netfilter: x_table: speedup compat operations

One iptables invocation with 135000 rules takes 35 seconds of cpu time
on a recent server, using a 32bit distro and a 64bit kernel.

We eventually trigger NMI/RCU watchdog.

INFO: rcu_sched_state detected stall on CPU 3 (t=6000 jiffies)

COMPAT mode has quadratic behavior and consume 16 bytes of memory per
rule.

Switch the xt_compat algos to use an array instead of list, and use a
binary search to locate an offset in the sorted array.

This halves memory need (8 bytes per rule), and removes quadratic
behavior [ O(N*N) -> O(N*log2(N)) ]

Time of iptables goes from 35 s to 150 ms.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
This commit is contained in:
Eric Dumazet
2010-12-18 18:35:15 +01:00
committed by Pablo Neira Ayuso
parent b017900aac
commit 255d0dc340
6 changed files with 57 additions and 35 deletions

View File

@@ -611,8 +611,9 @@ struct _compat_xt_align {
extern void xt_compat_lock(u_int8_t af);
extern void xt_compat_unlock(u_int8_t af);
extern int xt_compat_add_offset(u_int8_t af, unsigned int offset, short delta);
extern int xt_compat_add_offset(u_int8_t af, unsigned int offset, int delta);
extern void xt_compat_flush_offsets(u_int8_t af);
extern void xt_compat_init_offsets(u_int8_t af, unsigned int number);
extern int xt_compat_calc_jump(u_int8_t af, unsigned int offset);
extern int xt_compat_match_offset(const struct xt_match *match);