记一次 .NET 某打印服务 非托管内存泄漏分析

时间:2022-09-14 20:18:48

一:背景

1. 讲故事

前段时间有位朋友在微信上找到我,说他的程序出现了内存泄漏,能不能帮他看一下,这个问题还是比较经典的,加上好久没上非托管方面的东西了,这篇就和大家分享一下,话不多说,上 WinDbg 说话。

二:WinDbg 分析

1. 到底是哪里的泄漏

好的开始就是成功的一半,否则就南辕北辙了,对吧,还是用经典的 !address -summary 看一下内存排布情况。


0:000> !address -summary

--- Usage Summary ---------------- RgnCount ----------- Total Size -------- %ofBusy %ofTotal
Heap                                   1935          553b3000 (   1.332 GB)  70.57%   66.59%
Image                                  1022           c306000 ( 195.023 MB)  10.09%    9.52%
<unknown>                              1202           c09d000 ( 192.613 MB)   9.97%    9.41%
Stack                                   541           b280000 ( 178.500 MB)   9.24%    8.72%
Free                                   1158           73ab000 ( 115.668 MB)            5.65%
TEB                                     180            20f000 (   2.059 MB)   0.11%    0.10%
Other                                     8             5d000 ( 372.000 kB)   0.02%    0.02%
PEB                                       1              3000 (  12.000 kB)   0.00%    0.00%

--- State Summary ---------------- RgnCount ----------- Total Size -------- %ofBusy %ofTotal
MEM_COMMIT                             3077          643c6000 (   1.566 GB)  83.00%   78.31%
MEM_RESERVE                            1812          1487f000 ( 328.496 MB)  17.00%   16.04%
MEM_FREE                               1158           73ab000 ( 115.668 MB)            5.65%

从卦中可以看出,当前 MEM_COMMIT = 1.56 G, 并且 Heap= 1.3 G,既然超出了朋友的预期,很明显这是一个非托管内存泄漏,既然 NTHeap 出现了泄漏,那就挖一下看看,使用 !heap -s 观察一下各个heap句柄。


0:000> !heap -s

************************************************************************************************************************
                                              NT HEAP STATS BELOW
************************************************************************************************************************
LFH Key                   : 0xbb72f2a3
Termination on corruption : DISABLED
  Heap     Flags   Reserv  Commit  Virt   Free  List   UCR  Virt  Lock  Fast 
                    (k)     (k)    (k)     (k) length      blocks cont. heap 
-----------------------------------------------------------------------------
00770000 00000002   16576   9716  16364     33   195     5    0      0   LFH
006f0000 00001002    1292    148   1080     11     4     2    0      0   LFH
00a80000 00001002    3336   1972   3124     88    25     3    0      0   LFH
02460000 00001002      60      4     60      0     1     1    0      0      
023b0000 00041002      60      4     60      2     1     1    0      0      
02450000 00001002     272     24     60      1     3     1    0      0   LFH
04a40000 00041002    1292     80   1080      8     4     2    0      0   LFH
06e90000 00001002   64180  56660  63968   1434   473     9  624      7   LFH
09dc0000 00001002      60     12     60      3     2     1    0      0      
0a500000 00001002    7428   3772   7216     43    35     4    0      0   LFH
-----------------------------------------------------------------------------

从卦中的 Commit 列来看,内存占用都不大,最大的也不过 56M ,如果经验丰富的话,你会发现 Virt blocks 高达 624 个,明白 ntheap 的朋友应该知道,凡是大于 512kheapentry 都会单独安排到 VirtualAllocdBlocks 数组中,可以用 dt ntdll!_HEAP 06e90000 给show出来。


0:000> dt ntdll!_HEAP 06e90000
   ...
   +0x05c VirtualMemoryThreshold : 0xfe00
   +0x09c VirtualAllocdBlocks : _LIST_ENTRY [ 0x6ea4000 - 0x7c0d0000 ]
   ...

