]>
Commit | Line | Data |
---|---|---|
cec736d2 LP |
1 | /*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/ |
2 | ||
3 | /*** | |
4 | This file is part of systemd. | |
5 | ||
6 | Copyright 2011 Lennart Poettering | |
7 | ||
8 | systemd is free software; you can redistribute it and/or modify it | |
5430f7f2 LP |
9 | under the terms of the GNU Lesser General Public License as published by |
10 | the Free Software Foundation; either version 2.1 of the License, or | |
cec736d2 LP |
11 | (at your option) any later version. |
12 | ||
13 | systemd is distributed in the hope that it will be useful, but | |
14 | WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
5430f7f2 | 16 | Lesser General Public License for more details. |
cec736d2 | 17 | |
5430f7f2 | 18 | You should have received a copy of the GNU Lesser General Public License |
cec736d2 LP |
19 | along with systemd; If not, see <http://www.gnu.org/licenses/>. |
20 | ***/ | |
21 | ||
22 | #include <sys/mman.h> | |
23 | #include <errno.h> | |
24 | #include <sys/uio.h> | |
25 | #include <unistd.h> | |
26 | #include <sys/statvfs.h> | |
27 | #include <fcntl.h> | |
28 | #include <stddef.h> | |
d2edfae0 | 29 | #include <sys/xattr.h> |
fb0951b0 | 30 | |
cec736d2 LP |
31 | #include "journal-def.h" |
32 | #include "journal-file.h" | |
0284adc6 | 33 | #include "journal-authenticate.h" |
cec736d2 | 34 | #include "lookup3.h" |
807e17f0 | 35 | #include "compress.h" |
7560fffc | 36 | #include "fsprg.h" |
cec736d2 | 37 | |
4a92baf3 LP |
38 | #define DEFAULT_DATA_HASH_TABLE_SIZE (2047ULL*sizeof(HashItem)) |
39 | #define DEFAULT_FIELD_HASH_TABLE_SIZE (333ULL*sizeof(HashItem)) | |
cec736d2 | 40 | |
be19b7df | 41 | #define COMPRESSION_SIZE_THRESHOLD (512ULL) |
807e17f0 | 42 | |
babfc091 | 43 | /* This is the minimum journal file size */ |
253f59df | 44 | #define JOURNAL_FILE_SIZE_MIN (4ULL*1024ULL*1024ULL) /* 4 MiB */ |
babfc091 LP |
45 | |
46 | /* These are the lower and upper bounds if we deduce the max_use value | |
47 | * from the file system size */ | |
48 | #define DEFAULT_MAX_USE_LOWER (1ULL*1024ULL*1024ULL) /* 1 MiB */ | |
49 | #define DEFAULT_MAX_USE_UPPER (4ULL*1024ULL*1024ULL*1024ULL) /* 4 GiB */ | |
50 | ||
51 | /* This is the upper bound if we deduce max_size from max_use */ | |
71100051 | 52 | #define DEFAULT_MAX_SIZE_UPPER (128ULL*1024ULL*1024ULL) /* 128 MiB */ |
babfc091 LP |
53 | |
54 | /* This is the upper bound if we deduce the keep_free value from the | |
55 | * file system size */ | |
56 | #define DEFAULT_KEEP_FREE_UPPER (4ULL*1024ULL*1024ULL*1024ULL) /* 4 GiB */ | |
57 | ||
58 | /* This is the keep_free value when we can't determine the system | |
59 | * size */ | |
60 | #define DEFAULT_KEEP_FREE (1024ULL*1024ULL) /* 1 MB */ | |
61 | ||
dca6219e LP |
62 | /* n_data was the first entry we added after the initial file format design */ |
63 | #define HEADER_SIZE_MIN ALIGN64(offsetof(Header, n_data)) | |
cec736d2 | 64 | |
a4bcff5b LP |
65 | /* How many entries to keep in the entry array chain cache at max */ |
66 | #define CHAIN_CACHE_MAX 20 | |
67 | ||
a676e665 LP |
68 | /* How much to increase the journal file size at once each time we allocate something new. */ |
69 | #define FILE_SIZE_INCREASE (8ULL*1024ULL*1024ULL) /* 8MB */ | |
70 | ||
9588bc32 | 71 | static int journal_file_set_online(JournalFile *f) { |
26687bf8 OS |
72 | assert(f); |
73 | ||
74 | if (!f->writable) | |
75 | return -EPERM; | |
76 | ||
77 | if (!(f->fd >= 0 && f->header)) | |
78 | return -EINVAL; | |
79 | ||
80 | switch(f->header->state) { | |
81 | case STATE_ONLINE: | |
82 | return 0; | |
83 | ||
84 | case STATE_OFFLINE: | |
85 | f->header->state = STATE_ONLINE; | |
86 | fsync(f->fd); | |
87 | return 0; | |
88 | ||
89 | default: | |
90 | return -EINVAL; | |
91 | } | |
92 | } | |
93 | ||
94 | int journal_file_set_offline(JournalFile *f) { | |
95 | assert(f); | |
96 | ||
97 | if (!f->writable) | |
98 | return -EPERM; | |
99 | ||
100 | if (!(f->fd >= 0 && f->header)) | |
101 | return -EINVAL; | |
102 | ||
103 | if (f->header->state != STATE_ONLINE) | |
104 | return 0; | |
105 | ||
106 | fsync(f->fd); | |
107 | ||
108 | f->header->state = STATE_OFFLINE; | |
109 | ||
110 | fsync(f->fd); | |
111 | ||
112 | return 0; | |
113 | } | |
114 | ||
cec736d2 | 115 | void journal_file_close(JournalFile *f) { |
de190aef | 116 | assert(f); |
cec736d2 | 117 | |
feb12d3e | 118 | #ifdef HAVE_GCRYPT |
b0af6f41 | 119 | /* Write the final tag */ |
c586dbf1 | 120 | if (f->seal && f->writable) |
b0af6f41 | 121 | journal_file_append_tag(f); |
feb12d3e | 122 | #endif |
b0af6f41 | 123 | |
7560fffc | 124 | /* Sync everything to disk, before we mark the file offline */ |
16e9f408 LP |
125 | if (f->mmap && f->fd >= 0) |
126 | mmap_cache_close_fd(f->mmap, f->fd); | |
7560fffc | 127 | |
26687bf8 | 128 | journal_file_set_offline(f); |
cec736d2 | 129 | |
26687bf8 | 130 | if (f->header) |
d384c7a8 | 131 | munmap(f->header, PAGE_ALIGN(sizeof(Header))); |
cec736d2 | 132 | |
03e334a1 | 133 | safe_close(f->fd); |
cec736d2 | 134 | free(f->path); |
807e17f0 | 135 | |
16e9f408 LP |
136 | if (f->mmap) |
137 | mmap_cache_unref(f->mmap); | |
138 | ||
4743015d | 139 | ordered_hashmap_free_free(f->chain_cache); |
a4bcff5b | 140 | |
d89c8fdf | 141 | #if defined(HAVE_XZ) || defined(HAVE_LZ4) |
807e17f0 LP |
142 | free(f->compress_buffer); |
143 | #endif | |
144 | ||
7560fffc | 145 | #ifdef HAVE_GCRYPT |
baed47c3 LP |
146 | if (f->fss_file) |
147 | munmap(f->fss_file, PAGE_ALIGN(f->fss_file_size)); | |
b7c9ae91 LP |
148 | else if (f->fsprg_state) |
149 | free(f->fsprg_state); | |
150 | ||
151 | free(f->fsprg_seed); | |
7560fffc LP |
152 | |
153 | if (f->hmac) | |
154 | gcry_md_close(f->hmac); | |
155 | #endif | |
156 | ||
cec736d2 LP |
157 | free(f); |
158 | } | |
159 | ||
0ac38b70 | 160 | static int journal_file_init_header(JournalFile *f, JournalFile *template) { |
d89c8fdf | 161 | Header h = {}; |
cec736d2 LP |
162 | ssize_t k; |
163 | int r; | |
164 | ||
165 | assert(f); | |
166 | ||
7560fffc | 167 | memcpy(h.signature, HEADER_SIGNATURE, 8); |
23b0b2b2 | 168 | h.header_size = htole64(ALIGN64(sizeof(h))); |
cec736d2 | 169 | |
d89c8fdf ZJS |
170 | h.incompatible_flags |= htole32( |
171 | f->compress_xz * HEADER_INCOMPATIBLE_COMPRESSED_XZ | | |
172 | f->compress_lz4 * HEADER_INCOMPATIBLE_COMPRESSED_LZ4); | |
7560fffc | 173 | |
d89c8fdf ZJS |
174 | h.compatible_flags = htole32( |
175 | f->seal * HEADER_COMPATIBLE_SEALED); | |
7560fffc | 176 | |
cec736d2 LP |
177 | r = sd_id128_randomize(&h.file_id); |
178 | if (r < 0) | |
179 | return r; | |
180 | ||
0ac38b70 LP |
181 | if (template) { |
182 | h.seqnum_id = template->header->seqnum_id; | |
beec0085 | 183 | h.tail_entry_seqnum = template->header->tail_entry_seqnum; |
0ac38b70 LP |
184 | } else |
185 | h.seqnum_id = h.file_id; | |
cec736d2 LP |
186 | |
187 | k = pwrite(f->fd, &h, sizeof(h), 0); | |
188 | if (k < 0) | |
189 | return -errno; | |
190 | ||
191 | if (k != sizeof(h)) | |
192 | return -EIO; | |
193 | ||
194 | return 0; | |
195 | } | |
196 | ||
197 | static int journal_file_refresh_header(JournalFile *f) { | |
198 | int r; | |
de190aef | 199 | sd_id128_t boot_id; |
cec736d2 LP |
200 | |
201 | assert(f); | |
202 | ||
203 | r = sd_id128_get_machine(&f->header->machine_id); | |
204 | if (r < 0) | |
205 | return r; | |
206 | ||
de190aef | 207 | r = sd_id128_get_boot(&boot_id); |
cec736d2 LP |
208 | if (r < 0) |
209 | return r; | |
210 | ||
de190aef LP |
211 | if (sd_id128_equal(boot_id, f->header->boot_id)) |
212 | f->tail_entry_monotonic_valid = true; | |
213 | ||
214 | f->header->boot_id = boot_id; | |
215 | ||
26687bf8 | 216 | journal_file_set_online(f); |
b788cc23 | 217 | |
7560fffc | 218 | /* Sync the online state to disk */ |
a676e665 | 219 | fsync(f->fd); |
b788cc23 | 220 | |
cec736d2 LP |
221 | return 0; |
222 | } | |
223 | ||
224 | static int journal_file_verify_header(JournalFile *f) { | |
d89c8fdf ZJS |
225 | uint32_t flags; |
226 | ||
cec736d2 LP |
227 | assert(f); |
228 | ||
7560fffc | 229 | if (memcmp(f->header->signature, HEADER_SIGNATURE, 8)) |
cec736d2 LP |
230 | return -EBADMSG; |
231 | ||
7560fffc LP |
232 | /* In both read and write mode we refuse to open files with |
233 | * incompatible flags we don't know */ | |
d89c8fdf ZJS |
234 | flags = le32toh(f->header->incompatible_flags); |
235 | if (flags & ~HEADER_INCOMPATIBLE_SUPPORTED) { | |
236 | if (flags & ~HEADER_INCOMPATIBLE_ANY) | |
237 | log_debug("Journal file %s has unknown incompatible flags %"PRIx32, | |
238 | f->path, flags & ~HEADER_INCOMPATIBLE_ANY); | |
239 | flags = (flags & HEADER_INCOMPATIBLE_ANY) & ~HEADER_INCOMPATIBLE_SUPPORTED; | |
240 | if (flags) | |
241 | log_debug("Journal file %s uses incompatible flags %"PRIx32 | |
242 | " disabled at compilation time.", f->path, flags); | |
cec736d2 | 243 | return -EPROTONOSUPPORT; |
d89c8fdf | 244 | } |
cec736d2 | 245 | |
7560fffc LP |
246 | /* When open for writing we refuse to open files with |
247 | * compatible flags, too */ | |
d89c8fdf ZJS |
248 | flags = le32toh(f->header->compatible_flags); |
249 | if (f->writable && (flags & ~HEADER_COMPATIBLE_SUPPORTED)) { | |
250 | if (flags & ~HEADER_COMPATIBLE_ANY) | |
251 | log_debug("Journal file %s has unknown compatible flags %"PRIx32, | |
252 | f->path, flags & ~HEADER_COMPATIBLE_ANY); | |
253 | flags = (flags & HEADER_COMPATIBLE_ANY) & ~HEADER_COMPATIBLE_SUPPORTED; | |
254 | if (flags) | |
255 | log_debug("Journal file %s uses compatible flags %"PRIx32 | |
256 | " disabled at compilation time.", f->path, flags); | |
257 | return -EPROTONOSUPPORT; | |
7560fffc LP |
258 | } |
259 | ||
db11ac1a LP |
260 | if (f->header->state >= _STATE_MAX) |
261 | return -EBADMSG; | |
262 | ||
dca6219e LP |
263 | /* The first addition was n_data, so check that we are at least this large */ |
264 | if (le64toh(f->header->header_size) < HEADER_SIZE_MIN) | |
23b0b2b2 LP |
265 | return -EBADMSG; |
266 | ||
8088cbd3 | 267 | if (JOURNAL_HEADER_SEALED(f->header) && !JOURNAL_HEADER_CONTAINS(f->header, n_entry_arrays)) |
beec0085 LP |
268 | return -EBADMSG; |
269 | ||
db11ac1a LP |
270 | if ((le64toh(f->header->header_size) + le64toh(f->header->arena_size)) > (uint64_t) f->last_stat.st_size) |
271 | return -ENODATA; | |
272 | ||
273 | if (le64toh(f->header->tail_object_offset) > (le64toh(f->header->header_size) + le64toh(f->header->arena_size))) | |
274 | return -ENODATA; | |
275 | ||
7762e02b LP |
276 | if (!VALID64(le64toh(f->header->data_hash_table_offset)) || |
277 | !VALID64(le64toh(f->header->field_hash_table_offset)) || | |
278 | !VALID64(le64toh(f->header->tail_object_offset)) || | |
279 | !VALID64(le64toh(f->header->entry_array_offset))) | |
280 | return -ENODATA; | |
281 | ||
cec736d2 | 282 | if (f->writable) { |
ccdbaf91 | 283 | uint8_t state; |
cec736d2 LP |
284 | sd_id128_t machine_id; |
285 | int r; | |
286 | ||
287 | r = sd_id128_get_machine(&machine_id); | |
288 | if (r < 0) | |
289 | return r; | |
290 | ||
291 | if (!sd_id128_equal(machine_id, f->header->machine_id)) | |
292 | return -EHOSTDOWN; | |
293 | ||
de190aef | 294 | state = f->header->state; |
cec736d2 | 295 | |
71fa6f00 LP |
296 | if (state == STATE_ONLINE) { |
297 | log_debug("Journal file %s is already online. Assuming unclean closing.", f->path); | |
298 | return -EBUSY; | |
299 | } else if (state == STATE_ARCHIVED) | |
cec736d2 | 300 | return -ESHUTDOWN; |
71fa6f00 LP |
301 | else if (state != STATE_OFFLINE) { |
302 | log_debug("Journal file %s has unknown state %u.", f->path, state); | |
303 | return -EBUSY; | |
304 | } | |
cec736d2 LP |
305 | } |
306 | ||
d89c8fdf ZJS |
307 | f->compress_xz = JOURNAL_HEADER_COMPRESSED_XZ(f->header); |
308 | f->compress_lz4 = JOURNAL_HEADER_COMPRESSED_LZ4(f->header); | |
c586dbf1 | 309 | |
f1889c91 | 310 | f->seal = JOURNAL_HEADER_SEALED(f->header); |
7560fffc | 311 | |
cec736d2 LP |
312 | return 0; |
313 | } | |
314 | ||
315 | static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size) { | |
eda4b58b | 316 | uint64_t old_size, new_size; |
fec2aa2f | 317 | int r; |
cec736d2 LP |
318 | |
319 | assert(f); | |
320 | ||
cec736d2 | 321 | /* We assume that this file is not sparse, and we know that |
38ac38b2 | 322 | * for sure, since we always call posix_fallocate() |
cec736d2 LP |
323 | * ourselves */ |
324 | ||
325 | old_size = | |
23b0b2b2 | 326 | le64toh(f->header->header_size) + |
cec736d2 LP |
327 | le64toh(f->header->arena_size); |
328 | ||
bc85bfee | 329 | new_size = PAGE_ALIGN(offset + size); |
23b0b2b2 LP |
330 | if (new_size < le64toh(f->header->header_size)) |
331 | new_size = le64toh(f->header->header_size); | |
