示例程序2
在上一篇《Linux perf子系統的使用(一)——計數》已經講解了如何使用perf_event_open、read和ioctl對perf子系統進行編程。但有時我們並不需要計數,而是要採樣。比如這麼一個需求:統計一個程序中哪些函數最耗時間。嗯,這個功能確實可以通過perf record命令來做,但是perf record內部又是如何做到的呢?自己實現又是怎樣的呢?perf record是基於統計學原理的。假設以1000Hz的頻率對某個進程採樣,每次採樣記錄下該進程的IP寄存器的值(也就是下一條指令的地址)。通過分析該進程的可執行文件,是可以得知每次採樣的IP值處於哪個函數內部。OK,那麼我們相當於以1000Hz的頻率獲知進程當前所執行的函數。如果某個函數f()佔用了30%的時間,那麼所有采樣中,該函數出現的頻率也應該將近30%,只要採樣數量足夠多。這正是perf record的原理。所以,perf的採樣模式很有用~
但是,採樣比較複雜,主要表現在三點:1、採樣需要設置觸發源,也就是告訴kernel何時進行一次採樣;2、採樣需要設置信號,也就是告訴kernnel,採樣完成後通知誰;3、採樣值的讀取需要使用mmap,因爲採樣有異步性,需要一個環形隊列,另外也是出於性能的考慮。
直接上代碼吧,對照着官方手冊看,學習效率最高:
採集單個值
perf.c
//如果不加,則F_SETSIG未定義
#define _GNU_SOURCE 1
#include <stdio.h>
#include <fcntl.h>
#include <stdint.h>
#include <unistd.h>
#include <string.h>
#include <signal.h>
#include <sys/mman.h>
#include <sys/syscall.h>
#include <linux/perf_event.h>
//環形緩衝區大小,16頁,即16*4kB
#define RING_BUFFER_PAGES 16
//目前perf_event_open在glibc中沒有封裝,需要手工封裝一下
int perf_event_open(struct perf_event_attr *attr,pid_t pid,int cpu,int group_fd,unsigned long flags)
{
return syscall(__NR_perf_event_open,attr,pid,cpu,group_fd,flags);
}
//mmap共享內存的開始地址
void* rbuf;
//環形隊列中每一項元素
struct perf_my_sample
{
struct perf_event_header header;
uint64_t ip;
};
//下一條採樣記錄的相對於環形緩衝區開頭的偏移量
uint64_t next_offset=0;
//採樣完成後的信號處理函數
void sample_handler(int sig_num,siginfo_t *sig_info,void *context)
{
//計算出最新的採樣所在的位置(相對於rbuf的偏移量)
uint64_t offset=4096+next_offset;
//指向最新的採樣
struct perf_my_sample* sample=(void*)((uint8_t*)rbuf+offset);
//過濾一下記錄
if(sample->header.type==PERF_RECORD_SAMPLE)
{
//得到IP值
printf("%lx\n",sample->ip);
}
//共享內存開頭是一個struct perf_event_mmap_page,提供環形緩衝區的信息
struct perf_event_mmap_page* rinfo=rbuf;
//手工wrap一下data_head值,得到下一個記錄的偏移量
next_offset=rinfo->data_head%(RING_BUFFER_PAGES*4096);
}
//模擬的一個負載
void workload()
{
int i,c=0;
for(i=0;i<100000000;i++)
{
c+=i*i;
c-=i*100;
c+=i*i*i/100;
}
}
int main()
{
struct perf_event_attr attr;
memset(&attr,0,sizeof(struct perf_event_attr));
attr.size=sizeof(struct perf_event_attr);
//觸發源爲CPU時鐘
attr.type=PERF_TYPE_SOFTWARE;
attr.config=PERF_COUNT_SW_CPU_CLOCK;
//每100000個CPU時鐘採樣一次
attr.sample_period=100000;
//採樣目標是IP
attr.sample_type=PERF_SAMPLE_IP;
//初始化爲禁用
attr.disabled=1;
int fd=perf_event_open(&attr,0,-1,-1,0);
if(fd<0)
{
perror("Cannot open perf fd!");
return 1;
}
//創建1+16頁共享內存,應用程序只讀,讀取fd產生的內容
rbuf=mmap(0,(1+RING_BUFFER_PAGES)*4096,PROT_READ,MAP_SHARED,fd,0);
if(rbuf<0)
{
perror("Cannot mmap!");
return 1;
}
//這三個fcntl爲何一定這麼設置不明,但必須這樣
fcntl(fd,F_SETFL,O_RDWR|O_NONBLOCK|O_ASYNC);
fcntl(fd,F_SETSIG,SIGIO);
fcntl(fd,F_SETOWN,getpid());
//開始設置採樣完成後的信號通知
struct sigaction sig;
memset(&sig,0,sizeof(struct sigaction));
//由sample_handler來處理採樣完成事件
sig.sa_sigaction=sample_handler;
//要帶上siginfo_t參數(因爲perf子系統會傳入參數,包括fd)
sig.sa_flags=SA_SIGINFO;
if(sigaction(SIGIO,&sig,0)<0)
{
perror("Cannot sigaction");
return 1;
}
//開始監測
