Merge branch 'eb/no-pthreads'
[git] / bulk-checkin.c
1 /*
2  * Copyright (c) 2011, Google Inc.
3  */
4 #include "cache.h"
5 #include "bulk-checkin.h"
6 #include "csum-file.h"
7 #include "pack.h"
8 #include "strbuf.h"
9
10 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
11
12 static struct bulk_checkin_state {
13         unsigned plugged:1;
14
15         char *pack_tmp_name;
16         struct sha1file *f;
17         off_t offset;
18         struct pack_idx_option pack_idx_opts;
19
20         struct pack_idx_entry **written;
21         uint32_t alloc_written;
22         uint32_t nr_written;
23 } state;
24
25 static void finish_bulk_checkin(struct bulk_checkin_state *state)
26 {
27         unsigned char sha1[20];
28         struct strbuf packname = STRBUF_INIT;
29         int i;
30
31         if (!state->f)
32                 return;
33
34         if (state->nr_written == 0) {
35                 close(state->f->fd);
36                 unlink(state->pack_tmp_name);
37                 goto clear_exit;
38         } else if (state->nr_written == 1) {
39                 sha1close(state->f, sha1, CSUM_FSYNC);
40         } else {
41                 int fd = sha1close(state->f, sha1, 0);
42                 fixup_pack_header_footer(fd, sha1, state->pack_tmp_name,
43                                          state->nr_written, sha1,
44                                          state->offset);
45                 close(fd);
46         }
47
48         strbuf_addf(&packname, "%s/pack/pack-", get_object_directory());
49         finish_tmp_packfile(&packname, state->pack_tmp_name,
50                             state->written, state->nr_written,
51                             &state->pack_idx_opts, sha1);
52         for (i = 0; i < state->nr_written; i++)
53                 free(state->written[i]);
54
55 clear_exit:
56         free(state->written);
57         memset(state, 0, sizeof(*state));
58
59         strbuf_release(&packname);
60         /* Make objects we just wrote available to ourselves */
61         reprepare_packed_git();
62 }
63
64 static int already_written(struct bulk_checkin_state *state, unsigned char sha1[])
65 {
66         int i;
67
68         /* The object may already exist in the repository */
69         if (has_sha1_file(sha1))
70                 return 1;
71
72         /* Might want to keep the list sorted */
73         for (i = 0; i < state->nr_written; i++)
74                 if (!hashcmp(state->written[i]->sha1, sha1))
75                         return 1;
76
77         /* This is a new object we need to keep */
78         return 0;
79 }
80
81 /*
82  * Read the contents from fd for size bytes, streaming it to the
83  * packfile in state while updating the hash in ctx. Signal a failure
84  * by returning a negative value when the resulting pack would exceed
85  * the pack size limit and this is not the first object in the pack,
86  * so that the caller can discard what we wrote from the current pack
87  * by truncating it and opening a new one. The caller will then call
88  * us again after rewinding the input fd.
89  *
90  * The already_hashed_to pointer is kept untouched by the caller to
91  * make sure we do not hash the same byte when we are called
92  * again. This way, the caller does not have to checkpoint its hash
93  * status before calling us just in case we ask it to call us again
94  * with a new pack.
95  */
96 static int stream_to_pack(struct bulk_checkin_state *state,
97                           git_SHA_CTX *ctx, off_t *already_hashed_to,
98                           int fd, size_t size, enum object_type type,
99                           const char *path, unsigned flags)
100 {
101         git_zstream s;
102         unsigned char obuf[16384];
103         unsigned hdrlen;
104         int status = Z_OK;
105         int write_object = (flags & HASH_WRITE_OBJECT);
106         off_t offset = 0;
107
108         memset(&s, 0, sizeof(s));
109         git_deflate_init(&s, pack_compression_level);
110
111         hdrlen = encode_in_pack_object_header(type, size, obuf);
112         s.next_out = obuf + hdrlen;
113         s.avail_out = sizeof(obuf) - hdrlen;
114
115         while (status != Z_STREAM_END) {
116                 unsigned char ibuf[16384];
117
118                 if (size && !s.avail_in) {
119                         ssize_t rsize = size < sizeof(ibuf) ? size : sizeof(ibuf);
120                         if (read_in_full(fd, ibuf, rsize) != rsize)
121                                 die("failed to read %d bytes from '%s'",
122                                     (int)rsize, path);
123                         offset += rsize;
124                         if (*already_hashed_to < offset) {
125                                 size_t hsize = offset - *already_hashed_to;
126                                 if (rsize < hsize)
127                                         hsize = rsize;
128                                 if (hsize)
129                                         git_SHA1_Update(ctx, ibuf, hsize);
130                                 *already_hashed_to = offset;
131                         }
132                         s.next_in = ibuf;
133                         s.avail_in = rsize;
134                         size -= rsize;
135                 }
136
137                 status = git_deflate(&s, size ? 0 : Z_FINISH);
138
139                 if (!s.avail_out || status == Z_STREAM_END) {
140                         if (write_object) {
