本篇介绍
本篇介绍ebpf tracing 部分, 通过bcc(BPF Compiler Collection)获取多种渠道的系统信息.
probe 介绍
probe在很多书上叫做探针,是一种获取执行环境动态信息的方法,通过probe 就可以分析应用和系统的运行情况,对于分析应用和系统行为有很大帮助.
传统的probe是编写代码成内核模块ko,然后加载到内核中执行,这样如果probe代码本身有问题就会导致系统也出现问题,比如crash等.
而bpf由于有bpf verifier把关,就可以保证probe代码的可靠.
我们接下来讨论下probe的类型以及各种类型下bcc的使用.
kernel probes
通过在内核指令上设置一些动态标记来获取运行信息,对系统的负荷很小.当内核遇到这些标记后,会先执行对应的probe代码,执行完后再继续进行内核指令.这就如同是对某个函数进行了inline hook,进入函数后,先执行probe代码,再跳转回函数继续执行.
由于这种操作依赖内核的指令接口,而这些接口并没有标准的ABI 定义,因此容易遇到不同内核版本的兼容问题.
kernel probes 有两种,分别是kprobes和kretprobes,从名字上可以看出来,前者是用来监控指令入口,后者是指令出口.
接下来我们用c++ 演示下用BCC实现kprobes, demo 代码如下:
#include <string>
#include <iostream>
#include <stdlib.h>
#include <unistd.h>
#include <fstream>
#include <bcc/BPF.h>
const std::string BPF_PROGRAM = R"(
#include <linux/ptrace.h>
int do_sys_execve(struct pt_regs *ctx, char *filename, char *argv, char * envp) {
char comm[16] = {0};
bpf_get_current_comm(comm, sizeof(comm));
bpf_trace_printk("execve new process %s\n", comm);
return 0;
}
int ret_sys_execve(struct pt_regs *ctx) {
int return_value = PT_REGS_RC(ctx);
bpf_trace_printk("program return value is %d\n", return_value);
return 0;
}
)";
int main(){
ebpf::BPF bpf;
auto init_res = bpf.init(BPF_PROGRAM);
if (!init_res.ok()) {
std::cout<<init_res.msg() << std::endl;
return 1;
}
std::cout<<"bpf init ok"<<std::endl;
std::string execv_fnname = bpf.get_syscall_fnname("execve");
auto attach_res = bpf.attach_kprobe(execv_fnname, "do_sys_execve", 0, BPF_PROBE_ENTRY);
if (!attach_res.ok()) {
std::cout<<attach_res.msg()<<std::endl;
return 1;
}
attach_res = bpf.attach_kprobe(execv_fnname, "ret_sys_execve", 0, BPF_PROBE_RETURN);
if (!attach_res.ok()) {
std::cout<<attach_res.msg()<<std::endl;
return 1;
}
std::cout<<"bpf attach ok"<<std::endl;
std::ifstream pipe("/sys/kernel/debug/tracing/trace_pipe");
int count = 0;
while(true) {
std::string line;
if(std::getline(pipe,line)) {
std::cout<<line<<std::endl;
count++;
if (count > 50) {
bpf.detach_kprobe(execv_fnname, BPF_PROBE_ENTRY);
bpf.detach_kprobe(execv_fnname, BPF_PROBE_RETURN);
break;
}
} else {
sleep(1);
}
}
return 0;
}
编译如上代码需要拥有bcc环境, 一般是通过源码配置,具体可以参考bcc install md, 可能在配置时候也会出现一些问题, 具体看配置环境,我目前的机器是Ubuntu 24.04.1 LTS, 内核是Linux version 6.8.0-44-generic(我刚开始工作时, 使用的内核版本还是3.9, 现在都6.8了, 一丝唏嘘流过) ,配置好后, 编译即可运行:
bpf init ok
bpf attach ok
<...>-518742 [001] ...21 89067.095248: bpf_trace_printk: execve new process code
<...>-518742 [001] ...21 89067.095268: bpf_trace_printk: program return value is -2
<...>-518742 [001] ...21 89067.095286: bpf_trace_printk: execve new process code
<...>-518742 [001] ...21 89067.095292: bpf_trace_printk: program return value is -2
<...>-518742 [001] ...21 89067.095298: bpf_trace_printk: execve new process code
<...>-518742 [001] ...21 89067.095305: bpf_trace_printk: program return value is -2
<...>-518742 [001] ...21 89067.095311: bpf_trace_printk: execve new process code
<...>-518742 [001] ...21 89067.095317: bpf_trace_printk: program return value is -2
<...>-518742 [001] ...21 89067.095323: bpf_trace_printk: execve new process code
<...>-518742 [001] ...21 89067.095329: bpf_trace_printk: program return value is -2
<...>-518742 [001] ...21 89067.095335: bpf_trace_printk: execve new process code
<...>-518742 [001] ...21 89067.100773: bpf_trace_printk: program return value is 0
<...>-518749 [000] ...21 89067.625411: bpf_trace_printk: execve new process code
<...>-518749 [000] ...21 89067.625429: bpf_trace_printk: program return value is -2
<...>-518749 [000] ...21 89067.625438: bpf_trace_printk: execve new process code
<...>-518749 [000] ...21 89067.625442: bpf_trace_printk: program return value is -2
<...>-518749 [000] ...21 89067.625445: bpf_trace_printk: execve new process code
<...>-518749 [000] ...21 89067.625448: bpf_trace_printk: program return value is -2
<...>-518749 [000] ...21 89067.625451: bpf_trace_printk: execve new process code
<...>-518749 [000] ...21 89067.625454: bpf_trace_printk: program return value is -2
<...>-518749 [000] ...21 89067.625456: bpf_trace_printk: execve new process code
<...>-518749 [000] ...21 89067.629358: bpf_trace_printk: program return value is 0
<...>-518756 [003] ...21 89069.133218: bpf_trace_printk: execve new process code
<...>-518756 [003] ...21 89069.133239: bpf_trace_printk: program return value is -2
<...>-518756 [003] ...21 89069.133257: bpf_trace_printk: execve new process code
<...>-518756 [003] ...21 89069.133263: bpf_trace_printk: program return value is -2
上述方法涉及到3个关键函数:
StatusTuple init(const std::string& bpf_program,
const std::vector<std::string>& cflags = {},
const std::vector<USDT>& usdt = {});
StatusTuple attach_kprobe(const std::string& kernel_func,
const std::string& probe_func,
uint64_t kernel_func_offset = 0,
bpf_probe_attach_type = BPF_PROBE_ENTRY,
int maxactive = 0);
StatusTuple detach_kprobe(
const std::string& kernel_func,
bpf_probe_attach_type attach_type = BPF_PROBE_ENTRY);
差不多从函数参数上可以看出来对应的用法.
