#include <stdint.h>
#include <stdbool.h>
#include <stdlib.h>
+#include <stddef.h>
#include <sys/time.h>
#include <sys/mman.h>
#include <unistd.h>
typedef uint64_t tdb_len_t;
typedef uint64_t tdb_off_t;
-#ifndef offsetof
-#define offsetof(t,f) ((unsigned int)&((t *)0)->f)
-#endif
-
#define TDB_MAGIC_FOOD "TDB file\n"
#define TDB_VERSION ((uint64_t)(0x26011967 + 7))
#define TDB_MAGIC ((uint64_t)0x1999)
-#define TDB_FREE_MAGIC ((~(uint64_t)TDB_MAGIC) << 6)
+#define TDB_FREE_MAGIC ((uint64_t)0xFE)
+#define TDB_COALESCING_MAGIC ((uint64_t)0xFD)
#define TDB_HASH_MAGIC (0xA1ABE11A01092008ULL)
-#define TDB_RECOVERY_MAGIC (0xf53bc0e7U)
-#define TDB_RECOVERY_INVALID_MAGIC (0x0)
+#define TDB_RECOVERY_MAGIC (0xf53bc0e7ad124589ULL)
+#define TDB_RECOVERY_INVALID_MAGIC (0x0ULL)
#define TDB_OFF_ERR ((tdb_off_t)-1)
/* Hash chain locks. */
#define TDB_HASH_LOCK_START 3
-/* We start wih 256 hash buckets, and a 64k-sized zone. */
-#define INITIAL_HASH_BITS 8
-#define INITIAL_ZONE_BITS 16
-
-/* Try to create zones at least 32 times larger than allocations. */
-#define TDB_COMFORT_FACTOR_BITS 5
-
-/* We ensure buckets up to size 1 << (zone_bits - TDB_COMFORT_FACTOR_BITS). */
-/* FIXME: test this matches size_to_bucket! */
-#define BUCKETS_FOR_ZONE(zone_bits) ((zone_bits) + 2 - TDB_COMFORT_FACTOR_BITS)
+/* Range for hash locks. */
+#define TDB_HASH_LOCK_RANGE_BITS 30
+#define TDB_HASH_LOCK_RANGE (1 << TDB_HASH_LOCK_RANGE_BITS)
+
+/* We have 1024 entries in the top level. */
+#define TDB_TOPLEVEL_HASH_BITS 10
+/* And 64 entries in each sub-level: thus 64 bits exactly after 9 levels. */
+#define TDB_SUBLEVEL_HASH_BITS 6
+/* And 8 entries in each group, ie 8 groups per sublevel. */
+#define TDB_HASH_GROUP_BITS 3
+
+/* Extend file by least 100 times larger than needed. */
+#define TDB_EXTENSION_FACTOR 100
+
+/* We steal bits from the offsets to store hash info. */
+#define TDB_OFF_HASH_GROUP_MASK ((1ULL << TDB_HASH_GROUP_BITS) - 1)
+/* We steal this many upper bits, giving a maximum offset of 64 exabytes. */
+#define TDB_OFF_UPPER_STEAL 8
+#define TDB_OFF_UPPER_STEAL_EXTRA 7
+#define TDB_OFF_UPPER_STEAL_TRUNCBIT 1
+/* If this is set, hash is truncated (only 1 bit is valid). */
+#define TDB_OFF_HASH_TRUNCATED_BIT 56
+/* The bit number where we store next level of hash. */
+#define TDB_OFF_HASH_EXTRA_BIT 57
+/* Convenience mask to get actual offset. */
+#define TDB_OFF_MASK \
+ (((1ULL << (64 - TDB_OFF_UPPER_STEAL)) - 1) - TDB_OFF_HASH_GROUP_MASK)
+
+/* How many buckets in a free list: see size_to_bucket(). */
+#define TDB_FREE_BUCKETS (64 - TDB_OFF_UPPER_STEAL)
+
+/* We have to be able to fit a free record here. */
+#define TDB_MIN_DATA_LEN \
+ (sizeof(struct tdb_free_record) - sizeof(struct tdb_used_record))
#if !HAVE_BSWAP_64
static inline uint64_t bswap_64(uint64_t x)
magic: 16, (highest)
key_len_bits: 5,
extra_padding: 32
- hash_bits: 5,
- zone_bits: 6 (lowest)
+ hash_bits: 11
*/
uint64_t magic_and_meta;
/* The bottom key_len_bits*2 are key length, rest is data length. */
return (r->magic_and_meta >> 11) & 0xFFFFFFFF;
}
-static inline unsigned int rec_zone_bits(const struct tdb_used_record *r)
-{
- return r->magic_and_meta & ((1 << 6) - 1);
-}
-
