]>
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> | |
29 | ||
30 | #include "journal-def.h" | |
31 | #include "journal-file.h" | |
32 | #include "lookup3.h" | |
807e17f0 | 33 | #include "compress.h" |
7560fffc | 34 | #include "fsprg.h" |
cec736d2 | 35 | |
4a92baf3 LP |
36 | #define DEFAULT_DATA_HASH_TABLE_SIZE (2047ULL*sizeof(HashItem)) |
37 | #define DEFAULT_FIELD_HASH_TABLE_SIZE (333ULL*sizeof(HashItem)) | |
cec736d2 | 38 | |
1fa80181 | 39 | #define DEFAULT_WINDOW_SIZE (8ULL*1024ULL*1024ULL) |
cec736d2 | 40 | |
be19b7df | 41 | #define COMPRESSION_SIZE_THRESHOLD (512ULL) |
807e17f0 | 42 | |
babfc091 | 43 | /* This is the minimum journal file size */ |
b47ffcfd | 44 | #define JOURNAL_FILE_SIZE_MIN (64ULL*1024ULL) /* 64 KiB */ |
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 LP |
64 | |
65 | #define ALIGN64(x) (((x) + 7ULL) & ~7ULL) | |
66 | ||
dca6219e LP |
67 | #define JOURNAL_HEADER_CONTAINS(h, field) \ |
68 | (le64toh((h)->header_size) >= offsetof(Header, field) + sizeof((h)->field)) | |
69 | ||
7560fffc | 70 | static int journal_file_maybe_append_tag(JournalFile *f, uint64_t realtime); |
dca6219e | 71 | |
cec736d2 | 72 | void journal_file_close(JournalFile *f) { |
de190aef | 73 | int t; |
cec736d2 | 74 | |
de190aef | 75 | assert(f); |
cec736d2 | 76 | |
7560fffc LP |
77 | /* Sync everything to disk, before we mark the file offline */ |
78 | for (t = 0; t < _WINDOW_MAX; t++) | |
79 | if (f->windows[t].ptr) | |
80 | munmap(f->windows[t].ptr, f->windows[t].size); | |
81 | ||
82 | if (f->writable && f->fd >= 0) | |
83 | fdatasync(f->fd); | |
84 | ||
d384c7a8 | 85 | if (f->header) { |
cd96b3b8 LP |
86 | /* Mark the file offline. Don't override the archived state if it already is set */ |
87 | if (f->writable && f->header->state == STATE_ONLINE) | |
d384c7a8 | 88 | f->header->state = STATE_OFFLINE; |
cec736d2 | 89 | |
d384c7a8 MS |
90 | munmap(f->header, PAGE_ALIGN(sizeof(Header))); |
91 | } | |
cec736d2 | 92 | |
0ac38b70 LP |
93 | if (f->fd >= 0) |
94 | close_nointr_nofail(f->fd); | |
95 | ||
cec736d2 | 96 | free(f->path); |
807e17f0 LP |
97 | |
98 | #ifdef HAVE_XZ | |
99 | free(f->compress_buffer); | |
100 | #endif | |
101 | ||
7560fffc LP |
102 | #ifdef HAVE_GCRYPT |
103 | if (f->fsprg_header) | |
104 | munmap(f->fsprg_header, PAGE_ALIGN(f->fsprg_size)); | |
105 | ||
106 | if (f->hmac) | |
107 | gcry_md_close(f->hmac); | |
108 | #endif | |
109 | ||
cec736d2 LP |
110 | free(f); |
111 | } | |
112 | ||
0ac38b70 | 113 | static int journal_file_init_header(JournalFile *f, JournalFile *template) { |
cec736d2 LP |
114 | Header h; |
115 | ssize_t k; | |
116 | int r; | |
117 | ||
118 | assert(f); | |
119 | ||
120 | zero(h); | |
7560fffc | 121 | memcpy(h.signature, HEADER_SIGNATURE, 8); |
23b0b2b2 | 122 | h.header_size = htole64(ALIGN64(sizeof(h))); |
cec736d2 | 123 | |
7560fffc LP |
124 | h.incompatible_flags = |
125 | htole32(f->compress ? HEADER_INCOMPATIBLE_COMPRESSED : 0); | |
126 | ||
127 | h.compatible_flags = | |
128 | htole32(f->authenticate ? HEADER_COMPATIBLE_AUTHENTICATED : 0); | |
129 | ||
cec736d2 LP |
130 | r = sd_id128_randomize(&h.file_id); |
131 | if (r < 0) | |
132 | return r; | |
133 | ||
0ac38b70 LP |
134 | if (template) { |
135 | h.seqnum_id = template->header->seqnum_id; | |
dca6219e | 136 | h.tail_seqnum = template->header->tail_seqnum; |
0ac38b70 LP |
137 | } else |
138 | h.seqnum_id = h.file_id; | |
cec736d2 LP |
139 | |
140 | k = pwrite(f->fd, &h, sizeof(h), 0); | |
141 | if (k < 0) | |
142 | return -errno; | |
143 | ||
144 | if (k != sizeof(h)) | |
145 | return -EIO; | |
146 | ||
147 | return 0; | |
148 | } | |
149 | ||
150 | static int journal_file_refresh_header(JournalFile *f) { | |
151 | int r; | |
de190aef | 152 | sd_id128_t boot_id; |
cec736d2 LP |
153 | |
154 | assert(f); | |
155 | ||
156 | r = sd_id128_get_machine(&f->header->machine_id); | |
157 | if (r < 0) | |
158 | return r; | |
159 | ||
de190aef | 160 | r = sd_id128_get_boot(&boot_id); |
cec736d2 LP |
161 | if (r < 0) |
162 | return r; | |
163 | ||
de190aef LP |
164 | if (sd_id128_equal(boot_id, f->header->boot_id)) |
165 | f->tail_entry_monotonic_valid = true; | |
166 | ||
167 | f->header->boot_id = boot_id; | |
168 | ||
169 | f->header->state = STATE_ONLINE; | |
b788cc23 | 170 | |
7560fffc LP |
171 | /* Sync the online state to disk */ |
172 | msync(f->header, PAGE_ALIGN(sizeof(Header)), MS_SYNC); | |
173 | fdatasync(f->fd); | |
b788cc23 | 174 | |
cec736d2 LP |
175 | return 0; |
176 | } | |
177 | ||
178 | static int journal_file_verify_header(JournalFile *f) { | |
179 | assert(f); | |
180 | ||
7560fffc | 181 | if (memcmp(f->header->signature, HEADER_SIGNATURE, 8)) |
cec736d2 LP |
182 | return -EBADMSG; |
183 | ||
7560fffc LP |
184 | /* In both read and write mode we refuse to open files with |
185 | * incompatible flags we don't know */ | |
807e17f0 | 186 | #ifdef HAVE_XZ |
7560fffc | 187 | if ((le32toh(f->header->incompatible_flags) & ~HEADER_INCOMPATIBLE_COMPRESSED) != 0) |
807e17f0 LP |
188 | return -EPROTONOSUPPORT; |
189 | #else | |
cec736d2 LP |
190 | if (f->header->incompatible_flags != 0) |
191 | return -EPROTONOSUPPORT; | |
807e17f0 | 192 | #endif |
cec736d2 | 193 | |
7560fffc LP |
194 | /* When open for writing we refuse to open files with |
195 | * compatible flags, too */ | |
196 | if (f->writable) { | |
197 | #ifdef HAVE_GCRYPT | |
198 | if ((le32toh(f->header->compatible_flags) & ~HEADER_COMPATIBLE_AUTHENTICATED) != 0) | |
199 | return -EPROTONOSUPPORT; | |
200 | #else | |
201 | if (f->header->compatible_flags != 0) | |
202 | return -EPROTONOSUPPORT; | |
203 | #endif | |
204 | } | |
205 | ||
dca6219e LP |
206 | /* The first addition was n_data, so check that we are at least this large */ |
207 | if (le64toh(f->header->header_size) < HEADER_SIZE_MIN) | |
23b0b2b2 LP |
208 | return -EBADMSG; |
209 | ||
210 | if ((uint64_t) f->last_stat.st_size < (le64toh(f->header->header_size) + le64toh(f->header->arena_size))) | |
cec736d2 LP |
211 | return -ENODATA; |
212 | ||
213 | if (f->writable) { | |
ccdbaf91 | 214 | uint8_t state; |
cec736d2 LP |
215 | sd_id128_t machine_id; |
216 | int r; | |
217 | ||
218 | r = sd_id128_get_machine(&machine_id); | |
219 | if (r < 0) | |
220 | return r; | |
221 | ||
222 | if (!sd_id128_equal(machine_id, f->header->machine_id)) | |
223 | return -EHOSTDOWN; | |
224 | ||
de190aef | 225 | state = f->header->state; |
cec736d2 | 226 | |
71fa6f00 LP |
227 | if (state == STATE_ONLINE) { |
228 | log_debug("Journal file %s is already online. Assuming unclean closing.", f->path); | |
229 | return -EBUSY; | |
230 | } else if (state == STATE_ARCHIVED) | |
cec736d2 | 231 | return -ESHUTDOWN; |
71fa6f00 LP |
232 | else if (state != STATE_OFFLINE) { |
233 | log_debug("Journal file %s has unknown state %u.", f->path, state); | |
234 | return -EBUSY; | |
235 | } | |
cec736d2 LP |
236 | } |
237 | ||
7560fffc LP |
238 | f->compress = !!(le32toh(f->header->incompatible_flags) & HEADER_INCOMPATIBLE_COMPRESSED); |
239 | f->authenticate = !!(le32toh(f->header->compatible_flags) & HEADER_COMPATIBLE_AUTHENTICATED); | |
240 | ||
cec736d2 LP |
241 | return 0; |
242 | } | |
243 | ||
244 | static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size) { | |
cec736d2 | 245 | uint64_t old_size, new_size; |
fec2aa2f | 246 | int r; |
cec736d2 LP |
247 | |
248 | assert(f); | |
249 | ||
cec736d2 | 250 | /* We assume that this file is not sparse, and we know that |
38ac38b2 | 251 | * for sure, since we always call posix_fallocate() |
cec736d2 LP |
252 | * ourselves */ |
253 | ||
254 | old_size = | |
23b0b2b2 | 255 | le64toh(f->header->header_size) + |
cec736d2 LP |
256 | le64toh(f->header->arena_size); |
257 | ||
bc85bfee | 258 | new_size = PAGE_ALIGN(offset + size); |
23b0b2b2 LP |
259 | if (new_size < le64toh(f->header->header_size)) |
260 | new_size = le64toh(f->header->header_size); | |
bc85bfee LP |
261 | |
262 | if (new_size <= old_size) | |
cec736d2 LP |
263 | return 0; |
264 | ||
bc85bfee LP |
265 | if (f->metrics.max_size > 0 && |
266 | new_size > f->metrics.max_size) | |
267 | return -E2BIG; | |
cec736d2 | 268 | |
bc85bfee LP |
269 | if (new_size > f->metrics.min_size && |
270 | f->metrics.keep_free > 0) { | |
cec736d2 LP |
271 | struct statvfs svfs; |
272 | ||
273 | if (fstatvfs(f->fd, &svfs) >= 0) { | |
274 | uint64_t available; | |
275 | ||
276 | available = svfs.f_bfree * svfs.f_bsize; | |
277 | ||
bc85bfee LP |
278 | if (available >= f->metrics.keep_free) |
279 | available -= f->metrics.keep_free; | |
cec736d2 LP |
280 | else |
281 | available = 0; | |
282 | ||
283 | if (new_size - old_size > available) | |
284 | return -E2BIG; | |
285 | } | |
286 | } | |
287 | ||
bc85bfee LP |
288 | /* Note that the glibc fallocate() fallback is very |
289 | inefficient, hence we try to minimize the allocation area | |
290 | as we can. */ | |
fec2aa2f GV |
291 | r = posix_fallocate(f->fd, old_size, new_size - old_size); |
292 | if (r != 0) | |
293 | return -r; | |
cec736d2 LP |
294 | |
295 | if (fstat(f->fd, &f->last_stat) < 0) | |
296 | return -errno; | |
297 | ||
23b0b2b2 | 298 | f->header->arena_size = htole64(new_size - le64toh(f->header->header_size)); |
cec736d2 LP |
299 | |
300 | return 0; | |
301 | } | |
302 | ||
303 | static int journal_file_map( | |
304 | JournalFile *f, | |
305 | uint64_t offset, | |
306 | uint64_t size, | |
307 | void **_window, | |
308 | uint64_t *_woffset, | |
309 | uint64_t *_wsize, | |
310 | void **ret) { | |
311 | ||
312 | uint64_t woffset, wsize; | |
313 | void *window; | |
314 | ||
315 | assert(f); | |
316 | assert(size > 0); | |
317 | assert(ret); | |
318 | ||
319 | woffset = offset & ~((uint64_t) page_size() - 1ULL); | |
320 | wsize = size + (offset - woffset); | |
321 | wsize = PAGE_ALIGN(wsize); | |
322 | ||
2a59ea54 LP |
323 | /* Avoid SIGBUS on invalid accesses */ |
324 | if (woffset + wsize > (uint64_t) PAGE_ALIGN(f->last_stat.st_size)) | |
325 | return -EADDRNOTAVAIL; | |
326 | ||
cec736d2 LP |
327 | window = mmap(NULL, wsize, f->prot, MAP_SHARED, f->fd, woffset); |
328 | if (window == MAP_FAILED) | |
329 | return -errno; | |
330 | ||
331 | if (_window) | |
332 | *_window = window; | |
333 | ||
334 | if (_woffset) | |
335 | *_woffset = woffset; | |
336 | ||
337 | if (_wsize) | |
338 | *_wsize = wsize; | |
339 | ||
340 | *ret = (uint8_t*) window + (offset - woffset); | |
341 | ||
342 | return 0; | |
343 | } | |
344 | ||
de190aef | 345 | static int journal_file_move_to(JournalFile *f, int wt, uint64_t offset, uint64_t size, void **ret) { |
6c8a39b8 | 346 | void *p = NULL; |
cec736d2 LP |
347 | uint64_t delta; |
348 | int r; | |
de190aef | 349 | Window *w; |
cec736d2 LP |
350 | |
351 | assert(f); | |
352 | assert(ret); | |
de190aef LP |
353 | assert(wt >= 0); |
354 | assert(wt < _WINDOW_MAX); | |
cec736d2 | 355 | |
4bbdcdb3 LP |
356 | if (offset + size > (uint64_t) f->last_stat.st_size) { |
357 | /* Hmm, out of range? Let's refresh the fstat() data | |
358 | * first, before we trust that check. */ | |
359 | ||
360 | if (fstat(f->fd, &f->last_stat) < 0 || | |
361 | offset + size > (uint64_t) f->last_stat.st_size) | |
362 | return -EADDRNOTAVAIL; | |
363 | } | |
364 | ||
de190aef | 365 | w = f->windows + wt; |
cec736d2 | 366 | |
de190aef LP |
367 | if (_likely_(w->ptr && |
368 | w->offset <= offset && | |
369 | w->offset + w->size >= offset + size)) { | |
370 | ||
371 | *ret = (uint8_t*) w->ptr + (offset - w->offset); | |
cec736d2 LP |
372 | return 0; |
373 | } | |
374 | ||
de190aef LP |
375 | if (w->ptr) { |
376 | if (munmap(w->ptr, w->size) < 0) | |
cec736d2 LP |
377 | return -errno; |
378 | ||
de190aef LP |
379 | w->ptr = NULL; |
380 | w->size = w->offset = 0; | |
cec736d2 LP |
381 | } |
382 | ||
383 | if (size < DEFAULT_WINDOW_SIZE) { | |
384 | /* If the default window size is larger then what was | |
385 | * asked for extend the mapping a bit in the hope to | |
386 | * minimize needed remappings later on. We add half | |
387 | * the window space before and half behind the | |
388 | * requested mapping */ | |
389 | ||
1921a5cb | 390 | delta = (DEFAULT_WINDOW_SIZE - size) / 2; |
cec736d2 | 391 | |
a99c349d | 392 | if (delta > offset) |
cec736d2 LP |
393 | delta = offset; |
