25% developed

Linux Applications Debugging Techniques/Print Version

From Wikibooks, open books for an open world
Jump to navigation Jump to search

Linux Applications Debugging Techniques

Current, editable version of this book is available in Wikibooks, collection of open-content textbooks at URL:

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts. A copy of the license is included in the section entitled "GNU Free Documentation License".


A collection of notes about debugging applications under Linux and related topics. Probably useful for debugging dogs. Partly applicable to other Unices. Work in progress.


Aurelian Melinte

Table of Contents

  1. The debugger 25% developed  as of Feb 25, 2012
  2. The dynamic linker 0% developed  as of Mar 24, 2012
  3. Core files 25% developed  as of Feb 25, 2012
  4. The call stack 75% developed  as of Mar 9, 2012
  5. The interposition library 75% developed  as of Feb 25, 2012
  6. Memory issues 0% developed  as of Feb 25, 2012
    1. Leaks 75% developed  as of Feb 25, 2012
    2. Heap corruption 0% developed  as of Feb 25, 2012
    3. Stack corruption 0% developed  as of Feb 25, 2012
  7. Signals 0% developed  as of Mar 22, 2022
  8. Deadlocks 0% developed  as of Feb 25, 2012
  9. Race conditions 0% developed  as of Mar 03, 2012
  10. Resource leaks 0% developed  as of Apr 25, 2012
  11. Transactional memory 0% developed  as of Feb 11, 2013
  12. The compiler 0% developed  as of Dec 24, 2012
  13. Aiming for and measuring performance 0% developed  as of Feb 25, 2012
  14. Hardware emulators 0% developed  as of Feb 12, 2013
  15. Building the toolkit 25% developed  as of May 21, 2012
  16. Appendices 0% developed  as of Feb 25, 2012
  17. References and further reading 0% developed  as of Feb 25, 2012

Please add {{alphabetical}} only to book title pages.

The debugger


Someday some hard to reproduce issue will be found on a production machine. Typically, such a machine is difficult to access, has no development environment and nothing can be installed on it. At most it will have gdb, but very likely not. Typically also, the heaviest user will be the one to find the issue and typically still, the heaviest user is the one being the bigger -- money-wise. And that issue has to be root cause diagnosed and fixed.

Unless the application has been prepared for this forensic gathering moment, not much can be done to diagnose where the issue is. Thus, preparations should start with compilation:

  • Have a "symbol server" and make sure it is reacheable. Compile on the symbol server.
  • Compile the release with debugging symbols. Strip them if you do not want to ship them but keep them.
  • Ship gdbserver with the application for remote debugging.

These preparations will allow one to:

  • Debug the application running on any machine, including machines where there is no gdb installed.
  • Debug from any machine that has network visibility to the symbol server.

Also, think beforehand how would you debug on the machine:

  • Embed a breakpoint in the code, at hard of reach places of interest, then
    • Start the application
    • Attach to it with the debugger
    • Wait until the breakpoint is hit

The "symbol server"

One way to easily reach the right code from within the debugger is to build the binaries within an auto-mounted folder, each build in its own sub-folder. The same auto-mount share should be accessible from the machine you are debugging on.

  • Install autofs
  • In /etc/auto.master uncomment the line:
/net	-hosts
  • In /etc/exports export the folder:
# Build directory: rw for localhost, read-only for any other host
/home/amelinte/projects/lpt/lpt  localhost(rw,sync,no_subtree_check)  *(ro,root_squash,no_subtree_check)
  • As root: restart autofs & nfs and export the build share:
/etc/init.d/autofs restart
/etc/init.d/nfs-kernel-server restart
/usr/sbin/exportfs -a
  • Export the folder: edit /etc/exports
  • As root (RedHat): service autofs start

Finally, build the binaries within the automounted directory on the build machine (here the build machine is bear):

cd /net/bear/home/amelinte/projects/lpt/lpt.destructor
make clean

Notice the filename path that is resolved with the symbol information:

[0x413464] lpt::stack::call_stack<40ul>::call_stack(bool)+0x52
        At /net/bear/home/amelinte/projects/lpt/lpt.destructor/lpt/include/lpt/call_stack.hpp:74
        In binaries/bin/stacktest

If the symbols have been stripped from the binaries, point gdb to the folders where the symbols are with the debug-file-directory directive.


Source Code

To point the debugger to the source files:

(gdb) set substitute-path /from/path1 /to/path1
(gdb) set substitute-path /from/path2 /to/path2

Remote debugging

  • On the machine where the application runs (appmachine):
    • If gdbserver is not present, copy it over.
    • Start the application.
    • Start gdbserver: gdbserver gdbmachine:2345 --attach program
  • On gdbmachine:
    • At the gdb prompt, enter: target remote appmachine:2345

Sometimes you may have to tunnel over ssh:

  • On gdbmachine:
    • ssh -L 5432:appmachine:2345 user@appmachine
    • At the gdb prompt: target remote localhost:5432

Attaching to a process

Find out the PID of the process, then:

(gdb) attach 1045
Attaching to process 1045
Reading symbols from /usr/lib64/firefox-3.0.18/firefox...(no debugging symbols found)...done.
Reading symbols from /lib64/libpthread.so.0...(no debugging symbols found)...done.
[Thread debugging using libthread_db enabled]
[New Thread 0x448b4940 (LWP 1063)]
[New Thread 0x428b0940 (LWP 1054)]
(gdb) detach

Debugging programs that spawn multiple children

  • set detach-on-fork off
  • see "all-stop" vs "non-stop" modes in the GDB documentation and their related settings

Embedding breakpoints in the source

On x86 platforms:

#define EMBEDDED_BREAKPOINT  asm volatile ("int3;")

Or a more elaborate one:

    asm("0:"                              \
        ".pushsection embed-breakpoints;" \
        ".quad 0b;"                       \

This will break into the debugger on hard to reach conditions:

    if (cpuload > 3.1416 && started > 111*MINS*AGO && whatever-dynamic-condition) 

Data breakpoints (watchpoints)

Watchpoints can be implemented either in software either in hardware if the CPU supports it. Typically on an Intel processor there are eight debug registers out of which only four can be used for hardware breakpoints and this limits the number of watchpoints system wide.

Breakpoints conditional on caller

This requires gdb 7.9 or later, configured with python support:

(gdb) break funcA if $_caller_is("funcB")

The text user interface


GDB features a text user interface for code, disassembler and registers. For instance:

  • Ctrl-x 1 will show the code pane
  • Ctrl-x a will hide the TUI panes

None of the GUI interfaces to gdb (Qt Creator stands out for being intuitive and easy to use) can offer access to all of the gdb functionality.

curses gdb offers an improved TUI. A comprehensive list of debugger GUIs is available here.

Reverse debugging

As an example, reverse debugging is a functionality no GUI offers access to:

(gdb) l
1	/* 1 */ int var;
2	/* 2 */ int main (void) {
3	/* 3 */   int i; for (i=0;i<100;i++)
4	/* 4 */     var=i/10;
5	/* 5 */  return 0; }

(gdb) start
(gdb) record 
(gdb) adv 5
main () at history.c:5
5	/* 5 */  return 0; }

(gdb) watch var
Hardware watchpoint 2: var

(gdb) reverse-continue 
Hardware watchpoint 2: var

Old value = 9
New value = 8
0x00000000004004c3 in main () at history.c:4
4	/* 4 */     var=i/10;
(gdb) p i
$1 = 90

(gdb) reverse-continue 
Hardware watchpoint 2: var

Old value = 8
New value = 7
0x00000000004004c3 in main () at history.c:4
4	/* 4 */     var=i/10;

(gdb) p i
$2 = 80

See also

Register watch

You can watch registers. Note this will force the debugger to single step the debugged program and it will run very slowly:

(gdb) watch $ebp


As a note, in upcoming gdb releases, .gdbinit will be replaced by gdb-gdb.gdb:

 ^   ^   ^
 |   |   |---- It's a gdb script.
 |   |         If it were Python this would be .py.
 |   |
 |   --------- "-gdb" is a gdb convention, it's the suffix added to a file
 |             for auxiliary support.
 |             E.g., gdb will auto-load libstdc++.so-gdb.py (version elided)
 |             which contains the std c++ pretty-printers.
 ------------- This init script is for the program named "gdb".
               If this were for readelf the script would be named

C++ support

Canned gdb macros

gdb might need a bit of guidance with C++11 binaries:

(gdb) set cp-abi gnu-v3
(gdb) set language c++
(gdb) maintenance demangle _ZNSt16nested_exceptionD2Ev

One can eventually use templight to debug and profile templates.

Getting the subclass from the base class pointer
(gdb) python print(x.type)
SuperClass *
(gdb) python print(x.dynamic_type)
SubClass *


The dynamic linker


Resolved symbols

To find out which dynamic library is a symbol coming from:

$ LD_DEBUG_OUTPUT=sym.log LD_DEBUG=bindings /bin/ls

$ cat sym.log.7688 | grep malloc
      7688:     binding file /lib/i686/cmov/libc.so.6 [0] to /lib/i686/cmov/libc.so.6 [0]: normal symbol `__malloc_initialize_hook' [GLIBC_2.0]
      7688:     binding file /lib/i686/cmov/libc.so.6 [0] to /bin/ls [0]: normal symbol `malloc' [GLIBC_2.0]
      7688:     binding file /lib/i686/cmov/libc.so.6 [0] to /lib/i686/cmov/libc.so.6 [0]: normal symbol `__malloc_hook' [GLIBC_2.0]
      7688:     binding file /lib/ld-linux.so.2 [0] to /lib/i686/cmov/libc.so.6 [0]: normal symbol `malloc' [GLIBC_2.0]
      7688:     binding file /lib/i686/cmov/libc.so.6 [0] to /lib/i686/cmov/libc.so.6 [0]: normal symbol `malloc' [GLIBC_2.0]
      7688:     binding file /bin/ls [0] to /lib/i686/cmov/libc.so.6 [0]: normal symbol `malloc' [GLIBC_2.0]
$ LD_DEBUG=help /bin/ls
Valid options for the LD_DEBUG environment variable are:

  libs        display library search paths
  reloc       display relocation processing
  files       display progress for input file
  symbols     display symbol table processing
  bindings    display information about symbol binding
  versions    display version dependencies
  all         all previous options combined
  statistics  display relocation statistics
  unused      determined unused DSOs
  help        display this help message and exit

To direct the debugging output into a file instead of standard output
a filename can be specified using the LD_DEBUG_OUTPUT environment variable.


  • It is run under whatever credentials would be accorded to the application program; thus, for example, if a set-user-ID-root program is being executed, the dynamic linker will run with an effective user ID of root.

Core files

A core dump is a snapshot of the memory of the program, processor registers including program counter and stack pointer and other OS and memory management information, taken at a certain point in time. As such, they are invaluable for capturing the state of rare occurring races and abnormal conditions.

What is more, such rarities will be found usually on heavily used production or QA machines where gdb is not available, nor is access easy to the machine. Worse, the heaviest users are usually the biggest clients (moneywise...). As such, it is important to get as much forensic data as available, and plan for it.

One can force a core dump from within the program or from outside at chosen moments. What a core cannot tell is how the application ended up in that state: the core is no replacement for a good log. Verbose logs and core files go hand in glove.


For a process to be able to dump core, a few prerequisites have to be met:

  • the set core size limit should permit it (see the man page for ulimit). E.g.: ulimit -c unlimited. It can also be set from within the program.
  • the process to dump core should have write permissions to the folder where the core is to be dumped to (usually the current working directory of the process)

Where is my core?

Usually the core is dumped in the current working directory of the process. But the OS can be configured otherwise:

# cat /proc/sys/kernel/core_pattern

# sysctl -w "kernel.core_pattern=/var/cores/%h-%e-%p.core"

Dumping core from outside the program

One possibility is with gdb, if available. This will let the program running:

(gdb) attach <pid>
(gdb) generate-core-file <optional-filename>
(gdb) detach

Another possibility is to signal the process. This will terminate it, assuming the signal is not caught by a custom signal handler:

kill -s SIGABRT <pid>

Dumping core from within the program

Again, there are two possibilities: dump core and terminate the program or dump and continue:

void dump_core_and_terminate(void)
     * Alternative:
     *   char *p = NULL; *p = 0;

void dump_core_and_continue(void)
    pid_t child = fork();
    if (child < 0) {
        /*Parent: error*/
    else if (child == 0) {
        dump_core_and_terminate(); /*Child*/
    else {
        /*Parent: continue*/

Note: use dump_core_and_continue() with care: in a multi-threaded program, the forked child will have only a clone of the parent thread that called fork() [Butenhof Ch5; re: threads & fork]. This has number of implications, in particular with respect to mutexes, but the particular point here is that the core that the child will dump will contain information only for one thread. If you need to dump a core with all threads without aborting the process, try to use the google core dumper library, even if it has not been maintained for years.

Shared libraries

To obtain a good call stack, it is important that the gdb loads the same libraries that were loaded by the program that generated the core dump. If the machine we are analyzing the core has different libraries (or has them in different places) from the machine the core was dumped, then copy over the libraries to the analyzing machine, in a way that mirrors the dump machine. For instance:

$ tree .
|-- juggler-29964.core
|-- lib64
|   |-- ld-linux-x86-64.so.2
|   |-- libc.so.6
|   |-- libm.so.6
|   |-- libpthread.so.0
|   `-- librt.so.1

At the gdb prompt:

(gdb) set solib-absolute-prefix ./
(gdb) set solib-search-path .
(gdb) file ../../../../../threadpool/bin.v2/libs/threadpool/example/juggler/gcc-4.1.2/debug/link-static/threading-multi/juggler
Reading symbols from /home/aurelian_melinte/threadpool/threadpool-0_2_5-src/threadpool/bin.v2/libs/threadpool/example/juggler/gcc-4.1.2/debug/link-static/threading-multi/juggler...done.
(gdb) core-file juggler-29964.core 
Reading symbols from ./lib64/librt.so.1...(no debugging symbols found)...done.
Loaded symbols for ./lib64/librt.so.1
Reading symbols from ./lib64/libm.so.6...(no debugging symbols found)...done.
Loaded symbols for ./lib64/libm.so.6
Reading symbols from ./lib64/libpthread.so.0...(no debugging symbols found)...done.
Loaded symbols for ./lib64/libpthread.so.0
Reading symbols from ./lib64/libc.so.6...(no debugging symbols found)...done.
Loaded symbols for ./lib64/libc.so.6
Reading symbols from ./lib64/ld-linux-x86-64.so.2...(no debugging symbols found)...done.
Loaded symbols for ./lib64/ld-linux-x86-64.so.2
Core was generated by `../../../../bin.v2/libs/threadpool/example/juggler/gcc-4.1.2/debug/link-static/'.
Program terminated with signal 6, Aborted.
#0  0x0000003684030265 in raise () from ./lib64/libc.so.6
(gdb) frame 2
#2  0x0000000000404ae1 in dump_core_and_terminate () at juggler.cpp:30

Source Code

To point the debugger to the source files:

(gdb) set substitute-path /from/path1 /to/path1
(gdb) set substitute-path /from/path2 /to/path2


Here is a script that will generate a basic report per core file. Useful the days when cores are raining on you:


# A script to extract core-file informations

if [ $# -ne 1 ]
  echo "Usage: `basename $0` <for-binary-image>"
  exit -1

# Today and yesterdays cores
cores=`find . -name '*.core' -mtime -1`

#cores=`find . -name '*.core'`

for core in $cores 
  rm $gdblogfile
  bininfo=`ls -l $binimg`
  coreinfo=`ls -l $core`
  gdb -batch \
      -ex "set logging file $gdblogfile" \
      -ex "set logging on" \
      -ex "set pagination off" \
      -ex "printf \"**\n** Process info for $binimg - $core \n** Generated `date`\n\"" \
      -ex "printf \"**\n** $bininfo \n** $coreinfo\n**\n\"" \
      -ex "file $binimg" \
      -ex "core-file $core" \
      -ex "bt" \
      -ex "info proc" \
      -ex "printf \"*\n* Libraries \n*\n\"" \
      -ex "info sharedlib" \
      -ex "printf \"*\n* Memory map \n*\n\"" \
      -ex "info target" \
      -ex "printf \"*\n* Registers \n*\n\"" \
      -ex "info registers" \
      -ex "printf \"*\n* Current instructions \n*\n\"" -ex "x/16i \$pc" \
      -ex "printf \"*\n* Threads (full) \n*\n\"" \
      -ex "info threads" \
      -ex "bt" \
      -ex "thread apply all bt full" \
      -ex "printf \"*\n* Threads (basic) \n*\n\"" \
      -ex "info threads" \
      -ex "thread apply all bt" \
      -ex "printf \"*\n* Done \n*\n\"" \
      -ex "quit" 

An alternative worth exploring is btparser.

