Increasing robustness of executing a function every x seconds - c

I have written some code that tries to call a function (called worker) every x seconds (in this example, I chose 1s as the interval time). The code is a minimal working example that in reality is way more complex than this.
The code works when it is this simple, but I stumble across errors when running the more complex version for a longer period of time. Thus, I want to increase the robustness of this code an would like to get some ideas on how to do that.
Basically, the worker gets some data, processes it an writes it to a file. I open the file during every call to the worker. In the tests, after some time I get an error that the file cannot be opened anymore. In this regard I also noticed that this happens (maybe just by chance) everytime the worker execution time exceeds the interval time. Reason for this is the getter function which pulls data from remote and this can take some time, depending on the network traffic.
I've been thinking of trying a multithreaded approach, but I am not sure if this is worth the hassle. I would be grateful for any pointers on how to do this in a more robust way.
#include <unistd.h>
#include <stdbool.h>
#include <stdio.h>
#include <signal.h>
#include <stdint.h>
#include <time.h>
#define ALARM_INTERVAL_SEC 1
#define ALARM_INTERVAL_USEC 0
static bool running = true;
static struct itimerval alarm_interval;
static struct timeval previous_time;
static uint64_t loop_count = 0;
static FILE* testfile;
static void
signal_handler(int signum)
{
if (signum == SIGINT || signum == SIGTERM)
{
running = false;
}
}
static void
worker(int signum)
{
// Reset the alarm interval
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("Error: setitimer");
raise(SIGTERM);
return;
}
struct timeval current_time;
gettimeofday(&current_time, NULL);
printf("Loop count: %lu\n", loop_count);
printf("Loop time: %f us\n\n", (current_time.tv_sec - previous_time.tv_sec) * 1e6 +
(current_time.tv_usec - previous_time.tv_usec));
previous_time = current_time;
// convert time to human-readable format
char tmbuf[64];
char buf[64];
time_t nowtime = current_time.tv_sec;
struct tm *nowtm = localtime(&nowtime);
strftime(tmbuf, sizeof(tmbuf), "%Y-%m-%d %H:%M:%S", nowtm);
snprintf(buf, sizeof(buf), "%s.%06ld", tmbuf, current_time.tv_usec);
sleep(0.5);
// DO STH
testfile = fopen("testfile.txt", "ab+");
if(testfile == NULL)
{
printf("Error: open testfile");
raise(SIGTERM);
return;
}
fprintf(testfile, "[%s] Loop count: %lu\n", buf, loop_count);
fclose(testfile);
loop_count++;
}
int
main(int argc, char* argv[])
{
signal(SIGINT, signal_handler);
signal(SIGTERM, signal_handler);
signal(SIGALRM, worker);
// Set the alarm interval
alarm_interval.it_interval.tv_sec = 0;
alarm_interval.it_interval.tv_usec = 0;
alarm_interval.it_value.tv_sec = ALARM_INTERVAL_SEC;
alarm_interval.it_value.tv_usec = ALARM_INTERVAL_USEC;
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("Error: setitimer");
return -1;
}
gettimeofday(&previous_time, NULL);
while(running)
{
sleep(1);
}
alarm_interval.it_value.tv_sec = 0;
alarm_interval.it_value.tv_usec = 0;
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("Error: resetting itimer failed");
return -1;
}
return 0;
}

Related

What is the frequency of epoll_wait "wakeups"?

This question is elaboration and continuation of my previous question - On epoll_pwait, POSIX timers and X11 events. Most of X11 events is either delayed or dropped. I work with XCB window that displays some graphical information. The application has an insteractive input and had more or less static display. Now the requirenments had changed and I need to add some pereodic computations and have a interactive input.
The problem with the new requirenments is that computations goes at high rate. With that, I get inconsistent interaction with the XCB window. Input may lag behind, or change the rate of processing.
The setup is multiplex events with epoll_pwait. The events are signals, X11 events and recently added timers/timeout.
What I understand as of now, I need to separate user interaction from the computations. The problem with my setup, as of now, is that rate of X11 events changes in a way I can't explain.
So I decide to separate waiting on X11 events with the rest of the logic. Can you suggest a proper way to do it? Will having a X11 window in a separate thread/process/epoll set help?
And the actual question, as I look upon it now, is, what is a frequency of epoll_wait wakeups? I plan to have one epoll_wait in a loop. Maybe some processes to be wait on. I understand that epoll_wait will "wakeup" at some random points int time.
Update
My setup is close to this one:
#include <stdbool.h>
#include <string.h>
#include <errno.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <signal.h>
#include <time.h>
#include <math.h>
#include <sys/types.h>
#include <sys/ioctl.h>
#include <sys/signalfd.h>
#include <sys/epoll.h>
#include <sys/timerfd.h>
#include <sys/socket.h>
#include <xcb/xcb.h>
static struct timespec tm_p, tm_c, tm_diff;
static unsigned long nevents = 0;
static inline void timespec_diff(struct timespec *a, struct timespec *b, struct timespec *res) {
res->tv_sec = a->tv_sec - b->tv_sec;
res->tv_nsec = a->tv_nsec - b->tv_nsec;
if (res->tv_nsec < 0) {
--res->tv_sec;
res->tv_nsec += 1000000000L;
}
}
static double compute(){
double t = 1.0;
double eps = 1e-7;
double eps_2 = eps / 2.0;
clock_gettime(CLOCK_MONOTONIC, &tm_p);
while(t > eps){
t -= eps;
t += eps;
t -= eps_2 + eps_2;
}
clock_gettime(CLOCK_MONOTONIC, &tm_c);
timespec_diff(&tm_c, &tm_p, &tm_diff);
printf(" compute: %ld %f\n", tm_diff.tv_sec, tm_diff.tv_nsec / 1e9);
return (int)t;
}
/* defines for epoll */
#define MAXEVENTS 64
#define SET_EV(_ev,_fd,_events) _ev.data.fd = _fd; _ev.events = _events
static int xcb_get_atom(xcb_connection_t *c,
const char *name,
xcb_atom_t *atom)
{
xcb_intern_atom_cookie_t cookie;
xcb_generic_error_t *error;
xcb_intern_atom_reply_t *reply;
cookie = xcb_intern_atom(c, 0, strlen(name), name);
reply = xcb_intern_atom_reply(c, cookie, &error);
if(NULL == reply){
free(error);
return -1;
}
*atom = reply->atom;
free(reply);
return 0;
}
static int xcb_change_prop_wm_close(xcb_connection_t *c,
xcb_window_t window,
xcb_atom_t wm_p,
xcb_atom_t atom)
{
xcb_void_cookie_t cookie;
xcb_generic_error_t *error;
xcb_atom_enum_t type = XCB_ATOM_ATOM;
uint8_t format = 32;
uint32_t data_len = 1;
cookie = xcb_change_property_checked(c, /* xcb connection */
XCB_PROP_MODE_REPLACE, /* mode */
window, /* window */
wm_p, /* the property to change */
type, /* type of the property */
format, /* format(bits) */
data_len, /* number of elements(see format) */
