Re: [PATCHv2] x86: Use pad in pthread_unwind_buf to preserve shadow stack register

[Carlos O'Donell <carlos@redhat.com>]

Lets call this v2. Subject adjusted. Please keep incrementing the version
number on your patches to make the review easier by myself and others.

On 04/12/2018 06:50 PM, H.J. Lu wrote:
> From 1ce2b4f199070a63d9a60bd9b7ca9e33013d4208 Mon Sep 17 00:00:00 2001
> From: "H.J. Lu" <hjl.tools@gmail.com>
> Date: Sat, 24 Feb 2018 17:23:54 -0800
> Subject: [PATCH] x86: Use pad in pthread_unwind_buf to preserve shadow stack
>  register
> 
> The pad array in struct pthread_unwind_buf is used by setjmp to save
> shadow stack register.  We assert that size of struct pthread_unwind_buf
> is no less than offset of shadow stack pointer + shadow stack pointer
> size.
> 
> Since functions, like LIBC_START_MAIN, START_THREAD_DEFN as well as
> these with thread cancellation, call setjmp, but never return after
> __libc_unwind_longjmp, __libc_unwind_longjmp, which is defined as
> __libc_longjmp on x86, doesn't need to restore shadow stack register.
> __libc_longjmp, which is a private interface for thread cancellation
> implementation in libpthread, is changed to call __longjmp_cancel,
> instead of __longjmp.  __longjmp_cancel is a new internal function
> in libc, which is similar to __longjmp, but doesn't restore shadow
> stack register.
> 
> The compatibility longjmp and siglongjmp in libpthread.so are changed
> to call __libc_siglongjmp, instead of __libc_longjmp, so that they will
> restore shadow stack register.
> 
> Tested with build-many-glibcs.py.
> 
> 	* nptl/pthread_create.c (START_THREAD_DEFN): Clear previous
> 	handlers after setjmp.
> 	* setjmp/longjmp.c (__libc_longjmp): Don't define alias if
> 	defined.
> 	* sysdeps/unix/sysv/linux/x86/setjmpP.h: Include
> 	<libc-pointer-arith.h>.
> 	(_JUMP_BUF_SIGSET_BITS_PER_WORD): New.
> 	(_JUMP_BUF_SIGSET_NSIG): Changed to 96.
> 	(_JUMP_BUF_SIGSET_NWORDS): Changed to use ALIGN_UP and
> 	_JUMP_BUF_SIGSET_BITS_PER_WORD.
> 	* sysdeps/x86/Makefile (sysdep_routines): Add __longjmp_cancel.
> 	* sysdeps/x86/__longjmp_cancel.S: New file.
> 	* sysdeps/x86/longjmp.c: Likewise.
> 	* sysdeps/x86/nptl/pt-longjmp.c: Likewise.
> ---
>  nptl/pthread_create.c                 | 18 +++++--
>  setjmp/longjmp.c                      |  2 +
>  sysdeps/unix/sysv/linux/x86/setjmpP.h | 12 +++--
>  sysdeps/x86/Makefile                  |  4 ++
>  sysdeps/x86/__longjmp_cancel.S        | 20 ++++++++
>  sysdeps/x86/longjmp.c                 | 45 ++++++++++++++++
>  sysdeps/x86/nptl/pt-longjmp.c         | 97 +++++++++++++++++++++++++++++++++++
>  7 files changed, 191 insertions(+), 7 deletions(-)
>  create mode 100644 sysdeps/x86/__longjmp_cancel.S
>  create mode 100644 sysdeps/x86/longjmp.c
>  create mode 100644 sysdeps/x86/nptl/pt-longjmp.c
> 
> diff --git a/nptl/pthread_create.c b/nptl/pthread_create.c
> index caaf07c134..8b1f06599d 100644
> --- a/nptl/pthread_create.c
> +++ b/nptl/pthread_create.c
> @@ -427,12 +427,24 @@ START_THREAD_DEFN
>       compilers without that support we do use setjmp.  */
>    struct pthread_unwind_buf unwind_buf;
>  
> -  /* No previous handlers.  */
> +  int not_first_call;
> +  not_first_call = setjmp ((struct __jmp_buf_tag *) unwind_buf.cancel_jmp_buf);
> +
> +  /* No previous handlers.  NB: This must be done after setjmp since
> +     the private space in the unwind jump buffer may overlap space
> +     used by setjmp to store extra architecture-specific information
> +     which is never be used by the cancellation-specific