141                                 size_t written = s.next_out - obuf;
142
143                                 /* would we bust the size limit? */
144                                 if (state->nr_written &&
145                                     pack_size_limit_cfg &&
146                                     pack_size_limit_cfg < state->offset + written) {
147                                         git_deflate_abort(&s);
148                                         return -1;
149                                 }
150
151                                 sha1write(state->f, obuf, written);
152                                 state->offset += written;
153                         }
154                         s.next_out = obuf;
155                         s.avail_out = sizeof(obuf);
156                 }
157
158                 switch (status) {
159                 case Z_OK:
160                 case Z_BUF_ERROR:
161                 case Z_STREAM_END:
162                         continue;
163                 default:
164                         die("unexpected deflate failure: %d", status);
165                 }
166         }
167         git_deflate_end(&s);
168         return 0;
169 }
170
171 /* Lazily create backing packfile for the state */
172 static void prepare_to_stream(struct bulk_checkin_state *state,
173                               unsigned flags)
174 {
175         if (!(flags & HASH_WRITE_OBJECT) || state->f)
176                 return;
177
178         state->f = create_tmp_packfile(&state->pack_tmp_name);
179         reset_pack_idx_option(&state->pack_idx_opts);
180
181         /* Pretend we are going to write only one object */
182         state->offset = write_pack_header(state->f, 1);
183         if (!state->offset)
184                 die_errno("unable to write pack header");
185 }
186
187 static int deflate_to_pack(struct bulk_checkin_state *state,
188                            unsigned char result_sha1[],
189                            int fd, size_t size,
190                            enum object_type type, const char *path,
191                            unsigned flags)
192 {
193         off_t seekback, already_hashed_to;
194         git_SHA_CTX ctx;
195         unsigned char obuf[16384];
196         unsigned header_len;
197         struct sha1file_checkpoint checkpoint;
198         struct pack_idx_entry *idx = NULL;
199
200         seekback = lseek(fd, 0, SEEK_CUR);
201         if (seekback == (off_t) -1)
202                 return error("cannot find the current offset");
203
204         header_len = sprintf((char *)obuf, "%s %" PRIuMAX,
205                              typename(type), (uintmax_t)size) + 1;
206         git_SHA1_Init(&ctx);
207         git_SHA1_Update(&ctx, obuf, header_len);
208
209         /* Note: idx is non-NULL when we are writing */
210         if ((flags & HASH_WRITE_OBJECT) != 0)
211                 idx = xcalloc(1, sizeof(*idx));
212
213         already_hashed_to = 0;
214
215         while (1) {
216                 prepare_to_stream(state, flags);
217                 if (idx) {
218                         sha1file_checkpoint(state->f, &checkpoint);
219                         idx->offset = state->offset;
220                         crc32_begin(state->f);
221                 }
222                 if (!stream_to_pack(state, &ctx, &already_hashed_to,
223                                     fd, size, type, path, flags))
224                         break;
225                 /*
226                  * Writing this object to the current pack will make
227                  * it too big; we need to truncate it, start a new
228                  * pack, and write into it.
229                  */
230                 if (!idx)
231                         die("BUG: should not happen");
232                 sha1file_truncate(state->f, &checkpoint);
233                 state->offset = checkpoint.offset;
234                 finish_bulk_checkin(state);
235                 if (lseek(fd, seekback, SEEK_SET) == (off_t) -1)
236                         return error("cannot seek back");
237         }
238         git_SHA1_Final(result_sha1, &ctx);
239         if (!idx)
240                 return 0;
241
242         idx->crc32 = crc32_end(state->f);
243         if (already_written(state, result_sha1)) {
244                 sha1file_truncate(state->f, &checkpoint);
245                 state->offset = checkpoint.offset;
246                 free(idx);
247         } else {
248                 hashcpy(idx->sha1, result_sha1);
249                 ALLOC_GROW(state->written,
250                            state->nr_written + 1,
251                            state->alloc_written);
252                 state->written[state->nr_written++] = idx;
253         }
254         return 0;
255 }
256
257 int index_bulk_checkin(unsigned char *sha1,
258                        int fd, size_t size, enum object_type type,
259                        const char *path, unsigned flags)
260 {
261         int status = deflate_to_pack(&state, sha1, fd, size, type,
262                                      path, flags);
263         if (!state.plugged)
264                 finish_bulk_checkin(&state);
265         return status;
266 }
267
268 void plug_bulk_checkin(void)
269 {
270         state.plugged = 1;
271 }
272
273 void unplug_bulk_checkin(void)
274 {
275         state.plugged = 0;
276         if (state.f)
277                 finish_bulk_checkin(&state);
278 }