IPsec performace - netisr hits %100

[Mark Johnston]

On Sun, May 02, 2021 at 04:08:18PM +0300, Andrey V. Elsukov wrote:
> 30.04.2021 23:32, Mark Johnston пишет:
> > Second, netipsec unconditionally hands rx processing off to netisr
> > threads for some reason, that's why changing the dispatch policy doesn't
> > help.  Maybe it's to help avoid running out of kernel stack space or to
> > somehow avoid packet reordering in some case that is not clear to me.  I
> > tried a patch (see below) which eliminates this and it helped somewhat.
> > If anyone can provide an explanation for the current behaviour I'd
> > appreciate it.
> 
> Previously we have reports about kernel stack overflow during IPsec
> processing. In your example there is only one IPsec transform is
> configured, but it is possible to configure several in the bundle,
> AFAIR, it is limited to 4 transforms. E.g. if you configure ESP+AH - it
> is bundle of two transforms and this will grow kernel stack requirements.

Is it only a problem for synchronous crypto ops?  With hardware drivers
I'd expect the stack usage to be reset after each transform, since
completions are handled by a dedicated thread.  There is also the
net.inet.ipsec.async_crypto knob, which has a similar effect I think.