s/be//g

> +     __libc_unwind_longjmp.
> +
> +     The private space is allowed to overlap because the unwinder never
> +     has to return through any of the jumped-to call frames, and thus
> +     only a minimum amount of saved data need be stored, and for example,
> +     need not include the process signal mask information. This is all
> +     an optimization to reduce stack usage when pushing cancellation
> +     handlers.  */
>    unwind_buf.priv.data.prev = NULL;
>    unwind_buf.priv.data.cleanup = NULL;
>  
> -  int not_first_call;
> -  not_first_call = setjmp ((struct __jmp_buf_tag *) unwind_buf.cancel_jmp_buf);

OK.

>    if (__glibc_likely (! not_first_call))
>      {
>        /* Store the new cleanup handler info.  */
> diff --git a/setjmp/longjmp.c b/setjmp/longjmp.c
> index a2a7065a85..453889e103 100644
> --- a/setjmp/longjmp.c
> +++ b/setjmp/longjmp.c
> @@ -40,9 +40,11 @@ __libc_siglongjmp (sigjmp_buf env, int val)
>  }
>  
>  #ifndef __libc_siglongjmp
> +# ifndef __libc_longjmp
>  /* __libc_longjmp is a private interface for cancellation implementation
>     in libpthread.  */
>  strong_alias (__libc_siglongjmp, __libc_longjmp)
> +# endif

OK.

>  weak_alias (__libc_siglongjmp, _longjmp)
>  weak_alias (__libc_siglongjmp, longjmp)
>  weak_alias (__libc_siglongjmp, siglongjmp)
> diff --git a/sysdeps/unix/sysv/linux/x86/setjmpP.h b/sysdeps/unix/sysv/linux/x86/setjmpP.h
> index c0ed767a0d..24f87da204 100644
> --- a/sysdeps/unix/sysv/linux/x86/setjmpP.h
> +++ b/sysdeps/unix/sysv/linux/x86/setjmpP.h
> @@ -20,13 +20,17 @@
>  #define	_SETJMPP_H	1
>  
>  #include <bits/types/__sigset_t.h>
> +#include <libc-pointer-arith.h>
>  
> -/* The biggest signal number + 1.  As of kernel 4.14, x86 _NSIG is 64.
> -   Define it to 513 to leave some rooms for future use.  */
> -#define _JUMP_BUF_SIGSET_NSIG	513
> +/* Number of bits per long.  */
> +#define _JUMP_BUF_SIGSET_BITS_PER_WORD (8 * sizeof (unsigned long int))
> +/* The biggest signal number.  As of kernel 4.14, x86 _NSIG is 64.
> +   Define it to 96 to leave some rooms for future use.  */

Why 96?

Sure on x86_64 the cancel jump buf has 4 x void*'s and each is 64-bits,
so you have 128 signals worth of space and then the shadow stack pointer.

Does this work on x32?

On x32 you have only 4 void*'s in the private pad.

Your presently structured sigset_t looks like this:

typedef union
  {
    __sigset_t __saved_mask_compat;
    struct
      {
        __jmp_buf_sigset_t __saved_mask;
        /* Used for shadow stack pointer.  */
        unsigned long int __shadow_stack_pointer;
      } __saved;
  } __jmpbuf_arch_t;

Which means you have a sigset_t *before* the __shadow_stack_pointer.

On x32 to save a 64-bit shadow stack pointer, you'll only have 2 x 32-bit
words left? That's only space for 64 signals?

Are you counting on one additional 32-bit word of padding between the
__mask_was_saved and the rest of the long arguments?

If so, then this still needs spelling out in an a comment why 96 signals
works on both i686, x32, and x86_64.

Also it should be explained if 96 is the *maximum* we can do with the current
layout, or if more is possible. In which case picking 96 is *not* arbitrary.

