On Wed, 28 Sep 2022 19:58:23 +0100 Edward Cree wrote:

On 28/09/2022 19:32, Jakub Kicinski wrote:

quoted

I won't help with the indirect stuff, I fixed it once a while
back already and it keeps getting broken. It must be a case of 
the extack not being plumbed thru, or people being conservative
because the errors are not fatal, right? Solvable.  

The conceptual problem, as I see it, is that multiple hw drivers /
 driver instances might be trying to offload the same tunnel rule,
 because the ingress device isn't actually specified anywhere in
 the weird inside-out way TC tunnel rules work.
So how do you deal with the case where one driver succeeds and
 another fails to offload, or two fail with different rc and
 extack messages?

Let's solve practical problems first :) The cases with multiple devices
offloading are rare AFAIK.

But I really need to go and check what it does right now, because
 my information might be out of date — some of this driver code
 was first written two years ago so maybe it's since been solved.

quoted

The printf'ing? I recon something simple like adding a destructor 
for the message to the exack struct so you can allocate the message,   

What about just a flag to say "please kfree() the msg on destruct"?
I have a hard time imagining a destructor that would need to do
 anything different.

Yes, seems like that could be good enough. I was wondering if perhaps
someone would like to have a "static" buffer and manage ownership (given
most of the config happens under rtnl_lock) but perhaps that's
unnecessary complexity.

BTW we will prolly need two bits, one to indicate the creator will
actually call free and the other to mark that it's needed. Otherwise
we'd need to sift thru the stack and find all extack instances.
You can if you want to but...

quoted

or adding a small buffer in place (the messages aren't very long,
usually) come to mind.  

Also an option, yeah.  Downside is that it consumes that memory
 (I guess 80B or so?) for every netlink response whether it's using
 formatted extack or not; idk if people with lots of netlink
 traffic might start to care about that.

It's just a buffer on the stack, in the struct, the extack is
transformed into netlink attrs in the same way regardless.
Stack use is the only concern, no other impact on those not using it.

Should I rustle up an RFC patch for one of these, or post an RFD to
 the list to canvass other vendors' opinions?

Would be great! Maybe also grep the archive, cause this came up before.
Someone was against this in the past, perhaps, perhaps even me :)
But if it wasn't me we should CC them.