]>
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> | |
11689d2a | 29 | #include <linux/fs.h> |
fb0951b0 | 30 | |
f27a3864 | 31 | #include "btrfs-util.h" |
cec736d2 LP |
32 | #include "journal-def.h" |
33 | #include "journal-file.h" | |
0284adc6 | 34 | #include "journal-authenticate.h" |
cec736d2 | 35 | #include "lookup3.h" |
807e17f0 | 36 | #include "compress.h" |
3df3e884 | 37 | #include "random-util.h" |
cec736d2 | 38 | |
4a92baf3 LP |
39 | #define DEFAULT_DATA_HASH_TABLE_SIZE (2047ULL*sizeof(HashItem)) |
40 | #define DEFAULT_FIELD_HASH_TABLE_SIZE (333ULL*sizeof(HashItem)) | |
cec736d2 | 41 | |
be19b7df | 42 | #define COMPRESSION_SIZE_THRESHOLD (512ULL) |
807e17f0 | 43 | |
babfc091 | 44 | /* This is the minimum journal file size */ |
253f59df | 45 | #define JOURNAL_FILE_SIZE_MIN (4ULL*1024ULL*1024ULL) /* 4 MiB */ |
babfc091 LP |
46 | |
47 | /* These are the lower and upper bounds if we deduce the max_use value | |
48 | * from the file system size */ | |
49 | #define DEFAULT_MAX_USE_LOWER (1ULL*1024ULL*1024ULL) /* 1 MiB */ | |
50 | #define DEFAULT_MAX_USE_UPPER (4ULL*1024ULL*1024ULL*1024ULL) /* 4 GiB */ | |
51 | ||
52 | /* This is the upper bound if we deduce max_size from max_use */ | |
71100051 | 53 | #define DEFAULT_MAX_SIZE_UPPER (128ULL*1024ULL*1024ULL) /* 128 MiB */ |
babfc091 LP |
54 | |
55 | /* This is the upper bound if we deduce the keep_free value from the | |
56 | * file system size */ | |
57 | #define DEFAULT_KEEP_FREE_UPPER (4ULL*1024ULL*1024ULL*1024ULL) /* 4 GiB */ | |
58 | ||
59 | /* This is the keep_free value when we can't determine the system | |
60 | * size */ | |
61 | #define DEFAULT_KEEP_FREE (1024ULL*1024ULL) /* 1 MB */ | |
62 | ||
dca6219e LP |
63 | /* n_data was the first entry we added after the initial file format design */ |
64 | #define HEADER_SIZE_MIN ALIGN64(offsetof(Header, n_data)) | |
cec736d2 | 65 | |
a4bcff5b LP |
66 | /* How many entries to keep in the entry array chain cache at max */ |
67 | #define CHAIN_CACHE_MAX 20 | |
68 | ||
a676e665 LP |
69 | /* How much to increase the journal file size at once each time we allocate something new. */ |
70 | #define FILE_SIZE_INCREASE (8ULL*1024ULL*1024ULL) /* 8MB */ | |
71 | ||
2678031a LP |
72 | /* Reread fstat() of the file for detecting deletions at least this often */ |
73 | #define LAST_STAT_REFRESH_USEC (5*USEC_PER_SEC) | |
74 | ||
fa6ac760 LP |
75 | /* The mmap context to use for the header we pick as one above the last defined typed */ |
76 | #define CONTEXT_HEADER _OBJECT_TYPE_MAX | |
77 | ||
9588bc32 | 78 | static int journal_file_set_online(JournalFile *f) { |
26687bf8 OS |
79 | assert(f); |
80 | ||
81 | if (!f->writable) | |
82 | return -EPERM; | |
83 | ||
84 | if (!(f->fd >= 0 && f->header)) | |
85 | return -EINVAL; | |
86 | ||
fa6ac760 LP |
87 | if (mmap_cache_got_sigbus(f->mmap, f->fd)) |
88 | return -EIO; | |
89 | ||
26687bf8 OS |
90 | switch(f->header->state) { |
91 | case STATE_ONLINE: | |
92 | return 0; | |
93 | ||
94 | case STATE_OFFLINE: | |
95 | f->header->state = STATE_ONLINE; | |
96 | fsync(f->fd); | |
97 | return 0; | |
98 | ||
99 | default: | |
100 | return -EINVAL; | |
101 | } | |
102 | } | |
103 | ||
104 | int journal_file_set_offline(JournalFile *f) { | |
105 | assert(f); | |
106 | ||
107 | if (!f->writable) | |
108 | return -EPERM; | |
109 | ||
110 | if (!(f->fd >= 0 && f->header)) | |
111 | return -EINVAL; | |
112 | ||
113 | if (f->header->state != STATE_ONLINE) | |
114 | return 0; | |
115 | ||
116 | fsync(f->fd); | |
117 | ||
fa6ac760 LP |
118 | if (mmap_cache_got_sigbus(f->mmap, f->fd)) |
119 | return -EIO; | |
120 | ||
26687bf8 OS |
121 | f->header->state = STATE_OFFLINE; |
122 | ||
fa6ac760 LP |
123 | if (mmap_cache_got_sigbus(f->mmap, f->fd)) |
124 | return -EIO; | |
125 | ||
26687bf8 OS |
126 | fsync(f->fd); |
127 | ||
128 | return 0; | |
129 | } | |
130 | ||
cec736d2 | 131 | void journal_file_close(JournalFile *f) { |
de190aef | 132 | assert(f); |
cec736d2 | 133 | |
feb12d3e | 134 | #ifdef HAVE_GCRYPT |
b0af6f41 | 135 | /* Write the final tag */ |
c586dbf1 | 136 | if (f->seal && f->writable) |
b0af6f41 | 137 | journal_file_append_tag(f); |
feb12d3e | 138 | #endif |
b0af6f41 | 139 | |
26687bf8 | 140 | journal_file_set_offline(f); |
cec736d2 | 141 | |
fa6ac760 LP |
142 | if (f->mmap && f->fd >= 0) |
143 | mmap_cache_close_fd(f->mmap, f->fd); | |
cec736d2 | 144 | |
11689d2a LP |
145 | if (f->fd >= 0 && f->defrag_on_close) { |
146 | ||
147 | /* Be friendly to btrfs: turn COW back on again now, | |
148 | * and defragment the file. We won't write to the file | |
149 | * ever again, hence remove all fragmentation, and | |
150 | * reenable all the good bits COW usually provides | |
151 | * (such as data checksumming). */ | |
152 | ||
1ed8f8c1 | 153 | (void) chattr_fd(f->fd, 0, FS_NOCOW_FL); |
11689d2a LP |
154 | (void) btrfs_defrag_fd(f->fd); |
155 | } | |
f27a3864 | 156 | |
03e334a1 | 157 | safe_close(f->fd); |
cec736d2 | 158 | free(f->path); |
807e17f0 | 159 | |
16e9f408 LP |
160 | if (f->mmap) |
161 | mmap_cache_unref(f->mmap); | |
162 | ||
4743015d | 163 | ordered_hashmap_free_free(f->chain_cache); |
a4bcff5b | 164 | |
d89c8fdf | 165 | #if defined(HAVE_XZ) || defined(HAVE_LZ4) |
807e17f0 LP |
166 | free(f->compress_buffer); |
167 | #endif | |
168 | ||
7560fffc | 169 | #ifdef HAVE_GCRYPT |
baed47c3 LP |
170 | if (f->fss_file) |
171 | munmap(f->fss_file, PAGE_ALIGN(f->fss_file_size)); | |
b7c9ae91 LP |
172 | else if (f->fsprg_state) |
173 | free(f->fsprg_state); | |
174 | ||
175 | free(f->fsprg_seed); | |
7560fffc LP |
176 | |
177 | if (f->hmac) | |
178 | gcry_md_close(f->hmac); | |
179 | #endif | |
180 | ||
cec736d2 LP |
181 | free(f); |
182 | } | |
183 | ||
0ac38b70 | 184 | static int journal_file_init_header(JournalFile *f, JournalFile *template) { |
d89c8fdf | 185 | Header h = {}; |
cec736d2 LP |
186 | ssize_t k; |
187 | int r; | |
188 | ||
189 | assert(f); | |
190 | ||
7560fffc | 191 | memcpy(h.signature, HEADER_SIGNATURE, 8); |
23b0b2b2 | 192 | h.header_size = htole64(ALIGN64(sizeof(h))); |
cec736d2 | 193 | |
d89c8fdf ZJS |
194 | h.incompatible_flags |= htole32( |
195 | f->compress_xz * HEADER_INCOMPATIBLE_COMPRESSED_XZ | | |
196 | f->compress_lz4 * HEADER_INCOMPATIBLE_COMPRESSED_LZ4); | |
7560fffc | 197 | |
d89c8fdf ZJS |
198 | h.compatible_flags = htole32( |
199 | f->seal * HEADER_COMPATIBLE_SEALED); | |
7560fffc | 200 | |
cec736d2 LP |
201 | r = sd_id128_randomize(&h.file_id); |
202 | if (r < 0) | |
203 | return r; | |
204 | ||
0ac38b70 LP |
205 | if (template) { |
206 | h.seqnum_id = template->header->seqnum_id; | |
beec0085 | 207 | h.tail_entry_seqnum = template->header->tail_entry_seqnum; |
0ac38b70 LP |
208 | } else |
209 | h.seqnum_id = h.file_id; | |
cec736d2 LP |
210 | |
211 | k = pwrite(f->fd, &h, sizeof(h), 0); | |
212 | if (k < 0) | |
213 | return -errno; | |
214 | ||
215 | if (k != sizeof(h)) | |
216 | return -EIO; | |
217 | ||
218 | return 0; | |
219 | } | |
220 | ||
221 | static int journal_file_refresh_header(JournalFile *f) { | |
de190aef | 222 | sd_id128_t boot_id; |
fa6ac760 | 223 | int r; |
cec736d2 LP |
224 | |
225 | assert(f); | |
226 | ||
227 | r = sd_id128_get_machine(&f->header->machine_id); | |
228 | if (r < 0) | |
229 | return r; | |
230 | ||
de190aef | 231 | r = sd_id128_get_boot(&boot_id); |
cec736d2 LP |
232 | if (r < 0) |
233 | return r; | |
234 | ||
de190aef LP |
235 | if (sd_id128_equal(boot_id, f->header->boot_id)) |
236 | f->tail_entry_monotonic_valid = true; | |
237 | ||
238 | f->header->boot_id = boot_id; | |
239 | ||
fa6ac760 | 240 | r = journal_file_set_online(f); |
b788cc23 | 241 | |
7560fffc | 242 | /* Sync the online state to disk */ |
a676e665 | 243 | fsync(f->fd); |
b788cc23 | 244 | |
fa6ac760 | 245 | return r; |
cec736d2 LP |
246 | } |
247 | ||
248 | static int journal_file_verify_header(JournalFile *f) { | |
d89c8fdf ZJS |
249 | uint32_t flags; |
250 | ||
cec736d2 LP |
251 | assert(f); |
252 | ||
7560fffc | 253 | if (memcmp(f->header->signature, HEADER_SIGNATURE, 8)) |
cec736d2 LP |
254 | return -EBADMSG; |
255 | ||
7560fffc LP |
256 | /* In both read and write mode we refuse to open files with |
257 | * incompatible flags we don't know */ | |
d89c8fdf ZJS |
258 | flags = le32toh(f->header->incompatible_flags); |
259 | if (flags & ~HEADER_INCOMPATIBLE_SUPPORTED) { | |
260 | if (flags & ~HEADER_INCOMPATIBLE_ANY) | |
261 | log_debug("Journal file %s has unknown incompatible flags %"PRIx32, | |
262 | f->path, flags & ~HEADER_INCOMPATIBLE_ANY); | |
263 | flags = (flags & HEADER_INCOMPATIBLE_ANY) & ~HEADER_INCOMPATIBLE_SUPPORTED; | |
264 | if (flags) | |
265 | log_debug("Journal file %s uses incompatible flags %"PRIx32 | |
266 | " disabled at compilation time.", f->path, flags); | |
cec736d2 | 267 | return -EPROTONOSUPPORT; |
d89c8fdf | 268 | } |
cec736d2 | 269 | |
7560fffc LP |
270 | /* When open for writing we refuse to open files with |
271 | * compatible flags, too */ | |
d89c8fdf ZJS |
272 | flags = le32toh(f->header->compatible_flags); |
273 | if (f->writable && (flags & ~HEADER_COMPATIBLE_SUPPORTED)) { | |
274 | if (flags & ~HEADER_COMPATIBLE_ANY) | |
275 | log_debug("Journal file %s has unknown compatible flags %"PRIx32, | |
276 | f->path, flags & ~HEADER_COMPATIBLE_ANY); | |
277 | flags = (flags & HEADER_COMPATIBLE_ANY) & ~HEADER_COMPATIBLE_SUPPORTED; | |
278 | if (flags) | |
279 | log_debug("Journal file %s uses compatible flags %"PRIx32 | |
280 | " disabled at compilation time.", f->path, flags); | |
281 | return -EPROTONOSUPPORT; | |
7560fffc LP |
282 | } |
283 | ||
db11ac1a LP |
284 | if (f->header->state >= _STATE_MAX) |
285 | return -EBADMSG; | |
286 | ||
dca6219e LP |
287 | /* The first addition was n_data, so check that we are at least this large */ |
288 | if (le64toh(f->header->header_size) < HEADER_SIZE_MIN) | |
23b0b2b2 LP |
289 | return -EBADMSG; |
290 | ||
8088cbd3 | 291 | if (JOURNAL_HEADER_SEALED(f->header) && !JOURNAL_HEADER_CONTAINS(f->header, n_entry_arrays)) |
beec0085 LP |
292 | return -EBADMSG; |
293 | ||
db11ac1a LP |
294 | if ((le64toh(f->header->header_size) + le64toh(f->header->arena_size)) > (uint64_t) f->last_stat.st_size) |
295 | return -ENODATA; | |
296 | ||
297 | if (le64toh(f->header->tail_object_offset) > (le64toh(f->header->header_size) + le64toh(f->header->arena_size))) | |
298 | return -ENODATA; | |
299 | ||
7762e02b LP |
300 | if (!VALID64(le64toh(f->header->data_hash_table_offset)) || |
301 | !VALID64(le64toh(f->header->field_hash_table_offset)) || | |
302 | !VALID64(le64toh(f->header->tail_object_offset)) || | |
303 | !VALID64(le64toh(f->header->entry_array_offset))) | |
304 | return -ENODATA; | |
305 | ||
cec736d2 | 306 | if (f->writable) { |
ccdbaf91 | 307 | uint8_t state; |
cec736d2 LP |
308 | sd_id128_t machine_id; |
309 | int r; | |
310 | ||
311 | r = sd_id128_get_machine(&machine_id); | |
312 | if (r < 0) | |
313 | return r; | |
314 | ||
315 | if (!sd_id128_equal(machine_id, f->header->machine_id)) | |
316 | return -EHOSTDOWN; | |
317 | ||
de190aef | 318 | state = f->header->state; |
cec736d2 | 319 | |
71fa6f00 LP |
320 | if (state == STATE_ONLINE) { |
321 | log_debug("Journal file %s is already online. Assuming unclean closing.", f->path); | |
322 | return -EBUSY; | |
323 | } else if (state == STATE_ARCHIVED) | |
cec736d2 | 324 | return -ESHUTDOWN; |
71fa6f00 | 325 | else if (state != STATE_OFFLINE) { |
8facc349 | 326 | log_debug("Journal file %s has unknown state %i.", f->path, state); |
71fa6f00 LP |
327 | return -EBUSY; |
328 | } | |
cec736d2 LP |
329 | } |
330 | ||
d89c8fdf ZJS |
331 | f->compress_xz = JOURNAL_HEADER_COMPRESSED_XZ(f->header); |
332 | f->compress_lz4 = JOURNAL_HEADER_COMPRESSED_LZ4(f->header); | |
c586dbf1 | 333 | |
f1889c91 | 334 | f->seal = JOURNAL_HEADER_SEALED(f->header); |
7560fffc | 335 | |
cec736d2 LP |
336 | return 0; |
337 | } | |
338 | ||
2678031a LP |
339 | static int journal_file_fstat(JournalFile *f) { |
340 | assert(f); | |
341 | assert(f->fd >= 0); | |
342 | ||
343 | if (fstat(f->fd, &f->last_stat) < 0) | |
344 | return -errno; | |
345 | ||
346 | f->last_stat_usec = now(CLOCK_MONOTONIC); | |
347 | ||
348 | /* Refuse appending to files that are already deleted */ | |
349 | if (f->last_stat.st_nlink <= 0) | |
350 | return -EIDRM; | |
351 | ||
352 | return 0; | |
353 | } | |
354 | ||
cec736d2 | 355 | static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size) { |
eda4b58b | 356 | uint64_t old_size, new_size; |
fec2aa2f | 357 | int r; |
cec736d2 LP |
358 | |
359 | assert(f); | |
360 | ||
cec736d2 | 361 | /* We assume that this file is not sparse, and we know that |
38ac38b2 | 362 | * for sure, since we always call posix_fallocate() |
cec736d2 LP |
363 | * ourselves */ |
364 | ||
fa6ac760 LP |
365 | if (mmap_cache_got_sigbus(f->mmap, f->fd)) |
366 | return -EIO; | |
367 | ||
cec736d2 | 368 | old_size = |
23b0b2b2 | 369 | le64toh(f->header->header_size) + |
cec736d2 LP |
370 | le64toh(f->header->arena_size); |
371 | ||
bc85bfee | 372 | new_size = PAGE_ALIGN(offset + size); |
23b0b2b2 LP |
373 | if (new_size < le64toh(f->header->header_size)) |
374 | new_size = le64toh(f->header->header_size); | |
bc85bfee | 375 | |
2678031a LP |
376 | if (new_size <= old_size) { |
377 | ||
378 | /* We already pre-allocated enough space, but before | |
379 | * we write to it, let's check with fstat() if the | |
380 | * file got deleted, in order make sure we don't throw | |
381 | * away the data immediately. Don't check fstat() for | |
382 | * all writes though, but only once ever 10s. */ | |
