Some distributions (e.g. Fedora) ship only a shared RocksDB that exports
RocksDB::rocksdb-shared rather than RocksDB::rocksdb. The CMake target
selection now falls back to the shared target when the static one is absent.
Newer RocksDB also changed DB::Open and DB::OpenForReadOnly to take
std::unique_ptr<DB>* instead of DB**. IfcFile.cpp uses SFINAE tag dispatch
to build against both old and new APIs without version detection.
Fixes builds with newer GCC/libstdc++ that no longer provide <cstdint>,
<cstring>, <cfloat>, <memory>, <algorithm> etc. transitively. Also
disambiguates visit<> calls in taxonomy.h with the full namespace and
casts the character value in IfcCharacterDecoder to uint32_t to silence
ambiguous overload warnings.
Some distributions (e.g. Fedora) ship only a shared RocksDB that exports
RocksDB::rocksdb-shared rather than RocksDB::rocksdb. The CMake target
selection now falls back to the shared target when the static one is absent.
Newer RocksDB also changed DB::Open and DB::OpenForReadOnly to take
std::unique_ptr<DB>* instead of DB**. IfcFile.cpp uses SFINAE tag dispatch
to build against both old and new APIs without version detection.
Fixes builds with newer GCC/libstdc++ that no longer provide <cstdint>,
<cstring>, <cfloat>, <memory>, <algorithm> etc. transitively. Also
disambiguates visit<> calls in taxonomy.h with the full namespace and
casts the character value in IfcCharacterDecoder to uint32_t to silence
ambiguous overload warnings.
# What
Two upstream-shaped fixes to ifcopenshell's runtime plug-in discovery
so the schema/serializer plug-ins are findable in deployment layouts
other than a flat \`<prefix>/lib/\` (specifically: macOS .app bundles).
## (1) Stable anchor variable instead of a function pointer
\`schema_plugin_directory()\` used \`&load_schema_plugins\` as the
anchor whose containing module \`dladdr\` is asked to resolve. Function
addresses are not reliably equal to a single canonical location across
toolchains — on macOS arm64 with BonsaiViewer.app, \`&load_schema_plugins\`
took the address of a PLT/stub inside the consumer binary rather than
the actual symbol inside \`libIfcParse.dylib\`. \`dladdr\` then dutifully
returned the consumer's path and the loader started searching
\`BonsaiViewer.app/Contents/MacOS/\` for plug-ins that were never
installed there.
A variable doesn't suffer from this — it has exactly one canonical
address inside its defining dylib. Add \`ifcopenshell_libifcparse_anchor\`
(exported via IFC_PARSE_API) and use \`&that\` instead. Standard pattern
used by Boost.DLL, GStreamer, \`_dyld_get_image_*\`, etc.
## (2) Bundle-aware fallback search paths
The primary search path is \`dirname(libIfcParse)\`. That works for
flat installs (Linux \`lib/\`, Windows \`bin/\`) where plug-ins are
siblings of libIfcParse. macOS app bundles split the layout:
\`macdeployqt\` puts non-Qt @rpath deps in \`Contents/Frameworks/\`,
Apple convention asks for \`Contents/PlugIns/\`, and some install rules
co-locate libIfcParse with the exe in \`Contents/MacOS/\`. Plug-ins
typically end up in a sibling directory, not the same one.
In \`add_search_paths_or_default\`, after registering the primary path,
also register \`<parent>/PlugIns\`, \`<parent>/Frameworks\`, and
\`<parent>/MacOS\` on Apple platforms. \`discover_exact\` short-circuits
on the first hit so duplicates and missing directories are harmless.
# What this does NOT do
The plug-in dylibs still need to actually be inside the app bundle
somewhere for these fallbacks to find them — the upstream install
rules (\`install(TARGETS …)\` in \`src/ifcparse/CMakeLists.txt\`,
\`src/serializers/CMakeLists.txt\`, etc.) put them in \`<prefix>/lib/\`
which lives outside \`BonsaiViewer.app\`. That side of the fix is a
follow-up — either an explicit bundle-aware install destination on
the plug-in targets, or an install(CODE) sweep that mirrors them
into the bundle.
