Re: [PATCH 0/8] unwind, arm64: add sframe unwinder for kernel

[Song Liu]

On Thu, Feb 13, 2025 at 2:22 PM Puranjay Mohan <puranjay@xxxxxxxxxx> wrote:
>
> Song Liu <song@xxxxxxxxxx> writes:
>
> > On Thu, Feb 13, 2025 at 12:38 AM Puranjay Mohan <puranjay@xxxxxxxxxx> wrote:
> > [...]
> >>
> >> P.S. - The livepatch doesn't have copy_process() but only copy_signal(),
> >> yours had copy_process() somehow.
> >
> > In my build, copy_signal is inlined to copy_process, unless I add noinline.
> > If I do add noinline, the issue will not reproduce.
> >
> > I tried more combinations. The issue doesn't reproduce if I either
> > 1) add noinline to copy_signal, so we are not patching the whole
> >    copy_process function;
> > or
> > 2) Switch compiler from gcc 14.2.1 to gcc 11.5.0.
> >
> > So it appears something in gcc 14.2.1 is causing live patch to fail
> > for copy_process().
>
> So, can you test your RFC set (without SFRAME) with gcc 14.2.1, so we
> can be sure that it is not a sframe problem?
>
> And about having the .sframe section in the livepatch module, I realised
> that this set doesn't include support for reading/using sframe data from
> any module(livepatches included), so the patch I added for generating
> .sframe in kpatch is irrelevant because it is a no-op with the current setup.

Puranjay,

Could you please try the following?

1. Use gcc 11.4.1;
2. Add __always_inline to copy_signal();
3. Build kernel, and livepatch with the same test (we need to
    add __always_inline to the .patch file).
4. Run gdb livepatch-xxx.ko
5. In gdb do disassemble copy_process.

In my tests, both gcc-14.2.1 and gcc-11.5.0 generated a .ko file
that looks weird in gdb-disassemble. Specifically, readels shows
copy_process is about 5.5kB, but gdb-disassemble only shows
140 bytes or so for copy_process. clang doesn't seem to have
this problem.

I am really curious whether you have the same problem in your
setup.

Thanks,
Song