Re: [RFC PATCH -next v2 3/4] arm64/ftrace: support dynamically allocated trampolines

[Mark Rutland]

On Thu, Apr 21, 2022 at 01:06:48PM -0400, Steven Rostedt wrote:
> On Thu, 21 Apr 2022 17:27:40 +0100
> Mark Rutland <mark.rutland@xxxxxxx> wrote:
> 
> > We can initialize the ops pointer to a default ops that does the whole
> > __do_for_each_ftrace_ops() dance.
> 
> OK, I think I understand now. What you are doing is instead of creating a
> trampoline that has all the information in the trampoline, you add nops to
> all the functions where you can place the information in the nops (before
> the function), and then have the trampoline just read that information to
> find the ops pointer as well as the function to call.

FWIW, I had a go at mocking that up:

  https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/log/?h=arm64/ftrace/per-callsite-ops

Aside from some bodges required to ensure the patch site is suitably aligned
(which I think can be cleaned up somewhat), I don't think it looks that bad.

I wasn't sure how exactly to wire that up in the core code, so all the patch
sites are initialized with a default ops that calls
arch_ftrace_ops_list_func(), but it looks like it should be possible to wire
that up in the core with some refactoring.

> I guess you could have two trampolines as well. One that always calls the
> list loop, and one that calls the data stored in front of the function that
> was just called the trampoline. As it is always safe to call the loop
> function, you could have the call call that trampoline first, set up the
> specific data before the function, then call the trampoline that will read
> it. 

I was thinking we could just patch the ops with a default ops that called the
list loop, as my patches default them to.

> And same thing for tear down.

I wasn't sure how teardown was meant to work in general. When we want to
remove an ops structure, or a trampoline, how do we ensure those are no
longer in use before we remove them? I can see how we can synchronize
the updates to the kernel text, but I couldn't spot how we handle a
thread being in the middle of a trampoline.

Thanks,
Mark.