rep_record.c   [plain text]


/*-
 * See the file LICENSE for redistribution information.
 *
 * Copyright (c) 2001,2007 Oracle.  All rights reserved.
 *
 * $Id: rep_record.c,v 12.108 2007/07/10 15:40:16 bostic Exp $
 */

#include "db_config.h"

#include "db_int.h"
#include "dbinc/db_page.h"
#include "dbinc/db_am.h"
#include "dbinc/lock.h"
#include "dbinc/log.h"
#include "dbinc/mp.h"
#include "dbinc/txn.h"

static int __rep_collect_txn __P((DB_ENV *, DB_LSN *, LSN_COLLECTION *));
static int __rep_do_ckp __P((DB_ENV *, DBT *, REP_CONTROL *));
static int __rep_fire_newmaster __P((DB_ENV *, u_int32_t, int));
static int __rep_fire_startupdone __P((DB_ENV *, u_int32_t, int));
static int __rep_getnext __P((DB_ENV *));
static int __rep_lsn_cmp __P((const void *, const void *));
static int __rep_newfile __P((DB_ENV *, REP_CONTROL *, DBT *));
static int __rep_process_rec __P((DB_ENV *, REP_CONTROL *,
    DBT *, db_timespec *, DB_LSN *));
static int __rep_remfirst __P((DB_ENV *, DBT *, DBT *));
static int __rep_skip_msg __P((DB_ENV *, REP *, int, u_int32_t));

/* Used to consistently designate which messages ought to be received where. */

#define	MASTER_ONLY(rep, rp) do {					\
	if (!F_ISSET(rep, REP_F_MASTER)) {				\
		RPRINT(dbenv,						\
		    (dbenv, "Master record received on client"));	\
		REP_PRINT_MESSAGE(dbenv,				\
		    eid, rp, "rep_process_message", 0);		\
		ret = EINVAL;						\
		goto errlock;						\
	}								\
} while (0)

#define	CLIENT_ONLY(rep, rp) do {					\
	if (!F_ISSET(rep, REP_F_CLIENT)) {				\
		RPRINT(dbenv,						\
		    (dbenv, "Client record received on master"));	\
		/*							\
		 * Only broadcast DUPMASTER if leases are not		\
		 * in effect.  If I am an old master, using		\
		 * leases and I get a newer message, my leases		\
		 * had better all be expired.				\
		 */							\
		if (IS_USING_LEASES(dbenv))				\
			DB_ASSERT(dbenv,				\
			    __rep_lease_check(dbenv, 0) ==		\
			    DB_REP_LEASE_EXPIRED);			\
		else {							\
			REP_PRINT_MESSAGE(dbenv,			\
			    eid, rp, "rep_process_message", 0);		\
			(void)__rep_send_message(dbenv,	 DB_EID_BROADCAST, \
			    REP_DUPMASTER, NULL, NULL, 0, 0);		\
		}							\
		ret = DB_REP_DUPMASTER;					\
		goto errlock;						\
	}								\
} while (0)

/*
 * If a client is attempting to service a request it does not have,
 * call rep_skip_msg to skip this message and force a rerequest to the
 * sender.  We don't hold the mutex for the stats and may miscount.
 */
#define	CLIENT_REREQ do {						\
	if (F_ISSET(rep, REP_F_CLIENT)) {				\
		STAT(rep->stat.st_client_svc_req++);			\
		if (ret == DB_NOTFOUND) {				\
			STAT(rep->stat.st_client_svc_miss++);		\
			ret = __rep_skip_msg(dbenv, rep, eid, rp->rectype);\
		}							\
	}								\
} while (0)

#define	MASTER_UPDATE(dbenv, renv) do {					\
	REP_SYSTEM_LOCK(dbenv);						\
	F_SET((renv), DB_REGENV_REPLOCKED);				\
	(void)time(&(renv)->op_timestamp);				\
	REP_SYSTEM_UNLOCK(dbenv);					\
} while (0)

#define	RECOVERING_SKIP do {						\
	if (IS_REP_CLIENT(dbenv) && recovering) {			\
		/* Not holding region mutex, may miscount */		\
		STAT(rep->stat.st_msgs_recover++);			\
		ret = __rep_skip_msg(dbenv, rep, eid, rp->rectype);	\
		goto errlock;						\
	}								\
} while (0)

/*
 * If we're recovering the log we only want log records that are in the
 * range we need to recover.  Otherwise we can end up storing a huge
 * number of "new" records, only to truncate the temp database later after
 * we run recovery.  If we are actively delaying a sync-up, we also skip
 * all incoming log records until the application requests sync-up.
 */
#define	RECOVERING_LOG_SKIP do {					\
	if (F_ISSET(rep, REP_F_DELAY) ||				\
	    rep->master_id == DB_EID_INVALID ||				\
	    (recovering &&						\
	    (!F_ISSET(rep, REP_F_RECOVER_LOG) ||			\
	     LOG_COMPARE(&rp->lsn, &rep->last_lsn) > 0))) {		\
		/* Not holding region mutex, may miscount */		\
		STAT(rep->stat.st_msgs_recover++);			\
		ret = __rep_skip_msg(dbenv, rep, eid, rp->rectype);	\
		goto errlock;						\
	}								\
} while (0)

#define	ANYSITE(rep)

/*
 * __rep_process_message --
 *
 * This routine takes an incoming message and processes it.
 *
 * control: contains the control fields from the record
 * rec: contains the actual record
 * eid: the environment id of the sender of the message;
 * ret_lsnp: On DB_REP_ISPERM and DB_REP_NOTPERM returns, contains the
 *	lsn of the maximum permanent or current not permanent log record
 *	(respectively).
 *
 * PUBLIC: int __rep_process_message __P((DB_ENV *, DBT *, DBT *, int,
 * PUBLIC:     DB_LSN *));
 */
int
__rep_process_message(dbenv, control, rec, eid, ret_lsnp)
	DB_ENV *dbenv;
	DBT *control, *rec;
	int eid;
	DB_LSN *ret_lsnp;
{
	DB_LOG *dblp;
	DB_LSN last_lsn, lsn;
	DB_REP *db_rep;
	DB_THREAD_INFO *ip;
	DBT data_dbt;
	LOG *lp;
	REGENV *renv;
	REGINFO *infop;
	REP *rep;
	REP_CONTROL *rp, tmprp;
	REP_OLD_CONTROL *orp;
	u_int32_t egen, gen;
	int cmp, do_sync, recovering, ret;
	time_t savetime;

	PANIC_CHECK(dbenv);
	ENV_REQUIRES_CONFIG_XX(
	    dbenv, rep_handle, "DB_ENV->rep_process_message", DB_INIT_REP);

	/* Control argument must be non-Null. */
	if (control == NULL || control->size == 0) {
		__db_errx(dbenv,
	"DB_ENV->rep_process_message: control argument must be specified");
		return (EINVAL);
	}

	if (!IS_REP_MASTER(dbenv) && !IS_REP_CLIENT(dbenv)) {
		__db_errx(dbenv,
	"Environment not configured as replication master or client");
		return (EINVAL);
	}

	if ((ret = __dbt_usercopy(dbenv, control)) != 0 ||
	    (ret = __dbt_usercopy(dbenv, rec)) != 0) {
		__dbt_userfree(dbenv, control, rec, NULL);
		__db_errx(dbenv,
	"DB_ENV->rep_process_message: error retrieving DBT contents");
		return ret;
	}

	ret = 0;
	do_sync = 0;
	db_rep = dbenv->rep_handle;
	rep = db_rep->region;
	dblp = dbenv->lg_handle;
	lp = dblp->reginfo.primary;
	infop = dbenv->reginfo;
	renv = infop->primary;
	/*
	 * Casting this to REP_OLD_CONTROL is just kind of stylistic: the
	 * rep_version field of course has to be in the same offset in all
	 * versions in order for this to work.
	 */
	if (((REP_OLD_CONTROL *)control->data)->rep_version
	    <= DB_REPVERSION_45) {
		orp = (REP_OLD_CONTROL *)control->data;
		tmprp.rep_version = orp->rep_version;
		tmprp.log_version = orp->log_version;
		tmprp.lsn = orp->lsn;
		tmprp.rectype = orp->rectype;
		tmprp.gen = orp->gen;
		tmprp.flags = orp->flags;
		timespecclear(&tmprp.msg_time);
		rp = (REP_CONTROL *)&tmprp;
	} else
		rp = (REP_CONTROL *)control->data;
	if (ret_lsnp != NULL)
		ZERO_LSN(*ret_lsnp);

	ENV_ENTER(dbenv, ip);

	REP_PRINT_MESSAGE(dbenv, eid, rp, "rep_process_message", 0);
	/*
	 * Acquire the replication lock.
	 */
	REP_SYSTEM_LOCK(dbenv);
	if (F_ISSET(rep, REP_F_READY_MSG)) {
		/*
		 * If we're racing with a thread in rep_start, then
		 * just ignore the message and return.
		 */
		RPRINT(dbenv, (dbenv,
		    "Racing replication msg lockout, ignore message."));
		if (F_ISSET(rp, REPCTL_PERM))
			ret = DB_REP_IGNORE;
		REP_SYSTEM_UNLOCK(dbenv);
		/*
		 * If another client has sent a c2c request to us, it may be a
		 * long time before it resends the request (due to its dual data
		 * streams avoidance heuristic); let it know we can't serve the
		 * request just now.
		 */
		if (F_ISSET(rep, REP_F_CLIENT) && REP_MSG_REQ(rp->rectype)) {
			STAT(rep->stat.st_client_svc_req++);
			STAT(rep->stat.st_client_svc_miss++);
			(void)__rep_send_message(dbenv,
			    eid, REP_REREQUEST, NULL, NULL, 0, 0);
		}
		goto out;
	}
	rep->msg_th++;
	gen = rep->gen;
	recovering = F_ISSET(rep, REP_F_RECOVER_MASK);
	savetime = renv->rep_timestamp;