为了更好的输出 VirtualAllocdBlocks 数组,我们用 windbg 自带的 heap 分析命令。


0:000> !heap 06e90000 -m
Index   Address  Name      Debugging options enabled
  8:   06e90000 
    Segment at 06e90000 to 06e9f000 (0000f000 bytes committed)
    Segment at 078f0000 to 079ef000 (000ff000 bytes committed)
    Segment at 08870000 to 08a6f000 (001ff000 bytes committed)
    Segment at 0ec60000 to 0f05f000 (003f9000 bytes committed)
    Segment at 18660000 to 18e5f000 (007fa000 bytes committed)
    Segment at 26b20000 to 27aef000 (00fc0000 bytes committed)
    Segment at 45320000 to 462ef000 (00fcf000 bytes committed)
    Segment at 65bf0000 to 66bbf000 (008bf000 bytes committed)
    Flags:                00001002
    ForceFlags:           00000000
    Granularity:          8 bytes
    Segment Reserve:      03f70000
    Segment Commit:       00002000
    DeCommit Block Thres: 00000800
    DeCommit Total Thres: 00002000
    Total Free Size:      0002cd56
    Max. Allocation Size: 7ffdefff
    Lock Variable at:     06e90258
    Next TagIndex:        0000
    Maximum TagIndex:     0000
    Tag Entries:          00000000
    PsuedoTag Entries:    00000000
    Virtual Alloc List:   06e9009c
        06ea4000: 00200000 [commited 201000, unused 1000] - busy (b)
        070b2000: 00200000 [commited 201000, unused 1000] - busy (b)
        079f4000: 00200000 [commited 201000, unused 1000] - busy (b)
        07c0f000: 00200000 [commited 201000, unused 1000] - busy (b)
        0802b000: 00200000 [commited 201000, unused 1000] - busy (b)
        08238000: 00200000 [commited 201000, unused 1000] - busy (b)
        08444000: 00200000 [commited 201000, unused 1000] - busy (b)
        0865f000: 00200000 [commited 201000, unused 1000] - busy (b)
        0e20f000: 00200000 [commited 201000, unused 1000] - busy (b)
        0e42b000: 00200000 [commited 201000, unused 1000] - busy (b)
        0e635000: 00200000 [commited 201000, unused 1000] - busy (b)
        0e841000: 00200000 [commited 201000, unused 1000] - busy (b)
        0c661000: 00200000 [commited 201000, unused 1000] - busy (b)
        0c87e000: 00200000 [commited 201000, unused 1000] - busy (b)
        0ca8b000: 00200000 [commited 201000, unused 1000] - busy (b)
        0ea56000: 00200000 [commited 201000, unused 1000] - busy (b)
        0f062000: 00200000 [commited 201000, unused 1000] - busy (b)
        0f275000: 00200000 [commited 201000, unused 1000] - busy (b)
        ...

从卦中可以看到大量的 commited 201000, unused 1000 ,这里的 0x201000 转换一下大概就是 2M,以经验来说,这 2M 大概就是 pdf,image,bitmap 等这些玩意了,由于没有开启 pageheap 或 ust,没法追踪到底是什么东西分配的,到这里就没法进展下去了。

2. 到底是谁分配的 2M 数据

首先能进入 VirtualAllocdBlocks 数组自然是高层调用了 HeapAlloc 这类API,同时这个数据量高度怀疑是 Bitmap,Pdf 之类的大文件,很大可能是托管代码做了什么导致这个资源没有释放,接下来使用 !dumpheap -stat 看下托管堆。


0:000> !dumpheap -stat
Statistics:
      MT    Count    TotalSize Class Name
...
09ae7e48      627        15048 System.Drawing.Bitmap
6b267040      178       366680 System.Decimal[]
6b2cb4a0     1850       601588 System.String[]
6b2cdd14     1379       638190 System.Byte[]
6b2cac14    15919      1146764 System.String
09aec720    66332      1326640 System.Drawing.FontFamily
09ae8590    66074      2907256 System.Drawing.Font
Total 289300 objects

从卦中看,System.Drawing.Font 居然高达 6w 个,而且 System.Drawing.Bitmap 和 heap 上的 624 也非常接近,看样子就是 Bitmap 啦,那为什么这个 Bitmap 没有善终呢? 可以用 !frq -stat 观察下终结器队列的 Freachable Queue 情况。