&atom /* property data */
);
error = xcb_request_check(c, cookie);
if (error) {
free(error);
return -1;
}
return 0;
}
int main()
{
xcb_connection_t *c;
xcb_screen_t *screen;
xcb_window_t win;
xcb_atom_t a_wm_p;
xcb_atom_t wm_p_close;
struct epoll_event ep_ev, *ep_evs = NULL;
struct signalfd_siginfo siginf;
sigset_t mask_sigs, mask_osigs;
int sig_fd = -1, x11_fd = -1, ep_fd = -1, tm_fd = -1;
/* set up signals */
if(sigemptyset(&mask_sigs) < 0){
perror(" * sigemptyset(&mask_sigs)");
goto main_terminate;
}
if(sigaddset(&mask_sigs, SIGINT)){ /* these signals will be blocked. the signals will arrive */
perror(" * sigaddset(&mask_sigs, SIGINT)");
goto main_terminate;
}
if(sigaddset(&mask_sigs, SIGQUIT)){ /* to epoll and not to a default signal handler. */
perror(" * sigaddset(&mask_sigs, SIGQUIT)");
goto main_terminate;
}
/* save old sigmask, replace it with new sigmask */
if(sigprocmask(SIG_BLOCK, &mask_sigs, &mask_osigs) < 0){
perror(" * sigprocmask(SIG_BLOCK, &mask_sigs, &mask_osigs)");
goto main_terminate;
}
/* get signal file descriptor */
if((sig_fd = signalfd(-1, &mask_sigs, 0)) < 0){
perror(" * signalfd(-1, &mask_sigs, 0)");
goto main_terminate;
}
/* set signal fd as non-blocking */
{
int on = 1;
if(ioctl(sig_fd, FIONBIO, (char *)&on) < 0){
perror(" * ioctl(sig_fd, FIONBIO)");
goto main_terminate;
}
}
/* Open the connection to the X server */
c = xcb_connect (NULL, NULL);
/* Get the first screen */
screen = xcb_setup_roots_iterator (xcb_get_setup (c)).data;
/* Ask for our window's Id */
win = xcb_generate_id(c);
/* Create the window */
{
unsigned int cw_mask = XCB_CW_BORDER_PIXEL
| XCB_CW_EVENT_MASK
;
/* values must follow in the incresing order of the cw_mask constants */
unsigned int cw_values[] = {screen->white_pixel,
XCB_EVENT_MASK_KEY_PRESS|XCB_EVENT_MASK_KEY_RELEASE
};
xcb_create_window (c, /* Connection */
XCB_COPY_FROM_PARENT, /* depth (same as root)*/
win, /* window Id */
screen->root, /* parent window */
0, 0, /* x, y */
150, 150, /* width, height */
10, /* border_width */
XCB_WINDOW_CLASS_INPUT_OUTPUT, /* class */
screen->root_visual, /* visual */
cw_mask, cw_values); /* masks */
}
/* get x11 connection file descriptor */
x11_fd = xcb_get_file_descriptor(c);
/* atom WM_PROTOCOLS */
if(xcb_get_atom(c, "WM_PROTOCOLS", &a_wm_p) < 0){
fprintf(stderr, " %s:%s:%d\n", __FILE__, __func__, __LINE__);
return -1;
}
/* atom window close */
if(xcb_get_atom(c, "WM_DELETE_WINDOW", &wm_p_close) < 0){
fprintf(stderr, " %s:%s:%d\n", __FILE__, __func__, __LINE__);
return -1;
}
{ /* wm prop: intercept close */
if(xcb_change_prop_wm_close(c, win, a_wm_p, wm_p_close) < 0){
fprintf(stderr, " %s:%s:%d\n", __FILE__, __func__, __LINE__);
goto main_terminate;
}
}
/* create epoll set file descriptor */
if((ep_fd = epoll_create(1)) < 0){
perror(" * epoll_create");
goto main_terminate;
}
/* allocate events for epoll queue */
if(NULL == (ep_evs = (struct epoll_event*)calloc(MAXEVENTS,sizeof(ep_ev)))){
perror(" * calloc(MAXEVENTS)");
goto main_terminate;
}
{ /* fd timer */
struct itimerspec ts;
if((tm_fd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK)) < 0){
perror(" * timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK)");
goto main_terminate;
}
ts.it_value.tv_sec = 1;
ts.it_value.tv_nsec = 0;
ts.it_interval.tv_sec = 0;
ts.it_interval.tv_nsec = (unsigned long)10e6; /* 10 msec */
if(timerfd_settime(tm_fd, 0, &ts, NULL) < 0){
perror(" * timerfd_settime(tm_fd, 0, &ts, NULL)");
goto main_terminate;
}
}
/* add X11 event */
SET_EV(ep_ev,x11_fd,EPOLLIN);
if(epoll_ctl (ep_fd, EPOLL_CTL_ADD, x11_fd, &ep_ev) < 0){
perror(" * epoll_ctl x11_fd");
goto main_terminate;
}
/* add timer event */
SET_EV(ep_ev,tm_fd,EPOLLIN);
if(epoll_ctl (ep_fd, EPOLL_CTL_ADD, tm_fd, &ep_ev) < 0){
perror(" * epoll_ctl tm_fd");
goto main_terminate;
}
/* add signal event */
SET_EV(ep_ev,sig_fd,EPOLLIN);
if(epoll_ctl (ep_fd, EPOLL_CTL_ADD, sig_fd, &ep_ev) < 0){
perror(" * epoll_ctl sig_fd");
goto main_terminate;
}
/* window title */
const char *title = "epoll_pwait";
xcb_change_property (c,
XCB_PROP_MODE_REPLACE,
win,
XCB_ATOM_WM_NAME,
XCB_ATOM_STRING,
8,
strlen (title),
title );
/* Map the window on the screen */
xcb_map_window (c, win);
/* Make sure commands are sent before we pause, so window is shown */
xcb_flush (c);
while(1){
int n, i, fd, status;
bool f_compute = false;
bool f_exit_sig = false;
bool f_win_close = false;
n = epoll_pwait (ep_fd, ep_evs, MAXEVENTS, -1, &mask_sigs); /* wait, signal safe */
if(n < 0){
fprintf(stderr, " * main(): %s:%s:%d\n", __FILE__, __func__, __LINE__);
status = 1;
goto main_terminate;
}
for(i = 0; i < n; ++i){ /* service epoll events */
fd = ep_evs[i].data.fd;
if(fd == sig_fd){ /* signal */
status = read(fd, &siginf, sizeof(siginf));
if(status != sizeof(siginf)){
fprintf(stderr,"read != sizeof(siginf)");
goto main_terminate;
}
if(siginf.ssi_signo == SIGINT){
printf("got SIGINT\n");
f_exit_sig = true;
}else if(siginf.ssi_signo == SIGQUIT){
printf("got SIGQUIT\n");
f_exit_sig = true;
goto main_terminate;
}else {
printf("got unregistered signal\n");
}
}else if(fd == x11_fd){ /* x11 event */
xcb_generic_event_t *event;
while((event = xcb_poll_for_event(c))){
if (event && (event->response_type == 0)){ /* error recieved */
free(event);
goto main_terminate;
}
switch(event->response_type & ~0x80){
case XCB_CLIENT_MESSAGE: { /* window close */
xcb_client_message_event_t *ce = (xcb_client_message_event_t*)event;
if(ce->data.data32[0] == wm_p_close){ /* window should close */
f_win_close = true;
}
} break;
case XCB_KEY_PRESS: { /* keyboard key press */
printf("XCB_KEY_PRESS\n");
nevents++;
} break;
} /* end switch */
free(event);
} /* end while event loop */
}else if(fd == tm_fd){ /* timer event */
uint64_t overrun;
status = read(fd, &overrun, sizeof(uint64_t));
if(status != EAGAIN) {
//~ printf(" ! timer overruns: %lu\n", overrun);
}
f_compute = true;
}
} /* finish service epoll events */
if(f_exit_sig){ /* exit signal */
goto main_terminate;
}
if(f_win_close){ /* window close */
goto main_terminate;
}
if(f_compute){ /* do some computations */
compute();
xcb_flush(c);
}
} /* end while(1) */
main_terminate:
if(sig_fd != -1){
close(sig_fd);
}
if(tm_fd != -1){
close(tm_fd);
}
if(ep_fd != -1){
close(ep_fd);
}
if(ep_evs){
free(ep_evs);
}
xcb_disconnect(c);
if (sigprocmask(SIG_SETMASK, &mask_osigs, NULL) < 0){
perror(" * sigprocmask(SIG_SETMASK, &mask_osigs, NULL)");
}
printf("received %lu events\n", nevents);
return 0;
}
With the above, I can not reproduce input lag I do have in more verbose program. I test the above with xdotool, send some input X11 events, and my eye can't catch any visible input delay. All events get delivered. For now I can not post a full code that I have problem with.
