diff options
Diffstat (limited to 'storage/bdb/txn/txn.c')
-rw-r--r-- | storage/bdb/txn/txn.c | 1428 |
1 files changed, 1428 insertions, 0 deletions
diff --git a/storage/bdb/txn/txn.c b/storage/bdb/txn/txn.c new file mode 100644 index 00000000000..78c54791d06 --- /dev/null +++ b/storage/bdb/txn/txn.c @@ -0,0 +1,1428 @@ +/*- + * See the file LICENSE for redistribution information. + * + * Copyright (c) 1996-2002 + * Sleepycat Software. All rights reserved. + */ +/* + * Copyright (c) 1995, 1996 + * The President and Fellows of Harvard University. All rights reserved. + * + * This code is derived from software contributed to Berkeley by + * Margo Seltzer. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * 3. Neither the name of the University nor the names of its contributors + * may be used to endorse or promote products derived from this software + * without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + +#include "db_config.h" + +#ifndef lint +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 <sys/types.h> +#include <stdlib.h> + +#if TIME_WITH_SYS_TIME +#include <sys/time.h> +#include <time.h> +#else +#if HAVE_SYS_TIME_H +#include <sys/time.h> +#else +#include <time.h> +#endif +#endif + +#include <string.h> +#endif + +#include "db_int.h" +#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) + +/* + * __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 -- + * 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) + DB_ENV *dbenv; + DB_TXN *parent, **txnpp; + u_int32_t flags; +{ + DB_LOCKREGION *region; + DB_TXN *txn; + int ret; + + *txnpp = NULL; + PANIC_CHECK(dbenv); + ENV_REQUIRES_CONFIG(dbenv, dbenv->tx_handle, "txn_begin", DB_INIT_TXN); + + if ((ret = __db_fchk(dbenv, + "txn_begin", flags, + 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) + return (ret); + + if ((ret = __os_calloc(dbenv, 1, sizeof(DB_TXN), &txn)) != 0) + return (ret); + + 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)) + F_SET(txn, TXN_SYNC); + if (LF_ISSET(DB_TXN_NOWAIT)) + F_SET(txn, TXN_NOWAIT); + + if ((ret = __txn_begin_int(txn, 0)) != 0) + goto err; + + 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); +} + +/* + * __txn_xa_begin -- + * XA version of txn_begin. + * + * PUBLIC: int __txn_xa_begin __P((DB_ENV *, DB_TXN *)); + */ +int +__txn_xa_begin(dbenv, txn) + DB_ENV *dbenv; + DB_TXN *txn; +{ + PANIC_CHECK(dbenv); + + memset(txn, 0, sizeof(DB_TXN)); + + txn->mgrp = dbenv->tx_handle; + TAILQ_INIT(&txn->kids); + TAILQ_INIT(&txn->events); + + return (__txn_begin_int(txn, 0)); +} + +/* + * __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_int(txn, internal) + DB_TXN *txn; + int internal; +{ + DB_ENV *dbenv; + DB_LSN begin_lsn, null_lsn; + DB_TXNMGR *mgr; + DB_TXNREGION *region; + TXN_DETAIL *td; + size_t off; + u_int32_t id, *ids; + int nids, ret; + + mgr = txn->mgrp; + dbenv = mgr->dbenv; + region = mgr->reginfo.primary; + + /* + * We do not have to write begin records (and if we do not, then we + * 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 (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 we aren't still recovering prepared transactions. */ + if (!internal && region->stat.st_nrestores != 0) { + __db_err(dbenv, + "recovery of prepared but not yet committed transactions is incomplete"); + ret = EINVAL; + 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 err; + } + + /* Place transaction on active transaction list. */ + SH_TAILQ_INSERT_HEAD(®ion->active_txn, td, links, __txn_detail); + + id = ++region->last_txnid; + ++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; + ZERO_LSN(td->last_lsn); + td->status = TXN_RUNNING; + if (txn->parent != NULL) + td->parent = txn->parent->off; + 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 = (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 + * maximal grandparent in the lock table for deadlock detection. + */ + if (txn->parent != NULL && LOCKING_ON(dbenv)) + if ((ret = __lock_addfamilylocker(dbenv, + txn->parent->txnid, txn->txnid)) != 0) + return (ret); + + if (F_ISSET(txn, TXN_MALLOC)) { + MUTEX_THREAD_LOCK(dbenv, mgr->mutexp); + TAILQ_INSERT_TAIL(&mgr->txn_chain, txn, links); + MUTEX_THREAD_UNLOCK(dbenv, mgr->mutexp); + } + + return (0); + +err: R_UNLOCK(dbenv, &mgr->reginfo); + return (ret); +} + +/* + * __txn_commit -- + * Commit a