blob: d1d3d26ac6123d78e97c8c8be7f33f795470ca4f [file] [log] [blame]
# @lint-avoid-python-3-compatibility-imports
# profile Profile CPU usage by sampling stack traces at a timed interval.
# For Linux, uses BCC, BPF, perf_events. Embedded C.
# This is an efficient profiler, as stack traces are frequency counted in
# kernel context, rather than passing every stack to user space for frequency
# counting there. Only the unique stacks and counts are passed to user space
# at the end of the profile, greatly reducing the kernel<->user transfer.
# REQUIRES: Linux 4.9+ (BPF_PROG_TYPE_PERF_EVENT support). Under tools/old is
# a version of this tool that may work on Linux 4.6 - 4.8.
# Copyright 2016 Netflix, Inc.
# Licensed under the Apache License, Version 2.0 (the "License")
# THANKS: Alexei Starovoitov, who added proper BPF profiling support to Linux;
# Sasha Goldshtein, Andrew Birchall, and Evgeny Vereshchagin, who wrote much
# of the code here, borrowed from and; and
# Teng Qin, who added perf support in bcc.
# 15-Jul-2016 Brendan Gregg Created this.
# 20-Oct-2016 " " Switched to use the new 4.9 support.
from __future__ import print_function
from bcc import BPF, PerfType, PerfSWConfig
from sys import stderr
from time import sleep
import argparse
import signal
import os
import errno
import multiprocessing
import ctypes as ct
# Process Arguments
# arg validation
def positive_int(val):
ival = int(val)
except ValueError:
raise argparse.ArgumentTypeError("must be an integer")
if ival < 0:
raise argparse.ArgumentTypeError("must be positive")
return ival
def positive_nonzero_int(val):
ival = positive_int(val)
if ival == 0:
raise argparse.ArgumentTypeError("must be nonzero")
return ival
def stack_id_err(stack_id):
# -EFAULT in get_stackid normally means the stack-trace is not availible,
# Such as getting kernel stack trace in userspace code
return (stack_id < 0) and (stack_id != -errno.EFAULT)
# arguments
examples = """examples:
./profile # profile stack traces at 49 Hertz until Ctrl-C
./profile -F 99 # profile stack traces at 99 Hertz
./profile -c 1000000 # profile stack traces every 1 in a million events
./profile 5 # profile at 49 Hertz for 5 seconds only
./profile -f 5 # output in folded format for flame graphs
./profile -p 185 # only profile threads for PID 185
./profile -U # only show user space stacks (no kernel)
./profile -K # only show kernel space stacks (no user)
parser = argparse.ArgumentParser(
description="Profile CPU stack traces at a timed interval",
thread_group = parser.add_mutually_exclusive_group()
thread_group.add_argument("-p", "--pid", type=positive_int,
help="profile this PID only")
# TODO: add options for user/kernel threads only
stack_group = parser.add_mutually_exclusive_group()
stack_group.add_argument("-U", "--user-stacks-only", action="store_true",
help="show stacks from user space only (no kernel space stacks)")
stack_group.add_argument("-K", "--kernel-stacks-only", action="store_true",
help="show stacks from kernel space only (no user space stacks)")
sample_group = parser.add_mutually_exclusive_group()
sample_group.add_argument("-F", "--frequency", type=positive_int,
help="sample frequency, Hertz")
sample_group.add_argument("-c", "--count", type=positive_int,
help="sample period, number of events")
parser.add_argument("-d", "--delimited", action="store_true",
help="insert delimiter between kernel/user stacks")
parser.add_argument("-a", "--annotations", action="store_true",
help="add _[k] annotations to kernel frames")
parser.add_argument("-f", "--folded", action="store_true",
help="output folded format, one line per stack (for flame graphs)")
parser.add_argument("--stack-storage-size", default=16384,
help="the number of unique stack traces that can be stored and "
"displayed (default %(default)s)")
parser.add_argument("duration", nargs="?", default=99999999,
help="duration of trace, in seconds")
parser.add_argument("-C", "--cpu", type=int, default=-1,
help="cpu number to run profile on")
parser.add_argument("--ebpf", action="store_true",
# option logic
args = parser.parse_args()
pid = int( if is not None else -1
duration = int(args.duration)
debug = 0
need_delimiter = args.delimited and not (args.kernel_stacks_only or
# TODO: add stack depth, and interval
# Setup BPF
# define BPF program
bpf_text = """
#include <uapi/linux/ptrace.h>
#include <uapi/linux/bpf_perf_event.h>
#include <linux/sched.h>
struct key_t {
u32 pid;
u64 kernel_ip;
u64 kernel_ret_ip;
int user_stack_id;
int kernel_stack_id;
char name[TASK_COMM_LEN];
BPF_HASH(counts, struct key_t);
// This code gets a bit complex. Probably not suitable for casual hacking.
