[robobun]

Problem

• On Windows, Bun faults inside ExitProcess when it runs under Intel SDE: panic(main thread): Segmentation fault at address 0x…. The top frames are ntdll ← mi_scav_wake_one (WakeByAddressSingle, mimalloc src/scavenger.c:127) ← mi_arenas_page_free_prim ← Rust thread-local destructors. 6 of the last 12 builds of main print it in windows x64 - verify-baseline.
• ExitProcess terminates every other thread first. A thread-local destructor then frees a page, and mimalloc wakes its scavenger thread, which no longer exists. ntdll reads the wait block of that thread from its stack. Under SDE that stack is not mapped any more.
• The exit code of the emulated run stays 0. scripts/verify-baseline.ts prints PASS, and each crashing run takes about 27 s longer.

Fix

• Bump MIMALLOC_COMMIT to the fork change: the Windows mi_scav_wake_one returns before WakeByAddressSingle once RtlDllShutdownInProgress() is true.
• This is correct because ntdll sets that flag in LdrShutdownProcess, after the other threads are terminated and before the first callback runs. All four wakes of the scavenger go through this one function, so each exit path is covered.
• verify-baseline.ts fails a run that prints Bun's crash report, whatever its exit code is.
• Verified: in progress. This PR is a draft until the numbers are in (see Notes).

Background

• The scavenger is a thread of Bun's mimalloc fork that returns freed memory to the OS. On Windows it waits in WaitOnAddress.
• ntdll keeps the wait block of a WaitOnAddress waiter on the stack of that waiter, in a process-wide table. A wake walks the table.
• Considered mi_scavenger_stop() before ExitProcess: a join on every exit, and Bun has 9 exit sites. Considered TerminateProcess: fix(windows): use TerminateProcess to prevent NAPI module segfault on exit #27829 did that and was reverted. The guard in the fork also covers an exit() call from a native addon.

Downsides

• Numbers follow with the verification.

Notes

State of this draft. The first commits carry a temporary measurement: the emulated test file runs 7 times in verify-baseline.ts, first with the current pin and then with the new one. The measurement is removed before this leaves draft.

[robobun]

^{Updated 8:29 PM PT - Oct 20th, 2026}

✅ @robobun, your commit e896e1c9aa1ed442cc51b03f6a5500709149ccc0 passed in Build #129766! 🎉

---

🧪   To try this PR locally:

bunx bun-pr 46016

That installs a local version of the PR into your bun-46016 executable, so you can run:

bun-46016 --bun

[robobun]

Status: draft. The fix is oven-sh/mimalloc#61. This PR pins its head commit for verification and moves to the merged commit when that PR is merged.

How this was reproduced:

• The crash shows in the CI step windows x64 - verify-baseline (Bun under Intel SDE). Build 128726 of this PR (current mimalloc pin, 7 emulated runs, with the new crash gate): 4 of 7 runs print panic(main thread): Segmentation fault with exit code 0, and the step fails with FAIL: Bun printed a crash report, exit code 0.
• Build 129569 (new pin, same 7 runs): no run prints a crash report.
• All decoded traces have the same frames: ntdll+0x5e895 ← mi_scav_wake_one ← mi_arenas_page_free_prim ← Rust thread-local destructors ← ExitProcess ← bun_core::Global::exit. Sampled builds of main show the same fault since 2026-07-28.
• On native Windows (Server 2025 x64 and Windows 11 arm64) the test of the fork PR releases the stack of the terminated scavenger thread and then frees a block. Without the fix the wake ends in an access violation in 140 of 140 runs. With the fix 440 of 440 runs pass.

The PR still carries a temporary measurement (100 emulated runs and an exit code probe in the step). It is removed before the PR leaves draft.