Boost stacktrace
This is the most convenient option I've seen so far, because it:
can actually print out the line numbers.
It just makes calls to addr2line however, which is ugly and might be slow if your are taking too many traces.
demangles by default
Boost is header only, so no need to modify your build system most likely
boost_stacktrace.cpp
#include <iostream>
#define BOOST_STACKTRACE_USE_ADDR2LINE
#include <boost/stacktrace.hpp>
void my_func_2(void) {
std::cout << boost::stacktrace::stacktrace() << std::endl;
}
void my_func_1(double f) {
(void)f;
my_func_2();
}
void my_func_1(int i) {
(void)i;
my_func_2();
}
int main(int argc, char **argv) {
long long unsigned int n;
if (argc > 1) {
n = strtoul(argv[1], NULL, 0);
} else {
n = 1;
}
for (long long unsigned int i = 0; i < n; ++i) {
my_func_1(1); // line 28
my_func_1(2.0); // line 29
}
}
Unfortunately, it seems to be a more recent addition, and the package libboost-stacktrace-dev is not present in Ubuntu 16.04, only 18.04:
sudo apt-get install libboost-stacktrace-dev
g++ -fno-pie -ggdb3 -O0 -no-pie -o boost_stacktrace.out -std=c++11 \
-Wall -Wextra -pedantic-errors boost_stacktrace.cpp -ldl
./boost_stacktrace.out
We have to add -ldl at the end or else compilation fails.
Output:
0# boost::stacktrace::basic_stacktrace<std::allocator<boost::stacktrace::frame> >::basic_stacktrace() at /usr/include/boost/stacktrace/stacktrace.hpp:129
1# my_func_1(int) at /home/ciro/test/boost_stacktrace.cpp:18
2# main at /home/ciro/test/boost_stacktrace.cpp:29 (discriminator 2)
3# __libc_start_main in /lib/x86_64-linux-gnu/libc.so.6
4# _start in ./boost_stacktrace.out
0# boost::stacktrace::basic_stacktrace<std::allocator<boost::stacktrace::frame> >::basic_stacktrace() at /usr/include/boost/stacktrace/stacktrace.hpp:129
1# my_func_1(double) at /home/ciro/test/boost_stacktrace.cpp:13
2# main at /home/ciro/test/boost_stacktrace.cpp:27 (discriminator 2)
3# __libc_start_main in /lib/x86_64-linux-gnu/libc.so.6
4# _start in ./boost_stacktrace.out
The output and is further explained on the "glibc backtrace" section below, which is analogous.
Note how my_func_1(int) and my_func_1(float), which are mangled due to function overload, were nicely demangled for us.
Note that the first int calls is off by one line (28 instead of 27 and the second one is off by two lines (27 instead of 29). It was suggested in the comments that this is because the following instruction address is being considered, which makes 27 become 28, and 29 jump off the loop and become 27.
We then observe that with -O3, the output is completely mutilated:
0# boost::stacktrace::basic_stacktrace<std::allocator<boost::stacktrace::frame> >::size() const at /usr/include/boost/stacktrace/stacktrace.hpp:215
1# my_func_1(double) at /home/ciro/test/boost_stacktrace.cpp:12
2# __libc_start_main in /lib/x86_64-linux-gnu/libc.so.6
3# _start in ./boost_stacktrace.out
0# boost::stacktrace::basic_stacktrace<std::allocator<boost::stacktrace::frame> >::size() const at /usr/include/boost/stacktrace/stacktrace.hpp:215
1# main at /home/ciro/test/boost_stacktrace.cpp:31
2# __libc_start_main in /lib/x86_64-linux-gnu/libc.so.6
3# _start in ./boost_stacktrace.out
Backtraces are in general irreparably mutilated by optimizations. Tail call optimization is a notable example of that: What is tail call optimization?
Benchmark run on -O3:
time ./boost_stacktrace.out 1000 >/dev/null
Output:
real 0m43.573s
user 0m30.799s
sys 0m13.665s
So as expected, we see that this method is extremely slow likely to to external calls to addr2line, and is only going to be feasible if a limited number of calls are being made.
Each backtrace print seems to take hundreds of milliseconds, so be warned that if a backtrace happens very often, program performance will suffer significantly.