394 | ||
395 | offset -= delta; | |
a99c349d | 396 | size = DEFAULT_WINDOW_SIZE; |
cec736d2 LP |
397 | } else |
398 | delta = 0; | |
399 | ||
2a59ea54 | 400 | if (offset + size > (uint64_t) f->last_stat.st_size) |
1921a5cb | 401 | size = (uint64_t) f->last_stat.st_size - offset; |
2a59ea54 LP |
402 | |
403 | if (size <= 0) | |
404 | return -EADDRNOTAVAIL; | |
405 | ||
cec736d2 LP |
406 | r = journal_file_map(f, |
407 | offset, size, | |
de190aef LP |
408 | &w->ptr, &w->offset, &w->size, |
409 | &p); | |
cec736d2 LP |
410 | |
411 | if (r < 0) | |
412 | return r; | |
413 | ||
414 | *ret = (uint8_t*) p + delta; | |
415 | return 0; | |
416 | } | |
417 | ||
418 | static bool verify_hash(Object *o) { | |
de190aef | 419 | uint64_t h1, h2; |
cec736d2 LP |
420 | |
421 | assert(o); | |
422 | ||
807e17f0 | 423 | if (o->object.type == OBJECT_DATA && !(o->object.flags & OBJECT_COMPRESSED)) { |
cec736d2 | 424 | h1 = le64toh(o->data.hash); |
de190aef LP |
425 | h2 = hash64(o->data.payload, le64toh(o->object.size) - offsetof(Object, data.payload)); |
426 | } else if (o->object.type == OBJECT_FIELD) { | |
427 | h1 = le64toh(o->field.hash); | |
428 | h2 = hash64(o->field.payload, le64toh(o->object.size) - offsetof(Object, field.payload)); | |
429 | } else | |
430 | return true; | |
cec736d2 | 431 | |
de190aef | 432 | return h1 == h2; |
cec736d2 LP |
433 | } |
434 | ||
de190aef | 435 | int journal_file_move_to_object(JournalFile *f, int type, uint64_t offset, Object **ret) { |
cec736d2 LP |
436 | int r; |
437 | void *t; | |
438 | Object *o; | |
439 | uint64_t s; | |
440 | ||
441 | assert(f); | |
442 | assert(ret); | |
de190aef | 443 | assert(type < _OBJECT_TYPE_MAX); |
cec736d2 | 444 | |
de190aef | 445 | r = journal_file_move_to(f, type >= 0 ? type : WINDOW_UNKNOWN, offset, sizeof(ObjectHeader), &t); |
cec736d2 LP |
446 | if (r < 0) |
447 | return r; | |
448 | ||
449 | o = (Object*) t; | |
450 | s = le64toh(o->object.size); | |
451 | ||
452 | if (s < sizeof(ObjectHeader)) | |
453 | return -EBADMSG; | |
454 | ||
de190aef | 455 | if (type >= 0 && o->object.type != type) |
cec736d2 LP |
456 | return -EBADMSG; |
457 | ||
458 | if (s > sizeof(ObjectHeader)) { | |
de190aef | 459 | r = journal_file_move_to(f, o->object.type, offset, s, &t); |
cec736d2 LP |
460 | if (r < 0) |
461 | return r; | |
462 | ||
463 | o = (Object*) t; | |
464 | } | |
465 | ||
466 | if (!verify_hash(o)) | |
467 | return -EBADMSG; | |
468 | ||
469 | *ret = o; | |
470 | return 0; | |
471 | } | |
472 | ||
c2373f84 | 473 | static uint64_t journal_file_seqnum(JournalFile *f, uint64_t *seqnum) { |
cec736d2 LP |
474 | uint64_t r; |
475 | ||
476 | assert(f); | |
477 | ||
dca6219e | 478 | r = le64toh(f->header->tail_seqnum) + 1; |
c2373f84 LP |
479 | |
480 | if (seqnum) { | |
de190aef | 481 | /* If an external seqnum counter was passed, we update |
c2373f84 LP |
482 | * both the local and the external one, and set it to |
483 | * the maximum of both */ | |
484 | ||
485 | if (*seqnum + 1 > r) | |
486 | r = *seqnum + 1; | |
487 | ||
488 | *seqnum = r; | |
489 | } | |
490 | ||
dca6219e | 491 | f->header->tail_seqnum = htole64(r); |
cec736d2 | 492 | |
dca6219e LP |
493 | if (f->header->head_seqnum == 0) |
494 | f->header->head_seqnum = htole64(r); | |
de190aef | 495 | |
cec736d2 LP |
496 | return r; |
497 | } | |
498 | ||
de190aef | 499 | static int journal_file_append_object(JournalFile *f, int type, uint64_t size, Object **ret, uint64_t *offset) { |
cec736d2 LP |
500 | int r; |
501 | uint64_t p; | |
502 | Object *tail, *o; | |
503 | void *t; | |
504 | ||
505 | assert(f); | |
506 | assert(size >= sizeof(ObjectHeader)); | |
507 | assert(offset); | |
508 | assert(ret); | |
509 | ||
510 | p = le64toh(f->header->tail_object_offset); | |
cec736d2 | 511 | if (p == 0) |
23b0b2b2 | 512 | p = le64toh(f->header->header_size); |
cec736d2 | 513 | else { |
de190aef | 514 | r = journal_file_move_to_object(f, -1, p, &tail); |
cec736d2 LP |
515 | if (r < 0) |
516 | return r; | |
517 | ||
518 | p += ALIGN64(le64toh(tail->object.size)); | |
519 | } | |
520 | ||
521 | r = journal_file_allocate(f, p, size); | |
522 | if (r < 0) | |
523 | return r; | |
524 | ||
de190aef | 525 | r = journal_file_move_to(f, type, p, size, &t); |
cec736d2 LP |
526 | if (r < 0) |
527 | return r; | |
528 | ||
529 | o = (Object*) t; | |
530 | ||
531 | zero(o->object); | |
de190aef | 532 | o->object.type = type; |
cec736d2 LP |
533 | o->object.size = htole64(size); |
534 | ||
535 | f->header->tail_object_offset = htole64(p); | |
cec736d2 LP |
536 | f->header->n_objects = htole64(le64toh(f->header->n_objects) + 1); |
537 | ||
538 | *ret = o; | |
539 | *offset = p; | |
540 | ||
541 | return 0; | |
542 | } | |
543 | ||
de190aef | 544 | static int journal_file_setup_data_hash_table(JournalFile *f) { |
cec736d2 LP |
545 | uint64_t s, p; |
546 | Object *o; | |
547 | int r; | |
548 | ||
549 | assert(f); | |
550 | ||
dfabe643 | 551 | /* We estimate that we need 1 hash table entry per 768 of |
4a92baf3 LP |
552 | journal file and we want to make sure we never get beyond |
553 | 75% fill level. Calculate the hash table size for the | |
554 | maximum file size based on these metrics. */ | |
555 | ||
dfabe643 | 556 | s = (f->metrics.max_size * 4 / 768 / 3) * sizeof(HashItem); |
4a92baf3 LP |
557 | if (s < DEFAULT_DATA_HASH_TABLE_SIZE) |
558 | s = DEFAULT_DATA_HASH_TABLE_SIZE; | |
559 | ||
dfabe643 | 560 | log_info("Reserving %llu entries in hash table.", (unsigned long long) (s / sizeof(HashItem))); |
4a92baf3 | 561 | |
de190aef LP |
562 | r = journal_file_append_object(f, |
563 | OBJECT_DATA_HASH_TABLE, | |
564 | offsetof(Object, hash_table.items) + s, | |
565 | &o, &p); | |
cec736d2 LP |
566 | if (r < 0) |
567 | return r; | |
568 | ||
de190aef | 569 | memset(o->hash_table.items, 0, s); |
cec736d2 | 570 | |
de190aef LP |
571 | f->header->data_hash_table_offset = htole64(p + offsetof(Object, hash_table.items)); |
572 | f->header->data_hash_table_size = htole64(s); | |
cec736d2 LP |
573 | |
574 | return 0; | |
575 | } | |
576 | ||
de190aef | 577 | static int journal_file_setup_field_hash_table(JournalFile *f) { |
cec736d2 LP |
578 | uint64_t s, p; |
579 | Object *o; | |
580 | int r; | |
581 | ||
582 | assert(f); | |
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 | ||
de190aef | 592 | memset(o->hash_table.items, 0, 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 LP |
610 | r = journal_file_move_to(f, |
611 | WINDOW_DATA_HASH_TABLE, | |
612 | p, s, | |
613 | &t); | |
cec736d2 LP |
614 | if (r < 0) |
615 | return r; | |
616 | ||
de190aef | 617 | f->data_hash_table = t; |
cec736d2 LP |
618 | return 0; |
619 | } | |
620 | ||
de190aef | 621 | static int journal_file_map_field_hash_table(JournalFile *f) { |
cec736d2 LP |
622 | uint64_t s, p; |
623 | void *t; | |
624 | int r; | |
625 | ||
626 | assert(f); | |
627 | ||
de190aef LP |
628 | p = le64toh(f->header->field_hash_table_offset); |
629 | s = le64toh(f->header->field_hash_table_size); | |
cec736d2 | 630 | |
de190aef LP |
631 | r = journal_file_move_to(f, |
632 | WINDOW_FIELD_HASH_TABLE, | |
633 | p, s, | |
634 | &t); | |
cec736d2 LP |
635 | if (r < 0) |
636 | return r; | |
637 | ||
de190aef | 638 | f->field_hash_table = t; |
cec736d2 LP |
639 | return 0; |
640 | } | |
641 | ||
de190aef LP |
642 | static int journal_file_link_data(JournalFile *f, Object *o, uint64_t offset, uint64_t hash) { |
643 | uint64_t p, h; | |
cec736d2 LP |
644 | int r; |
645 | ||
646 | assert(f); | |
647 | assert(o); | |
648 | assert(offset > 0); | |
de190aef | 649 | assert(o->object.type == OBJECT_DATA); |
cec736d2 | 650 | |
48496df6 LP |
651 | /* This might alter the window we are looking at */ |
652 | ||
de190aef LP |
653 | o->data.next_hash_offset = o->data.next_field_offset = 0; |
654 | o->data.entry_offset = o->data.entry_array_offset = 0; | |
655 | o->data.n_entries = 0; | |
cec736d2 | 656 | |
de190aef | 657 | h = hash % (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)); |
8db4213e | 658 | p = le64toh(f->data_hash_table[h].tail_hash_offset); |
cec736d2 LP |
659 | if (p == 0) { |
660 | /* Only entry in the hash table is easy */ | |
de190aef | 661 | f->data_hash_table[h].head_hash_offset = htole64(offset); |
cec736d2 | 662 | } else { |
48496df6 LP |
663 | /* Move back to the previous data object, to patch in |
664 | * pointer */ | |
cec736d2 | 665 | |
de190aef | 666 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); |
cec736d2 LP |
667 | if (r < 0) |
668 | return r; | |
669 | ||
de190aef | 670 | o->data.next_hash_offset = htole64(offset); |
cec736d2 LP |
671 | } |
672 | ||
de190aef | 673 | f->data_hash_table[h].tail_hash_offset = htole64(offset); |
cec736d2 | 674 | |
dca6219e LP |
675 | if (JOURNAL_HEADER_CONTAINS(f->header, n_data)) |
676 | f->header->n_data = htole64(le64toh(f->header->n_data) + 1); | |
677 | ||
cec736d2 LP |
678 | return 0; |
679 | } | |
680 | ||
de190aef LP |
681 | int journal_file_find_data_object_with_hash( |
682 | JournalFile *f, | |
683 | const void *data, uint64_t size, uint64_t hash, | |
684 | Object **ret, uint64_t *offset) { | |
48496df6 | 685 | |
de190aef | 686 | uint64_t p, osize, h; |
cec736d2 LP |
687 | int r; |
688 | ||
689 | assert(f); | |
690 | assert(data || size == 0); | |
691 | ||
692 | osize = offsetof(Object, data.payload) + size; | |
693 | ||
bc85bfee LP |
694 | if (f->header->data_hash_table_size == 0) |
695 | return -EBADMSG; | |
696 | ||
de190aef LP |
697 | h = hash % (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)); |
698 | p = le64toh(f->data_hash_table[h].head_hash_offset); | |
cec736d2 | 699 | |
de190aef LP |
700 | while (p > 0) { |
701 | Object *o; | |
cec736d2 | 702 | |
de190aef | 703 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); |
cec736d2 LP |
704 | if (r < 0) |
705 | return r; | |
706 | ||
807e17f0 | 707 | if (le64toh(o->data.hash) != hash) |
85a131e8 | 708 | goto next; |
807e17f0 LP |
709 | |
710 | if (o->object.flags & OBJECT_COMPRESSED) { | |
711 | #ifdef HAVE_XZ | |
b785c858 | 712 | uint64_t l, rsize; |
cec736d2 | 713 | |
807e17f0 LP |
714 | l = le64toh(o->object.size); |
715 | if (l <= offsetof(Object, data.payload)) | |
cec736d2 LP |
716 | return -EBADMSG; |
717 | ||
807e17f0 LP |
718 | l -= offsetof(Object, data.payload); |
719 | ||
720 | if (!uncompress_blob(o->data.payload, l, &f->compress_buffer, &f->compress_buffer_size, &rsize)) | |
721 | return -EBADMSG; | |
722 | ||
b785c858 | 723 | if (rsize == size && |
807e17f0 LP |
724 | memcmp(f->compress_buffer, data, size) == 0) { |
725 | ||
726 | if (ret) | |
727 | *ret = o; | |
728 | ||
729 | if (offset) | |
730 | *offset = p; | |
731 | ||
732 | return 1; | |
733 | } | |
734 | #else | |
735 | return -EPROTONOSUPPORT; | |
736 | #endif | |
737 | ||
738 | } else if (le64toh(o->object.size) == osize && | |
739 | memcmp(o->data.payload, data, size) == 0) { | |
740 | ||
cec736d2 LP |
741 | if (ret) |
742 | *ret = o; | |
743 | ||
744 | if (offset) | |
745 | *offset = p; | |
746 | ||
de190aef | 747 | return 1; |
cec736d2 LP |
748 | } |
749 | ||
85a131e8 | 750 | next: |
cec736d2 LP |
751 | p = le64toh(o->data.next_hash_offset); |
752 | } | |
753 | ||
de190aef LP |
754 | return 0; |
755 | } | |
756 | ||
757 | int journal_file_find_data_object( | |
758 | JournalFile *f, | |
759 | const void *data, uint64_t size, | |
760 | Object **ret, uint64_t *offset) { | |
761 | ||
762 | uint64_t hash; | |
763 | ||
764 | assert(f); | |
765 | assert(data || size == 0); | |
766 | ||
767 | hash = hash64(data, size); | |
768 | ||
769 | return journal_file_find_data_object_with_hash(f, | |
770 | data, size, hash, | |
771 | ret, offset); | |
772 | } | |
773 | ||
48496df6 LP |
774 | static int journal_file_append_data( |
775 | JournalFile *f, | |
776 | const void *data, uint64_t size, | |
777 | Object **ret, uint64_t *offset) { | |
778 | ||
de190aef LP |
779 | uint64_t hash, p; |
780 | uint64_t osize; | |
781 | Object *o; | |
782 | int r; | |
807e17f0 | 783 | bool compressed = false; |
de190aef LP |
784 | |
785 | assert(f); | |
786 | assert(data || size == 0); | |
787 | ||
788 | hash = hash64(data, size); | |
789 | ||
790 | r = journal_file_find_data_object_with_hash(f, data, size, hash, &o, &p); | |
