mirror of
https://github.com/tobast/libunwind-eh_elf.git
synced 2024-11-24 16:27:37 +01:00
Fork of the libunwind official github repo — http://www.nongnu.org/libunwind/
c56fb8f99e
A recent commit added code to override the unwind location for the TOC pointer register r2: unsigned int *inst = (unw_word_t*)c->dwarf.ip; if (*inst == (0xE8410000 + 24)) { // @plt call, restoring R2 from CFA+24 c->dwarf.loc[UNW_PPC64_R2] = DWARF_LOC(c->dwarf.cfa + 24, 0); } It is correct that such code is needed, since DWARF CFI does not describe the unwind location for r2 on PowerPC. However, this particular bit of code has a number of issues, which are fixed in this patch. First of all, the location CFA+24 is correct only for the ELFv2 ABI. In the ELFv1 ABI, the TOC location is actually CFA+40. More problematically, attempting to read the current instruction by just dereferencing the address in c->dwarf.ip is wrong, and may often lead to crashes. In particular: - During remote unwinding, this is always wrong since we're in the wrong address space. I've used the fetch32 helper from remote.h to use the proper access_mem under the covers. - c->dwarf.ip may be NULL if we've reached the end-of-stack. I've fixed this by moving the c->dwarf.ip == 0 check down to after unwinding (instead of before), just like all other platforms do. - Even so, c->dwarf.ip may point to some random location if we've gotten confused during unwinding earlier. One likely cause for such confusion is that we did not find DWARF CFI for some earlier frame and attempted to use the stack backchain. The problem is that this code currently claims all registers remain unchanges in such a frame, which is generally wrong. In particular if the function actually saves and modifies r31, and this is used as frame pointer by a later frame, things will likely go quite wrong. While it is not really possibly to completely fix this, I've at least marked all registers as unavailable after unwinding a frame via stack backchain. Tested on powerpc64-linux and powerpc64le-linux. The patch fixes about a dozen test cases that were crashing before. Signed-off-by: Ulrich Weigand <ulrich.weigand@de.ibm.com> |
||
---|---|---|
aux | ||
doc | ||
include | ||
scripts | ||
src | ||
tests | ||
.gitignore | ||
acinclude.m4 | ||
AUTHORS | ||
autogen.sh | ||
ChangeLog | ||
configure.ac | ||
COPYING | ||
LICENSE | ||
Makefile.am | ||
NEWS | ||
README | ||
TODO |
-*- mode: Outline -*- This is version 1.0 of the unwind library. This library supports several architecture/operating-system combinations: Linux/x86-64: Works well. Linux/x86: Works well. Linux/ARM: Works well. Linux/IA-64: Fully tested and supported. Linux/PARISC: Works well, but C library missing unwind-info. HP-UX/IA-64: Mostly works but known to have some serious limitations. Linux/AArch64: Newly added. Linux/PPC64: Newly added. Linux/SuperH: Newly added. FreeBSD/i386: Newly added. FreeBSD/x86-64: Newly added (FreeBSD architecture is known as amd64). Linux/Tilegx: Newly added (64-bit mode only). * General Build Instructions In general, this library can be built and installed with the following commands: $ ./autogen.sh # Needed only for building from git. Depends on libtool. $ ./configure $ make $ make install prefix=PREFIX where PREFIX is the installation prefix. By default, a prefix of /usr/local is used, such that libunwind.a is installed in /usr/local/lib and unwind.h is installed in /usr/local/include. For testing, you may want to use a prefix of /usr/local instead. * Building with Intel compiler ** Version 8 and later Starting with version 8, the preferred name for the IA-64 Intel compiler is "icc" (same name as on x86). Thus, the configure-line should look like this: $ ./configure CC=icc CFLAGS="-g -O3 -ip" CXX=icc CCAS=gcc CCASFLAGS=-g \ LDFLAGS="-L$PWD/src/.libs" * Building on HP-UX For the time being, libunwind must be built with GCC on HP-UX. libunwind should be configured and installed on HP-UX like this: $ ./configure CFLAGS="-g -O2 -mlp64" CXXFLAGS="-g -O2 -mlp64" Caveat: Unwinding of 32-bit (ILP32) binaries is not supported at the moment. ** Workaround for older versions of GCC GCC v3.0 and GCC v3.2 ship with a bad version of sys/types.h. The workaround is to issue the following commands before running "configure": $ mkdir $top_dir/include/sys $ cp /usr/include/sys/types.h $top_dir/include/sys GCC v3.3.2 or later have been fixed and do not require this workaround. * Building for PowerPC64 / Linux For building for power64 you should use: $ ./configure CFLAGS="-g -O2 -m64" CXXFLAGS="-g -O2 -m64" If your