package.bbclass: inject "minidebuginfo" into packaged binaries

"Mini debuginfo" is a special section in ELF executables containing
minimal compressed debuginfo for non-exported symbols:

https://sourceware.org/gdb/onlinedocs/gdb/MiniDebugInfo.html

It lets debugging tools produce better stack traces, including local
function names, without incurring the space overhead of full debuginfo.
The feature was originally developed for Fedora:

https://fedoraproject.org/wiki/Features/MiniDebugInfo

but nowadays it is widely supported in the ecosystem, including in gdb
and elfutils (and therefore also in tools which use elfutils, such as
systemd-coredump).

This patch adds an optional extra step in package.bbclass to inject
minidebuginfo while stripping and splitting out debuginfo. It can be
enabled by setting PACKAGE_MINIDEBUGINFO=1. In my testing, this
increases the size of resulting binaries by roughly 5%.

The code for producing and re-injecting the minidebuginfo is my own
Python implementation but corresponds directly to the shell
implementation that RPM uses for doing the same:

https://github.com/rpm-software-management/rpm/blob/rpm-4.15.1-release/scripts/find-debuginfo.sh#L261

(From OE-Core rev: 4df992ce50c2d12e356b6d9fe7b23a6320c8b4df)

Signed-off-by: Dan Callaghan <dan.callaghan@opengear.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>

meta/classes/package.bbclass

@@ -245,6 +245,8 @@ python () {
         deps = ""
         for dep in (d.getVar('PACKAGE_DEPENDS') or "").split():
             deps += " %s:do_populate_sysroot" % dep
+        if d.getVar('PACKAGE_MINIDEBUGINFO') == '1':
+            deps += ' xz-native:do_populate_sysroot'
         d.appendVarFlag('do_package', 'depends', deps)
 
         # shlibs requires any DEPENDS to have already packaged for the *.list files
@@ -459,6 +461,83 @@ def splitstaticdebuginfo(file, dvar, debugstaticdir, debugstaticlibdir, debugsta
 
     return (file, sources)
 
+def inject_minidebuginfo(file, dvar, debugdir, debuglibdir, debugappend, debugsrcdir, d):
+    # Extract just the symbols from debuginfo into minidebuginfo,
+    # compress it with xz and inject it back into the binary in a .gnu_debugdata section.
+    # https://sourceware.org/gdb/onlinedocs/gdb/MiniDebugInfo.html
+
+    import subprocess
+
+    readelf = d.getVar('READELF')
+    nm = d.getVar('NM')
+    objcopy = d.getVar('OBJCOPY')
+
+    minidebuginfodir = d.expand('${WORKDIR}/minidebuginfo')
+
+    src = file[len(dvar):]
+    dest = debuglibdir + os.path.dirname(src) + debugdir + "/" + os.path.basename(src) + debugappend
+    debugfile = dvar + dest
+    minidebugfile = minidebuginfodir + src + '.minidebug'
+    bb.utils.mkdirhier(os.path.dirname(minidebugfile))
+
+    # If we didn't produce debuginfo for any reason, we can't produce minidebuginfo either
+    # so skip it.
+    if not os.path.exists(debugfile):
+        bb.debug(1, 'ELF file {} has no debuginfo, skipping minidebuginfo injection'.format(file))
+        return
+
+    # Find non-allocated PROGBITS, NOTE, and NOBITS sections in the debuginfo.
+    # We will exclude all of these from minidebuginfo to save space.
+    remove_section_names = []
+    for line in subprocess.check_output([readelf, '-W', '-S', debugfile], universal_newlines=True).splitlines():
+        fields = line.split()
+        if len(fields) < 8:
+            continue
+        name = fields[0]
+        type = fields[1]
+        flags = fields[7]
+        # .debug_ sections will be removed by objcopy -S so no need to explicitly remove them
+        if name.startswith('.debug_'):
+            continue
+        if 'A' not in flags and type in ['PROGBITS', 'NOTE', 'NOBITS']:
+            remove_section_names.append(name)
+
+    # List dynamic symbols in the binary. We can exclude these from minidebuginfo
+    # because they are always present in the binary.
+    dynsyms = set()
+    for line in subprocess.check_output([nm, '-D', file, '--format=posix', '--defined-only'], universal_newlines=True).splitlines():
+        dynsyms.add(line.split()[0])
+
+    # Find all function symbols from debuginfo which aren't in the dynamic symbols table.
+    # These are the ones we want to keep in minidebuginfo.
+    keep_symbols_file = minidebugfile + '.symlist'
+    found_any_symbols = False
+    with open(keep_symbols_file, 'w') as f:
+        for line in subprocess.check_output([nm, debugfile, '--format=sysv', '--defined-only'], universal_newlines=True).splitlines():
+            fields = line.split('|')
+            if len(fields) < 7:
+                continue
+            name = fields[0].strip()
+            type = fields[3].strip()
+            if type == 'FUNC' and name not in dynsyms:
+                f.write('{}\n'.format(name))
+                found_any_symbols = True
+
+    if not found_any_symbols:
+        bb.debug(1, 'ELF file {} contains no symbols, skipping minidebuginfo injection'.format(file))
+        return
+
+    bb.utils.remove(minidebugfile)
+    bb.utils.remove(minidebugfile + '.xz')
+
+    subprocess.check_call([objcopy, '-S'] +
+                          ['--remove-section={}'.format(s) for s in remove_section_names] +
+                          ['--keep-symbols={}'.format(keep_symbols_file), debugfile, minidebugfile])
+
+    subprocess.check_call(['xz', '--keep', minidebugfile])
+
+    subprocess.check_call([objcopy, '--add-section', '.gnu_debugdata={}.xz'.format(minidebugfile), file])
+
 def copydebugsources(debugsrcdir, sources, d):
     # The debug src information written out to sourcefile is further processed
     # and copied to the destination here.
@@ -1185,6 +1264,11 @@ python split_and_strip_files () {
 
         oe.utils.multiprocess_launch(oe.package.runstrip, sfiles, d)
 
+    # Build "minidebuginfo" and reinject it back into the stripped binaries
+    if d.getVar('PACKAGE_MINIDEBUGINFO') == '1':
+        oe.utils.multiprocess_launch(inject_minidebuginfo, list(elffiles), d,
+                                     extraargs=(dvar, debugdir, debuglibdir, debugappend, debugsrcdir, d))
+
     #
     # End of strip
     #