Tested on Ubuntu 19.10, GCC 9.2.1, boost 1.67.0.
glibc backtrace
Documented at: https://www.gnu.org/software/libc/manual/html_node/Backtraces.html
main.c
#include <stdio.h>
#include <stdlib.h>
/* Paste this on the file you want to debug. */
#include <stdio.h>
#include <execinfo.h>
void print_trace(void) {
char **strings;
size_t i, size;
enum Constexpr { MAX_SIZE = 1024 };
void *array[MAX_SIZE];
size = backtrace(array, MAX_SIZE);
strings = backtrace_symbols(array, size);
for (i = 0; i < size; i++)
printf("%s\n", strings[i]);
puts("");
free(strings);
}
void my_func_3(void) {
print_trace();
}
void my_func_2(void) {
my_func_3();
}
void my_func_1(void) {
my_func_3();
}
int main(void) {
my_func_1(); /* line 33 */
my_func_2(); /* line 34 */
return 0;
}
Compile:
gcc -fno-pie -ggdb3 -O3 -no-pie -o main.out -rdynamic -std=c99 \
-Wall -Wextra -pedantic-errors main.c
-rdynamic is the key required option.
Run:
./main.out
Outputs:
./main.out(print_trace+0x2d) [0x400a3d]
./main.out(main+0x9) [0x4008f9]
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf0) [0x7f35a5aad830]
./main.out(_start+0x29) [0x400939]
./main.out(print_trace+0x2d) [0x400a3d]
./main.out(main+0xe) [0x4008fe]
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf0) [0x7f35a5aad830]
./main.out(_start+0x29) [0x400939]
So we immediately see that an inlining optimization happened, and some functions were lost from the trace.
If we try to get the addresses:
addr2line -e main.out 0x4008f9 0x4008fe
we obtain:
/home/ciro/main.c:21
/home/ciro/main.c:36
which is completely off.
If we do the same with -O0 instead, ./main.out gives the correct full trace:
./main.out(print_trace+0x2e) [0x4009a4]
./main.out(my_func_3+0x9) [0x400a50]
./main.out(my_func_1+0x9) [0x400a68]
./main.out(main+0x9) [0x400a74]
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf0) [0x7f4711677830]
./main.out(_start+0x29) [0x4008a9]
./main.out(print_trace+0x2e) [0x4009a4]
./main.out(my_func_3+0x9) [0x400a50]
./main.out(my_func_2+0x9) [0x400a5c]
./main.out(main+0xe) [0x400a79]
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf0) [0x7f4711677830]
./main.out(_start+0x29) [0x4008a9]
and then:
addr2line -e main.out 0x400a74 0x400a79
gives:
/home/cirsan01/test/main.c:34
/home/cirsan01/test/main.c:35
so the lines are off by just one, TODO why? But this might still be usable.
Conclusion: backtraces can only possibly show perfectly with -O0. With optimizations, the original backtrace is fundamentally modified in the compiled code.
I couldn't find a simple way to automatically demangle C++ symbols with this however, here are some hacks:
Tested on Ubuntu 16.04, GCC 6.4.0, libc 2.23.
glibc backtrace_symbols_fd
This helper is a bit more convenient than backtrace_symbols, and produces basically identical output:
/* Paste this on the file you want to debug. */
#include <execinfo.h>
#include <stdio.h>
#include <unistd.h>
void print_trace(void) {
size_t i, size;
enum Constexpr { MAX_SIZE = 1024 };
void *array[MAX_SIZE];
size = backtrace(array, MAX_SIZE);
backtrace_symbols_fd(array, size, STDOUT_FILENO);
puts("");
}
Tested on Ubuntu 16.04, GCC 6.4.0, libc 2.23.
glibc backtrace with C++ demangling hack 1: -export-dynamic + dladdr
Adapted from: https://gist.github.com/fmela/591333/c64f4eb86037bb237862a8283df70cdfc25f01d3
This is a "hack" because it requires changing the ELF with -export-dynamic.
glibc_ldl.cpp
#include <dlfcn.h> // for dladdr
#include <cxxabi.h> // for __cxa_demangle
#include <cstdio>
#include <string>
#include <sstream>
#include <iostream>
// This function produces a stack backtrace with demangled function & method names.
