimage(struct apply_state *state,
3422                          struct image *image,
3423                          struct patch *patch, struct stat *st,
3424                          const struct cache_entry *ce)
3425 {
3426         struct strbuf buf = STRBUF_INIT;
3427         size_t len;
3428         char *img;
3429         struct patch *previous;
3430         int status;
3431
3432         previous = previous_patch(state, patch, &status);
3433         if (status)
3434                 return error(_("path %s has been renamed/deleted"),
3435                              patch->old_name);
3436         if (previous) {
3437                 /* We have a patched copy in memory; use that. */
3438                 strbuf_add(&buf, previous->result, previous->resultsize);
3439         } else {
3440                 status = load_patch_target(state, &buf, ce, st, patch,
3441                                            patch->old_name, patch->old_mode);
3442                 if (status < 0)
3443                         return status;
3444                 else if (status == SUBMODULE_PATCH_WITHOUT_INDEX) {
3445                         /*
3446                          * There is no way to apply subproject
3447                          * patch without looking at the index.
3448                          * NEEDSWORK: shouldn't this be flagged
3449                          * as an error???
3450                          */
3451                         free_fragment_list(patch->fragments);
3452                         patch->fragments = NULL;
3453                 } else if (status) {
3454                         return error(_("failed to read %s"), patch->old_name);
3455                 }
3456         }
3457
3458         img = strbuf_detach(&buf, &len);
3459         prepare_image(image, img, len, !patch->is_binary);
3460         return 0;
3461 }
3462
3463 static int three_way_merge(struct image *image,
3464                            char *path,
3465                            const struct object_id *base,
3466                            const struct object_id *ours,
3467                            const struct object_id *theirs)
3468 {
3469         mmfile_t base_file, our_file, their_file;
3470         mmbuffer_t result = { NULL };
3471         int status;
3472
3473         read_mmblob(&base_file, base);
3474         read_mmblob(&our_file, ours);
3475         read_mmblob(&their_file, theirs);
3476         status = ll_merge(&result, path,
3477                           &base_file, "base",
3478                           &our_file, "ours",
3479                           &their_file, "theirs", NULL);
3480         free(base_file.ptr);
3481         free(our_file.ptr);
3482         free(their_file.ptr);
3483         if (status < 0 || !result.ptr) {
3484                 free(result.ptr);
3485                 return -1;
3486         }
3487         clear_image(image);
3488         image->buf = result.ptr;
3489         image->len = result.size;
3490
3491         return status;
3492 }
3493
3494 /*
3495  * When directly falling back to add/add three-way merge, we read from
3496  * the current contents of the new_name.  In no cases other than that
3497  * this function will be called.
3498  */
3499 static int load_current(struct apply_state *state,
3500                         struct image *image,
3501                         struct patch *patch)
3502 {
3503         struct strbuf buf = STRBUF_INIT;
3504         int status, pos;
3505         size_t len;
3506         char *img;
3507         struct stat st;
3508         struct cache_entry *ce;
3509         char *name = patch->new_name;
3510         unsigned mode = patch->new_mode;
3511
3512         if (!patch->is_new)
3513                 die("BUG: patch to %s is not a creation", patch->old_name);
3514
3515         pos = cache_name_pos(name, strlen(name));
3516         if (pos < 0)
3517                 return error(_("%s: does not exist in index"), name);
3518         ce = active_cache[pos];
3519         if (lstat(name, &st)) {
3520                 if (errno != ENOENT)
3521                         return error_errno("%s", name);
3522                 if (checkout_target(&the_index, ce, &st))
3523                         return -1;
3524         }
3525         if (verify_index_match(ce, &st))
3526                 return error(_("%s: does not match index"), name);
3527
3528         status = load_patch_target(state, &buf, ce, &st, patch, name, mode);
3529         if (status < 0)
3530                 return status;
3531         else if (status)
3532                 return -1;
3533         img = strbuf_detach(&buf, &len);
3534         prepare_image(image, img, len, !patch->is_binary);
3535         return 0;
3536 }
3537
3538 static int try_threeway(struct apply_state *state,
3539                         struct image *image,
3540                         struct patch *patch,
3541                         struct stat *st,
3542                         const struct cache_entry *ce)
3543 {
3544         struct object_id pre_oid, post_oid, our_oid;
3545         struct strbuf buf = STRBUF_INIT;
3546         size_t len;
3547         int status;
3548         char *img;
3549         struct image tmp_image;
3550
3551         /* No point falling back to 3-way merge in these cases */
3552         if (patch->is_delete ||
3553             S_ISGITLINK(patch->old_mode) || S_ISGITLINK(patch->new_mode))
3554                 return -1;
3555
3556         /* Preimage the patch was prepared for */
3557         if (patch->is_new)
3558                 write_object_file("", 0, blob_type, &pre_oid);
3559         else if (get_oid(patch->old_sha1_prefix, &pre_oid) ||
3560                  read_blob_object(&buf, &pre_oid, patch->old_mode))
3561                 return error(_("repository lacks the necessary blob to fall back on 3-way merge."));
3562
3563         if (state->apply_verbosity > verbosity_silent)
3564                 fprintf(stderr, _("Falling back to three-way merge...\n"));
3565
3566         img = strbuf_detach(&buf, &len);
3567         prepare_image(&tmp_image, img, len, 1);
3568         /* Apply the patch to get the post image */
3569         if (apply_fragments(state, &tmp_image, patch) < 0) {
3570                 clear_image(&tmp_image);
3571                 return -1;
3572         }
3573         /* post_oid is theirs */
3574         write_object_file(tmp_image.buf, tmp_image.len, blob_type, &post_oid);
3575         clear_image(&tmp_image);
3576
3577         /* our_oid is ours */
3578         if (patch->is_new) {
3579                 if (load_current(state, &tmp_image, patch))
3580                         return error(_("cannot read the current contents of '%s'"),
3581                                      patch->new_name);
3582         } else {
3583                 if (load_preimage(state, &tmp_image, patch, st, ce))
3584                         return error(_("cannot read the current contents of '%s'"),
3585                                      patch->old_name);
3586         }
3587         write_object_file(tmp_image.buf, tmp_image.len, blob_type, &our_oid);
3588         clear_image(&tmp_image);
3589
3590         /* in-core three-way merge between post and our using pre as base */
3591         status = three_way_merge(image, patch->new_name,
3592                                  &pre_oid, &our_oid, &post_oid);
3593         if (status < 0) {
3594                 if (state->apply_verbosity > verbosity_silent)
3595                         fprintf(stderr,
3596                                 _("Failed to fall back on three-way merge...\n"));
3597                 return status;
3598         }
3599
3600         if (status) {
3601                 patch->conflicted_threeway = 1;
3602                 if (patch->is_new)
3603                         oidclr(&patch->threeway_stage[0]);
3604                 else
3605                         oidcpy(&patch->threeway_stage[0], &pre_oid);
3606                 oidcpy(&patch->threeway_stage[1], &our_oid);
3607                 oidcpy(&patch->threeway_stage[2], &post_oid);
3608                 if (state->apply_verbosity > verbosity_silent)
3609                         fprintf(stderr,
3610                                 _("Applied patch to '%s' with conflicts.\n"),
3611                                 patch->new_name);
3612         } else {
3613                 if (state->apply_verbosity > verbosity_silent)
3614                         fprintf(stderr,
3615                                 _("Applied patch to '%s' cleanly.\n"),
3616                                 patch->new_name);
3617         }
3618         return 0;
3619 }
3620
3621 static int apply_data(struct apply_state *state, struct patch *patch,
3622                       struct stat *st, const struct cache_entry *ce)
3623 {
3624         struct image image;
3625
3626         if (load_preimage(state, &image, patch, st, ce) < 0)
3627                 return -1;
3628
3629         if (patch->direct_to_threeway ||
3630             apply_fragments(state, &image, patch) < 0) {
3631                 /* Note: with --reject, apply_fragments() returns 0 */
3632                 if (!state->threeway || try_threeway(state, &image, patch, st, ce) < 0)
3633                         return -1;
3634         }
3635         patch->result = image.buf;
3636         patch->resultsize = image.len;
3637         add_to_fn_table(state, patch);
3638         free(image.line_allocated);
3639
3640         if (0 < patch->is_delete && patch->resultsize)
3641                 return error(_("removal patch leaves file contents"));
3642
3643         return 0;
3644 }
3645
3646 /*
3647  * If "patch" that we are looking at modifies or deletes what we have,
3648  * we would want it not to lose any local modification we have, either
3649  * in the working tree or in the index.
