Monday, June 18, 2012

Linux: ELF shared library versioning

Here is an amazingly helpful description of all aspects of shared library and symbol versioning for ELF libraries (on Linux, BSD, etc.):

For info on @ and @@:

Sunday, May 13, 2012

GCC demangling and stack traces

Tons of useful info:

In fact, it's so useful, I'm afraid to lose that page, so I've copied it here:

Call stack trace generation

Call stack trace ("backtrace") is very useful in debugging. Here are several ways to retrieve the backtrace in a user program.

(The contents are mostly from here and here)

Easiest approach: __builtin_return_address

GCC has a built-in function to retrieve call stack trace's addresses. For example
void do_backtrace()
    printf("Frame 0: PC=%p\n", __builtin_return_address(0));
    printf("Frame 1: PC=%p\n", __builtin_return_address(1));
    printf("Frame 2: PC=%p\n", __builtin_return_address(2));
    printf("Frame 3: PC=%p\n", __builtin_return_address(3));
__builtin_return_address(0) is always current function's address. On the other hand, __builtin_return_address(1), __builtin_return_address(2), ... may not be available on all platforms.

What to do with these addresses ?

Addresses can be mapped to the binary executable or dynamic link libraries. This is always doable even if the binary executable has been stripped off the symbols.

To see the mapping during runtime, parse the following plain-text file on the /proc file system:

A utility called pmap can do the same.

If the address belongs to a DLL, it is possible to obtain the function name since DLLs are usually not stripped.

Addresses can be mapped to function names. Even if a binary executable is compiled without -g option, it still contains function names. To see the function names in the binary executable, do

nm -C -n a.out
To see the function names programmatically in the binary executable during run-time, read later paragraphs.

Addresses can be mapped to line numbers in source files. This extra information (in DWARF format) is added to the binary executable if it is compiled with -g option. To see line numbers in source files, do

objdump -WL a.out
objdump --dwarf=decodedline a.out
or even better:
addr2line -ifC a.out 0x123456
where 0x123456 is the address of interest. To see line numbers in source files programmatically during run-time, read later paragraphs.

Approach 2: backtrace

backtrace and backtrace_symbols are functions in Glibc. To use backtrace_symbols, one must compile the program with -rdynamic option.

One does not need to compile with -g option (but -rdynamic option cannot be used together with -static option) since backtrace_symbols cannot retrieve line number information. Actually, one can even strip off the symbols, and the backtrace_symbols will still work. This is because when -rdynamic is used, all global symbols are also stored in .dynsym section in the ELF-formatted executable binary, and this section cannot be stripped away. (To see the content of .dynsym section, use readelf -s a.out command, or readelf -p .dynstr a.out command.)

backtrace_symbols obtains symbol information from .dynsym section.

(The main purpose of .dynsym section is for dynamic link libraries to expose their symbols so the runtime linker can find them.)

Here is the sample program:

#include <execinfo.h>
void do_backtrace()
    #define BACKTRACE_SIZ 100
    void    *array[BACKTRACE_SIZ];
    size_t  size, i;
    char    **strings;

    size = backtrace(array, BACKTRACE_SIZ);
    strings = backtrace_symbols(array, size);

    for (i = 0; i < size; ++i) {
        printf("%p : %s\n", array[i], strings[i]);


For C++ programs, to get demangled names, use abi::__cxa_demangle (include the header cxxabi.h)

Approach 3: Improved backtrace

The backtrace_symbols in Glibc uses dladdr to obtain function names, but it cannot retrieve line numbers. Jeff Muizelaar has an improved version here which can do line numbers.

If the user program is compiled without any special command-line options, then one can obtain function names (of course, provided the binary executable is not stripped.) Better yet, -rdynamic compiler option is not needed.

If the user program is compiled with -g option, one can obtain both line numbers and function names.

Note that to compile Jeff Muizelaar's backtrace_symbols implementation, make sure the following two macros are defined and appears as the first two lines of a user program (they must precede before all #include ...):

#define __USE_GNU
#define _GNU_SOURCE
and one needs Binary File Descriptor (BFD) library, which is now part of GNU binutils when linking Jeff's code to the user program.

Approach 4: libunwind

libunwind does pretty much what the original backtrace/backtrace_symbols do. Its main purpose, however, is to unwind the stack programmatically (even more powerful than setjmp/longjmp pair) through unw_step and unw_resume calls. One can also peek and modify the saved register values on stack via unw_get_reg, unw_get_freg, unw_set_reg, and unw_set_freg calls.

If one just wants to retrieve the backtrace, use the following code:

#include <libunwind.h>
void do_backtrace()
    unw_cursor_t    cursor;
    unw_context_t   context;

    unw_init_local(&cursor, &context);

    while (unw_step(&cursor) > 0) {
        unw_word_t  offset, pc;
        char        fname[64];

        unw_get_reg(&cursor, UNW_REG_IP, &pc);

        fname[0] = '\0';
        (void) unw_get_proc_name(&cursor, fname, sizeof(fname), &offset);

        printf ("%p : (%s+0x%x) [%p]\n", pc, fname, offset, pc);

and linked the user program with -lunwind -lunwind-x86_64.

There is no need to compile the user program with -g option.

Wednesday, April 25, 2012

Blekko's NoSQL Database

Here is an interesting blog on the Blekko NoSQL DB. Greg Lindahl (CTO of blekko) mentions cost/value of Amazon AWS and details of the implementation. In particular, the "Combinator", which relies on commutative and associative operations.

Saturday, March 31, 2012

Micheal Dunn: How console games can reach the Facebook audience

My brother (Micheal Dunn) gave a talk at Developer Day at Facebook's Game Developer Conference in March.

Tuesday, February 7, 2012

Git at large companies

That mentions Amazon, which has embraced git without prohibiting other VCSs. That helped us to solve several problems with scale (100k small repos vs. 3 large repos). I can't comment much, but I can re-post what was said:
Amazon uses Perforce at the moment, and for the most part developers are unhappy with it, as well as the team that has to support it (single giant server prone to outages which block up a couple thousand developers, etc). We're in the process of moving to Git for all of our source.
On the other hand, what you're describing as a problem (what Facebook is describing as going to be a problem) is less likely to be one for Amazon as, with some exceptions that are in the process of fixing the issue, the majority of software at Amazon is developed as a service. Services are segregated into their own package, with most services being broken up into cohesive subpackages (a service my team is building will probably have ~10-13 packages when done), and we have a dependency modeling system for packages baked into everything, from build through deploy, which eliminates most of the cognitive overhead of breaking our services up this way.
All of this translates very well into different Git repositories. What we lose is cohesive atomic commits across packages, which we do get with Perforce. The upshot is we have a team developing a system to handle that specific case.
There you go. SOA (Service Oriented Architecture) solves the repo problem along with a bunch of others.

The problems I've had with git at Amazon are minor, and we do have a very large code-base.