On Tue, Aug 19, 2008 at 09:16:42AM +0000, Jarek Poplawski wrote:

On Tue, Aug 19, 2008 at 06:55:04PM +1000, Herbert Xu wrote:

...

quoted

In fact we're not really adding anything new here, the qdiscs were
not accessed under RCU uniformly.  If you go back in the tree prior
to the multi-qdisc stuff, you'll find that only dev_queue_xmit works
under RCU.  qdisc_restart does not and therefore deferring the
destruction to RCU is pointless anyway.

So in fact we've already been relying on the fact that by the time
qdisc_destroy comes about nobody on the read side (i.e., the packet
transmission path) should have a reference to it.

Let's not discuss using such a qdisc by others but a possibility
that some common lists could be broken for readers from upper qdiscs.
(They were not deactivated.) Of course, if it's done properly with
some references before qdisc_destroy then it's all right. I'd prefer
to check this later yet.

So, what I was most suspicious of, cls_u32, looks like safe wrt. this.
Congratulations for good estimation of this.

But how about this part in qdisc_destroy(): 
        if (qdisc->parent)
                list_del(&qdisc->list);

If we do this with child qdisc from qdisc_graft() it's without
deactivation. The rest of the tree can be dequeued in the meantime
and call qdisc_tree_decrease_qlen() (like hfsc, tbf, netem), which
uses qdisc_lookup() to access this list. We list_del() under rtnl
lock only, they lookup under sch_tree_lock(). Is it a bit unsafe
or I miss something?

Thanks,
Jarek P.