3650  *
3651  * This also decides if a non-git patch is a creation patch or a
3652  * modification to an existing empty file.  We do not check the state
3653  * of the current tree for a creation patch in this function; the caller
3654  * check_patch() separately makes sure (and errors out otherwise) that
3655  * the path the patch creates does not exist in the current tree.
3656  */
3657 static int check_preimage(struct apply_state *state,
3658                           struct patch *patch,
3659                           struct cache_entry **ce,
3660                           struct stat *st)
3661 {
3662         const char *old_name = patch->old_name;
3663         struct patch *previous = NULL;
3664         int stat_ret = 0, status;
3665         unsigned st_mode = 0;
3666
3667         if (!old_name)
3668                 return 0;
3669
3670         assert(patch->is_new <= 0);
3671         previous = previous_patch(state, patch, &status);
3672
3673         if (status)
3674                 return error(_("path %s has been renamed/deleted"), old_name);
3675         if (previous) {
3676                 st_mode = previous->new_mode;
3677         } else if (!state->cached) {
3678                 stat_ret = lstat(old_name, st);
3679                 if (stat_ret && errno != ENOENT)
3680                         return error_errno("%s", old_name);
3681         }
3682
3683         if (state->check_index && !previous) {
3684                 int pos = cache_name_pos(old_name, strlen(old_name));
3685                 if (pos < 0) {
3686                         if (patch->is_new < 0)
3687                                 goto is_new;
3688                         return error(_("%s: does not exist in index"), old_name);
3689                 }
3690                 *ce = active_cache[pos];
3691                 if (stat_ret < 0) {
3692                         if (checkout_target(&the_index, *ce, st))
3693                                 return -1;
3694                 }
3695                 if (!state->cached && verify_index_match(*ce, st))
3696                         return error(_("%s: does not match index"), old_name);
3697                 if (state->cached)
3698                         st_mode = (*ce)->ce_mode;
3699         } else if (stat_ret < 0) {
3700                 if (patch->is_new < 0)
3701                         goto is_new;
3702                 return error_errno("%s", old_name);
3703         }
3704
3705         if (!state->cached && !previous)
3706                 st_mode = ce_mode_from_stat(*ce, st->st_mode);
3707
3708         if (patch->is_new < 0)
3709                 patch->is_new = 0;
3710         if (!patch->old_mode)
3711                 patch->old_mode = st_mode;
3712         if ((st_mode ^ patch->old_mode) & S_IFMT)
3713                 return error(_("%s: wrong type"), old_name);
3714         if (st_mode != patch->old_mode)
3715                 warning(_("%s has type %o, expected %o"),
3716                         old_name, st_mode, patch->old_mode);
3717         if (!patch->new_mode && !patch->is_delete)
3718                 patch->new_mode = st_mode;
3719         return 0;
3720
3721  is_new:
3722         patch->is_new = 1;
3723         patch->is_delete = 0;
3724         FREE_AND_NULL(patch->old_name);
3725         return 0;
3726 }
3727
3728
3729 #define EXISTS_IN_INDEX 1
3730 #define EXISTS_IN_WORKTREE 2
3731
3732 static int check_to_create(struct apply_state *state,
3733                            const char *new_name,
3734                            int ok_if_exists)
3735 {
3736         struct stat nst;
3737
3738         if (state->check_index &&
3739             cache_name_pos(new_name, strlen(new_name)) >= 0 &&
3740             !ok_if_exists)
3741                 return EXISTS_IN_INDEX;
3742         if (state->cached)
3743                 return 0;
3744
3745         if (!lstat(new_name, &nst)) {
3746                 if (S_ISDIR(nst.st_mode) || ok_if_exists)
3747                         return 0;
3748                 /*
3749                  * A leading component of new_name might be a symlink
3750                  * that is going to be removed with this patch, but
3751                  * still pointing at somewhere that has the path.
3752                  * In such a case, path "new_name" does not exist as
3753                  * far as git is concerned.
3754                  */
3755                 if (has_symlink_leading_path(new_name, strlen(new_name)))
3756                         return 0;
3757
3758                 return EXISTS_IN_WORKTREE;
3759         } else if (!is_missing_file_error(errno)) {
3760                 return error_errno("%s", new_name);
3761         }
3762         return 0;
3763 }
3764
3765 static uintptr_t register_symlink_changes(struct apply_state *state,
3766                                           const char *path,
3767                                           uintptr_t what)
3768 {
3769         struct string_list_item *ent;
3770
3771         ent = string_list_lookup(&state->symlink_changes, path);
3772         if (!ent) {
3773                 ent = string_list_insert(&state->symlink_changes, path);
3774                 ent->util = (void *)0;
3775         }
3776         ent->util = (void *)(what | ((uintptr_t)ent->util));
3777         return (uintptr_t)ent->util;
3778 }
3779
3780 static uintptr_t check_symlink_changes(struct apply_state *state, const char *path)
3781 {
3782         struct string_list_item *ent;
3783
3784         ent = string_list_lookup(&state->symlink_changes, path);
3785         if (!ent)
3786                 return 0;
3787         return (uintptr_t)ent->util;
3788 }
3789
3790 static void prepare_symlink_changes(struct apply_state *state, struct patch *patch)
3791 {
3792         for ( ; patch; patch = patch->next) {
3793                 if ((patch->old_name && S_ISLNK(patch->old_mode)) &&
3794                     (patch->is_rename || patch->is_delete))
3795                         /* the symlink at patch->old_name is removed */
3796                         register_symlink_changes(state, patch->old_name, APPLY_SYMLINK_GOES_AWAY);
3797
3798                 if (patch->new_name && S_ISLNK(patch->new_mode))
3799                         /* the symlink at patch->new_name is created or remains */
3800                         register_symlink_changes(state, patch->new_name, APPLY_SYMLINK_IN_RESULT);
3801         }
3802 }
3803
3804 static int path_is_beyond_symlink_1(struct apply_state *state, struct strbuf *name)
3805 {
3806         do {
3807                 unsigned int change;
3808
3809                 while (--name->len && name->buf[name->len] != '/')
3810                         ; /* scan backwards */
3811                 if (!name->len)
3812                         break;
3813                 name->buf[name->len] = '\0';
3814                 change = check_symlink_changes(state, name->buf);
3815                 if (change & APPLY_SYMLINK_IN_RESULT)
3816                         return 1;
3817                 if (change & APPLY_SYMLINK_GOES_AWAY)
3818                         /*
3819                          * This cannot be "return 0", because we may
3820                          * see a new one created at a higher level.