Canned user-defined commands

Same reporting functionality can be canned for gdb:

define procinfo
    printf "**\n** Process Info: \n**\n"
    info proc
    printf "*\n* Libraries \n*\n"
    info sharedlib
    printf "*\n* Memory Map \n*\n"
    info target
    printf "*\n* Registers \n*\n"
    info registers
    printf "*\n* Current Instructions \n*\n" 
    x/16i $pc
    printf "*\n* Threads (basic) \n*\n"
    info threads
    thread apply all bt
document procinfo
Infos about the debugee. 

define analyze
    printf "*\n* Threads (full) \n*\n"
    info threads
    thread apply all bt full


A script that will generate a basic report and a core file for a running process:


# A script to generate a core and a status report for a running process. 

if [ $# -ne 1 ]
  echo "Usage: `basename $0` <PID>"
  exit -1

rm $gdblogfile

gdb -batch \
      -ex "set logging file $gdblogfile" \
      -ex "set logging on" \
      -ex "set pagination off" \
      -ex "printf \"**\n** Process info for PID=$pid \n** Generated `date`\n\"" \
      -ex "printf \"**\n** Core: $corefile \n**\n\"" \
      -ex "attach $pid" \
      -ex "bt" \
      -ex "info proc" \
      -ex "printf \"*\n* Libraries \n*\n\"" \
      -ex "info sharedlib" \
      -ex "printf \"*\n* Memory map \n*\n\"" \
      -ex "info target" \
      -ex "printf \"*\n* Registers \n*\n\"" \
      -ex "info registers" \
      -ex "printf \"*\n* Current instructions \n*\n\"" -ex "x/16i \$pc" \
      -ex "printf \"*\n* Threads (full) \n*\n\"" \
      -ex "info threads" \
      -ex "bt" \
      -ex "thread apply all bt full" \
      -ex "printf \"*\n* Threads (basic) \n*\n\"" \
      -ex "info threads" \
      -ex "thread apply all bt" \
      -ex "printf \"*\n* Done \n*\n\"" \
      -ex "generate-core-file $corefile" \
      -ex "detach" \
      -ex "quit"

Thread Local Storage

TLS data is rather difficult to access with gdb in the core files, and __tls_get_addr() cannot be called.



The call stack


Sometimes we need the call stack at a certain point in the program. These are the API functions to get basic stack information:

#include <execinfo.h>

int backtrace(void **buffer, int size);
char **backtrace_symbols(void *const *buffer, int size);
void backtrace_symbols_fd(void *const *buffer, int size, int fd); 

#include <cxxabi.h>
char* __cxa_demangle(const char* __mangled_name, char* __output_buffer, size_t* __length, int* __status);

#include <dlfcn.h>
int dladdr(void *addr, Dl_info *info);


  • C++ symbols are still mangled. Use abi::__cxa_demangle() or something similar.
  • Some of the these functions do allocate memory - either temporarily either explicitly - and this might be a problem if the program is unstable already.
  • Some of the these functions do acquire locks (e.g. dladdr()).
  • backtrace* are expensive calls on some platforms (x86_64 for instance), the deeper the call stack is the more expensive the call is.
  • dladdr will fail to rsolve any symbol that is not exported. Hence, to get the most out of it:
    • Link with -rdynamic and -ldl

To extract more information (file, line number), use libbfd. Possible alternatives: libbacktrace; http://libcsdbg.sourceforge.net/

#define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#include <dlfcn.h>

#include <execinfo.h>
#include <signal.h>
#include <bfd.h>
#include <unistd.h>

/* globals retained across calls to resolve. */
static bfd* abfd = 0;
static asymbol **syms = 0;
static asection *text = 0;

static void resolve(char *address) {
    if (!abfd) {
 	  char ename[1024];
 	  int l = readlink("/proc/self/exe",ename,sizeof(ename));
 	  if (l == -1) {
 	    perror("failed to find executable\n");
 	  ename[l] = 0;


 	  abfd = bfd_openr(ename, 0);
 	  if (!abfd) {
 	      perror("bfd_openr failed: ");
 	  /* oddly, this is required for it to work... */

 	  unsigned storage_needed = bfd_get_symtab_upper_bound(abfd);
 	  syms = (asymbol **) malloc(storage_needed);
 	  unsigned cSymbols = bfd_canonicalize_symtab(abfd, syms);

 	  text = bfd_get_section_by_name(abfd, ".text");
    long offset = ((long)address) - text->vma;
    if (offset > 0) {
        const char *file;
        const char *func;
        unsigned line;
        if (bfd_find_nearest_line(abfd, text, syms, offset, &file, &func, &line) && file)
            printf("file: %s, line: %u, func %s\n",file,line,func);

A C++ wrapper

With this simple class, getting the stack is one line away:

class call_stack 

    static const int depth = 40;
    typedef std::array<void *, depth> stack_t;
    class const_iterator;    
    class frame
        frame(void *addr = 0) 
                : _addr(0)
                , _dladdr_ret(false)
                , _binary_name(0) 
                , _func_name(0)
                , _demangled_func_name(0)
                , _delta_sign('+')
                , _delta(0L)
                , _source_file_name(0)
                , _line_number(0)

        // frame(stack_t::iterator& it) : frame(*it) {} //C++0x
        frame(stack_t::const_iterator const& it) 
                : _addr(0)
                , _dladdr_ret(false)
                , _binary_name(0) 
                , _func_name(0)
                , _demangled_func_name(0)
                , _delta_sign('+')
                , _delta(0L)
                , _source_file_name(0)
                , _line_number(0)
        frame(frame const& other)
        frame& operator=(frame const& other)
            if (this != &other) {
            return *this;
        std::string as_string() const
            std::ostringstream s;
            s << "[" << std::hex << _addr << "] " 
              << demangled_function() 
              << " (" << binary_file() << _delta_sign << "0x" << std::hex << _delta << ")"
              << " in " << source_file() << ":" << line_number()
            return s.str();

        const void* addr() const               { return _addr; }
        const char* binary_file() const        { return safe(_binary_name); }
        const char* function() const           { return safe(_func_name); }
        const char* demangled_function() const { return safe(_demangled_func_name); }
        char        delta_sign() const         { return _delta_sign; }
        long        delta() const              { return _delta; }
        const char* source_file() const        { return safe(_source_file_name); } 
        int         line_number() const        { return  _line_number; }
        const char* safe(const char* p) const { return p ? p : "??"; }
        friend class const_iterator; // To call resolve()
        void resolve(const void * addr)
            if (_addr == addr)
            _addr = addr;
            _dladdr_ret = false; 
            _binary_name = 0; 
            _func_name = 0;
            if (_demangled_func_name) {
                _demangled_func_name = 0;
            _delta_sign = '+';
            _delta = 0L;
            _source_file_name = 0; 
            _line_number = 0;
            if (!_addr)
            _dladdr_ret = (::dladdr(_addr, &_info) != 0);
            if (_dladdr_ret)
                _binary_name = safe(_info.dli_fname);
                _func_name   = safe(_info.dli_sname);
                _delta_sign  = (_addr >=  _info.dli_saddr) ? '+' : '-';
                _delta = ::labs(static_cast<const char *>(_addr) - static_cast<const char *>(_info.dli_saddr));
                 int status = 0;
                 _demangled_func_name = abi::__cxa_demangle(_func_name, 0, 0, &status);
        const void* _addr;
        const char* _binary_name;
        const char* _func_name;
        const char* _demangled_func_name;
        char        _delta_sign;
        long        _delta;
        const char* _source_file_name; //TODO: libbfd
        int         _line_number;
        Dl_info     _info;
        bool        _dladdr_ret;
    }; //frame
    class const_iterator
            : public std::iterator< std::bidirectional_iterator_tag
                                  , ptrdiff_t 
        const_iterator(stack_t::const_iterator const& it)
                : _it(it)
                , _frame(it)
        bool operator==(const const_iterator& other) const 
            return _frame.addr() == other._frame.addr(); 

        bool operator!=(const const_iterator& x) const 
            return !(*this == x);

        const frame& operator*() const 
            return _frame;
        const frame* operator->() const 
            return &_frame;

        const_iterator& operator++() 
            return *this;
        const_iterator operator++(int) 
            const_iterator tmp = *this;
            return tmp;

        const_iterator& operator--() 
            return *this;
        const_iterator operator--(int) 
            const_iterator tmp = *this;
            return tmp;
        frame                    _frame;
        stack_t::const_iterator  _it;
    }; //const_iterator

    call_stack() : _num_frames(0)
        _num_frames = ::backtrace(_stack.data(), depth);
        assert(_num_frames >= 0 && _num_frames <= depth);
    std::string as_string()
        std::string s;
        const_iterator itEnd = end();
        for (const_iterator it = begin(); it != itEnd; ++it) {
            s += it->as_string();
            s += "\n";
        return std::move(s);
    virtual ~call_stack()
    const_iterator begin() const { return _stack.cbegin(); }
    const_iterator end() const   { return stack_t::const_iterator(&_stack[_num_frames]); }


    stack_t  _stack;
    int      _num_frames;

This class can be used with assertions, logging or exceptions.

The full code can be found at the LPT site and it offers three symbol resolver varieties. This code will require a C++11 compiler.

  • A "basic" address-only resolver, without memory side-effects
  • A libc resolver
  • A libbfd resolver

A boost-based version of the call stack utilities collection can be found at https://github.com/melintea/Boost-Call_stack.

A stack resolved with the libc resolver, embedded in an exception, with debug-compiled code would look like:

Exception PAPI_add_event: Event exists, but cannot be counted due to hardware resource limits
0x41e11f ??+0x41e11f
        At ??:0
        In binaries/bin/papitest
0x421ae0 lpt::stack::call_stack&lt;40ul&gt;::call_stack(bool)+0x52
        At ??:0
        In binaries/bin/papitest
0x420653 lpt::papi::papi_error::papi_error(std::string const&, int)+0x6f
        At ??:0
        In binaries/bin/papitest
0x426ffe lpt::papi::library&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt;::init()+0x722
        At ??:0
        In binaries/bin/papitest
0x426815 lpt::papi::library&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt;::library()+0x5b
        At ??:0
        In binaries/bin/papitest
0x4265d4 lpt::singleton&lt;lpt::papi::library&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt; &gt;::singleton()+0x18
        At ??:0
        In binaries/bin/papitest
0x42624e lpt::singleton&lt;lpt::papi::library&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt; &gt;::instance()+0x42
        At ??:0
        In binaries/bin/papitest
0x425b47 lpt::papi::thread&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt;::init()+0x25
        At ??:0
        In binaries/bin/papitest
0x425502 lpt::papi::thread&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt;::thread()+0x2e
        At ??:0
        In binaries/bin/papitest
0x424554 lpt::singleton&lt;lpt::papi::thread&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt; &gt;::singleton()+0x18
        At ??:0
        In binaries/bin/papitest
0x4230a5 lpt::singleton&lt;lpt::papi::thread&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt; &gt;::instance()+0x42
        At ??:0
        In binaries/bin/papitest
0x421d33 lpt::papi::counters&lt;lpt::papi::stdout_print, lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt;::counters(std::string const&)+0xc9
        At ??:0
        In binaries/bin/papitest
0x41e206 tests()+0x67
        At ??:0
        In binaries/bin/papitest
0x41edb8 ??+0x11
        At ??:0
        In binaries/bin/papitest
0x2ad66a7b9ead ??+0xfd
        At ??:0
        In /lib/x86_64-linux-gnu/libc.so.6
0x41dea9 ??+0x41dea9
        At ??:0
        In binaries/bin/papitest

Note how dladdr() fails to resolve some of the frames. Whereas a stack resolved with the bfd resolver would show file and line numbers and would resolve all frames:

Exception PAPI_add_event: Event exists, but cannot be counted due to hardware resource limits
[0x41e20f] backtrace+0x41e20f
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/detail/glibc_call_stack.hpp:61
        In binaries/bin/papitest
[0x421c4e] lpt::stack::call_stack&lt;40ul&gt;::call_stack(bool)+0x52
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/call_stack.hpp:74
        In binaries/bin/papitest
[0x4207a1] lpt::papi::papi_error::papi_error(std::string const&, int)+0x6f
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/papi.hpp:112
        In binaries/bin/papitest
[0x42716c] lpt::papi::library&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt;::init()+0x722
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/papi.hpp:371
        In binaries/bin/papitest
[0x426983] lpt::papi::library&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt;::library()+0x5b
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/papi.hpp:289
        In binaries/bin/papitest
[0x426742] lpt::singleton&lt;lpt::papi::library&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt; &gt;::singleton()+0x18
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/singleton.hpp:19
        In binaries/bin/papitest
[0x4263bc] lpt::singleton&lt;lpt::papi::library&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt; &gt;::instance()+0x42
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/singleton.hpp:28
        In binaries/bin/papitest
[0x425cb5] lpt::papi::thread&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt;::init()+0x25
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/papi.hpp:433
        In binaries/bin/papitest
[0x425670] lpt::papi::thread&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt;::thread()+0x2e
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/papi.hpp:409
        In binaries/bin/papitest
[0x4246c2] lpt::singleton&lt;lpt::papi::thread&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt; &gt;::singleton()+0x18
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/singleton.hpp:19
        In binaries/bin/papitest
[0x423213] lpt::singleton&lt;lpt::papi::thread&lt;lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt; &gt;::instance()+0x42
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/singleton.hpp:28
        In binaries/bin/papitest
[0x421ea1] lpt::papi::counters&lt;lpt::papi::stdout_print, lpt::papi::counting_per_thread, lpt::papi::multiplex_none&gt;::counters(std::string const&)+0xc9
        At /home/amelinte/projects/lpt/lpt/lpt/include/lpt/papi.hpp:646
        In binaries/bin/papitest
[0x41e2f6] tests()+0x67
        At /home/amelinte/projects/lpt/lpt/tests/papitest.cpp:51
        In binaries/bin/papitest
[0x41eea8] main+0x11
        At /home/amelinte/projects/lpt/lpt/tests/papitest.cpp:176
        In binaries/bin/papitest
[0x2b5f7aa1bead] __libc_start_main+0xfd
        At ??:0
        In /lib/x86_64-linux-gnu/libc.so.6
[0x41df99] _start+0x41df99
        At ??:0
        In binaries/bin/papitest

  • Do NOT use in asynchronous interrupts such as signal handlers.
  • It might or it might not work when the program is unstable - prefer a core dump if possible. Same for memory issues handlers .
  • Performance hit varies greatly with platform and compiler version.

