Closed
Description
In the following program, a number of threads are made, and then each thread forks of a child that sleeps forever and then immediately kills it. I would expect this program to succeed continuously, but it wedges on OSX occasionally, reporting a successful signal delivery, but failing to actually deliver the signal apparently.
#include <stdlib.h>
#include <pthread.h>
#include <signal.h>
#include <stdio.h>
#include <unistd.h>
#include <assert.h>
void *work(void *arg) {
#define M 100
int i;
for (i = 0; i < M; i++) {
int a = fork();
assert(a >= 0);
if (a == 0) {
sleep(1000000000);
return arg;
}
assert(kill(a, SIGTERM) == 0);
int ret;
assert(waitpid(a, &ret, 0) == a);
assert(!WIFEXITED(ret));
}
return arg;
}
int main() {
#define N 8
pthread_t c[N];
int i;
for (i = 0; i < N; i++) {
assert(pthread_create(&c[i], NULL, work, (void*) (size_t) i) == 0);
}
for (i = 0; i < N; i++) {
assert(pthread_join(c[i], NULL) == 0);
}
}This is essentially how we fork() in libnative, and it's how we're using fork from libgreen. Trying to investigate a solution to this, but I'm starting to think that multithreaded fork is just fundamentally broken on basically all platforms except linux.
This issue has appeared as various forms of flakiness on the bots, which is why I started investigating.