db_cam.c

这是国外的resip协议栈

	 *
	 * (XXX: It would be nice to avoid this extra get, and have the
	 * underlying put routines somehow pass us the old record
	 * since they need to traverse the tree anyway.  I'm saving
	 * this optimization for later, as it's a lot of work, and it
	 * would be hard to fit into this locking paradigm anyway.)
	 *
	 * The simple thing to do would be to go get the old record before
	 * we do anything else.  Unfortunately, though, doing so would
	 * violate our "secondary, then primary" lock acquisition
	 * ordering--even in the common case where no old primary record
	 * exists, we'll still acquire and keep a lock on the page where
	 * we're about to do the primary insert.
	 *
	 * To get around this, we do the following gyrations, which
	 * hopefully solve this problem in the common case:
	 *
	 * 1) If this is a c_put(DB_CURRENT), go ahead and get the
	 *    old record.  We already hold the lock on this page in
	 *    the primary, so no harm done, and we'll need the primary
	 *    key (which we weren't passed in this case) to do any
	 *    secondary puts anyway.
	 *
	 * 2) If we're doing a partial put, we need to perform the
	 *    get on the primary key right away, since we don't have
	 *    the whole datum that the secondary key is based on.
	 *    We may also need to pad out the record if the primary
	 *    has a fixed record length.
	 *
	 * 3) Loop through the secondary indices, putting into each a
	 *    new secondary key that corresponds to the new record.
	 *
	 * 4) If we haven't done so in (1) or (2), get the old primary
	 *    key/data pair.  If one does not exist--the common case--we're
	 *    done with secondary indices, and can go straight on to the
	 *    primary put.
	 *
	 * 5) If we do have an old primary key/data pair, however, we need
	 *    to loop through all the secondaries a second time and delete
	 *    the old secondary in each.
	 */
	memset(&pkey, 0, sizeof(DBT));
	have_oldrec = nodel = 0;

	/*
	 * Primary indices can't have duplicates, so only DB_CURRENT,
	 * DB_KEYFIRST, and DB_KEYLAST make any sense.  Other flags
	 * should have been caught by the checking routine, but
	 * add a sprinkling of paranoia.
	 */
	DB_ASSERT(flags == DB_CURRENT ||
	    flags == DB_KEYFIRST || flags == DB_KEYLAST);

	/*
	 * We'll want to use DB_RMW in a few places, but it's only legal
	 * when locking is on.
	 */
	rmw = STD_LOCKING(dbc_arg) ? DB_RMW : 0;

	if (flags == DB_CURRENT) {		/* Step 1. */
		/*
		 * This is safe to do on the cursor we already have;
		 * error or no, it won't move.
		 *
		 * We use DB_RMW for all of these gets because we'll be
		 * writing soon enough in the "normal" put code.  In
		 * transactional databases we'll hold those write locks
		 * even if we close the cursor we're reading with.
		 *
		 * The DB_KEYEMPTY return needs special handling -- if the
		 * cursor is on a deleted key, we return DB_NOTFOUND.
		 */
		ret = __db_c_get(dbc_arg, &pkey, &olddata, rmw | DB_CURRENT);
		if (ret == DB_KEYEMPTY)
			ret = DB_NOTFOUND;
		if (ret != 0)
			goto err;

		have_oldrec = 1; /* We've looked for the old record. */
	} else {
		/*
		 * Set pkey so we can use &pkey everywhere instead of key.
		 * If DB_CURRENT is set and there is a key at the current
		 * location, pkey will be overwritten before it's used.
		 */
		pkey.data = key->data;
		pkey.size = key->size;
	}

	/*
	 * Check for partial puts (step 2).
	 */
	if (F_ISSET(data, DB_DBT_PARTIAL)) {
		if (!have_oldrec && !nodel) {
			/*
			 * We're going to have to search the tree for the
			 * specified key.  Dup a cursor (so we have the same
			 * locking info) and do a c_get.
			 */
			if ((ret = __db_c_idup(dbc_arg, &pdbc, 0)) != 0)
				goto err;

