/*
the core of tdb_allocate - called when we have decided which
free list entry to use
Note that we try to allocate by grabbing data from the end of an existing record,
not the beginning. This is so the left merge in a free is more likely to be
able to free up the record without fragmentation
*/
static tdb_off_t tdb_allocate_ofs(struct tdb_context *tdb,
tdb_len_t length, tdb_off_t rec_ptr,
struct tdb_record *rec, tdb_off_t last_ptr)
{
#define MIN_REC_SIZE (sizeof(struct tdb_record) + sizeof(tdb_off_t) + 8)
if (rec->rec_len < length + MIN_REC_SIZE) {
/* we have to grab the whole record */
/* unlink it from the previous record */
if (tdb_ofs_write(tdb, last_ptr, &rec->next) == -1) {
return 0;
}
/* mark it not free */
rec->magic = TDB_MAGIC;
if (tdb_rec_write(tdb, rec_ptr, rec) == -1) {
return 0;
}
return rec_ptr;
}
/* we're going to just shorten the existing record */
rec->rec_len -= (length + sizeof(*rec));
if (tdb_rec_write(tdb, rec_ptr, rec) == -1) {
return 0;
}
if (update_tailer(tdb, rec_ptr, rec) == -1) {
return 0;
}
/* and setup the new record */
rec_ptr += sizeof(*rec) + rec->rec_len;
memset(rec, '\0', sizeof(*rec));
rec->rec_len = length;
rec->magic = TDB_MAGIC;
if (tdb_rec_write(tdb, rec_ptr, rec) == -1) {
return 0;
}
if (update_tailer(tdb, rec_ptr, rec) == -1) {
return 0;
}
return rec_ptr;
}
/* actually delete an entry in the database given the offset */
int tdb_do_delete(struct tdb_context *tdb, tdb_off_t rec_ptr, struct list_struct*rec)
{
tdb_off_t last_ptr, i;
struct list_struct lastrec;
if (tdb->read_only || tdb->traverse_read) return -1;
if (tdb_write_lock_record(tdb, rec_ptr) == -1) {
/* Someone traversing here: mark it as dead */
rec->magic = TDB_DEAD_MAGIC;
return tdb_rec_write(tdb, rec_ptr, rec);
}
if (tdb_write_unlock_record(tdb, rec_ptr) != 0)
return -1;
/* find previous record in hash chain */
if (tdb_ofs_read(tdb, TDB_HASH_TOP(rec->full_hash), &i) == -1)
return -1;
for (last_ptr = 0; i != rec_ptr; last_ptr = i, i = lastrec.next)
if (tdb_rec_read(tdb, i, &lastrec) == -1)
return -1;
/* unlink it: next ptr is at start of record. */
if (last_ptr == 0)
last_ptr = TDB_HASH_TOP(rec->full_hash);
if (tdb_ofs_write(tdb, last_ptr, &rec->next) == -1)
return -1;
/* recover the space */
if (tdb_free(tdb, rec_ptr, rec) == -1)
return -1;
return 0;
}
/* update an entry in place - this only works if the new data size
is <= the old data size and the key exists.
on failure return -1.
*/
static int tdb_update_hash(struct tdb_context *tdb, TDB_DATA key, u32 hash, TDB_DATA dbuf)
{
struct list_struct rec;
tdb_off_t rec_ptr;
/* find entry */
if (!(rec_ptr = tdb_find(tdb, key, hash, &rec)))
return -1;
/* must be long enough key, data and tailer */
if (rec.rec_len < key.dsize + dbuf.dsize + sizeof(tdb_off_t)) {
tdb->ecode = TDB_SUCCESS; /* Not really an error */
return -1;
}
if (tdb->methods->tdb_write(tdb, rec_ptr + sizeof(rec) + rec.key_len,
dbuf.dptr, dbuf.dsize) == -1)
return -1;
if (dbuf.dsize != rec.data_len) {
/* update size */
rec.data_len = dbuf.dsize;
return tdb_rec_write(tdb, rec_ptr, &rec);
}
return 0;
}
/* delete an entry in the database given a key */
static int tdb_delete_hash(struct tdb_context *tdb, TDB_DATA key, u32 hash)
{
tdb_off_t rec_ptr;
struct list_struct rec;
int ret;
if (tdb->max_dead_records != 0) {
/*
* Allow for some dead records per hash chain, mainly for
* tdb's with a very high create/delete rate like locking.tdb.
*/
if (tdb_lock(tdb, BUCKET(hash), F_WRLCK) == -1)
return -1;
if (tdb_count_dead(tdb, hash) >= tdb->max_dead_records) {
/*
* Don't let the per-chain freelist grow too large,
* delete all existing dead records
*/
tdb_purge_dead(tdb, hash);
}
if (!(rec_ptr = tdb_find(tdb, key, hash, &rec))) {
tdb_unlock(tdb, BUCKET(hash), F_WRLCK);
return -1;
}
/*
* Just mark the record as dead.