3821                          */
3822                         continue;
3823
3824                 /* otherwise, check the preimage */
3825                 if (state->check_index) {
3826                         struct cache_entry *ce;
3827
3828                         ce = cache_file_exists(name->buf, name->len, ignore_case);
3829                         if (ce && S_ISLNK(ce->ce_mode))
3830                                 return 1;
3831                 } else {
3832                         struct stat st;
3833                         if (!lstat(name->buf, &st) && S_ISLNK(st.st_mode))
3834                                 return 1;
3835                 }
3836         } while (1);
3837         return 0;
3838 }
3839
3840 static int path_is_beyond_symlink(struct apply_state *state, const char *name_)
3841 {
3842         int ret;
3843         struct strbuf name = STRBUF_INIT;
3844
3845         assert(*name_ != '\0');
3846         strbuf_addstr(&name, name_);
3847         ret = path_is_beyond_symlink_1(state, &name);
3848         strbuf_release(&name);
3849
3850         return ret;
3851 }
3852
3853 static int check_unsafe_path(struct patch *patch)
3854 {
3855         const char *old_name = NULL;
3856         const char *new_name = NULL;
3857         if (patch->is_delete)
3858                 old_name = patch->old_name;
3859         else if (!patch->is_new && !patch->is_copy)
3860                 old_name = patch->old_name;
3861         if (!patch->is_delete)
3862                 new_name = patch->new_name;
3863
3864         if (old_name && !verify_path(old_name))
3865                 return error(_("invalid path '%s'"), old_name);
3866         if (new_name && !verify_path(new_name))
3867                 return error(_("invalid path '%s'"), new_name);
3868         return 0;
3869 }
3870
3871 /*
3872  * Check and apply the patch in-core; leave the result in patch->result
3873  * for the caller to write it out to the final destination.
3874  */
3875 static int check_patch(struct apply_state *state, struct patch *patch)
3876 {
3877         struct stat st;
3878         const char *old_name = patch->old_name;
3879         const char *new_name = patch->new_name;
3880         const char *name = old_name ? old_name : new_name;
3881         struct cache_entry *ce = NULL;
3882         struct patch *tpatch;
3883         int ok_if_exists;
3884         int status;
3885
3886         patch->rejected = 1; /* we will drop this after we succeed */
3887
3888         status = check_preimage(state, patch, &ce, &st);
3889         if (status)
3890                 return status;
3891         old_name = patch->old_name;
3892
3893         /*
3894          * A type-change diff is always split into a patch to delete
3895          * old, immediately followed by a patch to create new (see
3896          * diff.c::run_diff()); in such a case it is Ok that the entry
3897          * to be deleted by the previous patch is still in the working
3898          * tree and in the index.
3899          *
3900          * A patch to swap-rename between A and B would first rename A
3901          * to B and then rename B to A.  While applying the first one,
3902          * the presence of B should not stop A from getting renamed to
3903          * B; ask to_be_deleted() about the later rename.  Removal of
3904          * B and rename from A to B is handled the same way by asking
3905          * was_deleted().
3906          */
3907         if ((tpatch = in_fn_table(state, new_name)) &&
3908             (was_deleted(tpatch) || to_be_deleted(tpatch)))
3909                 ok_if_exists = 1;
3910         else
3911                 ok_if_exists = 0;
3912
3913         if (new_name &&
3914             ((0 < patch->is_new) || patch->is_rename || patch->is_copy)) {
3915                 int err = check_to_create(state, new_name, ok_if_exists);
3916
3917                 if (err && state->threeway) {
3918                         patch->direct_to_threeway = 1;
3919                 } else switch (err) {
3920                 case 0:
3921                         break; /* happy */
3922                 case EXISTS_IN_INDEX:
3923                         return error(_("%s: already exists in index"), new_name);
3924                         break;
3925                 case EXISTS_IN_WORKTREE:
3926                         return error(_("%s: already exists in working directory"),
3927                                      new_name);
3928                 default:
3929                         return err;
3930                 }
3931
3932                 if (!patch->new_mode) {
3933                         if (0 < patch->is_new)
3934                                 patch->new_mode = S_IFREG | 0644;
3935                         else
3936                                 patch->new_mode = patch->old_mode;
3937                 }
3938         }
3939
3940         if (new_name && old_name) {
3941                 int same = !strcmp(old_name, new_name);
3942                 if (!patch->new_mode)
3943                         patch->new_mode = patch->old_mode;
3944                 if ((patch->old_mode ^ patch->new_mode) & S_IFMT) {
3945                         if (same)
3946                                 return error(_("new mode (%o) of %s does not "
3947                                                "match old mode (%o)"),
3948                                         patch->new_mode, new_name,
3949                                         patch->old_mode);
3950                         else
3951                                 return error(_("new mode (%o) of %s does not "
3952                                                "match old mode (%o) of %s"),
3953                                         patch->new_mode, new_name,
3954                                         patch->old_mode, old_name);
3955                 }
3956         }
3957
3958         if (!state->unsafe_paths && check_unsafe_path(patch))
3959                 return -128;
3960
3961         /*
3962          * An attempt to read from or delete a path that is beyond a
3963          * symbolic link will be prevented by load_patch_target() that
3964          * is called at the beginning of apply_data() so we do not
3965          * have to worry about a patch marked with "is_delete" bit
3966          * here.  We however need to make sure that the patch result
3967          * is not deposited to a path that is beyond a symbolic link
3968          * here.
3969          */
3970         if (!patch->is_delete && path_is_beyond_symlink(state, patch->new_name))
3971                 return error(_("affected file '%s' is beyond a symbolic link"),
3972                              patch->new_name);
3973
3974         if (apply_data(state, patch, &st, ce) < 0)
3975                 return error(_("%s: patch does not apply"), name);
3976         patch->rejected = 0;
3977         return 0;
3978 }
3979
3980 static int check_patch_list(struct apply_state *state, struct patch *patch)
3981 {
3982         int err = 0;
3983
3984         prepare_symlink_changes(state, patch);
3985         prepare_fn_table(state, patch);
3986         while (patch) {
3987                 int res;
3988                 if (state->apply_verbosity > verbosity_normal)
3989                         say_patch_name(stderr,
3990                                        _("Checking patch %s..."), patch);
3991                 res = check_patch(state, patch);
3992                 if (res == -128)
3993                         return -128;
3994                 err |= res;
3995                 patch = patch->next;
3996         }
3997         return err;
3998 }
3999
4000 static int read_apply_cache(struct apply_state *state)
4001 {
4002         if (state->index_file)
4003                 return read_cache_from(state->index_file);
4004         else
4005                 return read_cache();
4006 }
4007
4008 /* This function tries to read the object name from the current index */
4009 static int get_current_oid(struct apply_state *state, const char *path,
4010                            struct object_id *oid)
4011 {
4012         int pos;
4013
4014         if (read_apply_cache(state) < 0)
4015                 return -1;
4016         pos = cache_name_pos(path, strlen(path));
4017         if (pos < 0)
4018                 return -1;
4019         oidcpy(oid, &active_cache[pos]->oid);
4020         return 0;
4021 }
4022
4023 static int preimage_oid_in_gitlink_patch(struct patch *p, struct object_id *oid)
4024 {
4025         /*
4026          * A usable gitlink patch has only one fragment (hunk) that looks like:
4027          * @@ -1 +1 @@
4028          * -Subproject commit <old sha1>
4029          * +Subproject commit <new sha1>
4030          * or
4031          * @@ -1 +0,0 @@
4032          * -Subproject commit <old sha1>
4033          * for a removal patch.