383 | ||
384 | if (f->last_stat_usec + LAST_STAT_REFRESH_USEC > now(CLOCK_MONOTONIC)) | |
385 | return 0; | |
386 | ||
387 | return journal_file_fstat(f); | |
388 | } | |
389 | ||
390 | /* Allocate more space. */ | |
cec736d2 | 391 | |
a676e665 | 392 | if (f->metrics.max_size > 0 && new_size > f->metrics.max_size) |
bc85bfee | 393 | return -E2BIG; |
cec736d2 | 394 | |
a676e665 | 395 | if (new_size > f->metrics.min_size && f->metrics.keep_free > 0) { |
cec736d2 LP |
396 | struct statvfs svfs; |
397 | ||
398 | if (fstatvfs(f->fd, &svfs) >= 0) { | |
399 | uint64_t available; | |
400 | ||
401 | available = svfs.f_bfree * svfs.f_bsize; | |
402 | ||
bc85bfee LP |
403 | if (available >= f->metrics.keep_free) |
404 | available -= f->metrics.keep_free; | |
cec736d2 LP |
405 | else |
406 | available = 0; | |
407 | ||
408 | if (new_size - old_size > available) | |
409 | return -E2BIG; | |
410 | } | |
411 | } | |
412 | ||
eda4b58b LP |
413 | /* Increase by larger blocks at once */ |
414 | new_size = ((new_size+FILE_SIZE_INCREASE-1) / FILE_SIZE_INCREASE) * FILE_SIZE_INCREASE; | |
415 | if (f->metrics.max_size > 0 && new_size > f->metrics.max_size) | |
416 | new_size = f->metrics.max_size; | |
417 | ||
bc85bfee LP |
418 | /* Note that the glibc fallocate() fallback is very |
419 | inefficient, hence we try to minimize the allocation area | |
420 | as we can. */ | |
fec2aa2f GV |
421 | r = posix_fallocate(f->fd, old_size, new_size - old_size); |
422 | if (r != 0) | |
423 | return -r; | |
cec736d2 | 424 | |
23b0b2b2 | 425 | f->header->arena_size = htole64(new_size - le64toh(f->header->header_size)); |
cec736d2 | 426 | |
2678031a | 427 | return journal_file_fstat(f); |
cec736d2 LP |
428 | } |
429 | ||
78519831 | 430 | static unsigned type_to_context(ObjectType type) { |
d3d3208f | 431 | /* One context for each type, plus one catch-all for the rest */ |
69adae51 | 432 | assert_cc(_OBJECT_TYPE_MAX <= MMAP_CACHE_MAX_CONTEXTS); |
fa6ac760 | 433 | assert_cc(CONTEXT_HEADER < MMAP_CACHE_MAX_CONTEXTS); |
d05089d8 | 434 | return type > OBJECT_UNUSED && type < _OBJECT_TYPE_MAX ? type : 0; |
d3d3208f MS |
435 | } |
436 | ||
7a9dabea | 437 | static int journal_file_move_to(JournalFile *f, ObjectType type, bool keep_always, uint64_t offset, uint64_t size, void **ret) { |
2678031a LP |
438 | int r; |
439 | ||
cec736d2 | 440 | assert(f); |
cec736d2 LP |
441 | assert(ret); |
442 | ||
7762e02b LP |
443 | if (size <= 0) |
444 | return -EINVAL; | |
445 | ||
2a59ea54 | 446 | /* Avoid SIGBUS on invalid accesses */ |
4bbdcdb3 LP |
447 | if (offset + size > (uint64_t) f->last_stat.st_size) { |
448 | /* Hmm, out of range? Let's refresh the fstat() data | |
449 | * first, before we trust that check. */ | |
450 | ||
2678031a LP |
451 | r = journal_file_fstat(f); |
452 | if (r < 0) | |
453 | return r; | |
454 | ||
455 | if (offset + size > (uint64_t) f->last_stat.st_size) | |
4bbdcdb3 LP |
456 | return -EADDRNOTAVAIL; |
457 | } | |
458 | ||
7a9dabea | 459 | return mmap_cache_get(f->mmap, f->fd, f->prot, type_to_context(type), keep_always, offset, size, &f->last_stat, ret); |
cec736d2 LP |
460 | } |
461 | ||
16e9f408 LP |
462 | static uint64_t minimum_header_size(Object *o) { |
463 | ||
b8e891e6 | 464 | static const uint64_t table[] = { |
16e9f408 LP |
465 | [OBJECT_DATA] = sizeof(DataObject), |
466 | [OBJECT_FIELD] = sizeof(FieldObject), | |
467 | [OBJECT_ENTRY] = sizeof(EntryObject), | |
468 | [OBJECT_DATA_HASH_TABLE] = sizeof(HashTableObject), | |
469 | [OBJECT_FIELD_HASH_TABLE] = sizeof(HashTableObject), | |
470 | [OBJECT_ENTRY_ARRAY] = sizeof(EntryArrayObject), | |
471 | [OBJECT_TAG] = sizeof(TagObject), | |
472 | }; | |
473 | ||
474 | if (o->object.type >= ELEMENTSOF(table) || table[o->object.type] <= 0) | |
475 | return sizeof(ObjectHeader); | |
476 | ||
477 | return table[o->object.type]; | |
478 | } | |
479 | ||
78519831 | 480 | int journal_file_move_to_object(JournalFile *f, ObjectType type, uint64_t offset, Object **ret) { |
cec736d2 LP |
481 | int r; |
482 | void *t; | |
483 | Object *o; | |
484 | uint64_t s; | |
485 | ||
486 | assert(f); | |
487 | assert(ret); | |
488 | ||
db11ac1a LP |
489 | /* Objects may only be located at multiple of 64 bit */ |
490 | if (!VALID64(offset)) | |
491 | return -EFAULT; | |
492 | ||
7a9dabea | 493 | r = journal_file_move_to(f, type, false, offset, sizeof(ObjectHeader), &t); |
cec736d2 LP |
494 | if (r < 0) |
495 | return r; | |
496 | ||
497 | o = (Object*) t; | |
498 | s = le64toh(o->object.size); | |
499 | ||
500 | if (s < sizeof(ObjectHeader)) | |
501 | return -EBADMSG; | |
502 | ||
16e9f408 LP |
503 | if (o->object.type <= OBJECT_UNUSED) |
504 | return -EBADMSG; | |
505 | ||
506 | if (s < minimum_header_size(o)) | |
507 | return -EBADMSG; | |
508 | ||
d05089d8 | 509 | if (type > OBJECT_UNUSED && o->object.type != type) |
cec736d2 LP |
510 | return -EBADMSG; |
511 | ||
512 | if (s > sizeof(ObjectHeader)) { | |
7a9dabea | 513 | r = journal_file_move_to(f, type, false, offset, s, &t); |
cec736d2 LP |
514 | if (r < 0) |
515 | return r; | |
516 | ||
517 | o = (Object*) t; | |
518 | } | |
519 | ||
cec736d2 LP |
520 | *ret = o; |
521 | return 0; | |
522 | } | |
523 | ||
d98cc1f2 | 524 | static uint64_t journal_file_entry_seqnum(JournalFile *f, uint64_t *seqnum) { |
cec736d2 LP |
525 | uint64_t r; |
526 | ||
527 | assert(f); | |
528 | ||
beec0085 | 529 | r = le64toh(f->header->tail_entry_seqnum) + 1; |
c2373f84 LP |
530 | |
531 | if (seqnum) { | |
de190aef | 532 | /* If an external seqnum counter was passed, we update |
c2373f84 LP |
533 | * both the local and the external one, and set it to |
534 | * the maximum of both */ | |
535 | ||
536 | if (*seqnum + 1 > r) | |
537 | r = *seqnum + 1; | |
538 | ||
539 | *seqnum = r; | |
540 | } | |
541 | ||
beec0085 | 542 | f->header->tail_entry_seqnum = htole64(r); |
cec736d2 | 543 | |
beec0085 LP |
544 | if (f->header->head_entry_seqnum == 0) |
545 | f->header->head_entry_seqnum = htole64(r); | |
de190aef | 546 | |
cec736d2 LP |
547 | return r; |
548 | } | |
549 | ||
78519831 | 550 | int journal_file_append_object(JournalFile *f, ObjectType type, uint64_t size, Object **ret, uint64_t *offset) { |
cec736d2 LP |
551 | int r; |
552 | uint64_t p; | |
553 | Object *tail, *o; | |
554 | void *t; | |
555 | ||
556 | assert(f); | |
d05089d8 | 557 | assert(type > OBJECT_UNUSED && type < _OBJECT_TYPE_MAX); |
cec736d2 LP |
558 | assert(size >= sizeof(ObjectHeader)); |
559 | assert(offset); | |
560 | assert(ret); | |
561 | ||
26687bf8 OS |
562 | r = journal_file_set_online(f); |
563 | if (r < 0) | |
564 | return r; | |
565 | ||
cec736d2 | 566 | p = le64toh(f->header->tail_object_offset); |
cec736d2 | 567 | if (p == 0) |
23b0b2b2 | 568 | p = le64toh(f->header->header_size); |
cec736d2 | 569 | else { |
d05089d8 | 570 | r = journal_file_move_to_object(f, OBJECT_UNUSED, p, &tail); |
cec736d2 LP |
571 | if (r < 0) |
572 | return r; | |
573 | ||
574 | p += ALIGN64(le64toh(tail->object.size)); | |
575 | } | |
576 | ||
577 | r = journal_file_allocate(f, p, size); | |
578 | if (r < 0) | |
579 | return r; | |
580 | ||
fcde2389 | 581 | r = journal_file_move_to(f, type, false, p, size, &t); |
cec736d2 LP |
582 | if (r < 0) |
583 | return r; | |
584 | ||
585 | o = (Object*) t; | |
586 | ||
587 | zero(o->object); | |
de190aef | 588 | o->object.type = type; |
cec736d2 LP |
589 | o->object.size = htole64(size); |
590 | ||
591 | f->header->tail_object_offset = htole64(p); | |
cec736d2 LP |
592 | f->header->n_objects = htole64(le64toh(f->header->n_objects) + 1); |
593 | ||
594 | *ret = o; | |
595 | *offset = p; | |
596 | ||
597 | return 0; | |
598 | } | |
599 | ||
de190aef | 600 | static int journal_file_setup_data_hash_table(JournalFile *f) { |
cec736d2 LP |
601 | uint64_t s, p; |
602 | Object *o; | |
603 | int r; | |
604 | ||
605 | assert(f); | |
606 | ||
dfabe643 | 607 | /* We estimate that we need 1 hash table entry per 768 of |
4a92baf3 LP |
608 | journal file and we want to make sure we never get beyond |
609 | 75% fill level. Calculate the hash table size for the | |
610 | maximum file size based on these metrics. */ | |
611 | ||
dfabe643 | 612 | s = (f->metrics.max_size * 4 / 768 / 3) * sizeof(HashItem); |
4a92baf3 LP |
613 | if (s < DEFAULT_DATA_HASH_TABLE_SIZE) |
614 | s = DEFAULT_DATA_HASH_TABLE_SIZE; | |
615 | ||
507f22bd | 616 | log_debug("Reserving %"PRIu64" entries in hash table.", s / sizeof(HashItem)); |
4a92baf3 | 617 | |
de190aef LP |
618 | r = journal_file_append_object(f, |
619 | OBJECT_DATA_HASH_TABLE, | |
620 | offsetof(Object, hash_table.items) + s, | |
621 | &o, &p); | |
cec736d2 LP |
622 | if (r < 0) |
623 | return r; | |
624 | ||
29804cc1 | 625 | memzero(o->hash_table.items, s); |
cec736d2 | 626 | |
de190aef LP |
627 | f->header->data_hash_table_offset = htole64(p + offsetof(Object, hash_table.items)); |
628 | f->header->data_hash_table_size = htole64(s); | |
cec736d2 LP |
629 | |
630 | return 0; | |
631 | } | |
632 | ||
de190aef | 633 | static int journal_file_setup_field_hash_table(JournalFile *f) { |
cec736d2 LP |
634 | uint64_t s, p; |
635 | Object *o; | |
636 | int r; | |
637 | ||
638 | assert(f); | |
639 | ||
3c1668da LP |
640 | /* We use a fixed size hash table for the fields as this |
641 | * number should grow very slowly only */ | |
642 | ||
de190aef LP |
643 | s = DEFAULT_FIELD_HASH_TABLE_SIZE; |
644 | r = journal_file_append_object(f, | |
645 | OBJECT_FIELD_HASH_TABLE, | |
646 | offsetof(Object, hash_table.items) + s, | |
647 | &o, &p); | |
cec736d2 LP |
648 | if (r < 0) |
649 | return r; | |
650 | ||
29804cc1 | 651 | memzero(o->hash_table.items, s); |
cec736d2 | 652 | |
de190aef LP |
653 | f->header->field_hash_table_offset = htole64(p + offsetof(Object, hash_table.items)); |
654 | f->header->field_hash_table_size = htole64(s); | |
cec736d2 LP |
655 | |
656 | return 0; | |
657 | } | |
658 | ||
de190aef | 659 | static int journal_file_map_data_hash_table(JournalFile *f) { |
cec736d2 LP |
660 | uint64_t s, p; |
661 | void *t; | |
662 | int r; | |
663 | ||
664 | assert(f); | |
665 | ||
de190aef LP |
666 | p = le64toh(f->header->data_hash_table_offset); |
667 | s = le64toh(f->header->data_hash_table_size); | |
cec736d2 | 668 | |
de190aef | 669 | r = journal_file_move_to(f, |
16e9f408 | 670 | OBJECT_DATA_HASH_TABLE, |
fcde2389 | 671 | true, |
de190aef LP |
672 | p, s, |
673 | &t); | |
cec736d2 LP |
674 | if (r < 0) |
675 | return r; | |
676 | ||
de190aef | 677 | f->data_hash_table = t; |
cec736d2 LP |
678 | return 0; |
679 | } | |
680 | ||
de190aef | 681 | static int journal_file_map_field_hash_table(JournalFile *f) { |
cec736d2 LP |
682 | uint64_t s, p; |
683 | void *t; | |
684 | int r; | |
685 | ||
686 | assert(f); | |
687 | ||
de190aef LP |
688 | p = le64toh(f->header->field_hash_table_offset); |
689 | s = le64toh(f->header->field_hash_table_size); | |
cec736d2 | 690 | |
de190aef | 691 | r = journal_file_move_to(f, |
16e9f408 | 692 | OBJECT_FIELD_HASH_TABLE, |
fcde2389 | 693 | true, |
de190aef LP |
694 | p, s, |
695 | &t); | |
cec736d2 LP |
696 | if (r < 0) |
697 | return r; | |
698 | ||
de190aef | 699 | f->field_hash_table = t; |
cec736d2 LP |
700 | return 0; |
701 | } | |
702 | ||
3c1668da LP |
703 | static int journal_file_link_field( |
704 | JournalFile *f, | |
705 | Object *o, | |
706 | uint64_t offset, | |
707 | uint64_t hash) { | |
708 | ||
805d1486 | 709 | uint64_t p, h, m; |
3c1668da LP |
710 | int r; |
711 | ||
712 | assert(f); | |
713 | assert(o); | |
714 | assert(offset > 0); | |
715 | ||
716 | if (o->object.type != OBJECT_FIELD) | |
717 | return -EINVAL; | |
718 | ||
805d1486 LP |
719 | m = le64toh(f->header->field_hash_table_size) / sizeof(HashItem); |
720 | if (m <= 0) | |
721 | return -EBADMSG; | |
3c1668da | 722 | |
805d1486 | 723 | /* This might alter the window we are looking at */ |
3c1668da LP |
724 | o->field.next_hash_offset = o->field.head_data_offset = 0; |
725 | ||
805d1486 | 726 | h = hash % m; |
3c1668da LP |
727 | p = le64toh(f->field_hash_table[h].tail_hash_offset); |
728 | if (p == 0) | |
729 | f->field_hash_table[h].head_hash_offset = htole64(offset); | |
730 | else { | |
731 | r = journal_file_move_to_object(f, OBJECT_FIELD, p, &o); | |
732 | if (r < 0) | |
733 | return r; | |
734 | ||
735 | o->field.next_hash_offset = htole64(offset); | |
736 | } | |
737 | ||
738 | f->field_hash_table[h].tail_hash_offset = htole64(offset); | |
739 | ||
740 | if (JOURNAL_HEADER_CONTAINS(f->header, n_fields)) | |
741 | f->header->n_fields = htole64(le64toh(f->header->n_fields) + 1); | |
742 | ||
743 | return 0; | |
744 | } | |
745 | ||
746 | static int journal_file_link_data( | |
747 | JournalFile *f, | |
748 | Object *o, | |
749 | uint64_t offset, | |
750 | uint64_t hash) { | |
751 | ||
805d1486 | 752 | uint64_t p, h, m; |
cec736d2 LP |
753 | int r; |
754 | ||
755 | assert(f); | |
756 | assert(o); | |
757 | assert(offset > 0); | |
b588975f LP |
758 | |
759 | if (o->object.type != OBJECT_DATA) | |
760 | return -EINVAL; | |
cec736d2 | 761 | |
805d1486 LP |
762 | m = le64toh(f->header->data_hash_table_size) / sizeof(HashItem); |
763 | if (m <= 0) | |
764 | return -EBADMSG; | |
48496df6 | 765 | |
805d1486 | 766 | /* This might alter the window we are looking at */ |
de190aef LP |
767 | o->data.next_hash_offset = o->data.next_field_offset = 0; |
768 | o->data.entry_offset = o->data.entry_array_offset = 0; | |
769 | o->data.n_entries = 0; | |
cec736d2 | 770 | |
805d1486 | 771 | h = hash % m; |
8db4213e | 772 | p = le64toh(f->data_hash_table[h].tail_hash_offset); |
