记一次 .NET 某招聘网后端服务 内存暴涨分析
2021-10-27 一线码农的vlog
加入收藏
一：背景
1. 讲故事
前段时间有位朋友wx找到我，说他的程序存在内存阶段性暴涨，寻求如何解决，和朋友沟通下来，他的内存平时大概是5G 左右，在某些时点附近会暴涨到 10G+, 画个图大概就是这样。
所以接下来就是想办法给他找到那莫名奇妙的 5-6G 是个啥，上 windbg 说话。
二：Windbg 分析
1. 判断托管还是非托管
从描述上看大概率是托管层面的问题，但为了文章的完整性，我们还是用 !address -summary 和 !eeheap -gc 来看一下。
0:000> !address -summary
--- Usage Summary ---------------- RgnCount ----------- Total Size -------- %ofBusy %ofTotal
Free 1164 7f5`58f12000 ( 7.958 TB) 99.48%
<unknown> 6924 a`6de84000 ( 41.717 GB) 97.90% 0.51%
Stack 1123 0`16340000 ( 355.250 MB) 0.81% 0.00%
Image 4063 0`1607d000 ( 352.488 MB) 0.81% 0.00%
Heap 71 0`0c9ea000 ( 201.914 MB) 0.46% 0.00%
TEB 374 0`002ec000 ( 2.922 MB) 0.01% 0.00%
Other 13 0`001c6000 ( 1.773 MB) 0.00% 0.00%
PEB 1 0`00001000 ( 4.000 kB) 0.00% 0.00%
--- Type Summary (for busy) ------ RgnCount ----------- Total Size -------- %ofBusy %ofTotal
MEM_PRIVATE 5423 a`87200000 ( 42.111 GB) 98.83% 0.51%
MEM_IMAGE 7033 0`1e5d6000 ( 485.836 MB) 1.11% 0.01%
MEM_MAppED 113 0`01908000 ( 25.031 MB) 0.06% 0.00%
--- State Summary ---------------- RgnCount ----------- Total Size -------- %ofBusy %ofTotal
MEM_FREE 1164 7f5`58f12000 ( 7.958 TB) 99.48%
MEM_RESERVE 4165 8`1b873000 ( 32.430 GB) 76.11% 0.40%
MEM_COMMIT 8404 2`8b86b000 ( 10.180 GB) 23.89% 0.12%
0:000> !eeheap -gc
Number of GC Heaps: 32
------------------------------
Heap 0 (00000000004106d0)
generation 0 starts at 0x0000000082eb0e58
generation 1 starts at 0x0000000082d79b20
generation 2 starts at 0x000000007fff1000
ephemeral segment allocation context: none
segment begin allocated size
000000007fff0000 000000007fff1000 0000000083f80128 0x3f8f128(66646312)
Large object heap starts at 0x000000087fff1000
segment begin allocated size
000000087fff0000 000000087fff1000 0000000883fe4190 0x3ff3190(67056016)
0000000927ff0000 0000000927ff1000 000000092bfe2430 0x3ff1430(67048496)
0000000a81c50000 0000000a81c51000 0000000a8221c858 0x5cb858(6076504)
Heap Size: Size: 0xc53ef40 (206827328) bytes.
------------------------------
...
Heap 31 (0000000019c84130)
generation 0 starts at 0x0000000844fc5170
generation 1 starts at 0x0000000844f851f8
generation 2 starts at 0x000000083fff1000
ephemeral segment allocation context: none
segment begin allocated size
000000083fff0000 000000083fff1000 0000000845171ca0 0x5180ca0(85462176)
Large object heap starts at 0x00000008fbff1000
segment begin allocated size
00000008fbff0000 00000008fbff1000 00000008fffe2290 0x3ff1290(67048080)
000000094bff0000 000000094bff1000 000000094ea2ebb8 0x2a3dbb8(44293048)
000000096bff0000 000000096bff1000 000000096dbdec00 0x1bedc00(29285376)
Heap Size: Size: 0xd79d6e8 (226088680) bytes.
