db_dispatch.c

这是国外的resip协议栈

	DB_TXN *txn;
	DB_TXNLIST *elp;
	db_limbo_state state;
{
	DB *dbp;
	DB_MPOOLFILE *mpf;
	DBMETA *meta;
	DB_TXN *ctxn, *t;
	FNAME *fname;
	db_pgno_t last_pgno, pgno;
	int dbp_created, in_retry, ret, t_ret;

	ctxn = NULL;
	in_retry = 0;
	meta = NULL;
	mpf = NULL;
	ret = 0;
	for (; elp != NULL; elp = LIST_NEXT(elp, links)) {
		if (elp->type != TXNLIST_PGNO)
			continue;
retry:		dbp_created = 0;

		/*
		 * Pick the transaction in which to potentially
		 * log compensations.
		 */
		if (state == LIMBO_PREPARE)
			ctxn = txn;
		else if (!in_retry && state != LIMBO_RECOVER &&
		    state != LIMBO_TIMESTAMP && !T_RESTORED(txn) &&
		    (ret = __txn_compensate_begin(dbenv, &ctxn)) != 0)
			return (ret);

		/*
		 * Either use the compensating transaction or
		 * the one passed in, which will be null if recovering.
		 */
		t = ctxn == NULL ? txn : ctxn;

		/* First try to get a dbp by fileid. */
		ret = __dbreg_id_to_db(dbenv, t, &dbp, elp->u.p.fileid, 0);

		/*
		 * If the file was closed and reopened its id could change.
		 * Look it up the hard way.
		 */
		if (ret == DB_DELETED || ret == ENOENT ||
		    ((ret == 0 &&
		    memcmp(elp->u.p.uid, dbp->fileid, DB_FILE_ID_LEN) != 0))) {
			if ((ret = __dbreg_fid_to_fname(
			    dbenv->lg_handle, elp->u.p.uid, 0, &fname)) == 0)
				ret = __dbreg_id_to_db(
				     dbenv, t, &dbp, fname->id, 0);
		}
		/*
		 * File is being destroyed.  No need to worry about
		 * dealing with recovery of allocations.
		 */
		if (ret == DB_DELETED ||
		    (ret == 0 && F_ISSET(dbp, DB_AM_DISCARD)))
			goto next;

		if (ret != 0) {
			if ((ret = db_create(&dbp, dbenv, 0)) != 0)
				goto err;

			/*
			 * This tells the system not to lock, which is always
			 * OK, whether this is an abort or recovery.
			 */
			F_SET(dbp, DB_AM_COMPENSATE);
			dbp_created = 1;

			/* It is ok if the file is nolonger there. */
			ret = __db_open(dbp,
			    t, elp->u.p.fname, NULL, DB_UNKNOWN,
			    DB_ODDFILESIZE, __db_omode("rw----"), PGNO_BASE_MD);
			if (ret == ENOENT)
				goto next;
		}

		/*
		 * Verify that we are opening the same file that we were
		 * referring to when we wrote this log record.
		 */
		if (memcmp(elp->u.p.uid, dbp->fileid, DB_FILE_ID_LEN) != 0)
			goto next;

		mpf = dbp->mpf;
		last_pgno = PGNO_INVALID;

		if (meta == NULL && 
		    (ctxn == NULL || state == LIMBO_COMPENSATE)) {
			pgno = PGNO_BASE_MD;
			if ((ret = __memp_fget(mpf, &pgno, 0, &meta)) != 0)
				goto err;
			last_pgno = meta->free;
		}

		if (state == LIMBO_PREPARE) {
			if ((ret = __db_limbo_prepare(dbp, ctxn, elp)) != 0)
				goto err;
		} else
			ret = __db_limbo_fix(dbp,
			     ctxn, elp, &last_pgno, meta, state);
		/*
		 * If we were doing compensating transactions, then we are
		 * going to hope this error was due to running out of space.
		 * We'll change modes (into the sync the file mode) and keep
		 * trying.  If we weren't doing compensating transactions,
		 * then this is a real error and we're sunk.
		 */
		if (ret != 0) {
			if (ret == DB_RUNRECOVERY || ctxn == NULL)
				goto err;
			in_retry = 1;
			if ((ret = __txn_abort(ctxn)) != 0)
				goto err;
			ctxn = NULL;
			goto retry;
		}

		if (state == LIMBO_PREPARE)
			ctxn = NULL;

		else if (ctxn != NULL) {
			/*
			 * We only force compensation at the end of recovery.
			 * We want the txn_commit to be logged so turn
			 * off the recovery flag briefly.
			 */
			if (state == LIMBO_COMPENSATE)
				F_CLR(
				    (DB_LOG *)dbenv->lg_handle, DBLOG_RECOVER);
			ret = __txn_commit(ctxn, DB_TXN_NOSYNC);
			ctxn = NULL;
			if (state == LIMBO_COMPENSATE)
				F_SET(
				    (DB_LOG *)dbenv->lg_handle, DBLOG_RECOVER);
			if (ret != 0)
				goto retry;
		}

