ext == reinterpret_cast<uintptr_t>(nullptr) → true(已归还 cache 链表且 list head 为空,输出 OP_LOGW("cacheExt_ is nullptr, block has been returned to cache list (list head was empty)."))。
将 ext 强转 BlockHeader*,若其 magic_ == MAGIC → true(已归还 cache 链表的某节点,输出 OP_LOGW("cacheExt_ points to another block header in cache list, block has been returned already."))。
Backgroud(背景信息)
aclDestroyXxx系列 API(aclDestroyTensor/aclDestroyScalar/aclDestroyIntArray/aclDestroyFloatArray/aclDestroyBoolArray/aclDestroyTensorList/aclDestroyScalarList,共 7 个)当前实现仅为if (ptr == nullptr) return OK; delete ptr;(见src/nnopbase/common/api/acl_op_api.cpp)。当上层应用对同一个aclTensor*/aclScalar*/aclIntArray*等对象重复调用aclDestroyXxx时,第二次及之后会触发未定义行为(double-free):底层内存池BlockPool的 cache 链表会被破坏、或直接对已归还系统的内存再次std::free,进而引发堆损坏、段错误或难以定位的随机崩溃。由于这些 API 是 C 接口、被各类 host 应用与第三方框架频繁调用,问题在用户侧表现为"偶发崩溃 + 无任何 aclnn 侧日志",定位成本极高——既无法在 aclnn 层确认是否发生过 double-free,也无法定位到具体的 destroy 调用点。
Origin(信息来源)
本代码仓架构师和开发,基于内存重复释放的难题,设计的维测功能。
Benefit / Necessity(价值/作用)
OP_LOG_INFO级别)后,对同一个aclDestroyXxx入参地址的重复释放会输出OP_LOGW("possible double-free at addr %p."),把"用户侧偶发崩溃"转化为"aclnn 日志中的明确告警",定位时间从天级降到分钟级。unlikely(...) &&+std::atomic<bool>::load(memory_order_relaxed)守门,仅在调试模式开启时执行完整判定;生产模式(默认日志级别)下分支预测几乎零成本,不引入性能回归。OP_LOGW告警,不阻断delete(避免误报把正常用户路径打死),保持 API 行为向后兼容。BlockPoolcache,cacheExt_状态机可识别"已归还 cache 链表";同时覆盖同线程与跨线程 double-free 两个核心场景。应用场景:
OP_LOG_INFO即可拿到 aclnn 视角的告警;test_check_double_free.cpp,10 个TEST_F)。Design(设计方案)
触发开关(
src/nnopbase/composite_op/aclnn_engine/op_dfx.cpp/src/nnopbase/common/inc/op_dfx_internal.h):std::atomic<bool> g_aclnnDebugEnabled{false}与初始化函数InitAclnnDebugSwitch()(std::call_once保护)。InitL2Phase1Context()(op_executor.cpp)首次进入即调用InitAclnnDebugSwitch():若CheckLogLevelInner(OP_ID, OP_LOG_INFO) == 1则置g_aclnnDebugEnabled = true。IsAclnnDebugEnabled()用memory_order_relaxed读原子,配合unlikely实现热路径零成本。核心判定(
src/nnopbase/common/inc/bridge_pool.h+src/nnopbase/common/utils/bridge_pool.cpp新增bool CheckDoubleFree(void* addr)):addr == nullptr→false(合法的"删除空指针",不报)。BlockPool::InHugeMemRange(addr)→false(大页内存 free 是 no-op,不会 double-free)。BlockStore::BlockHeader* head = BlockStore::GetBlockHeader(addr),若head->magic_ != BlockStore::MAGIC→false(野指针 / 未初始化)。ext = head->cacheExt_:ext == BlockStore::NOT_IN_CACHE→false(冷路径或刚分配未入 cache)。ext == BlockCache::CurrentThreadCacheAddr()→false(本线程活跃 block,防误报)。ext == reinterpret_cast<uintptr_t>(nullptr)→true(已归还 cache 链表且 list head 为空,输出OP_LOGW("cacheExt_ is nullptr, block has been returned to cache list (list head was empty)."))。ext强转BlockHeader*,若其magic_ == MAGIC→true(已归还 cache 链表的某节点,输出OP_LOGW("cacheExt_ points to another block header in cache list, block has been returned already."))。ext指向分配方线程的BlockCachethis 指针)→false(防误报,4 字节偶然匹配 MAGIC 概率 1/2^32,可忽略)。接入点(
src/nnopbase/common/api/acl_op_api.cpp,7 个aclDestroyXxx):if (unlikely(op::internal::IsAclnnDebugEnabled()) && op::internal::CheckDoubleFree(const_cast<aclXxx*>(ptr))) { OP_LOGW("possible double-free at addr %p.", static_cast<const void*>(ptr)); } delete ptr;对
aclDestroyTensorList/aclDestroyScalarList,仅对外层 list 指针本身做一次预检,元素由循环内aclDestroyTensor/aclDestroyScalar各自覆盖。配套基础修复(
src/nnopbase/common/inc/block_pool.h,3 处std::malloc后):BlockHeader经std::malloc分配后未调用构造函数,magic_等非静态成员处于未初始化状态(UB)。本 PR 在MallocImpl/PoolAlloc/CacheFreeImpl三条路径补new (head) BlockStore::BlockHeader;,保证magic_有确定值,使CheckDoubleFree第 3 步的 magic 校验具备意义。static uintptr_t BlockCache::CurrentThreadCacheAddr(),返回当前线程BlockCache实例地址,供CheckDoubleFree活跃态短路使用。已知限制(best-effort,文档化):
FreeImpl直接std::free,CheckDoubleFree读已释放内存(UB),magic 几乎不可能匹配 → 漏报(符合设计,由ColdPath_Over64KB_NotReportedUT 文档化)。cacheExt_指向其他线程BlockCache实例内部偏移恰好 4 字节匹配 MAGIC,概率 1/2^32,可忽略。测试覆盖(
tests/nnopbase/ut/composite_op/test_check_double_free.cpp,10 个TEST_F):CurrentThreadCacheAddr_ReturnsStableNonZeroAddr/Nullptr_ReturnsFalse/WildPointer_MagicMismatch_ReturnsFalseSameThreadDoubleFree_DetectedCrossThreadAccess_NoFalsePositive/CrossThreadDoubleFree_DetectedHugeMemRange_NotReported/ColdPath_Over64KB_NotReported/ColdPath_NotInCacheShortCircuit/ActiveBlock_SameThread_NoFalsePositive关联 PR:https://gitcode.com/cann/opbase/pull/610