1#!/usr/bin/env python 2# 3# oomkill Trace oom_kill_process(). For Linux, uses BCC, eBPF. 4# 5# This traces the kernel out-of-memory killer, and prints basic details, 6# including the system load averages. This can provide more context on the 7# system state at the time of OOM: was it getting busier or steady, based 8# on the load averages? This tool may also be useful to customize for 9# investigations; for example, by adding other task_struct details at the time 10# of OOM. 11# 12# Copyright 2016 Netflix, Inc. 13# Licensed under the Apache License, Version 2.0 (the "License") 14# 15# 09-Feb-2016 Brendan Gregg Created this. 16 17from bcc import BPF 18from time import strftime 19 20# linux stats 21loadavg = "/proc/loadavg" 22 23# define BPF program 24bpf_text = """ 25#include <uapi/linux/ptrace.h> 26#include <linux/oom.h> 27 28struct data_t { 29 u32 fpid; 30 u32 tpid; 31 u64 pages; 32 char fcomm[TASK_COMM_LEN]; 33 char tcomm[TASK_COMM_LEN]; 34}; 35 36BPF_PERF_OUTPUT(events); 37 38void kprobe__oom_kill_process(struct pt_regs *ctx, struct oom_control *oc, const char *message) 39{ 40 struct task_struct *p = oc->chosen; 41 struct data_t data = {}; 42 u32 pid = bpf_get_current_pid_tgid() >> 32; 43 44 data.fpid = pid; 45 data.tpid = p->tgid; 46 data.pages = oc->totalpages; 47 bpf_get_current_comm(&data.fcomm, sizeof(data.fcomm)); 48 bpf_probe_read_kernel(&data.tcomm, sizeof(data.tcomm), p->comm); 49 events.perf_submit(ctx, &data, sizeof(data)); 50} 51""" 52 53# process event 54def print_event(cpu, data, size): 55 event = b["events"].event(data) 56 with open(loadavg) as stats: 57 avgline = stats.read().rstrip() 58 print(("%s Triggered by PID %d (\"%s\"), OOM kill of PID %d (\"%s\")" 59 ", %d pages, loadavg: %s") % (strftime("%H:%M:%S"), event.fpid, 60 event.fcomm.decode('utf-8', 'replace'), event.tpid, 61 event.tcomm.decode('utf-8', 'replace'), event.pages, avgline)) 62 63# initialize BPF 64b = BPF(text=bpf_text) 65print("Tracing OOM kills... Ctrl-C to stop.") 66b["events"].open_perf_buffer(print_event) 67while 1: 68 try: 69 b.perf_buffer_poll() 70 except KeyboardInterrupt: 71 exit() 72