]> git.ipfire.org Git - thirdparty/systemd.git/blame - src/journal/journal-file.c
projects / thirdparty / systemd.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
journald: increase rate limit burst rate
[thirdparty/systemd.git] / src / journal / journal-file.c
CommitLineData
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
9 under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 2 of the License, or
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
16 General Public License for more details.
17
18 You should have received a copy of the GNU General Public License
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"
cec736d2 34
de190aef
LP		 35#define DEFAULT_DATA_HASH_TABLE_SIZE (2047ULL*16ULL)
36#define DEFAULT_FIELD_HASH_TABLE_SIZE (2047ULL*16ULL)
cec736d2
LP		 37
38#define DEFAULT_WINDOW_SIZE (128ULL*1024ULL*1024ULL)
39
807e17f0
LP		 40#define COMPRESSION_SIZE_THRESHOLD (64ULL)
41
cec736d2
LP		 42static const char signature[] = { 'L', 'P', 'K', 'S', 'H', 'H', 'R', 'H' };
43
44#define ALIGN64(x) (((x) + 7ULL) & ~7ULL)
45
46void journal_file_close(JournalFile *f) {
de190aef 47 int t;
cec736d2 48
de190aef 49 assert(f);
cec736d2 50
de190aef
LP		 51 if (f->header && f->writable)
52 f->header->state = STATE_OFFLINE;
cec736d2 53
cec736d2 54
de190aef
LP		 55 for (t = 0; t < _WINDOW_MAX; t++)
56 if (f->windows[t].ptr)
57 munmap(f->windows[t].ptr, f->windows[t].size);
cec736d2 58
0ac38b70
LP		 59 if (f->fd >= 0)
60 close_nointr_nofail(f->fd);
61
cec736d2 62 free(f->path);
807e17f0
LP		 63
64#ifdef HAVE_XZ
65 free(f->compress_buffer);
66#endif
67
cec736d2
LP		 68 free(f);
69}
70
0ac38b70 71static int journal_file_init_header(JournalFile *f, JournalFile *template) {
cec736d2
LP		 72 Header h;
73 ssize_t k;
74 int r;
75
76 assert(f);
77
78 zero(h);
79 memcpy(h.signature, signature, 8);
80 h.arena_offset = htole64(ALIGN64(sizeof(h)));
cec736d2
LP		 81
82 r = sd_id128_randomize(&h.file_id);
83 if (r < 0)
84 return r;
85
0ac38b70
LP		 86 if (template) {
87 h.seqnum_id = template->header->seqnum_id;
88 h.seqnum = template->header->seqnum;
89 } else
90 h.seqnum_id = h.file_id;
cec736d2
LP		 91
92 k = pwrite(f->fd, &h, sizeof(h), 0);
93 if (k < 0)
94 return -errno;
95
96 if (k != sizeof(h))
97 return -EIO;
98
99 return 0;
100}
101
102static int journal_file_refresh_header(JournalFile *f) {
103 int r;
de190aef 104 sd_id128_t boot_id;
cec736d2
LP		 105
106 assert(f);
107
108 r = sd_id128_get_machine(&f->header->machine_id);
109 if (r < 0)
110 return r;
111
de190aef 112 r = sd_id128_get_boot(&boot_id);
cec736d2
LP		 113 if (r < 0)
114 return r;
115
de190aef
LP		 116 if (sd_id128_equal(boot_id, f->header->boot_id))
117 f->tail_entry_monotonic_valid = true;
118
119 f->header->boot_id = boot_id;
120
121 f->header->state = STATE_ONLINE;
cec736d2
LP		 122 return 0;
123}
124
125static int journal_file_verify_header(JournalFile *f) {
126 assert(f);
127
128 if (memcmp(f->header, signature, 8))
129 return -EBADMSG;
130
807e17f0
LP		 131#ifdef HAVE_XZ
132 if ((le64toh(f->header->incompatible_flags) & ~HEADER_INCOMPATIBLE_COMPRESSED) != 0)
133 return -EPROTONOSUPPORT;
134#else
cec736d2
LP		 135 if (f->header->incompatible_flags != 0)
136 return -EPROTONOSUPPORT;
807e17f0 137#endif
cec736d2
LP		 138
139 if ((uint64_t) f->last_stat.st_size < (le64toh(f->header->arena_offset) + le64toh(f->header->arena_size)))
140 return -ENODATA;
141
142 if (f->writable) {
143 uint32_t state;
144 sd_id128_t machine_id;
145 int r;
146
147 r = sd_id128_get_machine(&machine_id);
148 if (r < 0)
149 return r;
150
151 if (!sd_id128_equal(machine_id, f->header->machine_id))
152 return -EHOSTDOWN;
153
de190aef 154 state = f->header->state;
cec736d2
LP		 155
156 if (state == STATE_ONLINE)
157 log_debug("Journal file %s is already online. Assuming unclean closing. Ignoring.", f->path);
158 else if (state == STATE_ARCHIVED)
159 return -ESHUTDOWN;
160 else if (state != STATE_OFFLINE)
161 log_debug("Journal file %s has unknown state %u. Ignoring.", f->path, state);
162 }
163
164 return 0;
165}
166
167static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size) {
cec736d2
LP		 168 uint64_t old_size, new_size;
169
170 assert(f);
171
cec736d2 172 /* We assume that this file is not sparse, and we know that
38ac38b2 173 * for sure, since we always call posix_fallocate()
cec736d2
LP		 174 * ourselves */
175
176 old_size =
177 le64toh(f->header->arena_offset) +
178 le64toh(f->header->arena_size);
179
bc85bfee
LP		 180 new_size = PAGE_ALIGN(offset + size);
181 if (new_size < le64toh(f->header->arena_offset))
182 new_size = le64toh(f->header->arena_offset);
183
184 if (new_size <= old_size)
cec736d2
LP		 185 return 0;
186
bc85bfee
LP		 187 if (f->metrics.max_size > 0 &&
188 new_size > f->metrics.max_size)
189 return -E2BIG;
cec736d2 190
bc85bfee
LP		 191 if (new_size > f->metrics.min_size &&
192 f->metrics.keep_free > 0) {
cec736d2
LP		 193 struct statvfs svfs;
194
195 if (fstatvfs(f->fd, &svfs) >= 0) {
196 uint64_t available;
197
198 available = svfs.f_bfree * svfs.f_bsize;
199
bc85bfee
LP		 200 if (available >= f->metrics.keep_free)
201 available -= f->metrics.keep_free;
cec736d2
LP		 202 else
203 available = 0;
204
205 if (new_size - old_size > available)
206 return -E2BIG;
207 }
208 }
209
bc85bfee
LP		 210 /* Note that the glibc fallocate() fallback is very
211 inefficient, hence we try to minimize the allocation area
212 as we can. */
38ac38b2 213 if (posix_fallocate(f->fd, old_size, new_size - old_size) < 0)
cec736d2
LP		 214 return -errno;
215
216 if (fstat(f->fd, &f->last_stat) < 0)
217 return -errno;
218
bc85bfee 219 f->header->arena_size = new_size - htole64(f->header->arena_offset);
cec736d2
LP		 220
221 return 0;
222}
223
224static int journal_file_map(
225 JournalFile *f,
226 uint64_t offset,
227 uint64_t size,
228 void **_window,
229 uint64_t *_woffset,
230 uint64_t *_wsize,
231 void **ret) {
232
233 uint64_t woffset, wsize;
234 void *window;
235
236 assert(f);
237 assert(size > 0);
238 assert(ret);
239
240 woffset = offset & ~((uint64_t) page_size() - 1ULL);
241 wsize = size + (offset - woffset);
242 wsize = PAGE_ALIGN(wsize);
243
2a59ea54
LP		 244 /* Avoid SIGBUS on invalid accesses */