static inline uint32_t rec_hash(const struct tdb_used_record *r)
{
- return (r->magic_and_meta >> 6) & ((1 << 5) - 1);
+ return r->magic_and_meta & ((1 << 11) - 1);
}
static inline uint16_t rec_magic(const struct tdb_used_record *r)
}
struct tdb_free_record {
- uint64_t magic_and_meta; /* Bottom 6 bits are zone bits. */
- uint64_t data_len; /* Not counting these two fields. */
- /* This is why the minimum record size is 16 bytes. */
- uint64_t next, prev;
+ uint64_t magic_and_prev; /* TDB_OFF_UPPER_STEAL bits magic, then prev */
+ uint64_t flist_and_len; /* Len not counting these two fields. */
+ /* This is why the minimum record size is 8 bytes. */
+ uint64_t next;
};
-static inline unsigned int frec_zone_bits(const struct tdb_free_record *f)
+static inline uint64_t frec_prev(const struct tdb_free_record *f)
{
- return f->magic_and_meta & ((1 << 6) - 1);
+ return f->magic_and_prev & ((1ULL << (64 - TDB_OFF_UPPER_STEAL)) - 1);
}
static inline uint64_t frec_magic(const struct tdb_free_record *f)
{
- return f->magic_and_meta & ~((1ULL << 6) - 1);
+ return f->magic_and_prev >> (64 - TDB_OFF_UPPER_STEAL);
+}
+
+static inline uint64_t frec_len(const struct tdb_free_record *f)
+{
+ return f->flist_and_len & ((1ULL << (64 - TDB_OFF_UPPER_STEAL))-1);
+}
+
+static inline unsigned frec_flist(const struct tdb_free_record *f)
+{
+ return f->flist_and_len >> (64 - TDB_OFF_UPPER_STEAL);
}
-/* These parts can change while we have db open. */
-struct tdb_header_volatile {
- uint64_t generation; /* Makes sure it changes on every update. */
- uint64_t hash_bits; /* Entries in hash table. */
- uint64_t hash_off; /* Offset of hash table. */
+struct tdb_recovery_record {
+ uint64_t magic;
+ /* Length of record. */
+ uint64_t max_len;
+ /* Length used. */
+ uint64_t len;
+ /* Old length of file before transaction. */
+ uint64_t eof;
};
/* this is stored at the front of every database */
struct tdb_header {
- char magic_food[32]; /* for /etc/magic */
+ char magic_food[64]; /* for /etc/magic */
/* FIXME: Make me 32 bit? */
uint64_t version; /* version of the code */
uint64_t hash_test; /* result of hashing HASH_MAGIC. */
uint64_t hash_seed; /* "random" seed written at creation time. */
+ tdb_off_t free_list; /* (First) free list. */
+ tdb_off_t recovery; /* Transaction recovery area. */
+
+ tdb_off_t reserved[26];
+
+ /* Top level hash table. */
+ tdb_off_t hashtable[1ULL << TDB_TOPLEVEL_HASH_BITS];
+};
- struct tdb_header_volatile v;
+struct tdb_freelist {
+ struct tdb_used_record hdr;
+ tdb_off_t next;
+ tdb_off_t buckets[TDB_FREE_BUCKETS];
+};
- tdb_off_t reserved[19];
+/* Information about a particular (locked) hash entry. */
+struct hash_info {
+ /* Full hash value of entry. */
+ uint64_t h;
+ /* Start and length of lock acquired. */
+ tdb_off_t hlock_start;
+ tdb_len_t hlock_range;
+ /* Start of hash group. */
+ tdb_off_t group_start;
+ /* Bucket we belong in. */
+ unsigned int home_bucket;
+ /* Bucket we (or an empty space) were found in. */
+ unsigned int found_bucket;
+ /* How many bits of the hash are already used. */
+ unsigned int hash_used;
+ /* Current working group. */
+ tdb_off_t group[1 << TDB_HASH_GROUP_BITS];
};
-/* Each zone has its set of free lists at the head.
