1#!/usr/bin/env python
2# @lint-avoid-python-3-compatibility-imports
3#
4# runqlen    Summarize scheduler run queue length as a histogram.
5#            For Linux, uses BCC, eBPF.
6#
7# This counts the length of the run queue, excluding the currently running
8# thread, and shows it as a histogram.
9#
10# Also answers run queue occupancy.
11#
12# USAGE: runqlen [-h] [-T] [-Q] [-m] [-D] [interval] [count]
13#
14# REQUIRES: Linux 4.9+ (BPF_PROG_TYPE_PERF_EVENT support). Under tools/old is
15# a version of this tool that may work on Linux 4.6 - 4.8.
16#
17# Copyright 2016 Netflix, Inc.
18# Licensed under the Apache License, Version 2.0 (the "License")
19#
20# 12-Dec-2016   Brendan Gregg   Created this.
21
22from __future__ import print_function
23from bcc import BPF, PerfType, PerfSWConfig
24from time import sleep, strftime
25from tempfile import NamedTemporaryFile
26from os import open, close, dup, unlink, O_WRONLY
27import argparse
28
29# arguments
30examples = """examples:
31    ./runqlen            # summarize run queue length as a histogram
32    ./runqlen 1 10       # print 1 second summaries, 10 times
33    ./runqlen -T 1       # 1s summaries and timestamps
34    ./runqlen -O         # report run queue occupancy
35    ./runqlen -C         # show each CPU separately
36"""
37parser = argparse.ArgumentParser(
38    description="Summarize scheduler run queue length as a histogram",
39    formatter_class=argparse.RawDescriptionHelpFormatter,
40    epilog=examples)
41parser.add_argument("-T", "--timestamp", action="store_true",
42    help="include timestamp on output")
43parser.add_argument("-O", "--runqocc", action="store_true",
44    help="report run queue occupancy")
45parser.add_argument("-C", "--cpus", action="store_true",
46    help="print output for each CPU separately")
47parser.add_argument("interval", nargs="?", default=99999999,
48    help="output interval, in seconds")
49parser.add_argument("count", nargs="?", default=99999999,
50    help="number of outputs")
51parser.add_argument("--ebpf", action="store_true",
52    help=argparse.SUPPRESS)
53args = parser.parse_args()
54countdown = int(args.count)
55debug = 0
56frequency = 99
57
58# Linux 4.15 introduced a new field runnable_weight
59# in linux_src:kernel/sched/sched.h as
60#     struct cfs_rq {
61#         struct load_weight load;
62#         unsigned long runnable_weight;
63#         unsigned int nr_running, h_nr_running;
64#         ......
65#     }
66# and this tool requires to access nr_running to get
67# runqueue len information.
68#
69# The commit which introduces cfs_rq->runnable_weight
70# field also introduces the field sched_entity->runnable_weight
71# where sched_entity is defined in linux_src:include/linux/sched.h.
72#
73# To cope with pre-4.15 and 4.15/post-4.15 releases,
74# we run a simple BPF program to detect whether
75# field sched_entity->runnable_weight exists. The existence of
76# this field should infer the existence of cfs_rq->runnable_weight.
77#
78# This will need maintenance as the relationship between these
79# two fields may change in the future.
80#
81def check_runnable_weight_field():
82    # Define the bpf program for checking purpose
83    bpf_check_text = """
84#include <linux/sched.h>
85unsigned long dummy(struct sched_entity *entity)
86{
87    return entity->runnable_weight;
88}
89"""
90
91    # Get a temporary file name
92    tmp_file = NamedTemporaryFile(delete=False)
93    tmp_file.close();
94
95    # Duplicate and close stderr (fd = 2)
96    old_stderr = dup(2)
97    close(2)
98
99    # Open a new file, should get fd number 2
100    # This will avoid printing llvm errors on the screen
101    fd = open(tmp_file.name, O_WRONLY)
102    try:
103        t = BPF(text=bpf_check_text)
104        success_compile = True
105    except:
106        success_compile = False
107
108    # Release the fd 2, and next dup should restore old stderr
109    close(fd)
110    dup(old_stderr)
111    close(old_stderr)
112
113    # remove the temporary file and return
114    unlink(tmp_file.name)
115    return success_compile
116
117
118# define BPF program
119bpf_text = """
120#include <uapi/linux/ptrace.h>
121#include <linux/sched.h>
122
123// Declare enough of cfs_rq to find nr_running, since we can't #import the
124// header. This will need maintenance. It is from kernel/sched/sched.h:
125struct cfs_rq_partial {
126    struct load_weight load;
127    RUNNABLE_WEIGHT_FIELD
128    unsigned int nr_running, h_nr_running;
129};
130
131typedef struct cpu_key {
132    int cpu;
133    unsigned int slot;
134} cpu_key_t;
135STORAGE
136
137int do_perf_event()
138{
139    unsigned int len = 0;
140    pid_t pid = 0;
141    struct task_struct *task = NULL;
142    struct cfs_rq_partial *my_q = NULL;
143
144    // Fetch the run queue length from task->se.cfs_rq->nr_running. This is an
145    // unstable interface and may need maintenance. Perhaps a future version
146    // of BPF will support task_rq(p) or something similar as a more reliable
147    // interface.