这儿需要补充一个信息,probe函数中第一个参数总是struct pt_regs*, 这个是什么呢? 这个存储了cpu运行时对应进程的上下文信息,具体内容和处理器有关,比如arm,x86等. 比如我个人接触较多的arm64结构如下:
/*
* This struct defines the way the registers are stored on the stack during an
* exception. Note that sizeof(struct pt_regs) has to be a multiple of 16 (for
* stack alignment). struct user_pt_regs must form a prefix of struct pt_regs.
*/
struct pt_regs {
union {
struct user_pt_regs user_regs;
struct {
u64 regs[31];
u64 sp;
u64 pc;
u64 pstate;
};
};
u64 orig_x0;
#ifdef __AARCH64EB__
u32 unused2;
s32 syscallno;
#else
s32 syscallno;
u32 unused2;
#endif
u64 sdei_ttbr1;
/* Only valid when ARM64_HAS_GIC_PRIO_MASKING is enabled. */
u64 pmr_save;
u64 stackframe[2];
/* Only valid for some EL1 exceptions. */
u64 lockdep_hardirqs;
u64 exit_rcu;
};
从上面的demo也可以看出来, kprobe以来于内核的实现,存在一定的兼容性风险,接下来我们看一种稳定的方法.
Tracepoints
相比与kprobe,tracepoints位于内核中的固定位置,这就是为什么tracepoint叫做静态标记,这些标记在内核中会保证兼容,也就是老版本有的tracepoints 在新版本中也一定有, 不过也正因为这样,所以系统中tracepoints比较有限,不如kprobe那么灵活.
可以利用如下命令查看系统中支持的tracepoints:
root@shanks-ThinkPad-T460s:/sys/kernel/debug/tracing/events# ls
alarmtimer devlink hda_intel irq mdio page_isolation rv thp
amd_cpu dma_fence header_event irq_matrix mei pagemap sched timer
asoc drm header_page irq_vectors migrate page_pool scsi tlb
avc e1000e_trace huge_memory iwlwifi mmap percpu sd tls
block enable hwmon iwlwifi_data mmap_lock power signal udp
bpf_test_run error_report hyperv iwlwifi_io mmc printk skb v4l2
bpf_trace exceptions i2c iwlwifi_msg module pwm smbus vb2
bridge ext4 i915 iwlwifi_ucode mptcp qdisc sock vmalloc
cfg80211 fib initcall jbd2 msr qrtr sof vmscan
cgroup fib6 intel_avs kmem napi ras sof_intel vsyscall
clk filelock intel_iommu ksm neigh raw_syscalls spi watchdog
compaction filemap intel-sst libata net rcu swiotlb wbt
context_tracking fs_dax interconnect lock netlink regmap sync_trace workqueue
cpuhp ftrace iocost mac80211 nmi regulator syscalls writeback
cros_ec gpio iomap mac80211_msg notifier resctrl task x86_fpu
csd handshake iommu maple_tree nvme rpm tcp xdp
dev hda io_uring mce oom rseq thermal xen
devfreq hda_controller ipi mctp osnoise rtc thermal_power_allocator xhci-hcd
每个目录下分别存在enable和filter两个文件,前者用来使能当前tracepoints,后者用来事件过滤. 接下来我们看一个例子:
#include <string>
#include <iostream>
#include <stdlib.h>
#include <unistd.h>
#include <fstream>
#include <bcc/BPF.h>
const std::string BPF_PROGRAM = R"(
#include <linux/ptrace.h>
#include <linux/sched.h>
int on_sched_switch(struct tracepoint__sched__sched_switch *args) {
bpf_trace_printk("sched switch %s\n",args);
return 0;
}
)";
int main(){
ebpf::BPF bpf;
auto init_res = bpf.init(BPF_PROGRAM);
if (!init_res.ok()) {
std::cout<<init_res.msg() << std::endl;
return 1;
}
std::cout<<"bpf init ok"<<std::endl;
auto attach_res = bpf.attach_tracepoint("sched:sched_switch", "on_sched_switch");
if (!attach_res.ok()) {
std::cout<<attach_res.msg() << std::endl;
return 1;
}
std::cout<<"attach tracepoint ok"<<std::endl;
std::ifstream pipe("/sys/kernel/debug/tracing/trace_pipe");
int count = 0;
while(true) {
std::string line;
if(std::getline(pipe,line)) {
std::cout<<line<<std::endl;
count++;
if (count > 10) {
bpf.detach_tracepoint("sched:sched_switch");
break;
}
} else {
sleep(1);
}
}
return 0;
}
编译后运行,即可得到如下结果:
bpf init ok
attach tracepoint ok
<idle>-0 [003] d..2. 108816.680677: sched_switch: prev_comm=swapper/3 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=docker next_pid=652525 next_prio=120
docker-652525 [003] d..2. 108816.680683: sched_switch: prev_comm=docker prev_pid=652525 prev_prio=120 prev_state=S ==> next_comm=swapper/3 next_pid=0 next_prio=120
<idle>-0 [003] d..2. 108816.681607: sched_switch: prev_comm=swapper/3 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=docker next_pid=652525 next_prio=120
docker-652525 [003] d..2. 108816.681618: sched_switch: prev_comm=docker prev_pid=652525 prev_prio=120 prev_state=S ==> next_comm=swapper/3 next_pid=0 next_prio=120
<idle>-0 [003] d..2. 108816.681623: sched_switch: prev_comm=swapper/3 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=docker next_pid=652525 next_prio=120
docker-652525 [003] d..2. 108816.681649: sched_switch: prev_comm=docker prev_pid=652525 prev_prio=120 prev_state=S ==> next_comm=swapper/3 next_pid=0 next_prio=120
<idle>-0 [003] d..2. 108816.681710: sched_switch: prev_comm=swapper/3 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=docker next_pid=652525 next_prio=120
docker-652525 [003] d..2. 108816.681724: sched_switch: prev_comm=docker prev_pid=652525 prev_prio=120 prev_state=S ==> next_comm=swapper/3 next_pid=0 next_prio=120
<idle>-0 [003] d..2. 108816.681739: sched_switch: prev_comm=swapper/3 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=docker next_pid=652525 next_prio=120
docker-652525 [003] d..2. 108816.681750: sched_switch: prev_comm=docker prev_pid=652525 prev_prio=120 prev_state=S ==> next_comm=swapper/3 next_pid=0 next_prio=120
<idle>-0 [003] d..2. 108816.681783: sched_switch: prev_comm=swapper/3 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=docker next_pid=652525 next_prio=120
uprobes
看了内核态的操作后,很容易问用户态是否支持类似操作? 其实用户态也一样,也可以注册probes,也分为不稳定的动态probe和稳定的静态probe.