245 if (woffset + wsize > (uint64_t) PAGE_ALIGN(f->last_stat.st_size))
246 return -EADDRNOTAVAIL;
247
cec736d2
LP		 248 window = mmap(NULL, wsize, f->prot, MAP_SHARED, f->fd, woffset);
249 if (window == MAP_FAILED)
250 return -errno;
251
252 if (_window)
253 *_window = window;
254
255 if (_woffset)
256 *_woffset = woffset;
257
258 if (_wsize)
259 *_wsize = wsize;
260
261 *ret = (uint8_t*) window + (offset - woffset);
262
263 return 0;
264}
265
de190aef 266static int journal_file_move_to(JournalFile *f, int wt, uint64_t offset, uint64_t size, void **ret) {
cec736d2
LP		 267 void *p;
268 uint64_t delta;
269 int r;
de190aef 270 Window *w;
cec736d2
LP		 271
272 assert(f);
273 assert(ret);
de190aef
LP		 274 assert(wt >= 0);
275 assert(wt < _WINDOW_MAX);
cec736d2 276
de190aef 277 w = f->windows + wt;
cec736d2 278
de190aef
LP		 279 if (_likely_(w->ptr &&
280 w->offset <= offset &&
281 w->offset + w->size >= offset + size)) {
282
283 *ret = (uint8_t*) w->ptr + (offset - w->offset);
cec736d2
LP		 284 return 0;
285 }
286
de190aef
LP		 287 if (w->ptr) {
288 if (munmap(w->ptr, w->size) < 0)
cec736d2
LP		 289 return -errno;
290
de190aef
LP		 291 w->ptr = NULL;
292 w->size = w->offset = 0;
cec736d2
LP		 293 }
294
295 if (size < DEFAULT_WINDOW_SIZE) {
296 /* If the default window size is larger then what was
297 * asked for extend the mapping a bit in the hope to
298 * minimize needed remappings later on. We add half
299 * the window space before and half behind the
300 * requested mapping */
301
302 delta = PAGE_ALIGN((DEFAULT_WINDOW_SIZE - size) / 2);
303
304 if (offset < delta)
305 delta = offset;
306
307 offset -= delta;
308 size += (DEFAULT_WINDOW_SIZE - delta);
309 } else
310 delta = 0;
311
2a59ea54
LP		 312 if (offset > (uint64_t) f->last_stat.st_size)
313 return -EADDRNOTAVAIL;
314
315 if (offset + size > (uint64_t) f->last_stat.st_size)
316 size = PAGE_ALIGN((uint64_t) f->last_stat.st_size - offset);
317
318 if (size <= 0)
319 return -EADDRNOTAVAIL;
320
cec736d2
LP		 321 r = journal_file_map(f,
322 offset, size,
de190aef
LP		 323 &w->ptr, &w->offset, &w->size,
324 &p);
cec736d2
LP		 325
326 if (r < 0)
327 return r;
328
329 *ret = (uint8_t*) p + delta;
330 return 0;
331}
332
333static bool verify_hash(Object *o) {
de190aef 334 uint64_t h1, h2;
cec736d2
LP		 335
336 assert(o);
337
807e17f0 338 if (o->object.type == OBJECT_DATA && !(o->object.flags & OBJECT_COMPRESSED)) {
cec736d2 339 h1 = le64toh(o->data.hash);
de190aef
LP		 340 h2 = hash64(o->data.payload, le64toh(o->object.size) - offsetof(Object, data.payload));
341 } else if (o->object.type == OBJECT_FIELD) {
342 h1 = le64toh(o->field.hash);
343 h2 = hash64(o->field.payload, le64toh(o->object.size) - offsetof(Object, field.payload));
344 } else
345 return true;
cec736d2 346
de190aef 347 return h1 == h2;
cec736d2
LP		 348}
349
de190aef 350int journal_file_move_to_object(JournalFile *f, int type, uint64_t offset, Object **ret) {
cec736d2
LP		 351 int r;
352 void *t;
353 Object *o;
354 uint64_t s;
355
356 assert(f);
357 assert(ret);
de190aef 358 assert(type < _OBJECT_TYPE_MAX);
cec736d2 359
de190aef 360 r = journal_file_move_to(f, type >= 0 ? type : WINDOW_UNKNOWN, offset, sizeof(ObjectHeader), &t);
cec736d2
LP		 361 if (r < 0)
362 return r;
363
364 o = (Object*) t;
365 s = le64toh(o->object.size);
366
367 if (s < sizeof(ObjectHeader))
368 return -EBADMSG;
369
de190aef 370 if (type >= 0 && o->object.type != type)
cec736d2
LP		 371 return -EBADMSG;
372
373 if (s > sizeof(ObjectHeader)) {
de190aef 374 r = journal_file_move_to(f, o->object.type, offset, s, &t);
cec736d2
LP		 375 if (r < 0)
376 return r;
377
378 o = (Object*) t;
379 }
380
381 if (!verify_hash(o))
382 return -EBADMSG;
383
384 *ret = o;
385 return 0;
386}
387
c2373f84 388static uint64_t journal_file_seqnum(JournalFile *f, uint64_t *seqnum) {
cec736d2
LP		 389 uint64_t r;
390
391 assert(f);
392
393 r = le64toh(f->header->seqnum) + 1;
c2373f84
LP		 394
395 if (seqnum) {
de190aef 396 /* If an external seqnum counter was passed, we update
c2373f84
LP		 397 * both the local and the external one, and set it to
398 * the maximum of both */
399
400 if (*seqnum + 1 > r)
401 r = *seqnum + 1;
402
403 *seqnum = r;
404 }
405
cec736d2
LP		 406 f->header->seqnum = htole64(r);
407
de190aef
LP		 408 if (f->header->first_seqnum == 0)
409 f->header->first_seqnum = htole64(r);
410
cec736d2
LP		 411 return r;
412}
413
de190aef 414static int journal_file_append_object(JournalFile *f, int type, uint64_t size, Object **ret, uint64_t *offset) {
cec736d2
LP		 415 int r;
416 uint64_t p;
417 Object *tail, *o;
418 void *t;
419
420 assert(f);
421 assert(size >= sizeof(ObjectHeader));
422 assert(offset);
423 assert(ret);
424
425 p = le64toh(f->header->tail_object_offset);
cec736d2
LP		 426 if (p == 0)
427 p = le64toh(f->header->arena_offset);
428 else {
de190aef 429 r = journal_file_move_to_object(f, -1, p, &tail);
cec736d2
LP		 430 if (r < 0)
431 return r;
432
433 p += ALIGN64(le64toh(tail->object.size));
434 }
435
436 r = journal_file_allocate(f, p, size);
437 if (r < 0)
438 return r;
439
de190aef 440 r = journal_file_move_to(f, type, p, size, &t);
cec736d2
LP		 441 if (r < 0)
442 return r;
443
444 o = (Object*) t;
445
446 zero(o->object);
de190aef 447 o->object.type = type;
cec736d2
LP		 448 o->object.size = htole64(size);
449
450 f->header->tail_object_offset = htole64(p);
cec736d2
LP		 451 f->header->n_objects = htole64(le64toh(f->header->n_objects) + 1);
452
453 *ret = o;
454 *offset = p;
455
456 return 0;
457}
458
de190aef 459static int journal_file_setup_data_hash_table(JournalFile *f) {
cec736d2
LP		 460 uint64_t s, p;
461 Object *o;
462 int r;
463
464 assert(f);
465
de190aef
LP		 466 s = DEFAULT_DATA_HASH_TABLE_SIZE;
467 r = journal_file_append_object(f,
468 OBJECT_DATA_HASH_TABLE,
469 offsetof(Object, hash_table.items) + s,
470 &o, &p);
cec736d2
LP		 471 if (r < 0)
472 return r;
473
de190aef 474 memset(o->hash_table.items, 0, s);
cec736d2 475
de190aef
LP		 476 f->header->data_hash_table_offset = htole64(p + offsetof(Object, hash_table.items));
477 f->header->data_hash_table_size = htole64(s);
cec736d2
LP		 478
479 return 0;
480}
481
