一:背景
1. 讲故事
年初有位朋友找到我,说他们的管理系统不响应了,让我帮忙看下到底咋回事? 手上也有dump,那就来分析吧。
二:为什么没有响应
1. 线程池队列有积压吗?
朋友的系统是一个web系统,那web系统的无响应我们首先要关注的就是 线程池
,使用 !sos tpq
命令,参考输出如下:
0:000> !sos tpq
global work item queue________________________________
0x00000004010774C0 Microsoft.AspNetCore.Server.IIS.Core.IISHttpContextOfT<Microsoft.AspNetCore.Hosting.HostingApplication+Context>
0x0000000401077808 Microsoft.AspNetCore.Server.IIS.Core.IISHttpContextOfT<Microsoft.AspNetCore.Hosting.HostingApplication+Context>
....
0x000000030239DD78 Microsoft.AspNetCore.Server.IIS.Core.IISHttpContextOfT<Microsoft.AspNetCore.Hosting.HostingApplication+Context>
0x000000030239E0C0 Microsoft.AspNetCore.Server.IIS.Core.IISHttpContextOfT<Microsoft.AspNetCore.Hosting.HostingApplication+Context>
local per thread work items_____________________________________
0x0000000100A46410 System.Threading.Tasks.Task<System.Threading.Tasks.Task>
...
0x000000010133F8C0 System.Threading.Tasks.Task<System.Threading.Tasks.Task>
2 Work Microsoft.AspNetCore.Http.Connections.Internal.HttpConnectionContext+<>c.<WaitOnTasks>b__123_1
4 Work Microsoft.AspNetCore.Http.Connections.Internal.HttpConnectionContext+<>c.<WaitOnTasks>b__123_0
266 Work Microsoft.AspNetCore.SignalR.HubConnectionContext.AbortConnection
----
272
从卦中可以看到确实存在线程池积压的情况,那为什么会有积压呢?条件反射告诉我,是不是因为锁的原因,使用 !syncblk
观察。
0:000> !syncblk
Index SyncBlock MonitorHeld Recursion Owning Thread Info SyncBlock Owner
-----------------------------
Total 468
CCW 0
RCW 0
ComClassFactory 0
Free 120
从卦中看和锁没半毛钱关系,那就只能深入各个消费线程,看看这些线程为什么这么不给力。。。。
2. 线程都在干什么
要想观察各个线程都在做什么,可以用 ~*e !clrstack
观察各个线程调用栈,输出的调用栈太长,但仔细观察之后,发现很多线程都停留在 TryGetConnnection
上,截图如下:
从卦中可以看到大概有143个线程卡在 TryGetConnection 上,不知道 Connection 为啥取不到了,接下来观察问题代码,简化后如下:
private bool TryGetConnection(DbConnection owningObject, uint waitForMultipleObjectsTimeout, bool allowCreate, bool onlyOneCheckConnection, DbConnectionOptions userOptions, out DbConnectionInternal connection)
{
DbConnectionInternal dbConnectionInternal = null;
Transaction transaction = null;
if (HasTransactionAffinity)
{
dbConnectionInternal = GetFromTransactedPool(out transaction);
}
if (dbConnectionInternal == null)
{
Interlocked.Increment(ref _waitCount);
do
{
num = WaitHandle.WaitAny(_waitHandles.GetHandles(allowCreate), (int)waitForMultipleObjectsTimeout);
} while (dbConnectionInternal == null);
}
}
private DbConnectionInternal GetFromTransactedPool(out Transaction transaction)
{
transaction = ADP.GetCurrentTransaction();
DbConnectionInternal dbConnectionInternal = null;
if (null != transaction && _transactedConnectionPool != null)
{
dbConnectionInternal = _transactedConnectionPool.GetTransactedObject(transaction);
//....
}
return dbConnectionInternal;
}
internal DbConnectionInternal GetTransactedObject(Transaction transaction)
{
lock (_transactedCxns)
{
flag = _transactedCxns.TryGetValue(transaction, out value);
}
if (flag)
{
lock (value)
{
int num = value.Count - 1;
if (0 <= num)
{
dbConnectionInternal = value[num];
value.RemoveAt(num);
}
}
}
return dbConnectionInternal;
}
从卦中源码看是 _transactedCxns
返回 null 所致,看样子是_transactedCxns中的Conenction耗尽,接下来使用 !dso
从 DbConnectionPool 字段中去挖,输出如下:
0:298> !dso
OS Thread Id: 0x42c0 (298)
SP/REG Object Name
0005c087d020 0003ffc79528 Microsoft.Data.ProviderBase.DbConnectionPool
0:298> !do 0003ffc79528
Name: Microsoft.Data.ProviderBase.DbConnectionPool
MethodTable: 000007fe8dd714c8
EEClass: 000007fe8dd65438
Tracked Type: false
Size: 176(0xb0) bytes
File: D:\APP\EnergyDotNet\runtimes\win\lib\net6.0\Microsoft.Data.SqlClient.dll
Fields:
MT Field Offset Type VT Attr Value Name
000007fe8dd71370 40001a7 8c System.Int32 1 instance 1 _state
...