bc85bfee LP |
332 | |
333 | if (new_size <= old_size) | |
cec736d2 LP |
334 | return 0; |
335 | ||
a676e665 | 336 | if (f->metrics.max_size > 0 && new_size > f->metrics.max_size) |
bc85bfee | 337 | return -E2BIG; |
cec736d2 | 338 | |
a676e665 | 339 | if (new_size > f->metrics.min_size && f->metrics.keep_free > 0) { |
cec736d2 LP |
340 | struct statvfs svfs; |
341 | ||
342 | if (fstatvfs(f->fd, &svfs) >= 0) { | |
343 | uint64_t available; | |
344 | ||
345 | available = svfs.f_bfree * svfs.f_bsize; | |
346 | ||
bc85bfee LP |
347 | if (available >= f->metrics.keep_free) |
348 | available -= f->metrics.keep_free; | |
cec736d2 LP |
349 | else |
350 | available = 0; | |
351 | ||
352 | if (new_size - old_size > available) | |
353 | return -E2BIG; | |
354 | } | |
355 | } | |
356 | ||
eda4b58b LP |
357 | /* Increase by larger blocks at once */ |
358 | new_size = ((new_size+FILE_SIZE_INCREASE-1) / FILE_SIZE_INCREASE) * FILE_SIZE_INCREASE; | |
359 | if (f->metrics.max_size > 0 && new_size > f->metrics.max_size) | |
360 | new_size = f->metrics.max_size; | |
361 | ||
bc85bfee LP |
362 | /* Note that the glibc fallocate() fallback is very |
363 | inefficient, hence we try to minimize the allocation area | |
364 | as we can. */ | |
fec2aa2f GV |
365 | r = posix_fallocate(f->fd, old_size, new_size - old_size); |
366 | if (r != 0) | |
367 | return -r; | |
cec736d2 | 368 | |
eda4b58b LP |
369 | if (fstat(f->fd, &f->last_stat) < 0) |
370 | return -errno; | |
cec736d2 | 371 | |
23b0b2b2 | 372 | f->header->arena_size = htole64(new_size - le64toh(f->header->header_size)); |
cec736d2 LP |
373 | |
374 | return 0; | |
375 | } | |
376 | ||
78519831 | 377 | static unsigned type_to_context(ObjectType type) { |
d3d3208f | 378 | /* One context for each type, plus one catch-all for the rest */ |
69adae51 | 379 | assert_cc(_OBJECT_TYPE_MAX <= MMAP_CACHE_MAX_CONTEXTS); |
d05089d8 | 380 | return type > OBJECT_UNUSED && type < _OBJECT_TYPE_MAX ? type : 0; |
d3d3208f MS |
381 | } |
382 | ||
7a9dabea | 383 | static int journal_file_move_to(JournalFile *f, ObjectType type, bool keep_always, uint64_t offset, uint64_t size, void **ret) { |
cec736d2 | 384 | assert(f); |
cec736d2 LP |
385 | assert(ret); |
386 | ||
7762e02b LP |
387 | if (size <= 0) |
388 | return -EINVAL; | |
389 | ||
2a59ea54 | 390 | /* Avoid SIGBUS on invalid accesses */ |
4bbdcdb3 LP |
391 | if (offset + size > (uint64_t) f->last_stat.st_size) { |
392 | /* Hmm, out of range? Let's refresh the fstat() data | |
393 | * first, before we trust that check. */ | |
394 | ||
395 | if (fstat(f->fd, &f->last_stat) < 0 || | |
396 | offset + size > (uint64_t) f->last_stat.st_size) | |
397 | return -EADDRNOTAVAIL; | |
398 | } | |
399 | ||
7a9dabea | 400 | return mmap_cache_get(f->mmap, f->fd, f->prot, type_to_context(type), keep_always, offset, size, &f->last_stat, ret); |
cec736d2 LP |
401 | } |
402 | ||
16e9f408 LP |
403 | static uint64_t minimum_header_size(Object *o) { |
404 | ||
b8e891e6 | 405 | static const uint64_t table[] = { |
16e9f408 LP |
406 | [OBJECT_DATA] = sizeof(DataObject), |
407 | [OBJECT_FIELD] = sizeof(FieldObject), | |
408 | [OBJECT_ENTRY] = sizeof(EntryObject), | |
409 | [OBJECT_DATA_HASH_TABLE] = sizeof(HashTableObject), | |
410 | [OBJECT_FIELD_HASH_TABLE] = sizeof(HashTableObject), | |
411 | [OBJECT_ENTRY_ARRAY] = sizeof(EntryArrayObject), | |
412 | [OBJECT_TAG] = sizeof(TagObject), | |
413 | }; | |
414 | ||
415 | if (o->object.type >= ELEMENTSOF(table) || table[o->object.type] <= 0) | |
416 | return sizeof(ObjectHeader); | |
417 | ||
418 | return table[o->object.type]; | |
419 | } | |
420 | ||
78519831 | 421 | int journal_file_move_to_object(JournalFile *f, ObjectType type, uint64_t offset, Object **ret) { |
cec736d2 LP |
422 | int r; |
423 | void *t; | |
424 | Object *o; | |
425 | uint64_t s; | |
426 | ||
427 | assert(f); | |
428 | assert(ret); | |
429 | ||
db11ac1a LP |
430 | /* Objects may only be located at multiple of 64 bit */ |
431 | if (!VALID64(offset)) | |
432 | return -EFAULT; | |
433 | ||
7a9dabea | 434 | r = journal_file_move_to(f, type, false, offset, sizeof(ObjectHeader), &t); |
cec736d2 LP |
435 | if (r < 0) |
436 | return r; | |
437 | ||
438 | o = (Object*) t; | |
439 | s = le64toh(o->object.size); | |
440 | ||
441 | if (s < sizeof(ObjectHeader)) | |
442 | return -EBADMSG; | |
443 | ||
16e9f408 LP |
444 | if (o->object.type <= OBJECT_UNUSED) |
445 | return -EBADMSG; | |
446 | ||
447 | if (s < minimum_header_size(o)) | |
448 | return -EBADMSG; | |
449 | ||
d05089d8 | 450 | if (type > OBJECT_UNUSED && o->object.type != type) |
cec736d2 LP |
451 | return -EBADMSG; |
452 | ||
453 | if (s > sizeof(ObjectHeader)) { | |
7a9dabea | 454 | r = journal_file_move_to(f, type, false, offset, s, &t); |
cec736d2 LP |
455 | if (r < 0) |
456 | return r; | |
457 | ||
458 | o = (Object*) t; | |
459 | } | |
460 | ||
cec736d2 LP |
461 | *ret = o; |
462 | return 0; | |
463 | } | |
464 | ||
d98cc1f2 | 465 | static uint64_t journal_file_entry_seqnum(JournalFile *f, uint64_t *seqnum) { |
cec736d2 LP |
466 | uint64_t r; |
467 | ||
468 | assert(f); | |
469 | ||
beec0085 | 470 | r = le64toh(f->header->tail_entry_seqnum) + 1; |
c2373f84 LP |
471 | |
472 | if (seqnum) { | |
de190aef | 473 | /* If an external seqnum counter was passed, we update |
c2373f84 LP |
474 | * both the local and the external one, and set it to |
475 | * the maximum of both */ | |
476 | ||
477 | if (*seqnum + 1 > r) | |
478 | r = *seqnum + 1; | |
479 | ||
480 | *seqnum = r; | |
481 | } | |
482 | ||
beec0085 | 483 | f->header->tail_entry_seqnum = htole64(r); |
cec736d2 | 484 | |
beec0085 LP |
485 | if (f->header->head_entry_seqnum == 0) |
486 | f->header->head_entry_seqnum = htole64(r); | |
de190aef | 487 | |
cec736d2 LP |
488 | return r; |
489 | } | |
490 | ||
78519831 | 491 | int journal_file_append_object(JournalFile *f, ObjectType type, uint64_t size, Object **ret, uint64_t *offset) { |
cec736d2 LP |
492 | int r; |
493 | uint64_t p; | |
494 | Object *tail, *o; | |
495 | void *t; | |
496 | ||
497 | assert(f); | |
d05089d8 | 498 | assert(type > OBJECT_UNUSED && type < _OBJECT_TYPE_MAX); |
cec736d2 LP |
499 | assert(size >= sizeof(ObjectHeader)); |
500 | assert(offset); | |
501 | assert(ret); | |
502 | ||
26687bf8 OS |
503 | r = journal_file_set_online(f); |
504 | if (r < 0) | |
505 | return r; | |
506 | ||
cec736d2 | 507 | p = le64toh(f->header->tail_object_offset); |
cec736d2 | 508 | if (p == 0) |
23b0b2b2 | 509 | p = le64toh(f->header->header_size); |
cec736d2 | 510 | else { |
d05089d8 | 511 | r = journal_file_move_to_object(f, OBJECT_UNUSED, p, &tail); |
cec736d2 LP |
512 | if (r < 0) |
513 | return r; | |
514 | ||
515 | p += ALIGN64(le64toh(tail->object.size)); | |
516 | } | |
517 | ||
518 | r = journal_file_allocate(f, p, size); | |
519 | if (r < 0) | |
520 | return r; | |
521 | ||
fcde2389 | 522 | r = journal_file_move_to(f, type, false, p, size, &t); |
cec736d2 LP |
523 | if (r < 0) |
524 | return r; | |
525 | ||
526 | o = (Object*) t; | |
527 | ||
528 | zero(o->object); | |
de190aef | 529 | o->object.type = type; |
cec736d2 LP |
530 | o->object.size = htole64(size); |
531 | ||
532 | f->header->tail_object_offset = htole64(p); | |
cec736d2 LP |
533 | f->header->n_objects = htole64(le64toh(f->header->n_objects) + 1); |
534 | ||
535 | *ret = o; | |
536 | *offset = p; | |
537 | ||
538 | return 0; | |
539 | } | |
540 | ||
de190aef | 541 | static int journal_file_setup_data_hash_table(JournalFile *f) { |
cec736d2 LP |
542 | uint64_t s, p; |
543 | Object *o; | |
544 | int r; | |
545 | ||
546 | assert(f); | |
547 | ||
dfabe643 | 548 | /* We estimate that we need 1 hash table entry per 768 of |
4a92baf3 LP |
549 | journal file and we want to make sure we never get beyond |
550 | 75% fill level. Calculate the hash table size for the | |
551 | maximum file size based on these metrics. */ | |
552 | ||
dfabe643 | 553 | s = (f->metrics.max_size * 4 / 768 / 3) * sizeof(HashItem); |
4a92baf3 LP |
554 | if (s < DEFAULT_DATA_HASH_TABLE_SIZE) |
555 | s = DEFAULT_DATA_HASH_TABLE_SIZE; | |
556 | ||
507f22bd | 557 | log_debug("Reserving %"PRIu64" entries in hash table.", s / sizeof(HashItem)); |
4a92baf3 | 558 | |
de190aef LP |
559 | r = journal_file_append_object(f, |
560 | OBJECT_DATA_HASH_TABLE, | |
561 | offsetof(Object, hash_table.items) + s, | |
562 | &o, &p); | |
cec736d2 LP |
563 | if (r < 0) |
564 | return r; | |
565 | ||
29804cc1 | 566 | memzero(o->hash_table.items, s); |
cec736d2 | 567 | |
de190aef LP |
568 | f->header->data_hash_table_offset = htole64(p + offsetof(Object, hash_table.items)); |
569 | f->header->data_hash_table_size = htole64(s); | |
cec736d2 LP |
570 | |
571 | return 0; | |
572 | } | |
573 | ||
de190aef | 574 | static int journal_file_setup_field_hash_table(JournalFile *f) { |
cec736d2 LP |
575 | uint64_t s, p; |
576 | Object *o; | |
577 | int r; | |
578 | ||
579 | assert(f); | |
580 | ||
3c1668da LP |
581 | /* We use a fixed size hash table for the fields as this |
582 | * number should grow very slowly only */ | |
583 | ||
de190aef LP |
584 | s = DEFAULT_FIELD_HASH_TABLE_SIZE; |
585 | r = journal_file_append_object(f, | |
586 | OBJECT_FIELD_HASH_TABLE, | |
587 | offsetof(Object, hash_table.items) + s, | |
588 | &o, &p); | |
cec736d2 LP |
589 | if (r < 0) |
590 | return r; | |
591 | ||
29804cc1 | 592 | memzero(o->hash_table.items, s); |
cec736d2 | 593 | |
de190aef LP |
594 | f->header->field_hash_table_offset = htole64(p + offsetof(Object, hash_table.items)); |
595 | f->header->field_hash_table_size = htole64(s); | |
cec736d2 LP |
596 | |
597 | return 0; | |
598 | } | |
599 | ||
de190aef | 600 | static int journal_file_map_data_hash_table(JournalFile *f) { |
cec736d2 LP |
601 | uint64_t s, p; |
602 | void *t; | |
603 | int r; | |
604 | ||
605 | assert(f); | |
606 | ||
de190aef LP |
607 | p = le64toh(f->header->data_hash_table_offset); |
608 | s = le64toh(f->header->data_hash_table_size); | |
cec736d2 | 609 | |
de190aef | 610 | r = journal_file_move_to(f, |
16e9f408 | 611 | OBJECT_DATA_HASH_TABLE, |
fcde2389 | 612 | true, |
de190aef LP |
613 | p, s, |
614 | &t); | |
cec736d2 LP |
615 | if (r < 0) |
616 | return r; | |
617 | ||
de190aef | 618 | f->data_hash_table = t; |
cec736d2 LP |
619 | return 0; |
620 | } | |
621 | ||
de190aef | 622 | static int journal_file_map_field_hash_table(JournalFile *f) { |
cec736d2 LP |
623 | uint64_t s, p; |
624 | void *t; | |
625 | int r; | |
626 | ||
627 | assert(f); | |
628 | ||
de190aef LP |
629 | p = le64toh(f->header->field_hash_table_offset); |
630 | s = le64toh(f->header->field_hash_table_size); | |
cec736d2 | 631 | |
de190aef | 632 | r = journal_file_move_to(f, |
16e9f408 | 633 | OBJECT_FIELD_HASH_TABLE, |
fcde2389 | 634 | true, |
de190aef LP |
635 | p, s, |
636 | &t); | |
cec736d2 LP |
637 | if (r < 0) |
638 | return r; | |
639 | ||
de190aef | 640 | f->field_hash_table = t; |
cec736d2 LP |
641 | return 0; |
642 | } | |
643 | ||
3c1668da LP |
644 | static int journal_file_link_field( |
645 | JournalFile *f, | |
646 | Object *o, | |
647 | uint64_t offset, | |
648 | uint64_t hash) { | |
649 | ||
650 | uint64_t p, h; | |
651 | int r; | |
652 | ||
653 | assert(f); | |
654 | assert(o); | |
655 | assert(offset > 0); | |
656 | ||
657 | if (o->object.type != OBJECT_FIELD) | |
658 | return -EINVAL; | |
659 | ||
660 | /* This might alter the window we are looking at */ | |
661 | ||
662 | o->field.next_hash_offset = o->field.head_data_offset = 0; | |
663 | ||
664 | h = hash % (le64toh(f->header->field_hash_table_size) / sizeof(HashItem)); | |
665 | p = le64toh(f->field_hash_table[h].tail_hash_offset); | |
666 | if (p == 0) | |
667 | f->field_hash_table[h].head_hash_offset = htole64(offset); | |
668 | else { | |
669 | r = journal_file_move_to_object(f, OBJECT_FIELD, p, &o); | |
670 | if (r < 0) | |
671 | return r; | |
672 | ||
673 | o->field.next_hash_offset = htole64(offset); | |
674 | } | |
675 | ||
676 | f->field_hash_table[h].tail_hash_offset = htole64(offset); | |
677 | ||
678 | if (JOURNAL_HEADER_CONTAINS(f->header, n_fields)) | |
679 | f->header->n_fields = htole64(le64toh(f->header->n_fields) + 1); | |
680 | ||
681 | return 0; | |
682 | } | |
683 | ||
684 | static int journal_file_link_data( | |
685 | JournalFile *f, | |
686 | Object *o, | |
687 | uint64_t offset, | |
688 | uint64_t hash) { | |
689 | ||
de190aef | 690 | uint64_t p, h; |
cec736d2 LP |
691 | int r; |
692 | ||
693 | assert(f); | |
694 | assert(o); | |
695 | assert(offset > 0); | |
b588975f LP |
696 | |
697 | if (o->object.type != OBJECT_DATA) | |
698 | return -EINVAL; | |
cec736d2 | 699 | |
48496df6 LP |
