4 Trivial Database 2: private types and prototypes
5 Copyright (C) Rusty Russell 2010
7 This library is free software; you can redistribute it and/or
8 modify it under the terms of the GNU Lesser General Public
9 License as published by the Free Software Foundation; either
10 version 3 of the License, or (at your option) any later version.
12 This library is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 Lesser General Public License for more details.
17 You should have received a copy of the GNU Lesser General Public
18 License along with this library; if not, see <http://www.gnu.org/licenses/>.
22 #if HAVE_FILE_OFFSET_BITS
23 #define _FILE_OFFSET_BITS 64
38 #include <ccan/tdb2/tdb2.h>
39 #include <ccan/likely/likely.h>
40 #include <ccan/compiler/compiler.h>
41 #include <ccan/endian/endian.h>
47 /* #define TDB_TRACE 1 */
50 #define __STRING(x) #x
53 #ifndef __STRINGSTRING
54 #define __STRINGSTRING(x) __STRING(x)
58 #define __location__ __FILE__ ":" __STRINGSTRING(__LINE__)
61 typedef uint64_t tdb_len_t;
62 typedef uint64_t tdb_off_t;
64 #define TDB_MAGIC_FOOD "TDB file\n"
65 #define TDB_VERSION ((uint64_t)(0x26011967 + 7))
66 #define TDB_USED_MAGIC ((uint64_t)0x1999)
67 #define TDB_HTABLE_MAGIC ((uint64_t)0x1888)
68 #define TDB_CHAIN_MAGIC ((uint64_t)0x1777)
69 #define TDB_FTABLE_MAGIC ((uint64_t)0x1666)
70 #define TDB_FREE_MAGIC ((uint64_t)0xFE)
71 #define TDB_HASH_MAGIC (0xA1ABE11A01092008ULL)
72 #define TDB_RECOVERY_MAGIC (0xf53bc0e7ad124589ULL)
73 #define TDB_RECOVERY_INVALID_MAGIC (0x0ULL)
75 #define TDB_OFF_IS_ERR(off) unlikely(off >= (tdb_off_t)TDB_ERR_LAST)
77 /* Packing errors into pointers and v.v. */
78 #define TDB_PTR_IS_ERR(ptr) \
79 unlikely((unsigned long)(ptr) >= (unsigned long)TDB_ERR_LAST)
80 #define TDB_PTR_ERR(p) ((enum TDB_ERROR)(long)(p))
81 #define TDB_ERR_PTR(err) ((void *)(long)(err))
83 /* Common case of returning true, false or -ve error. */
84 typedef int tdb_bool_err;
86 /* Prevent others from opening the file. */
87 #define TDB_OPEN_LOCK 0
88 /* Doing a transaction. */
89 #define TDB_TRANSACTION_LOCK 1
91 #define TDB_EXPANSION_LOCK 2
92 /* Hash chain locks. */
93 #define TDB_HASH_LOCK_START 64
95 /* Range for hash locks. */
96 #define TDB_HASH_LOCK_RANGE_BITS 30
97 #define TDB_HASH_LOCK_RANGE (1 << TDB_HASH_LOCK_RANGE_BITS)
99 /* We have 1024 entries in the top level. */
100 #define TDB_TOPLEVEL_HASH_BITS 10
101 /* And 64 entries in each sub-level: thus 64 bits exactly after 9 levels. */
102 #define TDB_SUBLEVEL_HASH_BITS 6
103 /* And 8 entries in each group, ie 8 groups per sublevel. */
104 #define TDB_HASH_GROUP_BITS 3
105 /* This is currently 10: beyond this we chain. */
106 #define TDB_MAX_LEVELS (1+(64-TDB_TOPLEVEL_HASH_BITS) / TDB_SUBLEVEL_HASH_BITS)
108 /* Extend file by least 100 times larger than needed. */
109 #define TDB_EXTENSION_FACTOR 100
111 /* We steal bits from the offsets to store hash info. */
112 #define TDB_OFF_HASH_GROUP_MASK ((1ULL << TDB_HASH_GROUP_BITS) - 1)
113 /* We steal this many upper bits, giving a maximum offset of 64 exabytes. */
114 #define TDB_OFF_UPPER_STEAL 8
115 #define TDB_OFF_UPPER_STEAL_EXTRA 7
116 /* The bit number where we store extra hash bits. */
117 #define TDB_OFF_HASH_EXTRA_BIT 57
118 #define TDB_OFF_UPPER_STEAL_SUBHASH_BIT 56
120 /* Additional features we understand. Currently: none. */
121 #define TDB_FEATURE_MASK ((uint64_t)0)
123 /* The bit number where we store the extra hash bits. */
124 /* Convenience mask to get actual offset. */
125 #define TDB_OFF_MASK \