791 | if (r < 0) | |
792 | return r; | |
793 | else if (r > 0) { | |
794 | ||
795 | if (ret) | |
796 | *ret = o; | |
797 | ||
798 | if (offset) | |
799 | *offset = p; | |
800 | ||
801 | return 0; | |
802 | } | |
803 | ||
804 | osize = offsetof(Object, data.payload) + size; | |
805 | r = journal_file_append_object(f, OBJECT_DATA, osize, &o, &p); | |
cec736d2 LP |
806 | if (r < 0) |
807 | return r; | |
808 | ||
cec736d2 | 809 | o->data.hash = htole64(hash); |
807e17f0 LP |
810 | |
811 | #ifdef HAVE_XZ | |
812 | if (f->compress && | |
813 | size >= COMPRESSION_SIZE_THRESHOLD) { | |
814 | uint64_t rsize; | |
815 | ||
816 | compressed = compress_blob(data, size, o->data.payload, &rsize); | |
817 | ||
818 | if (compressed) { | |
819 | o->object.size = htole64(offsetof(Object, data.payload) + rsize); | |
820 | o->object.flags |= OBJECT_COMPRESSED; | |
821 | ||
807e17f0 LP |
822 | log_debug("Compressed data object %lu -> %lu", (unsigned long) size, (unsigned long) rsize); |
823 | } | |
824 | } | |
825 | #endif | |
826 | ||
64825d3c | 827 | if (!compressed && size > 0) |
807e17f0 | 828 | memcpy(o->data.payload, data, size); |
cec736d2 | 829 | |
de190aef | 830 | r = journal_file_link_data(f, o, p, hash); |
cec736d2 LP |
831 | if (r < 0) |
832 | return r; | |
833 | ||
48496df6 LP |
834 | /* The linking might have altered the window, so let's |
835 | * refresh our pointer */ | |
836 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); | |
837 | if (r < 0) | |
838 | return r; | |
839 | ||
cec736d2 LP |
840 | if (ret) |
841 | *ret = o; | |
842 | ||
843 | if (offset) | |
de190aef | 844 | *offset = p; |
cec736d2 LP |
845 | |
846 | return 0; | |
847 | } | |
848 | ||
849 | uint64_t journal_file_entry_n_items(Object *o) { | |
850 | assert(o); | |
7be3aa17 | 851 | assert(o->object.type == OBJECT_ENTRY); |
cec736d2 LP |
852 | |
853 | return (le64toh(o->object.size) - offsetof(Object, entry.items)) / sizeof(EntryItem); | |
854 | } | |
855 | ||
de190aef LP |
856 | static uint64_t journal_file_entry_array_n_items(Object *o) { |
857 | assert(o); | |
7be3aa17 | 858 | assert(o->object.type == OBJECT_ENTRY_ARRAY); |
de190aef LP |
859 | |
860 | return (le64toh(o->object.size) - offsetof(Object, entry_array.items)) / sizeof(uint64_t); | |
861 | } | |
862 | ||
863 | static int link_entry_into_array(JournalFile *f, | |
4fd052ae FC |
864 | le64_t *first, |
865 | le64_t *idx, | |
de190aef | 866 | uint64_t p) { |
cec736d2 | 867 | int r; |
de190aef LP |
868 | uint64_t n = 0, ap = 0, q, i, a, hidx; |
869 | Object *o; | |
870 | ||
cec736d2 | 871 | assert(f); |
de190aef LP |
872 | assert(first); |
873 | assert(idx); | |
874 | assert(p > 0); | |
cec736d2 | 875 | |
de190aef LP |
876 | a = le64toh(*first); |
877 | i = hidx = le64toh(*idx); | |
878 | while (a > 0) { | |
879 | ||
880 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &o); | |
881 | if (r < 0) | |
882 | return r; | |
cec736d2 | 883 | |
de190aef LP |
884 | n = journal_file_entry_array_n_items(o); |
885 | if (i < n) { | |
886 | o->entry_array.items[i] = htole64(p); | |
887 | *idx = htole64(hidx + 1); | |
888 | return 0; | |
889 | } | |
cec736d2 | 890 | |
de190aef LP |
891 | i -= n; |
892 | ap = a; | |
893 | a = le64toh(o->entry_array.next_entry_array_offset); | |
894 | } | |
895 | ||
896 | if (hidx > n) | |
897 | n = (hidx+1) * 2; | |
898 | else | |
899 | n = n * 2; | |
900 | ||
901 | if (n < 4) | |
902 | n = 4; | |
903 | ||
904 | r = journal_file_append_object(f, OBJECT_ENTRY_ARRAY, | |
905 | offsetof(Object, entry_array.items) + n * sizeof(uint64_t), | |
906 | &o, &q); | |
cec736d2 LP |
907 | if (r < 0) |
908 | return r; | |
909 | ||
de190aef | 910 | o->entry_array.items[i] = htole64(p); |
cec736d2 | 911 | |
de190aef | 912 | if (ap == 0) |
7be3aa17 | 913 | *first = htole64(q); |
cec736d2 | 914 | else { |
de190aef | 915 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, ap, &o); |
cec736d2 LP |
916 | if (r < 0) |
917 | return r; | |
918 | ||
de190aef LP |
919 | o->entry_array.next_entry_array_offset = htole64(q); |
920 | } | |
cec736d2 | 921 | |
de190aef LP |
922 | *idx = htole64(hidx + 1); |
923 | ||
924 | return 0; | |
925 | } | |
cec736d2 | 926 | |
de190aef | 927 | static int link_entry_into_array_plus_one(JournalFile *f, |
4fd052ae FC |
928 | le64_t *extra, |
929 | le64_t *first, | |
930 | le64_t *idx, | |
de190aef LP |
931 | uint64_t p) { |
932 | ||
933 | int r; | |
934 | ||
935 | assert(f); | |
936 | assert(extra); | |
937 | assert(first); | |
938 | assert(idx); | |
939 | assert(p > 0); | |
940 | ||
941 | if (*idx == 0) | |
942 | *extra = htole64(p); | |
943 | else { | |
4fd052ae | 944 | le64_t i; |
de190aef | 945 | |
7be3aa17 | 946 | i = htole64(le64toh(*idx) - 1); |
de190aef LP |
947 | r = link_entry_into_array(f, first, &i, p); |
948 | if (r < 0) | |
949 | return r; | |
cec736d2 LP |
950 | } |
951 | ||
de190aef LP |
952 | *idx = htole64(le64toh(*idx) + 1); |
953 | return 0; | |
954 | } | |
955 | ||
956 | static int journal_file_link_entry_item(JournalFile *f, Object *o, uint64_t offset, uint64_t i) { | |
957 | uint64_t p; | |
958 | int r; | |
959 | assert(f); | |
960 | assert(o); | |
961 | assert(offset > 0); | |
962 | ||
963 | p = le64toh(o->entry.items[i].object_offset); | |
964 | if (p == 0) | |
965 | return -EINVAL; | |
966 | ||
967 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); | |
cec736d2 LP |
968 | if (r < 0) |
969 | return r; | |
970 | ||
de190aef LP |
971 | return link_entry_into_array_plus_one(f, |
972 | &o->data.entry_offset, | |
973 | &o->data.entry_array_offset, | |
974 | &o->data.n_entries, | |
975 | offset); | |
cec736d2 LP |
976 | } |
977 | ||
978 | static int journal_file_link_entry(JournalFile *f, Object *o, uint64_t offset) { | |
de190aef | 979 | uint64_t n, i; |
cec736d2 LP |
980 | int r; |
981 | ||
982 | assert(f); | |
983 | assert(o); | |
984 | assert(offset > 0); | |
de190aef | 985 | assert(o->object.type == OBJECT_ENTRY); |
cec736d2 | 986 | |
b788cc23 LP |
987 | __sync_synchronize(); |
988 | ||
cec736d2 | 989 | /* Link up the entry itself */ |
de190aef LP |
990 | r = link_entry_into_array(f, |
991 | &f->header->entry_array_offset, | |
992 | &f->header->n_entries, | |
993 | offset); | |
994 | if (r < 0) | |
995 | return r; | |
cec736d2 | 996 | |
aaf53376 | 997 | /* log_debug("=> %s seqnr=%lu n_entries=%lu", f->path, (unsigned long) o->entry.seqnum, (unsigned long) f->header->n_entries); */ |
cec736d2 | 998 | |
de190aef | 999 | if (f->header->head_entry_realtime == 0) |
0ac38b70 | 1000 | f->header->head_entry_realtime = o->entry.realtime; |
cec736d2 | 1001 | |
0ac38b70 | 1002 | f->header->tail_entry_realtime = o->entry.realtime; |
de190aef LP |
1003 | f->header->tail_entry_monotonic = o->entry.monotonic; |
1004 | ||
1005 | f->tail_entry_monotonic_valid = true; | |
cec736d2 LP |
1006 | |
1007 | /* Link up the items */ | |
1008 | n = journal_file_entry_n_items(o); | |
1009 | for (i = 0; i < n; i++) { | |
1010 | r = journal_file_link_entry_item(f, o, offset, i); | |
1011 | if (r < 0) | |
1012 | return r; | |
1013 | } | |
1014 | ||
cec736d2 LP |
1015 | return 0; |
1016 | } | |
1017 | ||
1018 | static int journal_file_append_entry_internal( | |
1019 | JournalFile *f, | |
1020 | const dual_timestamp *ts, | |
1021 | uint64_t xor_hash, | |
1022 | const EntryItem items[], unsigned n_items, | |
de190aef | 1023 | uint64_t *seqnum, |
cec736d2 LP |
1024 | Object **ret, uint64_t *offset) { |
1025 | uint64_t np; | |
1026 | uint64_t osize; | |
1027 | Object *o; | |
1028 | int r; | |
1029 | ||
1030 | assert(f); | |
1031 | assert(items || n_items == 0); | |
de190aef | 1032 | assert(ts); |
cec736d2 LP |
1033 | |
1034 | osize = offsetof(Object, entry.items) + (n_items * sizeof(EntryItem)); | |
1035 | ||
de190aef | 1036 | r = journal_file_append_object(f, OBJECT_ENTRY, osize, &o, &np); |
cec736d2 LP |
1037 | if (r < 0) |
1038 | return r; | |
1039 | ||
de190aef | 1040 | o->entry.seqnum = htole64(journal_file_seqnum(f, seqnum)); |
cec736d2 | 1041 | memcpy(o->entry.items, items, n_items * sizeof(EntryItem)); |
de190aef LP |
1042 | o->entry.realtime = htole64(ts->realtime); |
1043 | o->entry.monotonic = htole64(ts->monotonic); | |
cec736d2 LP |
1044 | o->entry.xor_hash = htole64(xor_hash); |
1045 | o->entry.boot_id = f->header->boot_id; | |
1046 | ||
1047 | r = journal_file_link_entry(f, o, np); | |
1048 | if (r < 0) | |
1049 | return r; | |
1050 | ||
1051 | if (ret) | |
1052 | *ret = o; | |
1053 | ||
1054 | if (offset) | |
1055 | *offset = np; | |
1056 | ||
1057 | return 0; | |
1058 | } | |
1059 | ||
cf244689 | 1060 | void journal_file_post_change(JournalFile *f) { |
50f20cfd LP |
1061 | assert(f); |
1062 | ||
1063 | /* inotify() does not receive IN_MODIFY events from file | |
1064 | * accesses done via mmap(). After each access we hence | |
1065 | * trigger IN_MODIFY by truncating the journal file to its | |
1066 | * current size which triggers IN_MODIFY. */ | |
1067 | ||
bc85bfee LP |
1068 | __sync_synchronize(); |
1069 | ||
50f20cfd LP |
1070 | if (ftruncate(f->fd, f->last_stat.st_size) < 0) |
1071 | log_error("Failed to to truncate file to its own size: %m"); | |
1072 | } | |
1073 | ||
de190aef | 1074 | 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 |
1075 | unsigned i; |
1076 | EntryItem *items; | |
1077 | int r; | |
1078 | uint64_t xor_hash = 0; | |
de190aef | 1079 | struct dual_timestamp _ts; |
cec736d2 LP |
1080 | |
1081 | assert(f); | |
1082 | assert(iovec || n_iovec == 0); | |
1083 | ||
de190aef LP |
1084 | if (!f->writable) |
1085 | return -EPERM; | |
1086 | ||
1087 | if (!ts) { | |
1088 | dual_timestamp_get(&_ts); | |
1089 | ts = &_ts; | |
1090 | } | |
1091 | ||
1092 | if (f->tail_entry_monotonic_valid && | |
1093 | ts->monotonic < le64toh(f->header->tail_entry_monotonic)) | |
1094 | return -EINVAL; | |
1095 | ||
7560fffc LP |
1096 | r = journal_file_maybe_append_tag(f, ts->realtime); |
1097 | if (r < 0) | |
1098 | return r; | |
1099 | ||
64825d3c LP |
1100 | /* alloca() can't take 0, hence let's allocate at least one */ |
1101 | items = alloca(sizeof(EntryItem) * MAX(1, n_iovec)); | |
cec736d2 LP |
1102 | |
1103 | for (i = 0; i < n_iovec; i++) { | |
1104 | uint64_t p; | |
1105 | Object *o; | |
1106 | ||
1107 | r = journal_file_append_data(f, iovec[i].iov_base, iovec[i].iov_len, &o, &p); | |
1108 | if (r < 0) | |
cf244689 | 1109 | return r; |
cec736d2 LP |
1110 | |
1111 | xor_hash ^= le64toh(o->data.hash); | |
1112 | items[i].object_offset = htole64(p); | |
de7b95cd | 1113 | items[i].hash = o->data.hash; |
cec736d2 LP |
1114 | } |
1115 | ||
de190aef | 1116 | r = journal_file_append_entry_internal(f, ts, xor_hash, items, n_iovec, seqnum, ret, offset); |
cec736d2 | 1117 | |
50f20cfd LP |
1118 | journal_file_post_change(f); |
1119 | ||
cec736d2 LP |
1120 | return r; |
1121 | } | |
1122 | ||
de190aef LP |
1123 | static int generic_array_get(JournalFile *f, |
1124 | uint64_t first, | |
1125 | uint64_t i, | |
1126 | Object **ret, uint64_t *offset) { | |
1127 | ||
cec736d2 | 1128 | Object *o; |
6c8a39b8 | 1129 | uint64_t p = 0, a; |
cec736d2 LP |
1130 | int r; |
1131 | ||
1132 | assert(f); | |
1133 | ||
de190aef LP |
1134 | a = first; |
1135 | while (a > 0) { | |
1136 | uint64_t n; | |
cec736d2 | 1137 | |
de190aef LP |
1138 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &o); |
1139 | if (r < 0) | |
1140 | return r; | |
cec736d2 | 1141 | |
de190aef LP |
1142 | n = journal_file_entry_array_n_items(o); |
1143 | if (i < n) { | |
1144 | p = le64toh(o->entry_array.items[i]); | |
1145 | break; | |
cec736d2 LP |
1146 | } |
1147 | ||
de190aef LP |
1148 | i -= n; |
1149 | a = le64toh(o->entry_array.next_entry_array_offset); | |
1150 | } | |
1151 | ||
1152 | if (a <= 0 || p <= 0) | |
1153 | return 0; | |
1154 | ||
1155 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1156 | if (r < 0) | |
1157 | return r; | |
1158 | ||
1159 | if (ret) | |
1160 | *ret = o; | |
1161 | ||
1162 | if (offset) | |
1163 | *offset = p; | |
1164 | ||
1165 | return 1; | |
1166 | } | |
1167 | ||
1168 | static int generic_array_get_plus_one(JournalFile *f, | |
1169 | uint64_t extra, | |
1170 | uint64_t first, | |
1171 | uint64_t i, | |
1172 | Object **ret, uint64_t *offset) { | |
1173 | ||