From within gdb

A canned command to resolve a stack address from within gdb:

define addrtosym
    if $argc == 1
        printf "[%u]: ", $arg0
        #whatis/ptype EXPR
        #info frame ADDR
        info symbol $arg0
document addrtosym
Resolve the address (e.g. of one stack frame). Usage: addrtosym addr0


The interposition library

The dynamic liker allows for interception of any function call an application makes to any shared library it uses. As such, interposition is a powerful technique allowing to tune performance, collect runtime statistics, or debug the application without having to instrument the code of that application.

As an example, we can use an interposition library to trace calls, with arguments' values and return codes.

Call tracing

Note that part of code below is 32-bit x86 and gcc 4.1/4.2 specific.

Code instrumentation

In the library, we want to address the following points:

  • when a function/method is entered and exited.
  • what were the call arguments when the function is entered.
  • what was the return code when the function is exited.
  • optionally, where was the function called from.

The first one is easy: if requested, the compiler will instrument functions and methods so that when a function/method is entered, a call to an instrumentation function is made and when the function is exited, a similar instrumentation call is made:

    void __cyg_profile_func_enter(void *func, void *callsite); 
    void __cyg_profile_func_exit(void *func, void *callsite);

This is achieved by compiling the code with the -finstrument-functions flag. The above two functions can be used for instance to collect data for coverage; or for profiling. We will use them to print a trace of function calls. Furthermore, we can isolate these two functions and the supporting code in an interposition library of our own. This library can be loaded when and if needed, thus leaving the application code basically unchanged.

Now when the function is entered we can get the arguments of the call:

    void __cyg_profile_func_enter( void *func, void *callsite )
        char buf_func[CTRACE_BUF_LEN+1] = {0};
        char buf_file[CTRACE_BUF_LEN+1] = {0}; 
        char buf_args[ARG_BUF_LEN + 1] = {0}; 
        pthread_t self = (pthread_t)0;
        int *frame = NULL;
        int nargs = 0;
        self = pthread_self(); 
        frame = (int *)__builtin_frame_address(1); /*of the 'func'*/

        /*Which function*/
        libtrace_resolve (func, buf_func, CTRACE_BUF_LEN, NULL, 0);
        /*From where.  KO with optimizations. */
        libtrace_resolve (callsite, NULL, 0, buf_file, CTRACE_BUF_LEN);
        nargs = nchr(buf_func, ',') + 1; /*Last arg has no comma after*/ 
        nargs += is_cpp(buf_func);       /*'this'*/
        if (nargs > MAX_ARG_SHOW)
            nargs = MAX_ARG_SHOW;

        printf("T%p: %p %s %s [from %s]\n", 
               self, (int*)func, buf_func, 
               args(buf_args, ARG_BUF_LEN, nargs, frame), 

And when the function is exited, we get the return value:

    void __cyg_profile_func_exit( void *func, void *callsite )
        long ret = 0L; 
        char buf_func[CTRACE_BUF_LEN+1] = {0};
        char buf_file[CTRACE_BUF_LEN+1] = {0}; 
        pthread_t self = (pthread_t)0;

        self = pthread_self(); 

        /*Which function*/
        libtrace_resolve (func, buf_func, CTRACE_BUF_LEN, NULL, 0);
        printf("T%p: %p %s => %d\n", 
               self, (int*)func, buf_func, 


Since these two instrumentation functions are aware of addresses and we actually want the trace to be readable by humans, we need also a way to resolve symbol addresses to symbol names: this is what libtrace_resolve() does.

Binutils and libbfd

First, we have to have the symbols information handy. To achieve this, we compile our application with the -g flag. Then, we can map addresses to symbol names and this would normally require writing some code knowledgeable of the ELF format.

Luckily, the there is the binutils package which comes with a library that does just that: libbfd; and with a tool: addr2line. addr2line is a good example on how to use libbfd and I have simply used it to wrap around libbfd. The result is the libtrace_resolve() function.

Since the instrumentation functions are isolated in a stand-alone module, we tell this module the name of the instrumented executable through an environment variable (CTRACE_PROGRAM) that we set before running the program. This is needed to properly init libbfd to search for symbols.

Stack layout

To address the first point the work has been architecture-agnostic (actually libbfd is aware of the architecture, but things are hidden behind its API). However, to retrieve function arguments and return values we have to look at the stack, write a bit of architecture-specific code and exploit some gcc quirks. Again, the compilers I have used were gcc 4.1 and 4.2; later or previous versions might work differently. In short:

  • x86 dictates that stack grows down.
  • GCC dictates how the stack is used - a "typical" stack is depicted below.
  • each function has a stack frame marked by the ebp (base pointer) and esp (stack pointer) registers.
  • normally, we expect the eax register to contain the return code
               +------------+   |
               | arg 2      |   \
               +------------+    >- previous function's stack frame
               | arg 1      |   /
               +------------+   |
               | ret %eip   |   /
               | saved %ebp |   \
        %ebp-> +------------+   |
               |            |   |
               |   local    |   \
               | variables, |    >- current function's stack frame
               |    etc.    |   /
               |            |   |
               |            |   |
        %esp-> +------------+   /

In an ideal world, the code the compiler generates would make sure that upon instrumenting the exit of a function: the return value is set, then CPU registers pushed on the stack (to ensure the instrumentation function does not affects them), then call the instrumentation function and then pop the registers. This sequence of code would ensure we always get access to the

return value in the instrumentation function. The code generated by the compiler is a bit different...

Also, in practice, many of gcc's flags affect the stack layout and registers usage. The most obvious ones are:

  • -fomit-frame-pointer. This flag affects the stack offset where the arguments are to be found.
  • The optimization flags: -Ox; each of these flags aggregates a number of optimizations. These flags did not affected the stack, and, quite amazingly, arguments were always passed to functions through the stack, regardless of the optimization level. One would have expected that some arguments would pe passed through registers - in which case getting these arguments would have proven to be difficult to impossible. However, these flags did complicated recovering the return code. However, on some architectures, these flags will suck in the -fomit-frame-pointer optimization.
  • In any case, be wary: other flags you use to compile your application may reserve surprises.
Function arguments

In my tests with the compilers, all arguments were invariably passed through the stack. Hence this is trivial business, affected to a small extent by the -fomit-frame-pointer flag - this flag will change the offset at which arguments start.

How many arguments a function has, how many arguments are on the stack? One way to infer somehow the number of arguments is based on its signature (for C++, beware of the 'this' hidden argument) and this is the technique used in __cyg_profile_func_enter().

Once we know the offset where arguments start on the stack and how many of them there are, we just walk the stack to retrieve their values:

    char *args(char *buf, int len, int nargs, int *frame)
        int i; 
        int offset;

        memset(buf, 0, len); 

        snprintf(buf, len, "("); 
        offset = 1; 
        for (i=0; i < nargs && offset < len; i++) {
            offset += snprintf(buf+offset, len-offset, "%d%s", 
                             i==nargs-1 ? " ...)" : ", ");

        return buf; 
Function return values

Obtaining the return value proved to be possible only when using the -O0 flag.

Let's look what happens when this method

    class B {
        virtual int m1(int i, int j) {printf("B::m1()\n"); f1(i); return 20;}

is instrumented with -O0:

     080496a2 <_ZN1B2m1Eii>:
     80496a2:    55                       push   %ebp
     80496a3:    89 e5                    mov    %esp,%ebp
     80496a5:    53                       push   %ebx
     80496a6:    83 ec 24                 sub    $0x24,%esp
     80496a9:    8b 45 04                 mov    0x4(%ebp),%eax
     80496ac:    89 44 24 04              mov    %eax,0x4(%esp)
     80496b0:    c7 04 24 a2 96 04 08     movl   $0x80496a2,(%esp)
     80496b7:    e8 b0 f4 ff ff           call   8048b6c <__cyg_profile_func_enter@plt>
     80496bc:    c7 04 24 35 9c 04 08     movl   $0x8049c35,(%esp)
     80496c3:    e8 b4 f4 ff ff           call   8048b7c <puts@plt>
     80496c8:    8b 45 0c                 mov    0xc(%ebp),%eax
     80496cb:    89 04 24                 mov    %eax,(%esp)
     80496ce:    e8 9d f8 ff ff           call   8048f70 <_Z2f1i>

==>  80496d3:    bb 14 00 00 00           mov    $0x14,%ebx
     80496d8:    8b 45 04                 mov    0x4(%ebp),%eax
     80496db:    89 44 24 04              mov    %eax,0x4(%esp)
     80496df:    c7 04 24 a2 96 04 08     movl   $0x80496a2,(%esp)
==>  80496e6:    e8 81 f5 ff ff           call   8048c6c <__cyg_profile_func_exit@plt>
     80496eb:    89 5d f8                 mov    %ebx,0xfffffff8(%ebp)
==>  80496ee:    eb 27                    jmp    8049717 <_ZN1B2m1Eii+0x75>
     80496f0:    89 45 f4                 mov    %eax,0xfffffff4(%ebp)
     80496f3:    8b 5d f4                 mov    0xfffffff4(%ebp),%ebx
     80496f6:    8b 45 04                 mov    0x4(%ebp),%eax
     80496f9:    89 44 24 04              mov    %eax,0x4(%esp)
     80496fd:    c7 04 24 a2 96 04 08     movl   $0x80496a2,(%esp)
     8049704:    e8 63 f5 ff ff           call   8048c6c <__cyg_profile_func_exit@plt>
     8049709:    89 5d f4                 mov    %ebx,0xfffffff4(%ebp)
     804970c:    8b 45 f4                 mov    0xfffffff4(%ebp),%eax
     804970f:    89 04 24                 mov    %eax,(%esp)
     8049712:    e8 15 f5 ff ff           call   8048c2c <_Unwind_Resume@plt>

==>  8049717:    8b 45 f8                 mov    0xfffffff8(%ebp),%eax
     804971a:    83 c4 24                 add    $0x24,%esp
     804971d:    5b                       pop    %ebx
     804971e:    5d                       pop    %ebp
     804971f:    c3                       ret

Note how the return code is moved into the ebx register - a bit unexpected, since, traditionally, the eax register is used for return codes - and then the instrumentation function is called. Good to retrieve the return value but to avoid that the ebx register gets clobbered in the instrumentation function, we save it upon entering the function and we restore it upon exit.

When the compilation is done with some degree of optimization (-O1...3; shown here is -O2), the code changes:

    080498c0 <_ZN1B2m1Eii>:
     80498c0:    55                       push   %ebp
     80498c1:    89 e5                    mov    %esp,%ebp
     80498c3:    53                       push   %ebx
     80498c4:    83 ec 14                 sub    $0x14,%esp
     80498c7:    8b 45 04                 mov    0x4(%ebp),%eax
     80498ca:    c7 04 24 c0 98 04 08     movl   $0x80498c0,(%esp)
     80498d1:    89 44 24 04              mov    %eax,0x4(%esp)
     80498d5:    e8 12 f4 ff ff           call   8048cec <__cyg_profile_func_enter@plt>
     80498da:    c7 04 24 2d 9c 04 08     movl   $0x8049c2d,(%esp)
     80498e1:    e8 16 f4 ff ff           call   8048cfc <puts@plt>

     80498e6:    8b 45 0c                 mov    0xc(%ebp),%eax
     80498e9:    89 04 24                 mov    %eax,(%esp)
     80498ec:    e8 af f7 ff ff           call   80490a0 <_Z2f1i>
     80498f1:    8b 45 04                 mov    0x4(%ebp),%eax
     80498f4:    c7 04 24 c0 98 04 08     movl   $0x80498c0,(%esp)
     80498fb:    89 44 24 04              mov    %eax,0x4(%esp)
==>  80498ff:    e8 88 f3 ff ff           call   8048c8c <__cyg_profile_func_exit@plt>
     8049904:    83 c4 14                 add    $0x14,%esp
==>  8049907:    b8 14 00 00 00           mov    $0x14,%eax
     804990c:    5b                       pop    %ebx
     804990d:    5d                       pop    %ebp
==>  804990e:    c3                       ret    

     804990f:    89 c3                    mov    %eax,%ebx
     8049911:    8b 45 04                 mov    0x4(%ebp),%eax
     8049914:    c7 04 24 c0 98 04 08     movl   $0x80498c0,(%esp)
     804991b:    89 44 24 04              mov    %eax,0x4(%esp)
     804991f:    e8 68 f3 ff ff           call   8048c8c <__cyg_profile_func_exit@plt>
     8049924:    89 1c 24                 mov    %ebx,(%esp)
     8049927:    e8 f0 f3 ff ff           call   8048d1c <_Unwind_Resume@plt>
     804992c:    90                       nop    
     804992d:    90                       nop    
     804992e:    90                       nop    
     804992f:    90                       nop

Note how the instrumentation function gets called first and only then the eax register is set with the return value. Thus, if we absolutely want the return code, we are forced to compile with -O0.

Sample output

Finally, below are the results. At at shell prompt type:

$ export CTRACE_PROGRAM=./cpptraced
$ LD_PRELOAD=./libctrace.so ./cpptraced
T0xb7c0f6c0: 0x8048d34 main (0 ...) [from ]
./cpptraced: main(argc=1)
T0xb7c0ebb0: 0x80492d8 thread1(void*) (1 ...) [from ]
T0xb7c0ebb0: 0x80498b2 D (134605416 ...) [from cpptraced.cpp:91]
T0xb7c0ebb0: 0x8049630 B (134605416 ...) [from cpptraced.cpp:66]
T0xb7c0ebb0: 0x8049630 B => -1209622540 [from ]
T0xb7c0ebb0: 0x80498b2 D => -1209622540 [from ]
Hello World! It's me, thread #1!
./cpptraced: done.
T0xb7c0f6c0: 0x8048d34 main => -1212090144 [from ]
T0xb740dbb0: 0x8049000 thread2(void*) (2 ...) [from ]
T0xb740dbb0: 0x80498b2 D (134605432 ...) [from cpptraced.cpp:137]
T0xb740dbb0: 0x8049630 B (134605432 ...) [from cpptraced.cpp:66]
T0xb740dbb0: 0x8049630 B => -1209622540 [from ]
T0xb740dbb0: 0x80498b2 D => -1209622540 [from ]
Hello World! It's me, thread #2!
T0xb6c0cbb0: 0x8049166 thread3(void*) (3 ...) [from ]
T0xb6c0cbb0: 0x80498b2 D (134613288 ...) [from cpptraced.cpp:157]
T0xb6c0cbb0: 0x8049630 B (134613288 ...) [from cpptraced.cpp:66]
T0xb6c0cbb0: 0x8049630 B => -1209622540 [from ]
T0xb6c0cbb0: 0x80498b2 D => -1209622540 [from ]
Hello World! It's me, thread #3!
T0xb7c0ebb0: 0x80490dc wrap_strerror_r (134525680 ...) [from cpptraced.cpp:105]
T0xb7c0ebb0: 0x80490dc wrap_strerror_r => -1210887643 [from ]
T#1+M2 (Success)
T0xb740dbb0: 0x80495a0 D::m1(int, int) (134605432, 3, 4 ...) [from cpptraced.cpp:141]
T0xb740dbb0: 0x8049522 B::m2(int) (134605432, 14 ...) [from cpptraced.cpp:69]
T0xb740dbb0: 0x8048f70 f1 (14 ...) [from cpptraced.cpp:55]
f1 14
T0xb740dbb0: 0x8048ee0 f2(int) (74 ...) [from cpptraced.cpp:44]
f2 74
T0xb740dbb0: 0x8048e5e f3 (144 ...) [from cpptraced.cpp:36]
f3 144
T0xb740dbb0: 0x8048e5e f3 => 80 [from ]
T0xb740dbb0: 0x8048ee0 f2(int) => 70 [from ]
T0xb740dbb0: 0x8048f70 f1 => 60 [from ]
T0xb740dbb0: 0x8049522 B::m2(int) => 21 [from ]
T0xb740dbb0: 0x80495a0 D::m1(int, int) => 30 [from ]

Note how libbfd fails to resolve some addresses when the function gets inlined.