------------------------------
GC Heap Size: Size: 0x1f1986a88 (8348265096) bytes.
从卦中得知，10G的内存，托管堆吃掉了 8.3G，很明显托管层问题，知道大方向后，接下来就可以到托管堆看一看，根据过往经验程序肯定是生成了大量的类对象所致，上命令 !dumpheap -stat 。
0:000> !dumpheap -stat
Statistics:
MT Count TotalSize Class Name
...
000007fe9ddd5fc0 341280 30032640 System.ServiceModel.Description.MessagePartDescription
000007fe9c4865a0 866349 41584752 System.Xml.XmlDictionaryString
000007fe9defb098 937801 45014448 System.Xml.XmlDictionaryString
000007fe9c66bd28 105052 45086880 System.Collections.Generic.Dictionary`2+Entry[[System.String, mscorlib],[System.Xml.XmlDictionaryString, System.Runtime.Serialization]][]
000007fe9e0f4d20 113299 49050864 System.Collections.Generic.Dictionary`2+Entry[[System.String, mscorlib],[System.Xml.XmlDictionaryString, System.Runtime.Serialization]][]
00000000003c9190 44573 618414438 Free
000007fef8f6c168 428410 1209974642 System.Char[]
000007fef8f4f1b8 2849758 1246912848 System.Object[]
000007fef8f6f058 531963 1670620873 System.Byte[]
000007fef8f6aee0 2368431 2382587716 System.String
真是皂滑弄人，并没有命中过往经验，可以看出占用最大的都是些 Byte,String,Char,Object 基础类型，其实这些基础类型排查起来很难搞，要么不断的用 -min, -max 去筛选，要么就写一个脚本对它进行分组排序，蹩脚脚本如下：
"use strict";
/*
按 mt 对托管堆类型的size进行分组
*/
let platform = 64
let mtlist = ["000007fef8f4f1b8"];
let maxlimit = 100;
function initializeScript() { return [new host.apiVersionSupport(1, 7)]; }
function log(str) { host.diagnostics.debugLog(str + "n"); }
function exec(str) { log("n" + str); return host.namespace.Debugger.Utility.Control.ExecuteCommand(str); }
function invokeScript() { for (var mt of mtlist) { groupby_mtsize_inheap(mt); } }
//对某个类型按照size 进行分组
function groupby_mtsize_inheap(mt) {
var size_group = {};
var commandText = "!dumpheap -mt " + mt;
var output = exec(commandText);
for (var line of output) {
if (line == "" || line.indexOf("Address") > -1) continue;
if (line.indexOf("Statistics") > -1) break;
var size = parseInt(line.substring(Math.ceil(platform / 2) + 1).trim());
if (!size_group[size]) size_group[size] = 0;
size_group[size]++;
}
show_top10_format(mt, size_group);
}
function show_top10_format(mt, size_group) {
var maparr = [];
//转数组
for (var size in size_group) {
maparr.push({ "size": size, "count": size_group[size], "totalsize": (size * size_group[size]) });
}
maparr.sort(function (a, b) { return b.totalsize - a.totalsize });
var topTotalSize = 0;
//按size输出
for (var i = 0; i < Math.min(maparr.length, maxlimit); i++) {
var size = maparr[i].size;
var count = maparr[i].count;
var totalsize = Math.round(maparr[i].totalsize / 1024 / 1024, 2);
topTotalSize += totalsize
log("size=" + size + ",count=" + count + ",totalsize=" + totalsize + "M");
}
log("Total:" + topTotalSize + "M");
//show max
if (maparr.length > 0) {
var size = maparr[0].size;
var totalsize = Math.round(maparr[0].totalsize / 1024 / 1024, 2) + "M";
var output = exec("!dumpheap -mt " + mt + " -min 0n" + size + " -max 0n" + size + " -short").Take(maxlimit);
for (var line of output) {
log(line);
}
}
}
接下来把 string 的方法表地址传下去看看排序结果，简化输出如下：
!dumpheap -mt 000007fef8f6aee0
size=29285946,count=2,totalsize=56M
size=29285540,count=2,totalsize=56M
size=29285502,count=2,totalsize=56M
size=29285348,count=2,totalsize=56M
size=27455186,count=2,totalsize=52M
size=31116504,count=1,totalsize=30M
size=31116490,count=1,totalsize=30M
size=31116306,count=1,totalsize=30M
size=31115934,count=1,totalsize=30M
size=31115920,count=1,totalsize=30M
size=31115718,count=1,totalsize=30M
size=29286342,count=1,totalsize=28M
size=29285898,count=1,totalsize=28M
...
