Merge branch 'jk/maint-fetch-submodule-check-fix' into maint
[git] / pack-write.c
1 #include "cache.h"
2 #include "pack.h"
3 #include "csum-file.h"
4
5 void reset_pack_idx_option(struct pack_idx_option *opts)
6 {
7         memset(opts, 0, sizeof(*opts));
8         opts->version = 2;
9         opts->off32_limit = 0x7fffffff;
10 }
11
12 static int sha1_compare(const void *_a, const void *_b)
13 {
14         struct pack_idx_entry *a = *(struct pack_idx_entry **)_a;
15         struct pack_idx_entry *b = *(struct pack_idx_entry **)_b;
16         return hashcmp(a->sha1, b->sha1);
17 }
18
19 static int cmp_uint32(const void *a_, const void *b_)
20 {
21         uint32_t a = *((uint32_t *)a_);
22         uint32_t b = *((uint32_t *)b_);
23
24         return (a < b) ? -1 : (a != b);
25 }
26
27 static int need_large_offset(off_t offset, const struct pack_idx_option *opts)
28 {
29         uint32_t ofsval;
30
31         if ((offset >> 31) || (opts->off32_limit < offset))
32                 return 1;
33         if (!opts->anomaly_nr)
34                 return 0;
35         ofsval = offset;
36         return !!bsearch(&ofsval, opts->anomaly, opts->anomaly_nr,
37                          sizeof(ofsval), cmp_uint32);
38 }
39
40 /*
41  * On entry *sha1 contains the pack content SHA1 hash, on exit it is
42  * the SHA1 hash of sorted object names. The objects array passed in
43  * will be sorted by SHA1 on exit.
44  */
45 const char *write_idx_file(const char *index_name, struct pack_idx_entry **objects,
46                            int nr_objects, const struct pack_idx_option *opts,
47                            unsigned char *sha1)
48 {
49         struct sha1file *f;
50         struct pack_idx_entry **sorted_by_sha, **list, **last;
51         off_t last_obj_offset = 0;
52         uint32_t array[256];
53         int i, fd;
54         git_SHA_CTX ctx;
55         uint32_t index_version;
56
57         if (nr_objects) {
58                 sorted_by_sha = objects;
59                 list = sorted_by_sha;
60                 last = sorted_by_sha + nr_objects;
61                 for (i = 0; i < nr_objects; ++i) {
62                         if (objects[i]->offset > last_obj_offset)
63                                 last_obj_offset = objects[i]->offset;
64                 }
65                 qsort(sorted_by_sha, nr_objects, sizeof(sorted_by_sha[0]),
66                       sha1_compare);
67         }
68         else
69                 sorted_by_sha = list = last = NULL;
70
71         if (opts->flags & WRITE_IDX_VERIFY) {
72                 assert(index_name);
73                 f = sha1fd_check(index_name);
74         } else {
75                 if (!index_name) {
76                         static char tmpfile[PATH_MAX];
77                         fd = odb_mkstemp(tmpfile, sizeof(tmpfile), "pack/tmp_idx_XXXXXX");
78                         index_name = xstrdup(tmpfile);
79                 } else {
80                         unlink(index_name);
81                         fd = open(index_name, O_CREAT|O_EXCL|O_WRONLY, 0600);
82                 }
83                 if (fd < 0)
84                         die_errno("unable to create '%s'", index_name);
85                 f = sha1fd(fd, index_name);
86         }
87
88         /* if last object's offset is >= 2^31 we should use index V2 */
89         index_version = need_large_offset(last_obj_offset, opts) ? 2 : opts->version;
90
91         /* index versions 2 and above need a header */
92         if (index_version >= 2) {
93                 struct pack_idx_header hdr;
94                 hdr.idx_signature = htonl(PACK_IDX_SIGNATURE);
95                 hdr.idx_version = htonl(index_version);
96                 sha1write(f, &hdr, sizeof(hdr));
97         }
98
99         /*
100          * Write the first-level table (the list is sorted,
101          * but we use a 256-entry lookup to be able to avoid
102          * having to do eight extra binary search iterations).
