Question

I’m debugging a select loop that normally works OK but dies with segmentation fault under heavy load. I’ve figured out that the program is sometimes invoking FD_ISSET() for a (correct) descriptor that was not added to the select set. Like in a following snippet:

#include <sys/select.h>
#include <sys/time.h>
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>

void die(const char* msg)
{
    fprintf(stderr, "fatal %s", msg);
    exit(1);
}

int main(void)
{
    FILE* file = fopen("/tmp/test", "r");
    if (file == NULL)
        die("fopen");

    int file_fd = fileno(file);
    fd_set read_fds;
    int max_fd = 0;

    FD_ZERO(&read_fds);
    // Only stdin is added to read_fds.
    FD_SET(0, &read_fds);

    if (select(max_fd + 1, &read_fds, NULL, NULL, NULL) < 0)
        die("select");
    if (FD_ISSET(0, &read_fds))
        printf("Can read from 0");
    // !!! Here FD_ISSET is called with a valid descriptor that was 
    // not added to read_fds.
    if (FD_ISSET(file_fd, &read_fds))
        printf("Can read from file_fd");
    return 0;
}

It is obvious that the check marked with !!! should never return true, but is it possible that it can be the cause of the SEGFAULT? When I run this snippet under valgrind, no errors are reported, but when I run my load test under valgrind I’m ocasionnaly seing errors like:

==25513== Syscall param select(writefds) points to uninitialised byte(s)
==25513==    at 0x435DD2D: ___newselect_nocancel (syscall-template.S:82)

Answer 1

FD_ISSET() tests to see if a file descriptor is a part of the set read_fds. This means that FD_ISSET should not cause the segmentation fault.

Try checking for errno value set prior to calling the FD_ISSET. The select should be causing the segfault.

Also check that the file_fd value isn’t greater than FD_MAX.