> +#define _JUMP_BUF_SIGSET_NSIG	96
>  /* Number of longs to hold all signals.  */
>  #define _JUMP_BUF_SIGSET_NWORDS \
> -  ((_JUMP_BUF_SIGSET_NSIG - 1 + 7) / (8 * sizeof (unsigned long int)))
> +  (ALIGN_UP (_JUMP_BUF_SIGSET_NSIG, _JUMP_BUF_SIGSET_BITS_PER_WORD) \
> +   / _JUMP_BUF_SIGSET_BITS_PER_WORD)
>  
>  typedef struct
>    {
> diff --git a/sysdeps/x86/Makefile b/sysdeps/x86/Makefile
> index 0d0326c21a..d25d6f0ae4 100644
> --- a/sysdeps/x86/Makefile
> +++ b/sysdeps/x86/Makefile
> @@ -8,3 +8,7 @@ sysdep-dl-routines += dl-get-cpu-features
>  tests += tst-get-cpu-features
>  tests-static += tst-get-cpu-features-static
>  endif
> +
> +ifeq ($(subdir),setjmp)
> +sysdep_routines += __longjmp_cancel
> +endif
> diff --git a/sysdeps/x86/__longjmp_cancel.S b/sysdeps/x86/__longjmp_cancel.S
> new file mode 100644
> index 0000000000..b57dbfa376
> --- /dev/null
> +++ b/sysdeps/x86/__longjmp_cancel.S
> @@ -0,0 +1,20 @@
> +/* __longjmp_cancel for x86.
> +   Copyright (C) 2018 Free Software Foundation, Inc.
> +   This file is part of the GNU C Library.
> +
> +   The GNU C Library is free software; you can redistribute it and/or
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <http://www.gnu.org/licenses/>.  */
> +
> +#define __longjmp __longjmp_cancel
> +#include <__longjmp.S>

OK.

> diff --git a/sysdeps/x86/longjmp.c b/sysdeps/x86/longjmp.c
> new file mode 100644
> index 0000000000..a53f31e1dd
> --- /dev/null
> +++ b/sysdeps/x86/longjmp.c
> @@ -0,0 +1,45 @@
> +/* __libc_siglongjmp for x86.
> +   Copyright (C) 2018 Free Software Foundation, Inc.
> +   This file is part of the GNU C Library.
> +
> +   The GNU C Library is free software; you can redistribute it and/or
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <http://www.gnu.org/licenses/>.  */
> +
> +#define __libc_longjmp __redirect___libc_longjmp
> +#include <setjmp/longjmp.c>
> +#undef __libc_longjmp
> +
> +extern void __longjmp_cancel (__jmp_buf __env, int __val)
> +     __attribute__ ((__noreturn__)) attribute_hidden;
> +
> +/* Since __libc_longjmp is a private interface for cancellation
> +   implementation in libpthread, there is no need to restore shadow
> +   stack register.  */
> +
> +void
> +__libc_longjmp (sigjmp_buf env, int val)
> +{
> +  /* Perform any cleanups needed by the frames being unwound.  */
> +  _longjmp_unwind (env, val);
> +
> +  if (env[0].__mask_was_saved)
> +    /* Restore the saved signal mask.  */
> +    (void) __sigprocmask (SIG_SETMASK,
> +			  (sigset_t *) &env[0].__saved_mask,
> +			  (sigset_t *) NULL);
> +
> +  /* Call the machine-dependent function to restore machine state
> +     without shadow stack.  */
> +  __longjmp_cancel (env[0].__jmpbuf, val ?: 1);
> +}

OK.

> diff --git a/sysdeps/x86/nptl/pt-longjmp.c b/sysdeps/x86/nptl/pt-longjmp.c
> new file mode 100644
> index 0000000000..7eb8651cfe
> --- /dev/null
> +++ b/sysdeps/x86/nptl/pt-longjmp.c
> @@ -0,0 +1,97 @@
> +/* ABI compatibility for 'longjmp' and 'siglongjmp' symbols in libpthread ABI.
> +   X86 version.
> +   Copyright (C) 18 Free Software Foundation, Inc.
> +   This file is part of the GNU C Library.
> +
> +   The GNU C Library is free software; you can redistribute it and/or
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <http://www.gnu.org/licenses/>.  */
> +
> +/* <nptl/descr.h> has
> +
> +struct pthread_unwind_buf
> +{
> +  struct
> +  {
> +    __jmp_buf jmp_buf;
> +    int mask_was_saved;
> +  } cancel_jmp_buf[1];
> +
> +  union
> +  {
> +    void *pad[4];
> +    struct
> +    {
> +      struct pthread_unwind_buf *prev;
> +      struct _pthread_cleanup_buffer *cleanup;
> +      int canceltype;
> +    } data;
> +  } priv;
> +};
> +
> +  The pad array in struct pthread_unwind_buf is used by setjmp to save
> +  shadow stack register.  Assert that size of struct pthread_unwind_buf
> +  is no less than offset of shadow stack pointer plus shadow stack
> +  pointer size.
> +
> +  NB: setjmp is called in libpthread to save shadow stack register.  But
> +  __libc_unwind_longjmp doesn't restore shadow stack register since they
> +  never return after longjmp.  */