现在先介绍下不稳定的uprobes,这个操作需要依赖函数的签名,如果签名变了,那么probe就失效了. 接下来我们看一个例子, 目的是测量某个用户态函数的执行耗时,用户态对应的代码如下:
#include <iostream>
#include <unistd.h>
int foo1(int v) {
return v + 1;
}
int main() {
int v = 1;
v = foo1(v);
std::cout<<v<<std::endl;
return 0;
}
这儿测试的目标是foo1. 接下来我们看下如何不修改代码实现probe该函数, 对应的bcc代码如下:
#include <string>
#include <iostream>
#include <stdlib.h>
#include <unistd.h>
#include <fstream>
#include <bcc/BPF.h>
const std::string BPF_PROGRAM = R"(
#include <linux/ptrace.h>
#include <linux/sched.h>
BPF_HASH(cache, u64,u64);
int trace_start_time(struct pt_regs *ctx) {
u64 pid = bpf_get_current_pid_tgid();
u64 start_time_ns = bpf_ktime_get_ns();
cache.update(&pid, &start_time_ns);
return 0;
}
int print_duration(struct pt_regs *ctx) {
u64 pid = bpf_get_current_pid_tgid();
u64 *start_time_ns = cache.lookup(&pid);
if (start_time_ns == NULL) {
return 0;
}
u64 duration_ns = bpf_ktime_get_ns() - *start_time_ns;
bpf_trace_printk("call duration %d\n", duration_ns);
return 0;
}
)";
int main(){
ebpf::BPF bpf;
auto init_res = bpf.init(BPF_PROGRAM);
if (!init_res.ok()) {
std::cout<<init_res.msg() << std::endl;
return 1;
}
std::cout<<"bpf init ok"<<std::endl;
auto attach_res = bpf.attach_uprobe("./user.bin", "_Z4foo1i", "trace_start_time", 0, BPF_PROBE_ENTRY);
if (!attach_res.ok()) {
std::cout<<attach_res.msg() << std::endl;
return 1;
}
attach_res = bpf.attach_uprobe("./user.bin", "_Z4foo1i", "print_duration", 0, BPF_PROBE_RETURN);
if (!attach_res.ok()) {
std::cout<<attach_res.msg() << std::endl;
return 1;
}
std::cout<<"attach uprobe ok"<<std::endl;
std::ifstream pipe("/sys/kernel/debug/tracing/trace_pipe");
int count = 0;
while(true) {
std::string line;
if(std::getline(pipe,line)) {
std::cout<<line<<std::endl;
if (!line.empty()) {
count++;
}
if (count > 10) {
bpf.detach_uprobe("./user.bin", "_Z4foo1i", 0, BPF_PROBE_ENTRY);
bpf.detach_uprobe("./user.bin", "_Z4foo1i", 0, BPF_PROBE_RETURN);
break;
}
} else {
sleep(1);
}
}
return 0;
}
编译并运行:
bpf init ok
attach uprobe ok
<...>-696974 [003] ...11 115306.935932: bpf_trace_printk: call duration 16884
<...>-696982 [001] ...11 115307.566014: bpf_trace_printk: call duration 18168
<...>-696989 [001] ...11 115308.106875: bpf_trace_printk: call duration 47780
<...>-696990 [001] ...11 115308.660956: bpf_trace_printk: call duration 18897
<...>-696998 [001] ...11 115309.261987: bpf_trace_printk: call duration 55242
<...>-696999 [001] ...11 115309.876613: bpf_trace_printk: call duration 17145
<...>-697007 [003] ...11 115310.477063: bpf_trace_printk: call duration 49513
<...>-697015 [002] ...11 115311.181669: bpf_trace_printk: call duration 47936
<...>-697016 [003] ...11 115311.841587: bpf_trace_printk: call duration 17201
<...>-697024 [000] ...11 115312.381908: bpf_trace_printk: call duration 51744
<...>-697097 [000] ...11 115324.187386: bpf_trace_printk: call duration 48442
可以看到foo的执行耗时已经成功获取到了.实现uprobe的关键API如下:
StatusTuple attach_uprobe(const std::string& binary_path,
const std::string& symbol,
const std::string& probe_func,
uint64_t symbol_addr = 0,
bpf_probe_attach_type attach_type = BPF_PROBE_ENTRY,
pid_t pid = -1,
uint64_t symbol_offset = 0,
uint32_t ref_ctr_offset = 0);
StatusTuple detach_uprobe(const std::string& binary_path,
const std::string& symbol, uint64_t symbol_addr = 0,
bpf_probe_attach_type attach_type = BPF_PROBE_ENTRY,
pid_t pid = -1,
uint64_t symbol_offset = 0);
差不多也是从API 参数就可以看到作用.
USDT
User Statically Defined Tracepoints 是用户态的"tracepoints", 也可以稳定获取用户态代码的信息. usdt 很大程度上是参考的DTrace,一会儿可以从使用上看出来. 由于USDT是稳定的probe方法,因此就要求用户态代码显式定义USDT. 接下来我们看一个例子,用户态如何定义usdt.
