/*- * See the file LICENSE for redistribution information. * * Copyright (c) 1997-2009 Oracle. All rights reserved. * * $Id$ */ #include "db_config.h" #include "db_int.h" #include "dbinc/db_page.h" #include "dbinc/db_am.h" #include "dbinc/fop.h" #include "dbinc/log.h" #include "dbinc/mp.h" #include "dbinc/txn.h" static int __dbreg_check_master __P((ENV *, u_int8_t *, char *)); /* * __dbreg_add_dbentry -- * Adds a DB entry to the dbreg DB entry table. * * PUBLIC: int __dbreg_add_dbentry __P((ENV *, DB_LOG *, DB *, int32_t)); */ int __dbreg_add_dbentry(env, dblp, dbp, ndx) ENV *env; DB_LOG *dblp; DB *dbp; int32_t ndx; { int32_t i; int ret; ret = 0; MUTEX_LOCK(env, dblp->mtx_dbreg); /* * Check if we need to grow the table. Note, ndx is 0-based (the * index into the DB entry table) an dbentry_cnt is 1-based, the * number of available slots. */ if (dblp->dbentry_cnt <= ndx) { if ((ret = __os_realloc(env, (size_t)(ndx + DB_GROW_SIZE) * sizeof(DB_ENTRY), &dblp->dbentry)) != 0) goto err; /* Initialize the new entries. */ for (i = dblp->dbentry_cnt; i < ndx + DB_GROW_SIZE; i++) { dblp->dbentry[i].dbp = NULL; dblp->dbentry[i].deleted = 0; } dblp->dbentry_cnt = i; } DB_ASSERT(env, dblp->dbentry[ndx].dbp == NULL); dblp->dbentry[ndx].deleted = dbp == NULL; dblp->dbentry[ndx].dbp = dbp; err: MUTEX_UNLOCK(env, dblp->mtx_dbreg); return (ret); } /* * __dbreg_rem_dbentry * Remove an entry from the DB entry table. * * PUBLIC: int __dbreg_rem_dbentry __P((DB_LOG *, int32_t)); */ int __dbreg_rem_dbentry(dblp, ndx) DB_LOG *dblp; int32_t ndx; { MUTEX_LOCK(dblp->env, dblp->mtx_dbreg); if (dblp->dbentry_cnt > ndx) { dblp->dbentry[ndx].dbp = NULL; dblp->dbentry[ndx].deleted = 0; } MUTEX_UNLOCK(dblp->env, dblp->mtx_dbreg); return (0); } /* * __dbreg_log_files -- * Put a DBREG_CHKPNT/CLOSE log record for each open database. * * PUBLIC: int __dbreg_log_files __P((ENV *, u_int32_t)); */ int __dbreg_log_files(env, opcode) ENV *env; u_int32_t opcode; { DBT *dbtp, fid_dbt, t; DB_LOG *dblp; DB_LSN r_unused; FNAME *fnp; LOG *lp; int ret; dblp = env->lg_handle; lp = dblp->reginfo.primary; ret = 0; MUTEX_LOCK(env, lp->mtx_filelist); SH_TAILQ_FOREACH(fnp, &lp->fq, q, __fname) { /* This id was revoked by a switch in replication master. */ if (fnp->id == DB_LOGFILEID_INVALID) continue; if (fnp->fname_off == INVALID_ROFF) dbtp = NULL; else { memset(&t, 0, sizeof(t)); t.data = R_ADDR(&dblp->reginfo, fnp->fname_off); t.size = (u_int32_t)strlen(t.data) + 1; dbtp = &t; } memset(&fid_dbt, 0, sizeof(fid_dbt)); fid_dbt.data = fnp->ufid; fid_dbt.size = DB_FILE_ID_LEN; /* * Output DBREG_CHKPNT records which will be processed during * the OPENFILES pass of recovery. At the end of recovery we * want to output the files that were open so a future recovery * run will have the correct files open during a backward pass. * For this we output DBREG_RCLOSE records so the files will be * closed on the forward pass. */ if ((ret = __dbreg_register_log(env, NULL, &r_unused, F_ISSET(fnp, DB_FNAME_DURABLE) ? 