126 (((1ULL << (64 - TDB_OFF_UPPER_STEAL)) - 1) - TDB_OFF_HASH_GROUP_MASK)
128 /* How many buckets in a free list: see size_to_bucket(). */
129 #define TDB_FREE_BUCKETS (64 - TDB_OFF_UPPER_STEAL)
131 /* We have to be able to fit a free record here. */
132 #define TDB_MIN_DATA_LEN \
133 (sizeof(struct tdb_free_record) - sizeof(struct tdb_used_record))
135 /* Indicates this entry is not on an flist (can happen during coalescing) */
136 #define TDB_FTABLE_NONE ((1ULL << TDB_OFF_UPPER_STEAL) - 1)
138 struct tdb_used_record {
139 /* For on-disk compatibility, we avoid bitfields:
145 uint64_t magic_and_meta;
146 /* The bottom key_len_bits*2 are key length, rest is data length. */
147 uint64_t key_and_data_len;
150 static inline unsigned rec_key_bits(const struct tdb_used_record *r)
152 return ((r->magic_and_meta >> 43) & ((1 << 5)-1)) * 2;
155 static inline uint64_t rec_key_length(const struct tdb_used_record *r)
157 return r->key_and_data_len & ((1ULL << rec_key_bits(r)) - 1);
160 static inline uint64_t rec_data_length(const struct tdb_used_record *r)
162 return r->key_and_data_len >> rec_key_bits(r);
165 static inline uint64_t rec_extra_padding(const struct tdb_used_record *r)
167 return (r->magic_and_meta >> 11) & 0xFFFFFFFF;
170 static inline uint32_t rec_hash(const struct tdb_used_record *r)
172 return r->magic_and_meta & ((1 << 11) - 1);
175 static inline uint16_t rec_magic(const struct tdb_used_record *r)
177 return (r->magic_and_meta >> 48);
180 struct tdb_free_record {
181 uint64_t magic_and_prev; /* TDB_OFF_UPPER_STEAL bits magic, then prev */
182 uint64_t ftable_and_len; /* Len not counting these two fields. */
183 /* This is why the minimum record size is 8 bytes. */
187 static inline uint64_t frec_prev(const struct tdb_free_record *f)
189 return f->magic_and_prev & ((1ULL << (64 - TDB_OFF_UPPER_STEAL)) - 1);
192 static inline uint64_t frec_magic(const struct tdb_free_record *f)
194 return f->magic_and_prev >> (64 - TDB_OFF_UPPER_STEAL);
197 static inline uint64_t frec_len(const struct tdb_free_record *f)
199 return f->ftable_and_len & ((1ULL << (64 - TDB_OFF_UPPER_STEAL))-1);
202 static inline unsigned frec_ftable(const struct tdb_free_record *f)
204 return f->ftable_and_len >> (64 - TDB_OFF_UPPER_STEAL);
207 struct tdb_recovery_record {
209 /* Length of record (add this header to get total length). */
213 /* Old length of file before transaction. */
217 /* If we bottom out of the subhashes, we chain. */
219 tdb_off_t rec[1 << TDB_HASH_GROUP_BITS];
223 /* this is stored at the front of every database */
225 char magic_food[64]; /* for /etc/magic */
226 /* FIXME: Make me 32 bit? */
227 uint64_t version; /* version of the code */
228 uint64_t hash_test; /* result of hashing HASH_MAGIC. */
229 uint64_t hash_seed; /* "random" seed written at creation time. */
230 tdb_off_t free_table; /* (First) free table. */
231 tdb_off_t recovery; /* Transaction recovery area. */
233 uint64_t features_used; /* Features all writers understand */
234 uint64_t features_offered; /* Features offered */
236 uint64_t seqnum; /* Sequence number for TDB_SEQNUM */
238 tdb_off_t reserved[23];
240 /* Top level hash table. */
241 tdb_off_t hashtable[1ULL << TDB_TOPLEVEL_HASH_BITS];
244 struct tdb_freetable {
245 struct tdb_used_record hdr;
247 tdb_off_t buckets[TDB_FREE_BUCKETS];
250 /* Information about a particular (locked) hash entry. */
252 /* Full hash value of entry. */
254 /* Start and length of lock acquired. */
255 tdb_off_t hlock_start;
256 tdb_len_t hlock_range;
257 /* Start of hash group. */
258 tdb_off_t group_start;
