RCTIdentifierPool::dequeue() spins forever when the pool is exhausted, hanging the main thread and freezing the device

先從 packages/react-native/React/Fabric/Utils/RCTIdentifierPool.h 開始，然後檢查 RCTSurfaceTouchHandler.mm 中的 _registerTouches: 和 _unregisterTouches:。執行 issue 中的耗盡範例，或在可用時執行相關的現有測試。完成的標準是：當全部 11 個 slot 都已佔用時，dequeue() 會返回，lastIndex 已初始化，且正常的 allocate/free 行為維持不變。

Description

RCTIdentifierPool::dequeue() is an unbounded while (true) loop that never terminates
once every slot is occupied. When that happens the main thread spins at 100% CPU forever,
the app stops servicing scene updates, and iOS — which waits on the frontmost app — leaves
the whole device unresponsive until FrontBoard's watchdog kills the app (0x8BADF00D).

packages/react-native/React/Fabric/Utils/RCTIdentifierPool.h:

int dequeue() {
  while (true) {
    if (!usage[lastIndex]) {
      usage[lastIndex] = true;
      return lastIndex;
    }
    lastIndex = (lastIndex + 1) % size;   // nothing in the loop ever frees a slot
  }
}

Nothing inside the loop clears a bit in usage, so if all size bits are set the loop can
never exit. This is unchanged in 0.82, 0.85, 0.87 and main.

How the pool gets exhausted

RCTSurfaceTouchHandler uses RCTIdentifierPool<11>. Two paths leak a slot permanently:

1. _unregisterTouches: skips _identifierPool.enqueue(...) when the touch is not in
   _activeTouches — the continue after the RCTAssert. RCTAssert is compiled out in
   Release, so in production this leaks silently with no diagnostic at all.
2. _registerTouches: always calls dequeue(), but _activeTouches.emplace(touch, ...) is a
   no-op if that UITouch * key already exists, so the freshly taken identifier is orphaned.

The registry desync that drives (1) is already reported in #53303, where it surfaces as a
crash because the assert fires in Debug. In Release it is invisible and simply leaks.

Slots only reset on process restart, so this accumulates over the lifetime of the process.

Evidence from a production hang

iPad (A16) / iPadOS 26.6 / RN 0.81.1, New Architecture, process alive ~43 hours.

Two reports for the same pid:

• cpu_resource: 90 seconds cpu time over 113 seconds (80% cpu average), Num threads: 1,
  footprint 181 MB. Not memory pressure.
• Hang report: FRONTBOARD 0x8BADF00D — "scene-update watchdog transgression: app exhausted
  real (wall clock) time allowance of 10.00 seconds".

Symbolicated main thread:

-[RCTSurfaceTouchHandler _registerTouches:]         RCTSurfaceTouchHandler.mm:197
-[RCTSurfaceTouchHandler touchesBegan:withEvent:]   RCTSurfaceTouchHandler.mm:308
-[UIGestureRecognizer _componentsBegan:withEvent:]

Line 197 is activeTouch.touch.identifier = _identifierPool.dequeue();

Register state at the crash PC proves the pool was full:

Register	Value	Meaning
x9	2047 = 0b11111111111	all 11 slots occupied
x28	11	pool size
x27	0x2E8BA2E8BA2E8BA3	magic constant for % 11
x10 / x11	3 / 8	lastIndex, 1 << lastIndex

The PC was pinned to the tst/b.ne at the bottom of that loop in 14 of 15 CPU samples.

React Native Version

0.81.1 (verified unchanged in 0.82.0, 0.85.0, 0.87.0 and main)

Affected Platforms

iOS (New Architecture / Fabric)

Steps to reproduce

The loop is unbounded by inspection — no runtime reproducer is needed to see it cannot exit:

RCTIdentifierPool<11> pool;
for (int i = 0; i < 11; i++) pool.dequeue();  // fill every slot
pool.dequeue();                               // never returns

In the field it is reached by leaking 11 touch identifiers over a long-lived process.

Suggested fix

Bound the scan. After size steps lastIndex is back where it started, so a further
iteration cannot find anything new:

int dequeue() {
  for (size_t attempt = 0; attempt < size; attempt++) {
    if (!usage[lastIndex]) {
      usage[lastIndex] = true;
      return lastIndex;
    }
    lastIndex = (lastIndex + 1) % size;
  }
  // Every slot taken. Reclaim rather than hang: a reused touch identifier is a
  // transient glitch, an infinite loop takes the whole device down.
  usage.reset();
  usage[lastIndex] = true;
  return lastIndex;
}

This is behaviour-preserving wherever the current code terminates — I ran 200k randomised
allocate/free sequences against both and the returned identifiers are identical. They differ
only in the case that currently hangs forever.

int lastIndex; is also uninitialised and is used as a std::bitset subscript before any
assignment; int lastIndex{0}; would be worth including.

Fixing the leak in RCTSurfaceTouchHandler is worthwhile too, but the unbounded loop is what
turns a leaked identifier into an unresponsive device.