Re: [PATCH 4/7] arm: Add ftrace with regs support

From: Abel Vesa
Date: Wed Dec 07 2016 - 07:10:33 EST


On Wed, Dec 07, 2016 at 11:58:24AM +0000, Robin Murphy wrote:
> Hi Abel,
>
> On 06/12/16 17:06, Abel Vesa wrote:
> > This adds __ftrace_regs_caller which, unlike __ftrace_caller,
> > adds register saving/restoring and livepatch handling if
> > the pc register gets modified by klp_ftrace_handler.
> >
> > Signed-off-by: Abel Vesa <abelvesa@xxxxxxxxx>
> > ---
> > arch/arm/kernel/entry-ftrace.S | 49 ++++++++++++++++++++++++++++++++++++++++++
> > 1 file changed, 49 insertions(+)
> >
> > diff --git a/arch/arm/kernel/entry-ftrace.S b/arch/arm/kernel/entry-ftrace.S
> > index c73c403..b6ada5c 100644
> > --- a/arch/arm/kernel/entry-ftrace.S
> > +++ b/arch/arm/kernel/entry-ftrace.S
> > @@ -92,6 +92,46 @@
> > 2: mcount_exit
> > .endm
> >
> > +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> > +
> > +.macro __ftrace_regs_caller suffix
> > +
> > + stmdb sp!, {r0-r15}
>
> As the kbuild robot reported, you can't do this. For ARM, it's unknown
> what the value stored for r13 will be, and for a Thumb-2 kernel the
> whole instruction is flat out unpredictable (i.e. it might just undef or
> anything).
>
> > + mov r3, sp
> > +
> > + ldr r10, [sp, #60]
> > +
> > + mcount_get_lr r1 @ lr of instrumented func
> > + mcount_adjust_addr r0, lr @ instrumented function
> > +
> > + .globl ftrace_regs_call\suffix
> > +ftrace_regs_call\suffix:
> > + bl ftrace_stub
> > +
> > +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> > + .globl ftrace_regs_graph_call\suffix
> > +ftrace_regs_graph_call\suffix:
> > + mov r0, r0
> > +#endif
> > +#ifdef CONFIG_LIVEPATCH
> > + ldr r0, [sp, #60]
> > + cmp r0, r10
> > + beq ftrace_regs_caller_end
> > + ldmia sp!, {r0-r12}
> > + add sp, sp, #8
> > + ldmia sp!, {r11}
> > + sub sp, r12, #16
> > + str r11, [sp, #12]
> > + ldmia sp!, {r11, r12, lr, pc}
> > +#endif
> > +ftrace_regs_caller_end\suffix:
> > + ldmia sp!, {r0-r14}
>
> Again, the value of SP at this point is now unknown (regardless of
> whether what the value on memory was correct or not). Not good.
>
> Either use non-writeback addressing modes, or avoid saving/restoring SP
> at all (AFAICS it isn't necessary, since if the SP was changed in
> between, you then wouldn't know where to restore the saved registers
> from anyway).
>
> Robin.
>
Yep, as I said in the cover letter, I'll try to fix that in the next
iteration of this patch. You're right, sp doesn't need to be pushed or
popped. I'll send a another patch series which will include a fix for
that tomorrow, latest.

Thanks.
> > + add sp, sp, #4
> > + mov pc, lr
> > +.endm
> > +
> > +#endif
> > +
> > .macro __ftrace_caller suffix
> > mcount_enter
> >
> > @@ -212,6 +252,15 @@ UNWIND(.fnstart)
> > __ftrace_caller
> > UNWIND(.fnend)
> > ENDPROC(ftrace_caller)
> > +
> > +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> > +ENTRY(ftrace_regs_caller)
> > +UNWIND(.fnstart)
> > + __ftrace_regs_caller
> > +UNWIND(.fnend)
> > +ENDPROC(ftrace_regs_caller)
> > +#endif
> > +
> > #endif
> >
> > #ifdef CONFIG_FUNCTION_GRAPH_TRACER
> >
>