- *
- * For each zone we have a series of per-size buckets, and a final bucket for
- * "too big". */
-struct free_zone_header {
- /* How much does this zone cover? */
- uint64_t zone_bits;
- /* tdb_off_t buckets[free_buckets + 1] */
+struct traverse_info {
+ struct traverse_level {
+ tdb_off_t hashtable;
+ /* We ignore groups here, and treat it as a big array. */
+ unsigned entry;
+ unsigned int total_buckets;
+ } levels[64 / TDB_SUBLEVEL_HASH_BITS];
+ unsigned int num_levels;
+ unsigned int toplevel_group;
+ /* This makes delete-everything-inside-traverse work as expected. */
+ tdb_off_t prev;
};
enum tdb_lock_flags {
TDB_LOCK_WAIT = 1,
/* If set, don't log an error on failure. */
TDB_LOCK_PROBE = 2,
+ /* If set, don't check for recovery (used by recovery code). */
+ TDB_LOCK_NOCHECK = 4,
};
struct tdb_lock_type {
/* Mmap (if any), or malloc (for TDB_INTERNAL). */
void *map_ptr;
+ /* Are we accessing directly? (debugging check). */
+ int direct_access;
+
/* Open file descriptor (undefined for TDB_INTERNAL). */
int fd;
/* How much space has been mapped (<= current file size) */
tdb_len_t map_size;
- /* Opened read-only? */
+ /* Operating read-only? (Opened O_RDONLY, or in traverse_read) */
bool read_only;
+ /* mmap read only? */
+ int mmap_flags;
+
/* Error code for last tdb error. */
enum TDB_ERROR ecode;
- /* A cached copy of the header */
- struct tdb_header header;
- /* (for debugging). */
- bool header_uptodate;
-
/* the flags passed to tdb_open, for tdb_reopen. */
uint32_t flags;
/* Hash function. */
tdb_hashfn_t khash;
void *hash_priv;
+ uint64_t hash_seed;
/* Set if we are in a transaction. */
struct tdb_transaction *transaction;
- /* What zone of the tdb to use, for spreading load. */
- uint64_t zone_off;
- /* Cached copy of zone header. */
- struct free_zone_header zhdr;
+ /* What freelist are we using? */
+ uint64_t flist_off;
+ unsigned int flist;
/* IO methods: changes for transactions. */
const struct tdb_methods *methods;
uint64_t num_lockrecs;
struct tdb_lock_type *lockrecs;
+ struct tdb_attribute_stats *stats;
+
/* Single list of all TDBs, to avoid multiple opens. */
struct tdb_context *next;
dev_t device;
int (*write)(struct tdb_context *, tdb_off_t, const void *, tdb_len_t);
int (*oob)(struct tdb_context *, tdb_off_t, bool);
int (*expand_file)(struct tdb_context *, tdb_len_t);
+ void *(*direct)(struct tdb_context *, tdb_off_t, size_t);
};
/*
internal prototypes
*/
-/* tdb.c: */
-/* Returns true if header changed (and updates it). */
-bool header_changed(struct tdb_context *tdb);