de190aef 482static int journal_file_setup_field_hash_table(JournalFile *f) {
cec736d2
LP		 483 uint64_t s, p;
484 Object *o;
485 int r;
486
487 assert(f);
488
de190aef
LP		 489 s = DEFAULT_FIELD_HASH_TABLE_SIZE;
490 r = journal_file_append_object(f,
491 OBJECT_FIELD_HASH_TABLE,
492 offsetof(Object, hash_table.items) + s,
493 &o, &p);
cec736d2
LP		 494 if (r < 0)
495 return r;
496
de190aef 497 memset(o->hash_table.items, 0, s);
cec736d2 498
de190aef
LP		 499 f->header->field_hash_table_offset = htole64(p + offsetof(Object, hash_table.items));
500 f->header->field_hash_table_size = htole64(s);
cec736d2
LP		 501
502 return 0;
503}
504
de190aef 505static int journal_file_map_data_hash_table(JournalFile *f) {
cec736d2
LP		 506 uint64_t s, p;
507 void *t;
508 int r;
509
510 assert(f);
511
de190aef
LP		 512 p = le64toh(f->header->data_hash_table_offset);
513 s = le64toh(f->header->data_hash_table_size);
cec736d2 514
de190aef
LP		 515 r = journal_file_move_to(f,
516 WINDOW_DATA_HASH_TABLE,
517 p, s,
518 &t);
cec736d2
LP		 519 if (r < 0)
520 return r;
521
de190aef 522 f->data_hash_table = t;
cec736d2
LP		 523 return 0;
524}
525
de190aef 526static int journal_file_map_field_hash_table(JournalFile *f) {
cec736d2
LP		 527 uint64_t s, p;
528 void *t;
529 int r;
530
531 assert(f);
532
de190aef
LP		 533 p = le64toh(f->header->field_hash_table_offset);
534 s = le64toh(f->header->field_hash_table_size);
cec736d2 535
de190aef
LP		 536 r = journal_file_move_to(f,
537 WINDOW_FIELD_HASH_TABLE,
538 p, s,
539 &t);
cec736d2
LP		 540 if (r < 0)
541 return r;
542
de190aef 543 f->field_hash_table = t;
cec736d2
LP		 544 return 0;
545}
546
de190aef
LP		 547static int journal_file_link_data(JournalFile *f, Object *o, uint64_t offset, uint64_t hash) {
548 uint64_t p, h;
cec736d2
LP		 549 int r;
550
551 assert(f);
552 assert(o);
553 assert(offset > 0);
de190aef 554 assert(o->object.type == OBJECT_DATA);
cec736d2 555
de190aef
LP		 556 o->data.next_hash_offset = o->data.next_field_offset = 0;
557 o->data.entry_offset = o->data.entry_array_offset = 0;
558 o->data.n_entries = 0;
cec736d2 559
de190aef
LP		 560 h = hash % (le64toh(f->header->data_hash_table_size) / sizeof(HashItem));
561 p = le64toh(f->data_hash_table[h].head_hash_offset);
cec736d2
LP		 562 if (p == 0) {
563 /* Only entry in the hash table is easy */
de190aef 564 f->data_hash_table[h].head_hash_offset = htole64(offset);
cec736d2 565 } else {
cec736d2
LP		 566 /* Temporarily move back to the previous data object,
567 * to patch in pointer */
568
de190aef 569 r = journal_file_move_to_object(f, OBJECT_DATA, p, &o);
cec736d2
LP		 570 if (r < 0)
571 return r;
572
de190aef 573 o->data.next_hash_offset = htole64(offset);
cec736d2 574
de190aef 575 r = journal_file_move_to_object(f, OBJECT_DATA, offset, &o);
cec736d2
LP		 576 if (r < 0)
577 return r;
578 }
579
de190aef 580 f->data_hash_table[h].tail_hash_offset = htole64(offset);
cec736d2
LP		 581
582 return 0;
583}
584
de190aef
LP		 585int journal_file_find_data_object_with_hash(
586 JournalFile *f,
587 const void *data, uint64_t size, uint64_t hash,
588 Object **ret, uint64_t *offset) {
589 uint64_t p, osize, h;
cec736d2
LP		 590 int r;
591
592 assert(f);
593 assert(data || size == 0);
594
595 osize = offsetof(Object, data.payload) + size;
596
bc85bfee
LP		 597 if (f->header->data_hash_table_size == 0)
598 return -EBADMSG;
599
de190aef
LP		 600 h = hash % (le64toh(f->header->data_hash_table_size) / sizeof(HashItem));
601 p = le64toh(f->data_hash_table[h].head_hash_offset);
cec736d2 602
de190aef
LP		 603 while (p > 0) {
604 Object *o;
cec736d2 605
de190aef 606 r = journal_file_move_to_object(f, OBJECT_DATA, p, &o);
cec736d2
LP		 607 if (r < 0)
608 return r;
609
807e17f0 610 if (le64toh(o->data.hash) != hash)
85a131e8 611 goto next;
807e17f0
LP		 612
613 if (o->object.flags & OBJECT_COMPRESSED) {
614#ifdef HAVE_XZ
615 uint64_t l, rsize;
cec736d2 616
807e17f0
LP		 617 l = le64toh(o->object.size);
618 if (l <= offsetof(Object, data.payload))
cec736d2
LP		 619 return -EBADMSG;
620
807e17f0
LP		 621 l -= offsetof(Object, data.payload);
622
623 if (!uncompress_blob(o->data.payload, l, &f->compress_buffer, &f->compress_buffer_size, &rsize))
624 return -EBADMSG;
625
626 if (rsize == size &&
627 memcmp(f->compress_buffer, data, size) == 0) {
628
629 if (ret)
630 *ret = o;
631
632 if (offset)
633 *offset = p;
634
635 return 1;
636 }
637#else
638 return -EPROTONOSUPPORT;
639#endif
640
641 } else if (le64toh(o->object.size) == osize &&
642 memcmp(o->data.payload, data, size) == 0) {
643
cec736d2
LP		 644 if (ret)
645 *ret = o;
646
647 if (offset)
648 *offset = p;
649
de190aef 650 return 1;
cec736d2
LP		 651 }
652
85a131e8 653 next:
cec736d2
LP		 654 p = le64toh(o->data.next_hash_offset);
655 }
656
de190aef
LP		 657 return 0;
658}
659
660int journal_file_find_data_object(
661 JournalFile *f,
662 const void *data, uint64_t size,
663 Object **ret, uint64_t *offset) {
664
665 uint64_t hash;
666
667 assert(f);
668 assert(data || size == 0);
669
670 hash = hash64(data, size);
671
672 return journal_file_find_data_object_with_hash(f,
673 data, size, hash,
674 ret, offset);
675}
676
677static int journal_file_append_data(JournalFile *f, const void *data, uint64_t size, Object **ret, uint64_t *offset) {
678 uint64_t hash, p;
679 uint64_t osize;
680 Object *o;
681 int r;
807e17f0 682 bool compressed = false;
de190aef
LP		 683
684 assert(f);
685 assert(data || size == 0);
686
687 hash = hash64(data, size);
688
689 r = journal_file_find_data_object_with_hash(f, data, size, hash, &o, &p);
690 if (r < 0)
691 return r;
692 else if (r > 0) {
693
694 if (ret)
695 *ret = o;
696
697 if (offset)
698 *offset = p;
699
700 return 0;
701 }
702
703 osize = offsetof(Object, data.payload) + size;
704 r = journal_file_append_object(f, OBJECT_DATA, osize, &o, &p);
cec736d2
LP		 705 if (r < 0)
706 return r;
707
cec736d2 708 o->data.hash = htole64(hash);
807e17f0
LP		 709
710#ifdef HAVE_XZ
711 if (f->compress &&
712 size >= COMPRESSION_SIZE_THRESHOLD) {
713 uint64_t rsize;
714
715 compressed = compress_blob(data, size, o->data.payload, &rsize);
716
717 if (compressed) {
718 o->object.size = htole64(offsetof(Object, data.payload) + rsize);
719 o->object.flags |= OBJECT_COMPRESSED;