259 /* Bucket we belong in. */
260 unsigned int home_bucket;
261 /* Bucket we (or an empty space) were found in. */
262 unsigned int found_bucket;
263 /* How many bits of the hash are already used. */
264 unsigned int hash_used;
265 /* Current working group. */
266 tdb_off_t group[1 << TDB_HASH_GROUP_BITS];
269 struct traverse_info {
270 struct traverse_level {
272 /* We ignore groups here, and treat it as a big array. */
274 unsigned int total_buckets;
275 } levels[TDB_MAX_LEVELS + 1];
276 unsigned int num_levels;
277 unsigned int toplevel_group;
278 /* This makes delete-everything-inside-traverse work as expected. */
282 enum tdb_lock_flags {
283 /* WAIT == F_SETLKW, NOWAIT == F_SETLK */
286 /* If set, don't log an error on failure. */
288 /* If set, don't check for recovery (used by recovery code). */
289 TDB_LOCK_NOCHECK = 4,
293 struct tdb_context *owner;
299 /* This is only needed for tdb_access_commit, but used everywhere to
301 struct tdb_access_hdr {
302 struct tdb_access_hdr *next;
309 /* Single list of all TDBs, to detect multiple opens. */
310 struct tdb_file *next;
312 /* How many are sharing us? */
315 /* Mmap (if any), or malloc (for TDB_INTERNAL). */
318 /* How much space has been mapped (<= current file size) */
321 /* The file descriptor (-1 for TDB_INTERNAL). */
324 /* Lock information */
326 struct tdb_lock allrecord_lock;
328 struct tdb_lock *lockrecs;
330 /* Identity of this file. */
336 /* Filename of the database. */
339 /* Are we accessing directly? (debugging check). */
342 /* Operating read-only? (Opened O_RDONLY, or in traverse_read) */
345 /* mmap read only? */
348 /* the flags passed to tdb_open, for tdb_reopen. */
351 /* Logging function */
352 void (*log_fn)(struct tdb_context *tdb,
353 enum tdb_log_level level,
359 uint64_t (*hash_fn)(const void *key, size_t len, uint64_t seed, void *);
363 /* low level (fnctl) lock functions. */
364 int (*lock_fn)(int fd, int rw, off_t off, off_t len, bool w, void *);
365 int (*unlock_fn)(int fd, int rw, off_t off, off_t len, void *);
368 /* Set if we are in a transaction. */
369 struct tdb_transaction *transaction;
371 /* What free table are we using? */
372 tdb_off_t ftable_off;
375 /* IO methods: changes for transactions. */
376 const struct tdb_methods *methods;
378 /* Our statistics. */
379 struct tdb_attribute_stats stats;
381 /* Direct access information */
382 struct tdb_access_hdr *access;
384 /* Last error we returned. */
385 enum TDB_ERROR last_error;
387 /* The actual file information */
388 struct tdb_file *file;
392 enum TDB_ERROR (*tread)(struct tdb_context *, tdb_off_t, void *,
394 enum TDB_ERROR (*twrite)(struct tdb_context *, tdb_off_t, const void *,
396 enum TDB_ERROR (*oob)(struct tdb_context *, tdb_off_t, bool);
397 enum TDB_ERROR (*expand_file)(struct tdb_context *, tdb_len_t);
398 void *(*direct)(struct tdb_context *, tdb_off_t, size_t, bool);
405 tdb_bool_err first_in_hash(struct tdb_context *tdb,
406 struct traverse_info *tinfo,
407 TDB_DATA *kbuf, size_t *dlen);
409 tdb_bool_err next_in_hash(struct tdb_context *tdb,
410 struct traverse_info *tinfo,
411 TDB_DATA *kbuf, size_t *dlen);
413 /* Hash random memory. */
414 uint64_t tdb_hash(struct tdb_context *tdb, const void *ptr, size_t len);
417 uint64_t hash_record(struct tdb_context *tdb, tdb_off_t off);
419 /* Find and lock a hash entry (or where it would be). */
420 tdb_off_t find_and_lock(struct tdb_context *tdb,
424 struct tdb_used_record *rec,
425 struct traverse_info *tinfo);
427 enum TDB_ERROR replace_in_hash(struct tdb_context *tdb,
