On Thu, Nov 18, 2021 at 6:49 PM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Thu, Nov 18, 2021 at 2:33 AM Florent Revest <revest@xxxxxxxxxxxx> wrote:
> >
> > We have a use-case where we construct a skeleton with a sleepable
> > program and if it fails to load then we want to make it non-sleepable by
> > clearing BPF_F_SLEEPABLE.
>
> I'd say the better way to do this is to have two programs (that share
> the logic, of course) and pick one or another at runtime:
>
> static int whatever_logic(bool sleepable) { ... }
>
> SEC("fentry.s/whatever")
> int BPF_PROG(whatever_sleepable, ...)
> {
> return whatever_logic(true);
> }
>
> SEC("fentry/whatever")
> int BPF_PROG(whatever_nonsleepable, ...)
> {
> return whatever_logic(false);
> }
>
>
> Then at runtime you can bpf_program__autoload(..., false) for a
> variant you don't want to load.
Ah cool, thanks! That's a good idea :) it will also look cleaner.
> This clear_flags business seems too low-level and too limited. Next
> thing we'll be adding a few more bit manipulation variants (e.g, reset
> flags). Let's see how far you can get with the use of existing
> features. I'd set_extra_flags() to be almost never used, btw. And they
> shouldn't, if can be avoided. So I'm hesitant to keep extending
> operations around prog_flags.
I agree
> But given we just added set_extra_flags() and it's already too
> limiting, let's change set_extra flags to just set_flags() that will
> override the flags with whatever user provides. Then with
> bpf_program__flags() and bpf_program__set_flags() you can express
> whatever you want without adding extra APIs. Care to fix that?
Sure, I'll send a patch for this! :)
