In dquot_initialize(), we'd call dqget() only when i_dquot not
initialized, which can avoid 2 pair of dqget()/dqput() in most
case. It could relieve the global locks contenion caused by
dqget()/dqput().
--- linux-3.0/fs/quota/dquot.c.orig	2013-07-18 09:46:28.000000000 -0400
+++ linux-3.0/fs/quota/dquot.c	2013-07-18 09:49:20.000000000 -0400
@@ -1342,8 +1342,8 @@
 static void __dquot_initialize(struct inode *inode, int type)
 {
 	unsigned int id = 0;
-	int cnt;
-	struct dquot *got[MAXQUOTAS];
+	int cnt, ret = 0, dq_get = 0;
+	struct dquot *got[MAXQUOTAS] = { NULL, NULL };
 	struct super_block *sb = inode->i_sb;
 	qsize_t rsv;
 
@@ -1352,7 +1352,14 @@
 	if (!dquot_active(inode))
 		return;
 
-	/* First get references to structures we might need. */
+	/* In most case, the i_dquot should have been initialized, except
+	 * the newly allocated one. We'd always try to skip the dqget() and
+	 * dqput() calls to avoid unnecessary global lock contention. */
+	if (!(inode->i_state & I_NEW))
+		goto init_idquot;
+
+get_dquots:
+	dq_get = 1;
 	for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
 		got[cnt] = NULL;
 		if (type != -1 && cnt != type)
@@ -1367,7 +1374,7 @@
 		}
 		got[cnt] = dqget(sb, id, cnt);
 	}
-
+init_idquot:
 	spin_lock(&inode->i_lock);
 	if (IS_NOQUOTA(inode))
 		goto out_err;
@@ -1381,10 +1388,14 @@
 		/* Avoid races with quotaoff() */
 		if (!sb_has_quota_active(sb, cnt))
 			continue;
-		/* We could race with quotaon or dqget() could have failed */
-		if (!got[cnt])
-			continue;
 		if (!inode->i_dquot[cnt]) {
+			if (dq_get == 0) {
+				spin_unlock(&inode->i_lock);
+				goto get_dquots;
+			}
+			/* We could race with quotaon or dqget() could have failed */
+			if (!got[cnt])
+				continue;
 			inode->i_dquot[cnt] = got[cnt];
 			got[cnt] = NULL;
 			/*
@@ -1388,7 +1399,7 @@
 			 * did a write before quota was turned on
 			 */
 			rsv = inode_get_rsv_space(inode);
-			if (unlikely(rsv)) {
+			if (unlikely(rsv) && likely(inode->i_dquot[cnt])) {
 				spin_lock(&dq_data_lock);
 				dquot_resv_space(inode->i_dquot[cnt], rsv);
 				spin_unlock(&dq_data_lock);
@@ -1488,7 +1399,8 @@
 	spin_unlock(&inode->i_lock);
 
 	/* Drop unused references */
-	dqput_all(got);
+	if (dq_get)
+		dqput_all(got);
 }
 
 void dquot_initialize(struct inode *inode)