With linker sections GC enabled, we get a test failure in `nlist.c`:
```
nlist: nlist.c:72: main: Assertion `rc == 0' failed.
```
This turns out to be because several sections used by the test can be discarded:
```
ld: removing unused section '.text.func_pub' in file 'nlist.o'
ld: removing unused section '.bss.data_pub_uninit' in file 'nlist.o'
ld: removing unused section '.data.data_pub_init' in file 'nlist.o'
```
Reproduced with `CFLAGS="-Og -fdata-sections -ffunction-sections"` and
`LDFLAGS="-Wl,-O1 -Wl,--as-needed -Wl,--gc-sections -Wl,-z,start-stop-gc"`.
Additionally, `LDFLAGS="... -Wl,--print-gc-sections"` can help with diagnosing
which sections get removed.
We already mark these symbols as `used`, but we need `retain` [0] for them
to survive linker GC too.
[0] https://releases.llvm.org/18.1.0/tools/lld/docs/ELF/start-stop-gc.html#annotate-c-identifier-name-sectionsCloses: !29
Signed-off-by: Sam James <sam@gentoo.org>
Signed-off-by: Guillem Jover <guillem@hadrons.org>
Handle the three potential system scenarios:
- system time_t is time64
- system time_t is time32 and supports time64
- system time_t is time32 and does not support time64
Add the explicit time32 and time64 functions when necessary and map
them accordingly for each of these cases.
Explicitly select what to include as part of the target ABI, instead of
letting autoconfiguration potentially break ABI if the system grows
functionality provided by the library.
Make almost all the library selectable per target. Do not install manual
pages for interfaces not included in the library. Control inclusion of
symbols in map file via pre-processor macros, and move the comments
describing the ABI selection to configure.ac.
For now the header files are included as is and filtered through
pre-processor conditionals. Eventually they might get switched to be
autogenerated at build time.
On macOS, closefrom() only sets the close-on-exec flag, so we cannot
check whether all file descriptors were closed, which means that if
on entry our file descriptor table was filled after the 4th file
descriptor, then we might fail the assertions for the flags for odd
file descriptors which we expect to be closed.
This can easily happen when running the test suite in parallel mode
with «make -j8 check» for example.
Closes: #23
Some ld(1) implementations, such as the one on AIX, do not support using
/dev/null as the output filename for the compiled object.
Use an actual filename that we will then clean up.
This makes sure we include it when expected, alongside the man pages,
and the test cases, and do not accidentally break the ABI if the system
starts providing such interface.
If the system has configured a lower limit (either soft or hard) on the
number of open file descriptors, the test will fail. Make sure to check
whether we have exceeded that limit and adapt the max number of file
descriptors appropriately.
This code was added to cope with Hurd specific behavior, but it is
causing flakiness on containers on some Linux systems. Only enable
it where it is currently needed to try to get stability back on CI
systems.
Closes: #14
This will mean we cannot use sanitizer support on the Hurd, for which
this function was added to fix the test. But the sanitizer suppression
function attribute is not having any effect, so this is better than
nothing.
This function cannot be easily and (more importantly) correctly ported
without cooperation from the libc stdio layer. We already document that
users should be prepared to have the function not available on some
platforms and that they should ideally switch their code to other
more portable and better interfaces.
Instead of making the build fail, and requiring porters to add
exceptions for something that most probably cannot be ported correctly
anyway, simply print a warning and let it build. This will not be a
regression because on those systems libbsd would have never been built
before.
Prompted-by: Jens Finkhaeuser <jens@finkhaeuser.de>
This means we can add a trailing «\» to every element, so that they
can be removed without requiring modification of other lines, and can
be easily sorted.
Replace the old usage of $(nil) which could possibly end up with junk
added if such variable is ever defined, in the environment.
On the Hurd a small read(3) might end up (indirectly) copying the data
on the stack, which we will end up finding even when we have cleared
the buffer.
To avoid these side effects, we add a new function, that we force not
to be inlined, so that we can reuse the same stack space, that will
blank any possible stack side effects. This should be portable
regardless of stack growing up or down.
Diagnosis-by: Samuel Thibault <sthibault@debian.org>
- Remove trailing spaces.
- Declare file-scope functions and variables static.
- Declare functions with a proper prototype.
- Do not mix declarations and code for C90 conformance.
- Do not compare size_t and ssize_t variables.
This does not work with libcs that do not declare the structure
in a header file, like musl. And gets in the way of supporting
non-fd based streams.
Signed-off-by: Guillem Jover <guillem@hadrons.org>
This reverts commit beafad2657.
This test was already handled later on as part of the BUILD_LIBBSD_CTOR
conditional. Adding it to the unconditional set made the build fail when
the system does not have GNU .init_array section support.
Closes: #9
Analysis-by: Duncan Overbruck <mail@duncano.de>
When using the recent dlsym() based wrapper, we are not requiring any
symbol from libmd, as we resolve those dynamically at run-time. We were
ending up linking against libmd because in another part of the code we
require (depending on the architecture) the SHA512 functions for the
getentropy() local implementation. But that function might be provided
by the system libc on some systems, which means we end up not linking
against libmd at all.
To solve this we go back to the previous simpler solution of linking
directly, which had the main drawback of then making programs fail to
link when not specifying -lmd (on platforms that need it). And then
switch the .so link point from a symlink to a linker script, so that we
can inject the -lmd library as-needed. This is similar to what glibc is
doing.
Fixes: commit 31f034e386
Because some of the symbols are not otherwise referenced, GCC would like
to remove them.
Closes: !14
Signed-off-by: Guillem Jover <guillem@hadrons.org>
Switch from the previous versioned symbol implementation which required
users to also link against the message digest provider explicitly, or
they would fail to find the symbols, to an implementation that loads
the symbols from the linked library providing the functions using
dlsym(), thus preserving backwards compatibility.
This splits the implementation responsibilities, and reduces embedded
code copies, which was one of the driving points with this project to
start with, so it's nice to give a good example.
Commit e8d340de ("Remove a.out support from nlist()") introduced a copy
of the definition of nlist from a.out.h. However, as well as having
n_name inside n_un, on the various BSDs n_name could also be accessed
as a direct member of nlist, and this is made use of by FreeBSD's
usr.bin/netstat/main.c. Thus we should also add the same enclosing
anonymous union.
[guillem@hadrons.org:
- Add a minimal unit test. ]
Closes: !4
Signed-off-by: Guillem Jover <guillem@hadrons.org>
The NetBSD implementations have different prototypes to the ones coming
from OpenBSD, which will break builds, and have caused segfaults at
run-time. We provide now both interfaces with different prototypes as
different version nodes allow selecting them at compile-time, defaulting
for now to the OpenBSD one to avoid build-time breakage, while emitting
a compile-time warning. Later on, in 0.10.0, we will be switching the
compile-time default to the NetBSD version.
Ref: http://gnats.netbsd.org/44977
Fixes: https://bugs.debian.org/899282
On IA64 this is only the case in the ELF binary, but it gets normalized
when loaded at run-time.
Fixes: https://bugs.debian.org/881611
Signed-off-by: Guillem Jover <guillem@hadrons.org>
At least on IA64 and PowerPC 64-bit ELFv1, the functions are stored in
the .text sections but they are accessed through a function descriptor
stored in a data section, for example for PowerPC 64-bit ELFv1 that
section is called .opd.
We should take this into account when checking the n_type for the
functions we have requested information from nlist().
Rationale-by: James Clarke <jrtc27@jrtc27.com>