0:000> !frq -stat
Freachable Queue:
       Count      Total Size   Type
---------------------------------------------------------
         152            3648   System.Data.Odbc.CNativeBuffer
          76            2128   System.Data.Odbc.OdbcConnectionHandle
          77            1540   System.Transactions.SafeIUnknown
          76            1824   System.Data.Odbc.OdbcStatementHandle
        2432          145920   System.Windows.Forms.Control+ControlNativeWindow
         304            7296   System.Drawing.Bitmap
       66062         2906728   System.Drawing.Font
         258            5160   System.Drawing.FontFamily
         308            9856   System.Drawing.Graphics
         308            3696   System.Windows.Forms.ImageList+NativeImageList
           1              12   System.Drawing.Text.InstalledFontCollection
          12             240   System.Threading.ThreadPoolWorkQueueThreadLocals
           1              20   System.Security.Cryptography.SafeKeyHandle
           1              20   Microsoft.Win32.SafeHandles.SafeWaitHandle
           6             120   Microsoft.Win32.SafeHandles.SafeRegistryHandle
          12             624   System.Threading.Thread
        1577           69388   System.Threading.ReaderWriterLock
           1              20   System.Security.Cryptography.SafeProvHandle

71,664 objects, 3,158,240 bytes

我去,这个可终结队列居然高达 7.1w ,这是很有问题的,大概率当前的终结器线程瓦特了,接下来追查下 终结器线程 此时正在做什么 ?


0:000> !t
ThreadCount:      107
UnstartedThread:  0
BackgroundThread: 93
PendingThread:    0
DeadThread:       12
Hosted Runtime:   no
                                                                         Lock  
       ID OSID ThreadOBJ    State GC Mode     GC Alloc Context  Domain   Count Apt Exception
   0    1 138ac 0079fef0     26020 Preemptive  00000000:00000000 00798f38 1     STA 
   2    2 12b08 007adac0     2b220 Preemptive  00000000:00000000 00798f38 0     MTA (Finalizer) 
  ...

0:000> ~2s
eax=00000001 ebx=00000000 ecx=00000000 edx=00000000 esi=00000001 edi=00000001
eip=777b2f8c esp=0466eaf4 ebp=0466ec84 iopl=0         nv up ei pl nz na po nc
cs=0023  ss=002b  ds=002b  es=002b  fs=0053  gs=002b             efl=00000202
ntdll!NtWaitForMultipleObjects+0xc:
777b2f8c c21400          ret     14h
0:002> k
 # ChildEBP RetAddr      
00 0466ec84 762fc753     ntdll!NtWaitForMultipleObjects+0xc
01 0466ec84 7695d9aa     KERNELBASE!WaitForMultipleObjectsEx+0x103
02 0466ed34 7695c564     combase!MTAThreadWaitForCall+0x1ca [onecore\com\combase\dcomrem\channelb.cxx @ 7234] 
03 0466edc0 769a9923     combase!MTAThreadDispatchCrossApartmentCall+0xf4 [onecore\com\combase\dcomrem\chancont.cxx @ 229] 
04 (Inline) --------     combase!CSyncClientCall::SwitchAptAndDispatchCall+0x9e4 [onecore\com\combase\dcomrem\channelb.cxx @ 5856] 
05 0466efa0 769ab739     combase!CSyncClientCall::SendReceive2+0xad3 [onecore\com\combase\dcomrem\channelb.cxx @ 5459] 
06 (Inline) --------     combase!SyncClientCallRetryContext::SendReceiveWithRetry+0x29 [onecore\com\combase\dcomrem\callctrl.cxx @ 1542] 
07 (Inline) --------     combase!CSyncClientCall::SendReceiveInRetryContext+0x29 [onecore\com\combase\dcomrem\callctrl.cxx @ 565] 
...

从上面的 MTAThreadDispatchCrossApartmentCall 可知,这又是一个经典的 COM 释放问题导致终结器线程卡死。。。接下来仔细看下 线程列表的 STA 情况,可以发现有大量的线程是 STA 模式。

记一次 .NET 某打印服务 非托管内存泄漏分析

接下来就是将结果告诉朋友,为什么开的线程都是 STA 套件模式。

三:总结

总的来说,这次内存泄漏的原因在于朋友开了 STA 模式的线程,导致终结器线程卡死,进而导致 Bitmap 分配之后无法释放,最终引发非托管泄漏。

这个dump告诉我们,不要放弃,一定可以在绝望中找到希望。

记一次 .NET 某打印服务 非托管内存泄漏分析