/* We only need to lock individual bytes, but Linux merges consecutive locks
* so we lock in contiguous ranges. */
static int tdb_chainlock_gradual(struct tdb_context *tdb,
int ltype, enum tdb_lock_flags flags,
size_t off, size_t len)
{
int ret;
enum tdb_lock_flags nb_flags = (flags & ~TDB_LOCK_WAIT);
if (len <= 4) {
/* Single record. Just do blocking lock. */
return tdb_brlock(tdb, ltype, off, len, flags);
}
/* First we try non-blocking. */
ret = tdb_brlock(tdb, ltype, off, len, nb_flags);
if (ret == 0) {
return 0;
}
/* Try locking first half, then second. */
ret = tdb_chainlock_gradual(tdb, ltype, flags, off, len / 2);
if (ret == -1)
return -1;
ret = tdb_chainlock_gradual(tdb, ltype, flags,
off + len / 2, len - len / 2);
if (ret == -1) {
tdb_brunlock(tdb, ltype, off, len / 2);
return -1;
}
return 0;
}
/*
increment the tdb sequence number if the tdb has been opened using
the TDB_SEQNUM flag
*/
static void tdb_increment_seqnum(struct tdb_context *tdb)
{
if (!(tdb->flags & TDB_SEQNUM)) {
return;
}
if (tdb_brlock(tdb, TDB_SEQNUM_OFS, F_WRLCK, F_SETLKW, 1, 1) != 0) {
return;
}
tdb_increment_seqnum_nonblock(tdb);
tdb_brlock(tdb, TDB_SEQNUM_OFS, F_UNLCK, F_SETLKW, 1, 1);
}
/* record lock stops delete underneath */
int tdb_lock_record(struct tdb_context *tdb, tdb_off_t off)
{
if (tdb->allrecord_lock.count) {
return 0;
}
return off ? tdb_brlock(tdb, F_RDLCK, off, 1, TDB_LOCK_WAIT) : 0;
}
static int tdb_lock_and_recover(struct tdb_context *tdb)
{
int ret;
/* We need to match locking order in transaction commit. */
if (tdb_brlock(tdb, F_WRLCK, FREELIST_TOP, 0, TDB_LOCK_WAIT)) {
return -1;
}
if (tdb_brlock(tdb, F_WRLCK, OPEN_LOCK, 1, TDB_LOCK_WAIT)) {
tdb_brunlock(tdb, F_WRLCK, FREELIST_TOP, 0);
return -1;
}
ret = tdb_transaction_recover(tdb);
tdb_brunlock(tdb, F_WRLCK, OPEN_LOCK, 1);
tdb_brunlock(tdb, F_WRLCK, FREELIST_TOP, 0);
return ret;
}
/*
increment the tdb sequence number if the tdb has been opened using
the TDB_SEQNUM flag
*/
static void tdb_increment_seqnum(struct tdb_context *tdb)
{
tdb_off_t seqnum=0;
if (!(tdb->flags & TDB_SEQNUM)) {
return;
}
if (tdb_brlock(tdb, TDB_SEQNUM_OFS, F_WRLCK, F_SETLKW, 1, 1) != 0) {
return;
}
/* we ignore errors from this, as we have no sane way of
dealing with them.
*/
tdb_ofs_read(tdb, TDB_SEQNUM_OFS, &seqnum);
seqnum++;
tdb_ofs_write(tdb, TDB_SEQNUM_OFS, &seqnum);
tdb_brlock(tdb, TDB_SEQNUM_OFS, F_UNLCK, F_SETLKW, 1, 1);
}
/* lock an offset in the database. */
int tdb_nest_lock(struct tdb_context *tdb, uint32_t offset, int ltype,
enum tdb_lock_flags flags)
{
struct tdb_lock_type *new_lck;
if (offset >= lock_offset(tdb->hash_size)) {
tdb->ecode = TDB_ERR_LOCK;
TDB_LOG((tdb, TDB_DEBUG_ERROR,"tdb_lock: invalid offset %u for ltype=%d\n",
offset, ltype));
return -1;
}
if (tdb->flags & TDB_NOLOCK)
return 0;
new_lck = find_nestlock(tdb, offset);
if (new_lck) {
/*
* Just increment the in-memory struct, posix locks
* don't stack.
*/
new_lck->count++;
return 0;
}
if (tdb->num_lockrecs == tdb->lockrecs_array_length) {
new_lck = (struct tdb_lock_type *)realloc(
tdb->lockrecs,
sizeof(*tdb->lockrecs) * (tdb->num_lockrecs+1));
if (new_lck == NULL) {
errno = ENOMEM;
return -1;
}
tdb->lockrecs_array_length = tdb->num_lockrecs+1;
tdb->lockrecs = new_lck;
}
/* Since fcntl locks don't nest, we do a lock for the first one,
and simply bump the count for future ones */
if (tdb_brlock(tdb, ltype, offset, 1, flags)) {
return -1;
}
new_lck = &tdb->lockrecs[tdb->num_lockrecs];
new_lck->off = offset;
new_lck->count = 1;
new_lck->ltype = ltype;
tdb->num_lockrecs++;
return 0;
}
/*
Write locks override our own fcntl readlocks, so check it here.