			/* We should have gotten DB_CURRENT in step 1. */
			DB_ASSERT(flags != DB_CURRENT);

			ret = __db_c_get(pdbc, &pkey, &olddata, rmw | DB_SET);
			if (ret == DB_KEYEMPTY || ret == DB_NOTFOUND) {
				nodel = 1;
				ret = 0;
			}
			if ((t_ret = __db_c_close(pdbc)) != 0)
				ret = t_ret;
			if (ret != 0)
				goto err;

			have_oldrec = 1;
		}

		/*
		 * Now build the new datum from olddata and the partial
		 * data we were given.
		 */
		if ((ret =
		    __db_buildpartial(dbp, &olddata, data, &newdata)) != 0)
			goto err;
		ispartial = 1;
	} else
		ispartial = 0;

	/*
	 * Handle fixed-length records.  If the primary database has
	 * fixed-length records, we need to pad out the datum before
	 * we pass it into the callback function;  we always index the
	 * "real" record.
	 */
	if ((dbp->type == DB_RECNO && F_ISSET(dbp, DB_AM_FIXEDLEN)) ||
	    (dbp->type == DB_QUEUE)) {
		if (dbp->type == DB_QUEUE) {
			re_len = ((QUEUE *)dbp->q_internal)->re_len;
			re_pad = ((QUEUE *)dbp->q_internal)->re_pad;
		} else {
			re_len = ((BTREE *)dbp->bt_internal)->re_len;
			re_pad = ((BTREE *)dbp->bt_internal)->re_pad;
		}

		size = ispartial ? newdata.size : data->size;
		if (size > re_len) {
			ret = __db_rec_toobig(dbenv, size, re_len);
			goto err;
		} else if (size < re_len) {
			/*
			 * If we're not doing a partial put, copy
			 * data->data into newdata.data, then pad out
			 * newdata.data.
			 *
			 * If we're doing a partial put, the data
			 * we want are already in newdata.data;  we
			 * just need to pad.
			 *
			 * Either way, realloc is safe.
			 */
			if ((ret =
			    __os_realloc(dbenv, re_len, &newdata.data)) != 0)
				goto err;
			if (!ispartial)
				memcpy(newdata.data, data->data, size);
			memset((u_int8_t *)newdata.data + size, re_pad,
			    re_len - size);
			newdata.size = re_len;
			ispartial = 1;
		}
	}

	/*
	 * Loop through the secondaries.  (Step 3.)
	 *
	 * Note that __db_s_first and __db_s_next will take care of
	 * thread-locking and refcounting issues.
	 */
	for (sdbp = __db_s_first(dbp);
	    sdbp != NULL && ret == 0; ret = __db_s_next(&sdbp)) {
		/*
		 * Call the callback for this secondary, to get the
		 * appropriate secondary key.
		 */
		memset(&skey, 0, sizeof(DBT));
		if ((ret = sdbp->s_callback(sdbp,
		    &pkey, ispartial ? &newdata : data, &skey)) != 0) {
			if (ret == DB_DONOTINDEX)
				/*
				 * The callback returned a null value--don't
				 * put this key in the secondary.  Just
				 * move on to the next one--we'll handle
				 * any necessary deletes in step 5.
				 */
				continue;
			else
				goto err;
		}

		/*
		 * Open a cursor in this secondary.
		 *
		 * Use the same locker ID as our primary cursor, so that
		 * we're guaranteed that the locks don't conflict (e.g. in CDB
		 * or if we're subdatabases that share and want to lock a
		 * metadata page).
		 */
		if ((ret = __db_cursor_int(sdbp, dbc_arg->txn, sdbp->type,
		    PGNO_INVALID, 0, dbc_arg->locker, &sdbc)) != 0)
			goto err;

		/*
		 * If we're in CDB, updates will fail since the new cursor
		 * isn't a writer.  However, we hold the WRITE lock in the
		 * primary and will for as long as our new cursor lasts,
		 * and the primary and secondary share a lock file ID,
		 * so it's safe to consider this a WRITER.  The close
		 * routine won't try to put anything because we don't
		 * really have a lock.
		 */
		if (CDB_LOCKING(dbenv)) {
			DB_ASSERT(sdbc->mylock.off == LOCK_INVALID);
			F_SET(sdbc, DBC_WRITER);
		}

