]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Copyright (c) 2011, Google Inc. | |
3 | */ | |
4 | #include "git-compat-util.h" | |
5 | #include "bulk-checkin.h" | |
6 | #include "environment.h" | |
7 | #include "gettext.h" | |
8 | #include "hex.h" | |
9 | #include "lockfile.h" | |
10 | #include "repository.h" | |
11 | #include "csum-file.h" | |
12 | #include "pack.h" | |
13 | #include "strbuf.h" | |
14 | #include "tmp-objdir.h" | |
15 | #include "packfile.h" | |
16 | #include "object-file.h" | |
17 | #include "object-store-ll.h" | |
18 | ||
19 | static int odb_transaction_nesting; | |
20 | ||
21 | static struct tmp_objdir *bulk_fsync_objdir; | |
22 | ||
23 | static struct bulk_checkin_packfile { | |
24 | char *pack_tmp_name; | |
25 | struct hashfile *f; | |
26 | off_t offset; | |
27 | struct pack_idx_option pack_idx_opts; | |
28 | ||
29 | struct pack_idx_entry **written; | |
30 | uint32_t alloc_written; | |
31 | uint32_t nr_written; | |
32 | } bulk_checkin_packfile; | |
33 | ||
34 | static void finish_tmp_packfile(struct strbuf *basename, | |
35 | const char *pack_tmp_name, | |
36 | struct pack_idx_entry **written_list, | |
37 | uint32_t nr_written, | |
38 | struct pack_idx_option *pack_idx_opts, | |
39 | unsigned char hash[]) | |
40 | { | |
41 | char *idx_tmp_name = NULL; | |
42 | ||
43 | stage_tmp_packfiles(basename, pack_tmp_name, written_list, nr_written, | |
44 | NULL, pack_idx_opts, hash, &idx_tmp_name); | |
45 | rename_tmp_packfile_idx(basename, &idx_tmp_name); | |
46 | ||
47 | free(idx_tmp_name); | |
48 | } | |
49 | ||
50 | static void flush_bulk_checkin_packfile(struct bulk_checkin_packfile *state) | |
51 | { | |
52 | unsigned char hash[GIT_MAX_RAWSZ]; | |
53 | struct strbuf packname = STRBUF_INIT; | |
54 | int i; | |
55 | ||
56 | if (!state->f) | |
57 | return; | |
58 | ||
59 | if (state->nr_written == 0) { | |
60 | close(state->f->fd); | |
61 | unlink(state->pack_tmp_name); | |
62 | goto clear_exit; | |
63 | } else if (state->nr_written == 1) { | |
64 | finalize_hashfile(state->f, hash, FSYNC_COMPONENT_PACK, | |
65 | CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE); | |
66 | } else { | |
67 | int fd = finalize_hashfile(state->f, hash, FSYNC_COMPONENT_PACK, 0); | |
68 | fixup_pack_header_footer(fd, hash, state->pack_tmp_name, | |
69 | state->nr_written, hash, | |
70 | state->offset); | |
71 | close(fd); | |
72 | } | |
73 | ||
74 | strbuf_addf(&packname, "%s/pack/pack-%s.", get_object_directory(), | |
75 | hash_to_hex(hash)); | |
76 | finish_tmp_packfile(&packname, state->pack_tmp_name, | |
77 | state->written, state->nr_written, | |
78 | &state->pack_idx_opts, hash); | |
79 | for (i = 0; i < state->nr_written; i++) | |
80 | free(state->written[i]); | |
81 | ||
82 | clear_exit: | |
83 | free(state->written); | |
84 | memset(state, 0, sizeof(*state)); | |
85 | ||
86 | strbuf_release(&packname); | |
87 | /* Make objects we just wrote available to ourselves */ | |
88 | reprepare_packed_git(the_repository); | |
89 | } | |
90 | ||
91 | /* | |
92 | * Cleanup after batch-mode fsync_object_files. | |
93 | */ | |
94 | static void flush_batch_fsync(void) | |
95 | { | |
96 | struct strbuf temp_path = STRBUF_INIT; | |
97 | struct tempfile *temp; | |
98 | ||
99 | if (!bulk_fsync_objdir) | |
100 | return; | |
101 | ||
102 | /* | |
103 | * Issue a full hardware flush against a temporary file to ensure | |
