From d3805810d30fb0a900ad73eb9104368ca2222641 Mon Sep 17 00:00:00 2001 From: jbj Date: Fri, 15 Nov 2002 18:26:16 +0000 Subject: [PATCH] Apply patch.4.1.24.2: 2.Fix a bug where replicas could not remove log files because the checkpoint lsn was not being updated properly. CVS patchset: 5865 CVS date: 2002/11/15 18:26:16 --- db/dbinc_auto/int_def.in | 19 +- db/dbinc_auto/txn_ext.h | 2 +- db/rep/rep_record.c | 3 + db/txn/txn.c | 1175 ++++++++++++++++++++++++++++++++++------------ 4 files changed, 877 insertions(+), 322 deletions(-) diff --git a/db/dbinc_auto/int_def.in b/db/dbinc_auto/int_def.in index 3b4b5c6..e75e191 100644 --- a/db/dbinc_auto/int_def.in +++ b/db/dbinc_auto/int_def.in @@ -157,6 +157,7 @@ #define __db_pr __db_pr@DB_VERSION_UNIQUE_NAME@ #define __db_prdbt __db_prdbt@DB_VERSION_UNIQUE_NAME@ #define __db_prflags __db_prflags@DB_VERSION_UNIQUE_NAME@ +#define __db_dbtype_to_string __db_dbtype_to_string@DB_VERSION_UNIQUE_NAME@ #define __db_prheader __db_prheader@DB_VERSION_UNIQUE_NAME@ #define __db_prfooter __db_prfooter@DB_VERSION_UNIQUE_NAME@ #define __db_addrem_recover __db_addrem_recover@DB_VERSION_UNIQUE_NAME@ @@ -440,7 +441,9 @@ #define __dbreg_rem_dbentry __dbreg_rem_dbentry@DB_VERSION_UNIQUE_NAME@ #define __dbreg_open_files __dbreg_open_files@DB_VERSION_UNIQUE_NAME@ #define __dbreg_close_files __dbreg_close_files@DB_VERSION_UNIQUE_NAME@ +#define __dbreg_nofiles __dbreg_nofiles@DB_VERSION_UNIQUE_NAME@ #define __dbreg_id_to_db __dbreg_id_to_db@DB_VERSION_UNIQUE_NAME@ +#define __dbreg_id_to_db_int __dbreg_id_to_db_int@DB_VERSION_UNIQUE_NAME@ #define __dbreg_id_to_fname __dbreg_id_to_fname@DB_VERSION_UNIQUE_NAME@ #define __dbreg_fid_to_fname __dbreg_fid_to_fname@DB_VERSION_UNIQUE_NAME@ #define __dbreg_get_name __dbreg_get_name@DB_VERSION_UNIQUE_NAME@ @@ -449,6 +452,7 @@ #define __dbreg_push_id __dbreg_push_id@DB_VERSION_UNIQUE_NAME@ #define __dbreg_pop_id __dbreg_pop_id@DB_VERSION_UNIQUE_NAME@ #define __dbreg_pluck_id __dbreg_pluck_id@DB_VERSION_UNIQUE_NAME@ +#define __dbreg_print_dblist __dbreg_print_dblist@DB_VERSION_UNIQUE_NAME@ #define __db_shalloc_init __db_shalloc_init@DB_VERSION_UNIQUE_NAME@ #define __db_shalloc_size __db_shalloc_size@DB_VERSION_UNIQUE_NAME@ #define __db_shalloc __db_shalloc@DB_VERSION_UNIQUE_NAME@ @@ -831,6 +835,7 @@ #define __qam_sync __qam_sync@DB_VERSION_UNIQUE_NAME@ #define __qam_gen_filelist __qam_gen_filelist@DB_VERSION_UNIQUE_NAME@ #define __qam_extent_names __qam_extent_names@DB_VERSION_UNIQUE_NAME@ +#define __qam_exid __qam_exid@DB_VERSION_UNIQUE_NAME@ #define __qam_db_create __qam_db_create@DB_VERSION_UNIQUE_NAME@ #define __qam_db_close __qam_db_close@DB_VERSION_UNIQUE_NAME@ #define __db_prqueue __db_prqueue@DB_VERSION_UNIQUE_NAME@ @@ -1161,13 +1166,13 @@ #define __txn_discard __txn_discard@DB_VERSION_UNIQUE_NAME@ #define __txn_prepare __txn_prepare@DB_VERSION_UNIQUE_NAME@ #define __txn_id __txn_id@DB_VERSION_UNIQUE_NAME@ -#define __txn_free __txn_free@DB_VERSION_UNIQUE_NAME@ #define __txn_checkpoint __txn_checkpoint@DB_VERSION_UNIQUE_NAME@ #define __txn_getckp __txn_getckp@DB_VERSION_UNIQUE_NAME@ #define __txn_activekids __txn_activekids@DB_VERSION_UNIQUE_NAME@ #define __txn_force_abort __txn_force_abort@DB_VERSION_UNIQUE_NAME@ #define __txn_preclose __txn_preclose@DB_VERSION_UNIQUE_NAME@ #define __txn_reset __txn_reset@DB_VERSION_UNIQUE_NAME@ +#define __txn_updateckp __txn_updateckp@DB_VERSION_UNIQUE_NAME@ #define __txn_regop_log __txn_regop_log@DB_VERSION_UNIQUE_NAME@ #define __txn_regop_getpgnos __txn_regop_getpgnos@DB_VERSION_UNIQUE_NAME@ #define __txn_regop_print __txn_regop_print@DB_VERSION_UNIQUE_NAME@ @@ -1211,16 +1216,6 @@ #define __txn_lockevent __txn_lockevent@DB_VERSION_UNIQUE_NAME@ #define __txn_remlock __txn_remlock@DB_VERSION_UNIQUE_NAME@ #define __txn_doevents __txn_doevents@DB_VERSION_UNIQUE_NAME@ -#define __db_xa_open __db_xa_open@DB_VERSION_UNIQUE_NAME@ -#define __db_xa_close __db_xa_close@DB_VERSION_UNIQUE_NAME@ -#define __db_xa_start __db_xa_start@DB_VERSION_UNIQUE_NAME@ -#define __db_xa_end __db_xa_end@DB_VERSION_UNIQUE_NAME@ -#define __db_xa_prepare __db_xa_prepare@DB_VERSION_UNIQUE_NAME@ -#define __db_xa_commit __db_xa_commit@DB_VERSION_UNIQUE_NAME@ -#define __db_xa_recover __db_xa_recover@DB_VERSION_UNIQUE_NAME@ -#define __db_xa_rollback __db_xa_rollback@DB_VERSION_UNIQUE_NAME@ -#define __db_xa_forget __db_xa_forget@DB_VERSION_UNIQUE_NAME@ -#define __db_xa_complete __db_xa_complete@DB_VERSION_UNIQUE_NAME@ #define __db_xa_create __db_xa_create@DB_VERSION_UNIQUE_NAME@ #define __db_rmid_to_env __db_rmid_to_env@DB_VERSION_UNIQUE_NAME@ #define __db_xid_to_txn __db_xid_to_txn@DB_VERSION_UNIQUE_NAME@ @@ -1228,8 +1223,6 @@ #define __db_unmap_rmid __db_unmap_rmid@DB_VERSION_UNIQUE_NAME@ #define __db_map_xid __db_map_xid@DB_VERSION_UNIQUE_NAME@ #define __db_unmap_xid __db_unmap_xid@DB_VERSION_UNIQUE_NAME@ -#define __db_get_xa_txn __db_get_xa_txn@DB_VERSION_UNIQUE_NAME@ -#define __db_get_current_thread __db_get_current_thread@DB_VERSION_UNIQUE_NAME@ #define xdr___env_cachesize_msg xdr___env_cachesize_msg@DB_VERSION_UNIQUE_NAME@ #define xdr___env_cachesize_reply xdr___env_cachesize_reply@DB_VERSION_UNIQUE_NAME@ #define xdr___env_close_msg xdr___env_close_msg@DB_VERSION_UNIQUE_NAME@ diff --git a/db/dbinc_auto/txn_ext.h b/db/dbinc_auto/txn_ext.h index d297736..a53338d 100644 --- a/db/dbinc_auto/txn_ext.h +++ b/db/dbinc_auto/txn_ext.h @@ -14,13 +14,13 @@ int __txn_abort __P((DB_TXN *)); int __txn_discard __P((DB_TXN *, u_int32_t flags)); int __txn_prepare __P((DB_TXN *, u_int8_t *)); u_int32_t __txn_id __P((DB_TXN *)); -int __txn_free __P((DB_TXN *)); int __txn_checkpoint __P((DB_ENV *, u_int32_t, u_int32_t, u_int32_t)); int __txn_getckp __P((DB_ENV *, DB_LSN *)); int __txn_activekids __P((DB_ENV *, u_int32_t, DB_TXN *)); int __txn_force_abort __P((DB_ENV *, u_int8_t *)); int __txn_preclose __P((DB_ENV *)); int __txn_reset __P((DB_ENV *)); +void __txn_updateckp __P((DB_ENV *, DB_LSN *)); int __txn_regop_log __P((DB_ENV *, DB_TXN *, DB_LSN *, u_int32_t, u_int32_t, int32_t)); int __txn_regop_getpgnos __P((DB_ENV *, DBT *, DB_LSN *, db_recops, void *)); int __txn_regop_print __P((DB_ENV *, DBT *, DB_LSN *, db_recops, void *)); diff --git a/db/rep/rep_record.c b/db/rep/rep_record.c index 436dca1..5fe6843 100644 --- a/db/rep/rep_record.c +++ b/db/rep/rep_record.c @@ -1198,6 +1198,9 @@ gap_check: lp->wait_recs = 0; * replica get flushed now and again. */ ret = dbenv->log_flush(dbenv, &ckp_lsn); + /* Update the last_ckp in the txn region. */ + if (ret == 0) + __txn_updateckp(dbenv, &rp->lsn); break; case DB___txn_regop: if (!F_ISSET(dbenv, DB_ENV_REP_LOGSONLY)) diff --git a/db/txn/txn.c b/db/txn/txn.c index 0f6d894..5a0b43f 100644 --- a/db/txn/txn.c +++ b/db/txn/txn.c @@ -1,7 +1,7 @@ /*- * See the file LICENSE for redistribution information. * - * Copyright (c) 1996, 1997, 1998, 1999, 2000 + * Copyright (c) 1996-2002 * Sleepycat Software. All rights reserved. */ /* @@ -39,11 +39,12 @@ #include "db_config.h" #ifndef lint -static const char revid[] = "$Id: txn.c,v 11.61 2001/01/10 18:18:52 bostic Exp $"; +static const char revid[] = "Id: txn.c,v 11.179 2002/08/29 17:41:17 margo Exp "; #endif /* not lint */ #ifndef NO_SYSTEM_INCLUDES #include +#include #if TIME_WITH_SYS_TIME #include @@ -59,58 +60,129 @@ static const char revid[] = "$Id: txn.c,v 11.61 2001/01/10 18:18:52 bostic Exp $ #include #endif -#ifdef HAVE_RPC -#include "db_server.h" -#endif - #include "db_int.h" -#include "db_shash.h" -#include "txn.h" -#include "lock.h" -#include "log.h" -#include "db_dispatch.h" -#include "db_page.h" -#include "db_ext.h" - -#ifdef HAVE_RPC -#include "gen_client_ext.h" -#include "rpc_client_ext.h" -#endif +#include "dbinc/crypto.h" +#include "dbinc/hmac.h" +#include "dbinc/db_page.h" +#include "dbinc/db_shash.h" +#include "dbinc/hash.h" +#include "dbinc/lock.h" +#include "dbinc/log.h" +#include "dbinc/txn.h" + +#define SET_LOG_FLAGS(dbenv, txnp, lflags) \ + do { \ + lflags = DB_COMMIT | DB_PERMANENT; \ + if (F_ISSET(txnp, TXN_SYNC)) \ + lflags |= DB_FLUSH; \ + else if (!F_ISSET(txnp, TXN_NOSYNC) && \ + !F_ISSET(dbenv, DB_ENV_TXN_NOSYNC)) { \ + if (F_ISSET(dbenv, DB_ENV_TXN_WRITE_NOSYNC)) \ + lflags |= DB_WRNOSYNC; \ + else \ + lflags |= DB_FLUSH; \ + } \ + } while (0) -static int __txn_begin __P((DB_TXN *)); -static int __txn_isvalid __P((const DB_TXN *, TXN_DETAIL **, u_int32_t)); +/* + * __txn_isvalid enumerated types. We cannot simply use the transaction + * statuses, because different statuses need to be handled differently + * depending on the caller. + */ +typedef enum { + TXN_OP_ABORT, + TXN_OP_COMMIT, + TXN_OP_DISCARD, + TXN_OP_PREPARE +} txnop_t; + +static int __txn_begin_int __P((DB_TXN *, int)); +static int __txn_end __P((DB_TXN *, int)); +static int __txn_isvalid __P((const DB_TXN *, TXN_DETAIL **, txnop_t)); +static int __txn_set_timeout __P(( DB_TXN *, db_timeout_t, u_int32_t)); static int __txn_undo __P((DB_TXN *)); +#ifndef db_create /* + * txn_abort -- * txn_begin -- - * This is a wrapper to the actual begin process. Normal txn_begin() - * allocates a DB_TXN structure for the caller, while txn_xa_begin() does - * not. Other than that, both call into the common __txn_begin code(). + * txn_commit -- + * + * When we switched to methods in 4.0, we guessed txn_{abort,begin,commit} + * were the interfaces applications would likely use and not be willing to + * change, due to the sheer volume of the calls. Provide wrappers -- we + * could do txn_abort and txn_commit using macros, but not txn_begin, as + * the name of the field is txn_begin, we didn't want to modify it. + * + * The issue with txn_begin hits us in another way. If configured with the + * --with-uniquename option, we use #defines to re-define DB's interfaces + * to unique names. We can't do that for these functions because txn_begin + * is also a field name in the DB_ENV structure, and the #defines we use go + * at the end of the db.h file -- we get control too late to #define a field + * name. So, modify the script that generates the unique names #defines to + * not generate them for these three functions, and don't include the three + * functions in libraries built with that configuration option. + * + * EXTERN: int txn_abort __P((DB_TXN *)); + * EXTERN: int txn_begin __P((DB_ENV *, DB_TXN *, DB_TXN **, u_int32_t)); + * EXTERN: int txn_commit __P((DB_TXN *, u_int32_t)); + */ +int +txn_abort(txnp) + DB_TXN *txnp; +{ + return (txnp->abort(txnp)); +} + +int +txn_begin(dbenv, parent, txnpp, flags) + DB_ENV *dbenv; + DB_TXN *parent, **txnpp; + u_int32_t flags; +{ + return (dbenv->txn_begin(dbenv, parent, txnpp, flags)); +} + +int +txn_commit(txnp, flags) + DB_TXN *txnp; + u_int32_t flags; +{ + return (txnp->commit(txnp, flags)); +} +#endif /* !db_create */ + +/* + * __txn_begin -- + * This is a wrapper to the actual begin process. Normal transaction + * begin allocates a DB_TXN structure for the caller, while XA transaction + * begin does not. Other than that, both call into common __txn_begin_int + * code. * * Internally, we use TXN_DETAIL structures, but the DB_TXN structure * provides access to the transaction ID and the offset in the transaction * region of the TXN_DETAIL structure. + * + * PUBLIC: int __txn_begin __P((DB_ENV *, DB_TXN *, DB_TXN **, u_int32_t)); */ int -txn_begin(dbenv, parent, txnpp, flags) +__txn_begin(dbenv, parent, txnpp, flags) DB_ENV *dbenv; DB_TXN *parent, **txnpp; u_int32_t flags; { + DB_LOCKREGION *region; DB_TXN *txn; int ret; -#ifdef HAVE_RPC - if (F_ISSET(dbenv, DB_ENV_RPCCLIENT)) - return (__dbcl_txn_begin(dbenv, parent, txnpp, flags)); -#endif - + *txnpp = NULL; PANIC_CHECK(dbenv); - ENV_REQUIRES_CONFIG(dbenv, dbenv->tx_handle, DB_INIT_TXN); + ENV_REQUIRES_CONFIG(dbenv, dbenv->tx_handle, "txn_begin", DB_INIT_TXN); if ((ret = __db_fchk(dbenv, "txn_begin", flags, - DB_TXN_NOWAIT | DB_TXN_NOSYNC | DB_TXN_SYNC)) != 0) + DB_DIRTY_READ | DB_TXN_NOWAIT | + DB_TXN_NOSYNC | DB_TXN_SYNC)) != 0) return (ret); if ((ret = __db_fcchk(dbenv, "txn_begin", flags, DB_TXN_NOSYNC, DB_TXN_SYNC)) != 0) @@ -122,7 +194,10 @@ txn_begin(dbenv, parent, txnpp, flags) txn->mgrp = dbenv->tx_handle; txn->parent = parent; TAILQ_INIT(&txn->kids); + TAILQ_INIT(&txn->events); txn->flags = TXN_MALLOC; + if (LF_ISSET(DB_DIRTY_READ)) + F_SET(txn, TXN_DIRTY_READ); if (LF_ISSET(DB_TXN_NOSYNC)) F_SET(txn, TXN_NOSYNC); if (LF_ISSET(DB_TXN_SYNC)) @@ -130,15 +205,41 @@ txn_begin(dbenv, parent, txnpp, flags) if (LF_ISSET(DB_TXN_NOWAIT)) F_SET(txn, TXN_NOWAIT); - if ((ret = __txn_begin(txn)) != 0) { - __os_free(txn, sizeof(DB_TXN)); - txn = NULL; - } + if ((ret = __txn_begin_int(txn, 0)) != 0) + goto err; - if (txn != NULL && parent != NULL) + if (parent != NULL) TAILQ_INSERT_HEAD(&parent->kids, txn, klinks); + if (LOCKING_ON(dbenv)) { + region = ((DB_LOCKTAB *)dbenv->lk_handle)->reginfo.primary; + if (parent != NULL) { + ret = __lock_inherit_timeout(dbenv, + parent->txnid, txn->txnid); + /* No parent locker set yet. */ + if (ret == EINVAL) { + parent = NULL; + ret = 0; + } + if (ret != 0) + goto err; + } + + /* + * Parent is NULL if we have no parent + * or it has no timeouts set. + */ + if (parent == NULL && region->tx_timeout != 0) + if ((ret = __lock_set_timeout(dbenv, txn->txnid, + region->tx_timeout, DB_SET_TXN_TIMEOUT)) != 0) + goto err; + } + *txnpp = txn; + return (0); + +err: + __os_free(dbenv, txn); return (ret); } @@ -158,26 +259,60 @@ __txn_xa_begin(dbenv, txn) memset(txn, 0, sizeof(DB_TXN)); txn->mgrp = dbenv->tx_handle; + TAILQ_INIT(&txn->kids); + TAILQ_INIT(&txn->events); - return (__txn_begin(txn)); + return (__txn_begin_int(txn, 0)); } /* - * __txn_begin -- + * __txn_compensate_begin + * Begin an compensation transaction. This is a special interface + * that is used only for transactions that must be started to compensate + * for actions during an abort. Currently only used for allocations. + * + * PUBLIC: int __txn_compensate_begin __P((DB_ENV *, DB_TXN **txnp)); + */ +int +__txn_compensate_begin(dbenv, txnpp) + DB_ENV *dbenv; + DB_TXN **txnpp; +{ + DB_TXN *txn; + int ret; + + PANIC_CHECK(dbenv); + + if ((ret = __os_calloc(dbenv, 1, sizeof(DB_TXN), &txn)) != 0) + return (ret); + + txn->mgrp = dbenv->tx_handle; + TAILQ_INIT(&txn->kids); + TAILQ_INIT(&txn->events); + txn->flags = TXN_MALLOC; + F_SET(txn, TXN_COMPENSATE); + + *txnpp = txn; + return (__txn_begin_int(txn, 1)); +} + +/* + * __txn_begin_int -- * Normal DB version of txn_begin. */ static int -__txn_begin(txn) +__txn_begin_int(txn, internal) DB_TXN *txn; + int internal; { DB_ENV *dbenv; - DB_LSN begin_lsn; + DB_LSN begin_lsn, null_lsn; DB_TXNMGR *mgr; DB_TXNREGION *region; TXN_DETAIL *td; size_t off; - u_int32_t id; - int ret; + u_int32_t id, *ids; + int nids, ret; mgr = txn->mgrp; dbenv = mgr->dbenv; @@ -188,36 +323,73 @@ __txn_begin(txn) * need never write records for read-only transactions). However, * we do need to find the current LSN so that we can store it in the * transaction structure, so we can know where to take checkpoints. + * + * XXX + * We should set this value when we write the first log record, not + * here. */ - if (LOGGING_ON(dbenv) && - (ret = log_put(dbenv, &begin_lsn, NULL, DB_CURLSN)) != 0) - goto err2; + if (DBENV_LOGGING(dbenv)) + __log_txn_lsn(dbenv, &begin_lsn, NULL, NULL); R_LOCK(dbenv, &mgr->reginfo); + if (!F_ISSET(txn, TXN_COMPENSATE) && F_ISSET(region, TXN_IN_RECOVERY)) { + __db_err(dbenv, "operation not permitted during recovery"); + ret = EINVAL; + goto err; + } - /* Make sure that last_txnid is not going to wrap around. */ - if (region->last_txnid == TXN_INVALID) { + /* Make sure that we aren't still recovering prepared transactions. */ + if (!internal && region->stat.st_nrestores != 0) { __db_err(dbenv, -"txn_begin: transaction ID wrapped. Exit the database environment\nand restart the application as if application failure had occurred"); + "recovery of prepared but not yet committed transactions is incomplete"); ret = EINVAL; - goto err1; + goto err; + } + + /* + * Allocate a new transaction id. Our current valid range can span + * the maximum valid value, so check for it and wrap manually. + */ + if (region->last_txnid == TXN_MAXIMUM && + region->cur_maxid != TXN_MAXIMUM) + region->last_txnid = TXN_MINIMUM - 1; + + if (region->last_txnid == region->cur_maxid) { + if ((ret = __os_malloc(dbenv, + sizeof(u_int32_t) * region->maxtxns, &ids)) != 0) + goto err; + nids = 0; + for (td = SH_TAILQ_FIRST(®ion->active_txn, __txn_detail); + td != NULL; + td = SH_TAILQ_NEXT(td, links, __txn_detail)) + ids[nids++] = td->txnid; + region->last_txnid = TXN_MINIMUM - 1; + region->cur_maxid = TXN_MAXIMUM; + if (nids != 0) + __db_idspace(ids, nids, + ®ion->last_txnid, ®ion->cur_maxid); + __os_free(dbenv, ids); + if (DBENV_LOGGING(dbenv) && + (ret = __txn_recycle_log(dbenv, NULL, + &null_lsn, 0, region->last_txnid, region->cur_maxid)) != 0) + goto err; } /* Allocate a new transaction detail structure. */ if ((ret = __db_shalloc(mgr->reginfo.addr, sizeof(TXN_DETAIL), 0, &td)) != 0) { __db_err(dbenv, - "Unable to allocate memory for transaction detail"); - goto err1; + "Unable to allocate memory for transaction detail"); + goto err; } /* Place transaction on active transaction list. */ SH_TAILQ_INSERT_HEAD(®ion->active_txn, td, links, __txn_detail); id = ++region->last_txnid; - ++region->nbegins; - if (++region->nactive > region->maxnactive) - region->maxnactive = region->nactive; + ++region->stat.st_nbegins; + if (++region->stat.st_nactive > region->stat.st_maxnactive) + region->stat.st_maxnactive = region->stat.st_nactive; td->txnid = id; td->begin_lsn = begin_lsn; @@ -228,12 +400,20 @@ __txn_begin(txn) else td->parent = INVALID_ROFF; + td->flags = 0; off = R_OFFSET(&mgr->reginfo, td); R_UNLOCK(dbenv, &mgr->reginfo); ZERO_LSN(txn->last_lsn); txn->txnid = id; - txn->off = off; + txn->off = (u_int32_t)off; + + txn->abort = __txn_abort; + txn->commit = __txn_commit; + txn->discard = __txn_discard; + txn->id = __txn_id; + txn->prepare = __txn_prepare; + txn->set_timeout = __txn_set_timeout; /* * If this is a transaction family, we must link the child to the @@ -242,7 +422,7 @@ __txn_begin(txn) if (txn->parent != NULL && LOCKING_ON(dbenv)) if ((ret = __lock_addfamilylocker(dbenv, txn->parent->txnid, txn->txnid)) != 0) - goto err2; + return (ret); if (F_ISSET(txn, TXN_MALLOC)) { MUTEX_THREAD_LOCK(dbenv, mgr->mutexp); @@ -252,34 +432,33 @@ __txn_begin(txn) return (0); -err1: R_UNLOCK(dbenv, &mgr->reginfo); - -err2: return (ret); +err: R_UNLOCK(dbenv, &mgr->reginfo); + return (ret); } /* - * txn_commit -- + * __txn_commit -- * Commit a transaction. + * + * PUBLIC: int __txn_commit __P((DB_TXN *, u_int32_t)); */ int -txn_commit(txnp, flags) +__txn_commit(txnp, flags) DB_TXN *txnp; u_int32_t flags; { DB_ENV *dbenv; + DB_LOCKREQ request; DB_TXN *kid; - int is_commit, ret, t_ret; + TXN_DETAIL *td; + u_int32_t lflags; + int ret, t_ret; dbenv = txnp->mgrp->dbenv; -#ifdef HAVE_RPC - if (F_ISSET(dbenv, DB_ENV_RPCCLIENT)) - return (__dbcl_txn_commit(txnp, flags)); -#endif - PANIC_CHECK(dbenv); - if ((ret = __txn_isvalid(txnp, NULL, TXN_COMMITTED)) != 0) + if ((ret = __txn_isvalid(txnp, &td, TXN_OP_COMMIT)) != 0) return (ret); /* @@ -290,10 +469,10 @@ txn_commit(txnp, flags) * specifying the wrong flag for some reason. */ if (__db_fchk(dbenv, - "txn_commit", flags, DB_TXN_NOSYNC | DB_TXN_SYNC) != 0) + "DB_TXN->commit", flags, DB_TXN_NOSYNC | DB_TXN_SYNC) != 0) flags = DB_TXN_SYNC; if (__db_fcchk(dbenv, - "txn_commit", flags, DB_TXN_NOSYNC, DB_TXN_SYNC) != 0) + "DB_TXN->commit", flags, DB_TXN_NOSYNC, DB_TXN_SYNC) != 0) flags = DB_TXN_SYNC; if (LF_ISSET(DB_TXN_NOSYNC)) { F_CLR(txnp, TXN_SYNC); @@ -305,16 +484,33 @@ txn_commit(txnp, flags) } /* - * Commit any unresolved children. If there's an error, abort any - * unresolved children and the parent. + * Commit any unresolved children. If anyone fails to commit, + * then try to abort the rest of the kids and then abort the parent. + * Abort should never fail; if it does, we bail out immediately. */ while ((kid = TAILQ_FIRST(&txnp->kids)) != NULL) - if ((ret = txn_commit(kid, flags)) != 0) { + if ((ret = kid->commit(kid, flags)) != 0) while ((kid = TAILQ_FIRST(&txnp->kids)) != NULL) - (void)txn_abort(kid); - (void)txn_abort(txnp); - goto err; - } + if ((t_ret = kid->abort(kid)) != 0) + return (__db_panic(dbenv, t_ret)); + + /* + * Process any aborted pages from our children. + * We delay putting pages on the free list that are newly + * allocated and then aborted so that we can undo other + * allocations, if necessary, without worrying about + * these pages which were not on the free list before. + */ + if (txnp->txn_list != NULL) { + t_ret = __db_do_the_limbo(dbenv, NULL, txnp, txnp->txn_list); + __db_txnlist_end(dbenv, txnp->txn_list); + txnp->txn_list = NULL; + if (t_ret != 0 && ret == 0) + ret = t_ret; + } + + if (ret != 0) + goto err; /* * If there are any log records, write a log record and sync the log, @@ -322,27 +518,35 @@ txn_commit(txnp, flags) * we do not need to commit the child synchronously since it may still * abort (if its parent aborts), and otherwise its parent or ultimate * ancestor will write synchronously. - * - * I'd rather return a logging error than a flag-wrong error, so if - * the log routines fail, set "ret" without regard to previous value. */ - if (LOGGING_ON(dbenv) && !IS_ZERO_LSN(txnp->last_lsn)) { + if (DBENV_LOGGING(dbenv) && !IS_ZERO_LSN(txnp->last_lsn)) { if (txnp->parent == NULL) { - if ((t_ret = __txn_regop_log(dbenv, - txnp, &txnp->last_lsn, - (F_ISSET(dbenv, DB_ENV_TXN_NOSYNC) && - !F_ISSET(txnp, TXN_SYNC)) || - F_ISSET(txnp, TXN_NOSYNC) ? 0 : DB_FLUSH, - TXN_COMMIT, (int32_t)time(NULL))) != 0) { - ret = t_ret; + /* + * We are about to free all the read locks + * for this transaction below. Some of those + * locks might be handle locks which should + * not be freed, because they will be freed + * when the handle is closed. Check the + * events and preprocess any trades now so + * that we don't release the locks below. + */ + if ((ret = __txn_doevents(dbenv, txnp, 0, 1)) != 0) + goto err; + request.op = DB_LOCK_PUT_READ; + if (LOCKING_ON(dbenv) && (ret = dbenv->lock_vec( + dbenv, txnp->txnid, 0, &request, 1, NULL)) != 0) + goto err; + + SET_LOG_FLAGS(dbenv, txnp, lflags); + if ((ret = __txn_regop_log(dbenv, + txnp, &txnp->last_lsn, lflags, + TXN_COMMIT, (int32_t)time(NULL))) != 0) goto err; - } } else { /* Log the commit in the parent! */ - if ((t_ret = __txn_child_log(dbenv, + if ((ret = __txn_child_log(dbenv, txnp->parent, &txnp->parent->last_lsn, 0, txnp->txnid, &txnp->last_lsn)) != 0) { - ret = t_ret; goto err; } @@ -350,108 +554,206 @@ txn_commit(txnp, flags) } } - is_commit = 1; - if (0) { -err: is_commit = 0; - } - if ((t_ret = __txn_end(txnp, is_commit)) != 0 && ret == 0) + /* This is OK because __txn_end can only fail with a panic. */ + return (__txn_end(txnp, 1)); + +err: /* + * If we are prepared, then we "must" be able to commit. We + * panic here because even though the coordinator might be + * able to retry it is not clear it would know to do that. + * Otherwise we'll try to abort. If that is successful, + * then we return whatever was in ret (i.e., the reason we failed). + * If the abort was unsuccessful, then abort probably returned + * DB_RUNRECOVERY and we need to propagate that up. + */ + if (td->status == TXN_PREPARED) + return (__db_panic(dbenv, ret)); + + if ((t_ret = txnp->abort(txnp)) != 0) ret = t_ret; return (ret); } /* - * txn_abort -- + * __txn_abort -- * Abort a transaction. + * + * PUBLIC: int __txn_abort __P((DB_TXN *)); */ int -txn_abort(txnp) +__txn_abort(txnp) DB_TXN *txnp; { DB_ENV *dbenv; + DB_LOCKREQ request; DB_TXN *kid; - int ret, t_ret; + TXN_DETAIL *td; + u_int32_t lflags; + int ret; dbenv = txnp->mgrp->dbenv; -#ifdef HAVE_RPC - if (F_ISSET(dbenv, DB_ENV_RPCCLIENT)) - return (__dbcl_txn_abort(txnp)); -#endif + PANIC_CHECK(dbenv); + + /* Ensure that abort always fails fatally. */ + if ((ret = __txn_isvalid(txnp, &td, TXN_OP_ABORT)) != 0) + return (__db_panic(dbenv, ret)); + + /* + * Try to abort any unresolved children. + * + * Abort either succeeds or panics the region. As soon as we + * see any failure, we just get out of here and return the panic + * up. + */ + while ((kid = TAILQ_FIRST(&txnp->kids)) != NULL) + if ((ret = kid->abort(kid)) != 0) + return (ret); + + if (LOCKING_ON(dbenv)) { + /* + * We are about to free all the read locks for this transaction + * below. Some of those locks might be handle locks which + * should not be freed, because they will be freed when the + * handle is closed. Check the events and preprocess any + * trades now so that we don't release the locks below. + */ + if ((ret = __txn_doevents(dbenv, txnp, 0, 1)) != 0) + return (__db_panic(dbenv, ret)); + + /* Turn off timeouts. */ + if ((ret = __lock_set_timeout(dbenv, + txnp->txnid, 0, DB_SET_TXN_TIMEOUT)) != 0) + return (__db_panic(dbenv, ret)); + + if ((ret = __lock_set_timeout(dbenv, + txnp->txnid, 0, DB_SET_LOCK_TIMEOUT)) != 0) + return (__db_panic(dbenv, ret)); + + request.op = DB_LOCK_UPGRADE_WRITE; + if ((ret = dbenv->lock_vec( + dbenv, txnp->txnid, 0, &request, 1, NULL)) != 0) + return (__db_panic(dbenv, ret)); + } + if ((ret = __txn_undo(txnp)) != 0) + return (__db_panic(dbenv, ret)); + + /* + * Normally, we do not need to log aborts. However, if we + * are a distributed transaction (i.e., we have a prepare), + * then we log the abort so we know that this transaction + * was actually completed. + */ + SET_LOG_FLAGS(dbenv, txnp, lflags); + if (DBENV_LOGGING(dbenv) && td->status == TXN_PREPARED && + (ret = __txn_regop_log(dbenv, txnp, &txnp->last_lsn, + lflags, TXN_ABORT, (int32_t)time(NULL))) != 0) + return (__db_panic(dbenv, ret)); + + /* __txn_end always panics if it errors, so pass the return along. */ + return (__txn_end(txnp, 0)); +} + +/* + * __txn_discard -- + * Free the per-process resources associated with this txn handle. + * + * PUBLIC: int __txn_discard __P((DB_TXN *, u_int32_t flags)); + */ +int +__txn_discard(txnp, flags) + DB_TXN *txnp; + u_int32_t flags; +{ + DB_ENV *dbenv; + DB_TXN *freep; + TXN_DETAIL *td; + int ret; + + COMPQUIET(flags, 0); + + dbenv = txnp->mgrp->dbenv; + freep = NULL; PANIC_CHECK(dbenv); - if ((ret = __txn_isvalid(txnp, NULL, TXN_ABORTED)) != 0) + if ((ret = __txn_isvalid(txnp, &td, TXN_OP_DISCARD)) != 0) return (ret); - /* Abort any unresolved children. */ - while ((kid = TAILQ_FIRST(&txnp->kids)) != NULL) - if ((t_ret = txn_abort(kid)) != 0 && ret == 0) - ret = t_ret; + /* Should be no children. */ + DB_ASSERT(TAILQ_FIRST(&txnp->kids) == NULL); + DB_ASSERT(F_ISSET(td, TXN_RESTORED)); - if ((t_ret = __txn_undo(txnp)) != 0 && ret == 0) - ret = t_ret; + /* Free the space. */ + MUTEX_THREAD_LOCK(dbenv, txnp->mgrp->mutexp); + txnp->mgrp->n_discards++; + if (F_ISSET(txnp, TXN_MALLOC)) { + TAILQ_REMOVE(&txnp->mgrp->txn_chain, txnp, links); + freep = txnp; + } + MUTEX_THREAD_UNLOCK(dbenv, txnp->mgrp->mutexp); + if (freep != NULL) + __os_free(dbenv, freep); - if ((t_ret = __txn_end(txnp, 0)) != 0 && ret == 0) - ret = t_ret; - return (ret); + return (0); } /* - * txn_prepare -- + * __txn_prepare -- * Flush the log so a future commit is guaranteed to succeed. + * + * PUBLIC: int __txn_prepare __P((DB_TXN *, u_int8_t *)); */ int -txn_prepare(txnp) +__txn_prepare(txnp, gid) DB_TXN *txnp; + u_int8_t *gid; { DBT xid; DB_ENV *dbenv; DB_TXN *kid; TXN_DETAIL *td; + u_int32_t lflags; int ret; dbenv = txnp->mgrp->dbenv; -#ifdef HAVE_RPC - if (F_ISSET(dbenv, DB_ENV_RPCCLIENT)) - return (__dbcl_txn_prepare(txnp)); -#endif - PANIC_CHECK(dbenv); - if ((ret = __txn_isvalid(txnp, &td, TXN_PREPARED)) != 0) + if ((ret = __txn_isvalid(txnp, &td, TXN_OP_PREPARE)) != 0) return (ret); - /* Prepare any unresolved children. */ + /* Commit any unresolved children. */ while ((kid = TAILQ_FIRST(&txnp->kids)) != NULL) - if ((ret = txn_prepare(kid)) != 0) + if ((ret = kid->commit(kid, DB_TXN_NOSYNC)) != 0) return (ret); /* - * We indicate that a transaction is an XA transaction by putting - * a valid size in the xid.size fiels. XA requires that the transaction - * be either ENDED or SUSPENDED when prepare is called, so we know - * that if the xa_status isn't in one of those states, but we are - * calling prepare that we are not an XA transaction. + * In XA, the global transaction ID in the txn_detail structure is + * already set; in a non-XA environment, we must set it here. XA + * requires that the transaction be either ENDED or SUSPENDED when + * prepare is called, so we know that if the xa_status isn't in one + * of those states, then we are calling prepare directly and we need + * to fill in the td->xid. */ - - if (LOGGING_ON(dbenv)) { + if (DBENV_LOGGING(dbenv)) { memset(&xid, 0, sizeof(xid)); + if (td->xa_status != TXN_XA_ENDED && + td->xa_status != TXN_XA_SUSPENDED) + /* Regular prepare; fill in the gid. */ + memcpy(td->xid, gid, sizeof(td->xid)); + + xid.size = sizeof(td->xid); xid.data = td->xid; - xid.size = td->xa_status != TXN_XA_ENDED && - td->xa_status != TXN_XA_SUSPENDED ? 0 : sizeof(td->xid); + + SET_LOG_FLAGS(dbenv, txnp, lflags); if ((ret = __txn_xa_regop_log(dbenv, txnp, &txnp->last_lsn, - (F_ISSET(dbenv, DB_ENV_TXN_NOSYNC) && - !F_ISSET(txnp, TXN_SYNC)) || - F_ISSET(txnp, TXN_NOSYNC) ? 0 : DB_FLUSH, TXN_PREPARE, - &xid, td->format, td->gtrid, td->bqual, + lflags, TXN_PREPARE, &xid, td->format, td->gtrid, td->bqual, &td->begin_lsn)) != 0) { - __db_err(dbenv, "txn_prepare: log_write failed %s", + __db_err(dbenv, "DB_TXN->prepare: log_write failed %s", db_strerror(ret)); return (ret); } - if (txnp->parent != NULL) - F_SET(txnp->parent, TXN_CHILDCOMMIT); } MUTEX_THREAD_LOCK(dbenv, txnp->mgrp->mutexp); @@ -461,17 +763,36 @@ txn_prepare(txnp) } /* - * txn_id -- + * __txn_id -- * Return the transaction ID. + * + * PUBLIC: u_int32_t __txn_id __P((DB_TXN *)); */ u_int32_t -txn_id(txnp) +__txn_id(txnp) DB_TXN *txnp; { return (txnp->txnid); } /* + * __txn_set_timeout -- + * Set timeout values in the txn structure. + */ +static int +__txn_set_timeout(txnp, timeout, op) + DB_TXN *txnp; + db_timeout_t timeout; + u_int32_t op; +{ + if (op != DB_SET_TXN_TIMEOUT && op != DB_SET_LOCK_TIMEOUT) + return (__db_ferr(txnp->mgrp->dbenv, "DB_TXN->set_timeout", 0)); + + return (__lock_set_timeout( + txnp->mgrp->dbenv, txnp->txnid, timeout, op)); +} + +/* * __txn_isvalid -- * Return 0 if the txnp is reasonable, otherwise panic. */ @@ -479,12 +800,22 @@ static int __txn_isvalid(txnp, tdp, op) const DB_TXN *txnp; TXN_DETAIL **tdp; - u_int32_t op; + txnop_t op; { DB_TXNMGR *mgrp; + DB_TXNREGION *region; TXN_DETAIL *tp; mgrp = txnp->mgrp; + region = mgrp->reginfo.primary; + + /* Check for recovery. */ + if (!F_ISSET(txnp, TXN_COMPENSATE) && + F_ISSET(region, TXN_IN_RECOVERY)) { + __db_err(mgrp->dbenv, + "operation not permitted during recovery"); + goto err; + } /* Check for live cursors. */ if (txnp->cursors != 0) { @@ -492,25 +823,70 @@ __txn_isvalid(txnp, tdp, op) goto err; } - /* Check transaction's status. */ + /* Check transaction's state. */ tp = (TXN_DETAIL *)R_ADDR(&mgrp->reginfo, txnp->off); if (tdp != NULL) *tdp = tp; + /* Handle any operation specific checks. */ + switch (op) { + case TXN_OP_DISCARD: + /* + * Since we're just tossing the per-process space; there are + * a lot of problems with the transaction that we can tolerate. + */ + + /* Transaction is already been reused. */ + if (txnp->txnid != tp->txnid) + return (0); + + /* What we've got had better be a restored