Re: [RFC PATCH 3/3] net: ethernet: ti: cpsw: add XDP support

From: Ivan Khoronzhuk
Date: Thu Apr 18 2019 - 14:30:51 EST


On Thu, Apr 18, 2019 at 10:41:11AM -0700, Jakub Kicinski wrote:
On Thu, 18 Apr 2019 12:40:10 +0300, Ivan Khoronzhuk wrote:
>work? The stack's TX ring has a lock, and can be used from any CPU,
>while XDP TX rings are per-PCU, no?
Yes and no.
am572 has more queues then CPU num, How I can choose tx queue not based on CPU
num? It's always shared and has to have lock, and cpdma is done in this way.

Oh, I see, you already have another lock.

Here another thing bothering me, I send it to queue 0 always, instead of
taking cpu num. Not sure about this, but I expect to have some tx queue
not bind to cpu and didn't find a way it can be changed dynamically in
redirect.

Not sure I understand :)

I mean, is there a capability to choose tx queue while XDP_REDIRECT or XDP_TX.

From what I've seen it's taken by:
qidx = smp_processor_id();

What if I need to send classified traffic to hw queue 4 for instance...

--
Regards,
Ivan Khoronzhuk