	STAT(rep->stat.st_msgs_processed++);
	REP_SYSTEM_UNLOCK(dbenv);

	/*
	 * Check the version number for both rep and log.  If it is
	 * an old version we support, convert it.  Otherwise complain.
	 */
	if (rp->rep_version < DB_REPVERSION) {
		RPRINT(dbenv, (dbenv,
		    "Received record %lu with old rep version %lu",
		    (u_long)rp->rectype, (u_long)rp->rep_version));
		rp->rectype = __rep_msg_from_old(rp->rep_version, rp->rectype);
		/*
		 * We should have a valid new record type for all the old
		 * versions.
		 */
		DB_ASSERT(dbenv, rp->rectype != REP_INVALID);
		RPRINT(dbenv, (dbenv,
		    "Converted to record %lu with old rep version %lu",
		    (u_long)rp->rectype, (u_long)rp->rep_version));
	} else if (rp->rep_version > DB_REPVERSION) {
		__db_errx(dbenv,
		    "unexpected replication message version %lu, expected %d",
		    (u_long)rp->rep_version, DB_REPVERSION);
		ret = EINVAL;
		goto errlock;
	}

	if (rp->log_version < DB_LOGVERSION) {
		RPRINT(dbenv, (dbenv,
		    "Received record %lu with old log version %lu",
		    (u_long)rp->rectype, (u_long)rp->log_version));
	} else if (rp->log_version > DB_LOGVERSION) {
		__db_errx(dbenv,
		    "unexpected log record version %lu, expected %d",
		    (u_long)rp->log_version, DB_LOGVERSION);
		ret = EINVAL;
		goto errlock;
	}

	/*
	 * Check for generation number matching.  Ignore any old messages
	 * except requests that are indicative of a new client that needs
	 * to get in sync.
	 */
	if (rp->gen < gen && rp->rectype != REP_ALIVE_REQ &&
	    rp->rectype != REP_NEWCLIENT && rp->rectype != REP_MASTER_REQ &&
	    rp->rectype != REP_DUPMASTER) {
		/*
		 * We don't hold the rep mutex, and could miscount if we race.
		 */
		STAT(rep->stat.st_msgs_badgen++);
		if (F_ISSET(rp, REPCTL_PERM))
			ret = DB_REP_IGNORE;
		goto errlock;
	}

	if (rp->gen > gen) {
		/*
		 * If I am a master and am out of date with a lower generation
		 * number, I am in bad shape and should downgrade.
		 */
		if (F_ISSET(rep, REP_F_MASTER)) {
			STAT(rep->stat.st_dupmasters++);
			ret = DB_REP_DUPMASTER;
			/*
			 * Only broadcast DUPMASTER if leases are not
			 * in effect.  If I am an old master, using
			 * leases and I get a newer message, my leases
			 * had better all be expired.
			 */
			if (IS_USING_LEASES(dbenv))
				DB_ASSERT(dbenv,
				    __rep_lease_check(dbenv, 0) ==
				    DB_REP_LEASE_EXPIRED);
			else if (rp->rectype != REP_DUPMASTER)
				(void)__rep_send_message(dbenv,
				    DB_EID_BROADCAST, REP_DUPMASTER,
				    NULL, NULL, 0, 0);
			goto errlock;
		}

		/*
		 * I am a client and am out of date.  If this is an election,
		 * or a response from the first site I contacted, then I can
		 * accept the generation number and participate in future
		 * elections and communication. Otherwise, I need to hear about
		 * a new master and sync up.
		 */
		if (rp->rectype == REP_ALIVE ||
		    rp->rectype == REP_VOTE1 || rp->rectype == REP_VOTE2) {
			REP_SYSTEM_LOCK(dbenv);
			RPRINT(dbenv, (dbenv,
			    "Updating gen from %lu to %lu",
			    (u_long)gen, (u_long)rp->gen));
			rep->master_id = DB_EID_INVALID;
			gen = rep->gen = rp->gen;
			/*
			 * Updating of egen will happen when we process the
			 * message below for each message type.
			 */
			REP_SYSTEM_UNLOCK(dbenv);
			if (rp->rectype == REP_ALIVE)
				(void)__rep_send_message(dbenv,
				    DB_EID_BROADCAST, REP_MASTER_REQ, NULL,
				    NULL, 0, 0);
		} else if (rp->rectype != REP_NEWMASTER) {
			/*
			 * Ignore this message, retransmit if needed.
			 */
			if (__rep_check_doreq(dbenv, rep))
				(void)__rep_send_message(dbenv,
				    DB_EID_BROADCAST, REP_MASTER_REQ,
				    NULL, NULL, 0, 0);
			goto errlock;
		}
		/*
		 * If you get here, then you're a client and either you're
		 * in an election or you have a NEWMASTER or an ALIVE message
		 * whose processing will do the right thing below.
		 */
	}

	/*
	 * If the sender is part of an established group, so are we now.
	 */
	if (F_ISSET(rp, REPCTL_GROUP_ESTD)) {
		REP_SYSTEM_LOCK(dbenv);
#ifdef	DIAGNOSTIC
		if (!F_ISSET(rep, REP_F_GROUP_ESTD))
			RPRINT(dbenv, (dbenv,
			    "I am now part of an established group"));
#endif
		F_SET(rep, REP_F_GROUP_ESTD);
		REP_SYSTEM_UNLOCK(dbenv);
	}

	/*
	 * We need to check if we're in recovery and if we are
	 * then we need to ignore any messages except VERIFY*, VOTE*,
	 * NEW* and ALIVE_REQ, or backup related messages: UPDATE*,
	 * PAGE* and FILE*.  We need to also accept LOG messages
	 * if we're copying the log for recovery/backup.
	 */
	switch (rp->rectype) {
	case REP_ALIVE:
		/*
		 * Handle even if we're recovering.
		 */
		ANYSITE(rep);
		egen = *(u_int32_t *)rec->data;
		REP_SYSTEM_LOCK(dbenv);
		RPRINT(dbenv, (dbenv,
		    "Received ALIVE egen of %lu, mine %lu",
		    (u_long)egen, (u_long)rep->egen));
		if (egen > rep->egen) {
			/*
			 * We're changing egen, need to clear out any old
			 * election information.  We need to set the
			 * REP_F_EGENUPDATE flag here so that any thread
			 * waiting in rep_elect/rep_wait can distinguish
			 * this situation (and restart its election) from
			 * a current master saying it is still master and
			 * the egen getting incremented on that path.
			 */
			__rep_elect_done(dbenv, rep);
			rep->egen = egen;
			F_SET(rep, REP_F_EGENUPDATE);
		}
		REP_SYSTEM_UNLOCK(dbenv);
		break;
	case REP_ALIVE_REQ:
		/*
		 * Handle even if we're recovering.
		 */
		ANYSITE(rep);
		dblp = dbenv->lg_handle;
		LOG_SYSTEM_LOCK(dbenv);
		lsn = ((LOG *)dblp->reginfo.primary)->lsn;
		LOG_SYSTEM_UNLOCK(dbenv);
		REP_SYSTEM_LOCK(dbenv);
		egen = rep->egen;
		REP_SYSTEM_UNLOCK(dbenv);
		data_dbt.data = &egen;
		data_dbt.size = sizeof(egen);
		(void)__rep_send_message(dbenv,
		    eid, REP_ALIVE, &lsn, &data_dbt, 0, 0);
		break;
	case REP_ALL_REQ:
		RECOVERING_SKIP;
		ret = __rep_allreq(dbenv, rp, eid);
		CLIENT_REREQ;
		break;
	case REP_BULK_LOG:
		RECOVERING_LOG_SKIP;
		CLIENT_ONLY(rep, rp);
		ret = __rep_bulk_log(dbenv, rp, rec, savetime, ret_lsnp);
		break;
	case REP_BULK_PAGE:
		/*
		 * Handle even if we're recovering.
		 */
		CLIENT_ONLY(rep, rp);
		ret = __rep_bulk_page(dbenv, eid, rp, rec);
		break;
	case REP_DUPMASTER:
		/*
		 * Handle even if we're recovering.
		 */
		if (F_ISSET(rep, REP_F_MASTER))
			ret = DB_REP_DUPMASTER;
		break;
#ifdef NOTYET
	case REP_FILE: /* TODO */
		CLIENT_ONLY(rep, rp);
		break;
	case REP_FILE_REQ:
		ret = __rep_send_file(dbenv, rec, eid);
		break;
#endif
	case REP_FILE_FAIL:
		/*
		 * Handle even if we're recovering.
		 */
		CLIENT_ONLY(rep, rp);
		/*
		 * XXX
		 */
		break;
	case REP_LEASE_GRANT:
		/*
		 * Handle even if we're recovering.
		 */
		MASTER_ONLY(rep, rp);
		ret = __rep_lease_grant(dbenv, rp, rec, eid);
		break;
	case REP_LOG:
	case REP_LOG_MORE:
		RECOVERING_LOG_SKIP;
		CLIENT_ONLY(rep, rp);
		ret = __rep_log(dbenv, rp, rec, savetime, ret_lsnp);
		break;
	case REP_LOG_REQ:
		RECOVERING_SKIP;
		if (F_ISSET(rp, REPCTL_INIT))
			MASTER_UPDATE(dbenv, renv);
		ret = __rep_logreq(dbenv, rp, rec, eid);
		CLIENT_REREQ;
		break;
	case REP_NEWSITE:
		/*
		 * Handle even if we're recovering.
		 */
		/* We don't hold the rep mutex, and may miscount. */
		STAT(rep->stat.st_newsites++);

