On Tue, 2015-12-01 at 12:06 -0800, Eric Dumazet wrote: > On Tue, 2015-12-01 at 11:17 -0800, Cong Wang wrote: > > On Tue, Dec 1, 2015 at 11:09 AM, Eric Dumazet <eric.duma...@gmail.com> > > wrote: > > > On Tue, 2015-12-01 at 10:43 -0800, Cong Wang wrote: > > > > > >> This smells hacky... Another way to fix this is to hold the qdisc tree > > >> lock in mq_dump(), since it is not a hot path (comparing with > > >> enqueue/dequeue)? > > > > > > Really ? Which qdisc tree lock will protect you exactly ??? > > > > > > Whole point of MQ is that each TX queue has its own lock. > > > > > > So multiple cpus can call qdisc_tree_decrease_qlen() at the same time, > > > holding their own lock. > > > > > > Clearly modifying mq 'data' is wrong. > > > > Ah, yeah, but mq _seems_ also the only one who modifies sch->q.qlen > > in ->dump(), which is the root cause of this bug. I am wondering if it > > should > > just compute the qlen and return it without modifying sch->q.qlen. > > Sure, but then we still would get PAX underflows warnings ... > > Also need to take care of sch->qstats.drops += count; > > Also that would require a change of ->dump() api, since tc_fill_qdisc() > does : > > if (q->ops->dump && q->ops->dump(q, skb) < 0) > goto nla_put_failure; > qlen = q->q.qlen; > > Not sure it is worth the pain, changing signature of all ->dump() > handlers... > > > What about adding TCQ_F_NOPARENT and then : > > Note : Seems to be more invasive patch for net tree (need to properly > set TCQ_F_NOPARENT)
Hmm... it looks like we have a much more serious bug : qdisc_lookup() calls qdisc_match_from_root(dev->qdisc, handle) without proper lock being held, so we might actually crash the host, if qdisc_tree_decrease_qlen() happens at the time qdiscs are changed. qdisc_tree_decrease_qlen() needs serious care :( Damned. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
