Re: [PATCH v2 00/22] nolibc: usability improvements (errno, environ, auxv)

[Paul E. McKenney]

On Tue, Jan 10, 2023 at 08:24:12AM +0100, Willy Tarreau wrote:
> Hello Paul,
> 
> this 3rd series aims at generally improving usability and maintainance
> of nolibc. It needs to be applied on top of the s390 series:
> 
>   https://lore.kernel.org/lkml/20230109080910.26594-1-w@xxxxxx/
> 
> - first, I've encountered remaining problems related to section
>   reordering happening at certain optimization levels and that is also
>   arch-dependent. We were fortunate they didn't appear in rcutorture,
>   but I could reproduce them with ARM in Thumb mode at -O0. The problem
>   is that our out-of-block asm() statement changes the current section
>   to ".text" without the compiler knowing. Thus the compiler may believe
>   it's still in .bss and emit variables immediately after (e.g. errno),
>   which end up in the wrong section. Switching the section at the end
>   to .bss doesn't work either because at -Os I was seeing sys_brk()
>   placed immediately after and crashing as .bss is not executable. The
>   only safe solution is to turn _start to real functions. This was
>   tested on all archs at -O0,-O1,-O2,-O3,-Os and all of them now work.
> 
> - second, thumb mode support was in complete on ARM. Only thumb-2 was
>   supported and depending on how the toolchain is configured, passing
>   "-mthumb" would result in thumb-1 (if armv4/5 was the default) or
>   thumb-2 (when armv7 was the default). I discovered this when first
>   trying the toolchains from kernel.org because mine works in v7 by
>   default, hence thumb-2. The change only replaces a few instructions
>   that are not available in thumb-1 by their compatible equivalent. In
>   addition, thumb cannot be used at -O0 or with frame pointers in general
>   because register r7 is the frame pointer there, and cannot be assigned
>   by the compiler. That's bad because r7 carries the syscall number. Now
>   when thumb is detected, we simply use a slightly larger setup code
>   which uses r6 and swaps it with r7 when performing the call.
> 
> - third, the definitions of the (possibly wrong) arch-specific O_* values
>   were dropped in favor of those coming from asm/fcntl.h. Not only these
>   ones are correct, but doing so will avoid build redefinition warnings
>   should the file be included for whatever other reason.
> 
> - the errno, environ and the auxiliary vector were really a pain to
>   use. errno was declared as a static variable, showing a different
>   one to each build unit. environ had to be declared by the application,
>   and the auxv had to be both declared and found by the application if
>   needed. All three of them have now been declared as weak symbols,
>   and environ and _auxv are setup during startup, so that only one
>   instance of them exists across the whole binary, and that code
>   currently declaring them continues to work. This now means that code
>   not using them will not optimize them away anymore, but let's face
>   it, errno was always used and no relevant application manages to get
>   rid of .bss, so the amount of extra space is really just 8-16 bytes
>   total for a much better simplicity for the user.
> 
> - getauxval() and getpagesize() were added by Ammar Faizi (along with
>   the associated selftests).
> 
> This was tested on arm64/armv5/armv7/thumb1/thumb2/i386/x86_64/mips/riscv
> and s390 at all optimization levels. I could also verify that my original
> preinit code continues to build and work fine, so please consider queuing
> it.

Queued and pushed, thank you!

							Thanx, Paul

> Thank you!
> Willy
> 
> Changes since v1:
> - added missing s-o-b on the last 3 patches
> 
> Ammar Faizi (3):
>   nolibc/stdlib: Implement `getauxval(3)` function
>   nolibc/sys: Implement `getpagesize(2)` function
>   selftests/nolibc: Add `getpagesize(2)` selftest
> 
> Sven Schnelle (2):
>   tools/nolibc: export environ as a weak symbol on s390
>   tools/nolibc: add auxiliary vector retrieval for s390
> 
> Willy Tarreau (17):
>   tools/nolibc: make compiler and assembler agree on the section around
>     _start
>   tools/nolibc: enable support for thumb1 mode for ARM
>   tools/nolibc: support thumb mode with frame pointers on ARM
>   tools/nolibc: remove local definitions of O_* flags for open/fcntl
>   tools/nolibc: make errno a weak symbol instead of a static one
>   tools/nolibc: export environ as a weak symbol on x86_64
>   tools/nolibc: export environ as a weak symbol on i386
>   tools/nolibc: export environ as a weak symbol on arm64
>   tools/nolibc: export environ as a weak symbol on arm
>   tools/nolibc: export environ as a weak symbol on mips
>   tools/nolibc: export environ as a weak symbol on riscv
>   tools/nolibc: add auxiliary vector retrieval for i386
>   tools/nolibc: add auxiliary vector retrieval for x86_64
>   tools/nolibc: add auxiliary vector retrieval for arm64
>   tools/nolibc: add auxiliary vector retrieval for arm
>   tools/nolibc: add auxiliary vector retrieval for riscv
>   tools/nolibc: add auxiliary vector retrieval for mips
> 
>  tools/include/nolibc/arch-aarch64.h          |  52 +++----
>  tools/include/nolibc/arch-arm.h              | 138 ++++++++++++-------
>  tools/include/nolibc/arch-i386.h             |  60 ++++----
>  tools/include/nolibc/arch-mips.h             |  79 ++++++-----
>  tools/include/nolibc/arch-riscv.h            |  62 +++++----
>  tools/include/nolibc/arch-s390.h             |  70 +++++-----
>  tools/include/nolibc/arch-x86_64.h           |  52 +++----
>  tools/include/nolibc/errno.h                 |   4 +-
>  tools/include/nolibc/stdlib.h                |  27 ++++
>  tools/include/nolibc/sys.h                   |  22 +++
>  tools/testing/selftests/nolibc/nolibc-test.c |  30 ++++
>  11 files changed, 363 insertions(+), 233 deletions(-)
> 
> -- 
> 2.17.5
>