what is a frequency of epoll_wait wakeups?
I'll assume you are talking about the precision of wakeups. This surely depends on kernel details, but let's just make an experiment: We can write a small program that tests how long epoll_wait takes. This will test both a timeout of zero and a timeout of 1 ms since a timeout of zero might be treated specially.
Output on my system:
empty: avg 0.015636 μs, min 0.014 μs, max 0.065 μs
sleep 1 ms: avg 1108.55 μs, min 1014.72 μs, max 1577.42 μs
epoll_wait 0: avg 0.761084 μs, min 0.38 μs, max 37.787 μs
epoll_wait 1: avg 1108.97 μs, min 1017.22 μs, max 2602.4 μs
Doing nothing (empty) measures less than a microsecond, so the following results should be somewhat reliable. The minimum time is also not zero, so the clock has enough precision for what we are doing.
Sleeping 1 ms sleeps at least 1.014 ms, but there also was a case of 1.5 ms. I guess that means that wakeups are not all that precise.
Using epoll_wait() to do nothing take less than a microsecond. More than doing nothing, but this still does basically nothing, so perhaps this really just measures the syscall overhead...?
Sleeping 1 ms with epoll_wait() behaves more or less the same as sleeping for 1 ms with nanosleep().
If you want to improve this experiment, you could actually register some FDs via epoll_ctl(). It might be that the kernel handles "empty" epoll FDs specially.
#include <stdio.h>
#include <sys/epoll.h>
#include <time.h>
#include <unistd.h>
#define MEASUREMENT_COUNT 10000
#define NUM_EVENTS 42
static int epoll_fd;
double diff_micro_sec(struct timespec *a, struct timespec *b) {
double sec = a->tv_sec - b->tv_sec;
double ns = a->tv_nsec - b->tv_nsec;
return sec * 1e6 + ns / 1e3;
}
static void empty(void) {
}
static void sleep_one_ms(void) {
struct timespec spec;
spec.tv_sec = 0;
spec.tv_nsec = 1000 * 1000;
nanosleep(&spec, NULL);
}
static void epoll_wait_0(void) {
struct epoll_event events[NUM_EVENTS];
epoll_wait(epoll_fd, events, NUM_EVENTS, 0);
}
static void epoll_wait_1(void) {
struct epoll_event events[NUM_EVENTS];
epoll_wait(epoll_fd, events, NUM_EVENTS, 1);
}
static void do_it(const char *name, void (*func)(void)) {
double sum, min, max;
struct timespec a, b;
for (int i = 0; i < MEASUREMENT_COUNT; i++) {
double diff;
clock_gettime(CLOCK_MONOTONIC, &a);
func();
clock_gettime(CLOCK_MONOTONIC, &b);
diff = diff_micro_sec(&b, &a);
if (i == 0) {
sum = diff;
min = diff;
max = diff;
} else {
sum += diff;
if (diff < min)
min = diff;
if (diff > max)
max = diff;
}
}
printf("%14s: avg %g μs, min %g μs, max %g μs\n", name, sum / MEASUREMENT_COUNT, min, max);
}
int main() {
do_it("empty", empty);
do_it("sleep 1 ms", sleep_one_ms);
epoll_fd = epoll_create(1);
do_it("epoll_wait 0", epoll_wait_0);
do_it("epoll_wait 1", epoll_wait_1);
close(epoll_fd);
return 0;
}

Prevent msgrcv from waiting indefinitely

I have a multihreaded program whose 2 threads communicate with each other via a message queue. The first thread (sender) periodically sends a message, while the second thread (receiver) processes the information.
The sender has code similar to this:
// Create queue
key_t key = ftok("/tmp", 'B');
int msqid = msgget(key, 0664 | IPC_CREAT);
// Create message and send
struct request_msg req_msg;
req_msg.mtype = 1;
snprintf(req_msg.mtext, MSG_LENGTH, "Send this information");
msgsnd(msqid, &req_msg, strlen(req_msg.mtext) + 1, 0);
On the receiving thread, I do this:
// Subscribe to queue
key_t key = ftok("/tmp", 'B');
int msqid = msgget(key, 0664);
struct request_msg req_msg;
while(running)
{
msgrcv(msqid, &req_msg, sizeof(req_msg.mtext), 0, 0);
// Do sth with the message
}
As you can see, the receiver sits within a while loop that is controlled by a global variable named "running". Error handlers do set the boolean to false, if an error is encountered within the process. This works in most cases, but if an error occurs before being able to send a message to the queue, the receiver will not exit the while loop because it waits for a message before continuing and thus, checking the running variable. That means it will hang there forever, as the sender will not send anything for the rest of the runtime.
I would like to avoid this, but I do not know how to let msgrcv know that it cannot expect any more messages. I was unable to learn how msgrcv behaves if I kill the queue, assuming this is the easiest version. Maybe timeouts or sending some kind of termination message (possibly using the mtype member of the message struct) are also possible.
Please, let me know what the most robust solution to this problem is. Thanks!
EDIT: based on suggestions I have reworked the code to make the signal handlers action atomic.