		/*
		 * There are three cases here--
		 * 1) The secondary supports sorted duplicates.
		 *	If we attempt to put a secondary/primary pair
		 *	that already exists, that's a duplicate duplicate,
		 *	and c_put will return DB_KEYEXIST (see __db_duperr).
		 *	This will leave us with exactly one copy of the
		 *	secondary/primary pair, and this is just right--we'll
		 *	avoid deleting it later, as the old and new secondaries
		 *	will match (since the old secondary is the dup dup
		 *	that's already there).
		 * 2) The secondary supports duplicates, but they're not
		 *	sorted.  We need to avoid putting a duplicate
		 *	duplicate, because the matching old and new secondaries
		 *	will prevent us from deleting anything and we'll
		 *	wind up with two secondary records that point to the
		 *	same primary key.  Do a c_get(DB_GET_BOTH);  only
		 *	do the put if the secondary doesn't exist.
		 * 3) The secondary doesn't support duplicates at all.
		 *	In this case, secondary keys must be unique;  if
		 *	another primary key already exists for this
		 *	secondary key, we have to either overwrite it or
		 *	not put this one, and in either case we've
		 *	corrupted the secondary index.  Do a c_get(DB_SET).
		 *	If the secondary/primary pair already exists, do
		 *	nothing;  if the secondary exists with a different
		 *	primary, return an error;  and if the secondary
		 *	does not exist, put it.
		 */
		if (!F_ISSET(sdbp, DB_AM_DUP)) {
			/* Case 3. */
			memset(&oldpkey, 0, sizeof(DBT));
			F_SET(&oldpkey, DB_DBT_MALLOC);
			ret = __db_c_get(sdbc,
			    &skey, &oldpkey, rmw | DB_SET);
			if (ret == 0) {
				cmp = __bam_defcmp(sdbp, &oldpkey, &pkey);
				__os_ufree(dbenv, oldpkey.data);
				if (cmp != 0) {
					__db_err(dbenv, "%s%s",
			    "Put results in a non-unique secondary key in an ",
			    "index not configured to support duplicates");
					ret = EINVAL;
					goto skipput;
				}
			} else if (ret != DB_NOTFOUND && ret != DB_KEYEMPTY)
				goto skipput;
		} else if (!F_ISSET(sdbp, DB_AM_DUPSORT)) {
			/* Case 2. */
			memset(&tempskey, 0, sizeof(DBT));
			tempskey.data = skey.data;
			tempskey.size = skey.size;
			memset(&temppkey, 0, sizeof(DBT));
			temppkey.data = pkey.data;
			temppkey.size = pkey.size;
			ret = __db_c_get(sdbc, &tempskey, &temppkey,
			    rmw | DB_GET_BOTH);
			if (ret != DB_NOTFOUND && ret != DB_KEYEMPTY)
				goto skipput;
		}

		ret = __db_c_put(sdbc, &skey, &pkey, DB_UPDATE_SECONDARY);

		/*
		 * We don't know yet whether this was a put-overwrite that
		 * in fact changed nothing.  If it was, we may get DB_KEYEXIST.
		 * This is not an error.
		 */
		if (ret == DB_KEYEXIST)
			ret = 0;

skipput:	FREE_IF_NEEDED(sdbp, &skey)

		if ((t_ret = __db_c_close(sdbc)) != 0 && ret == 0)
			ret = t_ret;

		if (ret != 0)
			goto err;
	}
	if (ret != 0)
		goto err;

	/* If still necessary, go get the old primary key/data.  (Step 4.) */
	if (!have_oldrec) {
		/* See the comments in step 2.  This is real familiar. */
		if ((ret = __db_c_idup(dbc_arg, &pdbc, 0)) != 0)
			goto err;
		DB_ASSERT(flags != DB_CURRENT);
		pkey.data = key->data;
		pkey.size = key->size;
		ret = __db_c_get(pdbc, &pkey, &olddata, rmw | DB_SET);
		if (ret == DB_KEYEMPTY || ret == DB_NOTFOUND) {
			nodel = 1;
			ret = 0;
		}
		if ((t_ret = __db_c_close(pdbc)) != 0 && ret == 0)
			ret = t_ret;
		if (ret != 0)
			goto err;
		have_oldrec = 1;
	}