3c1668da | 773 | if (p == 0) |
cec736d2 | 774 | /* Only entry in the hash table is easy */ |
de190aef | 775 | f->data_hash_table[h].head_hash_offset = htole64(offset); |
3c1668da | 776 | else { |
48496df6 LP |
777 | /* Move back to the previous data object, to patch in |
778 | * pointer */ | |
cec736d2 | 779 | |
de190aef | 780 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); |
cec736d2 LP |
781 | if (r < 0) |
782 | return r; | |
783 | ||
de190aef | 784 | o->data.next_hash_offset = htole64(offset); |
cec736d2 LP |
785 | } |
786 | ||
de190aef | 787 | f->data_hash_table[h].tail_hash_offset = htole64(offset); |
cec736d2 | 788 | |
dca6219e LP |
789 | if (JOURNAL_HEADER_CONTAINS(f->header, n_data)) |
790 | f->header->n_data = htole64(le64toh(f->header->n_data) + 1); | |
791 | ||
cec736d2 LP |
792 | return 0; |
793 | } | |
794 | ||
3c1668da LP |
795 | int journal_file_find_field_object_with_hash( |
796 | JournalFile *f, | |
797 | const void *field, uint64_t size, uint64_t hash, | |
798 | Object **ret, uint64_t *offset) { | |
799 | ||
805d1486 | 800 | uint64_t p, osize, h, m; |
3c1668da LP |
801 | int r; |
802 | ||
803 | assert(f); | |
804 | assert(field && size > 0); | |
805 | ||
806 | osize = offsetof(Object, field.payload) + size; | |
807 | ||
805d1486 LP |
808 | m = le64toh(f->header->field_hash_table_size) / sizeof(HashItem); |
809 | ||
810 | if (m <= 0) | |
3c1668da LP |
811 | return -EBADMSG; |
812 | ||
805d1486 | 813 | h = hash % m; |
3c1668da LP |
814 | p = le64toh(f->field_hash_table[h].head_hash_offset); |
815 | ||
816 | while (p > 0) { | |
817 | Object *o; | |
818 | ||
819 | r = journal_file_move_to_object(f, OBJECT_FIELD, p, &o); | |
820 | if (r < 0) | |
821 | return r; | |
822 | ||
823 | if (le64toh(o->field.hash) == hash && | |
824 | le64toh(o->object.size) == osize && | |
825 | memcmp(o->field.payload, field, size) == 0) { | |
826 | ||
827 | if (ret) | |
828 | *ret = o; | |
829 | if (offset) | |
830 | *offset = p; | |
831 | ||
832 | return 1; | |
833 | } | |
834 | ||
835 | p = le64toh(o->field.next_hash_offset); | |
836 | } | |
837 | ||
838 | return 0; | |
839 | } | |
840 | ||
841 | int journal_file_find_field_object( | |
842 | JournalFile *f, | |
843 | const void *field, uint64_t size, | |
844 | Object **ret, uint64_t *offset) { | |
845 | ||
846 | uint64_t hash; | |
847 | ||
848 | assert(f); | |
849 | assert(field && size > 0); | |
850 | ||
851 | hash = hash64(field, size); | |
852 | ||
853 | return journal_file_find_field_object_with_hash(f, | |
854 | field, size, hash, | |
855 | ret, offset); | |
856 | } | |
857 | ||
de190aef LP |
858 | int journal_file_find_data_object_with_hash( |
859 | JournalFile *f, | |
860 | const void *data, uint64_t size, uint64_t hash, | |
861 | Object **ret, uint64_t *offset) { | |
48496df6 | 862 | |
805d1486 | 863 | uint64_t p, osize, h, m; |
cec736d2 LP |
864 | int r; |
865 | ||
866 | assert(f); | |
867 | assert(data || size == 0); | |
868 | ||
869 | osize = offsetof(Object, data.payload) + size; | |
870 | ||
805d1486 LP |
871 | m = le64toh(f->header->data_hash_table_size) / sizeof(HashItem); |
872 | if (m <= 0) | |
bc85bfee LP |
873 | return -EBADMSG; |
874 | ||
805d1486 | 875 | h = hash % m; |
de190aef | 876 | p = le64toh(f->data_hash_table[h].head_hash_offset); |
cec736d2 | 877 | |
de190aef LP |
878 | while (p > 0) { |
879 | Object *o; | |
cec736d2 | 880 | |
de190aef | 881 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); |
cec736d2 LP |
882 | if (r < 0) |
883 | return r; | |
884 | ||
807e17f0 | 885 | if (le64toh(o->data.hash) != hash) |
85a131e8 | 886 | goto next; |
807e17f0 | 887 | |
d89c8fdf | 888 | if (o->object.flags & OBJECT_COMPRESSION_MASK) { |
3b1a55e1 | 889 | #if defined(HAVE_XZ) || defined(HAVE_LZ4) |
fa1c4b51 | 890 | uint64_t l; |
a7f7d1bd | 891 | size_t rsize = 0; |
cec736d2 | 892 | |
807e17f0 LP |
893 | l = le64toh(o->object.size); |
894 | if (l <= offsetof(Object, data.payload)) | |
cec736d2 LP |
895 | return -EBADMSG; |
896 | ||
807e17f0 LP |
897 | l -= offsetof(Object, data.payload); |
898 | ||
d89c8fdf ZJS |
899 | r = decompress_blob(o->object.flags & OBJECT_COMPRESSION_MASK, |
900 | o->data.payload, l, &f->compress_buffer, &f->compress_buffer_size, &rsize, 0); | |
901 | if (r < 0) | |
902 | return r; | |
807e17f0 | 903 | |
b785c858 | 904 | if (rsize == size && |
807e17f0 LP |
905 | memcmp(f->compress_buffer, data, size) == 0) { |
906 | ||
907 | if (ret) | |
908 | *ret = o; | |
909 | ||
910 | if (offset) | |
911 | *offset = p; | |
912 | ||
913 | return 1; | |
914 | } | |
3b1a55e1 ZJS |
915 | #else |
916 | return -EPROTONOSUPPORT; | |
917 | #endif | |
807e17f0 LP |
918 | } else if (le64toh(o->object.size) == osize && |
919 | memcmp(o->data.payload, data, size) == 0) { | |
920 | ||
cec736d2 LP |
921 | if (ret) |
922 | *ret = o; | |
923 | ||
924 | if (offset) | |
925 | *offset = p; | |
926 | ||
de190aef | 927 | return 1; |
cec736d2 LP |
928 | } |
929 | ||
85a131e8 | 930 | next: |
cec736d2 LP |
931 | p = le64toh(o->data.next_hash_offset); |
932 | } | |
933 | ||
de190aef LP |
934 | return 0; |
935 | } | |
936 | ||
937 | int journal_file_find_data_object( | |
938 | JournalFile *f, | |
939 | const void *data, uint64_t size, | |
940 | Object **ret, uint64_t *offset) { | |
941 | ||
942 | uint64_t hash; | |
943 | ||
944 | assert(f); | |
945 | assert(data || size == 0); | |
946 | ||
947 | hash = hash64(data, size); | |
948 | ||
949 | return journal_file_find_data_object_with_hash(f, | |
950 | data, size, hash, | |
951 | ret, offset); | |
952 | } | |
953 | ||
3c1668da LP |
954 | static int journal_file_append_field( |
955 | JournalFile *f, | |
956 | const void *field, uint64_t size, | |
957 | Object **ret, uint64_t *offset) { | |
958 | ||
959 | uint64_t hash, p; | |
960 | uint64_t osize; | |
961 | Object *o; | |
962 | int r; | |
963 | ||
964 | assert(f); | |
965 | assert(field && size > 0); | |
966 | ||
967 | hash = hash64(field, size); | |
968 | ||
969 | r = journal_file_find_field_object_with_hash(f, field, size, hash, &o, &p); | |
970 | if (r < 0) | |
971 | return r; | |
972 | else if (r > 0) { | |
973 | ||
974 | if (ret) | |
975 | *ret = o; | |
976 | ||
977 | if (offset) | |
978 | *offset = p; | |
979 | ||
980 | return 0; | |
981 | } | |
982 | ||
983 | osize = offsetof(Object, field.payload) + size; | |
984 | r = journal_file_append_object(f, OBJECT_FIELD, osize, &o, &p); | |
8c92d4bb LP |
985 | if (r < 0) |
986 | return r; | |
3c1668da LP |
987 | |
988 | o->field.hash = htole64(hash); | |
989 | memcpy(o->field.payload, field, size); | |
990 | ||
991 | r = journal_file_link_field(f, o, p, hash); | |
992 | if (r < 0) | |
993 | return r; | |
994 | ||
995 | /* The linking might have altered the window, so let's | |
996 | * refresh our pointer */ | |
997 | r = journal_file_move_to_object(f, OBJECT_FIELD, p, &o); | |
998 | if (r < 0) | |
999 | return r; | |
1000 | ||
1001 | #ifdef HAVE_GCRYPT | |
1002 | r = journal_file_hmac_put_object(f, OBJECT_FIELD, o, p); | |
1003 | if (r < 0) | |
1004 | return r; | |
1005 | #endif | |
1006 | ||
1007 | if (ret) | |
1008 | *ret = o; | |
1009 | ||
1010 | if (offset) | |
1011 | *offset = p; | |
1012 | ||
1013 | return 0; | |
1014 | } | |
1015 | ||
48496df6 LP |
1016 | static int journal_file_append_data( |
1017 | JournalFile *f, | |
1018 | const void *data, uint64_t size, | |
1019 | Object **ret, uint64_t *offset) { | |
1020 | ||
de190aef LP |
1021 | uint64_t hash, p; |
1022 | uint64_t osize; | |
1023 | Object *o; | |
d89c8fdf | 1024 | int r, compression = 0; |
3c1668da | 1025 | const void *eq; |
de190aef LP |
1026 | |
1027 | assert(f); | |
1028 | assert(data || size == 0); | |
1029 | ||
1030 | hash = hash64(data, size); | |
1031 | ||
1032 | r = journal_file_find_data_object_with_hash(f, data, size, hash, &o, &p); | |
1033 | if (r < 0) | |
1034 | return r; | |
1035 | else if (r > 0) { | |
1036 | ||
1037 | if (ret) | |
1038 | *ret = o; | |
1039 | ||
1040 | if (offset) | |
1041 | *offset = p; | |
1042 | ||
1043 | return 0; | |
1044 | } | |
1045 | ||
1046 | osize = offsetof(Object, data.payload) + size; | |
1047 | r = journal_file_append_object(f, OBJECT_DATA, osize, &o, &p); | |
cec736d2 LP |
1048 | if (r < 0) |
1049 | return r; | |
1050 | ||
cec736d2 | 1051 | o->data.hash = htole64(hash); |
807e17f0 | 1052 | |
d89c8fdf ZJS |
1053 | #if defined(HAVE_XZ) || defined(HAVE_LZ4) |
1054 | if (f->compress_xz && | |
807e17f0 | 1055 | size >= COMPRESSION_SIZE_THRESHOLD) { |
a7f7d1bd | 1056 | size_t rsize = 0; |
807e17f0 | 1057 | |
d89c8fdf | 1058 | compression = compress_blob(data, size, o->data.payload, &rsize); |
807e17f0 | 1059 | |
d89c8fdf | 1060 | if (compression) { |
807e17f0 | 1061 | o->object.size = htole64(offsetof(Object, data.payload) + rsize); |
d89c8fdf | 1062 | o->object.flags |= compression; |
807e17f0 | 1063 | |
fa1c4b51 | 1064 | log_debug("Compressed data object %"PRIu64" -> %zu using %s", |
d89c8fdf | 1065 | size, rsize, object_compressed_to_string(compression)); |
807e17f0 LP |
1066 | } |
1067 | } | |
1068 | #endif | |
1069 | ||
d89c8fdf | 1070 | if (!compression && size > 0) |
807e17f0 | 1071 | memcpy(o->data.payload, data, size); |
cec736d2 | 1072 | |
de190aef | 1073 | r = journal_file_link_data(f, o, p, hash); |
cec736d2 LP |
1074 | if (r < 0) |
1075 | return r; | |
1076 | ||
48496df6 LP |
1077 | /* The linking might have altered the window, so let's |
1078 | * refresh our pointer */ | |
1079 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); | |
1080 | if (r < 0) | |
1081 | return r; | |
1082 | ||
08c6f819 SL |
1083 | if (!data) |
1084 | eq = NULL; | |
1085 | else | |
1086 | eq = memchr(data, '=', size); | |
3c1668da | 1087 | if (eq && eq > data) { |
748db592 | 1088 | Object *fo = NULL; |
3c1668da | 1089 | uint64_t fp; |
3c1668da LP |
1090 | |
1091 | /* Create field object ... */ | |
1092 | r = journal_file_append_field(f, data, (uint8_t*) eq - (uint8_t*) data, &fo, &fp); | |
1093 | if (r < 0) | |
1094 | return r; | |
1095 | ||
1096 | /* ... and link it in. */ | |
1097 | o->data.next_field_offset = fo->field.head_data_offset; | |
1098 | fo->field.head_data_offset = le64toh(p); | |
1099 | } | |
1100 | ||
5996c7c2 LP |
1101 | #ifdef HAVE_GCRYPT |
1102 | r = journal_file_hmac_put_object(f, OBJECT_DATA, o, p); | |
1103 | if (r < 0) | |
1104 | return r; | |
1105 | #endif | |
1106 | ||
cec736d2 LP |
1107 | if (ret) |
1108 | *ret = o; | |
1109 | ||
1110 | if (offset) | |
de190aef | 1111 | *offset = p; |
cec736d2 LP |
1112 | |
1113 | return 0; | |
1114 | } | |
1115 | ||
1116 | uint64_t journal_file_entry_n_items(Object *o) { | |
1117 | assert(o); | |
b588975f LP |
1118 | |
1119 | if (o->object.type != OBJECT_ENTRY) | |
1120 | return 0; | |
cec736d2 LP |
1121 | |
1122 | return (le64toh(o->object.size) - offsetof(Object, entry.items)) / sizeof(EntryItem); | |
1123 | } | |
1124 | ||
0284adc6 | 1125 | uint64_t journal_file_entry_array_n_items(Object *o) { |
de190aef | 1126 | assert(o); |
b588975f LP |
1127 | |
1128 | if (o->object.type != OBJECT_ENTRY_ARRAY) | |
1129 | return 0; | |
de190aef LP |
1130 | |
1131 | return (le64toh(o->object.size) - offsetof(Object, entry_array.items)) / sizeof(uint64_t); | |
1132 | } | |
1133 | ||
fb9a24b6 LP |
1134 | uint64_t journal_file_hash_table_n_items(Object *o) { |
1135 | assert(o); | |
b588975f LP |
1136 | |
1137 | if (o->object.type != OBJECT_DATA_HASH_TABLE && | |
1138 | o->object.type != OBJECT_FIELD_HASH_TABLE) | |
1139 | return 0; | |
fb9a24b6 LP |
1140 | |
1141 | return (le64toh(o->object.size) - offsetof(Object, hash_table.items)) / sizeof(HashItem); | |
1142 | } | |
1143 | ||
de190aef | 1144 | static int link_entry_into_array(JournalFile *f, |
4fd052ae FC |
1145 | le64_t *first, |
1146 | le64_t *idx, | |
de190aef | 1147 | uint64_t p) { |
cec736d2 | 1148 | int r; |
de190aef LP |
1149 | uint64_t n = 0, ap = 0, q, i, a, hidx; |
1150 | Object *o; | |
1151 | ||
cec736d2 | 1152 | assert(f); |
de190aef LP |
1153 | assert(first); |
1154 | assert(idx); | |
1155 | assert(p > 0); | |
cec736d2 | 1156 | |
de190aef LP |
1157 | a = le64toh(*first); |
1158 | i = hidx = le64toh(*idx); | |
1159 | while (a > 0) { | |
1160 | ||
1161 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &o); | |
1162 | if (r < 0) | |
1163 | return r; | |
cec736d2 | 1164 | |
de190aef LP |
1165 | n = journal_file_entry_array_n_items(o); |
1166 | if (i < n) { | |
1167 | o->entry_array.items[i] = htole64(p); | |
1168 | *idx = htole64(hidx + 1); | |
1169 | return 0; | |
1170 | } | |
cec736d2 | 1171 | |
de190aef LP |
1172 | i -= n; |
1173 | ap = a; | |
1174 | a = le64toh(o->entry_array.next_entry_array_offset); | |
1175 | } | |
1176 | ||
1177 | if (hidx > n) | |
1178 | n = (hidx+1) * 2; | |
1179 | else | |
1180 | n = n * 2; | |
1181 | ||
1182 | if (n < 4) | |
1183 | n = 4; | |
1184 | ||
1185 | r = journal_file_append_object(f, OBJECT_ENTRY_ARRAY, | |
1186 | offsetof(Object, entry_array.items) + n * sizeof(uint64_t), | |
1187 | &o, &q); | |
cec736d2 LP |
1188 | if (r < 0) |
1189 | return r; | |
1190 | ||
feb12d3e | 1191 | #ifdef HAVE_GCRYPT |
5996c7c2 | 1192 | r = journal_file_hmac_put_object(f, OBJECT_ENTRY_ARRAY, o, q); |
b0af6f41 LP |
1193 | if (r < 0) |
1194 | return r; | |
feb12d3e | 1195 | #endif |
b0af6f41 | 1196 | |
de190aef | 1197 | o->entry_array.items[i] = htole64(p); |
cec736d2 | 1198 | |
de190aef | 1199 | if (ap == 0) |
7be3aa17 | 1200 | *first = htole64(q); |
cec736d2 | 1201 | else { |
de190aef | 1202 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, ap, &o); |