Memory issues

Linux Applications Debugging Techniques/Memory issues


What to look for

Memory can be allocated through many API calls:

  1. malloc()
  2. calloc()
  3. realloc()
  4. memalign()
  5. posix_memalign()
  6. valloc()
  7. mmap()
  8. brk() / sbrk()

To return memory to the OS:

  1. free()
  2. munmap()


Valgrind should be the first stop for any memory related issue. However:

  1. it slows down the program by at least one order of magnitude. In particular server C++ programs can be slowed down 15-20 times.
  2. from experience, some versions might have difficulties tracking mmap() allocated memory.
  3. on amd64, the vex disassembler is likely to fail (v3.7) rather sooner than later so valgrind is of no use for any medium or intensive usage.
  4. you need to write suppressions to filter down the issues reported.

If these are real drawbacks, lighter solutions are available.

LD_LIBRARY_PATH=/path/to/valgrind/libs:$LD_LIBRARY_PATH /path/to/valgrind 
    -v \
    --error-limit=no \
    --num-callers=40 \
    --fullpath-after= \
    --track-origins=yes \
    --log-file=/path/to/valgrind.log \
    --leak-check=full \
    --show-reachable=yes \
    --vex-iropt-precise-memory-exns=yes \
    /path/to/program program-args


Note: Mudflap has been removed from GCC 4.9 and later [1]


Memleax debugs memory leak of a running process by attaching it, without recompiling program or restarting target process!!!

It's very convenient and suitable for production environment.

It works on GNU/Linux-x86_64 and FreeBSD-amd64.

gcc/clang flags

See https://lemire.me/blog/2016/04/20/no-more-leaks-with-sanitize-flags-in-gcc-and-clang/


See https://github.com/iovisor/bcc/blob/master/tools/memleak.py in https://github.com/iovisor/bcc.


Libmemleak finds memory leaks that cause a process to slowly increase the amount of memory it uses, also without the need to recompile the program as it can be LD_PRELOAD-ed when starting the program under test. Unlike valgrind it hardly slows down the process under test. Leaks are reported on a per backtrace basis. This is sometimes very important because often the caller, deeper into the backtrace, is responsible for the leak (not freeing it) while the actual place where the allocation happens wouldn't tell you anything.

It's been tested on GNU/Linux-x86_64.

Sample report
hello: Now: 287;        Backtraces: 77;         allocations: 650036;    total memory: 83,709,180 bytes.
backtrace 50 (value_n: 104636.00); [ 178, 238>(  60): 25957 allocations (1375222 total,  1.9%), size 3311982; 432.62 allocations/s, 55199 bytes/s
backtrace 50 (value_n: 104636.00); [  55, 178>( 123): 52722 allocations (2793918 total,  1.9%), size 6734135; 428.63 allocations/s, 54749 bytes/s
backtrace 49 (value_n: 58296.00); [ 178, 238>(  60): 14520 allocations (1382814 total,  1.1%), size 1860716; 242.00 allocations/s, 31011 bytes/s
backtrace 49 (value_n: 58296.00); [  55, 178>( 123): 29256 allocations (2794155 total,  1.0%), size 3744938; 237.85 allocations/s, 30446 bytes/s

libmemleak> dump 49
 #0  00007f84b862d33b  in malloc at src/memleak.c:1008
 #1  00000000004014da  in do_work(int)
 #2  000000000040101c  in thread_entry0(void*)
 #3  00007f84b7e7070a  in start_thread
 #4  00007f84b7b9f82d  in ?? at /build/glibc-Qz8a69/glibc-2.23/misc/../sysdeps/unix/sysv/linux/x86_64/clone.S:111
libmemleak> dump 50
 #0  00007f84b862d33b  in malloc at src/memleak.c:1008
 #1  00000000004014da  in do_work(int)
 #2  0000000000401035  in thread_entry1(void*)
 #3  00007f84b7e7070a  in start_thread
 #4  00007f84b7b9f82d  in ?? at /build/glibc-Qz8a69/glibc-2.23/misc/../sysdeps/unix/sysv/linux/x86_64/clone.S:111


The GNU C library comes with a built-in functionality to help detecting memory issues: mtrace(). One of the shortcomings: it does not log the call stacks of the memory allocations it tracks. We can build an interposition library to augment mtrace().

The basics

The malloc implementation in the GNU C library provides a simple but powerful way to detect memory leaks and obtain some information to find the location where the leaks occurs, and this, with rather minimal speed penalties for the program.

Getting started is as simple as it can be:

  • #include mcheck.h in your code.
  • Call mtrace() to install hooks for malloc(), realloc(), free() and memalign(). From this point on, all memory manipulations by these functions will be tracked. Note there are other untracked ways to allocate memory.
  • Call muntrace() to uninstall the tracking handlers.
  • Recompile.
      #include <mcheck.h>
21    mtrace();
25    std::string* pstr = new std::string("leak");
27    char *leak = (char*)malloc(1024);
32    muntrace();

Under the hood, mtrace() installs the four hooks mentioned above. The information collected through the hooks is written to a log file.

Note: there are other ways to allocate memory, notably mmap(). These allocations will not be reported, unfortunately.


  • Set the MALLOC_TRACE environment variable with the memory log name.
  • Run the program.
  • Run the memory log through mtrace.
$ MALLOC_TRACE=logs/mtrace.plain.log  ./dleaker
$ mtrace  dleaker  logs/mtrace.plain.log  >  logs/mtrace.plain.leaks.log
$ cat logs/mtrace.plain.leaks.log

Memory not freed:
   Address     Size     Caller
0x081e2390      0x4  at 0x400fa727
0x081e23a0     0x11  at 0x400fa727
0x081e23b8    0x400  at /home/amelinte/projects/articole/memtrace/memtrace.v3/main.cpp:27

One of the leaks (the malloc() call) was precisely traced to the exact file and line number. However, the other leaks at line 25, while detected, we do not know where they occur. The two memory allocations for the std::string are buried deep inside the C++ library. We would need the stack trace for these two leaks to pinpoint the place in our code.

We can use GDB (or the trace_call macro) to get the allocations' stacks:

$ gdb ./dleaker
(gdb) set env MALLOC_TRACE=./logs/gdb.mtrace.log

(gdb) b __libc_malloc
Make breakpoint pending on future shared library load? (y or [n]) 
Breakpoint 1 (__libc_malloc) pending.

(gdb) run
Starting program: /home/amelinte/projects/articole/memtrace/memtrace.v3/dleaker 
Breakpoint 2 at 0xb7cf28d6
Pending breakpoint "__libc_malloc" resolved

Breakpoint 2, 0xb7cf28d6 in malloc () from /lib/i686/cmov/libc.so.6
(gdb) command
Type commands for when breakpoint 2 is hit, one per line.
End with a line saying just "end".
(gdb) c


Breakpoint 2, 0xb7cf28d6 in malloc () from /lib/i686/cmov/libc.so.6
#0  0xb7cf28d6 in malloc () from /lib/i686/cmov/libc.so.6
#1  0xb7ebb727 in operator new () from /usr/lib/libstdc++.so.6
#2  0x08048a14 in main () at main.cpp:25             <== new std::string("leak");
Breakpoint 2, 0xb7cf28d6 in malloc () from /lib/i686/cmov/libc.so.6
#0  0xb7cf28d6 in malloc () from /lib/i686/cmov/libc.so.6
#1  0xb7ebb727 in operator new () from /usr/lib/libstdc++.so.6   <== mangled: _Znwj
#2  0xb7e95c01 in std::string::_Rep::_S_create () from /usr/lib/libstdc++.so.6
#3  0xb7e96f05 in ?? () from /usr/lib/libstdc++.so.6
#4  0xb7e970b7 in std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string () from /usr/lib/libstdc++.so.6
#5  0x08048a58 in main () at main.cpp:25           <== new std::string("leak");

Breakpoint 2, 0xb7cf28d6 in malloc () from /lib/i686/cmov/libc.so.6
#0  0xb7cf28d6 in malloc () from /lib/i686/cmov/libc.so.6

#1  0x08048a75 in main () at main.cpp:27            <== malloc(leak);
A couple of improvements

It would be good to have mtrace() itself dump the allocation stack and dispense with gdb. The modified mtrace() would have to supplement the information with:

  • The stack trace for each allocation.
  • Demangled function names.
  • File name and line number.

Additionally, we can put the code in a library, to free the program from being instrumented with mtrace(). In this case, all we have to do is interpose the library when we want to trace memory allocations (and pay the performance price).

Note: getting all this information at runtime, particularly in a human-readable form will have a performance impact on the program, unlike the plain vanilla mtrace() supplied with glibc.

The stack trace

A good start would be to use another API function: backtrace_symbols_fd(). This would print the stack directly to the log file. Perfect for a C program but C++ symbols are mangled:

@ /usr/lib/libstdc++.so.6:(_Znwj+27)[0xb7f1f727] + 0x9d3f3b0 0x4
**[ Stack: 8
/usr/lib/libstdc++.so.6(_Znwj+0x27)[0xb7f1f727]           <=== here
**] Stack

For C++ we would have to get the stack (backtrace_symbols()), resolve each address (dladdr()) and demangle each symbol name (abi::__cxa_demangle()).

  • Memory allocation is one of these basic operation everything else builds on. One needs to allocate memory to load libraries and executables; needs to allocate memory to track memory allocations; and we hook onto it very early in the life of a process: the first pre-loaded library is the memory tracking library. Thus, any API call we make inside this interposition library can reserve surprises, especially in multi-threaded environments.
  • The API functions we use to trace the stack can allocate memory. These allocations are also going through the hooks we installed. As we trace the new allocation, the hooks are activated again and another allocation is made as we trace this new allocation. We will run out of stack in this infinite loop. We break out of this pitfall by using a per-thread flag.
  • The API functions we use to trace the stack can deadlock. Suppose we would use a lock while in our trace. We lock the trace lock and we call dladdr() which in turn tries to lock a dynamic linker internal lock. If on another thread dlopen() is called while we trace, dlopen() locks the same linker lock, then allocates memory: this will trigger the memory hooks and we now have the dlopen() thread wait on the trace lock with the linker lock taken. Deadlock.
  • On some platforms (gcc 4.7.2 amd64) TLS calls would trip the memalign hook. This could result in an infinite recursion if the memalign hook in its turn, accesses a TLS variable.
What we got

Let's try again with our new library:

$ MALLOC_TRACE=logs/mtrace.stack.log LD_PRELOAD=./libmtrace.so ./dleaker
$ mtrace dleaker logs/mtrace.stack.log > logs/mtrace.stack.leaks.log
$ cat logs/mtrace.stack.leaks.log

Memory not freed:
   Address     Size     Caller
0x08bf89b0      0x4  at 0x400ff727    <=== here
0x08bf89e8     0x11  at 0x400ff727
0x08bf8a00    0x400  at /home/amelinte/projects/articole/memtrace/memtrace.v3/main.cpp:27

Apparently, not much of an improvement: the summary still does not get us back to line 25 in main.cpp. However, if we search for address 8bf89b0 in the trace log, we find this:

@ /usr/lib/libstdc++.so.6:(_Znwj+27)[0x400ff727] + 0x8bf89b0 0x4     <=== here
**[ Stack: 8
[0x40022251]  (./libmtrace.so+40022251)
[0x40022b43]  (./libmtrace.so+40022b43)
[0x400231e8]  (./libmtrace.so+400231e8)
[0x401cf905] __libc_malloc (/lib/i686/cmov/libc.so.6+35)
[0x400ff727] operator new(unsigned int) (/usr/lib/libstdc++.so.6+27) <== was: _Znwj
[0x80489cf] __gxx_personality_v0 (./dleaker+27f)
[0x40178450] __libc_start_main (/lib/i686/cmov/libc.so.6+e0)         <=== here
[0x8048791] __gxx_personality_v0 (./dleaker+41)
**] Stack

This is good, but having file and line information would be better.

File and line

Here we have a few possibilities:

  • Run the address (e.g. 0x40178450 above) through the addr2line tool. If the address is in a shared object that the program loaded, it might not resolve properly.
  • If we have a core dump of the program, we can ask gdb to resolve the address. Or we can attach to the running program and resolve the address.
  • Use the API described here. The downside is that it takes a quite heavy toll on the performance of the program.


Building on libmtrace, we can go one step further and have an interposition library track the memory allocations made by the program. The library generates a report on demand, much like Valgrind does.

Libmemleak is significantly faster than valgrind but also has limited functionality (only leak detection).


libmemtrace has two drawbacks:

  • The log file will quickly reach gigs
  • You are left grepping the log to figure out what leaks when

A better solution would be to have an interposition library to collect memory operations information and to generate a report on-demand.

For mmap()/munmap() we have no choice but hook these directly. Thus, an call from within the application would first hit the hooks in libmemleak, then go to libc. For malloc()realloc()/memalign()/free() we have two options:

  • Use mtrace()/muntrace() as before, to install hooks that will be called from within libc. Thus, a malloc() call would first go through libc which will then call the hooks in libmemtrace. This leave us at the mercy of libc.
  • The second solution is to hook these like m(un)map.

The second solution also frees mtrace()/muntrace() for on-demand report generation:

  • A first call to mtrace() will kick in data collection.
  • Subsequent calls to mtrace() will generate reports.
  • muntrace() will stop data collection and will generate a final report.
  • MALLOC_TRACE is not needed.

The application can then sprinkle its code with mtrace() calls at strategic places to avoid reporting much noise. These calls will do nothing in a normal operation as long as MALLOC_TRACE is not set. Or, the application can be completely ignorant of the ongoing data collection (no mtrace() calls within the application code) and libmemleak can start collecting as early as being loaded and generate one report upon being unloaded.

To control the libmemleak functionality, an environment variable - MEMLEAK_CONFIG - has to be set before loading the library:

export MEMLEAK_CONFIG=mtraceinit
  • mtraceinit will instruct the library to start collecting data upon being loaded. Default is off and the application has to be instrumented with m(un)trace calls.

