I'm working on a project that contains many libraries (built as shared objects) that use functionality from each other. In normal usage (in this project) these are loaded dynamically by python. I have no idea how the python loader resolves dependencies between these libraries, but if one tries to build an application that uses the plain dynamic linker to link the libraries, things become difficult. The reason: these libraries have no DT_NEEDED tags so the dynamic linker has no dependency information.


Suppose we have a library libA.so that uses functionality from libB.so to do some of its work. Now suppose you have an application exe that calls functions in libA.so, but not libB.so. The person invoking the linker to link exe knows to build with -lA because they know that exe uses libA.so. However this person may not know to link with -lB: their program that they wrote is exe, and they have no idea how the internals of libA.so work, nor should they care. The solution is to put a DT_NEEDED tag into libA.so to indicate that it needs libB.so. With this tag, the linker command to build exe only needs to specify -lA, and the linker will read the DT_NEEDED tag, and link in libB.so automatically.

This tag is created when libA.so is linked: its build command needs to include -lB. These tags can be queried with the objdump command. For instance:

dima@shorty:/tmp$ objdump -p /usr/lib/x86_64-linux-gnu/libopencv_highgui.so.2.4 | grep NEEDED

  NEEDED               libopencv_imgproc.so.2.4
  NEEDED               libdl.so.2
  NEEDED               libpthread.so.0
  NEEDED               librt.so.1
  NEEDED               libtbb.so.2
  NEEDED               libatomic.so.1
  NEEDED               libz.so.1
  NEEDED               libjpeg.so.62
  NEEDED               libpng16.so.16
  NEEDED               libtiff.so.5
  NEEDED               libImath-2_2.so.12
  NEEDED               libIlmImf-2_2.so.22
  NEEDED               libIex-2_2.so.12
  NEEDED               libHalf.so.12
  NEEDED               libIlmThread-2_2.so.12
  NEEDED               libgtk-x11-2.0.so.0
  NEEDED               libgdk-x11-2.0.so.0
  NEEDED               libpangocairo-1.0.so.0
  NEEDED               libatk-1.0.so.0
  NEEDED               libcairo.so.2
  NEEDED               libgdk_pixbuf-2.0.so.0
  NEEDED               libgio-2.0.so.0
  NEEDED               libpangoft2-1.0.so.0
  NEEDED               libpango-1.0.so.0
  NEEDED               libgobject-2.0.so.0
  NEEDED               libglib-2.0.so.0
  NEEDED               libfontconfig.so.1
  NEEDED               libfreetype.so.6
  NEEDED               libgthread-2.0.so.0
  NEEDED               libdc1394.so.22
  NEEDED               libv4l1.so.0
  NEEDED               libavcodec.so.57
  NEEDED               libavformat.so.57
  NEEDED               libavutil.so.55
  NEEDED               libswscale.so.4
  NEEDED               libopencv_core.so.2.4
  NEEDED               libstdc++.so.6
  NEEDED               libm.so.6
  NEEDED               libgcc_s.so.1
  NEEDED               libc.so.6

So a user building an application that uses libopencv_highgui.so doesn't need to know to link in all this stuff when building their executable.


So I have a bunch of libraries that have no DT_NEEDED tags, and I want to change their build invocations to add those tags. But which tags do I need? There are many libraries, and it would take a lot of work to look through each one and make a list of libraries that it depends on. I wrote this script to do that work for me:



# Invoke this tool thusly:
#     nm -o -D *.so | perl infer_dt_needed.pl

use strict;
use warnings;

use feature ':5.10';

my %undef_syms_in_libs;
my %defining_lib_for_sym;
my %library_dep;

    my ($lib,$type,$sym) = /^(.*?):           # symbol name:
                            (?:[0-9a-fA-F]+)? # possibly an address
                            ([a-zA-Z])        # symbol type
                            (.*?)             # symbol name
      or next;

    $lib =~ s/^lib//;
    $lib =~ s/\.so.*//;

    # read in each symbol, and make a record of each defined and undefined
    # function. I look for very specific symbol types here, and this set is
    # sufficient for my application. Look at the 'nm' manpage for the other
    # possible symbols that may be necessary for other applications
    if( length($type) )
        if($type eq 'T')
            $defining_lib_for_sym{$sym} = $lib;
        elsif($type eq 'U')
            $undef_syms_in_libs{$lib} //= [];
            push @{$undef_syms_in_libs{$lib}}, $sym;

# I go through each undefined symbol, find the library that provides it, and if
# found, record the dependency between those libraries
for my $lib (keys %undef_syms_in_libs)
    for my $sym ( @{$undef_syms_in_libs{$lib}} )
        if ($defining_lib_for_sym{$sym})
            $library_dep{$lib} //= {};
            $library_dep{$lib}{$defining_lib_for_sym{$sym}} = 1;

# print out the dependencies
for my $lib(keys %library_dep)
    say "lib$lib.so:" . join('', map {" -l$_"} keys %{$library_dep{$lib}} );

To use this tool, read the symbol table from all my shared libraries, and feed this list into the tool. The tool finds the connection between libraries that use a particular symbol and libraries that provide it. In the trivial example above, the invocation and output would look like this:

$ nm -o -D libA.so libB.so | perl infer_dt_needed.pl

libA.so: -lB

The nm tool reads the symbol table, and the -o -D options are more or less required for this usage. The tool only looks at a very specific set of symbols (nm types T and U), which is sufficient for my purposes. Read the nm manpage and update the script if you need more.

With many libraries and many symbols this sort of automated bookeeping is invaluable. The resulting list can be used to update the Makefiles to put the appropriate tags in place.