Hello!
Today I found two important things.
1. Mesa seems to use both __attribute__((__destructor__)) AND atexit(3)
at same time, but destructor call is not executed or gdb does not catch
destructor call (?)
2. Crash seems to occurs only in
/usr/X11R7/lib/modules/dri/swrast_dri.so (swrast driver). When I tried
"modesetting" Intel Broxton "HD Graphics 500 (APL 2)") it works (after
several unrelated tweaks)
Ad 1: Looking closely on source at
https://cvsweb.netbsd.org/bsdweb.cgi/xsrc/external/mit/MesaLib/dist/src/mesa/main/context.c?rev=1.9;content-type=text%2Fplain
we can see both:
static void __attribute__((__destructor__))
one_time_fini(void)
{
if (api_init_mask) {
glsl_type_singleton_decref();
_mesa_locale_fini();
}
}
static void
one_time_init(void)
{
// ...
atexit(one_time_fini);
// ...
}
But when I place deferred brakpoint at one_time_fini() only final
handler call is caught (that one which cause SIGSEGV) where GDB properly
warns:
warning: Temporarily disabling breakpoints for unloaded shared library
"/usr/X11R7/lib/modules/dri/swrast_dri.so"
Thread 1 "" received signal SIGSEGV, Segmentation fault.
So now I have 2 mysteries:
1. why there is used both __destructor__ and atexit(3) in Mesa's context.c
2. why my gdb does not catch __destructor__ call - is it called or not?
Regards
--Henryk Paluch
On 12/21/25 09:04, Henryk Paluch wrote:
Hello!
New example with __cxa_atexit works for me.
Let me summarize things:
Problem: any program using new Mesa from (xsrc/external/mit/MesaLib/
dist/src/mesa/main/context.c r1.9), including NetBSD 11.0_BETA will
crash on exit, because Mesa now calls atexit(3) from dlopen(3) DSO that
is not supported in NetBSD. In base system this affects at least:
- glxinfo
- glxgears
And possibly some ports. Unfortunately I had not enough time to build
any major browser (Firefox, Chromium) so can't confirm if these are also
affected.
Possible solutions:
1. fix atexit(3) to work properly from DSOs (as done on most other systems)
2. use __attribute__((destructor))
3. use __cxa_atexit
So we have to decide what solution is best. Personally I think that this
problem is blocker for 11.0_RELEASE. We can't ship system where major X
programs just crash on exit.
Regards
--Henryk Paluch
On 12/21/25 00:54, RVP wrote:
On Sat, 20 Dec 2025, RVP wrote:
Shared libs. should use __cxa_atexit() with a non-NULL `dso' parameter.
This works on FreeBSD, OpenBSD, Linux and even OpenIndiana. (Anything
which
supports C++ would be my bet.)
-RVP
---START---
==> dl-atexit.c <==
#include <dlfcn.h>
#include <stdio.h>
#include <stdlib.h>
static void
diemain(void)
{
printf("%s: atexit handler\n", __func__);
}
int
main(void)
{
extern const void* const __dso_handle;
printf("__dso_handle = %p\n", __dso_handle);
atexit(diemain);
void *h = dlopen("libfoo.so", RTLD_LOCAL | RTLD_LAZY);
if (h == NULL) {
fprintf(stderr, "dlopen(): %s\n", dlerror());
return 1;
}
void (*foo)(void) = dlsym(h, "foo");
if (foo == NULL) {
fprintf(stderr, "dlsym(foo): %s\n", dlerror());
return 1;
}
printf("%s: calling foo()\n", __func__);
foo();
printf("%s: calling dlclose()\n", __func__);
dlclose(h);
printf("%s: exiting...\n", __func__);
return 0;
}
==> libfoo.c <==
#include <stdio.h>
#include <stdlib.h>
extern int __cxa_atexit(void (*)(void*), void*, const void*);
extern const void* const __dso_handle;
static void
// die(void)
die(void* v)
{
printf("%s: atexit handler\n", __func__);
}
void
foo(void)
{
printf("__dso_handle = %p\n", __dso_handle);
// atexit(die);
__cxa_atexit(die, NULL, __dso_handle);
printf("%s: atexit handler die() @ %p\n", __func__, die);
}
==> Makefile <==
LDFLAGS= -fuse-linker-plugin -Wl,--enable-new-dtags,-z,relro,-z,now
dl-atexit: dl-atexit.c libfoo.so
cc -pie -fPIE -o dl-atexit dl-atexit.c -Wl,-rpath=$$(pwd) ${LDFLAGS}
libfoo.so: libfoo.c
cc -shared -DPIC -fPIC -o libfoo.so libfoo.c ${LDFLAGS}
run: dl-atexit
./dl-atexit
clean:
rm -f dl-atexit *.so *.core
---END---