cec736d2 LP |
1203 | if (r < 0) |
1204 | return r; | |
1205 | ||
de190aef LP |
1206 | o->entry_array.next_entry_array_offset = htole64(q); |
1207 | } | |
cec736d2 | 1208 | |
2dee23eb LP |
1209 | if (JOURNAL_HEADER_CONTAINS(f->header, n_entry_arrays)) |
1210 | f->header->n_entry_arrays = htole64(le64toh(f->header->n_entry_arrays) + 1); | |
1211 | ||
de190aef LP |
1212 | *idx = htole64(hidx + 1); |
1213 | ||
1214 | return 0; | |
1215 | } | |
cec736d2 | 1216 | |
de190aef | 1217 | static int link_entry_into_array_plus_one(JournalFile *f, |
4fd052ae FC |
1218 | le64_t *extra, |
1219 | le64_t *first, | |
1220 | le64_t *idx, | |
de190aef LP |
1221 | uint64_t p) { |
1222 | ||
1223 | int r; | |
1224 | ||
1225 | assert(f); | |
1226 | assert(extra); | |
1227 | assert(first); | |
1228 | assert(idx); | |
1229 | assert(p > 0); | |
1230 | ||
1231 | if (*idx == 0) | |
1232 | *extra = htole64(p); | |
1233 | else { | |
4fd052ae | 1234 | le64_t i; |
de190aef | 1235 | |
7be3aa17 | 1236 | i = htole64(le64toh(*idx) - 1); |
de190aef LP |
1237 | r = link_entry_into_array(f, first, &i, p); |
1238 | if (r < 0) | |
1239 | return r; | |
cec736d2 LP |
1240 | } |
1241 | ||
de190aef LP |
1242 | *idx = htole64(le64toh(*idx) + 1); |
1243 | return 0; | |
1244 | } | |
1245 | ||
1246 | static int journal_file_link_entry_item(JournalFile *f, Object *o, uint64_t offset, uint64_t i) { | |
1247 | uint64_t p; | |
1248 | int r; | |
1249 | assert(f); | |
1250 | assert(o); | |
1251 | assert(offset > 0); | |
1252 | ||
1253 | p = le64toh(o->entry.items[i].object_offset); | |
1254 | if (p == 0) | |
1255 | return -EINVAL; | |
1256 | ||
1257 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); | |
cec736d2 LP |
1258 | if (r < 0) |
1259 | return r; | |
1260 | ||
de190aef LP |
1261 | return link_entry_into_array_plus_one(f, |
1262 | &o->data.entry_offset, | |
1263 | &o->data.entry_array_offset, | |
1264 | &o->data.n_entries, | |
1265 | offset); | |
cec736d2 LP |
1266 | } |
1267 | ||
1268 | static int journal_file_link_entry(JournalFile *f, Object *o, uint64_t offset) { | |
de190aef | 1269 | uint64_t n, i; |
cec736d2 LP |
1270 | int r; |
1271 | ||
1272 | assert(f); | |
1273 | assert(o); | |
1274 | assert(offset > 0); | |
b588975f LP |
1275 | |
1276 | if (o->object.type != OBJECT_ENTRY) | |
1277 | return -EINVAL; | |
cec736d2 | 1278 | |
b788cc23 LP |
1279 | __sync_synchronize(); |
1280 | ||
cec736d2 | 1281 | /* Link up the entry itself */ |
de190aef LP |
1282 | r = link_entry_into_array(f, |
1283 | &f->header->entry_array_offset, | |
1284 | &f->header->n_entries, | |
1285 | offset); | |
1286 | if (r < 0) | |
1287 | return r; | |
cec736d2 | 1288 | |
507f22bd | 1289 | /* log_debug("=> %s seqnr=%"PRIu64" n_entries=%"PRIu64, f->path, o->entry.seqnum, f->header->n_entries); */ |
cec736d2 | 1290 | |
de190aef | 1291 | if (f->header->head_entry_realtime == 0) |
0ac38b70 | 1292 | f->header->head_entry_realtime = o->entry.realtime; |
cec736d2 | 1293 | |
0ac38b70 | 1294 | f->header->tail_entry_realtime = o->entry.realtime; |
de190aef LP |
1295 | f->header->tail_entry_monotonic = o->entry.monotonic; |
1296 | ||
1297 | f->tail_entry_monotonic_valid = true; | |
cec736d2 LP |
1298 | |
1299 | /* Link up the items */ | |
1300 | n = journal_file_entry_n_items(o); | |
1301 | for (i = 0; i < n; i++) { | |
1302 | r = journal_file_link_entry_item(f, o, offset, i); | |
1303 | if (r < 0) | |
1304 | return r; | |
1305 | } | |
1306 | ||
cec736d2 LP |
1307 | return 0; |
1308 | } | |
1309 | ||
1310 | static int journal_file_append_entry_internal( | |
1311 | JournalFile *f, | |
1312 | const dual_timestamp *ts, | |
1313 | uint64_t xor_hash, | |
1314 | const EntryItem items[], unsigned n_items, | |
de190aef | 1315 | uint64_t *seqnum, |
cec736d2 LP |
1316 | Object **ret, uint64_t *offset) { |
1317 | uint64_t np; | |
1318 | uint64_t osize; | |
1319 | Object *o; | |
1320 | int r; | |
1321 | ||
1322 | assert(f); | |
1323 | assert(items || n_items == 0); | |
de190aef | 1324 | assert(ts); |
cec736d2 LP |
1325 | |
1326 | osize = offsetof(Object, entry.items) + (n_items * sizeof(EntryItem)); | |
1327 | ||
de190aef | 1328 | r = journal_file_append_object(f, OBJECT_ENTRY, osize, &o, &np); |
cec736d2 LP |
1329 | if (r < 0) |
1330 | return r; | |
1331 | ||
d98cc1f2 | 1332 | o->entry.seqnum = htole64(journal_file_entry_seqnum(f, seqnum)); |
cec736d2 | 1333 | memcpy(o->entry.items, items, n_items * sizeof(EntryItem)); |
de190aef LP |
1334 | o->entry.realtime = htole64(ts->realtime); |
1335 | o->entry.monotonic = htole64(ts->monotonic); | |
cec736d2 LP |
1336 | o->entry.xor_hash = htole64(xor_hash); |
1337 | o->entry.boot_id = f->header->boot_id; | |
1338 | ||
feb12d3e | 1339 | #ifdef HAVE_GCRYPT |
5996c7c2 | 1340 | r = journal_file_hmac_put_object(f, OBJECT_ENTRY, o, np); |
b0af6f41 LP |
1341 | if (r < 0) |
1342 | return r; | |
feb12d3e | 1343 | #endif |
b0af6f41 | 1344 | |
cec736d2 LP |
1345 | r = journal_file_link_entry(f, o, np); |
1346 | if (r < 0) | |
1347 | return r; | |
1348 | ||
1349 | if (ret) | |
1350 | *ret = o; | |
1351 | ||
1352 | if (offset) | |
1353 | *offset = np; | |
1354 | ||
1355 | return 0; | |
1356 | } | |
1357 | ||
cf244689 | 1358 | void journal_file_post_change(JournalFile *f) { |
50f20cfd LP |
1359 | assert(f); |
1360 | ||
1361 | /* inotify() does not receive IN_MODIFY events from file | |
1362 | * accesses done via mmap(). After each access we hence | |
1363 | * trigger IN_MODIFY by truncating the journal file to its | |
1364 | * current size which triggers IN_MODIFY. */ | |
1365 | ||
bc85bfee LP |
1366 | __sync_synchronize(); |
1367 | ||
50f20cfd | 1368 | if (ftruncate(f->fd, f->last_stat.st_size) < 0) |
56f64d95 | 1369 | log_error_errno(errno, "Failed to truncate file to its own size: %m"); |
50f20cfd LP |
1370 | } |
1371 | ||
1f2da9ec LP |
1372 | static int entry_item_cmp(const void *_a, const void *_b) { |
1373 | const EntryItem *a = _a, *b = _b; | |
1374 | ||
1375 | if (le64toh(a->object_offset) < le64toh(b->object_offset)) | |
1376 | return -1; | |
1377 | if (le64toh(a->object_offset) > le64toh(b->object_offset)) | |
1378 | return 1; | |
1379 | return 0; | |
1380 | } | |
1381 | ||
de190aef | 1382 | 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 |
1383 | unsigned i; |
1384 | EntryItem *items; | |
1385 | int r; | |
1386 | uint64_t xor_hash = 0; | |
de190aef | 1387 | struct dual_timestamp _ts; |
cec736d2 LP |
1388 | |
1389 | assert(f); | |
1390 | assert(iovec || n_iovec == 0); | |
1391 | ||
de190aef LP |
1392 | if (!ts) { |
1393 | dual_timestamp_get(&_ts); | |
1394 | ts = &_ts; | |
1395 | } | |
1396 | ||
1397 | if (f->tail_entry_monotonic_valid && | |
1398 | ts->monotonic < le64toh(f->header->tail_entry_monotonic)) | |
1399 | return -EINVAL; | |
1400 | ||
feb12d3e | 1401 | #ifdef HAVE_GCRYPT |
7560fffc LP |
1402 | r = journal_file_maybe_append_tag(f, ts->realtime); |
1403 | if (r < 0) | |
1404 | return r; | |
feb12d3e | 1405 | #endif |
7560fffc | 1406 | |
64825d3c | 1407 | /* alloca() can't take 0, hence let's allocate at least one */ |
9607d947 | 1408 | items = alloca(sizeof(EntryItem) * MAX(1u, n_iovec)); |
cec736d2 LP |
1409 | |
1410 | for (i = 0; i < n_iovec; i++) { | |
1411 | uint64_t p; | |
1412 | Object *o; | |
1413 | ||
1414 | r = journal_file_append_data(f, iovec[i].iov_base, iovec[i].iov_len, &o, &p); | |
1415 | if (r < 0) | |
cf244689 | 1416 | return r; |
cec736d2 LP |
1417 | |
1418 | xor_hash ^= le64toh(o->data.hash); | |
1419 | items[i].object_offset = htole64(p); | |
de7b95cd | 1420 | items[i].hash = o->data.hash; |
cec736d2 LP |
1421 | } |
1422 | ||
1f2da9ec LP |
1423 | /* Order by the position on disk, in order to improve seek |
1424 | * times for rotating media. */ | |
7ff7394d | 1425 | qsort_safe(items, n_iovec, sizeof(EntryItem), entry_item_cmp); |
1f2da9ec | 1426 | |
de190aef | 1427 | r = journal_file_append_entry_internal(f, ts, xor_hash, items, n_iovec, seqnum, ret, offset); |
cec736d2 | 1428 | |
fa6ac760 LP |
1429 | /* If the memory mapping triggered a SIGBUS then we return an |
1430 | * IO error and ignore the error code passed down to us, since | |
1431 | * it is very likely just an effect of a nullified replacement | |
1432 | * mapping page */ | |
1433 | ||
1434 | if (mmap_cache_got_sigbus(f->mmap, f->fd)) | |
1435 | r = -EIO; | |
1436 | ||
50f20cfd LP |
1437 | journal_file_post_change(f); |
1438 | ||
cec736d2 LP |
1439 | return r; |
1440 | } | |
1441 | ||
a4bcff5b | 1442 | typedef struct ChainCacheItem { |
fb099c8d | 1443 | uint64_t first; /* the array at the beginning of the chain */ |
a4bcff5b LP |
1444 | uint64_t array; /* the cached array */ |
1445 | uint64_t begin; /* the first item in the cached array */ | |
1446 | uint64_t total; /* the total number of items in all arrays before this one in the chain */ | |
f268980d | 1447 | uint64_t last_index; /* the last index we looked at, to optimize locality when bisecting */ |
a4bcff5b LP |
1448 | } ChainCacheItem; |
1449 | ||
1450 | static void chain_cache_put( | |
4743015d | 1451 | OrderedHashmap *h, |
a4bcff5b LP |
1452 | ChainCacheItem *ci, |
1453 | uint64_t first, | |
1454 | uint64_t array, | |
1455 | uint64_t begin, | |
f268980d LP |
1456 | uint64_t total, |
1457 | uint64_t last_index) { | |
a4bcff5b LP |
1458 | |
1459 | if (!ci) { | |
34741aa3 LP |
1460 | /* If the chain item to cache for this chain is the |
1461 | * first one it's not worth caching anything */ | |
1462 | if (array == first) | |
1463 | return; | |
1464 | ||
29433089 | 1465 | if (ordered_hashmap_size(h) >= CHAIN_CACHE_MAX) { |
4743015d | 1466 | ci = ordered_hashmap_steal_first(h); |
29433089 LP |
1467 | assert(ci); |
1468 | } else { | |
a4bcff5b LP |
1469 | ci = new(ChainCacheItem, 1); |
1470 | if (!ci) | |
1471 | return; | |
1472 | } | |
1473 | ||
1474 | ci->first = first; | |
1475 | ||
4743015d | 1476 | if (ordered_hashmap_put(h, &ci->first, ci) < 0) { |
a4bcff5b LP |
1477 | free(ci); |
1478 | return; | |
1479 | } | |
1480 | } else | |
1481 | assert(ci->first == first); | |
1482 | ||
1483 | ci->array = array; | |
1484 | ci->begin = begin; | |
1485 | ci->total = total; | |
f268980d | 1486 | ci->last_index = last_index; |
a4bcff5b LP |
1487 | } |
1488 | ||
f268980d LP |
1489 | static int generic_array_get( |
1490 | JournalFile *f, | |
1491 | uint64_t first, | |
1492 | uint64_t i, | |
1493 | Object **ret, uint64_t *offset) { | |
de190aef | 1494 | |
cec736d2 | 1495 | Object *o; |
a4bcff5b | 1496 | uint64_t p = 0, a, t = 0; |
cec736d2 | 1497 | int r; |
a4bcff5b | 1498 | ChainCacheItem *ci; |
cec736d2 LP |
1499 | |
1500 | assert(f); | |
1501 | ||
de190aef | 1502 | a = first; |
a4bcff5b LP |
1503 | |
1504 | /* Try the chain cache first */ | |
4743015d | 1505 | ci = ordered_hashmap_get(f->chain_cache, &first); |
a4bcff5b LP |
1506 | if (ci && i > ci->total) { |
1507 | a = ci->array; | |
1508 | i -= ci->total; | |
1509 | t = ci->total; | |
1510 | } | |
1511 | ||
de190aef | 1512 | while (a > 0) { |
a4bcff5b | 1513 | uint64_t k; |
cec736d2 | 1514 | |
de190aef LP |
1515 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &o); |
1516 | if (r < 0) | |
1517 | return r; | |
cec736d2 | 1518 | |
a4bcff5b LP |
1519 | k = journal_file_entry_array_n_items(o); |
1520 | if (i < k) { | |
de190aef | 1521 | p = le64toh(o->entry_array.items[i]); |
a4bcff5b | 1522 | goto found; |
cec736d2 LP |
1523 | } |
1524 | ||
a4bcff5b LP |
1525 | i -= k; |
1526 | t += k; | |
de190aef LP |
1527 | a = le64toh(o->entry_array.next_entry_array_offset); |
1528 | } | |
1529 | ||
a4bcff5b LP |
1530 | return 0; |
1531 | ||
1532 | found: | |
1533 | /* Let's cache this item for the next invocation */ | |
af13a6b0 | 1534 | chain_cache_put(f->chain_cache, ci, first, a, le64toh(o->entry_array.items[0]), t, i); |
de190aef LP |
1535 | |
1536 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1537 | if (r < 0) | |
1538 | return r; | |
1539 | ||
1540 | if (ret) | |
1541 | *ret = o; | |
1542 | ||
1543 | if (offset) | |
1544 | *offset = p; | |
1545 | ||
1546 | return 1; | |
1547 | } | |
1548 | ||
f268980d LP |
1549 | static int generic_array_get_plus_one( |
1550 | JournalFile *f, | |
1551 | uint64_t extra, | |
1552 | uint64_t first, | |
1553 | uint64_t i, | |
1554 | Object **ret, uint64_t *offset) { | |
de190aef LP |
1555 | |
1556 | Object *o; | |
1557 | ||
1558 | assert(f); | |
1559 | ||
1560 | if (i == 0) { | |
1561 | int r; | |
1562 | ||
1563 | r = journal_file_move_to_object(f, OBJECT_ENTRY, extra, &o); | |
cec736d2 LP |
1564 | if (r < 0) |
1565 | return r; | |
1566 | ||
de190aef LP |
1567 | if (ret) |
1568 | *ret = o; | |
cec736d2 | 1569 | |
de190aef LP |
1570 | if (offset) |
1571 | *offset = extra; | |
cec736d2 | 1572 | |
de190aef | 1573 | return 1; |
cec736d2 LP |
1574 | } |
1575 | ||
de190aef LP |
1576 | return generic_array_get(f, first, i-1, ret, offset); |
1577 | } | |
cec736d2 | 1578 | |
de190aef LP |
1579 | enum { |
1580 | TEST_FOUND, | |
1581 | TEST_LEFT, | |
1582 | TEST_RIGHT | |
1583 | }; | |
cec736d2 | 1584 | |
f268980d LP |
1585 | static int generic_array_bisect( |
1586 | JournalFile *f, | |
1587 | uint64_t first, | |
1588 | uint64_t n, | |
1589 | uint64_t needle, | |
1590 | int (*test_object)(JournalFile *f, uint64_t p, uint64_t needle), | |
1591 | direction_t direction, | |
1592 | Object **ret, | |
1593 | uint64_t *offset, | |
1594 | uint64_t *idx) { | |
1595 | ||
1596 | uint64_t a, p, t = 0, i = 0, last_p = 0, last_index = (uint64_t) -1; | |
de190aef LP |
1597 | bool subtract_one = false; |
1598 | Object *o, *array = NULL; | |
1599 | int r; | |
