c++ stack trace from unhandled exception? [duplicate]

Question

This question has been asked before and there have been windows-specific answers but no satisfactory gcc answer. I can use set_terminate() to set a function that will be called (in place of terminate()) when an unhandled exception is thrown. I know how to use the backtrace library to generate a stack trace from a given point in the program. However, this won't help when my terminate-replacement is called since at that point the stack has been unwound.

Yet if I simply allow the program to abort(), it will produce a core-dump which contains the full stack information from the point at which the exception was thrown. So the information is there -- but is there a programmatic way to get it, for example so it can be logged, rather than having to examine a core file?

Impermissible

Look at this answer. – Unmoving

Answer 1

Edited Answer:

You can use std::set_terminate

#include <cstdlib>
#include <iostream>
#include <stdexcept>

#include <execinfo.h>

void
handler()
{
    void *trace_elems[20];
    int trace_elem_count(backtrace( trace_elems, 20 ));
    char **stack_syms(backtrace_symbols( trace_elems, trace_elem_count ));
    for ( int i = 0 ; i < trace_elem_count ; ++i )
    {
        std::cout << stack_syms[i] << "\n";
    }
    free( stack_syms );

    exit(1);
}   

int foo()
{
    throw std::runtime_error( "hello" );
}   

void bar()
{
    foo();
}

void baz()
{
    bar();
}

int
main()
{
    std::set_terminate( handler );
    baz();
}

giving this output:

samm@macmini ~> ./a.out 
./a.out [0x10000d20]
/usr/lib/libstdc++.so.6 [0xf9bb8c8]
/usr/lib/libstdc++.so.6 [0xf9bb90c]
/usr/lib/libstdc++.so.6 [0xf9bbaa0]
./a.out [0x10000c18]
./a.out [0x10000c70]
./a.out [0x10000ca0]
./a.out [0x10000cdc]
/lib/libc.so.6 [0xfe4dd80]
/lib/libc.so.6 [0xfe4dfc0]
samjmill@bgqfen4 ~> 

assuming you have debug symbols in your binary, you can then use addr2line to construct a prettier stack trace postmortem

samm@macmini ~> addr2line 0x10000c18
/home/samm/foo.cc:23
samm@macmini ~> 

original answer is below

---

I've done this in the past using boost::error_info to inject the stack trace using backtrace from execinfo.h into an exception that is thrown.

typedef boost::error_info<struct tag_stack_str,std::string> stack_info;

Then when catching the exceptions, you can do

} catch ( const std::exception& e ) {                                                                                                            
    if ( std::string const *stack boost::get_error_info<stack_error_info>(e) ) {                    
        std::cout << stack << std::endl;
    }
}

Uttasta

Yes, as I said I know how to get the trace at the current point. Clearly if I stored it into every exception that would be that. But the question assumes that I haven't done that -- perhaps due to laziness, perhaps because I don't have access to the source code. – Impermissible

what do you expect the stack trace to look like after the stack is unwound due to not catching an exception? – Uttasta

Like the stack trace I would get from applying gdb to the core file if I didn't intercept terminate() -- as clearly stated in my question. – Impermissible

Ah, very interesting! So I assume you are claiming that if I run those a.out addresses through addr2line I'll get the function calls on the stack! Perfect! – Impermissible

So it's easier than I thought. This is just the usual way to get the stack from inside the program. I assumed it wouldn't show anything after unwinding, and that is true if you catch the exception and then call backtrace(). But when terminate is called, the stack magically reappears. – Impermissible

@Impermissible I think the terminate handler is called prior to unwinding, does it not? In which case there is no magic (the throw does not unwind, its the fact that there is something there to catch it). If you were to catch it - you would lose that backtrace. Also about addr2line - its very useful - I used to use it from a post-mortem script and format the output so its much more readable.<p>one final note ... if you just compile with -rdynamic you wont need sddr2line, though with addr2line you can demangle C++ so there sre benefits there too – Inspire

Sam, your example program is evil. It changed both your user and host names ;-) – Benitobenjamen

What if the exception is thrown by a static lib like db_driver.a? – Gatepost

How do you set the boost::error_info to a std::exception? – Unmoving

Answer 2

Yet if I simply allow the program to abort(), it will produce a core-dump which contains the full stack information from the point at which the exception was thrown. So the information is there -- but is there a programmatic way to get it, for example so it can be logged, rather than having to examine a core file?

I doubt my experience would fit your needs but here it goes anyway.

I was overloading abort(): either by adding my own object file before the libc or using LD_PRELOAD. In my own version of abort() I was starting the debugger telling it to attach to the process (well, I surely know my PID) and dump the stack trace into a file (commands were passed to the debugger via command line). After debugger had finished, terminate the process with e.g. _exit(100).

