Question about inline, notrace, and livepatch

Question about inline, notrace, and livepatch

[Song Liu]

We hit the following hiccup a couple times in the past few months:

A function is marked as "inline", but the compiler decided not to inline it.
Since "inline" implies "notrace" since [1], this function doesn't have
fentry/mcount. When we built livepatch for this function, kpatch-build
failed with:

   xxx.o: function yyy has no fentry/mcount call, unable to patch

This is not a deal breaker, as we can usually modify the patch to work
around it. But I wonder whether we still need "inline" to imply "notrace".
Can we remove this to make livepatch a little easier?

Thanks,
Song

[1] 93b3cca1ccd3 ("ftrace: Make all inline tags also include notrace")

Re: Question about inline, notrace, and livepatch

[Steven Rostedt]

On Tue, 2 May 2023 10:40:28 -0700
Song Liu <song@kernel.org> wrote:

> We hit the following hiccup a couple times in the past few months:
> 
> A function is marked as "inline", but the compiler decided not to inline it.
> Since "inline" implies "notrace" since [1], this function doesn't have
> fentry/mcount. When we built livepatch for this function, kpatch-build
> failed with:
> 
>    xxx.o: function yyy has no fentry/mcount call, unable to patch
> 
> This is not a deal breaker, as we can usually modify the patch to work
> around it. But I wonder whether we still need "inline" to imply "notrace".
> Can we remove this to make livepatch a little easier?

The history behind this is that there were cases that functions that were
inlined, were in critical paths that could cause crashes if they were
traced. In testing they never triggered because the developer's compiler
inlined them. Then on someone else's machine, the compiler decided not to
inline the function and the system crashed. It was hell to debug because I
was not able to reproduce the issue, as my compiler would always keep the
function inlined!

But a lot has changed since then. I've implemented the
"ftrace_test_recursion_trylock()" that catches pretty much all recursive
bugs (and can ever report when they happen). So I may be open to removing
the "notrace" from "inline".

-- Steve

Re: Question about inline, notrace, and livepatch

[Song Liu]

+ Yonghong,

On Tue, May 2, 2023 at 10:56 AM Steven Rostedt <rostedt@goodmis.org> wrote:
>
> On Tue, 2 May 2023 10:40:28 -0700
> Song Liu <song@kernel.org> wrote:
>
> > We hit the following hiccup a couple times in the past few months:
> >
> > A function is marked as "inline", but the compiler decided not to inline it.
> > Since "inline" implies "notrace" since [1], this function doesn't have
> > fentry/mcount. When we built livepatch for this function, kpatch-build
> > failed with:
> >
> >    xxx.o: function yyy has no fentry/mcount call, unable to patch
> >
> > This is not a deal breaker, as we can usually modify the patch to work
> > around it. But I wonder whether we still need "inline" to imply "notrace".
> > Can we remove this to make livepatch a little easier?
>
> The history behind this is that there were cases that functions that were
> inlined, were in critical paths that could cause crashes if they were
> traced. In testing they never triggered because the developer's compiler
> inlined them. Then on someone else's machine, the compiler decided not to
> inline the function and the system crashed. It was hell to debug because I
> was not able to reproduce the issue, as my compiler would always keep the
> function inlined!
>
> But a lot has changed since then. I've implemented the
> "ftrace_test_recursion_trylock()" that catches pretty much all recursive
> bugs (and can ever report when they happen). So I may be open to removing
> the "notrace" from "inline".

Thanks for the history! Let's try to remove this constraint.

Song