transaction. */ + if (!F_ISSET(tp, TXN_RESTORED)) { + __db_err(mgrp->dbenv, "not a restored transaction"); + return (__db_panic(mgrp->dbenv, EINVAL)); + } + + return (0); + case TXN_OP_PREPARE: + if (txnp->parent != NULL) { + /* + * This is not fatal, because you could imagine an + * application that simply prepares everybody because + * it doesn't distinguish between children and parents. + * I'm not arguing this is good, but I could imagine + * someone doing it. + */ + __db_err(mgrp->dbenv, + "Prepare disallowed on child transactions"); + return (EINVAL); + } + break; + case TXN_OP_ABORT: + case TXN_OP_COMMIT: + default: + break; + } + switch (tp->status) { + case TXN_PREPARED: + if (op == TXN_OP_PREPARE) { + __db_err(mgrp->dbenv, "transaction already prepared"); + /* + * Txn_prepare doesn't blow away the user handle, so + * in this case, give the user the opportunity to + * abort or commit. + */ + return (EINVAL); + } + break; + case TXN_RUNNING: + break; case TXN_ABORTED: case TXN_COMMITTED: default: __db_err(mgrp->dbenv, "transaction already %s", tp->status == TXN_COMMITTED ? "committed" : "aborted"); goto err; - case TXN_PREPARED: - if (op == TXN_PREPARED) { - __db_err(mgrp->dbenv, "transaction already prepared"); - goto err; - } - case TXN_RUNNING: - break; } return (0); @@ -526,10 +902,8 @@ err: /* /* * __txn_end -- * Internal transaction end routine. - * - * PUBLIC: int __txn_end __P((DB_TXN *, int)); */ -int +static int __txn_end(txnp, is_commit) DB_TXN *txnp; int is_commit; @@ -539,46 +913,59 @@ __txn_end(txnp, is_commit) DB_TXNMGR *mgr; DB_TXNREGION *region; TXN_DETAIL *tp; - int ret; + int do_closefiles, ret; mgr = txnp->mgrp; dbenv = mgr->dbenv; region = mgr->reginfo.primary; + do_closefiles = 0; + + /* Process commit events. */ + if ((ret = __txn_doevents(dbenv, txnp, is_commit, 0)) != 0) + return (__db_panic(dbenv, ret)); /* Release the locks. */ request.op = txnp->parent == NULL || is_commit == 0 ? DB_LOCK_PUT_ALL : DB_LOCK_INHERIT; - if (LOCKING_ON(dbenv)) { - ret = lock_vec(dbenv, txnp->txnid, 0, &request, 1, NULL); - if (ret != 0 && (ret != DB_LOCK_DEADLOCK || is_commit)) { - __db_err(dbenv, "%s: release locks failed %s", - is_commit ? "txn_commit" : "txn_abort", - db_strerror(ret)); - __db_panic(dbenv, ret); - } - } + /* + * __txn_end cannot return an simple error, we MUST return + * success/failure from commit or abort, ignoring any internal + * errors. So, we panic if something goes wrong. We can't + * deadlock here because we're not acquiring any new locks, + * so DB_LOCK_DEADLOCK is just as fatal as any other error. + */ + if (LOCKING_ON(dbenv) && (ret = dbenv->lock_vec( + dbenv, txnp->txnid, DB_LOCK_FREE_LOCKER, &request, 1, NULL)) != 0) + return (__db_panic(dbenv, ret)); /* End the transaction. */ R_LOCK(dbenv, &mgr->reginfo); tp = (TXN_DETAIL *)R_ADDR(&mgr->reginfo, txnp->off); SH_TAILQ_REMOVE(®ion->active_txn, tp, links, __txn_detail); + if (F_ISSET(tp, TXN_RESTORED)) { + region->stat.st_nrestores--; + do_closefiles = region->stat.st_nrestores == 0; + } + __db_shalloc_free(mgr->reginfo.addr, tp); if (is_commit) - region->ncommits++; + region->stat.st_ncommits++; else - region->naborts++; - --region->nactive; + region->stat.st_naborts++; + --region->stat.st_nactive; R_UNLOCK(dbenv, &mgr->reginfo); /* - * The transaction cannot get more locks, remove its locker info. + * The transaction cannot get more locks, remove its locker info, + * if any. */ - if (LOCKING_ON(dbenv)) - __lock_freefamilylocker(dbenv->lk_handle, txnp->txnid); + if (LOCKING_ON(dbenv) && (ret = + __lock_freefamilylocker(dbenv->lk_handle, txnp->txnid)) != 0) + return (__db_panic(dbenv, ret)); if (txnp->parent != NULL) TAILQ_REMOVE(&txnp->parent->kids, txnp, klinks); @@ -588,9 +975,16 @@ __txn_end(txnp, is_commit) TAILQ_REMOVE(&mgr->txn_chain, txnp, links); MUTEX_THREAD_UNLOCK(dbenv, mgr->mutexp); - __os_free(txnp, sizeof(*txnp)); + __os_free(dbenv, txnp); } + if (do_closefiles) { + F_SET((DB_LOG *)dbenv->lg_handle, DBLOG_RECOVER); + (void)__dbreg_close_files(dbenv); + F_CLR((DB_LOG *)dbenv->lg_handle, DBLOG_RECOVER); + mgr->n_discards = 0; + (void)dbenv->txn_checkpoint(dbenv, 0, 0, DB_FORCE); + } return (0); } @@ -605,16 +999,19 @@ __txn_undo(txnp) { DBT rdbt; DB_ENV *dbenv; + DB_LOGC *logc; DB_LSN key_lsn; + DB_TXN *ptxn; DB_TXNMGR *mgr; + int ret, t_ret; void *txnlist; - int ret, threaded; mgr = txnp->mgrp; dbenv = mgr->dbenv; + logc = NULL; txnlist = NULL; - if (!LOGGING_ON(dbenv)) + if (!DBENV_LOGGING(dbenv)) return (0); /* @@ -623,53 +1020,66 @@ __txn_undo(txnp) * allocation here and use DB_DBT_USERMEM. */ memset(&rdbt, 0, sizeof(rdbt)); - threaded = F_ISSET(dbenv, DB_ENV_THREAD) ? 1 : 0; - if (threaded) - F_SET(&rdbt, DB_DBT_MALLOC); key_lsn = txnp->last_lsn; - /* Allocate a transaction list for children or aborted page creates. */ - if ((ret = __db_txnlist_init(dbenv, &txnlist)) != 0) + /* + * Allocate a txnlist for children and aborted page allocs. + * We need to associate the list with the maximal parent + * so that aborted pages are recovered when that transaction + * is commited or aborted. + */ + for (ptxn = txnp->parent; ptxn != NULL && ptxn->parent != NULL;) + ptxn = ptxn->parent; + + if (ptxn != NULL && ptxn->txn_list != NULL) + txnlist = ptxn->txn_list; + else if (txnp->txn_list != NULL) + txnlist = txnp->txn_list; + else if ((ret = __db_txnlist_init(dbenv, 0, 0, NULL, &txnlist)) != 0) return (ret); + else if (ptxn != NULL) + ptxn->txn_list = txnlist; if (F_ISSET(txnp, TXN_CHILDCOMMIT) && - (ret = __db_txnlist_lsninit(dbenv, - txnlist, &txnp->last_lsn)) != 0) + (ret = __db_txnlist_lsninit(dbenv, txnlist, &txnp->last_lsn)) != 0) return (ret); - for (ret = 0; ret == 0 && !IS_ZERO_LSN(key_lsn);) { + if ((ret = dbenv->log_cursor(dbenv, &logc, 0)) != 0) + goto err; + while (ret == 0 && !IS_ZERO_LSN(key_lsn)) { /* * The dispatch routine returns the lsn of the record * before the current one in the key_lsn argument. */ - if ((ret = log_get(dbenv, &key_lsn, &rdbt, DB_SET)) == 0) { - ret = __db_dispatch(dbenv, - &rdbt, &key_lsn, DB_TXN_ABORT, txnlist); - if (threaded && rdbt.data != NULL) { - __os_free(rdbt.data, rdbt.size); - rdbt.data = NULL; - } + if ((ret = logc->get(logc, &key_lsn, &rdbt, DB_SET)) == 0) { + ret = __db_dispatch(dbenv, dbenv->recover_dtab, + dbenv->recover_dtab_size, &rdbt, &key_lsn, + DB_TXN_ABORT, txnlist); if (F_ISSET(txnp, TXN_CHILDCOMMIT)) (void)__db_txnlist_lsnadd(dbenv, txnlist, &key_lsn, 0); } - if (ret != 0) { + if (ret == DB_SURPRISE_KID) { + if ((ret = __db_txnlist_lsninit( + dbenv, txnlist, &key_lsn)) == 0) + F_SET(txnp, TXN_CHILDCOMMIT); + } else if (ret != 0) { __db_err(txnp->mgrp->dbenv, - "txn_abort: Log undo failed for LSN: %lu %lu: %s", + "DB_TXN->abort: Log undo failed for LSN: %lu %lu: %s", (u_long)key_lsn.file, (u_long)key_lsn.offset, db_strerror(ret)); - if (txnlist != NULL) - __db_txnlist_end(dbenv, txnlist); - return (ret); + goto err; } } - if (txnlist != NULL) { - __db_do_the_limbo(dbenv, txnlist); - __db_txnlist_end(dbenv, txnlist); - } + ret = __db_do_the_limbo(dbenv, ptxn, txnp, txnlist); +err: if (logc != NULL && (t_ret = logc->close(logc, 0)) != 0 && ret == 0) + ret = t_ret; + + if (ptxn == NULL && txnlist != NULL) + __db_txnlist_end(dbenv, txnlist); return (ret); } @@ -682,150 +1092,115 @@ __txn_undo(txnp) * its first LSN. This is the lowest LSN we can checkpoint, since any record * written after since that point may be involved in a transaction and may * therefore need to be undone in the case of an abort. + * + * PUBLIC: int __txn_checkpoint + * PUBLIC: __P((DB_ENV *, u_int32_t, u_int32_t, u_int32_t)); */ int -txn_checkpoint(dbenv, kbytes, minutes, flags) +__txn_checkpoint(dbenv, kbytes, minutes, flags) DB_ENV *dbenv; u_int32_t kbytes, minutes, flags; { - DB_LOG *dblp; - DB_LSN ckp_lsn, sync_lsn, last_ckp; + DB_LSN ckp_lsn, last_ckp; DB_TXNMGR *mgr; DB_TXNREGION *region; - LOG *lp; TXN_DETAIL *txnp; time_t last_ckp_time, now; u_int32_t bytes, mbytes; - int interval, ret; + int ret; -#ifdef HAVE_RPC - if (F_ISSET(dbenv, DB_ENV_RPCCLIENT)) - return (__dbcl_txn_checkpoint(dbenv, kbytes, minutes)); -#endif PANIC_CHECK(dbenv); - ENV_REQUIRES_CONFIG(dbenv, dbenv->tx_handle, DB_INIT_TXN); - - mgr = dbenv->tx_handle; - region = mgr->reginfo.primary; - dblp = dbenv->lg_handle; - lp = dblp->reginfo.primary; + ENV_REQUIRES_CONFIG(dbenv, + dbenv->tx_handle, "txn_checkpoint", DB_INIT_TXN); /* - * Check if we need to checkpoint. + * On a replication client, all transactions are read-only; therefore, + * a checkpoint is a null-op. + * + * We permit txn_checkpoint, instead of just rendering it illegal, + * so that an application can just let a checkpoint thread continue + * to operate as it gets promoted or demoted between being a + * master and a client. */ - ZERO_LSN(ckp_lsn); + if (F_ISSET(dbenv, DB_ENV_REP_CLIENT)) + return (0); - if (LF_ISSET(DB_FORCE)) - goto do_ckp; + mgr = dbenv->tx_handle; + region = mgr->reginfo.primary; - R_LOCK(dbenv, &dblp->reginfo); - mbytes = lp->stat.st_wc_mbytes; /* - * We add the current buffer offset so as to count bytes that - * have not yet been written, but are sitting in the log buffer. + * The checkpoint LSN is an LSN such that all transactions begun before + * it are complete. Our first guess (corrected below based on the list + * of active transactions) is the last-written LSN. */ - bytes = lp->stat.st_wc_bytes + lp->b_off; - ckp_lsn = lp->lsn; - R_UNLOCK(dbenv, &dblp->reginfo); - - /* Don't checkpoint a quiescent database. */ - if (bytes == 0 && mbytes == 0) - return (0); + __log_txn_lsn(dbenv, &ckp_lsn, &mbytes, &bytes); - if (kbytes != 0 && mbytes * 1024 + bytes / 1024 >= (u_int32_t)kbytes) - goto do_ckp; + if (!LF_ISSET(DB_FORCE)) { + /* Don't checkpoint a quiescent database. */ + if (bytes == 0 && mbytes == 0) + return (0); - if (minutes != 0) { - (void)time(&now); + if (kbytes != 0 && + mbytes * 1024 + bytes / 1024 >= (u_int32_t)kbytes) + goto do_ckp; - R_LOCK(dbenv, &mgr->reginfo); - last_ckp_time = region->time_ckp; - R_UNLOCK(dbenv, &mgr->reginfo); + if (minutes != 0) { + (void)time(&now); - if (now - last_ckp_time >= (time_t)(minutes * 60)) - goto do_ckp; - } + R_LOCK(dbenv, &mgr->reginfo); + last_ckp_time = region->time_ckp; + R_UNLOCK(dbenv, &mgr->reginfo); - /* - * If we checked time and data and didn't go to checkpoint, - * we're done. - */ - if (minutes != 0 || kbytes != 0) - return (0); + if (now - last_ckp_time >= (time_t)(minutes * 60)) + goto do_ckp; + } -do_ckp: - if (IS_ZERO_LSN(ckp_lsn)) { - R_LOCK(dbenv, &dblp->reginfo); - ckp_lsn = lp->lsn; - R_UNLOCK(dbenv, &dblp->reginfo); + /* + * If we checked time and data and didn't go to checkpoint, + * we're done. + */ + if (minutes != 0 || kbytes != 0) + return (0); } - /* - * We have to find an LSN such that all transactions begun - * before that LSN are complete. - */ +do_ckp: /* Look through the active transactions for the lowest begin LSN. */ R_LOCK(dbenv, &mgr->reginfo); - - if (IS_ZERO_LSN(region->pending_ckp)) { - for (txnp = - SH_TAILQ_FIRST(®ion->active_txn, __txn_detail); - txnp != NULL; - txnp = SH_TAILQ_NEXT(txnp, links, __txn_detail)) { - - /* - * Look through the active transactions for the - * lowest begin lsn. - */ - if (!IS_ZERO_LSN(txnp->begin_lsn) && - log_compare(&txnp->begin_lsn, &ckp_lsn) < 0) - ckp_lsn = txnp->begin_lsn; - } - region->pending_ckp = ckp_lsn; - } else - ckp_lsn = region->pending_ckp; - + for (txnp = SH_TAILQ_FIRST(®ion->active_txn, __txn_detail); + txnp != NULL; + txnp = SH_TAILQ_NEXT(txnp, links, __txn_detail)) + if (!IS_ZERO_LSN(txnp->begin_lsn) && + log_compare(&txnp->begin_lsn, &ckp_lsn) < 0) + ckp_lsn = txnp->begin_lsn; R_UNLOCK(dbenv, &mgr->reginfo); + if (MPOOL_ON(dbenv) && (ret = dbenv->memp_sync(dbenv, NULL)) != 0) { + __db_err(dbenv, + "txn_checkpoint: failed to flush the buffer cache %s", + db_strerror(ret)); + return (ret); + } + /* - * Try three times to sync the mpool buffers up to the specified LSN, - * sleeping 1, 2 and 4 seconds between attempts. + * Because we can't be a replication client here, and because + * recovery (somewhat unusually) calls txn_checkpoint and expects + * it to write a log message, LOGGING_ON is the correct macro here. */ - if (MPOOL_ON(dbenv)) - for (interval = 1;;) { - /* - * memp_sync may change the lsn you pass it, so don't - * pass it the actual ckp_lsn, pass it a local instead. - */ - sync_lsn = ckp_lsn; - if ((ret = memp_sync(dbenv, &sync_lsn)) == 0) - break; - - /* - * ret == DB_INCOMPLETE means there are still buffers - * to flush, the checkpoint is not complete. - */ - if (ret == DB_INCOMPLETE) { - if (interval > 