4034          */
4035         struct fragment *hunk = p->fragments;
4036         static const char heading[] = "-Subproject commit ";
4037         char *preimage;
4038
4039         if (/* does the patch have only one hunk? */
4040             hunk && !hunk->next &&
4041             /* is its preimage one line? */
4042             hunk->oldpos == 1 && hunk->oldlines == 1 &&
4043             /* does preimage begin with the heading? */
4044             (preimage = memchr(hunk->patch, '\n', hunk->size)) != NULL &&
4045             starts_with(++preimage, heading) &&
4046             /* does it record full SHA-1? */
4047             !get_oid_hex(preimage + sizeof(heading) - 1, oid) &&
4048             preimage[sizeof(heading) + GIT_SHA1_HEXSZ - 1] == '\n' &&
4049             /* does the abbreviated name on the index line agree with it? */
4050             starts_with(preimage + sizeof(heading) - 1, p->old_sha1_prefix))
4051                 return 0; /* it all looks fine */
4052
4053         /* we may have full object name on the index line */
4054         return get_oid_hex(p->old_sha1_prefix, oid);
4055 }
4056
4057 /* Build an index that contains the just the files needed for a 3way merge */
4058 static int build_fake_ancestor(struct apply_state *state, struct patch *list)
4059 {
4060         struct patch *patch;
4061         struct index_state result = { NULL };
4062         static struct lock_file lock;
4063         int res;
4064
4065         /* Once we start supporting the reverse patch, it may be
4066          * worth showing the new sha1 prefix, but until then...
4067          */
4068         for (patch = list; patch; patch = patch->next) {
4069                 struct object_id oid;
4070                 struct cache_entry *ce;
4071                 const char *name;
4072
4073                 name = patch->old_name ? patch->old_name : patch->new_name;
4074                 if (0 < patch->is_new)
4075                         continue;
4076
4077                 if (S_ISGITLINK(patch->old_mode)) {
4078                         if (!preimage_oid_in_gitlink_patch(patch, &oid))
4079                                 ; /* ok, the textual part looks sane */
4080                         else
4081                                 return error(_("sha1 information is lacking or "
4082                                                "useless for submodule %s"), name);
4083                 } else if (!get_oid_blob(patch->old_sha1_prefix, &oid)) {
4084                         ; /* ok */
4085                 } else if (!patch->lines_added && !patch->lines_deleted) {
4086                         /* mode-only change: update the current */
4087                         if (get_current_oid(state, patch->old_name, &oid))
4088                                 return error(_("mode change for %s, which is not "
4089                                                "in current HEAD"), name);
4090                 } else
4091                         return error(_("sha1 information is lacking or useless "
4092                                        "(%s)."), name);
4093
4094                 ce = make_cache_entry(patch->old_mode, oid.hash, name, 0, 0);
4095                 if (!ce)
4096                         return error(_("make_cache_entry failed for path '%s'"),
4097                                      name);
4098                 if (add_index_entry(&result, ce, ADD_CACHE_OK_TO_ADD)) {
4099                         free(ce);
4100                         return error(_("could not add %s to temporary index"),
4101                                      name);
4102                 }
4103         }
4104
4105         hold_lock_file_for_update(&lock, state->fake_ancestor, LOCK_DIE_ON_ERROR);
4106         res = write_locked_index(&result, &lock, COMMIT_LOCK);
4107         discard_index(&result);
4108
4109         if (res)
4110                 return error(_("could not write temporary index to %s"),
4111                              state->fake_ancestor);
4112
4113         return 0;
4114 }
4115
4116 static void stat_patch_list(struct apply_state *state, struct patch *patch)
4117 {
4118         int files, adds, dels;
4119
4120         for (files = adds = dels = 0 ; patch ; patch = patch->next) {
4121                 files++;
4122                 adds += patch->lines_added;
4123                 dels += patch->lines_deleted;
4124                 show_stats(state, patch);
4125         }
4126
4127         print_stat_summary(stdout, files, adds, dels);
4128 }
4129
4130 static void numstat_patch_list(struct apply_state *state,
4131                                struct patch *patch)
4132 {
4133         for ( ; patch; patch = patch->next) {
4134                 const char *name;
4135                 name = patch->new_name ? patch->new_name : patch->old_name;
4136                 if (patch->is_binary)
4137                         printf("-\t-\t");
4138                 else
4139                         printf("%d\t%d\t", patch->lines_added, patch->lines_deleted);
4140                 write_name_quoted(name, stdout, state->line_termination);
4141         }
4142 }
4143
4144 static void show_file_mode_name(const char *newdelete, unsigned int mode, const char *name)
4145 {
4146         if (mode)
4147                 printf(" %s mode %06o %s\n", newdelete, mode, name);
4148         else
4149                 printf(" %s %s\n", newdelete, name);
4150 }
4151
4152 static void show_mode_change(struct patch *p, int show_name)
4153 {
4154         if (p->old_mode && p->new_mode && p->old_mode != p->new_mode) {
4155                 if (show_name)
4156                         printf(" mode change %06o => %06o %s\n",
4157                                p->old_mode, p->new_mode, p->new_name);
4158                 else
4159                         printf(" mode change %06o => %06o\n",
4160                                p->old_mode, p->new_mode);
4161         }
4162 }
4163
4164 static void show_rename_copy(struct patch *p)
4165 {
4166         const char *renamecopy = p->is_rename ? "rename" : "copy";
4167         const char *old_name, *new_name;
4168
4169         /* Find common prefix */
4170         old_name = p->old_name;
4171         new_name = p->new_name;
4172         while (1) {
4173                 const char *slash_old, *slash_new;
4174                 slash_old = strchr(old_name, '/');
4175                 slash_new = strchr(new_name, '/');
4176                 if (!slash_old ||
4177                     !slash_new ||
4178                     slash_old - old_name != slash_new - new_name ||
4179                     memcmp(old_name, new_name, slash_new - new_name))
4180                         break;
4181                 old_name = slash_old + 1;
4182                 new_name = slash_new + 1;
4183         }
4184         /* p->old_name thru old_name is the common prefix, and old_name and new_name
4185          * through the end of names are renames
4186          */
4187         if (old_name != p->old_name)
4188                 printf(" %s %.*s{%s => %s} (%d%%)\n", renamecopy,
4189                        (int)(old_name - p->old_name), p->old_name,
4190                        old_name, new_name, p->score);
4191         else
4192                 printf(" %s %s => %s (%d%%)\n", renamecopy,
4193                        p->old_name, p->new_name, p->score);
4194         show_mode_change(p, 0);
4195 }
4196
4197 static void summary_patch_list(struct patch *patch)
4198 {
