X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb%2Flock.c;h=2a6f1230c9d7308b83592771ec03de5d7f5a95f1;hp=e3759a3f8dddd38465fa4da4fdaea992cecb264c;hb=729fc5b931c8eb26f192399d8aa7064f9af058df;hpb=acf1a21e440b2908df0ce57c48dab1aca9e23cad diff --git a/ccan/tdb/lock.c b/ccan/tdb/lock.c index e3759a3f..2a6f1230 100644 --- a/ccan/tdb/lock.c +++ b/ccan/tdb/lock.c @@ -119,6 +119,12 @@ static int fcntl_unlock(struct tdb_context *tdb, int rw, off_t off, off_t len) return fcntl(tdb->fd, F_SETLKW, &fl); } +/* list -1 is the alloc list, otherwise a hash chain. */ +static tdb_off_t lock_offset(int list) +{ + return FREELIST_TOP + 4*list; +} + /* a byte range locking function - return 0 on success this functions locks/unlocks 1 byte at the specified offset. @@ -146,6 +152,14 @@ int tdb_brlock(struct tdb_context *tdb, return -1; } + /* Sanity check */ + if (tdb->transaction && offset >= lock_offset(-1) && len != 0) { + tdb->ecode = TDB_ERR_RDONLY; + TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_brlock attempted in transaction at offset %d rw_type=%d flags=%d len=%d\n", + offset, rw_type, flags, (int)len)); + return -1; + } + do { ret = fcntl_lock(tdb, rw_type, offset, len, flags & TDB_LOCK_WAIT); @@ -197,13 +211,29 @@ int tdb_brunlock(struct tdb_context *tdb, 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 tdb_allrecord_upgrade(struct tdb_context *tdb) { int count = 1000; + + if (tdb->allrecord_lock.count != 1) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "tdb_allrecord_upgrade failed: count %u too high\n", + tdb->allrecord_lock.count)); + return -1; + } + + if (tdb->allrecord_lock.off != 1) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, + "tdb_allrecord_upgrade failed: already upgraded?\n")); + return -1; + } + while (count--) { struct timeval tv; - if (tdb_brlock(tdb, F_WRLCK, offset, len, + if (tdb_brlock(tdb, F_WRLCK, FREELIST_TOP, 0, TDB_LOCK_WAIT|TDB_LOCK_PROBE) == 0) { + tdb->allrecord_lock.ltype = F_WRLCK; + tdb->allrecord_lock.off = 0; return 0; } if (errno != EDEADLK) { @@ -214,14 +244,21 @@ int tdb_brlock_upgrade(struct tdb_context *tdb, tdb_off_t offset, size_t len) 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)); + TDB_LOG((tdb, TDB_DEBUG_TRACE,"tdb_allrecord_upgrade failed\n")); return -1; } -/* list -1 is the alloc list, otherwise a hash chain. */ -static tdb_off_t lock_offset(int list) +static struct tdb_lock_type *find_nestlock(struct tdb_context *tdb, + tdb_off_t offset) { - return FREELIST_TOP + 4*list; + unsigned int i; + + for (i=0; inum_lockrecs; i++) { + if (tdb->lockrecs[i].off == offset) { + return &tdb->lockrecs[i]; + } + } + return NULL; } /* lock an offset in the database. */ @@ -229,7 +266,6 @@ int tdb_nest_lock(struct tdb_context *tdb, uint32_t offset, int ltype, enum tdb_lock_flags flags) { struct tdb_lock_type *new_lck; - int i; if (offset >= lock_offset(tdb->header.hash_size)) { tdb->ecode = TDB_ERR_LOCK; @@ -240,24 +276,14 @@ int tdb_nest_lock(struct tdb_context *tdb, uint32_t offset, int ltype, if (tdb->flags & TDB_NOLOCK) return 0; - for (i=0; inum_lockrecs; i++) { - if (tdb->lockrecs[i].off == offset) { - if (tdb->lockrecs[i].count == 0) { - /* - * Can't happen, see tdb_unlock(). It