Note this is meant to be F_SETLK, *not* F_SETLKW, as it's not
an error to fail to get the lock here.
*/
int tdb_write_lock_record(struct tdb_context *tdb, tdb_off_t off)
{
struct tdb_traverse_lock *i;
for (i = &tdb->travlocks; i; i = i->next)
if (i->off == off)
return -1;
if (tdb->allrecord_lock.count) {
if (tdb->allrecord_lock.ltype == F_WRLCK) {
return 0;
}
return -1;
}
return tdb_brlock(tdb, F_WRLCK, off, 1, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
}
/*
upgrade a read lock to a write lock. This needs to be handled in a
special way as some OSes (such as solaris) have too conservative
deadlock detection and claim a deadlock when progress can be
made. For those OSes we may loop for a while.
*/
int tdb_brlock_upgrade(struct tdb_context *tdb, tdb_off_t offset, size_t len)
{
int count = 1000;
while (count--) {
struct timeval tv;
if (tdb_brlock(tdb, offset, F_WRLCK, F_SETLKW, 1, len) == 0) {
return 0;
}
if (errno != EDEADLK) {
break;
}
/* sleep for as short a time as we can - more portable than usleep() */
tv.tv_sec = 0;
tv.tv_usec = 1;
select(0, NULL, NULL, NULL, &tv);
}
TDB_LOG((tdb, TDB_DEBUG_TRACE,"tdb_brlock_upgrade failed at offset %d\n", offset));
return -1;
}
static int tdb_brlock_retry(struct tdb_context *tdb,
int rw_type, tdb_off_t offset, size_t len,
enum tdb_lock_flags flags)
{
int count = 1000;
while (count--) {
struct timeval tv;
int ret;
ret = tdb_brlock(tdb, rw_type, offset, len, flags);
if (ret == 0) {
return 0;
}
if (errno != EDEADLK) {
break;
}
/* sleep for as short a time as we can - more portable than usleep() */
tv.tv_sec = 0;
tv.tv_usec = 1;
select(0, NULL, NULL, NULL, &tv);
}
return -1;
}
/* lock/unlock entire database. It can only be upgradable if you have some
* other way of guaranteeing exclusivity (ie. transaction write lock).
* We do the locking gradually to avoid being starved by smaller locks. */
int tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
enum tdb_lock_flags flags, bool upgradable)
{
int ret;
switch (tdb_allrecord_check(tdb, ltype, flags, upgradable)) {
case -1:
return -1;
case 0:
return 0;
}
/* We cover two kinds of locks:
* 1) Normal chain locks. Taken for almost all operations.
* 2) Individual records locks. Taken after normal or free
* chain locks.
*
* It is (1) which cause the starvation problem, so we're only
* gradual for that. */
if (tdb_have_mutexes(tdb)) {
ret = tdb_mutex_allrecord_lock(tdb, ltype, flags);
} else {
ret = tdb_chainlock_gradual(tdb, ltype, flags, FREELIST_TOP,
tdb->hash_size * 4);
}
if (ret == -1) {
return -1;
}
/* Grab individual record locks. */
if (tdb_brlock(tdb, ltype, lock_offset(tdb->hash_size), 0,
flags) == -1) {
if (tdb_have_mutexes(tdb)) {
tdb_mutex_allrecord_unlock(tdb);
} else {
tdb_brunlock(tdb, ltype, FREELIST_TOP,
tdb->hash_size * 4);
}
return -1;
}
tdb->allrecord_lock.count = 1;
/* If it's upgradable, it's actually exclusive so we can treat
* it as a write lock. */
tdb->allrecord_lock.ltype = upgradable ? F_WRLCK : ltype;
tdb->allrecord_lock.off = upgradable;
if (tdb_needs_recovery(tdb)) {
bool mark = flags & TDB_LOCK_MARK_ONLY;
tdb_allrecord_unlock(tdb, ltype, mark);
if (mark) {
tdb->ecode = TDB_ERR_LOCK;
TDB_LOG((tdb, TDB_DEBUG_ERROR,
"tdb_lockall_mark cannot do recovery\n"));
return -1;
}
if (tdb_lock_and_recover(tdb) == -1) {
return -1;
}
return tdb_allrecord_lock(tdb, ltype, flags, upgradable);
}
return 0;
}