4199         struct patch *p;
4200
4201         for (p = patch; p; p = p->next) {
4202                 if (p->is_new)
4203                         show_file_mode_name("create", p->new_mode, p->new_name);
4204                 else if (p->is_delete)
4205                         show_file_mode_name("delete", p->old_mode, p->old_name);
4206                 else {
4207                         if (p->is_rename || p->is_copy)
4208                                 show_rename_copy(p);
4209                         else {
4210                                 if (p->score) {
4211                                         printf(" rewrite %s (%d%%)\n",
4212                                                p->new_name, p->score);
4213                                         show_mode_change(p, 0);
4214                                 }
4215                                 else
4216                                         show_mode_change(p, 1);
4217                         }
4218                 }
4219         }
4220 }
4221
4222 static void patch_stats(struct apply_state *state, struct patch *patch)
4223 {
4224         int lines = patch->lines_added + patch->lines_deleted;
4225
4226         if (lines > state->max_change)
4227                 state->max_change = lines;
4228         if (patch->old_name) {
4229                 int len = quote_c_style(patch->old_name, NULL, NULL, 0);
4230                 if (!len)
4231                         len = strlen(patch->old_name);
4232                 if (len > state->max_len)
4233                         state->max_len = len;
4234         }
4235         if (patch->new_name) {
4236                 int len = quote_c_style(patch->new_name, NULL, NULL, 0);
4237                 if (!len)
4238                         len = strlen(patch->new_name);
4239                 if (len > state->max_len)
4240                         state->max_len = len;
4241         }
4242 }
4243
4244 static int remove_file(struct apply_state *state, struct patch *patch, int rmdir_empty)
4245 {
4246         if (state->update_index) {
4247                 if (remove_file_from_cache(patch->old_name) < 0)
4248                         return error(_("unable to remove %s from index"), patch->old_name);
4249         }
4250         if (!state->cached) {
4251                 if (!remove_or_warn(patch->old_mode, patch->old_name) && rmdir_empty) {
4252                         remove_path(patch->old_name);
4253                 }
4254         }
4255         return 0;
4256 }
4257
4258 static int add_index_file(struct apply_state *state,
4259                           const char *path,
4260                           unsigned mode,
4261                           void *buf,
4262                           unsigned long size)
4263 {
4264         struct stat st;
4265         struct cache_entry *ce;
4266         int namelen = strlen(path);
4267         unsigned ce_size = cache_entry_size(namelen);
4268
4269         if (!state->update_index)
4270                 return 0;
4271
4272         ce = xcalloc(1, ce_size);
4273         memcpy(ce->name, path, namelen);
4274         ce->ce_mode = create_ce_mode(mode);
4275         ce->ce_flags = create_ce_flags(0);
4276         ce->ce_namelen = namelen;
4277         if (S_ISGITLINK(mode)) {
4278                 const char *s;
4279
4280                 if (!skip_prefix(buf, "Subproject commit ", &s) ||
4281                     get_oid_hex(s, &ce->oid)) {
4282                         free(ce);
4283                        return error(_("corrupt patch for submodule %s"), path);
4284                 }
4285         } else {
4286                 if (!state->cached) {
4287                         if (lstat(path, &st) < 0) {
4288                                 free(ce);
4289                                 return error_errno(_("unable to stat newly "
4290                                                      "created file '%s'"),
4291                                                    path);
4292                         }
4293                         fill_stat_cache_info(ce, &st);
4294                 }
4295                 if (write_object_file(buf, size, blob_type, &ce->oid) < 0) {
4296                         free(ce);
4297                         return error(_("unable to create backing store "
4298                                        "for newly created file %s"), path);
4299                 }
4300         }
4301         if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD) < 0) {
4302                 free(ce);
4303                 return error(_("unable to add cache entry for %s"), path);
4304         }
4305
4306         return 0;
4307 }
4308
4309 /*
4310  * Returns:
4311  *  -1 if an unrecoverable error happened
4312  *   0 if everything went well
4313  *   1 if a recoverable error happened
4314  */
4315 static int try_create_file(const char *path, unsigned int mode, const char *buf, unsigned long size)
4316 {
4317         int fd, res;
4318         struct strbuf nbuf = STRBUF_INIT;
4319
4320         if (S_ISGITLINK(mode)) {
4321                 struct stat st;
4322                 if (!lstat(path, &st) && S_ISDIR(st.st_mode))
4323                         return 0;
4324                 return !!mkdir(path, 0777);
4325         }
4326
4327         if (has_symlinks && S_ISLNK(mode))
4328                 /* Although buf:size is counted string, it also is NUL
4329                  * terminated.
4330                  */
4331                 return !!symlink(buf, path);
4332
4333         fd = open(path, O_CREAT | O_EXCL | O_WRONLY, (mode & 0100) ? 0777 : 0666);
4334         if (fd < 0)
4335                 return 1;
4336
4337         if (convert_to_working_tree(path, buf, size, &nbuf)) {
4338                 size = nbuf.len;
4339                 buf  = nbuf.buf;
4340         }
4341
4342         res = write_in_full(fd, buf, size) < 0;
4343         if (res)
4344                 error_errno(_("failed to write to '%s'"), path);
4345         strbuf_release(&nbuf);
4346
4347         if (close(fd) < 0 && !res)
4348                 return error_errno(_("closing file '%s'"), path);
4349
4350         return res ? -1 : 0;
4351 }
4352
4353 /*
4354  * We optimistically assume that the directories exist,
4355  * which is true 99% of the time anyway. If they don't,
4356  * we create them and try again.
4357  *
4358  * Returns:
4359  *   -1 on error
4360  *   0 otherwise
4361  */
4362 static int create_one_file(struct apply_state *state,
4363                            char *path,
4364                            unsigned mode,
4365                            const char *buf,
4366                            unsigned long size)
4367 {
4368         int res;
4369
4370         if (state->cached)
4371                 return 0;
4372
4373         res = try_create_file(path, mode, buf, size);
4374         if (res < 0)
4375                 return -1;
4376         if (!res)
4377                 return 0;
4378
4379         if (errno == ENOENT) {
4380                 if (safe_create_leading_directories(path))
4381                         return 0;
4382                 res = try_create_file(path, mode, buf, size);
4383                 if (res < 0)
4384                         return -1;
4385                 if (!res)
4386                         return 0;
4387         }
4388
4389         if (errno == EEXIST || errno == EACCES) {
4390                 /* We may be trying to create a file where a directory
4391                  * used to be.
