A GPU plugin replaces a VM's state transitions wholesale. Three of the four
seams reached one by dlsym'ing launcher names out of a library nothing had
identified, and all four would fall back to a bare filename — which makes the
dynamic loader's default search path, and in bridgevm's case the working
directory of whoever started the node, the thing that decides which code
computes consensus state.
A library exporting the five bridgevm launcher names and nothing else,
compiled from twenty lines of C that return 0:
before, LUX_GPU_PLUGIN_DIR=/tmp/rogue AutoBackend = hip
before, no env, cwd holds the .so AutoBackend = hip
after, either AutoBackend = none
The plugin header says a loader must check abi_version and vtbl_size. aivm,
bridgevm and mpcvm now do so before resolving a single launcher: a name was
the only thing being matched, and a name is not a contract.
Bare filenames and the CWD entry are gone. What is left is LUX_GPU_PLUGIN_DIR
and LUXCPP_PREFIX — places an operator named. Unset means no plugin, which is
the configuration every box without a GPU is already in.
Cherry-picked from 634df5e, whose quantumvm hunk is dropped: that seam's file
was deleted in 409d0e4 and the chain no longer loads a plugin.
Co-authored-by: Hanzo Dev <[email protected]>
A CMP group with no shares sums to the identity point, which marshals to a
well-formed thirty-three bytes. Handed back as the custody key it made an
empty config look like a finished keygen: the node reported itself ready to
attest and bridge_getInfo agreed, on a chain that could verify nothing.
The rest is subtraction. A settlement record now holds only the fact that
the transfer is settled, because the block that settled it is already named
by the chain. movedToday reports what it read or why it could not, instead
of reporting zero. RemoveSigner returns what happened rather than an error
it never had. Marshal cannot fail, so Bytes has no failure to handle. The
gorilla-rpc registration was a second way to serve one API. The route table
NewHTTPHandler built was a second copy of the one CreateHandlers names. The
logger is set at Initialize, so guarding every log call against its absence
was guarding against ourselves.
How often the watcher polls is data on the watcher, as the releaser's
backoff already is.
OSS public chains module must not reference the private GPU plugin
implementation by repo name or filesystem path. Two leak types fixed:
1. Hard compile-time dependency in quantumvm:
`#cgo CFLAGS: -I${SRCDIR}/../../../lux-private/gpu-kernels/include`
Vendor the public ABI v14 header (backend_plugin.h + crypto.h, 1475
LOC total) into chains/internal/luxgpu/include/lux/gpu/ and point
CFLAGS at the in-tree path. Plugin DSO is still dlopen'd at runtime
— only the public C ABI surface is vendored.
2. Soft leakage — `LUX_PRIVATE_GPU_KERNELS_DIR` env vocabulary, dev-tree
probe paths to ~/work/lux-private/gpu-kernels/build/..., prose
mentions of "lux-private/gpu-kernels" in package comments.
All replaced with `LUX_GPU_PLUGIN_DIR` (universal env) and generic
"GPU plugin" prose.
Bridges affected: aivm, bridgevm, dexvm, quantumvm, thresholdvm.
All 139 tests PASS post-scrub under both build tags. No functional
behavior change — only path / env / prose.
Mirrors the evm/cevm pattern (cevm_cgo.go / cevm_nocgo.go +
backend_cgo.go). Three new files in bridgevm/:
backend.go — Backend enum, layout structs (byte-equal to
ops/bridgevm/cuda/bridgevm_kernels_common.cuh),
ErrGPUNotAvailable, GPUBackend interface,
unsafe.Sizeof asserts in init().
bridgevm_gpu.go cgo — direct dlopen(3) + dlsym(3) for the per-backend
plugin libluxgpu_backend_<bk>.{so,dylib}.
Probes cuda → hip → metal → vulkan → webgpu in
order; first plugin exposing all five
lux_<bk>_bridgevm_* launchers wins.
bridgevm_gpu_nocgo.go — stub returning ErrGPUNotAvailable on every method.
Direct dlopen (NOT pkg-config) so the public chains module builds without
the private gpu-kernels repo present. Search path: \$LUX_GPU_PLUGIN_DIR,
\$LUXCPP_PREFIX/lib/lux-gpu, \$LUXCPP_PREFIX/lib, the dev tree under
~/work/lux-private/gpu-kernels/build/backends/<bk>, CWD, and finally the
system loader's default.
Five launcher signatures dispatched via cgo trampolines (one per shape) —
matches the C ABI in include/lux/gpu/bridgevm.h. Layout structs include
explicit '_pad*' fields to keep Go sizeof() byte-equal to the C++
alignas(16) layout; the init() sizeof asserts catch any drift at process
start with a per-struct panic message.
vm.go / bridge_signing.go / mpc.go are NOT touched — this is a thin
substrate; the round applier opts in later when it's ready to route
through ActiveGPUBackend().
bridgevm_gpu_test.go — four tests under //go:build cgo:
TestActiveBackend — AutoBackend() returns a known tag.
TestStubReturnsErrGPUNotAvailable
— BackendNone path returns the sentinel.
TestLayoutSizesMatchHeader — runtime-equivalent of the C++ static_asserts.
TestSignerApplyZeroFixture — dlopen + dlsym + zero-fixture round trip
through lux_<best>_bridgevm_signer_apply.
Builds verified under both CGO_ENABLED=0 (nocgo path active) and
CGO_ENABLED=1 -tags cgo (cgo path active). All four cgo-tagged tests
pass on the local dev box (Vulkan plugin auto-detected from the
gpu-kernels build tree).
zeekay