-
-/* Commit header to disk. */
-int write_header(struct tdb_context *tdb);
+/* hash.c: */
+void tdb_hash_init(struct tdb_context *tdb);
/* Hash random memory. */
uint64_t tdb_hash(struct tdb_context *tdb, const void *ptr, size_t len);
-/* offset of hash table entry for this list/hash value */
-tdb_off_t hash_off(struct tdb_context *tdb, uint64_t list);
+/* Hash on disk. */
+uint64_t hash_record(struct tdb_context *tdb, tdb_off_t off);
+
+/* Find and lock a hash entry (or where it would be). */
+tdb_off_t find_and_lock(struct tdb_context *tdb,
+ struct tdb_data key,
+ int ltype,
+ struct hash_info *h,
+ struct tdb_used_record *rec,
+ struct traverse_info *tinfo);
+
+int replace_in_hash(struct tdb_context *tdb,
+ struct hash_info *h,
+ tdb_off_t new_off);
+
+int add_to_hash(struct tdb_context *tdb, struct hash_info *h,
+ tdb_off_t new_off);
+int delete_from_hash(struct tdb_context *tdb, struct hash_info *h);
+
+/* For tdb_check */
+bool is_subhash(tdb_off_t val);
/* free.c: */
-int tdb_zone_init(struct tdb_context *tdb);
+int tdb_flist_init(struct tdb_context *tdb);
+
+/* check.c needs these to iterate through free lists. */
+tdb_off_t first_flist(struct tdb_context *tdb);
+tdb_off_t next_flist(struct tdb_context *tdb, tdb_off_t flist);
/* If this fails, try tdb_expand. */
tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
uint64_t hash, bool growing);
/* Put this record in a free list. */
-int add_free_record(struct tdb_context *tdb, unsigned int zone_bits,
+int add_free_record(struct tdb_context *tdb,
tdb_off_t off, tdb_len_t len_with_header);
/* Set up header for a used record. */
-int set_header(struct tdb_context *tdb,
- struct tdb_used_record *rec,
- uint64_t keylen, uint64_t datalen,
- uint64_t actuallen, uint64_t hash,
- unsigned int zone_bits);
+int set_used_header(struct tdb_context *tdb,
+ struct tdb_used_record *rec,
+ uint64_t keylen, uint64_t datalen,
+ uint64_t actuallen, unsigned hashlow);
/* Used by tdb_check to verify. */
-unsigned int size_to_bucket(unsigned int free_buckets, tdb_len_t data_len);
-tdb_off_t bucket_off(tdb_off_t zone_off, tdb_off_t bucket);
+unsigned int size_to_bucket(tdb_len_t data_len);
+tdb_off_t bucket_off(tdb_off_t flist_off, unsigned bucket);
+
+/* Used by tdb_summary */
+size_t dead_space(struct tdb_context *tdb, tdb_off_t off);
/* io.c: */
/* Initialize tdb->methods. */
void tdb_munmap(struct tdb_context *tdb);
void tdb_mmap(struct tdb_context *tdb);
-/* Either make a copy into pad and return that, or return ptr into mmap.