int do_perf_event(struct bpf_perf_event_data *ctx) {
u32 pid = bpf_get_current_pid_tgid() >> 32;
return 0;
// create map key
struct key_t key = {.pid = pid};
bpf_get_current_comm(&, sizeof(;
// get stacks
key.user_stack_id = USER_STACK_GET;
key.kernel_stack_id = KERNEL_STACK_GET;
if (key.kernel_stack_id >= 0) {
// populate extras to fix the kernel stack
u64 ip = PT_REGS_IP(&ctx->regs);
u64 page_offset;
// if ip isn't sane, leave key ips as zero for later checking
#if defined(CONFIG_X86_64) && defined(__PAGE_OFFSET_BASE)
// x64, 4.16, ..., 4.11, etc., but some earlier kernel didn't have it
page_offset = __PAGE_OFFSET_BASE;
#elif defined(CONFIG_X86_64) && defined(__PAGE_OFFSET_BASE_L4)
// x64, 4.17, and later
page_offset = __PAGE_OFFSET_BASE_L5;
page_offset = __PAGE_OFFSET_BASE_L4;
// earlier x86_64 kernels, e.g., 4.6, comes here
// arm64, s390, powerpc, x86_32
page_offset = PAGE_OFFSET;
if (ip > page_offset) {
key.kernel_ip = ip;
return 0;
# set thread filter
thread_context = ""
perf_filter = "-a"
if is not None:
thread_context = "PID %s" %
thread_filter = 'pid == %s' %
perf_filter = '-p %s' %
thread_context = "all threads"
thread_filter = '1'
bpf_text = bpf_text.replace('THREAD_FILTER', thread_filter)
# set stack storage size
bpf_text = bpf_text.replace('STACK_STORAGE_SIZE', str(args.stack_storage_size))
# handle stack args
kernel_stack_get = "stack_traces.get_stackid(&ctx->regs, 0)"
user_stack_get = "stack_traces.get_stackid(&ctx->regs, BPF_F_USER_STACK)"
stack_context = ""
if args.user_stacks_only:
stack_context = "user"
kernel_stack_get = "-1"
elif args.kernel_stacks_only:
stack_context = "kernel"
user_stack_get = "-1"
stack_context = "user + kernel"
bpf_text = bpf_text.replace('USER_STACK_GET', user_stack_get)
bpf_text = bpf_text.replace('KERNEL_STACK_GET', kernel_stack_get)
sample_freq = 0
sample_period = 0
if args.frequency:
sample_freq = args.frequency
elif args.count:
sample_period = args.count
# If user didn't specify anything, use default 49Hz sampling
sample_freq = 49
sample_context = "%s%d %s" % (("", sample_freq, "Hertz") if sample_freq
else ("every ", sample_period, "events"))
# header
if not args.folded:
print("Sampling at %s of %s by %s stack" %
(sample_context, thread_context, stack_context), end="")
if args.cpu >= 0:
print(" on CPU#{}".format(args.cpu), end="")
if duration < 99999999:
print(" for %d secs." % duration)
print("... Hit Ctrl-C to end.")
if debug or args.ebpf:
if args.ebpf:
# initialize BPF & perf_events
b = BPF(text=bpf_text)
ev_config=PerfSWConfig.CPU_CLOCK, fn_name="do_perf_event",
sample_period=sample_period, sample_freq=sample_freq, cpu=args.cpu)
# signal handler
def signal_ignore(signal, frame):
# Output Report
# collect samples
except KeyboardInterrupt:
# as cleanup can take some time, trap Ctrl-C:
signal.signal(signal.SIGINT, signal_ignore)
if not args.folded:
def aksym(addr):
if args.annotations:
return b.ksym(addr) + "_[k]".encode()
return b.ksym(addr)
# output stacks
missing_stacks = 0
has_enomem = False
counts = b.get_table("counts")
stack_traces = b.get_table("stack_traces")
need_delimiter = args.delimited and not (args.kernel_stacks_only or
for k, v in sorted(counts.items(), key=lambda counts: counts[1].value):
# handle get_stackid errors
if not args.user_stacks_only and stack_id_err(k.kernel_stack_id):
missing_stacks += 1
has_enomem = has_enomem or k.kernel_stack_id == -errno.ENOMEM
if not args.kernel_stacks_only and stack_id_err(k.user_stack_id):
missing_stacks += 1
has_enomem = has_enomem or k.user_stack_id == -errno.ENOMEM
user_stack = [] if k.user_stack_id < 0 else \
kernel_tmp = [] if k.kernel_stack_id < 0 else \
# fix kernel stack
kernel_stack = []
if k.kernel_stack_id >= 0:
for addr in kernel_tmp:
# the later IP checking
if k.kernel_ip:
kernel_stack.insert(0, k.kernel_ip)
if args.folded:
# print folded stack output
user_stack = list(user_stack)
kernel_stack = list(kernel_stack)
line = []
# if we failed to get the stack is, such as due to no space (-ENOMEM) or
# hash collision (-EEXIST), we still print a placeholder for consistency
if not args.kernel_stacks_only:
if stack_id_err(k.user_stack_id):
line.append("[Missed User Stack]")
line.extend([b.sym(addr, for addr in reversed(user_stack)])
if not args.user_stacks_only:
line.extend(["-"] if (need_delimiter and k.kernel_stack_id >= 0 and k.user_stack_id >= 0) else [])
if stack_id_err(k.kernel_stack_id):
line.append("[Missed Kernel Stack]")
line.extend([b.ksym(addr) for addr in reversed(kernel_stack)])
print("%s %d" % (b";".join(line).decode('utf-8', 'replace'), v.value))
# print default multi-line stack output
if not args.user_stacks_only:
if stack_id_err(k.kernel_stack_id):
print(" [Missed Kernel Stack]")
for addr in kernel_stack:
print(" %s" % aksym(addr))
if not args.kernel_stacks_only:
if need_delimiter and k.user_stack_id >= 0 and k.kernel_stack_id >= 0:
print(" --")
if stack_id_err(k.user_stack_id):
print(" [Missed User Stack]")
for addr in user_stack:
print(" %s" % b.sym(addr,'utf-8', 'replace'))
print(" %-16s %s (%d)" % ("-",'utf-8', 'replace'),
print(" %d\n" % v.value)
# check missing
if missing_stacks > 0:
enomem_str = "" if not has_enomem else \
" Consider increasing --stack-storage-size."
print("WARNING: %d stack traces could not be displayed.%s" %
(missing_stacks, enomem_str),