os/linux/elf: avoid using ldd for listing dynamic dependencies

[alebcay]

• Have you followed the guidelines in our Contributing document?
• Have you checked to ensure there aren't other open Pull Requests for the same change?
• Have you added an explanation of what your changes do and why you'd like us to include them?
• Have you written new tests for your changes? Here's an example.
• Have you successfully run brew style with your changes locally?
• Have you successfully run brew typecheck with your changes locally?
• Have you successfully run brew tests with your changes locally?

---

Current approach is to query ldd and capture its output (which includes resolved paths to shared libraries), then filters based on the needed libraries determined through checking the dynamic section.

New approach is to search for each needed library listed in the dynamic section, trying to locate it via runpath/rpath and the system ld.so.conf paths.

With the new approach, we avoid passing (potentially untrusted) files to ldd. Instead, we analyze the ELF file statically (inspect the dynamic section) and only try to perform lookups for the shared libraries indicated in the dynamic section.

This change is intended to modify how the dynamic dependencies lookup happens, but should not alter any internal or public-facing API (only private methods/classes changed) - for instance, the output semantics, format, and type of Pathname("/some/path/to/an/executable").dynamically_linked_libraries should remain unchanged.

[Bo98]

Looks good!

Just as a couple questions:

• Have you tested this on a few formulae?
• What's the performance difference? Faster/slower and by how much?

[alebcay]

So far I've spot-tested a few binaries/libraries, like zsh binary from the zsh formula and libleptonica.so.6 from leptonica, to confirm that the output of Pathname(<path to ELF file>).dynamically_linked_libraries is the same. Open to any ideas on how to test further, especially for covering the case where DF_1_NODEFLIB is set.

As far as performance, the performance seems to be about the same. Testing with the following:

def snippet()= 100.times { Pathname("/home/linuxbrew/.linuxbrew/Cellar/leptonica/1.84.1/lib/libleptonica.so.6").dynamically_linked_libraries }
Benchmark.measure { snippet }

Before:

=> #<Benchmark::Tms:0x00000040084902a8
 @cstime=1.1855609999999928,
 @cutime=7.002927,
 @label="",
 @real=13.041561631999684,
 @stime=0.21336699999999809,
 @total=11.345018999999986,
 @utime=2.943163999999996>

After:

=> #<Benchmark::Tms:0x0000004008601da8
@cstime=1.219255,
@cutime=7.094854,
@label="",
@real=13.061064489999808,
@stime=0.18731199999999998,
@total=11.323314,
@utime=2.821893>

The above "after" result is with no caching mechanism implemented.

As mentioned earlier, there is possibly room for caching-related improvements here:

• Cache the output of ld.so --list-diagnostics
  • Per brew invocation, i.e. memoize (don't persist to disk)
  • Persistent cache - need to figure out what is the right interval for invalidation (this feels a bit overkill)
• Read/parse/use ld.so.cache before trying to check library paths (and memoize/cache this)

[MikeMcQuaid]

Nice work so far, a few suggested tweaks, feel free to push back on any.

Changes addressed

[MikeMcQuaid]

Looks good, thanks again @alebcay!

[carlocab]

Might need a revert. Homebrew/homebrew-core#169088

On mobile at the moment, so would appreciate if someone else got to it.