104 | * that all objects are durable before any renames occur. The code in | |
105 | * fsync_loose_object_bulk_checkin has already issued a writeout | |
106 | * request, but it has not flushed any writeback cache in the storage | |
107 | * hardware or any filesystem logs. This fsync call acts as a barrier | |
108 | * to ensure that the data in each new object file is durable before | |
109 | * the final name is visible. | |
110 | */ | |
111 | strbuf_addf(&temp_path, "%s/bulk_fsync_XXXXXX", get_object_directory()); | |
112 | temp = xmks_tempfile(temp_path.buf); | |
113 | fsync_or_die(get_tempfile_fd(temp), get_tempfile_path(temp)); | |
114 | delete_tempfile(&temp); | |
115 | strbuf_release(&temp_path); | |
116 | ||
117 | /* | |
118 | * Make the object files visible in the primary ODB after their data is | |
119 | * fully durable. | |
120 | */ | |
121 | tmp_objdir_migrate(bulk_fsync_objdir); | |
122 | bulk_fsync_objdir = NULL; | |
123 | } | |
124 | ||
125 | static int already_written(struct bulk_checkin_packfile *state, struct object_id *oid) | |
126 | { | |
127 | int i; | |
128 | ||
129 | /* The object may already exist in the repository */ | |
130 | if (repo_has_object_file(the_repository, oid)) | |
131 | return 1; | |
132 | ||
133 | /* Might want to keep the list sorted */ | |
134 | for (i = 0; i < state->nr_written; i++) | |
135 | if (oideq(&state->written[i]->oid, oid)) | |
136 | return 1; | |
137 | ||
138 | /* This is a new object we need to keep */ | |
139 | return 0; | |
140 | } | |
141 | ||
142 | /* | |
143 | * Read the contents from fd for size bytes, streaming it to the | |
144 | * packfile in state while updating the hash in ctx. Signal a failure | |
145 | * by returning a negative value when the resulting pack would exceed | |
146 | * the pack size limit and this is not the first object in the pack, | |
147 | * so that the caller can discard what we wrote from the current pack | |
148 | * by truncating it and opening a new one. The caller will then call | |
149 | * us again after rewinding the input fd. | |
150 | * | |
151 | * The already_hashed_to pointer is kept untouched by the caller to | |
152 | * make sure we do not hash the same byte when we are called | |
153 | * again. This way, the caller does not have to checkpoint its hash | |
154 | * status before calling us just in case we ask it to call us again | |
155 | * with a new pack. | |
156 | */ | |
157 | static int stream_blob_to_pack(struct bulk_checkin_packfile *state, | |
158 | git_hash_ctx *ctx, off_t *already_hashed_to, | |
159 | int fd, size_t size, const char *path, | |
160 | unsigned flags) | |
161 | { | |
162 | git_zstream s; | |
163 | unsigned char ibuf[16384]; | |
164 | unsigned char obuf[16384]; | |
165 | unsigned hdrlen; | |
166 | int status = Z_OK; | |
167 | int write_object = (flags & HASH_WRITE_OBJECT); | |
168 | off_t offset = 0; | |
169 | ||
170 | git_deflate_init(&s, pack_compression_level); | |
171 | ||
172 | hdrlen = encode_in_pack_object_header(obuf, sizeof(obuf), OBJ_BLOB, size); | |
173 | s.next_out = obuf + hdrlen; | |
174 | s.avail_out = sizeof(obuf) - hdrlen; | |
175 | ||
176 | while (status != Z_STREAM_END) { | |
177 | if (size && !s.avail_in) { | |
178 | ssize_t rsize = size < sizeof(ibuf) ? size : sizeof(ibuf); | |
179 | ssize_t read_result = read_in_full(fd, ibuf, rsize); | |
180 | if (read_result < 0) | |
181 | die_errno("failed to read from '%s'", path); | |
182 | if (read_result != rsize) | |
