db_join.c

这是linux下运行的mysql软件包,可用于linux 下安装 php + mysql + apach 的网络配置

					 * that cursor is sorted, then any
					 * other sorted cursors can be safely
					 * reset to the first duplicate
					 * duplicate in the current set if we
					 * have a pointer to it (we can't just
					 * leave them be, or we'll miss
					 * duplicate duplicates in the outer
					 * relation).
					 *
					 * If the first cursor is unsorted, or
					 * if cursor j is unsorted, we can
					 * make no assumptions about what
					 * we're looking for next or where it
					 * will be, so we reset to the very
					 * beginning (setting workcurs NULL
					 * will achieve this next go-round).
					 *
					 * XXX: This is likely to break
					 * horribly if any two cursors are
					 * both sorted, but have different
					 * specified sort functions.  For,
					 * now, we dismiss this as pathology
					 * and let strange things happen--we
					 * can't make rope childproof.
					 */
					if ((ret = jc->j_workcurs[j]->c_close(
					    jc->j_workcurs[j])) != 0)
						goto err;
					if (!SORTED_SET(jc, 0) ||
					    !SORTED_SET(jc, j) ||
					    jc->j_fdupcurs[j] == NULL)
						/*
						 * Unsafe conditions;
						 * reset fully.
						 */
						jc->j_workcurs[j] = NULL;
					else
						/* Partial reset suffices. */
						if ((jc->j_fdupcurs[j]->c_dup(
						    jc->j_fdupcurs[j],
						    &jc->j_workcurs[j],
						    DB_POSITIONI)) != 0)
							goto err;
					jc->j_exhausted[j] = 0;
				}
				goto retry;
				/* NOTREACHED */
			}

			/*
			 * We're about to advance the cursor and need to
			 * reset all of the workcurs[j] where j>i, so that
			 * we don't miss any duplicate duplicates.
			 */
			for (j = i + 1;
			    jc->j_workcurs[j] != NULL;
			    j++) {
				if ((ret = jc->j_workcurs[j]->c_close(
				    jc->j_workcurs[j])) != 0)
					goto err;
				jc->j_exhausted[j] = 0;
				if (jc->j_fdupcurs[j] != NULL &&
				    (ret = jc->j_fdupcurs[j]->c_dup(
				    jc->j_fdupcurs[j], &jc->j_workcurs[j],
				    DB_POSITIONI)) != 0)
					goto err;
				else
					jc->j_workcurs[j] = NULL;
			}
			goto retry2;
			/* NOTREACHED */
		}

		if (ret == ENOMEM) {
			jc->j_key.ulen <<= 1;
			if ((ret = __os_realloc(dbp->dbenv, jc->j_key.ulen,
			    &jc->j_key.data)) != 0) {
mem_err:			__db_err(dbp->dbenv,
				    "Allocation failed for join key, len = %lu",
				    (u_long)jc->j_key.ulen);
				goto err;
			}
			goto retry2;
		}

		if (ret != 0)
			goto err;

		/*
		 * If we made it this far, we've found a matching
		 * datum in cursor i.  Mark the current cursor
		 * unexhausted, so we don't miss any duplicate
		 * duplicates the next go-round--unless this is the
		 * very last cursor, in which case there are none to
		 * miss, and we'll need that exhausted flag to finally
		 * get a DB_NOTFOUND and move on to the next datum in
		 * the outermost cursor.
		 */
		if (i + 1 != jc->j_ncurs)
			jc->j_exhausted[i] = 0;
		else
			jc->j_exhausted[i] = 1;

		/*
		 * If jc->j_fdupcurs[i] is NULL and the ith cursor's dups are
		 * sorted, then we're here for the first time since advancing
		 * cursor 0, and we have a new datum of interest.
		 * jc->j_workcurs[i] points to the beginning of a set of
		 * duplicate duplicates;  store this into jc->j_fdupcurs[i].
		 */
		if (SORTED_SET(jc, i) && jc->j_fdupcurs[i] == NULL && (ret =
		    cp->c_dup(cp, &jc->j_fdupcurs[i], DB_POSITIONI)) != 0)
			goto err;

	}

err:	if (ret != 0)
		return (ret);

	if (0) {
samekey:	/*
		 * Get the key we tried and failed to return last time;
		 * it should be the current datum of all the secondary cursors.
		 */
		if ((ret = jc->j_workcurs[0]->c_real_get(jc->j_workcurs[0],
		    &jc->j_key, key_n, DB_CURRENT | opmods)) != 0)
			return (ret);
		F_CLR(jc, JOIN_RETRY);
	}