ioctl(fd,PERF_EVENT_IOC_RESET,0);
ioctl(fd,PERF_EVENT_IOC_ENABLE,0);
workload();
//停止監測
ioctl(fd,PERF_EVENT_IOC_DISABLE,0);
munmap(rbuf,(1+RING_BUFFER_PAGES)*4096);
close(fd);
return 0;
}
可以看到一下子比計數模式複雜多了。採樣模式是要基於計數模式的——選擇一個“參考計數器”,並設置一個閾值,每當這個“參考計數器”達到閾值時,觸發一次採樣。每次採樣,kernel會把值放入隊列的末尾。如何得知kernenl完成了一次最新的採樣了呢?一種方法就是定時輪詢,另一種就是響應信號。
如何讀取mmap共享內存中的值呢?首先,共享內存開頭是一個struct perf_event_mmap_page,提供環形緩衝區的信息,對我們最重要的字段就是data_head,官方手冊的介紹是這樣的:
注意,data_head一直遞增,不回滾!!所以需要手動處理wrap。另外一個需要注意的地方是,每次事件響應中,得到的data_head是下一次採樣的隊列頭部,所以需要自己保存一個副本next_offset,以供下次使用。
這個struct perf_event_mmap_page獨佔共享內存的第一頁。後面必須跟2n頁,n自己決定。這2n頁用來存放採樣記錄。每一條記錄的結構體如下:
因爲我只選擇了採樣IP,即PERF_SAMPLE_IP,所以這個結構體就退化爲了:
struct perf_my_sample
{
struct perf_event_header header;
uint64_t ip;
};
另外一個需要注意的地方是mmap中的第三個參數,是PROT_READ,表示應用程序只讀。如果設置爲了PROT_READ|PROT_WRITE,那麼讀取的過程就不一樣了:
這樣相當於和kernel做一個同步操作,效率務必下降。而且由於SIGIO這個信號是不可靠信號,所以如果某次採樣完成的通知沒有被截獲,那麼就可能產生死鎖。
gcc perf.c -o perf
sudo ./perf
運行上面的代碼,產生如下輸出:
爲了驗證採集到的IP值是否正確,可以反彙編一下:
objdump -d ./perf
可以看到採集到的IP值全部落在workload這個函數的地址範圍內。
採集多個值
要採多個值的話,也很方便:
//如果不加,則F_SETSIG未定義
#define _GNU_SOURCE 1
#include <stdio.h>
#include <fcntl.h>
#include <stdint.h>
#include <unistd.h>
#include <string.h>
#include <signal.h>
#include <sys/mman.h>
#include <sys/syscall.h>
#include <linux/perf_event.h>
//環形緩衝區大小,16頁,即16*4kB
#define RING_BUFFER_PAGES 2
//目前perf_event_open在glibc中沒有封裝,需要手工封裝一下
int perf_event_open(struct perf_event_attr *attr,pid_t pid,int cpu,int group_fd,unsigned long flags)
{
return syscall(__NR_perf_event_open,attr,pid,cpu,group_fd,flags);
}
//mmap共享內存的開始地址
void* rbuf;
//環形隊列中每一項元素
struct perf_my_sample
{
struct perf_event_header header;
uint64_t ip;
uint64_t nr;
uint64_t ips[0];
};
//下一條採樣記錄的相對於環形緩衝區開頭的偏移量
uint64_t next_offset=0;
//採樣完成後的信號處理函數
void sample_handler(int sig_num,siginfo_t *sig_info,void *context)
{
//計算出最新的採樣所在的位置(相對於rbuf的偏移量)
uint64_t offset=4096+next_offset;
//指向最新的採樣
struct perf_my_sample* sample=(void*)((uint8_t*)rbuf+offset);
//過濾一下記錄
if(sample->header.type==PERF_RECORD_SAMPLE)
{
//得到IP值
printf("IP: %lx\n",sample->ip);
if(sample->nr<1024)
{
//得到調用鏈長度
printf("Call Depth: %lu\n",sample->nr);
//遍歷調用鏈
int i;
for(i=0;i<sample->nr;i++)
printf(" %lx\n",sample->ips[i]);
}
}
//共享內存開頭是一個struct perf_event_mmap_page,提供環形緩衝區的信息
struct perf_event_mmap_page* rinfo=rbuf;
//手工wrap一下data_head值,得到下一個記錄的偏移量
next_offset=rinfo->data_head%(RING_BUFFER_PAGES*4096);
}
//模擬的一個負載
void workload()
{
int i,c=0;
for(i=0;i<1000000000;i++)
{
c+=i*i;
c-=i*100;
c+=i*i*i/100;
}
}
int main()
{
struct perf_event_attr attr;
memset(&attr,0,sizeof(struct perf_event_attr));
attr.size=sizeof(struct perf_event_attr);
//觸發源爲CPU時鐘
attr.type=PERF_TYPE_SOFTWARE;
attr.config=PERF_COUNT_SW_CPU_CLOCK;
//每100000個CPU時鐘採樣一次
attr.sample_period=100000;
//採樣目標是IP
attr.sample_type=PERF_SAMPLE_IP|PERF_SAMPLE_CALLCHAIN;
//初始化爲禁用
attr.disabled=1;