1174 | Object *o; | |
1175 | ||
1176 | assert(f); | |
1177 | ||
1178 | if (i == 0) { | |
1179 | int r; | |
1180 | ||
1181 | r = journal_file_move_to_object(f, OBJECT_ENTRY, extra, &o); | |
cec736d2 LP |
1182 | if (r < 0) |
1183 | return r; | |
1184 | ||
de190aef LP |
1185 | if (ret) |
1186 | *ret = o; | |
cec736d2 | 1187 | |
de190aef LP |
1188 | if (offset) |
1189 | *offset = extra; | |
cec736d2 | 1190 | |
de190aef | 1191 | return 1; |
cec736d2 LP |
1192 | } |
1193 | ||
de190aef LP |
1194 | return generic_array_get(f, first, i-1, ret, offset); |
1195 | } | |
cec736d2 | 1196 | |
de190aef LP |
1197 | enum { |
1198 | TEST_FOUND, | |
1199 | TEST_LEFT, | |
1200 | TEST_RIGHT | |
1201 | }; | |
cec736d2 | 1202 | |
de190aef LP |
1203 | static int generic_array_bisect(JournalFile *f, |
1204 | uint64_t first, | |
1205 | uint64_t n, | |
1206 | uint64_t needle, | |
1207 | int (*test_object)(JournalFile *f, uint64_t p, uint64_t needle), | |
1208 | direction_t direction, | |
1209 | Object **ret, | |
1210 | uint64_t *offset, | |
1211 | uint64_t *idx) { | |
1212 | ||
1213 | uint64_t a, p, t = 0, i = 0, last_p = 0; | |
1214 | bool subtract_one = false; | |
1215 | Object *o, *array = NULL; | |
1216 | int r; | |
cec736d2 | 1217 | |
de190aef LP |
1218 | assert(f); |
1219 | assert(test_object); | |
cec736d2 | 1220 | |
de190aef LP |
1221 | a = first; |
1222 | while (a > 0) { | |
1223 | uint64_t left, right, k, lp; | |
1224 | ||
1225 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &array); | |
cec736d2 LP |
1226 | if (r < 0) |
1227 | return r; | |
1228 | ||
de190aef LP |
1229 | k = journal_file_entry_array_n_items(array); |
1230 | right = MIN(k, n); | |
1231 | if (right <= 0) | |
1232 | return 0; | |
cec736d2 | 1233 | |
de190aef LP |
1234 | i = right - 1; |
1235 | lp = p = le64toh(array->entry_array.items[i]); | |
1236 | if (p <= 0) | |
1237 | return -EBADMSG; | |
cec736d2 | 1238 | |
de190aef LP |
1239 | r = test_object(f, p, needle); |
1240 | if (r < 0) | |
1241 | return r; | |
cec736d2 | 1242 | |
de190aef LP |
1243 | if (r == TEST_FOUND) |
1244 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1245 | ||
1246 | if (r == TEST_RIGHT) { | |
1247 | left = 0; | |
1248 | right -= 1; | |
1249 | for (;;) { | |
1250 | if (left == right) { | |
1251 | if (direction == DIRECTION_UP) | |
1252 | subtract_one = true; | |
1253 | ||
1254 | i = left; | |
1255 | goto found; | |
1256 | } | |
1257 | ||
1258 | assert(left < right); | |
1259 | ||
1260 | i = (left + right) / 2; | |
1261 | p = le64toh(array->entry_array.items[i]); | |
1262 | if (p <= 0) | |
1263 | return -EBADMSG; | |
1264 | ||
1265 | r = test_object(f, p, needle); | |
1266 | if (r < 0) | |
1267 | return r; | |
cec736d2 | 1268 | |
de190aef LP |
1269 | if (r == TEST_FOUND) |
1270 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1271 | ||
1272 | if (r == TEST_RIGHT) | |
1273 | right = i; | |
1274 | else | |
1275 | left = i + 1; | |
1276 | } | |
1277 | } | |
1278 | ||
cbdca852 LP |
1279 | if (k > n) { |
1280 | if (direction == DIRECTION_UP) { | |
1281 | i = n; | |
1282 | subtract_one = true; | |
1283 | goto found; | |
1284 | } | |
1285 | ||
cec736d2 | 1286 | return 0; |
cbdca852 | 1287 | } |
cec736d2 | 1288 | |
de190aef LP |
1289 | last_p = lp; |
1290 | ||
1291 | n -= k; | |
1292 | t += k; | |
1293 | a = le64toh(array->entry_array.next_entry_array_offset); | |
cec736d2 LP |
1294 | } |
1295 | ||
1296 | return 0; | |
de190aef LP |
1297 | |
1298 | found: | |
1299 | if (subtract_one && t == 0 && i == 0) | |
1300 | return 0; | |
1301 | ||
1302 | if (subtract_one && i == 0) | |
1303 | p = last_p; | |
1304 | else if (subtract_one) | |
1305 | p = le64toh(array->entry_array.items[i-1]); | |
1306 | else | |
1307 | p = le64toh(array->entry_array.items[i]); | |
1308 | ||
1309 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1310 | if (r < 0) | |
1311 | return r; | |
1312 | ||
1313 | if (ret) | |
1314 | *ret = o; | |
1315 | ||
1316 | if (offset) | |
1317 | *offset = p; | |
1318 | ||
1319 | if (idx) | |
cbdca852 | 1320 | *idx = t + i + (subtract_one ? -1 : 0); |
de190aef LP |
1321 | |
1322 | return 1; | |
cec736d2 LP |
1323 | } |
1324 | ||
de190aef LP |
1325 | static int generic_array_bisect_plus_one(JournalFile *f, |
1326 | uint64_t extra, | |
1327 | uint64_t first, | |
1328 | uint64_t n, | |
1329 | uint64_t needle, | |
1330 | int (*test_object)(JournalFile *f, uint64_t p, uint64_t needle), | |
1331 | direction_t direction, | |
1332 | Object **ret, | |
1333 | uint64_t *offset, | |
1334 | uint64_t *idx) { | |
1335 | ||
cec736d2 | 1336 | int r; |
cbdca852 LP |
1337 | bool step_back = false; |
1338 | Object *o; | |
cec736d2 LP |
1339 | |
1340 | assert(f); | |
de190aef | 1341 | assert(test_object); |
cec736d2 | 1342 | |
de190aef LP |
1343 | if (n <= 0) |
1344 | return 0; | |
cec736d2 | 1345 | |
de190aef LP |
1346 | /* This bisects the array in object 'first', but first checks |
1347 | * an extra */ | |
de190aef LP |
1348 | r = test_object(f, extra, needle); |
1349 | if (r < 0) | |
1350 | return r; | |
a536e261 LP |
1351 | |
1352 | if (r == TEST_FOUND) | |
1353 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1354 | ||
cbdca852 LP |
1355 | /* if we are looking with DIRECTION_UP then we need to first |
1356 | see if in the actual array there is a matching entry, and | |
1357 | return the last one of that. But if there isn't any we need | |
1358 | to return this one. Hence remember this, and return it | |
1359 | below. */ | |
1360 | if (r == TEST_LEFT) | |
1361 | step_back = direction == DIRECTION_UP; | |
de190aef | 1362 | |
cbdca852 LP |
1363 | if (r == TEST_RIGHT) { |
1364 | if (direction == DIRECTION_DOWN) | |
1365 | goto found; | |
1366 | else | |
1367 | return 0; | |
a536e261 | 1368 | } |
cec736d2 | 1369 | |
de190aef LP |
1370 | r = generic_array_bisect(f, first, n-1, needle, test_object, direction, ret, offset, idx); |
1371 | ||
cbdca852 LP |
1372 | if (r == 0 && step_back) |
1373 | goto found; | |
1374 | ||
ecf68b1d | 1375 | if (r > 0 && idx) |
de190aef LP |
1376 | (*idx) ++; |
1377 | ||
1378 | return r; | |
cbdca852 LP |
1379 | |
1380 | found: | |
1381 | r = journal_file_move_to_object(f, OBJECT_ENTRY, extra, &o); | |
1382 | if (r < 0) | |
1383 | return r; | |
1384 | ||
1385 | if (ret) | |
1386 | *ret = o; | |
1387 | ||
1388 | if (offset) | |
1389 | *offset = extra; | |
1390 | ||
1391 | if (idx) | |
1392 | *idx = 0; | |
1393 | ||
1394 | return 1; | |
1395 | } | |
1396 | ||
1397 | static int test_object_offset(JournalFile *f, uint64_t p, uint64_t needle) { | |
1398 | assert(f); | |
1399 | assert(p > 0); | |
1400 | ||
1401 | if (p == needle) | |
1402 | return TEST_FOUND; | |
1403 | else if (p < needle) | |
1404 | return TEST_LEFT; | |
1405 | else | |
1406 | return TEST_RIGHT; | |
1407 | } | |
1408 | ||
1409 | int journal_file_move_to_entry_by_offset( | |
1410 | JournalFile *f, | |
1411 | uint64_t p, | |
1412 | direction_t direction, | |
1413 | Object **ret, | |
1414 | uint64_t *offset) { | |
1415 | ||
1416 | return generic_array_bisect(f, | |
1417 | le64toh(f->header->entry_array_offset), | |
1418 | le64toh(f->header->n_entries), | |
1419 | p, | |
1420 | test_object_offset, | |
1421 | direction, | |
1422 | ret, offset, NULL); | |
de190aef LP |
1423 | } |
1424 | ||
cbdca852 | 1425 | |
de190aef LP |
1426 | static int test_object_seqnum(JournalFile *f, uint64_t p, uint64_t needle) { |
1427 | Object *o; | |
1428 | int r; | |
1429 | ||
1430 | assert(f); | |
1431 | assert(p > 0); | |
1432 | ||
1433 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
cec736d2 LP |
1434 | if (r < 0) |
1435 | return r; | |
1436 | ||
de190aef LP |
1437 | if (le64toh(o->entry.seqnum) == needle) |
1438 | return TEST_FOUND; | |
1439 | else if (le64toh(o->entry.seqnum) < needle) | |
1440 | return TEST_LEFT; | |
1441 | else | |
1442 | return TEST_RIGHT; | |
1443 | } | |
cec736d2 | 1444 | |
de190aef LP |
1445 | int journal_file_move_to_entry_by_seqnum( |
1446 | JournalFile *f, | |
1447 | uint64_t seqnum, | |
1448 | direction_t direction, | |
1449 | Object **ret, | |
1450 | uint64_t *offset) { | |
1451 | ||
1452 | return generic_array_bisect(f, | |
1453 | le64toh(f->header->entry_array_offset), | |
1454 | le64toh(f->header->n_entries), | |
1455 | seqnum, | |
1456 | test_object_seqnum, | |
1457 | direction, | |
1458 | ret, offset, NULL); | |
1459 | } | |
cec736d2 | 1460 | |
de190aef LP |
1461 | static int test_object_realtime(JournalFile *f, uint64_t p, uint64_t needle) { |
1462 | Object *o; | |
1463 | int r; | |
1464 | ||
1465 | assert(f); | |
1466 | assert(p > 0); | |
1467 | ||
1468 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1469 | if (r < 0) | |
1470 | return r; | |
1471 | ||
1472 | if (le64toh(o->entry.realtime) == needle) | |
1473 | return TEST_FOUND; | |
1474 | else if (le64toh(o->entry.realtime) < needle) | |
1475 | return TEST_LEFT; | |
1476 | else | |
1477 | return TEST_RIGHT; | |
cec736d2 LP |
1478 | } |
1479 | ||
de190aef LP |
1480 | int journal_file_move_to_entry_by_realtime( |
1481 | JournalFile *f, | |
1482 | uint64_t realtime, | |
1483 | direction_t direction, | |
1484 | Object **ret, | |
1485 | uint64_t *offset) { | |
1486 | ||
1487 | return generic_array_bisect(f, | |
1488 | le64toh(f->header->entry_array_offset), | |
1489 | le64toh(f->header->n_entries), | |
1490 | realtime, | |
1491 | test_object_realtime, | |
1492 | direction, | |
1493 | ret, offset, NULL); | |
1494 | } | |
1495 | ||
1496 | static int test_object_monotonic(JournalFile *f, uint64_t p, uint64_t needle) { | |
1497 | Object *o; | |
1498 | int r; | |
1499 | ||
1500 | assert(f); | |
1501 | assert(p > 0); | |
1502 | ||
1503 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1504 | if (r < 0) | |
1505 | return r; | |
1506 | ||
1507 | if (le64toh(o->entry.monotonic) == needle) | |
1508 | return TEST_FOUND; | |
1509 | else if (le64toh(o->entry.monotonic) < needle) | |
1510 | return TEST_LEFT; | |
1511 | else | |
1512 | return TEST_RIGHT; | |
1513 | } | |
1514 | ||
1515 | int journal_file_move_to_entry_by_monotonic( | |
1516 | JournalFile *f, | |
1517 | sd_id128_t boot_id, | |
1518 | uint64_t monotonic, | |
1519 | direction_t direction, | |
1520 | Object **ret, | |
1521 | uint64_t *offset) { | |
1522 | ||
10b6f904 | 1523 | char t[9+32+1] = "_BOOT_ID="; |
de190aef LP |
1524 | Object *o; |
1525 | int r; | |
1526 | ||
cbdca852 | 1527 | assert(f); |
de190aef | 1528 | |
cbdca852 | 1529 | sd_id128_to_string(boot_id, t + 9); |
de190aef LP |
1530 | r = journal_file_find_data_object(f, t, strlen(t), &o, NULL); |
1531 | if (r < 0) | |
1532 | return r; | |
cbdca852 | 1533 | if (r == 0) |
de190aef LP |
1534 | return -ENOENT; |
1535 | ||
1536 | return generic_array_bisect_plus_one(f, | |
1537 | le64toh(o->data.entry_offset), | |
1538 | le64toh(o->data.entry_array_offset), | |
1539 | le64toh(o->data.n_entries), | |
1540 | monotonic, | |
1541 | test_object_monotonic, | |
1542 | direction, | |
1543 | ret, offset, NULL); | |
1544 | } | |
1545 | ||
de190aef LP |
1546 | int journal_file_next_entry( |
1547 | JournalFile *f, | |
1548 | Object *o, uint64_t p, | |
1549 | direction_t direction, | |
1550 | Object **ret, uint64_t *offset) { | |
1551 | ||
1552 | uint64_t i, n; | |
cec736d2 LP |
1553 | int r; |
1554 | ||
1555 | assert(f); | |
de190aef LP |
1556 | assert(p > 0 || !o); |
1557 | ||
1558 | n = le64toh(f->header->n_entries); | |
1559 | if (n <= 0) | |
1560 | return 0; | |
cec736d2 LP |
1561 | |
1562 | if (!o) | |
de190aef | 1563 | i = direction == DIRECTION_DOWN ? 0 : n - 1; |
cec736d2 | 1564 | else { |
de190aef | 1565 | if (o->object.type != OBJECT_ENTRY) |
cec736d2 LP |
1566 | return -EINVAL; |
1567 | ||
de190aef LP |
1568 | r = generic_array_bisect(f, |
1569 | le64toh(f->header->entry_array_offset), | |
1570 | le64toh(f->header->n_entries), | |
1571 | p, | |
1572 | test_object_offset, | |
1573 | DIRECTION_DOWN, | |
1574 | NULL, NULL, | |
1575 | &i); | |
1576 | if (r <= 0) | |
1577 | return r; | |
1578 | ||
1579 | if (direction == DIRECTION_DOWN) { | |
1580 | if (i >= n - 1) | |
1581 | return 0; | |
1582 | ||
1583 | i++; | |
1584 | } else { | |
1585 | if (i <= 0) | |
1586 | return 0; | |
1587 | ||
1588 | i--; | |
1589 | } | |
cec736d2 LP |
1590 | } |