700 | /* This might alter the window we are looking at */ |
701 | ||
de190aef LP |
702 | o->data.next_hash_offset = o->data.next_field_offset = 0; |
703 | o->data.entry_offset = o->data.entry_array_offset = 0; | |
704 | o->data.n_entries = 0; | |
cec736d2 | 705 | |
de190aef | 706 | h = hash % (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)); |
8db4213e | 707 | p = le64toh(f->data_hash_table[h].tail_hash_offset); |
3c1668da | 708 | if (p == 0) |
cec736d2 | 709 | /* Only entry in the hash table is easy */ |
de190aef | 710 | f->data_hash_table[h].head_hash_offset = htole64(offset); |
3c1668da | 711 | else { |
48496df6 LP |
712 | /* Move back to the previous data object, to patch in |
713 | * pointer */ | |
cec736d2 | 714 | |
de190aef | 715 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); |
cec736d2 LP |
716 | if (r < 0) |
717 | return r; | |
718 | ||
de190aef | 719 | o->data.next_hash_offset = htole64(offset); |
cec736d2 LP |
720 | } |
721 | ||
de190aef | 722 | f->data_hash_table[h].tail_hash_offset = htole64(offset); |
cec736d2 | 723 | |
dca6219e LP |
724 | if (JOURNAL_HEADER_CONTAINS(f->header, n_data)) |
725 | f->header->n_data = htole64(le64toh(f->header->n_data) + 1); | |
726 | ||
cec736d2 LP |
727 | return 0; |
728 | } | |
729 | ||
3c1668da LP |
730 | int journal_file_find_field_object_with_hash( |
731 | JournalFile *f, | |
732 | const void *field, uint64_t size, uint64_t hash, | |
733 | Object **ret, uint64_t *offset) { | |
734 | ||
735 | uint64_t p, osize, h; | |
736 | int r; | |
737 | ||
738 | assert(f); | |
739 | assert(field && size > 0); | |
740 | ||
741 | osize = offsetof(Object, field.payload) + size; | |
742 | ||
743 | if (f->header->field_hash_table_size == 0) | |
744 | return -EBADMSG; | |
745 | ||
746 | h = hash % (le64toh(f->header->field_hash_table_size) / sizeof(HashItem)); | |
747 | p = le64toh(f->field_hash_table[h].head_hash_offset); | |
748 | ||
749 | while (p > 0) { | |
750 | Object *o; | |
751 | ||
752 | r = journal_file_move_to_object(f, OBJECT_FIELD, p, &o); | |
753 | if (r < 0) | |
754 | return r; | |
755 | ||
756 | if (le64toh(o->field.hash) == hash && | |
757 | le64toh(o->object.size) == osize && | |
758 | memcmp(o->field.payload, field, size) == 0) { | |
759 | ||
760 | if (ret) | |
761 | *ret = o; | |
762 | if (offset) | |
763 | *offset = p; | |
764 | ||
765 | return 1; | |
766 | } | |
767 | ||
768 | p = le64toh(o->field.next_hash_offset); | |
769 | } | |
770 | ||
771 | return 0; | |
772 | } | |
773 | ||
774 | int journal_file_find_field_object( | |
775 | JournalFile *f, | |
776 | const void *field, uint64_t size, | |
777 | Object **ret, uint64_t *offset) { | |
778 | ||
779 | uint64_t hash; | |
780 | ||
781 | assert(f); | |
782 | assert(field && size > 0); | |
783 | ||
784 | hash = hash64(field, size); | |
785 | ||
786 | return journal_file_find_field_object_with_hash(f, | |
787 | field, size, hash, | |
788 | ret, offset); | |
789 | } | |
790 | ||
de190aef LP |
791 | int journal_file_find_data_object_with_hash( |
792 | JournalFile *f, | |
793 | const void *data, uint64_t size, uint64_t hash, | |
794 | Object **ret, uint64_t *offset) { | |
48496df6 | 795 | |
de190aef | 796 | uint64_t p, osize, h; |
cec736d2 LP |
797 | int r; |
798 | ||
799 | assert(f); | |
800 | assert(data || size == 0); | |
801 | ||
802 | osize = offsetof(Object, data.payload) + size; | |
803 | ||
bc85bfee LP |
804 | if (f->header->data_hash_table_size == 0) |
805 | return -EBADMSG; | |
806 | ||
de190aef LP |
807 | h = hash % (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)); |
808 | p = le64toh(f->data_hash_table[h].head_hash_offset); | |
cec736d2 | 809 | |
de190aef LP |
810 | while (p > 0) { |
811 | Object *o; | |
cec736d2 | 812 | |
de190aef | 813 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); |
cec736d2 LP |
814 | if (r < 0) |
815 | return r; | |
816 | ||
807e17f0 | 817 | if (le64toh(o->data.hash) != hash) |
85a131e8 | 818 | goto next; |
807e17f0 | 819 | |
d89c8fdf | 820 | if (o->object.flags & OBJECT_COMPRESSION_MASK) { |
3b1a55e1 | 821 | #if defined(HAVE_XZ) || defined(HAVE_LZ4) |
fa1c4b51 ZJS |
822 | uint64_t l; |
823 | size_t rsize; | |
cec736d2 | 824 | |
807e17f0 LP |
825 | l = le64toh(o->object.size); |
826 | if (l <= offsetof(Object, data.payload)) | |
cec736d2 LP |
827 | return -EBADMSG; |
828 | ||
807e17f0 LP |
829 | l -= offsetof(Object, data.payload); |
830 | ||
d89c8fdf ZJS |
831 | r = decompress_blob(o->object.flags & OBJECT_COMPRESSION_MASK, |
832 | o->data.payload, l, &f->compress_buffer, &f->compress_buffer_size, &rsize, 0); | |
833 | if (r < 0) | |
834 | return r; | |
807e17f0 | 835 | |
b785c858 | 836 | if (rsize == size && |
807e17f0 LP |
837 | memcmp(f->compress_buffer, data, size) == 0) { |
838 | ||
839 | if (ret) | |
840 | *ret = o; | |
841 | ||
842 | if (offset) | |
843 | *offset = p; | |
844 | ||
845 | return 1; | |
846 | } | |
3b1a55e1 ZJS |
847 | #else |
848 | return -EPROTONOSUPPORT; | |
849 | #endif | |
807e17f0 LP |
850 | } else if (le64toh(o->object.size) == osize && |
851 | memcmp(o->data.payload, data, size) == 0) { | |
852 | ||
cec736d2 LP |
853 | if (ret) |
854 | *ret = o; | |
855 | ||
856 | if (offset) | |
857 | *offset = p; | |
858 | ||
de190aef | 859 | return 1; |
cec736d2 LP |
860 | } |
861 | ||
85a131e8 | 862 | next: |
cec736d2 LP |
863 | p = le64toh(o->data.next_hash_offset); |
864 | } | |
865 | ||
de190aef LP |
866 | return 0; |
867 | } | |
868 | ||
869 | int journal_file_find_data_object( | |
870 | JournalFile *f, | |
871 | const void *data, uint64_t size, | |
872 | Object **ret, uint64_t *offset) { | |
873 | ||
874 | uint64_t hash; | |
875 | ||
876 | assert(f); | |
877 | assert(data || size == 0); | |
878 | ||
879 | hash = hash64(data, size); | |
880 | ||
881 | return journal_file_find_data_object_with_hash(f, | |
882 | data, size, hash, | |
883 | ret, offset); | |
884 | } | |
885 | ||
3c1668da LP |
886 | static int journal_file_append_field( |
887 | JournalFile *f, | |
888 | const void *field, uint64_t size, | |
889 | Object **ret, uint64_t *offset) { | |
890 | ||
891 | uint64_t hash, p; | |
892 | uint64_t osize; | |
893 | Object *o; | |
894 | int r; | |
895 | ||
896 | assert(f); | |
897 | assert(field && size > 0); | |
898 | ||
899 | hash = hash64(field, size); | |
900 | ||
901 | r = journal_file_find_field_object_with_hash(f, field, size, hash, &o, &p); | |
902 | if (r < 0) | |
903 | return r; | |
904 | else if (r > 0) { | |
905 | ||
906 | if (ret) | |
907 | *ret = o; | |
908 | ||
909 | if (offset) | |
910 | *offset = p; | |
911 | ||
912 | return 0; | |
913 | } | |
914 | ||
915 | osize = offsetof(Object, field.payload) + size; | |
916 | r = journal_file_append_object(f, OBJECT_FIELD, osize, &o, &p); | |
8c92d4bb LP |
917 | if (r < 0) |
918 | return r; | |
3c1668da LP |
919 | |
920 | o->field.hash = htole64(hash); | |
921 | memcpy(o->field.payload, field, size); | |
922 | ||
923 | r = journal_file_link_field(f, o, p, hash); | |
924 | if (r < 0) | |
925 | return r; | |
926 | ||
927 | /* The linking might have altered the window, so let's | |
928 | * refresh our pointer */ | |
929 | r = journal_file_move_to_object(f, OBJECT_FIELD, p, &o); | |
930 | if (r < 0) | |
931 | return r; | |
932 | ||
933 | #ifdef HAVE_GCRYPT | |
934 | r = journal_file_hmac_put_object(f, OBJECT_FIELD, o, p); | |
935 | if (r < 0) | |
936 | return r; | |
937 | #endif | |
938 | ||
939 | if (ret) | |
940 | *ret = o; | |
941 | ||
942 | if (offset) | |
943 | *offset = p; | |
944 | ||
945 | return 0; | |
946 | } | |
947 | ||
48496df6 LP |
948 | static int journal_file_append_data( |
949 | JournalFile *f, | |
950 | const void *data, uint64_t size, | |
951 | Object **ret, uint64_t *offset) { | |
952 | ||
de190aef LP |
953 | uint64_t hash, p; |
954 | uint64_t osize; | |
955 | Object *o; | |
d89c8fdf | 956 | int r, compression = 0; |
3c1668da | 957 | const void *eq; |
de190aef LP |
958 | |
959 | assert(f); | |
960 | assert(data || size == 0); | |
961 | ||
962 | hash = hash64(data, size); | |
963 | ||
964 | r = journal_file_find_data_object_with_hash(f, data, size, hash, &o, &p); | |
965 | if (r < 0) | |
966 | return r; | |
967 | else if (r > 0) { | |
968 | ||
969 | if (ret) | |
970 | *ret = o; | |
971 | ||
972 | if (offset) | |
973 | *offset = p; | |
974 | ||
975 | return 0; | |
976 | } | |
977 | ||
978 | osize = offsetof(Object, data.payload) + size; | |
979 | r = journal_file_append_object(f, OBJECT_DATA, osize, &o, &p); | |
cec736d2 LP |
980 | if (r < 0) |
981 | return r; | |
982 | ||
cec736d2 | 983 | o->data.hash = htole64(hash); |
807e17f0 | 984 | |
d89c8fdf ZJS |
985 | #if defined(HAVE_XZ) || defined(HAVE_LZ4) |
986 | if (f->compress_xz && | |
807e17f0 | 987 | size >= COMPRESSION_SIZE_THRESHOLD) { |
fa1c4b51 | 988 | size_t rsize; |
807e17f0 | 989 | |
d89c8fdf | 990 | compression = compress_blob(data, size, o->data.payload, &rsize); |
807e17f0 | 991 | |
d89c8fdf | 992 | if (compression) { |
807e17f0 | 993 | o->object.size = htole64(offsetof(Object, data.payload) + rsize); |
d89c8fdf | 994 | o->object.flags |= compression; |
807e17f0 | 995 | |
fa1c4b51 | 996 | log_debug("Compressed data object %"PRIu64" -> %zu using %s", |
d89c8fdf | 997 | size, rsize, object_compressed_to_string(compression)); |
807e17f0 LP |
998 | } |
999 | } | |
1000 | #endif | |
1001 | ||
d89c8fdf | 1002 | if (!compression && size > 0) |
807e17f0 | 1003 | memcpy(o->data.payload, data, size); |
cec736d2 | 1004 | |
de190aef | 1005 | r = journal_file_link_data(f, o, p, hash); |
cec736d2 LP |
1006 | if (r < 0) |
1007 | return r; | |
1008 | ||
48496df6 LP |
1009 | /* The linking might have altered the window, so let's |
1010 | * refresh our pointer */ | |
1011 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); | |
1012 | if (r < 0) | |
1013 | return r; | |
1014 | ||
08c6f819 SL |
1015 | if (!data) |
1016 | eq = NULL; | |
1017 | else | |
1018 | eq = memchr(data, '=', size); | |
3c1668da | 1019 | if (eq && eq > data) { |
748db592 | 1020 | Object *fo = NULL; |
3c1668da | 1021 | uint64_t fp; |
3c1668da LP |
1022 | |
1023 | /* Create field object ... */ | |
1024 | r = journal_file_append_field(f, data, (uint8_t*) eq - (uint8_t*) data, &fo, &fp); | |
1025 | if (r < 0) | |
1026 | return r; | |
1027 | ||
1028 | /* ... and link it in. */ | |
1029 | o->data.next_field_offset = fo->field.head_data_offset; | |
1030 | fo->field.head_data_offset = le64toh(p); | |
1031 | } | |
1032 | ||
5996c7c2 LP |
1033 | #ifdef HAVE_GCRYPT |
1034 | r = journal_file_hmac_put_object(f, OBJECT_DATA, o, p); | |
1035 | if (r < 0) | |
1036 | return r; | |
1037 | #endif | |
1038 | ||
cec736d2 LP |
1039 | if (ret) |
1040 | *ret = o; | |
1041 | ||
1042 | if (offset) | |
de190aef | 1043 | *offset = p; |
cec736d2 LP |
1044 | |
1045 | return 0; | |
1046 | } | |
1047 | ||
1048 | uint64_t journal_file_entry_n_items(Object *o) { | |
1049 | assert(o); | |
b588975f LP |
1050 | |
1051 | if (o->object.type != OBJECT_ENTRY) | |
1052 | return 0; | |
cec736d2 LP |
1053 | |
1054 | return (le64toh(o->object.size) - offsetof(Object, entry.items)) / sizeof(EntryItem); | |
1055 | } | |
1056 | ||
0284adc6 | 1057 | uint64_t journal_file_entry_array_n_items(Object *o) { |
de190aef | 1058 | assert(o); |
b588975f LP |
1059 | |
1060 | if (o->object.type != OBJECT_ENTRY_ARRAY) | |
1061 | return 0; | |
de190aef LP |
1062 | |
1063 | return (le64toh(o->object.size) - offsetof(Object, entry_array.items)) / sizeof(uint64_t); | |
1064 | } | |
1065 | ||
fb9a24b6 LP |
1066 | uint64_t journal_file_hash_table_n_items(Object *o) { |
1067 | assert(o); | |
b588975f LP |
1068 | |
1069 | if (o->object.type != OBJECT_DATA_HASH_TABLE && | |
1070 | o->object.type != OBJECT_FIELD_HASH_TABLE) | |
1071 | return 0; | |
fb9a24b6 LP |
1072 | |
1073 | return (le64toh(o->object.size) - offsetof(Object, hash_table.items)) / sizeof(HashItem); | |
1074 | } | |
1075 | ||
de190aef | 1076 | static int link_entry_into_array(JournalFile *f, |
4fd052ae FC |
1077 | le64_t *first, |
1078 | le64_t *idx, | |
de190aef | 1079 | uint64_t p) { |
cec736d2 | 1080 | int r; |
de190aef LP |
1081 | uint64_t n = 0, ap = 0, q, i, a, hidx; |
1082 | Object *o; | |
1083 | ||
cec736d2 | 1084 | assert(f); |
de190aef LP |
1085 | assert(first); |
1086 | assert(idx); | |
1087 | assert(p > 0); | |
cec736d2 | 1088 | |
de190aef LP |
1089 | a = le64toh(*first); |
1090 | i = hidx = le64toh(*idx); | |
1091 | while (a > 0) { | |
1092 | ||
1093 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &o); | |
1094 | if (r < 0) | |
1095 | return r; | |
cec736d2 | 1096 | |
de190aef LP |
1097 | n = journal_file_entry_array_n_items(o); |
1098 | if (i < n) { | |