std::string backtrace(int skip = 1)
{
void *callstack[128];
const int nMaxFrames = sizeof(callstack) / sizeof(callstack[0]);
char buf[1024];
int nFrames = backtrace(callstack, nMaxFrames);
char **symbols = backtrace_symbols(callstack, nFrames);
std::ostringstream trace_buf;
for (int i = skip; i < nFrames; i++) {
Dl_info info;
if (dladdr(callstack[i], &info)) {
char *demangled = NULL;
int status;
demangled = abi::__cxa_demangle(info.dli_sname, NULL, 0, &status);
std::snprintf(
buf,
sizeof(buf),
"%-3d %*p %s + %zd\n",
i,
(int)(2 + sizeof(void*) * 2),
callstack[i],
status == 0 ? demangled : info.dli_sname,
(char *)callstack[i] - (char *)info.dli_saddr
);
free(demangled);
} else {
std::snprintf(buf, sizeof(buf), "%-3d %*p\n",
i, (int)(2 + sizeof(void*) * 2), callstack[i]);
}
trace_buf << buf;
std::snprintf(buf, sizeof(buf), "%s\n", symbols[i]);
trace_buf << buf;
}
free(symbols);
if (nFrames == nMaxFrames)
trace_buf << "[truncated]\n";
return trace_buf.str();
}
void my_func_2(void) {
std::cout << backtrace() << std::endl;
}
void my_func_1(double f) {
(void)f;
my_func_2();
}
void my_func_1(int i) {
(void)i;
my_func_2();
}
int main() {
my_func_1(1);
my_func_1(2.0);
}
Compile and run:
g++ -fno-pie -ggdb3 -O0 -no-pie -o glibc_ldl.out -std=c++11 -Wall -Wextra \
-pedantic-errors -fpic glibc_ldl.cpp -export-dynamic -ldl
./glibc_ldl.out
output:
1 0x40130a my_func_2() + 41
./glibc_ldl.out(_Z9my_func_2v+0x29) [0x40130a]
2 0x40139e my_func_1(int) + 16
./glibc_ldl.out(_Z9my_func_1i+0x10) [0x40139e]
3 0x4013b3 main + 18
./glibc_ldl.out(main+0x12) [0x4013b3]
4 0x7f7594552b97 __libc_start_main + 231
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xe7) [0x7f7594552b97]
5 0x400f3a _start + 42
./glibc_ldl.out(_start+0x2a) [0x400f3a]
1 0x40130a my_func_2() + 41
./glibc_ldl.out(_Z9my_func_2v+0x29) [0x40130a]
2 0x40138b my_func_1(double) + 18
./glibc_ldl.out(_Z9my_func_1d+0x12) [0x40138b]
3 0x4013c8 main + 39
./glibc_ldl.out(main+0x27) [0x4013c8]
4 0x7f7594552b97 __libc_start_main + 231
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xe7) [0x7f7594552b97]
5 0x400f3a _start + 42
./glibc_ldl.out(_start+0x2a) [0x400f3a]
Tested on Ubuntu 18.04.
glibc backtrace with C++ demangling hack 2: parse backtrace output
Shown at: https://panthema.net/2008/0901-stacktrace-demangled/
This is a hack because it requires parsing.
TODO get it to compile and show it here.
libunwind
TODO does this have any advantage over glibc backtrace? Very similar output, also requires modifying the build command, but not part of glibc so requires an extra package installation.