		/* This is a rebroadcast; simply tell the application. */
		if (F_ISSET(rep, REP_F_MASTER)) {
			dblp = dbenv->lg_handle;
			lp = dblp->reginfo.primary;
			LOG_SYSTEM_LOCK(dbenv);
			lsn = lp->lsn;
			LOG_SYSTEM_UNLOCK(dbenv);
			(void)__rep_send_message(dbenv,
			    eid, REP_NEWMASTER, &lsn, NULL, 0, 0);
			if (IS_USING_LEASES(dbenv))
				ret = __rep_lease_refresh(dbenv);
		}
		ret = DB_REP_NEWSITE;
		break;
	case REP_NEWCLIENT:
		/*
		 * Handle even if we're recovering.
		 */
		/*
		 * This message was received and should have resulted in the
		 * application entering the machine ID in its machine table.
		 * We respond to this with an ALIVE to send relevant information
		 * to the new client (if we are a master, we'll send a
		 * NEWMASTER, so we only need to send the ALIVE if we're a
		 * client).  But first, broadcast the new client's record to
		 * all the clients.
		 */
		(void)__rep_send_message(dbenv,
		    DB_EID_BROADCAST, REP_NEWSITE, &rp->lsn, rec, 0, 0);

		ret = DB_REP_NEWSITE;

		if (F_ISSET(rep, REP_F_CLIENT)) {
			REP_SYSTEM_LOCK(dbenv);
			egen = rep->egen;
			if (eid == rep->master_id)
				rep->master_id = DB_EID_INVALID;
			REP_SYSTEM_UNLOCK(dbenv);
			data_dbt.data = &egen;
			data_dbt.size = sizeof(egen);
			(void)__rep_send_message(dbenv, DB_EID_BROADCAST,
			    REP_ALIVE, &rp->lsn, &data_dbt, 0, 0);
			break;
		}
		/* FALLTHROUGH */
	case REP_MASTER_REQ:
		RECOVERING_SKIP;
		if (F_ISSET(rep, REP_F_MASTER)) {
			LOG_SYSTEM_LOCK(dbenv);
			lsn = lp->lsn;
			LOG_SYSTEM_UNLOCK(dbenv);
			(void)__rep_send_message(dbenv,
			    DB_EID_BROADCAST, REP_NEWMASTER, &lsn, NULL, 0, 0);
			if (IS_USING_LEASES(dbenv))
				ret = __rep_lease_refresh(dbenv);
		}
		/*
		 * If there is no master, then we could get into a state
		 * where an old client lost the initial ALIVE message and
		 * is calling an election under an old gen and can
		 * never get to the current gen.
		 */
		if (F_ISSET(rep, REP_F_CLIENT) && rp->gen < gen) {
			REP_SYSTEM_LOCK(dbenv);
			egen = rep->egen;
			if (eid == rep->master_id)
				rep->master_id = DB_EID_INVALID;
			REP_SYSTEM_UNLOCK(dbenv);
			data_dbt.data = &egen;
			data_dbt.size = sizeof(egen);
			(void)__rep_send_message(dbenv, eid,
			    REP_ALIVE, &rp->lsn, &data_dbt, 0, 0);
		}
		break;
	case REP_NEWFILE:
		RECOVERING_LOG_SKIP;
		CLIENT_ONLY(rep, rp);
		ret = __rep_apply(dbenv, rp, rec, ret_lsnp, NULL, &last_lsn);
		if (ret == DB_REP_LOGREADY)
			ret = __rep_logready(dbenv, rep, savetime, &last_lsn);
		break;
	case REP_NEWMASTER:
		/*
		 * Handle even if we're recovering.
		 */
		ANYSITE(rep);
		if (F_ISSET(rep, REP_F_MASTER) &&
		    eid != rep->eid) {
			/* We don't hold the rep mutex, and may miscount. */
			STAT(rep->stat.st_dupmasters++);
			ret = DB_REP_DUPMASTER;
			if (IS_USING_LEASES(dbenv))
				DB_ASSERT(dbenv,
				    __rep_lease_check(dbenv, 0) ==
				    DB_REP_LEASE_EXPIRED);
			else
				(void)__rep_send_message(dbenv,
				    DB_EID_BROADCAST, REP_DUPMASTER,
				    NULL, NULL, 0, 0);
			break;
		}
		if ((ret =
		    __rep_new_master(dbenv, rp, eid)) == DB_REP_NEWMASTER)
			ret = __rep_fire_newmaster(dbenv, rp->gen, eid);
		break;
	case REP_PAGE:
	case REP_PAGE_MORE:
		/*
		 * Handle even if we're recovering.
		 */
		CLIENT_ONLY(rep, rp);
		ret = __rep_page(dbenv, eid, rp, rec);
		if (ret == DB_REP_PAGEDONE)
			ret = 0;
		break;
	case REP_PAGE_FAIL:
		/*
		 * Handle even if we're recovering.
		 */
		CLIENT_ONLY(rep, rp);
		ret = __rep_page_fail(dbenv, eid, rec);
		break;
	case REP_PAGE_REQ:
		RECOVERING_SKIP;
		MASTER_UPDATE(dbenv, renv);
		ret = __rep_page_req(dbenv, eid, rec);
		CLIENT_REREQ;
		break;
	case REP_REREQUEST:
		/*
		 * Handle even if we're recovering.  Don't do a master
		 * check.
		 */
		CLIENT_ONLY(rep, rp);
		/*
		 * Don't hold any mutex, may miscount.
		 */
		STAT(rep->stat.st_client_rerequests++);
		ret = __rep_resend_req(dbenv, 1);
		break;
	case REP_START_SYNC:
		RECOVERING_SKIP;
		MUTEX_LOCK(dbenv, rep->mtx_clientdb);
		cmp = LOG_COMPARE(&rp->lsn, &lp->ready_lsn);
		/*
		 * The comparison needs to be <= because the LSN in
		 * the message can be the LSN of the first outstanding
		 * txn, which may be the LSN immediately after the
		 * previous commit.  The ready_lsn is the LSN of the
		 * next record expected.  In that case, the LSNs
		 * could be equal and the client has the commit and
		 * wants to sync. [SR #15338]
		 */
		if (cmp <= 0) {
			MUTEX_UNLOCK(dbenv, rep->mtx_clientdb);
			do_sync = 1;
		} else {
			STAT(rep->stat.st_startsync_delayed++);
			/*
			 * There are cases where keeping the first ckp_lsn
			 * LSN is advantageous and cases where keeping
			 * a later LSN is better.  If random, earlier
			 * log records are missing, keeping the later
			 * LSN seems to be better.  That is what we'll
			 * do for now.
			 */
			if (LOG_COMPARE(&rp->lsn, &rep->ckp_lsn) > 0)
				rep->ckp_lsn = rp->lsn;
			RPRINT(dbenv, (dbenv,
    "Delayed START_SYNC memp_sync due to missing records."));
			RPRINT(dbenv, (dbenv,
    "ready LSN [%lu][%lu], ckp_lsn [%lu][%lu]",
		    (u_long)lp->ready_lsn.file, (u_long)lp->ready_lsn.offset,
		    (u_long)rep->ckp_lsn.file, (u_long)rep->ckp_lsn.offset));
			MUTEX_UNLOCK(dbenv, rep->mtx_clientdb);
		}
		break;
	case REP_UPDATE:
		/*
		 * Handle even if we're recovering.
		 */
		CLIENT_ONLY(rep, rp);
		ret = __rep_update_setup(dbenv, eid, rp, rec);
		break;
	case REP_UPDATE_REQ:
		/*
		 * Handle even if we're recovering.
		 */
		MASTER_ONLY(rep, rp);
		infop = dbenv->reginfo;
		renv = infop->primary;
		MASTER_UPDATE(dbenv, renv);
		ret = __rep_update_req(dbenv, eid);
		break;
	case REP_VERIFY:
		if (recovering) {
			MUTEX_LOCK(dbenv, rep->mtx_clientdb);
			cmp = LOG_COMPARE(&lp->verify_lsn, &rp->lsn);
			MUTEX_UNLOCK(dbenv, rep->mtx_clientdb);
			/*
			 * If this is not the verify record I want, skip it.
			 */
			if (cmp != 0) {
				ret = __rep_skip_msg(
				    dbenv, rep, eid, rp->rectype);
				break;
			}
		}
		CLIENT_ONLY(rep, rp);
		ret = __rep_verify(dbenv, rp, rec, eid, savetime);
		break;
	case REP_VERIFY_FAIL:
		/*
		 * Handle even if we're recovering.
		 */
		CLIENT_ONLY(rep, rp);
		ret = __rep_verify_fail(dbenv, rp, eid);
		break;
	case REP_VERIFY_REQ:
		RECOVERING_SKIP;
		ret = __rep_verify_req(dbenv, rp, eid);
		CLIENT_REREQ;
		break;
	case REP_VOTE1:
		/*
		 * Handle even if we're recovering.
		 */
		ret = __rep_vote1(dbenv, rp, rec, eid);
		break;
	case REP_VOTE2:
		/*
		 * Handle even if we're recovering.
		 */
		ret = __rep_vote2(dbenv, rec, eid);
		break;
	default:
		__db_errx(dbenv,
	"DB_ENV->rep_process_message: unknown replication message: type %lu",
		   (u_long)rp->rectype);
		ret = EINVAL;
		break;
	}

errlock:
	REP_SYSTEM_LOCK(dbenv);
	rep->msg_th--;
	REP_SYSTEM_UNLOCK(dbenv);
	if (do_sync) {
		MUTEX_LOCK(dbenv, rep->mtx_ckp);
		lsn = rp->lsn;
		/*
		 * This is the REP_START_SYNC sync, and so we permit it to be
		 * interrupted.
		 */
		ret = __memp_sync(
		    dbenv, DB_SYNC_CHECKPOINT | DB_SYNC_INTERRUPT_OK, &lsn);
		MUTEX_UNLOCK(dbenv, rep->mtx_ckp);
		RPRINT(dbenv, (dbenv, "ALIVE: Completed sync [%lu][%lu]",
		    (u_long)lsn.file, (u_long)lsn.offset));
	}
out:
	if (ret == 0 && F_ISSET(rp, REPCTL_PERM)) {
		if (ret_lsnp != NULL)
			*ret_lsnp = rp->lsn;
		ret = DB_REP_NOTPERM;
	}
	__dbt_userfree(dbenv, control, rec, NULL);
	ENV_LEAVE(dbenv, ip);
	return (ret);
}