int fd=perf_event_open(&attr,0,-1,-1,0);
if(fd<0)
{
perror("Cannot open perf fd!");
return 1;
}
//創建1+16頁共享內存,應用程序只讀,讀取fd產生的內容
rbuf=mmap(0,(1+RING_BUFFER_PAGES)*4096,PROT_READ,MAP_SHARED,fd,0);
if(rbuf<0)
{
perror("Cannot mmap!");
return 1;
}
//這三個fcntl爲何一定這麼設置不明,但必須這樣
fcntl(fd,F_SETFL,O_RDWR|O_NONBLOCK|O_ASYNC);
fcntl(fd,F_SETSIG,SIGIO);
fcntl(fd,F_SETOWN,getpid());
//開始設置採樣完成後的信號通知
struct sigaction sig;
memset(&sig,0,sizeof(struct sigaction));
//由sample_handler來處理採樣完成事件
sig.sa_sigaction=sample_handler;
//要帶上siginfo_t參數(因爲perf子系統會傳入參數,包括fd)
sig.sa_flags=SA_SIGINFO;
if(sigaction(SIGIO,&sig,0)<0)
{
perror("Cannot sigaction");
return 1;
}
//開始監測
ioctl(fd,PERF_EVENT_IOC_RESET,0);
ioctl(fd,PERF_EVENT_IOC_ENABLE,0);
workload();
//停止監測
ioctl(fd,PERF_EVENT_IOC_DISABLE,0);
munmap(rbuf,(1+RING_BUFFER_PAGES)*4096);
close(fd);
return 0;
}
示例程序2
在上一篇《Linux perf子系統的使用(二)——採樣(signal方式)》中,我使用了信號來接收採樣完成通知,並在回調函數中讀取最新的採樣值。雖說回調方式有很多優點,但是並不是太通用。更加糟糕的是,信號會打斷幾乎所有的系統調用,使得本來的程序邏輯被破壞。另一個很糟糕的點是,如果一個進程中需要開多個採樣器,那麼就要共享同一個事件回調函數,破壞了封裝性。
因此,最好有一個阻塞式輪詢的辦法。嗯,這就是今天要講的東西——通過poll()函數等待採樣完成。
其實poll()輪詢的實現比信號的方式簡單,只要把perf_event_open()返回的文件描述符當做普通的文件描述符傳入poll()就可以了。在創建perf文件描述附時,唯一需要注意的就是需要手動設置wakeup_events的值。wakeup_events決定了多少次採樣以後進行一次通知(poll模式下就是讓poll返回),一般設置爲1。
直接上代碼吧,和《Linux perf子系統的使用(二)——採樣(signal方式)》中的代碼比較一下就一目瞭然了。
perf_poll.cpp
#include <poll.h>
#include <errno.h>
#include <stdio.h>
#include <stdint.h>
#include <assert.h>
#include <signal.h>
#include <string.h>
#include <unistd.h>
#include <signal.h>
#include <sys/mman.h>
#include <sys/ioctl.h>
#include <sys/syscall.h>
#include <linux/perf_event.h>
// the number of pages to hold ring buffer
#define RING_BUFFER_PAGES 8
// a wrapper for perf_event_open()
static int perf_event_open(struct perf_event_attr *attr,
pid_t pid,int cpu,int group_fd,unsigned long flags)
{
return syscall(__NR_perf_event_open,attr,pid,cpu,group_fd,flags);
}
static bool sg_running=true;
// to receive SIGINT to stop sampling and exit
static void on_closing(int signum)
{
sg_running=false;
}
int main()
{
// 這裏我強行指定了一個值
pid_t pid=6268;
// create a perf fd
struct perf_event_attr attr;
memset(&attr,0,sizeof(struct perf_event_attr));
attr.size=sizeof(struct perf_event_attr);
// disable at init time
attr.disabled=1;
// set what is the event
attr.type=PERF_TYPE_SOFTWARE;
attr.config=PERF_COUNT_SW_CPU_CLOCK;
// how many clocks to trigger sampling
attr.sample_period=1000000;
// what to sample is IP
attr.sample_type=PERF_SAMPLE_IP;
// notify every 1 overflow
attr.wakeup_events=1;
// open perf fd
int perf_fd=perf_event_open(&attr,pid,-1,-1,0);
if(perf_fd<0)
{
perror("perf_event_open() failed!");
return errno;
}
// create a shared memory to read samples from kernel
void* shared_mem=mmap(0,(1+RING_BUFFER_PAGES)*4096,PROT_READ,MAP_SHARED,perf_fd,0);
if(shared_mem==0)
{
perror("mmap() failed!");
return errno;
}
// reset and enable
