Edit: /usr/sbin/runqlat-bpfcc (9526B)
#! /usr/bin/python3
# @lint-avoid-python-3-compatibility-imports
#
# runqlat Run queue (scheduler) latency as a histogram.
# For Linux, uses BCC, eBPF.
#
# USAGE: runqlat [-h] [-T] [-m] [-P] [-L] [-p PID] [interval] [count]
#
# This measures the time a task spends waiting on a run queue for a turn
# on-CPU, and shows this time as a histogram. This time should be small, but a
# task may need to wait its turn due to CPU load.
#
# This measures two types of run queue latency:
# 1. The time from a task being enqueued on a run queue to its context switch
# and execution. This traces ttwu_do_wakeup(), wake_up_new_task() ->
# finish_task_switch() with either raw tracepoints (if supported) or kprobes
# and instruments the run queue latency after a voluntary context switch.
# 2. The time from when a task was involuntary context switched and still
# in the runnable state, to when it next executed. This is instrumented
# from finish_task_switch() alone.
#
# Copyright 2016 Netflix, Inc.
# Licensed under the Apache License, Version 2.0 (the "License")
#
# 07-Feb-2016 Brendan Gregg Created this.
from __future__ import print_function
from bcc import BPF
from time import sleep, strftime
import argparse
# arguments
examples = """examples:
./runqlat # summarize run queue latency as a histogram
./runqlat 1 10 # print 1 second summaries, 10 times
./runqlat -mT 1 # 1s summaries, milliseconds, and timestamps
./runqlat -P # show each PID separately
./runqlat -p 185 # trace PID 185 only
"""
parser = argparse.ArgumentParser(
description="Summarize run queue (scheduler) latency as a histogram",
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog=examples)
parser.add_argument("-T", "--timestamp", action="store_true",
help="include timestamp on output")
parser.add_argument("-m", "--milliseconds", action="store_true",
help="millisecond histogram")
parser.add_argument("-P", "--pids", action="store_true",
help="print a histogram per process ID")
# PID options are --pid and --pids, so namespaces should be --pidns (not done
# yet) and --pidnss:
parser.add_argument("--pidnss", action="store_true",
help="print a histogram per PID namespace")
parser.add_argument("-L", "--tids", action="store_true",
help="print a histogram per thread ID")
parser.add_argument("-p", "--pid",
help="trace this PID only")
parser.add_argument("interval", nargs="?", default=99999999,
help="output interval, in seconds")
parser.add_argument("count", nargs="?", default=99999999,
help="number of outputs")
parser.add_argument("--ebpf", action="store_true",
help=argparse.SUPPRESS)
args = parser.parse_args()
countdown = int(args.count)
debug = 0
# define BPF program
bpf_text = """
#include
#include
#include
#include
#include
typedef struct pid_key {
u32 id;
u64 slot;
} pid_key_t;
typedef struct pidns_key {
u32 id;
u64 slot;
} pidns_key_t;
BPF_HASH(start, u32);
STORAGE
// record enqueue timestamp
static int trace_enqueue(u32 tgid, u32 pid)
{
if (FILTER || pid == 0)
return 0;
u64 ts = bpf_ktime_get_ns();
start.update(&pid, &ts);
return 0;
}
static __always_inline unsigned int pid_namespace(struct task_struct *task)
{
/* pids[] was removed from task_struct since commit 2c4704756cab7cfa031ada4dab361562f0e357c0
* Using the macro INIT_PID_LINK as a conditional judgment.
*/
#ifdef INIT_PID_LINK
struct pid_link pids;
unsigned int level;
struct upid upid;
struct ns_common ns;
/* get the pid namespace by following task_active_pid_ns(),
* pid->numbers[pid->level].ns
*/
bpf_probe_read_kernel(&pids, sizeof(pids), &task->pids[PIDTYPE_PID]);
bpf_probe_read_kernel(&level, sizeof(level), &pids.pid->level);
bpf_probe_read_kernel(&upid, sizeof(upid), &pids.pid->numbers[level]);
bpf_probe_read_kernel(&ns, sizeof(ns), &upid.ns->ns);
return ns.inum;
#else
struct pid *pid;
unsigned int level;
struct upid upid;
struct ns_common ns;
/* get the pid namespace by following task_active_pid_ns(),
* pid->numbers[pid->level].ns
*/
bpf_probe_read_kernel(&pid, sizeof(pid), &task->thread_pid);
bpf_probe_read_kernel(&level, sizeof(level), &pid->level);
bpf_probe_read_kernel(&upid, sizeof(upid), &pid->numbers[level]);
bpf_probe_read_kernel(&ns, sizeof(ns), &upid.ns->ns);
return ns.inum;
#endif
}
"""
bpf_text_kprobe = """