/*
 * __rep_apply --
 *
 * Handle incoming log records on a client, applying when possible and
 * entering into the bookkeeping table otherwise.  This routine manages
 * the state of the incoming message stream -- processing records, via
 * __rep_process_rec, when possible and enqueuing in the __db.rep.db
 * when necessary.  As gaps in the stream are filled in, this is where
 * we try to process as much as possible from __db.rep.db to catch up.
 *
 * PUBLIC: int __rep_apply __P((DB_ENV *, REP_CONTROL *,
 * PUBLIC:     DBT *, DB_LSN *, int *, DB_LSN *));
 */
int
__rep_apply(dbenv, rp, rec, ret_lsnp, is_dupp, last_lsnp)
	DB_ENV *dbenv;
	REP_CONTROL *rp;
	DBT *rec;
	DB_LSN *ret_lsnp;
	int *is_dupp;
	DB_LSN *last_lsnp;
{
	DB_REP *db_rep;
	DBT control_dbt, key_dbt;
	DBT rec_dbt;
	DB *dbp;
	DB_LOG *dblp;
	DB_LSN max_lsn, save_lsn;
	LOG *lp;
	REP *rep;
	db_timespec max_ts;
	u_int32_t gen;
	int cmp, event, master, ret, set_apply, t_ret;

	COMPQUIET(gen, 0);
	COMPQUIET(master, DB_EID_INVALID);

	db_rep = dbenv->rep_handle;
	rep = db_rep->region;
	event = ret = set_apply = 0;
	memset(&control_dbt, 0, sizeof(control_dbt));
	memset(&rec_dbt, 0, sizeof(rec_dbt));
	ZERO_LSN(max_lsn);
	timespecclear(&max_ts);
	cmp = -2;		/* OOB value that LOG_COMPARE can't return. */

	dblp = dbenv->lg_handle;
	MUTEX_LOCK(dbenv, rep->mtx_clientdb);
	/*
	 * Lazily open the temp db.  Always set the startup flag to 0
	 * because it was initialized from rep_start.
	 */
	if (db_rep->rep_db == NULL &&
	    (ret = __rep_client_dbinit(dbenv, 0, REP_DB)) != 0) {
		MUTEX_UNLOCK(dbenv, rep->mtx_clientdb);
		goto out;
	}
	dbp = db_rep->rep_db;
	lp = dblp->reginfo.primary;
	REP_SYSTEM_LOCK(dbenv);
	if (F_ISSET(rep, REP_F_RECOVER_LOG) &&
	    LOG_COMPARE(&lp->ready_lsn, &rep->first_lsn) < 0)
		lp->ready_lsn = rep->first_lsn;
	cmp = LOG_COMPARE(&rp->lsn, &lp->ready_lsn);
	/*
	 * If we are going to skip or process any message other
	 * than a duplicate, make note of it if we're in an
	 * election so that the election can rerequest proactively.
	 */
	if (F_ISSET(rep, REP_F_READY_APPLY) && cmp >= 0)
		F_SET(rep, REP_F_SKIPPED_APPLY);

	/*
	 * If we're in the middle of processing a NEWFILE, we've dropped
	 * the mutex and if this matches it is a duplicate record.  We
	 * do not want this call taking the "matching" code below because
	 * we may then process later records in the temp db and the
	 * original NEWFILE may not have the log file ready.  It will
	 * process those temp db items when it completes.
	 */
	if (F_ISSET(rep, REP_F_NEWFILE) && cmp == 0)
		cmp = -1;

	if (cmp == 0) {
		/*
		 * If we are in an election (i.e. we've sent a vote
		 * with an LSN in it), then we drop the next record
		 * we're expecting.  When we find a master, we'll
		 * either go into sync, or if it was an existing
		 * master, rerequest this one record (later records
		 * are accumulating in the temp db).
		 *
		 * We can simply return here, and rep_process_message
		 * will set NOTPERM if necessary for this record.
		 */
		if (F_ISSET(rep, REP_F_READY_APPLY)) {
			/*
			 * We will simply return now.  All special return
			 * processing should be ignored because the special
			 * values are just initialized.  Variables like
			 * max_lsn are still 0.
			 */
			RPRINT(dbenv, (dbenv,
			    "rep_apply: In election. Ignoring [%lu][%lu]",
			    (u_long)rp->lsn.file, (u_long)rp->lsn.offset));
			REP_SYSTEM_UNLOCK(dbenv);
			MUTEX_UNLOCK(dbenv, rep->mtx_clientdb);
			goto out;
		}
		rep->apply_th++;
		set_apply = 1;
		RPRINT(dbenv, (dbenv,
		    "rep_apply: Set apply_th %d", rep->apply_th));
		REP_SYSTEM_UNLOCK(dbenv);
		if ((ret =
		    __rep_process_rec(dbenv, rp, rec, &max_ts, &max_lsn)) != 0)
			goto err;
		/*
		 * If we get the record we are expecting, reset
		 * the count of records we've received and are applying
		 * towards the request interval.
		 */
		lp->rcvd_recs = 0;
		ZERO_LSN(lp->max_wait_lsn);

		while (ret == 0 &&
		    LOG_COMPARE(&lp->ready_lsn, &lp->waiting_lsn) == 0) {
			/*
			 * We just filled in a gap in the log record stream.
			 * Write subsequent records to the log.
			 */
gap_check:
			if ((ret =
			    __rep_remfirst(dbenv, &control_dbt, &rec_dbt)) != 0)
				goto err;

			rp = (REP_CONTROL *)control_dbt.data;
			rec = &rec_dbt;
			if ((ret = __rep_process_rec(dbenv,
			    rp, rec, &max_ts, &max_lsn)) != 0)
				goto err;

			--rep->stat.st_log_queued;

			/*
			 * Since we just filled a gap in the log stream, and
			 * we're writing subsequent records to the log, we want
			 * to use rcvd_recs and wait_recs so that we will
			 * request the next gap if we end up with a gap and
			 * a lot of records still in the temp db, but not
			 * request if it is near the end of the temp db and
			 * likely to arrive on its own shortly.  We want to
			 * avoid requesting the record in that case.  Also
			 * reset max_wait_lsn because the next gap is a
			 * fresh gap.
			 */
			lp->rcvd_recs = rep->stat.st_log_queued;
			lp->wait_recs = rep->request_gap;

			if ((ret = __rep_getnext(dbenv)) == DB_NOTFOUND) {
				lp->rcvd_recs = 0;
				ret = 0;
				break;
			} else if (ret != 0)
				goto err;
		}

		/*
		 * Check if we're at a gap in the table and if so, whether we
		 * need to ask for any records.
		 */
		if (!IS_ZERO_LSN(lp->waiting_lsn) &&
		    LOG_COMPARE(&lp->ready_lsn, &lp->waiting_lsn) != 0) {
			/*
			 * We got a record and processed it, but we may
			 * still be waiting for more records.  If we
			 * filled a gap we keep a count of how many other
			 * records are in the temp database and if we should
			 * request the next gap at this time.
			 */
			if (__rep_check_doreq(dbenv, rep) && (ret =
			    __rep_loggap_req(dbenv, rep, &rp->lsn, 0)) != 0)
				goto err;
		} else {
			lp->wait_recs = 0;
			ZERO_LSN(lp->max_wait_lsn);
		}

	} else if (cmp > 0) {
		/*
		 * The LSN is higher than the one we were waiting for.
		 * This record isn't in sequence; add it to the temporary
		 * database, update waiting_lsn if necessary, and perform
		 * calculations to determine if we should issue requests
		 * for new records.
		 */
		REP_SYSTEM_UNLOCK(dbenv);
		memset(&key_dbt, 0, sizeof(key_dbt));
		key_dbt.data = rp;
		key_dbt.size = sizeof(*rp);
		if (lp->wait_recs == 0) {
			/*
			 * This is a new gap. Initialize the number of
			 * records that we should wait before requesting
			 * that it be resent.  We grab the limits out of
			 * the rep without the mutex.
			 */
			lp->wait_recs = rep->request_gap;
			lp->rcvd_recs = 0;
			ZERO_LSN(lp->max_wait_lsn);
		}
		if (__rep_check_doreq(dbenv, rep) &&
		    (ret = __rep_loggap_req(dbenv, rep, &rp->lsn, 0) != 0))
			goto err;

		ret = __db_put(dbp, NULL, &key_dbt, rec, DB_NOOVERWRITE);
		if (ret == 0) {
			rep->stat.st_log_queued++;
#ifdef HAVE_STATISTICS
			STAT(rep->stat.st_log_queued_total++);
			if (rep->stat.st_log_queued_max <
			    rep->stat.st_log_queued)
				rep->stat.st_log_queued_max =
				    rep->stat.st_log_queued;
#endif
		}

		if (ret == DB_KEYEXIST)
			ret = 0;
		if (ret != 0)
			goto done;

		if (IS_ZERO_LSN(lp->waiting_lsn) ||
		    LOG_COMPARE(&rp->lsn, &lp->waiting_lsn) < 0)
			lp->waiting_lsn = rp->lsn;

