Native Memory Tracking 与 RSS的差异问题

一问题现象

前一段时间用nmt查看jvm进程的栈区占用的内存大小。测试代码如下

public class ThreadOOM {public static void main(String[] args) {int i = 1;while (i < 3000) {Thread thread = new TestThread();thread.start();System.out.println("thread : " + i);i++;}}
}class TestThread extends Thread {@Overridepublic void run() {while (true) {try {Thread.sleep(Long.MAX_VALUE);} catch (InterruptedException e) {throw new RuntimeException(e);}}}
}

启动命令

nohup java -Xms2G -Xmx2G -XX:MaxMetaspaceSize=512M -XX:NativeMemoryTracking=detail ThreadOOM &

用native memory tracking查看内存占用

jcmd 37898 VM.native_memory scale=MB
37898:Native Memory Tracking:Total: reserved=9366MB, committed=8211MB
-                 Java Heap (reserved=2048MB, committed=2048MB)(mmap: reserved=2048MB, committed=2048MB)-                     Class (reserved=1039MB, committed=12MB)(classes #433)(malloc=7MB #3218)(mmap: reserved=1032MB, committed=5MB)-                    Thread (reserved=6046MB, committed=6046MB)(thread #3017)(stack: reserved=6032MB, committed=6032MB)(malloc=10MB #18096)(arena=3MB #6029)-                      Code (reserved=130MB, committed=3MB)(mmap: reserved=130MB, committed=2MB)-                        GC (reserved=83MB, committed=83MB)(malloc=8MB #123)(mmap: reserved=75MB, committed=75MB)-                  Internal (reserved=17MB, committed=17MB)(malloc=17MB #34406)-                    Symbol (reserved=1MB, committed=1MB)(malloc=1MB #110)-    Native Memory Tracking (reserved=1MB, committed=1MB)(tracking overhead=1MB)

显示线程占用了6G左右，jvm总共committed了8G左右。
使用top查看，常驻物理内存（RES）才占用了139M，这个和nmt显示的差距太大了吧！commited内存不就应该是RES的大小吗？
在这里插入图片描述

二 jdk8申请内存的源码分析

我看的jdk的源码：https://github.com/openjdk/jdk
分支: jdk8-b120
文件位置: hotspot/src/os/linux/vm/os_linux.cpp

reserve内存

char* os::reserve_memory(size_t bytes, char* requested_addr,size_t alignment_hint) {return anon_mmap(requested_addr, bytes, (requested_addr != NULL));
}static char* anon_mmap(char* requested_addr, size_t bytes, bool fixed) {char * addr;int flags;flags = MAP_PRIVATE | MAP_NORESERVE | MAP_ANONYMOUS;if (fixed) {assert((uintptr_t)requested_addr % os::Linux::page_size() == 0, "unaligned address");flags |= MAP_FIXED;}// Map uncommitted pages PROT_READ and PROT_WRITE, change access// to PROT_EXEC if executable when we commit the page.addr = (char*)::mmap(requested_addr, bytes, PROT_READ|PROT_WRITE,flags, -1, 0);if (addr != MAP_FAILED) {if ((address)addr + bytes > _highest_vm_reserved_address) {_highest_vm_reserved_address = (address)addr + bytes;}}return addr == MAP_FAILED ? NULL : addr;
}

commit内存

// NOTE: Linux kernel does not really reserve the pages for us.
//       All it does is to check if there are enough free pages
//       left at the time of mmap(). This could be a potential
//       problem.
bool os::commit_memory(char* addr, size_t size, bool exec) {int prot = exec ? PROT_READ|PROT_WRITE|PROT_EXEC : PROT_READ|PROT_WRITE;uintptr_t res = (uintptr_t) ::mmap(addr, size, prot,MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0);return res != (uintptr_t) MAP_FAILED;
}

不管是reserve还是commit内存，背后都是调用mmap函数

三 mmap函数分析

函数原型

void *mmap(void *start, size_t length, int prot, int flags, int fd, off_t offset);

mmap主要做文件映射，也可以用来为进程申请内存。jdk显然是用来申请内存空间。但是这个系统函数调用后，os并不会立刻分配物理内存，而是等对申请到的内存块进行具体的读写之后再进行物理内存page实际分配。

测试代码

#include <sys/mman.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>int main() {int test = 0;size_t initial_size = 1024*1024*50;  // 初始大小为 50MBsize_t expanded_size = 1024*1024*512; // 扩展大小为 512MB// 创建映射区域void *ptr = mmap(NULL, initial_size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);if (ptr == MAP_FAILED) {perror("mmap");exit(EXIT_FAILURE);}printf("Initial size: %zuKB\n", initial_size / 1024);scanf("%d", &test);// 使用 mremap 扩展映射区域的大小void *new_ptr = mremap(ptr, initial_size, expanded_size, MREMAP_MAYMOVE);if (new_ptr == MAP_FAILED) {perror("mremap");exit(EXIT_FAILURE);}printf("Expanded size: %zuKB\n", expanded_size / 1024);scanf("%d", &test);// 使用新的映射区域进行读写操作...//使用10Mmemset(new_ptr, 0, 1024 * 1024 * 10);scanf("%d", &test);// 使用100Mmemset(new_ptr, 0, 1024 * 1024 * 100);scanf("%d", &test);// 解除映射if (munmap(new_ptr, expanded_size) == -1) {perror("munmap");exit(EXIT_FAILURE);}return 0;
}

3.1 执行mmap函数

mmap函数执行后，查看top输出，虚拟内存52M接近申请的50M，而RES仅有1M
在这里插入图片描述

3.2 执行mremap

mremap执行后，查看top输出，虚拟内存涨到了514MB，接近扩容申请的512MB，RES常驻内存不变
在这里插入图片描述

3.3 执行第一个memset

接着执行第一个memset，进行内存写入。这次发现虚拟内存不变，而RES物理内存增长了10MB，这和memset的内存大小一致
在这里插入图片描述

3.4 执行第2个memset

接着执行第二个memset，写入100MB（指针位置没有变化）。虚存没有变化，RES增加了90MB。
在这里插入图片描述
使用pmp命令分析

lvsheng@lvsheng:/proc/36287$ pmap -x  41422
41422:   ./mmap
Address           Kbytes     RSS   Dirty Mode  Mapping
0000c8e2d33a0000       4       4       0 r-x-- mmap
0000c8e2d33bf000       4       4       4 r---- mmap
0000c8e2d33c0000       4       4       4 rw--- mmap
0000c8e309014000     132       4       4 rw---   [ anon ]
0000e16bc8c00000  524288  102400  102400 rw---   [ anon ]
0000e16bebe20000    1640    1088       0 r-x-- libc.so.6
0000e16bebfba000      76       0       0 ----- libc.so.6
0000e16bebfcd000      12      12      12 r---- libc.so.6
0000e16bebfd0000       8       8       8 rw--- libc.so.6
0000e16bebfd2000      48      16      16 rw---   [ anon ]
0000e16bebfdf000     156     156       0 r-x-- ld-linux-aarch64.so.1
0000e16bec018000       8       8       8 rw---   [ anon ]
0000e16bec01a000       8       0       0 r----   [ anon ]
0000e16bec01c000       4       4       0 r-x--   [ anon ]
0000e16bec01d000       8       8       8 r---- ld-linux-aarch64.so.1
0000e16bec01f000       8       8       8 rw--- ld-linux-aarch64.so.1
0000fffff236c000     132      12      12 rw---   [ stack ]
---------------- ------- ------- -------
total kB          526540  103736  102484