	/*
	 * ret == 0;  we have a key to return.
	 *
	 * If DB_DBT_USERMEM or DB_DBT_MALLOC is set, we need to copy the key
	 * back into the dbt we were given for the key; call __db_retcopy.
	 * Otherwise, assert that we do not need to copy anything and proceed.
	 */
	DB_ASSERT(F_ISSET(
	    key_arg, DB_DBT_USERMEM | DB_DBT_MALLOC) || key_n == key_arg);

	if (F_ISSET(key_arg, DB_DBT_USERMEM | DB_DBT_MALLOC) &&
	    (ret = __db_retcopy(dbp->dbenv,
	    key_arg, key_n->data, key_n->size, NULL, NULL)) != 0) {
		/*
		 * The retcopy failed, most commonly because we have a user
		 * buffer for the key which is too small. Set things up to
		 * retry next time, and return.
		 */
		F_SET(jc, JOIN_RETRY);
		return (ret);
	}

	/*
	 * If DB_JOIN_ITEM is set, we return it; otherwise we do the lookup
	 * in the primary and then return.
	 *
	 * Note that we use key_arg here;  it is safe (and appropriate)
	 * to do so.
	 */
	if (operation == DB_JOIN_ITEM)
		return (0);

	/*
	 * If data_arg->flags == 0--that is, if DB is managing the
	 * data DBT's memory--it's not safe to just pass the DBT
	 * through to the primary get call, since we don't want that
	 * memory to belong to the primary DB handle (and if the primary
	 * is free-threaded, it can't anyway).
	 *
	 * Instead, use memory that is managed by the join cursor, in
	 * jc->j_rdata.
	 */
	if (!F_ISSET(data_arg, DB_DBT_MALLOC | DB_DBT_REALLOC | DB_DBT_USERMEM))
		db_manage_data = 1;
	else
		db_manage_data = 0;
	if ((ret = __db_join_primget(jc->j_primary,
	    jc->j_curslist[0]->txn, jc->j_curslist[0]->locker, key_arg,
	    db_manage_data ? &jc->j_rdata : data_arg, opmods)) != 0) {
		if (ret == DB_NOTFOUND)
			/*
			 * If ret == DB_NOTFOUND, the primary and secondary
			 * are out of sync;  every item in each secondary
			 * should correspond to something in the primary,
			 * or we shouldn't have done the join this way.
			 * Wail.
			 */
			ret = __db_secondary_corrupt(jc->j_primary);
		else
			/*
			 * The get on the primary failed for some other
			 * reason, most commonly because we're using a user
			 * buffer that's not big enough.  Flag our failure
			 * so we can return the same key next time.
			 */
			F_SET(jc, JOIN_RETRY);
	}
	if (db_manage_data && ret == 0) {
		data_arg->data = jc->j_rdata.data;
		data_arg->size = jc->j_rdata.size;
	}

	return (ret);
}

static int
__db_join_close(dbc)
	DBC *dbc;
{
	DB *dbp;
	DB_ENV *dbenv;
	JOIN_CURSOR *jc;
	int ret, t_ret;
	u_int32_t i;

	jc = (JOIN_CURSOR *)dbc->internal;
	dbp = dbc->dbp;
	dbenv = dbp->dbenv;
	ret = t_ret = 0;

	/*
	 * Remove from active list of join cursors.  Note that this
	 * must happen before any action that can fail and return, or else
	 * __db_close may loop indefinitely.
	 */
	MUTEX_THREAD_LOCK(dbenv, dbp->mutexp);
	TAILQ_REMOVE(&dbp->join_queue, dbc, links);
	MUTEX_THREAD_UNLOCK(dbenv, dbp->mutexp);

	PANIC_CHECK(dbenv);

	/*
	 * Close any open scratch cursors.  In each case, there may
	 * not be as many outstanding as there are cursors in
	 * curslist, but we want to close whatever's there.
	 *
	 * If any close fails, there's no reason not to close everything else;
	 * we'll just return the error code of the last one to fail.  There's
	 * not much the caller can do anyway, since these cursors only exist
	 * hanging off a db-internal data structure that they shouldn't be
	 * mucking with.
	 */
	for (i = 0; i < jc->j_ncurs; i++) {
		if (jc->j_workcurs[i] != NULL && (t_ret =
		    jc->j_workcurs[i]->c_close(jc->j_workcurs[i])) != 0)
			ret = t_ret;
		if (jc->j_fdupcurs[i] != NULL && (t_ret =
		    jc->j_fdupcurs[i]->c_close(jc->j_fdupcurs[i])) != 0)
			ret = t_ret;
	}

