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.
fsync/msync calls are made.
- if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
- tdb_add_flags() transaction is enabled.
- The default is that transaction nesting is not allowed and an attempt
- to create a nested transaction will fail with TDB_ERR_NESTING.
+ tdb_add_flags() transaction nesting is enabled.
+ It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
+ The default is that transaction nesting is allowed.
+ Note: this default may change in future versions of tdb.
Beware. when transactions are nested a transaction successfully
completed with tdb_transaction_commit() can be silently unrolled later.
+
+ - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
+ tdb_add_flags() transaction nesting is disabled.
+ It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
+ An attempt create a nested transaction will fail with TDB_ERR_NESTING.
+ The default is that transaction nesting is allowed.
+ Note: this default may change in future versions of tdb.
*/
bool prepared;
tdb_off_t magic_offset;
+ /* set when the OPEN_LOCK has been taken */
+ bool open_lock_taken;
+
/* old file size before transaction */
tdb_len_t old_map_size;
if (tdb->transaction->magic_offset) {
const struct tdb_methods *methods = tdb->transaction->io_methods;
- uint32_t zero = 0;
+ 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;
}
}
+ if (tdb->transaction->open_lock_taken) {
+ tdb_brunlock(tdb, F_WRLCK, OPEN_LOCK, 1);
+ tdb->transaction->open_lock_taken = false;
+ }
+
/* remove any global lock created during the transaction */
- if (tdb->global_lock.count != 0) {
- tdb_brunlock(tdb, tdb->global_lock.ltype,
+ if (tdb->allrecord_lock.count != 0) {
+ tdb_brunlock(tdb, tdb->allrecord_lock.ltype,
FREELIST_TOP, 4*tdb->header.hash_size);
- tdb->global_lock.count = 0;
+ tdb->allrecord_lock.count = 0;
}
/* remove any locks created during the transaction */
if (tdb->num_locks != 0) {
for (i=0;i<tdb->num_lockrecs;i++) {
tdb_brunlock(tdb, tdb->lockrecs[i].ltype,
- FREELIST_TOP+4*tdb->lockrecs[i].list, 1);
+ tdb->lockrecs[i].off, 1);
}
tdb->num_locks = 0;
tdb->num_lockrecs = 0;
return 0;
}
- if (tdb->num_locks != 0 || tdb->global_lock.count) {
+ if (tdb->num_locks != 0 || tdb->allrecord_lock.count) {
/* the caller must not have any locks when starting a
transaction as otherwise we'll be screwed by lack
of nested locks in posix */
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->num_locks || tdb->allrecord_lock.count) {
tdb->ecode = TDB_ERR_LOCK;
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
_tdb_transaction_cancel(tdb, F_RDLCK);
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, F_WRLCK, GLOBAL_LOCK, 1, TDB_LOCK_WAIT) == -1) {
- TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
+ if (tdb_brlock(tdb, F_WRLCK, OPEN_LOCK, 1, TDB_LOCK_WAIT) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
tdb->ecode = TDB_ERR_LOCK;
_tdb_transaction_cancel(tdb, F_WRLCK);
return -1;
}
+ tdb->transaction->open_lock_taken = true;
+
if (!(tdb->flags & TDB_NOSYNC)) {
/* 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_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
_tdb_transaction_cancel(tdb, F_WRLCK);
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_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
_tdb_transaction_cancel(tdb, F_WRLCK);
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, F_WRLCK);
- tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
return -1;
return -1;
}
- tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 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)
projects / ccan / blobdiff