Patrick McHardy wrote:

Jarek Poplawski wrote:

quoted

After implementing qdisc->ops->peek() there is no more calling
qdisc_tree_decrease_qlen() without rtnl_lock(), so qdisc_list_lock
added by commit: f6e0b239a2657ea8cb67f0d83d0bfdbfd19a481b "pkt_sched:
Fix qdisc list locking" can be removed.

I might be misunderstanding something, but qdisc_tree_decrease_qlen()
doesn't need rtnl_lock() but the sch_tree_lock() since it changes
q.qlen.

OK I see now, rtnl_lock() was taken to protect the parent qdisc lookup.
In that case it seems fine, all callers are holding both sch_tree_lock()
and rtnl_lock().