1099 | o->entry_array.items[i] = htole64(p); | |
1100 | *idx = htole64(hidx + 1); | |
1101 | return 0; | |
1102 | } | |
cec736d2 | 1103 | |
de190aef LP |
1104 | i -= n; |
1105 | ap = a; | |
1106 | a = le64toh(o->entry_array.next_entry_array_offset); | |
1107 | } | |
1108 | ||
1109 | if (hidx > n) | |
1110 | n = (hidx+1) * 2; | |
1111 | else | |
1112 | n = n * 2; | |
1113 | ||
1114 | if (n < 4) | |
1115 | n = 4; | |
1116 | ||
1117 | r = journal_file_append_object(f, OBJECT_ENTRY_ARRAY, | |
1118 | offsetof(Object, entry_array.items) + n * sizeof(uint64_t), | |
1119 | &o, &q); | |
cec736d2 LP |
1120 | if (r < 0) |
1121 | return r; | |
1122 | ||
feb12d3e | 1123 | #ifdef HAVE_GCRYPT |
5996c7c2 | 1124 | r = journal_file_hmac_put_object(f, OBJECT_ENTRY_ARRAY, o, q); |
b0af6f41 LP |
1125 | if (r < 0) |
1126 | return r; | |
feb12d3e | 1127 | #endif |
b0af6f41 | 1128 | |
de190aef | 1129 | o->entry_array.items[i] = htole64(p); |
cec736d2 | 1130 | |
de190aef | 1131 | if (ap == 0) |
7be3aa17 | 1132 | *first = htole64(q); |
cec736d2 | 1133 | else { |
de190aef | 1134 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, ap, &o); |
cec736d2 LP |
1135 | if (r < 0) |
1136 | return r; | |
1137 | ||
de190aef LP |
1138 | o->entry_array.next_entry_array_offset = htole64(q); |
1139 | } | |
cec736d2 | 1140 | |
2dee23eb LP |
1141 | if (JOURNAL_HEADER_CONTAINS(f->header, n_entry_arrays)) |
1142 | f->header->n_entry_arrays = htole64(le64toh(f->header->n_entry_arrays) + 1); | |
1143 | ||
de190aef LP |
1144 | *idx = htole64(hidx + 1); |
1145 | ||
1146 | return 0; | |
1147 | } | |
cec736d2 | 1148 | |
de190aef | 1149 | static int link_entry_into_array_plus_one(JournalFile *f, |
4fd052ae FC |
1150 | le64_t *extra, |
1151 | le64_t *first, | |
1152 | le64_t *idx, | |
de190aef LP |
1153 | uint64_t p) { |
1154 | ||
1155 | int r; | |
1156 | ||
1157 | assert(f); | |
1158 | assert(extra); | |
1159 | assert(first); | |
1160 | assert(idx); | |
1161 | assert(p > 0); | |
1162 | ||
1163 | if (*idx == 0) | |
1164 | *extra = htole64(p); | |
1165 | else { | |
4fd052ae | 1166 | le64_t i; |
de190aef | 1167 | |
7be3aa17 | 1168 | i = htole64(le64toh(*idx) - 1); |
de190aef LP |
1169 | r = link_entry_into_array(f, first, &i, p); |
1170 | if (r < 0) | |
1171 | return r; | |
cec736d2 LP |
1172 | } |
1173 | ||
de190aef LP |
1174 | *idx = htole64(le64toh(*idx) + 1); |
1175 | return 0; | |
1176 | } | |
1177 | ||
1178 | static int journal_file_link_entry_item(JournalFile *f, Object *o, uint64_t offset, uint64_t i) { | |
1179 | uint64_t p; | |
1180 | int r; | |
1181 | assert(f); | |
1182 | assert(o); | |
1183 | assert(offset > 0); | |
1184 | ||
1185 | p = le64toh(o->entry.items[i].object_offset); | |
1186 | if (p == 0) | |
1187 | return -EINVAL; | |
1188 | ||
1189 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); | |
cec736d2 LP |
1190 | if (r < 0) |
1191 | return r; | |
1192 | ||
de190aef LP |
1193 | return link_entry_into_array_plus_one(f, |
1194 | &o->data.entry_offset, | |
1195 | &o->data.entry_array_offset, | |
1196 | &o->data.n_entries, | |
1197 | offset); | |
cec736d2 LP |
1198 | } |
1199 | ||
1200 | static int journal_file_link_entry(JournalFile *f, Object *o, uint64_t offset) { | |
de190aef | 1201 | uint64_t n, i; |
cec736d2 LP |
1202 | int r; |
1203 | ||
1204 | assert(f); | |
1205 | assert(o); | |
1206 | assert(offset > 0); | |
b588975f LP |
1207 | |
1208 | if (o->object.type != OBJECT_ENTRY) | |
1209 | return -EINVAL; | |
cec736d2 | 1210 | |
b788cc23 LP |
1211 | __sync_synchronize(); |
1212 | ||
cec736d2 | 1213 | /* Link up the entry itself */ |
de190aef LP |
1214 | r = link_entry_into_array(f, |
1215 | &f->header->entry_array_offset, | |
1216 | &f->header->n_entries, | |
1217 | offset); | |
1218 | if (r < 0) | |
1219 | return r; | |
cec736d2 | 1220 | |
507f22bd | 1221 | /* log_debug("=> %s seqnr=%"PRIu64" n_entries=%"PRIu64, f->path, o->entry.seqnum, f->header->n_entries); */ |
cec736d2 | 1222 | |
de190aef | 1223 | if (f->header->head_entry_realtime == 0) |
0ac38b70 | 1224 | f->header->head_entry_realtime = o->entry.realtime; |
cec736d2 | 1225 | |
0ac38b70 | 1226 | f->header->tail_entry_realtime = o->entry.realtime; |
de190aef LP |
1227 | f->header->tail_entry_monotonic = o->entry.monotonic; |
1228 | ||
1229 | f->tail_entry_monotonic_valid = true; | |
cec736d2 LP |
1230 | |
1231 | /* Link up the items */ | |
1232 | n = journal_file_entry_n_items(o); | |
1233 | for (i = 0; i < n; i++) { | |
1234 | r = journal_file_link_entry_item(f, o, offset, i); | |
1235 | if (r < 0) | |
1236 | return r; | |
1237 | } | |
1238 | ||
cec736d2 LP |
1239 | return 0; |
1240 | } | |
1241 | ||
1242 | static int journal_file_append_entry_internal( | |
1243 | JournalFile *f, | |
1244 | const dual_timestamp *ts, | |
1245 | uint64_t xor_hash, | |
1246 | const EntryItem items[], unsigned n_items, | |
de190aef | 1247 | uint64_t *seqnum, |
cec736d2 LP |
1248 | Object **ret, uint64_t *offset) { |
1249 | uint64_t np; | |
1250 | uint64_t osize; | |
1251 | Object *o; | |
1252 | int r; | |
1253 | ||
1254 | assert(f); | |
1255 | assert(items || n_items == 0); | |
de190aef | 1256 | assert(ts); |
cec736d2 LP |
1257 | |
1258 | osize = offsetof(Object, entry.items) + (n_items * sizeof(EntryItem)); | |
1259 | ||
de190aef | 1260 | r = journal_file_append_object(f, OBJECT_ENTRY, osize, &o, &np); |
cec736d2 LP |
1261 | if (r < 0) |
1262 | return r; | |
1263 | ||
d98cc1f2 | 1264 | o->entry.seqnum = htole64(journal_file_entry_seqnum(f, seqnum)); |
cec736d2 | 1265 | memcpy(o->entry.items, items, n_items * sizeof(EntryItem)); |
de190aef LP |
1266 | o->entry.realtime = htole64(ts->realtime); |
1267 | o->entry.monotonic = htole64(ts->monotonic); | |
cec736d2 LP |
1268 | o->entry.xor_hash = htole64(xor_hash); |
1269 | o->entry.boot_id = f->header->boot_id; | |
1270 | ||
feb12d3e | 1271 | #ifdef HAVE_GCRYPT |
5996c7c2 | 1272 | r = journal_file_hmac_put_object(f, OBJECT_ENTRY, o, np); |
b0af6f41 LP |
1273 | if (r < 0) |
1274 | return r; | |
feb12d3e | 1275 | #endif |
b0af6f41 | 1276 | |
cec736d2 LP |
1277 | r = journal_file_link_entry(f, o, np); |
1278 | if (r < 0) | |
1279 | return r; | |
1280 | ||
1281 | if (ret) | |
1282 | *ret = o; | |
1283 | ||
1284 | if (offset) | |
1285 | *offset = np; | |
1286 | ||
1287 | return 0; | |
1288 | } | |
1289 | ||
cf244689 | 1290 | void journal_file_post_change(JournalFile *f) { |
50f20cfd LP |
1291 | assert(f); |
1292 | ||
1293 | /* inotify() does not receive IN_MODIFY events from file | |
1294 | * accesses done via mmap(). After each access we hence | |
1295 | * trigger IN_MODIFY by truncating the journal file to its | |
1296 | * current size which triggers IN_MODIFY. */ | |
1297 | ||
bc85bfee LP |
1298 | __sync_synchronize(); |
1299 | ||
50f20cfd | 1300 | if (ftruncate(f->fd, f->last_stat.st_size) < 0) |
56f64d95 | 1301 | log_error_errno(errno, "Failed to truncate file to its own size: %m"); |
50f20cfd LP |
1302 | } |
1303 | ||
1f2da9ec LP |
1304 | static int entry_item_cmp(const void *_a, const void *_b) { |
1305 | const EntryItem *a = _a, *b = _b; | |
1306 | ||
1307 | if (le64toh(a->object_offset) < le64toh(b->object_offset)) | |
1308 | return -1; | |
1309 | if (le64toh(a->object_offset) > le64toh(b->object_offset)) | |
1310 | return 1; | |
1311 | return 0; | |
1312 | } | |
1313 | ||
de190aef | 1314 | int journal_file_append_entry(JournalFile *f, const dual_timestamp *ts, const struct iovec iovec[], unsigned n_iovec, uint64_t *seqnum, Object **ret, uint64_t *offset) { |
cec736d2 LP |
1315 | unsigned i; |
1316 | EntryItem *items; | |
1317 | int r; | |
1318 | uint64_t xor_hash = 0; | |
de190aef | 1319 | struct dual_timestamp _ts; |
cec736d2 LP |
1320 | |
1321 | assert(f); | |
1322 | assert(iovec || n_iovec == 0); | |
1323 | ||
de190aef LP |
1324 | if (!ts) { |
1325 | dual_timestamp_get(&_ts); | |
1326 | ts = &_ts; | |
1327 | } | |
1328 | ||
1329 | if (f->tail_entry_monotonic_valid && | |
1330 | ts->monotonic < le64toh(f->header->tail_entry_monotonic)) | |
1331 | return -EINVAL; | |
1332 | ||
feb12d3e | 1333 | #ifdef HAVE_GCRYPT |
7560fffc LP |
1334 | r = journal_file_maybe_append_tag(f, ts->realtime); |
1335 | if (r < 0) | |
1336 | return r; | |
feb12d3e | 1337 | #endif |
7560fffc | 1338 | |
64825d3c | 1339 | /* alloca() can't take 0, hence let's allocate at least one */ |
9607d947 | 1340 | items = alloca(sizeof(EntryItem) * MAX(1u, n_iovec)); |
cec736d2 LP |
1341 | |
1342 | for (i = 0; i < n_iovec; i++) { | |
1343 | uint64_t p; | |
1344 | Object *o; | |
1345 | ||
1346 | r = journal_file_append_data(f, iovec[i].iov_base, iovec[i].iov_len, &o, &p); | |
1347 | if (r < 0) | |
cf244689 | 1348 | return r; |
cec736d2 LP |
1349 | |
1350 | xor_hash ^= le64toh(o->data.hash); | |
1351 | items[i].object_offset = htole64(p); | |
de7b95cd | 1352 | items[i].hash = o->data.hash; |
cec736d2 LP |
1353 | } |
1354 | ||
1f2da9ec LP |
1355 | /* Order by the position on disk, in order to improve seek |
1356 | * times for rotating media. */ | |
7ff7394d | 1357 | qsort_safe(items, n_iovec, sizeof(EntryItem), entry_item_cmp); |
1f2da9ec | 1358 | |
de190aef | 1359 | r = journal_file_append_entry_internal(f, ts, xor_hash, items, n_iovec, seqnum, ret, offset); |
cec736d2 | 1360 | |
50f20cfd LP |
1361 | journal_file_post_change(f); |
1362 | ||
cec736d2 LP |
1363 | return r; |
1364 | } | |
1365 | ||
a4bcff5b | 1366 | typedef struct ChainCacheItem { |
fb099c8d | 1367 | uint64_t first; /* the array at the beginning of the chain */ |
a4bcff5b LP |
1368 | uint64_t array; /* the cached array */ |
1369 | uint64_t begin; /* the first item in the cached array */ | |
1370 | uint64_t total; /* the total number of items in all arrays before this one in the chain */ | |
f268980d | 1371 | uint64_t last_index; /* the last index we looked at, to optimize locality when bisecting */ |
a4bcff5b LP |
1372 | } ChainCacheItem; |
1373 | ||
1374 | static void chain_cache_put( | |
4743015d | 1375 | OrderedHashmap *h, |
a4bcff5b LP |
1376 | ChainCacheItem *ci, |
1377 | uint64_t first, | |
1378 | uint64_t array, | |
1379 | uint64_t begin, | |
f268980d LP |
1380 | uint64_t total, |
1381 | uint64_t last_index) { | |
a4bcff5b LP |
1382 | |
1383 | if (!ci) { | |
34741aa3 LP |
1384 | /* If the chain item to cache for this chain is the |
1385 | * first one it's not worth caching anything */ | |
1386 | if (array == first) | |
1387 | return; | |
1388 | ||
29433089 | 1389 | if (ordered_hashmap_size(h) >= CHAIN_CACHE_MAX) { |
4743015d | 1390 | ci = ordered_hashmap_steal_first(h); |
29433089 LP |
1391 | assert(ci); |
1392 | } else { | |
a4bcff5b LP |
1393 | ci = new(ChainCacheItem, 1); |
1394 | if (!ci) | |
1395 | return; | |
1396 | } | |
1397 | ||
1398 | ci->first = first; | |
1399 | ||
4743015d | 1400 | if (ordered_hashmap_put(h, &ci->first, ci) < 0) { |
a4bcff5b LP |
1401 | free(ci); |
1402 | return; | |
1403 | } | |
1404 | } else | |
1405 | assert(ci->first == first); | |
1406 | ||
1407 | ci->array = array; | |
1408 | ci->begin = begin; | |
1409 | ci->total = total; | |
f268980d | 1410 | ci->last_index = last_index; |
a4bcff5b LP |
1411 | } |
1412 | ||
f268980d LP |
1413 | static int generic_array_get( |
1414 | JournalFile *f, | |
1415 | uint64_t first, | |
1416 | uint64_t i, | |
1417 | Object **ret, uint64_t *offset) { | |
de190aef | 1418 | |
cec736d2 | 1419 | Object *o; |
a4bcff5b | 1420 | uint64_t p = 0, a, t = 0; |
cec736d2 | 1421 | int r; |
a4bcff5b | 1422 | ChainCacheItem *ci; |
cec736d2 LP |
1423 | |
1424 | assert(f); | |
1425 | ||
de190aef | 1426 | a = first; |
a4bcff5b LP |
1427 | |
1428 | /* Try the chain cache first */ | |
4743015d | 1429 | ci = ordered_hashmap_get(f->chain_cache, &first); |
a4bcff5b LP |
1430 | if (ci && i > ci->total) { |
1431 | a = ci->array; | |
1432 | i -= ci->total; | |
1433 | t = ci->total; | |
1434 | } | |
1435 | ||
de190aef | 1436 | while (a > 0) { |
a4bcff5b | 1437 | uint64_t k; |
cec736d2 | 1438 | |
de190aef LP |
1439 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &o); |
1440 | if (r < 0) | |
1441 | return r; | |
cec736d2 | 1442 | |
a4bcff5b LP |
1443 | k = journal_file_entry_array_n_items(o); |
1444 | if (i < k) { | |
de190aef | 1445 | p = le64toh(o->entry_array.items[i]); |
a4bcff5b | 1446 | goto found; |
cec736d2 LP |
1447 | } |
1448 | ||
a4bcff5b LP |
1449 | i -= k; |
1450 | t += k; | |
de190aef LP |
1451 | a = le64toh(o->entry_array.next_entry_array_offset); |
1452 | } | |
1453 | ||
a4bcff5b LP |
1454 | return 0; |
1455 | ||
1456 | found: | |
1457 | /* Let's cache this item for the next invocation */ | |
af13a6b0 | 1458 | chain_cache_put(f->chain_cache, ci, first, a, le64toh(o->entry_array.items[0]), t, i); |