a4bcff5b | 1600 | ChainCacheItem *ci; |
cec736d2 | 1601 | |
de190aef LP |
1602 | assert(f); |
1603 | assert(test_object); | |
cec736d2 | 1604 | |
a4bcff5b | 1605 | /* Start with the first array in the chain */ |
de190aef | 1606 | a = first; |
a4bcff5b | 1607 | |
4743015d | 1608 | ci = ordered_hashmap_get(f->chain_cache, &first); |
a4bcff5b LP |
1609 | if (ci && n > ci->total) { |
1610 | /* Ah, we have iterated this bisection array chain | |
1611 | * previously! Let's see if we can skip ahead in the | |
1612 | * chain, as far as the last time. But we can't jump | |
1613 | * backwards in the chain, so let's check that | |
1614 | * first. */ | |
1615 | ||
1616 | r = test_object(f, ci->begin, needle); | |
1617 | if (r < 0) | |
1618 | return r; | |
1619 | ||
1620 | if (r == TEST_LEFT) { | |
f268980d | 1621 | /* OK, what we are looking for is right of the |
a4bcff5b LP |
1622 | * begin of this EntryArray, so let's jump |
1623 | * straight to previously cached array in the | |
1624 | * chain */ | |
1625 | ||
1626 | a = ci->array; | |
1627 | n -= ci->total; | |
1628 | t = ci->total; | |
f268980d | 1629 | last_index = ci->last_index; |
a4bcff5b LP |
1630 | } |
1631 | } | |
1632 | ||
de190aef LP |
1633 | while (a > 0) { |
1634 | uint64_t left, right, k, lp; | |
1635 | ||
1636 | r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &array); | |
cec736d2 LP |
1637 | if (r < 0) |
1638 | return r; | |
1639 | ||
de190aef LP |
1640 | k = journal_file_entry_array_n_items(array); |
1641 | right = MIN(k, n); | |
1642 | if (right <= 0) | |
1643 | return 0; | |
cec736d2 | 1644 | |
de190aef LP |
1645 | i = right - 1; |
1646 | lp = p = le64toh(array->entry_array.items[i]); | |
1647 | if (p <= 0) | |
1648 | return -EBADMSG; | |
cec736d2 | 1649 | |
de190aef LP |
1650 | r = test_object(f, p, needle); |
1651 | if (r < 0) | |
1652 | return r; | |
cec736d2 | 1653 | |
de190aef LP |
1654 | if (r == TEST_FOUND) |
1655 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1656 | ||
1657 | if (r == TEST_RIGHT) { | |
1658 | left = 0; | |
1659 | right -= 1; | |
f268980d LP |
1660 | |
1661 | if (last_index != (uint64_t) -1) { | |
1662 | assert(last_index <= right); | |
1663 | ||
1664 | /* If we cached the last index we | |
1665 | * looked at, let's try to not to jump | |
1666 | * too wildly around and see if we can | |
1667 | * limit the range to look at early to | |
1668 | * the immediate neighbors of the last | |
1669 | * index we looked at. */ | |
1670 | ||
1671 | if (last_index > 0) { | |
1672 | uint64_t x = last_index - 1; | |
1673 | ||
1674 | p = le64toh(array->entry_array.items[x]); | |
1675 | if (p <= 0) | |
1676 | return -EBADMSG; | |
1677 | ||
1678 | r = test_object(f, p, needle); | |
1679 | if (r < 0) | |
1680 | return r; | |
1681 | ||
1682 | if (r == TEST_FOUND) | |
1683 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1684 | ||
1685 | if (r == TEST_RIGHT) | |
1686 | right = x; | |
1687 | else | |
1688 | left = x + 1; | |
1689 | } | |
1690 | ||
1691 | if (last_index < right) { | |
1692 | uint64_t y = last_index + 1; | |
1693 | ||
1694 | p = le64toh(array->entry_array.items[y]); | |
1695 | if (p <= 0) | |
1696 | return -EBADMSG; | |
1697 | ||
1698 | r = test_object(f, p, needle); | |
1699 | if (r < 0) | |
1700 | return r; | |
1701 | ||
1702 | if (r == TEST_FOUND) | |
1703 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1704 | ||
1705 | if (r == TEST_RIGHT) | |
1706 | right = y; | |
1707 | else | |
1708 | left = y + 1; | |
1709 | } | |
f268980d LP |
1710 | } |
1711 | ||
de190aef LP |
1712 | for (;;) { |
1713 | if (left == right) { | |
1714 | if (direction == DIRECTION_UP) | |
1715 | subtract_one = true; | |
1716 | ||
1717 | i = left; | |
1718 | goto found; | |
1719 | } | |
1720 | ||
1721 | assert(left < right); | |
de190aef | 1722 | i = (left + right) / 2; |
f268980d | 1723 | |
de190aef LP |
1724 | p = le64toh(array->entry_array.items[i]); |
1725 | if (p <= 0) | |
1726 | return -EBADMSG; | |
1727 | ||
1728 | r = test_object(f, p, needle); | |
1729 | if (r < 0) | |
1730 | return r; | |
cec736d2 | 1731 | |
de190aef LP |
1732 | if (r == TEST_FOUND) |
1733 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1734 | ||
1735 | if (r == TEST_RIGHT) | |
1736 | right = i; | |
1737 | else | |
1738 | left = i + 1; | |
1739 | } | |
1740 | } | |
1741 | ||
2173cbf8 | 1742 | if (k >= n) { |
cbdca852 LP |
1743 | if (direction == DIRECTION_UP) { |
1744 | i = n; | |
1745 | subtract_one = true; | |
1746 | goto found; | |
1747 | } | |
1748 | ||
cec736d2 | 1749 | return 0; |
cbdca852 | 1750 | } |
cec736d2 | 1751 | |
de190aef LP |
1752 | last_p = lp; |
1753 | ||
1754 | n -= k; | |
1755 | t += k; | |
f268980d | 1756 | last_index = (uint64_t) -1; |
de190aef | 1757 | a = le64toh(array->entry_array.next_entry_array_offset); |
cec736d2 LP |
1758 | } |
1759 | ||
1760 | return 0; | |
de190aef LP |
1761 | |
1762 | found: | |
1763 | if (subtract_one && t == 0 && i == 0) | |
1764 | return 0; | |
1765 | ||
a4bcff5b | 1766 | /* Let's cache this item for the next invocation */ |
af13a6b0 | 1767 | chain_cache_put(f->chain_cache, ci, first, a, le64toh(array->entry_array.items[0]), t, subtract_one ? (i > 0 ? i-1 : (uint64_t) -1) : i); |
a4bcff5b | 1768 | |
de190aef LP |
1769 | if (subtract_one && i == 0) |
1770 | p = last_p; | |
1771 | else if (subtract_one) | |
1772 | p = le64toh(array->entry_array.items[i-1]); | |
1773 | else | |
1774 | p = le64toh(array->entry_array.items[i]); | |
1775 | ||
1776 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1777 | if (r < 0) | |
1778 | return r; | |
1779 | ||
1780 | if (ret) | |
1781 | *ret = o; | |
1782 | ||
1783 | if (offset) | |
1784 | *offset = p; | |
1785 | ||
1786 | if (idx) | |
cbdca852 | 1787 | *idx = t + i + (subtract_one ? -1 : 0); |
de190aef LP |
1788 | |
1789 | return 1; | |
cec736d2 LP |
1790 | } |
1791 | ||
f268980d LP |
1792 | static int generic_array_bisect_plus_one( |
1793 | JournalFile *f, | |
1794 | uint64_t extra, | |
1795 | uint64_t first, | |
1796 | uint64_t n, | |
1797 | uint64_t needle, | |
1798 | int (*test_object)(JournalFile *f, uint64_t p, uint64_t needle), | |
1799 | direction_t direction, | |
1800 | Object **ret, | |
1801 | uint64_t *offset, | |
1802 | uint64_t *idx) { | |
de190aef | 1803 | |
cec736d2 | 1804 | int r; |
cbdca852 LP |
1805 | bool step_back = false; |
1806 | Object *o; | |
cec736d2 LP |
1807 | |
1808 | assert(f); | |
de190aef | 1809 | assert(test_object); |
cec736d2 | 1810 | |
de190aef LP |
1811 | if (n <= 0) |
1812 | return 0; | |
cec736d2 | 1813 | |
de190aef LP |
1814 | /* This bisects the array in object 'first', but first checks |
1815 | * an extra */ | |
de190aef LP |
1816 | r = test_object(f, extra, needle); |
1817 | if (r < 0) | |
1818 | return r; | |
a536e261 LP |
1819 | |
1820 | if (r == TEST_FOUND) | |
1821 | r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT; | |
1822 | ||
cbdca852 LP |
1823 | /* if we are looking with DIRECTION_UP then we need to first |
1824 | see if in the actual array there is a matching entry, and | |
1825 | return the last one of that. But if there isn't any we need | |
1826 | to return this one. Hence remember this, and return it | |
1827 | below. */ | |
1828 | if (r == TEST_LEFT) | |
1829 | step_back = direction == DIRECTION_UP; | |
de190aef | 1830 | |
cbdca852 LP |
1831 | if (r == TEST_RIGHT) { |
1832 | if (direction == DIRECTION_DOWN) | |
1833 | goto found; | |
1834 | else | |
1835 | return 0; | |
a536e261 | 1836 | } |
cec736d2 | 1837 | |
de190aef LP |
1838 | r = generic_array_bisect(f, first, n-1, needle, test_object, direction, ret, offset, idx); |
1839 | ||
cbdca852 LP |
1840 | if (r == 0 && step_back) |
1841 | goto found; | |
1842 | ||
ecf68b1d | 1843 | if (r > 0 && idx) |
de190aef LP |
1844 | (*idx) ++; |
1845 | ||
1846 | return r; | |
cbdca852 LP |
1847 | |
1848 | found: | |
1849 | r = journal_file_move_to_object(f, OBJECT_ENTRY, extra, &o); | |
1850 | if (r < 0) | |
1851 | return r; | |
1852 | ||
1853 | if (ret) | |
1854 | *ret = o; | |
1855 | ||
1856 | if (offset) | |
1857 | *offset = extra; | |
1858 | ||
1859 | if (idx) | |
1860 | *idx = 0; | |
1861 | ||
1862 | return 1; | |
1863 | } | |
1864 | ||
44a6b1b6 | 1865 | _pure_ static int test_object_offset(JournalFile *f, uint64_t p, uint64_t needle) { |
cbdca852 LP |
1866 | assert(f); |
1867 | assert(p > 0); | |
1868 | ||
1869 | if (p == needle) | |
1870 | return TEST_FOUND; | |
1871 | else if (p < needle) | |
1872 | return TEST_LEFT; | |
1873 | else | |
1874 | return TEST_RIGHT; | |
1875 | } | |
1876 | ||
de190aef LP |
1877 | static int test_object_seqnum(JournalFile *f, uint64_t p, uint64_t needle) { |
1878 | Object *o; | |
1879 | int r; | |
1880 | ||
1881 | assert(f); | |
1882 | assert(p > 0); | |
1883 | ||
1884 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
cec736d2 LP |
1885 | if (r < 0) |
1886 | return r; | |
1887 | ||
de190aef LP |
1888 | if (le64toh(o->entry.seqnum) == needle) |
1889 | return TEST_FOUND; | |
1890 | else if (le64toh(o->entry.seqnum) < needle) | |
1891 | return TEST_LEFT; | |
1892 | else | |
1893 | return TEST_RIGHT; | |
1894 | } | |
cec736d2 | 1895 | |
de190aef LP |
1896 | int journal_file_move_to_entry_by_seqnum( |
1897 | JournalFile *f, | |
1898 | uint64_t seqnum, | |
1899 | direction_t direction, | |
1900 | Object **ret, | |
1901 | uint64_t *offset) { | |
1902 | ||
1903 | return generic_array_bisect(f, | |
1904 | le64toh(f->header->entry_array_offset), | |
1905 | le64toh(f->header->n_entries), | |
1906 | seqnum, | |
1907 | test_object_seqnum, | |
1908 | direction, | |
1909 | ret, offset, NULL); | |
1910 | } | |
cec736d2 | 1911 | |
de190aef LP |
1912 | static int test_object_realtime(JournalFile *f, uint64_t p, uint64_t needle) { |
1913 | Object *o; | |
1914 | int r; | |
1915 | ||
1916 | assert(f); | |
1917 | assert(p > 0); | |
1918 | ||
1919 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1920 | if (r < 0) | |
1921 | return r; | |
1922 | ||
1923 | if (le64toh(o->entry.realtime) == needle) | |
1924 | return TEST_FOUND; | |
1925 | else if (le64toh(o->entry.realtime) < needle) | |
1926 | return TEST_LEFT; | |
1927 | else | |
1928 | return TEST_RIGHT; | |
cec736d2 LP |
1929 | } |
1930 | ||
de190aef LP |
1931 | int journal_file_move_to_entry_by_realtime( |
1932 | JournalFile *f, | |
1933 | uint64_t realtime, | |
1934 | direction_t direction, | |
1935 | Object **ret, | |
1936 | uint64_t *offset) { | |
1937 | ||
1938 | return generic_array_bisect(f, | |
1939 | le64toh(f->header->entry_array_offset), | |
1940 | le64toh(f->header->n_entries), | |
1941 | realtime, | |
1942 | test_object_realtime, | |
1943 | direction, | |
1944 | ret, offset, NULL); | |
1945 | } | |
1946 | ||
1947 | static int test_object_monotonic(JournalFile *f, uint64_t p, uint64_t needle) { | |
1948 | Object *o; | |
1949 | int r; | |
1950 | ||
1951 | assert(f); | |
1952 | assert(p > 0); | |
1953 | ||
1954 | r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o); | |
1955 | if (r < 0) | |
1956 | return r; | |
1957 | ||
1958 | if (le64toh(o->entry.monotonic) == needle) | |
1959 | return TEST_FOUND; | |
1960 | else if (le64toh(o->entry.monotonic) < needle) | |
1961 | return TEST_LEFT; | |
1962 | else | |
1963 | return TEST_RIGHT; | |
1964 | } | |
1965 | ||
2a560338 | 1966 | static int find_data_object_by_boot_id( |
47838ab3 ZJS |
1967 | JournalFile *f, |
1968 | sd_id128_t boot_id, | |
1969 | Object **o, | |
1970 | uint64_t *b) { | |
2a560338 | 1971 | |
47838ab3 ZJS |
1972 | char t[sizeof("_BOOT_ID=")-1 + 32 + 1] = "_BOOT_ID="; |
1973 | ||
1974 | sd_id128_to_string(boot_id, t + 9); | |
1975 | return journal_file_find_data_object(f, t, sizeof(t) - 1, o, b); | |
1976 | } | |
1977 | ||
de190aef LP |
1978 | int journal_file_move_to_entry_by_monotonic( |
1979 | JournalFile *f, | |
1980 | sd_id128_t boot_id, | |
1981 | uint64_t monotonic, | |
1982 | direction_t direction, | |
1983 | Object **ret, | |
1984 | uint64_t *offset) { | |
1985 | ||
de190aef LP |
1986 | Object *o; |
1987 | int r; | |
1988 | ||
cbdca852 | 1989 | assert(f); |
de190aef | 1990 | |
47838ab3 | 1991 | r = find_data_object_by_boot_id(f, boot_id, &o, NULL); |
de190aef LP |
1992 | if (r < 0) |
1993 | return r; | |
cbdca852 | 1994 | if (r == 0) |
de190aef LP |
1995 | return -ENOENT; |
1996 | ||
1997 | return generic_array_bisect_plus_one(f, | |
1998 | le64toh(o->data.entry_offset), | |
1999 | le64toh(o->data.entry_array_offset), | |
2000 | le64toh(o->data.n_entries), | |
2001 | monotonic, | |
2002 | test_object_monotonic, | |
2003 | direction, | |
2004 | ret, offset, NULL); | |
2005 | } | |
2006 | ||
1fc605b0 | 2007 | void journal_file_reset_location(JournalFile *f) { |
6573ef05 | 2008 | f->location_type = LOCATION_HEAD; |
1fc605b0 | 2009 | f->current_offset = 0; |
6573ef05 MS |
2010 | f->current_seqnum = 0; |
2011 | f->current_realtime = 0; | |
2012 | f->current_monotonic = 0; | |
2013 | zero(f->current_boot_id); | |
2014 | f->current_xor_hash = 0; | |
2015 | } | |
2016 | ||
950c07d4 | 2017 | void journal_file_save_location(JournalFile *f, Object *o, uint64_t offset) { |
6573ef05 MS |
2018 | f->location_type = LOCATION_SEEK; |
2019 | f->current_offset = offset; | |
2020 | f->current_seqnum = le64toh(o->entry.seqnum); | |
2021 | f->current_realtime = le64toh(o->entry.realtime); | |
2022 | f->current_monotonic = le64toh(o->entry.monotonic); | |
2023 | f->current_boot_id = o->entry.boot_id; | |
2024 | f->current_xor_hash = le64toh(o->entry.xor_hash); | |
1fc605b0 MS |
2025 | } |
2026 | ||
d8ae66d7 MS |
