1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 2023 Red Hat Inc, Daniel Bristot de Oliveira <bristot@kernel.org>
4 */
5
6 #define _GNU_SOURCE
7 #include <sched.h>
8 #include <fcntl.h>
9 #include <stdlib.h>
10 #include <unistd.h>
11 #include <stdio.h>
12 #include <errno.h>
13 #include <string.h>
14 #include <tracefs.h>
15 #include <pthread.h>
16 #include <sys/wait.h>
17 #include <sys/prctl.h>
18
19 #include "utils.h"
20 #include "timerlat_u.h"
21
22 /*
23 * This is the user-space main for the tool timerlatu/ threads.
24 *
25 * It is as simple as this:
26 * - set affinity
27 * - set priority
28 * - open tracer fd
29 * - spin
30 * - close
31 */
timerlat_u_main(int cpu,struct timerlat_u_params * params)32 static int timerlat_u_main(int cpu, struct timerlat_u_params *params)
33 {
34 struct sched_param sp = { .sched_priority = 95 };
35 char buffer[1024];
36 int timerlat_fd;
37 cpu_set_t set;
38 int retval;
39
40 /*
41 * This all is only setting up the tool.
42 */
43 CPU_ZERO(&set);
44 CPU_SET(cpu, &set);
45
46 retval = sched_setaffinity(gettid(), sizeof(set), &set);
47 if (retval == -1) {
48 debug_msg("Error setting user thread affinity %d, is the CPU online?\n", cpu);
49 exit(1);
50 }
51
52 if (!params->sched_param) {
53 retval = sched_setscheduler(0, SCHED_FIFO, &sp);
54 if (retval < 0) {
55 err_msg("Error setting timerlat u default priority: %s\n", strerror(errno));
56 exit(1);
57 }
58 } else {
59 retval = __set_sched_attr(getpid(), params->sched_param);
60 if (retval) {
61 /* __set_sched_attr prints an error message, so */
62 exit(0);
63 }
64 }
65
66 if (params->cgroup_name) {
67 retval = set_pid_cgroup(gettid(), params->cgroup_name);
68 if (!retval) {
69 err_msg("Error setting timerlat u cgroup pid\n");
70 pthread_exit(&retval);
71 }
72 }
73
74 /*
75 * This is the tool's loop. If you want to use as base for your own tool...
76 * go ahead.
77 */
78 snprintf(buffer, sizeof(buffer), "osnoise/per_cpu/cpu%d/timerlat_fd", cpu);
79
80 timerlat_fd = tracefs_instance_file_open(NULL, buffer, O_RDONLY);
81 if (timerlat_fd < 0) {
82 err_msg("Error opening %s:%s\n", buffer, strerror(errno));
83 exit(1);
84 }
85
86 debug_msg("User-space timerlat pid %d on cpu %d\n", gettid(), cpu);
87
88 /* add should continue with a signal handler */
89 while (true) {
90 retval = read(timerlat_fd, buffer, 1024);
91 if (retval < 0)
92 break;
93 }
94
95 close(timerlat_fd);
96
97 debug_msg("Leaving timerlat pid %d on cpu %d\n", gettid(), cpu);
98 exit(0);
99 }
100
101 /*
102 * timerlat_u_send_kill - send a kill signal for all processes
103 *
104 * Return the number of processes that received the kill.
105 */
timerlat_u_send_kill(pid_t * procs,int nr_cpus)106 static int timerlat_u_send_kill(pid_t *procs, int nr_cpus)
107 {
108 int killed = 0;
109 int i, retval;
110
111 for (i = 0; i < nr_cpus; i++) {
112 if (!procs[i])
113 continue;
114 retval = kill(procs[i], SIGKILL);
115 if (!retval)
116 killed++;
117 else
118 err_msg("Error killing child process %d\n", procs[i]);
119 }
120
121 return killed;
122 }
123
124 /**
125 * timerlat_u_dispatcher - dispatch one timerlatu/ process per monitored CPU
126 *
127 * This is a thread main that will fork one new process for each monitored
128 * CPU. It will wait for:
129 *
130 * - rtla to tell to kill the child processes
131 * - some child process to die, and the cleanup all the processes
132 *
133 * whichever comes first.
134 *
135 */
timerlat_u_dispatcher(void * data)136 void *timerlat_u_dispatcher(void *data)
137 {
138 int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
139 struct timerlat_u_params *params = data;
140 char proc_name[128];
141 int procs_count = 0;
142 int retval = 1;
143 pid_t *procs;
144 int wstatus;
145 pid_t pid;
146 int i;
147
148 debug_msg("Dispatching timerlat u procs\n");
149
150 procs = calloc(nr_cpus, sizeof(pid_t));
151 if (!procs)
152 pthread_exit(&retval);
153
154 for (i = 0; i < nr_cpus; i++) {
155 if (params->set && !CPU_ISSET(i, params->set))
156 continue;
157
158 pid = fork();
159
160 /* child */
161 if (!pid) {
162
163 /*
164 * rename the process
165 */
166 snprintf(proc_name, sizeof(proc_name), "timerlatu/%d", i);
167 pthread_setname_np(pthread_self(), proc_name);
168 prctl(PR_SET_NAME, (unsigned long)proc_name, 0, 0, 0);
169
170 timerlat_u_main(i, params);
171 /* timerlat_u_main should exit()! Anyways... */
172 pthread_exit(&retval);
173 }
174
175 /* parent */
176 if (pid == -1) {
177 timerlat_u_send_kill(procs, nr_cpus);
178 debug_msg("Failed to create child processes");
179 pthread_exit(&retval);
180 }
181
182 procs_count++;
183 procs[i] = pid;
184 }
185
186 while (params->should_run) {
187 /* check if processes died */
188 pid = waitpid(-1, &wstatus, WNOHANG);
189 if (pid != 0) {
190 for (i = 0; i < nr_cpus; i++) {
191 if (procs[i] == pid) {
192 procs[i] = 0;
193 procs_count--;
194 }
195 }
196
197 if (!procs_count)
198 break;
199 }
200
201 sleep(1);
202 }
203
204 timerlat_u_send_kill(procs, nr_cpus);
205
206 while (procs_count) {
207 pid = waitpid(-1, &wstatus, 0);
208 if (pid == -1) {
209 err_msg("Failed to monitor child processes");
210 pthread_exit(&retval);
211 }
212 for (i = 0; i < nr_cpus; i++) {
213 if (procs[i] == pid) {
214 procs[i] = 0;
215 procs_count--;
216 }
217 }
218 }
219
220 params->stopped_running = 1;
221
222 free(procs);
223 retval = 0;
224 pthread_exit(&retval);
225
226 }
227