1591 | ||
de190aef LP |
1592 | /* And jump to it */ |
1593 | return generic_array_get(f, | |
1594 | le64toh(f->header->entry_array_offset), | |
1595 | i, | |
1596 | ret, offset); | |
1597 | } | |
cec736d2 | 1598 | |
de190aef LP |
1599 | int journal_file_skip_entry( |
1600 | JournalFile *f, | |
1601 | Object *o, uint64_t p, | |
1602 | int64_t skip, | |
1603 | Object **ret, uint64_t *offset) { | |
1604 | ||
1605 | uint64_t i, n; | |
1606 | int r; | |
1607 | ||
1608 | assert(f); | |
1609 | assert(o); | |
1610 | assert(p > 0); | |
1611 | ||
1612 | if (o->object.type != OBJECT_ENTRY) | |
1613 | return -EINVAL; | |
1614 | ||
1615 | r = generic_array_bisect(f, | |
1616 | le64toh(f->header->entry_array_offset), | |
1617 | le64toh(f->header->n_entries), | |
1618 | p, | |
1619 | test_object_offset, | |
1620 | DIRECTION_DOWN, | |
1621 | NULL, NULL, | |
1622 | &i); | |
1623 | if (r <= 0) | |
cec736d2 LP |
1624 | return r; |
1625 | ||
de190aef LP |
1626 | /* Calculate new index */ |
1627 | if (skip < 0) { | |
1628 | if ((uint64_t) -skip >= i) | |
1629 | i = 0; | |
1630 | else | |
1631 | i = i - (uint64_t) -skip; | |
1632 | } else | |
1633 | i += (uint64_t) skip; | |
cec736d2 | 1634 | |
de190aef LP |
1635 | n = le64toh(f->header->n_entries); |
1636 | if (n <= 0) | |
1637 | return -EBADMSG; | |
cec736d2 | 1638 | |
de190aef LP |
1639 | if (i >= n) |
1640 | i = n-1; | |
1641 | ||
1642 | return generic_array_get(f, | |
1643 | le64toh(f->header->entry_array_offset), | |
1644 | i, | |
1645 | ret, offset); | |
cec736d2 LP |
1646 | } |
1647 | ||
de190aef LP |
1648 | int journal_file_next_entry_for_data( |
1649 | JournalFile *f, | |
1650 | Object *o, uint64_t p, | |
1651 | uint64_t data_offset, | |
1652 | direction_t direction, | |
1653 | Object **ret, uint64_t *offset) { | |
1654 | ||
1655 | uint64_t n, i; | |
cec736d2 | 1656 | int r; |
de190aef | 1657 | Object *d; |
cec736d2 LP |
1658 | |
1659 | assert(f); | |
de190aef | 1660 | assert(p > 0 || !o); |
cec736d2 | 1661 | |
de190aef | 1662 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); |
466ccd92 | 1663 | if (r < 0) |
de190aef | 1664 | return r; |
cec736d2 | 1665 | |
de190aef LP |
1666 | n = le64toh(d->data.n_entries); |
1667 | if (n <= 0) | |
1668 | return n; | |
cec736d2 | 1669 | |
de190aef LP |
1670 | if (!o) |
1671 | i = direction == DIRECTION_DOWN ? 0 : n - 1; | |
1672 | else { | |
1673 | if (o->object.type != OBJECT_ENTRY) | |
1674 | return -EINVAL; | |
cec736d2 | 1675 | |
de190aef LP |
1676 | r = generic_array_bisect_plus_one(f, |
1677 | le64toh(d->data.entry_offset), | |
1678 | le64toh(d->data.entry_array_offset), | |
1679 | le64toh(d->data.n_entries), | |
1680 | p, | |
1681 | test_object_offset, | |
1682 | DIRECTION_DOWN, | |
1683 | NULL, NULL, | |
1684 | &i); | |
1685 | ||
1686 | if (r <= 0) | |
cec736d2 LP |
1687 | return r; |
1688 | ||
de190aef LP |
1689 | if (direction == DIRECTION_DOWN) { |
1690 | if (i >= n - 1) | |
1691 | return 0; | |
cec736d2 | 1692 | |
de190aef LP |
1693 | i++; |
1694 | } else { | |
1695 | if (i <= 0) | |
1696 | return 0; | |
cec736d2 | 1697 | |
de190aef LP |
1698 | i--; |
1699 | } | |
cec736d2 | 1700 | |
de190aef | 1701 | } |
cec736d2 | 1702 | |
de190aef LP |
1703 | return generic_array_get_plus_one(f, |
1704 | le64toh(d->data.entry_offset), | |
1705 | le64toh(d->data.entry_array_offset), | |
1706 | i, | |
1707 | ret, offset); | |
1708 | } | |
cec736d2 | 1709 | |
cbdca852 LP |
1710 | int journal_file_move_to_entry_by_offset_for_data( |
1711 | JournalFile *f, | |
1712 | uint64_t data_offset, | |
1713 | uint64_t p, | |
1714 | direction_t direction, | |
1715 | Object **ret, uint64_t *offset) { | |
1716 | ||
1717 | int r; | |
1718 | Object *d; | |
1719 | ||
1720 | assert(f); | |
1721 | ||
1722 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); | |
1723 | if (r < 0) | |
1724 | return r; | |
1725 | ||
1726 | return generic_array_bisect_plus_one(f, | |
1727 | le64toh(d->data.entry_offset), | |
1728 | le64toh(d->data.entry_array_offset), | |
1729 | le64toh(d->data.n_entries), | |
1730 | p, | |
1731 | test_object_offset, | |
1732 | direction, | |
1733 | ret, offset, NULL); | |
1734 | } | |
1735 | ||
1736 | int journal_file_move_to_entry_by_monotonic_for_data( | |
1737 | JournalFile *f, | |
1738 | uint64_t data_offset, | |
1739 | sd_id128_t boot_id, | |
1740 | uint64_t monotonic, | |
1741 | direction_t direction, | |
1742 | Object **ret, uint64_t *offset) { | |
1743 | ||
1744 | char t[9+32+1] = "_BOOT_ID="; | |
1745 | Object *o, *d; | |
1746 | int r; | |
1747 | uint64_t b, z; | |
1748 | ||
1749 | assert(f); | |
1750 | ||
1751 | /* First, seek by time */ | |
1752 | sd_id128_to_string(boot_id, t + 9); | |
1753 | r = journal_file_find_data_object(f, t, strlen(t), &o, &b); | |
1754 | if (r < 0) | |
1755 | return r; | |
1756 | if (r == 0) | |
1757 | return -ENOENT; | |
1758 | ||
1759 | r = generic_array_bisect_plus_one(f, | |
1760 | le64toh(o->data.entry_offset), | |
1761 | le64toh(o->data.entry_array_offset), | |
1762 | le64toh(o->data.n_entries), | |
1763 | monotonic, | |
1764 | test_object_monotonic, | |
1765 | direction, | |
1766 | NULL, &z, NULL); | |
1767 | if (r <= 0) | |
1768 | return r; | |
1769 | ||
1770 | /* And now, continue seeking until we find an entry that | |
1771 | * exists in both bisection arrays */ | |
1772 | ||
1773 | for (;;) { | |
1774 | Object *qo; | |
1775 | uint64_t p, q; | |
1776 | ||
1777 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); | |
1778 | if (r < 0) | |
1779 | return r; | |
1780 | ||
1781 | r = generic_array_bisect_plus_one(f, | |
1782 | le64toh(d->data.entry_offset), | |
1783 | le64toh(d->data.entry_array_offset), | |
1784 | le64toh(d->data.n_entries), | |
1785 | z, | |
1786 | test_object_offset, | |
1787 | direction, | |
1788 | NULL, &p, NULL); | |
1789 | if (r <= 0) | |
1790 | return r; | |
1791 | ||
1792 | r = journal_file_move_to_object(f, OBJECT_DATA, b, &o); | |
1793 | if (r < 0) | |
1794 | return r; | |
1795 | ||
1796 | r = generic_array_bisect_plus_one(f, | |
1797 | le64toh(o->data.entry_offset), | |
1798 | le64toh(o->data.entry_array_offset), | |
1799 | le64toh(o->data.n_entries), | |
1800 | p, | |
1801 | test_object_offset, | |
1802 | direction, | |
1803 | &qo, &q, NULL); | |
1804 | ||
1805 | if (r <= 0) | |
1806 | return r; | |
1807 | ||
1808 | if (p == q) { | |
1809 | if (ret) | |
1810 | *ret = qo; | |
1811 | if (offset) | |
1812 | *offset = q; | |
1813 | ||
1814 | return 1; | |
1815 | } | |
1816 | ||
1817 | z = q; | |
1818 | } | |
1819 | ||
1820 | return 0; | |
1821 | } | |
1822 | ||
de190aef LP |
1823 | int journal_file_move_to_entry_by_seqnum_for_data( |
1824 | JournalFile *f, | |
1825 | uint64_t data_offset, | |
1826 | uint64_t seqnum, | |
1827 | direction_t direction, | |
1828 | Object **ret, uint64_t *offset) { | |
cec736d2 | 1829 | |
de190aef LP |
1830 | Object *d; |
1831 | int r; | |
cec736d2 | 1832 | |
91a31dde LP |
1833 | assert(f); |
1834 | ||
de190aef | 1835 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); |
91a31dde | 1836 | if (r < 0) |
de190aef | 1837 | return r; |
cec736d2 | 1838 | |
de190aef LP |
1839 | return generic_array_bisect_plus_one(f, |
1840 | le64toh(d->data.entry_offset), | |
1841 | le64toh(d->data.entry_array_offset), | |
1842 | le64toh(d->data.n_entries), | |
1843 | seqnum, | |
1844 | test_object_seqnum, | |
1845 | direction, | |
1846 | ret, offset, NULL); | |
1847 | } | |
cec736d2 | 1848 | |
de190aef LP |
1849 | int journal_file_move_to_entry_by_realtime_for_data( |
1850 | JournalFile *f, | |
1851 | uint64_t data_offset, | |
1852 | uint64_t realtime, | |
1853 | direction_t direction, | |
1854 | Object **ret, uint64_t *offset) { | |
1855 | ||
1856 | Object *d; | |
1857 | int r; | |
1858 | ||
91a31dde LP |
1859 | assert(f); |
1860 | ||
de190aef | 1861 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); |
91a31dde | 1862 | if (r < 0) |
de190aef LP |
1863 | return r; |
1864 | ||
1865 | return generic_array_bisect_plus_one(f, | |
1866 | le64toh(d->data.entry_offset), | |
1867 | le64toh(d->data.entry_array_offset), | |
1868 | le64toh(d->data.n_entries), | |
1869 | realtime, | |
1870 | test_object_realtime, | |
1871 | direction, | |
1872 | ret, offset, NULL); | |
cec736d2 LP |
1873 | } |
1874 | ||
7560fffc LP |
1875 | static void *fsprg_state(JournalFile *f) { |
1876 | uint64_t a, b; | |
1877 | assert(f); | |
1878 | ||
1879 | if (!f->authenticate) | |
1880 | return NULL; | |
1881 | ||
1882 | a = le64toh(f->fsprg_header->header_size); | |
1883 | b = le64toh(f->fsprg_header->state_size); | |
1884 | ||
1885 | if (a + b > f->fsprg_size) | |
1886 | return NULL; | |
1887 | ||
1888 | return (uint8_t*) f->fsprg_header + a; | |
1889 | } | |
1890 | ||
1891 | static int journal_file_append_tag(JournalFile *f) { | |
1892 | Object *o; | |
1893 | uint64_t p; | |
1894 | int r; | |
1895 | ||
1896 | assert(f); | |
1897 | ||
1898 | if (!f->authenticate) | |
1899 | return 0; | |
1900 | ||
1901 | if (!f->hmac_running) | |
1902 | return 0; | |
1903 | ||
1904 | log_debug("Writing tag for epoch %llu\n", (unsigned long long) FSPRG_GetEpoch(fsprg_state(f))); | |
1905 | ||
1906 | assert(f->hmac); | |
1907 | ||
1908 | r = journal_file_append_object(f, OBJECT_TAG, sizeof(struct TagObject), &o, &p); | |
1909 | if (r < 0) | |
1910 | return r; | |
1911 | ||
1912 | /* Get the HMAC tag and store it in the object */ | |
1913 | memcpy(o->tag.tag, gcry_md_read(f->hmac, 0), TAG_LENGTH); | |
1914 | f->hmac_running = false; | |
1915 | ||
1916 | return 0; | |
1917 | } | |
1918 | ||
1919 | static int journal_file_hmac_start(JournalFile *f) { | |
1920 | uint8_t key[256 / 8]; /* Let's pass 256 bit from FSPRG to HMAC */ | |
1921 | ||
1922 | assert(f); | |
1923 | ||
1924 | if (!f->authenticate) | |
1925 | return 0; | |
1926 | ||
1927 | if (f->hmac_running) | |
1928 | return 0; | |
1929 | ||
1930 | /* Prepare HMAC for next cycle */ | |
1931 | gcry_md_reset(f->hmac); | |
1932 | FSPRG_GetKey(fsprg_state(f), key, sizeof(key), 0); | |
1933 | gcry_md_setkey(f->hmac, key, sizeof(key)); | |
1934 | ||
1935 | f->hmac_running = true; | |
1936 | ||
1937 | return 0; | |
1938 | } | |
1939 | ||
1940 | static int journal_file_get_epoch(JournalFile *f, uint64_t realtime, uint64_t *epoch) { | |
1941 | uint64_t t; | |
1942 | ||
1943 | assert(f); | |
1944 | assert(epoch); | |
1945 | assert(f->authenticate); | |
1946 | ||
1947 | if (le64toh(f->fsprg_header->fsprg_start_usec) == 0 || | |
1948 | le64toh(f->fsprg_header->fsprg_interval_usec) == 0) | |
1949 | return -ENOTSUP; | |
1950 | ||
1951 | if (realtime < le64toh(f->fsprg_header->fsprg_start_usec)) | |
1952 | return -ESTALE; | |
1953 | ||
1954 | t = realtime - le64toh(f->fsprg_header->fsprg_start_usec); | |
1955 | t = t / le64toh(f->fsprg_header->fsprg_interval_usec); | |
1956 | ||
1957 | *epoch = t; | |
1958 | return 0; | |
1959 | } | |
1960 | ||
1961 | static int journal_file_need_evolve(JournalFile *f, uint64_t realtime) { | |
1962 | uint64_t goal, epoch; | |
1963 | int r; | |
1964 | assert(f); | |
1965 | ||
1966 | if (!f->authenticate) | |
1967 | return 0; | |
1968 | ||
1969 | r = journal_file_get_epoch(f, realtime, &goal); | |
1970 | if (r < 0) | |
1971 | return r; | |
1972 | ||
1973 | epoch = FSPRG_GetEpoch(fsprg_state(f)); | |
1974 | if (epoch > goal) | |
1975 | return -ESTALE; | |
1976 | ||
1977 | return epoch != goal; | |
1978 | } | |
1979 | ||
1980 | static int journal_file_evolve(JournalFile *f, uint64_t realtime) { | |
1981 | uint64_t goal, epoch; | |
1982 | int r; | |
1983 | ||
1984 | assert(f); | |
1985 | ||
1986 | if (!f->authenticate) | |
1987 | return 0; | |
1988 | ||
1989 | r = journal_file_get_epoch(f, realtime, &goal); | |
1990 | if (r < 0) | |
1991 | return r; | |
1992 | ||
1993 | epoch = FSPRG_GetEpoch(fsprg_state(f)); | |
1994 | if (epoch < goal) | |
1995 | log_debug("Evolving FSPRG key from epoch %llu to %llu.", (unsigned long long) epoch, (unsigned long long) goal); | |
1996 | ||
1997 | for (;;) { | |
1998 | if (epoch > goal) | |
1999 | return -ESTALE; | |
2000 | if (epoch == goal) | |
2001 | return 0; | |
2002 | ||
2003 | FSPRG_Evolve(fsprg_state(f)); | |
2004 | epoch = FSPRG_GetEpoch(fsprg_state(f)); | |
2005 | } | |
2006 | } | |
2007 | ||