#include <stdbool.h> // bool data type
#include <stdio.h>
#include <signal.h>
#include <stdint.h>
#include <pthread.h>
#include <stdlib.h>
#include <unistd.h>
#define ALARM_INTERVAL_SEC 1
#define ALARM_INTERVAL_USEC 0
struct message
{
uint64_t iteration;
char req_time[28];
};
static volatile bool running = true;
static volatile bool work = false;
static struct itimerval alarm_interval;
static struct timeval previous_time;
static uint64_t loop_count = 0;
static struct message msg;
pthread_mutex_t mutexmsg;
pthread_cond_t data_updated_cv;
static void
termination_handler(int signum)
{
running = false;
}
static void
alarm_handler(int signum)
{
work = true;
}
static void
write_msg(void)
{
// Reset the alarm interval
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("setitimer");
raise(SIGTERM);
return;
}
struct timeval current_time;
gettimeofday(&current_time, NULL);
printf("\nLoop count: %lu\n", loop_count);
printf("Loop time: %f us\n", (current_time.tv_sec - previous_time.tv_sec) * 1e6 +
(current_time.tv_usec - previous_time.tv_usec));
previous_time = current_time;
// format timeval struct
char tmbuf[64];
time_t nowtime = current_time.tv_sec;
struct tm *nowtm = localtime(&nowtime);
strftime(tmbuf, sizeof(tmbuf), "%Y-%m-%d %H:%M:%S", nowtm);
// write values
pthread_mutex_lock(&mutexmsg);
msg.iteration = loop_count;
snprintf(msg.req_time, sizeof(msg.req_time), "%s.%06ld", tmbuf, current_time.tv_usec);
pthread_cond_signal(&data_updated_cv);
pthread_mutex_unlock(&mutexmsg);
loop_count++;
}
static void*
process_msg(void *args)
{
while(1)
{
pthread_mutex_lock(&mutexmsg);
printf("Waiting for condition\n");
pthread_cond_wait(&data_updated_cv, &mutexmsg);
printf("Condition fulfilled\n");
if(!running)
{
break;
}
struct timeval process_time;
gettimeofday(&process_time, NULL);
char tmbuf[64];
char buf[64];
time_t nowtime = process_time.tv_sec;
struct tm *nowtm = localtime(&nowtime);
strftime(tmbuf, sizeof(tmbuf), "%Y-%m-%d %H:%M:%S", nowtm);
snprintf(buf, sizeof(buf), "%s.%06ld", tmbuf, process_time.tv_usec);
// something that takes longer than the interval time
// sleep(1);
printf("[%s] Req time: %s loop cnt: %lu\n", buf, msg.req_time, msg.iteration);
pthread_mutex_unlock(&mutexmsg);
}
pthread_exit(NULL);
}
int
main(int argc, char* argv[])
{
pthread_t thread_id;
pthread_attr_t attr;
// for portability, set thread explicitly as joinable
pthread_attr_init(&attr);
pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
if(pthread_create(&thread_id, NULL, process_msg, NULL) != 0)
{
perror("pthread_create");
exit(1);
}
pthread_attr_destroy(&attr);
// signal handling setup
struct sigaction t;
t.sa_handler = termination_handler;
sigemptyset(&t.sa_mask);
t.sa_flags = 0;
sigaction(SIGINT, &t, NULL);
sigaction(SIGTERM, &t, NULL);
struct sigaction a;
a.sa_handler = alarm_handler;
sigemptyset(&a.sa_mask);
a.sa_flags = 0;
sigaction(SIGALRM, &a, NULL);
// Set the alarm interval
alarm_interval.it_interval.tv_sec = 0;
alarm_interval.it_interval.tv_usec = 0;
alarm_interval.it_value.tv_sec = ALARM_INTERVAL_SEC;
alarm_interval.it_value.tv_usec = ALARM_INTERVAL_USEC;
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("setitimer");
exit(1);
}
gettimeofday(&previous_time, NULL);
while(1)
{
// suspending main thread until a signal is caught
pause();
if(!running)
{
// signal the worker thread to stop execution
pthread_mutex_lock(&mutexmsg);
pthread_cond_signal(&data_updated_cv);
pthread_mutex_unlock(&mutexmsg);
break;
}
if(work)
{
write_msg();
work = false;
}
}
// suspend thread until the worker thread joins back in
pthread_join(thread_id, NULL);
// reset the timer
alarm_interval.it_value.tv_sec = 0;
alarm_interval.it_value.tv_usec = 0;
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("setitimer");
exit(1);
}
printf("EXIT\n");
pthread_exit(NULL);
}
You have not justified the use of a message queue other than as a synchronization primitive. You could be passing the message via a variable and an atomic flag to indicate message readiness. This answer then describes how to implement thread suspension and resuming using a condition variable. That’s how it’d be typically done between threads, although of course is not the only way.
I do not know how to let msgrcv know that it cannot expect any more messages
No need for that. Just send a message that tells the thread to finish! The running variable doesn’t belong: you are trying to communicate with the other thread, so do it the way you chose to: message it!
I have spent the last day to read a lot about threading and mutexes and tried to get my example program to work. It does, but unfortunately, it gets stuck when I try to shut it down via Ctrl+C. Reason being (again) that this time, that the worker thread waits for a signal from the main thread that won't send a signal anymore.
#Rachid K. and #Unslander Monica: if you want to take a look again, is this more state of the art code for doing this? Also, I think I have to use pthread_cond_timedwait instead of pthread_cond_wait to avoid the termination deadlock. Could you tell me how to handle that exactly?
Note that the program does simply periodically (interval 1 s) hand a timestamp and a loop counter to the processing thread, that prints out the data. The output also shows when the print got called.
Thanks again!