de190aef LP |
1459 | |
1460 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1461 | if (r < 0) | |
1462 | return r; | |
1463 | ||
1464 | if (ret) | |
1465 | *ret = o; | |
1466 | ||
1467 | if (offset) | |
1468 | *offset = p; | |
1469 | ||
1470 | return 1; | |
1471 | } | |
1472 | ||
f268980d LP |
1473 | static int generic_array_get_plus_one( |
1474 | JournalFile *f, | |
1475 | uint64_t extra, | |
1476 | uint64_t first, | |
1477 | uint64_t i, | |
1478 | Object **ret, uint64_t *offset) { | |
de190aef LP |
1479 | |
1480 | Object *o; | |
1481 | ||
1482 | assert(f); | |
1483 | ||
1484 | if (i == 0) { | |
1485 | int r; | |
1486 | ||
1487 | r = journal_file_move_to_object(f, OBJECT_ENTRY, extra, &o); | |
cec736d2 LP |
1488 | if (r < 0) |
1489 | return r; | |
1490 | ||
de190aef LP |
1491 | if (ret) |
1492 | *ret = o; | |
cec736d2 | 1493 | |
de190aef LP |
1494 | if (offset) |
1495 | *offset = extra; | |
cec736d2 | 1496 | |
de190aef | 1497 | return 1; |
cec736d2 LP |
1498 | } |
1499 | ||
de190aef LP |
1500 | return generic_array_get(f, first, i-1, ret, offset); |
1501 | } | |
cec736d2 | 1502 | |
de190aef LP |
1503 | enum { |
1504 | TEST_FOUND, | |
1505 | TEST_LEFT, | |
1506 | TEST_RIGHT | |
1507 | }; | |
cec736d2 | 1508 | |
f268980d LP |
1509 | static int generic_array_bisect( |
1510 | JournalFile *f, | |
1511 | uint64_t first, | |
1512 | uint64_t n, | |
1513 | uint64_t needle, | |
1514 | int (*test_object)(JournalFile *f, uint64_t p, uint64_t needle), | |
1515 | direction_t direction, | |
1516 | Object **ret, | |
1517 | uint64_t *offset, | |
1518 | uint64_t *idx) { | |
1519 | ||
1520 | uint64_t a, p, t = 0, i = 0, last_p = 0, last_index = (uint64_t) -1; | |
de190aef LP |
1521 | bool subtract_one = false; |
1522 | Object *o, *array = NULL; | |
1523 | int r; | |
a4bcff5b | 1524 | ChainCacheItem *ci; |
cec736d2 | 1525 | |
de190aef LP |
1526 | assert(f); |
1527 | assert(test_object); | |
cec736d2 | 1528 | |
a4bcff5b | 1529 | /* Start with the first array in the chain */ |
de190aef | 1530 | a = first; |
a4bcff5b | 1531 | |
4743015d | 1532 | ci = ordered_hashmap_get(f->chain_cache, &first); |
a4bcff5b LP |
1533 | if (ci && n > ci->total) { |
1534 | /* Ah, we have iterated this bisection array chain | |
1535 | * previously! Let's see if we can skip ahead in the | |
1536 | * chain, as far as the last time. But we can't jump | |
1537 | * backwards in the chain, so let's check that | |
1538 | * first. */ | |
1539 | ||
1540 | r = test_object(f, ci->begin, needle); | |
1541 | if (r < 0) | |
1542 | return r; | |
1543 | ||
1544 | if (r == TEST_LEFT) { | |
f268980d | 1545 | /* OK, what we are looking for is right of the |
a4bcff5b LP |
1546 | * begin of this EntryArray, so let's jump |
1547 | * straight to previously cached array in the | |
1548 | * chain */ | |
1549 | ||
1550 | a = ci->array; | |
1551 | n -= ci->total; | |
1552 | t = ci->total; | |
f268980d | 1553 | last_index = ci->last_index; |
a4bcff5b LP |
1554 | } |
1555 | } | |
1556 | ||
de190aef LP |
1557 | while (a > 0) { |
1558 | uint64_t left, right, k, lp; | |
1559 | ||
1560 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &array); | |
cec736d2 LP |
1561 | if (r < 0) |
1562 | return r; | |
1563 | ||
de190aef LP |
1564 | k = journal_file_entry_array_n_items(array); |
1565 | right = MIN(k, n); | |
1566 | if (right <= 0) | |
1567 | return 0; | |
cec736d2 | 1568 | |
de190aef LP |
1569 | i = right - 1; |
1570 | lp = p = le64toh(array->entry_array.items[i]); | |
1571 | if (p <= 0) | |
1572 | return -EBADMSG; | |
cec736d2 | 1573 | |
de190aef LP |
1574 | r = test_object(f, p, needle); |
1575 | if (r < 0) | |
1576 | return r; | |
cec736d2 | 1577 | |
de190aef LP |
1578 | if (r == TEST_FOUND) |
1579 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1580 | ||
1581 | if (r == TEST_RIGHT) { | |
1582 | left = 0; | |
1583 | right -= 1; | |
f268980d LP |
1584 | |
1585 | if (last_index != (uint64_t) -1) { | |
1586 | assert(last_index <= right); | |
1587 | ||
1588 | /* If we cached the last index we | |
1589 | * looked at, let's try to not to jump | |
1590 | * too wildly around and see if we can | |
1591 | * limit the range to look at early to | |
1592 | * the immediate neighbors of the last | |
1593 | * index we looked at. */ | |
1594 | ||
1595 | if (last_index > 0) { | |
1596 | uint64_t x = last_index - 1; | |
1597 | ||
1598 | p = le64toh(array->entry_array.items[x]); | |
1599 | if (p <= 0) | |
1600 | return -EBADMSG; | |
1601 | ||
1602 | r = test_object(f, p, needle); | |
1603 | if (r < 0) | |
1604 | return r; | |
1605 | ||
1606 | if (r == TEST_FOUND) | |
1607 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1608 | ||
1609 | if (r == TEST_RIGHT) | |
1610 | right = x; | |
1611 | else | |
1612 | left = x + 1; | |
1613 | } | |
1614 | ||
1615 | if (last_index < right) { | |
1616 | uint64_t y = last_index + 1; | |
1617 | ||
1618 | p = le64toh(array->entry_array.items[y]); | |
1619 | if (p <= 0) | |
1620 | return -EBADMSG; | |
1621 | ||
1622 | r = test_object(f, p, needle); | |
1623 | if (r < 0) | |
1624 | return r; | |
1625 | ||
1626 | if (r == TEST_FOUND) | |
1627 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1628 | ||
1629 | if (r == TEST_RIGHT) | |
1630 | right = y; | |
1631 | else | |
1632 | left = y + 1; | |
1633 | } | |
f268980d LP |
1634 | } |
1635 | ||
de190aef LP |
1636 | for (;;) { |
1637 | if (left == right) { | |
1638 | if (direction == DIRECTION_UP) | |
1639 | subtract_one = true; | |
1640 | ||
1641 | i = left; | |
1642 | goto found; | |
1643 | } | |
1644 | ||
1645 | assert(left < right); | |
de190aef | 1646 | i = (left + right) / 2; |
f268980d | 1647 | |
de190aef LP |
1648 | p = le64toh(array->entry_array.items[i]); |
1649 | if (p <= 0) | |
1650 | return -EBADMSG; | |
1651 | ||
1652 | r = test_object(f, p, needle); | |
1653 | if (r < 0) | |
1654 | return r; | |
cec736d2 | 1655 | |
de190aef LP |
1656 | if (r == TEST_FOUND) |
1657 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1658 | ||
1659 | if (r == TEST_RIGHT) | |
1660 | right = i; | |
1661 | else | |
1662 | left = i + 1; | |
1663 | } | |
1664 | } | |
1665 | ||
2173cbf8 | 1666 | if (k >= n) { |
cbdca852 LP |
1667 | if (direction == DIRECTION_UP) { |
1668 | i = n; | |
1669 | subtract_one = true; | |
1670 | goto found; | |
1671 | } | |
1672 | ||
cec736d2 | 1673 | return 0; |
cbdca852 | 1674 | } |
cec736d2 | 1675 | |
de190aef LP |
1676 | last_p = lp; |
1677 | ||
1678 | n -= k; | |
1679 | t += k; | |
f268980d | 1680 | last_index = (uint64_t) -1; |
de190aef | 1681 | a = le64toh(array->entry_array.next_entry_array_offset); |
cec736d2 LP |
1682 | } |
1683 | ||
1684 | return 0; | |
de190aef LP |
1685 | |
1686 | found: | |
1687 | if (subtract_one && t == 0 && i == 0) | |
1688 | return 0; | |
1689 | ||
a4bcff5b | 1690 | /* Let's cache this item for the next invocation */ |
af13a6b0 | 1691 | chain_cache_put(f->chain_cache, ci, first, a, le64toh(array->entry_array.items[0]), t, subtract_one ? (i > 0 ? i-1 : (uint64_t) -1) : i); |
a4bcff5b | 1692 | |
de190aef LP |
1693 | if (subtract_one && i == 0) |
1694 | p = last_p; | |
1695 | else if (subtract_one) | |
1696 | p = le64toh(array->entry_array.items[i-1]); | |
1697 | else | |
1698 | p = le64toh(array->entry_array.items[i]); | |
1699 | ||
1700 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1701 | if (r < 0) | |
1702 | return r; | |
1703 | ||
1704 | if (ret) | |
1705 | *ret = o; | |
1706 | ||
1707 | if (offset) | |
1708 | *offset = p; | |
1709 | ||
1710 | if (idx) | |
cbdca852 | 1711 | *idx = t + i + (subtract_one ? -1 : 0); |
de190aef LP |
1712 | |
1713 | return 1; | |
cec736d2 LP |
1714 | } |
1715 | ||
f268980d LP |
1716 | |
1717 | static int generic_array_bisect_plus_one( | |
1718 | JournalFile *f, | |
1719 | uint64_t extra, | |
1720 | uint64_t first, | |
1721 | uint64_t n, | |
1722 | uint64_t needle, | |
1723 | int (*test_object)(JournalFile *f, uint64_t p, uint64_t needle), | |
1724 | direction_t direction, | |
1725 | Object **ret, | |
1726 | uint64_t *offset, | |
1727 | uint64_t *idx) { | |
de190aef | 1728 | |
cec736d2 | 1729 | int r; |
cbdca852 LP |
1730 | bool step_back = false; |
1731 | Object *o; | |
cec736d2 LP |
1732 | |
1733 | assert(f); | |
de190aef | 1734 | assert(test_object); |
cec736d2 | 1735 | |
de190aef LP |
1736 | if (n <= 0) |
1737 | return 0; | |
cec736d2 | 1738 | |
de190aef LP |
1739 | /* This bisects the array in object 'first', but first checks |
1740 | * an extra */ | |
de190aef LP |
1741 | r = test_object(f, extra, needle); |
1742 | if (r < 0) | |
1743 | return r; | |
a536e261 LP |
1744 | |
1745 | if (r == TEST_FOUND) | |
1746 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1747 | ||
cbdca852 LP |
1748 | /* if we are looking with DIRECTION_UP then we need to first |
1749 | see if in the actual array there is a matching entry, and | |
1750 | return the last one of that. But if there isn't any we need | |
1751 | to return this one. Hence remember this, and return it | |
1752 | below. */ | |
1753 | if (r == TEST_LEFT) | |
1754 | step_back = direction == DIRECTION_UP; | |
de190aef | 1755 | |
cbdca852 LP |
1756 | if (r == TEST_RIGHT) { |
1757 | if (direction == DIRECTION_DOWN) | |
1758 | goto found; | |
1759 | else | |
1760 | return 0; | |
a536e261 | 1761 | } |
cec736d2 | 1762 | |
de190aef LP |
1763 | r = generic_array_bisect(f, first, n-1, needle, test_object, direction, ret, offset, idx); |
1764 | ||
cbdca852 LP |
1765 | if (r == 0 && step_back) |
1766 | goto found; | |
1767 | ||
ecf68b1d | 1768 | if (r > 0 && idx) |
de190aef LP |
1769 | (*idx) ++; |
1770 | ||
1771 | return r; | |
cbdca852 LP |
1772 | |
1773 | found: | |
1774 | r = journal_file_move_to_object(f, OBJECT_ENTRY, extra, &o); | |
1775 | if (r < 0) | |
1776 | return r; | |
1777 | ||
1778 | if (ret) | |
1779 | *ret = o; | |
1780 | ||
1781 | if (offset) | |
1782 | *offset = extra; | |
1783 | ||
1784 | if (idx) | |
1785 | *idx = 0; | |
1786 | ||
1787 | return 1; | |
1788 | } | |
1789 | ||
44a6b1b6 | 1790 | _pure_ static int test_object_offset(JournalFile *f, uint64_t p, uint64_t needle) { |
cbdca852 LP |
1791 | assert(f); |
1792 | assert(p > 0); | |
1793 | ||
1794 | if (p == needle) | |
1795 | return TEST_FOUND; | |
1796 | else if (p < needle) | |
1797 | return TEST_LEFT; | |
1798 | else | |
1799 | return TEST_RIGHT; | |
1800 | } | |
1801 | ||
de190aef LP |
1802 | static int test_object_seqnum(JournalFile *f, uint64_t p, uint64_t needle) { |
1803 | Object *o; | |
1804 | int r; | |
1805 | ||
1806 | assert(f); | |
1807 | assert(p > 0); | |
1808 | ||
1809 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
cec736d2 LP |
1810 | if (r < 0) |
1811 | return r; | |
1812 | ||
de190aef LP |
1813 | if (le64toh(o->entry.seqnum) == needle) |
1814 | return TEST_FOUND; | |
1815 | else if (le64toh(o->entry.seqnum) < needle) | |
1816 | return TEST_LEFT; | |
1817 | else | |
1818 | return TEST_RIGHT; | |
1819 | } | |
cec736d2 | 1820 | |
de190aef LP |
1821 | int journal_file_move_to_entry_by_seqnum( |
1822 | JournalFile *f, | |
1823 | uint64_t seqnum, | |
1824 | direction_t direction, | |
1825 | Object **ret, | |
1826 | uint64_t *offset) { | |
1827 | ||
1828 | return generic_array_bisect(f, | |
1829 | le64toh(f->header->entry_array_offset), | |
1830 | le64toh(f->header->n_entries), | |
1831 | seqnum, | |
1832 | test_object_seqnum, | |
1833 | direction, | |
1834 | ret, offset, NULL); | |
1835 | } | |
cec736d2 | 1836 | |
de190aef LP |
1837 | static int test_object_realtime(JournalFile *f, uint64_t p, uint64_t needle) { |
1838 | Object *o; | |
1839 | int r; | |
1840 | ||
1841 | assert(f); | |
1842 | assert(p > 0); | |
1843 | ||
1844 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1845 | if (r < 0) | |
1846 | return r; | |
1847 | ||
1848 | if (le64toh(o->entry.realtime) == needle) | |
1849 | return TEST_FOUND; | |
1850 | else if (le64toh(o->entry.realtime) < needle) | |
1851 | return TEST_LEFT; | |
1852 | else | |
1853 | return TEST_RIGHT; | |
cec736d2 LP |
1854 | } |
1855 | ||
de190aef LP |
1856 | int journal_file_move_to_entry_by_realtime( |
1857 | JournalFile *f, | |
1858 | uint64_t realtime, | |
1859 | direction_t direction, | |
1860 | Object **ret, | |
1861 | uint64_t *offset) { | |
1862 | ||
1863 | return generic_array_bisect(f, | |
1864 | le64toh(f->header->entry_array_offset), | |
1865 | le64toh(f->header->n_entries), | |
1866 | realtime, | |
1867 | test_object_realtime, | |
1868 | direction, | |
1869 | ret, offset, NULL); | |
1870 | } | |
1871 | ||
1872 | static int test_object_monotonic(JournalFile *f, uint64_t p, uint64_t needle) { | |
1873 | Object *o; | |
1874 | int r; | |
1875 | ||