- * Converts endian (ie. will use pad in that case). */
-void *tdb_get(struct tdb_context *tdb, tdb_off_t off, void *pad, size_t len);
-
/* Either alloc a copy, or give direct access. Release frees or noop. */
const void *tdb_access_read(struct tdb_context *tdb,
tdb_off_t off, tdb_len_t len, bool convert);
+void *tdb_access_write(struct tdb_context *tdb,
+ tdb_off_t off, tdb_len_t len, bool convert);
+
+/* Is this pointer direct? (Otherwise it's malloced) */
+bool is_direct(const struct tdb_context *tdb, const void *p);
+
+/* Release result of tdb_access_read/write. */
void tdb_access_release(struct tdb_context *tdb, const void *p);
+/* Commit result of tdb_acces_write. */
+int tdb_access_commit(struct tdb_context *tdb, void *p);
/* Convenience routine to get an offset. */
tdb_off_t tdb_read_off(struct tdb_context *tdb, tdb_off_t off);
/* Clear an ondisk area. */
int zero_out(struct tdb_context *tdb, tdb_off_t off, tdb_len_t len);
-/* Return a non-zero offset in this array, or num. */
-tdb_off_t tdb_find_nonzero_off(struct tdb_context *tdb, tdb_off_t off,
- uint64_t num);
+/* Return a non-zero offset between >= start < end in this array (or end). */
+tdb_off_t tdb_find_nonzero_off(struct tdb_context *tdb,
+ tdb_off_t base,
+ uint64_t start,
+ uint64_t end);
/* Return a zero offset in this array, or num. */
tdb_off_t tdb_find_zero_off(struct tdb_context *tdb, tdb_off_t off,
int tdb_read_convert(struct tdb_context *tdb, tdb_off_t off,
void *rec, size_t len);
-/* Hash on disk. */
-uint64_t hash_record(struct tdb_context *tdb, tdb_off_t off);
+/* Adds a stat, if it's in range. */
+void add_stat_(struct tdb_context *tdb, uint64_t *stat, size_t val);
+#define add_stat(tdb, statname, val) \
+ do { \
+ if (unlikely((tdb)->stats)) \
+ add_stat_((tdb), &(tdb)->stats->statname, (val)); \
+ } while (0)
/* lock.c: */
void tdb_lock_init(struct tdb_context *tdb);
-/* Lock/unlock a particular hash list. */
-tdb_off_t tdb_lock_list(struct tdb_context *tdb, uint64_t hash,
- int ltype, enum tdb_lock_flags waitflag);
-int tdb_unlock_list(struct tdb_context *tdb, tdb_off_t list, int ltype);
+/* Lock/unlock a range of hashes. */
+int tdb_lock_hashes(struct tdb_context *tdb,
+ tdb_off_t hash_lock, tdb_len_t hash_range,
+ int ltype, enum tdb_lock_flags waitflag);
+int tdb_unlock_hashes(struct tdb_context *tdb,
+ tdb_off_t hash_lock,
+ tdb_len_t hash_range, int ltype);
/* Lock/unlock a particular free bucket. */
int tdb_lock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off,
enum tdb_lock_flags waitflag);
void tdb_unlock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off);
-/* Do we have any locks? */
-bool tdb_has_locks(struct tdb_context *tdb);
+/* Serialize transaction start. */
+int tdb_transaction_lock(struct tdb_context *tdb, int ltype);
+int tdb_transaction_unlock(struct tdb_context *tdb, int ltype);
+
+/* Do we have any hash locks (ie. via tdb_chainlock) ? */
+bool tdb_has_hash_locks(struct tdb_context *tdb);
/* Lock entire database. */
int tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
enum tdb_lock_flags flags, bool upgradable);
int tdb_allrecord_unlock(struct tdb_context *tdb, int ltype);
+int tdb_allrecord_upgrade(struct tdb_context *tdb);
/* Serialize db open. */
-int tdb_lock_open(struct tdb_context *tdb);
+int tdb_lock_open(struct tdb_context *tdb, enum tdb_lock_flags flags);
void tdb_unlock_open(struct tdb_context *tdb);
+bool tdb_has_open_lock(struct tdb_context *tdb);
/* Serialize db expand. */
int tdb_lock_expand(struct tdb_context *tdb, int ltype);