should - * be an assert. - */ - TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_lock: " - "lck->count == 0 for offset %u", - offset)); - } - /* - * Just increment the in-memory struct, posix locks - * don't stack. - */ - tdb->lockrecs[i].count++; - 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; } new_lck = (struct tdb_lock_type *)realloc( @@ -271,24 +297,56 @@ int tdb_nest_lock(struct tdb_context *tdb, uint32_t offset, int ltype, /* Since fcntl locks don't nest, we do a lock for the first one, and simply bump the count for future ones */ - if (tdb->methods->brlock(tdb, ltype, offset, 1, flags)) { + if (tdb_brlock(tdb, ltype, offset, 1, flags)) { return -1; } - tdb->num_locks++; - tdb->lockrecs[tdb->num_lockrecs].off = offset; tdb->lockrecs[tdb->num_lockrecs].count = 1; tdb->lockrecs[tdb->num_lockrecs].ltype = ltype; - tdb->num_lockrecs += 1; + tdb->num_lockrecs++; return 0; } -/* lock a list in the database. list -1 is the alloc list */ -int tdb_lock(struct tdb_context *tdb, int list, int ltype) +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; +} + +static bool have_data_locks(const struct tdb_context *tdb) +{ + unsigned int i; + + for (i = 0; i < tdb->num_lockrecs; i++) { + if (tdb->lockrecs[i].off >= lock_offset(-1)) + return true; + } + return false; +} + +static int tdb_lock_list(struct tdb_context *tdb, int list, int ltype, + enum tdb_lock_flags waitflag) { int ret; + bool check = false; /* a allrecord lock allows us to avoid per chain locks */ if (tdb->allrecord_lock.count && @@ -300,9 +358,28 @@ int tdb_lock(struct tdb_context *tdb, int list, int ltype) tdb->ecode = TDB_ERR_LOCK; ret = -1; } else { - ret = tdb_nest_lock(tdb, lock_offset(list), ltype, - TDB_LOCK_WAIT); + /* Only check when we grab first data lock. */ + check = !have_data_locks(tdb); + ret = tdb_nest_lock(tdb, lock_offset(list), ltype, waitflag); + + if (ret == 0 && check && tdb_needs_recovery(tdb)) { + tdb_nest_unlock(tdb, lock_offset(list), ltype, false); + + if (tdb_lock_and_recover(tdb) == -1) { + return -1; + } + return tdb_lock_list(tdb, list, ltype, waitflag); + } } + return ret; +} + +/* lock a list in the database. list -1 is the alloc list */ +int tdb_lock(struct tdb_context *tdb, int list, int ltype) +{ + int ret; + + ret = tdb_lock_list(tdb, list, ltype, TDB_LOCK_WAIT); if (ret) { TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_lock failed on list %d " "ltype=%d (%s)\n", list, ltype, strerror(errno))); @@ -313,18 +390,7 @@ int tdb_lock(struct tdb_context *tdb, int list, int ltype) /* lock a list in the database. list -1 is the alloc list. non-blocking lock */ int tdb_lock_nonblock(struct tdb_context *tdb, int list, int ltype) { - /* a allrecord lock allows us to avoid per chain locks */ - if (tdb->allrecord_lock.count && - (ltype == tdb->allrecord_lock.ltype || ltype == F_RDLCK)) { - return 0; - } - - if (tdb->allrecord_lock.count) { - tdb->ecode = TDB_ERR_LOCK; - return -1; - } - - return tdb_nest_lock(tdb, lock_offset(list), ltype, TDB_LOCK_NOWAIT); + return tdb_lock_list(tdb, list, ltype, TDB_LOCK_NOWAIT); } @@ -332,8 +398,7 @@ int tdb_nest_unlock(struct tdb_context *tdb, uint32_t offset, int ltype, bool mark_lock) { int ret = -1; - int i; - struct tdb_lock_type *lck = NULL; + struct tdb_lock_type *lck; if (tdb->flags & TDB_NOLOCK) return 0; @@ -344,13 +409,7 @@ int tdb_nest_unlock(struct tdb_context *tdb, uint32_t offset, int ltype, return ret; } - for (i=0; inum_lockrecs; i++) { - if (tdb->lockrecs[i].off == offset) { - lck = &tdb->lockrecs[i]; - break; - } - } - + lck = find_nestlock(tdb, offset); if ((lck == NULL) || (lck->count == 0)) { TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_unlock: count is 0\n")); return -1; @@ -371,19 +430,14 @@ int tdb_nest_unlock(struct tdb_context *tdb, uint32_t offset, int ltype, if (mark_lock) { ret = 0; } else { - ret = tdb->methods->brunlock(tdb, ltype, offset, 1); + ret = tdb_brunlock(tdb, ltype, offset, 1); } - tdb->num_locks--; /* * Shrink the array by overwriting the element just unlocked with the * last array element. */ - - if (tdb->num_lockrecs > 1) { - *lck = tdb->lockrecs[tdb->num_lockrecs-1]; - } - tdb->num_lockrecs -= 1; + *lck = tdb->lockrecs[--tdb->num_lockrecs]; /* * We don't bother with realloc when the array shrinks, but if we have @@ -420,21 +474,7 @@ int tdb_unlock(struct tdb_context *tdb, int list, int ltype) */ int tdb_transaction_lock(struct tdb_context *tdb, int ltype) { - if (tdb->allrecord_lock.count) { - return 0; - } - if (tdb->transaction_lock_count > 0) { - tdb->transaction_lock_count++; - return 0; - } - - if (tdb->methods->brlock(tdb, ltype, TRANSACTION_LOCK, 1, TDB_LOCK_WAIT) == -1) { - TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_lock: failed to get transaction lock\n")); - tdb->ecode = TDB_ERR_LOCK; - return -1; - } - tdb->transaction_lock_count++; - return 0; + return tdb_nest_lock(tdb, TRANSACTION_LOCK, ltype, TDB_LOCK_WAIT); } /* @@ -442,27 +482,14 @@ int tdb_transaction_lock(struct tdb_context *tdb, int ltype) */ int tdb_transaction_unlock(struct tdb_context *tdb, int ltype) { - int ret; - if (tdb->allrecord_lock.count) { - return 0; - } - if (tdb->transaction_lock_count > 1) { - tdb->transaction_lock_count--; - return 0; - } - ret = tdb->methods->brunlock(tdb, ltype, TRANSACTION_LOCK, 1); - if (ret == 0) { - tdb->transaction_lock_count = 0; - } - return ret; + return tdb_nest_unlock(tdb, TRANSACTION_LOCK, ltype, false); } - - -/* lock/unlock entire database */ -static int _tdb_lockall(struct tdb_context *tdb, int ltype, - enum tdb_lock_flags flags) +/* lock/unlock entire database. It can only be upgradable if you have some + * other way of guaranteeing exclusivity (ie. transaction write lock). */ +int tdb_allrecord_lock(struct tdb_context *tdb, int ltype, + enum tdb_lock_flags flags, bool upgradable) { /* There are no locks on read-only dbs */ if (tdb->read_only || tdb->traverse_read) { @@ -480,16 +507,20 @@ static int _tdb_lockall(struct tdb_context *tdb, int ltype, tdb->ecode = TDB_ERR_LOCK; return -1; } - - if (tdb->num_locks != 0) { + + if (tdb_have_extra_locks(tdb)) { /* can't combine global and chain locks */ tdb->ecode = TDB_ERR_LOCK; return -1; } - if (tdb->methods->brlock(tdb, ltype, - FREELIST_TOP, 4*tdb->header.hash_size, - flags)) { + if (upgradable && ltype != F_RDLCK) { + /* tdb error: you can't upgrade a write lock! */ + tdb->ecode = TDB_ERR_LOCK; + return -1; + } + + if (tdb_brlock(tdb, ltype, FREELIST_TOP, 0, flags)) { if (flags & TDB_LOCK_WAIT) { TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_lockall failed (%s)\n", strerror(errno))); } @@ -497,7 +528,25 @@ static int _tdb_lockall(struct tdb_context *tdb, int ltype, } tdb->allrecord_lock.count = 1; - tdb->allrecord_lock.ltype = ltype; + /* 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; } @@ -505,7 +554,7 @@ static int _tdb_lockall(struct tdb_context *tdb, int ltype, /* unlock entire db */ -static int _tdb_unlockall(struct tdb_context *tdb, int ltype, bool mark_lock) +int tdb_allrecord_unlock(struct tdb_context *tdb, int ltype, bool mark_lock) { /* There are no locks on read-only dbs */ if (tdb->read_only || tdb->traverse_read) { @@ -513,7 +562,14 @@ static int _tdb_unlockall(struct tdb_context *tdb, int ltype, bool mark_lock) return -1; } - if (tdb->allrecord_lock.ltype != ltype || tdb->allrecord_lock.count == 0) { + if (tdb->allrecord_lock.count == 0) { + tdb->ecode = TDB_ERR_LOCK; + return -1; + } + + /* Upgradable locks are marked as write locks. */ + if (tdb->allrecord_lock.ltype != ltype + && (!tdb->allrecord_lock.off || ltype != F_RDLCK)) { tdb->ecode = TDB_ERR_LOCK; return -1; } @@ -523,9 +579,7 @@ static int _tdb_unlockall(struct tdb_context *tdb, int ltype, bool mark_lock) return 0; } - if (!mark_lock && - tdb->methods->brunlock(tdb, ltype, - FREELIST_TOP, 4*tdb->header.hash_size)) { + if (!mark_lock && tdb_brunlock(tdb, ltype, FREELIST_TOP, 0)) { TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_unlockall failed (%s)\n", strerror(errno))); return -1; } @@ -540,27 +594,27 @@ static int _tdb_unlockall(struct tdb_context *tdb, int ltype, bool mark_lock) int tdb_lockall(struct tdb_context *tdb) { tdb_trace(tdb, "tdb_lockall"); - return _tdb_lockall(tdb, F_WRLCK, TDB_LOCK_WAIT); + return tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false); } /* lock entire database with write lock - mark only */ int tdb_lockall_mark(struct tdb_context *tdb) { tdb_trace(tdb, "tdb_lockall_mark"); - return _tdb_lockall(tdb, F_WRLCK, TDB_LOCK_MARK_ONLY); + return tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_MARK_ONLY, false); } /* unlock entire database with write lock - unmark only */ int tdb_lockall_unmark(struct tdb_context *tdb) { tdb_trace(tdb, "tdb_lockall_unmark"); - return _tdb_unlockall(tdb, F_WRLCK, true); + return tdb_allrecord_unlock(tdb, F_WRLCK, true); } /* lock entire database with write lock - nonblocking varient */ int tdb_lockall_nonblock(struct tdb_context *tdb) { - int ret = _tdb_lockall(tdb, F_WRLCK, TDB_LOCK_NOWAIT); + int ret = tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_NOWAIT, false); tdb_trace_ret(tdb, "tdb_lockall_nonblock", ret); return ret; } @@ -569,20 +623,20 @@ int tdb_lockall_nonblock(struct tdb_context *tdb) int tdb_unlockall(struct tdb_context *tdb) { tdb_trace(tdb, "tdb_unlockall"); - return _tdb_unlockall(tdb, F_WRLCK, false); + return tdb_allrecord_unlock(tdb, F_WRLCK, false); } /* lock entire database with read lock */ int tdb_lockall_read(struct tdb_context *tdb) { tdb_trace(tdb, "tdb_lockall_read"); - return _tdb_lockall(tdb, F_RDLCK, TDB_LOCK_WAIT); + return tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, false); } /* lock entire database with read lock - nonblock varient */ int tdb_lockall_read_nonblock(struct tdb_context *tdb) { - int ret = _tdb_lockall(tdb, F_RDLCK, TDB_LOCK_NOWAIT); + int ret = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_NOWAIT, false); tdb_trace_ret(tdb, "tdb_lockall_read_nonblock", ret); return ret; } @@ -591,7 +645,7 @@ int tdb_lockall_read_nonblock(struct tdb_context *tdb) int tdb_unlockall_read(struct tdb_context *tdb) { tdb_trace(tdb, "tdb_unlockall_read"); - return _tdb_unlockall(tdb, F_RDLCK, false); + return tdb_allrecord_unlock(tdb, F_RDLCK, false); } /* lock/unlock one hash chain. This is meant to be used to reduce @@ -658,7 +712,7 @@ int tdb_lock_record(struct tdb_context *tdb, tdb_off_t off) if (tdb->allrecord_lock.count) { return 0; } - return off ? tdb->methods->brlock(tdb, F_RDLCK, off, 1, TDB_LOCK_WAIT) : 0; + return off ? tdb_brlock(tdb, F_RDLCK, off, 1, TDB_LOCK_WAIT) : 0; } /* @@ -672,12 +726,21 @@ int tdb_write_lock_record(struct tdb_context *tdb, tdb_off_t off) for (i = &tdb->travlocks; i; i = i->next) if (i->off == off) return -1; - return tdb->methods->brlock(tdb, F_WRLCK, off, 1, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE); + 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); } int tdb_write_unlock_record(struct tdb_context *tdb, tdb_off_t off) { - return tdb->methods->brunlock(tdb, F_WRLCK, off, 1); + if (tdb->allrecord_lock.count) { + return 0; + } + return tdb_brunlock(tdb, F_WRLCK, off, 1); } /* fcntl locks don't stack: avoid unlocking someone else's */ @@ -695,5 +758,56 @@ int tdb_unlock_record(struct tdb_context *tdb, tdb_off_t off) for (i = &tdb->travlocks; i; i = i->next) if (i->off == off) count++; - return (count == 1 ? tdb->methods->brunlock(tdb, F_RDLCK, off, 1) : 0); + return (count == 1 ? tdb_brunlock(tdb, F_RDLCK, off, 1) : 0); +} + +bool tdb_have_extra_locks(struct tdb_context *tdb) +{ + unsigned int extra = tdb->num_lockrecs; + + /* A transaction holds the lock for all records. */ + if (!tdb->transaction && tdb->allrecord_lock.count) { + return true; + } + + /* We always hold the active lock if CLEAR_IF_FIRST. */ + if (find_nestlock(tdb, ACTIVE_LOCK)) { + extra--; + } + + /* In a transaction, we expect to hold the transaction lock */ + if (tdb->transaction && find_nestlock(tdb, TRANSACTION_LOCK)) { + extra--; + } + + return extra; +} + +/* The transaction code uses this to remove all locks. Note that this + may include OPEN_LOCK. */ +void tdb_release_extra_locks(struct tdb_context *tdb) +{ + unsigned int i, extra = 0; + + if (tdb->allrecord_lock.count != 0) { + tdb_brunlock(tdb, tdb->allrecord_lock.ltype, FREELIST_TOP, 0); + tdb->allrecord_lock.count = 0; + } + + for (i=0;inum_lockrecs;i++) { + struct tdb_lock_type *lck = &tdb->lockrecs[i]; + + /* Don't release transaction or active locks! */ + if (tdb->transaction && lck->off == TRANSACTION_LOCK) { + tdb->lockrecs[extra++] = *lck; + } else if (lck->off == ACTIVE_LOCK) { + tdb->lockrecs[extra++] = *lck; + } else { + tdb_brunlock(tdb, lck->ltype, lck->off, 1); + } + } + tdb->num_lockrecs = extra; + if (tdb->num_lockrecs == 0) { + SAFE_FREE(tdb->lockrecs); + } }