Nils Goroll

> > But yes, I am aware and I was specifically referring to this argument in the last paragraph of the initial comment. > > And I quoted the last...

As the rest of the discussion looks like it might be resolved, I will only respond to the last paragraph: > we should avoid the complications of a just-in-time allocation...

> So does a small systematic allocation (56B by default) during the `struct req` setup. It is not like I had not considered this option. My worry is that it...

Sure is the workspace allocation failure handling doable. The point here is that if we run into the ws overflow at _this_ point, a subsequent request will already fail, and...

I disagree that it is the better option, because it charges all users with a workspace allocation which will not be needed in most cases. The `struct reqtop` case is...

> fixed tiny workspace overhead It is not fixed. It depends on `max_restarts`. > paying dividends as soon as you set up a resp for vcl_deliver No. The straight path...

@dridi On the `ReqFiniObjcoreStash()` / `ocstash_fini()` / `ocstash_clear()` call site (called "the cleanup" in what follows to avoid the bikeshed): I agree that calling it in the middle of `Req_Rollback()`...

@dridi thank you for your excellent summary. I am now going to continue here and would like to ask you to please apologize my confusion when I wrote "But calling...

@dridi I have now taken your version of the patch in slightly modified form and added you as Co-Auther by the de-facto standard form (I should have looked this up...

> So I still think this is a good idea, but we need to deal with `set HEADER = HEADER` to actually reap benefits. b00092.vtc shows how copying from an...