Thus, all the hooks have to do is to call into the reporting:

extern "C" void *__libc_malloc(size_t size); 
extern "C" void *malloc(size_t size)
    void *ptr = __libc_malloc(size);
    if ( _capture_on) {
            libmemleak::alloc(ptr, size);
    return ptr;

extern "C" void __libc_free(void *ptr);
extern "C" void free(void *ptr)
    if (_capture_on) {
        libmemleak::free(ptr, 0); // Call to reporting
    else {
        serror(ptr, "Untraced free", __FILE__, __LINE__);

extern "C" void mtrace ()
    // Make sure not to track memory when globals get destructed
    static std::atomic<bool> _atexit(false);
    if (!_atexit.load(std::memory_order_acquire)) {
        int ret = atexit(muntrace);
        assert(0 == ret);
        _atexit.store(true, std::memory_order_release);

    if (!_capture_on) {
        _capture_on = true; 
    else {
Sample report
// Leaks since previous report

// Ordered by Num Total Bytes
// Stack Key,  Num Total Bytes,  Num Allocs,  Num Delta Bytes
   5163ae4c,   1514697,          5000,        42420

11539977 total bytes, since previous report: 42420 bytes
Max tracked: stacks=6, allocations=25011

// All known allocations

// Key   Total Bytes    Allocations
4945512: 84983 bytes in 5000 allocations 
bbc54f2: 1029798 bytes in 10000 allocations 

bbc54f2: 1029798 bytes in 10000 allocations 
[0x4005286a] lpt::stack::detail::call_stack<lpt::stack::bfd::source_resolver>::call_stack() (binaries/lib/libmemleak_mtrace_hooks.so+0x66) in crtstuff.c:0
[0x4005238d] _pstack::_pstack() (binaries/lib/libmemleak_mtrace_hooks.so+0x4b) in crtstuff.c:0
[0x4004f8dd] libmemleak::alloc(void*, unsigned long long) (binaries/lib/libmemleak_mtrace_hooks.so+0x75) in crtstuff.c:0
[0x4004ee7c] ?? (binaries/lib/libmemleak_mtrace_hooks.so+0x4004ee7c) in crtstuff.c:0
[0x402f5905] ?? (/lib/i686/cmov/libc.so.6+0x35) in ??:0
[0x401a02b7] operator new(unsigned int) (/opt/lpt/gcc-4.7.0-bin/lib/libstdc++.so.6+0x27) in crtstuff.c:0
[0x8048e3b] ?? (binaries/bin/1001leakseach+0x323) in /home/amelinte/projects/articole/lpt/lpt/tests/1001leakseach.cpp:68
[0x8048e48] ?? (binaries/bin/1001leakseach+0x330) in /home/amelinte/projects/articole/lpt/lpt/tests/1001leakseach.cpp:74
[0x8048e61] ?? (binaries/bin/1001leakseach+0x349) in /home/amelinte/projects/articole/lpt/lpt/tests/1001leakseach.cpp:82
[0x8048eab] ?? (binaries/bin/1001leakseach+0x393) in /home/amelinte/projects/articole/lpt/lpt/tests/1001leakseach.cpp:90
[0x401404fb] ?? (/lib/i686/cmov/libpthread.so.0+0x401404fb) in ??:0
[0x4035e60e] ?? (/lib/i686/cmov/libc.so.6+0x5e) in ??:0

// Crosstalk: leaked bytes per stack frame
1029798 bytes: [0x8048e3b] ?? (binaries/bin/1001leakseach+0x323) in /home/amelinte/projects/articole/lpt/lpt/tests/1001leakseach.cpp:68

// Mem Address, Stack Key, Bytes
   0x8ce7988,   bbc54f2,   4

This report took 44 ms to generate.


The mallinfo() API is rumored to be deprecated. But, if available, it is very useful:

#include <malloc.h>

namespace process {

class memory

    memory() : _meminfo(::mallinfo()) {}

    int total() const
        return _meminfo.hblkhd + _meminfo.uordblks;

    bool operator==(memory const& other) const
        return total() == other.total();

    bool operator!=(memory const& other) const
        return total() != other.total();

    bool operator<(memory const& other) const
        return total() < other.total();

    bool operator<=(memory const& other) const
        return total() <= other.total();

    bool operator>(memory const& other) const
        return total() > other.total();

    bool operator>=(memory const& other) const
        return total() >= other.total();


    struct mallinfo _meminfo;

} //process
#include <iostream>
#include <string>
#include <cassert>

int main()

    process::memory first;

        void* p = ::malloc(1025);
        process::memory second;
        std::cout << "Mem diff: " << second.total() - first.total() << std::endl;
        assert(second > first);

        process::memory third;
        std::cout << "Mem diff: " << third.total() - first.total() << std::endl;
        assert(third == first);
        std::string s("abc");
        process::memory second;
        std::cout << "Mem diff: " << second.total() - first.total() << std::endl;
        assert(second > first);

    process::memory fourth;
    assert(first == fourth);

    return 0;


Coarse grained information can be obtained from /proc:

# Based on:
# http://stackoverflow.com/questions/131303/linux-how-to-measure-actual-memory-usage-of-an-application-or-process
# Returns total memory used by process $1 in kb.
# See /proc/PID/smaps; This file is only present if the CONFIG_MMU 
# kernel configuration option is enabled


for line in $(</proc/$1/smaps)
   [[ $line =~ ^Private_Clean:\s+(\S+) ]] && ((pkb += ${.sh.match[1]}))
   [[ $line =~ ^Private_Dirty:\s+(\S+) ]] && ((pkb += ${.sh.match[1]}))
   [[ $line =~ ^Shared_Clean:\s+(\S+) ]]  && ((skb += ${.sh.match[1]}))
   [[ $line =~ ^Shared_Dirty:\s+(\S+) ]]  && ((skb += ${.sh.match[1]}))
   [[ $line =~ ^Size:\s+(\S+) ]]          && ((szkb += ${.sh.match[1]}))
   [[ $line =~ ^Pss:\s+(\S+) ]]           && ((psskb += ${.sh.match[1]}))

((tkb += pkb))
((tkb += skb))
#((tkb += psskb))

echo "Total private:        $pkb kb"
echo "Total shared:         $skb kb"
echo "Total proc prop:      $psskb kb Pss"
echo "Priv + shared:        $tkb kb"
echo "Size:                 $szkb kb"

pmap -d $1 | tail -n 1


Various tools

Heap corruption

Electric Fence

Electric Fence is still the reference for dealing with heap corruption, even if not maintined for a while. RedHat ships a version that can be used as an interposition library.

Drawback: might not work with code that uses mmap() to allocate memory.


Duma is a fork of Electric Fence.

glibc builtin

man (3) malloc: Recent versions of Linux libc (later than 5.4.23) and GNU libc (2.x) include a malloc implementation which is tunable via environment variables. When MALLOC_CHECK_ is set, a special (less efficient) implementation is used which is designed to be tolerant against simple errors, such as double calls of free() with the same argument, or overruns of a single byte (off-by-one bugs). Not all such errors can be protected against, however, and memory leaks can result. If MALLOC_CHECK_ is set to 0, any detected heap corruption is silently ignored and an error message is not generated; if set to 1, the error message is printed on stderr, but the program is not aborted; if set to 2, abort() is called immediately, but the error message is not generated; if set to 3, the error message is printed on stderr and program is aborted. This can be useful because otherwise a crash may happen much later, and the true cause for the problem is then very hard to track down.

C++ stdlib builtin

Compile with -D_GLIBCXX_DEBUG - it turns on debugging checks in the C++ standard library.


A fast memory error detector via -fsanitize=address , with a recent gcc (since 4.8) or clang (since 3.1). Memory access instructions will be instrumented to detect heap-, stack-, and global-buffer overflow as well as use-after-free bugs. To get nicer stacktraces, use -fno-omit-frame-pointer. The AddressSanitizer is available on IA-32/x86-64/x32/PowerPC/PowerPC64 GNU/Linux and on x86-64 Darwin.


Starting with gcc 4.9 and later, you can use the ubsan sanitizer for bounds checking.

gdb heisenbugs

Memory-related bugs that don't happen when using GDB but do without GDB are often caused by GDB disabling ASLR. This is usually useful for reproducibility of debugging sessions, but may indeed make address space layout too predictable and fail to reproduce a bug. Try using the GDB command "set disable-randomization off" (without quotes) to undo this behavior and thus start your program in a normal, randomized, environment.


Stack corruption

Stack corruption is rather hard to diagnose. Luckily, gcc 4.x can instrument the code to check for stack corruption:

  • -fstack-protector Add stack protection to functions that have “alloca” or have a (signed or unsigned) char array with size > 8 (SSP_BUFFER_SIZE)
  • -fstack-protector-strong To more functions, see below
  • -fstack-protector-all To ALL functions

gcc will add guard variables and code to check for buffer overflows upon exiting a function. A quick example:

/* Compile with: gcc -ggdb -fstack-protector-all stacktest.c */

#include <stdio.h>
#include <string.h>
void bar(char* str) 
    char buf[4];
    strcpy(buf, str);
void foo() 
    printf("It survived!");
int main(void) 
    bar("Longer than 4.");
    return 0;

When run, the program will dump core:

$ ./a.out 
*** stack smashing detected ***: ./a.out terminated
Aborted (core dumped)
Core was generated by `./a.out'.
Program terminated with signal 6, Aborted.
#0  0x0000003684030265 in raise () from /lib64/libc.so.6
(gdb) bt full
#0  0x0000003684030265 in raise () from /lib64/libc.so.6
No symbol table info available.
#1  0x0000003684031d10 in abort () from /lib64/libc.so.6
No symbol table info available.
#2  0x000000368406a84b in __libc_message () from /lib64/libc.so.6
No symbol table info available.
#3  0x00000036840e8ebf in __stack_chk_fail () from /lib64/libc.so.6
No symbol table info available.
#4  0x0000000000400584 in bar (str=0x400715 "Longer than 4.") at stacktest.c:10
        buf =           "Long"
#5  0x00000000004005e3 in main () at stacktest.c:19
No locals.


Added by google to gcc.

Benefit - gain big performance while sacrificing little security (for scenarios using -fstack-protector-all)

Background - some times stack-protector is too-simple while stack-protector-all over-kills, for example, to build one of our core systems, we forcibly add "-fstack-protector-all" to all compile commands, which brings big performance penalty (due to extra stack guard/check insns on function prologue and epilogue) on both atom and arm. To use "-fstack-protector" is just regarded as not secure enough (only "protects" <2% functions) by the system secure team. "-fstack-protector-strong" hits the balance between "-fstack-protector" and "-fstack-protector-all".

Adds the check to a function:

  • if any of its local variable’s address is taken, as part of the RHS of an assignment
  • or if any of its local variable’s address is taken as part of a function argument.
  • or if it has an array, regardless of array type or length
  • or if it has a struct/union which contains an array, regardless of array type or length.
  • or if function has register local variables

See http://gcc.gnu.org/ml/gcc-patches/2012-06/msg00974.html

Further reading

Starting with gcc 4.9 and later, you can use the ubsan sanitizer for bounds checking.



Searching for a deadlock means reconstructing the graph of dependencies between threads and resources (mutexes, semaphores, condition variables, etc.) - who owns what and who wants to acquire what. A typical deadlock would look like a loop in that graph. The task is tedious, as some of the parameters we are looking for have been optimized by the compiler into registers.

Below is an analysis of an x86_64 deadlock. On this platform, register r8 is the one containing the first argument: the address of the mutex:

(gdb) thread apply all bt
Thread 4 (Thread 0x419bc940 (LWP 12275)):
#0  0x0000003684c0d4c4 in __lll_lock_wait () from /lib64/libpthread.so.0
#1  0x0000003684c08e1a in _L_lock_1034 () from /lib64/libpthread.so.0
#2  0x0000003684c08cdc in pthread_mutex_lock () from /lib64/libpthread.so.0
#3  0x0000000000400a50 in thread1 (threadid=0x1) at deadlock.c:66
#4  0x0000003684c0673d in start_thread () from /lib64/libpthread.so.0
#5  0x00000036840d3d1d in clone () from /lib64/libc.so.6

Thread 3 (Thread 0x421bd940 (LWP 12276)):
#0  0x0000003684c0d4c4 in __lll_lock_wait () from /lib64/libpthread.so.0
#1  0x0000003684c08e1a in _L_lock_1034 () from /lib64/libpthread.so.0
#2  0x0000003684c08cdc in pthread_mutex_lock () from /lib64/libpthread.so.0
#3  0x0000000000400c07 in thread2 (threadid=0x2) at deadlock.c:111
#4  0x0000003684c0673d in start_thread () from /lib64/libpthread.so.0
#5  0x00000036840d3d1d in clone () from /lib64/libc.so.6

(gdb) thread 4
[Switching to thread 4 (Thread 0x419bc940 (LWP 12275))]#2  0x0000003684c08cdc in pthread_mutex_lock ()
   from /lib64/libpthread.so.0

(gdb) frame 2
#2  0x0000003684c08cdc in pthread_mutex_lock () from /lib64/libpthread.so.0

(gdb) info reg
r8             0x6015a0	6296992

(gdb) p *(pthread_mutex_t*)0x6015a0
$3 = {
  __data = {
    __lock = 2, 
    __count = 0, 
    __owner = 12276,   <== LWP 12276 is Thread 3
    __nusers = 1, 
    __kind = 0,        <== non-recursive
    __spins = 0, 
    __list = {
      __prev = 0x0, 
      __next = 0x0
  __size =     "\002\000\000\000\000\000\000\000\364/\000\000\001", '\000' <repeats 26 times>, 
  __align = 2

(gdb) thread 3
[Switching to thread 3 (Thread 0x421bd940 (LWP 12276))]#0  0x0000003684c0d4c4 in __lll_lock_wait ()
   from /lib64/libpthread.so.0
(gdb) bt
#0  0x0000003684c0d4c4 in __lll_lock_wait () from /lib64/libpthread.so.0
#1  0x0000003684c08e1a in _L_lock_1034 () from /lib64/libpthread.so.0
#2  0x0000003684c08cdc in pthread_mutex_lock () from /lib64/libpthread.so.0
#3  0x0000000000400c07 in thread2 (threadid=0x2) at deadlock.c:111
#4  0x0000003684c0673d in start_thread () from /lib64/libpthread.so.0
#5  0x00000036840d3d1d in clone () from /lib64/libc.so.6

(gdb) info reg
r8             0x6015e0	6297056

(gdb) p *(pthread_mutex_t*)0x6015e0
$4 = {
  __data = {
    __lock = 2, 
    __count = 0, 
    __owner = 12275,  <=== Thread 4
    __nusers = 1, 
    __kind = 0, 
    __spins = 0, 
    __list = {
      __prev = 0x0, 
      __next = 0x0
  __size =     "\002\000\000\000\000\000\000\000\363/\000\000\001", '\000' <repeats 26 times>, 
  __align = 2

Threads 3 and 4 are deadlocking over two mutexes.

Note: If gdb is unable to find the symbol pthread_mutex_t because it has not loaded the symbol table for pthreadtypes.h, you can still print the individual members of the struct as follows:

(gdb) print *((int*)(0x6015e0))
$4 = 2
(gdb) print *((int*)(0x6015e0)+1)
$5 = 0
(gdb) print *((int*)(0x6015e0)+2)
$6 = 12275


An interposition library can be built to automate deadlock analysis. A significant number of APIs have to be interposed and even then there are cases that would go unnoticed by the library. For instance, one creative way to deadlock two threads without involving any userland locking mechanism is to have each thread join the other. Thus, interposition tools have limited diagnostic functionality.

There are a number of other tools available:

  • gdb-automatic-deadlock-detector - script adds new command 'blocked' to GDB. This command analyze all threads and display which threads are waiting for other threads. It also shows deadlocks between threads.
  • Userspace lockdep. Incipient work.
  • Locksmith. Basic.
  • Valgrind Helgrind. Does not suffer from the terrible slowdown other Valgrind tools have (in particular the memory analysis ones) but does not survive long on an amd64 platform.


pthreads has some built-in support for certain synchronization mechanisms (e.g. PTHREAD_MUTEX_ERRORCHECK mutexes).

Also, there's a POSIX mutex construction attribute for robust mutexes (a recoverable mutex that was held by a thread whose process has died): an error return code indicates the earlier owner died; and lock acquisition implies acquired responsibility for dealing with any cleanup.


On lockdep-enabled kernels, pressing Alt+SysRq+d will dump information about all the locks known to the kernel.

Race conditions

Valgrind Helgrind

  • [v 3.7] On amd64 platforms it does not survive for long because of the vex disassembler.

Valgrind Drd

  • Same.

ThreadSanitizer (tsan)

  • Usually better than Helgrind, as it does not depend on valgrind anymore.


  • C++0x/11 synchronization modeler/unit tests tool.



Resource leaks

Zombie threads

Any thread that has terminated but has not been joined or detached will leak OS resources until the process terminates. Unfortunately, neither /proc nor gdb will show you these zombie threads, at least not on some kernels.