		/*
		 * If this is permanent; let the caller know that we have
		 * not yet written it to disk, but we've accepted it.
		 */
		if (ret == 0 && F_ISSET(rp, REPCTL_PERM)) {
			max_lsn = rp->lsn;
			ret = DB_REP_NOTPERM;
		}
		goto done;
	} else {
		STAT(rep->stat.st_log_duplicated++);
		REP_SYSTEM_UNLOCK(dbenv);
		if (is_dupp != NULL)
			*is_dupp = 1;
		if (F_ISSET(rp, REPCTL_PERM))
			max_lsn = lp->max_perm_lsn;
		/*
		 * We check REPCTL_LEASE here, because this client may
		 * have leases configured but the master may not (especially
		 * in a mixed version group.  If the master has leases
		 * configured, all clients must also.
		 */
		if (IS_USING_LEASES(dbenv) &&
		    F_ISSET(rp, REPCTL_LEASE) &&
		    timespecisset(&rp->msg_time)) {
			if (timespeccmp(&rp->msg_time, &lp->max_lease_ts, >))
				max_ts = rp->msg_time;
			else
				max_ts = lp->max_lease_ts;
		}
		goto done;
	}

	/* Check if we need to go back into the table. */
	if (ret == 0 && LOG_COMPARE(&lp->ready_lsn, &lp->waiting_lsn) == 0)
		goto gap_check;

done:
err:	/*
	 * In case of a race, to make sure only one thread can get
	 * DB_REP_LOGREADY, zero out rep->last_lsn to show that we've gotten to
	 * this point.
	 */
	REP_SYSTEM_LOCK(dbenv);
	if (ret == 0 &&
	    F_ISSET(rep, REP_F_RECOVER_LOG) &&
	    !IS_ZERO_LSN(rep->last_lsn) &&
	    LOG_COMPARE(&lp->ready_lsn, &rep->last_lsn) >= 0) {
		*last_lsnp = max_lsn;
		ZERO_LSN(rep->last_lsn);
		ZERO_LSN(max_lsn);
		ret = DB_REP_LOGREADY;
	}
	/*
	 * Only decrement if we were actually applying log records.
	 * We do not care if we processed a dup record or put one
	 * in the temp db.
	 */
	if (set_apply) {
		rep->apply_th--;
		RPRINT(dbenv, (dbenv,
		    "rep_apply: Decrement apply_th %d", rep->apply_th));
	}

	if (ret == 0 && !F_ISSET(rep, REP_F_RECOVER_LOG) &&
	    !IS_ZERO_LSN(max_lsn)) {
		if (ret_lsnp != NULL)
			*ret_lsnp = max_lsn;
		ret = DB_REP_ISPERM;
		DB_ASSERT(dbenv, LOG_COMPARE(&max_lsn, &lp->max_perm_lsn) >= 0);
		lp->max_perm_lsn = max_lsn;
	}

	/*
	 * Start-up is complete when we process (or have already processed) up
	 * to the end of the replication group's log.  In case we miss that
	 * message, as a back-up, we also recognize start-up completion when we
	 * actually process a live log record.  Having cmp==0 here (with a good
	 * "ret" value) implies we actually processed the record.
	 */
	if ((ret == 0 || ret == DB_REP_ISPERM) &&
	    rep->stat.st_startup_complete == 0 &&
	    ((cmp <= 0 && F_ISSET(rp, REPCTL_LOG_END)) ||
	    (cmp == 0 && !F_ISSET(rp, REPCTL_RESEND)))) {
		rep->stat.st_startup_complete = 1;
		event = 1;
		gen = rep->gen;
		master = rep->master_id;
	}
	REP_SYSTEM_UNLOCK(dbenv);
	/*
	 * If we've processed beyond the needed LSN for a pending
	 * start sync, start it now.  We can compare >= here
	 * because ready_lsn is the next record we expect.
	 * Since ckp_lsn can point to the last commit record itself,
	 * but if it does and ready_lsn == commit (i.e. we haven't
	 * written the commit yet), we can still start to sync
	 * because we're guaranteed no additional buffers can
	 * be dirtied.
	 */
	if (!IS_ZERO_LSN(rep->ckp_lsn) &&
	    LOG_COMPARE(&lp->ready_lsn, &rep->ckp_lsn) >= 0) {
		save_lsn = rep->ckp_lsn;
		ZERO_LSN(rep->ckp_lsn);
	} else
		ZERO_LSN(save_lsn);

	/*
	 * If this is a perm record, we are using leases, update the lease
	 * grant.  We must hold the clientdb mutex.  We must not hold
	 * the region mutex because rep_update_grant will acquire it.
	 */
	if (ret == DB_REP_ISPERM && IS_USING_LEASES(dbenv) &&
	    timespecisset(&max_ts)) {
		if ((ret = __rep_update_grant(dbenv, &max_ts)) != 0)
			return (ret);
		if (timespeccmp(&max_ts, &lp->max_lease_ts, >))
			lp->max_lease_ts = max_ts;
	}

	MUTEX_UNLOCK(dbenv, rep->mtx_clientdb);
	if (!IS_ZERO_LSN(save_lsn)) {
		/*
		 * Now call memp_sync holding only the ckp mutex.
		 */
		MUTEX_LOCK(dbenv, rep->mtx_ckp);
		RPRINT(dbenv, (dbenv,
		    "Starting delayed __memp_sync call [%lu][%lu]",
		    (u_long)save_lsn.file, (u_long)save_lsn.offset));
		ret = __memp_sync(dbenv, DB_SYNC_CHECKPOINT, &save_lsn);
		MUTEX_UNLOCK(dbenv, rep->mtx_ckp);
	}
	if (event) {
		RPRINT(dbenv, (dbenv,
		    "Start-up is done [%lu][%lu]",
		    (u_long)rp->lsn.file, (u_long)rp->lsn.offset));

		if ((t_ret = __rep_fire_startupdone(dbenv, gen, master)) != 0) {
			DB_ASSERT(dbenv, ret == 0 || ret == DB_REP_ISPERM);
			/* Failure trumps either of those values. */
			ret = t_ret;
			goto out;
		}
	}
	if (ret == 0 && rp->rectype == REP_NEWFILE && lp->db_log_autoremove)
		__log_autoremove(dbenv);
	if (control_dbt.data != NULL)
		__os_ufree(dbenv, control_dbt.data);
	if (rec_dbt.data != NULL)
		__os_ufree(dbenv, rec_dbt.data);

out:
	switch (ret) {
	case 0:
		break;
	case DB_REP_ISPERM:
		RPRINT(dbenv, (dbenv, "Returning ISPERM [%lu][%lu], cmp = %d",
		    (u_long)max_lsn.file, (u_long)max_lsn.offset, cmp));
		break;
	case DB_REP_LOGREADY:
		RPRINT(dbenv, (dbenv,
		    "Returning LOGREADY up to [%lu][%lu], cmp = %d",
		    (u_long)last_lsnp->file,
		    (u_long)last_lsnp->offset, cmp));
		break;
	case DB_REP_NOTPERM:
		if (!F_ISSET(rep, REP_F_RECOVER_LOG) &&
		    !IS_ZERO_LSN(max_lsn) && ret_lsnp != NULL)
			*ret_lsnp = max_lsn;

		RPRINT(dbenv, (dbenv, "Returning NOTPERM [%lu][%lu], cmp = %d",
		    (u_long)max_lsn.file, (u_long)max_lsn.offset, cmp));
		break;
	default:
		RPRINT(dbenv, (dbenv, "Returning %d [%lu][%lu], cmp = %d", ret,
		    (u_long)max_lsn.file, (u_long)max_lsn.offset, cmp));
		break;
	}

	return (ret);
}

/*
 * __rep_process_txn --
 *
 * This is the routine that actually gets a transaction ready for
 * processing.
 *
 * PUBLIC: int __rep_process_txn __P((DB_ENV *, DBT *));
 */
int
__rep_process_txn(dbenv, rec)
	DB_ENV *dbenv;
	DBT *rec;
{
	DBT data_dbt, *lock_dbt;
	DB_LOCKER *locker;
	DB_LOCKREQ req, *lvp;
	DB_LOGC *logc;
	DB_LSN prev_lsn, *lsnp;
	DB_REP *db_rep;
	DB_TXNHEAD *txninfo;
	LSN_COLLECTION lc;
	REP *rep;
	__txn_regop_args *txn_args;
	__txn_regop_42_args *txn42_args;
	__txn_xa_regop_args *prep_args;
	u_int32_t rectype;
	u_int i;
	int ret, t_ret;

	db_rep = dbenv->rep_handle;
	rep = db_rep->region;
	logc = NULL;
	txn_args = NULL;
	txn42_args = NULL;
	prep_args = NULL;
	txninfo = NULL;

	memset(&data_dbt, 0, sizeof(data_dbt));
	if (F_ISSET(dbenv, DB_ENV_THREAD))
		F_SET(&data_dbt, DB_DBT_REALLOC);

	/*
	 * There are two phases:  First, we have to traverse backwards through
	 * the log records gathering the list of all LSNs in the transaction.
	 * Once we have this information, we can loop through and then apply it.
	 *
	 * We may be passed a prepare (if we're restoring a prepare on upgrade)
	 * instead of a commit (the common case).  Check which it is and behave
	 * appropriately.
	 */
	memcpy(&rectype, rec->data, sizeof(rectype));
	memset(&lc, 0, sizeof(lc));
	if (rectype == DB___txn_regop) {
		/*
		 * We're the end of a transaction.  Make sure this is
		 * really a commit and not an abort!
		 */
		if (rep->version >= DB_REPVERSION_44) {
			if ((ret = __txn_regop_read(dbenv, rec->data,
			    &txn_args)) != 0)
				return (ret);
			if (txn_args->opcode != TXN_COMMIT) {
				__os_free(dbenv, txn_args);
				return (0);
			}
			prev_lsn = txn_args->prev_lsn;
			lock_dbt = &txn_args->locks;
		} else {
			if ((ret = __txn_regop_42_read(dbenv, rec->data,
			    &txn42_args)) != 0)
				return (ret);
			if (txn42_args->opcode != TXN_COMMIT) {
				__os_free(dbenv, txn42_args);
				return (0);
			}
			prev_lsn = txn42_args->prev_lsn;
			lock_dbt = &txn42_args->locks;
		}
	} else {
		/* We're a prepare. */
		DB_ASSERT(dbenv, rectype == DB___txn_xa_regop);

		if ((ret =
		    __txn_xa_regop_read(dbenv, rec->data, &prep_args)) != 0)
			return (ret);
		prev_lsn = prep_args->prev_lsn;
		lock_dbt = &prep_args->locks;
	}