183 | die("failed to read %d bytes from '%s'", | |
184 | (int)rsize, path); | |
185 | offset += rsize; | |
186 | if (*already_hashed_to < offset) { | |
187 | size_t hsize = offset - *already_hashed_to; | |
188 | if (rsize < hsize) | |
189 | hsize = rsize; | |
190 | if (hsize) | |
191 | the_hash_algo->update_fn(ctx, ibuf, hsize); | |
192 | *already_hashed_to = offset; | |
193 | } | |
194 | s.next_in = ibuf; | |
195 | s.avail_in = rsize; | |
196 | size -= rsize; | |
197 | } | |
198 | ||
199 | status = git_deflate(&s, size ? 0 : Z_FINISH); | |
200 | ||
201 | if (!s.avail_out || status == Z_STREAM_END) { | |
202 | if (write_object) { | |
203 | size_t written = s.next_out - obuf; | |
204 | ||
205 | /* would we bust the size limit? */ | |
206 | if (state->nr_written && | |
207 | pack_size_limit_cfg && | |
208 | pack_size_limit_cfg < state->offset + written) { | |
209 | git_deflate_abort(&s); | |
210 | return -1; | |
211 | } | |
212 | ||
213 | hashwrite(state->f, obuf, written); | |
214 | state->offset += written; | |
215 | } | |
216 | s.next_out = obuf; | |
217 | s.avail_out = sizeof(obuf); | |
218 | } | |
219 | ||
220 | switch (status) { | |
221 | case Z_OK: | |
222 | case Z_BUF_ERROR: | |
223 | case Z_STREAM_END: | |
224 | continue; | |
225 | default: | |
226 | die("unexpected deflate failure: %d", status); | |
227 | } | |
228 | } | |
229 | git_deflate_end(&s); | |
230 | return 0; | |
231 | } | |
232 | ||
233 | /* Lazily create backing packfile for the state */ | |
234 | static void prepare_to_stream(struct bulk_checkin_packfile *state, | |
235 | unsigned flags) | |
236 | { | |
237 | if (!(flags & HASH_WRITE_OBJECT) || state->f) | |
238 | return; | |
239 | ||
240 | state->f = create_tmp_packfile(&state->pack_tmp_name); | |
241 | reset_pack_idx_option(&state->pack_idx_opts); | |
242 | ||
243 | /* Pretend we are going to write only one object */ | |
244 | state->offset = write_pack_header(state->f, 1); | |
245 | if (!state->offset) | |
246 | die_errno("unable to write pack header"); | |
247 | } | |
248 | ||
249 | static int deflate_blob_to_pack(struct bulk_checkin_packfile *state, | |
250 | struct object_id *result_oid, | |
251 | int fd, size_t size, | |
252 | const char *path, unsigned flags) | |
253 | { | |
254 | off_t seekback, already_hashed_to; | |
255 | git_hash_ctx ctx; | |
256 | unsigned char obuf[16384]; | |
257 | unsigned header_len; | |
258 | struct hashfile_checkpoint checkpoint = {0}; | |
259 | struct pack_idx_entry *idx = NULL; | |
260 | ||
261 | seekback = lseek(fd, 0, SEEK_CUR); | |
262 | if (seekback == (off_t) -1) | |
263 | return error("cannot find the current offset"); | |
264 | ||
265 | header_len = format_object_header((char *)obuf, sizeof(obuf), | |
266 | OBJ_BLOB, size); | |
267 | the_hash_algo->init_fn(&ctx); | |
268 | the_hash_algo->update_fn(&ctx, obuf, header_len); | |
269 | the_hash_algo->init_fn(&checkpoint.ctx); | |
270 | ||
271 | /* Note: idx is non-NULL when we are writing */ | |
272 | if ((flags & HASH_WRITE_OBJECT) != 0) | |
273 | CALLOC_ARRAY(idx, 1); | |
274 | ||
275 | already_hashed_to = 0; | |
276 | ||
277 | while (1) { | |
278 | prepare_to_stream(state, flags); | |
279 | if (idx) { | |
280 | hashfile_checkpoint(state->f, &checkpoint); | |
281 | idx->offset = state->offset; | |
282 | crc32_begin(state->f); | |
283 | } | |
284 | if (!stream_blob_to_pack(state, &ctx, &already_hashed_to, | |
285 | fd, size, path, flags)) | |
