cvehound-spatch

A prebuilt Coccinelle spatch, packaged as a Python wheel so that CVEhound works without asking you to install OCaml.

You probably do not want this package on its own. Install it through CVEhound, which picks it up automatically:

pip install 'cvehound[spatch]'

CVEhound still runs happily against a system spatch; this package only removes the need for one. Resolution order is --spatch / config → $CVEHOUND_SPATCH → this package → PATH.

What is in the wheel

Four files inside the cvehound_spatch/ package directory, 20-25 MB:

They sit next to each other on purpose: spatch locates its data files relative to the real path of its own executable, so the directory is relocatable and callers need no --iso-file/--macro-file-builtins arguments.

Wheels are built for Linux x86_64 and aarch64 against glibc 2.28 (manylinux_2_28), and for macOS arm64 against a macOS 11 deployment target (macosx_11_0_arm64). The binary links against nothing but the platform's own libc — glibc on Linux, libSystem on macOS. On any other platform the wheel simply does not install, and CVEhound falls back to PATH. Installing needs pip 20.3 or newer, which is where manylinux_2_28 support landed.

One runtime requirement is not in the wheel: diff (diffutils). Coccinelle renders what a rule matched by shelling out to it, and on a system without it spatch prints an internal error and still exits 0 — a match becomes silence. Every distribution has it; minimal container images sometimes do not.

Using it directly

import cvehound_spatch

cvehound_spatch.spatch_path()      # PosixPath('.../cvehound_spatch/spatch')
cvehound_spatch.COCCINELLE_VERSION # '1.3.3'
cvehound_spatch.COCCINELLE_COMMIT  # the exact commit it was built from

python -m cvehound_spatch --path         # where the binary is
python -m cvehound_spatch --build-info   # what it was built from
python -m cvehound_spatch --version      # anything else is passed to spatch

Choices made

This is not a general-purpose coccinelle build. It is the smallest, fastest build that runs CVEhound's rules, and every deviation from upstream defaults was decided by measurement:

• No Python scripting (--disable-python). Coccinelle's Python support dlopens libpython by heuristic search, which fails on interpreters that ship no shared libpython — a whole class of "works on my machine". CVEhound's rules report by starring lines (* context mode) and need no scripting at all, so the support is dead weight; dropping it also cuts per-invocation startup from ~50 ms to ~9 ms, which matters when a scan runs 500+ invocations. A rule that did use script:python fails loudly here (exit 255), never silently.

• No OCaml scripting (--disable-ocaml). Runtime script:ocaml compilation would require an OCaml toolchain on the user's machine — the opposite of the point of this package.

• No PCRE syntax (--disable-pcre-syntax). No rule uses =~; the Str fallback stays available.

• OCaml 5.3. The 5.x runtime is 20-25 % faster than 4.x on the heavy rules, by far the largest effect found. Verdicts are identical, and coccinelle's own test suite scores the same on both.

• OCaml built --without-zstd (ocaml-option-no-compression). From 5.1 the runtime links libzstd wherever configure finds it, and a wheel cannot ship a Homebrew dylib — on macOS this put /opt/homebrew/opt/zstd straight into otool -L. The manylinux image happens to carry no zstd headers, so the Linux build was getting the same result by luck; asking for it makes the invariant explicit on both. Nothing is given up: compressed marshalling is opt-in, and the compilation artifacts zstd would shrink are not shipped.

• flambda -O3, but no BOLT. Both were built and measured twice. The first study (six heavy rules, OCaml 4.14, one spatch exec per rule) put flambda at 2-3 % and rejected it. Re-measured on the real corpus — 490 rules, each compared against itself, both binary orders — flambda is worth 4.5 % of CPU, and 5.3 % when spatch is run as a fork-per-request server. The earlier number was an artifact of a workload whose cost sat in a few heavy rules; inlining pays across the many small and medium ones. It costs +11.4 MB, which buys +428 minor page faults per exec and 0.05 ms — a real cost, and a negligible one.