103          */
104         for (i = 0; i < 256; i++) {
105                 struct pack_idx_entry **next = list;
106                 while (next < last) {
107                         struct pack_idx_entry *obj = *next;
108                         if (obj->sha1[0] != i)
109                                 break;
110                         next++;
111                 }
112                 array[i] = htonl(next - sorted_by_sha);
113                 list = next;
114         }
115         sha1write(f, array, 256 * 4);
116
117         /* compute the SHA1 hash of sorted object names. */
118         git_SHA1_Init(&ctx);
119
120         /*
121          * Write the actual SHA1 entries..
122          */
123         list = sorted_by_sha;
124         for (i = 0; i < nr_objects; i++) {
125                 struct pack_idx_entry *obj = *list++;
126                 if (index_version < 2) {
127                         uint32_t offset = htonl(obj->offset);
128                         sha1write(f, &offset, 4);
129                 }
130                 sha1write(f, obj->sha1, 20);
131                 git_SHA1_Update(&ctx, obj->sha1, 20);
132         }
133
134         if (index_version >= 2) {
135                 unsigned int nr_large_offset = 0;
136
137                 /* write the crc32 table */
138                 list = sorted_by_sha;
139                 for (i = 0; i < nr_objects; i++) {
140                         struct pack_idx_entry *obj = *list++;
141                         uint32_t crc32_val = htonl(obj->crc32);
142                         sha1write(f, &crc32_val, 4);
143                 }
144
145                 /* write the 32-bit offset table */
146                 list = sorted_by_sha;
147                 for (i = 0; i < nr_objects; i++) {
148                         struct pack_idx_entry *obj = *list++;
149                         uint32_t offset;
150
151                         offset = (need_large_offset(obj->offset, opts)
152                                   ? (0x80000000 | nr_large_offset++)
153                                   : obj->offset);
154                         offset = htonl(offset);
155                         sha1write(f, &offset, 4);
156                 }
157
158                 /* write the large offset table */
159                 list = sorted_by_sha;
160                 while (nr_large_offset) {
161                         struct pack_idx_entry *obj = *list++;
162                         uint64_t offset = obj->offset;
163                         uint32_t split[2];
164
165                         if (!need_large_offset(offset, opts))
166                                 continue;
167                         split[0] = htonl(offset >> 32);
168                         split[1] = htonl(offset & 0xffffffff);
169                         sha1write(f, split, 8);
170                         nr_large_offset--;
171                 }
172         }
173
174         sha1write(f, sha1, 20);
175         sha1close(f, NULL, ((opts->flags & WRITE_IDX_VERIFY)
176                             ? CSUM_CLOSE : CSUM_FSYNC));
177         git_SHA1_Final(sha1, &ctx);
178         return index_name;
179 }
180
181 /*
182  * Update pack header with object_count and compute new SHA1 for pack data
183  * associated to pack_fd, and write that SHA1 at the end.  That new SHA1
184  * is also returned in new_pack_sha1.
185  *
186  * If partial_pack_sha1 is non null, then the SHA1 of the existing pack
187  * (without the header update) is computed and validated against the
188  * one provided in partial_pack_sha1.  The validation is performed at
189  * partial_pack_offset bytes in the pack file.  The SHA1 of the remaining
190  * data (i.e. from partial_pack_offset to the end) is then computed and
191  * returned in partial_pack_sha1.
192  *
193  * Note that new_pack_sha1 is updated last, so both new_pack_sha1 and
194  * partial_pack_sha1 can refer to the same buffer if the caller is not
195  * interested in the resulting SHA1 of pack data above partial_pack_offset.
