On Fri, Mar 21, 2025 at 02:30:40PM -0700, Mark Millard wrote: > Under: > > # uname -apKU > FreeBSD 7950X3D-ZFS 15.0-CURRENT FreeBSD 15.0-CURRENT #4 > main-n275926-a54a240c1b57-dirty: Thu Mar 13 00:44:25 PDT 2025 > root@7950X3D-ZFS:/usr/obj/BUILDs/main-ZNV4-nodbg-clang/usr/main-src/amd64.amd64/sys/GENERIC-NODBG > amd64 amd64 1500034 1500034 > > Looking, I see: > > # man -K "libc, -l" | more > /usr/share/man/man2/_Fork.2.gz: Standard C Library (libc, -lc) > /usr/share/man/man2/__syscall.2.gz: Standard C Library (libc, -lc) > /usr/share/man/man2/_exit.2.gz: Standard C Library (libc, -lc) > /usr/share/man/man2/_umtx_op.2.gz: Standard C Library (libc, -lc) > . . . > > But: > > # man -K "libsys, -l" | more > # > > (So nothing references libsys in a similar way.) > > # readelf -drs /lib/libsys.so.7 | sort -k8,8 | grep "\<_*errno\>" > 633: 000000000001d328 4 OBJECT GLOBAL DEFAULT 28 errno@FBSD_1.0 (2) > > # readelf -drs /lib/libc.so.7 | sort -k8,8 | grep "\<_*errno\>" > # > > # man errno > INTRO(2) FreeBSD System Calls Manual INTRO(2) > > NAME > intro, errno – introduction to system calls and their error numbers > > LIBRARY > Standard C Library (libc, -lc) > > SYNOPSIS > #include <sys/syscall.h> > #include <errno.h> > . . . > > (So "libc, -lc" is still referenced for errno in the man page.) > > (errno is just used as an example above.)
First, libsys does not participate in the 'official' ABI of the FreeBSD userspace. This is indicated, in particular, by the man pages, which clearly and correctly state that syscall stubs symbols are supposed to be provided by libc, and apps must link to libc to get them. Libsys is the internal implementation detail which ABI is the subject to change if it is ever exposed in more formal way. I particularly do not think that traditional Unix ABI/API with 0/-1 and errno is good fit for libsys. Second, errno itself is very complicated and irrelevant thing. The symbol really should not be referenced by any ABI-compliant binary or dso, and it is provided only for formal ABI compat. Since long time, errno is defined as *__error(), and this is the current ABI for it. Since errno is an object and exported from dso (lets put aside libc vs libsys ATM), the reference to it ends up with the same sized object allocated in the referencing object, and with the COPY relocation. Now, and this is the worst part, it is indeed impossible to keep ABI of errno (in the part that the symbol is copied from libc and not libsys) with the libsys split, and still allow the main thread to directly access the errno location to get correct syscall error values. As result, errno is copied from libsys. So formally we somewhat break ABI, but I do think that this is the least possible breakage to still have (future) benefits of libsys.