*/
rec.magic = TDB_DEAD_MAGIC;
ret = tdb_rec_write(tdb, rec_ptr, &rec);
}
else {
if (!(rec_ptr = tdb_find_lock_hash(tdb, key, hash, F_WRLCK,
&rec)))
return -1;
ret = tdb_do_delete(tdb, rec_ptr, &rec);
}
if (ret == 0) {
tdb_increment_seqnum(tdb);
}
if (tdb_unlock(tdb, BUCKET(rec.full_hash), F_WRLCK) != 0)
TDB_LOG((tdb, TDB_DEBUG_WARNING, "tdb_delete: WARNING tdb_unlock failed!\n"));
return ret;
}
/* update an entry in place - this only works if the new data size
is <= the old data size and the key exists.
on failure return -1.
*/
static int tdb_update_hash(struct tdb_context *tdb, TDB_DATA key, uint32_t hash, TDB_DATA dbuf)
{
struct tdb_record rec;
tdb_off_t rec_ptr;
/* find entry */
if (!(rec_ptr = tdb_find(tdb, key, hash, &rec)))
return -1;
/* it could be an exact duplicate of what is there - this is
* surprisingly common (eg. with a ldb re-index). */
if (rec.key_len == key.dsize &&
rec.data_len == dbuf.dsize &&
rec.full_hash == hash &&
tdb_parse_record(tdb, key, tdb_update_hash_cmp, &dbuf) == 0) {
return 0;
}
/* must be long enough key, data and tailer */
if (rec.rec_len < key.dsize + dbuf.dsize + sizeof(tdb_off_t)) {
tdb->ecode = TDB_SUCCESS; /* Not really an error */
return -1;
}
if (tdb->methods->tdb_write(tdb, rec_ptr + sizeof(rec) + rec.key_len,
dbuf.dptr, dbuf.dsize) == -1)
return -1;
if (dbuf.dsize != rec.data_len) {
/* update size */
rec.data_len = dbuf.dsize;
return tdb_rec_write(tdb, rec_ptr, &rec);
}
return 0;
}
/* Add an element into the freelist. Merge adjacent records if
necessary. */
int tdb_free(struct tdb_context *tdb, tdb_off_t offset, struct tdb_record *rec)
{
/* Allocation and tailer lock */
if (tdb_lock(tdb, -1, F_WRLCK) != 0)
return -1;
/* set an initial tailer, so if we fail we don't leave a bogus record */
if (update_tailer(tdb, offset, rec) != 0) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed!\n"));
goto fail;
}
#if USE_RIGHT_MERGES
/* Look right first (I'm an Australian, dammit) */
if (offset + sizeof(*rec) + rec->rec_len + sizeof(*rec) <= tdb->map_size) {
tdb_off_t right = offset + sizeof(*rec) + rec->rec_len;
struct tdb_record r;
if (tdb->methods->tdb_read(tdb, right, &r, sizeof(r), DOCONV()) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: right read failed at %u\n", right));
goto left;
}
/* If it's free, expand to include it. */
if (r.magic == TDB_FREE_MAGIC) {
if (remove_from_freelist(tdb, right, r.next) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: right free failed at %u\n", right));
goto left;
}
rec->rec_len += sizeof(r) + r.rec_len;
if (update_tailer(tdb, offset, rec) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed at %u\n", offset));
goto fail;
}
}
}
left:
#endif
/* Look left */
if (offset - sizeof(tdb_off_t) > TDB_DATA_START(tdb->header.hash_size)) {
tdb_off_t left = offset - sizeof(tdb_off_t);
struct tdb_record l;
tdb_off_t leftsize;
/* Read in tailer and jump back to header */
if (tdb_ofs_read(tdb, left, &leftsize) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: left offset read failed at %u\n", left));
goto update;
}
/* it could be uninitialised data */
if (leftsize == 0 || leftsize == TDB_PAD_U32) {
goto update;
}
left = offset - leftsize;
if (leftsize > offset ||
left < TDB_DATA_START(tdb->header.hash_size)) {
goto update;
}
/* Now read in the left record */
if (tdb->methods->tdb_read(tdb, left, &l, sizeof(l), DOCONV()) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: left read failed at %u (%u)\n", left, leftsize));
goto update;
}
/* If it's free, expand to include it. */
if (l.magic == TDB_FREE_MAGIC) {
/* we now merge the new record into the left record, rather than the other
way around. This makes the operation O(1) instead of O(n). This change
prevents traverse from being O(n^2) after a lot of deletes */
l.rec_len += sizeof(*rec) + rec->rec_len;
if (tdb_rec_write(tdb, left, &l) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_left failed at %u\n", left));
goto fail;
}
