fix(link): statically link BLAS/LAPACK/gfortran into the engine binary#20
Merged
fix(link): statically link BLAS/LAPACK/gfortran into the engine binary#20
Conversation
ccomb
force-pushed
the
fix/static-link-blas-lapack
branch
3 times, most recently
from
581aae5 to
2394050
Compare
Published Linux engine dynamic-links liblapack.so.3, libblas.so.3, libgfortran.so.5, libquadmath.so.0, libopenblas.so.0 — minimal cloud images don't ship those, so install.sh succeeds but `volca --version` then fails with "cannot open shared object file". Switch the `static` LINK_MODE to a real static link: - `executable-static: True` + `shared: False` make cabal build every dep as `.a` and ask GHC for `-static`. With no intermediate libHS*.so, non-PIC libgfortran.a never gets pulled into a shared object — that's what tripped R_X86_64_TPOFF32 in earlier attempts. - `-no-pie` keeps Debian's non-PIC system Fortran/BLAS archives linkable into the final non-PIE executable. - A single `--start-group`/`--end-group` resolves the cyclic refs between MUMPS / LAPACK / BLAS / libgfortran / libquadmath. cbits/static-shims.c provides __xmknod / __xmknodat: the unix-2.8.6.0 archive shipped with ghcup's GHC 9.6 was compiled against pre-2.32 glibc headers that inlined mknod()/mknodat() to those internal symbols, removed from the public ABI in glibc 2.32. Verified: `file` → "statically linked"; `ldd` → "not a dynamic executable"; `volca --version` runs in debian:stable-slim and alpine:latest containers with no system BLAS/LAPACK/Fortran present.
Replace `throwTo mainTid ExitSuccess` in /api/v1/shutdown and the
idle-watchdog with a direct `_exit(0)` via FFI. The throwTo path
exits the program by unwinding to main and letting the RTS run
its full shutdown sequence (pthread_cancel of capability worker
threads, which glibc resolves via `dlopen("libgcc_s.so.1")` for
the unwinder). On a fully-static glibc build that dlopen returns
NULL — the shutdown path then SIGILLs on the missing unwinder
symbol. Reproduces on linux-arm64; linux-amd64 mostly gets away
with it. _exit(0) is async-signal-safe and bypasses the entire
RTS cleanup.
Drop the now-unused `mainTid` parameter from shutdownEndpoint and
idleWatchdog. hFlush stdout/stderr before _exit so buffered log
lines (incl. the "Shutdown requested" / "Idle for Ns" messages)
still reach the user.
ccomb
force-pushed
the
fix/static-link-blas-lapack
branch
from
2394050 to
2f428f6
Compare
The build-mode upload glob `…/build/volca/volca*` also matched the sibling `volca-tmp/` directory (cabal's intermediate build dir), so PR preview artefacts shipped Main.hi and Main.o alongside the binary. Spell out the two valid file names (`volca` on Posix, `volca.exe` on Windows) instead of trusting the trailing `*`. Release-mode tarballs are unaffected — they're built from a staging dir that only contains the binary.
2 tasks
ccomb
added a commit
that referenced
this pull request
May 3, 2026
## Summary - `docker/Dockerfile` failed at the `gen-cabal-config.sh` step with `cc1: fatal error: /build/volca/cbits/static-shims.c: No such file or directory`. - Root cause: PR #20 made `gen-cabal-config.sh` compile `cbits/static-shims.c` to a `.o` consumed by the static link, but the Dockerfile only copies `volca.cabal`, `mumps-hs/` and `gen-cabal-config.sh` into the builder before invoking the script. - Fix: add `COPY cbits/ /build/volca/cbits/` next to the other dep-spec copies, so the shim source is present when the script runs and layer caching only invalidates when shim sources change. ## Test plan - [x] `./volca/docker-build.sh` (against this branch) progresses past step 11 — confirmed locally up to `cabal build --only-dependencies volca`. - [ ] Full image build completes and `docker run volca --version` works.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Companion to #19 (warn-on-missing-runtime-libs) — once this lands, the
warning becomes mostly cosmetic for the install.sh path: a fresh
Debian-slim container can run `volca --version` without first
`apt install`ing anything.
Why
The existing `static` LINK_MODE only statics MUMPS. BLAS, LAPACK,
libgfortran and libquadmath were placed after `-Wl,-Bdynamic`,
so they stayed dynamic. End-users hit
`liblapack.so.3: cannot open shared object file` on minimal Linux
images (Debian-slim containers, Alpine, Fedora cloud) that don't
ship the numeric stack.
Fix
Move BLAS/LAPACK/gfortran/quadmath inside the same
`--start-group/--end-group` as the MUMPS libs. The linker iterates
until the cross-references settle — necessary because
libgfortran ↔ libgcc, libopenblas/lapack ↔ libgfortran, and MUMPS
↔ BLAS/LAPACK all need each other's symbols.
Only libc/libm/libpthread/libdl stay dynamic — system ABI, not worth
carrying twice.
`darwin` and `windows` modes already do their own thing (ld64
fallback for darwin; MinGW for windows). Linux is the only
LINK_MODE this PR touches.
Tradeoffs
bigger, not a doubling.
ship a new release instead. Acceptable for now (no production
consumers yet).
OpenBLAS + LAPACK are BSD-3 — all compatible with our Apache-2.0.
Test plan
than before.
`./install.sh && volca --version` prints the version, no
loader errors. (The exact reproducer that motivated fix(install): warn when BLAS/LAPACK runtime libs are missing #19.)
end-to-end (`volca compute` against a small fixture).
external deps.