2008 | static int journal_file_maybe_append_tag(JournalFile *f, uint64_t realtime) { | |
2009 | int r; | |
2010 | ||
2011 | assert(f); | |
2012 | ||
2013 | if (!f->authenticate) | |
2014 | return 0; | |
2015 | ||
2016 | r = journal_file_need_evolve(f, realtime); | |
2017 | if (r <= 0) | |
2018 | return 0; | |
2019 | ||
2020 | r = journal_file_append_tag(f); | |
2021 | if (r < 0) | |
2022 | return r; | |
2023 | ||
2024 | r = journal_file_evolve(f, realtime); | |
2025 | if (r < 0) | |
2026 | return r; | |
2027 | ||
2028 | r = journal_file_hmac_start(f); | |
2029 | if (r < 0) | |
2030 | return r; | |
2031 | ||
2032 | return 0; | |
2033 | } | |
2034 | ||
2035 | static int journal_file_hmac_put_object(JournalFile *f, int type, uint64_t p) { | |
2036 | int r; | |
2037 | Object *o; | |
2038 | ||
2039 | assert(f); | |
2040 | ||
2041 | if (!f->authenticate) | |
2042 | return 0; | |
2043 | ||
2044 | r = journal_file_hmac_start(f); | |
2045 | if (r < 0) | |
2046 | return r; | |
2047 | ||
2048 | r = journal_file_move_to_object(f, type, p, &o); | |
2049 | if (r < 0) | |
2050 | return r; | |
2051 | ||
2052 | gcry_md_write(f->hmac, o, offsetof(ObjectHeader, payload)); | |
2053 | ||
2054 | switch (o->object.type) { | |
2055 | ||
2056 | case OBJECT_DATA: | |
2057 | /* All but: entry_array_offset, n_entries are mutable */ | |
2058 | gcry_md_write(f->hmac, &o->data.hash, offsetof(DataObject, entry_array_offset) - offsetof(DataObject, hash)); | |
2059 | gcry_md_write(f->hmac, o->data.payload, le64toh(o->object.size) - offsetof(DataObject, payload)); | |
2060 | break; | |
2061 | ||
2062 | case OBJECT_ENTRY: | |
2063 | /* All */ | |
2064 | gcry_md_write(f->hmac, &o->entry.seqnum, le64toh(o->object.size) - offsetof(EntryObject, seqnum)); | |
2065 | break; | |
2066 | ||
2067 | case OBJECT_FIELD_HASH_TABLE: | |
2068 | case OBJECT_DATA_HASH_TABLE: | |
2069 | case OBJECT_ENTRY_ARRAY: | |
2070 | /* Nothing: everything is mutable */ | |
2071 | break; | |
2072 | ||
2073 | case OBJECT_TAG: | |
2074 | /* All */ | |
2075 | gcry_md_write(f->hmac, o->tag.tag, le64toh(o->object.size) - offsetof(TagObject, tag)); | |
2076 | break; | |
2077 | ||
2078 | default: | |
2079 | return -EINVAL; | |
2080 | } | |
2081 | ||
2082 | return 0; | |
2083 | } | |
2084 | ||
2085 | static int journal_file_hmac_put_header(JournalFile *f) { | |
2086 | int r; | |
2087 | ||
2088 | assert(f); | |
2089 | ||
2090 | if (!f->authenticate) | |
2091 | return 0; | |
2092 | ||
2093 | r = journal_file_hmac_start(f); | |
2094 | if (r < 0) | |
2095 | return r; | |
2096 | ||
2097 | /* All but state+reserved, boot_id, arena_size, | |
2098 | * tail_object_offset, n_objects, n_entries, tail_seqnum, | |
2099 | * head_entry_realtime, tail_entry_realtime, | |
2100 | * tail_entry_monotonic, n_data, n_fields, header_tag */ | |
2101 | ||
2102 | gcry_md_write(f->hmac, f->header->signature, offsetof(Header, state) - offsetof(Header, signature)); | |
2103 | gcry_md_write(f->hmac, &f->header->file_id, offsetof(Header, boot_id) - offsetof(Header, file_id)); | |
2104 | gcry_md_write(f->hmac, &f->header->seqnum_id, offsetof(Header, arena_size) - offsetof(Header, seqnum_id)); | |
2105 | gcry_md_write(f->hmac, &f->header->data_hash_table_offset, offsetof(Header, tail_object_offset) - offsetof(Header, data_hash_table_offset)); | |
2106 | gcry_md_write(f->hmac, &f->header->head_seqnum, offsetof(Header, head_entry_realtime) - offsetof(Header, head_seqnum)); | |
2107 | ||
2108 | return 0; | |
2109 | } | |
2110 | ||
2111 | static int journal_file_load_fsprg(JournalFile *f) { | |
2112 | int r, fd = -1; | |
2113 | char *p = NULL; | |
2114 | struct stat st; | |
2115 | FSPRGHeader *m = NULL; | |
2116 | sd_id128_t machine; | |
2117 | ||
2118 | assert(f); | |
2119 | ||
2120 | if (!f->authenticate) | |
2121 | return 0; | |
2122 | ||
2123 | r = sd_id128_get_machine(&machine); | |
2124 | if (r < 0) | |
2125 | return r; | |
2126 | ||
2127 | if (asprintf(&p, "/var/log/journal/" SD_ID128_FORMAT_STR "/fsprg", | |
2128 | SD_ID128_FORMAT_VAL(machine)) < 0) | |
2129 | return -ENOMEM; | |
2130 | ||
2131 | fd = open(p, O_RDWR|O_CLOEXEC|O_NOCTTY, 0600); | |
2132 | if (fd < 0) { | |
2133 | log_error("Failed to open %s: %m", p); | |
2134 | r = -errno; | |
2135 | goto finish; | |
2136 | } | |
2137 | ||
2138 | if (fstat(fd, &st) < 0) { | |
2139 | r = -errno; | |
2140 | goto finish; | |
2141 | } | |
2142 | ||
2143 | if (st.st_size < (off_t) sizeof(FSPRGHeader)) { | |
2144 | r = -ENODATA; | |
2145 | goto finish; | |
2146 | } | |
2147 | ||
2148 | m = mmap(NULL, PAGE_ALIGN(sizeof(FSPRGHeader)), PROT_READ, MAP_SHARED, fd, 0); | |
2149 | if (m == MAP_FAILED) { | |
2150 | m = NULL; | |
2151 | r = -errno; | |
2152 | goto finish; | |
2153 | } | |
2154 | ||
2155 | if (memcmp(m->signature, FSPRG_HEADER_SIGNATURE, 8) != 0) { | |
2156 | r = -EBADMSG; | |
2157 | goto finish; | |
2158 | } | |
2159 | ||
2160 | if (m->incompatible_flags != 0) { | |
2161 | r = -EPROTONOSUPPORT; | |
2162 | goto finish; | |
2163 | } | |
2164 | ||
2165 | if (le64toh(m->header_size) < sizeof(FSPRGHeader)) { | |
2166 | r = -EBADMSG; | |
2167 | goto finish; | |
2168 | } | |
2169 | ||
2170 | if (le64toh(m->state_size) != FSPRG_stateinbytes(m->secpar)) { | |
2171 | r = -EBADMSG; | |
2172 | goto finish; | |
2173 | } | |
2174 | ||
2175 | f->fsprg_size = le64toh(m->header_size) + le64toh(m->state_size); | |
2176 | if ((uint64_t) st.st_size < f->fsprg_size) { | |
2177 | r = -ENODATA; | |
2178 | goto finish; | |
2179 | } | |
2180 | ||
2181 | if (!sd_id128_equal(machine, m->machine_id)) { | |
2182 | r = -EHOSTDOWN; | |
2183 | goto finish; | |
2184 | } | |
2185 | ||
2186 | if (le64toh(m->fsprg_start_usec) <= 0 || | |
2187 | le64toh(m->fsprg_interval_usec) <= 0) { | |
2188 | r = -EBADMSG; | |
2189 | goto finish; | |
2190 | } | |
2191 | ||
2192 | f->fsprg_header = mmap(NULL, PAGE_ALIGN(f->fsprg_size), PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0); | |
2193 | if (f->fsprg_header == MAP_FAILED) { | |
2194 | f->fsprg_header = NULL; | |
2195 | r = -errno; | |
2196 | goto finish; | |
2197 | } | |
2198 | ||
2199 | r = 0; | |
2200 | ||
2201 | finish: | |
2202 | if (m) | |
2203 | munmap(m, PAGE_ALIGN(sizeof(FSPRGHeader))); | |
2204 | ||
2205 | if (fd >= 0) | |
2206 | close_nointr_nofail(fd); | |
2207 | ||
2208 | free(p); | |
2209 | return r; | |
2210 | } | |
2211 | ||
2212 | static int journal_file_setup_hmac(JournalFile *f) { | |
2213 | gcry_error_t e; | |
2214 | ||
2215 | if (!f->authenticate) | |
2216 | return 0; | |
2217 | ||
2218 | e = gcry_md_open(&f->hmac, GCRY_MD_SHA256, GCRY_MD_FLAG_HMAC); | |
2219 | if (e != 0) | |
2220 | return -ENOTSUP; | |
2221 | ||
2222 | return 0; | |
2223 | } | |
2224 | ||
2225 | static int journal_file_append_first_tag(JournalFile *f) { | |
2226 | int r; | |
2227 | uint64_t p; | |
2228 | ||
2229 | if (!f->authenticate) | |
2230 | return 0; | |
2231 | ||
2232 | log_debug("Calculating first tag..."); | |
2233 | ||
2234 | r = journal_file_hmac_put_header(f); | |
2235 | if (r < 0) | |
2236 | return r; | |
2237 | ||
2238 | p = le64toh(f->header->field_hash_table_offset); | |
2239 | if (p < offsetof(Object, hash_table.items)) | |
2240 | return -EINVAL; | |
2241 | p -= offsetof(Object, hash_table.items); | |
2242 | ||
2243 | r = journal_file_hmac_put_object(f, OBJECT_FIELD_HASH_TABLE, p); | |
2244 | if (r < 0) | |
2245 | return r; | |
2246 | ||
2247 | p = le64toh(f->header->data_hash_table_offset); | |
2248 | if (p < offsetof(Object, hash_table.items)) | |
2249 | return -EINVAL; | |
2250 | p -= offsetof(Object, hash_table.items); | |
2251 | ||
2252 | r = journal_file_hmac_put_object(f, OBJECT_DATA_HASH_TABLE, p); | |
2253 | if (r < 0) | |
2254 | return r; | |
2255 | ||
2256 | r = journal_file_append_tag(f); | |
2257 | if (r < 0) | |
2258 | return r; | |
2259 | ||
2260 | return 0; | |
2261 | } | |
2262 | ||
cec736d2 | 2263 | void journal_file_dump(JournalFile *f) { |
cec736d2 LP |
2264 | Object *o; |
2265 | int r; | |
2266 | uint64_t p; | |
2267 | ||
2268 | assert(f); | |
2269 | ||
dca6219e | 2270 | journal_file_print_header(f); |
cec736d2 | 2271 | |
23b0b2b2 | 2272 | p = le64toh(f->header->header_size); |
cec736d2 | 2273 | while (p != 0) { |
de190aef | 2274 | r = journal_file_move_to_object(f, -1, p, &o); |
cec736d2 LP |
2275 | if (r < 0) |
2276 | goto fail; | |
2277 | ||
2278 | switch (o->object.type) { | |
2279 | ||
2280 | case OBJECT_UNUSED: | |
2281 | printf("Type: OBJECT_UNUSED\n"); | |
2282 | break; | |
2283 | ||
2284 | case OBJECT_DATA: | |
2285 | printf("Type: OBJECT_DATA\n"); | |
2286 | break; | |
2287 | ||
2288 | case OBJECT_ENTRY: | |
3fbf9cbb LP |
2289 | printf("Type: OBJECT_ENTRY %llu %llu %llu\n", |
2290 | (unsigned long long) le64toh(o->entry.seqnum), | |
2291 | (unsigned long long) le64toh(o->entry.monotonic), | |
2292 | (unsigned long long) le64toh(o->entry.realtime)); | |
cec736d2 LP |
2293 | break; |
2294 | ||
de190aef LP |
2295 | case OBJECT_FIELD_HASH_TABLE: |
2296 | printf("Type: OBJECT_FIELD_HASH_TABLE\n"); | |
cec736d2 LP |
2297 | break; |
2298 | ||
de190aef LP |
2299 | case OBJECT_DATA_HASH_TABLE: |
2300 | printf("Type: OBJECT_DATA_HASH_TABLE\n"); | |
2301 | break; | |
2302 | ||
2303 | case OBJECT_ENTRY_ARRAY: | |
2304 | printf("Type: OBJECT_ENTRY_ARRAY\n"); | |
cec736d2 | 2305 | break; |
8144056f | 2306 | |
7560fffc LP |
2307 | case OBJECT_TAG: |
2308 | printf("Type: OBJECT_TAG\n"); | |
8144056f | 2309 | break; |
cec736d2 LP |
2310 | } |
2311 | ||
807e17f0 LP |
2312 | if (o->object.flags & OBJECT_COMPRESSED) |
2313 | printf("Flags: COMPRESSED\n"); | |
2314 | ||
cec736d2 LP |
2315 | if (p == le64toh(f->header->tail_object_offset)) |
2316 | p = 0; | |
2317 | else | |
2318 | p = p + ALIGN64(le64toh(o->object.size)); | |
2319 | } | |
2320 | ||
2321 | return; | |
2322 | fail: | |
2323 | log_error("File corrupt"); | |
2324 | } | |
2325 | ||
dca6219e LP |
2326 | void journal_file_print_header(JournalFile *f) { |
2327 | char a[33], b[33], c[33]; | |
2328 | char x[FORMAT_TIMESTAMP_MAX], y[FORMAT_TIMESTAMP_MAX]; | |
2329 | ||
2330 | assert(f); | |
2331 | ||
2332 | printf("File Path: %s\n" | |
2333 | "File ID: %s\n" | |
2334 | "Machine ID: %s\n" | |
2335 | "Boot ID: %s\n" | |
2336 | "Sequential Number ID: %s\n" | |
dc36ac67 LP |
2337 | "State: %s\n" |
2338 | "Compatible Flags:%s%s\n" | |
2339 | "Incompatible Flags:%s%s\n" | |
dca6219e LP |
2340 | "Header size: %llu\n" |
2341 | "Arena size: %llu\n" | |
2342 | "Data Hash Table Size: %llu\n" | |
2343 | "Field Hash Table Size: %llu\n" | |
2344 | "Objects: %llu\n" | |
2345 | "Entry Objects: %llu\n" | |
2346 | "Rotate Suggested: %s\n" | |
2347 | "Head Sequential Number: %llu\n" | |
2348 | "Tail Sequential Number: %llu\n" | |
2349 | "Head Realtime Timestamp: %s\n" | |
2350 | "Tail Realtime Timestamp: %s\n", | |
2351 | f->path, | |
2352 | sd_id128_to_string(f->header->file_id, a), | |
2353 | sd_id128_to_string(f->header->machine_id, b), | |
2354 | sd_id128_to_string(f->header->boot_id, c), | |
2355 | sd_id128_to_string(f->header->seqnum_id, c), | |
dc36ac67 LP |
2356 | f->header->state == STATE_OFFLINE ? "offline" : |
2357 | f->header->state == STATE_ONLINE ? "online" : | |
2358 | f->header->state == STATE_ARCHIVED ? "archived" : "unknown", | |
7560fffc LP |
2359 | (f->header->compatible_flags & HEADER_COMPATIBLE_AUTHENTICATED) ? " AUTHENTICATED" : "", |
2360 | (f->header->compatible_flags & ~HEADER_COMPATIBLE_AUTHENTICATED) ? " ???" : "", | |
dc36ac67 LP |
2361 | (f->header->incompatible_flags & HEADER_INCOMPATIBLE_COMPRESSED) ? " COMPRESSED" : "", |
2362 | (f->header->incompatible_flags & ~HEADER_INCOMPATIBLE_COMPRESSED) ? " ???" : "", | |
dca6219e LP |
2363 | (unsigned long long) le64toh(f->header->header_size), |
2364 | (unsigned long long) le64toh(f->header->arena_size), | |
2365 | (unsigned long long) le64toh(f->header->data_hash_table_size) / sizeof(HashItem), | |
2366 | (unsigned long long) le64toh(f->header->field_hash_table_size) / sizeof(HashItem), | |
2367 | (unsigned long long) le64toh(f->header->n_objects), | |
2368 | (unsigned long long) le64toh(f->header->n_entries), | |
2369 | yes_no(journal_file_rotate_suggested(f)), | |
2370 | (unsigned long long) le64toh(f->header->head_seqnum), | |