		/*
		 * This is where we handle the case where we're explicitly
		 * putting together a free list.  We need to decide whether
		 * we have to write the meta-data page, and if we do, then
		 * we need to sync it as well.
		 */
		else if (last_pgno == meta->free) {
			/* No change to page; just put the page back. */
			if ((ret = __memp_fput(mpf, meta, 0)) != 0)
				goto err;
			meta = NULL;
		} else {
			/*
			 * These changes are unlogged so we cannot have the
			 * metapage pointing at pages that are not on disk.
			 * Therefore, we flush the new free list, then update
			 * the metapage.  We have to put the meta-data page
			 * first so that it isn't pinned when we try to sync.
			 */
			if (!IS_RECOVERING(dbenv) && !T_RESTORED(txn))
				__db_err(dbenv, "Flushing free list to disk");
			if ((ret = __memp_fput(mpf, meta, 0)) != 0)
				goto err;
			meta = NULL;
			/*
			 * If the sync fails then we cannot flush the
			 * newly allocated pages.  That is, the file
			 * cannot be extended. Don't let the metapage
			 * point at them.
			 * We may lose these pages from the file if it
			 * can be extended later.  If there is never
			 * space for the pages, then things will be ok.
			 */
			if ((ret = __db_sync(dbp)) == 0) {
				pgno = PGNO_BASE_MD;
				if ((ret =
				    __memp_fget(mpf, &pgno, 0, &meta)) != 0)
					goto err;
				meta->free = last_pgno;
				if ((ret = __memp_fput(mpf,
				     meta, DB_MPOOL_DIRTY)) != 0)
					goto err;
				meta = NULL;
			} else {
				__db_err(dbenv,
				    "%s: %s", dbp->fname, db_strerror(ret));
				__db_err(dbenv, "%s: %s %s", dbp->fname,
				    "allocation flush failed, some free pages",
				    "may not appear in the free list");
				ret = 0;
			}
		}

next:
		/*
		 * If we get here, either we have processed the list
		 * or the db file has been deleted or could not be opened.
		 */
		if (ctxn != NULL &&
		    (t_ret = __txn_abort(ctxn)) != 0 && ret == 0)
			ret = t_ret;

		if (dbp_created &&
		    (t_ret = __db_close(dbp, txn, DB_NOSYNC)) != 0 && ret == 0)
			ret = t_ret;
		dbp = NULL;
		if (state != LIMBO_PREPARE && state != LIMBO_TIMESTAMP) {
			__os_free(dbenv, elp->u.p.fname);
			__os_free(dbenv, elp->u.p.pgno_array);
		}
		if (ret == ENOENT)
			ret = 0;
		else if (ret != 0)
			goto err;
	}

err:	if (meta != NULL)
		(void)__memp_fput(mpf, meta, 0);
	return (ret);
}

/*
 * __db_limbo_fix --
 *	Process a single limbo entry which describes all the page allocations
 * for a single file.
 */
static int
__db_limbo_fix(dbp, ctxn, elp, lastp, meta, state)
	DB *dbp;
	DB_TXN *ctxn;
	DB_TXNLIST *elp;
	db_pgno_t *lastp;
	DBMETA *meta;
	db_limbo_state state;
{
	DBC *dbc;
	DBT ldbt;
	DB_MPOOLFILE *mpf;
	PAGE *freep, *pagep;
	db_pgno_t next, pgno;
	u_int32_t i;
	int ret, t_ret;