148    task = (struct task_struct *)bpf_get_current_task();
149    my_q = (struct cfs_rq_partial *)task->se.cfs_rq;
150    len = my_q->nr_running;
151
152    // Calculate run queue length by subtracting the currently running task,
153    // if present. len 0 == idle, len 1 == one running task.
154    if (len > 0)
155        len--;
156
157    STORE
158
159    return 0;
160}
161"""
162
163# code substitutions
164if args.cpus:
165    bpf_text = bpf_text.replace('STORAGE',
166        'BPF_HISTOGRAM(dist, cpu_key_t);')
167    bpf_text = bpf_text.replace('STORE', 'cpu_key_t key = {.slot = len}; ' +
168        'key.cpu = bpf_get_smp_processor_id(); ' +
169        'dist.increment(key);')
170else:
171    bpf_text = bpf_text.replace('STORAGE',
172        'BPF_HISTOGRAM(dist, unsigned int);')
173    bpf_text = bpf_text.replace('STORE', 'dist.increment(len);')
174
175if check_runnable_weight_field():
176    bpf_text = bpf_text.replace('RUNNABLE_WEIGHT_FIELD', 'unsigned long runnable_weight;')
177else:
178    bpf_text = bpf_text.replace('RUNNABLE_WEIGHT_FIELD', '')
179
180if debug or args.ebpf:
181    print(bpf_text)
182    if args.ebpf:
183        exit()
184
185# initialize BPF & perf_events
186b = BPF(text=bpf_text)
187b.attach_perf_event(ev_type=PerfType.SOFTWARE,
188    ev_config=PerfSWConfig.CPU_CLOCK, fn_name="do_perf_event",
189    sample_period=0, sample_freq=frequency)
190
191print("Sampling run queue length... Hit Ctrl-C to end.")
192
193# output
194exiting = 0 if args.interval else 1
195dist = b.get_table("dist")
196while (1):
197    try:
198        sleep(int(args.interval))
199    except KeyboardInterrupt:
200        exiting = 1
201
202    print()
203    if args.timestamp:
204        print("%-8s\n" % strftime("%H:%M:%S"), end="")
205
206    if args.runqocc:
207        if args.cpus:
208            # run queue occupancy, per-CPU summary
209            idle = {}
210            queued = {}
211            cpumax = 0
212            for k, v in dist.items():
213                if k.cpu > cpumax:
214                    cpumax = k.cpu
215            for c in range(0, cpumax + 1):
216                idle[c] = 0
217                queued[c] = 0
218            for k, v in dist.items():
219                if k.slot == 0:
220                    idle[k.cpu] += v.value
221                else:
222                    queued[k.cpu] += v.value
223            for c in range(0, cpumax + 1):
224                samples = idle[c] + queued[c]
225                if samples:
226                    runqocc = float(queued[c]) / samples
227                else:
228                    runqocc = 0
229                print("runqocc, CPU %-3d %6.2f%%" % (c, 100 * runqocc))
230
231        else:
232            # run queue occupancy, system-wide summary
233            idle = 0
234            queued = 0
235            for k, v in dist.items():
236                if k.value == 0:
237                    idle += v.value
238                else:
239                    queued += v.value
240            samples = idle + queued
241            if samples:
242                runqocc = float(queued) / samples
243            else:
244                runqocc = 0
245            print("runqocc: %0.2f%%" % (100 * runqocc))
246
247    else:
248        # run queue length histograms
249        dist.print_linear_hist("runqlen", "cpu")
250
251    dist.clear()
252
253    countdown -= 1
254    if exiting or countdown == 0:
255        exit()
256