Calling tdb_traverse inside a transaction led to the transaction lock being
held indefinitely. This was caused by the tdb_transaction_lock/unlock inside
tdb_traverse: The transaction code holds the global lock at offset
TRANSACTION_LOCK. The call to tdb_transaction_lock does nothing because the
transaction_lock is already being held. tdb_transaction_unlock inside tdb_wrap
resets tdb->have_transaction_lock but does not release the kernel-level fcntl
lock. transaction_commit later on does not release that fcntl lock either,
because tdb->have_transaction_lock was already reset by tdb_transaction().
This patch does fix that problem for me. An alternative would be to make
tdb->have_transaction_lock a counter that can cope with proper nesting, maybe
in other places as well.
Volker
(This used to be commit
89543005fe2e4934b3c560c937d49304a32a7fc2)
{
struct tdb_traverse_lock tl = { NULL, 0, 0, F_WRLCK };
int ret;
+ int in_transaction = (tdb->transaction != NULL);
if (tdb->read_only || tdb->traverse_read) {
return tdb_traverse_read(tdb, fn, private_data);
}
- if (tdb_transaction_lock(tdb, F_WRLCK)) {
- return -1;
+ if (!in_transaction) {
+ if (tdb_transaction_lock(tdb, F_WRLCK)) {
+ return -1;
+ }
}
tdb->traverse_write++;
ret = tdb_traverse_internal(tdb, fn, private_data, &tl);
tdb->traverse_write--;
- tdb_transaction_unlock(tdb);
+ if (!in_transaction) {
+ tdb_transaction_unlock(tdb);
+ }
return ret;
}