720
721 f->header->incompatible_flags = htole32(le32toh(f->header->incompatible_flags) | HEADER_INCOMPATIBLE_COMPRESSED);
722
723 log_debug("Compressed data object %lu -> %lu", (unsigned long) size, (unsigned long) rsize);
724 }
725 }
726#endif
727
728 if (!compressed)
729 memcpy(o->data.payload, data, size);
cec736d2 730
de190aef 731 r = journal_file_link_data(f, o, p, hash);
cec736d2
LP		 732 if (r < 0)
733 return r;
734
735 if (ret)
736 *ret = o;
737
738 if (offset)
de190aef 739 *offset = p;
cec736d2
LP		 740
741 return 0;
742}
743
744uint64_t journal_file_entry_n_items(Object *o) {
745 assert(o);
746 assert(o->object.type == htole64(OBJECT_ENTRY));
747
748 return (le64toh(o->object.size) - offsetof(Object, entry.items)) / sizeof(EntryItem);
749}
750
de190aef
LP		 751static uint64_t journal_file_entry_array_n_items(Object *o) {
752 assert(o);
753 assert(o->object.type == htole64(OBJECT_ENTRY_ARRAY));
754
755 return (le64toh(o->object.size) - offsetof(Object, entry_array.items)) / sizeof(uint64_t);
756}
757
758static int link_entry_into_array(JournalFile *f,
759 uint64_t *first,
760 uint64_t *idx,
761 uint64_t p) {
cec736d2 762 int r;
de190aef
LP		 763 uint64_t n = 0, ap = 0, q, i, a, hidx;
764 Object *o;
765
cec736d2 766 assert(f);
de190aef
LP		 767 assert(first);
768 assert(idx);
769 assert(p > 0);
cec736d2 770
de190aef
LP		 771 a = le64toh(*first);
772 i = hidx = le64toh(*idx);
773 while (a > 0) {
774
775 r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &o);
776 if (r < 0)
777 return r;
cec736d2 778
de190aef
LP		 779 n = journal_file_entry_array_n_items(o);
780 if (i < n) {
781 o->entry_array.items[i] = htole64(p);
782 *idx = htole64(hidx + 1);
783 return 0;
784 }
cec736d2 785
de190aef
LP		 786 i -= n;
787 ap = a;
788 a = le64toh(o->entry_array.next_entry_array_offset);
789 }
790
791 if (hidx > n)
792 n = (hidx+1) * 2;
793 else
794 n = n * 2;
795
796 if (n < 4)
797 n = 4;
798
799 r = journal_file_append_object(f, OBJECT_ENTRY_ARRAY,
800 offsetof(Object, entry_array.items) + n * sizeof(uint64_t),
801 &o, &q);
cec736d2
LP		 802 if (r < 0)
803 return r;
804
de190aef 805 o->entry_array.items[i] = htole64(p);
cec736d2 806
de190aef
LP		 807 if (ap == 0)
808 *first = q;
cec736d2 809 else {
de190aef 810 r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, ap, &o);
cec736d2
LP		 811 if (r < 0)
812 return r;
813
de190aef
LP		 814 o->entry_array.next_entry_array_offset = htole64(q);
815 }
cec736d2 816
de190aef
LP		 817 *idx = htole64(hidx + 1);
818
819 return 0;
820}
cec736d2 821
de190aef
LP		 822static int link_entry_into_array_plus_one(JournalFile *f,
823 uint64_t *extra,
824 uint64_t *first,
825 uint64_t *idx,
826 uint64_t p) {
827
828 int r;
829
830 assert(f);
831 assert(extra);
832 assert(first);
833 assert(idx);
834 assert(p > 0);
835
836 if (*idx == 0)
837 *extra = htole64(p);
838 else {
839 uint64_t i;
840
841 i = le64toh(*idx) - 1;
842 r = link_entry_into_array(f, first, &i, p);
843 if (r < 0)
844 return r;
cec736d2
LP		 845 }
846
de190aef
LP		 847 *idx = htole64(le64toh(*idx) + 1);
848 return 0;
849}
850
851static int journal_file_link_entry_item(JournalFile *f, Object *o, uint64_t offset, uint64_t i) {
852 uint64_t p;
853 int r;
854 assert(f);
855 assert(o);
856 assert(offset > 0);
857
858 p = le64toh(o->entry.items[i].object_offset);
859 if (p == 0)
860 return -EINVAL;
861
862 r = journal_file_move_to_object(f, OBJECT_DATA, p, &o);
cec736d2
LP		 863 if (r < 0)
864 return r;
865
de190aef
LP		 866 return link_entry_into_array_plus_one(f,
867 &o->data.entry_offset,
868 &o->data.entry_array_offset,
869 &o->data.n_entries,
870 offset);
cec736d2
LP		 871}
872
873static int journal_file_link_entry(JournalFile *f, Object *o, uint64_t offset) {
de190aef 874 uint64_t n, i;
cec736d2
LP		 875 int r;
876
877 assert(f);
878 assert(o);
879 assert(offset > 0);
de190aef 880 assert(o->object.type == OBJECT_ENTRY);
cec736d2
LP		 881
882 /* Link up the entry itself */
de190aef
LP		 883 r = link_entry_into_array(f,
884 &f->header->entry_array_offset,
885 &f->header->n_entries,
886 offset);
887 if (r < 0)
888 return r;
cec736d2 889
bc85bfee 890 log_error("=> %s seqnr=%lu n_entries=%lu", f->path, (unsigned long) o->entry.seqnum, (unsigned long) f->header->n_entries);
cec736d2 891
de190aef 892 if (f->header->head_entry_realtime == 0)
0ac38b70 893 f->header->head_entry_realtime = o->entry.realtime;
cec736d2 894
0ac38b70 895 f->header->tail_entry_realtime = o->entry.realtime;
de190aef
LP		 896 f->header->tail_entry_monotonic = o->entry.monotonic;
897
898 f->tail_entry_monotonic_valid = true;
cec736d2
LP		 899
900 /* Link up the items */
901 n = journal_file_entry_n_items(o);
902 for (i = 0; i < n; i++) {
903 r = journal_file_link_entry_item(f, o, offset, i);
904 if (r < 0)
905 return r;
906 }
907
cec736d2
LP		 908 return 0;
909}
910
911static int journal_file_append_entry_internal(
912 JournalFile *f,
913 const dual_timestamp *ts,
914 uint64_t xor_hash,
915 const EntryItem items[], unsigned n_items,
de190aef 916 uint64_t *seqnum,
cec736d2
LP		 917 Object **ret, uint64_t *offset) {
918 uint64_t np;
919 uint64_t osize;
920 Object *o;
921 int r;
922
923 assert(f);
924 assert(items || n_items == 0);
de190aef 925 assert(ts);
cec736d2
LP		 926
927 osize = offsetof(Object, entry.items) + (n_items * sizeof(EntryItem));
928
de190aef 929 r = journal_file_append_object(f, OBJECT_ENTRY, osize, &o, &np);
cec736d2
LP		 930 if (r < 0)
931 return r;
932
de190aef 933 o->entry.seqnum = htole64(journal_file_seqnum(f, seqnum));
cec736d2 934 memcpy(o->entry.items, items, n_items * sizeof(EntryItem));
de190aef
LP		 935 o->entry.realtime = htole64(ts->realtime);
936 o->entry.monotonic = htole64(ts->monotonic);
cec736d2
LP		 937 o->entry.xor_hash = htole64(xor_hash);
938 o->entry.boot_id = f->header->boot_id;
939
940 r = journal_file_link_entry(f, o, np);
941 if (r < 0)
942 return r;
943
944 if (ret)
945 *ret = o;
946
947 if (offset)
948 *offset = np;
949
950 return 0;
951}
952
50f20cfd
LP		 953static void journal_file_post_change(JournalFile *f) {
954 assert(f);
955
956 /* inotify() does not receive IN_MODIFY events from file
957 * accesses done via mmap(). After each access we hence