431 enum TDB_ERROR add_to_hash(struct tdb_context *tdb, struct hash_info *h,
434 enum TDB_ERROR delete_from_hash(struct tdb_context *tdb, struct hash_info *h);
437 bool is_subhash(tdb_off_t val);
440 enum TDB_ERROR tdb_ftable_init(struct tdb_context *tdb);
442 /* check.c needs these to iterate through free lists. */
443 tdb_off_t first_ftable(struct tdb_context *tdb);
444 tdb_off_t next_ftable(struct tdb_context *tdb, tdb_off_t ftable);
446 /* This returns space or -ve error number. */
447 tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
448 uint64_t hash, unsigned magic, bool growing);
450 /* Put this record in a free list. */
451 enum TDB_ERROR add_free_record(struct tdb_context *tdb,
452 tdb_off_t off, tdb_len_t len_with_header,
453 enum tdb_lock_flags waitflag,
456 /* Set up header for a used/ftable/htable/chain record. */
457 enum TDB_ERROR set_header(struct tdb_context *tdb,
458 struct tdb_used_record *rec,
459 unsigned magic, uint64_t keylen, uint64_t datalen,
460 uint64_t actuallen, unsigned hashlow);
462 /* Used by tdb_check to verify. */
463 unsigned int size_to_bucket(tdb_len_t data_len);
464 tdb_off_t bucket_off(tdb_off_t ftable_off, unsigned bucket);
466 /* Used by tdb_summary */
467 tdb_off_t dead_space(struct tdb_context *tdb, tdb_off_t off);
470 /* Initialize tdb->methods. */
471 void tdb_io_init(struct tdb_context *tdb);
473 /* Convert endian of the buffer if required. */
474 void *tdb_convert(const struct tdb_context *tdb, void *buf, tdb_len_t size);
476 /* Unmap and try to map the tdb. */
477 void tdb_munmap(struct tdb_file *file);
478 void tdb_mmap(struct tdb_context *tdb);
480 /* Either alloc a copy, or give direct access. Release frees or noop. */
481 const void *tdb_access_read(struct tdb_context *tdb,
482 tdb_off_t off, tdb_len_t len, bool convert);
483 void *tdb_access_write(struct tdb_context *tdb,
484 tdb_off_t off, tdb_len_t len, bool convert);
486 /* Release result of tdb_access_read/write. */
487 void tdb_access_release(struct tdb_context *tdb, const void *p);
488 /* Commit result of tdb_acces_write. */
489 enum TDB_ERROR tdb_access_commit(struct tdb_context *tdb, void *p);
491 /* Convenience routine to get an offset. */
492 tdb_off_t tdb_read_off(struct tdb_context *tdb, tdb_off_t off);
494 /* Write an offset at an offset. */
495 enum TDB_ERROR tdb_write_off(struct tdb_context *tdb, tdb_off_t off,
498 /* Clear an ondisk area. */
499 enum TDB_ERROR zero_out(struct tdb_context *tdb, tdb_off_t off, tdb_len_t len);
501 /* Return a non-zero offset between >= start < end in this array (or end). */
502 tdb_off_t tdb_find_nonzero_off(struct tdb_context *tdb,
507 /* Return a zero offset in this array, or num. */
508 tdb_off_t tdb_find_zero_off(struct tdb_context *tdb, tdb_off_t off,
511 /* Allocate and make a copy of some offset. */
512 void *tdb_alloc_read(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t len);
514 /* Writes a converted copy of a record. */
515 enum TDB_ERROR tdb_write_convert(struct tdb_context *tdb, tdb_off_t off,
516 const void *rec, size_t len);
518 /* Reads record and converts it */
519 enum TDB_ERROR tdb_read_convert(struct tdb_context *tdb, tdb_off_t off,
520 void *rec, size_t len);
522 /* Bump the seqnum (caller checks for tdb->flags & TDB_SEQNUM) */
523 void tdb_inc_seqnum(struct tdb_context *tdb);
526 /* Lock/unlock a range of hashes. */
527 enum TDB_ERROR tdb_lock_hashes(struct tdb_context *tdb,
528 tdb_off_t hash_lock, tdb_len_t hash_range,
529 int ltype, enum tdb_lock_flags waitflag);