2027 | int journal_file_compare_locations(JournalFile *af, JournalFile *bf) { |
2028 | assert(af); | |
2029 | assert(bf); | |
2030 | assert(af->location_type == LOCATION_SEEK); | |
2031 | assert(bf->location_type == LOCATION_SEEK); | |
2032 | ||
2033 | /* If contents and timestamps match, these entries are | |
2034 | * identical, even if the seqnum does not match */ | |
2035 | if (sd_id128_equal(af->current_boot_id, bf->current_boot_id) && | |
2036 | af->current_monotonic == bf->current_monotonic && | |
2037 | af->current_realtime == bf->current_realtime && | |
2038 | af->current_xor_hash == bf->current_xor_hash) | |
2039 | return 0; | |
2040 | ||
2041 | if (sd_id128_equal(af->header->seqnum_id, bf->header->seqnum_id)) { | |
2042 | ||
2043 | /* If this is from the same seqnum source, compare | |
2044 | * seqnums */ | |
2045 | if (af->current_seqnum < bf->current_seqnum) | |
2046 | return -1; | |
2047 | if (af->current_seqnum > bf->current_seqnum) | |
2048 | return 1; | |
2049 | ||
2050 | /* Wow! This is weird, different data but the same | |
2051 | * seqnums? Something is borked, but let's make the | |
2052 | * best of it and compare by time. */ | |
2053 | } | |
2054 | ||
2055 | if (sd_id128_equal(af->current_boot_id, bf->current_boot_id)) { | |
2056 | ||
2057 | /* If the boot id matches, compare monotonic time */ | |
2058 | if (af->current_monotonic < bf->current_monotonic) | |
2059 | return -1; | |
2060 | if (af->current_monotonic > bf->current_monotonic) | |
2061 | return 1; | |
2062 | } | |
2063 | ||
2064 | /* Otherwise, compare UTC time */ | |
2065 | if (af->current_realtime < bf->current_realtime) | |
2066 | return -1; | |
2067 | if (af->current_realtime > bf->current_realtime) | |
2068 | return 1; | |
2069 | ||
2070 | /* Finally, compare by contents */ | |
2071 | if (af->current_xor_hash < bf->current_xor_hash) | |
2072 | return -1; | |
2073 | if (af->current_xor_hash > bf->current_xor_hash) | |
2074 | return 1; | |
2075 | ||
2076 | return 0; | |
2077 | } | |
2078 | ||
de190aef LP |
2079 | int journal_file_next_entry( |
2080 | JournalFile *f, | |
f534928a | 2081 | uint64_t p, |
de190aef LP |
2082 | direction_t direction, |
2083 | Object **ret, uint64_t *offset) { | |
2084 | ||
fb099c8d | 2085 | uint64_t i, n, ofs; |
cec736d2 LP |
2086 | int r; |
2087 | ||
2088 | assert(f); | |
de190aef LP |
2089 | |
2090 | n = le64toh(f->header->n_entries); | |
2091 | if (n <= 0) | |
2092 | return 0; | |
cec736d2 | 2093 | |
f534928a | 2094 | if (p == 0) |
de190aef | 2095 | i = direction == DIRECTION_DOWN ? 0 : n - 1; |
cec736d2 | 2096 | else { |
de190aef LP |
2097 | r = generic_array_bisect(f, |
2098 | le64toh(f->header->entry_array_offset), | |
2099 | le64toh(f->header->n_entries), | |
2100 | p, | |
2101 | test_object_offset, | |
2102 | DIRECTION_DOWN, | |
2103 | NULL, NULL, | |
2104 | &i); | |
2105 | if (r <= 0) | |
2106 | return r; | |
2107 | ||
2108 | if (direction == DIRECTION_DOWN) { | |
2109 | if (i >= n - 1) | |
2110 | return 0; | |
2111 | ||
2112 | i++; | |
2113 | } else { | |
2114 | if (i <= 0) | |
2115 | return 0; | |
2116 | ||
2117 | i--; | |
2118 | } | |
cec736d2 LP |
2119 | } |
2120 | ||
de190aef | 2121 | /* And jump to it */ |
fb099c8d ZJS |
2122 | r = generic_array_get(f, |
2123 | le64toh(f->header->entry_array_offset), | |
2124 | i, | |
2125 | ret, &ofs); | |
2126 | if (r <= 0) | |
2127 | return r; | |
2128 | ||
2129 | if (p > 0 && | |
2130 | (direction == DIRECTION_DOWN ? ofs <= p : ofs >= p)) { | |
2131 | log_debug("%s: entry array corrupted at entry %"PRIu64, | |
2132 | f->path, i); | |
2133 | return -EBADMSG; | |
2134 | } | |
2135 | ||
2136 | if (offset) | |
2137 | *offset = ofs; | |
2138 | ||
2139 | return 1; | |
de190aef | 2140 | } |
cec736d2 | 2141 | |
de190aef LP |
2142 | int journal_file_next_entry_for_data( |
2143 | JournalFile *f, | |
2144 | Object *o, uint64_t p, | |
2145 | uint64_t data_offset, | |
2146 | direction_t direction, | |
2147 | Object **ret, uint64_t *offset) { | |
2148 | ||
2149 | uint64_t n, i; | |
cec736d2 | 2150 | int r; |
de190aef | 2151 | Object *d; |
cec736d2 LP |
2152 | |
2153 | assert(f); | |
de190aef | 2154 | assert(p > 0 || !o); |
cec736d2 | 2155 | |
de190aef | 2156 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); |
466ccd92 | 2157 | if (r < 0) |
de190aef | 2158 | return r; |
cec736d2 | 2159 | |
de190aef LP |
2160 | n = le64toh(d->data.n_entries); |
2161 | if (n <= 0) | |
2162 | return n; | |
cec736d2 | 2163 | |
de190aef LP |
2164 | if (!o) |
2165 | i = direction == DIRECTION_DOWN ? 0 : n - 1; | |
2166 | else { | |
2167 | if (o->object.type != OBJECT_ENTRY) | |
2168 | return -EINVAL; | |
cec736d2 | 2169 | |
de190aef LP |
2170 | r = generic_array_bisect_plus_one(f, |
2171 | le64toh(d->data.entry_offset), | |
2172 | le64toh(d->data.entry_array_offset), | |
2173 | le64toh(d->data.n_entries), | |
2174 | p, | |
2175 | test_object_offset, | |
2176 | DIRECTION_DOWN, | |
2177 | NULL, NULL, | |
2178 | &i); | |
2179 | ||
2180 | if (r <= 0) | |
cec736d2 LP |
2181 | return r; |
2182 | ||
de190aef LP |
2183 | if (direction == DIRECTION_DOWN) { |
2184 | if (i >= n - 1) | |
2185 | return 0; | |
cec736d2 | 2186 | |
de190aef LP |
2187 | i++; |
2188 | } else { | |
2189 | if (i <= 0) | |
2190 | return 0; | |
cec736d2 | 2191 | |
de190aef LP |
2192 | i--; |
2193 | } | |
cec736d2 | 2194 | |
de190aef | 2195 | } |
cec736d2 | 2196 | |
de190aef LP |
2197 | return generic_array_get_plus_one(f, |
2198 | le64toh(d->data.entry_offset), | |
2199 | le64toh(d->data.entry_array_offset), | |
2200 | i, | |
2201 | ret, offset); | |
2202 | } | |
cec736d2 | 2203 | |
cbdca852 LP |
2204 | int journal_file_move_to_entry_by_offset_for_data( |
2205 | JournalFile *f, | |
2206 | uint64_t data_offset, | |
2207 | uint64_t p, | |
2208 | direction_t direction, | |
2209 | Object **ret, uint64_t *offset) { | |
2210 | ||
2211 | int r; | |
2212 | Object *d; | |
2213 | ||
2214 | assert(f); | |
2215 | ||
2216 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); | |
2217 | if (r < 0) | |
2218 | return r; | |
2219 | ||
2220 | return generic_array_bisect_plus_one(f, | |
2221 | le64toh(d->data.entry_offset), | |
2222 | le64toh(d->data.entry_array_offset), | |
2223 | le64toh(d->data.n_entries), | |
2224 | p, | |
2225 | test_object_offset, | |
2226 | direction, | |
2227 | ret, offset, NULL); | |
2228 | } | |
2229 | ||
2230 | int journal_file_move_to_entry_by_monotonic_for_data( | |
2231 | JournalFile *f, | |
2232 | uint64_t data_offset, | |
2233 | sd_id128_t boot_id, | |
2234 | uint64_t monotonic, | |
2235 | direction_t direction, | |
2236 | Object **ret, uint64_t *offset) { | |
2237 | ||
cbdca852 LP |
2238 | Object *o, *d; |
2239 | int r; | |
2240 | uint64_t b, z; | |
2241 | ||
2242 | assert(f); | |
2243 | ||
2244 | /* First, seek by time */ | |
47838ab3 | 2245 | r = find_data_object_by_boot_id(f, boot_id, &o, &b); |
cbdca852 LP |
2246 | if (r < 0) |
2247 | return r; | |
2248 | if (r == 0) | |
2249 | return -ENOENT; | |
2250 | ||
2251 | r = generic_array_bisect_plus_one(f, | |
2252 | le64toh(o->data.entry_offset), | |
2253 | le64toh(o->data.entry_array_offset), | |
2254 | le64toh(o->data.n_entries), | |
2255 | monotonic, | |
2256 | test_object_monotonic, | |
2257 | direction, | |
2258 | NULL, &z, NULL); | |
2259 | if (r <= 0) | |
2260 | return r; | |
2261 | ||
2262 | /* And now, continue seeking until we find an entry that | |
2263 | * exists in both bisection arrays */ | |
2264 | ||
2265 | for (;;) { | |
2266 | Object *qo; | |
2267 | uint64_t p, q; | |
2268 | ||
2269 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); | |
2270 | if (r < 0) | |
2271 | return r; | |
2272 | ||
2273 | r = generic_array_bisect_plus_one(f, | |
2274 | le64toh(d->data.entry_offset), | |
2275 | le64toh(d->data.entry_array_offset), | |
2276 | le64toh(d->data.n_entries), | |
2277 | z, | |
2278 | test_object_offset, | |
2279 | direction, | |
2280 | NULL, &p, NULL); | |
2281 | if (r <= 0) | |
2282 | return r; | |
2283 | ||
2284 | r = journal_file_move_to_object(f, OBJECT_DATA, b, &o); | |
2285 | if (r < 0) | |
2286 | return r; | |
2287 | ||
2288 | r = generic_array_bisect_plus_one(f, | |
2289 | le64toh(o->data.entry_offset), | |
2290 | le64toh(o->data.entry_array_offset), | |
2291 | le64toh(o->data.n_entries), | |
2292 | p, | |
2293 | test_object_offset, | |
2294 | direction, | |
2295 | &qo, &q, NULL); | |
2296 | ||
2297 | if (r <= 0) | |
2298 | return r; | |
2299 | ||
2300 | if (p == q) { | |
2301 | if (ret) | |
2302 | *ret = qo; | |
2303 | if (offset) | |
2304 | *offset = q; | |
2305 | ||
2306 | return 1; | |
2307 | } | |
2308 | ||
2309 | z = q; | |
2310 | } | |
cbdca852 LP |
2311 | } |
2312 | ||
de190aef LP |
2313 | int journal_file_move_to_entry_by_seqnum_for_data( |
2314 | JournalFile *f, | |
2315 | uint64_t data_offset, | |
2316 | uint64_t seqnum, | |
2317 | direction_t direction, | |
2318 | Object **ret, uint64_t *offset) { | |
cec736d2 | 2319 | |
de190aef LP |
2320 | Object *d; |
2321 | int r; | |
cec736d2 | 2322 | |
91a31dde LP |
2323 | assert(f); |
2324 | ||
de190aef | 2325 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); |
91a31dde | 2326 | if (r < 0) |
de190aef | 2327 | return r; |
cec736d2 | 2328 | |
de190aef LP |
2329 | return generic_array_bisect_plus_one(f, |
2330 | le64toh(d->data.entry_offset), | |
2331 | le64toh(d->data.entry_array_offset), | |
2332 | le64toh(d->data.n_entries), | |
2333 | seqnum, | |
2334 | test_object_seqnum, | |
2335 | direction, | |
2336 | ret, offset, NULL); | |
2337 | } | |
cec736d2 | 2338 | |
de190aef LP |
2339 | int journal_file_move_to_entry_by_realtime_for_data( |
2340 | JournalFile *f, | |
2341 | uint64_t data_offset, | |
2342 | uint64_t realtime, | |
2343 | direction_t direction, | |
2344 | Object **ret, uint64_t *offset) { | |
2345 | ||
2346 | Object *d; | |
2347 | int r; | |
2348 | ||
91a31dde LP |
2349 | assert(f); |
2350 | ||
de190aef | 2351 | r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d); |
91a31dde | 2352 | if (r < 0) |
de190aef LP |
2353 | return r; |
2354 | ||
2355 | return generic_array_bisect_plus_one(f, | |
2356 | le64toh(d->data.entry_offset), | |
2357 | le64toh(d->data.entry_array_offset), | |
2358 | le64toh(d->data.n_entries), | |
2359 | realtime, | |
2360 | test_object_realtime, | |
2361 | direction, | |
2362 | ret, offset, NULL); | |
cec736d2 LP |
2363 | } |
2364 | ||
0284adc6 | 2365 | void journal_file_dump(JournalFile *f) { |
7560fffc | 2366 | Object *o; |
7560fffc | 2367 | int r; |
0284adc6 | 2368 | uint64_t p; |
7560fffc LP |
2369 | |
2370 | assert(f); | |
2371 | ||
0284adc6 | 2372 | journal_file_print_header(f); |
7560fffc | 2373 | |
0284adc6 LP |
2374 | p = le64toh(f->header->header_size); |
2375 | while (p != 0) { | |
d05089d8 | 2376 | r = journal_file_move_to_object(f, OBJECT_UNUSED, p, &o); |
0284adc6 LP |
2377 | if (r < 0) |
2378 | goto fail; | |
7560fffc | 2379 | |
0284adc6 | 2380 | switch (o->object.type) { |
d98cc1f2 | 2381 | |
0284adc6 LP |
2382 | case OBJECT_UNUSED: |
2383 | printf("Type: OBJECT_UNUSED\n"); | |
2384 | break; | |
d98cc1f2 | 2385 | |
0284adc6 LP |
2386 | case OBJECT_DATA: |
2387 | printf("Type: OBJECT_DATA\n"); | |
2388 | break; | |
7560fffc | 2389 | |
3c1668da LP |
2390 | case OBJECT_FIELD: |
2391 | printf("Type: OBJECT_FIELD\n"); | |
2392 | break; | |
2393 | ||
0284adc6 | 2394 | case OBJECT_ENTRY: |
507f22bd ZJS |
2395 | printf("Type: OBJECT_ENTRY seqnum=%"PRIu64" monotonic=%"PRIu64" realtime=%"PRIu64"\n", |
2396 | le64toh(o->entry.seqnum), | |
2397 | le64toh(o->entry.monotonic), | |
2398 | le64toh(o->entry.realtime)); | |
0284adc6 | 2399 | break; |
7560fffc | 2400 | |
0284adc6 LP |
2401 | case OBJECT_FIELD_HASH_TABLE: |
2402 | printf("Type: OBJECT_FIELD_HASH_TABLE\n"); | |
2403 | break; | |
7560fffc | 2404 | |
0284adc6 LP |
2405 | case OBJECT_DATA_HASH_TABLE: |
2406 | printf("Type: OBJECT_DATA_HASH_TABLE\n"); | |
2407 | break; | |
7560fffc | 2408 | |
0284adc6 LP |
2409 | case OBJECT_ENTRY_ARRAY: |
2410 | printf("Type: OBJECT_ENTRY_ARRAY\n"); | |
2411 | break; | |
7560fffc | 2412 | |
0284adc6 | 2413 | case OBJECT_TAG: |
507f22bd ZJS |
2414 | printf("Type: OBJECT_TAG seqnum=%"PRIu64" epoch=%"PRIu64"\n", |
2415 | le64toh(o->tag.seqnum), | |
2416 | le64toh(o->tag.epoch)); | |
0284adc6 | 2417 | break; |
3c1668da LP |
2418 | |
2419 | default: | |
8facc349 | 2420 | printf("Type: unknown (%i)\n", o->object.type); |
3c1668da | 2421 | break; |
0284adc6 | 2422 | } |
7560fffc | 2423 | |
d89c8fdf ZJS |
2424 | if (o->object.flags & OBJECT_COMPRESSION_MASK) |
2425 | printf("Flags: %s\n", | |
2426 | object_compressed_to_string(o->object.flags & OBJECT_COMPRESSION_MASK)); | |
7560fffc | 2427 | |
0284adc6 LP |
2428 | if (p == le64toh(f->header->tail_object_offset)) |
2429 | p = 0; | |
2430 | else | |
2431 | p = p + ALIGN64(le64toh(o->object.size)); | |
2432 | } | |
7560fffc | 2433 | |
0284adc6 LP |
2434 | return; |
2435 | fail: | |
2436 | log_error("File corrupt"); | |
7560fffc LP |
2437 | } |
2438 | ||
718fe4b1 ZJS |
2439 | static const char* format_timestamp_safe(char *buf, size_t l, usec_t t) { |
2440 | const char *x; | |
2441 | ||
2442 | x = format_timestamp(buf, l, t); | |
2443 | if (x) | |
2444 | return x; | |
2445 | return " --- "; | |
2446 | } | |
2447 | ||
0284adc6 | 2448 | void journal_file_print_header(JournalFile *f) { |
2765b7bb | 2449 | char a[33], b[33], c[33], d[33]; |
ed375beb | 2450 | char x[FORMAT_TIMESTAMP_MAX], y[FORMAT_TIMESTAMP_MAX], z[FORMAT_TIMESTAMP_MAX]; |
a1a03e30 LP |
2451 | struct stat st; |
2452 | char bytes[FORMAT_BYTES_MAX]; | |