	/*
	 * Loop through the entries for this txnlist element and
	 * either link them into the free list or write a compensating
	 * record for each.
	 */
	dbc = NULL;
	mpf = dbp->mpf;
	pagep = NULL;
	ret = 0;

	for (i = 0; i < elp->u.p.nentries; i++) {
		pgno = elp->u.p.pgno_array[i];

		if (pgno == PGNO_INVALID)
			continue;

		if ((ret =
		    __memp_fget(mpf, &pgno, DB_MPOOL_CREATE, &pagep)) != 0) {
			if (ret != ENOSPC)
				goto err;
			continue;
		}

		if (state == LIMBO_COMPENSATE || IS_ZERO_LSN(LSN(pagep))) {
			if (ctxn == NULL) {
				/*
				 * If this is a fatal recovery which
				 * spans a previous crash this page may
				 * be on the free list already.
				 */
				for (next = *lastp; next != 0; ) {
					if (next == pgno)
						break;
					if ((ret = __memp_fget(mpf,
					    &next, 0, &freep)) != 0)
						goto err;
					next = NEXT_PGNO(freep);
					if ((ret =
					    __memp_fput(mpf, freep, 0)) != 0)
						goto err;
				}

				if (next != pgno) {
					P_INIT(pagep, dbp->pgsize, pgno,
					    PGNO_INVALID, *lastp, 0, P_INVALID);
					/* Make the lsn non-zero but generic. */
					INIT_LSN(LSN(pagep));
					*lastp = pgno;
				}
			} else if (state == LIMBO_COMPENSATE) {
				/*
				 * Generate a log record for what we did on the
				 * LIMBO_TIMESTAMP pass.  All pages here are
				 * free so P_OVERHEAD is sufficient.
				 */
				ZERO_LSN(pagep->lsn);
				memset(&ldbt, 0, sizeof(ldbt));
				ldbt.data = pagep;
				ldbt.size = P_OVERHEAD(dbp);
				if ((ret = __db_pg_new_log(dbp, ctxn,
				     &LSN(meta), 0, pagep->pgno,
				     &LSN(meta), PGNO_BASE_MD,
				     &ldbt, pagep->next_pgno)) != 0)
					goto err;
			} else {
				if (dbc == NULL && (ret =
				    __db_cursor(dbp, ctxn, &dbc, 0)) != 0)
						goto err;
				/*
				 * If the dbp is compensating (because we
				 * opened it), the dbc will automatically be
				 * marked compensating, but in case we didn't
				 * do the open, we have to mark it explicitly.
				 */
				F_SET(dbc, DBC_COMPENSATE);
				ret = __db_free(dbc, pagep);
				pagep = NULL;

				/*
				 * On any error, we hope that the error was
				 * caused due to running out of space, and we
				 * switch modes, doing the processing where we
				 * sync out files instead of doing compensating
				 * transactions.  If this was a real error and
				 * not out of space, we assume that some other
				 * call will fail real soon.
				 */
				if (ret != 0) {
					/* Assume that this is out of space. */
					(void)__db_c_close(dbc);
					dbc = NULL;
					goto err;
				}
			}
		}
		else
			elp->u.p.pgno_array[i] = PGNO_INVALID;

		if (pagep != NULL) {
			ret = __memp_fput(mpf, pagep, DB_MPOOL_DIRTY);
			pagep = NULL;
		}
		if (ret != 0)
			goto err;
	}

err:	if (pagep != NULL &&
	    (t_ret = __memp_fput(mpf, pagep, DB_MPOOL_DIRTY)) != 0 && ret == 0)
		ret = t_ret;
	if (dbc != NULL && (t_ret = __db_c_close(dbc)) != 0 && ret == 0)
		ret = t_ret;
	return (ret);
}

static int
__db_limbo_prepare(dbp, txn, elp)
	DB *dbp;
	DB_TXN *txn;
	DB_TXNLIST *elp;
{
	DB_LSN lsn;
	DB_MPOOLFILE *mpf;
	PAGE *pagep;
	db_pgno_t pgno;
	u_int32_t i;
	int ret, t_ret;

