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
25 #include <ccan/likely/likely.h>
26 #include <ccan/compiler/compiler.h>
27 #include <ccan/endian/endian.h>
32 #include "system/filesys.h"
33 #include "system/time.h"
34 #include "system/shmem.h"
35 #include "system/select.h"
36 #include "system/wait.h"
57 /* #define TDB_TRACE 1 */
60 #define __STRING(x) #x
63 #ifndef __STRINGSTRING
64 #define __STRINGSTRING(x) __STRING(x)
68 #define __location__ __FILE__ ":" __STRINGSTRING(__LINE__)
71 typedef uint64_t tdb_len_t
;
72 typedef uint64_t tdb_off_t
;
74 #define TDB_MAGIC_FOOD "TDB file\n"
75 #define TDB_VERSION ((uint64_t)(0x26011967 + 7))
76 #define TDB_USED_MAGIC ((uint64_t)0x1999)
77 #define TDB_HTABLE_MAGIC ((uint64_t)0x1888)
78 #define TDB_CHAIN_MAGIC ((uint64_t)0x1777)
79 #define TDB_FTABLE_MAGIC ((uint64_t)0x1666)
80 #define TDB_FREE_MAGIC ((uint64_t)0xFE)
81 #define TDB_HASH_MAGIC (0xA1ABE11A01092008ULL)
82 #define TDB_RECOVERY_MAGIC (0xf53bc0e7ad124589ULL)
83 #define TDB_RECOVERY_INVALID_MAGIC (0x0ULL)
85 #define TDB_OFF_IS_ERR(off) unlikely(off >= (tdb_off_t)TDB_ERR_LAST)
87 /* Packing errors into pointers and v.v. */
88 #define TDB_PTR_IS_ERR(ptr) \
89 unlikely((unsigned long)(ptr) >= (unsigned long)TDB_ERR_LAST)
90 #define TDB_PTR_ERR(p) ((enum TDB_ERROR)(long)(p))
91 #define TDB_ERR_PTR(err) ((void *)(long)(err))
93 /* Common case of returning true, false or -ve error. */
94 typedef int tdb_bool_err
;
96 /* Prevent others from opening the file. */
97 #define TDB_OPEN_LOCK 0
98 /* Doing a transaction. */
99 #define TDB_TRANSACTION_LOCK 1
100 /* Expanding file. */
101 #define TDB_EXPANSION_LOCK 2
102 /* Hash chain locks. */
103 #define TDB_HASH_LOCK_START 64
105 /* Range for hash locks. */
106 #define TDB_HASH_LOCK_RANGE_BITS 30
107 #define TDB_HASH_LOCK_RANGE (1 << TDB_HASH_LOCK_RANGE_BITS)
109 /* We have 1024 entries in the top level. */
110 #define TDB_TOPLEVEL_HASH_BITS 10
111 /* And 64 entries in each sub-level: thus 64 bits exactly after 9 levels. */
112 #define TDB_SUBLEVEL_HASH_BITS 6
113 /* And 8 entries in each group, ie 8 groups per sublevel. */
114 #define TDB_HASH_GROUP_BITS 3
115 /* This is currently 10: beyond this we chain. */
116 #define TDB_MAX_LEVELS (1+(64-TDB_TOPLEVEL_HASH_BITS) / TDB_SUBLEVEL_HASH_BITS)
118 /* Extend file by least 100 times larger than needed. */
119 #define TDB_EXTENSION_FACTOR 100
121 /* We steal bits from the offsets to store hash info. */
122 #define TDB_OFF_HASH_GROUP_MASK ((1ULL << TDB_HASH_GROUP_BITS) - 1)
123 /* We steal this many upper bits, giving a maximum offset of 64 exabytes. */
124 #define TDB_OFF_UPPER_STEAL 8
125 #define TDB_OFF_UPPER_STEAL_EXTRA 7
126 /* The bit number where we store extra hash bits. */
127 #define TDB_OFF_HASH_EXTRA_BIT 57
128 #define TDB_OFF_UPPER_STEAL_SUBHASH_BIT 56
130 /* Additional features we understand. Currently: none. */
131 #define TDB_FEATURE_MASK ((uint64_t)0)
133 /* The bit number where we store the extra hash bits. */
134 /* Convenience mask to get actual offset. */
135 #define TDB_OFF_MASK \
136 (((1ULL << (64 - TDB_OFF_UPPER_STEAL)) - 1) - TDB_OFF_HASH_GROUP_MASK)
138 /* How many buckets in a free list: see size_to_bucket(). */
139 #define TDB_FREE_BUCKETS (64 - TDB_OFF_UPPER_STEAL)
141 /* We have to be able to fit a free record here. */
142 #define TDB_MIN_DATA_LEN \
143 (sizeof(struct tdb_free_record) - sizeof(struct tdb_used_record))
145 /* Indicates this entry is not on an flist (can happen during coalescing) */