	/* Get locks. */
	if ((ret = __lock_id(dbenv, NULL, &locker)) != 0)
		goto err1;

	if ((ret =
	      __lock_get_list(dbenv, locker, 0, DB_LOCK_WRITE, lock_dbt)) != 0)
		goto err;

	/* Phase 1.  Get a list of the LSNs in this transaction, and sort it. */
	if ((ret = __rep_collect_txn(dbenv, &prev_lsn, &lc)) != 0)
		goto err;
	qsort(lc.array, lc.nlsns, sizeof(DB_LSN), __rep_lsn_cmp);

	/*
	 * The set of records for a transaction may include dbreg_register
	 * records.  Create a txnlist so that they can keep track of file
	 * state between records.
	 */
	if ((ret = __db_txnlist_init(dbenv, 0, 0, NULL, &txninfo)) != 0)
		goto err;

	/* Phase 2: Apply updates. */
	if ((ret = __log_cursor(dbenv, &logc)) != 0)
		goto err;
	for (lsnp = &lc.array[0], i = 0; i < lc.nlsns; i++, lsnp++) {
		if ((ret = __logc_get(logc, lsnp, &data_dbt, DB_SET)) != 0) {
			__db_errx(dbenv, "failed to read the log at [%lu][%lu]",
			    (u_long)lsnp->file, (u_long)lsnp->offset);
			goto err;
		}
		if ((ret = __db_dispatch(dbenv, dbenv->recover_dtab,
		    dbenv->recover_dtab_size, &data_dbt, lsnp,
		    DB_TXN_APPLY, txninfo)) != 0) {
			__db_errx(dbenv, "transaction failed at [%lu][%lu]",
			    (u_long)lsnp->file, (u_long)lsnp->offset);
			goto err;
		}
	}

err:	memset(&req, 0, sizeof(req));
	req.op = DB_LOCK_PUT_ALL;
	if ((t_ret =
	     __lock_vec(dbenv, locker, 0, &req, 1, &lvp)) != 0 && ret == 0)
		ret = t_ret;

	if ((t_ret = __lock_id_free(dbenv, locker)) != 0 && ret == 0)
		ret = t_ret;

err1:	if (txn_args != NULL)
		__os_free(dbenv, txn_args);
	if (txn42_args != NULL)
		__os_free(dbenv, txn42_args);
	if (prep_args != NULL)
		__os_free(dbenv, prep_args);
	if (lc.array != NULL)
		__os_free(dbenv, lc.array);

	if (logc != NULL && (t_ret = __logc_close(logc)) != 0 && ret == 0)
		ret = t_ret;

	if (txninfo != NULL)
		__db_txnlist_end(dbenv, txninfo);

	if (F_ISSET(&data_dbt, DB_DBT_REALLOC) && data_dbt.data != NULL)
		__os_ufree(dbenv, data_dbt.data);

#ifdef HAVE_STATISTICS
	if (ret == 0)
		/*
		 * We don't hold the rep mutex, and could miscount if we race.
		 */
		rep->stat.st_txns_applied++;
#endif

	return (ret);
}

/*
 * __rep_collect_txn
 *	Recursive function that will let us visit every entry in a transaction
 *	chain including all child transactions so that we can then apply
 *	the entire transaction family at once.
 */
static int
__rep_collect_txn(dbenv, lsnp, lc)
	DB_ENV *dbenv;
	DB_LSN *lsnp;
	LSN_COLLECTION *lc;
{
	__txn_child_args *argp;
	DB_LOGC *logc;
	DB_LSN c_lsn;
	DBT data;
	u_int32_t rectype;
	u_int nalloc;
	int ret, t_ret;

	memset(&data, 0, sizeof(data));
	F_SET(&data, DB_DBT_REALLOC);

	if ((ret = __log_cursor(dbenv, &logc)) != 0)
		return (ret);

	while (!IS_ZERO_LSN(*lsnp) &&
	    (ret = __logc_get(logc, lsnp, &data, DB_SET)) == 0) {
		memcpy(&rectype, data.data, sizeof(rectype));
		if (rectype == DB___txn_child) {
			if ((ret = __txn_child_read(dbenv,
			    data.data, &argp)) != 0)
				goto err;
			c_lsn = argp->c_lsn;
			*lsnp = argp->prev_lsn;
			__os_free(dbenv, argp);
			ret = __rep_collect_txn(dbenv, &c_lsn, lc);
		} else {
			if (lc->nalloc < lc->nlsns + 1) {
				nalloc = lc->nalloc == 0 ? 20 : lc->nalloc * 2;
				if ((ret = __os_realloc(dbenv,
				    nalloc * sizeof(DB_LSN), &lc->array)) != 0)
					goto err;
				lc->nalloc = nalloc;
			}
			lc->array[lc->nlsns++] = *lsnp;

			/*
			 * Explicitly copy the previous lsn.  The record
			 * starts with a u_int32_t record type, a u_int32_t
			 * txn id, and then the DB_LSN (prev_lsn) that we
			 * want.  We copy explicitly because we have no idea
			 * what kind of record this is.
			 */
			memcpy(lsnp, (u_int8_t *)data.data +
			    sizeof(u_int32_t) + sizeof(u_int32_t),
			    sizeof(DB_LSN));
		}

		if (ret != 0)
			goto err;
	}
	if (ret != 0)
		__db_errx(dbenv, "collect failed at: [%lu][%lu]",
		    (u_long)lsnp->file, (u_long)lsnp->offset);

err:	if ((t_ret = __logc_close(logc)) != 0 && ret == 0)
		ret = t_ret;
	if (data.data != NULL)
		__os_ufree(dbenv, data.data);
	return (ret);
}

/*
 * __rep_lsn_cmp --
 *	qsort-type-compatible wrapper for LOG_COMPARE.
 */
static int
__rep_lsn_cmp(lsn1, lsn2)
	const void *lsn1, *lsn2;
{

	return (LOG_COMPARE((DB_LSN *)lsn1, (DB_LSN *)lsn2));
}

/*
 * __rep_newfile --
 *	NEWFILE messages have the LSN of the last record in the previous
 * log file.  When applying a NEWFILE message, make sure we haven't already
 * swapped files.  Assume caller hold mtx_clientdb.
 */
static int
__rep_newfile(dbenv, rc, rec)
	DB_ENV *dbenv;
	REP_CONTROL *rc;
	DBT *rec;
{
	DB_LOG *dblp;
	DB_LSN tmplsn;
	DB_REP *db_rep;
	LOG *lp;
	REP *rep;
	u_int32_t version;
	int ret;

	dblp = dbenv->lg_handle;
	lp = dblp->reginfo.primary;
	db_rep = dbenv->rep_handle;
	rep = db_rep->region;

	/*
	 * If a newfile is already in progress, just ignore.
	 */
	if (F_ISSET(rep, REP_F_NEWFILE))
		return (0);
	if (rc->lsn.file + 1 > lp->ready_lsn.file) {
		if (rec == NULL || rec->size == 0) {
			RPRINT(dbenv, (dbenv,
"Rep_newfile: Old-style NEWFILE msg.  Use control msg log version: %lu",
    (u_long) rc->log_version));
			version = rc->log_version;
		} else
			version = *(u_int32_t *)rec->data;
		RPRINT(dbenv, (dbenv, "Rep_newfile: File %lu vers %lu",
		    (u_long)rc->lsn.file + 1, (u_long)version));

		/*
		 * We drop the mtx_clientdb mutex during
		 * the file operation, and then reacquire it when
		 * we're done.  We avoid colliding with new incoming
		 * log records because lp->ready_lsn is not getting
		 * updated and there is no real log record at this
		 * ready_lsn.  We avoid colliding with a duplicate
		 * NEWFILE message by setting an in-progress flag.
		 */
		REP_SYSTEM_LOCK(dbenv);
		F_SET(rep, REP_F_NEWFILE);
		REP_SYSTEM_UNLOCK(dbenv);
		MUTEX_UNLOCK(dbenv, rep->mtx_clientdb);
		LOG_SYSTEM_LOCK(dbenv);
		ret = __log_newfile(dblp, &tmplsn, 0, version);
		LOG_SYSTEM_UNLOCK(dbenv);
		MUTEX_LOCK(dbenv, rep->mtx_clientdb);
		REP_SYSTEM_LOCK(dbenv);
		F_CLR(rep, REP_F_NEWFILE);
		REP_SYSTEM_UNLOCK(dbenv);
		if (ret == 0)
			lp->ready_lsn = tmplsn;
		return (ret);
	} else
		/* We've already applied this NEWFILE.  Just ignore it. */
		return (0);
}

/*
 * __rep_do_ckp --
 * Perform the memp_sync necessary for this checkpoint without holding the
 * REP->mtx_clientdb.  Callers of this function must hold REP->mtx_clientdb
 * and must not be holding the region mutex.
 */
static int
__rep_do_ckp(dbenv, rec, rp)
	DB_ENV *dbenv;
	DBT *rec;
	REP_CONTROL *rp;
{
	__txn_ckp_args *ckp_args;
	DB_LSN ckp_lsn;
	REP *rep;
	int ret;