# What this also un-does
Reverts the BonsaiViewer-only \`install(CODE)\` hack that was about to
copy \`ifcopenshell.*.dylib\` from \`lib/\` into
\`BonsaiViewer.app/Contents/MacOS/\` — superseded by the loader-side
fix above, which lets us put the plug-ins anywhere sane inside the
bundle without further consumer-side stitching.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Loading a federated project (.ifcfed) with several models segfaults
non-deterministically on a fresh start. The viewer's SceneLoader spawns
one background std::thread per model in startDataSourceLoad() to
construct an ifcopenshell::file; with cached sidecars all models reach
that point near-simultaneously, so multiple threads parse different IFC
files at once. Parsing touches the process-wide schema singleton, which
was not thread-safe in two places.
Race 1 — concurrent schema population
-------------------------------------
schema_registry::get() lazily runs the schema's get_() function (e.g.
Ifc4::get_schema() -> IFC4_populate_schema()) and mutates entries_ with
no lock. Two threads calling schema_by_name("IFC4") at once both run
IFC4_populate_schema() concurrently, which fills global arrays
(IFC4_types[], strings[]). One thread reads a slot the other is still
writing.
Core-dump evidence (gdb thread apply all bt):
Thread 1 SIGSEGV in IFC4_populate_schema Ifc4-schema.cpp:1989
<- Ifc4::get_schema
<- schema_registry::get schema.cpp:241
<- schema_by_name("IFC4")
<- ifcopenshell::file::file (NWCH-PIR-SS...ifc)
<- SceneLoader::startDataSourceLoad lambda SceneLoader.cpp:315
Thread 3 also in IFC4_populate_schema (entity ctor for
"IfcMaterialProfileSetUsageTapering")
<- Ifc4::get_schema
<- schema_registry::get schema.cpp:241
<- ifcopenshell::file::file (NWCH-PIR-PT...ifc)
<- SceneLoader::startDataSourceLoad lambda
Two threads inside IFC4_populate_schema() at the same time is the race.
Fix: guard schema_registry's bind()/get()/names()/clear() with a
recursive_mutex (recursive because get() re-enters bind() via
load_schema_plugin(), and a freshly populated schema registers itself
through register_schema()). get() is serialized, so only the first
thread populates the schema; the rest block briefly and then observe
the finished result. Returned schema pointers are stable for the
process lifetime, so holding the lock only across get() is sufficient.
Race 2 — lazy all_attributes_ cache filled during parsing
---------------------------------------------------------
entity::all_attributes() lazily fills a `mutable` optional cache on the
shared schema entity the first time it is accessed — and that first
access happens during parsing (parse_context::construct), not during
schema population. With race 1 fixed, two parser threads still raced
here: both saw the cache empty, both did all_attributes_.emplace() and
std::copy() into it, corrupting the vector.
Core-dump evidence after the race-1 fix:
Thread 1 SIGSEGV in attribute::type_of_attribute (this=0xe130...55c)
<- std::transform(first=0x4, last=0xb0d1...) <-- garbage
iterators into a corrupt std::vector
<- parse_context::construct over
decl->as_entity()->all_attributes() file.cpp:249
<- instance_streamer::read_instance
<- ifcopenshell::file::file (NWCH-PIR-PT...ifc)
<- SceneLoader::startDataSourceLoad lambda
The begin pointer 0x4 is a half-written vector being read mid-resize by
another thread.
Fix: force every entity's all_attributes_ cache in the
schema_definition constructor, while construction is still
single-threaded. The schema is then genuinely immutable after
construction, so concurrent parsing needs no hot-path lock.
Both crashes reproduce reliably on a fresh start at native speed but
vanish under gdb (which serializes thread scheduling) — the classic
signature of a data race. With both fixes the federated load completes
cleanly.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Read-only handles reject Flush/CompactRange, so the destructor's status
assertion always fired on shutdown when the streamer's sidecar was
opened with read_only=true. Track the flag and skip the write path; also
guard against a null db when the initial open failed.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Thomas Krijnen