286 | break; | |
287 | /* | |
288 | * Writing this object to the current pack will make | |
289 | * it too big; we need to truncate it, start a new | |
290 | * pack, and write into it. | |
291 | */ | |
292 | if (!idx) | |
293 | BUG("should not happen"); | |
294 | hashfile_truncate(state->f, &checkpoint); | |
295 | state->offset = checkpoint.offset; | |
296 | flush_bulk_checkin_packfile(state); | |
297 | if (lseek(fd, seekback, SEEK_SET) == (off_t) -1) | |
298 | return error("cannot seek back"); | |
299 | } | |
300 | the_hash_algo->final_oid_fn(result_oid, &ctx); | |
301 | if (!idx) | |
302 | return 0; | |
303 | ||
304 | idx->crc32 = crc32_end(state->f); | |
305 | if (already_written(state, result_oid)) { | |
306 | hashfile_truncate(state->f, &checkpoint); | |
307 | state->offset = checkpoint.offset; | |
308 | free(idx); | |
309 | } else { | |
310 | oidcpy(&idx->oid, result_oid); | |
311 | ALLOC_GROW(state->written, | |
312 | state->nr_written + 1, | |
313 | state->alloc_written); | |
314 | state->written[state->nr_written++] = idx; | |
315 | } | |
316 | return 0; | |
317 | } | |
318 | ||
319 | void prepare_loose_object_bulk_checkin(void) | |
320 | { | |
321 | /* | |
322 | * We lazily create the temporary object directory | |
323 | * the first time an object might be added, since | |
324 | * callers may not know whether any objects will be | |
325 | * added at the time they call begin_odb_transaction. | |
326 | */ | |
327 | if (!odb_transaction_nesting || bulk_fsync_objdir) | |
328 | return; | |
329 | ||
330 | bulk_fsync_objdir = tmp_objdir_create("bulk-fsync"); | |
331 | if (bulk_fsync_objdir) | |
332 | tmp_objdir_replace_primary_odb(bulk_fsync_objdir, 0); | |
333 | } | |
334 | ||
335 | void fsync_loose_object_bulk_checkin(int fd, const char *filename) | |
336 | { | |
337 | /* | |
338 | * If we have an active ODB transaction, we issue a call that | |
339 | * cleans the filesystem page cache but avoids a hardware flush | |
340 | * command. Later on we will issue a single hardware flush | |
341 | * before renaming the objects to their final names as part of | |
342 | * flush_batch_fsync. | |
343 | */ | |
344 | if (!bulk_fsync_objdir || | |
345 | git_fsync(fd, FSYNC_WRITEOUT_ONLY) < 0) { | |
346 | if (errno == ENOSYS) | |
347 | warning(_("core.fsyncMethod = batch is unsupported on this platform")); | |
348 | fsync_or_die(fd, filename); | |
349 | } | |
350 | } | |
351 | ||
352 | int index_blob_bulk_checkin(struct object_id *oid, | |
353 | int fd, size_t size, | |
354 | const char *path, unsigned flags) | |
355 | { | |
356 | int status = deflate_blob_to_pack(&bulk_checkin_packfile, oid, fd, size, | |
357 | path, flags); | |
358 | if (!odb_transaction_nesting) | |
359 | flush_bulk_checkin_packfile(&bulk_checkin_packfile); | |
360 | return status; | |
361 | } | |
362 | ||
363 | void begin_odb_transaction(void) | |
364 | { | |
365 | odb_transaction_nesting += 1; | |
366 | } | |
367 | ||
368 | void flush_odb_transaction(void) | |
369 | { | |
370 | flush_batch_fsync(); | |
371 | flush_bulk_checkin_packfile(&bulk_checkin_packfile); | |
372 | } | |
373 | ||
374 | void end_odb_transaction(void) | |
375 | { | |
376 | odb_transaction_nesting -= 1; | |
377 | if (odb_transaction_nesting < 0) | |
378 | BUG("Unbalanced ODB transaction nesting"); | |
379 | ||
380 | if (odb_transaction_nesting) | |
381 | return; | |
382 | ||
383 | flush_odb_transaction(); | |
384 | } |