#include <sys/sdt.h>
#include <unistd.h>
int main() {
DTRACE_PROBE(hello-usdt, probe-main);
return 0;
}
就这么简单, 定义后, 直接编译成elf,利用readelf 命令就可以看到对应的usdt信息.
shanks@shanks-ThinkPad-T460s:~/Documents/01code/ebpf/chapter03$ readelf -n ./usdt.bin
Displaying notes found in: .note.gnu.property
Owner Data size Description
GNU 0x00000020 NT_GNU_PROPERTY_TYPE_0
Properties: x86 feature: IBT, SHSTK
x86 ISA needed: x86-64-baseline
Displaying notes found in: .note.gnu.build-id
Owner Data size Description
GNU 0x00000014 NT_GNU_BUILD_ID (unique build ID bitstring)
Build ID: 7054072f29c24f0617dae26545233b7107d84668
Displaying notes found in: .note.ABI-tag
Owner Data size Description
GNU 0x00000010 NT_GNU_ABI_TAG (ABI version tag)
OS: Linux, ABI: 3.2.0
Displaying notes found in: .note.stapsdt
Owner Data size Description
stapsdt 0x0000002f NT_STAPSDT (SystemTap probe descriptors)
Provider: hello-usdt
Name: probe-main
Location: 0x0000000000001131, Base: 0x0000000000002004, Semaphore: 0x0000000000000000
Arguments:
接下来我们看下如何利用bcc 进行probe:
#include <string>
#include <iostream>
#include <stdlib.h>
#include <unistd.h>
#include <fstream>
#include <bcc/BPF.h>
const std::string BPF_PROGRAM = R"(
#include <linux/ptrace.h>
#include <linux/sched.h>
int trace_usdt_exec(struct pt_regs *ctx) {
u64 pid = bpf_get_current_pid_tgid();
bpf_trace_printk("usdt process running with pid %d\n", pid);
return 0;
}
)";
int main(){
ebpf::USDT usdt("./usdt.bin", "hello-usdt", "probe-main", "trace_usdt_exec");
ebpf::BPF bpf;
auto res = bpf.init(BPF_PROGRAM, {}, {usdt});
if (!res.ok()) {
std::cout<<res.msg()<<std::endl;
return 1;
}
std::cout<<"bpf init ok"<<std::endl;
res = bpf.attach_usdt(usdt);
if (!res.ok()) {
std::cout<<res.msg() << std::endl;
return 1;
}
std::cout<<"bpf attach usdt ok"<<std::endl;
std::ifstream pipe("/sys/kernel/debug/tracing/trace_pipe");
int count = 0;
while(true) {
std::string line;
if(std::getline(pipe,line)) {
std::cout<<line<<std::endl;
if (!line.empty()) {
count++;
}
if (count > 10) {
bpf.detach_usdt(usdt);
break;
}
} else {
sleep(1);
}
}
return 0;
}
同样编译后,运行即可看到结果如下:
bpf init ok
bpf attach usdt ok
<...>-732507 [003] ...11 120226.668339: bpf_trace_printk: usdt process running with pid 732507
<...>-732510 [003] ...11 120227.210096: bpf_trace_printk: usdt process running with pid 732510
<...>-732519 [002] ...11 120227.753714: bpf_trace_printk: usdt process running with pid 732519
<...>-732529 [001] ...11 120228.306039: bpf_trace_printk: usdt process running with pid 732529
<...>-732532 [003] ...11 120228.834264: bpf_trace_printk: usdt process running with pid 732532
<...>-732535 [002] ...11 120229.384237: bpf_trace_printk: usdt process running with pid 732535
<...>-732545 [002] ...11 120229.972352: bpf_trace_printk: usdt process running with pid 732545
<...>-732555 [003] ...11 120230.823713: bpf_trace_printk: usdt process running with pid 732555
<...>-732558 [003] ...11 120231.482447: bpf_trace_printk: usdt process running with pid 732558
<...>-732567 [000] ...11 120232.159065: bpf_trace_printk: usdt process running with pid 732567
<...>-732577 [000] ...11 120232.824309: bpf_trace_printk: usdt process running with pid 732577
这儿的关键函数是:
StatusTuple attach_usdt(const USDT& usdt, pid_t pid = -1);
StatusTuple detach_usdt(const USDT& usdt, pid_t pid = -1);
性能统计
看了上面内容可能会觉得ebpf很强大,实际上更强大. ebpf还可以进行性能统计! 这儿简单介绍下on cpu 和off cpu这两个在统计性能时非常重要的概念.
on-cpu: 就是cpu花到当前目标应用上的时间
off-cput: 就是cpu没花到目标应用上的时间
可能这样理解有点直白，实际上很有学问，比如这儿就可以提出问题，如何增加目标应用在on-cpu上的时间？
本节的目的主要是讲ebpf在性能统计上的用法，所以就简单提一下，后续有机会再专门讨论。先看一个统计内存引用命中率的case：
#include <string>
#include <iostream>
#include <stdlib.h>
#include <unistd.h>
#include <fstream>
#include <bcc/BPF.h>
#include <linux/perf_event.h>
#include <iomanip>
const std::string BPF_PROGRAM = R"(
#include <linux/ptrace.h>
#include <uapi/linux/bpf_perf_event.h>
struct event_t {
int cpu;
int pid;
char name[16];
};
BPF_HASH(ref_count, struct event_t);
BPF_HASH(miss_count, struct event_t);
static inline __attribute__((always_inline)) void get_key(struct event_t *key) {
key->cpu = bpf_get_smp_processor_id();
key->pid = bpf_get_current_pid_tgid();
bpf_get_current_comm(key->name, sizeof(key->name));
}
int on_cache_miss(struct bpf_perf_event_data *ctx) {
struct event_t key = {};
get_key(&key);
u64 zero = 0, *val;
val = miss_count.lookup_or_try_init(&key, &zero);
if (val) {
*val += ctx->sample_period;
}
return 0;info_t
}
int on_cache_ref(struct bpf_perf_event_data *ctx) {
struct event_t key = {};
get_key(&key);
u64 zero = 0, *val;
val = ref_count.lookup_or_try_init(&key, &zero);
if (val) {
*val += ctx->sample_period;
}
return 0;
}
)";
struct event_t {
int cpu;
int pid;
char name[16];
};
int main(){
ebpf::BPF bpf;
auto init_res = bpf.init(BPF_PROGRAM);
if (!init_res.ok()) {
std::cout<<init_res.msg() << std::endl;
return 1;
}
std::cout<<"bpf init ok"<<std::endl;
auto res = bpf.attach_perf_event(PERF_TYPE_HARDWARE,PERF_COUNT_HW_CACHE_REFERENCES, "on_cache_ref", 100, 0);
if (!res.ok()) {
std::cout<<res.msg()<<std::endl;
return 1;
}
res = bpf.attach_perf_event(PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_MISSES, "on_cache_miss", 100, 0);
if (!res.ok()) {
std::cout<<res.msg()<<std::endl;
return 1;
}
std::cout<<"bpf attach ok"<<std::endl;
std::cout<<"probing ..."<<std::endl;
sleep(5);
bpf.detach_perf_event(PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_REFERENCES);
bpf.detach_perf_event(PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_MISSES);
auto refs = bpf.get_hash_table<event_t, uint64_t>("ref_count");
auto misses = bpf.get_hash_table<event_t, uint64_t>("miss_count");
for (auto it : refs.get_table_offline()) {
uint64_t hit;
auto miss = misses[it.first];
hit = miss <= it.second? it.second -miss:0;
double ratio = double(hit) / double(it.second) * 100.0;
std::cout<<"pid " << std::setw(8) << std::setfill(' ') << it.first.pid;
std::cout<<std::setw(20) << std::setfill(' ')<< std::left << " (" +
std::string(it.first.name) + ")" <<std::right;
std::cout<<"on cpu "<<std::setw(2) << std::setfill(' ') << it.first.cpu;
std::cout<<" hit rate "<< std::setprecision(4)<<ratio<<"% ";
std::cout<<"(" << hit << "/" << it.second<<")" << std::endl;
}
return 0;
}
运行结果如下：
bpf init ok
bpf attach ok
probing ...