4392                  */
4393                 struct stat st;
4394                 if (!lstat(path, &st) && (!S_ISDIR(st.st_mode) || !rmdir(path)))
4395                         errno = EEXIST;
4396         }
4397
4398         if (errno == EEXIST) {
4399                 unsigned int nr = getpid();
4400
4401                 for (;;) {
4402                         char newpath[PATH_MAX];
4403                         mksnpath(newpath, sizeof(newpath), "%s~%u", path, nr);
4404                         res = try_create_file(newpath, mode, buf, size);
4405                         if (res < 0)
4406                                 return -1;
4407                         if (!res) {
4408                                 if (!rename(newpath, path))
4409                                         return 0;
4410                                 unlink_or_warn(newpath);
4411                                 break;
4412                         }
4413                         if (errno != EEXIST)
4414                                 break;
4415                         ++nr;
4416                 }
4417         }
4418         return error_errno(_("unable to write file '%s' mode %o"),
4419                            path, mode);
4420 }
4421
4422 static int add_conflicted_stages_file(struct apply_state *state,
4423                                        struct patch *patch)
4424 {
4425         int stage, namelen;
4426         unsigned ce_size, mode;
4427         struct cache_entry *ce;
4428
4429         if (!state->update_index)
4430                 return 0;
4431         namelen = strlen(patch->new_name);
4432         ce_size = cache_entry_size(namelen);
4433         mode = patch->new_mode ? patch->new_mode : (S_IFREG | 0644);
4434
4435         remove_file_from_cache(patch->new_name);
4436         for (stage = 1; stage < 4; stage++) {
4437                 if (is_null_oid(&patch->threeway_stage[stage - 1]))
4438                         continue;
4439                 ce = xcalloc(1, ce_size);
4440                 memcpy(ce->name, patch->new_name, namelen);
4441                 ce->ce_mode = create_ce_mode(mode);
4442                 ce->ce_flags = create_ce_flags(stage);
4443                 ce->ce_namelen = namelen;
4444                 oidcpy(&ce->oid, &patch->threeway_stage[stage - 1]);
4445                 if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD) < 0) {
4446                         free(ce);
4447                         return error(_("unable to add cache entry for %s"),
4448                                      patch->new_name);
4449                 }
4450         }
4451
4452         return 0;
4453 }
4454
4455 static int create_file(struct apply_state *state, struct patch *patch)
4456 {
4457         char *path = patch->new_name;
4458         unsigned mode = patch->new_mode;
4459         unsigned long size = patch->resultsize;
4460         char *buf = patch->result;
4461
4462         if (!mode)
4463                 mode = S_IFREG | 0644;
4464         if (create_one_file(state, path, mode, buf, size))
4465                 return -1;
4466
4467         if (patch->conflicted_threeway)
4468                 return add_conflicted_stages_file(state, patch);
4469         else
4470                 return add_index_file(state, path, mode, buf, size);
4471 }
4472
4473 /* phase zero is to remove, phase one is to create */
4474 static int write_out_one_result(struct apply_state *state,
4475                                 struct patch *patch,
4476                                 int phase)
4477 {
4478         if (patch->is_delete > 0) {
4479                 if (phase == 0)
4480                         return remove_file(state, patch, 1);
4481                 return 0;
4482         }
4483         if (patch->is_new > 0 || patch->is_copy) {
4484                 if (phase == 1)
4485                         return create_file(state, patch);
4486                 return 0;
4487         }
4488         /*
4489          * Rename or modification boils down to the same
4490          * thing: remove the old, write the new
4491          */
4492         if (phase == 0)
4493                 return remove_file(state, patch, patch->is_rename);
4494         if (phase == 1)
4495                 return create_file(state, patch);
4496         return 0;
4497 }
4498
4499 static int write_out_one_reject(struct apply_state *state, struct patch *patch)
4500 {
4501         FILE *rej;
4502         char namebuf[PATH_MAX];
4503         struct fragment *frag;
4504         int cnt = 0;
4505         struct strbuf sb = STRBUF_INIT;
4506
4507         for (cnt = 0, frag = patch->fragments; frag; frag = frag->next) {
4508                 if (!frag->rejected)
4509                         continue;
4510                 cnt++;
4511         }
4512
4513         if (!cnt) {
4514                 if (state->apply_verbosity > verbosity_normal)
4515                         say_patch_name(stderr,
4516                                        _("Applied patch %s cleanly."), patch);
4517                 return 0;
4518         }
4519
4520         /* This should not happen, because a removal patch that leaves
4521          * contents are marked "rejected" at the patch level.
4522          */
4523         if (!patch->new_name)
4524                 die(_("internal error"));
4525
4526         /* Say this even without --verbose */
4527         strbuf_addf(&sb, Q_("Applying patch %%s with %d reject...",
4528                             "Applying patch %%s with %d rejects...",
4529                             cnt),
4530                     cnt);
4531         if (state->apply_verbosity > verbosity_silent)
4532                 say_patch_name(stderr, sb.buf, patch);
4533         strbuf_release(&sb);
4534
4535         cnt = strlen(patch->new_name);
4536         if (ARRAY_SIZE(namebuf) <= cnt + 5) {
4537                 cnt = ARRAY_SIZE(namebuf) - 5;
4538                 warning(_("truncating .rej filename to %.*s.rej"),
4539                         cnt - 1, patch->new_name);
4540         }
4541         memcpy(namebuf, patch->new_name, cnt);
4542         memcpy(namebuf + cnt, ".rej", 5);
4543
4544         rej = fopen(namebuf, "w");
4545         if (!rej)
4546                 return error_errno(_("cannot open %s"), namebuf);
4547
4548         /* Normal git tools never deal with .rej, so do not pretend
4549          * this is a git patch by saying --git or giving extended
4550          * headers.  While at it, maybe please "kompare" that wants
4551          * the trailing TAB and some garbage at the end of line ;-).
4552          */
4553         fprintf(rej, "diff a/%s b/%s\t(rejected hunks)\n",
4554                 patch->new_name, patch->new_name);
4555         for (cnt = 1, frag = patch->fragments;
4556              frag;
4557              cnt++, frag = frag->next) {
4558                 if (!frag->rejected) {
4559                         if (state->apply_verbosity > verbosity_silent)
4560                                 fprintf_ln(stderr, _("Hunk #%d applied cleanly."), cnt);
4561                         continue;
4562                 }
4563                 if (state->apply_verbosity > verbosity_silent)
4564                         fprintf_ln(stderr, _("Rejected hunk #%d."), cnt);
4565                 fprintf(rej, "%.*s", frag->size, frag->patch);
4566                 if (frag->patch[frag->size-1] != '\n')
4567                         fputc('\n', rej);
4568         }
4569         fclose(rej);
4570         return -1;
4571 }
4572
4573 /*
4574  * Returns:
4575  *  -1 if an error happened
4576  *   0 if the patch applied cleanly
4577  *   1 if the patch did not apply cleanly
4578  */
4579 static int write_out_results(struct apply_state *state, struct patch *list)
4580 {
4581         int phase;
4582         int errs = 0;
4583         struct patch *l;
4584         struct string_list cpath = STRING_LIST_INIT_DUP;
4585
4586         for (phase = 0; phase < 2; phase++) {
4587                 l = list;
4588                 while (l) {
4589                         if (l->rejected)
4590                                 errs = 1;
4591                         else {
4592                                 if (write_out_one_result(state, l, phase)) {
4593                                         string_list_clear(&cpath, 0);
4594                                         return -1;
4595                                 }
4596                                 if (phase == 1) {
4597                                         if (write_out_one_reject(state, l))
4598                                                 errs = 1;
4599                                         if (l->conflicted_threeway) {
4600                                                 string_list_append(&cpath, l->new_name);
4601                                                 errs = 1;
4602                                         }
4603                                 }
4604                         }
4605                         l = l->next;
4606                 }
4607         }
4608
4609         if (cpath.nr) {
4610                 struct string_list_item *item;
4611
4612                 string_list_sort(&cpath);
4613                 if (state->apply_verbosity > verbosity_silent) {
4614                         for_each_string_list_item(item, &cpath)
4615                                 fprintf(stderr, "U %s\n", item->string);
4616                 }
4617                 string_list_clear(&cpath, 0);
4618
4619                 rerere(0);
4620         }
4621
4622         return errs;
4623 }
4624
4625 /*
4626  * Try to apply a patch.
