Quelle  pidfd_info_test.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0

#define _GNU_SOURCE
#include <errno.h>
#include <fcntl.h>
#include <limits.h>
#include <linux/types.h>
#include <poll.h>
#include <pthread.h>
#include <sched.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <syscall.h>
#include <sys/prctl.h>
#include <sys/wait.h>
#include <unistd.h>
#include <sys/socket.h>
#include <linux/kcmp.h>
#include <sys/stat.h>

#include "pidfd.h"
#include "../kselftest_harness.h"

FIXTURE(pidfd_info)
{
 pid_t child_pid1;
 int child_pidfd1;

 pid_t child_pid2;
 int child_pidfd2;

 pid_t child_pid3;
 int child_pidfd3;

 pid_t child_pid4;
 int child_pidfd4;
};

FIXTURE_SETUP(pidfd_info)
{
 int ret;
 int ipc_sockets[2];
 char c;

 ret = socketpair(AF_LOCAL, SOCK_STREAM | SOCK_CLOEXEC, 0, ipc_sockets);
 EXPECT_EQ(ret, 0);

 self->child_pid1 = create_child(&self->child_pidfd1, 0);
 EXPECT_GE(self->child_pid1, 0);

 if (self->child_pid1 == 0) {
  close(ipc_sockets[0]);

  if (write_nointr(ipc_sockets[1], "1", 1) < 0)
   _exit(EXIT_FAILURE);

  close(ipc_sockets[1]);

  pause();
  _exit(EXIT_SUCCESS);
 }

 EXPECT_EQ(close(ipc_sockets[1]), 0);
 ASSERT_EQ(read_nointr(ipc_sockets[0], &c, 1), 1);
 EXPECT_EQ(close(ipc_sockets[0]), 0);

 /* SIGKILL but don't reap. */
 EXPECT_EQ(sys_pidfd_send_signal(self->child_pidfd1, SIGKILL, NULL, 0), 0);

 ret = socketpair(AF_LOCAL, SOCK_STREAM | SOCK_CLOEXEC, 0, ipc_sockets);
 EXPECT_EQ(ret, 0);

 self->child_pid2 = create_child(&self->child_pidfd2, 0);
 EXPECT_GE(self->child_pid2, 0);

 if (self->child_pid2 == 0) {
  close(ipc_sockets[0]);

  if (write_nointr(ipc_sockets[1], "1", 1) < 0)
   _exit(EXIT_FAILURE);

  close(ipc_sockets[1]);

  pause();
  _exit(EXIT_SUCCESS);
 }

 EXPECT_EQ(close(ipc_sockets[1]), 0);
 ASSERT_EQ(read_nointr(ipc_sockets[0], &c, 1), 1);
 EXPECT_EQ(close(ipc_sockets[0]), 0);

 /* SIGKILL and reap. */
 EXPECT_EQ(sys_pidfd_send_signal(self->child_pidfd2, SIGKILL, NULL, 0), 0);
 EXPECT_EQ(sys_waitid(P_PID, self->child_pid2, NULL, WEXITED), 0);

 self->child_pid3 = create_child(&self->child_pidfd3, CLONE_NEWUSER | CLONE_NEWPID);
 EXPECT_GE(self->child_pid3, 0);

 if (self->child_pid3 == 0)
  _exit(EXIT_SUCCESS);

 self->child_pid4 = create_child(&self->child_pidfd4, CLONE_NEWUSER | CLONE_NEWPID);
 EXPECT_GE(self->child_pid4, 0);

 if (self->child_pid4 == 0)
  _exit(EXIT_SUCCESS);

 EXPECT_EQ(sys_waitid(P_PID, self->child_pid4, NULL, WEXITED), 0);
}

FIXTURE_TEARDOWN(pidfd_info)
{
 sys_pidfd_send_signal(self->child_pidfd1, SIGKILL, NULL, 0);
 if (self->child_pidfd1 >= 0)
  EXPECT_EQ(0, close(self->child_pidfd1));

 sys_waitid(P_PID, self->child_pid1, NULL, WEXITED);

 sys_pidfd_send_signal(self->child_pidfd2, SIGKILL, NULL, 0);
 if (self->child_pidfd2 >= 0)
  EXPECT_EQ(0, close(self->child_pidfd2));

 sys_waitid(P_PID, self->child_pid2, NULL, WEXITED);
 sys_waitid(P_PID, self->child_pid3, NULL, WEXITED);
 sys_waitid(P_PID, self->child_pid4, NULL, WEXITED);
}