pid 22363 (code) on cpu 1 hit rate 0% (0/7100)
pid 3912 (gmain) on cpu 3 hit rate 73.91% (1700/2300)
pid 7843 (code) on cpu 0 hit rate 4.231% (1100/26000)
pid 7887 (code) on cpu 3 hit rate 0% (0/5000)
pid 723596 (kworker/u9:1) on cpu 3 hit rate 42.42% (22100/52100)
pid 723596 (kworker/u9:1) on cpu 1 hit rate 74.21% (70200/94600)
pid 782016 (docker) on cpu 3 hit rate 26.04% (4400/16900)
pid 1590 (containerd) on cpu 1 hit rate 77.78% (7700/9900)
pid 782020 (docker) on cpu 1 hit rate 0% (0/91800)
pid 782013 (docker) on cpu 0 hit rate 0% (0/30700)
pid 4356 (HangWatcher) on cpu 0 hit rate 10% (300/3000)
pid 782040 (code) on cpu 1 hit rate 0% (0/57500)
pid 1610 (containerd) on cpu 2 hit rate 39.19% (5800/14800)
pid 782039 (docker) on cpu 3 hit rate 0% (0/5000)
pid 97688 (code) on cpu 0 hit rate 33.33% (1100/3300)
pid 97663 (code) on cpu 1 hit rate 1.918% (4100/213800)
pid 529775 (cpptools-srv) on cpu 0 hit rate 71.84% (7400/10300)
pid 97845 (code) on cpu 0 hit rate 34.51% (3900/11300)
pid 7947 (code) on cpu 0 hit rate 0% (0/12500)
pid 205757 (HangWatcher) on cpu 0 hit rate 52.38% (3300/6300)
pid 782033 (code) on cpu 0 hit rate 2.41% (800/33200)
pid 780635 (kworker/0:0) on cpu 0 hit rate 64.47% (24500/38000)
pid 18 (migration/0) on cpu 0 hit rate 50% (500/1000)
pid 3236 (JS Helper) on cpu 0 hit rate 63.33% (1900/3000)
pid 2326 (dockerd) on cpu 1 hit rate 88.89% (2400/2700)
pid 782046 (docker) on cpu 3 hit rate 76.19% (1600/2100)
pid 782017 (docker) on cpu 1 hit rate 14.88% (1800/12100)
pid 782030 (docker) on cpu 1 hit rate 47.06% (1600/3400)
pid 3726 (gvfs-afc-volume) on cpu 1 hit rate 35.71% (1500/4200)
pid 3226 (gnome-shell) on cpu 2 hit rate 18.91% (201300/1064400)
pid 35 (migration/3) on cpu 3 hit rate 81.48% (2200/2700)
pid 733067 (cpptools) on cpu 1 hit rate 82.35% (1400/1700)
pid 1753 (rtkit-daemon) on cpu 0 hit rate 39.29% (1100/2800)
pid 782037 (docker) on cpu 1 hit rate 80.49% (3300/4100)
pid 64 (kworker/3:1H) on cpu 3 hit rate 7.143% (800/11200)
pid 756219 (kworker/u8:2) on cpu 0 hit rate 28.1% (3400/12100)
pid 782035 (docker) on cpu 2 hit rate 50.63% (4000/7900)
pid 782028 (docker) on cpu 2 hit rate 77.57% (8300/10700)
pid 529963 (cpptools-srv) on cpu 1 hit rate 64.71% (6600/10200)
pid 1610 (containerd) on cpu 3 hit rate 1.347% (400/29700)
pid 97456 (Compositor) on cpu 2 hit rate 43.33% (2600/6000)
pid 782034 (docker) on cpu 3 hit rate 57.36% (7400/12900)
pid 4068 (sogoupinyinServ) on cpu 0 hit rate 19.98% (35800/179200)
pid 8508 (code) on cpu 1 hit rate 50% (1900/3800)
pid 4251 (chrome) on cpu 2 hit rate 0% (0/31500)
pid 782014 (docker) on cpu 0 hit rate 79.31% (6900/8700)
pid 5785 (gpg-agent) on cpu 1 hit rate 61.11% (2200/3600)
pid 8503 (code) on cpu 2 hit rate 42.11% (800/1900)
pid 1048 (NetworkManager) on cpu 1 hit rate 0% (0/2700)
pid 0 (swapper/1) on cpu 1 hit rate 79.07% (2133300/2697900)
pid 7883 (code) on cpu 2 hit rate 53.47% (5400/10100)
pid 7883 (code) on cpu 3 hit rate 62.07% (1800/2900)
pid 781990 (kworker/u9:2) on cpu 1 hit rate 79.04% (36200/45800)
pid 1610 (containerd) on cpu 1 hit rate 36.23% (22100/61000)
pid 132 (kworker/0:1H) on cpu 0 hit rate 0% (0/1500)
pid 4300 (chrome) on cpu 2 hit rate 0% (0/4000)
pid 97663 (code) on cpu 3 hit rate 0.1463% (200/136700)
pid 782026 (docker) on cpu 3 hit rate 0% (0/6400)
pid 17 (rcu_preempt) on cpu 0 hit rate 77.66% (22600/29100)
pid 98376 (cpptools) on cpu 2 hit rate 33.93% (1900/5600)
pid 782028 (docker) on cpu 1 hit rate 96.77% (3000/3100)
pid 257 (jbd2/nvme0n1p2-) on cpu 0 hit rate 0% (0/8500)
pid 782016 (docker) on cpu 1 hit rate 0% (0/1300)
pid 97446 (code) on cpu 1 hit rate 0% (0/15000)
pid 97674 (code) on cpu 3 hit rate 27.04% (6300/23300)
pid 1207 (gmain) on cpu 0 hit rate 0% (0/2500)
pid 782030 (docker) on cpu 2 hit rate 48.55% (6700/13800)
pid 97674 (code) on cpu 1 hit rate 86.49% (3200/3700)
pid 782015 (docker) on cpu 1 hit rate 50% (200/400)
pid 3226 (gnome-shell) on cpu 0 hit rate 52.27% (27600/52800)
pid 5804 (gnome-terminal-) on cpu 1 hit rate 4.598% (16300/354500)