958 * trigger IN_MODIFY by truncating the journal file to its
959 * current size which triggers IN_MODIFY. */
960
bc85bfee
LP		 961 __sync_synchronize();
962
50f20cfd
LP		 963 if (ftruncate(f->fd, f->last_stat.st_size) < 0)
964 log_error("Failed to to truncate file to its own size: %m");
965}
966
de190aef 967int 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		 968 unsigned i;
969 EntryItem *items;
970 int r;
971 uint64_t xor_hash = 0;
de190aef 972 struct dual_timestamp _ts;
cec736d2
LP		 973
974 assert(f);
975 assert(iovec || n_iovec == 0);
976
de190aef
LP		 977 if (!f->writable)
978 return -EPERM;
979
980 if (!ts) {
981 dual_timestamp_get(&_ts);
982 ts = &_ts;
983 }
984
985 if (f->tail_entry_monotonic_valid &&
986 ts->monotonic < le64toh(f->header->tail_entry_monotonic))
987 return -EINVAL;
988
989 if (ts->realtime < le64toh(f->header->tail_entry_realtime))
990 return -EINVAL;
991
cec736d2
LP		 992 items = new(EntryItem, n_iovec);
993 if (!items)
994 return -ENOMEM;
995
996 for (i = 0; i < n_iovec; i++) {
997 uint64_t p;
998 Object *o;
999
1000 r = journal_file_append_data(f, iovec[i].iov_base, iovec[i].iov_len, &o, &p);
1001 if (r < 0)
1002 goto finish;
1003
1004 xor_hash ^= le64toh(o->data.hash);
1005 items[i].object_offset = htole64(p);
de7b95cd 1006 items[i].hash = o->data.hash;
cec736d2
LP		 1007 }
1008
de190aef 1009 r = journal_file_append_entry_internal(f, ts, xor_hash, items, n_iovec, seqnum, ret, offset);
cec736d2 1010
50f20cfd
LP		 1011 journal_file_post_change(f);
1012
cec736d2
LP		 1013finish:
1014 free(items);
1015
1016 return r;
1017}
1018
de190aef
LP		 1019static int generic_array_get(JournalFile *f,
1020 uint64_t first,
1021 uint64_t i,
1022 Object **ret, uint64_t *offset) {
1023
cec736d2 1024 Object *o;
de190aef 1025 uint64_t p, a;
cec736d2
LP		 1026 int r;
1027
1028 assert(f);
1029
de190aef
LP		 1030 a = first;
1031 while (a > 0) {
1032 uint64_t n;
cec736d2 1033
de190aef
LP		 1034 r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &o);
1035 if (r < 0)
1036 return r;
cec736d2 1037
de190aef
LP		 1038 n = journal_file_entry_array_n_items(o);
1039 if (i < n) {
1040 p = le64toh(o->entry_array.items[i]);
1041 break;
cec736d2
LP		 1042 }
1043
de190aef
LP		 1044 i -= n;
1045 a = le64toh(o->entry_array.next_entry_array_offset);
1046 }
1047
1048 if (a <= 0 || p <= 0)
1049 return 0;
1050
1051 r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o);
1052 if (r < 0)
1053 return r;
1054
1055 if (ret)
1056 *ret = o;
1057
1058 if (offset)
1059 *offset = p;
1060
1061 return 1;
1062}
1063
1064static int generic_array_get_plus_one(JournalFile *f,
1065 uint64_t extra,
1066 uint64_t first,
1067 uint64_t i,
1068 Object **ret, uint64_t *offset) {
1069
1070 Object *o;
1071
1072 assert(f);
1073
1074 if (i == 0) {
1075 int r;
1076
1077 r = journal_file_move_to_object(f, OBJECT_ENTRY, extra, &o);
cec736d2
LP		 1078 if (r < 0)
1079 return r;
1080
de190aef
LP		 1081 if (ret)
1082 *ret = o;
cec736d2 1083
de190aef
LP		 1084 if (offset)
1085 *offset = extra;
cec736d2 1086
de190aef 1087 return 1;
cec736d2
LP		 1088 }
1089
de190aef
LP		 1090 return generic_array_get(f, first, i-1, ret, offset);
1091}
cec736d2 1092
de190aef
LP		 1093enum {
1094 TEST_FOUND,
1095 TEST_LEFT,
1096 TEST_RIGHT
1097};
cec736d2 1098
de190aef
LP		 1099static int generic_array_bisect(JournalFile *f,
1100 uint64_t first,
1101 uint64_t n,
1102 uint64_t needle,
1103 int (*test_object)(JournalFile *f, uint64_t p, uint64_t needle),
1104 direction_t direction,
1105 Object **ret,
1106 uint64_t *offset,
1107 uint64_t *idx) {
1108
1109 uint64_t a, p, t = 0, i = 0, last_p = 0;
1110 bool subtract_one = false;
1111 Object *o, *array = NULL;
1112 int r;
cec736d2 1113
de190aef
LP		 1114 assert(f);
1115 assert(test_object);
cec736d2 1116
de190aef
LP		 1117 a = first;
1118 while (a > 0) {
1119 uint64_t left, right, k, lp;
1120
1121 r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &array);
cec736d2
LP		 1122 if (r < 0)
1123 return r;
1124
de190aef
LP		 1125 k = journal_file_entry_array_n_items(array);
1126 right = MIN(k, n);
1127 if (right <= 0)
1128 return 0;
cec736d2 1129
de190aef
LP		 1130 i = right - 1;
1131 lp = p = le64toh(array->entry_array.items[i]);
1132 if (p <= 0)
1133 return -EBADMSG;
cec736d2 1134
de190aef
LP		 1135 r = test_object(f, p, needle);
1136 if (r < 0)
1137 return r;
cec736d2 1138
de190aef
LP		 1139 if (r == TEST_FOUND)
1140 r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT;
1141
1142 if (r == TEST_RIGHT) {
1143 left = 0;
1144 right -= 1;
1145 for (;;) {
1146 if (left == right) {
1147 if (direction == DIRECTION_UP)
1148 subtract_one = true;
1149
1150 i = left;
1151 goto found;
1152 }
1153
1154 assert(left < right);
1155
1156 i = (left + right) / 2;
1157 p = le64toh(array->entry_array.items[i]);
1158 if (p <= 0)
1159 return -EBADMSG;
1160
1161 r = test_object(f, p, needle);
1162 if (r < 0)
1163 return r;
cec736d2 1164
de190aef
LP		 1165 if (r == TEST_FOUND)
1166 r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT;
1167
1168 if (r == TEST_RIGHT)
1169 right = i;
1170 else
1171 left = i + 1;
1172 }
1173 }
1174
1175 if (k > n)
cec736d2
LP		 1176 return 0;
1177
de190aef
LP		 1178 last_p = lp;
1179
1180 n -= k;
1181 t += k;
1182 a = le64toh(array->entry_array.next_entry_array_offset);
cec736d2
LP		 1183 }
1184
1185 return 0;
de190aef
LP		 1186
1187found:
1188 if (subtract_one && t == 0 && i == 0)
1189 return 0;
1190
1191 if (subtract_one && i == 0)
1192 p = last_p;
1193 else if (subtract_one)
1194 p = le64toh(array->entry_array.items[i-1]);
1195 else
1196 p = le64toh(array->entry_array.items[i]);
1197
1198 r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o);
1199 if (r < 0)
1200 return r;
1201
1202 if (ret)
1203 *ret = o;
1204
1205 if (offset)
1206 *offset = p;
1207
1208 if (idx)
1209 *idx = t + i - (subtract_one ? 1 : 0);
1210
1211 return 1;
cec736d2
LP		 1212}
1213
de190aef
LP		 1214static int generic_array_bisect_plus_one(JournalFile *f,
1215 uint64_t extra,
1216 uint64_t first,
1217 uint64_t n,
1218 uint64_t needle,
1219 int (*test_object)(JournalFile *f, uint64_t p, uint64_t needle),
1220 direction_t direction,
1221 Object **ret,
1222 uint64_t *offset,