Total:1198M
可以看到，有不少大 size 的 string，那这些string到底是个啥，这里我随便抽几个导出到txt看看。
0:000> !dumpheap -mt 000007fef8f6aee0 -min 0n31116490 -max 0n31116490 -short
0000000a61c51000
0:000> !do 0000000a61c51000
Name: System.String
MethodTable: 000007fef8f6aee0
EEClass: 000007fef88d3720
Size: 31116490(0x1daccca) bytes
File: C:windowsMicrosoft.NETassemblyGAC_64mscorlibv4.0_4.0.0.0__b77a5c561934e089mscorlib.dll
String: <String is invalid or too large to print>
Fields:
MT Field Offset Type VT Attr Value Name
000007fef8f6dc90 40000aa 8 System.Int32 1 instance 15558232 m_stringLength
000007fef8f6c1c8 40000ab c System.Char 1 instance 50 m_firstChar
000007fef8f6aee0 40000ac 18 System.String 0 shared static Empty
>> Domain:Value 00000000003fb620:NotInit 000000001ca30bd0:NotInit 000000001f7b21a0:NotInit 000000001f8940c0:NotInit 0000000027dc46b0:NotInit 00000000281bd720:NotInit 00000000282b7ee0:NotInit <<
0:000> .writemem D:dumpsxxxxstring.txt 0000000a61c51000 L?0x1daccca
Writing 1daccca bytes..........
从内容看其实就是 pdf 的 base64 编码，以同样的方式调研 char[] 和 byte[] 类型，发现大多也都是 pdf，猜测程序在处理 pdf 的过程中，进行了 byte[],char[],string 之间的切换，所以这些对象理论上大多属于无根对象，其实通过 !heapstat -iu 也能看到那大约 5.5G 的无根对象正等待GC回收。
0:000> !heapstat -iu
Heap Gen0 Gen1 Gen2 LOH
Heap0 17625808 1274680 47745824 140181016
...
Total 357486256 28100616 2229673376 5733004848
Free space: Percentage
Heap0 3962240 24 11211224 298616SOH: 22% LOH: 0%
Heap1 5625856 144 9857168 302152SOH: 27% LOH: 0%
...
Heap31 1448576 24 19957312 218024SOH: 25% LOH: 0%
Total 181492784 1136 431825856 5183128
Unrooted objects: Percentage
Heap0 12163928 243584 42872 137153536SOH: 18% LOH: 97%
...
Heap31 236832 239272 1435840 139770656SOH: 2% LOH: 99%
Total 164954952 7948448 29066480 5530423784
三：总结
本次内存阶段性暴涨的事故，主要还是程序接收了上游过多的 pdf文件，毕竟这些都是大对象，还进行了 char[] ，string，byte[] 的切换，造成短时间内过大的内存占用。
最后就是我个人的解决建议：
- 针对大量的pdf，能否借用第三方的 oss 软件来规避一些不必要的内存占用。
- 清洗服务是否可以做些限流或者使用服务均摊的方式。
后来听朋友说，他做了筛选过滤以及一些业务流程优化解决了这个问题，我想现实中肯定有很多朋友遇到过这类问题，欢迎大家留言补充您的解决方案。