4627  *
4628  * Returns:
4629  *  -128 if a bad error happened (like patch unreadable)
4630  *  -1 if patch did not apply and user cannot deal with it
4631  *   0 if the patch applied
4632  *   1 if the patch did not apply but user might fix it
4633  */
4634 static int apply_patch(struct apply_state *state,
4635                        int fd,
4636                        const char *filename,
4637                        int options)
4638 {
4639         size_t offset;
4640         struct strbuf buf = STRBUF_INIT; /* owns the patch text */
4641         struct patch *list = NULL, **listp = &list;
4642         int skipped_patch = 0;
4643         int res = 0;
4644
4645         state->patch_input_file = filename;
4646         if (read_patch_file(&buf, fd) < 0)
4647                 return -128;
4648         offset = 0;
4649         while (offset < buf.len) {
4650                 struct patch *patch;
4651                 int nr;
4652
4653                 patch = xcalloc(1, sizeof(*patch));
4654                 patch->inaccurate_eof = !!(options & APPLY_OPT_INACCURATE_EOF);
4655                 patch->recount =  !!(options & APPLY_OPT_RECOUNT);
4656                 nr = parse_chunk(state, buf.buf + offset, buf.len - offset, patch);
4657                 if (nr < 0) {
4658                         free_patch(patch);
4659                         if (nr == -128) {
4660                                 res = -128;
4661                                 goto end;
4662                         }
4663                         break;
4664                 }
4665                 if (state->apply_in_reverse)
4666                         reverse_patches(patch);
4667                 if (use_patch(state, patch)) {
4668                         patch_stats(state, patch);
4669                         *listp = patch;
4670                         listp = &patch->next;
4671                 }
4672                 else {
4673                         if (state->apply_verbosity > verbosity_normal)
4674                                 say_patch_name(stderr, _("Skipped patch '%s'."), patch);
4675                         free_patch(patch);
4676                         skipped_patch++;
4677                 }
4678                 offset += nr;
4679         }
4680
4681         if (!list && !skipped_patch) {
4682                 error(_("unrecognized input"));
4683                 res = -128;
4684                 goto end;
4685         }
4686
4687         if (state->whitespace_error && (state->ws_error_action == die_on_ws_error))
4688                 state->apply = 0;
4689
4690         state->update_index = state->check_index && state->apply;
4691         if (state->update_index && !is_lock_file_locked(&state->lock_file)) {
4692                 if (state->index_file)
4693                         hold_lock_file_for_update(&state->lock_file,
4694                                                   state->index_file,
4695                                                   LOCK_DIE_ON_ERROR);
4696                 else
4697                         hold_locked_index(&state->lock_file, LOCK_DIE_ON_ERROR);
4698         }
4699
4700         if (state->check_index && read_apply_cache(state) < 0) {
4701                 error(_("unable to read index file"));
4702                 res = -128;
4703                 goto end;
4704         }
4705
4706         if (state->check || state->apply) {
4707                 int r = check_patch_list(state, list);
4708                 if (r == -128) {
4709                         res = -128;
4710                         goto end;
4711                 }
4712                 if (r < 0 && !state->apply_with_reject) {
4713                         res = -1;
4714                         goto end;
4715                 }
4716         }
4717
4718         if (state->apply) {
4719                 int write_res = write_out_results(state, list);
4720                 if (write_res < 0) {
4721                         res = -128;
4722                         goto end;
4723                 }
4724                 if (write_res > 0) {
4725                         /* with --3way, we still need to write the index out */
4726                         res = state->apply_with_reject ? -1 : 1;
4727                         goto end;
4728                 }
4729         }
4730
4731         if (state->fake_ancestor &&
4732             build_fake_ancestor(state, list)) {
4733                 res = -128;
4734                 goto end;
4735         }
4736
4737         if (state->diffstat && state->apply_verbosity > verbosity_silent)
4738                 stat_patch_list(state, list);
4739
4740         if (state->numstat && state->apply_verbosity > verbosity_silent)
4741                 numstat_patch_list(state, list);
4742
4743         if (state->summary && state->apply_verbosity > verbosity_silent)
4744                 summary_patch_list(list);
4745
4746 end:
4747         free_patch_list(list);
4748         strbuf_release(&buf);
4749         string_list_clear(&state->fn_table, 0);
4750         return res;
4751 }
4752
4753 static int apply_option_parse_exclude(const struct option *opt,
4754                                       const char *arg, int unset)
4755 {
4756         struct apply_state *state = opt->value;
4757         add_name_limit(state, arg, 1);
4758         return 0;
4759 }
4760
4761 static int apply_option_parse_include(const struct option *opt,
4762                                       const char *arg, int unset)
4763 {
4764         struct apply_state *state = opt->value;
4765         add_name_limit(state, arg, 0);
4766         state->has_include = 1;
4767         return 0;
4768 }
4769
4770 static int apply_option_parse_p(const struct option *opt,
4771                                 const char *arg,
4772                                 int unset)
4773 {
4774         struct apply_state *state = opt->value;
4775         state->p_value = atoi(arg);
4776         state->p_value_known = 1;
4777         return 0;
4778 }
4779
4780 static int apply_option_parse_space_change(const struct option *opt,
4781                                            const char *arg, int unset)
4782 {
4783         struct apply_state *state = opt->value;
4784         if (unset)
4785                 state->ws_ignore_action = ignore_ws_none;
4786         else
4787                 state->ws_ignore_action = ignore_ws_change;
4788         return 0;
4789 }
4790
4791 static int apply_option_parse_whitespace(const struct option *opt,
4792                                          const char *arg, int unset)
4793 {
4794         struct apply_state *state = opt->value;
4795         state->whitespace_option = arg;
4796         if (parse_whitespace_option(state, arg))
4797                 exit(1);
4798         return 0;
4799 }
4800
4801 static int apply_option_parse_directory(const struct option *opt,
4802                                         const char *arg, int unset)
4803 {
4804         struct apply_state *state = opt->value;
4805         strbuf_reset(&state->root);
4806         strbuf_addstr(&state->root, arg);
4807         strbuf_complete(&state->root, '/');
4808         return 0;
4809 }
4810
4811 int apply_all_patches(struct apply_state *state,
4812                       int argc,
4813                       const char **argv,
4814                       int options)
4815 {
4816         int i;
4817         int res;
4818         int errs = 0;
4819         int read_stdin = 1;
4820
4821         for (i = 0; i < argc; i++) {
4822                 const char *arg = argv[i];
4823                 char *to_free = NULL;
4824                 int fd;
4825
4826                 if (!strcmp(arg, "-")) {
4827                         res = apply_patch(state, 0, "<stdin>", options);
4828                         if (res < 0)
4829                                 goto end;
4830                         errs |= res;
4831                         read_stdin = 0;
4832                         continue;
4833                 } else
4834                         arg = to_free = prefix_filename(state->prefix, arg);
4835
4836                 fd = open(arg, O_RDONLY);
4837                 if (fd < 0) {
4838                         error(_("can't open patch '%s': %s"), arg, strerror(errno));
4839                         res = -128;
4840                         free(to_free);
4841                         goto end;
4842                 }
4843                 read_stdin = 0;
4844                 set_default_whitespace_mode(state);
4845                 res = apply_patch(state, fd, arg, options);