pid 3256 (gnome-shell) on cpu 1 hit rate 0% (0/53200)
pid 97673 (code) on cpu 3 hit rate 0% (0/13000)
pid 1613 (containerd) on cpu 0 hit rate 30.89% (35400/114600)
pid 97684 (code) on cpu 0 hit rate 64.29% (4500/7000)
pid 756219 (kworker/u8:2) on cpu 1 hit rate 28.14% (6500/23100)
pid 7911 (code) on cpu 0 hit rate 0% (0/16300)
pid 779807 (kworker/u9:4) on cpu 1 hit rate 71.52% (75100/105000)
pid 782041 (docker) on cpu 3 hit rate 79.37% (5000/6300)
pid 733065 (cpptools) on cpu 0 hit rate 0% (0/100)
pid 16 (ksoftirqd/0) on cpu 0 hit rate 54.29% (1900/3500)
pid 782036 (docker) on cpu 3 hit rate 0% (0/6500)
pid 17 (rcu_preempt) on cpu 1 hit rate 64.39% (8500/13200)
pid 36 (ksoftirqd/3) on cpu 3 hit rate 60% (600/1000)
pid 8483 (code) on cpu 3 hit rate 0% (0/7000)
pid 23 (migration/1) on cpu 1 hit rate 87.5% (700/800)
pid 189281 (HangWatcher) on cpu 0 hit rate 8% (400/5000)
pid 782043 (docker) on cpu 2 hit rate 37.29% (6600/17700)
pid 551770 (cpptools-srv) on cpu 0 hit rate 0% (0/4300)
pid 782017 (docker) on cpu 3 hit rate 66.67% (1800/2700)
pid 781984 (tracker-extract) on cpu 1 hit rate 0% (0/2400)
pid 782035 (docker) on cpu 1 hit rate 31.82% (2800/8800)
pid 626 (irq/132-iwlwifi) on cpu 2 hit rate 0% (0/5700)
pid 8526 (code) on cpu 3 hit rate 61.9% (2600/4200)
pid 3235 (JS Helper) on cpu 1 hit rate 78.12% (2500/3200)
pid 29 (migration/2) on cpu 2 hit rate 66.67% (1600/2400)
pid 760435 (bluetoothd) on cpu 2 hit rate 7.647% (3900/51000)
pid 1590 (containerd) on cpu 0 hit rate 39.39% (1300/3300)
pid 3238 (KMS thread) on cpu 0 hit rate 57.74% (577000/999300)
pid 5804 (gnome-terminal-) on cpu 3 hit rate 22.27% (31200/140100)
pid 3226 (gnome-shell) on cpu 1 hit rate 38.27% (63100/164900)
pid 782014 (docker) on cpu 1 hit rate 66.96% (7700/11500)
pid 529521 (cpptools-srv) on cpu 0 hit rate 55.06% (4900/8900)
pid 717380 (kworker/2:2) on cpu 2 hit rate 56.07% (13400/23900)
pid 782015 (docker) on cpu 0 hit rate 1.562% (100/6400)
pid 1584 (containerd) on cpu 1 hit rate 53.14% (44000/82800)
pid 24 (ksoftirqd/1) on cpu 1 hit rate 46.63% (9700/20800)
pid 782019 (docker) on cpu 0 hit rate 31.25% (4500/14400)
pid 7851 (Chrome_IOThread) on cpu 1 hit rate 0% (0/7200)
pid 782013 (code) on cpu 3 hit rate 0% (0/33900)
pid 210799 (code) on cpu 0 hit rate 0% (0/11000)
pid 7974 (code) on cpu 3 hit rate 0% (0/5000)
pid 782024 (docker) on cpu 3 hit rate 39.13% (900/2300)
pid 4309 (chrome) on cpu 1 hit rate 0% (0/8400)
pid 97450 (Chrome_ChildIOT) on cpu 3 hit rate 0% (0/9400)
pid 4297 (chrome) on cpu 0 hit rate 14.29% (1200/8400)
pid 760435 (bluetoothd) on cpu 3 hit rate 0% (0/78800)
pid 17 (rcu_preempt) on cpu 3 hit rate 68.29% (14000/20500)
pid 7910 (CacheThread_Blo) on cpu 3 hit rate 0% (0/400)
pid 8492 (code) on cpu 0 hit rate 0% (0/138200)
pid 5804 (gnome-terminal-) on cpu 2 hit rate 27.8% (50200/180600)
pid 8504 (code) on cpu 0 hit rate 61.02% (3600/5900)
pid 8503 (code) on cpu 1 hit rate 50.43% (17600/34900)
pid 5804 (gnome-terminal-) on cpu 0 hit rate 22.74% (6800/29900)
pid 98373 (cpptools) on cpu 1 hit rate 0% (0/3000)
pid 782044 (docker) on cpu 0 hit rate 18.56% (3100/16700)
pid 3256 (gnome-shell) on cpu 0 hit rate 47.1% (12200/25900)
pid 122 (kworker/2:1H) on cpu 2 hit rate 51.69% (19900/38500)
pid 781987 (pool-tracker-ex) on cpu 0 hit rate 1.099% (100/9100)
pid 3256 (gnome-shell) on cpu 2 hit rate 52.11% (43300/83100)
pid 0 (swapper/0) on cpu 0 hit rate 75.81% (2063500/2722100)
pid 7926 (code) on cpu 0 hit rate 37.96% (4100/10800)
pid 768320 (kworker/1:2) on cpu 1 hit rate 40.46% (40300/99600)
pid 779807 (kworker/u9:4) on cpu 2 hit rate 76.34% (117800/154300)
pid 781999 (perf_event.bin) on cpu 3 hit rate 50% (7300/14600)
pid 782021 (docker) on cpu 0 hit rate 0% (0/1800)
pid 782031 (docker) on cpu 0 hit rate 32.2% (7600/23600)
pid 157242 (cpptools) on cpu 0 hit rate 0% (0/2100)
pid 547196 (cpptools) on cpu 1 hit rate 8.209% (1100/13400)
pid 7884 (code) on cpu 3 hit rate 0% (0/2000)
pid 744 (systemd-oomd) on cpu 0 hit rate 7.031% (1800/25600)
pid 1002 (avahi-daemon) on cpu 1 hit rate 0% (0/18400)