ioctl(perf_fd,PERF_EVENT_IOC_RESET,0);
ioctl(perf_fd,PERF_EVENT_IOC_ENABLE,0);
// the offset from the head of ring-buffer where the next sample is
uint64_t next_offset=0;
// poll perf_fd
struct pollfd perf_poll;
perf_poll.fd=perf_fd;
perf_poll.events=POLLIN;
signal(SIGINT,on_closing);
while(sg_running)
{
if(poll(&perf_poll,1,-1)<0)
{
perror("poll() failed!");
break;
}
// the pointer to the completed sample
struct sample
{
struct perf_event_header header;
uint64_t ip;
}*
sample=(struct sample*)((uint8_t*)shared_mem+4096+next_offset);
// the pointer to the info structure of ring-buffer
struct perf_event_mmap_page* info=(struct perf_event_mmap_page*)shared_mem;
// update the offset, wrap the offset
next_offset=info->data_head%(RING_BUFFER_PAGES*4096);
// allow only the PERF_RECORD_SAMPLE
if(sample->header.type!=PERF_RECORD_SAMPLE)
continue;
printf("%lx\n",sample->ip);
}
printf("clean up\n");
// disable
ioctl(perf_fd,PERF_EVENT_IOC_DISABLE,0);
// unmap shared memory
munmap(shared_mem,(1+RING_BUFFER_PAGES)*4096);
// close perf fd
close(perf_fd);
return 0;
}
可以看到除了獲取通知的部分由signal改爲poll()以外,幾乎沒有改動。
g++ perf_poll.cpp -o perf_poll
sudo ./perf_poll
==2017年7月28日補充
首先,爲了方便以後的使用,我把perf採樣callchain的功能封裝成了一個C++的類,它能夠針對一個特定的pid進行採樣,支持帶有超時的輪詢。接口聲明如下:
CallChainSampler.h
#ifndef CALLCHAINSAMPLER_H
#define CALLCHAINSAMPLER_H
#include <stdint.h>
#include <unistd.h>
// a class to sample the callchain of a process
class CallChainSampler
{
public:
// the structure of a sampled callchain
struct callchain
{
// the timestamp when sampling
uint64_t time;
// the pid and tid
uint32_t pid,tid;
// the depth of callchain, or called the length
uint64_t depth;
// <depth>-array, each items is an IP register value
const uint64_t* ips;
};
// constructor
// pid: the process's id
// period: how many clocks to trigger a sample
// pages: how many pages (4K) allocated for the ring-buffer to hold samples
CallChainSampler(pid_t pid,uint64_t period,uint32_t pages);
// destructor
~CallChainSampler();
// start sampling
void start();
// stop sampling
void stop();
// wait and get the next sample
// timeout: the max milliseconds that will block
// max_depth: the max depth of the call chain
// callchain: the sampled callchain to be outputed
// return: if get before timeout, return 0,
// if timeout, return -1
// if an error occurs, return errno
// ATTENTION: the field [ips] in callchain should be used immediately,
// don't hold it for too long time
int sample(int32_t timeout,uint64_t max_depth,struct callchain* callchain);
private:
// the perf file descriptor
int fd;
// the mmap area
void* mem;
// how many pages to hold the ring-buffer
uint32_t pages;
// the offset in the ring-buffer where the next sample is
uint64_t offset;
};
#endif
實現基本就是把上面的C代碼封裝一下:
CallChainSampler.cpp
#include "CallChainSampler.h"
#include <poll.h>
#include <errno.h>
