User Leak, libc++ leak or false positive - c++

Question

I am building a dynamic library on mac in C++11 using the clang compiler and libc++ standard library. When I run valgrind on my test code which links to my dynamic library I get one block of memory that is definitely lost. Here is the valgrind report:
==45659== 36 bytes in 1 blocks are definitely lost in loss record 57 of 228
==45659== at 0x66BB: malloc (vg_replace_malloc.c:300)
==45659== by 0x31EAB0: __Balloc_D2A (in /usr/lib/system/libsystem_c.dylib)
==45659== by 0x31F2A5: __d2b_D2A (in /usr/lib/system/libsystem_c.dylib)
==45659== by 0x31BED6: __dtoa (in /usr/lib/system/libsystem_c.dylib)
==45659== by 0x3438A9: __vfprintf (in /usr/lib/system/libsystem_c.dylib)
==45659== by 0x36A2DA: __v2printf (in /usr/lib/system/libsystem_c.dylib)
==45659== by 0x34FF66: _vsnprintf (in /usr/lib/system/libsystem_c.dylib)
==45659== by 0x34FFC5: vsnprintf_l (in /usr/lib/system/libsystem_c.dylib)
==45659== by 0x34057A: snprintf_l (in /usr/lib/system/libsystem_c.dylib)
==45659== by 0x10C75A: std::__1::num_put<char, std::__1::ostreambuf_iterator<char, std::__1::char_traits<char> > >::do_put(std::__1::ostreambuf_iterator<char, std::__1::char_traits<char> >, std::__1::ios_base&, char, double) const (in /usr/lib/libc++.1.dylib)
==45659== by 0xF3221: std::__1::basic_ostream<char, std::__1::char_traits<char> >::operator<<(double) (in /usr/lib/libc++.1.dylib)
==45659== by 0x12102: lmpsdata::header_data::write_dimension(std::__1::basic_ofstream<char, std::__1::char_traits<char> >&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > const&) (header_data.cpp:75)
==45659==
==45659== LEAK SUMMARY:
==45659== definitely lost: 36 bytes in 1 blocks
==45659== indirectly lost: 0 bytes in 0 blocks
==45659== possibly lost: 0 bytes in 0 blocks
==45659== still reachable: 18,340 bytes in 215 blocks
==45659== suppressed: 25,274 bytes in 374 blocks
That particular section of code does not deal with any memory I have dynamically allocated, only STL objects and streams are used. I have attached both the header file and the method reported by valgrind as causing the leak. Can someone please explain what is going on here? I am absolutely baffled.
header_data.h
#ifndef ____header_data__
#define ____header_data__
#include <string>
#include <fstream>
#include <cstdint>
#include <vector>
#include <map>
namespace lmpsdata {
class header_data
{//LAMMPS header information
//only the point particle header information has been implemented
//since the current atom base class is designed for point particles only
public:
header_data():xdim(2), ydim(2), zdim(2), tiltdim(3) {}
//methods
void read(const std::string &, const std::string &);
void write(std::ofstream &, const std::string &); //this write command will replace the one
std::string check_header_keyword(const std::string &, bool &);
uint64_t get(const std::string&);
std::vector<double>& get_vector(const std::string&);
void set(const std::string&, uint64_t);
void set_vector(const std::string&, std::vector<double>&);
private:
//methods
void read_dimension(const std::string&, const std::string&);
void write_dimension(std::ofstream&, const std::string&);
//members
uint64_t atomnum;
uint64_t bondnum;
uint64_t anglenum;
uint64_t dihedralnum;
uint64_t impropernum;
uint64_t atomtypenum;
uint64_t bondtypenum;
uint64_t angletypenum;
uint64_t dihedraltypenum;
uint64_t impropertypenum;
uint64_t extrabondnum;
std::vector<double> xdim;
std::vector<double> ydim;
std::vector<double> zdim;
std::vector<double> tiltdim;//for use with triclinic system
std::map<std::string, uint64_t&> int_map {
{"atoms", atomnum},
{"bonds", bondnum},
{"angles", anglenum},
{"dihedrals", dihedralnum},
{"impropers", impropernum},
{"atom types", atomtypenum},
{"bond types", bondtypenum},
{"angle types", angletypenum},
{"dihedral types", dihedraltypenum},
{"improper types", impropertypenum},
{"extra bond per atom", extrabondnum},
};
std::map<std::string, std::vector<double>&> v_map {
{"xlo xhi", xdim},
{"ylo yhi", ydim},
{"zlo zhi", zdim},
{"xy xz yz", tiltdim}
};
};
}
#endif /* defined(____header_data__) */
lmpsdata.cpp only write_dimension method and beginning of file is shown
#include "header_data.h"
#include <stdexcept>
using namespace lmpsdata;
void header_data::write_dimension(std::ofstream &file, const std::string& keyword)
{
std::vector<double>& data = v_map.at(keyword);
for (auto value: data) {
file << value << " ";
}
}
If more information is needed please let me know.