#include <stdbool.h> // bool data type
#include <stdio.h>
#include <signal.h>
#include <stdint.h>
#include <pthread.h>
#include <stdlib.h>
#define ALARM_INTERVAL_SEC 1
#define ALARM_INTERVAL_USEC 0
static bool running = true;
static struct itimerval alarm_interval;
static struct timeval previous_time;
static uint64_t loop_count = 0;
struct message
{
uint64_t iteration;
char req_time[28];
} msg;
pthread_mutex_t mutexmsg;
pthread_cond_t data_updated_cv;
static void
signal_handler(int signum)
{
if (signum == SIGINT || signum == SIGTERM)
{
running = false;
}
}
static void
write_msg(int signum)
{
if(!running)
{
return;
}
// Reset the alarm interval
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("setitimer");
raise(SIGTERM);
return;
}
struct timeval current_time;
gettimeofday(&current_time, NULL);
printf("\nLoop count: %lu\n", loop_count);
printf("Loop time: %f us\n", (current_time.tv_sec - previous_time.tv_sec) * 1e6 +
(current_time.tv_usec - previous_time.tv_usec));
previous_time = current_time;
// format timeval struct
char tmbuf[64];
time_t nowtime = current_time.tv_sec;
struct tm *nowtm = localtime(&nowtime);
strftime(tmbuf, sizeof(tmbuf), "%Y-%m-%d %H:%M:%S", nowtm);
// write values
pthread_mutex_lock(&mutexmsg);
msg.iteration = loop_count;
snprintf(msg.req_time, sizeof(msg.req_time), "%s.%06ld", tmbuf, current_time.tv_usec);
pthread_cond_signal(&data_updated_cv);
pthread_mutex_unlock(&mutexmsg);
loop_count++;
}
static void*
process_msg(void *args)
{
while(running)
{
pthread_mutex_lock(&mutexmsg);
printf("Waiting for condition\n");
pthread_cond_wait(&data_updated_cv, &mutexmsg);
printf("Condition fulfilled\n");
struct timeval process_time;
gettimeofday(&process_time, NULL);
char tmbuf[64];
char buf[64];
time_t nowtime = process_time.tv_sec;
struct tm *nowtm = localtime(&nowtime);
strftime(tmbuf, sizeof(tmbuf), "%Y-%m-%d %H:%M:%S", nowtm);
snprintf(buf, sizeof(buf), "%s.%06ld", tmbuf, process_time.tv_usec);
printf("[%s] Message req time: %s loop cnt: %lu\n", buf, msg.req_time, msg.iteration);
pthread_mutex_unlock(&mutexmsg);
}
pthread_exit(NULL);
}
int
main(int argc, char* argv[])
{
pthread_t thread_id;
pthread_attr_t attr;
// for portability, set thread explicitly as joinable
pthread_attr_init(&attr);
pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
if(pthread_create(&thread_id, NULL, process_msg, NULL) != 0)
{
perror("pthread_create");
exit(1);
}
pthread_attr_destroy(&attr);
// signal handling setup
struct sigaction s;
s.sa_handler = signal_handler;
sigemptyset(&s.sa_mask);
s.sa_flags = 0;
sigaction(SIGINT, &s, NULL);
sigaction(SIGTERM, &s, NULL);
struct sigaction a;
a.sa_handler = write_msg;
sigemptyset(&a.sa_mask);
a.sa_flags = 0;
sigaction(SIGALRM, &a, NULL);
// Set the alarm interval
alarm_interval.it_interval.tv_sec = 0;
alarm_interval.it_interval.tv_usec = 0;
alarm_interval.it_value.tv_sec = ALARM_INTERVAL_SEC;
alarm_interval.it_value.tv_usec = ALARM_INTERVAL_USEC;
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("setitimer");
exit(1);
}
gettimeofday(&previous_time, NULL);
// suspend thread until the worker thread joins back in
pthread_join(thread_id, NULL);
// reset the timer
alarm_interval.it_value.tv_sec = 0;
alarm_interval.it_value.tv_usec = 0;
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("setitimer");
exit(1);
}
pthread_exit(NULL);
return 0;
}
On the receiving thread, I do this:
...
while(running)
{
msgrcv(msqid, &req_msg, sizeof(req_msg.mtext), 0, 0);
Hopefully, in reality you do more than that.
Because you're not checking any error status in the code you've posted. And that's flat-out wrong for a blocking function call that is likely specified to never be restarted on receipt of a signal (as is true on Linux and Solaris). Per Linux `signal(2):
The following interfaces are never restarted after being interrupted
by a signal handler, regardless of the use of SA_RESTART; they
always fail with the error EINTR when interrupted by a signal
handler:
...
System V IPC interfaces: msgrcv(2), msgsnd(2), semop(2), and
semtimedop(2).
and Solaris sigaction():
SA_RESTART
If set and the signal is caught, functions that are interrupted by the execution of this signal's handler are transparently restarted by the system, namely fcntl(2), ioctl(2), wait(3C), waitid(2), and the following functions on slow devices like terminals: getmsg() and getpmsg() (see getmsg(2)); putmsg() and putpmsg() (see putmsg(2)); pread(), read(), and readv() (see read(2)); pwrite(), write(), and writev() (see write(2)); recv(), recvfrom(), and recvmsg() (see recv(3SOCKET)); and send(), sendto(), and sendmsg() (see send(3SOCKET)). Otherwise, the function returns an EINTR error.
So your code need to look more like this in order to handle both errors and signal interrupts:
volatile sig_atomic_t running;
...
while(running)
{
errno = 0;
ssize_t result = msgrcv(msqid, &req_msg, sizeof(req_msg.mtext), 0, 0);
if ( result == ( ssize_t ) -1 )
{
// if the call failed or no longer running
// break the loop
if ( ( errno != EINTR ) || !running )
{
break;
}
// the call was interrupted by a signal
continue
}
...
}
And that opens up the opportunity to use a alarm() and a SIGALRM signal handler to set running to 0 for use as a timeout:
volatile sig_atomic_t running;
void handler( int sig );
{
running = 0;
}
...
struct sigaction sa;
memset( &sa, 0, sizeof( sa ) );
sa.sa_handler = handler;
sigaction( SIGALRM, &sa, NULL );
while(running)
{
// 10-sec timeout
alarm( 10 );
errno = 0;
ssize_t result = msgrcv( msqid, &req_msg, sizeof(req_msg.mtext), 0, 0 );
// save errno as alarm() can munge it
int saved_errno = errno;
// clear alarm if it hasn't fired yet
alarm( 0 );
if ( result == ( ssize_t ) -1 )
{
// if the call failed or no longer running
// break the loop
if ( ( saved_errno != EINTR ) || !running )
{
break;
}
// the call was interrupted by a signal
continue
}
...
}
That can almost certainly be improved upon - the logic is rather complex to catch all the corner cases and there's likely a simpler way to do it.
Answer to the new proposal in the question
The cyclic timer should be rearmed in the event loop of the main thread for better visibility (subjective proposal);
When the secondary thread goes out of its loop, it must release the
mutex otherwise the main thread will enter into a deadlock (waiting
for the mutex which was locked by the terminated secondary thread).
