David S. Miller wrote:
Looks like qdisc destruction has some expensive algorithms.
Any quick ideas about the root culprit at least in the hfsc
case? He says htb does it too.
hfsc_destroy_qdisc takes O(n) time wrt. the number of classes,
but 5-6 seconds is still long. If all these classes contain inner
qdiscs other than the default, I guess removing the classes from
dev->qdisc_list in qdisc_destroy takes up most of the time, with
n O(n) operations. The __qdisc_destroy rcu callback also calls
reset before destroy, I don't know any qdisc where this is really
neccessary. Without inner qdiscs, I need to see the script first to
judge what's going wrong. Tomasz ?
BTW: The lockless loopback patch broke qdisc_destroy in multiple ways.
The rcu callback doesn't do any locking, to add locking all
read/write_lock(qdisc_tree_lock) need to be changed to
read/write_lock_bh because the callback is called from a tasklet,
until now all changes to the tree structure were made in
process-context. Additionally it invalidates the assumption made by
dev_shutdown that qdisc_destroy will destroy all qdiscs and clear
dev->qdisc_list immediately. Since qdisc->dev is not refcounted
netdev_wait_allrefs won't notice when the rcu callback hasn't destroyed
all qdiscs yet and free the device, but qdisc_destroy called from
ops->destroy called from the callback will still access the memory.
Patch coming up soon.
Regards
Patrick
Begin forwarded message:
Date: Tue, 27 Jul 2004 11:47:02 +0200
From: Tomasz Paszkowski <tomasz.paszkowski@xxxxxxxx>
To: linux-kernel@xxxxxxxxxxxxxxx
Subject: hfsc and huge set of rules
Hello,
I'am running hfsc qdisc with huge set of rules loaded.
root@hades:/home/system/scr/etc/hfsc_rebuild# cat tc.batch | grep hfsc | wc -l
27884
Always when I delete the root qdisc (qdisc del dev eth0 root)
the machine stop responding for about 5-6 seconds. As I think it's due the
hfsc_destory_qdisc is executed in main kernel thread. Similar problem is
present also in htb scheduler.
Is there any quick solution to solve this problem ?
|