pid 782034 (docker) on cpu 1 hit rate 91.18% (3100/3400)
pid 782022 (docker) on cpu 0 hit rate 62.73% (6900/11000)
pid 782021 (docker) on cpu 2 hit rate 80.92% (12300/15200)
pid 782042 (docker) on cpu 2 hit rate 86.11% (6200/7200)
pid 782028 (docker) on cpu 3 hit rate 25.64% (1000/3900)
pid 5335 (chrome) on cpu 2 hit rate 12.17% (1400/11500)
pid 98376 (cpptools) on cpu 3 hit rate 12% (300/2500)
pid 8492 (code) on cpu 3 hit rate 2.302% (2700/117300)
pid 30 (ksoftirqd/2) on cpu 2 hit rate 57.14% (800/1400)
pid 8505 (code) on cpu 1 hit rate 44.74% (1700/3800)
pid 0 (swapper/3) on cpu 3 hit rate 79.53% (1942800/2443000)
pid 782027 (code) on cpu 0 hit rate 0% (0/15000)
pid 4173 (GUsbEventThread) on cpu 0 hit rate 33.87% (2100/6200)
pid 760435 (bluetoothd) on cpu 0 hit rate 11.59% (2400/20700)
pid 529521 (cpptools-srv) on cpu 2 hit rate 0% (0/1000)
pid 97450 (Chrome_ChildIOT) on cpu 1 hit rate 54% (2700/5000)
pid 223024 (HangWatcher) on cpu 2 hit rate 60.27% (4400/7300)
pid 7883 (code) on cpu 0 hit rate 0% (0/1500)
pid 7887 (code) on cpu 2 hit rate 8.125% (1300/16000)
pid 7843 (code) on cpu 2 hit rate 0% (0/31800)
pid 7978 (Chrome_ChildIOT) on cpu 1 hit rate 18.92% (1400/7400)
pid 767096 (cpptools-srv) on cpu 0 hit rate 0% (0/7100)
pid 3237 (JS Helper) on cpu 0 hit rate 56.14% (3200/5700)
pid 104929 (code) on cpu 2 hit rate 0% (0/1600)
pid 8483 (code) on cpu 1 hit rate 17.19% (1100/6400)
pid 782027 (docker) on cpu 0 hit rate 0% (0/30000)
pid 4081 (QDBusConnection) on cpu 3 hit rate 2.419% (300/12400)
pid 21523 (cpptools) on cpu 2 hit rate 44.78% (6000/13400)
pid 782040 (docker) on cpu 1 hit rate 0% (0/114900)
pid 782036 (docker) on cpu 1 hit rate 77.22% (6100/7900)
pid 782035 (docker) on cpu 3 hit rate 0% (0/12700)
pid 1754 (rtkit-daemon) on cpu 0 hit rate 0% (0/400)
pid 4305 (Chrome_ChildIOT) on cpu 0 hit rate 0% (0/3700)
pid 756219 (kworker/u8:2) on cpu 3 hit rate 0% (0/4500)
pid 98373 (cpptools) on cpu 0 hit rate 12% (600/5000)
pid 3806 (gmain) on cpu 0 hit rate 20% (600/3000)
pid 782033 (docker) on cpu 2 hit rate 0% (0/78500)
pid 17 (rcu_preempt) on cpu 2 hit rate 66.44% (9700/14600)
pid 723596 (kworker/u9:1) on cpu 0 hit rate 58.14% (2500/4300)
pid 97663 (code) on cpu 2 hit rate 0% (0/11500)
pid 1590 (containerd) on cpu 2 hit rate 5.967% (2500/41900)
pid 97674 (code) on cpu 0 hit rate 34.04% (6400/18800)
pid 1127 (gmain) on cpu 1 hit rate 0% (0/2800)
pid 647386 (kworker/3:0) on cpu 3 hit rate 76.49% (104100/136100)
pid 4302 (HangWatcher) on cpu 1 hit rate 21.43% (600/2800)
pid 4171 (gmain) on cpu 3 hit rate 0% (0/6400)
pid 1613 (containerd) on cpu 2 hit rate 36.92% (2400/6500)
pid 21520 (cpptools) on cpu 2 hit rate 24.78% (2800/11300)
pid 98381 (cpptools) on cpu 2 hit rate 21.88% (2100/9600)
pid 8526 (code) on cpu 2 hit rate 51.38% (5600/10900)
pid 760435 (bluetoothd) on cpu 1 hit rate 10.54% (6300/59800)
pid 781990 (kworker/u9:2) on cpu 2 hit rate 87.27% (9600/11000)
pid 548101 (cpptools-srv) on cpu 2 hit rate 33.33% (800/2400)
pid 223016 (chrome) on cpu 2 hit rate 0% (0/5100)
pid 7843 (code) on cpu 3 hit rate 48.76% (5900/12100)
pid 97446 (code) on cpu 0 hit rate 0% (0/124300)
pid 779807 (kworker/u9:4) on cpu 3 hit rate 16.24% (4400/27100)
pid 97685 (code) on cpu 0 hit rate 61.76% (4200/6800)
pid 7843 (code) on cpu 1 hit rate 0% (0/5000)
pid 530384 (cpptools) on cpu 3 hit rate 91.3% (2100/2300)
pid 782023 (docker) on cpu 2 hit rate 0% (0/11000)
pid 723596 (kworker/u9:1) on cpu 2 hit rate 71.65% (73800/103000)
pid 3912 (gmain) on cpu 1 hit rate 0% (0/600)
pid 0 (swapper/2) on cpu 2 hit rate 79.31% (1658100/2090600)
pid 21684 (cpptools-srv) on cpu 3 hit rate 28.57% (1600/5600)
pid 781466 (DartWorker) on cpu 2 hit rate 0% (0/6500)
pid 3463 (snap-store) on cpu 2 hit rate 0% (0/3700)
pid 1584 (containerd) on cpu 3 hit rate 49.72% (8900/17900)
pid 7886 (code) on cpu 0 hit rate 46.15% (600/1300)
pid 1584 (containerd) on cpu 2 hit rate 60.89% (30200/49600)
pid 1056 (gmain) on cpu 0 hit rate 67.74% (2100/3100)
pid 782029 (docker) on cpu 1 hit rate 0% (0/19600)
