ARM Ftrace Function Graph Fails With UNWINDER_FRAME_POINTER

Ard Biesheuvel ardb at kernel.org
Fri Dec 1 22:53:58 PST 2023


On Fri, 1 Dec 2023 at 23:59, Justin Chen <justin.chen at broadcom.com> wrote:
>
>
>
> On 12/1/23 10:07 AM, Steven Rostedt wrote:
> > On Fri, 1 Dec 2023 09:25:59 -0800
> > Justin Chen <justin.chen at broadcom.com> wrote:
> >
> >>> It appears the sub instruction at 0x6dd0 correctly accounts for the
> >>> extra 8 bytes, so the frame pointer is valid. So it is our assumption
> >>> that there are no gaps between the stack frames is invalid.
> >>
> >> Thanks for the assistance. The gap between the stack frame depends on
> >> the function. Most do not have a gap. Some have 8 (as shown above), some
> >> have 12. A single assumption here is not going to work. I'm having a
> >> hard time finding out the reasoning for this gap. I tried disabling a
> >> bunch of gcc flags as well as -O2 and the gap still exists.
> >
> > That code was originally added because of some strange things that gcc did
> > with mcount (for example, it made a copy of the stack frame that it passed
> > to mcount, where the function graph tracer replaced the copy of the return
> > stack making the shadow stack go out of sync and crash). This was very hard
> > to debug and I added this code to detect it if it happened again.
> >
> > Well it's been over a decade since that happened (2009).
> >
> >    71e308a239c09 ("function-graph: add stack frame test")
> >
> > I'm happy assuming that the compiler folks are aware of our tricks with
> > hijacking return calls and I don't expect it to happen again. We can just
> > rip out those checks. That is, if it's only causing false positives, I
> > don't think it's worth keeping around.
> >
> > Has it detected any real issues on the Arm platforms?
> >
> > -- Steve
>
> I am not familiar enough to make a call. But from my limited testing
> with ARM, I didn't see any issues. If you would like me to, I can submit
> a patch to remove the check entirely. Or maybe only disable it for ARM?
>

Please try the fix I proposed first.



More information about the linux-arm-kernel mailing list