So, here is the last proposal with the above fixes/enhancements:
#include <stdbool.h> // bool data type
#include <stdio.h>
#include <signal.h>
#include <stdint.h>
#include <pthread.h>
#include <stdlib.h>
#include <unistd.h>
#include <sys/time.h>
#define ALARM_INTERVAL_SEC 1
#define ALARM_INTERVAL_USEC 0
struct message
{
uint64_t iteration;
char req_time[28];
};
static volatile bool running = true;
static volatile bool work = false;
static struct itimerval alarm_interval;
static struct timeval previous_time;
static uint64_t loop_count = 0;
static struct message msg;
pthread_mutex_t mutexmsg;
pthread_cond_t data_updated_cv;
static void
termination_handler(int signum)
{
running = false;
}
static void
alarm_handler(int signum)
{
work = true;
}
static void
write_msg(void)
{
struct timeval current_time;
gettimeofday(&current_time, NULL);
printf("\nLoop count: %lu\n", loop_count);
printf("Loop time: %f us\n", (current_time.tv_sec - previous_time.tv_sec) * 1e6 +
(current_time.tv_usec - previous_time.tv_usec));
previous_time = current_time;
// format timeval struct
char tmbuf[64];
time_t nowtime = current_time.tv_sec;
struct tm *nowtm = localtime(&nowtime);
strftime(tmbuf, sizeof(tmbuf), "%Y-%m-%d %H:%M:%S", nowtm);
// write values
pthread_mutex_lock(&mutexmsg);
msg.iteration = loop_count;
snprintf(msg.req_time, sizeof(msg.req_time), "%s.%06ld", tmbuf, current_time.tv_usec);
pthread_cond_signal(&data_updated_cv);
pthread_mutex_unlock(&mutexmsg);
loop_count++;
}
static void*
process_msg(void *args)
{
while(1)
{
pthread_mutex_lock(&mutexmsg);
printf("Waiting for condition\n");
pthread_cond_wait(&data_updated_cv, &mutexmsg);
printf("Condition fulfilled\n");
if(!running)
{
pthread_mutex_unlock(&mutexmsg); // <----- To avoid deadlock
break;
}
struct timeval process_time;
gettimeofday(&process_time, NULL);
char tmbuf[64];
char buf[64];
time_t nowtime = process_time.tv_sec;
struct tm *nowtm = localtime(&nowtime);
strftime(tmbuf, sizeof(tmbuf), "%Y-%m-%d %H:%M:%S", nowtm);
snprintf(buf, sizeof(buf), "%s.%06ld", tmbuf, process_time.tv_usec);
// something that takes longer than the interval time
//sleep(2);
printf("[%s] Req time: %s loop cnt: %lu\n", buf, msg.req_time, msg.iteration);
pthread_mutex_unlock(&mutexmsg);
}
printf("Thread exiting...\n");
pthread_exit(NULL);
}
int
main(int argc, char* argv[])
{
pthread_t thread_id;
pthread_attr_t attr;
// for portability, set thread explicitly as joinable
pthread_attr_init(&attr);
pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
if(pthread_create(&thread_id, NULL, process_msg, NULL) != 0)
{
perror("pthread_create");
exit(1);
}
pthread_attr_destroy(&attr);
// signal handling setup
struct sigaction t;
t.sa_handler = termination_handler;
sigemptyset(&t.sa_mask);
t.sa_flags = 0;
sigaction(SIGINT, &t, NULL);
sigaction(SIGTERM, &t, NULL);
struct sigaction a;
a.sa_handler = alarm_handler;
sigemptyset(&a.sa_mask);
a.sa_flags = 0;
sigaction(SIGALRM, &a, NULL);
// Set the alarm interval
alarm_interval.it_interval.tv_sec = 0;
alarm_interval.it_interval.tv_usec = 0;
alarm_interval.it_value.tv_sec = ALARM_INTERVAL_SEC;
alarm_interval.it_value.tv_usec = ALARM_INTERVAL_USEC;
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("setitimer");
exit(1);
}
gettimeofday(&previous_time, NULL);
while(1)
{
// Reset the alarm interval <-------- Rearm the timer in the main loop
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("setitimer");
raise(SIGTERM);
break;
}
// suspending main thread until a signal is caught
pause();
if(!running)
{
// signal the worker thread to stop execution
pthread_mutex_lock(&mutexmsg);
pthread_cond_signal(&data_updated_cv);
pthread_mutex_unlock(&mutexmsg);
break;
}
if(work)
{
write_msg();
work = false;
}
}
// suspend thread until the worker thread joins back in
pthread_join(thread_id, NULL);
// reset the timer
alarm_interval.it_value.tv_sec = 0;
alarm_interval.it_value.tv_usec = 0;
if(setitimer(ITIMER_REAL, &alarm_interval, NULL) < 0)
{
perror("setitimer");
exit(1);
}
printf("EXIT\n");
pthread_exit(NULL);
}
==================================================================
Answer to the original question
It is possible to use a conditional variable to wait for a signal from the senders. This makes the receiver wake up and check for messages in the message queue by passing IPC_NOWAIT in the flags parameter of msgrcv(). To end the communication, an "End of communication" message can be posted. It is also possible to use pthread_con_timedwait() to wake up periodically and check a "end of communication" or "end of receiver condition" (e.g. by checking your global "running" variable).
Receiver side:
// Mutex initialization
pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
// Condition variable initialization
pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
[...]
while (1) {
// Lock the mutex
pthread_mutex_lock(&mutex);
// Check for messages (non blocking thanks to IPC_NOWAIT)
rc = msgrcv(msqid, &req_msg, sizeof(req_msg.mtext), 0, IPC_NOWAIT);
if (rc == -1) {
if (errno == ENOMSG) {
// message queue empty
// Wait for message notification
pthread_cond_wait(&cond, &mutex); // <--- pthread_cond_timedwait() can be used to wake up and check for the end of communication or senders...
} else {
// Error
}
}
// Handle the message, end of communication (e.g. "running" variable)...
// Release the lock (so that the sender can post something in the queue)
pthread_mutex_unlock(&mutex);
}
Sender side:
// Prepare the message
[...]
// Take the lock
pthread_mutex_lock(&mutex);
// Send the message
msgsnd(msqid, &req_msg, strlen(req_msg.mtext) + 1, 0);
// Wake up the receiver
pthread_cond_signal(&cond);
// Release the lock
pthread_mutex_unlock(&mutex);
N.B.: SYSV message queues are obsolete. It is preferable to use the brand new Posix services.

timerfd won't be ready to read when using epoll

I'm intending to use epoll to check out timerfd and fire some actions.
Code is blow:
#include <time.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/timerfd.h>
#include <stdint.h>
#include <unistd.h>
#include <sys/epoll.h>
int main(int argc, char const *argv[])
{
struct timespec now;
clock_gettime(CLOCK_MONOTONIC, &now);
int timerfd;
timerfd = timerfd_create(CLOCK_MONOTONIC, 0);
struct itimerspec new_value;
new_value.it_value.tv_sec = 1;
new_value.it_interval.tv_sec = 1;
timerfd_settime(timerfd, 0, &new_value, NULL);
// uint64_t buff;
// while(true) {
// read(timerfd, &buff, sizeof(uint64_t));
// printf("%s\n", "ding");
// }
// code above works fine.
struct epoll_event ev, events[10];
int epollfd;
epollfd = epoll_create1(0);
if (epollfd == -1) {
perror("epoll_create1");
exit(EXIT_FAILURE);
}
ev.events = EPOLLIN;
ev.data.fd = timerfd;
if (epoll_ctl(epollfd, EPOLL_CTL_ADD, timerfd, &ev) == -1) {
perror("epoll_ctl: timerfd");
exit(EXIT_FAILURE);
}
int num;
printf("start\n");
while(true) {
num = epoll_wait(epollfd, events, 10, -1);
printf("%d\n", num);
uint64_t buff;
read(timerfd, &buff, sizeof(uint64_t));
printf("%s\n", "ding");
}
return 0;
}
When using timerfd seperately, it works fine. Every second will print "ding". But when adding epoll to observe timerfd, progrom will block at epoll_wait for ever.
I'v tryed using EPOLLET, but noting changed. What's wrong with this code?
Your itimerspec is not properly initialized, so depending on what particular garbage values it contains, timerfd_settime() might fail. To detect that, do error checking:
if (timerfd_settime(timerfd, 0, &new_value, NULL) != 0) {
perror("settime");
exit(-1);
}
Another way to debug this is to run your program under the strace, program, and you will see which, if any, system calls that fails.