Answer 1

You have not mentioned the exact versions of OSX and valgrind that you're using. I've not been able to entirely reproduce this on the version that I'm using (OSX 10.10; valgrind HEAD==Valgrind-3.11.0.SVN).
This is not in the C++ standard library, but in the C library. You should be able to reproduce it (almost identically) with the simple code:
#include <stdio.h>
#include <xlocale.h>
int
main(int argc, char **argv)
{
locale_t loc = newlocale(LC_NUMERIC_MASK, "C", (locale_t)0);
double d = 22.22;
char buffer[1024];
snprintf_l(buffer, 1024, loc, "%f\n", d);
printf("%s", buffer);
freelocale(loc);
}
When run with valgrind --show-leak-kinds=all --leak-check=full ./leak I see some 'still reachable' leaks (yours should show actual leaks in this case):
==26151== 32 bytes in 1 blocks are still reachable in loss record 28 of 85
==26151== at 0x10000850B: malloc (in /usr/local/Cellar/valgrind/HEAD/lib/valgrind/vgpreload_memcheck-amd64-darwin.so)
==26151== by 0x1002BC7DF: __Balloc_D2A (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002B9533: __rv_alloc_D2A (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002B9B3A: __dtoa (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002E1D52: __vfprintf (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x10030A9AE: __v2printf (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002EF154: _vsnprintf (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002EF1B3: vsnprintf_l (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002DF5F7: snprintf_l (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x100000ECA: main (leak.cpp:10)
==26151==
==26151== 36 bytes in 1 blocks are still reachable in loss record 30 of 85
==26151== at 0x10000850B: malloc (in /usr/local/Cellar/valgrind/HEAD/lib/valgrind/vgpreload_memcheck-amd64-darwin.so)
==26151== by 0x1002BC7DF: __Balloc_D2A (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002BD055: __d2b_D2A (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002B986B: __dtoa (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002E1D52: __vfprintf (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x10030A9AE: __v2printf (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002EF154: _vsnprintf (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002EF1B3: vsnprintf_l (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002DF5F7: snprintf_l (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x100000ECA: main (leak.cpp:10)
==26151==
==26151== 80 bytes in 1 blocks are still reachable in loss record 47 of 85
==26151== at 0x10000850B: malloc (in /usr/local/Cellar/valgrind/HEAD/lib/valgrind/vgpreload_memcheck-amd64-darwin.so)
==26151== by 0x1002BC736: __Balloc_D2A (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002BD055: __d2b_D2A (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002B986B: __dtoa (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002E1D52: __vfprintf (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x10030A9AE: __v2printf (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002EF154: _vsnprintf (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002EF1B3: vsnprintf_l (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x1002DF5F7: snprintf_l (in /usr/lib/system/libsystem_c.dylib)
==26151== by 0x100000ECA: main (leak.cpp:10)
I'd put in a suppression for it, as it's almost undoubtably not a real leak and is completely out of your control.
looking at the published OSX libc source, I can't see where the bug is present - 320 and later all look OK; with the underlying code appearing to allocate and free where necessary. The bug may have been introduced and fixed, though, as I didn't perform an exhaustive trawl of all the sources; the file in question is vfprintf.c, you're looking at the dtoaresult assignments & frees.
To determine your version of libc, you can do:
$ otool -l /usr/lib/system/libsystem_c.dylib | grep -A5 ID_
On my system I get the output:
cmd LC_ID_DYLIB
cmdsize 64
name /usr/lib/system/libsystem_c.dylib (offset 24)
time stamp 1 Thu Jan 1 01:00:01 1970
current version 1044.10.1
compatibility version 1.0.0
997, I think, is mavericks (10.9). I don't know if they introduced a leak in the code at some point in the mavericks timeframe and then fixed it - there doesn't appear to be a code path in the __vfprintf routine that actually leaks in any of the published source.