.tsan-suppressions.txt

# ThreadSanitizer suppressions for @photostructure/fs-metadata
#
# DISCIPLINE: suppress only exact external functions that cannot be rebuilt
# with instrumentation.
#
# CRITICAL -- do NOT add `race:node::`, `race:uv_`, `race:uv__`, `race:napi_` or
# `race:Napi::`. TSan's `race:` rules match against ANY frame in the reported
# stacks, and every AsyncWorker race unwinds through
# node::ThreadPoolWork::ScheduleWork -> napi -> FSMeta::...  Such a rule
# therefore silences the addon's OWN data races -- the exact defects this job
# exists to find. This was verified empirically: with `race:node::` present, a
# deliberately injected unguarded `static int` increment in
# LinuxMetadataWorker::Execute() was NOT reported. With the rules below it IS
# reported. Keep it that way.
#
# Format: <suppression_type>:<function pattern>  (substring match)
# https://github.com/google/sanitizers/wiki/ThreadSanitizerSuppressions

# Stock Node 26's uninstrumented V8 background paths produce allocator races.
# Keep these pinned to the exact observed class/function: namespace-wide V8
# rules can mask first-party binding calls whose JavaScript stacks enter V8.
race:v8::internal::CancelableTaskManager::CancelAndWait
race:v8::base::RegionAllocator::Split
race:BaselineBatchCompilerJob
race:v8::internal::compiler::CompilationDependencies
race:v8::internal::Scavenger::
race:v8::internal::ScavengerWeakObjectsProcessor::ProcessJSWeakRefs
race:v8::internal::ScavengerCollector::CollectGarbage
race:v8::internal::PooledPage
race:v8::internal::CompactionSpaceCollection::
race:v8::platform::tracing::TracingController::GetCategoryGroupEnabled
race:absl::container_internal::PrepareInsertSmallNonSoo
race:v8::internal::CodeRangeAddressHint::NotifyFreedCodeRange
race:v8::internal::MutablePage::ReleaseAllocatedMemoryNeededForWritableChunk
race:v8::internal::MutablePage::MutablePage
race:v8::internal::FutexWaitListNode::NotifyWake
race:v8::internal::GlobalBackingStoreRegistry::Purge
race:std::_Hashtable<v8::internal::MemoryChunk
race:v8::internal::NormalPage::ReleaseFreeListCategories
race:cppgc::internal::HeapRegistry::UnregisterHeap

# libuv, likewise compiled into `node` without instrumentation. This is a
# specific false positive between two libuv worker threads at startup: io_uring
# setup (uv__iou_init) reads an fd that uv__slurp/uv__open_cloexec wrote.
#
# NOTE the function-level precision. `race:uv_` or `race:uv__` would ALSO match
# `uv_thread_create_ex`, which appears in the thread-creation stack of every
# libuv-threadpool race -- including ours -- and would silently mask first-party
# findings. Keep these rules pinned to exact function names.
race:uv__iou_init

# Stock Node 26.5.0 can destroy a Worker environment's libuv rwlock while
# TSan still observes the main thread's final unlock. Both reported accesses
# are inside Node's uninstrumented libuv (uv_rwlock_destroy / wrunlock), and
# this addon neither creates nor uses uv_rwlock_t. Pin the suppression to the
# destructive access: unlike `race:uv_`, this cannot match ordinary addon
# worker or thread-creation stacks.
race:uv_rwlock_destroy