[PATCH v6 0/5] glibc: Add SFrame support for stack backtracing

[Adhemerval Zanella Netto]

On 10/06/25 04:33, claudiu.zissulescu-ianculescu@oracle.com wrote:
> From: Claudiu Zissulescu <claudiu.zissulescu-ianculescu@oracle.com>
> 
> Hello,
> 
> Please find a new update of glibc SFrame stack backtracer support.
> 
> What is new:
>  - Update SFrame stack backtracer to the latest developments present on binutils.
>  - all SFrame stack backtracer errors will trigger Dwarf unwinder.
> 
> Older comments:
>  - Fix Bigendian handling in 32b systems.
>  - Added a new target dependent hook to detect the outermost frame.
>  - The SFrame backtracer is falling back to Dwarf unwinder if it
>    detects any posibility of a truncate trace.
>  - Various typo/text fixes
> 
> ABI impact notes:
> * Changes in 'struct dl_find_object'
>  - A new member 'void *dlfo_sframe' is inserted to the structure.
>  
>  - The member '__dlfo_reserved' is changed, removing one element, which
>    is a decrease of 64bits.
>  
>  - The size of the type has not changed, as the new added field (i.e.,
>    dlfo_sframe) size is of the same size as of the one element removed
>    from struct memeber __dlfo_reserved. For 32bit systems, a padding is
>    added to compensate for pointer size change.
> 
>  - A new macro DLF_SFRAME is introduced.  It helps controlling the
>    newly introduced flag bit in dlfo_flags member. This may implies
>    changing specific ABI tests (e.g. glibc: update
>    conform/data/dlfcn.h-data file)
> 
>  - The function 'dl_find_object' is called within libc' backtrace
>    routine using GLRO macro, thus avoinding any PLT calls.
>  
> * Possible issues
>  - Automatic ABI checkers may mark this change as an ABI change, thus,
>    human review and updating of the scripts is required.
>  
>  - For unknown programs which directly access the '__dlfo_reserved'
>    field, this change is incompatible. As note, this field should not
>    be used by anyone as it is reserved for further changes of the
>    dl_find_object structure.
> 
> Other changes and observations:
> 
>  - The getPC and getSP helper functions are moved to sframe.c file as
>    they may require machine dependent changes like for s390.
> 
>  - The dlfo_sframe member is always present.
> 
>  - The dlfo_flags member is used, bit 0 signalize the presence of
>    sframe information.
> 
>  - Update sframe_fre_get_cfa_offset and sframe_fre_get_ra_offset
>    functions to set an error pointer if return offset is unavailable.
> 
>  - Updated NEWS file, and other documents.
> 
>  - Cleanup Aarch64 specific function.
> 
>  - Compiling the x86 tst-quadmod1.S test file with --gsframe will hit
>    binutils' assembler Bug #32879.
> 
>  - Fixed 32bit compilation issues.

It is not clear to me how this was tested, since there is no regression
tests.  Is it because with --enable-sframe the -Wa,--gsframe will be
used by default for all tests and thus the backtrace ones will use
SFRAME? 

If it were the case I think we will need to add tests on backtrace for
*both* cases, by disabling -Wa,--gsframe on some and adding
-Wa,--gsframe on another to get full coverage independent whether
--enable-sframe is used.

> 
> Thank you,
> Claudiu
> 
> Claudiu Zissulescu (5):
>   elf: Add SFrame support to _dl_find_object function
>   x86: Add SFrame support for x86 architecture
>   aarch64: Add SFrame support for aarch64 architecture
>   elf: Add SFrame stack tracing
>   configure: Add --enable-sframe option
> 
>  INSTALL                                       |   5 +
>  Makeconfig                                    |   8 +-
>  NEWS                                          |   4 +
>  bits/dl_find_object.h                         |   3 +
>  config.make.in                                |   1 +
>  configure                                     | 129 ++++
>  configure.ac                                  |  44 ++
>  conform/data/dlfcn.h-data                     |   3 +
>  debug/backtrace.c                             |  33 +-
>  dlfcn/dlfcn.h                                 |   7 +-
>  elf/dl-find_object.h                          |  80 ++-
>  manual/dynlink.texi                           |  25 +-
>  manual/install.texi                           |   5 +
>  sysdeps/arm/bits/dl_find_object.h             |   1 +
>  sysdeps/generic/Makefile                      |   1 +
>  sysdeps/generic/sframe-read.c                 | 567 ++++++++++++++++++
>  sysdeps/generic/sframe-read.h                 | 106 ++++
>  sysdeps/generic/sframe.c                      | 187 ++++++
>  sysdeps/generic/sframe.h                      | 367 ++++++++++++
>  sysdeps/generic/uw-sigframe.h                 |  17 +
>  sysdeps/unix/sysv/linux/aarch64/uw-sigframe.h |  75 +++
>  sysdeps/unix/sysv/linux/x86/uw-sigframe.h     |  74 +++
>  sysdeps/x86/bits/dl_find_object.h             |   1 +
>  23 files changed, 1731 insertions(+), 12 deletions(-)
>  create mode 100644 sysdeps/generic/sframe-read.c
>  create mode 100644 sysdeps/generic/sframe-read.h
>  create mode 100644 sysdeps/generic/sframe.c
>  create mode 100644 sysdeps/generic/sframe.h
>  create mode 100644 sysdeps/generic/uw-sigframe.h
>  create mode 100644 sysdeps/unix/sysv/linux/aarch64/uw-sigframe.h
>  create mode 100644 sysdeps/unix/sysv/linux/x86/uw-sigframe.h
>