[Bug] Two ring-buffer allocator defects in pto_ring_buffer.h: heap wrap-around off-by-one and DepListPool sentinel collision

[chenshengxin2026]

Platform

a2a3 (Ascend 910B/C hardware)

Runtime Variant

tensormap_and_ringbuffer

Description

Two related allocation logic defects in src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_ring_buffer.h:

Bug 1 — Heap wrap-around: strict > should be >= in try_bump_heap (line ~270)

PTO2TaskAllocator::try_bump_heap uses tail > alloc_size (strict greater-than) when checking whether to wrap the heap pointer to the beginning of the ring. When tail == alloc_size there is exactly enough space at [0, alloc_size), but the condition incorrectly rejects it, so the allocator spins until deadlock.

Fix: change tail > alloc_size → tail >= alloc_size.

Bug 2 — DepListPool: top % capacity returns the NULL sentinel slot (index 0)

PTO2DepListPool::alloc() computes the physical index as idx = top % capacity (line ~444). When top is a multiple of capacity (e.g. top=8, capacity=8) the result is 0 — the same slot that init() reserved as the NULL sentinel. Handing out &entries_[0] overwrites the sentinel, breaking pto2_dep_pool_get() for callers that rely on a nullptr chain terminator at index 0.

Fix: offset the usable range so index 0 is never handed out (e.g. idx = (top % capacity) + 1 with capacity reduced by 1, or start top at 1).

Steps to Reproduce

1. Run the ring-buffer unit tests that exercise allocation wrap-around and dep-list pool full-cycle allocation (added in PR Add: unit tests for C++ runtime modules and Python utilities #427):

   ctest -R test_ring_buffer
   

2. The HeapWrapAround test case deadlocks (Bug 1).
3. The DepListPoolSentinelCollision test case fails with sentinel corruption (Bug 2).

Expected Behavior

• try_bump_heap should wrap the heap pointer to the start of the ring when tail == alloc_size, allowing allocation to succeed.
• PTO2DepListPool::alloc() should never return &entries_[0] (the NULL sentinel slot); the sentinel must remain intact for the lifetime of the pool.

Actual Behavior

• Bug 1: When tail == alloc_size the allocator fails to wrap, enters an infinite spin-wait, and the test times out / deadlocks.
• Bug 2: When the pool has been fully cycled (top reaches a multiple of capacity), alloc() returns &entries_[0], overwriting the NULL sentinel and corrupting the dep-list chain termination logic.

Git Commit ID

9aa9681

Host Platform

Linux (aarch64)

Additional Context

Both defects were discovered while writing unit tests for the ring buffer subsystem. Related PR: #427