to reduce this to 3 or even 2 with some more work.
- check for a valid recovery record on open of the tdb, while the
- global lock is held. Automatically recover from the transaction
+ open lock is held. Automatically recover from the transaction
recovery area if needed, then continue with the open as
usual. This allows for smooth crash recovery with no administrator
intervention.
/*
brlock during a transaction - ignore them
*/
-static int transaction_brlock(struct tdb_context *tdb, tdb_off_t offset,
- int rw_type, int lck_type, int probe, size_t len)
+static int transaction_brlock(struct tdb_context *tdb,
+ int rw_type, tdb_off_t offset, size_t len,
+ enum tdb_lock_flags flags)
+{
+ /* FIXME: We actually grab the open lock during a transaction. */
+ if (offset == OPEN_LOCK)
+ return tdb_brlock(tdb, rw_type, offset, len, flags);
+ return 0;
+}
+
+static int transaction_brunlock(struct tdb_context *tdb,
+ int rw_type, tdb_off_t offset, size_t len)
{
return 0;
}
transaction_next_hash_chain,
transaction_oob,
transaction_expand_file,
- transaction_brlock
+ transaction_brlock,
+ transaction_brunlock
};
return 0;
}
- if (tdb->num_locks != 0 || tdb->global_lock.count) {
+ if (tdb_have_extra_locks(tdb)) {
/* the caller must not have any locks when starting a
transaction as otherwise we'll be screwed by lack
of nested locks in posix */
/* get a read lock from the freelist to the end of file. This
is upgraded to a write lock during the commit */
- if (tdb_brlock(tdb, FREELIST_TOP, F_RDLCK, F_SETLKW, 0, 0) == -1) {
+ if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
- tdb->ecode = TDB_ERR_LOCK;
- goto fail;
+ goto fail_allrecord_lock;
}
/* setup a copy of the hash table heads so the hash scan in
return 0;
fail:
- tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
- tdb_transaction_unlock(tdb);
+ tdb_allrecord_unlock(tdb, F_RDLCK, false);
+fail_allrecord_lock:
+ tdb_transaction_unlock(tdb, F_WRLCK);
SAFE_FREE(tdb->transaction->blocks);
SAFE_FREE(tdb->transaction->hash_heads);
SAFE_FREE(tdb->transaction);
return 0;
}
- if (fsync(tdb->fd) != 0) {
+ if (fdatasync(tdb->fd) != 0) {
tdb->ecode = TDB_ERR_IO;
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
return -1;
}
-int _tdb_transaction_cancel(struct tdb_context *tdb)
+static int _tdb_transaction_cancel(struct tdb_context *tdb)
{
int i, ret = 0;
if (tdb->transaction->magic_offset) {
const struct tdb_methods *methods = tdb->transaction->io_methods;
- uint32_t zero = 0;
+ const uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
/* remove the recovery marker */
- if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &zero, 4) == -1 ||
+ if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
ret = -1;
}
}
- /* remove any global lock created during the transaction */
- if (tdb->global_lock.count != 0) {
- tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 4*tdb->header.hash_size);
- tdb->global_lock.count = 0;
- }
-
- /* remove any locks created during the transaction */
- if (tdb->num_locks != 0) {
- for (i=0;i<tdb->num_lockrecs;i++) {
- tdb_brlock(tdb,FREELIST_TOP+4*tdb->lockrecs[i].list,
- F_UNLCK,F_SETLKW, 0, 1);
- }
- tdb->num_locks = 0;
- tdb->num_lockrecs = 0;
- SAFE_FREE(tdb->lockrecs);
- }
+ /* This also removes the OPEN_LOCK, if we have it. */
+ tdb_release_extra_locks(tdb);
/* restore the normal io methods */
tdb->methods = tdb->transaction->io_methods;
- tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
- tdb_transaction_unlock(tdb);
+ tdb_transaction_unlock(tdb, F_WRLCK);
SAFE_FREE(tdb->transaction->hash_heads);
SAFE_FREE(tdb->transaction);
rec.rec_len = 0;
- if (recovery_head != 0 &&
- methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
- TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
- return -1;
+ if (recovery_head != 0) {
+ if (methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
+ return -1;
+ }
+ /* ignore invalid recovery regions: can happen in crash */
+ if (rec.magic != TDB_RECOVERY_MAGIC &&
+ rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
+ recovery_head = 0;
+ }
}
*recovery_size = tdb_recovery_size(tdb);
rec = (struct tdb_record *)data;
memset(rec, 0, sizeof(*rec));
- rec->magic = 0;
+ rec->magic = TDB_RECOVERY_INVALID_MAGIC;
rec->data_len = recovery_size;
rec->rec_len = recovery_max_size;
rec->key_len = old_map_size;
/* if there are any locks pending then the caller has not
nested their locks properly, so fail the transaction */
- if (tdb->num_locks || tdb->global_lock.count) {
+ if (tdb_have_extra_locks(tdb)) {
tdb->ecode = TDB_ERR_LOCK;
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
_tdb_transaction_cancel(tdb);
}
/* upgrade the main transaction lock region to a write lock */
- if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
+ if (tdb_allrecord_upgrade(tdb) == -1) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
- tdb->ecode = TDB_ERR_LOCK;
_tdb_transaction_cancel(tdb);
return -1;
}
- /* get the global lock - this prevents new users attaching to the database
+ /* get the open lock - this prevents new users attaching to the database
during the commit */
- if (tdb_brlock(tdb, GLOBAL_LOCK, F_WRLCK, F_SETLKW, 0, 1) == -1) {
- TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
- tdb->ecode = TDB_ERR_LOCK;
+ if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
_tdb_transaction_cancel(tdb);
return -1;
}
/* write the recovery data to the end of the file */
if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
- tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
_tdb_transaction_cancel(tdb);
return -1;
}
tdb->transaction->old_map_size) == -1) {
tdb->ecode = TDB_ERR_IO;
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
- tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
_tdb_transaction_cancel(tdb);
return -1;
}
methods->tdb_oob(tdb, tdb->map_size + 1, 1);
}
- /* Keep the global lock until the actual commit */
+ /* Keep the open lock until the actual commit */
return 0;
}
tdb_transaction_recover(tdb);
_tdb_transaction_cancel(tdb);
- tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
return -1;
return -1;
}
- tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
-
/*
TODO: maybe write to some dummy hdr field, or write to magic
offset without mmap, before the last sync, instead of the
/*
recover from an aborted transaction. Must be called with exclusive
- database write access already established (including the global
+ database write access already established (including the open
lock to prevent new processes attaching)
*/
int tdb_transaction_recover(struct tdb_context *tdb)