[GH-ISSUE #2197] [Bug]: 华为昇腾，开启ascend_use_fabric_mem时，使用dummy-client卸载ssd，出现unmap 崩溃 #604

New issue

Closed

opened 2026-10-29 21:17:06 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-29 21:17:06 +00:00

Owner

Originally created by @xuyucheng220 on GitHub (May 23, 2026).
Original GitHub issue: https://github.com/kvcache-ai/Mooncake/issues/2197

Bug Report

初步分析原因：
mooncake-store/src/real_client.cpp中在 FabricMem=1 时跳过RegisterLocalMemory：
if (!globalConfig().ascend_use_fabric_mem) {
auto result = client_->RegisterLocalMemory(mapped_va, mem_size, "npu",
false, true);
if (!result.has_value()) {
LOG(ERROR) << "Failed to register memory";
(void)aclrtIpcMemClose(ipc_key_bytes.data());
return tl::unexpected(result.error());
}
LOG(INFO) << "RegisterLocalMemory for ipc shm, va:" << mapped_va;
}

但是Unmap 时 teardown_ascend_shm_buffer始终调 unregisterLocalMemory，没有判断是否开启了fabricMem：
void RealClient::teardown_ascend_shm_buffer(MappedShm &shm) {
#ifdef USE_ASCEND_DIRECT
(void)ContextManager::getInstance().setCurrentContextByPhysicalId(
shm.device_id);
if (shm.shm_size > 0 && client_) {
auto res = client_->unregisterLocalMemory(shm.shm_buffer, true);
if (!res) {
LOG(WARNING)
<< "Failed to unregister local memory for shared memory: "
<< shm.shm_name << ", error: " << toString(res.error());
}
}

Before submitting...

Ensure you searched for relevant issues and read the [documentation]

Originally created by @xuyucheng220 on GitHub (May 23, 2026). Original GitHub issue: https://github.com/kvcache-ai/Mooncake/issues/2197 ### Bug Report 初步分析原因： mooncake-store/src/real_client.cpp中在 FabricMem=1 时跳过RegisterLocalMemory： if (!globalConfig().ascend_use_fabric_mem) { auto result = client_->RegisterLocalMemory(mapped_va, mem_size, "npu", false, true); if (!result.has_value()) { LOG(ERROR) << "Failed to register memory"; (void)aclrtIpcMemClose(ipc_key_bytes.data()); return tl::unexpected(result.error()); } LOG(INFO) << "RegisterLocalMemory for ipc shm, va:" << mapped_va; } 但是Unmap 时 teardown_ascend_shm_buffer始终调 unregisterLocalMemory，没有判断是否开启了fabricMem： void RealClient::teardown_ascend_shm_buffer(MappedShm &shm) { #ifdef USE_ASCEND_DIRECT (void)ContextManager::getInstance().setCurrentContextByPhysicalId( shm.device_id); if (shm.shm_size > 0 && client_) { auto res = client_->unregisterLocalMemory(shm.shm_buffer, true); if (!res) { LOG(WARNING) << "Failed to unregister local memory for shared memory: " << shm.shm_name << ", error: " << toString(res.error()); } } ### Before submitting... - [ ] Ensure you searched for relevant issues and read the [documentation]

bsakaguchi

2026-10-29 21:17:06 +00:00

closed this issue
added the
bug
label

bsakaguchi commented

2026-10-29 21:17:07 +00:00

Author

Owner

@ascend-direct-dev commented on GitHub (May 25, 2026):

unregisterLocalMemory has special handling — if the memory was not registered, it is skipped and will not cause problems. However, we suggest identifying the root cause.

@ascend-direct-dev commented on GitHub (May 25, 2026): unregisterLocalMemory has special handling — if the memory was not registered, it is skipped and will not cause problems. However, we suggest identifying the root cause.

bsakaguchi referenced this issue

2026-10-29 21:29:27 +00:00

[PR #604] [MERGED] [TransferEngine] Fix address already in use #1681