  BOLT is worth a further ~1.3 points and is not shipped. It needs a representative workload to train a profile on — a kernel tree and the rule corpus — which this build does not have and would have to grow, and its benefit measures at zero once spatch shares parsed ASTs between rules, which is the direction CVEhound is heading. The pipeline works and the numbers are real (L1i misses −17 %, ITLB −11 %, cycles only −0.7 % because the workload is not frontend-bound), so it is written down rather than built in; if that ever changes it is there to collect. (OCaml has neither LTO nor usable PGO, so there is nothing else to try.)

• Bounds checks kept. Coccinelle builds with -unsafe by default; removing the checks is worth about 1 % here, which is not a good trade against running a parser over untrusted sources.

• Patches on top of 1.3.3. One is a performance regression, submitted upstream and carried here until it is in a release: an atoms-only fallback when the file prefilter's CNF conversion hits max_cnf. Giving up there returned "no query", which worth_trying reads as "try every file", so the semantic patch ran over the whole tree with no prefiltering at all; it now degrades to a one-clause query over the atoms, which is still sound because the formula is negation-free — coccinelle#420.

  Four more change how spatch can be driven. The first three are the named capabilities this build's FEATURES advertises; the fourth is invisible to callers and needs no flag:

  • --zygote, a fork-per-request server mode. It warms standard.iso and standard.h once and then forks per request, so a scan of many rules pays process startup once instead of per rule. Every request still runs in its own process, so no engine global survives from one rule to the next.
  • exit 124 on an engine timeout. A fired --timeout used to escape as an uncaught Common.Timeout, i.e. exit 2, indistinguishable from a crash; it now exits 124 the way timeout(1) does, keeping the exception name in the message for callers that already scrape it.
  • A shareable AST cache. --cache-prefix entries are written through a temporary name and renamed, value before dependencies, so parallel scans sharing one cache cannot read a torn entry; standard.h is now part of the cache key.
  • Cheaper per-file overhead: the standard.iso parse and standard.h extraction are memoised, and the cache path no longer forks /bin/sh to run mkdir -p once per file — which a cache hit was also paying.

Provenance and reproducing a build

Sources come from the cvehound branch of the coccinelle fork — coccinelle 1.3.3 plus the patches above, nothing else. Every wheel records the exact commit in BUILD-INFO and in COCCINELLE_COMMIT, and the capabilities it was probed to have in FEATURES:

>>> import cvehound_spatch
>>> cvehound_spatch.FEATURES
frozenset({'exit124', 'shared-cache', 'zygote'})

FEATURES exists because these capabilities cannot be detected from outside: --zygote is dispatched before spatch parses its arguments, so it appears in no --help output, and the version banner is identical with and without it. make_wheel.py probes the binary at pack time rather than asserting, so a build from a ref that predates a feature advertises nothing instead of lying, and a consumer reading getattr(cvehound_spatch, 'FEATURES', frozenset()) degrades correctly against any older wheel.

To rebuild locally on Linux (needs podman or docker):

./build-in-container.sh          # -> dist/bundle-<arch>/
./make_wheel.py --bundle dist/bundle-$(uname -m)

On macOS there is no container to be portable against, so build.sh runs directly on the host and installs what it needs through Homebrew:

./build.sh                       # -> dist/bundle-<arch>/
./make_wheel.py --bundle dist/bundle-$(uname -m)

build.sh is the build on both — inside the manylinux_2_28 image on Linux, on the host on macOS — and it fails the build if the binary picks up a dependency beyond the platform's libc or turns out to have Python support. On Linux it also rejects a binary needing glibc newer than 2.28; on macOS it rejects one whose LC_BUILD_VERSION disagrees with the deployment target the wheel tag promises, or whose ad-hoc signature is not valid.

Verify a built wheel with python tests/smoke.py after installing it — the same script CI runs, including a parse check over every CVEhound rule.

Versioning

The package version is the coccinelle version it contains: 1.3.3. Packaging changes and rebuilds that keep the same coccinelle release bump a post-release segment (1.3.3.post1), and a new coccinelle release gives a new version.

License

The wheel ships coccinelle's binary, so it is distributed under the GPL-2.0 (see LICENSE), the same license as coccinelle itself. The packaging scripts in this repository are under the same terms. CVEhound itself is GPL-3.0 and simply executes this binary as a separate program.