196  */
197 void fixup_pack_header_footer(int pack_fd,
198                          unsigned char *new_pack_sha1,
199                          const char *pack_name,
200                          uint32_t object_count,
201                          unsigned char *partial_pack_sha1,
202                          off_t partial_pack_offset)
203 {
204         int aligned_sz, buf_sz = 8 * 1024;
205         git_SHA_CTX old_sha1_ctx, new_sha1_ctx;
206         struct pack_header hdr;
207         char *buf;
208
209         git_SHA1_Init(&old_sha1_ctx);
210         git_SHA1_Init(&new_sha1_ctx);
211
212         if (lseek(pack_fd, 0, SEEK_SET) != 0)
213                 die_errno("Failed seeking to start of '%s'", pack_name);
214         if (read_in_full(pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
215                 die_errno("Unable to reread header of '%s'", pack_name);
216         if (lseek(pack_fd, 0, SEEK_SET) != 0)
217                 die_errno("Failed seeking to start of '%s'", pack_name);
218         git_SHA1_Update(&old_sha1_ctx, &hdr, sizeof(hdr));
219         hdr.hdr_entries = htonl(object_count);
220         git_SHA1_Update(&new_sha1_ctx, &hdr, sizeof(hdr));
221         write_or_die(pack_fd, &hdr, sizeof(hdr));
222         partial_pack_offset -= sizeof(hdr);
223
224         buf = xmalloc(buf_sz);
225         aligned_sz = buf_sz - sizeof(hdr);
226         for (;;) {
227                 ssize_t m, n;
228                 m = (partial_pack_sha1 && partial_pack_offset < aligned_sz) ?
229                         partial_pack_offset : aligned_sz;
230                 n = xread(pack_fd, buf, m);
231                 if (!n)
232                         break;
233                 if (n < 0)
234                         die_errno("Failed to checksum '%s'", pack_name);
235                 git_SHA1_Update(&new_sha1_ctx, buf, n);
236
237                 aligned_sz -= n;
238                 if (!aligned_sz)
239                         aligned_sz = buf_sz;
240
241                 if (!partial_pack_sha1)
242                         continue;
243
244                 git_SHA1_Update(&old_sha1_ctx, buf, n);
245                 partial_pack_offset -= n;
246                 if (partial_pack_offset == 0) {
247                         unsigned char sha1[20];
248                         git_SHA1_Final(sha1, &old_sha1_ctx);
249                         if (hashcmp(sha1, partial_pack_sha1) != 0)
250                                 die("Unexpected checksum for %s "
251                                     "(disk corruption?)", pack_name);
252                         /*
253                          * Now let's compute the SHA1 of the remainder of the
254                          * pack, which also means making partial_pack_offset
255                          * big enough not to matter anymore.
256                          */
257                         git_SHA1_Init(&old_sha1_ctx);
258                         partial_pack_offset = ~partial_pack_offset;
259                         partial_pack_offset -= MSB(partial_pack_offset, 1);
260                 }
261         }
262         free(buf);
263
264         if (partial_pack_sha1)
265                 git_SHA1_Final(partial_pack_sha1, &old_sha1_ctx);
266         git_SHA1_Final(new_pack_sha1, &new_sha1_ctx);
267         write_or_die(pack_fd, new_pack_sha1, 20);
268         fsync_or_die(pack_fd, pack_name);
269 }
270
271 char *index_pack_lockfile(int ip_out)
272 {
273         char packname[46];
274
275         /*
276          * The first thing we expect from index-pack's output
277          * is "pack\t%40s\n" or "keep\t%40s\n" (46 bytes) where
278          * %40s is the newly created pack SHA1 name.  In the "keep"
279          * case, we need it to remove the corresponding .keep file
280          * later on.  If we don't get that then tough luck with it.
281          */
282         if (read_in_full(ip_out, packname, 46) == 46 && packname[45] == '\n' &&
283             memcmp(packname, "keep\t", 5) == 0) {
284                 char path[PATH_MAX];
285                 packname[45] = 0;
286                 snprintf(path, sizeof(path), "%s/pack/pack-%s.keep",
287                          get_object_directory(), packname + 5);
288                 return xstrdup(path);
289         }
290         return NULL;
291 }
292
293 /*
294  * The per-object header is a pretty dense thing, which is
295  *  - first byte: low four bits are "size", then three bits of "type",
296  *    and the high bit is "size continues".
297  *  - each byte afterwards: low seven bits are size continuation,
298  *    with the high bit being "size continues"
299  */
300 int encode_in_pack_object_header(enum object_type type, uintmax_t size, unsigned char *hdr)
301 {
302         int n = 1;
303         unsigned char c;
304
305         if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
306                 die("bad type %d", type);
307
308         c = (type << 4) | (size & 15);
309         size >>= 4;
310         while (size) {
311                 *hdr++ = c | 0x80;
312                 c = size & 0x7f;
313                 size >>= 7;
314                 n++;
315         }
316         *hdr = c;
317         return n;
318 }