That was on Linux using GDB. On Solaris I routinely employ similar trick but due to unavailability of a sane debugger I use the pstack tool: system("pstack <PID>").

Tweeter

Answer 3

You can use libunwind (just add -lunwind to linker parameters) (tested with clang++ 3.6):

demagle.hpp:

#pragma once

char const *
get_demangled_name(char const * const symbol) noexcept;

demangle.cpp:

#include "demangle.hpp"

#include <memory>

#include <cstdlib>

#include <cxxabi.h>

namespace
{

#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wglobal-constructors"
#pragma clang diagnostic ignored "-Wexit-time-destructors"
std::unique_ptr< char, decltype(std::free) & > demangled_name{nullptr, std::free};
#pragma clang diagnostic pop

}

char const *
get_demangled_name(char const * const symbol) noexcept
{
    if (!symbol) {
        return "<null>";
    }
    int status = -4;
    demangled_name.reset(abi::__cxa_demangle(symbol, demangled_name.release(), nullptr, &status));
    return ((status == 0) ? demangled_name.get() : symbol);
}

backtrace.hpp:

#pragma once
#include <ostream>

void
backtrace(std::ostream & _out) noexcept;

backtrace.cpp:

#include "backtrace.hpp"

#include <iostream>
#include <iomanip>
#include <limits>
#include <ostream>

#include <cstdint>

#define UNW_LOCAL_ONLY
#include <libunwind.h>

namespace
{

void
print_reg(std::ostream & _out, unw_word_t reg) noexcept
{
    constexpr std::size_t address_width = std::numeric_limits< std::uintptr_t >::digits / 4;
    _out << "0x" << std::setfill('0') << std::setw(address_width) << reg;
}

char symbol[1024];

}

void
backtrace(std::ostream & _out) noexcept
{
    unw_cursor_t cursor;
    unw_context_t context;
    unw_getcontext(&context);
    unw_init_local(&cursor, &context);
    _out << std::hex << std::uppercase;
    while (0 < unw_step(&cursor)) {
        unw_word_t ip = 0;
        unw_get_reg(&cursor, UNW_REG_IP, &ip);
        if (ip == 0) {
            break;
        }
        unw_word_t sp = 0;
        unw_get_reg(&cursor, UNW_REG_SP, &sp);
        print_reg(_out, ip);
        _out << ": (SP:";
        print_reg(_out, sp);
        _out << ") ";
        unw_word_t offset = 0;
        if (unw_get_proc_name(&cursor, symbol, sizeof(symbol), &offset) == 0) {
            _out << "(" << get_demangled_name(symbol) << " + 0x" << offset << ")\n\n";
        } else {
            _out << "-- error: unable to obtain symbol name for this frame\n\n";
        }
    }
    _out << std::flush;
}

backtrace_on_terminate.hpp:

#include "demangle.hpp"
#include "backtrace.hpp"

#include <iostream>
#include <type_traits>
#include <exception>
#include <memory>
#include <typeinfo>

#include <cstdlib>

#include <cxxabi.h>

namespace
{

[[noreturn]]
void
backtrace_on_terminate() noexcept;

static_assert(std::is_same< std::terminate_handler, decltype(&backtrace_on_terminate) >{});

#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wglobal-constructors"
#pragma clang diagnostic ignored "-Wexit-time-destructors"
std::unique_ptr< std::remove_pointer_t< std::terminate_handler >, decltype(std::set_terminate) & > terminate_handler{std::set_terminate(backtrace_on_terminate), std::set_terminate};
#pragma clang diagnostic pop

[[noreturn]]
void
backtrace_on_terminate() noexcept
{
    std::set_terminate(terminate_handler.release()); // to avoid infinite looping if any
    backtrace(std::clog);
    if (std::exception_ptr ep = std::current_exception()) {
        try {
            std::rethrow_exception(ep);
        } catch (std::exception const & e) {
            std::clog << "backtrace: unhandled exception std::exception:what(): " << e.what() << std::endl;
        } catch (...) {
            if (std::type_info * et = abi::__cxa_current_exception_type()) {
                std::clog << "backtrace: unhandled exception type: " << get_demangled_name(et->name()) << std::endl;
            } else {
                std::clog << "backtrace: unhandled unknown exception" << std::endl;
            }
        }
    }
    std::_Exit(EXIT_FAILURE);
}

}

There is good article concerning the issue.

Chopper

its not really: just add -lunwind to linker parameters. You need to have the library on your system first. – Uncaused

@AleksanderFular Surely, it is implied. – Chopper