transaction. + * + * PUBLIC: int __txn_commit __P((DB_TXN *, u_int32_t)); + */ +int +__txn_commit(txnp, flags) + DB_TXN *txnp; + u_int32_t flags; +{ + DB_ENV *dbenv; + DB_LOCKREQ request; + DB_TXN *kid; + TXN_DETAIL *td; + u_int32_t lflags; + int ret, t_ret; + + dbenv = txnp->mgrp->dbenv; + + PANIC_CHECK(dbenv); + + if ((ret = __txn_isvalid(txnp, &td, TXN_OP_COMMIT)) != 0) + return (ret); + + /* + * We clear flags that are incorrect, ignoring any flag errors, and + * default to synchronous operations. By definition, transaction + * handles are dead when we return, and this error should never + * happen, but we don't want to fail in the field 'cause the app is + * specifying the wrong flag for some reason. + */ + if (__db_fchk(dbenv, + "DB_TXN->commit", flags, DB_TXN_NOSYNC | DB_TXN_SYNC) != 0) + flags = DB_TXN_SYNC; + if (__db_fcchk(dbenv, + "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); + F_SET(txnp, TXN_NOSYNC); + } + if (LF_ISSET(DB_TXN_SYNC)) { + F_CLR(txnp, TXN_NOSYNC); + F_SET(txnp, TXN_SYNC); + } + + /* + * 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 = kid->commit(kid, flags)) != 0) + while ((kid = TAILQ_FIRST(&txnp->kids)) != NULL) + 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, + * else do no log writes. If the commit is for a child transaction, + * 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. + */ + if (DBENV_LOGGING(dbenv) && !IS_ZERO_LSN(txnp->last_lsn)) { + if (txnp->parent == NULL) { + /* + * 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 ((ret = __txn_child_log(dbenv, + txnp->parent, &txnp->parent->last_lsn, + 0, txnp->txnid, &txnp->last_lsn)) != 0) { + goto err; + } + + F_SET(txnp->parent, TXN_CHILDCOMMIT); + } + } + + /* 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 -- + * Abort a transaction. + * + * PUBLIC: int __txn_abort __P((DB_TXN *)); + */ +int +__txn_abort(txnp) + DB_TXN *txnp; +{ + DB_ENV *dbenv; + DB_LOCKREQ request; + DB_TXN *kid; + TXN_DETAIL *td; + u_int32_t lflags; + int ret; + + dbenv = txnp->mgrp->dbenv; + + 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, &td, TXN_OP_DISCARD)) != 0) + return (ret); + + /* Should be no children. */ + DB_ASSERT(TAILQ_FIRST(&txnp->kids) == NULL); + DB_ASSERT(F_ISSET(td, TXN_RESTORED)); + + /* 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); + + return (0); +} + +/* + * __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, 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; + + PANIC_CHECK(dbenv); + + if ((ret = __txn_isvalid(txnp, &td, TXN_OP_PREPARE)) != 0) + return (ret); + + /* Commit any unresolved children. */ + while ((kid = TAILQ_FIRST(&txnp->kids)) != NULL) + if ((ret = kid->commit(kid, DB_TXN_NOSYNC)) != 0) + return (ret); + + /* + * 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 (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; + + SET_LOG_FLAGS(dbenv, txnp, lflags); + if ((ret = __txn_xa_regop_log(dbenv, txnp, &txnp->last_lsn, + lflags, TXN_PREPARE, &xid, td->format, td->gtrid, td->bqual, + &td->begin_lsn)) != 0) { + __db_err(dbenv, "DB_TXN->prepare: log_write failed %s", + db_strerror(ret)); + return (ret); + } + } + + MUTEX_THREAD_LOCK(dbenv, txnp->mgrp->mutexp); + td->status = TXN_PREPARED; + MUTEX_THREAD_UNLOCK(dbenv, txnp->mgrp->mutexp); + return (0); +} + +/* + * __txn_id -- + * Return the transaction ID. + * + * PUBLIC: u_int32_t __txn_id __P((DB_TXN *)); + */ +u_int32_t +__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. + */ +static int +__txn_isvalid(txnp, tdp, op) + const DB_TXN *txnp; + TXN_DETAIL **tdp; + 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) { + __db_err(mgrp->dbenv, "transaction has active cursors"); + goto err; + } + + /* 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; + } + + return (0); + +err: /* + * If there's a serious problem with the transaction, panic. TXN + * handles are dead by definition when we return, and if you use + * a cursor you forgot to close, we have no idea what will happen. + */ + return (__db_panic(mgrp->dbenv, EINVAL)); +} + +/* + * __txn_end -- + * Internal transaction end routine. + */ +static int +__txn_end(txnp, is_commit) + DB_TXN *txnp; + int is_commit; +{ + DB_ENV *dbenv; + DB_LOCKREQ request; + DB_TXNMGR *mgr; + DB_TXNREGION *region; + TXN_DETAIL *tp; + 