1/*- 2 * See the file LICENSE for redistribution information. 3 * 4 * Copyright (c) 1996,2008 Oracle. All rights reserved. 5 * 6 * $Id: txn.h,v 12.20 2008/01/08 20:58:18 bostic Exp $ 7 */ 8 9#ifndef _DB_TXN_H_ 10#define _DB_TXN_H_ 11 12#include "dbinc/xa.h" 13 14#if defined(__cplusplus) 15extern "C" { 16#endif 17 18/* Operation parameters to the delayed commit processing code. */ 19typedef enum { 20 TXN_CLOSE, /* Close a DB handle whose close had failed. */ 21 TXN_REMOVE, /* Remove a file. */ 22 TXN_TRADE, /* Trade lockers. */ 23 TXN_TRADED /* Already traded; downgrade lock. */ 24} TXN_EVENT_T; 25 26struct __db_txnregion; typedef struct __db_txnregion DB_TXNREGION; 27struct __txn_logrec; typedef struct __txn_logrec DB_TXNLOGREC; 28 29/* 30 * !!! 31 * TXN_MINIMUM = (DB_LOCK_MAXID + 1) but this makes compilers complain. 32 */ 33#define TXN_MINIMUM 0x80000000 34#define TXN_MAXIMUM 0xffffffff /* Maximum number of txn ids. */ 35#define TXN_INVALID 0 /* Invalid transaction ID. */ 36 37#define DEF_MAX_TXNS 100 /* Default max transactions. */ 38#define TXN_NSLOTS 4 /* Initial slots to hold DB refs */ 39 40/* 41 * Internal data maintained in shared memory for each transaction. 42 */ 43typedef struct __txn_detail { 44 u_int32_t txnid; /* current transaction id 45 used to link free list also */ 46 pid_t pid; /* Process owning txn */ 47 db_threadid_t tid; /* Thread owning txn */ 48 49 DB_LSN last_lsn; /* Last LSN written for this txn. */ 50 DB_LSN begin_lsn; /* LSN of begin record. */ 51 roff_t parent; /* Offset of transaction's parent. */ 52 roff_t name; /* Offset of txn name. */ 53 54 u_int32_t nlog_dbs; /* Number of databases used. */ 55 u_int32_t nlog_slots; /* Number of allocated slots. */ 56 roff_t log_dbs; /* Databases used. */ 57 58 DB_LSN read_lsn; /* Read LSN for MVCC. */ 59 DB_LSN visible_lsn; /* LSN at which this transaction's 60 changes are visible. */ 61 db_mutex_t mvcc_mtx; /* Version mutex. */ 62 u_int32_t mvcc_ref; /* Number of buffers created by this 63 transaction still in cache. */ 64 65 SH_TAILQ_HEAD(__tdkids) kids; /* Linked list of child txn detail. */ 66 SH_TAILQ_ENTRY klinks; 67 68 /* TXN_{ABORTED, COMMITTED PREPARED, RUNNING} */ 69 u_int32_t status; /* status of the transaction */ 70 71#define TXN_DTL_COLLECTED 0x1 /* collected during txn_recover */ 72#define TXN_DTL_RESTORED 0x2 /* prepared txn restored */ 73#define TXN_DTL_INMEMORY 0x4 /* uses in memory logs */ 74 u_int32_t flags; 75 76 /* TXN_XA_{ABORTED, DEADLOCKED, ENDED, PREPARED, STARTED, SUSPENDED} */ 77 SH_TAILQ_ENTRY links; /* active/free/snapshot list */ 78 79 u_int32_t xa_status; /* XA status */ 80 81 /* 82 * XID (xid_t) structure: because these fields are logged, the 83 * sizes have to be explicit. 84 */ 85 u_int8_t xid[XIDDATASIZE]; /* XA global transaction id */ 86 u_int32_t bqual; /* bqual_length from XID */ 87 u_int32_t gtrid; /* gtrid_length from XID */ 88 int32_t format; /* XA format */ 89 roff_t slots[TXN_NSLOTS]; /* Initial DB slot allocation. */ 90} TXN_DETAIL; 91 92/* 93 * DB_TXNMGR -- 94 * The transaction manager encapsulates the transaction system. 95 */ 96struct __db_txnmgr { 97 /* 98 * These fields need to be protected for multi-threaded support. 99 * 100 * Lock list of active transactions (including the content of each 101 * TXN_DETAIL structure on the list). 102 */ 103 db_mutex_t mutex; 104 /* List of active transactions. */ 105 TAILQ_HEAD(_chain, __db_txn) txn_chain; 106 107 u_int32_t n_discards; /* Number of txns discarded. */ 108 109 /* These fields are never updated after creation, so not protected. */ 110 ENV *env; /* Environment. */ 111 REGINFO reginfo; /* Region information. */ 112}; 113 114/* Macros to lock/unlock the transaction region as a whole. */ 115#define TXN_SYSTEM_LOCK(env) \ 116 MUTEX_LOCK(env, ((DB_TXNREGION *) \ 117 (env)->tx_handle->reginfo.primary)->mtx_region) 118#define TXN_SYSTEM_UNLOCK(env) \ 119 MUTEX_UNLOCK(env, ((DB_TXNREGION *) \ 120 (env)->tx_handle->reginfo.primary)->mtx_region) 121 122/* 123 * DB_TXNREGION -- 124 * The primary transaction data structure in the shared memory region. 