程序內(nèi)存一直在泄漏,原來是異步死循環(huán)了 !
一、背景
1. 講故事
上個月有位朋友找到我,說他的程序出現(xiàn)了內(nèi)存泄漏,不知道如何進一步分析,截圖如下:
朋友這段話已經(jīng)說的非常言簡意賅了,那就上 windbg 說話吧。
二、Windbg 分析
1. 到底是哪一方面的泄漏
根據(jù)朋友描述,程序運行一段時間后,內(nèi)存就炸了,應該沒造成人員傷亡,不然也不會跟我wx聊天了,這里可以用 .time 看看當前的 process 跑了多久。
- 0:000> .time
- Debug session time: Thu Oct 21 14:54:39.000 2021 (UTC + 8:00)
- System Uptime: 6 days 4:37:27.851
- Process Uptime: 0 days 0:40:14.000
- Kernel time: 0 days 0:01:55.000
- User time: 0 days 0:07:33.000
看的出來,這個 dump 是在程序跑了 40min 之后抓的,接下來我們比較一下 process 的內(nèi)存和 gc堆 占比, 看看到底是哪一塊的泄漏。
- 0:000> !address -summary
- --- State Summary ---------------- RgnCount ----------- Total Size -------- %ofBusy %ofTotal
- MEM_FREE 327 7dfc`c665a000 ( 125.987 TB) 98.43%
- MEM_RESERVE 481 201`e91a2000 ( 2.007 TB) 99.74% 1.57%
- MEM_COMMIT 2307 1`507f4000 ( 5.258 GB) 0.26% 0.00%
- 0:000> !eeheap -gc
- Number of GC Heaps: 2
- ------------------------------
- GC Allocated Heap Size: Size: 0x139923528 (5260850472) bytes.
- GC Committed Heap Size: Size: 0x13bf23000 (5300695040) bytes.
2. 到底是什么占用了如此大的內(nèi)存
知道是 托管層 的泄漏,感覺一下子就幸福起來了,接下來用 !dumpheap -stat 看看有沒有什么大對象可挖。
- 0:000> !dumpheap -stat
- Statistics:
- MT Count TotalSize Class Name
- 00007ffdeb1fc400 5362921 128710104 xxxBLLs.xxx.BundleBiz+<>c__DisplayClass20_0
- 00007ffdeaeff140 5362929 171613728 System.Collections.Generic.List`1[[xxx.xxx, xxx]]
- 00007ffdeaeff640 5362957 171615272 xxx.BLLs.Plan.Dto.xxx[]
- 00007ffde8171e18 16146362 841456072 System.String
- 00007ffdeb210098 5362921 1415811144 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.Threading.Tasks.VoidTaskResult, System.Private.CoreLib],[xxx.BundleBiz+<DistributionBundle>d__20, xxx]]
- 00007ffdea9ca260 5362921 2359685240 xxx.Bundle
從輸出看,內(nèi)存主要被 xxx.Bundle 和 AsyncTaskMethodBuilder 兩大類對象給吃掉了,數(shù)量都高達 536w,這里有一個非常有意思的地方,如果你了解異步,我相信你一看就能看出 AsyncTaskMethodBuilder + VoidTaskResult 是干嘛的,按照經(jīng)驗,這位朋友應該誤入了 異步無限遞歸 ,那怎么去挖呢?接著往下看。
3. 尋找問題代碼
看到上面的 xxx.BundleBiz+
雖然找到了源碼,但代碼是 ILSpy 反編譯出來的異步狀態(tài)機,接下來的一個問題是,如何根據(jù)狀態(tài)機代碼反向?qū)ふ业?await ,async 代碼呢?在 ILSpy 中有一個 used by 功能,在這里可以用起來了。
雙擊 used by 就能看到真正的調(diào)用代碼,簡化后如下:
- public async Task DistributionBundle(List<Bundle> list, List<xxx> bwdList, xxx item, List<xxx> sumDetails, List<xxx> details, BundleParameter bundleParameter, IEnumerable<dynamic> labels)
- {
- int num = 0;
- foreach (xxx detail in sumDetails)
- {
- IEnumerable<xxx> woDetails = details.Where((xxx w) => w.Size == detail.Size && w.Color == detail.Color);
- foreach (xxx item2 in woDetails)
- {
- xxx
- }
- woDetails = woDetails.OrderBy((xxx s) => s.Seq).ToList();
- num++;
- xxx
- Bundle bundle = new Bundle();
- Bundle bundle2 = bundle;
- bundle2.BundleId = await _repo.CreateBundleId();
- foreach (xxx item3 in woDetails)
- {
- item3.TaskQty = item3.WoQty + Math.Ceiling(item3.WoQty * item3.OverCutRate);
- decimal value = default(decimal);
- }
- await DistributionBundle(list, bwdList, item, sumDetails, details, bundleParameter, labels);
- }
- }
仔細看上面這段代碼, 我去, await DistributionBundle(list, bwdList, item, sumDetails, details, bundleParameter, labels); 又調(diào)用了自身,看樣子是某種條件下陷入了一個死遞歸。
有些朋友可能要問,除了經(jīng)驗之外,能從 dump 中分析出來嗎?當然可以,從 500w+ 中抽一個看看它的 !gcroot 即可。
- 0:000> !DumpHeap /d -mt 00007ffdeb210098
- Address MT Size
- 000001a297913a68 00007ffdeb210098 264
- 000001a297913b70 00007ffdeb210098 264