One way to get them is with a gdb canned command:

define trace_call
    b $arg0
    bt full
document trace_call
Trace specified call with call stack to screen. Example:
    set breakpoint pending on
    set pagination off
    set logging on
    trace_call __pthread_create_2_1
Using host libthread_db library "/lib/i686/cmov/libthread_db.so.1".
(gdb) trace_call __pthread_create_2_1
Function "__pthread_create_2_1" not defined.
Breakpoint 1 (__pthread_create_2_1) pending.
(gdb) trace_call __pthread_create_2_0
Function "__pthread_create_2_0" not defined.
Breakpoint 2 (__pthread_create_2_0) pending.
(gdb) r
Starting program: /home/amelinte/projects/articole/wikibooks/debug/plock foo bar bax
[Thread debugging using libthread_db enabled]
Breakpoint 3 at 0xb7f9b746
Pending breakpoint "__pthread_create_2_1" resolved
Breakpoint 4 at 0xb7f9c395
Pending breakpoint "__pthread_create_2_0" resolved
[New Thread 0xb7e48ad0 (LWP 8635)]
[Switching to Thread 0xb7e48ad0 (LWP 8635)]

Breakpoint 3, 0xb7f9b746 in pthread_create@@GLIBC_2.1 () from /lib/i686/cmov/libpthread.so.0
#0  0xb7f9b746 in pthread_create@@GLIBC_2.1 () from /lib/i686/cmov/libpthread.so.0
No symbol table info available.
#1  0x08048a7f in main (argc=4, argv=0xbfceb714) at plock.c:97
        s = 0
        tnum = 0
        opt = -1
        num_threads = 3
        tinfo = (struct thread_info *) 0x833b008
        attr = {__size = '\0' <repeats 13 times>, "\020", '\0' <repeats 21 times>, __align = 0}
        stack_size = -1
        res = (void *) 0x0
[New Thread 0xb7e47b90 (LWP 8638)]
Thread 1: top of stack near 0xb7e473c8; argv_string=foo

Another way is to use (again) an interposition library:

 *  Hook library. Usage: 
 *    gcc -c -g -Wall -fPIC libhook.c -o libhook.o 
 *    ld -o libhook.so libhook.o -shared -ldl
 *    LD_PRELOAD=./libhook.so program arguments
 *  Copyright 2012 Aurelian Melinte. 
 *  Released under GPL 3.0 or later. 

#define _GNU_SOURCE
#include <dlfcn.h>

#include <signal.h>
#include <execinfo.h>

#include <errno.h>
#include <stdlib.h>
#include <stdio.h>  /*printf*/
#include <unistd.h>

#include <pthread.h>

#include <assert.h>

typedef int (*lp_pthread_mutex_func)(pthread_mutex_t *mutex);
typedef int (*pthread_create_func)(pthread_t *thread, 
                                   const pthread_attr_t *attr,
								   void *(*start_routine) (void *), void *arg);
static pthread_create_func  _pthread_create_hook = NULL;

static int
hook_one(pthread_create_func *fptr, const char *fname)
    char *msg = NULL;

    assert(fname != NULL);

    if (*fptr == NULL) {
        printf("dlsym : wrapping %s\n", fname);
        *fptr = dlsym(RTLD_NEXT, fname);
        printf("next_%s = %p\n", fname, *fptr);
        if ((*fptr == NULL) || ((msg = dlerror()) != NULL)) {
            printf("dlsym %s failed : %s\n", fname, msg);
            return -1;
        } else {
            printf("dlsym: wrapping %s done\n", fname);
            return 0;
    } else {
        return 0;

static void
    if (_pthread_create_hook == NULL) {
        int rc = hook_one(&_pthread_create_hook, "pthread_create"); 
        if (NULL == _pthread_create_hook || rc != 0) {
            printf("Failed to hook.\n");


pthread_create(pthread_t *thread, 
               const pthread_attr_t *attr,
               void *(*start_routine) (void *), void *arg)
#define SIZE 40
    void *buffer[SIZE] = {0};
	int nptrs = 0;

    int rc = EINVAL; 
	rc = _pthread_create_hook(thread, attr, start_routine, arg);

    printf("*** pthread_create:\n");
    nptrs = backtrace(buffer, SIZE);
    backtrace_symbols_fd(buffer, nptrs, STDOUT_FILENO);

    return rc; 

void _init()  __attribute__((constructor));
    printf("*** _init().\n");

void  _fini()  __attribute__((destructor)); 
    printf("*** _fini().\n");

The output is a bit rough but it can be refined down to file and line by replacing backtrace_symbols_fd() with appropriate code:

*** pthread_create:

File descriptors

As just about anything is a file (folders, sockets, pipes, etc.), just about anything can result in a file descriptor that needs to be closed. /proc can help:

# tree /proc/26041
|-- fd                  # Open files descriptors
|   |-- 0 -> /dev/pts/21
|   |-- 1 -> /dev/pts/21
|   |-- 2 -> /dev/pts/21
|   `-- 3 -> socket:[113497835]
|-- fdinfo
|   |-- 0
|   |-- 1
|   |-- 2
|   `-- 3

The trace_call command for gdb can help with the call stack.

If gdb is not available on the machine, an interposition library hooking open(), pipe(), socket(), etc. can be built.

Other tools that can be used:

  • lsof
  • fuser


Which process is using a port? As root:

# netstat -tlnp
Active Internet connections (only servers)
Proto Recv-Q Send-Q Local Address               Foreign Address             State       PID/Program name   
tcp        0      0     *                   LISTEN      -                   
tcp        0      0    *                   LISTEN      3438/python         
# lsof
COMMAND     PID             USER   FD      TYPE             DEVICE       SIZE       NODE NAME
init          1             root  cwd       DIR              253,0       4096          2 /
python     3438             root    4u     IPv4              11416                   TCP localhost.localdomain:2207 (LISTEN)

# lsof -i :2207
python  3438 root    4u  IPv4  11416       TCP localhost.localdomain:2207 (LISTEN)

Other tools:

  • fuser


For semaphores, shared memory and message queues.

  • ipcs
  • ipcrm
# ipcs -spt
------ Semaphore Operation/Change Times --------
semid    owner      last-op                    last-changed              
187826177 aurelian_m  Fri Feb 10 09:37:26 2012   Fri Feb 10 09:33:39 2012  
187858946 aurelian_m  Fri Feb 10 09:52:11 2012   Fri Feb 10 09:50:44 2012

DYI: an interposition resource counter

libmemleak can be easily modified to keep track of whatever resources are leaking. Hook the right API (e.g. open()/close()).

Transactional memory


  • If writing alters a cache line used in a transaction that transaction will be aborted.

The compiler

When requested to optimize (-O2 and above), the compiler is granted a "total license" to modify it: is treats undefined behavior (by the standard) as it cannot happen. Thus, the resulting code behaves differently in release-optimized mode than debug mode. Such differences in behavior cannot be found by static analyzers or code reviews.

Signed Integer Overflow

1. This is C99 undefined behavior and the compiler assumes overflow cannot occur. Thus, any checks for overflow are discarded and the following is an infinite loop in optimized code:

int i, j=0;
for (i = 1; i > 0; i += i) {

2. The following:

    (i * 2000) / 1000

is optimized into:

    i * 2
  • Use -Wstrict-overflow=N with N>=3 to signal cases where code is optimized away.
  • Use -fwrapv

If lucky enough to use gcc 4.9 and later, you can use the ubsan sanitizer:

  • compile and link the program with -fsanitize=undefined option
  • use other available flags as needed.

Unsigned Wrap Around

The compiler assumes unsigned integers do not wrap. Keep the unsigned variables in the range [INT_MIN/2, INT_MAX/2] to avoid surprises.

"Dead Code" Removed

The memset() call could be removed because the compiler deems buf unused at that point in the code and after:

void do_something(void {
    char buf[123];
    ... use buf...
    /* Clear it. But removed by gcc: */
    memset(buf, 0, sizeof(buf));
  • Use #pragma optimize() directives to force the code in.
  • Use volatile.

volatile Pitfalls

Code can be moved around:

volatile int flag = 0;
char buf[123];
void init_buf() {
    for (size_t i=0; i<123; ++i) {
        buf[i] = 0; //Do something
    flag = 1; // Can move!

could be optimized into:

volatile int flag = 0;
char buf[123];
void init_buf() {
    flag = 1; // Moved!
    for (size_t i=0; i<123; ++i) {
        //Do something
  • Use compiler intrinsic barriers to prevent the flag moving.

Loops could be optimized into one read call only:

void *ptr = address;
while ( *((volatile int*)ptr) & flag ) {}


  • Use restrict
  • Use -fno-delete-null-pointer-checks. Null pointer checks are deleted if placed after the first use of the pointer.



Aiming for and measuring performance

gprof & -pg

To profile the application with gprof:

  • Compile the code with -pg
  • Link with -pg
  • Run the application. This creates a file gmon.out in the current folder of the application.
  • At the prompt, in the folder where gmon.out lives: gprof path-to-application


The Performance Application Programming Interface (PAPI) offers the programmer access to the performance counter hardware found in most major microprocessors. With a decent C++ wrapper, measuring branch mispredictions and cache misses (and much more) is literally one line of code away.

By default, these are the events that papi::counters<Print_policy> is looking for:

static  const events_type events = {  
      {PAPI_TOT_INS, "Total instructions"} 
    , {PAPI_TOT_CYC, "Total cpu cycles"}
    , {PAPI_L1_DCM,  "L1 load  misses"}
//  , {PAPI_L1_STM,  "L1 store  missess"}
    , {PAPI_L2_DCM,  "L2 load  misses"}
//  , {PAPI_L2_STM,  "L2 store  missess"}
    , {PAPI_BR_MSP,  "Branch mispredictions"}

The counters class is parameterized with a Print_policy instructing what to do with the counters once they get out of scope.

As an example, lets look a bit at these lines:

   const int nlines = 196608;
   const int ncols  = 64;
   char ctrash[nlines][ncols];
       int x;
       papi::counters<papi::stdout_print> pc("by column"); //<== the famous one-line
       for (int c = 0; c < ncols; ++c) {
           for (int l = 0; l < nlines; ++l) {
               x = ctrash[l][c];

The code just loops over an array but in the wrong order: the innermost loop iterates on the outer index. While the result is the same whether we loop over the first index first or over the last one, theorically, to preserve cache locality, the innermost loop should iterate over the innermost index. This should make a big difference for the time it takes to iterate over the array:

       int x;
       papi::counters<papi::stdout_print> pc("By line");
       for (int l = 0; l < nlines; ++l) {
           for (int c = 0; c < ncols; ++c) {
               x = ctrash[l][c];

papi::counters is a class wrapping around PAPI functionality. It will take a snaphost of some performance counters (in our case, we are interested in cache misses and in branch mispredictions) when a counters object is instantiated and another snapshot when the object is destroyed. Then it will print out the differences.

A first measure, with non-optimized code (-O0), shows the following:

Delta by column:

 PAPI_TOT_INS (Total instructions): 188744788 (380506167-191761379)
 PAPI_TOT_CYC (Total cpu cycles): 92390347 (187804288-95413941)
 PAPI_L1_DCM (L1 load  misses): 28427 (30620-2193)                 <==
 PAPI_L2_DCM (L2 load  misses): 102 (1269-1167)
 PAPI_BR_MSP (Branch mispredictions): 176 (207651-207475)          <==

Delta By line:

 PAPI_TOT_INS (Total instructions): 190909841 (191734047-824206)
 PAPI_TOT_CYC (Total cpu cycles): 94460862 (95387664-926802)
 PAPI_L1_DCM (L1 load  misses): 403 (2046-1643)                    <==
 PAPI_L2_DCM (L2 load  misses): 21 (1081-1060)
 PAPI_BR_MSP (Branch mispredictions): 205934 (207350-1416)         <==

While the cache misses have indeed improved, branch mispredictions exploded. Not exactly a good tradeoff. Down in the pipeline of the processor, a comparison operation translates into a branch operation. Something is funny with the unoptimized code the compiler generated.

Typically, branch machine code is generated directly by if/else and ternary operators; and indirectly by virtual calls and by calls though pointers

Maybe the optimized code (-O2) is behaving better? Or maybe not:

Delta by column:

 PAPI_TOT_INS (Total instructions): 329 (229368-229039)
 PAPI_TOT_CYC (Total cpu cycles): 513 (186217-185704)
 PAPI_L1_DCM (L1 load  misses): 2 (1523-1521)
 PAPI_L2_DCM (L2 load  misses): 0 (993-993)
 PAPI_BR_MSP (Branch mispredictions): 7 (1287-1280)

Delta By line:

 PAPI_TOT_INS (Total instructions): 330 (209614-209284)
 PAPI_TOT_CYC (Total cpu cycles): 499 (173487-172988)
 PAPI_L1_DCM (L1 load  misses): 2 (1498-1496)
 PAPI_L2_DCM (L2 load  misses): 0 (992-992)
 PAPI_BR_MSP (Branch mispredictions): 7 (1225-1218)

This time the compiler optimized the loops out! It figured we do not really use the data in the array, so it got rid of. Completely!

Let's see how this code behaves:

       int x;
       papi::counters<papi::stdout_print> pc("by column");
       for (int c = 0; c < ncols; ++c) {
           for (int l = 0; l < nlines; ++l) {
               x = ctrash[l][c];
               ctrash[l][c] = x + 1;

Delta by column:

 PAPI_TOT_INS (Total instructions): 62918492 (63167552-249060)
 PAPI_TOT_CYC (Total cpu cycles): 224705473 (224904307-198834)
 PAPI_L1_DCM (L1 load  misses): 12415661 (12417203-1542)
 PAPI_L2_DCM (L2 load  misses): 9654638 (9655632-994)
 PAPI_BR_MSP (Branch mispredictions): 14217 (15558-1341)

Delta By line:

 PAPI_TOT_INS (Total instructions): 51904854 (115092642-63187788)
 PAPI_TOT_CYC (Total cpu cycles): 25914254 (250864272-224950018)
 PAPI_L1_DCM (L1 load  misses): 197104 (12614449-12417345)
 PAPI_L2_DCM (L2 load  misses): 6330 (9662090-9655760)
 PAPI_BR_MSP (Branch mispredictions): 296 (16066-15770)

Both cache misses and branch mispredictions improved by at least an order of magnitude. A run with unoptimized code will show the same order of improvement.



OProfile offers access to the same hardware counters as PAPI but without having to instrument the code:

  • It is coarser grained than PAPI - at function level.
  • Some out of the box kernels (RedHat) are not OProfile-friendly.
  • You need root access.

# A script to OProfile a program. 
# Must be run as root. 

if [ $# -ne 1 ]
  echo "Usage: `basename $0` <for-binary-image>"
  exit -1

opcontrol --stop
opcontrol --shutdown

# Out of the box RedHat kernels are OProfile repellent.
opcontrol --no-vmlinux
opcontrol --reset

# List of events for platform to be found in /usr/share/oprofile/<>/events
opcontrol --event=L2_CACHE_MISSES:1000

opcontrol --start


opcontrol --stop
opcontrol --dump

rm $binimg.opreport.log
opreport > $binimg.opreport.log

rm $binimg.opreport.sym
opreport -l > $binimg.opreport.sym

opcontrol --shutdown
opcontrol --deinit
echo "Done"


perf is a kernel-based subsystem that provide a framework for performance analysis of the impact of the progams being run on the kernel. It covers: hardware (CPU/PMU, Performance Monitoring Unit) features; and software features (software counters, tracepoints).

perf tutorial

perf list

Lists the events available on a particular machine. These events vary based on the performance monitoring hardware and the software configuration of the system.