	/*
	 * Loop through the entries for this txnlist element and
	 * output a prepare record for them.
	 */
	pagep = NULL;
	ret = 0;
	mpf = dbp->mpf;

	for (i = 0; i < elp->u.p.nentries; i++) {
		pgno = elp->u.p.pgno_array[i];

		if ((ret =
		    __memp_fget(mpf, &pgno, DB_MPOOL_CREATE, &pagep)) != 0) {
			if (ret != ENOSPC)
				return (ret);
			continue;
		}

		if (IS_ZERO_LSN(LSN(pagep)))
			ret = __db_pg_prepare_log(dbp, txn, &lsn, 0, pgno);

		if ((t_ret = __memp_fput(mpf, pagep, 0)) != 0 && ret == 0)
			ret = t_ret;

		if (ret != 0)
			return (ret);
	}

	return (0);
}

#define	DB_TXNLIST_MAX_PGNO	8		/* A nice even number. */

/*
 * __db_txnlist_pgnoadd --
 *	Find the txnlist entry for a file and add this pgno, or add the list
 *	entry for the file and then add the pgno.
 */
static int
__db_txnlist_pgnoadd(dbenv, hp, fileid, uid, fname, pgno)
	DB_ENV *dbenv;
	DB_TXNHEAD *hp;
	int32_t fileid;
	u_int8_t uid[DB_FILE_ID_LEN];
	char *fname;
	db_pgno_t pgno;
{
	DB_TXNLIST *elp;
	size_t len;
	u_int32_t hash, status;
	int ret;

	elp = NULL;

	if ((ret = __db_txnlist_find_internal(dbenv, hp,
	    TXNLIST_PGNO, 0, uid, &elp, 0, &status)) != 0 && ret != DB_NOTFOUND)
		goto err;

	if (ret == DB_NOTFOUND || status != TXN_OK) {
		if ((ret =
		    __os_malloc(dbenv, sizeof(DB_TXNLIST), &elp)) != 0)
			goto err;
		memcpy(&hash, uid, sizeof(hash));
		LIST_INSERT_HEAD(
		    &hp->head[DB_TXNLIST_MASK(hp, hash)], elp, links);
		memcpy(elp->u.p.uid, uid, DB_FILE_ID_LEN);

		len = strlen(fname) + 1;
		if ((ret = __os_malloc(dbenv, len, &elp->u.p.fname)) != 0)
			goto err;
		memcpy(elp->u.p.fname, fname, len);

		elp->u.p.maxentry = 0;
		elp->u.p.locked = 0;
		elp->type = TXNLIST_PGNO;
		if ((ret = __os_malloc(dbenv,
		    8 * sizeof(db_pgno_t), &elp->u.p.pgno_array)) != 0)
			goto err;
		elp->u.p.maxentry = DB_TXNLIST_MAX_PGNO;
		elp->u.p.nentries = 0;
	} else if (elp->u.p.nentries == elp->u.p.maxentry) {
		elp->u.p.maxentry <<= 1;
		if ((ret = __os_realloc(dbenv, elp->u.p.maxentry *
		    sizeof(db_pgno_t), &elp->u.p.pgno_array)) != 0)
			goto err;
	}

	elp->u.p.pgno_array[elp->u.p.nentries++] = pgno;
	/* Update to the latest fileid.  Limbo will find it faster. */
	elp->u.p.fileid = fileid;

	return (0);

err:	return (ret);
}
#endif

#ifdef DEBUG
/*
 * __db_txnlist_print --
 *	Print out the transaction list.
 *
 * PUBLIC: void __db_txnlist_print __P((void *));
 */
void
__db_txnlist_print(listp)
	void *listp;
{
	DB_TXNHEAD *hp;
	DB_TXNLIST *p;
	u_int32_t i;
	char *txntype;

	hp = (DB_TXNHEAD *)listp;

	printf("Maxid: %lu Generation: %lu\n",
	    (u_long)hp->maxid, (u_long)hp->generation);
	for (i = 0; i < hp->nslots; i++)
		for (p = LIST_FIRST(&hp->head[i]);
		    p != NULL; p = LIST_NEXT(p, links)) {
			if (p->type != TXNLIST_TXNID) {
				printf("Unrecognized type: %d\n", p->type);
				continue;
			}
			switch (p->u.t.status) {
			case TXN_OK:
				txntype = "OK";
				break;
			case TXN_COMMIT:
				txntype = "commit";
				break;
			case TXN_PREPARE:
				txntype = "prepare";
				break;
			case TXN_ABORT:
				txntype = "abort";
				break;
			case TXN_IGNORE:
				txntype = "ignore";
				break;
			case TXN_EXPECTED:
				txntype = "expected";
				break;
			case TXN_UNEXPECTED:
				txntype = "unexpected";
				break;
			default:
				txntype = "UNKNOWN";
				break;
			}
			printf("TXNID: %lx(%lu): %s\n",
			    (u_long)p->u.t.txnid,
			    (u_long)p->u.t.generation, txntype);
		}
}
#endif