2371 | (unsigned long long) le64toh(f->header->tail_seqnum), | |
2372 | format_timestamp(x, sizeof(x), le64toh(f->header->head_entry_realtime)), | |
2373 | format_timestamp(y, sizeof(y), le64toh(f->header->tail_entry_realtime))); | |
2374 | ||
2375 | if (JOURNAL_HEADER_CONTAINS(f->header, n_data)) | |
2376 | printf("Data Objects: %llu\n" | |
2377 | "Data Hash Table Fill: %.1f%%\n", | |
2378 | (unsigned long long) le64toh(f->header->n_data), | |
2379 | 100.0 * (double) le64toh(f->header->n_data) / ((double) (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)))); | |
2380 | ||
2381 | if (JOURNAL_HEADER_CONTAINS(f->header, n_fields)) | |
2382 | printf("Field Objects: %llu\n" | |
2383 | "Field Hash Table Fill: %.1f%%\n", | |
2384 | (unsigned long long) le64toh(f->header->n_fields), | |
2385 | 100.0 * (double) le64toh(f->header->n_fields) / ((double) (le64toh(f->header->field_hash_table_size) / sizeof(HashItem)))); | |
2386 | } | |
2387 | ||
cec736d2 LP |
2388 | int journal_file_open( |
2389 | const char *fname, | |
2390 | int flags, | |
2391 | mode_t mode, | |
7560fffc LP |
2392 | bool compress, |
2393 | bool authenticate, | |
4a92baf3 | 2394 | JournalMetrics *metrics, |
0ac38b70 | 2395 | JournalFile *template, |
cec736d2 LP |
2396 | JournalFile **ret) { |
2397 | ||
2398 | JournalFile *f; | |
2399 | int r; | |
2400 | bool newly_created = false; | |
2401 | ||
2402 | assert(fname); | |
2403 | ||
2404 | if ((flags & O_ACCMODE) != O_RDONLY && | |
2405 | (flags & O_ACCMODE) != O_RDWR) | |
2406 | return -EINVAL; | |
2407 | ||
9447a7f1 LP |
2408 | if (!endswith(fname, ".journal")) |
2409 | return -EINVAL; | |
2410 | ||
cec736d2 LP |
2411 | f = new0(JournalFile, 1); |
2412 | if (!f) | |
2413 | return -ENOMEM; | |
2414 | ||
0ac38b70 | 2415 | f->fd = -1; |
0ac38b70 | 2416 | f->mode = mode; |
cec736d2 | 2417 | |
7560fffc LP |
2418 | f->flags = flags; |
2419 | f->prot = prot_from_flags(flags); | |
2420 | f->writable = (flags & O_ACCMODE) != O_RDONLY; | |
2421 | f->compress = compress; | |
2422 | f->authenticate = authenticate; | |
15944db8 | 2423 | |
cec736d2 LP |
2424 | f->path = strdup(fname); |
2425 | if (!f->path) { | |
2426 | r = -ENOMEM; | |
2427 | goto fail; | |
2428 | } | |
2429 | ||
0ac38b70 LP |
2430 | f->fd = open(f->path, f->flags|O_CLOEXEC, f->mode); |
2431 | if (f->fd < 0) { | |
2432 | r = -errno; | |
2433 | goto fail; | |
2434 | } | |
2435 | ||
cec736d2 LP |
2436 | if (fstat(f->fd, &f->last_stat) < 0) { |
2437 | r = -errno; | |
2438 | goto fail; | |
2439 | } | |
2440 | ||
2441 | if (f->last_stat.st_size == 0 && f->writable) { | |
2442 | newly_created = true; | |
2443 | ||
7560fffc LP |
2444 | /* Try to load the FSPRG state, and if we can't, then |
2445 | * just don't do authentication */ | |
2446 | r = journal_file_load_fsprg(f); | |
2447 | if (r < 0) | |
2448 | f->authenticate = false; | |
2449 | ||
0ac38b70 | 2450 | r = journal_file_init_header(f, template); |
cec736d2 LP |
2451 | if (r < 0) |
2452 | goto fail; | |
2453 | ||
2454 | if (fstat(f->fd, &f->last_stat) < 0) { | |
2455 | r = -errno; | |
2456 | goto fail; | |
2457 | } | |
2458 | } | |
2459 | ||
dca6219e | 2460 | if (f->last_stat.st_size < (off_t) HEADER_SIZE_MIN) { |
cec736d2 LP |
2461 | r = -EIO; |
2462 | goto fail; | |
2463 | } | |
2464 | ||
2465 | f->header = mmap(NULL, PAGE_ALIGN(sizeof(Header)), prot_from_flags(flags), MAP_SHARED, f->fd, 0); | |
2466 | if (f->header == MAP_FAILED) { | |
2467 | f->header = NULL; | |
2468 | r = -errno; | |
2469 | goto fail; | |
2470 | } | |
2471 | ||
2472 | if (!newly_created) { | |
2473 | r = journal_file_verify_header(f); | |
2474 | if (r < 0) | |
2475 | goto fail; | |
7560fffc LP |
2476 | |
2477 | r = journal_file_load_fsprg(f); | |
2478 | if (r < 0) | |
2479 | goto fail; | |
cec736d2 LP |
2480 | } |
2481 | ||
2482 | if (f->writable) { | |
4a92baf3 LP |
2483 | if (metrics) { |
2484 | journal_default_metrics(metrics, f->fd); | |
2485 | f->metrics = *metrics; | |
2486 | } else if (template) | |
2487 | f->metrics = template->metrics; | |
2488 | ||
cec736d2 LP |
2489 | r = journal_file_refresh_header(f); |
2490 | if (r < 0) | |
2491 | goto fail; | |
7560fffc LP |
2492 | |
2493 | r = journal_file_setup_hmac(f); | |
2494 | if (r < 0) | |
2495 | goto fail; | |
cec736d2 LP |
2496 | } |
2497 | ||
2498 | if (newly_created) { | |
de190aef | 2499 | r = journal_file_setup_field_hash_table(f); |
cec736d2 LP |
2500 | if (r < 0) |
2501 | goto fail; | |
2502 | ||
de190aef | 2503 | r = journal_file_setup_data_hash_table(f); |
cec736d2 LP |
2504 | if (r < 0) |
2505 | goto fail; | |
7560fffc LP |
2506 | |
2507 | r = journal_file_append_first_tag(f); | |
2508 | if (r < 0) | |
2509 | goto fail; | |
cec736d2 LP |
2510 | } |
2511 | ||
de190aef | 2512 | r = journal_file_map_field_hash_table(f); |
cec736d2 LP |
2513 | if (r < 0) |
2514 | goto fail; | |
2515 | ||
de190aef | 2516 | r = journal_file_map_data_hash_table(f); |
cec736d2 LP |
2517 | if (r < 0) |
2518 | goto fail; | |
2519 | ||
2520 | if (ret) | |
2521 | *ret = f; | |
2522 | ||
2523 | return 0; | |
2524 | ||
2525 | fail: | |
2526 | journal_file_close(f); | |
2527 | ||
2528 | return r; | |
2529 | } | |
0ac38b70 | 2530 | |
7560fffc | 2531 | int journal_file_rotate(JournalFile **f, bool compress, bool authenticate) { |
0ac38b70 LP |
2532 | char *p; |
2533 | size_t l; | |
2534 | JournalFile *old_file, *new_file = NULL; | |
2535 | int r; | |
2536 | ||
2537 | assert(f); | |
2538 | assert(*f); | |
2539 | ||
2540 | old_file = *f; | |
2541 | ||
2542 | if (!old_file->writable) | |
2543 | return -EINVAL; | |
2544 | ||
2545 | if (!endswith(old_file->path, ".journal")) | |
2546 | return -EINVAL; | |
2547 | ||
2548 | l = strlen(old_file->path); | |
2549 | ||
9447a7f1 | 2550 | p = new(char, l + 1 + 32 + 1 + 16 + 1 + 16 + 1); |
0ac38b70 LP |
2551 | if (!p) |
2552 | return -ENOMEM; | |
2553 | ||
2554 | memcpy(p, old_file->path, l - 8); | |
2555 | p[l-8] = '@'; | |
2556 | sd_id128_to_string(old_file->header->seqnum_id, p + l - 8 + 1); | |
2557 | snprintf(p + l - 8 + 1 + 32, 1 + 16 + 1 + 16 + 8 + 1, | |
2558 | "-%016llx-%016llx.journal", | |
dca6219e | 2559 | (unsigned long long) le64toh((*f)->header->tail_seqnum), |
0ac38b70 LP |
2560 | (unsigned long long) le64toh((*f)->header->tail_entry_realtime)); |
2561 | ||
2562 | r = rename(old_file->path, p); | |
2563 | free(p); | |
2564 | ||
2565 | if (r < 0) | |
2566 | return -errno; | |
2567 | ||
ccdbaf91 | 2568 | old_file->header->state = STATE_ARCHIVED; |
0ac38b70 | 2569 | |
7560fffc | 2570 | r = journal_file_open(old_file->path, old_file->flags, old_file->mode, compress, authenticate, NULL, old_file, &new_file); |
0ac38b70 LP |
2571 | journal_file_close(old_file); |
2572 | ||
2573 | *f = new_file; | |
2574 | return r; | |
2575 | } | |
2576 | ||
9447a7f1 LP |
2577 | int journal_file_open_reliably( |
2578 | const char *fname, | |
2579 | int flags, | |
2580 | mode_t mode, | |
7560fffc LP |
2581 | bool compress, |
2582 | bool authenticate, | |
4a92baf3 | 2583 | JournalMetrics *metrics, |
9447a7f1 LP |
2584 | JournalFile *template, |
2585 | JournalFile **ret) { | |
2586 | ||
2587 | int r; | |
2588 | size_t l; | |
2589 | char *p; | |
2590 | ||
7560fffc | 2591 | r = journal_file_open(fname, flags, mode, compress, authenticate, metrics, template, ret); |
0071d9f1 LP |
2592 | if (r != -EBADMSG && /* corrupted */ |
2593 | r != -ENODATA && /* truncated */ | |
2594 | r != -EHOSTDOWN && /* other machine */ | |
a1a1898f LP |
2595 | r != -EPROTONOSUPPORT && /* incompatible feature */ |
2596 | r != -EBUSY && /* unclean shutdown */ | |
2597 | r != -ESHUTDOWN /* already archived */) | |
9447a7f1 LP |
2598 | return r; |
2599 | ||
2600 | if ((flags & O_ACCMODE) == O_RDONLY) | |
2601 | return r; | |
2602 | ||
2603 | if (!(flags & O_CREAT)) | |
2604 | return r; | |
2605 | ||
7560fffc LP |
2606 | if (!endswith(fname, ".journal")) |
2607 | return r; | |
2608 | ||
5c70eab4 LP |
2609 | /* The file is corrupted. Rotate it away and try it again (but only once) */ |
2610 | ||
9447a7f1 LP |
2611 | l = strlen(fname); |
2612 | if (asprintf(&p, "%.*s@%016llx-%016llx.journal~", | |
2613 | (int) (l-8), fname, | |
2614 | (unsigned long long) now(CLOCK_REALTIME), | |
2615 | random_ull()) < 0) | |
2616 | return -ENOMEM; | |
2617 | ||
2618 | r = rename(fname, p); | |
2619 | free(p); | |
2620 | if (r < 0) | |
2621 | return -errno; | |
2622 | ||
a1a1898f | 2623 | log_warning("File %s corrupted or uncleanly shut down, renaming and replacing.", fname); |
9447a7f1 | 2624 | |
7560fffc | 2625 | return journal_file_open(fname, flags, mode, compress, authenticate, metrics, template, ret); |
9447a7f1 LP |
2626 | } |
2627 | ||
0ac38b70 LP |
2628 | struct vacuum_info { |
2629 | off_t usage; | |
2630 | char *filename; | |
2631 | ||
2632 | uint64_t realtime; | |
2633 | sd_id128_t seqnum_id; | |
2634 | uint64_t seqnum; | |
5c70eab4 LP |
2635 | |
2636 | bool have_seqnum; | |
0ac38b70 LP |
2637 | }; |
2638 | ||
2639 | static int vacuum_compare(const void *_a, const void *_b) { | |
2640 | const struct vacuum_info *a, *b; | |
2641 | ||
2642 | a = _a; | |
2643 | b = _b; | |
2644 | ||
5c70eab4 LP |
2645 | if (a->have_seqnum && b->have_seqnum && |
2646 | sd_id128_equal(a->seqnum_id, b->seqnum_id)) { | |
0ac38b70 LP |
2647 | if (a->seqnum < b->seqnum) |
2648 | return -1; | |
2649 | else if (a->seqnum > b->seqnum) | |
2650 | return 1; | |
2651 | else | |
2652 | return 0; | |
2653 | } | |
2654 | ||
2655 | if (a->realtime < b->realtime) | |
2656 | return -1; | |
2657 | else if (a->realtime > b->realtime) | |
2658 | return 1; | |
5c70eab4 | 2659 | else if (a->have_seqnum && b->have_seqnum) |
0ac38b70 | 2660 | return memcmp(&a->seqnum_id, &b->seqnum_id, 16); |
5c70eab4 LP |
2661 | else |
2662 | return strcmp(a->filename, b->filename); | |
0ac38b70 LP |
2663 | } |
2664 | ||
2665 | int journal_directory_vacuum(const char *directory, uint64_t max_use, uint64_t min_free) { | |
2666 | DIR *d; | |
2667 | int r = 0; | |
2668 | struct vacuum_info *list = NULL; | |
2669 | unsigned n_list = 0, n_allocated = 0, i; | |
2670 | uint64_t sum = 0; | |
2671 | ||
2672 | assert(directory); | |
2673 | ||
2674 | if (max_use <= 0) | |
babfc091 | 2675 | return 0; |
0ac38b70 LP |
2676 | |
2677 | d = opendir(directory); | |
2678 | if (!d) | |
2679 | return -errno; | |
2680 | ||
2681 | for (;;) { | |
2682 | int k; | |
2683 | struct dirent buf, *de; | |
2684 | size_t q; | |
2685 | struct stat st; | |
2686 | char *p; | |
7ea07dcd | 2687 | unsigned long long seqnum = 0, realtime; |
0ac38b70 | 2688 | sd_id128_t seqnum_id; |
5c70eab4 | 2689 | bool have_seqnum; |
0ac38b70 LP |
2690 | |
2691 | k = readdir_r(d, &buf, &de); | |
2692 | if (k != 0) { | |
2693 | r = -k; | |
2694 | goto finish; | |
2695 | } | |
2696 | ||
2697 | if (!de) | |
2698 | break; | |
2699 | ||
5c70eab4 LP |
2700 | if (fstatat(dirfd(d), de->d_name, &st, AT_SYMLINK_NOFOLLOW) < 0) |
2701 | continue; | |
2702 | ||
2703 | if (!S_ISREG(st.st_mode)) | |
0ac38b70 LP |
2704 | continue; |
2705 | ||
2706 | q = strlen(de->d_name); | |
2707 | ||
5c70eab4 | 2708 | if (endswith(de->d_name, ".journal")) { |
0ac38b70 | 2709 | |
5c70eab4 | 2710 | /* Vacuum archived files */ |
0ac38b70 | 2711 | |
5c70eab4 LP |
2712 | if (q < 1 + 32 + 1 + 16 + 1 + 16 + 8) |
2713 | continue; | |
0ac38b70 | 2714 | |
5c70eab4 LP |
2715 | if (de->d_name[q-8-16-1] != '-' || |
2716 | de->d_name[q-8-16-1-16-1] != '-' || | |
2717 | de->d_name[q-8-16-1-16-1-32-1] != '@') | |
2718 | continue; | |
0ac38b70 | 2719 | |
5c70eab4 LP |
2720 | p = strdup(de->d_name); |
2721 | if (!p) { | |
2722 | r = -ENOMEM; | |
2723 | goto finish; | |
2724 | } | |
0ac38b70 | 2725 | |
5c70eab4 LP |
2726 | de->d_name[q-8-16-1-16-1] = 0; |
2727 | if (sd_id128_from_string(de->d_name + q-8-16-1-16-1-32, &seqnum_id) < 0) { | |
2728 | free(p); | |
2729 | continue; | |
2730 | } | |
2731 | ||
2732 | if (sscanf(de->d_name + q-8-16-1-16, "%16llx-%16llx.journal", &seqnum, &realtime) != 2) { | |
2733 | free(p); | |
2734 | continue; | |
2735 | } | |
2736 | ||
2737 | have_seqnum = true; | |
2738 | ||
2739 | } else if (endswith(de->d_name, ".journal~")) { | |
2740 | unsigned long long tmp; | |
2741 | ||
2742 | /* Vacuum corrupted files */ | |
2743 | ||
