Re: [PATCH net-next 1/3] net: provide macros for commonly copied lockless queue stop/wake code

[Jakub Kicinski <kuba@kernel.org>]

On Tue, 4 Apr 2023 14:39:11 +0800 Herbert Xu wrote:
> Thanks for adding me to this thread as otherwise I would've surely
> missed it.
> 
> I see where the confusion is coming from.  The key is that we weren't
> trying to stop every single race, because not all of them are fatal.
> 
> In particular, we tolerate the race where a wake is done when it
> shouldn't be because the network stack copes with that by requeueing
> the skb onto the qdisc.
> 
> So it's a trade-off.  We could make our code water-tight, but then
> we would be incurring a penalty for every skb.  With our current
> approach, the penalty is only incurred in the unlikely event of a
> race which results in the unlucky skb being requeued.
> 
> The race that we do want to stop is a queue being stuck in a stopped
> state when it shouldn't because that indeed is fatal.
> 
> Going back to the proposed helpers, we only need one mb because
> that's all we need to fix the stuck/stopped queue race.

Thanks, I'm impressed you still remember the details :)

I'll leave it racy in the next version. Re-using the BQL barrier
is a bit more tricky on the xmit path than I thought. I'll just
document that false-positive wake ups are possible.