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

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

[Florian Weimer via Libc-alpha]

* 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?

Thanks,
Florian
-- 
Red Hat GmbH, https://de.redhat.com/ , Registered seat: Grasbrunn,
Commercial register: Amtsgericht Muenchen, HRB 153243,
Managing Directors: Charles Cachera, Brian Klemm, Laurie Krebs, Michael O'Neill

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

[Rich Felker]

On Fri, Jan 22, 2021 at 12:27:14PM +0100, Florian Weimer 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?

Maybe I'm missing something but I don't see how this would break musl;
we just inspect the PC in the mcontext, which I don't see any changes
to and which should still point to the next instruction of the
interrupted context. I don't have a test environment though so I'll
have to wait for feedback from ppc users to be sure. Are there any
further details on how it's breaking glibc?

Rich

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

[Raoni Fassina Firmino via Libc-alpha]

On Fri, Jan 22, 2021 at 09:44:05AM -0500, Rich Felker wrote:
> Maybe I'm missing something but I don't see how this would break musl;
> we just inspect the PC in the mcontext, which I don't see any changes
> to and which should still point to the next instruction of the
> interrupted context. I don't have a test environment though so I'll
> have to wait for feedback from ppc users to be sure. Are there any
> further details on how it's breaking glibc?

For glibc, backtrace() compares the return-address from each stack frame
to the value of `__kernel_sigtramp_rt64` to identify the frame with the
mcontext information, but now the return-address is not the start of the
routine, but the middle of it, so it fails to catch this special frame.


o/
Raoni Fassina

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

[Rich Felker]

On Fri, Jan 22, 2021 at 03:19:22PM -0300, Raoni Fassina Firmino wrote:
> On Fri, Jan 22, 2021 at 09:44:05AM -0500, Rich Felker wrote:
> > Maybe I'm missing something but I don't see how this would break musl;
> > we just inspect the PC in the mcontext, which I don't see any changes
> > to and which should still point to the next instruction of the
> > interrupted context. I don't have a test environment though so I'll
> > have to wait for feedback from ppc users to be sure. Are there any
> > further details on how it's breaking glibc?
> 
> For glibc, backtrace() compares the return-address from each stack frame
> to the value of `__kernel_sigtramp_rt64` to identify the frame with the
> mcontext information, but now the return-address is not the start of the
> routine, but the middle of it, so it fails to catch this special frame.

Is there a reason it's backtracing rather than just looking at the
interrupted context (pointed to by the third argument to the signal
handler)?

Rich

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

[Raoni Fassina Firmino via Libc-alpha]

On Fri, Jan 22, 2021 at 01:31:27PM -0500, Rich Felker wrote:
> On Fri, Jan 22, 2021 at 03:19:22PM -0300, Raoni Fassina Firmino wrote:
> > On Fri, Jan 22, 2021 at 09:44:05AM -0500, Rich Felker wrote:
> > > Maybe I'm missing something but I don't see how this would break musl;
> > > we just inspect the PC in the mcontext, which I don't see any changes
> > > to and which should still point to the next instruction of the
> > > interrupted context. I don't have a test environment though so I'll
> > > have to wait for feedback from ppc users to be sure. Are there any
> > > further details on how it's breaking glibc?
> > 
> > For glibc, backtrace() compares the return-address from each stack frame
> > to the value of `__kernel_sigtramp_rt64` to identify the frame with the
> > mcontext information, but now the return-address is not the start of the
> > routine, but the middle of it, so it fails to catch this special frame.
> 
> Is there a reason it's backtracing rather than just looking at the
> interrupted context (pointed to by the third argument to the signal
> handler)?

The regression is exposed in the backtrace() routine. More precisely,
when the backtrace() is called from inside a signal handling. What I
described is the way backtrace() uses to identify this special
situation. What is failling in glibc is the test for this.


o/
Raoni Fassina

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

[Raoni Fassina Firmino via Libc-alpha]

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