146 #define TDB_FTABLE_NONE ((1ULL << TDB_OFF_UPPER_STEAL) - 1)
148 struct tdb_used_record
{
149 /* For on-disk compatibility, we avoid bitfields:
155 uint64_t magic_and_meta
;
156 /* The bottom key_len_bits*2 are key length, rest is data length. */
157 uint64_t key_and_data_len
;
160 static inline unsigned rec_key_bits(const struct tdb_used_record
*r
)
162 return ((r
->magic_and_meta
>> 43) & ((1 << 5)-1)) * 2;
165 static inline uint64_t rec_key_length(const struct tdb_used_record
*r
)
167 return r
->key_and_data_len
& ((1ULL << rec_key_bits(r
)) - 1);
170 static inline uint64_t rec_data_length(const struct tdb_used_record
*r
)
172 return r
->key_and_data_len
>> rec_key_bits(r
);
175 static inline uint64_t rec_extra_padding(const struct tdb_used_record
*r
)
177 return (r
->magic_and_meta
>> 11) & 0xFFFFFFFF;
180 static inline uint32_t rec_hash(const struct tdb_used_record
*r
)
182 return r
->magic_and_meta
& ((1 << 11) - 1);
185 static inline uint16_t rec_magic(const struct tdb_used_record
*r
)
187 return (r
->magic_and_meta
>> 48);
190 struct tdb_free_record
{
191 uint64_t magic_and_prev
; /* TDB_OFF_UPPER_STEAL bits magic, then prev */
192 uint64_t ftable_and_len
; /* Len not counting these two fields. */
193 /* This is why the minimum record size is 8 bytes. */
197 static inline uint64_t frec_prev(const struct tdb_free_record
*f
)
199 return f
->magic_and_prev
& ((1ULL << (64 - TDB_OFF_UPPER_STEAL
)) - 1);
202 static inline uint64_t frec_magic(const struct tdb_free_record
*f
)
204 return f
->magic_and_prev
>> (64 - TDB_OFF_UPPER_STEAL
);
207 static inline uint64_t frec_len(const struct tdb_free_record
*f
)
209 return f
->ftable_and_len
& ((1ULL << (64 - TDB_OFF_UPPER_STEAL
))-1);
212 static inline unsigned frec_ftable(const struct tdb_free_record
*f
)
214 return f
->ftable_and_len
>> (64 - TDB_OFF_UPPER_STEAL
);
217 struct tdb_recovery_record
{
219 /* Length of record (add this header to get total length). */
223 /* Old length of file before transaction. */
227 /* If we bottom out of the subhashes, we chain. */
229 tdb_off_t rec
[1 << TDB_HASH_GROUP_BITS
];
233 /* this is stored at the front of every database */
235 char magic_food
[64]; /* for /etc/magic */
236 /* FIXME: Make me 32 bit? */
237 uint64_t version
; /* version of the code */
238 uint64_t hash_test
; /* result of hashing HASH_MAGIC. */
239 uint64_t hash_seed
; /* "random" seed written at creation time. */
240 tdb_off_t free_table
; /* (First) free table. */
241 tdb_off_t recovery
; /* Transaction recovery area. */
243 uint64_t features_used
; /* Features all writers understand */
244 uint64_t features_offered
; /* Features offered */
246 uint64_t seqnum
; /* Sequence number for TDB_SEQNUM */
248 tdb_off_t reserved
[23];
250 /* Top level hash table. */
251 tdb_off_t hashtable
[1ULL << TDB_TOPLEVEL_HASH_BITS
];
254 struct tdb_freetable
{
255 struct tdb_used_record hdr
;
257 tdb_off_t buckets
[TDB_FREE_BUCKETS
];
260 /* Information about a particular (locked) hash entry. */
262 /* Full hash value of entry. */
264 /* Start and length of lock acquired. */
265 tdb_off_t hlock_start
;
266 tdb_len_t hlock_range
;
267 /* Start of hash group. */
268 tdb_off_t group_start
;
269 /* Bucket we belong in. */
270 unsigned int home_bucket
;
271 /* Bucket we (or an empty space) were found in. */
272 unsigned int found_bucket
;
273 /* How many bits of the hash are already used. */
274 unsigned int hash_used
;
275 /* Current working group. */
276 tdb_off_t group
[1 << TDB_HASH_GROUP_BITS
];
279 struct traverse_info
{