if (update_tailer(tdb, left, &l) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed at %u\n", offset));
goto fail;
}
tdb_unlock(tdb, -1, F_WRLCK);
return 0;
}
}
update:
/* Now, prepend to free list */
rec->magic = TDB_FREE_MAGIC;
if (tdb_ofs_read(tdb, FREELIST_TOP, &rec->next) == -1 ||
tdb_rec_write(tdb, offset, rec) == -1 ||
tdb_ofs_write(tdb, FREELIST_TOP, &offset) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free record write failed at offset=%d\n", offset));
goto fail;
}
/* And we're done. */
tdb_unlock(tdb, -1, F_WRLCK);
return 0;
fail:
tdb_unlock(tdb, -1, F_WRLCK);
return -1;
}
/* store an element in the database, replacing any existing element
with the same key
return 0 on success, -1 on failure
*/
int tdb_store(struct tdb_context *tdb, TDB_DATA key, TDB_DATA dbuf, int flag)
{
struct list_struct rec;
u32 hash;
tdb_off_t rec_ptr;
char *p = NULL;
int ret = -1;
if (tdb->read_only || tdb->traverse_read) {
tdb->ecode = TDB_ERR_RDONLY;
return -1;
}
/* find which hash bucket it is in */
hash = tdb->hash_fn(&key);
if (tdb_lock(tdb, BUCKET(hash), F_WRLCK) == -1)
return -1;
/* check for it existing, on insert. */
if (flag == TDB_INSERT) {
if (tdb_exists_hash(tdb, key, hash)) {
tdb->ecode = TDB_ERR_EXISTS;
goto fail;
}
} else {
/* first try in-place update, on modify or replace. */
if (tdb_update_hash(tdb, key, hash, dbuf) == 0) {
goto done;
}
if (tdb->ecode == TDB_ERR_NOEXIST &&
flag == TDB_MODIFY) {
/* if the record doesn't exist and we are in TDB_MODIFY mode then
we should fail the store */
goto fail;
}
}
/* reset the error code potentially set by the tdb_update() */
tdb->ecode = TDB_SUCCESS;
/* delete any existing record - if it doesn't exist we don't
care. Doing this first reduces fragmentation, and avoids
coalescing with `allocated' block before it's updated. */
if (flag != TDB_INSERT)
tdb_delete_hash(tdb, key, hash);
/* Copy key+value *before* allocating free space in case malloc
fails and we are left with a dead spot in the tdb. */
if (!(p = (char *)malloc(key.dsize + dbuf.dsize))) {
tdb->ecode = TDB_ERR_OOM;
goto fail;
}
memcpy(p, key.dptr, key.dsize);
if (dbuf.dsize)
memcpy(p+key.dsize, dbuf.dptr, dbuf.dsize);
if (tdb->max_dead_records != 0) {
/*
* Allow for some dead records per hash chain, look if we can
* find one that can hold the new record. We need enough space
* for key, data and tailer. If we find one, we don't have to
* consult the central freelist.
*/
rec_ptr = tdb_find_dead(
tdb, hash, &rec,
key.dsize + dbuf.dsize + sizeof(tdb_off_t));
if (rec_ptr != 0) {
rec.key_len = key.dsize;
rec.data_len = dbuf.dsize;
rec.full_hash = hash;
rec.magic = TDB_MAGIC;
if (tdb_rec_write(tdb, rec_ptr, &rec) == -1
|| tdb->methods->tdb_write(
tdb, rec_ptr + sizeof(rec),
p, key.dsize + dbuf.dsize) == -1) {
goto fail;
}
goto done;
}
}
/*
* We have to allocate some space from the freelist, so this means we
* have to lock it. Use the chance to purge all the DEAD records from
* the hash chain under the freelist lock.
*/
if (tdb_lock(tdb, -1, F_WRLCK) == -1) {
goto fail;
}
if ((tdb->max_dead_records != 0)
&& (tdb_purge_dead(tdb, hash) == -1)) {
tdb_unlock(tdb, -1, F_WRLCK);
goto fail;
}
/* we have to allocate some space */
rec_ptr = tdb_allocate(tdb, key.dsize + dbuf.dsize, &rec);
tdb_unlock(tdb, -1, F_WRLCK);
if (rec_ptr == 0) {
goto fail;
}
/* Read hash top into next ptr */
if (tdb_ofs_read(tdb, TDB_HASH_TOP(hash), &rec.next) == -1)
goto fail;
rec.key_len = key.dsize;
rec.data_len = dbuf.dsize;
rec.full_hash = hash;
rec.magic = TDB_MAGIC;
/* write out and point the top of the hash chain at it */
if (tdb_rec_write(tdb, rec_ptr, &rec) == -1
|| tdb->methods->tdb_write(tdb, rec_ptr+sizeof(rec), p, key.dsize+dbuf.dsize)==-1
|| tdb_ofs_write(tdb, TDB_HASH_TOP(hash), &rec_ptr) == -1) {
/* Need to tdb_unallocate() here */
goto fail;
}
done:
ret = 0;
fail:
if (ret == 0) {
tdb_increment_seqnum(tdb);
}
SAFE_FREE(p);
tdb_unlock(tdb, BUCKET(hash), F_WRLCK);
return ret;
}