1876 | assert(f); | |
1877 | assert(p > 0); | |
1878 | ||
1879 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1880 | if (r < 0) | |
1881 | return r; | |
1882 | ||
1883 | if (le64toh(o->entry.monotonic) == needle) | |
1884 | return TEST_FOUND; | |
1885 | else if (le64toh(o->entry.monotonic) < needle) | |
1886 | return TEST_LEFT; | |
1887 | else | |
1888 | return TEST_RIGHT; | |
1889 | } | |
1890 | ||
47838ab3 ZJS |
1891 | static inline int find_data_object_by_boot_id( |
1892 | JournalFile *f, | |
1893 | sd_id128_t boot_id, | |
1894 | Object **o, | |
1895 | uint64_t *b) { | |
1896 | char t[sizeof("_BOOT_ID=")-1 + 32 + 1] = "_BOOT_ID="; | |
1897 | ||
1898 | sd_id128_to_string(boot_id, t + 9); | |
1899 | return journal_file_find_data_object(f, t, sizeof(t) - 1, o, b); | |
1900 | } | |
1901 | ||
de190aef LP |
1902 | int journal_file_move_to_entry_by_monotonic( |
1903 | JournalFile *f, | |
1904 | sd_id128_t boot_id, | |
1905 | uint64_t monotonic, | |
1906 | direction_t direction, | |
1907 | Object **ret, | |
1908 | uint64_t *offset) { | |
1909 | ||
de190aef LP |
1910 | Object *o; |
1911 | int r; | |
1912 | ||
cbdca852 | 1913 | assert(f); |
de190aef | 1914 | |
47838ab3 | 1915 | r = find_data_object_by_boot_id(f, boot_id, &o, NULL); |
de190aef LP |
1916 | if (r < 0) |
1917 | return r; | |
cbdca852 | 1918 | if (r == 0) |
de190aef LP |
1919 | return -ENOENT; |
1920 | ||
1921 | return generic_array_bisect_plus_one(f, | |
1922 | le64toh(o->data.entry_offset), | |
1923 | le64toh(o->data.entry_array_offset), | |
1924 | le64toh(o->data.n_entries), | |
1925 | monotonic, | |
1926 | test_object_monotonic, | |
1927 | direction, | |
1928 | ret, offset, NULL); | |
1929 | } | |
1930 | ||
1fc605b0 MS |
1931 | void journal_file_reset_location(JournalFile *f) { |
1932 | f->current_offset = 0; | |
1933 | } | |
1934 | ||
de190aef LP |
1935 | int journal_file_next_entry( |
1936 | JournalFile *f, | |
1937 | Object *o, uint64_t p, | |
1938 | direction_t direction, | |
1939 | Object **ret, uint64_t *offset) { | |
1940 | ||
fb099c8d | 1941 | uint64_t i, n, ofs; |
cec736d2 LP |
1942 | int r; |
1943 | ||
1944 | assert(f); | |
de190aef LP |
1945 | assert(p > 0 || !o); |
1946 | ||
1947 | n = le64toh(f->header->n_entries); | |
1948 | if (n <= 0) | |
1949 | return 0; | |
cec736d2 LP |
1950 | |
1951 | if (!o) | |
de190aef | 1952 | i = direction == DIRECTION_DOWN ? 0 : n - 1; |
cec736d2 | 1953 | else { |
de190aef | 1954 | if (o->object.type != OBJECT_ENTRY) |
cec736d2 LP |
1955 | return -EINVAL; |
1956 | ||
de190aef LP |
1957 | r = generic_array_bisect(f, |
1958 | le64toh(f->header->entry_array_offset), | |
1959 | le64toh(f->header->n_entries), | |
1960 | p, | |
1961 | test_object_offset, | |
1962 | DIRECTION_DOWN, | |
1963 | NULL, NULL, | |
1964 | &i); | |
1965 | if (r <= 0) | |
1966 | return r; | |
1967 | ||
1968 | if (direction == DIRECTION_DOWN) { | |
1969 | if (i >= n - 1) | |
1970 | return 0; | |
1971 | ||
1972 | i++; | |
1973 | } else { | |
1974 | if (i <= 0) | |
1975 | return 0; | |
1976 | ||
1977 | i--; | |
1978 | } | |
cec736d2 LP |
1979 | } |
1980 | ||
de190aef | 1981 | /* And jump to it */ |
fb099c8d ZJS |
1982 | r = generic_array_get(f, |
1983 | le64toh(f->header->entry_array_offset), | |
1984 | i, | |
1985 | ret, &ofs); | |
1986 | if (r <= 0) | |
1987 | return r; | |
1988 | ||
1989 | if (p > 0 && | |
1990 | (direction == DIRECTION_DOWN ? ofs <= p : ofs >= p)) { | |
1991 | log_debug("%s: entry array corrupted at entry %"PRIu64, | |
1992 | f->path, i); | |
1993 | return -EBADMSG; | |
1994 | } | |
1995 | ||
1996 | if (offset) | |
1997 | *offset = ofs; | |
1998 | ||
1999 | return 1; | |
de190aef | 2000 | } |
cec736d2 | 2001 | |
de190aef LP |
2002 | int journal_file_next_entry_for_data( |
2003 | JournalFile *f, | |
2004 | Object *o, uint64_t p, | |
2005 | uint64_t data_offset, | |
2006 | direction_t direction, | |
2007 | Object **ret, uint64_t *offset) { | |
2008 | ||
2009 | uint64_t n, i; | |
cec736d2 | 2010 | int r; |
de190aef | 2011 | Object *d; |
cec736d2 LP |
2012 | |
2013 | assert(f); | |
de190aef | 2014 | assert(p > 0 || !o); |
cec736d2 | 2015 | |
de190aef | 2016 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); |
466ccd92 | 2017 | if (r < 0) |
de190aef | 2018 | return r; |
cec736d2 | 2019 | |
de190aef LP |
2020 | n = le64toh(d->data.n_entries); |
2021 | if (n <= 0) | |
2022 | return n; | |
cec736d2 | 2023 | |
de190aef LP |
2024 | if (!o) |
2025 | i = direction == DIRECTION_DOWN ? 0 : n - 1; | |
2026 | else { | |
2027 | if (o->object.type != OBJECT_ENTRY) | |
2028 | return -EINVAL; | |
cec736d2 | 2029 | |
de190aef LP |
2030 | r = generic_array_bisect_plus_one(f, |
2031 | le64toh(d->data.entry_offset), | |
2032 | le64toh(d->data.entry_array_offset), | |
2033 | le64toh(d->data.n_entries), | |
2034 | p, | |
2035 | test_object_offset, | |
2036 | DIRECTION_DOWN, | |
2037 | NULL, NULL, | |
2038 | &i); | |
2039 | ||
2040 | if (r <= 0) | |
cec736d2 LP |
2041 | return r; |
2042 | ||
de190aef LP |
2043 | if (direction == DIRECTION_DOWN) { |
2044 | if (i >= n - 1) | |
2045 | return 0; | |
cec736d2 | 2046 | |
de190aef LP |
2047 | i++; |
2048 | } else { | |
2049 | if (i <= 0) | |
2050 | return 0; | |
cec736d2 | 2051 | |
de190aef LP |
2052 | i--; |
2053 | } | |
cec736d2 | 2054 | |
de190aef | 2055 | } |
cec736d2 | 2056 | |
de190aef LP |
2057 | return generic_array_get_plus_one(f, |
2058 | le64toh(d->data.entry_offset), | |
2059 | le64toh(d->data.entry_array_offset), | |
2060 | i, | |
2061 | ret, offset); | |
2062 | } | |
cec736d2 | 2063 | |
cbdca852 LP |
2064 | int journal_file_move_to_entry_by_offset_for_data( |
2065 | JournalFile *f, | |
2066 | uint64_t data_offset, | |
2067 | uint64_t p, | |
2068 | direction_t direction, | |
2069 | Object **ret, uint64_t *offset) { | |
2070 | ||
2071 | int r; | |
2072 | Object *d; | |
2073 | ||
2074 | assert(f); | |
2075 | ||
2076 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); | |
2077 | if (r < 0) | |
2078 | return r; | |
2079 | ||
2080 | return generic_array_bisect_plus_one(f, | |
2081 | le64toh(d->data.entry_offset), | |
2082 | le64toh(d->data.entry_array_offset), | |
2083 | le64toh(d->data.n_entries), | |
2084 | p, | |
2085 | test_object_offset, | |
2086 | direction, | |
2087 | ret, offset, NULL); | |
2088 | } | |
2089 | ||
2090 | int journal_file_move_to_entry_by_monotonic_for_data( | |
2091 | JournalFile *f, | |
2092 | uint64_t data_offset, | |
2093 | sd_id128_t boot_id, | |
2094 | uint64_t monotonic, | |
2095 | direction_t direction, | |
2096 | Object **ret, uint64_t *offset) { | |
2097 | ||
cbdca852 LP |
2098 | Object *o, *d; |
2099 | int r; | |
2100 | uint64_t b, z; | |
2101 | ||
2102 | assert(f); | |
2103 | ||
2104 | /* First, seek by time */ | |
47838ab3 | 2105 | r = find_data_object_by_boot_id(f, boot_id, &o, &b); |
cbdca852 LP |
2106 | if (r < 0) |
2107 | return r; | |
2108 | if (r == 0) | |
2109 | return -ENOENT; | |
2110 | ||
2111 | r = generic_array_bisect_plus_one(f, | |
2112 | le64toh(o->data.entry_offset), | |
2113 | le64toh(o->data.entry_array_offset), | |
2114 | le64toh(o->data.n_entries), | |
2115 | monotonic, | |
2116 | test_object_monotonic, | |
2117 | direction, | |
2118 | NULL, &z, NULL); | |
2119 | if (r <= 0) | |
2120 | return r; | |
2121 | ||
2122 | /* And now, continue seeking until we find an entry that | |
2123 | * exists in both bisection arrays */ | |
2124 | ||
2125 | for (;;) { | |
2126 | Object *qo; | |
2127 | uint64_t p, q; | |
2128 | ||
2129 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); | |
2130 | if (r < 0) | |
2131 | return r; | |
2132 | ||
2133 | r = generic_array_bisect_plus_one(f, | |
2134 | le64toh(d->data.entry_offset), | |
2135 | le64toh(d->data.entry_array_offset), | |
2136 | le64toh(d->data.n_entries), | |
2137 | z, | |
2138 | test_object_offset, | |
2139 | direction, | |
2140 | NULL, &p, NULL); | |
2141 | if (r <= 0) | |
2142 | return r; | |
2143 | ||
2144 | r = journal_file_move_to_object(f, OBJECT_DATA, b, &o); | |
2145 | if (r < 0) | |
2146 | return r; | |
2147 | ||
2148 | r = generic_array_bisect_plus_one(f, | |
2149 | le64toh(o->data.entry_offset), | |
2150 | le64toh(o->data.entry_array_offset), | |
2151 | le64toh(o->data.n_entries), | |
2152 | p, | |
2153 | test_object_offset, | |
2154 | direction, | |
2155 | &qo, &q, NULL); | |
2156 | ||
2157 | if (r <= 0) | |
2158 | return r; | |
2159 | ||
2160 | if (p == q) { | |
2161 | if (ret) | |
2162 | *ret = qo; | |
2163 | if (offset) | |
2164 | *offset = q; | |
2165 | ||
2166 | return 1; | |
2167 | } | |
2168 | ||
2169 | z = q; | |
2170 | } | |
cbdca852 LP |
2171 | } |
2172 | ||
de190aef LP |
2173 | int journal_file_move_to_entry_by_seqnum_for_data( |
2174 | JournalFile *f, | |
2175 | uint64_t data_offset, | |
2176 | uint64_t seqnum, | |
2177 | direction_t direction, | |
2178 | Object **ret, uint64_t *offset) { | |
cec736d2 | 2179 | |
de190aef LP |
2180 | Object *d; |
2181 | int r; | |
cec736d2 | 2182 | |
91a31dde LP |
2183 | assert(f); |
2184 | ||
de190aef | 2185 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); |
91a31dde | 2186 | if (r < 0) |
de190aef | 2187 | return r; |
cec736d2 | 2188 | |
de190aef LP |
2189 | return generic_array_bisect_plus_one(f, |
2190 | le64toh(d->data.entry_offset), | |
2191 | le64toh(d->data.entry_array_offset), | |
2192 | le64toh(d->data.n_entries), | |
2193 | seqnum, | |
2194 | test_object_seqnum, | |
2195 | direction, | |
2196 | ret, offset, NULL); | |
2197 | } | |
cec736d2 | 2198 | |
de190aef LP |
2199 | int journal_file_move_to_entry_by_realtime_for_data( |
2200 | JournalFile *f, | |
2201 | uint64_t data_offset, | |
2202 | uint64_t realtime, | |
2203 | direction_t direction, | |
2204 | Object **ret, uint64_t *offset) { | |
2205 | ||
2206 | Object *d; | |
2207 | int r; | |
2208 | ||
91a31dde LP |
2209 | assert(f); |
2210 | ||
de190aef | 2211 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); |
91a31dde | 2212 | if (r < 0) |
de190aef LP |
2213 | return r; |
2214 | ||
2215 | return generic_array_bisect_plus_one(f, | |
2216 | le64toh(d->data.entry_offset), | |
2217 | le64toh(d->data.entry_array_offset), | |
2218 | le64toh(d->data.n_entries), | |
2219 | realtime, | |
2220 | test_object_realtime, | |
2221 | direction, | |
2222 | ret, offset, NULL); | |
cec736d2 LP |
2223 | } |
2224 | ||
0284adc6 | 2225 | void journal_file_dump(JournalFile *f) { |
7560fffc | 2226 | Object *o; |
7560fffc | 2227 | int r; |
0284adc6 | 2228 | uint64_t p; |
7560fffc LP |
2229 | |
2230 | assert(f); | |
2231 | ||
0284adc6 | 2232 | journal_file_print_header(f); |
7560fffc | 2233 | |
0284adc6 LP |
2234 | p = le64toh(f->header->header_size); |
2235 | while (p != 0) { | |
d05089d8 | 2236 | r = journal_file_move_to_object(f, OBJECT_UNUSED, p, &o); |
0284adc6 LP |
2237 | if (r < 0) |
2238 | goto fail; | |
7560fffc | 2239 | |
0284adc6 | 2240 | switch (o->object.type) { |
d98cc1f2 | 2241 | |
0284adc6 LP |
2242 | case OBJECT_UNUSED: |
2243 | printf("Type: OBJECT_UNUSED\n"); | |
2244 | break; | |
d98cc1f2 | 2245 | |
0284adc6 LP |
2246 | case OBJECT_DATA: |
2247 | printf("Type: OBJECT_DATA\n"); | |
2248 | break; | |
7560fffc | 2249 | |
3c1668da LP |
2250 | case OBJECT_FIELD: |
2251 | printf("Type: OBJECT_FIELD\n"); | |
2252 | break; | |
2253 | ||
0284adc6 | 2254 | case OBJECT_ENTRY: |
507f22bd ZJS |
2255 | printf("Type: OBJECT_ENTRY seqnum=%"PRIu64" monotonic=%"PRIu64" realtime=%"PRIu64"\n", |
2256 | le64toh(o->entry.seqnum), | |
2257 | le64toh(o->entry.monotonic), | |
2258 | le64toh(o->entry.realtime)); | |
0284adc6 | 2259 | break; |
7560fffc | 2260 | |
0284adc6 LP |
2261 | case OBJECT_FIELD_HASH_TABLE: |
2262 | printf("Type: OBJECT_FIELD_HASH_TABLE\n"); | |
2263 | break; | |
7560fffc | 2264 | |
0284adc6 LP |
2265 | case OBJECT_DATA_HASH_TABLE: |
2266 | printf("Type: OBJECT_DATA_HASH_TABLE\n"); | |
2267 | break; | |
7560fffc | 2268 | |
0284adc6 LP |
2269 | case OBJECT_ENTRY_ARRAY: |
2270 | printf("Type: OBJECT_ENTRY_ARRAY\n"); | |
2271 | break; | |
7560fffc | 2272 | |
0284adc6 | 2273 | case OBJECT_TAG: |
507f22bd ZJS |
2274 | printf("Type: OBJECT_TAG seqnum=%"PRIu64" epoch=%"PRIu64"\n", |
2275 | le64toh(o->tag.seqnum), | |
2276 | le64toh(o->tag.epoch)); | |
0284adc6 | 2277 | break; |
3c1668da LP |
2278 | |
2279 | default: | |
2280 | printf("Type: unknown (%u)\n", o->object.type); | |
2281 | break; | |
0284adc6 | 2282 | } |
7560fffc | 2283 | |
d89c8fdf ZJS |
2284 | if (o->object.flags & OBJECT_COMPRESSION_MASK) |
2285 | printf("Flags: %s\n", | |
2286 | object_compressed_to_string(o->object.flags & OBJECT_COMPRESSION_MASK)); | |
7560fffc | 2287 | |
0284adc6 LP |
2288 | if (p == le64toh(f->header->tail_object_offset)) |
