Re: [PATCH net-next] once: add DO_ONCE_SLOW() for sleepable contexts

[Willy Tarreau <w@1wt.eu>]

On Sun, Oct 02, 2022 at 12:50:38AM +0200, Jason A. Donenfeld wrote:
> > > This patch adds DO_ONCE_SLOW() which uses a mutex instead of a spinlock
> > > for operations where we prefer to stay in process context.
> > 
> > That's a nice improvement I think. I was wondering if, for this special
> > case, we *really* need an exclusive DO_ONCE(). I mean, we're getting
> > random bytes, we really do not care if two CPUs change them in parallel
> > provided that none uses them before the table is entirely filled. Thus
> > that could probably end up as something like:
> > 
> >     if (!atomic_read(&done)) {
> >         get_random_bytes(array);
> >         atomic_set(&done, 1);
> >     }
> 
> If you don't care about the tables being consistent between CPUs, then
> yea, sure, that seems like a reasonable approach, and I like not
> polluting once.{c,h} with some _SLOW() special cases.

I don't see this as pollution, it possibly is a nice addition for certain
use cases or early fast paths where the risk of contention is high.

> If you don't want
> the atomic read in there you could also do the same pattern with a
> static branch, like what DO_ONCE() does:
> 
>    if (static_branch_unlikely(&need_bytes)) {
>       get_random_bytes(array);
>       static_branch_disable(&need_bytes);
>    }
> 
> Anyway, same thing as your suggestion more or less.

What I don't know in fact is if the code patching itself can be
responsible for a measurable part of the extra time Christophe noticed.
Anyway at least Christophe now has a few approaches to try, let's first
see if any of them fixes the regression.

Willy