int trace_wake_up_new_task(struct pt_regs *ctx, struct task_struct *p)
{
return trace_enqueue(p->tgid, p->pid);
}
int trace_ttwu_do_wakeup(struct pt_regs *ctx, struct rq *rq, struct task_struct *p,
int wake_flags)
{
return trace_enqueue(p->tgid, p->pid);
}
// calculate latency
int trace_run(struct pt_regs *ctx, struct task_struct *prev)
{
u32 pid, tgid;
// ivcsw: treat like an enqueue event and store timestamp
if (prev->STATE_FIELD == TASK_RUNNING) {
tgid = prev->tgid;
pid = prev->pid;
if (!(FILTER || pid == 0)) {
u64 ts = bpf_ktime_get_ns();
start.update(&pid, &ts);
}
}
tgid = bpf_get_current_pid_tgid() >> 32;
pid = bpf_get_current_pid_tgid();
if (FILTER || pid == 0)
return 0;
u64 *tsp, delta;
// fetch timestamp and calculate delta
tsp = start.lookup(&pid);
if (tsp == 0) {
return 0; // missed enqueue
}
delta = bpf_ktime_get_ns() - *tsp;
FACTOR
// store as histogram
STORE
start.delete(&pid);
return 0;
}
"""
bpf_text_raw_tp = """
RAW_TRACEPOINT_PROBE(sched_wakeup)
{
// TP_PROTO(struct task_struct *p)
struct task_struct *p = (struct task_struct *)ctx->args[0];
return trace_enqueue(p->tgid, p->pid);
}
RAW_TRACEPOINT_PROBE(sched_wakeup_new)
{
// TP_PROTO(struct task_struct *p)
struct task_struct *p = (struct task_struct *)ctx->args[0];
return trace_enqueue(p->tgid, p->pid);
}
RAW_TRACEPOINT_PROBE(sched_switch)
{
// TP_PROTO(bool preempt, struct task_struct *prev, struct task_struct *next)
struct task_struct *prev = (struct task_struct *)ctx->args[1];
struct task_struct *next = (struct task_struct *)ctx->args[2];
u32 pid, tgid;
// ivcsw: treat like an enqueue event and store timestamp
if (prev->STATE_FIELD == TASK_RUNNING) {
tgid = prev->tgid;
pid = prev->pid;
if (!(FILTER || pid == 0)) {
u64 ts = bpf_ktime_get_ns();
start.update(&pid, &ts);
}
}
tgid = next->tgid;
pid = next->pid;
if (FILTER || pid == 0)
return 0;
u64 *tsp, delta;
// fetch timestamp and calculate delta
tsp = start.lookup(&pid);
if (tsp == 0) {
return 0; // missed enqueue
}
delta = bpf_ktime_get_ns() - *tsp;
FACTOR
// store as histogram
STORE
start.delete(&pid);
return 0;
}
"""
is_support_raw_tp = BPF.support_raw_tracepoint()
if is_support_raw_tp:
bpf_text += bpf_text_raw_tp
else:
bpf_text += bpf_text_kprobe
# code substitutions
if BPF.kernel_struct_has_field(b'task_struct', b'__state') == 1:
bpf_text = bpf_text.replace('STATE_FIELD', '__state')
else:
bpf_text = bpf_text.replace('STATE_FIELD', 'state')
if args.pid:
# pid from userspace point of view is thread group from kernel pov
bpf_text = bpf_text.replace('FILTER', 'tgid != %s' % args.pid)
else:
bpf_text = bpf_text.replace('FILTER', '0')
if args.milliseconds:
bpf_text = bpf_text.replace('FACTOR', 'delta /= 1000000;')
label = "msecs"
else:
bpf_text = bpf_text.replace('FACTOR', 'delta /= 1000;')
label = "usecs"
if args.pids or args.tids:
section = "pid"
pid = "tgid"
if args.tids:
pid = "pid"
section = "tid"
bpf_text = bpf_text.replace('STORAGE',
'BPF_HISTOGRAM(dist, pid_key_t);')
bpf_text = bpf_text.replace('STORE',
'pid_key_t key = {}; key.id = ' + pid + '; key.slot = bpf_log2l(delta); ' +
'dist.increment(key);')
elif args.pidnss:
section = "pidns"
bpf_text = bpf_text.replace('STORAGE',
'BPF_HISTOGRAM(dist, pidns_key_t);')
bpf_text = bpf_text.replace('STORE', 'pidns_key_t key = ' +
'{.id = pid_namespace(prev), ' +
'.slot = bpf_log2l(delta)}; dist.atomic_increment(key);')
else:
section = ""
bpf_text = bpf_text.replace('STORAGE', 'BPF_HISTOGRAM(dist);')
bpf_text = bpf_text.replace('STORE',
'dist.atomic_increment(bpf_log2l(delta));')
if debug or args.ebpf:
print(bpf_text)
if args.ebpf:
exit()
# load BPF program
b = BPF(text=bpf_text)
if not is_support_raw_tp:
b.attach_kprobe(event="ttwu_do_wakeup", fn_name="trace_ttwu_do_wakeup")
b.attach_kprobe(event="wake_up_new_task", fn_name="trace_wake_up_new_task")
b.attach_kprobe(event_re=r'^finish_task_switch$|^finish_task_switch\.isra\.\d$',
fn_name="trace_run")
print("Tracing run queue latency... Hit Ctrl-C to end.")
# output
exiting = 0 if args.interval else 1
dist = b.get_table("dist")
while (1):
try:
sleep(int(args.interval))
except KeyboardInterrupt:
exiting = 1
print()
if args.timestamp:
print("%-8s\n" % strftime("%H:%M:%S"), end="")
dist.print_log2_hist(label, section, section_print_fn=int)
dist.clear()
countdown -= 1
if exiting or countdown == 0:
exit()