[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-ID: <4AE7279B.7070300@gmail.com>
Date: Tue, 27 Oct 2009 18:02:19 +0100
From: Eric Dumazet <eric.dumazet@...il.com>
To: "David S. Miller" <davem@...emloft.net>
CC: Linux Netdev List <netdev@...r.kernel.org>
Subject: [net-next-2.6 PATCH 0/6] net: Speedup netdevice unregisters
netdevice unregisters are serialized and call synchronize_{net|rcu}()
three times per device.
This means it can take a long time to remove a device, if many virtual
devices are attached. (vlan, macvlan, tunnels, ...)
This patch series partially solve the problem by batching several devices in a list,
so that two synchronize_net() calls can be factorized.
Some results on my dev machine (2x4 cpus, HZ=1000, PREEMPT):
modprobe dummy numdummies=10000
# unregister 10000 DOWN netdevices
time rmmod dummy
Before patches :
real 2m0.303s (12 ms per device)
user 0m0.001s
sys 0m0.117s
After patches 1,2,3 :
real 0m1.111s
user 0m0.001s
sys 0m0.922s (mostly sysfs overhead)
modprobe dummy
ip link set dummy0 up
for f in `seq 1 1000`
do
ip link add link dummy0 dummy0.$f type vlan id $f
ip link set dummy0.$f up
done
# Dismantle of 1000 UP vlans (but no IP address)
time rmmod dummy
Before patches :
real 0m40.410s (40 ms per vlan)
user 0m0.000s
sys 0m0.022s
After patches 1,2,3 :
real 0m20.990s
user 0m0.001s
sys 0m0.011s
After patch 4 (vlan: Use unregister_netdevice_many())
real 0m4.392s (-> 4.3 ms per vlan)
user 0m0.000s
sys 0m0.067s
We still have a synchronize_rcu() in dev_deactivate() that could be
factorized in followup patches if there is some interest.
PATCH 1/6 : net: Introduce unregister_netdevice_queue()
PATCH 2/6 : net: Introduce unregister_netdevice_many()
PATCH 3/6 : net: Add a list_head parameter to dellink() method
PATCH 4/6 : vlan: Optimize multiple unregistration
PATCH 5/6 : ipip: Optimize multiple unregistration
PATCH 6/6 : gre: Optimize multiple unregistration
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Powered by blists - more mailing lists