X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb%2Ftransaction.c;h=ded3fa836fc721a07932c54a6f741b89c8329066;hp=e55285f94cc6609a2afa720d39fe66d2261a36fe;hb=e9989ef0758b3b660abf444d014c5b51a3a7c1b4;hpb=1bc7c4367132a36c05ac1b3afcbec70a6e26540d diff --git a/ccan/tdb/transaction.c b/ccan/tdb/transaction.c index e55285f9..ded3fa83 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. */ @@ -144,14 +152,6 @@ static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf, { uint32_t blk; - /* Only a commit is allowed on a prepared transaction */ - if (tdb->transaction->prepared) { - tdb->ecode = TDB_ERR_EINVAL; - TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: transaction already prepared, read not allowed\n")); - tdb->transaction->transaction_error = 1; - return -1; - } - /* break it down into block sized ops */ while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) { tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size); @@ -411,8 +411,18 @@ static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size, /* brlock during a transaction - ignore them */ -static int transaction_brlock(struct tdb_context *tdb, tdb_off_t offset, - int rw_type, int lck_type, int probe, size_t len) +static int transaction_brlock(struct tdb_context *tdb, + int rw_type, tdb_off_t offset, size_t len, + enum tdb_lock_flags flags) +{ + /* FIXME: We actually grab the open lock during a transaction. */ + if (offset == OPEN_LOCK) + return tdb_brlock(tdb, rw_type, offset, len, flags); + return 0; +} + +static int transaction_brunlock(struct tdb_context *tdb, + int rw_type, tdb_off_t offset, size_t len) { return 0; } @@ -423,7 +433,8 @@ static const struct tdb_methods transaction_methods = { transaction_next_hash_chain, transaction_oob, transaction_expand_file, - transaction_brlock + transaction_brlock, + transaction_brunlock }; /* @@ -455,7 +466,8 @@ static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t return 0; } -int tdb_transaction_cancel_internal(struct tdb_context *tdb) +/* ltype is F_WRLCK after prepare. */ +int _tdb_transaction_cancel(struct tdb_context *tdb, int ltype) { int i, ret = 0; @@ -482,38 +494,24 @@ int tdb_transaction_cancel_internal(struct tdb_context *tdb) 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; } } - /* remove any global lock created during the transaction */ - if (tdb->global_lock.count != 0) { - tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 4*tdb->header.hash_size); - tdb->global_lock.count = 0; - } - - /* remove any locks created during the transaction */ - if (tdb->num_locks != 0) { - for (i=0;inum_lockrecs;i++) { - tdb_brlock(tdb,FREELIST_TOP+4*tdb->lockrecs[i].list, - F_UNLCK,F_SETLKW, 0, 1); - } - tdb->num_locks = 0; - tdb->num_lockrecs = 0; - SAFE_FREE(tdb->lockrecs); - } + /* This also removes the OPEN_LOCK, if we have it. */ + tdb_release_extra_locks(tdb); /* restore the normal io methods */ tdb->methods = tdb->transaction->io_methods; - tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0); - tdb_transaction_unlock(tdb); + tdb_brunlock(tdb, ltype, FREELIST_TOP, 0); + tdb_transaction_unlock(tdb, F_WRLCK); SAFE_FREE(tdb->transaction->hash_heads); SAFE_FREE(tdb->transaction); @@ -546,7 +544,7 @@ int tdb_transaction_start(struct tdb_context *tdb) return 0; } - if (tdb->num_locks != 0 || tdb->global_lock.count) { + if (tdb_have_extra_locks(tdb)) { /* the caller must not have any locks when starting a transaction as otherwise we'll be screwed by lack of nested locks in posix */ @@ -585,7 +583,7 @@ int tdb_transaction_start(struct tdb_context *tdb) /* get a read lock from the freelist to the end of file. This is upgraded to a write lock during the commit */ - if (tdb_brlock(tdb, FREELIST_TOP, F_RDLCK, F_SETLKW, 0, 0) == -1) { + if (tdb_brlock(tdb, F_RDLCK, FREELIST_TOP, 0, TDB_LOCK_WAIT) == -1) { TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n")); tdb->ecode = TDB_ERR_LOCK; goto fail; @@ -621,8 +619,8 @@ int tdb_transaction_start(struct tdb_context *tdb) return 0; fail: - tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0); - tdb_transaction_unlock(tdb); + tdb_brunlock(tdb, F_RDLCK, FREELIST_TOP, 0); + tdb_transaction_unlock(tdb, F_WRLCK); SAFE_FREE(tdb->transaction->blocks); SAFE_FREE(tdb->transaction->hash_heads); SAFE_FREE(tdb->transaction); @@ -634,9 +632,12 @@ fail: cancel the current transaction */ int tdb_transaction_cancel(struct tdb_context *tdb) -{ +{ + int ltype = F_RDLCK; tdb_trace(tdb, "tdb_transaction_cancel"); - return tdb_transaction_cancel_internal(tdb); + if (tdb->transaction && tdb->transaction->prepared) + ltype = F_WRLCK; + return _tdb_transaction_cancel(tdb, ltype); } /* @@ -675,7 +676,7 @@ 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; @@ -761,7 +762,7 @@ 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; @@ -781,10 +782,10 @@ 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; @@ -864,7 +865,7 @@ 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")); @@ -885,7 +886,7 @@ static int