illumos gate

Kernel doesn't require to call thread_join() on threads, instead it uses reaper-thread. It brought LWP <defunc> leaking in libfakekernel, due to nobody calls thread_join() or thr_join() for almost all <defunc> threads.

This patch modifies thread_exit() wrapper:
"orders" to reap <defunc> thread to the call next ensuing.

Testing done:

#include <sys/cmn_err.h>
#include <sys/thread.h>
#include <sys/zone.h>
#include <sys/proc.h>
#include <atomic.h>
#include <unistd.h>

#define    N 10
extern int rand(void);

static void
worker(void)
{
    (void) sleep(rand() % N + 1);
    thread_exit();
}

int
main(int argc, char **argv)
{
    unsigned i;
    kthread_t *tid;

    for (i = 0; i < 100000; i++) {
        tid = thread_create(NULL, 0, worker, NULL, 0, NULL, TS_RUN, 0);
        printf("tid: %ld\n", (uintptr_t)tid);
        (void) sleep(1);
    }
    return (0);
}

Used this test program to generate threads and observed with mdb -k '::threadlist' the number of threads. Number of threads was ~10 all times - so the exited threads got reaped.