280 struct traverse_level
{
282 /* We ignore groups here, and treat it as a big array. */
284 unsigned int total_buckets
;
285 } levels
[TDB_MAX_LEVELS
+ 1];
286 unsigned int num_levels
;
287 unsigned int toplevel_group
;
288 /* This makes delete-everything-inside-traverse work as expected. */
292 enum tdb_lock_flags
{
293 /* WAIT == F_SETLKW, NOWAIT == F_SETLK */
296 /* If set, don't log an error on failure. */
298 /* If set, don't check for recovery (used by recovery code). */
299 TDB_LOCK_NOCHECK
= 4,
303 struct tdb_context
*owner
;
309 /* This is only needed for tdb_access_commit, but used everywhere to
311 struct tdb_access_hdr
{
312 struct tdb_access_hdr
*next
;
319 /* Single list of all TDBs, to detect multiple opens. */
320 struct tdb_file
*next
;
322 /* How many are sharing us? */
325 /* Mmap (if any), or malloc (for TDB_INTERNAL). */
328 /* How much space has been mapped (<= current file size) */
331 /* The file descriptor (-1 for TDB_INTERNAL). */
334 /* Lock information */
336 struct tdb_lock allrecord_lock
;
338 struct tdb_lock
*lockrecs
;
340 /* Identity of this file. */
346 /* Filename of the database. */
349 /* Are we accessing directly? (debugging check). */
352 /* Operating read-only? (Opened O_RDONLY, or in traverse_read) */
355 /* mmap read only? */
358 /* the flags passed to tdb_open, for tdb_reopen. */
361 /* Logging function */
362 void (*log_fn
)(struct tdb_context
*tdb
,
363 enum tdb_log_level level
,
369 uint64_t (*hash_fn
)(const void *key
, size_t len
, uint64_t seed
, void *);
373 /* low level (fnctl) lock functions. */
374 int (*lock_fn
)(int fd
, int rw
, off_t off
, off_t len
, bool w
, void *);
375 int (*unlock_fn
)(int fd
, int rw
, off_t off
, off_t len
, void *);
378 /* Set if we are in a transaction. */
379 struct tdb_transaction
*transaction
;
381 /* What free table are we using? */
382 tdb_off_t ftable_off
;
385 /* IO methods: changes for transactions. */
386 const struct tdb_methods
*methods
;
388 /* Our statistics. */
389 struct tdb_attribute_stats stats
;
391 /* Direct access information */
392 struct tdb_access_hdr
*access
;
394 /* Last error we returned. */
395 enum TDB_ERROR last_error
;
397 /* The actual file information */
398 struct tdb_file
*file
;
402 enum TDB_ERROR (*tread
)(struct tdb_context
*, tdb_off_t
, void *,
404 enum TDB_ERROR (*twrite
)(struct tdb_context
*, tdb_off_t
, const void *,
406 enum TDB_ERROR (*oob
)(struct tdb_context
*, tdb_off_t
, bool);
407 enum TDB_ERROR (*expand_file
)(struct tdb_context
*, tdb_len_t
);
408 void *(*direct
)(struct tdb_context
*, tdb_off_t
, size_t, bool);
415 tdb_bool_err
first_in_hash(struct tdb_context
*tdb
,
416 struct traverse_info
*tinfo
,
417 TDB_DATA
*kbuf
, size_t *dlen
);
419 tdb_bool_err
next_in_hash(struct tdb_context
*tdb
,
420 struct traverse_info
*tinfo
,
421 TDB_DATA
*kbuf
, size_t *dlen
);
423 /* Hash random memory. */
424 uint64_t tdb_hash(struct tdb_context
*tdb
, const void *ptr
, size_t len
);
427 uint64_t hash_record(struct tdb_context
*tdb
, tdb_off_t off
);
429 /* Find and lock a hash entry (or where it would be). */
430 tdb_off_t
find_and_lock(struct tdb_context
*tdb
,
434 struct tdb_used_record
*rec
,
435 struct traverse_info
*tinfo
);
437 enum TDB_ERROR
replace_in_hash(struct tdb_context
*tdb
,
441 enum TDB_ERROR
add_to_hash(struct tdb_context
*tdb
, struct hash_info
*h
,
444 enum TDB_ERROR
delete_from_hash(struct tdb_context
*tdb
, struct hash_info
*h
);
447 bool is_subhash(tdb_off_t val
);
450 enum TDB_ERROR