Suggest:
~~~
The pad array in struct pthread_unwind_buf is used by setjmp to save
the shadow stack register.  Assert that the size of struct pthread_unwind_buf
is no less than the offset of the shadow stack pointer plus the shadow stack
pointer size.

NB: We use setjmp in thread cancellation and this saves the shadow stack
register, but __libc_unwind_longjmp doesn't restore the shadow stack register
since cancellation never returns after longjmp.
~~~

> +
> +#include <pthreadP.h>
> +#include <jmp_buf-ssp.h>
> +
> +#ifdef __x86_64__
> +# define SHADOW_STACK_POINTER_SIZE 8
> +#else
> +# define SHADOW_STACK_POINTER_SIZE 4
> +#endif
> +
> +_Static_assert ((sizeof (struct pthread_unwind_buf)
> +		 >= (SHADOW_STACK_POINTER_OFFSET
> +		     + SHADOW_STACK_POINTER_SIZE)),
> +		"size of struct pthread_unwind_buf < "
> +		"(SHADOW_STACK_POINTER_OFFSET + SHADOW_STACK_POINTER_SIZE)");

This assertion is too loose.

The assertion you need is that the shadow stack pointer itself falls within
the range of the private padding. This would have caught the previous bug
with the rounded up size.

Please adjust the assertion to be as tight as possible or add new assertions.

Completely untested, but just to show what I'm thinking:

_Static_assert ((offsetof (struct pthread_unwind_buf, cancel_jump_buf[0].mask_was_saved)
		 + sizeof (int) < SHADOW_STACK_POINTER_OFFSET
		 && (offsetof (struct pthread_unwind_buf, priv.pad[4])
		    > (SHADOW_STACK_POINTER_OFFSET + SHADOW_STACK_POINTER_SIZE),
		"Shadow stack pointer is not within private storage of pthread_unwind_buf.");

> +
> +#include <shlib-compat.h>
> +
> +/* libpthread once had its own longjmp (and siglongjmp alias), though there
> +   was no apparent reason for it.  There is no use in having a separate
> +   symbol in libpthread, but the historical ABI requires it.  For static
> +   linking, there is no need to provide anything here--the libc version
> +   will be linked in.  For shared library ABI compatibility, there must be
> +   longjmp and siglongjmp symbols in libpthread.so.
> +
> +   With an IFUNC resolver, it would be possible to avoid the indirection,
> +   but the IFUNC resolver might run before the __libc_longjmp symbol has
> +   been relocated, in which case the IFUNC resolver would not be able to
> +   provide the correct address.  */
> +
> +#if SHLIB_COMPAT (libpthread, GLIBC_2_0, GLIBC_2_22)
> +
> +static void __attribute__ ((noreturn, used))
> +longjmp_compat (jmp_buf env, int val)
> +{
> +  /* NB: We call __libc_siglongjmp,  instead of __libc_longjmp, since
> +     __libc_longjmp is a private interface for cancellation which
> +     doesn't restore shadow stack register.  */
> +  __libc_siglongjmp (env, val);
> +}
> +
> +strong_alias (longjmp_compat, longjmp_alias)
> +compat_symbol (libpthread, longjmp_alias, longjmp, GLIBC_2_0);
> +
> +strong_alias (longjmp_alias, siglongjmp_alias)
> +compat_symbol (libpthread, siglongjmp_alias, siglongjmp, GLIBC_2_0);
> +
> +#endif
> -- 2.14.3

Looking forward to a v3.

-- 
Cheers,
Carlos.