X-Git-Url: https://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb%2Ftransaction.c;h=ad07ad9fed1006dc41714c440395881ddc7465a3;hp=34137251bce51f2697184fbd8dc515315b9d64c6;hb=05f69a4723f3a48502a699805c90c16489bf0525;hpb=3f011090d857a625f9eccfca6dd33d1b897417a8 diff --git a/ccan/tdb/transaction.c b/ccan/tdb/transaction.c index 34137251..ad07ad9f 100644 --- a/ccan/tdb/transaction.c +++ b/ccan/tdb/transaction.c @@ -76,7 +76,7 @@ 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. @@ -86,12 +86,20 @@ 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. */ @@ -127,6 +135,9 @@ struct tdb_transaction { 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; @@ -483,16 +494,21 @@ int _tdb_transaction_cancel(struct tdb_context *tdb, int ltype) 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, @@ -680,7 +696,7 @@ static int tdb_recovery_allocate(struct tdb_context *tdb, tdb_off_t *recovery_offset, tdb_len_t *recovery_max_size) { - struct list_struct rec; + struct tdb_record rec; const struct tdb_methods *methods = tdb->transaction->io_methods; tdb_off_t recovery_head; @@ -766,7 +782,7 @@ static int transaction_setup_recovery(struct tdb_context *tdb, tdb_len_t recovery_size; unsigned char *data, *p; const struct tdb_methods *methods = tdb->transaction->io_methods; - struct list_struct *rec; + struct tdb_record *rec; tdb_off_t recovery_offset, recovery_max_size; tdb_off_t old_map_size = tdb->transaction->old_map_size; uint32_t magic, tailer; @@ -786,10 +802,10 @@ static int transaction_setup_recovery(struct tdb_context *tdb, return -1; } - rec = (struct list_struct *)data; + 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; @@ -869,7 +885,7 @@ static int transaction_setup_recovery(struct tdb_context *tdb, magic = TDB_RECOVERY_MAGIC; CONVERT(magic); - *magic_offset = recovery_offset + offsetof(struct list_struct, magic); + *magic_offset = recovery_offset + offsetof(struct tdb_record, magic); if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) { TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n")); @@ -943,20 +959,21 @@ static int _tdb_transaction_prepare_commit(struct tdb_context *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, 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; } @@ -971,7 +988,6 @@ static int _tdb_transaction_prepare_commit(struct tdb_context *tdb) 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; } @@ -979,7 +995,7 @@ static int _tdb_transaction_prepare_commit(struct tdb_context *tdb) 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; } @@ -1061,7 +1077,6 @@ int tdb_transaction_commit(struct tdb_context *tdb) 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; @@ -1077,8 +1092,6 @@ int tdb_transaction_commit(struct tdb_context *tdb) 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 @@ -1110,7 +1123,7 @@ int tdb_transaction_commit(struct tdb_context *tdb) /* 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) @@ -1118,7 +1131,7 @@ int tdb_transaction_recover(struct tdb_context *tdb) tdb_off_t recovery_head, recovery_eof; unsigned char *data, *p; uint32_t zero = 0; - struct list_struct rec; + struct tdb_record rec; /* find the recovery area */ if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) { @@ -1205,7 +1218,7 @@ int tdb_transaction_recover(struct tdb_context *tdb) } /* remove the recovery magic */ - if (tdb_ofs_write(tdb, recovery_head + offsetof(struct list_struct, magic), + if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic), &zero) == -1) { TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n")); tdb->ecode = TDB_ERR_IO;