000007fe8cd994b0 40001ad 94 System.Int32 1 instance 143 _waitCount
000007fe8e012c68 40001b4 78 ...tedConnectionPool 0 instance 00000003ffc7a3e8 _transactedConnectionPool
...
000007fe8cd994b0 40001b6 9c System.Int32 1 instance 100 _totalObjects
...
0:298> !DumpObj /d 00000003ffc7a520
Name: System.Collections.Generic.Dictionary`2[[System.Transactions.Transaction, System.Transactions.Local],[Microsoft.Data.ProviderBase.DbConnectionPool+TransactedConnectionList, Microsoft.Data.SqlClient]]
MethodTable: 000007fe8e0146a0
EEClass: 000007fe8cd8af30
Tracked Type: false
Size: 80(0x50) bytes
File: C:\Program Files\dotnet\shared\Microsoft.NETCore.App\6.0.22\System.Private.CoreLib.dll
Fields:
MT Field Offset Type VT Attr Value Name
000007fe8cda8480 4002065 8 System.Int32[] 0 instance 0000000000000000 _buckets
000007fe8e3255d8 4002066 10 ...ivate.CoreLib]][] 0 instance 0000000000000000 _entries
000007fe8cd9ca08 4002067 30 System.UInt64 1 instance 0 _fastModMultiplier
000007fe8cd994b0 4002068 38 System.Int32 1 instance 0 _count
000007fe8cd994b0 4002069 3c System.Int32 1 instance 0 _freeList
000007fe8cd994b0 400206a 40 System.Int32 1 instance 0 _freeCount
000007fe8cd994b0 400206b 44 System.Int32 1 instance 0 _version
000007fe8cdb7900 400206c 18 ...Private.CoreLib]] 0 instance 0000000000000000 _comparer
000007fe8da70c90 400206d 20 ...Private.CoreLib]] 0 instance 0000000000000000 _keys
000007fe8d2a1c60 400206e 28 ...Private.CoreLib]] 0 instance 0000000000000000 _values
从上面的卦中数据可知三点信息:
100 _totalObjects
当前的线程池存着100个Connection。0 _count
当前100个Connection全部耗尽。143 _waitCount
表示当前有 143 个线程在获取 Connection 上进行等待。
3. 池中之物都去了哪里
要想找到这个答案,继续观察线程栈,比如搜索TDS传输层方法 Microsoft.Data.SqlClient.TdsParserStateObject.TryReadByte
,可以看到刚好是 100 个,上层大多是 xxxx.GetRoomNosInDltParams
方法,截图如下:
挖掘各个线程栈,大概都是下面的sql,格式化如下:
SELECT [HotelId],
[RoomId],
[SubRoomId],
[SupplierId],
[RoomNo],
[RoomCategory],
[SellPrice],
[RoomNoState],
[CheckInDate]
FROM [xxx]
WHERE ((((((([RoomId] IN (673,674)) AND( [SellPrice] > @SellPrice1 ))
AND ( [CheckInDate] >= @CheckInDate2 ))
AND ( [CheckInDate] <= @CheckInDate3 ))
AND ( [RoomCategory] <=@constant5))
AND ( [RoomNoState] =@constant7))
AND NOT ([RoomNo] LIKE @MethodConst8+'%') )
反编译源码发现有一个 useCached 字段形同虚设,导致每次都是从 数据库 读取,截图如下:
到这里基本上就能推测出来的,程序的卡死主要是 Connection 耗尽,优化建议如下:
- SQL 加上 nolock,避免锁问题。
- GetRoomNosInDltParams 方法尽量使用缓存。
- 观察数据库层的锁和负载情况。
- 设置更大的数据库连接池。
三:总结
这次卡死的生产事故,是大量数据库的慢请求导致SDK侧的数据库连接池(100)耗尽所致,如果有数据库侧的监控工具,我想一眼就能找到答案。