power support altivec registers: $ ./configure CFLAGS="-g -O2 -m64 -maltivec" CXXFLAGS="-g -O2 -m64 -maltivec" To check if your processor has support for vector registers (altivec): cat /proc/cpuinfo | grep altivec and should have something like this: cpu : PPC970, altivec supported If libunwind seems to not work (backtracing failing), try to compile it with -O0, without optimizations. There are some compiler problems depending on the version of your gcc. * Building on FreeBSD General building instructions apply. To build and execute several tests, you need libexecinfo library available in ports as devel/libexecinfo. Development of the port was done of FreeBSD 8.0-STABLE. The library was build with the system compiler that is modified version of gcc 4.2.1, as well as the gcc 4.4.3. * Regression Testing After building the library, you can run a set of regression tests with: $ make check ** Expected results on IA-64 Linux Unless you have a very recent C library and compiler installed, it is currently expected to have the following tests fail on IA-64 Linux: Gtest-init (should pass starting with glibc-2.3.x/gcc-3.4) Ltest-init (should pass starting with glibc-2.3.x/gcc-3.4) test-ptrace (should pass starting with glibc-2.3.x/gcc-3.4) run-ia64-test-dyn1 (should pass starting with glibc-2.3.x) This does not mean that libunwind cannot be used with older compilers or C libraries, it just means that for certain corner cases, unwinding will fail. Since they're corner cases, it is not likely for applications to trigger them. Note: If you get lots of errors in Gia64-test-nat and Lia64-test-nat, it's almost certainly a sign of an old assembler. The GNU assembler used to encode previous-stack-pointer-relative offsets incorrectly. This bug was fixed on 21-Sep-2004 so any later assembler will be fine. ** Expected results on x86 Linux The following tests are expected to fail on x86 Linux: Gtest-resume-sig (fails to get SIGUSR2) Ltest-resume-sig (likewise) Gtest-dyn1 (no dynamic unwind info support yet) Ltest-dyn1 (no dynamic unwind info support yet) test-setjmp (longjmp() not implemented yet) run-check-namespace (no _Ux86_getcontext yet) test-ptrace ** Expected results on x86-64 Linux The following tests are expected to fail on x86-64 Linux: Gtest-dyn1 (no dynamic unwind info support yet) Ltest-dyn1 (no dynamic unwind info support yet) Gtest-init (see http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18743) Ltest-init (likewise) test-async-sig (crashes due to bad unwind-info?) test-setjmp (longjmp() not implemented yet) run-check-namespace (no _Ux86_64_getcontext yet) run-ptrace-mapper (??? investigate) run-ptrace-misc (see http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18748 and http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18749) ** Expected results on PARISC Linux Caveat: GCC v3.4 or newer is needed on PA-RISC Linux. Earlier versions of the compiler failed to generate the exception-handling program header (GNU_EH_FRAME) needed for unwinding. The following tests are expected to fail on x86-64 Linux: Gtest-bt (backtrace truncated at kill() due to lack of unwind-info) Ltest-bt (likewise) Gtest-resume-sig (Gresume.c:my_rt_sigreturn() is wrong somehow) Ltest-resume-sig (likewise) Gtest-init (likewise) Ltest-init (likewise) Gtest-dyn1 (no dynamic unwind info support yet) Ltest-dyn1 (no dynamic unwind info support yet) test-setjmp (longjmp() not implemented yet) run-check-namespace (toolchain doesn't support HIDDEN yet) ** Expected results on HP-UX "make check" is currently unsupported for HP-UX. You can try to run it, but most tests will fail (and some may fail to terminate). The only test programs that are known to work at this time are: tests/bt tests/Gperf-simple tests/test-proc-info tests/test-static-link tests/Gtest-init tests/Ltest-init tests/Gtest-resume-sig tests/Ltest-resume-sig ** Expected results on PPC64 Linux "make check" should run with no more than 10 out of 24 tests failed. * Performance Testing This distribution includes a few simple performance tests which give some idea of the basic cost of various libunwind operations. After building the library, you can run these tests with the following commands: $ cd tests $ make perf * Contacting the Developers Please direct all questions regarding this library to: libunwind-devel@nongnu.org You can do this by sending a mail to libunwind-request@nongnu.org with a body of: subscribe libunwind-devel or you can subscribe and manage your subscription via the web-interface at: https://savannah.nongnu.org/mail/?group=libunwind