tdb_ftable_init(struct tdb_context
*tdb
);
452 /* check.c needs these to iterate through free lists. */
453 tdb_off_t
first_ftable(struct tdb_context
*tdb
);
454 tdb_off_t
next_ftable(struct tdb_context
*tdb
, tdb_off_t ftable
);
456 /* This returns space or -ve error number. */
457 tdb_off_t
alloc(struct tdb_context
*tdb
, size_t keylen
, size_t datalen
,
458 uint64_t hash
, unsigned magic
, bool growing
);
460 /* Put this record in a free list. */
461 enum TDB_ERROR
add_free_record(struct tdb_context
*tdb
,
462 tdb_off_t off
, tdb_len_t len_with_header
,
463 enum tdb_lock_flags waitflag
,
466 /* Set up header for a used/ftable/htable/chain record. */
467 enum TDB_ERROR
set_header(struct tdb_context
*tdb
,
468 struct tdb_used_record
*rec
,
469 unsigned magic
, uint64_t keylen
, uint64_t datalen
,
470 uint64_t actuallen
, unsigned hashlow
);
472 /* Used by tdb_check to verify. */
473 unsigned int size_to_bucket(tdb_len_t data_len
);
474 tdb_off_t
bucket_off(tdb_off_t ftable_off
, unsigned bucket
);
476 /* Used by tdb_summary */
477 tdb_off_t
dead_space(struct tdb_context
*tdb
, tdb_off_t off
);
480 /* Initialize tdb->methods. */
481 void tdb_io_init(struct tdb_context
*tdb
);
483 /* Convert endian of the buffer if required. */
484 void *tdb_convert(const struct tdb_context
*tdb
, void *buf
, tdb_len_t size
);
486 /* Unmap and try to map the tdb. */
487 void tdb_munmap(struct tdb_file
*file
);
488 void tdb_mmap(struct tdb_context
*tdb
);
490 /* Either alloc a copy, or give direct access. Release frees or noop. */
491 const void *tdb_access_read(struct tdb_context
*tdb
,
492 tdb_off_t off
, tdb_len_t len
, bool convert
);
493 void *tdb_access_write(struct tdb_context
*tdb
,
494 tdb_off_t off
, tdb_len_t len
, bool convert
);
496 /* Release result of tdb_access_read/write. */
497 void tdb_access_release(struct tdb_context
*tdb
, const void *p
);
498 /* Commit result of tdb_acces_write. */
499 enum TDB_ERROR
tdb_access_commit(struct tdb_context
*tdb
, void *p
);
501 /* Convenience routine to get an offset. */
502 tdb_off_t
tdb_read_off(struct tdb_context
*tdb
, tdb_off_t off
);
504 /* Write an offset at an offset. */
505 enum TDB_ERROR
tdb_write_off(struct tdb_context
*tdb
, tdb_off_t off
,
508 /* Clear an ondisk area. */
509 enum TDB_ERROR
zero_out(struct tdb_context
*tdb
, tdb_off_t off
, tdb_len_t len
);
511 /* Return a non-zero offset between >= start < end in this array (or end). */
512 tdb_off_t
tdb_find_nonzero_off(struct tdb_context
*tdb
,
517 /* Return a zero offset in this array, or num. */
518 tdb_off_t
tdb_find_zero_off(struct tdb_context
*tdb
, tdb_off_t off
,
521 /* Allocate and make a copy of some offset. */
522 void *tdb_alloc_read(struct tdb_context
*tdb
, tdb_off_t offset
, tdb_len_t len
);
524 /* Writes a converted copy of a record. */
525 enum TDB_ERROR
tdb_write_convert(struct tdb_context
*tdb
, tdb_off_t off
,
526 const void *rec
, size_t len
);
528 /* Reads record and converts it */
529 enum TDB_ERROR
tdb_read_convert(struct tdb_context
*tdb
, tdb_off_t off
,
530 void *rec
, size_t len
);
532 /* Bump the seqnum (caller checks for tdb->flags & TDB_SEQNUM) */
533 void tdb_inc_seqnum(struct tdb_context
*tdb
);
536 /* Lock/unlock a range of hashes. */
537 enum TDB_ERROR
tdb_lock_hashes(struct tdb_context
*tdb
,
538 tdb_off_t hash_lock
, tdb_len_t hash_range
,
539 int ltype
, enum tdb_lock_flags waitflag
);
540 enum TDB_ERROR
tdb_unlock_hashes(struct tdb_context
*tdb
,
542 tdb_len_t hash_range
, int ltype
);
544 /* For closing the file. */
