On 01/03/2018 03:41 PM, Cong Wang wrote: > On Wed, Jan 3, 2018 at 10:09 AM, John Fastabend > <john.fastab...@gmail.com> wrote: >> On 01/02/2018 08:41 PM, Cong Wang wrote: >>> Hi, John >>> >>> While reviewing your ptr_ring fix again today, it looks like your >>> "lockless" qdisc patchset breaks dev->tx_queue_len behavior. >>> >>> Before your patchset, dev->tx_queue_len is merely an integer to read, >>> after your patchset, the skb array has to be resized when >>> dev->tx_queue_len changes, but I don't see any qdisc code handles >>> this... >>> >>> Also, because of that, I doubt __skb_array_empty() in >>> pfifo_fast_dequeue() can be safe any more even with your ptr_ring fix. >>> >>> What am I missing? >>> >> >> I dropped support for tx_queue_len changes after qdisc has been >> created. The only check is at init time when building the qdisc. > > This is where it breaks. > > >> >> Before this series teql and pfifo_fast were the only qdiscs that >> used tx_queue_len other qdiscs used other mechanisms or copied >> tx_queue_len at init time. So the API is inconsistent. > > Yeah, pfifo_fast was able to drop based on latest value of tx_queue_len > before your patchset, this is why I am complaining. >
Yep good complaint. > >> >> OK, but arguably its kAPI now and needs to be supported on live >> qdiscs. So couple options drop the __skb_array_empty() check, >> stop supporting changes on running qdiscs, or do a qdisc swap >> with the new array. > > I don't think we can break the old behavior of tx_queue_len change > for pfifo_fast, people may already rely on it. > Agreed needed for legacy support. > Doing a swap seems reasonable. > >> >> I'm tempted to make the qdisc swap work, still need benchmarks >> I guess without the empty check. Either way to get it working >> we need a callback from tx_queue_len code paths. > > Right, probably need a new ops in Qdisc_ops. > Maybe instead of a Qdisc op just do a direct call to avoid encouraging users to use this code path. Either way is probably fine we can just watch any future patches and have users add a specific attribute for it like codel. > >> >> Unfortunately, I guess someone somewhere probably uses pfifo_fast >> and changes there queue length with a script after creating the >> qdisc and expects it to work. >> > > This is my concern as well. I will work on some patches, this doesn't > look trivial to solve at all. How about a dev_deactivate_many() that instead of replacing with noop qdisc replaces with new updated qdisc. Seems like it might work. Thanks, John > > Thanks. >