The relevant structs looks like this:
struct timespec {
time_t tv_sec;
long tv_nsec;
};
struct itimerspec {
struct timespec it_interval;
struct timespec it_value;
};
You have to initialize both these members completely, and your program will work reliably:
new_value.it_value.tv_sec = 1;
new_value.it_value.tv_nsec = 0;
new_value.it_interval.tv_sec = 1;
new_value.it_interval.tv_nsec = 0;

Repeating timerfd event works with epoll and not with poll

I am implementing a timer using timerfd. This is a relative timer that I just need to repeat forever at the rate it is set to. I want to poll on this event and originally tried using poll. When I did this, I would see the timer event the first time and then never again. However, when I changed to using epoll (no change at all to how the timerfd was set up) it works as expected.
Here is the code with poll:
#include <sys/timerfd.h>
#include <sys/poll.h>
#include <sys/epoll.h>
#include <stdio.h>
#include <stdlib.h>
#include <strings.h>
#include <unistd.h>
int main(int ac, char *av[])
{
struct pollfd p;
int timerfd;
struct itimerspec timerValue;
/* clear pollfd */
bzero(&p, sizeof(p));
/* set timerfd */
timerfd = timerfd_create(CLOCK_REALTIME, 0);
if (timerfd < 0) {
printf("failed to create timer fd\n");
exit(1);
}
bzero(&timerValue, sizeof(timerValue));
timerValue.it_value.tv_sec = 1;
timerValue.it_value.tv_nsec = 0;
timerValue.it_interval.tv_sec = 1;
timerValue.it_interval.tv_nsec = 0;
/* set events */
p.fd = timerfd;
p.revents = 0;
p.events = POLLIN;
/* start timer */
if (timerfd_settime(timerfd, 0, &timerValue, NULL) < 0) {
printf("could not start timer\n");
exit(1);
}
/* wait for events */
while (1) {
int numEvents = poll(&p, 1, -1);
if (numEvents > 0) {
int timersElapsed = 0;
(void) read(p.fd, &timersElapsed, 8);
printf("timers elapsed: %d\n", timersElapsed);
}
}
exit(0);
}
And here is the code with epoll:
#include <sys/timerfd.h>
#include <sys/poll.h>
#include <sys/epoll.h>
#include <stdio.h>
#include <stdlib.h>
#include <strings.h>
#include <unistd.h>
int main(int ac, char *av[])
{
struct epoll_event epollEvent;
struct epoll_event newEvents;
int timerfd;
int epollfd;
struct itimerspec timerValue;
/* set timerfd */
timerfd = timerfd_create(CLOCK_MONOTONIC, 0);
if (timerfd < 0) {
printf("failed to create timer fd\n");
exit(1);
}
bzero(&timerValue, sizeof(timerValue));
timerValue.it_value.tv_sec = 1;
timerValue.it_value.tv_nsec = 0;
timerValue.it_interval.tv_sec = 1;
timerValue.it_interval.tv_nsec = 0;
/* set events */
epollfd = epoll_create1(0);
epollEvent.events = EPOLLIN;
epollEvent.data.fd = timerfd;
epoll_ctl(epollfd, EPOLL_CTL_ADD, timerfd, &epollEvent);
/* start timer */
if (timerfd_settime(timerfd, 0, &timerValue, NULL) < 0) {
printf("could not start timer\n");
exit(1);
}
/* wait for events */
while (1) {
int numEvents = epoll_wait(epollfd, &newEvents, 1, 0);
if (numEvents > 0) {
int timersElapsed = 0;
(void) read(epollEvent.data.fd, &timersElapsed, 8);
printf("timers elapsed: %d\n", timersElapsed);
}
}
exit(0);
}
Any idea what I might be doing wrong with poll? Maybe it is not meant to be used this way with a timerfd? Thank you.
Ok, this is an old question, but nevertheless. The problem lies in these lines of code:
int timersElapsed = 0;
(void) read(p.fd, &timersElapsed, 8);
printf("timers elapsed: %d\n", timersElapsed);
int timersElapsed is 4 bytes. Reading 8 bytes into this results in a stack overflow, giving unpredictable behaviour.
Changing timersElapsed to a long int and fixing the printf did the trick for me.
long int timersElapsed = 0;
(void) read(p.fd, &timersElapsed, 8);
printf("timers elapsed: %ld\n", timersElapsed);
This appears to be an issue with Fedora (or my installation of Fedora). That system is running 3.16, and poll() does not work.
However, on a separate Ubuntu installation with 3.13, the poll() code above works just fine. As I will be using Ubuntu in the future anyway, I will not try to track down the issue on Fedora. Though I am curious if others are seeing this same issue on Fedora systems.
I faced the same problem.
After debugging, the root cause in poll example is that
timerValue should be declared as uint64_t.
- int timersElapsed = 0;
+ uint64_t timersElapsed = 0;
The man page of timerfd_create() describes this.
Operating on a timer file descriptor
The file descriptor returned by timerfd_create() supports the following
operations:
read(2)
If the timer has already expired one or more times since its settings
were last modified using timerfd_settime(), or since the last suc‐
cessful read(2), then the buffer given to read(2) returns an unsigned
8-byte integer (uint64_t) containing the number of expirations that
have occurred. (The returned value is in host byte order—that is,
the native byte order for integers on the host machine.)

How do you make a precise countdown timer using clock_gettime?

Could somebody please explain how to make a countdown timer using clock_gettime, under Linux. I know you can use the clock() function to get cpu time, and multiply it by CLOCKS_PER_SEC to get actual time, but I'm told the clock() function is not well suited for this.
So far I have attempted this (a billion is to pause for one second)
#include <stdio.h>
#include <time.h>
#define BILLION 1000000000
int main()
{
struct timespec rawtime;
clock_gettime(CLOCK_MONOTONIC_RAW, &rawtime);
unsigned long int current = ( rawtime.tv_sec + rawtime.tv_nsec );
unsigned long int end = (( rawtime.tv_sec + rawtime.tv_nsec ) + BILLION );
while ( current < end )
{
clock_gettime(CLOCK_MONOTONIC_RAW, &rawtime);
current = ( rawtime.tv_sec + rawtime.tv_nsec );
}
return 0;
}
I know this wouldn't be very useful on its own, but once I've found out how to time correctly I can use this in my projects. I know that sleep() can be used for this purpose, but I want to code the timer myself so that I can better integrate it in my projects - such as the possibility of it returning the time left, as opposed to pausing the whole program.
Please, do not do that. You're burning CPU power for nothing in a busy loop.
Why not use the nanosleep() function instead? It's perfectly suited to the use case you outlined. Or, if you want an easier interface, perhaps something like
#define _POSIX_C_SOURCE 200809L
#include <time.h>
#include <errno.h>
/* Sleep for the specified number of seconds,
* and return the time left over.