125 */ 126struct __db_txnregion { 127 db_mutex_t mtx_region; /* Region mutex. */ 128 129 u_int32_t maxtxns; /* maximum number of active TXNs */ 130 u_int32_t last_txnid; /* last transaction id given out */ 131 u_int32_t cur_maxid; /* current max unused id. */ 132 133 db_mutex_t mtx_ckp; /* Single thread checkpoints. */ 134 DB_LSN last_ckp; /* lsn of the last checkpoint */ 135 time_t time_ckp; /* time of last checkpoint */ 136 137 DB_TXN_STAT stat; /* Statistics for txns. */ 138 139#define TXN_IN_RECOVERY 0x01 /* environment is being recovered */ 140 u_int32_t flags; 141 /* active TXN list */ 142 SH_TAILQ_HEAD(__active) active_txn; 143 SH_TAILQ_HEAD(__mvcc) mvcc_txn; 144}; 145 146/* 147 * DB_TXNLOGREC -- 148 * An in-memory, linked-list copy of a log record. 149 */ 150struct __txn_logrec { 151 STAILQ_ENTRY(__txn_logrec) links;/* Linked list. */ 152 153 u_int8_t data[1]; /* Log record. */ 154}; 155 156/* 157 * Log record types. Note that these are *not* alphabetical. This is 158 * intentional so that we don't change the meaning of values between 159 * software upgrades. 160 * 161 * EXPECTED, UNEXPECTED, IGNORE, and OK are used in the txnlist functions. 162 * Here is an explanation of how the statuses are used. 163 * 164 * TXN_OK 165 * BEGIN records for transactions found on the txnlist during 166 * OPENFILES (BEGIN records are those with a prev_lsn of 0,0) 167 * 168 * TXN_COMMIT 169 * Transaction committed and should be rolled forward. 170 * 171 * TXN_ABORT 172 * This transaction's changes must be undone. Either there was 173 * never a prepare or commit record for this transaction OR there 174 * was a commit, but we are recovering to a timestamp or particular 175 * LSN and that point is before this transaction's commit. 176 * 177 * TXN_PREPARE 178 * Prepare record, but no commit record is in the log. 179 * 180 * TXN_IGNORE 181 * Generic meaning is that this transaction should not be 182 * processed during later recovery passes. We use it in a 183 * number of different manners: 184 * 185 * 1. We never saw its BEGIN record. Therefore, the logs have 186 * been reclaimed and we *know* that this transaction doesn't 187 * need to be aborted, because in order for it to be 188 * reclaimed, there must have been a subsequent checkpoint 189 * (and any dirty pages for this transaction made it to 190 * disk). 191 * 192 * 2. This is a child transaction that created a database. 193 * For some reason, we don't want to recreate that database 194 * (i.e., it already exists or some other database created 195 * after it exists). 196 * 197 * 3. During recovery open of subdatabases, if the master check fails, 198 * we use a TXN_IGNORE on the create of the subdb in the nested 199 * transaction. 200 * 201 * 4. During a remove, the file with the name being removed isn't 202 * the file for which we are recovering a remove. 203 * 204 * TXN_EXPECTED 205 * After a successful open during recovery, we update the 206 * transaction's status to TXN_EXPECTED. The open was done 207 * in the parent, but in the open log record, we record the 208 * child transaction's ID if we also did a create. When there 209 * is a valid ID in that field, we use it and mark the child's 210 * status as TXN_EXPECTED (indicating that we don't need to redo 211 * a create for this file). 212 * 213 * When recovering a remove, if we don't find or can't open 214 * the file, the child (which does the remove) gets marked 215 * EXPECTED (indicating that we don't need to redo the remove). 216 * 217 * TXN_UNEXPECTED 218 * During recovery, we attempted an open that should have succeeded 219 * and we got ENOENT, so like with the EXPECTED case, we indicate 220 * in the child that we got the UNEXPECTED return so that we do redo 221 * the creating/deleting operation. 222 * 223 */ 224#define TXN_OK 0 225#define TXN_COMMIT 1 226#define TXN_PREPARE 2 227#define TXN_ABORT 3 228#define TXN_IGNORE 4 229#define TXN_EXPECTED 5 230#define TXN_UNEXPECTED 6 231 232#if defined(__cplusplus) 233} 234#endif 235 236#include "dbinc_auto/txn_auto.h" 237#include "dbinc_auto/txn_ext.h" 238#include "dbinc_auto/xa_ext.h" 239#endif /* !_DB_TXN_H_ */ 240