	__os_free(dbenv, jc->j_exhausted);
	__os_free(dbenv, jc->j_curslist);
	__os_free(dbenv, jc->j_workcurs);
	__os_free(dbenv, jc->j_fdupcurs);
	__os_free(dbenv, jc->j_key.data);
	if (jc->j_rdata.data != NULL)
		__os_ufree(dbenv, jc->j_rdata.data);
	__os_free(dbenv, jc);
	__os_free(dbenv, dbc);

	return (ret);
}

/*
 * __db_join_getnext --
 *	This function replaces the DBC_CONTINUE and DBC_KEYSET
 *	functionality inside the various cursor get routines.
 *
 *	If exhausted == 0, we're not done with the current datum;
 *	return it if it matches "matching", otherwise search
 *	using DB_GET_BOTHC (which is faster than iteratively doing
 *	DB_NEXT_DUP) forward until we find one that does.
 *
 *	If exhausted == 1, we are done with the current datum, so just
 *	leap forward to searching NEXT_DUPs.
 *
 *	If no matching datum exists, returns DB_NOTFOUND, else 0.
 */
static int
__db_join_getnext(dbc, key, data, exhausted, opmods)
	DBC *dbc;
	DBT *key, *data;
	u_int32_t exhausted, opmods;
{
	int ret, cmp;
	DB *dbp;
	DBT ldata;
	int (*func) __P((DB *, const DBT *, const DBT *));

	dbp = dbc->dbp;
	func = (dbp->dup_compare == NULL) ? __bam_defcmp : dbp->dup_compare;

	switch (exhausted) {
	case 0:
		/*
		 * We don't want to step on data->data;  use a new
		 * DBT and malloc so we don't step on dbc's rdata memory.
		 */
		memset(&ldata, 0, sizeof(DBT));
		F_SET(&ldata, DB_DBT_MALLOC);
		if ((ret = dbc->c_real_get(dbc,
		    key, &ldata, opmods | DB_CURRENT)) != 0)
			break;
		cmp = func(dbp, data, &ldata);
		if (cmp == 0) {
			/*
			 * We have to return the real data value.  Copy
			 * it into data, then free the buffer we malloc'ed
			 * above.
			 */
			if ((ret = __db_retcopy(dbp->dbenv, data, ldata.data,
			    ldata.size, &data->data, &data->size)) != 0)
				return (ret);
			__os_ufree(dbp->dbenv, ldata.data);
			return (0);
		}

		/*
		 * Didn't match--we want to fall through and search future
		 * dups.  We just forget about ldata and free
		 * its buffer--data contains the value we're searching for.
		 */
		__os_ufree(dbp->dbenv, ldata.data);
		/* FALLTHROUGH */
	case 1:
		ret = dbc->c_real_get(dbc, key, data, opmods | DB_GET_BOTHC);
		break;
	default:
		ret = EINVAL;
		break;
	}

	return (ret);
}

/*
 * __db_join_cmp --
 *	Comparison function for sorting DBCs in cardinality order.
 */
static int
__db_join_cmp(a, b)
	const void *a, *b;
{
	DBC *dbca, *dbcb;
	db_recno_t counta, countb;

	/* In case c_count fails, pretend cursors are equal. */
	counta = countb = 0;

	dbca = *((DBC * const *)a);
	dbcb = *((DBC * const *)b);

	if (dbca->c_count(dbca, &counta, 0) != 0 ||
	    dbcb->c_count(dbcb, &countb, 0) != 0)
		return (0);

	return (counta - countb);
}

/*
 * __db_join_primget --
 *	Perform a DB->get in the primary, being careful not to use a new
 * locker ID if we're doing CDB locking.
 */
static int
__db_join_primget(dbp, txn, lockerid, key, data, flags)
	DB *dbp;
	DB_TXN *txn;
	u_int32_t lockerid;
	DBT *key, *data;
	u_int32_t flags;
{
	DBC *dbc;
	int dirty, ret, rmw, t_ret;

	/*
	 * The only allowable flags here are the two flags copied into
	 * "opmods" in __db_join_get, DB_RMW and DB_DIRTY_READ.  The former
	 * is an op on the c_get call, the latter on the cursor call.
	 * It's a DB bug if we allow any other flags down in here.
	 */
	rmw = LF_ISSET(DB_RMW);
	dirty = LF_ISSET(DB_DIRTY_READ);
	LF_CLR(DB_RMW | DB_DIRTY_READ);
	DB_ASSERT(flags == 0);

	if ((ret = __db_icursor(dbp,
	    txn, dbp->type, PGNO_INVALID, 0, lockerid, &dbc)) != 0)
		return (ret);

	if (dirty ||
	    (txn != NULL && F_ISSET(txn, TXN_DIRTY_READ)))
		F_SET(dbc, DBC_DIRTY_READ);
	F_SET(dbc, DBC_TRANSIENT);

	/*
	 * This shouldn't be necessary, thanks to the fact that join cursors
	 * swap in their own DB_DBT_REALLOC'ed buffers, but just for form's
	 * sake, we mirror what __db_get does.
	 */
	SET_RET_MEM(dbc, dbp);

	ret = dbc->c_get(dbc, key, data, DB_SET | rmw);

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

	return (ret);
}