Web lists-archives.com

Autodetection of packages that need visibility annotations

Hi all,

Linux shared libraries by default allow for runtime interposition of
symbols. This comes in handy for customization of malloc but might
also have undesired consequences. Firstly, it badly affects
performance of generated code as
1) compiler is not allowed to optimize (inline/clone/split) functions
in fear that they will be interposed
2) calls to intra-library functions have to go through PLT
3) ld.so symbol resolution takes longer time (due to larger dynamic
symbol tables)
Also interposition hurts code correctness as it increases the chance of
1) inadvertent symbol collisions (see Flameeys'
https://flameeyes.blog/2013/02/23/redundant-symbols/ and other related
posts on his blog)
2) clients relying on private symbols (and later breaking and
complaining when such symbols are removed)

Recommended way to fix interposition is to build libraries with
`-fvisibility=hidden` and use [GNU visibility
annotations](https://gcc.gnu.org/wiki/Visibility) to mark public
symbols. In effort to promote a wider use of those in open-source
projects, I've recently made a simple tool to help locate Debian
packages which would benefit the most from such annotations
(https://github.com/yugr/ShlibVisibilityChecker). The tool (based on
libclang) compares symbols exported from shlibs with those declared in
package headers and reports the difference i.e. exported symbols which
aren't present in headers. The idea is that such symbols are likely to
be internal library symbols and thus should be marked as hidden.

Here's an example usage for libacl1 package:

# # Needs to be run in chroot
# make clean all
# scripts/debiancheck libacl1
Binary symbols not in public interface of acl:
For a total of 14 (25%).

Tool, albeit heuristic and imprecise, turned out to be useful - it
could successfully process ~60% of packages I tried (main reason for
failures were missing #includes in packages' public headers which are
arguably package bugs). Some resulting upstream bugs are in

The main reason for writing this mail is that albeit I'd be happy to
do more visibility patching myself, Debian has 50K packages and it
would be infeasible for me, both in terms of computing power and
personal time to handle this task on my own (not to mention that
upstreams may be hesitant to take patches from unknown contributors).
Can I suggest that interested maintainers try the tool on packages
that they maintain and add visibility annotations if they turn out to
provide noticeable savings (or at least bug upstream projects to
enable visibility themselves)? I'd obviously be happy to fix any
arising bugs or add missing features to the tool.

-Yury Gribov