0 : DB_LOG_NOT_DURABLE, opcode, dbtp, &fid_dbt, fnp->id, fnp->s_type, fnp->meta_pgno, TXN_INVALID)) != 0) break; } MUTEX_UNLOCK(env, lp->mtx_filelist); return (ret); } /* * __dbreg_close_files -- * Remove the id's of open files and actually close those * files that were opened by the recovery daemon. We sync the * file, unless its mpf pointer has been NULLed by a db_remove or * db_rename. We may not have flushed the log_register record that * closes the file. * * PUBLIC: int __dbreg_close_files __P((ENV *, int)); */ int __dbreg_close_files(env, do_restored) ENV *env; int do_restored; { DB *dbp; DB_LOG *dblp; int ret, t_ret; int32_t i; /* If we haven't initialized logging, we have nothing to do. */ if (!LOGGING_ON(env)) return (0); dblp = env->lg_handle; ret = 0; MUTEX_LOCK(env, dblp->mtx_dbreg); for (i = 0; i < dblp->dbentry_cnt; i++) { /* * We only want to close dbps that recovery opened. Any * dbps that weren't opened by recovery but show up here * are about to be unconditionally removed from the table. * Before doing so, we need to revoke their log fileids * so that we don't end up leaving around FNAME entries * for dbps that shouldn't have them. */ if ((dbp = dblp->dbentry[i].dbp) != NULL) { /* * It's unsafe to call DB->close or revoke_id * while holding the thread lock, because * we'll call __dbreg_rem_dbentry and grab it again. * * Just drop it. Since dbreg ids go monotonically * upward, concurrent opens should be safe, and the * user should have no business closing files while * we're in this loop anyway--we're in the process of * making all outstanding dbps invalid. */ /* * If we only want to close those FNAMES marked * as restored, check now. */ if (do_restored && !F_ISSET(dbp->log_filename, DB_FNAME_RESTORED)) continue; MUTEX_UNLOCK(env, dblp->mtx_dbreg); if (F_ISSET(dbp, DB_AM_RECOVER)) t_ret = __db_close(dbp, NULL, dbp->mpf == NULL ? DB_NOSYNC : 0); else t_ret = __dbreg_revoke_id( dbp, 0, DB_LOGFILEID_INVALID); if (ret == 0) ret = t_ret; MUTEX_LOCK(env, dblp->mtx_dbreg); } dblp->dbentry[i].deleted = 0; dblp->dbentry[i].dbp = NULL; } MUTEX_UNLOCK(env, dblp->mtx_dbreg); return (ret); } /* * __dbreg_close_file -- * Close a database file opened by recovery. * PUBLIC: int __dbreg_close_file __P((ENV *, FNAME *)); */ int __dbreg_close_file(env, fnp) ENV *env; FNAME *fnp; { DB *dbp; DB_LOG *dblp; dblp = env->lg_handle; dbp = dblp->dbentry[fnp->id].dbp; if (dbp == NULL) return (0); DB_ASSERT(env, dbp->log_filename == fnp); DB_ASSERT(env, F_ISSET(dbp, DB_AM_RECOVER)); return (__db_close(dbp, NULL, DB_NOSYNC)); } /* * __dbreg_mark_restored -- * Mark files when we change replication roles and there are outstanding * prepared txns that may use these files. These will be invalidated later * when all outstanding prepared txns are resolved. * * PUBLIC: int __dbreg_mark_restored __P((ENV *)); */ int __dbreg_mark_restored(env) ENV *env; { DB_LOG *dblp; FNAME *fnp; LOG *lp; /* If we haven't initialized logging, we have nothing to do. */ if (!LOGGING_ON(env)) return (0); dblp = env->lg_handle; lp = dblp->reginfo.primary; MUTEX_LOCK(env, lp->mtx_filelist); SH_TAILQ_FOREACH(fnp, &lp->fq, q, __fname) if (fnp->id != DB_LOGFILEID_INVALID) F_SET(fnp, DB_FNAME_RESTORED); MUTEX_UNLOCK(env, lp->mtx_filelist); return (0); } /* * __dbreg_invalidate_files -- * Invalidate files when we change replication roles. Save the * id so that another process will be able to clean up the information * when it notices. * * PUBLIC: int __dbreg_invalidate_files __P((ENV *, int)); */ int __dbreg_invalidate_files(env, do_restored) ENV *env; int do_restored; { DB_LOG *dblp; FNAME *fnp; LOG *lp; int ret; /* If we haven't initialized