#include <assert.h>
#include <string.h>
#include <stdexcept>
#include <sys/time.h>
#include <sys/mman.h>
#include <sys/ioctl.h>
#include <sys/syscall.h>
#include <linux/perf_event.h>
// a wrapper for perf_event_open()
static int perf_event_open(struct perf_event_attr *attr,
pid_t pid,int cpu,int group_fd,unsigned long flags)
{
return syscall(__NR_perf_event_open,attr,pid,cpu,group_fd,flags);
}
// a tool function to get the time in ms
static uint64_t get_milliseconds()
{
struct timeval now;
assert(gettimeofday(&now,0)==0);
return now.tv_sec*1000+now.tv_usec/1000;
}
#define min(a,b) ((a)<(b)?(a):(b))
CallChainSampler::CallChainSampler(pid_t pid,uint64_t period,uint32_t pages)
{
// create a perf fd
struct perf_event_attr attr;
memset(&attr,0,sizeof(struct perf_event_attr));
attr.size=sizeof(struct perf_event_attr);
// disable at init time
attr.disabled=1;
// set what is the event
attr.type=PERF_TYPE_SOFTWARE;
attr.config=PERF_COUNT_SW_CPU_CLOCK;
// how many clocks to trigger sampling
attr.sample_period=period;
// what to sample is IP
attr.sample_type=PERF_SAMPLE_TIME|PERF_SAMPLE_TID|PERF_SAMPLE_CALLCHAIN;
// notify every 1 overflow
attr.wakeup_events=1;
// open perf fd
fd=perf_event_open(&attr,pid,-1,-1,0);
if(fd<0)
throw std::runtime_error("perf_event_open() failed!");
// create a shared memory to read samples from kernel
mem=mmap(0,(1+pages)*4096,PROT_READ,MAP_SHARED,fd,0);
if(mem==0)
throw std::runtime_error("mmap() failed!");
this->pages=pages;
// the offset of next sample
offset=0;
}
CallChainSampler::~CallChainSampler()
{
stop();
// unmap shared memory
munmap(mem,(1+pages)*4096);
// close perf fd
close(fd);
}
void CallChainSampler::start()
{
// enable
ioctl(fd,PERF_EVENT_IOC_ENABLE,0);
}
void CallChainSampler::stop()
{
// disable
ioctl(fd,PERF_EVENT_IOC_DISABLE,0);
}
int CallChainSampler::sample(int32_t timeout,uint64_t max_depth,struct callchain* callchain)
{
if(callchain==0)
throw std::runtime_error("arg <callchain> is NULL!");
// the poll sturct
struct pollfd pfd;
pfd.fd=fd;
pfd.events=POLLIN;
// the time when start
uint64_t start=get_milliseconds();
while(1)
{
// the current time
uint64_t now=get_milliseconds();
// the milliseconds to wait
int32_t to_wait;
if(timeout<0)
to_wait=-1;
else
{
to_wait=timeout-(int32_t)(now-start);
if(to_wait<0)
return -1;
}
// wait next sample
int ret=poll(&pfd,1,to_wait);
if(ret==0)
return -1;
else if(ret==-1)
return errno;
// the pointer to the completed sample
struct sample
{
struct perf_event_header header;
uint32_t pid,tid;
uint64_t time;
uint64_t nr;
uint64_t ips[0];
}*
sample=(struct sample*)((uint8_t*)mem+4096+offset);
// the pointer to the info structure of ring-buffer
struct perf_event_mmap_page* info=(struct perf_event_mmap_page*)mem;
// update the offset, wrap the offset
offset=info->data_head%(pages*4096);
// allow only the PERF_RECORD_SAMPLE