1223 uint64_t *idx) {
1224
cec736d2
LP		 1225 int r;
1226
1227 assert(f);
de190aef 1228 assert(test_object);
cec736d2 1229
de190aef
LP		 1230 if (n <= 0)
1231 return 0;
cec736d2 1232
de190aef
LP		 1233 /* This bisects the array in object 'first', but first checks
1234 * an extra */
de190aef
LP		 1235 r = test_object(f, extra, needle);
1236 if (r < 0)
1237 return r;
1238 else if (r == TEST_FOUND) {
1239 Object *o;
1240
1241 r = journal_file_move_to_object(f, OBJECT_ENTRY, extra, &o);
1242 if (r < 0)
1243 return r;
1244
1245 if (ret)
1246 *ret = o;
cec736d2 1247
de190aef
LP		 1248 if (offset)
1249 *offset = extra;
440ee366
LP		 1250
1251 if (idx)
1252 *idx = 0;
1253
1254 return 1;
de190aef 1255 } else if (r == TEST_RIGHT)
cec736d2
LP		 1256 return 0;
1257
de190aef
LP		 1258 r = generic_array_bisect(f, first, n-1, needle, test_object, direction, ret, offset, idx);
1259
1260 if (r > 0)
1261 (*idx) ++;
1262
1263 return r;
1264}
1265
1266static int test_object_seqnum(JournalFile *f, uint64_t p, uint64_t needle) {
1267 Object *o;
1268 int r;
1269
1270 assert(f);
1271 assert(p > 0);
1272
1273 r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o);
cec736d2
LP		 1274 if (r < 0)
1275 return r;
1276
de190aef
LP		 1277 if (le64toh(o->entry.seqnum) == needle)
1278 return TEST_FOUND;
1279 else if (le64toh(o->entry.seqnum) < needle)
1280 return TEST_LEFT;
1281 else
1282 return TEST_RIGHT;
1283}
cec736d2 1284
de190aef
LP		 1285int journal_file_move_to_entry_by_seqnum(
1286 JournalFile *f,
1287 uint64_t seqnum,
1288 direction_t direction,
1289 Object **ret,
1290 uint64_t *offset) {
1291
1292 return generic_array_bisect(f,
1293 le64toh(f->header->entry_array_offset),
1294 le64toh(f->header->n_entries),
1295 seqnum,
1296 test_object_seqnum,
1297 direction,
1298 ret, offset, NULL);
1299}
cec736d2 1300
de190aef
LP		 1301static int test_object_realtime(JournalFile *f, uint64_t p, uint64_t needle) {
1302 Object *o;
1303 int r;
1304
1305 assert(f);
1306 assert(p > 0);
1307
1308 r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o);
1309 if (r < 0)
1310 return r;
1311
1312 if (le64toh(o->entry.realtime) == needle)
1313 return TEST_FOUND;
1314 else if (le64toh(o->entry.realtime) < needle)
1315 return TEST_LEFT;
1316 else
1317 return TEST_RIGHT;
cec736d2
LP		 1318}
1319
de190aef
LP		 1320int journal_file_move_to_entry_by_realtime(
1321 JournalFile *f,
1322 uint64_t realtime,
1323 direction_t direction,
1324 Object **ret,
1325 uint64_t *offset) {
1326
1327 return generic_array_bisect(f,
1328 le64toh(f->header->entry_array_offset),
1329 le64toh(f->header->n_entries),
1330 realtime,
1331 test_object_realtime,
1332 direction,
1333 ret, offset, NULL);
1334}
1335
1336static int test_object_monotonic(JournalFile *f, uint64_t p, uint64_t needle) {
1337 Object *o;
1338 int r;
1339
1340 assert(f);
1341 assert(p > 0);
1342
1343 r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o);
1344 if (r < 0)
1345 return r;
1346
1347 if (le64toh(o->entry.monotonic) == needle)
1348 return TEST_FOUND;
1349 else if (le64toh(o->entry.monotonic) < needle)
1350 return TEST_LEFT;
1351 else
1352 return TEST_RIGHT;
1353}
1354
1355int journal_file_move_to_entry_by_monotonic(
1356 JournalFile *f,
1357 sd_id128_t boot_id,
1358 uint64_t monotonic,
1359 direction_t direction,
1360 Object **ret,
1361 uint64_t *offset) {
1362
1363 char t[8+32+1] = "_BOOT_ID=";
1364 Object *o;
1365 int r;
1366
1367 sd_id128_to_string(boot_id, t + 8);
1368
1369 r = journal_file_find_data_object(f, t, strlen(t), &o, NULL);
1370 if (r < 0)
1371 return r;
1372 else if (r == 0)
1373 return -ENOENT;
1374
1375 return generic_array_bisect_plus_one(f,
1376 le64toh(o->data.entry_offset),
1377 le64toh(o->data.entry_array_offset),
1378 le64toh(o->data.n_entries),
1379 monotonic,
1380 test_object_monotonic,
1381 direction,
1382 ret, offset, NULL);
1383}
1384
1385static int test_object_offset(JournalFile *f, uint64_t p, uint64_t needle) {
1386 assert(f);
1387 assert(p > 0);
1388
1389 if (p == needle)
1390 return TEST_FOUND;
1391 else if (p < needle)
1392 return TEST_LEFT;
1393 else
1394 return TEST_RIGHT;
1395}
1396
1397int journal_file_next_entry(
1398 JournalFile *f,
1399 Object *o, uint64_t p,
1400 direction_t direction,
1401 Object **ret, uint64_t *offset) {
1402
1403 uint64_t i, n;
cec736d2
LP		 1404 int r;
1405
1406 assert(f);
de190aef
LP		 1407 assert(p > 0 || !o);
1408
1409 n = le64toh(f->header->n_entries);
1410 if (n <= 0)
1411 return 0;
cec736d2
LP		 1412
1413 if (!o)
de190aef 1414 i = direction == DIRECTION_DOWN ? 0 : n - 1;
cec736d2 1415 else {
de190aef 1416 if (o->object.type != OBJECT_ENTRY)
cec736d2
LP		 1417 return -EINVAL;
1418
de190aef
LP		 1419 r = generic_array_bisect(f,
1420 le64toh(f->header->entry_array_offset),
1421 le64toh(f->header->n_entries),
1422 p,
1423 test_object_offset,
1424 DIRECTION_DOWN,
1425 NULL, NULL,
1426 &i);
1427 if (r <= 0)
1428 return r;
1429
1430 if (direction == DIRECTION_DOWN) {
1431 if (i >= n - 1)
1432 return 0;
1433
1434 i++;
1435 } else {
1436 if (i <= 0)
1437 return 0;
1438
1439 i--;
1440 }
cec736d2
LP		 1441 }
1442
de190aef
LP		 1443 /* And jump to it */
1444 return generic_array_get(f,
1445 le64toh(f->header->entry_array_offset),
1446 i,
1447 ret, offset);
1448}
cec736d2 1449
de190aef
LP		 1450int journal_file_skip_entry(
1451 JournalFile *f,
1452 Object *o, uint64_t p,
1453 int64_t skip,
1454 Object **ret, uint64_t *offset) {
1455
1456 uint64_t i, n;
1457 int r;
1458
1459 assert(f);
1460 assert(o);
1461 assert(p > 0);
1462
1463 if (o->object.type != OBJECT_ENTRY)
1464 return -EINVAL;
1465
1466 r = generic_array_bisect(f,
1467 le64toh(f->header->entry_array_offset),
1468 le64toh(f->header->n_entries),
1469 p,
1470 test_object_offset,
1471 DIRECTION_DOWN,
1472 NULL, NULL,
1473 &i);
1474 if (r <= 0)
cec736d2
LP		 1475 return r;
1476
de190aef
LP		 1477 /* Calculate new index */
1478 if (skip < 0) {
1479 if ((uint64_t) -skip >= i)
1480 i = 0;
1481 else
1482 i = i - (uint64_t) -skip;
1483 } else
1484 i += (uint64_t) skip;
cec736d2 1485
de190aef
LP		 1486 n = le64toh(f->header->n_entries);
1487 if (n <= 0)
1488 return -EBADMSG;
cec736d2 1489
de190aef
LP		 1490 if (i >= n)
1491 i = n-1;
1492
