u

Author: shulaoda

libmimalloc-sys/build.rs

@@ -6,40 +6,46 @@ use cmake::Config;
 /// `MI_TLS_MODEL_THREAD_LOCAL` with `MI_TLS_RECURSE_GUARD` instead of the
 /// upstream default `MI_TLS_MODEL_FIXED_SLOT`.
 ///
-/// Background:
-/// - FIXED_SLOT (default) hardcodes the per-thread `theap` pointer at
-///   TCB[108]/[109]. When multiple statically-linked instances of
-///   mimalloc-safe live in the same process (e.g. multiple Node.js
-///   napi addons), every instance reads and writes the same slot on
-///   any thread that calls into more than one of them (the main JS
-///   thread in particular) — instances overwrite each other's `theap`
-///   pointers and corrupt each other's heaps.
-/// - The other obvious workaround, `MI_TLS_MODEL_DYNAMIC_PTHREADS`
-///   (activated by passing `-DMI_HAS_TLS_SLOT=0`), turned out to have
-///   poor macOS testing for long-lived non-tokio threads — notably
-///   rayon workers pulled in via `oxc_cfg → oxc_index → rayon` — and
-///   caused SIGABRT crashes in rolldown's CLI tests.
+/// This fixes two distinct SIGABRT failure modes on macOS:
 ///
-/// THREAD_LOCAL + RECURSE_GUARD avoids both failure modes:
-/// - `__thread` + `mi_decl_hidden` (already how upstream declares
-///   `__mi_theap_default` / `__mi_theap_cached` in `init.c`) gives
-///   per-image per-thread storage allocated by dyld's TLV system — no
-///   FIXED_SLOT sharing across instances.
-/// - RECURSE_GUARD uses a non-TLS `_mi_process_is_initialized` bool to
-///   short-circuit the fast path before process init completes, so a
-///   first `__thread` access triggering dyld's TLV malloc cannot
-///   recurse into mimalloc itself.
-/// - `MI_HAS_TLS_SLOT` stays at default 1, so the *thread-id* fast
-///   path (`prim.h` line ~304) still uses `mi_prim_tls_slot(0)`
-///   (Apple's system-defined thread-id TSD slot, shared semantically
-///   across all consumers — no conflict).
+/// Mode A — FIXED_SLOT (upstream default, 100% crash):
+///   TCB[108]/[109] is a hardcoded slot shared by all loaded images.
+///   When a second napi addon loads, its mimalloc reads the first
+///   addon's heap pointer from TCB[108], sees it's non-NULL, and
+///   skips its own initialization (`init.c:717` early-return). The
+///   second addon then runs with the first's heap data but its own
+///   uninitialized internal state — the mismatch causes an immediate
+///   SIGABRT on load.
+///
+/// Mode B — DYNAMIC_PTHREADS (#67's workaround, 5-15% crash):
+///   `-DMI_HAS_TLS_SLOT=0` routes Apple to DYNAMIC_PTHREADS, which
+///   stores heap pointers via `pthread_setspecific`. This fixes Mode A,
+///   but at process exit the destructor (`init.c:1190`) deletes the
+///   pthread key and zeros it. The one-time `mi_atomic_do_once` flag
+///   that guards key creation never resets, so the key can never be
+///   recreated. Background threads (rayon workers from `oxc_cfg →
+///   oxc_index → rayon`) are still alive during exit — their next
+///   alloc finds key=0, allocates a heap but can't store it, and
+///   returns NULL → SIGABRT.
+///
+/// THREAD_LOCAL + RECURSE_GUARD avoids both:
+/// - `__thread` + `mi_decl_hidden` gives per-image per-thread storage
+///   via dyld TLV. Different addons have different variables (fixes
+///   Mode A). No pthread key involved, so nothing to delete on exit
+///   (fixes Mode B).
+/// - RECURSE_GUARD short-circuits the fast path with a non-TLS bool
+///   (`_mi_process_is_initialized`) before process init completes,
+///   preventing recursion when dyld's TLV allocator calls `calloc`
+///   on first `__thread` access.
+/// - `MI_HAS_TLS_SLOT` stays at default 1, so the thread-id fast path
+///   still uses `mi_prim_tls_slot(0)` (Apple's system TSD slot — no
+///   conflict).
 ///
 /// Idempotent: detects an already-patched file by a marker comment and
 /// returns early. Panics loudly if the upstream selector block can't be
 /// located — that means mimalloc changed `prim.h` and this patch needs
-/// to be updated. Leaves the submodule's working tree marked dirty
-/// after a fresh clone+build; reset with `git submodule update --force`
-/// if needed (next build will re-apply the patch).
+/// to be updated. Submodule stays pristine in git; `git submodule
+/// update --force` resets it, next build re-applies the patch.
 fn patch_apple_tls_model_for_v3() {
     let prim_h = "c_src/mimalloc3/include/mimalloc/prim.h";
     let content = std::fs::read_to_string(prim_h).unwrap_or_else(|e| {