*/
double dsleep(const double seconds)
{
struct timespec req, rem;
/* No sleep? */
if (seconds <= 0.0)
return 0.0;
/* Convert to seconds and nanoseconds. */
req.tv_sec = (time_t)seconds;
req.tv_nsec = (long)((seconds - (double)req.tv_sec) * 1000000000.0);
/* Take care of any rounding errors. */
if (req.tv_nsec < 0L)
req.tv_nsec = 0L;
else
if (req.tv_nsec > 999999999L)
req.tv_nsec = 999999999L;
/* Do the nanosleep. */
if (nanosleep(&req, &rem) != -1)
return 0.0;
/* Error? */
if (errno != EINTR)
return 0.0;
/* Return remainder. */
return (double)rem.tv_sec + (double)rem.tv_nsec / 1000000000.0;
}
The difference is that using this one the CPU is free to do something else, rather than spin like a crazed squirrel on speed.
This is not an answer, but an example of how to use signals and a POSIX timer to implement a timeout timer; intended as a response to the OP's followup question in a comment to the accepted answer.
#define _POSIX_C_SOURCE 200809L
#include <stdlib.h>
#include <signal.h>
#include <time.h>
#include <errno.h>
#include <string.h>
#include <stdio.h>
/* Timeout timer.
*/
static timer_t timeout_timer;
static volatile sig_atomic_t timeout_state = 0;
static volatile sig_atomic_t timeout_armed = 2;
static const int timeout_signo = SIGALRM;
#define TIMEDOUT() (timeout_state != 0)
/* Timeout signal handler.
*/
static void timeout_handler(int signo, siginfo_t *info, void *context __attribute__((unused)))
{
if (timeout_armed == 1)
if (signo == timeout_signo && info && info->si_code == SI_TIMER)
timeout_state = ~0;
}
/* Unset timeout.
* Returns nonzero if timeout had expired, zero otherwise.
*/
static int timeout_unset(void)
{
struct itimerspec t;
const int retval = timeout_state;
/* Not armed? */
if (timeout_armed != 1)
return retval;
/* Disarm. */
t.it_value.tv_sec = 0;
t.it_value.tv_nsec = 0;
t.it_interval.tv_sec = 0;
t.it_interval.tv_nsec = 0;
timer_settime(timeout_timer, 0, &t, NULL);
return retval;
}
/* Set timeout (in wall clock seconds).
* Cancels any pending timeouts.
*/
static int timeout_set(const double seconds)
{
struct itimerspec t;
/* Uninitialized yet? */
if (timeout_armed == 2) {
struct sigaction act;
struct sigevent evt;
/* Use timeout_handler() for timeout_signo signal. */
sigemptyset(&act.sa_mask);
act.sa_sigaction = timeout_handler;
act.sa_flags = SA_SIGINFO;
if (sigaction(timeout_signo, &act, NULL) == -1)
return errno;
/* Create a monotonic timer, delivering timeout_signo signal. */
evt.sigev_value.sival_ptr = NULL;
evt.sigev_signo = timeout_signo;
evt.sigev_notify = SIGEV_SIGNAL;
if (timer_create(CLOCK_MONOTONIC, &evt, &timeout_timer) == -1)
return errno;
/* Timeout is initialzied but unarmed. */
timeout_armed = 0;
}
/* Disarm timer, if armed. */
if (timeout_armed == 1) {
/* Set zero timeout, disarming the timer. */
t.it_value.tv_sec = 0;
t.it_value.tv_nsec = 0;
t.it_interval.tv_sec = 0;
t.it_interval.tv_nsec = 0;
if (timer_settime(timeout_timer, 0, &t, NULL) == -1)
return errno;
timeout_armed = 0;
}
/* Clear timeout state. It should be safe (no pending signals). */
timeout_state = 0;
/* Invalid timeout? */
if (seconds <= 0.0)
return errno = EINVAL;
/* Set new timeout. Check for underflow/overflow. */
t.it_value.tv_sec = (time_t)seconds;
t.it_value.tv_nsec = (long)((seconds - (double)t.it_value.tv_sec) * 1000000000.0);
if (t.it_value.tv_nsec < 0L)
t.it_value.tv_nsec = 0L;
else
if (t.it_value.tv_nsec > 999999999L)
t.it_value.tv_nsec = 999999999L;
/* Set it repeat once every millisecond, just in case the initial
* interrupt is missed. */
t.it_interval.tv_sec = 0;
t.it_interval.tv_nsec = 1000000L;
if (timer_settime(timeout_timer, 0, &t, NULL) == -1)
return errno;
timeout_armed = 1;
return 0;
}
int main(void)
{
char *line = NULL;
size_t size = 0;
ssize_t len;
fprintf(stderr, "Please supply input. The program will exit automatically if\n");
fprintf(stderr, "it takes more than five seconds for the next line to arrive.\n");
fflush(stderr);
while (1) {
if (timeout_set(5.0)) {
const char *const errmsg = strerror(errno);
fprintf(stderr, "Cannot set timeout: %s.\n", errmsg);
return 1;
}
len = getline(&line, &size, stdin);
if (len == (ssize_t)-1)
break;
if (len < (ssize_t)1) {
/* This should never occur (except for -1, of course). */
errno = EIO;
break;
}
/* We do not want *output* to be interrupted,
* so we cancel the timeout. */
timeout_unset();
if (fwrite(line, (size_t)len, 1, stdout) != 1) {
fprintf(stderr, "Error writing to standard output.\n");
fflush(stderr);
return 1;
}
fflush(stdout);
/* Next line. */
}
/* Remember to cancel the timeout. Also check it. */
if (timeout_unset())
fprintf(stderr, "Timed out.\n");
else
if (ferror(stdin) || !feof(stdin))
fprintf(stderr, "Error reading standard input.\n");
else
fprintf(stderr, "End of input.\n");
fflush(stderr);
/* Free line buffer. */
free(line);
line = NULL;
size = 0;
/* Done. */
return 0;
}
If you save the above as timer.c, you can compile it using e.g.
gcc -W -Wall -O3 -std=c99 -pedantic timer.c -lrt -o timer
and run it using ./timer.
If you read the code above carefully, you'll see that it is actually a periodic timer signal (at millisecond intervals), with a variable delay before the first signal. That is just a technique I like to use to make sure I don't miss the signal. (The signal repeats until the timeout is unset.)
Note that although you can do computation in an signal handler, you should only use functions that are async-signal-safe; see man 7 signal. Also, only the sig_atomic_t type is atomic wrt. normal single-threaded code and a signal handler. So, it is better to just use the signal as an indicator, and do the actual code in your own program.
If you wanted to e.g. update monster coordinates in a signal handler, it is possible but a bit tricky. I'd use three arrays containing the monster information, and use GCC __sync_bool_compare_and_swap() to update the array pointers -- very much the same technique as triple-buffering in graphics.
If you need more than one concurrent timeout, you could use multiple timers (there is a number of them available), but the best option is to define timeout slots. (You can use generation counters to detect "forgotten" timeouts, and so on.) Whenever a new timeout is set or unset, you update the timeout to reflect the next timeout that expires. It's a bit more code, but really a straightforward extension of the above.

Resources