Code adapted from: https://eli.thegreenplace.net/2015/programmatic-access-to-the-call-stack-in-c/
main.c
/* This must be on top. */
#define _XOPEN_SOURCE 700
#include <stdio.h>
#include <stdlib.h>
/* Paste this on the file you want to debug. */
#define UNW_LOCAL_ONLY
#include <libunwind.h>
#include <stdio.h>
void print_trace() {
char sym[256];
unw_context_t context;
unw_cursor_t cursor;
unw_getcontext(&context);
unw_init_local(&cursor, &context);
while (unw_step(&cursor) > 0) {
unw_word_t offset, pc;
unw_get_reg(&cursor, UNW_REG_IP, &pc);
if (pc == 0) {
break;
}
printf("0x%lx:", pc);
if (unw_get_proc_name(&cursor, sym, sizeof(sym), &offset) == 0) {
printf(" (%s+0x%lx)\n", sym, offset);
} else {
printf(" -- error: unable to obtain symbol name for this frame\n");
}
}
puts("");
}
void my_func_3(void) {
print_trace();
}
void my_func_2(void) {
my_func_3();
}
void my_func_1(void) {
my_func_3();
}
int main(void) {
my_func_1(); /* line 46 */
my_func_2(); /* line 47 */
return 0;
}
Compile and run:
sudo apt-get install libunwind-dev
gcc -fno-pie -ggdb3 -O3 -no-pie -o main.out -std=c99 \
-Wall -Wextra -pedantic-errors main.c -lunwind
Either #define _XOPEN_SOURCE 700 must be on top, or we must use -std=gnu99:
Run:
./main.out
Output:
0x4007db: (main+0xb)
0x7f4ff50aa830: (__libc_start_main+0xf0)
0x400819: (_start+0x29)
0x4007e2: (main+0x12)
0x7f4ff50aa830: (__libc_start_main+0xf0)
0x400819: (_start+0x29)
and:
addr2line -e main.out 0x4007db 0x4007e2
gives:
/home/ciro/main.c:34
/home/ciro/main.c:49
With -O0:
0x4009cf: (my_func_3+0xe)
0x4009e7: (my_func_1+0x9)
0x4009f3: (main+0x9)
0x7f7b84ad7830: (__libc_start_main+0xf0)
0x4007d9: (_start+0x29)
0x4009cf: (my_func_3+0xe)
0x4009db: (my_func_2+0x9)
0x4009f8: (main+0xe)
0x7f7b84ad7830: (__libc_start_main+0xf0)
0x4007d9: (_start+0x29)
and:
addr2line -e main.out 0x4009f3 0x4009f8
gives:
/home/ciro/main.c:47
/home/ciro/main.c:48
Tested on Ubuntu 16.04, GCC 6.4.0, libunwind 1.1.
libunwind with C++ name demangling
Code adapted from: https://eli.thegreenplace.net/2015/programmatic-access-to-the-call-stack-in-c/
unwind.cpp
#define UNW_LOCAL_ONLY
#include <cxxabi.h>
#include <libunwind.h>
#include <cstdio>
#include <cstdlib>
#include <iostream>
void backtrace() {
unw_cursor_t cursor;
unw_context_t context;
// Initialize cursor to current frame for local unwinding.
unw_getcontext(&context);
unw_init_local(&cursor, &context);
// Unwind frames one by one, going up the frame stack.
while (unw_step(&cursor) > 0) {
unw_word_t offset, pc;
unw_get_reg(&cursor, UNW_REG_IP, &pc);
if (pc == 0) {
break;
}
std::printf("0x%lx:", pc);
char sym[256];
if (unw_get_proc_name(&cursor, sym, sizeof(sym), &offset) == 0) {
char* nameptr = sym;
int status;
char* demangled = abi::__cxa_demangle(sym, nullptr, nullptr, &status);
if (status == 0) {
nameptr = demangled;
}
std::printf(" (%s+0x%lx)\n", nameptr, offset);
std::free(demangled);
} else {
std::printf(" -- error: unable to obtain symbol name for this frame\n");
}
}
}
void my_func_2(void) {
backtrace();
std::cout << std::endl; // line 43
}
void my_func_1(double f) {
(void)f;
my_func_2();
}
void my_func_1(int i) {
(void)i;
my_func_2();
} // line 54
int main() {
my_func_1(1);
my_func_1(2.0);
}
Compile and run:
sudo apt-get install libunwind-dev
g++ -fno-pie -ggdb3 -O0 -no-pie -o unwind.out -std=c++11 \
-Wall -Wextra -pedantic-errors unwind.cpp -lunwind -pthread
./unwind.out
Output:
0x400c80: (my_func_2()+0x9)
0x400cb7: (my_func_1(int)+0x10)
0x400ccc: (main+0x12)
0x7f4c68926b97: (__libc_start_main+0xe7)
0x400a3a: (_start+0x2a)
0x400c80: (my_func_2()+0x9)
0x400ca4: (my_func_1(double)+0x12)
0x400ce1: (main+0x27)
0x7f4c68926b97: (__libc_start_main+0xe7)
0x400a3a: (_start+0x2a)
and then we can find the lines of my_func_2 and my_func_1(int) with:
addr2line -e unwind.out 0x400c80 0x400cb7
which gives:
/home/ciro/test/unwind.cpp:43
/home/ciro/test/unwind.cpp:54
TODO: why are the lines off by one?