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; + + /* + * __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->stat.st_ncommits++; + else + region->stat.st_naborts++; + --region->stat.st_nactive; + + R_UNLOCK(dbenv, &mgr->reginfo); + + /* + * The transaction cannot get more locks, remove its locker info, + * if any. + */ + 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); + + /* Free the space. */ + if (F_ISSET(txnp, TXN_MALLOC)) { + MUTEX_THREAD_LOCK(dbenv, mgr->mutexp); + TAILQ_REMOVE(&mgr->txn_chain, txnp, links); + MUTEX_THREAD_UNLOCK(dbenv, mgr->mutexp); + + __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); +} + +/* + * __txn_undo -- + * Undo the transaction with id txnid. Returns 0 on success and + * errno on failure. + */ +static int +__txn_undo(txnp) + DB_TXN *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; + + mgr = txnp->mgrp; + dbenv = mgr->dbenv; + logc = NULL; + txnlist = NULL; + + if (!DBENV_LOGGING(dbenv)) + return (0); + + /* + * This is the simplest way to code this, but if the mallocs during + * recovery turn out to be a performance issue, we can do the + * allocation here and use DB_DBT_USERMEM. + */ + memset(&rdbt, 0, sizeof(rdbt)); + + key_lsn = txnp->last_lsn; + + /* + * 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) + return (ret); + + 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 = 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 == 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, + "DB_TXN->abort: Log undo failed for LSN: %lu %lu: %s", + (u_long)key_lsn.file, (u_long)key_lsn.offset, + db_strerror(ret)); + goto err; + } + } + + 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); +} + +/* + * Transaction checkpoint. + * If either kbytes or minutes is non-zero, then we only take the checkpoint + * more than "minutes" minutes have passed since the last checkpoint or if + * more than "kbytes" of log data have been written since the last checkpoint. + * When taking a checkpoint, find the oldest active transaction and figure out + * 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) + DB_ENV *dbenv; + u_int32_t kbytes, minutes, flags; +{ + DB_LSN ckp_lsn, last_ckp; + DB_TXNMGR *mgr; + DB_TXNREGION *region; + TXN_DETAIL *txnp; + time_t last_ckp_time, now; + u_int32_t bytes, mbytes; + int ret; + + PANIC_CHECK(dbenv); + ENV_REQUIRES_CONFIG(dbenv, + dbenv->tx_handle, "txn_checkpoint", DB_INIT_TXN); + + /* + * 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. + */ + if (F_ISSET(dbenv, DB_ENV_REP_CLIENT)) + return (0); + + mgr = dbenv->tx_handle; + region = mgr->reginfo.primary; + + /* + * 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. + */ + __log_txn_lsn(dbenv, &ckp_lsn, &mbytes, &bytes); + + if (!LF_ISSET(DB_FORCE)) { + /* Don't checkpoint a quiescent database. */ + if (bytes == 0 && mbytes == 0) + return (0); + + if (kbytes != 0 && + mbytes * 1024 + bytes / 1024 >= (u_int32_t)kbytes) + goto do_ckp; + + if (minutes != 0) { + (void)time(&now); + + R_LOCK(dbenv, &mgr->reginfo); + last_ckp_time = region->time_ckp; + R_UNLOCK(dbenv, &mgr->reginfo); + + if (now - last_ckp_time >= (time_t)(minutes * 60)) + goto do_ckp; + } + + /* + * If we checked time and data and didn't go to checkpoint, + * we're done. + */ + if (minutes != 0 || kbytes != 0) + return (0); + } + +do_ckp: /* Look through the active transactions for the lowest begin LSN. */ + R_LOCK(dbenv, &mgr->reginfo); + 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); + } + + /* + * 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 (LOGGING_ON(dbenv)) { + R_LOCK(dbenv, &mgr->reginfo); + last_ckp = region->last_ckp; + R_UNLOCK(dbenv, &mgr->reginfo); + + /* + * 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", + (long)ckp_lsn.file, (long)ckp_lsn.offset, + db_strerror(ret)); + return (ret); + } + + __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. + * + * PUBLIC: int __txn_activekids __P((DB_ENV *, u_int32_t, DB_TXN *)); + */ +int +__txn_activekids(dbenv, rectype, txnp) + DB_ENV *dbenv; + u_int32_t rectype; + DB_TXN *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 (F_ISSET(txnp, TXN_COMPENSATE) || rectype == DB___txn_child) + return (0); + + if (TAILQ_FIRST(&txnp->kids) != NULL) { + __db_err(dbenv, "Child transaction is active"); + return (EPERM); + } + 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); +} |