530 enum TDB_ERROR tdb_unlock_hashes(struct tdb_context *tdb,
532 tdb_len_t hash_range, int ltype);
534 /* For closing the file. */
535 void tdb_lock_cleanup(struct tdb_context *tdb);
537 /* Lock/unlock a particular free bucket. */
538 enum TDB_ERROR tdb_lock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off,
539 enum tdb_lock_flags waitflag);
540 void tdb_unlock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off);
542 /* Serialize transaction start. */
543 enum TDB_ERROR tdb_transaction_lock(struct tdb_context *tdb, int ltype);
544 void tdb_transaction_unlock(struct tdb_context *tdb, int ltype);
546 /* Do we have any hash locks (ie. via tdb_chainlock) ? */
547 bool tdb_has_hash_locks(struct tdb_context *tdb);
549 /* Lock entire database. */
550 enum TDB_ERROR tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
551 enum tdb_lock_flags flags, bool upgradable);
552 void tdb_allrecord_unlock(struct tdb_context *tdb, int ltype);
553 enum TDB_ERROR tdb_allrecord_upgrade(struct tdb_context *tdb);
555 /* Serialize db open. */
556 enum TDB_ERROR tdb_lock_open(struct tdb_context *tdb,
557 int ltype, enum tdb_lock_flags flags);
558 void tdb_unlock_open(struct tdb_context *tdb, int ltype);
559 bool tdb_has_open_lock(struct tdb_context *tdb);
561 /* Serialize db expand. */
562 enum TDB_ERROR tdb_lock_expand(struct tdb_context *tdb, int ltype);
563 void tdb_unlock_expand(struct tdb_context *tdb, int ltype);
564 bool tdb_has_expansion_lock(struct tdb_context *tdb);
566 /* If it needs recovery, grab all the locks and do it. */
567 enum TDB_ERROR tdb_lock_and_recover(struct tdb_context *tdb);
569 /* Default lock and unlock functions. */
570 int tdb_fcntl_lock(int fd, int rw, off_t off, off_t len, bool waitflag, void *);
571 int tdb_fcntl_unlock(int fd, int rw, off_t off, off_t len, void *);
574 enum TDB_ERROR tdb_transaction_recover(struct tdb_context *tdb);
575 tdb_bool_err tdb_needs_recovery(struct tdb_context *tdb);
578 enum TDB_ERROR COLD tdb_logerr(struct tdb_context *tdb,
579 enum TDB_ERROR ecode,
580 enum tdb_log_level level,
581 const char *fmt, ...);
584 void tdb_trace(struct tdb_context *tdb, const char *op);
585 void tdb_trace_seqnum(struct tdb_context *tdb, uint32_t seqnum, const char *op);
586 void tdb_trace_open(struct tdb_context *tdb, const char *op,
587 unsigned hash_size, unsigned tdb_flags, unsigned open_flags);
588 void tdb_trace_ret(struct tdb_context *tdb, const char *op, int ret);
589 void tdb_trace_retrec(struct tdb_context *tdb, const char *op, TDB_DATA ret);
590 void tdb_trace_1rec(struct tdb_context *tdb, const char *op,
592 void tdb_trace_1rec_ret(struct tdb_context *tdb, const char *op,
593 TDB_DATA rec, int ret);
594 void tdb_trace_1rec_retrec(struct tdb_context *tdb, const char *op,
595 TDB_DATA rec, TDB_DATA ret);
596 void tdb_trace_2rec_flag_ret(struct tdb_context *tdb, const char *op,
597 TDB_DATA rec1, TDB_DATA rec2, unsigned flag,
599 void tdb_trace_2rec_retrec(struct tdb_context *tdb, const char *op,
600 TDB_DATA rec1, TDB_DATA rec2, TDB_DATA ret);
602 #define tdb_trace(tdb, op)
603 #define tdb_trace_seqnum(tdb, seqnum, op)
604 #define tdb_trace_open(tdb, op, hash_size, tdb_flags, open_flags)
605 #define tdb_trace_ret(tdb, op, ret)
606 #define tdb_trace_retrec(tdb, op, ret)
607 #define tdb_trace_1rec(tdb, op, rec)
608 #define tdb_trace_1rec_ret(tdb, op, rec, ret)
609 #define tdb_trace_1rec_retrec(tdb, op, rec, ret)
610 #define tdb_trace_2rec_flag_ret(tdb, op, rec1, rec2, flag, ret)
611 #define tdb_trace_2rec_retrec(tdb, op, rec1, rec2, ret)
612 #endif /* !TDB_TRACE */