Hi,
I'm experiencing problems with uClibc-ng when creating static binaries that uses threads. I originally reported this to the buildroot mailing list but got redirected here instead.
The problem is when using a combination of stack protection, threads and static binaries. The program will during start up get a sigsegv in __pthread_initialize_minimal_internal(), and gdb says the instruction is a
mov %fs:0x28,%rax
where %fs is 0x00. The reason for that is that the pthread library is compiled with stack protection but __pthread_initialize_minimal_internal is called before __libc_setup_tls in static binaries.
I have created an ugly patch to verify my hypothesis (attached). With that patch the test program works, but what is the proper way to solve this issue?
With uClibc 0.9.33.x the same thing happens. With glibc the program works.
Best regards, Daniel Fahlgren
Hi Daniel, Daniel Fahlgren wrote,
Hi,
I'm experiencing problems with uClibc-ng when creating static binaries that uses threads. I originally reported this to the buildroot mailing list but got redirected here instead.
I have seen your comprehensive bug report.
The problem is when using a combination of stack protection, threads and static binaries. The program will during start up get a sigsegv in __pthread_initialize_minimal_internal(), and gdb says the instruction is a
mov %fs:0x28,%rax
where %fs is 0x00. The reason for that is that the pthread library is compiled with stack protection but __pthread_initialize_minimal_internal is called before __libc_setup_tls in static binaries.
I have created an ugly patch to verify my hypothesis (attached). With that patch the test program works, but what is the proper way to solve this issue?
Does this happen only when uClibc itself is compiled with stack protection? Or just when the pthread app is statically linked and compiled with -fstack-protector?
With uClibc 0.9.33.x the same thing happens. With glibc the program works.
I can reproduce the bug. Moving __libc_setup_tls from NPTL init to __uClibc_main.c fixes it. Why do you think it is ugly to do this?
Does attached patch works for you? Just a cleaned up version of your fix. It doesn't produce any regressions running the testsuite.
best regards Waldemar
Hi Waldemar,
On Mon, 2015-09-14 at 06:52 +0200, Waldemar Brodkorb wrote:
...
Does this happen only when uClibc itself is compiled with stack protection? Or just when the pthread app is statically linked and compiled with -fstack-protector?
This only happen when libpthread.a is compiled with stack protection and the binary is statically linked. I have tried compiling the app both with and without -fstack-protector-all but that doesn't change the behavior. A libpthread.a not compiled with stack protection also does not show this behavior.
With uClibc 0.9.33.x the same thing happens. With glibc the program works.
I can reproduce the bug. Moving __libc_setup_tls from NPTL init to __uClibc_main.c fixes it. Why do you think it is ugly to do this?
With my patch __libc_setup_tls() was called twice. I also haven't thought it through if if is ok to simply move it from the pthread code or if something else is depending on it being there.
Does attached patch works for you? Just a cleaned up version of your fix. It doesn't produce any regressions running the testsuite.
Yes, that patch works just fine. I have tested it on my limited test case as well as in the environment where I first encountered the bug.
Thanks and best regards, Daniel Fahlgren
Hi Daniel, Daniel Fahlgren wrote,
Hi Waldemar,
On Mon, 2015-09-14 at 06:52 +0200, Waldemar Brodkorb wrote:
...
Does this happen only when uClibc itself is compiled with stack protection? Or just when the pthread app is statically linked and compiled with -fstack-protector?
This only happen when libpthread.a is compiled with stack protection and the binary is statically linked. I have tried compiling the app both with and without -fstack-protector-all but that doesn't change the behavior. A libpthread.a not compiled with stack protection also does not show this behavior.
With uClibc 0.9.33.x the same thing happens. With glibc the program works.
I can reproduce the bug. Moving __libc_setup_tls from NPTL init to __uClibc_main.c fixes it. Why do you think it is ugly to do this?
With my patch __libc_setup_tls() was called twice. I also haven't thought it through if if is ok to simply move it from the pthread code or if something else is depending on it being there.
Does attached patch works for you? Just a cleaned up version of your fix. It doesn't produce any regressions running the testsuite.
Yes, that patch works just fine. I have tested it on my limited test case as well as in the environment where I first encountered the bug.
I still get segfaults when using __UCLIBC_HAS_SSP__ while compiling uClibc-ng and then use the results to statically link your test app with -fstack-protector-all. Gdb shows: (gdb) bt #0 0x000000000040cf1b in _dl_aux_init ()
So it seems not completely solved, yet. Any idea?
best regards Waldemar
Hi Waldemar,
On tor, 2015-09-17 at 20:02 +0200, Waldemar Brodkorb wrote:
I still get segfaults when using __UCLIBC_HAS_SSP__ while compiling uClibc-ng and then use the results to statically link your test app with -fstack-protector-all. Gdb shows: (gdb) bt #0 0x000000000040cf1b in _dl_aux_init ()
So it seems not completely solved, yet. Any idea?
Strange, I fail to reproduce this issue when using buildroot. How to you build uClibc-ng and what kind of OS / environment / toolchain / etc do you have?
Best regards, Daniel
Hi Daniel, Daniel Fahlgren wrote,
Hi Waldemar,
On tor, 2015-09-17 at 20:02 +0200, Waldemar Brodkorb wrote:
I still get segfaults when using __UCLIBC_HAS_SSP__ while compiling uClibc-ng and then use the results to statically link your test app with -fstack-protector-all. Gdb shows: (gdb) bt #0 0x000000000040cf1b in _dl_aux_init ()
So it seems not completely solved, yet. Any idea?
Strange, I fail to reproduce this issue when using buildroot. How to you build uClibc-ng and what kind of OS / environment / toolchain / etc do you have?
I am using OpenADK for my runtime testing before pushing changes to uClibc-ng repository. But the problem seems ADK specific, when SSP is enabled. Need to analyze it further, but the fix will be included in 1.0.7, I promise :)
best regards Waldemar
-
Daniel Fahlgren -
Waldemar Brodkorb