1493 return generic_array_get(f,
1494 le64toh(f->header->entry_array_offset),
1495 i,
1496 ret, offset);
cec736d2
LP		 1497}
1498
de190aef
LP		 1499int journal_file_next_entry_for_data(
1500 JournalFile *f,
1501 Object *o, uint64_t p,
1502 uint64_t data_offset,
1503 direction_t direction,
1504 Object **ret, uint64_t *offset) {
1505
1506 uint64_t n, i;
cec736d2 1507 int r;
de190aef 1508 Object *d;
cec736d2
LP		 1509
1510 assert(f);
de190aef 1511 assert(p > 0 || !o);
cec736d2 1512
de190aef 1513 r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d);
466ccd92 1514 if (r < 0)
de190aef 1515 return r;
cec736d2 1516
de190aef
LP		 1517 n = le64toh(d->data.n_entries);
1518 if (n <= 0)
1519 return n;
cec736d2 1520
de190aef
LP		 1521 if (!o)
1522 i = direction == DIRECTION_DOWN ? 0 : n - 1;
1523 else {
1524 if (o->object.type != OBJECT_ENTRY)
1525 return -EINVAL;
cec736d2 1526
de190aef
LP		 1527 r = generic_array_bisect_plus_one(f,
1528 le64toh(d->data.entry_offset),
1529 le64toh(d->data.entry_array_offset),
1530 le64toh(d->data.n_entries),
1531 p,
1532 test_object_offset,
1533 DIRECTION_DOWN,
1534 NULL, NULL,
1535 &i);
1536
1537 if (r <= 0)
cec736d2
LP		 1538 return r;
1539
de190aef
LP		 1540 if (direction == DIRECTION_DOWN) {
1541 if (i >= n - 1)
1542 return 0;
cec736d2 1543
de190aef
LP		 1544 i++;
1545 } else {
1546 if (i <= 0)
1547 return 0;
cec736d2 1548
de190aef
LP		 1549 i--;
1550 }
cec736d2 1551
de190aef 1552 }
cec736d2 1553
de190aef
LP		 1554 return generic_array_get_plus_one(f,
1555 le64toh(d->data.entry_offset),
1556 le64toh(d->data.entry_array_offset),
1557 i,
1558 ret, offset);
1559}
cec736d2 1560
de190aef
LP		 1561int journal_file_move_to_entry_by_seqnum_for_data(
1562 JournalFile *f,
1563 uint64_t data_offset,
1564 uint64_t seqnum,
1565 direction_t direction,
1566 Object **ret, uint64_t *offset) {
cec736d2 1567
de190aef
LP		 1568 Object *d;
1569 int r;
cec736d2 1570
de190aef
LP		 1571 r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d);
1572 if (r <= 0)
1573 return r;
cec736d2 1574
de190aef
LP		 1575 return generic_array_bisect_plus_one(f,
1576 le64toh(d->data.entry_offset),
1577 le64toh(d->data.entry_array_offset),
1578 le64toh(d->data.n_entries),
1579 seqnum,
1580 test_object_seqnum,
1581 direction,
1582 ret, offset, NULL);
1583}
cec736d2 1584
de190aef
LP		 1585int journal_file_move_to_entry_by_realtime_for_data(
1586 JournalFile *f,
1587 uint64_t data_offset,
1588 uint64_t realtime,
1589 direction_t direction,
1590 Object **ret, uint64_t *offset) {
1591
1592 Object *d;
1593 int r;
1594
1595 r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d);
1596 if (r <= 0)
1597 return r;
1598
1599 return generic_array_bisect_plus_one(f,
1600 le64toh(d->data.entry_offset),
1601 le64toh(d->data.entry_array_offset),
1602 le64toh(d->data.n_entries),
1603 realtime,
1604 test_object_realtime,
1605 direction,
1606 ret, offset, NULL);
cec736d2
LP		 1607}
1608
1609void journal_file_dump(JournalFile *f) {
1610 char a[33], b[33], c[33];
1611 Object *o;
1612 int r;
1613 uint64_t p;
1614
1615 assert(f);
1616
de190aef
LP		 1617 printf("File Path: %s\n"
1618 "File ID: %s\n"
cec736d2
LP		 1619 "Machine ID: %s\n"
1620 "Boot ID: %s\n"
de190aef
LP		 1621 "Arena size: %llu\n"
1622 "Objects: %lu\n"
1623 "Entries: %lu\n",
1624 f->path,
cec736d2
LP		 1625 sd_id128_to_string(f->header->file_id, a),
1626 sd_id128_to_string(f->header->machine_id, b),
1627 sd_id128_to_string(f->header->boot_id, c),
de190aef
LP		 1628 (unsigned long long) le64toh(f->header->arena_size),
1629 (unsigned long) le64toh(f->header->n_objects),
1630 (unsigned long) le64toh(f->header->n_entries));
cec736d2 1631
de190aef 1632 p = le64toh(f->header->arena_offset);
cec736d2 1633 while (p != 0) {
de190aef 1634 r = journal_file_move_to_object(f, -1, p, &o);
cec736d2
LP		 1635 if (r < 0)
1636 goto fail;
1637
1638 switch (o->object.type) {
1639
1640 case OBJECT_UNUSED:
1641 printf("Type: OBJECT_UNUSED\n");
1642 break;
1643
1644 case OBJECT_DATA:
1645 printf("Type: OBJECT_DATA\n");
1646 break;
1647
1648 case OBJECT_ENTRY:
3fbf9cbb
LP		 1649 printf("Type: OBJECT_ENTRY %llu %llu %llu\n",
1650 (unsigned long long) le64toh(o->entry.seqnum),
1651 (unsigned long long) le64toh(o->entry.monotonic),
1652 (unsigned long long) le64toh(o->entry.realtime));
cec736d2
LP		 1653 break;
1654
de190aef
LP		 1655 case OBJECT_FIELD_HASH_TABLE:
1656 printf("Type: OBJECT_FIELD_HASH_TABLE\n");
cec736d2
LP		 1657 break;
1658
de190aef
LP		 1659 case OBJECT_DATA_HASH_TABLE:
1660 printf("Type: OBJECT_DATA_HASH_TABLE\n");
1661 break;
1662
1663 case OBJECT_ENTRY_ARRAY:
1664 printf("Type: OBJECT_ENTRY_ARRAY\n");
cec736d2
LP		 1665 break;
1666 }
1667
807e17f0
LP		 1668 if (o->object.flags & OBJECT_COMPRESSED)
1669 printf("Flags: COMPRESSED\n");
1670
cec736d2
LP		 1671 if (p == le64toh(f->header->tail_object_offset))
1672 p = 0;
1673 else
1674 p = p + ALIGN64(le64toh(o->object.size));
1675 }
1676
1677 return;
1678fail:
1679 log_error("File corrupt");
1680}
1681
1682int journal_file_open(
1683 const char *fname,
1684 int flags,
1685 mode_t mode,
0ac38b70 1686 JournalFile *template,
cec736d2
LP		 1687 JournalFile **ret) {
1688
1689 JournalFile *f;
1690 int r;
1691 bool newly_created = false;
1692
1693 assert(fname);
1694
1695 if ((flags & O_ACCMODE) != O_RDONLY &&
1696 (flags & O_ACCMODE) != O_RDWR)
1697 return -EINVAL;
1698
1699 f = new0(JournalFile, 1);
1700 if (!f)
1701 return -ENOMEM;
1702
0ac38b70
LP		 1703 f->fd = -1;
1704 f->flags = flags;
1705 f->mode = mode;
cec736d2
LP		 1706 f->writable = (flags & O_ACCMODE) != O_RDONLY;
1707 f->prot = prot_from_flags(flags);
1708
bc85bfee
LP		 1709 f->metrics.max_size = DEFAULT_MAX_SIZE;
1710 f->metrics.min_size = DEFAULT_MIN_SIZE;
1711 f->metrics.keep_free = DEFAULT_KEEP_FREE;
1712
cec736d2
LP		 1713 f->path = strdup(fname);
1714 if (!f->path) {
1715 r = -ENOMEM;
1716 goto fail;
1717 }
1718
0ac38b70
LP		 1719 f->fd = open(f->path, f->flags|O_CLOEXEC, f->mode);
1720 if (f->fd < 0) {
1721 r = -errno;
1722 goto fail;
1723 }
1724
cec736d2
LP		 1725 if (fstat(f->fd, &f->last_stat) < 0) {
1726 r = -errno;
1727 goto fail;
1728 }
1729
1730 if (f->last_stat.st_size == 0 && f->writable) {
1731 newly_created = true;