- 0:000> !gcroot 000001a297913a68
- Thread 5ac:
- 000000470B1EE4E0 00007FFE45103552 System.Threading.Tasks.Task.SpinThenBlockingWait(Int32, System.Threading.CancellationToken) [/_/src/System.Private.CoreLib/shared/System/Threading/Tasks/Task.cs @ 2922]
- rbp+10: 000000470b1ee550
- -> 000001A297A25D88 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.Threading.Tasks.VoidTaskResult, System.Private.CoreLib],[Microsoft.Extensions.Hosting.HostingAbstractionsHostExtensions+<RunAsync>d__4, Microsoft.Extensions.Hosting.Abstractions]]
- -> 000001A29796D8C0 Microsoft.Extensions.Hosting.Internal.Host
- ...
- -> 000001A298213248 System.Data.SqlClient.TdsParserStateObjectNative
- -> 000001A32E6AB700 System.Threading.Tasks.TaskFactory`1+<>c__DisplayClass38_0`1[[System.Data.SqlClient.SqlDataReader, System.Data.SqlClient],[System.Data.CommandBehavior, System.Data.Common]]
- -> 000001A32E6AB728 System.Threading.Tasks.Task`1[[System.Data.SqlClient.SqlDataReader, System.Data.SqlClient]]
- -> 000001A32E6ABB18 System.Threading.Tasks.StandardTaskContinuation
- -> 000001A32E6ABA80 System.Threading.Tasks.ContinuationTaskFromResultTask`1[[System.Data.SqlClient.SqlDataReader, System.Data.SqlClient]]
- -> 000001A32E6AB6C0 System.Action`1[[System.Threading.Tasks.Task`1[[System.Data.SqlClient.SqlDataReader, System.Data.SqlClient]], System.Private.CoreLib]]
- -> 000001A32E6AB428 System.Data.SqlClient.SqlCommand+<>c__DisplayClass130_0
- ...
- -> 000001A32E6ABC08 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.String, System.Private.CoreLib],[Dapper.SqlMapper+<QueryRowAsync>d__34`1[[System.String, System.Private.CoreLib]], Dapper]]
- -> 000001A32E6ABD20 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.String, System.Private.CoreLib],[xxx.DALs.xxx.BundleRepo+<CreateBundleId>d__12, xxx]]
- -> 000001A32E6ABD98 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.Threading.Tasks.VoidTaskResult, System.Private.CoreLib],[xxx.BundleBiz+<DistributionBundle>d__20, xxx]]
- -> 000001A32E6A6BD8 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.Threading.Tasks.VoidTaskResult, System.Private.CoreLib],[xxx.BundleBiz+<DistributionBundle>d__20, xxx]]
- -> 000001A433250520 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.Threading.Tasks.VoidTaskResult, System.Private.CoreLib],[xxx.BundleBiz+<DistributionBundle>d__20, xxx]]
- -> 000001A32E69E0F8 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.Threading.Tasks.VoidTaskResult, System.Private.CoreLib],[xxx.BundleBiz+<DistributionBundle>d__20, xxx]]
- -> 000001A433247D28 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.Threading.Tasks.VoidTaskResult, System.Private.CoreLib],[xxx.BundleBiz+<DistributionBundle>d__20, xxx]]
- -> 000001A433246330 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.Threading.Tasks.VoidTaskResult, System.Private.CoreLib],[xxx.BundleBiz+<DistributionBundle>d__20, xxx]]
- -> 000001A32E69A568 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.Threading.Tasks.VoidTaskResult, System.Private.CoreLib],[xxx.BundleBiz+<DistributionBundle>d__20, xxx]]
- -> 000001A433245408 System.Runtime.CompilerServices.AsyncTaskMethodBuilder`1+AsyncStateMachineBox`1[[System.Threading.Tasks.VoidTaskResult, System.Private.CoreLib],[xxx.BundleBiz+<DistributionBundle>d__20, xxx]]
- ...
