[musl] Re: [PATCH v2] powerpc/64/signal: balance return predictor stack in signal trampoline

[Raoni Fassina Firmino <raoni@linux.ibm.com>]

On Fri, Jan 22, 2021 at 12:27:14PM +0100, AL glibc-alpha wrote:
> * Nicholas Piggin:
> 
> > diff --git a/arch/powerpc/kernel/vdso64/sigtramp.S b/arch/powerpc/kernel/vdso64/sigtramp.S
> > index a8cc0409d7d2..bbf68cd01088 100644
> > --- a/arch/powerpc/kernel/vdso64/sigtramp.S
> > +++ b/arch/powerpc/kernel/vdso64/sigtramp.S
> > @@ -6,6 +6,7 @@
> >   * Copyright (C) 2004 Benjamin Herrenschmuidt (benh@kernel.crashing.org), IBM Corp.
> >   * Copyright (C) 2004 Alan Modra (amodra@au.ibm.com)), IBM Corp.
> >   */
> > +#include <asm/cache.h>		/* IFETCH_ALIGN_BYTES */
> >  #include <asm/processor.h>
> >  #include <asm/ppc_asm.h>
> >  #include <asm/unistd.h>
> > @@ -14,21 +15,17 @@
> >  
> >  	.text
> >  
> > -/* The nop here is a hack.  The dwarf2 unwind routines subtract 1 from
> > -   the return address to get an address in the middle of the presumed
> > -   call instruction.  Since we don't have a call here, we artificially
> > -   extend the range covered by the unwind info by padding before the
> > -   real start.  */
> > -	nop
> >  	.balign 8
> > +	.balign IFETCH_ALIGN_BYTES
> >  V_FUNCTION_BEGIN(__kernel_sigtramp_rt64)
> > -.Lsigrt_start = . - 4
> > +.Lsigrt_start:
> > +	bctrl	/* call the handler */
> >  	addi	r1, r1, __SIGNAL_FRAMESIZE
> >  	li	r0,__NR_rt_sigreturn
> >  	sc
> >  .Lsigrt_end:
> >  V_FUNCTION_END(__kernel_sigtramp_rt64)
> > -/* The ".balign 8" above and the following zeros mimic the old stack
> > +/* The .balign 8 above and the following zeros mimic the old stack
> >     trampoline layout.  The last magic value is the ucontext pointer,
> >     chosen in such a way that older libgcc unwind code returns a zero
> >     for a sigcontext pointer.  */
> 
> As far as I understand it, this breaks cancellation handling on musl and
> future glibc because it is necessary to look at the signal delivery
> location to see if a system call sequence has result in an action, and
> that location is no longer in user code after this change.
> 
> We have a glibc test in preparation of our change, and it started
> failing:
> 
>   Linux 5.10 breaks sigcontext_get_pc on powerpc64
>   <https://sourceware.org/bugzilla/show_bug.cgi?id=27223>
> 
> Isn't it possible to avoid the return predictor desynchronization by
> adding the appropriate hint?

I also caught this regression, I believe it was introduced in the kernel
5.9.

I don't know enough to comment on Florian suggestion, but I am working
on some possible fixes:

On the kernel side we can keep `__kernel_sigtramp_rt64` in the original
place after `bctrl` and add a new symbol so the kernel can jump to the
right place before `bctrl`.  This would ensure backward compatibility.

On the other side, this change exposed how fragile `backtrace()` is to
any changes in the trampoline code, which the libc has no control over
in this case. So maybe there is something that can be improved in how
backtrace decides that the return-address is to the trampoline. My fist
option is to test a range, after `__kernel_sigtramp_rt64` so see if the
address is inside the routine. This would be better if we can know the
size of the function, I know that the vdso.so has this info in the elf
but I don't know if it is exposed to the glibc.

As Nicholas mentioned in his patch, GDB seems to keep working just fine,
it is seems that GDB uses some heuristics to match the surround code of
the return address to identify that it is the trampoline code. So maybe
other option is to do something similar.


o/
Raoni Fassina