	/* Crack the log record and extract the checkpoint LSN. */
	if ((ret = __txn_ckp_read(dbenv, rec->data, &ckp_args)) != 0)
		return (ret);
	ckp_lsn = ckp_args->ckp_lsn;
	__os_free(dbenv, ckp_args);

	rep = dbenv->rep_handle->region;

	MUTEX_UNLOCK(dbenv, rep->mtx_clientdb);
	DB_TEST_WAIT(dbenv, dbenv->test_check);

	/*
	 * Sync the memory pool.
	 *
	 * This is the real PERM lock record/ckp.  We cannot return ISPERM
	 * if we haven't truly completed the checkpoint, so we don't allow
	 * this call to be interrupted.
	 *
	 * We may be overlapping our log record with an in-progress startsync
	 * of this checkpoint; suppress the max_write settings on any running
	 * cache-flush operation so it completes quickly.
	 */
	(void)__memp_set_config(dbenv, DB_MEMP_SUPPRESS_WRITE, 1);
	MUTEX_LOCK(dbenv, rep->mtx_ckp);
	ret = __memp_sync(dbenv, DB_SYNC_CHECKPOINT, &ckp_lsn);
	MUTEX_UNLOCK(dbenv, rep->mtx_ckp);
	(void)__memp_set_config(dbenv, DB_MEMP_SUPPRESS_WRITE, 0);

	/* Update the last_ckp in the txn region. */
	if (ret == 0)
		ret = __txn_updateckp(dbenv, &rp->lsn);
	else {
		__db_errx(dbenv, "Error syncing ckp [%lu][%lu]",
		    (u_long)ckp_lsn.file, (u_long)ckp_lsn.offset);
		ret = __db_panic(dbenv, ret);
	}

	MUTEX_LOCK(dbenv, rep->mtx_clientdb);
	return (ret);
}

/*
 * __rep_remfirst --
 * Remove the first entry from the __db.rep.db
 */
static int
__rep_remfirst(dbenv, cntrl, rec)
	DB_ENV *dbenv;
	DBT *cntrl;
	DBT *rec;
{
	DB *dbp;
	DBC *dbc;
	DB_REP *db_rep;
	int ret, t_ret;

	db_rep = dbenv->rep_handle;
	dbp = db_rep->rep_db;

	if ((ret = __db_cursor(dbp, NULL, &dbc, 0)) != 0)
		return (ret);

	/* The DBTs need to persist through another call. */
	F_SET(cntrl, DB_DBT_REALLOC);
	F_SET(rec, DB_DBT_REALLOC);
	if ((ret = __dbc_get(dbc, cntrl, rec, DB_RMW | DB_FIRST)) == 0)
		ret = __dbc_del(dbc, 0);
	if ((t_ret = __dbc_close(dbc)) != 0 && ret == 0)
		ret = t_ret;

	return (ret);
}

/*
 * __rep_getnext --
 * Get the next record out of the __db.rep.db table.
 */
static int
__rep_getnext(dbenv)
	DB_ENV *dbenv;
{
	DB *dbp;
	DB_REP *db_rep;
	DB_LOG *dblp;
	DBC *dbc;
	DBT lsn_dbt, nextrec_dbt;
	LOG *lp;
	REP_CONTROL *rp;
	int ret, t_ret;

	dblp = dbenv->lg_handle;
	lp = dblp->reginfo.primary;

	db_rep = dbenv->rep_handle;
	dbp = db_rep->rep_db;

	if ((ret = __db_cursor(dbp, NULL, &dbc, 0)) != 0)
		return (ret);

	/*
	 * Update waiting_lsn.  We need to move it
	 * forward to the LSN of the next record
	 * in the queue.
	 *
	 * If the next item in the database is a log
	 * record--the common case--we're not
	 * interested in its contents, just in its LSN.
	 * Optimize by doing a partial get of the data item.
	 */
	memset(&nextrec_dbt, 0, sizeof(nextrec_dbt));
	F_SET(&nextrec_dbt, DB_DBT_PARTIAL);
	nextrec_dbt.ulen = nextrec_dbt.dlen = 0;

	memset(&lsn_dbt, 0, sizeof(lsn_dbt));
	ret = __dbc_get(dbc, &lsn_dbt, &nextrec_dbt, DB_FIRST);
	if (ret != DB_NOTFOUND && ret != 0)
		goto err;

	if (ret == DB_NOTFOUND) {
		ZERO_LSN(lp->waiting_lsn);
		/*
		 * Whether or not the current record is
		 * simple, there's no next one, and
		 * therefore we haven't got anything
		 * else to do right now.  Break out.
		 */
		goto err;
	}
	rp = (REP_CONTROL *)lsn_dbt.data;
	lp->waiting_lsn = rp->lsn;

err:	if ((t_ret = __dbc_close(dbc)) != 0 && ret == 0)
		ret = t_ret;
	return (ret);
}

/*
 * __rep_process_rec --
 *
 * Given a record in 'rp', process it.  In the case of a NEWFILE, that means
 * potentially switching files.  In the case of a checkpoint, it means doing
 * the checkpoint, and in other cases, it means simply writing the record into
 * the log.
 */
static int
__rep_process_rec(dbenv, rp, rec, ret_tsp, ret_lsnp)
	DB_ENV *dbenv;
	REP_CONTROL *rp;
	DBT *rec;
	db_timespec *ret_tsp;
	DB_LSN *ret_lsnp;
{
	DB *dbp;
	DB_REP *db_rep;
	DBT control_dbt, key_dbt, rec_dbt;
	REP *rep;
	u_int32_t rectype, txnid;
	int ret, t_ret;

	db_rep = dbenv->rep_handle;
	rep = db_rep->region;
	dbp = db_rep->rep_db;
	ret = 0;

	if (rp->rectype == REP_NEWFILE) {
		ret = __rep_newfile(dbenv, rp, rec);
		return (0);
	}

	memcpy(&rectype, rec->data, sizeof(rectype));
	memset(&control_dbt, 0, sizeof(control_dbt));
	memset(&rec_dbt, 0, sizeof(rec_dbt));

	/*
	 * We write all records except for checkpoint records here.
	 * All non-checkpoint records need to appear in the log before
	 * we take action upon them (i.e., we enforce write-ahead logging).
	 * However, we can't write the checkpoint record here until the
	 * data buffers are actually written to disk, else we are creating
	 * an invalid log -- one that says all data before a certain point
	 * has been written to disk.
	 *
	 * If two threads are both processing the same checkpoint record
	 * (because, for example, it was resent and the original finally
	 * arrived), we handle that below by checking for the existence of
	 * the log record when we add it to the replication database.
	 *
	 * Any log records that arrive while we are processing the checkpoint
	 * are added to the bookkeeping database because ready_lsn is not yet
	 * updated to point after the checkpoint record.
	 */
	if (rectype != DB___txn_ckp || F_ISSET(rep, REP_F_RECOVER_LOG)) {
		if ((ret = __log_rep_put(dbenv, &rp->lsn, rec, 0)) != 0)
			return (ret);
		STAT(rep->stat.st_log_records++);
		if (F_ISSET(rep, REP_F_RECOVER_LOG)) {
			*ret_lsnp = rp->lsn;
			goto out;
		}
	}

	switch (rectype) {
	case DB___dbreg_register:
		/*
		 * DB opens occur in the context of a transaction, so we can
		 * simply handle them when we process the transaction.  Closes,
		 * however, are not transaction-protected, so we have to handle
		 * them here.
		 *
		 * It should be unsafe for the master to do a close of a file
		 * that was opened in an active transaction, so we should be
		 * guaranteed to get the ordering right.
		 *
		 * !!!
		 * The txn ID is the second 4-byte field of the log record.
		 * We should really be calling __dbreg_register_read() and
		 * working from the __dbreg_register_args structure, but this
		 * is considerably faster and the order of the fields won't
		 * change.
		 */
		memcpy(&txnid, (u_int8_t *)
		    rec->data + sizeof(u_int32_t), sizeof(u_int32_t));
		if (txnid == TXN_INVALID)
			ret = __db_dispatch(dbenv, dbenv->recover_dtab,
			    dbenv->recover_dtab_size, rec, &rp->lsn,
			    DB_TXN_APPLY, NULL);
		break;
	case DB___txn_regop:
		/*
		 * If an application is doing app-specific recovery
		 * and acquires locks while applying a transaction,
		 * it can deadlock.  Any other locks held by this
		 * thread should have been discarded in the
		 * __rep_process_txn error path, so if we simply
		 * retry, we should eventually succeed.
		 */
		do {
			ret = 0;
			if (!F_ISSET(db_rep, DBREP_OPENFILES)) {
				ret = __txn_openfiles(dbenv, NULL, 1);
				F_SET(db_rep, DBREP_OPENFILES);
			}
			if (ret == 0)
				ret = __rep_process_txn(dbenv, rec);
		} while (ret == DB_LOCK_DEADLOCK);

		/* Now flush the log unless we're running TXN_NOSYNC. */
		if (ret == 0 && !F_ISSET(dbenv, DB_ENV_TXN_NOSYNC))
			ret = __log_flush(dbenv, NULL);
		if (ret != 0) {
			__db_errx(dbenv, "Error processing txn [%lu][%lu]",
			    (u_long)rp->lsn.file, (u_long)rp->lsn.offset);
			ret = __db_panic(dbenv, ret);
		}
		break;
	case DB___txn_xa_regop:
		ret = __log_flush(dbenv, NULL);
		/*
		 * Save the biggest prepared LSN we've seen.
		 */
		rep->max_prep_lsn = rp->lsn;
		RPRINT(dbenv, (dbenv, "process_rec: prepare at [%lu][%lu]",
		    (u_long)rep->max_prep_lsn.file,
		    (u_long)rep->max_prep_lsn.offset));
		break;
	case DB___txn_ckp:
		/*
		 * We do not want to hold the REP->mtx_clientdb mutex while
		 * syncing the mpool, so if we get a checkpoint record we are
		 * supposed to process, add it to the __db.rep.db, do the
		 * memp_sync and then go back and process it later, when the
		 * sync has finished.  If this record is already in the table,
		 * then some other thread will process it, so simply return
		 * REP_NOTPERM.
		 */
		memset(&key_dbt, 0, sizeof(key_dbt));
		key_dbt.data = rp;
		key_dbt.size = sizeof(*rp);