從調(diào)用棧來看,代碼貌似是從數(shù)據(jù)庫讀取記錄的過程中陷入死循環(huán)的。
4. 為什么沒有出現(xiàn)棧溢出
一看到無限循環(huán),我相信很多朋友肯定要問,為啥沒出現(xiàn)堆棧溢出,畢竟默認的線程??臻g僅僅 1M 而已,從 !gcroot 上看,這些引用都是掛在 5ac 線程上,也就是下面輸出的 主線程 ,而且主線程棧也非常干凈。
- 0:000> !t
- ThreadCount: 30
- UnstartedThread: 0
- BackgroundThread: 24
- PendingThread: 0
- DeadThread: 5
- Hosted Runtime: no
- Lock
- DBG ID OSID ThreadOBJ State GC Mode GC Alloc Context Domain Count Apt Exception
- 0 1 5ac 000001A29752CDF0 202a020 Preemptive 0000000000000000:0000000000000000 000001a29754c570 0 MTA
- 4 2 1e64 000001A29752A490 2b220 Preemptive 0000000000000000:0000000000000000 000001a29754c570 0 MTA (Finalizer)
- ...
- 0:000> !clrstack
- OS Thread Id: 0x5ac (0)
- Child SP IP Call Site
- 000000470B1EE1D0 00007ffe5eb30544 [GCFrame: 000000470b1ee1d0]
- 000000470B1EE318 00007ffe5eb30544 [HelperMethodFrame_1OBJ: 000000470b1ee318] System.Threading.Monitor.ObjWait(Boolean, Int32, System.Object)
- 000000470B1EE440 00007ffe45103c25 System.Threading.ManualResetEventSlim.Wait(Int32, System.Threading.CancellationToken)
- 000000470B1EE4E0 00007ffe45103552 System.Threading.Tasks.Task.SpinThenBlockingWait(Int32, System.Threading.CancellationToken) [/_/src/System.Private.CoreLib/shared/System/Threading/Tasks/Task.cs @ 2922]
- 000000470B1EE550 00007ffe451032cf System.Threading.Tasks.Task.InternalWaitCore(Int32, System.Threading.CancellationToken) [/_/src/System.Private.CoreLib/shared/System/Threading/Tasks/Task.cs @ 2861]
- 000000470B1EE5D0 00007ffe45121b04 System.Runtime.CompilerServices.TaskAwaiter.HandleNonSuccessAndDebuggerNotification(System.Threading.Tasks.Task) [/_/src/System.Private.CoreLib/shared/System/Runtime/CompilerServices/TaskAwaiter.cs @ 143]
- 000000470B1EE600 00007ffe4510482d System.Runtime.CompilerServices.TaskAwaiter.GetResult() [/_/src/System.Private.CoreLib/shared/System/Runtime/CompilerServices/TaskAwaiter.cs @ 106]
- 000000470B1EE630 00007ffe4de36595 Microsoft.Extensions.Hosting.HostingAbstractionsHostExtensions.Run(Microsoft.Extensions.Hosting.IHost) [/_/src/Hosting/Abstractions/src/HostingAbstractionsHostExtensions.cs @ 49]
- 000000470B1EE660 00007ffde80f3b4b xxx.Program.Main(System.String[])
- 000000470B1EE8B8 00007ffe47c06c93 [GCFrame: 000000470b1ee8b8]
- 000000470B1EEE50 00007ffe47c06c93 [GCFrame: 000000470b1eee50]
如果你稍微了解一點異步的玩法,你應該知道這其中有一個 IO完成端口 的概念,它可以實現(xiàn) 句柄 和 ThreadPool 的綁定,無限遞歸只不過是進了 IO完成端口 的待回調(diào)隊列中而已,理論上和??臻g沒什么關(guān)系,也就不會出現(xiàn)棧溢出了。
三、總結(jié)
本次內(nèi)存泄漏的事故主要還是因為程序員的大意,也許是長期的 996 給弄恍惚了 ??????,有了這些信息,修正起來相信會非常簡單。
本文轉(zhuǎn)載自微信公眾號「一線碼農(nóng)聊技術(shù)」,可以通過以下二維碼關(guān)注。轉(zhuǎn)載本文請聯(lián)系一線碼農(nóng)聊技術(shù)公眾號。