void tdb_unlock_expand(struct tdb_context *tdb, int ltype);
+bool tdb_has_expansion_lock(struct tdb_context *tdb);
+/* If it needs recovery, grab all the locks and do it. */
+int tdb_lock_and_recover(struct tdb_context *tdb);
-#if 0
-/* Low-level locking primitives. */
-int tdb_nest_lock(struct tdb_context *tdb, tdb_off_t offset, int ltype,
- enum tdb_lock_flags flags);
-int tdb_nest_unlock(struct tdb_context *tdb, tdb_off_t offset, int ltype);
+/* traverse.c: */
+int first_in_hash(struct tdb_context *tdb, int ltype,
+ struct traverse_info *tinfo,
+ TDB_DATA *kbuf, size_t *dlen);
+int next_in_hash(struct tdb_context *tdb, int ltype,
+ struct traverse_info *tinfo,
+ TDB_DATA *kbuf, size_t *dlen);
-int tdb_munmap(struct tdb_context *tdb);
-void tdb_mmap(struct tdb_context *tdb);
-int tdb_lock(struct tdb_context *tdb, int list, int ltype);
-int tdb_lock_nonblock(struct tdb_context *tdb, int list, int ltype);
-bool tdb_have_locks(struct tdb_context *tdb);
-int tdb_unlock(struct tdb_context *tdb, int list, int ltype);
-int tdb_brlock(struct tdb_context *tdb,
- int rw_type, tdb_off_t offset, size_t len,
- enum tdb_lock_flags flags);
-int tdb_brunlock(struct tdb_context *tdb,
- int rw_type, tdb_off_t offset, size_t len);
-bool tdb_have_extra_locks(struct tdb_context *tdb);
-void tdb_release_extra_locks(struct tdb_context *tdb);
-int tdb_transaction_lock(struct tdb_context *tdb, int ltype);
-int tdb_transaction_unlock(struct tdb_context *tdb, int ltype);
-int tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
- enum tdb_lock_flags flags, bool upgradable);
-int tdb_allrecord_unlock(struct tdb_context *tdb, int ltype);
-int tdb_allrecord_upgrade(struct tdb_context *tdb);
-int tdb_write_lock_record(struct tdb_context *tdb, tdb_off_t off);
-int tdb_write_unlock_record(struct tdb_context *tdb, tdb_off_t off);
-int tdb_ofs_read(struct tdb_context *tdb, tdb_off_t offset, tdb_off_t *d);
-int tdb_ofs_write(struct tdb_context *tdb, tdb_off_t offset, tdb_off_t *d);
-int tdb_free(struct tdb_context *tdb, tdb_off_t offset, struct tdb_record *rec);
-tdb_off_t tdb_allocate(struct tdb_context *tdb, tdb_len_t length, struct tdb_record *rec);
-int tdb_ofs_read(struct tdb_context *tdb, tdb_off_t offset, tdb_off_t *d);
-int tdb_ofs_write(struct tdb_context *tdb, tdb_off_t offset, tdb_off_t *d);
-int tdb_lock_record(struct tdb_context *tdb, tdb_off_t off);
-int tdb_unlock_record(struct tdb_context *tdb, tdb_off_t off);
+/* transaction.c: */
+int tdb_transaction_recover(struct tdb_context *tdb);
bool tdb_needs_recovery(struct tdb_context *tdb);
-int tdb_rec_read(struct tdb_context *tdb, tdb_off_t offset, struct tdb_record *rec);
-int tdb_rec_write(struct tdb_context *tdb, tdb_off_t offset, struct tdb_record *rec);
-int tdb_do_delete(struct tdb_context *tdb, tdb_off_t rec_ptr, struct tdb_record *rec);
-unsigned char *tdb_alloc_read(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t len);
-int tdb_parse_data(struct tdb_context *tdb, TDB_DATA key,
- tdb_off_t offset, tdb_len_t len,
- int (*parser)(TDB_DATA key, TDB_DATA data,
- void *private_data),
- void *private_data);
-tdb_off_t tdb_find_lock_hash(struct tdb_context *tdb, TDB_DATA key, uint32_t hash, int locktype,
- struct tdb_record *rec);
-void tdb_io_init(struct tdb_context *tdb);
-int tdb_expand(struct tdb_context *tdb, tdb_off_t size);
-int tdb_rec_free_read(struct tdb_context *tdb, tdb_off_t off,
- struct tdb_record *rec);
-#endif
#ifdef TDB_TRACE
void tdb_trace(struct tdb_context *tdb, const char *op);