$ perf list

List of pre-defined events (to be used in -e):
  cpu-cycles OR cycles                               [Hardware event]
  stalled-cycles-frontend OR idle-cycles-frontend    [Hardware event]
  stalled-cycles-backend OR idle-cycles-backend      [Hardware event]
  instructions                                       [Hardware event]
  cache-references                                   [Hardware event]
  cache-misses                                       [Hardware event]
  branch-instructions OR branches                    [Hardware event]
  branch-misses                                      [Hardware event]
  bus-cycles                                         [Hardware event]

  cpu-clock                                          [Software event]
  task-clock                                         [Software event]
  page-faults OR faults                              [Software event]
  minor-faults                                       [Software event]
  major-faults                                       [Software event]
  context-switches OR cs                             [Software event]
  cpu-migrations OR migrations                       [Software event]
  alignment-faults                                   [Software event]
  emulation-faults                                   [Software event]

  L1-dcache-loads                                    [Hardware cache event]
  L1-dcache-load-misses                              [Hardware cache event]
  L1-dcache-stores                                   [Hardware cache event]
  L1-dcache-store-misses                             [Hardware cache event]
  L1-dcache-prefetches                               [Hardware cache event]
  L1-dcache-prefetch-misses                          [Hardware cache event]
  L1-icache-loads                                    [Hardware cache event]
  L1-icache-load-misses                              [Hardware cache event]
  L1-icache-prefetches                               [Hardware cache event]
  L1-icache-prefetch-misses                          [Hardware cache event]
  LLC-loads                                          [Hardware cache event]
  LLC-load-misses                                    [Hardware cache event]
  LLC-stores                                         [Hardware cache event]
  LLC-store-misses                                   [Hardware cache event]
  LLC-prefetches                                     [Hardware cache event]
  LLC-prefetch-misses                                [Hardware cache event]
  dTLB-loads                                         [Hardware cache event]
  dTLB-load-misses                                   [Hardware cache event]
  dTLB-stores                                        [Hardware cache event]
  dTLB-store-misses                                  [Hardware cache event]
  dTLB-prefetches                                    [Hardware cache event]
  dTLB-prefetch-misses                               [Hardware cache event]
  iTLB-loads                                         [Hardware cache event]
  iTLB-load-misses                                   [Hardware cache event]
  branch-loads                                       [Hardware cache event]
  branch-load-misses                                 [Hardware cache event]
  node-loads                                         [Hardware cache event]
  node-load-misses                                   [Hardware cache event]
  node-stores                                        [Hardware cache event]
  node-store-misses                                  [Hardware cache event]
  node-prefetches                                    [Hardware cache event]
  node-prefetch-misses                               [Hardware cache event]

  rNNN (see 'perf list --help' on how to encode it)  [Raw hardware event descriptor]

  mem:<addr>[:access]                                [Hardware breakpoint]

Note: Running this as root will give out an extended list of events; some of the events (tracepoints?) require root privileges.

perf stat <options> <cmd>

Gathers overall statistics for common performance events, including instructions executed and clock cycles consumed. There are option flags to gather statistics on events other than the default measurement events.

$ g++ -std=c++11 -ggdb -fno-omit-frame-pointer perftest.cpp -o perftest

$ perf stat ./perftest
 Performance counter stats for './perftest':

        379.991103 task-clock                #    0.996 CPUs utilized          
                62 context-switches          #    0.000 M/sec                  
                 0 CPU-migrations            #    0.000 M/sec                  
             6,436 page-faults               #    0.017 M/sec                  
       984,969,006 cycles                    #    2.592 GHz                     [83.27%]
       663,592,329 stalled-cycles-frontend   #   67.37% frontend cycles idle    [83.17%]
       473,904,165 stalled-cycles-backend    #   48.11% backend  cycles idle    [66.42%]
     1,010,613,552 instructions              #    1.03  insns per cycle        
                                             #    0.66  stalled cycles per insn [83.23%]
        53,831,403 branches                  #  141.665 M/sec                   [84.14%]
           401,518 branch-misses             #    0.75% of all branches         [83.48%]

       0.381602838 seconds time elapsed

$ perf stat --event=L1-dcache-load-misses ./perftest
 Performance counter stats for './perftest':

        12,942,048 L1-dcache-load-misses                                       

       0.373719009 seconds time elapsed
perf record

Records performance data into a file which can be later analyzed using perf report.

$ perf record --event=L1-dcache-load-misses ./perftest

[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.025 MB perf.data (~1078 samples) ]

$ ls -al
-rw-------  1 amelinte amelinte  27764 Feb 17 17:23 perf.data
perf report

Reads the performance data from a file and analyzes the recorded data.

$ perf report --stdio
# ========
# captured on: Sun Feb 17 17:23:34 2013
# hostname : bear
# os release : 3.2.0-4-amd64
# perf version : 3.2.17
# arch : x86_64
# nrcpus online : 4
# nrcpus avail : 4
# cpudesc : Intel(R) Core(TM) i3 CPU M 390 @ 2.67GHz
# cpuid : GenuineIntel,6,37,5
# total memory : 3857640 kB
# cmdline : /usr/bin/perf_3.2 record --event=L1-dcache-load-misses ./perftest 
# event : name = L1-dcache-load-misses, type = 3, config = 0x10000, config1 = 0x0, config2 = 0x0, excl_usr = 0, excl_kern = 0, id = { 
# HEADER_CPU_TOPOLOGY info available, use -I to display
# HEADER_NUMA_TOPOLOGY info available, use -I to display
# ========
# Events: 274  L1-dcache-load-misses
# Overhead         Command   Shared Object      Symbol
# ........         ........  .................  ..................
#   95.93 percent  perftest  perftest           [.] 0xd35           
#    1.06 percent  perftest  [kernel.kallsyms]  [k] clear_page_c
#    0.82 percent  perftest  [kernel.kallsyms]  [k] run_timer_softirq
#    0.42 percent  perftest  [kernel.kallsyms]  [k] trylock_page
#    0.41 percent  perftest  [kernel.kallsyms]  [k] __rcu_pending
#    0.41 percent  perftest  [kernel.kallsyms]  [k] update_curr
#    0.33 percent  perftest  [kernel.kallsyms]  [k] do_raw_spin_lock
#    0.26 percent  perftest  [kernel.kallsyms]  [k] __flush_tlb_one
#    0.18 percent  perftest  [kernel.kallsyms]  [k] flush_old_exec
#    0.06 percent  perftest  [kernel.kallsyms]  [k] __free_one_page
#    0.05 percent  perftest  [kernel.kallsyms]  [k] free_swap_cache
#    0.05 percent  perftest  [kernel.kallsyms]  [k] zone_statistics
#    0.04 percent  perftest  [kernel.kallsyms]  [k] alloc_pages_vma
#    0.01 percent  perftest  [kernel.kallsyms]  [k] mm_init
#    0.01 percent  perftest  [kernel.kallsyms]  [k] vfs_read
#    0.00 percent  perftest  [kernel.kallsyms]  [k] __cond_resched
#    0.00 percent  perftest  [kernel.kallsyms]  [k] finish_task_switch
# (For a higher level overview, try: perf report --sort comm,dso)

$ perf record -g ./perftest
$ perf report -g --stdio
# Overhead   Command      Shared Object                  Symbol
# ........  ........  .................  ......................
    97.23%  perftest  perftest           [.] 0xc75           
            --- 0x400d2c
perf annotate

Reads the input file and displays an annotated version of the code. If the object file has debug symbols then the source code will be displayed alongside assembly code. If there is no debug info in the object, then annotated assembly is displayed. Broken!?

$ perf annotate -i ./perf.data -d ./perftest --stdio -f
The ./perf.data file has no samples!
perf top

Performance top tool look-alike. It generates and displays a performance counter profile in realtime.

perf archive
perf buildid-cache
perf buildid-list
perf diff
perf inject
perf kmem
perf kvm
perf probe
perf sched
perf script
perf timechart

Valgrind: cachegrind

Cachegrind simulates a machine with two level ([I1 & D1] and L2) cache and branch (mis)prediction. Useful in that it annotates the code down to line level. Can significantly differ from machine's real CPU. Will not go very far on an AMD64 CPU (vex disassembler issues). Extremely slow, typically slows down the application 12-15 times.

DIY: libhitcount

libmemleak can be easily modified to keep track of what calls into a particular point in the code. Just insert an mtrace() call in that place.


How things scale up

L1 cache reference ......................... 0.5 ns
Branch mispredict ............................ 5 ns
L2 cache reference ........................... 7 ns
Mutex lock/unlock ........................... 25 ns
Main memory reference ...................... 100 ns             
Compress 1K bytes with Zippy ............. 3,000 ns  =   3 µs
Send 2K bytes over 1 Gbps network ....... 20,000 ns  =  20 µs
SSD random read ........................ 150,000 ns  = 150 µs
Read 1 MB sequentially from memory ..... 250,000 ns  = 250 µs
Round trip within same datacenter ...... 500,000 ns  = 0.5 ms
Read 1 MB sequentially from SSD* ..... 1,000,000 ns  =   1 ms
Disk seek ........................... 10,000,000 ns  =  10 ms
Read 1 MB sequentially from disk .... 20,000,000 ns  =  20 ms
Send packet CA->Netherlands->CA .... 150,000,000 ns  = 150 ms
Operation           Cost (ns)     Ratio
Clock period        0.6           1.0
Best-case CAS       37.9          63.2
Best-case lock      65.6          109.3
Single cache miss   139.5         232.5
CAS cache miss      306.0         510.0
Comms Fabric        3,000         5,000
Global Comms        130,000,000   216,000,000
Table 2.1: Performance of Synchronization Mechanisms on 4-CPU 1.8GHz AMD Opteron 844 System
  • Source: Paul E. McKenney

Notes on Hyper Threading

Other tools

Hardware emulators


umockdev mocks hardware devices for creating unit tests for libraries and programs that handle Linux hardware devices. It also provides tools to record the properties and behaviour of particular devices. Useful for reproducing failure scenarios encountered in the wild.

Building the toolkit

The code detailed herein is slowly melted into a toolkit and is available at the LPT site.


A C++11/C++0x compiler

If your distribution does not have such a compiler, compile one from scratch. First compile the gcc prerequsites, available here:

  • cloog
  • ppl
  • mpc
  • mpfr
  • gmp

Then compile gcc itself:

./configure \
    --enable-languages=c,c++ \
    --prefix=/opt/lpt/gcc-4.7.0-bin \
    --with-mpfr=/opt/lpt/mpfr-2.4.2-bin \
    --with-mpc=/opt/lpt/mpc-0.8.1-bin \
    --with-gmp=/opt/lpt/gmp-4.3.2-bin \
    --with-ppl=/opt/lpt/ppl-0.11-bin \
LD_LIBRARY_PATH=/opt/lpt/mpfr-2.4.2-bin/lib:/opt/lpt/mpc-0.8.1-bin/lib:/opt/lpt/gmp-4.3.2-bin/lib:/opt/lpt/ppl-0.11-bin/lib:/opt/lpt/cloog-ppl-0.15.11-bin/lib   make
make install


This one is a straightforward ./configure && make. No need to install.

-fPIC libbfd

The call stack code has a dependency on libbfd. If you wish to link libbfd statically to cut short a dependency on the deployment machine - which may or may not have a dynamic libbfd - you need an libbfd.a that is -fPIC compiled. This flag is needed because the static library will be linked into a shared object. The plain libbfd.a coming with the platforms might not be an -fPIC one, in which case you have to compile your own.

cd binutils-2.22/bfd
export CFLAGS = -fPIC

No need to install it either.


Finally, we need to adjust the LPT Makefile:

CXX11_HOME       = /opt/lpt/gcc-4.7.0-bin
PAPI_HOME        = /opt/lpt/papi-4.4.0/src
LIBBFD_FPIC_HOME = /opt/lpt/binutils-2.22/bfd



Things to watch for

Interrupted calls

A number of API functions return an error code if the call was interrupted by a signal. Usually this is not an error by itself and the call should be restarted. For instance:

int raccept( int s, struct sockaddr *addr, socklen_t *addrlen )
     int rc;
     do {
         rc = accept( s, addr, addrlen );
     } while ( rc == -1 && errno == EINTR );
     return rc;

The list of interruptible function differs from Unix-like platform to platform. For Linux see signal(7). Note that sem_wait() & co. are on list.

Spurious wake-ups

Threads waiting on a pthreads condition variable can be waken up even if the condition hs not been met. Upon waking up, the condition should be explicitly checked and return waiting if it is not met.

Code review checklist

  1. Functional impact
    1. Which modules are affected; which functionality is impacted.
    2. Which modules & functionality is consumed by the changed code.
    3. Which modules/functionality is using it.
    4. Side effects.
  2. Architecture
    1. Scalability
    2. Stability
    3. Functionality
    4. Robustness
  3. Technicalities
    1. Generic
      1. First pass: logic. Correctness. What does it (tries to) do.
        1. Side effects.
        2. Why is code there?
      2. Second pass: error handling & odd conditions.
        1. Side effects.
        2. Are returned codes checked. Can the code handle them.
        3. Arguments:
          1. Border/out of bounds/strange values handled.
          2. Arg type: convenient?
          3. Where is the data coming from? Three levels upstream.
          4. Where are the outputs going? Three levels downstream.
            1. What are the possible/permitted outputs? Are all handled downstream?
        4. Variables: initialized?
        5. (over/under)flows:
          1. buffers.
          2. integral types.
        6. Asserted invariants & assumptions.
        7. Loops: off-by-one? infinite?
        8. Recursion: base case?
      3. Third pass: thread safety.
      4. Localization?
      5. Duplicate information?
      6. Math:
        1. Small values processed first?
        2. NaN checks/std::isnan & friends.
        3. Division by zero.
    2. C
      1. return: resource handling.
      2. Use the N functions (ex. s'n'printf vs. sprintf).
      3. printf & friends: do formatters match type and number of arguments?
      4. Non re-entrant/non thread-safe API
      5. Unsafe/race condition afflicted API ( mktemp() )
      6. enums:
        1. all cases handled?
        2. default returns/logs/throws error
      7. Unspecified/undefined/implementation-defined/non-standard behavior or extensions
    3. C++
      1. Fourth pass: exception handling:
        1. In destructors.
        2. Exception catching boundaries?
        3. Are resources RAII guarded.
      2. Fifth pass: performance.
        1. C++11: perfect forwarding & temporaries.
        2. How does it scale?
      3. Initialized data members.
      4. The big 6 (default constructor, destructor, copy/move constructors/operators).
      5. Virtual destructor?
      6. Resource handling: RAII.
      7. Where is created & destructed? Lifetime management?
      8. Thorough const & volatile.
      9. Are predicates stateless.
      10. Use size_t for std::string positions.
      11. Why casts?
      12. Do not trace under lock
      13. Do not destroy objects under lock
  4. Platform specific
    1. POSIX/SUS
      1. Interrupted calls.
      2. Spurious wake-ups.
      3. Calls known to have race conditions.



A pseudo-filesystem exposing information about running processes:

# tree /proc/26041
|-- cmdline             # Command line
|-- cwd ->              /current/working/folder/for/PID
|-- environ             # Program environment variables
|-- exe -> /bin/su
|-- fd                  # Open files descriptors
|   |-- 0 -> /dev/pts/21
|   |-- 1 -> /dev/pts/21
|   |-- 2 -> /dev/pts/21
|   `-- 3 -> socket:[113497835]
|-- fdinfo
|   |-- 0
|   |-- 1
|   |-- 2
|   `-- 3
|-- latency
|-- limits
|-- maps
|-- mem
|-- mountinfo
|-- mounts
|-- mountstats

# cat /proc/26041/status
VmPeak:   103276 kB     # Max virtual memory reached
VmSize:   103196 kB     # Current VM
VmLck:         0 kB
VmHWM:      1492 kB
VmRSS:      1488 kB     # Live memory used
Threads:        1
sysstat, sar

Other tools


Given an address in an executable or an offset in a section of a relocatable object, addr2line translates it into file name and line number.


A tool to demangle symbol names.

  • Disassemble binary, with source code: objdump -C -S -r -R -l <binary>


  • HLE: Hardware Lock Elision
  • HT: Hyper Threading
  • RTM: Restricted Transactional Memory
  • TSX: Intel Transactional Synchronization Extension ([3], [4])

References and further reading

Books, articles, presentations



GNU Free Documentation License

Version 1.3, 3 November 2008 Copyright (C) 2000, 2001, 2002, 2007, 2008 Free Software Foundation, Inc. <http://fsf.org/>

Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed.


