Pablo Neira Ayuso wrote:
This patch adds dynamic message size calculation for ctnetlink. This
reduces CPU consumption since the overhead in the message trimming
is removed.
+static inline size_t
+ctnetlink_calculate_room_size(const struct nf_conn *ct, unsigned long events)
+{
+ size_t size = NLMSG_SPACE(sizeof(struct nfgenmsg));
+
+ size += calculate_tuple_room_size(ct) * 2 + /* original and reply */
+ nla_total_size(sizeof(u_int32_t)) + /* status */
+ nla_total_size(sizeof(u_int32_t)); /* id */
+
+#ifdef CONFIG_NF_CONNTRACK_MARK
+ if (events & IPCT_MARK || ct->mark)
+ size += nla_total_size(sizeof(u_int32_t));
+#endif
+
+ if (events & IPCT_DESTROY) {
+ const struct nf_conn_counter *acct;
+
+ acct = nf_conn_acct_find(ct);
+ if (acct) {
+ size += nla_total_size(0) * 2 +
+ nla_total_size(sizeof(u_int64_t)) * 2 * 2;
+ }
+ return size;
+ }
+...
Holger's patch pre-calculates things where possible, which reduces
run-time overhead.
I'm not sure about this. Instead, I noticed with oprofile that we spend
more cycles to calculate the message size. I think it's more like a
trade-off, reducing the message size to what we need reduces the chances
to hit ENOBUFS.
"More cycles" is to be expected for calculation (which is not done
so far), but summed up, I always suspected a measurement error and
I'd be surprised if this is correct.
So we so far have your numbers, which showed a regression, and no
numbers from Holger, which is slightly better :)
OK seriously, we need *some* numbers showing an improvement since I
have basically zero base to decide between your patches, besides the
fact that its to be expected that Holger's will be slightly faster.
--
To unsubscribe from this list: send the line "unsubscribe netfilter-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
