diff options
Diffstat (limited to 'db2/common/db_apprec.c')
| -rw-r--r-- | db2/common/db_apprec.c | 129 |
1 files changed, 83 insertions, 46 deletions
diff --git a/db2/common/db_apprec.c b/db2/common/db_apprec.c index 188c6b9..7a42e13 100644 --- a/db2/common/db_apprec.c +++ b/db2/common/db_apprec.c @@ -11,12 +11,13 @@ static const char copyright[] = "@(#) Copyright (c) 1997\n\ Sleepycat Software Inc. All rights reserved.\n"; -static const char sccsid[] = "@(#)db_apprec.c 10.19 (Sleepycat) 11/23/97"; +static const char sccsid[] = "@(#)db_apprec.c 10.23 (Sleepycat) 1/17/98"; #endif #ifndef NO_SYSTEM_INCLUDES #include <sys/types.h> +#include <errno.h> #include <time.h> #include <string.h> #include <stdlib.h> @@ -44,9 +45,9 @@ __db_apprec(dbenv, flags) { DBT data; DB_LOG *lp; - DB_LSN ckp_lsn, first_lsn, lsn, tmp_lsn; + DB_LSN ckp_lsn, first_lsn, lsn; time_t now; - int first_flag, is_thread, ret; + int is_thread, ret; void *txninfo; lp = dbenv->lg_info; @@ -63,73 +64,109 @@ __db_apprec(dbenv, flags) F_CLR(lp, DB_AM_THREAD); /* - * Read forward through the log, opening the appropriate files so that - * we can call recovery routines. In general, we start at the last - * checkpoint prior to the last checkpointed LSN. For catastrophic - * recovery, we begin at the first LSN that appears in any log file - * (log_get figures this out for us when we pass it the DB_FIRST flag). + * Recovery is done in three passes: + * Pass #1: + * Read forward through the log from the last checkpoint to the + * end of the log, opening and closing files so that at the end + * of the log we have the "current" set of files open. + * Pass #2: + * Read backward through the log undoing any uncompleted TXNs. + * If doing catastrophic recovery, we read to the beginning of + * the log, otherwise, to the most recent checkpoint that occurs + * before the most recent checkpoint LSN, which is returned by + * __log_findckp(). During this pass, checkpoint file information + * is ignored, and file openings and closings are undone. + * Pass #3: + * Read forward through the log from the LSN found in pass #2, + * redoing any committed TXNs. During this pass, checkpoint + * file information is ignored, and file openings and closings + * are redone. */ - if (LF_ISSET(DB_RECOVER_FATAL)) - first_flag = DB_FIRST; - else { - if ((ret = __log_findckp(lp, &lsn)) == DB_NOTFOUND) - goto out; - first_flag = DB_SET; - } - /* If we're a threaded application, we have to allocate space. */ + /* + * Find the last checkpoint in the log. This is the point from which + * we want to begin pass #1 (the TXN_OPENFILES pass). + */ memset(&data, 0, sizeof(data)); - if ((ret = log_get(lp, &lsn, &data, first_flag)) != 0) { - __db_err(dbenv, "Failure: unable to get log record"); - if (first_flag == DB_SET) - __db_err(dbenv, "Retrieving LSN %lu %lu", - (u_long)lsn.file, (u_long)lsn.offset); - else - __db_err(dbenv, "Retrieving first LSN"); - goto out; + if ((ret = log_get(lp, &ckp_lsn, &data, DB_CHECKPOINT)) != 0) { + /* + * If we don't find a checkpoint, start from the beginning. + * If that fails, we're done. Note, we require that there + * be log records if we're performing recovery, and fail if + * there aren't. + */ + if ((ret = log_get(lp, &ckp_lsn, &data, DB_FIRST)) != 0) { + __db_err(dbenv, "First log record not found"); + if (ret == DB_NOTFOUND) + ret = EINVAL; + goto out; + } } - first_lsn = lsn; + /* + * Now, ckp_lsn is either the lsn of the last checkpoint or the lsn + * of the first record in the log. Begin the TXN_OPENFILES pass from + * that lsn, and proceed to the end of the log. + */ + lsn = ckp_lsn; for (;;) { ret = __db_dispatch(lp, &data, &lsn, TXN_OPENFILES, txninfo); if (ret != 0 && ret != DB_TXN_CKP) goto msgerr; - if ((ret = - log_get(dbenv->lg_info, &lsn, &data, DB_NEXT)) != 0) { - if (ret != DB_NOTFOUND) - goto out; - break; + if ((ret = log_get(lp, &lsn, &data, DB_NEXT)) != 0) { + if (ret == DB_NOTFOUND) + break; + goto out; } } /* - * Initialize the ckp_lsn to 0,0. If we never find a valid - * checkpoint in the log, then leaving ckp_lsn at 0,0 is correct. + * Pass #2. + * + * Before we can begin pass #2, backward roll phase, we determine how + * far back in the log to recover. If we are doing catastrophic + * recovery, then we go as far back as we have files. If we are + * doing normal recovery, we go as back to the most recent checkpoint + * that occurs before the most recent checkpoint LSN. */ - ZERO_LSN(ckp_lsn); + if (LF_ISSET(DB_RECOVER_FATAL)) { + ZERO_LSN(first_lsn); + } else + if ((ret = __log_findckp(lp, &first_lsn)) == DB_NOTFOUND) { + /* + * If recovery was specified, there must be log files. + * If we don't find one, it's an error. (This should + * have been caught above, when a log_get() of DB_FIRST + * or DB_CHECKPOINT succeeded, but paranoia is good.) + */ + ret = EINVAL; + goto out; + } for (ret = log_get(lp, &lsn, &data, DB_LAST); ret == 0 && log_compare(&lsn, &first_lsn) > 0; - ret = log_get(lp,&lsn, &data, DB_PREV)) { - tmp_lsn = lsn; + ret = log_get(lp, &lsn, &data, DB_PREV)) { ret = __db_dispatch(lp, &data, &lsn, TXN_BACKWARD_ROLL, txninfo); - if (ret == DB_TXN_CKP) { - if (IS_ZERO_LSN(ckp_lsn)) - ckp_lsn = tmp_lsn; - ret = 0; - } else if (ret != 0) - goto msgerr; + if (ret != 0) + if (ret != DB_TXN_CKP) + goto msgerr; + else + ret = 0; } if (ret != 0 && ret != DB_NOTFOUND) goto out; + /* + * Pass #3. + */ for (ret = log_get(lp, &lsn, &data, DB_NEXT); ret == 0; ret = log_get(lp, &lsn, &data, DB_NEXT)) { ret = __db_dispatch(lp, &data, &lsn, TXN_FORWARD_ROLL, txninfo); - if (ret == DB_TXN_CKP) - ret = 0; - else if (ret != 0) - goto msgerr; + if (ret != 0) + if (ret != DB_TXN_CKP) + goto msgerr; + else + ret = 0; } if (ret != DB_NOTFOUND) goto out; @@ -149,7 +186,7 @@ __db_apprec(dbenv, flags) goto out; if (dbenv->db_verbose) { - __db_err(lp->dbenv, "Recovery complete at %s", ctime(&now)); + __db_err(lp->dbenv, "Recovery complete at %.24s", ctime(&now)); __db_err(lp->dbenv, "%s %lu %s [%lu][%lu]", "Maximum transaction id", (u_long)dbenv->tx_info->region->last_txnid, |