2289 | p = 0; | |
2290 | else | |
2291 | p = p + ALIGN64(le64toh(o->object.size)); | |
2292 | } | |
7560fffc | 2293 | |
0284adc6 LP |
2294 | return; |
2295 | fail: | |
2296 | log_error("File corrupt"); | |
7560fffc LP |
2297 | } |
2298 | ||
718fe4b1 ZJS |
2299 | static const char* format_timestamp_safe(char *buf, size_t l, usec_t t) { |
2300 | const char *x; | |
2301 | ||
2302 | x = format_timestamp(buf, l, t); | |
2303 | if (x) | |
2304 | return x; | |
2305 | return " --- "; | |
2306 | } | |
2307 | ||
0284adc6 | 2308 | void journal_file_print_header(JournalFile *f) { |
2765b7bb | 2309 | char a[33], b[33], c[33], d[33]; |
ed375beb | 2310 | char x[FORMAT_TIMESTAMP_MAX], y[FORMAT_TIMESTAMP_MAX], z[FORMAT_TIMESTAMP_MAX]; |
a1a03e30 LP |
2311 | struct stat st; |
2312 | char bytes[FORMAT_BYTES_MAX]; | |
7560fffc LP |
2313 | |
2314 | assert(f); | |
7560fffc | 2315 | |
0284adc6 LP |
2316 | printf("File Path: %s\n" |
2317 | "File ID: %s\n" | |
2318 | "Machine ID: %s\n" | |
2319 | "Boot ID: %s\n" | |
2320 | "Sequential Number ID: %s\n" | |
2321 | "State: %s\n" | |
2322 | "Compatible Flags:%s%s\n" | |
d89c8fdf | 2323 | "Incompatible Flags:%s%s%s\n" |
507f22bd ZJS |
2324 | "Header size: %"PRIu64"\n" |
2325 | "Arena size: %"PRIu64"\n" | |
2326 | "Data Hash Table Size: %"PRIu64"\n" | |
2327 | "Field Hash Table Size: %"PRIu64"\n" | |
0284adc6 | 2328 | "Rotate Suggested: %s\n" |
507f22bd ZJS |
2329 | "Head Sequential Number: %"PRIu64"\n" |
2330 | "Tail Sequential Number: %"PRIu64"\n" | |
0284adc6 | 2331 | "Head Realtime Timestamp: %s\n" |
3223f44f | 2332 | "Tail Realtime Timestamp: %s\n" |
ed375beb | 2333 | "Tail Monotonic Timestamp: %s\n" |
507f22bd ZJS |
2334 | "Objects: %"PRIu64"\n" |
2335 | "Entry Objects: %"PRIu64"\n", | |
0284adc6 LP |
2336 | f->path, |
2337 | sd_id128_to_string(f->header->file_id, a), | |
2338 | sd_id128_to_string(f->header->machine_id, b), | |
2339 | sd_id128_to_string(f->header->boot_id, c), | |
2765b7bb | 2340 | sd_id128_to_string(f->header->seqnum_id, d), |
3223f44f LP |
2341 | f->header->state == STATE_OFFLINE ? "OFFLINE" : |
2342 | f->header->state == STATE_ONLINE ? "ONLINE" : | |
2343 | f->header->state == STATE_ARCHIVED ? "ARCHIVED" : "UNKNOWN", | |
8088cbd3 | 2344 | JOURNAL_HEADER_SEALED(f->header) ? " SEALED" : "", |
d89c8fdf ZJS |
2345 | (le32toh(f->header->compatible_flags) & ~HEADER_COMPATIBLE_ANY) ? " ???" : "", |
2346 | JOURNAL_HEADER_COMPRESSED_XZ(f->header) ? " COMPRESSED-XZ" : "", | |
2347 | JOURNAL_HEADER_COMPRESSED_LZ4(f->header) ? " COMPRESSED-LZ4" : "", | |
2348 | (le32toh(f->header->incompatible_flags) & ~HEADER_INCOMPATIBLE_ANY) ? " ???" : "", | |
507f22bd ZJS |
2349 | le64toh(f->header->header_size), |
2350 | le64toh(f->header->arena_size), | |
2351 | le64toh(f->header->data_hash_table_size) / sizeof(HashItem), | |
2352 | le64toh(f->header->field_hash_table_size) / sizeof(HashItem), | |
fb0951b0 | 2353 | yes_no(journal_file_rotate_suggested(f, 0)), |
507f22bd ZJS |
2354 | le64toh(f->header->head_entry_seqnum), |
2355 | le64toh(f->header->tail_entry_seqnum), | |
718fe4b1 ZJS |
2356 | format_timestamp_safe(x, sizeof(x), le64toh(f->header->head_entry_realtime)), |
2357 | format_timestamp_safe(y, sizeof(y), le64toh(f->header->tail_entry_realtime)), | |
ed375beb | 2358 | format_timespan(z, sizeof(z), le64toh(f->header->tail_entry_monotonic), USEC_PER_MSEC), |
507f22bd ZJS |
2359 | le64toh(f->header->n_objects), |
2360 | le64toh(f->header->n_entries)); | |
7560fffc | 2361 | |
0284adc6 | 2362 | if (JOURNAL_HEADER_CONTAINS(f->header, n_data)) |
507f22bd | 2363 | printf("Data Objects: %"PRIu64"\n" |
0284adc6 | 2364 | "Data Hash Table Fill: %.1f%%\n", |
507f22bd | 2365 | le64toh(f->header->n_data), |
0284adc6 | 2366 | 100.0 * (double) le64toh(f->header->n_data) / ((double) (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)))); |
7560fffc | 2367 | |
0284adc6 | 2368 | if (JOURNAL_HEADER_CONTAINS(f->header, n_fields)) |
507f22bd | 2369 | printf("Field Objects: %"PRIu64"\n" |
0284adc6 | 2370 | "Field Hash Table Fill: %.1f%%\n", |
507f22bd | 2371 | le64toh(f->header->n_fields), |
0284adc6 | 2372 | 100.0 * (double) le64toh(f->header->n_fields) / ((double) (le64toh(f->header->field_hash_table_size) / sizeof(HashItem)))); |
3223f44f LP |
2373 | |
2374 | if (JOURNAL_HEADER_CONTAINS(f->header, n_tags)) | |
507f22bd ZJS |
2375 | printf("Tag Objects: %"PRIu64"\n", |
2376 | le64toh(f->header->n_tags)); | |
3223f44f | 2377 | if (JOURNAL_HEADER_CONTAINS(f->header, n_entry_arrays)) |
507f22bd ZJS |
2378 | printf("Entry Array Objects: %"PRIu64"\n", |
2379 | le64toh(f->header->n_entry_arrays)); | |
a1a03e30 LP |
2380 | |
2381 | if (fstat(f->fd, &st) >= 0) | |
2382 | printf("Disk usage: %s\n", format_bytes(bytes, sizeof(bytes), (off_t) st.st_blocks * 512ULL)); | |
7560fffc LP |
2383 | } |
2384 | ||
0284adc6 LP |
2385 | int journal_file_open( |
2386 | const char *fname, | |
2387 | int flags, | |
2388 | mode_t mode, | |
2389 | bool compress, | |
baed47c3 | 2390 | bool seal, |
0284adc6 LP |
2391 | JournalMetrics *metrics, |
2392 | MMapCache *mmap_cache, | |
2393 | JournalFile *template, | |
2394 | JournalFile **ret) { | |
7560fffc | 2395 | |
0284adc6 LP |
2396 | JournalFile *f; |
2397 | int r; | |
2398 | bool newly_created = false; | |
7560fffc | 2399 | |
0284adc6 | 2400 | assert(fname); |
0559d3a5 | 2401 | assert(ret); |
7560fffc | 2402 | |
0284adc6 LP |
2403 | if ((flags & O_ACCMODE) != O_RDONLY && |
2404 | (flags & O_ACCMODE) != O_RDWR) | |
2405 | return -EINVAL; | |
7560fffc | 2406 | |
a0108012 LP |
2407 | if (!endswith(fname, ".journal") && |
2408 | !endswith(fname, ".journal~")) | |
0284adc6 | 2409 | return -EINVAL; |
7560fffc | 2410 | |
0284adc6 LP |
2411 | f = new0(JournalFile, 1); |
2412 | if (!f) | |
2413 | return -ENOMEM; | |
7560fffc | 2414 | |
0284adc6 LP |
2415 | f->fd = -1; |
2416 | f->mode = mode; | |
7560fffc | 2417 | |
0284adc6 LP |
2418 | f->flags = flags; |
2419 | f->prot = prot_from_flags(flags); | |
2420 | f->writable = (flags & O_ACCMODE) != O_RDONLY; | |
92261977 | 2421 | #if defined(HAVE_LZ4) |
d89c8fdf ZJS |
2422 | f->compress_lz4 = compress; |
2423 | #elif defined(HAVE_XZ) | |
2424 | f->compress_xz = compress; | |
48b61739 | 2425 | #endif |
49a32d43 | 2426 | #ifdef HAVE_GCRYPT |
baed47c3 | 2427 | f->seal = seal; |
49a32d43 | 2428 | #endif |
7560fffc | 2429 | |
0284adc6 LP |
2430 | if (mmap_cache) |
2431 | f->mmap = mmap_cache_ref(mmap_cache); | |
2432 | else { | |
84168d80 | 2433 | f->mmap = mmap_cache_new(); |
0284adc6 LP |
2434 | if (!f->mmap) { |
2435 | r = -ENOMEM; | |
2436 | goto fail; | |
2437 | } | |
2438 | } | |
7560fffc | 2439 | |
0284adc6 LP |
2440 | f->path = strdup(fname); |
2441 | if (!f->path) { | |
2442 | r = -ENOMEM; | |
2443 | goto fail; | |
2444 | } | |
7560fffc | 2445 | |
4743015d | 2446 | f->chain_cache = ordered_hashmap_new(&uint64_hash_ops); |
a4bcff5b LP |
2447 | if (!f->chain_cache) { |
2448 | r = -ENOMEM; | |
2449 | goto fail; | |
2450 | } | |
2451 | ||
0284adc6 LP |
2452 | f->fd = open(f->path, f->flags|O_CLOEXEC, f->mode); |
2453 | if (f->fd < 0) { | |
2454 | r = -errno; | |
2455 | goto fail; | |
7560fffc | 2456 | } |
7560fffc | 2457 | |
0284adc6 LP |
2458 | if (fstat(f->fd, &f->last_stat) < 0) { |
2459 | r = -errno; | |
2460 | goto fail; | |
2461 | } | |
7560fffc | 2462 | |
0284adc6 | 2463 | if (f->last_stat.st_size == 0 && f->writable) { |
fb0951b0 LP |
2464 | uint64_t crtime; |
2465 | ||
2466 | /* Let's attach the creation time to the journal file, | |
2467 | * so that the vacuuming code knows the age of this | |
2468 | * file even if the file might end up corrupted one | |
2469 | * day... Ideally we'd just use the creation time many | |
2470 | * file systems maintain for each file, but there is | |
2471 | * currently no usable API to query this, hence let's | |
2472 | * emulate this via extended attributes. If extended | |
2473 | * attributes are not supported we'll just skip this, | |
7517e174 | 2474 | * and rely solely on mtime/atime/ctime of the file. */ |
fb0951b0 LP |
2475 | |
2476 | crtime = htole64((uint64_t) now(CLOCK_REALTIME)); | |
2477 | fsetxattr(f->fd, "user.crtime_usec", &crtime, sizeof(crtime), XATTR_CREATE); | |
7560fffc | 2478 | |
feb12d3e | 2479 | #ifdef HAVE_GCRYPT |
0284adc6 | 2480 | /* Try to load the FSPRG state, and if we can't, then |
baed47c3 | 2481 | * just don't do sealing */ |
49a32d43 LP |
2482 | if (f->seal) { |
2483 | r = journal_file_fss_load(f); | |
2484 | if (r < 0) | |
2485 | f->seal = false; | |
2486 | } | |
feb12d3e | 2487 | #endif |
7560fffc | 2488 | |
0284adc6 LP |
2489 | r = journal_file_init_header(f, template); |
2490 | if (r < 0) | |
2491 | goto fail; | |
7560fffc | 2492 | |
0284adc6 LP |
2493 | if (fstat(f->fd, &f->last_stat) < 0) { |
2494 | r = -errno; | |
2495 | goto fail; | |
2496 | } | |
fb0951b0 LP |
2497 | |
2498 | newly_created = true; | |
0284adc6 | 2499 | } |
7560fffc | 2500 | |
0284adc6 LP |
2501 | if (f->last_stat.st_size < (off_t) HEADER_SIZE_MIN) { |
2502 | r = -EIO; | |
2503 | goto fail; | |
2504 | } | |
7560fffc | 2505 | |
0284adc6 LP |
2506 | f->header = mmap(NULL, PAGE_ALIGN(sizeof(Header)), prot_from_flags(flags), MAP_SHARED, f->fd, 0); |
2507 | if (f->header == MAP_FAILED) { | |
2508 | f->header = NULL; | |
2509 | r = -errno; | |
2510 | goto fail; | |
2511 | } | |
7560fffc | 2512 | |
0284adc6 LP |
2513 | if (!newly_created) { |
2514 | r = journal_file_verify_header(f); | |
2515 | if (r < 0) | |
2516 | goto fail; | |
2517 | } | |
7560fffc | 2518 | |
feb12d3e | 2519 | #ifdef HAVE_GCRYPT |
0284adc6 | 2520 | if (!newly_created && f->writable) { |
baed47c3 | 2521 | r = journal_file_fss_load(f); |
0284adc6 LP |
2522 | if (r < 0) |
2523 | goto fail; | |
2524 | } | |
feb12d3e | 2525 | #endif |
cec736d2 LP |
2526 | |
2527 | if (f->writable) { | |
4a92baf3 LP |
2528 | if (metrics) { |
2529 | journal_default_metrics(metrics, f->fd); | |
2530 | f->metrics = *metrics; | |
2531 | } else if (template) | |
2532 | f->metrics = template->metrics; | |
2533 | ||
cec736d2 LP |
2534 | r = journal_file_refresh_header(f); |
2535 | if (r < 0) | |
2536 | goto fail; | |
2537 | } | |
2538 | ||
feb12d3e | 2539 | #ifdef HAVE_GCRYPT |
baed47c3 | 2540 | r = journal_file_hmac_setup(f); |
14d10188 LP |
2541 | if (r < 0) |
2542 | goto fail; | |
feb12d3e | 2543 | #endif |
14d10188 | 2544 | |
cec736d2 | 2545 | if (newly_created) { |
de190aef | 2546 | r = journal_file_setup_field_hash_table(f); |
cec736d2 LP |
2547 | if (r < 0) |
2548 | goto fail; | |
2549 | ||
de190aef | 2550 | r = journal_file_setup_data_hash_table(f); |
cec736d2 LP |
2551 | if (r < 0) |
2552 | goto fail; | |
7560fffc | 2553 | |
feb12d3e | 2554 | #ifdef HAVE_GCRYPT |
7560fffc LP |
2555 | r = journal_file_append_first_tag(f); |
2556 | if (r < 0) | |
2557 | goto fail; | |
feb12d3e | 2558 | #endif |
cec736d2 LP |
2559 | } |
2560 | ||
de190aef | 2561 | r = journal_file_map_field_hash_table(f); |
cec736d2 LP |
2562 | if (r < 0) |
2563 | goto fail; | |
2564 | ||
de190aef | 2565 | r = journal_file_map_data_hash_table(f); |
cec736d2 LP |
2566 | if (r < 0) |
2567 | goto fail; | |
2568 | ||
0559d3a5 | 2569 | *ret = f; |
cec736d2 LP |
2570 | return 0; |
2571 | ||
2572 | fail: | |
2573 | journal_file_close(f); | |
2574 | ||
2575 | return r; | |
2576 | } | |
0ac38b70 | 2577 | |
baed47c3 | 2578 | int journal_file_rotate(JournalFile **f, bool compress, bool seal) { |
57535f47 | 2579 | _cleanup_free_ char *p = NULL; |
0ac38b70 LP |
2580 | size_t l; |
2581 | JournalFile *old_file, *new_file = NULL; | |
2582 | int r; | |
2583 | ||
2584 | assert(f); | |
2585 | assert(*f); | |
2586 | ||
2587 | old_file = *f; | |
2588 | ||
2589 | if (!old_file->writable) | |
2590 | return -EINVAL; | |
2591 | ||
2592 | if (!endswith(old_file->path, ".journal")) | |
2593 | return -EINVAL; | |
2594 | ||
2595 | l = strlen(old_file->path); | |
57535f47 ZJS |
2596 | r = asprintf(&p, "%.*s@" SD_ID128_FORMAT_STR "-%016"PRIx64"-%016"PRIx64".journal", |
2597 | (int) l - 8, old_file->path, | |
2598 | SD_ID128_FORMAT_VAL(old_file->header->seqnum_id), | |
2599 | le64toh((*f)->header->head_entry_seqnum), | |
2600 | le64toh((*f)->header->head_entry_realtime)); | |
2601 | if (r < 0) | |
0ac38b70 LP |
2602 | return -ENOMEM; |
2603 | ||
0ac38b70 | 2604 | r = rename(old_file->path, p); |
0ac38b70 LP |
2605 | if (r < 0) |
2606 | return -errno; | |
2607 | ||
ccdbaf91 | 2608 | old_file->header->state = STATE_ARCHIVED; |
0ac38b70 | 2609 | |
baed47c3 | 2610 | r = journal_file_open(old_file->path, old_file->flags, old_file->mode, compress, seal, NULL, old_file->mmap, old_file, &new_file); |
0ac38b70 LP |