The purpose of this License is to make a manual, textbook, or other functional and useful document "free" in the sense of freedom: to assure everyone the effective freedom to copy and redistribute it, with or without modifying it, either commercially or noncommercially. Secondarily, this License preserves for the author and publisher a way to get credit for their work, while not being considered responsible for modifications made by others.

This License is a kind of "copyleft", which means that derivative works of the document must themselves be free in the same sense. It complements the GNU General Public License, which is a copyleft license designed for free software.

We have designed this License in order to use it for manuals for free software, because free software needs free documentation: a free program should come with manuals providing the same freedoms that the software does. But this License is not limited to software manuals; it can be used for any textual work, regardless of subject matter or whether it is published as a printed book. We recommend this License principally for works whose purpose is instruction or reference.


This License applies to any manual or other work, in any medium, that contains a notice placed by the copyright holder saying it can be distributed under the terms of this License. Such a notice grants a world-wide, royalty-free license, unlimited in duration, to use that work under the conditions stated herein. The "Document", below, refers to any such manual or work. Any member of the public is a licensee, and is addressed as "you". You accept the license if you copy, modify or distribute the work in a way requiring permission under copyright law.

A "Modified Version" of the Document means any work containing the Document or a portion of it, either copied verbatim, or with modifications and/or translated into another language.

A "Secondary Section" is a named appendix or a front-matter section of the Document that deals exclusively with the relationship of the publishers or authors of the Document to the Document's overall subject (or to related matters) and contains nothing that could fall directly within that overall subject. (Thus, if the Document is in part a textbook of mathematics, a Secondary Section may not explain any mathematics.) The relationship could be a matter of historical connection with the subject or with related matters, or of legal, commercial, philosophical, ethical or political position regarding them.

The "Invariant Sections" are certain Secondary Sections whose titles are designated, as being those of Invariant Sections, in the notice that says that the Document is released under this License. If a section does not fit the above definition of Secondary then it is not allowed to be designated as Invariant. The Document may contain zero Invariant Sections. If the Document does not identify any Invariant Sections then there are none.

The "Cover Texts" are certain short passages of text that are listed, as Front-Cover Texts or Back-Cover Texts, in the notice that says that the Document is released under this License. A Front-Cover Text may be at most 5 words, and a Back-Cover Text may be at most 25 words.

A "Transparent" copy of the Document means a machine-readable copy, represented in a format whose specification is available to the general public, that is suitable for revising the document straightforwardly with generic text editors or (for images composed of pixels) generic paint programs or (for drawings) some widely available drawing editor, and that is suitable for input to text formatters or for automatic translation to a variety of formats suitable for input to text formatters. A copy made in an otherwise Transparent file format whose markup, or absence of markup, has been arranged to thwart or discourage subsequent modification by readers is not Transparent. An image format is not Transparent if used for any substantial amount of text. A copy that is not "Transparent" is called "Opaque".

Examples of suitable formats for Transparent copies include plain ASCII without markup, Texinfo input format, LaTeX input format, SGML or XML using a publicly available DTD, and standard-conforming simple HTML, PostScript or PDF designed for human modification. Examples of transparent image formats include PNG, XCF and JPG. Opaque formats include proprietary formats that can be read and edited only by proprietary word processors, SGML or XML for which the DTD and/or processing tools are not generally available, and the machine-generated HTML, PostScript or PDF produced by some word processors for output purposes only.

The "Title Page" means, for a printed book, the title page itself, plus such following pages as are needed to hold, legibly, the material this License requires to appear in the title page. For works in formats which do not have any title page as such, "Title Page" means the text near the most prominent appearance of the work's title, preceding the beginning of the body of the text.

The "publisher" means any person or entity that distributes copies of the Document to the public.

A section "Entitled XYZ" means a named subunit of the Document whose title either is precisely XYZ or contains XYZ in parentheses following text that translates XYZ in another language. (Here XYZ stands for a specific section name mentioned below, such as "Acknowledgements", "Dedications", "Endorsements", or "History".) To "Preserve the Title" of such a section when you modify the Document means that it remains a section "Entitled XYZ" according to this definition.

The Document may include Warranty Disclaimers next to the notice which states that this License applies to the Document. These Warranty Disclaimers are considered to be included by reference in this License, but only as regards disclaiming warranties: any other implication that these Warranty Disclaimers may have is void and has no effect on the meaning of this License.


You may copy and distribute the Document in any medium, either commercially or noncommercially, provided that this License, the copyright notices, and the license notice saying this License applies to the Document are reproduced in all copies, and that you add no other conditions whatsoever to those of this License. You may not use technical measures to obstruct or control the reading or further copying of the copies you make or distribute. However, you may accept compensation in exchange for copies. If you distribute a large enough number of copies you must also follow the conditions in section 3.

You may also lend copies, under the same conditions stated above, and you may publicly display copies.


If you publish printed copies (or copies in media that commonly have printed covers) of the Document, numbering more than 100, and the Document's license notice requires Cover Texts, you must enclose the copies in covers that carry, clearly and legibly, all these Cover Texts: Front-Cover Texts on the front cover, and Back-Cover Texts on the back cover. Both covers must also clearly and legibly identify you as the publisher of these copies. The front cover must present the full title with all words of the title equally prominent and visible. You may add other material on the covers in addition. Copying with changes limited to the covers, as long as they preserve the title of the Document and satisfy these conditions, can be treated as verbatim copying in other respects.

If the required texts for either cover are too voluminous to fit legibly, you should put the first ones listed (as many as fit reasonably) on the actual cover, and continue the rest onto adjacent pages.

If you publish or distribute Opaque copies of the Document numbering more than 100, you must either include a machine-readable Transparent copy along with each Opaque copy, or state in or with each Opaque copy a computer-network location from which the general network-using public has access to download using public-standard network protocols a complete Transparent copy of the Document, free of added material. If you use the latter option, you must take reasonably prudent steps, when you begin distribution of Opaque copies in quantity, to ensure that this Transparent copy will remain thus accessible at the stated location until at least one year after the last time you distribute an Opaque copy (directly or through your agents or retailers) of that edition to the public.

It is requested, but not required, that you contact the authors of the Document well before redistributing any large number of copies, to give them a chance to provide you with an updated version of the Document.


You may copy and distribute a Modified Version of the Document under the conditions of sections 2 and 3 above, provided that you release the Modified Version under precisely this License, with the Modified Version filling the role of the Document, thus licensing distribution and modification of the Modified Version to whoever possesses a copy of it. In addition, you must do these things in the Modified Version:

  1. Use in the Title Page (and on the covers, if any) a title distinct from that of the Document, and from those of previous versions (which should, if there were any, be listed in the History section of the Document). You may use the same title as a previous version if the original publisher of that version gives permission.
  2. List on the Title Page, as authors, one or more persons or entities responsible for authorship of the modifications in the Modified Version, together with at least five of the principal authors of the Document (all of its principal authors, if it has fewer than five), unless they release you from this requirement.
  3. State on the Title page the name of the publisher of the Modified Version, as the publisher.
  4. Preserve all the copyright notices of the Document.
  5. Add an appropriate copyright notice for your modifications adjacent to the other copyright notices.
  6. Include, immediately after the copyright notices, a license notice giving the public permission to use the Modified Version under the terms of this License, in the form shown in the Addendum below.
  7. Preserve in that license notice the full lists of Invariant Sections and required Cover Texts given in the Document's license notice.
  8. Include an unaltered copy of this License.
  9. Preserve the section Entitled "History", Preserve its Title, and add to it an item stating at least the title, year, new authors, and publisher of the Modified Version as given on the Title Page. If there is no section Entitled "History" in the Document, create one stating the title, year, authors, and publisher of the Document as given on its Title Page, then add an item describing the Modified Version as stated in the previous sentence.
  10. Preserve the network location, if any, given in the Document for public access to a Transparent copy of the Document, and likewise the network locations given in the Document for previous versions it was based on. These may be placed in the "History" section. You may omit a network location for a work that was published at least four years before the Document itself, or if the original publisher of the version it refers to gives permission.
  11. For any section Entitled "Acknowledgements" or "Dedications", Preserve the Title of the section, and preserve in the section all the substance and tone of each of the contributor acknowledgements and/or dedications given therein.
  12. Preserve all the Invariant Sections of the Document, unaltered in their text and in their titles. Section numbers or the equivalent are not considered part of the section titles.
  13. Delete any section Entitled "Endorsements". Such a section may not be included in the Modified version.
  14. Do not retitle any existing section to be Entitled "Endorsements" or to conflict in title with any Invariant Section.
  15. Preserve any Warranty Disclaimers.

If the Modified Version includes new front-matter sections or appendices that qualify as Secondary Sections and contain no material copied from the Document, you may at your option designate some or all of these sections as invariant. To do this, add their titles to the list of Invariant Sections in the Modified Version's license notice. These titles must be distinct from any other section titles.

You may add a section Entitled "Endorsements", provided it contains nothing but endorsements of your Modified Version by various parties—for example, statements of peer review or that the text has been approved by an organization as the authoritative definition of a standard.

You may add a passage of up to five words as a Front-Cover Text, and a passage of up to 25 words as a Back-Cover Text, to the end of the list of Cover Texts in the Modified Version. Only one passage of Front-Cover Text and one of Back-Cover Text may be added by (or through arrangements made by) any one entity. If the Document already includes a cover text for the same cover, previously added by you or by arrangement made by the same entity you are acting on behalf of, you may not add another; but you may replace the old one, on explicit permission from the previous publisher that added the old one.

The author(s) and publisher(s) of the Document do not by this License give permission to use their names for publicity for or to assert or imply endorsement of any Modified Version.


You may combine the Document with other documents released under this License, under the terms defined in section 4 above for modified versions, provided that you include in the combination all of the Invariant Sections of all of the original documents, unmodified, and list them all as Invariant Sections of your combined work in its license notice, and that you preserve all their Warranty Disclaimers.

The combined work need only contain one copy of this License, and multiple identical Invariant Sections may be replaced with a single copy. If there are multiple Invariant Sections with the same name but different contents, make the title of each such section unique by adding at the end of it, in parentheses, the name of the original author or publisher of that section if known, or else a unique number. Make the same adjustment to the section titles in the list of Invariant Sections in the license notice of the combined work.

In the combination, you must combine any sections Entitled "History" in the various original documents, forming one section Entitled "History"; likewise combine any sections Entitled "Acknowledgements", and any sections Entitled "Dedications". You must delete all sections Entitled "Endorsements".


You may make a collection consisting of the Document and other documents released under this License, and replace the individual copies of this License in the various documents with a single copy that is included in the collection, provided that you follow the rules of this License for verbatim copying of each of the documents in all other respects.

You may extract a single document from such a collection, and distribute it individually under this License, provided you insert a copy of this License into the extracted document, and follow this License in all other respects regarding verbatim copying of that document.


A compilation of the Document or its derivatives with other separate and independent documents or works, in or on a volume of a storage or distribution medium, is called an "aggregate" if the copyright resulting from the compilation is not used to limit the legal rights of the compilation's users beyond what the individual works permit. When the Document is included in an aggregate, this License does not apply to the other works in the aggregate which are not themselves derivative works of the Document.

If the Cover Text requirement of section 3 is applicable to these copies of the Document, then if the Document is less than one half of the entire aggregate, the Document's Cover Texts may be placed on covers that bracket the Document within the aggregate, or the electronic equivalent of covers if the Document is in electronic form. Otherwise they must appear on printed covers that bracket the whole aggregate.


Translation is considered a kind of modification, so you may distribute translations of the Document under the terms of section 4. Replacing Invariant Sections with translations requires special permission from their copyright holders, but you may include translations of some or all Invariant Sections in addition to the original versions of these Invariant Sections. You may include a translation of this License, and all the license notices in the Document, and any Warranty Disclaimers, provided that you also include the original English version of this License and the original versions of those notices and disclaimers. In case of a disagreement between the translation and the original version of this License or a notice or disclaimer, the original version will prevail.

If a section in the Document is Entitled "Acknowledgements", "Dedications", or "History", the requirement (section 4) to Preserve its Title (section 1) will typically require changing the actual title.


You may not copy, modify, sublicense, or distribute the Document except as expressly provided under this License. Any attempt otherwise to copy, modify, sublicense, or distribute it is void, and will automatically terminate your rights under this License.

However, if you cease all violation of this License, then your license from a particular copyright holder is reinstated (a) provisionally, unless and until the copyright holder explicitly and finally terminates your license, and (b) permanently, if the copyright holder fails to notify you of the violation by some reasonable means prior to 60 days after the cessation.

Moreover, your license from a particular copyright holder is reinstated permanently if the copyright holder notifies you of the violation by some reasonable means, this is the first time you have received notice of violation of this License (for any work) from that copyright holder, and you cure the violation prior to 30 days after your receipt of the notice.

Termination of your rights under this section does not terminate the licenses of parties who have received copies or rights from you under this License. If your rights have been terminated and not permanently reinstated, receipt of a copy of some or all of the same material does not give you any rights to use it.


The Free Software Foundation may publish new, revised versions of the GNU Free Documentation License from time to time. Such new versions will be similar in spirit to the present version, but may differ in detail to address new problems or concerns. See http://www.gnu.org/copyleft/.

Each version of the License is given a distinguishing version number. If the Document specifies that a particular numbered version of this License "or any later version" applies to it, you have the option of following the terms and conditions either of that specified version or of any later version that has been published (not as a draft) by the Free Software Foundation. If the Document does not specify a version number of this License, you may choose any version ever published (not as a draft) by the Free Software Foundation. If the Document specifies that a proxy can decide which future versions of this License can be used, that proxy's public statement of acceptance of a version permanently authorizes you to choose that version for the Document.


"Massive Multiauthor Collaboration Site" (or "MMC Site") means any World Wide Web server that publishes copyrightable works and also provides prominent facilities for anybody to edit those works. A public wiki that anybody can edit is an example of such a server. A "Massive Multiauthor Collaboration" (or "MMC") contained in the site means any set of copyrightable works thus published on the MMC site.

"CC-BY-SA" means the Creative Commons Attribution-Share Alike 3.0 license published by Creative Commons Corporation, a not-for-profit corporation with a principal place of business in San Francisco, California, as well as future copyleft versions of that license published by that same organization.

"Incorporate" means to publish or republish a Document, in whole or in part, as part of another Document.

An MMC is "eligible for relicensing" if it is licensed under this License, and if all works that were first published under this License somewhere other than this MMC, and subsequently incorporated in whole or in part into the MMC, (1) had no cover texts or invariant sections, and (2) were thus incorporated prior to November 1, 2008.

The operator of an MMC Site may republish an MMC contained in the site under CC-BY-SA on the same site at any time before August 1, 2009, provided the MMC is eligible for relicensing.

How to use this License for your documents

To use this License in a document you have written, include a copy of the License in the document and put the following copyright and license notices just after the title page:

Copyright (c) YEAR YOUR NAME.
Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License, Version 1.3
or any later version published by the Free Software Foundation;
with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts.
A copy of the license is included in the section entitled "GNU
Free Documentation License".

If you have Invariant Sections, Front-Cover Texts and Back-Cover Texts, replace the "with...Texts." line with this:

with the Invariant Sections being LIST THEIR TITLES, with the
Front-Cover Texts being LIST, and with the Back-Cover Texts being LIST.

If you have Invariant Sections without Cover Texts, or some other combination of the three, merge those two alternatives to suit the situation.

If your document contains nontrivial examples of program code, we recommend releasing these examples in parallel under your choice of free software license, such as the GNU General Public License, to permit their use in free software.

  1. {https://gcc.gnu.org/gcc-4.9/changes.html}