1732
0ac38b70 1733 r = journal_file_init_header(f, template);
cec736d2
LP		 1734 if (r < 0)
1735 goto fail;
1736
1737 if (fstat(f->fd, &f->last_stat) < 0) {
1738 r = -errno;
1739 goto fail;
1740 }
1741 }
1742
1743 if (f->last_stat.st_size < (off_t) sizeof(Header)) {
1744 r = -EIO;
1745 goto fail;
1746 }
1747
1748 f->header = mmap(NULL, PAGE_ALIGN(sizeof(Header)), prot_from_flags(flags), MAP_SHARED, f->fd, 0);
1749 if (f->header == MAP_FAILED) {
1750 f->header = NULL;
1751 r = -errno;
1752 goto fail;
1753 }
1754
1755 if (!newly_created) {
1756 r = journal_file_verify_header(f);
1757 if (r < 0)
1758 goto fail;
1759 }
1760
1761 if (f->writable) {
1762 r = journal_file_refresh_header(f);
1763 if (r < 0)
1764 goto fail;
1765 }
1766
1767 if (newly_created) {
1768
de190aef 1769 r = journal_file_setup_field_hash_table(f);
cec736d2
LP		 1770 if (r < 0)
1771 goto fail;
1772
de190aef 1773 r = journal_file_setup_data_hash_table(f);
cec736d2
LP		 1774 if (r < 0)
1775 goto fail;
1776 }
1777
de190aef 1778 r = journal_file_map_field_hash_table(f);
cec736d2
LP		 1779 if (r < 0)
1780 goto fail;
1781
de190aef 1782 r = journal_file_map_data_hash_table(f);
cec736d2
LP		 1783 if (r < 0)
1784 goto fail;
1785
1786 if (ret)
1787 *ret = f;
1788
1789 return 0;
1790
1791fail:
1792 journal_file_close(f);
1793
1794 return r;
1795}
0ac38b70
LP		 1796
1797int journal_file_rotate(JournalFile **f) {
1798 char *p;
1799 size_t l;
1800 JournalFile *old_file, *new_file = NULL;
1801 int r;
1802
1803 assert(f);
1804 assert(*f);
1805
1806 old_file = *f;
1807
1808 if (!old_file->writable)
1809 return -EINVAL;
1810
1811 if (!endswith(old_file->path, ".journal"))
1812 return -EINVAL;
1813
1814 l = strlen(old_file->path);
1815
1816 p = new(char, l + 1 + 16 + 1 + 32 + 1 + 16 + 1);
1817 if (!p)
1818 return -ENOMEM;
1819
1820 memcpy(p, old_file->path, l - 8);
1821 p[l-8] = '@';
1822 sd_id128_to_string(old_file->header->seqnum_id, p + l - 8 + 1);
1823 snprintf(p + l - 8 + 1 + 32, 1 + 16 + 1 + 16 + 8 + 1,
1824 "-%016llx-%016llx.journal",
1825 (unsigned long long) le64toh((*f)->header->seqnum),
1826 (unsigned long long) le64toh((*f)->header->tail_entry_realtime));
1827
1828 r = rename(old_file->path, p);
1829 free(p);
1830
1831 if (r < 0)
1832 return -errno;
1833
1834 old_file->header->state = le32toh(STATE_ARCHIVED);
1835
1836 r = journal_file_open(old_file->path, old_file->flags, old_file->mode, old_file, &new_file);
1837 journal_file_close(old_file);
1838
1839 *f = new_file;
1840 return r;
1841}
1842
1843struct vacuum_info {
1844 off_t usage;
1845 char *filename;
1846
1847 uint64_t realtime;
1848 sd_id128_t seqnum_id;
1849 uint64_t seqnum;
1850};
1851
1852static int vacuum_compare(const void *_a, const void *_b) {
1853 const struct vacuum_info *a, *b;
1854
1855 a = _a;
1856 b = _b;
1857
1858 if (sd_id128_equal(a->seqnum_id, b->seqnum_id)) {
1859 if (a->seqnum < b->seqnum)
1860 return -1;
1861 else if (a->seqnum > b->seqnum)
1862 return 1;
1863 else
1864 return 0;
1865 }
1866
1867 if (a->realtime < b->realtime)
1868 return -1;
1869 else if (a->realtime > b->realtime)
1870 return 1;
1871 else
1872 return memcmp(&a->seqnum_id, &b->seqnum_id, 16);
1873}
1874
1875int journal_directory_vacuum(const char *directory, uint64_t max_use, uint64_t min_free) {
1876 DIR *d;
1877 int r = 0;
1878 struct vacuum_info *list = NULL;
1879 unsigned n_list = 0, n_allocated = 0, i;
1880 uint64_t sum = 0;
1881
1882 assert(directory);
1883
1884 if (max_use <= 0)
1885 max_use = DEFAULT_MAX_USE;
1886
1887 d = opendir(directory);
1888 if (!d)
1889 return -errno;
1890
1891 for (;;) {
1892 int k;
1893 struct dirent buf, *de;
1894 size_t q;
1895 struct stat st;
1896 char *p;
1897 unsigned long long seqnum, realtime;
1898 sd_id128_t seqnum_id;
1899
1900 k = readdir_r(d, &buf, &de);
1901 if (k != 0) {
1902 r = -k;
1903 goto finish;
1904 }
1905
1906 if (!de)
1907 break;
1908
1909 if (!dirent_is_file_with_suffix(de, ".journal"))
1910 continue;
1911
1912 q = strlen(de->d_name);
1913
1914 if (q < 1 + 32 + 1 + 16 + 1 + 16 + 8)
1915 continue;
1916
1917 if (de->d_name[q-8-16-1] != '-' ||
1918 de->d_name[q-8-16-1-16-1] != '-' ||
1919 de->d_name[q-8-16-1-16-1-32-1] != '@')
1920 continue;
1921
1922 if (fstatat(dirfd(d), de->d_name, &st, AT_SYMLINK_NOFOLLOW) < 0)
1923 continue;
1924
1925 if (!S_ISREG(st.st_mode))
1926 continue;
1927
1928 p = strdup(de->d_name);
1929 if (!p) {
1930 r = -ENOMEM;
1931 goto finish;
1932 }
1933
1934 de->d_name[q-8-16-1-16-1] = 0;
1935 if (sd_id128_from_string(de->d_name + q-8-16-1-16-1-32, &seqnum_id) < 0) {
1936 free(p);
1937 continue;
1938 }
1939
1940 if (sscanf(de->d_name + q-8-16-1-16, "%16llx-%16llx.journal", &seqnum, &realtime) != 2) {
1941 free(p);
1942 continue;
1943 }
1944
1945 if (n_list >= n_allocated) {
1946 struct vacuum_info *j;
1947
1948 n_allocated = MAX(n_allocated * 2U, 8U);
1949 j = realloc(list, n_allocated * sizeof(struct vacuum_info));
1950 if (!j) {
1951 free(p);
1952 r = -ENOMEM;
1953 goto finish;
1954 }
1955
1956 list = j;
1957 }
1958
1959 list[n_list].filename = p;
1960 list[n_list].usage = (uint64_t) st.st_blksize * (uint64_t) st.st_blocks;
1961 list[n_list].seqnum = seqnum;
1962 list[n_list].realtime = realtime;
1963 list[n_list].seqnum_id = seqnum_id;
1964
1965 sum += list[n_list].usage;
1966
1967 n_list ++;
1968 }
1969
1970 qsort(list, n_list, sizeof(struct vacuum_info), vacuum_compare);
1971
1972 for(i = 0; i < n_list; i++) {
1973 struct statvfs ss;
1974
1975 if (fstatvfs(dirfd(d), &ss) < 0) {
1976 r = -errno;
1977 goto finish;
1978 }
1979
1980 if (sum <= max_use &&
1981 (uint64_t) ss.f_bavail * (uint64_t) ss.f_bsize >= min_free)
1982 break;
1983
1984 if (unlinkat(dirfd(d), list[i].filename, 0) >= 0) {
1985 log_debug("Deleted archived journal %s/%s.", directory, list[i].filename);
1986 sum -= list[i].usage;
1987 } else if (errno != ENOENT)
1988 log_warning("Failed to delete %s/%s: %m", directory, list[i].filename);
1989 }
1990
1991finish:
1992 for (i = 0; i < n_list; i++)
1993 free(list[i].filename);
1994
1995 free(list);
1996
de190aef
LP		 1997 if (d)
1998 closedir(d);
1999
0ac38b70
LP		 2000 return r;
2001}
Unnamed repository; edit this file 'description' to name the repository.