Tested on Ubuntu 18.04, GCC 7.4.0, libunwind 1.2.1.
GDB automation
We can also do this with GDB without recompiling by using: How to do an specific action when a certain breakpoint is hit in GDB?
Although if you are going to print the backtrace a lot, this will likely be less fast than the other options, but maybe we can reach native speeds with compile code, but I'm lazy to test it out now: How to call assembly in gdb?
main.cpp
void my_func_2(void) {}
void my_func_1(double f) {
my_func_2();
}
void my_func_1(int i) {
my_func_2();
}
int main() {
my_func_1(1);
my_func_1(2.0);
}
main.gdb
start
break my_func_2
commands
silent
backtrace
printf "\n"
continue
end
continue
Compile and run:
g++ -ggdb3 -o main.out main.cpp
gdb -nh -batch -x main.gdb main.out
Output:
Temporary breakpoint 1 at 0x1158: file main.cpp, line 12.
Temporary breakpoint 1, main () at main.cpp:12
12 my_func_1(1);
Breakpoint 2 at 0x555555555129: file main.cpp, line 1.
#0 my_func_2 () at main.cpp:1
#1 0x0000555555555151 in my_func_1 (i=1) at main.cpp:8
#2 0x0000555555555162 in main () at main.cpp:12
#0 my_func_2 () at main.cpp:1
#1 0x000055555555513e in my_func_1 (f=2) at main.cpp:4
#2 0x000055555555516f in main () at main.cpp:13
[Inferior 1 (process 14193) exited normally]
TODO I wanted to do this with just -ex from the command line to not have to create main.gdb but I couldn't get the commands to work there.
Tested in Ubuntu 19.04, GDB 8.2.
Linux kernel
How to print the current thread stack trace inside the Linux kernel?
libdwfl
This was originally mentioned at: https://stackoverflow.com/a/60713161/895245 and it might be the best method, but I have to benchmark a bit more, but please go upvote that answer.
TODO: I tried to minimize the code in that answer, which was working, to a single function, but it is segfaulting, let me know if anyone can find why.
dwfl.cpp
#include <cassert>
#include <iostream>
#include <memory>
#include <sstream>
#include <string>
#include <cxxabi.h> // __cxa_demangle
#include <elfutils/libdwfl.h> // Dwfl*
#include <execinfo.h> // backtrace
#include <unistd.h> // getpid
// https://stackoverflow.com/questions/281818/unmangling-the-result-of-stdtype-infoname
std::string demangle(const char* name) {
int status = -4;
std::unique_ptr<char, void(*)(void*)> res {
abi::__cxa_demangle(name, NULL, NULL, &status),
std::free
};
return (status==0) ? res.get() : name ;
}
std::string debug_info(Dwfl* dwfl, void* ip) {
std::string function;
int line = -1;
char const* file;
uintptr_t ip2 = reinterpret_cast<uintptr_t>(ip);
Dwfl_Module* module = dwfl_addrmodule(dwfl, ip2);
char const* name = dwfl_module_addrname(module, ip2);
function = name ? demangle(name) : "<unknown>";
if (Dwfl_Line* dwfl_line = dwfl_module_getsrc(module, ip2)) {
Dwarf_Addr addr;
file = dwfl_lineinfo(dwfl_line, &addr, &line, nullptr, nullptr, nullptr);
}
std::stringstream ss;
ss << ip << ' ' << function;
if (file)
ss << " at " << file << ':' << line;
ss << std::endl;
return ss.str();
}
std::string stacktrace() {
// Initialize Dwfl.