2611 | journal_file_close(old_file); |
2612 | ||
2613 | *f = new_file; | |
2614 | return r; | |
2615 | } | |
2616 | ||
9447a7f1 LP |
2617 | int journal_file_open_reliably( |
2618 | const char *fname, | |
2619 | int flags, | |
2620 | mode_t mode, | |
7560fffc | 2621 | bool compress, |
baed47c3 | 2622 | bool seal, |
4a92baf3 | 2623 | JournalMetrics *metrics, |
27370278 | 2624 | MMapCache *mmap_cache, |
9447a7f1 LP |
2625 | JournalFile *template, |
2626 | JournalFile **ret) { | |
2627 | ||
2628 | int r; | |
2629 | size_t l; | |
ed375beb | 2630 | _cleanup_free_ char *p = NULL; |
9447a7f1 | 2631 | |
baed47c3 | 2632 | r = journal_file_open(fname, flags, mode, compress, seal, |
27370278 | 2633 | metrics, mmap_cache, template, ret); |
0071d9f1 LP |
2634 | if (r != -EBADMSG && /* corrupted */ |
2635 | r != -ENODATA && /* truncated */ | |
2636 | r != -EHOSTDOWN && /* other machine */ | |
a1a1898f LP |
2637 | r != -EPROTONOSUPPORT && /* incompatible feature */ |
2638 | r != -EBUSY && /* unclean shutdown */ | |
2639 | r != -ESHUTDOWN /* already archived */) | |
9447a7f1 LP |
2640 | return r; |
2641 | ||
2642 | if ((flags & O_ACCMODE) == O_RDONLY) | |
2643 | return r; | |
2644 | ||
2645 | if (!(flags & O_CREAT)) | |
2646 | return r; | |
2647 | ||
7560fffc LP |
2648 | if (!endswith(fname, ".journal")) |
2649 | return r; | |
2650 | ||
5c70eab4 LP |
2651 | /* The file is corrupted. Rotate it away and try it again (but only once) */ |
2652 | ||
9447a7f1 | 2653 | l = strlen(fname); |
9bf3b535 | 2654 | if (asprintf(&p, "%.*s@%016llx-%016" PRIx64 ".journal~", |
57535f47 | 2655 | (int) l - 8, fname, |
9447a7f1 | 2656 | (unsigned long long) now(CLOCK_REALTIME), |
9bf3b535 | 2657 | random_u64()) < 0) |
9447a7f1 LP |
2658 | return -ENOMEM; |
2659 | ||
2660 | r = rename(fname, p); | |
9447a7f1 LP |
2661 | if (r < 0) |
2662 | return -errno; | |
2663 | ||
a1a1898f | 2664 | log_warning("File %s corrupted or uncleanly shut down, renaming and replacing.", fname); |
9447a7f1 | 2665 | |
baed47c3 | 2666 | return journal_file_open(fname, flags, mode, compress, seal, |
27370278 | 2667 | metrics, mmap_cache, template, ret); |
9447a7f1 LP |
2668 | } |
2669 | ||
cf244689 LP |
2670 | int journal_file_copy_entry(JournalFile *from, JournalFile *to, Object *o, uint64_t p, uint64_t *seqnum, Object **ret, uint64_t *offset) { |
2671 | uint64_t i, n; | |
2672 | uint64_t q, xor_hash = 0; | |
2673 | int r; | |
2674 | EntryItem *items; | |
2675 | dual_timestamp ts; | |
2676 | ||
2677 | assert(from); | |
2678 | assert(to); | |
2679 | assert(o); | |
2680 | assert(p); | |
2681 | ||
2682 | if (!to->writable) | |
2683 | return -EPERM; | |
2684 | ||
2685 | ts.monotonic = le64toh(o->entry.monotonic); | |
2686 | ts.realtime = le64toh(o->entry.realtime); | |
2687 | ||
cf244689 | 2688 | n = journal_file_entry_n_items(o); |
4faa7004 TA |
2689 | /* alloca() can't take 0, hence let's allocate at least one */ |
2690 | items = alloca(sizeof(EntryItem) * MAX(1u, n)); | |
cf244689 LP |
2691 | |
2692 | for (i = 0; i < n; i++) { | |
4fd052ae FC |
2693 | uint64_t l, h; |
2694 | le64_t le_hash; | |
cf244689 LP |
2695 | size_t t; |
2696 | void *data; | |
2697 | Object *u; | |
2698 | ||
2699 | q = le64toh(o->entry.items[i].object_offset); | |
2700 | le_hash = o->entry.items[i].hash; | |
2701 | ||
2702 | r = journal_file_move_to_object(from, OBJECT_DATA, q, &o); | |
2703 | if (r < 0) | |
2704 | return r; | |
2705 | ||
2706 | if (le_hash != o->data.hash) | |
2707 | return -EBADMSG; | |
2708 | ||
2709 | l = le64toh(o->object.size) - offsetof(Object, data.payload); | |
2710 | t = (size_t) l; | |
2711 | ||
2712 | /* We hit the limit on 32bit machines */ | |
2713 | if ((uint64_t) t != l) | |
2714 | return -E2BIG; | |
2715 | ||
d89c8fdf | 2716 | if (o->object.flags & OBJECT_COMPRESSION_MASK) { |
3b1a55e1 | 2717 | #if defined(HAVE_XZ) || defined(HAVE_LZ4) |
fa1c4b51 | 2718 | size_t rsize; |
cf244689 | 2719 | |
d89c8fdf ZJS |
2720 | r = decompress_blob(o->object.flags & OBJECT_COMPRESSION_MASK, |
2721 | o->data.payload, l, &from->compress_buffer, &from->compress_buffer_size, &rsize, 0); | |
2722 | if (r < 0) | |
2723 | return r; | |
cf244689 LP |
2724 | |
2725 | data = from->compress_buffer; | |
2726 | l = rsize; | |
3b1a55e1 ZJS |
2727 | #else |
2728 | return -EPROTONOSUPPORT; | |
2729 | #endif | |
cf244689 LP |
2730 | } else |
2731 | data = o->data.payload; | |
2732 | ||
2733 | r = journal_file_append_data(to, data, l, &u, &h); | |
2734 | if (r < 0) | |
2735 | return r; | |
2736 | ||
2737 | xor_hash ^= le64toh(u->data.hash); | |
2738 | items[i].object_offset = htole64(h); | |
2739 | items[i].hash = u->data.hash; | |
2740 | ||
2741 | r = journal_file_move_to_object(from, OBJECT_ENTRY, p, &o); | |
2742 | if (r < 0) | |
2743 | return r; | |
2744 | } | |
2745 | ||
2746 | return journal_file_append_entry_internal(to, &ts, xor_hash, items, n, seqnum, ret, offset); | |
2747 | } | |
babfc091 LP |
2748 | |
2749 | void journal_default_metrics(JournalMetrics *m, int fd) { | |
2750 | uint64_t fs_size = 0; | |
2751 | struct statvfs ss; | |
a7bc2c2a | 2752 | char a[FORMAT_BYTES_MAX], b[FORMAT_BYTES_MAX], c[FORMAT_BYTES_MAX], d[FORMAT_BYTES_MAX]; |
babfc091 LP |
2753 | |
2754 | assert(m); | |
2755 | assert(fd >= 0); | |
2756 | ||
2757 | if (fstatvfs(fd, &ss) >= 0) | |
2758 | fs_size = ss.f_frsize * ss.f_blocks; | |
2759 | ||
2760 | if (m->max_use == (uint64_t) -1) { | |
2761 | ||
2762 | if (fs_size > 0) { | |
2763 | m->max_use = PAGE_ALIGN(fs_size / 10); /* 10% of file system size */ | |
2764 | ||
2765 | if (m->max_use > DEFAULT_MAX_USE_UPPER) | |
2766 | m->max_use = DEFAULT_MAX_USE_UPPER; | |
2767 | ||
2768 | if (m->max_use < DEFAULT_MAX_USE_LOWER) | |
2769 | m->max_use = DEFAULT_MAX_USE_LOWER; | |
2770 | } else | |
2771 | m->max_use = DEFAULT_MAX_USE_LOWER; | |
2772 | } else { | |
2773 | m->max_use = PAGE_ALIGN(m->max_use); | |
2774 | ||
2775 | if (m->max_use < JOURNAL_FILE_SIZE_MIN*2) | |
2776 | m->max_use = JOURNAL_FILE_SIZE_MIN*2; | |
2777 | } | |
2778 | ||
2779 | if (m->max_size == (uint64_t) -1) { | |
2780 | m->max_size = PAGE_ALIGN(m->max_use / 8); /* 8 chunks */ | |
2781 | ||
2782 | if (m->max_size > DEFAULT_MAX_SIZE_UPPER) | |
2783 | m->max_size = DEFAULT_MAX_SIZE_UPPER; | |
2784 | } else | |
2785 | m->max_size = PAGE_ALIGN(m->max_size); | |
2786 | ||
2787 | if (m->max_size < JOURNAL_FILE_SIZE_MIN) | |
2788 | m->max_size = JOURNAL_FILE_SIZE_MIN; | |
2789 | ||
2790 | if (m->max_size*2 > m->max_use) | |
2791 | m->max_use = m->max_size*2; | |
2792 | ||
2793 | if (m->min_size == (uint64_t) -1) | |
2794 | m->min_size = JOURNAL_FILE_SIZE_MIN; | |
2795 | else { | |
2796 | m->min_size = PAGE_ALIGN(m->min_size); | |
2797 | ||
2798 | if (m->min_size < JOURNAL_FILE_SIZE_MIN) | |
2799 | m->min_size = JOURNAL_FILE_SIZE_MIN; | |
2800 | ||
2801 | if (m->min_size > m->max_size) | |
2802 | m->max_size = m->min_size; | |
2803 | } | |
2804 | ||
2805 | if (m->keep_free == (uint64_t) -1) { | |
2806 | ||
2807 | if (fs_size > 0) { | |
8621b110 | 2808 | m->keep_free = PAGE_ALIGN(fs_size * 3 / 20); /* 15% of file system size */ |
babfc091 LP |
2809 | |
2810 | if (m->keep_free > DEFAULT_KEEP_FREE_UPPER) | |
2811 | m->keep_free = DEFAULT_KEEP_FREE_UPPER; | |
2812 | ||
2813 | } else | |
2814 | m->keep_free = DEFAULT_KEEP_FREE; | |
2815 | } | |
2816 | ||
2b43f939 LP |
2817 | log_debug("Fixed max_use=%s max_size=%s min_size=%s keep_free=%s", |
2818 | format_bytes(a, sizeof(a), m->max_use), | |
2819 | format_bytes(b, sizeof(b), m->max_size), | |
2820 | format_bytes(c, sizeof(c), m->min_size), | |
2821 | format_bytes(d, sizeof(d), m->keep_free)); | |
babfc091 | 2822 | } |
08984293 LP |
2823 | |
2824 | int journal_file_get_cutoff_realtime_usec(JournalFile *f, usec_t *from, usec_t *to) { | |
08984293 LP |
2825 | assert(f); |
2826 | assert(from || to); | |
2827 | ||
2828 | if (from) { | |
162566a4 LP |
2829 | if (f->header->head_entry_realtime == 0) |
2830 | return -ENOENT; | |
08984293 | 2831 | |
162566a4 | 2832 | *from = le64toh(f->header->head_entry_realtime); |
08984293 LP |
2833 | } |
2834 | ||
2835 | if (to) { | |
162566a4 LP |
2836 | if (f->header->tail_entry_realtime == 0) |
2837 | return -ENOENT; | |
08984293 | 2838 | |
162566a4 | 2839 | *to = le64toh(f->header->tail_entry_realtime); |
08984293 LP |
2840 | } |
2841 | ||
2842 | return 1; | |
2843 | } | |
2844 | ||
2845 | int journal_file_get_cutoff_monotonic_usec(JournalFile *f, sd_id128_t boot_id, usec_t *from, usec_t *to) { | |
08984293 LP |
2846 | Object *o; |
2847 | uint64_t p; | |
2848 | int r; | |
2849 | ||
2850 | assert(f); | |
2851 | assert(from || to); | |
2852 | ||
47838ab3 | 2853 | r = find_data_object_by_boot_id(f, boot_id, &o, &p); |
08984293 LP |
2854 | if (r <= 0) |
2855 | return r; | |
2856 | ||
2857 | if (le64toh(o->data.n_entries) <= 0) | |
2858 | return 0; | |
2859 | ||
2860 | if (from) { | |
2861 | r = journal_file_move_to_object(f, OBJECT_ENTRY, le64toh(o->data.entry_offset), &o); | |
2862 | if (r < 0) | |
2863 | return r; | |
2864 | ||
2865 | *from = le64toh(o->entry.monotonic); | |
2866 | } | |
2867 | ||
2868 | if (to) { | |
2869 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); | |
2870 | if (r < 0) | |
2871 | return r; | |
2872 | ||
2873 | r = generic_array_get_plus_one(f, | |
2874 | le64toh(o->data.entry_offset), | |
2875 | le64toh(o->data.entry_array_offset), | |
2876 | le64toh(o->data.n_entries)-1, | |
2877 | &o, NULL); | |
2878 | if (r <= 0) | |
2879 | return r; | |
2880 | ||
2881 | *to = le64toh(o->entry.monotonic); | |
2882 | } | |
2883 | ||
2884 | return 1; | |
2885 | } | |
dca6219e | 2886 | |
fb0951b0 | 2887 | bool journal_file_rotate_suggested(JournalFile *f, usec_t max_file_usec) { |
dca6219e LP |
2888 | assert(f); |
2889 | ||
2890 | /* If we gained new header fields we gained new features, | |
2891 | * hence suggest a rotation */ | |
361f9cbc LP |
2892 | if (le64toh(f->header->header_size) < sizeof(Header)) { |
2893 | log_debug("%s uses an outdated header, suggesting rotation.", f->path); | |
dca6219e | 2894 | return true; |
361f9cbc | 2895 | } |
dca6219e LP |
2896 | |
2897 | /* Let's check if the hash tables grew over a certain fill | |
2898 | * level (75%, borrowing this value from Java's hash table | |
2899 | * implementation), and if so suggest a rotation. To calculate | |
2900 | * the fill level we need the n_data field, which only exists | |
2901 | * in newer versions. */ | |
2902 | ||
2903 | if (JOURNAL_HEADER_CONTAINS(f->header, n_data)) | |
361f9cbc | 2904 | if (le64toh(f->header->n_data) * 4ULL > (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)) * 3ULL) { |
507f22bd | 2905 | log_debug("Data hash table of %s has a fill level at %.1f (%"PRIu64" of %"PRIu64" items, %llu file size, %"PRIu64" bytes per hash table item), suggesting rotation.", |
361f9cbc LP |
2906 | f->path, |
2907 | 100.0 * (double) le64toh(f->header->n_data) / ((double) (le64toh(f->header->data_hash_table_size) / sizeof(HashItem))), | |
507f22bd ZJS |
2908 | le64toh(f->header->n_data), |
2909 | le64toh(f->header->data_hash_table_size) / sizeof(HashItem), | |
2910 | (unsigned long long) f->last_stat.st_size, | |
2911 | f->last_stat.st_size / le64toh(f->header->n_data)); | |
dca6219e | 2912 | return true; |
361f9cbc | 2913 | } |
dca6219e LP |
2914 | |
2915 | if (JOURNAL_HEADER_CONTAINS(f->header, n_fields)) | |
361f9cbc | 2916 | if (le64toh(f->header->n_fields) * 4ULL > (le64toh(f->header->field_hash_table_size) / sizeof(HashItem)) * 3ULL) { |
507f22bd | 2917 | log_debug("Field hash table of %s has a fill level at %.1f (%"PRIu64" of %"PRIu64" items), suggesting rotation.", |
361f9cbc LP |
2918 | f->path, |
2919 | 100.0 * (double) le64toh(f->header->n_fields) / ((double) (le64toh(f->header->field_hash_table_size) / sizeof(HashItem))), | |
507f22bd ZJS |
2920 | le64toh(f->header->n_fields), |
2921 | le64toh(f->header->field_hash_table_size) / sizeof(HashItem)); | |
dca6219e | 2922 | return true; |
361f9cbc | 2923 | } |
dca6219e | 2924 | |
0598fd4a LP |
2925 | /* Are the data objects properly indexed by field objects? */ |
2926 | if (JOURNAL_HEADER_CONTAINS(f->header, n_data) && | |
2927 | JOURNAL_HEADER_CONTAINS(f->header, n_fields) && | |
2928 | le64toh(f->header->n_data) > 0 && | |
2929 | le64toh(f->header->n_fields) == 0) | |
2930 | return true; | |
2931 | ||
fb0951b0 LP |
2932 | if (max_file_usec > 0) { |
2933 | usec_t t, h; | |
2934 | ||
2935 | h = le64toh(f->header->head_entry_realtime); | |
2936 | t = now(CLOCK_REALTIME); | |
2937 | ||
2938 | if (h > 0 && t > h + max_file_usec) | |
2939 | return true; | |
2940 | } | |
2941 | ||
dca6219e LP |
2942 | return false; |
2943 | } |