2744 | if (q < 1 + 16 + 1 + 16 + 8 + 1) | |
2745 | continue; | |
0ac38b70 | 2746 | |
5c70eab4 LP |
2747 | if (de->d_name[q-1-8-16-1] != '-' || |
2748 | de->d_name[q-1-8-16-1-16-1] != '@') | |
2749 | continue; | |
2750 | ||
2751 | p = strdup(de->d_name); | |
2752 | if (!p) { | |
2753 | r = -ENOMEM; | |
2754 | goto finish; | |
2755 | } | |
2756 | ||
2757 | if (sscanf(de->d_name + q-1-8-16-1-16, "%16llx-%16llx.journal~", &realtime, &tmp) != 2) { | |
2758 | free(p); | |
2759 | continue; | |
2760 | } | |
2761 | ||
2762 | have_seqnum = false; | |
2763 | } else | |
0ac38b70 | 2764 | continue; |
0ac38b70 LP |
2765 | |
2766 | if (n_list >= n_allocated) { | |
2767 | struct vacuum_info *j; | |
2768 | ||
2769 | n_allocated = MAX(n_allocated * 2U, 8U); | |
2770 | j = realloc(list, n_allocated * sizeof(struct vacuum_info)); | |
2771 | if (!j) { | |
2772 | free(p); | |
2773 | r = -ENOMEM; | |
2774 | goto finish; | |
2775 | } | |
2776 | ||
2777 | list = j; | |
2778 | } | |
2779 | ||
2780 | list[n_list].filename = p; | |
a3a52c0f | 2781 | list[n_list].usage = 512UL * (uint64_t) st.st_blocks; |
0ac38b70 LP |
2782 | list[n_list].seqnum = seqnum; |
2783 | list[n_list].realtime = realtime; | |
2784 | list[n_list].seqnum_id = seqnum_id; | |
5c70eab4 | 2785 | list[n_list].have_seqnum = have_seqnum; |
0ac38b70 LP |
2786 | |
2787 | sum += list[n_list].usage; | |
2788 | ||
2789 | n_list ++; | |
2790 | } | |
2791 | ||
64825d3c LP |
2792 | if (n_list > 0) |
2793 | qsort(list, n_list, sizeof(struct vacuum_info), vacuum_compare); | |
0ac38b70 LP |
2794 | |
2795 | for(i = 0; i < n_list; i++) { | |
2796 | struct statvfs ss; | |
2797 | ||
2798 | if (fstatvfs(dirfd(d), &ss) < 0) { | |
2799 | r = -errno; | |
2800 | goto finish; | |
2801 | } | |
2802 | ||
2803 | if (sum <= max_use && | |
2804 | (uint64_t) ss.f_bavail * (uint64_t) ss.f_bsize >= min_free) | |
2805 | break; | |
2806 | ||
2807 | if (unlinkat(dirfd(d), list[i].filename, 0) >= 0) { | |
e7bf07b3 | 2808 | log_info("Deleted archived journal %s/%s.", directory, list[i].filename); |
0ac38b70 LP |
2809 | sum -= list[i].usage; |
2810 | } else if (errno != ENOENT) | |
2811 | log_warning("Failed to delete %s/%s: %m", directory, list[i].filename); | |
2812 | } | |
2813 | ||
2814 | finish: | |
2815 | for (i = 0; i < n_list; i++) | |
2816 | free(list[i].filename); | |
2817 | ||
2818 | free(list); | |
2819 | ||
de190aef LP |
2820 | if (d) |
2821 | closedir(d); | |
2822 | ||
0ac38b70 LP |
2823 | return r; |
2824 | } | |
cf244689 LP |
2825 | |
2826 | int journal_file_copy_entry(JournalFile *from, JournalFile *to, Object *o, uint64_t p, uint64_t *seqnum, Object **ret, uint64_t *offset) { | |
2827 | uint64_t i, n; | |
2828 | uint64_t q, xor_hash = 0; | |
2829 | int r; | |
2830 | EntryItem *items; | |
2831 | dual_timestamp ts; | |
2832 | ||
2833 | assert(from); | |
2834 | assert(to); | |
2835 | assert(o); | |
2836 | assert(p); | |
2837 | ||
2838 | if (!to->writable) | |
2839 | return -EPERM; | |
2840 | ||
2841 | ts.monotonic = le64toh(o->entry.monotonic); | |
2842 | ts.realtime = le64toh(o->entry.realtime); | |
2843 | ||
2844 | if (to->tail_entry_monotonic_valid && | |
2845 | ts.monotonic < le64toh(to->header->tail_entry_monotonic)) | |
2846 | return -EINVAL; | |
2847 | ||
cf244689 LP |
2848 | n = journal_file_entry_n_items(o); |
2849 | items = alloca(sizeof(EntryItem) * n); | |
2850 | ||
2851 | for (i = 0; i < n; i++) { | |
4fd052ae FC |
2852 | uint64_t l, h; |
2853 | le64_t le_hash; | |
cf244689 LP |
2854 | size_t t; |
2855 | void *data; | |
2856 | Object *u; | |
2857 | ||
2858 | q = le64toh(o->entry.items[i].object_offset); | |
2859 | le_hash = o->entry.items[i].hash; | |
2860 | ||
2861 | r = journal_file_move_to_object(from, OBJECT_DATA, q, &o); | |
2862 | if (r < 0) | |
2863 | return r; | |
2864 | ||
2865 | if (le_hash != o->data.hash) | |
2866 | return -EBADMSG; | |
2867 | ||
2868 | l = le64toh(o->object.size) - offsetof(Object, data.payload); | |
2869 | t = (size_t) l; | |
2870 | ||
2871 | /* We hit the limit on 32bit machines */ | |
2872 | if ((uint64_t) t != l) | |
2873 | return -E2BIG; | |
2874 | ||
2875 | if (o->object.flags & OBJECT_COMPRESSED) { | |
2876 | #ifdef HAVE_XZ | |
2877 | uint64_t rsize; | |
2878 | ||
2879 | if (!uncompress_blob(o->data.payload, l, &from->compress_buffer, &from->compress_buffer_size, &rsize)) | |
2880 | return -EBADMSG; | |
2881 | ||
2882 | data = from->compress_buffer; | |
2883 | l = rsize; | |
2884 | #else | |
2885 | return -EPROTONOSUPPORT; | |
2886 | #endif | |
2887 | } else | |
2888 | data = o->data.payload; | |
2889 | ||
2890 | r = journal_file_append_data(to, data, l, &u, &h); | |
2891 | if (r < 0) | |
2892 | return r; | |
2893 | ||
2894 | xor_hash ^= le64toh(u->data.hash); | |
2895 | items[i].object_offset = htole64(h); | |
2896 | items[i].hash = u->data.hash; | |
2897 | ||
2898 | r = journal_file_move_to_object(from, OBJECT_ENTRY, p, &o); | |
2899 | if (r < 0) | |
2900 | return r; | |
2901 | } | |
2902 | ||
2903 | return journal_file_append_entry_internal(to, &ts, xor_hash, items, n, seqnum, ret, offset); | |
2904 | } | |
babfc091 LP |
2905 | |
2906 | void journal_default_metrics(JournalMetrics *m, int fd) { | |
2907 | uint64_t fs_size = 0; | |
2908 | struct statvfs ss; | |
a7bc2c2a | 2909 | char a[FORMAT_BYTES_MAX], b[FORMAT_BYTES_MAX], c[FORMAT_BYTES_MAX], d[FORMAT_BYTES_MAX]; |
babfc091 LP |
2910 | |
2911 | assert(m); | |
2912 | assert(fd >= 0); | |
2913 | ||
2914 | if (fstatvfs(fd, &ss) >= 0) | |
2915 | fs_size = ss.f_frsize * ss.f_blocks; | |
2916 | ||
2917 | if (m->max_use == (uint64_t) -1) { | |
2918 | ||
2919 | if (fs_size > 0) { | |
2920 | m->max_use = PAGE_ALIGN(fs_size / 10); /* 10% of file system size */ | |
2921 | ||
2922 | if (m->max_use > DEFAULT_MAX_USE_UPPER) | |
2923 | m->max_use = DEFAULT_MAX_USE_UPPER; | |
2924 | ||
2925 | if (m->max_use < DEFAULT_MAX_USE_LOWER) | |
2926 | m->max_use = DEFAULT_MAX_USE_LOWER; | |
2927 | } else | |
2928 | m->max_use = DEFAULT_MAX_USE_LOWER; | |
2929 | } else { | |
2930 | m->max_use = PAGE_ALIGN(m->max_use); | |
2931 | ||
2932 | if (m->max_use < JOURNAL_FILE_SIZE_MIN*2) | |
2933 | m->max_use = JOURNAL_FILE_SIZE_MIN*2; | |
2934 | } | |
2935 | ||
2936 | if (m->max_size == (uint64_t) -1) { | |
2937 | m->max_size = PAGE_ALIGN(m->max_use / 8); /* 8 chunks */ | |
2938 | ||
2939 | if (m->max_size > DEFAULT_MAX_SIZE_UPPER) | |
2940 | m->max_size = DEFAULT_MAX_SIZE_UPPER; | |
2941 | } else | |
2942 | m->max_size = PAGE_ALIGN(m->max_size); | |
2943 | ||
2944 | if (m->max_size < JOURNAL_FILE_SIZE_MIN) | |
2945 | m->max_size = JOURNAL_FILE_SIZE_MIN; | |
2946 | ||
2947 | if (m->max_size*2 > m->max_use) | |
2948 | m->max_use = m->max_size*2; | |
2949 | ||
2950 | if (m->min_size == (uint64_t) -1) | |
2951 | m->min_size = JOURNAL_FILE_SIZE_MIN; | |
2952 | else { | |
2953 | m->min_size = PAGE_ALIGN(m->min_size); | |
2954 | ||
2955 | if (m->min_size < JOURNAL_FILE_SIZE_MIN) | |
2956 | m->min_size = JOURNAL_FILE_SIZE_MIN; | |
2957 | ||
2958 | if (m->min_size > m->max_size) | |
2959 | m->max_size = m->min_size; | |
2960 | } | |
2961 | ||
2962 | if (m->keep_free == (uint64_t) -1) { | |
2963 | ||
2964 | if (fs_size > 0) { | |
2965 | m->keep_free = PAGE_ALIGN(fs_size / 20); /* 5% of file system size */ | |
2966 | ||
2967 | if (m->keep_free > DEFAULT_KEEP_FREE_UPPER) | |
2968 | m->keep_free = DEFAULT_KEEP_FREE_UPPER; | |
2969 | ||
2970 | } else | |
2971 | m->keep_free = DEFAULT_KEEP_FREE; | |
2972 | } | |
2973 | ||
e7bf07b3 LP |
2974 | log_info("Fixed max_use=%s max_size=%s min_size=%s keep_free=%s", |
2975 | format_bytes(a, sizeof(a), m->max_use), | |
2976 | format_bytes(b, sizeof(b), m->max_size), | |
2977 | format_bytes(c, sizeof(c), m->min_size), | |
2978 | format_bytes(d, sizeof(d), m->keep_free)); | |
babfc091 | 2979 | } |
08984293 LP |
2980 | |
2981 | int journal_file_get_cutoff_realtime_usec(JournalFile *f, usec_t *from, usec_t *to) { | |
08984293 LP |
2982 | assert(f); |
2983 | assert(from || to); | |
2984 | ||
2985 | if (from) { | |
162566a4 LP |
2986 | if (f->header->head_entry_realtime == 0) |
2987 | return -ENOENT; | |
08984293 | 2988 | |
162566a4 | 2989 | *from = le64toh(f->header->head_entry_realtime); |
08984293 LP |
2990 | } |
2991 | ||
2992 | if (to) { | |
162566a4 LP |
2993 | if (f->header->tail_entry_realtime == 0) |
2994 | return -ENOENT; | |
08984293 | 2995 | |
162566a4 | 2996 | *to = le64toh(f->header->tail_entry_realtime); |
08984293 LP |
2997 | } |
2998 | ||
2999 | return 1; | |
3000 | } | |
3001 | ||
3002 | int journal_file_get_cutoff_monotonic_usec(JournalFile *f, sd_id128_t boot_id, usec_t *from, usec_t *to) { | |
3003 | char t[9+32+1] = "_BOOT_ID="; | |
3004 | Object *o; | |
3005 | uint64_t p; | |
3006 | int r; | |
3007 | ||
3008 | assert(f); | |
3009 | assert(from || to); | |
3010 | ||
3011 | sd_id128_to_string(boot_id, t + 9); | |
3012 | ||
3013 | r = journal_file_find_data_object(f, t, strlen(t), &o, &p); | |
3014 | if (r <= 0) | |
3015 | return r; | |
3016 | ||
3017 | if (le64toh(o->data.n_entries) <= 0) | |
3018 | return 0; | |
3019 | ||
3020 | if (from) { | |
3021 | r = journal_file_move_to_object(f, OBJECT_ENTRY, le64toh(o->data.entry_offset), &o); | |
3022 | if (r < 0) | |
3023 | return r; | |
3024 | ||
3025 | *from = le64toh(o->entry.monotonic); | |
3026 | } | |
3027 | ||
3028 | if (to) { | |
3029 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); | |
3030 | if (r < 0) | |
3031 | return r; | |
3032 | ||
3033 | r = generic_array_get_plus_one(f, | |
3034 | le64toh(o->data.entry_offset), | |
3035 | le64toh(o->data.entry_array_offset), | |
3036 | le64toh(o->data.n_entries)-1, | |
3037 | &o, NULL); | |
3038 | if (r <= 0) | |
3039 | return r; | |
3040 | ||
3041 | *to = le64toh(o->entry.monotonic); | |
3042 | } | |
3043 | ||
3044 | return 1; | |
3045 | } | |
dca6219e LP |
3046 | |
3047 | bool journal_file_rotate_suggested(JournalFile *f) { | |
3048 | assert(f); | |
3049 | ||
3050 | /* If we gained new header fields we gained new features, | |
3051 | * hence suggest a rotation */ | |
361f9cbc LP |
3052 | if (le64toh(f->header->header_size) < sizeof(Header)) { |
3053 | log_debug("%s uses an outdated header, suggesting rotation.", f->path); | |
dca6219e | 3054 | return true; |
361f9cbc | 3055 | } |
dca6219e LP |
3056 | |
3057 | /* Let's check if the hash tables grew over a certain fill | |
3058 | * level (75%, borrowing this value from Java's hash table | |
3059 | * implementation), and if so suggest a rotation. To calculate | |
3060 | * the fill level we need the n_data field, which only exists | |
3061 | * in newer versions. */ | |
3062 | ||
3063 | if (JOURNAL_HEADER_CONTAINS(f->header, n_data)) | |
361f9cbc LP |
3064 | if (le64toh(f->header->n_data) * 4ULL > (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)) * 3ULL) { |
3065 | log_debug("Data hash table of %s has a fill level at %.1f (%llu of %llu items, %llu file size, %llu bytes per hash table item), suggesting rotation.", | |
3066 | f->path, | |
3067 | 100.0 * (double) le64toh(f->header->n_data) / ((double) (le64toh(f->header->data_hash_table_size) / sizeof(HashItem))), | |
3068 | (unsigned long long) le64toh(f->header->n_data), | |
3069 | (unsigned long long) (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)), | |
3070 | (unsigned long long) (f->last_stat.st_size), | |
3071 | (unsigned long long) (f->last_stat.st_size / le64toh(f->header->n_data))); | |
dca6219e | 3072 | return true; |
361f9cbc | 3073 | } |
dca6219e LP |
3074 | |
3075 | if (JOURNAL_HEADER_CONTAINS(f->header, n_fields)) | |
361f9cbc LP |
3076 | if (le64toh(f->header->n_fields) * 4ULL > (le64toh(f->header->field_hash_table_size) / sizeof(HashItem)) * 3ULL) { |
3077 | log_debug("Field hash table of %s has a fill level at %.1f (%llu of %llu items), suggesting rotation.", | |
3078 | f->path, | |
3079 | 100.0 * (double) le64toh(f->header->n_fields) / ((double) (le64toh(f->header->field_hash_table_size) / sizeof(HashItem))), | |
3080 | (unsigned long long) le64toh(f->header->n_fields), | |
3081 | (unsigned long long) (le64toh(f->header->field_hash_table_size) / sizeof(HashItem))); | |
dca6219e | 3082 | return true; |
361f9cbc | 3083 | } |
dca6219e LP |
3084 | |
3085 | return false; | |
3086 | } |