		/*
		 * We want to put this record into the tmp DB only if
		 * it doesn't exist, so use DB_NOOVERWRITE.
		 */
		ret = __db_put(dbp, NULL, &key_dbt, rec, DB_NOOVERWRITE);
		if (ret == DB_KEYEXIST) {
			if (ret_lsnp != NULL)
				*ret_lsnp = rp->lsn;
			ret = DB_REP_NOTPERM;
		}
		if (ret != 0)
			break;

		/*
		 * Now, do the checkpoint.  Regardless of
		 * whether the checkpoint succeeds or not,
		 * we need to remove the record we just put
		 * in the temporary database.  If the
		 * checkpoint failed, return an error.  We
		 * will act like we never received the
		 * checkpoint.
		 */
		if ((ret = __rep_do_ckp(dbenv, rec, rp)) == 0)
			ret = __log_rep_put(dbenv, &rp->lsn, rec,
			    DB_LOG_CHKPNT);
		if ((t_ret = __rep_remfirst(dbenv,
		    &control_dbt, &rec_dbt)) != 0 && ret == 0)
			ret = t_ret;
		/*
		 * If we're successful putting the log record in the
		 * log, flush it for a checkpoint.
		 */
		if (ret == 0)
			ret = __log_flush(dbenv, NULL);
		break;
	default:
		break;
	}

out:
	if (ret == 0 && F_ISSET(rp, REPCTL_PERM))
		*ret_lsnp = rp->lsn;
	if (IS_USING_LEASES(dbenv) &&
	    F_ISSET(rp, REPCTL_LEASE))
		*ret_tsp = rp->msg_time;
	/*
	 * Set ret_lsnp before flushing the log because if the
	 * flush fails, we've still written the record to the
	 * log and the LSN has been entered.
	 */
	if (ret == 0 && F_ISSET(rp, REPCTL_FLUSH))
		ret = __log_flush(dbenv, NULL);
	if (control_dbt.data != NULL)
		__os_ufree(dbenv, control_dbt.data);
	if (rec_dbt.data != NULL)
		__os_ufree(dbenv, rec_dbt.data);

	return (ret);
}

/*
 * __rep_resend_req --
 *	We might have dropped a message, we need to resend our request.
 *	The request we send is dependent on what recovery state we're in.
 *	The caller holds no locks.
 *
 * PUBLIC: int __rep_resend_req __P((DB_ENV *, int));
 */
int
__rep_resend_req(dbenv, rereq)
	DB_ENV *dbenv;
	int rereq;
{

	DB_LOG *dblp;
	DB_LSN lsn;
	DB_REP *db_rep;
	LOG *lp;
	REP *rep;
	int ret;
	u_int32_t gapflags, repflags;

	db_rep = dbenv->rep_handle;
	rep = db_rep->region;
	dblp = dbenv->lg_handle;
	lp = dblp->reginfo.primary;
	ret = 0;

	repflags = rep->flags;
	/*
	 * If we are delayed we do not rerequest anything.
	 */
	if (FLD_ISSET(repflags, REP_F_DELAY))
		return (ret);
	gapflags = rereq ? REP_GAP_REREQUEST : 0;

	if (FLD_ISSET(repflags, REP_F_RECOVER_VERIFY)) {
		MUTEX_LOCK(dbenv, rep->mtx_clientdb);
		lsn = lp->verify_lsn;
		MUTEX_UNLOCK(dbenv, rep->mtx_clientdb);
		if (!IS_ZERO_LSN(lsn))
			(void)__rep_send_message(dbenv, rep->master_id,
			    REP_VERIFY_REQ, &lsn, NULL, 0, DB_REP_REREQUEST);
	} else if (FLD_ISSET(repflags, REP_F_RECOVER_UPDATE)) {
		/*
		 * UPDATE_REQ only goes to the master.
		 */
		(void)__rep_send_message(dbenv, rep->master_id,
		    REP_UPDATE_REQ, NULL, NULL, 0, 0);
	} else if (FLD_ISSET(repflags, REP_F_RECOVER_PAGE)) {
		REP_SYSTEM_LOCK(dbenv);
		ret = __rep_pggap_req(dbenv, rep, NULL, gapflags);
		REP_SYSTEM_UNLOCK(dbenv);
	} else {
		MUTEX_LOCK(dbenv, rep->mtx_clientdb);
		ret = __rep_loggap_req(dbenv, rep, NULL, gapflags);
		MUTEX_UNLOCK(dbenv, rep->mtx_clientdb);
	}

	return (ret);
}

/*
 * __rep_check_doreq --
 * PUBLIC: int __rep_check_doreq __P((DB_ENV *, REP *));
 *
 * Check if we need to send another request.  If so, compare with
 * the request limits the user might have set.  This assumes the
 * caller holds the REP->mtx_clientdb mutex.  Returns 1 if a request
 * needs to be made, and 0 if it does not.
 */
int
__rep_check_doreq(dbenv, rep)
	DB_ENV *dbenv;
	REP *rep;
{

	DB_LOG *dblp;
	LOG *lp;
	int req;

	dblp = dbenv->lg_handle;
	lp = dblp->reginfo.primary;
	req = ++lp->rcvd_recs >= lp->wait_recs;
	if (req) {
		lp->wait_recs *= 2;
		if (lp->wait_recs > rep->max_gap)
			lp->wait_recs = rep->max_gap;
		lp->rcvd_recs = 0;
	}
	return (req);
}

/*
 * __rep_skip_msg -
 *
 *	If we're in recovery we want to skip/ignore the message, but
 *	we also need to see if we need to re-request any retransmissions.
 */
static int
__rep_skip_msg(dbenv, rep, eid, rectype)
	DB_ENV *dbenv;
	REP *rep;
	int eid;
	u_int32_t rectype;
{
	int do_req, ret;

	ret = 0;
	/*
	 * If we have a request message from a client then immediately
	 * send a REP_REREQUEST back to that client since we're skipping it.
	 */
	if (F_ISSET(rep, REP_F_CLIENT) && REP_MSG_REQ(rectype))
		do_req = 1;
	else {
		/* Check for need to retransmit. */
		MUTEX_LOCK(dbenv, rep->mtx_clientdb);
		do_req = __rep_check_doreq(dbenv, rep);
		MUTEX_UNLOCK(dbenv, rep->mtx_clientdb);
	}
	/*
	 * Don't respond to a MASTER_REQ with
	 * a MASTER_REQ or REREQUEST.
	 */
	if (do_req && rectype != REP_MASTER_REQ) {
		/*
		 * There are three cases:
		 * 1.  If we don't know who the master is, then send MASTER_REQ.
		 * 2.  If the message we're skipping came from the master,
		 * then we need to rerequest.
		 * 3.  If the message didn't come from a master (i.e. client
		 * to client), then send a rerequest back to the sender so
		 * the sender can rerequest it elsewhere, if we are a client.
		 */
		if (rep->master_id == DB_EID_INVALID)	/* Case 1. */
			(void)__rep_send_message(dbenv,
			    DB_EID_BROADCAST, REP_MASTER_REQ, NULL, NULL, 0, 0);
		else if (eid == rep->master_id)		/* Case 2. */
			ret = __rep_resend_req(dbenv, 0);
		else if (F_ISSET(rep, REP_F_CLIENT))	/* Case 3. */
			(void)__rep_send_message(dbenv,
			    eid, REP_REREQUEST, NULL, NULL, 0, 0);
	}
	return (ret);
}

static int
__rep_fire_newmaster(dbenv, gen, master)
	DB_ENV *dbenv;
	u_int32_t gen;
	int master;
{
	DB_REP *db_rep;
	REP *rep;

	db_rep = dbenv->rep_handle;
	rep = db_rep->region;

	REP_EVENT_LOCK(dbenv);
	/*
	 * The firing of this event should be idempotent with respect to a
	 * particular generation number.
	 */
	if (rep->newmaster_event_gen < gen) {
		__rep_fire_event(dbenv, DB_EVENT_REP_NEWMASTER, &master);
		rep->newmaster_event_gen = gen;
	}
	REP_EVENT_UNLOCK(dbenv);
	return (0);
}

static int
__rep_fire_startupdone(dbenv, gen, master)
	DB_ENV *dbenv;
	u_int32_t gen;
	int master;
{
	DB_REP *db_rep;
	REP *rep;

	db_rep = dbenv->rep_handle;
	rep = db_rep->region;

	REP_EVENT_LOCK(dbenv);
	/*
	 * Usually NEWMASTER will already have been fired.  But if not, fire
	 * it here now, to ensure the application receives events in the
	 * expected order.
	 */
	if (rep->newmaster_event_gen < gen) {
		__rep_fire_event(dbenv, DB_EVENT_REP_NEWMASTER, &master);
		rep->newmaster_event_gen = gen;
	}

	/*
	 * Caller already ensures that it only tries to fire STARTUPDONE once
	 * per generation.  If we did not want to rely on that, we could add a
	 * simple boolean flag (to the set of data protected by the mtx_event).
	 * The precise meaning of that flag would be "STARTUPDONE has been fired
	 * for the generation value stored in `newmaster_event_gen'".  Then the
	 * more accurate test here would be simply to check that flag, and fire
	 * the event (and set the flag) if it were not already set.
	 */
	if (rep->newmaster_event_gen == gen)
		__rep_fire_event(dbenv, DB_EVENT_REP_STARTUPDONE, NULL);
	REP_EVENT_UNLOCK(dbenv);
	return (0);
}