7560fffc LP |
2453 | |
2454 | assert(f); | |
7560fffc | 2455 | |
0284adc6 LP |
2456 | printf("File Path: %s\n" |
2457 | "File ID: %s\n" | |
2458 | "Machine ID: %s\n" | |
2459 | "Boot ID: %s\n" | |
2460 | "Sequential Number ID: %s\n" | |
2461 | "State: %s\n" | |
2462 | "Compatible Flags:%s%s\n" | |
d89c8fdf | 2463 | "Incompatible Flags:%s%s%s\n" |
507f22bd ZJS |
2464 | "Header size: %"PRIu64"\n" |
2465 | "Arena size: %"PRIu64"\n" | |
2466 | "Data Hash Table Size: %"PRIu64"\n" | |
2467 | "Field Hash Table Size: %"PRIu64"\n" | |
0284adc6 | 2468 | "Rotate Suggested: %s\n" |
507f22bd ZJS |
2469 | "Head Sequential Number: %"PRIu64"\n" |
2470 | "Tail Sequential Number: %"PRIu64"\n" | |
0284adc6 | 2471 | "Head Realtime Timestamp: %s\n" |
3223f44f | 2472 | "Tail Realtime Timestamp: %s\n" |
ed375beb | 2473 | "Tail Monotonic Timestamp: %s\n" |
507f22bd ZJS |
2474 | "Objects: %"PRIu64"\n" |
2475 | "Entry Objects: %"PRIu64"\n", | |
0284adc6 LP |
2476 | f->path, |
2477 | sd_id128_to_string(f->header->file_id, a), | |
2478 | sd_id128_to_string(f->header->machine_id, b), | |
2479 | sd_id128_to_string(f->header->boot_id, c), | |
2765b7bb | 2480 | sd_id128_to_string(f->header->seqnum_id, d), |
3223f44f LP |
2481 | f->header->state == STATE_OFFLINE ? "OFFLINE" : |
2482 | f->header->state == STATE_ONLINE ? "ONLINE" : | |
2483 | f->header->state == STATE_ARCHIVED ? "ARCHIVED" : "UNKNOWN", | |
8088cbd3 | 2484 | JOURNAL_HEADER_SEALED(f->header) ? " SEALED" : "", |
d89c8fdf ZJS |
2485 | (le32toh(f->header->compatible_flags) & ~HEADER_COMPATIBLE_ANY) ? " ???" : "", |
2486 | JOURNAL_HEADER_COMPRESSED_XZ(f->header) ? " COMPRESSED-XZ" : "", | |
2487 | JOURNAL_HEADER_COMPRESSED_LZ4(f->header) ? " COMPRESSED-LZ4" : "", | |
2488 | (le32toh(f->header->incompatible_flags) & ~HEADER_INCOMPATIBLE_ANY) ? " ???" : "", | |
507f22bd ZJS |
2489 | le64toh(f->header->header_size), |
2490 | le64toh(f->header->arena_size), | |
2491 | le64toh(f->header->data_hash_table_size) / sizeof(HashItem), | |
2492 | le64toh(f->header->field_hash_table_size) / sizeof(HashItem), | |
fb0951b0 | 2493 | yes_no(journal_file_rotate_suggested(f, 0)), |
507f22bd ZJS |
2494 | le64toh(f->header->head_entry_seqnum), |
2495 | le64toh(f->header->tail_entry_seqnum), | |
718fe4b1 ZJS |
2496 | format_timestamp_safe(x, sizeof(x), le64toh(f->header->head_entry_realtime)), |
2497 | format_timestamp_safe(y, sizeof(y), le64toh(f->header->tail_entry_realtime)), | |
ed375beb | 2498 | format_timespan(z, sizeof(z), le64toh(f->header->tail_entry_monotonic), USEC_PER_MSEC), |
507f22bd ZJS |
2499 | le64toh(f->header->n_objects), |
2500 | le64toh(f->header->n_entries)); | |
7560fffc | 2501 | |
0284adc6 | 2502 | if (JOURNAL_HEADER_CONTAINS(f->header, n_data)) |
507f22bd | 2503 | printf("Data Objects: %"PRIu64"\n" |
0284adc6 | 2504 | "Data Hash Table Fill: %.1f%%\n", |
507f22bd | 2505 | le64toh(f->header->n_data), |
0284adc6 | 2506 | 100.0 * (double) le64toh(f->header->n_data) / ((double) (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)))); |
7560fffc | 2507 | |
0284adc6 | 2508 | if (JOURNAL_HEADER_CONTAINS(f->header, n_fields)) |
507f22bd | 2509 | printf("Field Objects: %"PRIu64"\n" |
0284adc6 | 2510 | "Field Hash Table Fill: %.1f%%\n", |
507f22bd | 2511 | le64toh(f->header->n_fields), |
0284adc6 | 2512 | 100.0 * (double) le64toh(f->header->n_fields) / ((double) (le64toh(f->header->field_hash_table_size) / sizeof(HashItem)))); |
3223f44f LP |
2513 | |
2514 | if (JOURNAL_HEADER_CONTAINS(f->header, n_tags)) | |
507f22bd ZJS |
2515 | printf("Tag Objects: %"PRIu64"\n", |
2516 | le64toh(f->header->n_tags)); | |
3223f44f | 2517 | if (JOURNAL_HEADER_CONTAINS(f->header, n_entry_arrays)) |
507f22bd ZJS |
2518 | printf("Entry Array Objects: %"PRIu64"\n", |
2519 | le64toh(f->header->n_entry_arrays)); | |
a1a03e30 LP |
2520 | |
2521 | if (fstat(f->fd, &st) >= 0) | |
2522 | printf("Disk usage: %s\n", format_bytes(bytes, sizeof(bytes), (off_t) st.st_blocks * 512ULL)); | |
7560fffc LP |
2523 | } |
2524 | ||
fc68c929 LP |
2525 | static int journal_file_warn_btrfs(JournalFile *f) { |
2526 | unsigned attrs; | |
2527 | int r; | |
2528 | ||
2529 | assert(f); | |
2530 | ||
2531 | /* Before we write anything, check if the COW logic is turned | |
2532 | * off on btrfs. Given our write pattern that is quite | |
2533 | * unfriendly to COW file systems this should greatly improve | |
2534 | * performance on COW file systems, such as btrfs, at the | |
2535 | * expense of data integrity features (which shouldn't be too | |
2536 | * bad, given that we do our own checksumming). */ | |
2537 | ||
2538 | r = btrfs_is_filesystem(f->fd); | |
2539 | if (r < 0) | |
2540 | return log_warning_errno(r, "Failed to determine if journal is on btrfs: %m"); | |
2541 | if (!r) | |
2542 | return 0; | |
2543 | ||
2544 | r = read_attr_fd(f->fd, &attrs); | |
2545 | if (r < 0) | |
2546 | return log_warning_errno(r, "Failed to read file attributes: %m"); | |
2547 | ||
2548 | if (attrs & FS_NOCOW_FL) { | |
2549 | log_debug("Detected btrfs file system with copy-on-write disabled, all is good."); | |
2550 | return 0; | |
2551 | } | |
2552 | ||
2553 | log_notice("Creating journal file %s on a btrfs file system, and copy-on-write is enabled. " | |
2554 | "This is likely to slow down journal access substantially, please consider turning " | |
2555 | "off the copy-on-write file attribute on the journal directory, using chattr +C.", f->path); | |
2556 | ||
2557 | return 1; | |
2558 | } | |
2559 | ||
0284adc6 LP |
2560 | int journal_file_open( |
2561 | const char *fname, | |
2562 | int flags, | |
2563 | mode_t mode, | |
2564 | bool compress, | |
baed47c3 | 2565 | bool seal, |
0284adc6 LP |
2566 | JournalMetrics *metrics, |
2567 | MMapCache *mmap_cache, | |
2568 | JournalFile *template, | |
2569 | JournalFile **ret) { | |
7560fffc | 2570 | |
fa6ac760 | 2571 | bool newly_created = false; |
0284adc6 | 2572 | JournalFile *f; |
fa6ac760 | 2573 | void *h; |
0284adc6 | 2574 | int r; |
7560fffc | 2575 | |
0284adc6 | 2576 | assert(fname); |
0559d3a5 | 2577 | assert(ret); |
7560fffc | 2578 | |
0284adc6 LP |
2579 | if ((flags & O_ACCMODE) != O_RDONLY && |
2580 | (flags & O_ACCMODE) != O_RDWR) | |
2581 | return -EINVAL; | |
7560fffc | 2582 | |
a0108012 LP |
2583 | if (!endswith(fname, ".journal") && |
2584 | !endswith(fname, ".journal~")) | |
0284adc6 | 2585 | return -EINVAL; |
7560fffc | 2586 | |
0284adc6 LP |
2587 | f = new0(JournalFile, 1); |
2588 | if (!f) | |
2589 | return -ENOMEM; | |
7560fffc | 2590 | |
0284adc6 LP |
2591 | f->fd = -1; |
2592 | f->mode = mode; | |
7560fffc | 2593 | |
0284adc6 LP |
2594 | f->flags = flags; |
2595 | f->prot = prot_from_flags(flags); | |
2596 | f->writable = (flags & O_ACCMODE) != O_RDONLY; | |
92261977 | 2597 | #if defined(HAVE_LZ4) |
d89c8fdf ZJS |
2598 | f->compress_lz4 = compress; |
2599 | #elif defined(HAVE_XZ) | |
2600 | f->compress_xz = compress; | |
48b61739 | 2601 | #endif |
49a32d43 | 2602 | #ifdef HAVE_GCRYPT |
baed47c3 | 2603 | f->seal = seal; |
49a32d43 | 2604 | #endif |
7560fffc | 2605 | |
0284adc6 LP |
2606 | if (mmap_cache) |
2607 | f->mmap = mmap_cache_ref(mmap_cache); | |
2608 | else { | |
84168d80 | 2609 | f->mmap = mmap_cache_new(); |
0284adc6 LP |
2610 | if (!f->mmap) { |
2611 | r = -ENOMEM; | |
2612 | goto fail; | |
2613 | } | |
2614 | } | |
7560fffc | 2615 | |
0284adc6 LP |
2616 | f->path = strdup(fname); |
2617 | if (!f->path) { | |
2618 | r = -ENOMEM; | |
2619 | goto fail; | |
2620 | } | |
7560fffc | 2621 | |
4743015d | 2622 | f->chain_cache = ordered_hashmap_new(&uint64_hash_ops); |
a4bcff5b LP |
2623 | if (!f->chain_cache) { |
2624 | r = -ENOMEM; | |
2625 | goto fail; | |
2626 | } | |
2627 | ||
0284adc6 LP |
2628 | f->fd = open(f->path, f->flags|O_CLOEXEC, f->mode); |
2629 | if (f->fd < 0) { | |
2630 | r = -errno; | |
2631 | goto fail; | |
7560fffc | 2632 | } |
7560fffc | 2633 | |
2678031a LP |
2634 | r = journal_file_fstat(f); |
2635 | if (r < 0) | |
0284adc6 | 2636 | goto fail; |
7560fffc | 2637 | |
0284adc6 | 2638 | if (f->last_stat.st_size == 0 && f->writable) { |
11689d2a | 2639 | |
fc68c929 | 2640 | (void) journal_file_warn_btrfs(f); |
11689d2a | 2641 | |
fb0951b0 LP |
2642 | /* Let's attach the creation time to the journal file, |
2643 | * so that the vacuuming code knows the age of this | |
2644 | * file even if the file might end up corrupted one | |
2645 | * day... Ideally we'd just use the creation time many | |
2646 | * file systems maintain for each file, but there is | |
2647 | * currently no usable API to query this, hence let's | |
2648 | * emulate this via extended attributes. If extended | |
2649 | * attributes are not supported we'll just skip this, | |
7517e174 | 2650 | * and rely solely on mtime/atime/ctime of the file. */ |
fb0951b0 | 2651 | |
d61b600d | 2652 | fd_setcrtime(f->fd, 0); |
7560fffc | 2653 | |
feb12d3e | 2654 | #ifdef HAVE_GCRYPT |
0284adc6 | 2655 | /* Try to load the FSPRG state, and if we can't, then |
baed47c3 | 2656 | * just don't do sealing */ |
49a32d43 LP |
2657 | if (f->seal) { |
2658 | r = journal_file_fss_load(f); | |
2659 | if (r < 0) | |
2660 | f->seal = false; | |
2661 | } | |
feb12d3e | 2662 | #endif |
7560fffc | 2663 | |
0284adc6 LP |
2664 | r = journal_file_init_header(f, template); |
2665 | if (r < 0) | |
2666 | goto fail; | |
7560fffc | 2667 | |
2678031a LP |
2668 | r = journal_file_fstat(f); |
2669 | if (r < 0) | |
0284adc6 | 2670 | goto fail; |
fb0951b0 LP |
2671 | |
2672 | newly_created = true; | |
0284adc6 | 2673 | } |
7560fffc | 2674 | |
0284adc6 LP |
2675 | if (f->last_stat.st_size < (off_t) HEADER_SIZE_MIN) { |
2676 | r = -EIO; | |
2677 | goto fail; | |
2678 | } | |
7560fffc | 2679 | |
fa6ac760 | 2680 | r = mmap_cache_get(f->mmap, f->fd, f->prot, CONTEXT_HEADER, true, 0, PAGE_ALIGN(sizeof(Header)), &f->last_stat, &h); |
977eaa1e | 2681 | if (r < 0) |
0284adc6 | 2682 | goto fail; |
7560fffc | 2683 | |
fa6ac760 LP |
2684 | f->header = h; |
2685 | ||
0284adc6 LP |
2686 | if (!newly_created) { |
2687 | r = journal_file_verify_header(f); | |
2688 | if (r < 0) | |
2689 | goto fail; | |
2690 | } | |
7560fffc | 2691 | |
feb12d3e | 2692 | #ifdef HAVE_GCRYPT |
0284adc6 | 2693 | if (!newly_created && f->writable) { |
baed47c3 | 2694 | r = journal_file_fss_load(f); |
0284adc6 LP |
2695 | if (r < 0) |
2696 | goto fail; | |
2697 | } | |
feb12d3e | 2698 | #endif |
cec736d2 LP |
2699 | |
2700 | if (f->writable) { | |
4a92baf3 LP |
2701 | if (metrics) { |
2702 | journal_default_metrics(metrics, f->fd); | |
2703 | f->metrics = *metrics; | |
2704 | } else if (template) | |
2705 | f->metrics = template->metrics; | |
2706 | ||
cec736d2 LP |
2707 | r = journal_file_refresh_header(f); |
2708 | if (r < 0) | |
2709 | goto fail; | |
2710 | } | |
2711 | ||
feb12d3e | 2712 | #ifdef HAVE_GCRYPT |
baed47c3 | 2713 | r = journal_file_hmac_setup(f); |
14d10188 LP |
2714 | if (r < 0) |
2715 | goto fail; | |
feb12d3e | 2716 | #endif |
14d10188 | 2717 | |
cec736d2 | 2718 | if (newly_created) { |
de190aef | 2719 | r = journal_file_setup_field_hash_table(f); |
cec736d2 LP |
2720 | if (r < 0) |
2721 | goto fail; | |
2722 | ||
de190aef | 2723 | r = journal_file_setup_data_hash_table(f); |
cec736d2 LP |
2724 | if (r < 0) |
2725 | goto fail; | |
7560fffc | 2726 | |
feb12d3e | 2727 | #ifdef HAVE_GCRYPT |
7560fffc LP |
2728 | r = journal_file_append_first_tag(f); |
2729 | if (r < 0) | |
2730 | goto fail; | |
feb12d3e | 2731 | #endif |
cec736d2 LP |
2732 | } |
2733 | ||
de190aef | 2734 | r = journal_file_map_field_hash_table(f); |
cec736d2 LP |
2735 | if (r < 0) |
2736 | goto fail; | |
2737 | ||
de190aef | 2738 | r = journal_file_map_data_hash_table(f); |
cec736d2 LP |
2739 | if (r < 0) |
2740 | goto fail; | |
2741 | ||
fa6ac760 LP |
2742 | if (mmap_cache_got_sigbus(f->mmap, f->fd)) { |
2743 | r = -EIO; | |
2744 | goto fail; | |
2745 | } | |
2746 | ||
0559d3a5 | 2747 | *ret = f; |
cec736d2 LP |
2748 | return 0; |
2749 | ||
2750 | fail: | |
fa6ac760 LP |
2751 | if (f->fd >= 0 && mmap_cache_got_sigbus(f->mmap, f->fd)) |
2752 | r = -EIO; | |
2753 | ||
cec736d2 LP |
2754 | journal_file_close(f); |
2755 | ||
2756 | return r; | |
2757 | } | |
0ac38b70 | 2758 | |
baed47c3 | 2759 | int journal_file_rotate(JournalFile **f, bool compress, bool seal) { |
57535f47 | 2760 | _cleanup_free_ char *p = NULL; |
0ac38b70 LP |
2761 | size_t l; |
2762 | JournalFile *old_file, *new_file = NULL; | |
2763 | int r; | |
2764 | ||
2765 | assert(f); | |
2766 | assert(*f); | |
2767 | ||
2768 | old_file = *f; | |
2769 | ||
2770 | if (!old_file->writable) | |
2771 | return -EINVAL; | |
2772 | ||
2773 | if (!endswith(old_file->path, ".journal")) | |
2774 | return -EINVAL; | |
2775 | ||
2776 | l = strlen(old_file->path); | |
57535f47 ZJS |
2777 | r = asprintf(&p, "%.*s@" SD_ID128_FORMAT_STR "-%016"PRIx64"-%016"PRIx64".journal", |
2778 | (int) l - 8, old_file->path, | |
2779 | SD_ID128_FORMAT_VAL(old_file->header->seqnum_id), | |
2780 | le64toh((*f)->header->head_entry_seqnum), | |
2781 | le64toh((*f)->header->head_entry_realtime)); | |
2782 | if (r < 0) | |
0ac38b70 LP |
2783 | return -ENOMEM; |
2784 | ||
2678031a LP |
2785 | /* Try to rename the file to the archived version. If the file |
2786 | * already was deleted, we'll get ENOENT, let's ignore that | |
2787 | * case. */ | |
0ac38b70 | 2788 | r = rename(old_file->path, p); |
2678031a | 2789 | if (r < 0 && errno != ENOENT) |
0ac38b70 LP |
2790 | return -errno; |
2791 | ||
ccdbaf91 | 2792 | old_file->header->state = STATE_ARCHIVED; |
0ac38b70 | 2793 | |
f27a3864 LP |
