From ff4ca605388d143025085ddb52d4145c54f8b49f Mon Sep 17 00:00:00 2001
From: Yuriy Vostrikov <vostrikov@corp.mail.ru>
Date: Fri, 19 Nov 2010 14:49:02 +0300
Subject: [PATCH] [box] Use vector io only if tuple is big enough.
---
mod/silverbox/box.c | 29 +++++++++++++++++++++++++----
1 file changed, 25 insertions(+), 4 deletions(-)
diff --git a/mod/silverbox/box.c b/mod/silverbox/box.c
index ecd4dbe8b9..66d6d4419a 100644
--- a/mod/silverbox/box.c
+++ b/mod/silverbox/box.c
@@ -62,6 +62,21 @@ const struct field ASTERISK = {
/* hooks */
typedef int (*box_hook_t) (struct box_txn * txn);
+
+/*
+ For tuples of size below this threshold, when sending a tuple
+ to the client, make a deep copy of the tuple for the duration
+ of sending rather than increment a reference counter.
+ This is necessary to avoid excessive page splits when taking
+ a snapshot: many small tuples can be accessed by clients
+ immediately after the snapshot process has forked off,
+ thus incrementing tuple ref count, and causing the OS to
+ create a copy of the memory page for the forked
+ child.
+*/
+
+const int BOX_REF_THRESHOLD = 8196;
+
struct namespace namespace[256];
struct box_snap_row {
@@ -887,12 +902,18 @@ prepare_update_fields(struct box_txn *txn, struct tbuf *data)
static void
tuple_add_iov(struct box_txn *txn, struct box_tuple *tuple)
{
- tuple_txn_ref(txn, tuple);
+ size_t len;
- add_iov(&tuple->bsize,
- tuple->bsize +
+ len = tuple->bsize +
field_sizeof(struct box_tuple, bsize) +
- field_sizeof(struct box_tuple, cardinality));
+ field_sizeof(struct box_tuple, cardinality);
+
+ if (len > BOX_REF_THRESHOLD) {
+ tuple_txn_ref(txn, tuple);
+ add_iov(&tuple->bsize, len);
+ } else {
+ add_iov_dup(&tuple->bsize, len);
+ }
}
static int __noinline__
--
GitLab