545 void tdb_lock_cleanup(struct tdb_context
*tdb
);
547 /* Lock/unlock a particular free bucket. */
548 enum TDB_ERROR
tdb_lock_free_bucket(struct tdb_context
*tdb
, tdb_off_t b_off
,
549 enum tdb_lock_flags waitflag
);
550 void tdb_unlock_free_bucket(struct tdb_context
*tdb
, tdb_off_t b_off
);
552 /* Serialize transaction start. */
553 enum TDB_ERROR
tdb_transaction_lock(struct tdb_context
*tdb
, int ltype
);
554 void tdb_transaction_unlock(struct tdb_context
*tdb
, int ltype
);
556 /* Do we have any hash locks (ie. via tdb_chainlock) ? */
557 bool tdb_has_hash_locks(struct tdb_context
*tdb
);
559 /* Lock entire database. */
560 enum TDB_ERROR
tdb_allrecord_lock(struct tdb_context
*tdb
, int ltype
,
561 enum tdb_lock_flags flags
, bool upgradable
);
562 void tdb_allrecord_unlock(struct tdb_context
*tdb
, int ltype
);
563 enum TDB_ERROR
tdb_allrecord_upgrade(struct tdb_context
*tdb
);
565 /* Serialize db open. */
566 enum TDB_ERROR
tdb_lock_open(struct tdb_context
*tdb
,
567 int ltype
, enum tdb_lock_flags flags
);
568 void tdb_unlock_open(struct tdb_context
*tdb
, int ltype
);
569 bool tdb_has_open_lock(struct tdb_context
*tdb
);
571 /* Serialize db expand. */
572 enum TDB_ERROR
tdb_lock_expand(struct tdb_context
*tdb
, int ltype
);
573 void tdb_unlock_expand(struct tdb_context
*tdb
, int ltype
);
574 bool tdb_has_expansion_lock(struct tdb_context
*tdb
);
576 /* If it needs recovery, grab all the locks and do it. */
577 enum TDB_ERROR
tdb_lock_and_recover(struct tdb_context
*tdb
);
579 /* Default lock and unlock functions. */
580 int tdb_fcntl_lock(int fd
, int rw
, off_t off
, off_t len
, bool waitflag
, void *);
581 int tdb_fcntl_unlock(int fd
, int rw
, off_t off
, off_t len
, void *);
584 enum TDB_ERROR
tdb_transaction_recover(struct tdb_context
*tdb
);
585 tdb_bool_err
tdb_needs_recovery(struct tdb_context
*tdb
);
588 enum TDB_ERROR COLD
tdb_logerr(struct tdb_context
*tdb
,
589 enum TDB_ERROR ecode
,
590 enum tdb_log_level level
,
591 const char *fmt
, ...);
594 void tdb_trace(struct tdb_context
*tdb
, const char *op
);
595 void tdb_trace_seqnum(struct tdb_context
*tdb
, uint32_t seqnum
, const char *op
);
596 void tdb_trace_open(struct tdb_context
*tdb
, const char *op
,
597 unsigned hash_size
, unsigned tdb_flags
, unsigned open_flags
);
598 void tdb_trace_ret(struct tdb_context
*tdb
, const char *op
, int ret
);
599 void tdb_trace_retrec(struct tdb_context
*tdb
, const char *op
, TDB_DATA ret
);
600 void tdb_trace_1rec(struct tdb_context
*tdb
, const char *op
,
602 void tdb_trace_1rec_ret(struct tdb_context
*tdb
, const char *op
,
603 TDB_DATA rec
, int ret
);
604 void tdb_trace_1rec_retrec(struct tdb_context
*tdb
, const char *op
,
605 TDB_DATA rec
, TDB_DATA ret
);
606 void tdb_trace_2rec_flag_ret(struct tdb_context
*tdb
, const char *op
,
607 TDB_DATA rec1
, TDB_DATA rec2
, unsigned flag
,
609 void tdb_trace_2rec_retrec(struct tdb_context
*tdb
, const char *op
,
610 TDB_DATA rec1
, TDB_DATA rec2
, TDB_DATA ret
);
612 #define tdb_trace(tdb, op)
613 #define tdb_trace_seqnum(tdb, seqnum, op)
614 #define tdb_trace_open(tdb, op, hash_size, tdb_flags, open_flags)
615 #define tdb_trace_ret(tdb, op, ret)
616 #define tdb_trace_retrec(tdb, op, ret)
617 #define tdb_trace_1rec(tdb, op, rec)
618 #define tdb_trace_1rec_ret(tdb, op, rec, ret)
619 #define tdb_trace_1rec_retrec(tdb, op, rec, ret)
620 #define tdb_trace_2rec_flag_ret(tdb, op, rec1, rec2, flag, ret)
621 #define tdb_trace_2rec_retrec(tdb, op, rec1, rec2, ret)
622 #endif /* !TDB_TRACE */