4) - return (ret); - - (void)__os_sleep(dbenv, interval, 0); - interval *= 2; - } else { - __db_err(dbenv, - "txn_checkpoint: failure in memp_sync %s", - db_strerror(ret)); - return (ret); - } - } - if (LOGGING_ON(dbenv)) { R_LOCK(dbenv, &mgr->reginfo); last_ckp = region->last_ckp; - ZERO_LSN(region->pending_ckp); R_UNLOCK(dbenv, &mgr->reginfo); - if ((ret = __txn_ckp_log(dbenv, - NULL, &ckp_lsn, DB_CHECKPOINT, &ckp_lsn, + /* + * Put out records for the open files before we log + * the checkpoint. The records are certain to be at + * or after ckp_lsn, but before the checkpoint record + * itself, so they're sure to be included if we start + * recovery from the ckp_lsn contained in this + * checkpoint. + */ + if ((ret = __dbreg_open_files(dbenv)) != 0 || + (ret = __txn_ckp_log(dbenv, + NULL, &ckp_lsn, DB_PERMANENT | DB_FLUSH, &ckp_lsn, &last_ckp, (int32_t)time(NULL))) != 0) { __db_err(dbenv, "txn_checkpoint: log failed at LSN [%ld %ld] %s", @@ -834,15 +1209,41 @@ do_ckp: return (ret); } - R_LOCK(dbenv, &mgr->reginfo); - region->last_ckp = ckp_lsn; - (void)time(®ion->time_ckp); - R_UNLOCK(dbenv, &mgr->reginfo); + __txn_updateckp(dbenv, &ckp_lsn); } return (0); } /* + * __txn_getckp -- + * Get the LSN of the last transaction checkpoint. + * + * PUBLIC: int __txn_getckp __P((DB_ENV *, DB_LSN *)); + */ +int +__txn_getckp(dbenv, lsnp) + DB_ENV *dbenv; + DB_LSN *lsnp; +{ + DB_LSN lsn; + DB_TXNMGR *mgr; + DB_TXNREGION *region; + + mgr = dbenv->tx_handle; + region = mgr->reginfo.primary; + + R_LOCK(dbenv, &mgr->reginfo); + lsn = region->last_ckp; + R_UNLOCK(dbenv, &mgr->reginfo); + + if (IS_ZERO_LSN(lsn)) + return (DB_NOTFOUND); + + *lsnp = lsn; + return (0); +} + +/* * __txn_activekids -- * Return if this transaction has any active children. * @@ -858,7 +1259,7 @@ __txn_activekids(dbenv, rectype, txnp) * On a child commit, we know that there are children (i.e., the * commiting child at the least. In that case, skip this check. */ - if (rectype == DB_txn_child) + if (F_ISSET(txnp, TXN_COMPENSATE) || rectype == DB___txn_child) return (0); if (TAILQ_FIRST(&txnp->kids) != NULL) { @@ -867,3 +1268,161 @@ __txn_activekids(dbenv, rectype, txnp) } return (0); } + +/* + * __txn_force_abort -- + * Force an abort record into the log if the commit record + * failed to get to disk. + * + * PUBLIC: int __txn_force_abort __P((DB_ENV *, u_int8_t *)); + */ +int +__txn_force_abort(dbenv, buffer) + DB_ENV *dbenv; + u_int8_t *buffer; +{ + DB_CIPHER *db_cipher; + HDR *hdr; + u_int32_t offset, opcode, rec_len, rec_type, sum_len; + u_int8_t *bp, *key, chksum[DB_MAC_KEY]; + size_t hdrsize; + int ret; + + db_cipher = dbenv->crypto_handle; + + /* + * This routine depends on the layout of HDR and the __txn_regop + * record in txn.src. We are passed the beginning of the commit + * record in the log buffer and overwrite the commit with an abort + * and recalculate the checksum. We may be passed a txn_xa_regop + * that is, an XA prepare), there's no need to overwrite that one. + */ + hdr = (HDR *)buffer; + memcpy(&rec_type, hdr, sizeof(rec_type)); + if (rec_type == DB___txn_xa_regop) + return (0); + + offset = sizeof(u_int32_t) + sizeof(u_int32_t) + sizeof(DB_LSN); + rec_len = offset + sizeof(u_int32_t) + sizeof(int32_t); + if (CRYPTO_ON(dbenv)) { + key = db_cipher->mac_key; + hdrsize = HDR_CRYPTO_SZ; + sum_len = DB_MAC_KEY; + if ((ret = db_cipher->decrypt(dbenv, db_cipher->data, + &hdr->iv[0], buffer + hdrsize, rec_len)) != 0) + return (__db_panic(dbenv, ret)); + } else { + key = NULL; + hdrsize = HDR_NORMAL_SZ; + sum_len = sizeof(u_int32_t); + } + bp = buffer + hdrsize + offset; + opcode = TXN_ABORT; + memcpy(bp, &opcode, sizeof(opcode)); + + if (CRYPTO_ON(dbenv) && + (ret = db_cipher->encrypt(dbenv, + db_cipher->data, &hdr->iv[0], buffer + hdrsize, rec_len)) != 0) + return (__db_panic(dbenv, ret)); + + __db_chksum(buffer + hdrsize, rec_len, key, chksum); + memcpy(buffer + SSZ(HDR, chksum), &chksum, sum_len); + + return (0); +} + +/* + * __txn_preclose + * Before we can close an environment, we need to check if we + * were in the midst of taking care of restored transactions. If + * so, then we need to close the files that we opened. + * + * PUBLIC: int __txn_preclose __P((DB_ENV *)); + */ +int +__txn_preclose(dbenv) + DB_ENV *dbenv; +{ + DB_TXNMGR *mgr; + DB_TXNREGION *region; + int do_closefiles, ret; + + mgr = (DB_TXNMGR *)dbenv->tx_handle; + region = mgr->reginfo.primary; + do_closefiles = 0; + + R_LOCK(dbenv, &mgr->reginfo); + if (region != NULL && + region->stat.st_nrestores + <= mgr->n_discards && mgr->n_discards != 0) + do_closefiles = 1; + R_UNLOCK(dbenv, &mgr->reginfo); + + if (do_closefiles) { + /* + * Set the DBLOG_RECOVER flag while closing these + * files so they do not create additional log records + * that will confuse future recoveries. + */ + F_SET((DB_LOG *)dbenv->lg_handle, DBLOG_RECOVER); + ret = __dbreg_close_files(dbenv); + F_CLR((DB_LOG *)dbenv->lg_handle, DBLOG_RECOVER); + } else + ret = 0; + + return (ret); +} + +/* + * __txn_reset -- + * Reset the last txnid to its minimum value, and log the reset. + * + * PUBLIC: int __txn_reset __P((DB_ENV *)); + */ +int +__txn_reset(dbenv) + DB_ENV *dbenv; +{ + DB_LSN scrap; + DB_TXNREGION *region; + + region = ((DB_TXNMGR *)dbenv->tx_handle)->reginfo.primary; + region->last_txnid = TXN_MINIMUM; + + DB_ASSERT(LOGGING_ON(dbenv)); + return (__txn_recycle_log(dbenv, + NULL, &scrap, 0, TXN_MINIMUM, TXN_MAXIMUM)); +} + +/* + * __txn_updateckp -- + * Update the last_ckp field in the transaction region. This happens + * at the end of a normal checkpoint and also when a replication client + * receives a checkpoint record. + * + * PUBLIC: void __txn_updateckp __P((DB_ENV *, DB_LSN *)); + */ +void +__txn_updateckp(dbenv, lsnp) + DB_ENV *dbenv; + DB_LSN *lsnp; +{ + DB_TXNMGR *mgr; + DB_TXNREGION *region; + + mgr = dbenv->tx_handle; + region = mgr->reginfo.primary; + + /* + * We want to make sure last_ckp only moves forward; since + * we drop locks above and in log_put, it's possible + * for two calls to __txn_ckp_log to finish in a different + * order from how they were called. + */ + R_LOCK(dbenv, &mgr->reginfo); + if (log_compare(®ion->last_ckp, lsnp) < 0) { + region->last_ckp = *lsnp; + (void)time(®ion->time_ckp); + } + R_UNLOCK(dbenv, &mgr->reginfo); +} -- 2.7.4