if(sample->header.type!=PERF_RECORD_SAMPLE)
continue;
// fill the result
callchain->time=sample->time;
callchain->pid=sample->pid;
callchain->tid=sample->tid;
callchain->depth=min(max_depth,sample->nr);
callchain->ips=sample->ips;
return 0;
}
}
最後要補充一個我最新的發現!perf_event_open()裏面傳入的pid,本質上是一個線程id,也就是tid。它只能監控一個線程,而無法監控一個進程中的所有線程。所以要用到實際項目中,肯定得配合使用epoll來監控所有的線程。
測試代碼如下:
#include <stdio.h>
#include <signal.h>
#include <stdlib.h>
#include "CallChainSampler.h"
CallChainSampler* sampler;
// to receive SIGINT to stop sampling and exit
static void on_closing(int signum)
{
delete sampler;
exit(0);
}
int main()
{
// create a sampler, pid=5281, 10000 clocks trigger a sample
// and allocate 128 pages to hold the ring-buffer
sampler=new CallChainSampler(5281,10000,128);
signal(SIGINT,on_closing);
sampler->start();
for(int i=0;i<10000;i++)
{
CallChainSampler::callchain callchain;
// sample, max depth of callchain is 256
int ret=sampler->sample(-1,256,&callchain);
printf("%d\n",ret);
if(ret==0)
{
// successful sample, print it out
printf("time=%lu\n",callchain.time);
printf("pid,tid=%d,%d\n",callchain.pid,callchain.tid);
printf("stack:\n");
for(int j=0;j<callchain.depth;j++)
printf("[%d] %lx\n",j,callchain.ips[j]);
}
}
return 0;
}
示例程序3
#define _GNU_SOURCE
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <sys/syscall.h>
#include <string.h>
#include <sys/ioctl.h>
#include <linux/perf_event.h>
#include <sys/mman.h>
#include <linux/hw_breakpoint.h>
#include <asm/unistd.h>
#include <errno.h>
#include <stdint.h>
#include <inttypes.h>
#ifndef MAP_FAILED
#define MAP_FAILED ((void *)-1)
#endif
struct perf_sample_event {
struct perf_event_header hdr;
uint64_t sample_id; // if PERF_SAMPLE_IDENTIFIER
uint64_t ip; // if PERF_SAMPLE_IP
uint32_t pid, tid; // if PERF_SAMPLE_TID
uint64_t time; // if PERF_SAMPLE_TIME
uint64_t addr; // if PERF_SAMPLE_ADDR
uint64_t id; // if PERF_SAMPLE_ID
uint64_t stream_id; // if PERF_SAMPLE_STREAM_ID
uint32_t cpu, res; // if PERF_SAMPLE_CPU
uint64_t period; // if PERF_SAMPLE_PERIOD
struct read_format *v; // if PERF_SAMPLE_READ
uint64_t nr; // if PERF_SAMPLE_CALLCHAIN
uint64_t *ips; // if PERF_SAMPLE_CALLCHAIN
uint32_t size_raw; // if PERF_SAMPLE_RAW
char *data_raw; // if PERF_SAMPLE_RAW
uint64_t bnr; // if PERF_SAMPLE_BRANCH_STACK
struct perf_branch_entry *lbr; // if PERF_SAMPLE_BRANCH_STACK
uint64_t abi; // if PERF_SAMPLE_REGS_USER
uint64_t *regs; // if PERF_SAMPLE_REGS_USER
uint64_t size_stack; // if PERF_SAMPLE_STACK_USER
char *data_stack; // if PERF_SAMPLE_STACK_USER
uint64_t dyn_size_stack; // if PERF_SAMPLE_STACK_USER
uint64_t weight; // if PERF_SAMPLE_WEIGHT
uint64_t data_src; // if PERF_SAMPLE_DATA_SRC
uint64_t transaction; // if PERF_SAMPLE_TRANSACTION
uint64_t abi_intr; // if PERF_SAMPLE_REGS_INTR
uint64_t *regs_intr; // if PERF_SAMPLE_REGS_INTR
};
int fib(int n) {