pid 552090 (cpptools-srv) on cpu 1 hit rate 60% (3000/5000)
pid 782013 (docker) on cpu 3 hit rate 0% (0/85000)
pid 2359 (nmbd) on cpu 0 hit rate 0% (0/2300)
pid 4310 (HangWatcher) on cpu 2 hit rate 12.5% (400/3200)
pid 782044 (docker) on cpu 1 hit rate 0% (0/5300)
pid 780275 (kworker/u8:3) on cpu 2 hit rate 15.79% (1500/9500)
pid 3226 (gnome-shell) on cpu 3 hit rate 53.48% (514000/961100)
pid 626 (irq/132-iwlwifi) on cpu 3 hit rate 0% (0/10300)
pid 548101 (cpptools-srv) on cpu 3 hit rate 0% (0/1800)
pid 782041 (docker) on cpu 0 hit rate 36.78% (3200/8700)
pid 7885 (code) on cpu 2 hit rate 0% (0/5000)
pid 118 (kworker/1:1H) on cpu 1 hit rate 0% (0/10000)
pid 8503 (code) on cpu 3 hit rate 0% (0/3400)
pid 545505 (cpptools-srv) on cpu 2 hit rate 14.71% (500/3400)
pid 21528 (cpptools) on cpu 0 hit rate 34.76% (5700/16400)
pid 7888 (code) on cpu 3 hit rate 0% (0/5000)
pid 782020 (code) on cpu 1 hit rate 0.7194% (300/41700)
pid 756219 (kworker/u8:2) on cpu 2 hit rate 0% (0/1200)
pid 7914 (Chrome_ChildIOT) on cpu 0 hit rate 0% (0/3700)
pid 782015 (docker) on cpu 2 hit rate 78% (3900/5000)
pid 782029 (docker) on cpu 2 hit rate 0% (0/3000)
pid 4273 (HangWatcher) on cpu 0 hit rate 72.73% (4000/5500)
pid 545775 (cpptools-srv) on cpu 2 hit rate 73.33% (5500/7500)
pid 782033 (docker) on cpu 0 hit rate 0% (0/35000)
pid 782042 (docker) on cpu 3 hit rate 28.23% (5900/20900)
pid 782037 (docker) on cpu 0 hit rate 27.37% (2600/9500)
pid 782015 (docker) on cpu 3 hit rate 74.07% (2000/2700)
pid 97845 (code) on cpu 2 hit rate 65.79% (5000/7600)
pid 782034 (docker) on cpu 0 hit rate 67.35% (6600/9800)
pid 3256 (gnome-shell) on cpu 3 hit rate 18.69% (4000/21400)
pid 782025 (docker) on cpu 0 hit rate 0% (0/1500)
pid 1010 (irqbalance) on cpu 2 hit rate 5.426% (700/12900)
pid 782027 (docker) on cpu 3 hit rate 0.6472% (400/61800)
里面最核心的函数如下：
StatusTuple attach_perf_event(uint32_t ev_type, uint32_t ev_config,
const std::string& probe_func,
uint64_t sample_period, uint64_t sample_freq,
pid_t pid = -1, int cpu = -1,
int group_fd = -1);
StatusTuple detach_perf_event(uint32_t ev_type, uint32_t ev_config);
接下来我们再看一个例子，这个例子是统计启动最多的进程名称，这次会用到 perf buffer, 代码如下：
#include <string>
#include <iostream>
#include <stdlib.h>
#include <unistd.h>
#include <fstream>
#include <signal.h>
#include <bcc/BPF.h>
#include <linux/perf_event.h>
#include <iomanip>
const std::string BPF_PROGRAM = R"(
#include <linux/ptrace.h>
BPF_PERF_OUTPUT(events);
int do_sys_execve(struct pt_regs *ctx, void *filename, void *argv, void *envp) {
char comm[16] = {0};
bpf_get_current_comm(comm, sizeof(comm));
events.perf_submit(ctx, comm, sizeof(comm));
return 0;
};
)";
std::map<std::string, int64_t> exec_map;
bool cancel = false;
void signal_handler(int s) {
std::cout<<"receive cancel signal" << std::endl;
cancel = true;
}
void handle_output(void *cb, void *data, int data_size) {
char *comm = static_cast<char *>(data);
exec_map[comm] += 1;
}
int main(){
ebpf::BPF bpf;
auto init_res = bpf.init(BPF_PROGRAM);
if (!init_res.ok()) {
std::cout<<init_res.msg() << std::endl;
return 1;
}
std::cout<<"bpf init ok"<<std::endl;
std::string fnname = bpf.get_syscall_fnname("execve");
auto res = bpf.attach_kprobe(fnname, "do_sys_execve");
if (!res.ok()) {
std::cout<<res.msg()<<std::endl;
return 1;
}
res = bpf.open_perf_buffer("events", &handle_output);
if (!res.ok()) {
std::cout<<res.msg()<<std::endl;
return 1;
}
std::cout<<"open perf buffer ok"<<std::endl;
signal(SIGINT, signal_handler);
while (!cancel) {
bpf.poll_perf_buffer("events");
}
for (auto [key, value]: exec_map) {
std::cout<<key << " " << value<< std::endl;
}
return 0;
}
输出如下：
bpf init ok
open perf buffer ok
^Creceive cancel signal
bash 3328
code 220
这儿的open_perf_buffer需要介绍下，我们知道在性能分析的时候如果需要从内核往用户态传输大量的数据，这时候的性能就会很差，bpf是通过perf buffer解决的这个问题，当创建perf events buffer后，内核每发生一次对应的事件，都会激活用户态注册的回调，用户态的回调负责统计，这样就不需要传输大量数据了，因为统计数据本来就在用户态，不需要从内核态发送给用户态了。
poll_perf_buffer 就类似一个io多路复用，阻塞到该buffer上，等待时间的发生。每发生一次事件，都会被唤醒一次。这儿的case中，我们是注册了中断信号，运行后，等待段时间，按下ctrl c，就可以知道哪个应用被创建的次数最多了。
网友评论