Dwfl* dwfl = nullptr;
{
Dwfl_Callbacks callbacks = {};
char* debuginfo_path = nullptr;
callbacks.find_elf = dwfl_linux_proc_find_elf;
callbacks.find_debuginfo = dwfl_standard_find_debuginfo;
callbacks.debuginfo_path = &debuginfo_path;
dwfl = dwfl_begin(&callbacks);
assert(dwfl);
int r;
r = dwfl_linux_proc_report(dwfl, getpid());
assert(!r);
r = dwfl_report_end(dwfl, nullptr, nullptr);
assert(!r);
static_cast<void>(r);
}
// Loop over stack frames.
std::stringstream ss;
{
void* stack[512];
int stack_size = ::backtrace(stack, sizeof stack / sizeof *stack);
for (int i = 0; i < stack_size; ++i) {
ss << i << ": ";
// Works.
ss << debug_info(dwfl, stack[i]);
#if 0
// TODO intended to do the same as above, but segfaults,
// so possibly UB In above function that does not blow up by chance?
void *ip = stack[i];
std::string function;
int line = -1;
char const* file;
uintptr_t ip2 = reinterpret_cast<uintptr_t>(ip);
Dwfl_Module* module = dwfl_addrmodule(dwfl, ip2);
char const* name = dwfl_module_addrname(module, ip2);
function = name ? demangle(name) : "<unknown>";
// TODO if I comment out this line it does not blow up anymore.
if (Dwfl_Line* dwfl_line = dwfl_module_getsrc(module, ip2)) {
Dwarf_Addr addr;
file = dwfl_lineinfo(dwfl_line, &addr, &line, nullptr, nullptr, nullptr);
}
ss << ip << ' ' << function;
if (file)
ss << " at " << file << ':' << line;
ss << std::endl;
#endif
}
}
dwfl_end(dwfl);
return ss.str();
}
void my_func_2() {
std::cout << stacktrace() << std::endl;
std::cout.flush();
}
void my_func_1(double f) {
(void)f;
my_func_2();
}
void my_func_1(int i) {
(void)i;
my_func_2();
}
int main(int argc, char **argv) {
long long unsigned int n;
if (argc > 1) {
n = strtoul(argv[1], NULL, 0);
} else {
n = 1;
}
for (long long unsigned int i = 0; i < n; ++i) {
my_func_1(1);
my_func_1(2.0);
}
}
Compile and run:
sudo apt install libdw-dev
g++ -fno-pie -ggdb3 -O0 -no-pie -o dwfl.out -std=c++11 -Wall -Wextra -pedantic-errors dwfl.cpp -ldw
./dwfl.out
Output:
0: 0x402b74 stacktrace[abi:cxx11]() at /home/ciro/test/dwfl.cpp:65
1: 0x402ce0 my_func_2() at /home/ciro/test/dwfl.cpp:100
2: 0x402d7d my_func_1(int) at /home/ciro/test/dwfl.cpp:112
3: 0x402de0 main at /home/ciro/test/dwfl.cpp:123
4: 0x7f7efabbe1e3 __libc_start_main at ../csu/libc-start.c:342
5: 0x40253e _start at ../csu/libc-start.c:-1
0: 0x402b74 stacktrace[abi:cxx11]() at /home/ciro/test/dwfl.cpp:65
1: 0x402ce0 my_func_2() at /home/ciro/test/dwfl.cpp:100
2: 0x402d66 my_func_1(double) at /home/ciro/test/dwfl.cpp:107
3: 0x402df1 main at /home/ciro/test/dwfl.cpp:121
4: 0x7f7efabbe1e3 __libc_start_main at ../csu/libc-start.c:342
5: 0x40253e _start at ../csu/libc-start.c:-1
Benchmark run:
g++ -fno-pie -ggdb3 -O3 -no-pie -o dwfl.out -std=c++11 -Wall -Wextra -pedantic-errors dwfl.cpp -ldw
time ./dwfl.out 1000 >/dev/null
Output:
real 0m3.751s
user 0m2.822s
sys 0m0.928s
So we see that this method is 10x faster than Boost's stacktrace, and might therefore be applicable to more use cases.
Tested in Ubuntu 19.10 amd64, libdw-dev 0.176-1.1.
See also