	/*
	 * If we don't follow this goto, we do in fact have an old record
	 * we may need to go delete.  (Step 5).
	 */
	if (nodel)
		goto skip_s_update;

	for (sdbp = __db_s_first(dbp);
	    sdbp != NULL && ret == 0; ret = __db_s_next(&sdbp)) {
		/*
		 * Call the callback for this secondary to get the
		 * old secondary key.
		 */
		memset(&oldskey, 0, sizeof(DBT));
		if ((ret = sdbp->s_callback(sdbp,
		    &pkey, &olddata, &oldskey)) != 0) {
			if (ret == DB_DONOTINDEX)
				/*
				 * The callback returned a null value--there's
				 * nothing to delete.  Go on to the next
				 * secondary.
				 */
				continue;
			else
				goto err;
		}
		memset(&skey, 0, sizeof(DBT));
		if ((ret = sdbp->s_callback(sdbp,
		    &pkey, ispartial ? &newdata : data, &skey)) != 0 &&
		    ret != DB_DONOTINDEX)
			goto err;

		/*
		 * If there is no new secondary key, or if the old secondary
		 * key is different from the new secondary key, then
		 * we need to delete the old one.
		 *
		 * Note that bt_compare is (and must be) set no matter
		 * what access method we're in.
		 */
		sdbc = NULL;
		if (ret == DB_DONOTINDEX ||
		    ((BTREE *)sdbp->bt_internal)->bt_compare(sdbp,
		    &oldskey, &skey) != 0) {
			if ((ret = __db_cursor_int(
			    sdbp, dbc_arg->txn, sdbp->type,
			    PGNO_INVALID, 0, dbc_arg->locker, &sdbc)) != 0)
				goto err;
			if (CDB_LOCKING(dbenv)) {
				DB_ASSERT(sdbc->mylock.off == LOCK_INVALID);
				F_SET(sdbc, DBC_WRITER);
			}

			/*
			 * Don't let c_get(DB_GET_BOTH) stomp on
			 * our data.  Use a temp DBT instead.
			 */
			memset(&tempskey, 0, sizeof(DBT));
			tempskey.data = oldskey.data;
			tempskey.size = oldskey.size;
			memset(&temppkey, 0, sizeof(DBT));
			temppkey.data = pkey.data;
			temppkey.size = pkey.size;
			if ((ret = __db_c_get(sdbc,
			    &tempskey, &temppkey, rmw | DB_GET_BOTH)) == 0)
				ret = __db_c_del(sdbc, DB_UPDATE_SECONDARY);
			else if (ret == DB_NOTFOUND)
				ret = __db_secondary_corrupt(dbp);
		}

		FREE_IF_NEEDED(sdbp, &skey);
		FREE_IF_NEEDED(sdbp, &oldskey);
		if (sdbc != NULL && (t_ret = __db_c_close(sdbc)) != 0 &&
		    ret == 0)
			ret = t_ret;
		if (ret != 0)
			goto err;
	}

	/* Secondary index updates are now done.  On to the "real" stuff. */

skip_s_update:
	/*
	 * If we have an off-page duplicates cursor, and the operation applies
	 * to it, perform the operation.  Duplicate the cursor and call the
	 * underlying function.
	 *
	 * Off-page duplicate trees are locked in the primary tree, that is,
	 * we acquire a write lock in the primary tree and no locks in the
	 * off-page dup tree.  If the put operation is done in an off-page
	 * duplicate tree, call the primary cursor's upgrade routine first.
	 */
	if (dbc_arg->internal->opd != NULL &&
	    (flags == DB_AFTER || flags == DB_BEFORE || flags == DB_CURRENT)) {
		/*
		 * A special case for hash off-page duplicates.  Hash doesn't
		 * support (and is documented not to support) put operations
		 * relative to a cursor which references an already deleted
		 * item.  For consistency, apply the same criteria to off-page
		 * duplicates as well.
		 */
		if (dbc_arg->dbtype == DB_HASH && F_ISSET(
		    ((BTREE_CURSOR *)(dbc_arg->internal->opd->internal)),
		    C_DELETED)) {
			ret = DB_NOTFOUND;
			goto err;
		}

		if ((ret = dbc_arg->c_am_writelock(dbc_arg)) != 0)
			return (ret);
		if ((ret = __db_c_dup(dbc_arg, &dbc_n, DB_POSITION)) != 0)
			goto err;
		opd = dbc_n->internal->opd;
		if ((ret = opd->c_am_put(
		    opd, key, data, flags, NULL)) != 0)
			goto err;
		goto done;
	}