Re: [PATCH net-next] net: sched: remove unnecessay lock protection for skb_bad_txq/gso_skb

On Mon, Mar 15, 2021 at 2:29 AM Yunsheng Lin [off-list ref] wrote:

Currently qdisc_lock(q) is taken before enqueuing and dequeuing
for lockless qdisc's skb_bad_txq/gso_skb queue, qdisc->seqlock is
also taken, which can provide the same protection as qdisc_lock(q).

This patch removes the unnecessay qdisc_lock(q) protection for
lockless qdisc' skb_bad_txq/gso_skb queue.

And dev_reset_queue() takes the qdisc->seqlock for lockless qdisc
besides taking the qdisc_lock(q) when doing the qdisc reset,
some_qdisc_is_busy() takes both qdisc->seqlock and qdisc_lock(q)
when checking qdisc status. It is unnecessary to take both lock
while the fast path only take one lock, so this patch also changes
it to only take qdisc_lock(q) for locked qdisc, and only take
qdisc->seqlock for lockless qdisc.

Since qdisc->seqlock is taken for lockless qdisc when calling
qdisc_is_running() in some_qdisc_is_busy(), use qdisc->running
to decide if the lockless qdisc is running.

What's the benefit here? Since qdisc->q.lock is also per-qdisc,
so there is no actual contention to take it when we already acquire
q->seqlock, right?

Also, is ->seqlock supposed to be used for protecting skb_bad_txq
etc.? From my understanding, it was introduced merely for replacing
__QDISC_STATE_RUNNING. If you want to extend it, you probably
have to rename it too.

Thanks.