2794 | /* Currently, btrfs is not very good with out write patterns |
2795 | * and fragments heavily. Let's defrag our journal files when | |
2796 | * we archive them */ | |
2797 | old_file->defrag_on_close = true; | |
2798 | ||
baed47c3 | 2799 | r = journal_file_open(old_file->path, old_file->flags, old_file->mode, compress, seal, NULL, old_file->mmap, old_file, &new_file); |
0ac38b70 LP |
2800 | journal_file_close(old_file); |
2801 | ||
2802 | *f = new_file; | |
2803 | return r; | |
2804 | } | |
2805 | ||
9447a7f1 LP |
2806 | int journal_file_open_reliably( |
2807 | const char *fname, | |
2808 | int flags, | |
2809 | mode_t mode, | |
7560fffc | 2810 | bool compress, |
baed47c3 | 2811 | bool seal, |
4a92baf3 | 2812 | JournalMetrics *metrics, |
27370278 | 2813 | MMapCache *mmap_cache, |
9447a7f1 LP |
2814 | JournalFile *template, |
2815 | JournalFile **ret) { | |
2816 | ||
2817 | int r; | |
2818 | size_t l; | |
ed375beb | 2819 | _cleanup_free_ char *p = NULL; |
9447a7f1 | 2820 | |
baed47c3 | 2821 | r = journal_file_open(fname, flags, mode, compress, seal, |
27370278 | 2822 | metrics, mmap_cache, template, ret); |
288359db ZJS |
2823 | if (!IN_SET(r, |
2824 | -EBADMSG, /* corrupted */ | |
2825 | -ENODATA, /* truncated */ | |
2826 | -EHOSTDOWN, /* other machine */ | |
2827 | -EPROTONOSUPPORT, /* incompatible feature */ | |
2828 | -EBUSY, /* unclean shutdown */ | |
2829 | -ESHUTDOWN, /* already archived */ | |
2830 | -EIO, /* IO error, including SIGBUS on mmap */ | |
2831 | -EIDRM /* File has been deleted */)) | |
9447a7f1 LP |
2832 | return r; |
2833 | ||
2834 | if ((flags & O_ACCMODE) == O_RDONLY) | |
2835 | return r; | |
2836 | ||
2837 | if (!(flags & O_CREAT)) | |
2838 | return r; | |
2839 | ||
7560fffc LP |
2840 | if (!endswith(fname, ".journal")) |
2841 | return r; | |
2842 | ||
5c70eab4 LP |
2843 | /* The file is corrupted. Rotate it away and try it again (but only once) */ |
2844 | ||
9447a7f1 | 2845 | l = strlen(fname); |
d587eca5 | 2846 | if (asprintf(&p, "%.*s@%016"PRIx64 "-%016"PRIx64 ".journal~", |
57535f47 | 2847 | (int) l - 8, fname, |
d587eca5 | 2848 | now(CLOCK_REALTIME), |
9bf3b535 | 2849 | random_u64()) < 0) |
9447a7f1 LP |
2850 | return -ENOMEM; |
2851 | ||
2852 | r = rename(fname, p); | |
9447a7f1 LP |
2853 | if (r < 0) |
2854 | return -errno; | |
2855 | ||
f27a3864 LP |
2856 | /* btrfs doesn't cope well with our write pattern and |
2857 | * fragments heavily. Let's defrag all files we rotate */ | |
11689d2a LP |
2858 | |
2859 | (void) chattr_path(p, false, FS_NOCOW_FL); | |
f27a3864 LP |
2860 | (void) btrfs_defrag(p); |
2861 | ||
a1a1898f | 2862 | log_warning("File %s corrupted or uncleanly shut down, renaming and replacing.", fname); |
9447a7f1 | 2863 | |
baed47c3 | 2864 | return journal_file_open(fname, flags, mode, compress, seal, |
27370278 | 2865 | metrics, mmap_cache, template, ret); |
9447a7f1 LP |
2866 | } |
2867 | ||
cf244689 LP |
2868 | int journal_file_copy_entry(JournalFile *from, JournalFile *to, Object *o, uint64_t p, uint64_t *seqnum, Object **ret, uint64_t *offset) { |
2869 | uint64_t i, n; | |
2870 | uint64_t q, xor_hash = 0; | |
2871 | int r; | |
2872 | EntryItem *items; | |
2873 | dual_timestamp ts; | |
2874 | ||
2875 | assert(from); | |
2876 | assert(to); | |
2877 | assert(o); | |
2878 | assert(p); | |
2879 | ||
2880 | if (!to->writable) | |
2881 | return -EPERM; | |
2882 | ||
2883 | ts.monotonic = le64toh(o->entry.monotonic); | |
2884 | ts.realtime = le64toh(o->entry.realtime); | |
2885 | ||
cf244689 | 2886 | n = journal_file_entry_n_items(o); |
4faa7004 TA |
2887 | /* alloca() can't take 0, hence let's allocate at least one */ |
2888 | items = alloca(sizeof(EntryItem) * MAX(1u, n)); | |
cf244689 LP |
2889 | |
2890 | for (i = 0; i < n; i++) { | |
4fd052ae FC |
2891 | uint64_t l, h; |
2892 | le64_t le_hash; | |
cf244689 LP |
2893 | size_t t; |
2894 | void *data; | |
2895 | Object *u; | |
2896 | ||
2897 | q = le64toh(o->entry.items[i].object_offset); | |
2898 | le_hash = o->entry.items[i].hash; | |
2899 | ||
2900 | r = journal_file_move_to_object(from, OBJECT_DATA, q, &o); | |
2901 | if (r < 0) | |
2902 | return r; | |
2903 | ||
2904 | if (le_hash != o->data.hash) | |
2905 | return -EBADMSG; | |
2906 | ||
2907 | l = le64toh(o->object.size) - offsetof(Object, data.payload); | |
2908 | t = (size_t) l; | |
2909 | ||
2910 | /* We hit the limit on 32bit machines */ | |
2911 | if ((uint64_t) t != l) | |
2912 | return -E2BIG; | |
2913 | ||
d89c8fdf | 2914 | if (o->object.flags & OBJECT_COMPRESSION_MASK) { |
3b1a55e1 | 2915 | #if defined(HAVE_XZ) || defined(HAVE_LZ4) |
a7f7d1bd | 2916 | size_t rsize = 0; |
cf244689 | 2917 | |
d89c8fdf ZJS |
2918 | r = decompress_blob(o->object.flags & OBJECT_COMPRESSION_MASK, |
2919 | o->data.payload, l, &from->compress_buffer, &from->compress_buffer_size, &rsize, 0); | |
2920 | if (r < 0) | |
2921 | return r; | |
cf244689 LP |
2922 | |
2923 | data = from->compress_buffer; | |
2924 | l = rsize; | |
3b1a55e1 ZJS |
2925 | #else |
2926 | return -EPROTONOSUPPORT; | |
2927 | #endif | |
cf244689 LP |
2928 | } else |
2929 | data = o->data.payload; | |
2930 | ||
2931 | r = journal_file_append_data(to, data, l, &u, &h); | |
2932 | if (r < 0) | |
2933 | return r; | |
2934 | ||
2935 | xor_hash ^= le64toh(u->data.hash); | |
2936 | items[i].object_offset = htole64(h); | |
2937 | items[i].hash = u->data.hash; | |
2938 | ||
2939 | r = journal_file_move_to_object(from, OBJECT_ENTRY, p, &o); | |
2940 | if (r < 0) | |
2941 | return r; | |
2942 | } | |
2943 | ||
fa6ac760 LP |
2944 | r = journal_file_append_entry_internal(to, &ts, xor_hash, items, n, seqnum, ret, offset); |
2945 | ||
2946 | if (mmap_cache_got_sigbus(to->mmap, to->fd)) | |
2947 | return -EIO; | |
2948 | ||
2949 | return r; | |
cf244689 | 2950 | } |
babfc091 LP |
2951 | |
2952 | void journal_default_metrics(JournalMetrics *m, int fd) { | |
2953 | uint64_t fs_size = 0; | |
2954 | struct statvfs ss; | |
a7bc2c2a | 2955 | char a[FORMAT_BYTES_MAX], b[FORMAT_BYTES_MAX], c[FORMAT_BYTES_MAX], d[FORMAT_BYTES_MAX]; |
babfc091 LP |
2956 | |
2957 | assert(m); | |
2958 | assert(fd >= 0); | |
2959 | ||
2960 | if (fstatvfs(fd, &ss) >= 0) | |
2961 | fs_size = ss.f_frsize * ss.f_blocks; | |
2962 | ||
2963 | if (m->max_use == (uint64_t) -1) { | |
2964 | ||
2965 | if (fs_size > 0) { | |
2966 | m->max_use = PAGE_ALIGN(fs_size / 10); /* 10% of file system size */ | |
2967 | ||
2968 | if (m->max_use > DEFAULT_MAX_USE_UPPER) | |
2969 | m->max_use = DEFAULT_MAX_USE_UPPER; | |
2970 | ||
2971 | if (m->max_use < DEFAULT_MAX_USE_LOWER) | |
2972 | m->max_use = DEFAULT_MAX_USE_LOWER; | |
2973 | } else | |
2974 | m->max_use = DEFAULT_MAX_USE_LOWER; | |
2975 | } else { | |
2976 | m->max_use = PAGE_ALIGN(m->max_use); | |
2977 | ||
2978 | if (m->max_use < JOURNAL_FILE_SIZE_MIN*2) | |
2979 | m->max_use = JOURNAL_FILE_SIZE_MIN*2; | |
2980 | } | |
2981 | ||
2982 | if (m->max_size == (uint64_t) -1) { | |
2983 | m->max_size = PAGE_ALIGN(m->max_use / 8); /* 8 chunks */ | |
2984 | ||
2985 | if (m->max_size > DEFAULT_MAX_SIZE_UPPER) | |
2986 | m->max_size = DEFAULT_MAX_SIZE_UPPER; | |
2987 | } else | |
2988 | m->max_size = PAGE_ALIGN(m->max_size); | |
2989 | ||
2990 | if (m->max_size < JOURNAL_FILE_SIZE_MIN) | |
2991 | m->max_size = JOURNAL_FILE_SIZE_MIN; | |
2992 | ||
2993 | if (m->max_size*2 > m->max_use) | |
2994 | m->max_use = m->max_size*2; | |
2995 | ||
2996 | if (m->min_size == (uint64_t) -1) | |
2997 | m->min_size = JOURNAL_FILE_SIZE_MIN; | |
2998 | else { | |
2999 | m->min_size = PAGE_ALIGN(m->min_size); | |
3000 | ||
3001 | if (m->min_size < JOURNAL_FILE_SIZE_MIN) | |
3002 | m->min_size = JOURNAL_FILE_SIZE_MIN; | |
3003 | ||
3004 | if (m->min_size > m->max_size) | |
3005 | m->max_size = m->min_size; | |
3006 | } | |
3007 | ||
3008 | if (m->keep_free == (uint64_t) -1) { | |
3009 | ||
3010 | if (fs_size > 0) { | |
8621b110 | 3011 | m->keep_free = PAGE_ALIGN(fs_size * 3 / 20); /* 15% of file system size */ |
babfc091 LP |
3012 | |
3013 | if (m->keep_free > DEFAULT_KEEP_FREE_UPPER) | |
3014 | m->keep_free = DEFAULT_KEEP_FREE_UPPER; | |
3015 | ||
3016 | } else | |
3017 | m->keep_free = DEFAULT_KEEP_FREE; | |
3018 | } | |
3019 | ||
2b43f939 LP |
3020 | log_debug("Fixed max_use=%s max_size=%s min_size=%s keep_free=%s", |
3021 | format_bytes(a, sizeof(a), m->max_use), | |
3022 | format_bytes(b, sizeof(b), m->max_size), | |
3023 | format_bytes(c, sizeof(c), m->min_size), | |
3024 | format_bytes(d, sizeof(d), m->keep_free)); | |
babfc091 | 3025 | } |
08984293 LP |
3026 | |
3027 | int journal_file_get_cutoff_realtime_usec(JournalFile *f, usec_t *from, usec_t *to) { | |
08984293 LP |
3028 | assert(f); |
3029 | assert(from || to); | |
3030 | ||
3031 | if (from) { | |
162566a4 LP |
3032 | if (f->header->head_entry_realtime == 0) |
3033 | return -ENOENT; | |
08984293 | 3034 | |
162566a4 | 3035 | *from = le64toh(f->header->head_entry_realtime); |
08984293 LP |
3036 | } |
3037 | ||
3038 | if (to) { | |
162566a4 LP |
3039 | if (f->header->tail_entry_realtime == 0) |
3040 | return -ENOENT; | |
08984293 | 3041 | |
162566a4 | 3042 | *to = le64toh(f->header->tail_entry_realtime); |
08984293 LP |
3043 | } |
3044 | ||
3045 | return 1; | |
3046 | } | |
3047 | ||
3048 | int journal_file_get_cutoff_monotonic_usec(JournalFile *f, sd_id128_t boot_id, usec_t *from, usec_t *to) { | |
08984293 LP |
3049 | Object *o; |
3050 | uint64_t p; | |
3051 | int r; | |
3052 | ||
3053 | assert(f); | |
3054 | assert(from || to); | |
3055 | ||
47838ab3 | 3056 | r = find_data_object_by_boot_id(f, boot_id, &o, &p); |
08984293 LP |
3057 | if (r <= 0) |
3058 | return r; | |
3059 | ||
3060 | if (le64toh(o->data.n_entries) <= 0) | |
3061 | return 0; | |
3062 | ||
3063 | if (from) { | |
3064 | r = journal_file_move_to_object(f, OBJECT_ENTRY, le64toh(o->data.entry_offset), &o); | |
3065 | if (r < 0) | |
3066 | return r; | |
3067 | ||
3068 | *from = le64toh(o->entry.monotonic); | |
3069 | } | |
3070 | ||
3071 | if (to) { | |
3072 | r = journal_file_move_to_object(f, OBJECT_DATA, p, &o); | |
3073 | if (r < 0) | |
3074 | return r; | |
3075 | ||
3076 | r = generic_array_get_plus_one(f, | |
3077 | le64toh(o->data.entry_offset), | |
3078 | le64toh(o->data.entry_array_offset), | |
3079 | le64toh(o->data.n_entries)-1, | |
3080 | &o, NULL); | |
3081 | if (r <= 0) | |
3082 | return r; | |
3083 | ||
3084 | *to = le64toh(o->entry.monotonic); | |
3085 | } | |
3086 | ||
3087 | return 1; | |
3088 | } | |
dca6219e | 3089 | |
fb0951b0 | 3090 | bool journal_file_rotate_suggested(JournalFile *f, usec_t max_file_usec) { |
dca6219e LP |
3091 | assert(f); |
3092 | ||
3093 | /* If we gained new header fields we gained new features, | |
3094 | * hence suggest a rotation */ | |
361f9cbc LP |
3095 | if (le64toh(f->header->header_size) < sizeof(Header)) { |
3096 | log_debug("%s uses an outdated header, suggesting rotation.", f->path); | |
dca6219e | 3097 | return true; |
361f9cbc | 3098 | } |
dca6219e LP |
3099 | |
3100 | /* Let's check if the hash tables grew over a certain fill | |
3101 | * level (75%, borrowing this value from Java's hash table | |
3102 | * implementation), and if so suggest a rotation. To calculate | |
3103 | * the fill level we need the n_data field, which only exists | |
3104 | * in newer versions. */ | |
3105 | ||
3106 | if (JOURNAL_HEADER_CONTAINS(f->header, n_data)) | |
361f9cbc | 3107 | if (le64toh(f->header->n_data) * 4ULL > (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)) * 3ULL) { |
507f22bd | 3108 | log_debug("Data hash table of %s has a fill level at %.1f (%"PRIu64" of %"PRIu64" items, %llu file size, %"PRIu64" bytes per hash table item), suggesting rotation.", |
361f9cbc LP |
3109 | f->path, |
3110 | 100.0 * (double) le64toh(f->header->n_data) / ((double) (le64toh(f->header->data_hash_table_size) / sizeof(HashItem))), | |
507f22bd ZJS |
3111 | le64toh(f->header->n_data), |
3112 | le64toh(f->header->data_hash_table_size) / sizeof(HashItem), | |
3113 | (unsigned long long) f->last_stat.st_size, | |
3114 | f->last_stat.st_size / le64toh(f->header->n_data)); | |
dca6219e | 3115 | return true; |
361f9cbc | 3116 | } |
dca6219e LP |
3117 | |
3118 | if (JOURNAL_HEADER_CONTAINS(f->header, n_fields)) | |
361f9cbc | 3119 | if (le64toh(f->header->n_fields) * 4ULL > (le64toh(f->header->field_hash_table_size) / sizeof(HashItem)) * 3ULL) { |
507f22bd | 3120 | log_debug("Field hash table of %s has a fill level at %.1f (%"PRIu64" of %"PRIu64" items), suggesting rotation.", |
361f9cbc LP |
3121 | f->path, |
3122 | 100.0 * (double) le64toh(f->header->n_fields) / ((double) (le64toh(f->header->field_hash_table_size) / sizeof(HashItem))), | |
507f22bd ZJS |
3123 | le64toh(f->header->n_fields), |
3124 | le64toh(f->header->field_hash_table_size) / sizeof(HashItem)); | |
dca6219e | 3125 | return true; |
361f9cbc | 3126 | } |
dca6219e | 3127 | |
0598fd4a LP |
3128 | /* Are the data objects properly indexed by field objects? */ |
3129 | if (JOURNAL_HEADER_CONTAINS(f->header, n_data) && | |
3130 | JOURNAL_HEADER_CONTAINS(f->header, n_fields) && | |
3131 | le64toh(f->header->n_data) > 0 && | |
3132 | le64toh(f->header->n_fields) == 0) | |
3133 | return true; | |
3134 | ||
fb0951b0 LP |
3135 | if (max_file_usec > 0) { |
3136 | usec_t t, h; | |
3137 | ||
3138 | h = le64toh(f->header->head_entry_realtime); | |
3139 | t = now(CLOCK_REALTIME); | |
3140 | ||
3141 | if (h > 0 && t > h + max_file_usec) | |
3142 | return true; | |
3143 | } | |
3144 | ||
dca6219e LP |
3145 | return false; |
3146 | } |