4846                 close(fd);
4847                 free(to_free);
4848                 if (res < 0)
4849                         goto end;
4850                 errs |= res;
4851         }
4852         set_default_whitespace_mode(state);
4853         if (read_stdin) {
4854                 res = apply_patch(state, 0, "<stdin>", options);
4855                 if (res < 0)
4856                         goto end;
4857                 errs |= res;
4858         }
4859
4860         if (state->whitespace_error) {
4861                 if (state->squelch_whitespace_errors &&
4862                     state->squelch_whitespace_errors < state->whitespace_error) {
4863                         int squelched =
4864                                 state->whitespace_error - state->squelch_whitespace_errors;
4865                         warning(Q_("squelched %d whitespace error",
4866                                    "squelched %d whitespace errors",
4867                                    squelched),
4868                                 squelched);
4869                 }
4870                 if (state->ws_error_action == die_on_ws_error) {
4871                         error(Q_("%d line adds whitespace errors.",
4872                                  "%d lines add whitespace errors.",
4873                                  state->whitespace_error),
4874                               state->whitespace_error);
4875                         res = -128;
4876                         goto end;
4877                 }
4878                 if (state->applied_after_fixing_ws && state->apply)
4879                         warning(Q_("%d line applied after"
4880                                    " fixing whitespace errors.",
4881                                    "%d lines applied after"
4882                                    " fixing whitespace errors.",
4883                                    state->applied_after_fixing_ws),
4884                                 state->applied_after_fixing_ws);
4885                 else if (state->whitespace_error)
4886                         warning(Q_("%d line adds whitespace errors.",
4887                                    "%d lines add whitespace errors.",
4888                                    state->whitespace_error),
4889                                 state->whitespace_error);
4890         }
4891
4892         if (state->update_index) {
4893                 res = write_locked_index(&the_index, &state->lock_file, COMMIT_LOCK);
4894                 if (res) {
4895                         error(_("Unable to write new index file"));
4896                         res = -128;
4897                         goto end;
4898                 }
4899         }
4900
4901         res = !!errs;
4902
4903 end:
4904         rollback_lock_file(&state->lock_file);
4905
4906         if (state->apply_verbosity <= verbosity_silent) {
4907                 set_error_routine(state->saved_error_routine);
4908                 set_warn_routine(state->saved_warn_routine);
4909         }
4910
4911         if (res > -1)
4912                 return res;
4913         return (res == -1 ? 1 : 128);
4914 }
4915
4916 int apply_parse_options(int argc, const char **argv,
4917                         struct apply_state *state,
4918                         int *force_apply, int *options,
4919                         const char * const *apply_usage)
4920 {
4921         struct option builtin_apply_options[] = {
4922                 { OPTION_CALLBACK, 0, "exclude", state, N_("path"),
4923                         N_("don't apply changes matching the given path"),
4924                         0, apply_option_parse_exclude },
4925                 { OPTION_CALLBACK, 0, "include", state, N_("path"),
4926                         N_("apply changes matching the given path"),
4927                         0, apply_option_parse_include },
4928                 { OPTION_CALLBACK, 'p', NULL, state, N_("num"),
4929                         N_("remove <num> leading slashes from traditional diff paths"),
4930                         0, apply_option_parse_p },
4931                 OPT_BOOL(0, "no-add", &state->no_add,
4932                         N_("ignore additions made by the patch")),
4933                 OPT_BOOL(0, "stat", &state->diffstat,
4934                         N_("instead of applying the patch, output diffstat for the input")),
4935                 OPT_NOOP_NOARG(0, "allow-binary-replacement"),
4936                 OPT_NOOP_NOARG(0, "binary"),
4937                 OPT_BOOL(0, "numstat", &state->numstat,
4938                         N_("show number of added and deleted lines in decimal notation")),
4939                 OPT_BOOL(0, "summary", &state->summary,
4940                         N_("instead of applying the patch, output a summary for the input")),
4941                 OPT_BOOL(0, "check", &state->check,
4942                         N_("instead of applying the patch, see if the patch is applicable")),
4943                 OPT_BOOL(0, "index", &state->check_index,
4944                         N_("make sure the patch is applicable to the current index")),
4945                 OPT_BOOL(0, "cached", &state->cached,
4946                         N_("apply a patch without touching the working tree")),
4947                 OPT_BOOL_F(0, "unsafe-paths", &state->unsafe_paths,
4948                            N_("accept a patch that touches outside the working area"),
4949                            PARSE_OPT_NOCOMPLETE),
4950                 OPT_BOOL(0, "apply", force_apply,
4951                         N_("also apply the patch (use with --stat/--summary/--check)")),
4952                 OPT_BOOL('3', "3way", &state->threeway,
4953                          N_( "attempt three-way merge if a patch does not apply")),
4954                 OPT_FILENAME(0, "build-fake-ancestor", &state->fake_ancestor,
4955                         N_("build a temporary index based on embedded index information")),
4956                 /* Think twice before adding "--nul" synonym to this */
4957                 OPT_SET_INT('z', NULL, &state->line_termination,
4958                         N_("paths are separated with NUL character"), '\0'),
4959                 OPT_INTEGER('C', NULL, &state->p_context,
4960                                 N_("ensure at least <n> lines of context match")),
4961                 { OPTION_CALLBACK, 0, "whitespace", state, N_("action"),
4962                         N_("detect new or modified lines that have whitespace errors"),
4963                         0, apply_option_parse_whitespace },
4964                 { OPTION_CALLBACK, 0, "ignore-space-change", state, NULL,
4965                         N_("ignore changes in whitespace when finding context"),
4966                         PARSE_OPT_NOARG, apply_option_parse_space_change },
4967                 { OPTION_CALLBACK, 0, "ignore-whitespace", state, NULL,
4968                         N_("ignore changes in whitespace when finding context"),
4969                         PARSE_OPT_NOARG, apply_option_parse_space_change },
4970                 OPT_BOOL('R', "reverse", &state->apply_in_reverse,
4971                         N_("apply the patch in reverse")),
4972                 OPT_BOOL(0, "unidiff-zero", &state->unidiff_zero,
4973                         N_("don't expect at least one line of context")),
4974                 OPT_BOOL(0, "reject", &state->apply_with_reject,
4975                         N_("leave the rejected hunks in corresponding *.rej files")),
4976                 OPT_BOOL(0, "allow-overlap", &state->allow_overlap,
4977                         N_("allow overlapping hunks")),
4978                 OPT__VERBOSE(&state->apply_verbosity, N_("be verbose")),
4979                 OPT_BIT(0, "inaccurate-eof", options,
4980                         N_("tolerate incorrectly detected missing new-line at the end of file"),
4981                         APPLY_OPT_INACCURATE_EOF),
4982                 OPT_BIT(0, "recount", options,
4983                         N_("do not trust the line counts in the hunk headers"),
4984                         APPLY_OPT_RECOUNT),
4985                 { OPTION_CALLBACK, 0, "directory", state, N_("root"),
4986                         N_("prepend <root> to all filenames"),
4987                         0, apply_option_parse_directory },
4988                 OPT_END()
4989         };
4990
4991         return parse_options(argc, argv, state->prefix, builtin_apply_options, apply_usage, 0);
4992 }