if (n == 0) {
return 0;
} else if (n == 1 || n == 2) {
return 1;
} else {
return fib(n-1) + fib(n-2);
}
}
void do_something() {
int i;
char* ptr;
ptr = malloc(100*1024*1024);
for (i = 0; i < 100*1024*1024; i++) {
ptr[i] = (char) (i & 0xff); // pagefault
}
free(ptr);
}
void insertion_sort(int *nums, size_t n) {
int i = 1;
while (i < n) {
int j = i;
while (j > 0 && nums[j-1] > nums[j]) {
int tmp = nums[j];
nums[j] = nums[j-1];
nums[j-1] = tmp;
j -= 1;
}
i += 1;
}
}
static void process_ring_buffer_events(struct perf_event_mmap_page *data, int page_size) {
struct perf_event_header *header = (uintptr_t) data + page_size + data->data_tail;
void *end = (uintptr_t) data + page_size + data->data_head;
while (header != end) {
if (header->type == PERF_RECORD_SAMPLE) {
struct perf_sample_event *event = (uintptr_t) header;
uint64_t ip = event->ip;
printf("PERF_RECORD_SAMPLE found with ip: %lld\n", ip);
uint64_t size_stack = event->size_stack;
char *data_stack = (uintptr_t) event->data_stack;
if (data_stack > 0) {
printf("PERF_RECORD_SAMPLE has size stack: %lld at location: %lld\n", size_stack, data_stack);
}
} else {
printf("other type %d found!", header->type);
}
header = (uintptr_t) header + header->size;
}
}
int main(int argc, char* argv[]) {
struct perf_event_attr pea;
int fd1, fd2;
uint64_t id1, id2;
uint64_t val1, val2;
char buf[4096];
const int NUM_MMAP_PAGES = (1U << 4) + 1;
// const int NUM_MMAP_PAGES = 17;
int i;
int some_nums[1000];
for (int i=0; i < 1000; i++) {
some_nums[i] = 1000-i;
}
memset(&pea, 0, sizeof(struct perf_event_attr));
pea.type = PERF_TYPE_SOFTWARE;
pea.size = sizeof(struct perf_event_attr);
pea.config = PERF_COUNT_SW_CPU_CLOCK;
pea.disabled = 1;
pea.exclude_kernel = 1;
pea.exclude_hv = 0;
pea.sample_period = 1;
pea.precise_ip = 3;
pea.sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_STACK_USER;
// pea.sample_type = PERF_SAMPLE_IP;
pea.sample_stack_user = 10000;
fd1 = syscall(__NR_perf_event_open, &pea, 0, -1, -1, 0);
printf("size of perf_event_mmap_page struct is %d\n", sizeof(struct perf_event_mmap_page));
int page_size = (int) sysconf(_SC_PAGESIZE);
printf("page size in general is: %d\n", page_size);
// Map the ring buffer into memory
struct perf_event_mmap_page *pages = mmap(NULL, page_size * NUM_MMAP_PAGES, PROT_READ
| PROT_WRITE, MAP_SHARED, fd1, 0);
if (pages == MAP_FAILED) {
perror("Error mapping ring buffer");
return 1;
}
ioctl(fd1, PERF_EVENT_IOC_RESET, PERF_IOC_FLAG_GROUP);
ioctl(fd1, PERF_EVENT_IOC_ENABLE, PERF_IOC_FLAG_GROUP);
// do_something();
// fib(40);
size_t n = sizeof(some_nums)/sizeof(some_nums[0]);
insertion_sort(&some_nums, n);
ioctl(fd1, PERF_EVENT_IOC_DISABLE, PERF_IOC_FLAG_GROUP);
printf("head of perf ring buffer is at: %d", pages->data_head);
process_ring_buffer_events(pages, page_size);
munmap(pages, page_size * NUM_MMAP_PAGES);
return 0;
}