Re: [PATCH net] net: fix raising a softirq on the current cpu with rps enabled

On Mon, Mar 27, 2023 at 1:35 AM Eric Dumazet [off-list ref] wrote:

quoted

Forgive me. Really I need some coffee. I made a mistake. This line
above should be:

+               if (!test_bit(NAPI_STATE_SCHED, &mysd->backlog.state))

But the whole thing doesn't feel right. I need a few days to dig into
this part until Eric can help me with more of it.

I am still traveling, and this is weekend time :/

Thanks for your time, Eric, really appreciate it.

It should not be too hard to read net/core/dev.c and remember that not
_all_ drivers (or some core networking functions) use the NAPI model.

eg look at netif_rx() and ask yourself why your patch is buggy.

Yes, it is. In my last email I sent yesterday I encountered one issue
which exactly happened when I started hundreds of iperf processes
transmitting data to loopback.
It got stuck :( So I realized it is the non-napi case that triggers
such a problem.

Just look at callers of enqueue_to_backlog() and ask yourself if all
of them are called from net_rx_action()

[The answer is no, just in case you wonder]

In order to add your optimization, more work is needed, like adding
new parameters so that we do not miss critical
__raise_softirq_irqoff(NET_RX_SOFTIRQ) when _needed_.

Thanks, I need to do more work/study on it.

We keep going circles around softirq deficiencies, I feel you are
trying to fix a second-order 'issue'.

Right, going circles gives me a headache.

Real cause is elsewhere, look at recent patches from Jakub.

After you pointed out, I searched and found there is indeed one patchset in 2022

The tile like this:

[PATCH 0/3] softirq: uncontroversial change

Thanks,
Jason

Thanks.