Le 29/09/2018 à 10:57, Sebastian Moeller a écrit :
>>> So in kernel 4.17 I believe landed flow offloading support. This 
>>> accomplishes something similar to SFE while being a proper upstream 
>>> solution. This is already implemented in OpenWrt.
>> flow offloading for NAT is different from SFE :
>> In the first case, you offload the NATing algorithm work to the hardware
>> that support it, which narrow a lot the supported targets.
>> In the second case, you optimize the NATing algorithm Linux kernel
>> implementation for all targets, it's a software and generic optimization.  
>       As far as I can tell, SFE ans friends gain their speed advantage mainly 
> by being less generic that the main kernel, so speaking of an optimization 
> seems problematic. Because the reached better routing/NATing/whatever 
> typically is quite fickle and might for example not be compatible with 
> traffic shaping...

It's an implementation issue of the forwarding fastpath inside the Linux
kernel. Someone already stated here that other firmware that propose
forwarding fastpath as an option that is not compatible with other
features like netfilter, which is a showstopper, exists.

Honestly, I've not followed lately the status of openfastpath or SFE or
friends but I'm pretty their end goal is not to enable fastpath in the
Linux networking stack at the cost of other features ... but to offer an
alternative to the default slowpath in the first place, and then to
enable the fastpath for flows that have the criterion to "follow" it.   


Jérôme Benoit aka fraggle
Piment Noir - https://piment-noir.org
OpenPGP Key ID : 27B535D3
Key fingerprint : B799 BBF6 8EC8 911B B8D7 CDBC C3B1 92C6 27B5 35D3

Attachment: signature.asc
Description: OpenPGP digital signature

openwrt-devel mailing list

Reply via email to