transaction_setup_recovery(struct tdb_context *tdb, return 0; } -static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb) +static int _tdb_transaction_prepare_commit(struct tdb_context *tdb) { const struct tdb_methods *methods; @@ -896,14 +897,14 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb) if (tdb->transaction->prepared) { tdb->ecode = TDB_ERR_EINVAL; - tdb_transaction_cancel(tdb); + _tdb_transaction_cancel(tdb, F_WRLCK); TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n")); return -1; } if (tdb->transaction->transaction_error) { tdb->ecode = TDB_ERR_IO; - tdb_transaction_cancel_internal(tdb); + _tdb_transaction_cancel(tdb, F_RDLCK); TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n")); return -1; } @@ -914,11 +915,6 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb) return 0; } -#ifdef TDB_TRACE - /* store seqnum now, before reading becomes illegal. */ - tdb_ofs_read(tdb, TDB_SEQNUM_OFS, &tdb->transaction_prepare_seqnum); -#endif - /* check for a null transaction */ if (tdb->transaction->blocks == NULL) { return 0; @@ -928,10 +924,10 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb) /* 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_have_extra_locks(tdb)) { tdb->ecode = TDB_ERR_LOCK; TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n")); - tdb_transaction_cancel_internal(tdb); + _tdb_transaction_cancel(tdb, F_RDLCK); return -1; } @@ -939,16 +935,15 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb) if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) { TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n")); tdb->ecode = TDB_ERR_LOCK; - tdb_transaction_cancel_internal(tdb); + _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, GLOBAL_LOCK, F_WRLCK, F_SETLKW, 0, 1) == -1) { - TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n")); - tdb->ecode = TDB_ERR_LOCK; - tdb_transaction_cancel_internal(tdb); + if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) { + TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n")); + _tdb_transaction_cancel(tdb, F_WRLCK); return -1; } @@ -956,8 +951,7 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb) /* 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_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1); - tdb_transaction_cancel_internal(tdb); + _tdb_transaction_cancel(tdb, F_WRLCK); return -1; } } @@ -971,15 +965,14 @@ static int tdb_transaction_prepare_commit_internal(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_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1); - tdb_transaction_cancel_internal(tdb); + _tdb_transaction_cancel(tdb, F_WRLCK); return -1; } tdb->map_size = tdb->transaction->old_map_size; 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; } @@ -990,7 +983,7 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb) int tdb_transaction_prepare_commit(struct tdb_context *tdb) { tdb_trace(tdb, "tdb_transaction_prepare_commit"); - return tdb_transaction_prepare_commit_internal(tdb); + return _tdb_transaction_prepare_commit(tdb); } /* @@ -1007,13 +1000,7 @@ int tdb_transaction_commit(struct tdb_context *tdb) return -1; } - /* If we've prepared, can't read seqnum. */ - if (tdb->transaction->prepared) { - tdb_trace_seqnum(tdb, tdb->transaction_prepare_seqnum, - "tdb_transaction_commit"); - } else { - tdb_trace(tdb, "tdb_transaction_commit"); - } + tdb_trace(tdb, "tdb_transaction_commit"); if (tdb->transaction->transaction_error) { tdb->ecode = TDB_ERR_IO; @@ -1030,12 +1017,12 @@ int tdb_transaction_commit(struct tdb_context *tdb) /* check for a null transaction */ if (tdb->transaction->blocks == NULL) { - tdb_transaction_cancel_internal(tdb); + _tdb_transaction_cancel(tdb, F_RDLCK); return 0; } if (!tdb->transaction->prepared) { - int ret = tdb_transaction_prepare_commit_internal(tdb); + int ret = _tdb_transaction_prepare_commit(tdb); if (ret) return ret; } @@ -1066,8 +1053,7 @@ int tdb_transaction_commit(struct tdb_context *tdb) tdb->methods = methods; tdb_transaction_recover(tdb); - tdb_transaction_cancel_internal(tdb); - tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1); + _tdb_transaction_cancel(tdb, F_WRLCK); TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n")); return -1; @@ -1083,8 +1069,6 @@ int tdb_transaction_commit(struct tdb_context *tdb) return -1; } - tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1); - /* TODO: maybe write to some dummy hdr field, or write to magic offset without mmap, before the last sync, instead of the @@ -1104,7 +1088,7 @@ int tdb_transaction_commit(struct tdb_context *tdb) /* use a transaction cancel to free memory and remove the transaction locks */ - tdb_transaction_cancel_internal(tdb); + _tdb_transaction_cancel(tdb, F_WRLCK); if (need_repack) { return tdb_repack(tdb); @@ -1116,7 +1100,7 @@ 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) @@ -1124,7 +1108,7 @@ 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) { @@ -1211,7 +1195,7 @@ 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;