logging, we have nothing to do. */ if (!LOGGING_ON(env)) return (0); dblp = env->lg_handle; lp = dblp->reginfo.primary; ret = 0; MUTEX_LOCK(env, lp->mtx_filelist); SH_TAILQ_FOREACH(fnp, &lp->fq, q, __fname) { /* * Normally, skip any file with DB_FNAME_RESTORED * set. If do_restored is set, only invalidate * those files with the flag set and skip all others. */ if (F_ISSET(fnp, DB_FNAME_RESTORED) && !do_restored) continue; if (!F_ISSET(fnp, DB_FNAME_RESTORED) && do_restored) continue; if (fnp->id != DB_LOGFILEID_INVALID) { if ((ret = __dbreg_log_close(env, fnp, NULL, DBREG_RCLOSE)) != 0) goto err; fnp->old_id = fnp->id; fnp->id = DB_LOGFILEID_INVALID; } } err: MUTEX_UNLOCK(env, lp->mtx_filelist); return (ret); } /* * __dbreg_id_to_db -- * Return the DB corresponding to the specified dbreg id. * * PUBLIC: int __dbreg_id_to_db __P((ENV *, DB_TXN *, DB **, int32_t, int)); */ int __dbreg_id_to_db(env, txn, dbpp, ndx, tryopen) ENV *env; DB_TXN *txn; DB **dbpp; int32_t ndx; int tryopen; { DB_LOG *dblp; FNAME *fname; int ret; char *name; dblp = env->lg_handle; ret = 0; MUTEX_LOCK(env, dblp->mtx_dbreg); /* * We take a final parameter that indicates whether we should attempt * to open the file if no mapping is found. During recovery, the * recovery routines all want to try to open the file (and this is * called from __dbreg_id_to_db), however, if we have a multi-process * environment where some processes may not have the files open, * then we also get called from __dbreg_assign_id and it's OK if * there is no mapping. * * Under failchk, a process different than the one issuing DB * operations may abort a transaction. In this case, the "recovery" * routines are run by a process that does not necessarily have the * file open, so we we must open the file explicitly. */ if (ndx >= dblp->dbentry_cnt || (!dblp->dbentry[ndx].deleted && dblp->dbentry[ndx].dbp == NULL)) { if (!tryopen || F_ISSET(dblp, DBLOG_RECOVER)) { ret = ENOENT; goto err; } /* * __dbreg_id_to_fname acquires the mtx_filelist mutex, which * we can't safely acquire while we hold the thread lock. We * no longer need it anyway--the dbentry table didn't have what * we needed. */ MUTEX_UNLOCK(env, dblp->mtx_dbreg); if (__dbreg_id_to_fname(dblp, ndx, 0, &fname) != 0) /* * With transactional opens, we may actually have * closed this file in the transaction in which * case this will fail too. Then it's up to the * caller to reopen the file. */ return (ENOENT); /* * Note that we're relying on fname not to change, even though * we released the mutex that protects it (mtx_filelist) inside * __dbreg_id_to_fname. This should be a safe assumption, the * other process that has the file open shouldn't be closing it * while we're trying to abort. */ name = fname->fname_off == INVALID_ROFF ? NULL : R_ADDR(&dblp->reginfo, fname->fname_off); /* * At this point, we are not holding the thread lock, so exit * directly instead of going through the exit code at the * bottom. If the __dbreg_do_open succeeded, then we don't need * to do any of the remaining error checking at the end of this * routine. * If TXN_INVALID is passed then no txnlist is needed. */ if ((ret = __dbreg_do_open(env, txn, dblp, fname->ufid, name, fname->s_type, ndx, fname->meta_pgno, NULL, TXN_INVALID, F_ISSET(fname, DB_FNAME_INMEM) ? DBREG_REOPEN : DBREG_OPEN)) != 0) return (ret); *dbpp = dblp->dbentry[ndx].dbp; return (*dbpp == NULL ? DB_DELETED : 0); } /* * Return DB_DELETED if the file has been deleted (it's not an error). */ if (dblp->dbentry[ndx].deleted) { ret = DB_DELETED; goto err; } /* It's an error if we don't have a corresponding writeable DB. */ if ((*dbpp = dblp->dbentry[ndx].dbp) == NULL) ret = ENOENT; else /* * If we are in recovery, then set that the file has * been written. It is possible to run recovery, * find all the pages in their post update state * in the OS buffer pool, put a checkpoint in the log * and then crash the system without forcing the pages * to disk. If this is an in-memory file, we may not have * an mpf yet. */ if ((*dbpp)->mpf != NULL && (*dbpp)->mpf->mfp != NULL) (*dbpp)->mpf->mfp->file_written = 1; err: MUTEX_UNLOCK(env, dblp->mtx_dbreg); return (ret); } /* * __dbreg_id_to_fname -- * Traverse the shared-memory region looking for the entry that * matches the passed dbreg id. Returns 0 on success; -1 on error. * * PUBLIC: int __dbreg_id_to_fname __P((DB_LOG *, int32_t, int, FNAME **)); */ int __dbreg_id_to_fname(dblp, id, have_lock, fnamep) DB_LOG *dblp; int32_t id; int have_lock; FNAME **fnamep; { ENV *env; FNAME *fnp; LOG *lp; int ret; env = dblp->env; lp = dblp->reginfo.primary; ret = -1; if (!have_lock) MUTEX_LOCK(env, lp->mtx_filelist); SH_TAILQ_FOREACH(fnp, &lp->fq, q, __fname) if (fnp->id == id) { *fnamep = fnp; ret = 0; break; } if (!have_lock) MUTEX_UNLOCK(env, lp->mtx_filelist); return (ret); } /* * __dbreg_fid_to_fname -- * Traverse the shared-memory region looking for the entry that * matches the passed file unique id. Returns 0 on success; -1 on error. * * PUBLIC: int __dbreg_fid_to_fname __P((DB_LOG *, u_int8_t *, int, FNAME **)); */ int __dbreg_fid_to_fname(dblp, fid, have_lock, fnamep) DB_LOG *dblp; u_int8_t *fid; int have_lock; FNAME **fnamep; { ENV *env; FNAME *fnp; LOG *lp; int ret; env = dblp->env; lp = dblp->reginfo.primary; ret = -1; if (!have_lock) MUTEX_LOCK(env, lp->mtx_filelist); SH_TAILQ_FOREACH(fnp, &lp->fq, q, __fname) if (memcmp(fnp->ufid, fid, DB_FILE_ID_LEN) == 0) { *fnamep = fnp; ret = 0; break; } if (!have_lock) MUTEX_UNLOCK(env, lp->mtx_filelist); return (ret); } /* * __dbreg_get_name * * Interface to get name of registered files. This is mainly diagnostic * and the name passed could be transient unless there is something * ensuring that the file cannot be closed. * * PUBLIC: int __dbreg_get_name __P((ENV *, u_int8_t *, char **, char **)); */ int __dbreg_get_name(env, fid, fnamep, dnamep) ENV *env; u_int8_t *fid; char **fnamep, **dnamep; { DB_LOG *dblp; FNAME *fnp; dblp = env->lg_handle; if (dblp != NULL && __dbreg_fid_to_fname(dblp, fid, 0, &fnp) == 0) { *fnamep = fnp->fname_off == INVALID_ROFF ? NULL : R_ADDR(&dblp->reginfo, fnp->fname_off); *dnamep = fnp->dname_off == INVALID_ROFF ? NULL : R_ADDR(&dblp->reginfo, fnp->dname_off); return (0); } *fnamep = *dnamep = NULL; return (-1); } /* * __dbreg_do_open -- * Open files referenced in the log. This is the part of the open that * is not protected by the thread mutex. * PUBLIC: int __dbreg_do_open __P((ENV *, * PUBLIC: DB_TXN *, DB_LOG *, u_int8_t *, char *, DBTYPE, * PUBLIC: int32_t, db_pgno_t, void *, u_int32_t, u_int32_t)); */ int __dbreg_do_open(env, txn, lp, uid, name, ftype, ndx, meta_pgno, info, id, opcode) ENV *env; DB_TXN *txn; DB_LOG *lp; u_int8_t *uid; char *name; DBTYPE ftype; int32_t ndx; db_pgno_t meta_pgno; void *info; u_int32_t id, opcode; { DB *dbp; u_int32_t cstat, ret_stat; int ret, t_ret, try_inmem; char *dname, *fname; cstat = TXN_EXPECTED; fname = name; dname = NULL; try_inmem = 0; retry_inmem: if ((ret = __db_create_internal(&dbp, lp->env, 0)) != 0) return (ret); /* * We can open files under a number of different scenarios. * First, we can open a file during a normal txn_abort, if that file * was opened and closed during the transaction (as is the master * database of a sub-database). * Second, we might be aborting a transaction in a process other than * the one that did it (failchk). * Third, we might be in recovery. * In case 3, there is no locking, so there is no issue. * In cases 1 and 2, we are guaranteed to already hold any locks * that we need, since we're still in the same transaction, so by * setting DB_AM_RECOVER, we guarantee that we don't log and that * we don't try to acquire locks on behalf of a different locker id. */ F_SET(dbp, DB_AM_RECOVER); if (meta_pgno != PGNO_BASE_MD) { memcpy(dbp->fileid, uid, DB_FILE_ID_LEN); dbp->meta_pgno = meta_pgno; } if (opcode == DBREG_PREOPEN) { dbp->type = ftype; if ((ret = __dbreg_setup(dbp, name, NULL, id)) != 0) goto err; MAKE_INMEM(dbp); goto skip_open; } if (opcode == DBREG_REOPEN || try_inmem) { MAKE_INMEM(dbp); fname = NULL; dname = name; } if ((ret = __db_open(dbp, NULL, txn, fname, dname, ftype, DB_DURABLE_UNKNOWN | DB_ODDFILESIZE, DB_MODE_600, meta_pgno)) == 0) { skip_open: /* * Verify that we are opening the same file that we were * referring to when we wrote this log record. */ if ((meta_pgno != PGNO_BASE_MD && __dbreg_check_master(env, uid, name) != 0) || memcmp(uid, dbp->fileid, DB_FILE_ID_LEN) != 0) cstat = TXN_UNEXPECTED; else cstat = TXN_EXPECTED; /* Assign the specific dbreg id to this dbp. */ if ((ret = __dbreg_assign_id(dbp, ndx, 0)) != 0) goto err; /* * If we successfully opened this file, then we need to * convey that information to the txnlist so that we * know how to handle the subtransaction that created * the file system object. */ if (id != TXN_INVALID) ret = __db_txnlist_update(env, info, id, cstat, NULL, &ret_stat, 1); err: if (cstat == TXN_UNEXPECTED) goto not_right; return (ret); } else if (ret == ENOENT) { /* * If the open failed with ENOENT, retry it as a named in-mem * database. Some record types do not distinguish between a * named in-memory database and one on-disk. Therefore, an * internal init via replication that is trying to open and * access this as a named in-mem database will not find it * on-disk, and we need to try to open it in-memory too. * But don't do this for [P]REOPEN, since we're already * handling those cases specially, above. */ if (try_inmem == 0 && opcode != DBREG_PREOPEN && opcode != DBREG_REOPEN) { if ((ret = __db_close(dbp, NULL, DB_NOSYNC)) != 0) return (ret); try_inmem = 1; goto retry_inmem; } else if (try_inmem != 0) CLR_INMEM(dbp); /* * If it exists neither on disk nor in memory * record that the open failed in the txnlist. */ if (id != TXN_INVALID && (ret = __db_txnlist_update(env, info, id, TXN_UNEXPECTED, NULL, &ret_stat, 1)) != 0) goto not_right; /* * If this is file is missing then we may have crashed * without writing the corresponding close, record * the open so recovery will write a close record * with its checkpoint. */ if ((opcode == DBREG_CHKPNT || opcode == DBREG_OPEN) && dbp->log_filename == NULL && (ret = __dbreg_setup(dbp, name, NULL, id)) != 0) return (ret); ret = __dbreg_assign_id(dbp, ndx, 1); return (ret); } not_right: if ((t_ret = __db_close(dbp, NULL, DB_NOSYNC)) != 0) return (ret == 0 ? t_ret : ret); /* Add this file as deleted. */ if ((t_ret = __dbreg_add_dbentry(env, lp, NULL, ndx)) != 0 && ret == 0) ret = t_ret; return (ret); } static int __dbreg_check_master(env, uid, name) ENV *env; u_int8_t *uid; char *name; { DB *dbp; int ret; ret = 0; if ((ret = __db_create_internal(&dbp, env, 0)) != 0) return (ret); F_SET(dbp, DB_AM_RECOVER); ret = __db_open(dbp, NULL, NULL, name, NULL, DB_BTREE, 0, DB_MODE_600, PGNO_BASE_MD); if (ret == 0 && memcmp(uid, dbp->fileid, DB_FILE_ID_LEN) != 0) ret = EINVAL; (void)__db_close(dbp, NULL, 0); return (ret); } /* * __dbreg_lazy_id -- * When a replication client gets upgraded to being a replication master, * it may have database handles open that have not been assigned an ID, but * which have become legal to use for logging. * * This function lazily allocates a new ID for such a function, in a * new transaction created for the purpose. We need to do this in a new * transaction because we definitely wish to commit the dbreg_register, but * at this point we have no way of knowing whether the log record that incited * us to call this will be part of a committed transaction. * * We first revoke any old id this handle may have had. That can happen * if a master becomes a client and then becomes a master again and * there are other processes with valid open handles to this env. * * PUBLIC: int __dbreg_lazy_id __P((DB *)); */ int __dbreg_lazy_id(dbp) DB *dbp; { DB_LOG *dblp; DB_TXN *txn; ENV *env; FNAME *fnp; LOG *lp; int32_t id; int ret; env = dbp->env; DB_ASSERT(env, IS_REP_MASTER(env)); env = dbp->env; dblp = env->lg_handle; lp = dblp->reginfo.primary; fnp = dbp->log_filename; /* The mtx_filelist protects the FNAME list and id management. */ MUTEX_LOCK(env, lp->mtx_filelist); if (fnp->id != DB_LOGFILEID_INVALID) { MUTEX_UNLOCK(env, lp->mtx_filelist); return (0); } id = DB_LOGFILEID_INVALID; /* * When we became master we moved the fnp->id to old_id in * every FNAME structure that was open. If our id was changed, * we need to revoke and give back that id. */ if (fnp->old_id != DB_LOGFILEID_INVALID && (ret = __dbreg_revoke_id(dbp, 1, DB_LOGFILEID_INVALID)) != 0) goto err; if ((ret = __txn_begin(env, NULL, NULL, &txn, 0)) != 0) goto err; if ((ret = __dbreg_get_id(dbp, txn, &id)) != 0) { (void)__txn_abort(txn); goto err; } if ((ret = __txn_commit(txn, DB_TXN_NOSYNC)) != 0) goto err; /* * All DB related logging routines check the id value *without* * holding the mtx_filelist to know whether we need to call * dbreg_lazy_id to begin with. We must set the ID after a * *successful* commit so that there is no possibility of a second * modification call finding a valid ID in the dbp before the * dbreg_register and commit records are in the log. * If there was an error, then we call __dbreg_revoke_id to * remove the entry from the lists. */ fnp->id = id; err: if (ret != 0 && id != DB_LOGFILEID_INVALID) (void)__dbreg_revoke_id(dbp, 1, id); MUTEX_UNLOCK(env, lp->mtx_filelist); return (ret); }