TEST_F(pidfd_info, sigkill_exit)
{
 struct pidfd_info info = {
  .mask = PIDFD_INFO_CGROUPID,
 };

 /* Process has exited but not been reaped so this must work. */
 ASSERT_EQ(ioctl(self->child_pidfd1, PIDFD_GET_INFO, &info), 0);

 info.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(self->child_pidfd1, PIDFD_GET_INFO, &info), 0);
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_CREDS));
 /* Process has exited but not been reaped, so no PIDFD_INFO_EXIT information yet. */
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_EXIT));
}

TEST_F(pidfd_info, sigkill_reaped)
{
 struct pidfd_info info = {
  .mask = PIDFD_INFO_CGROUPID,
 };

 /* Process has already been reaped and PIDFD_INFO_EXIT hasn't been set. */
 ASSERT_NE(ioctl(self->child_pidfd2, PIDFD_GET_INFO, &info), 0);
 ASSERT_EQ(errno, ESRCH);

 info.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(self->child_pidfd2, PIDFD_GET_INFO, &info), 0);
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_CREDS));
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_EXIT));
 ASSERT_TRUE(WIFSIGNALED(info.exit_code));
 ASSERT_EQ(WTERMSIG(info.exit_code), SIGKILL);
}

TEST_F(pidfd_info, success_exit)
{
 struct pidfd_info info = {
  .mask = PIDFD_INFO_CGROUPID,
 };

 /* Process has exited but not been reaped so this must work. */
 ASSERT_EQ(ioctl(self->child_pidfd3, PIDFD_GET_INFO, &info), 0);

 info.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(self->child_pidfd3, PIDFD_GET_INFO, &info), 0);
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_CREDS));
 /* Process has exited but not been reaped, so no PIDFD_INFO_EXIT information yet. */
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_EXIT));
}

TEST_F(pidfd_info, success_reaped)
{
 struct pidfd_info info = {
  .mask = PIDFD_INFO_CGROUPID,
 };

 /* Process has already been reaped and PIDFD_INFO_EXIT hasn't been set. */
 ASSERT_NE(ioctl(self->child_pidfd4, PIDFD_GET_INFO, &info), 0);
 ASSERT_EQ(errno, ESRCH);

 info.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(self->child_pidfd4, PIDFD_GET_INFO, &info), 0);
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_CREDS));
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_EXIT));
 ASSERT_TRUE(WIFEXITED(info.exit_code));
 ASSERT_EQ(WEXITSTATUS(info.exit_code), 0);
}

TEST_F(pidfd_info, success_reaped_poll)
{
 struct pidfd_info info = {
  .mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT,
 };
 struct pollfd fds = {};
 int nevents;

 fds.events = POLLIN;
 fds.fd = self->child_pidfd2;

 nevents = poll(&fds, 1, -1);
 ASSERT_EQ(nevents, 1);
 ASSERT_TRUE(!!(fds.revents & POLLIN));
 ASSERT_TRUE(!!(fds.revents & POLLHUP));

 ASSERT_EQ(ioctl(self->child_pidfd2, PIDFD_GET_INFO, &info), 0);
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_CREDS));
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_EXIT));
 ASSERT_TRUE(WIFSIGNALED(info.exit_code));
 ASSERT_EQ(WTERMSIG(info.exit_code), SIGKILL);
}

static void *pidfd_info_pause_thread(void *arg)
{
 pid_t pid_thread = gettid();
 int ipc_socket = *(int *)arg;

 /* Inform the grand-parent what the tid of this thread is. */
 if (write_nointr(ipc_socket, &pid_thread, sizeof(pid_thread)) != sizeof(pid_thread))
  return NULL;

 close(ipc_socket);

 /* Sleep untill we're killed. */
 pause();
 return NULL;
}

TEST_F(pidfd_info, thread_group)
{
 pid_t pid_leader, pid_poller, pid_thread;
 pthread_t thread;
 int nevents, pidfd_leader, pidfd_thread, pidfd_leader_thread, ret;
 int ipc_sockets[2];
 struct pollfd fds = {};
 struct pidfd_info info = {
  .mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT,
 }, info2;

 ret = socketpair(AF_LOCAL, SOCK_STREAM | SOCK_CLOEXEC, 0, ipc_sockets);
 EXPECT_EQ(ret, 0);

 pid_leader = create_child(&pidfd_leader, 0);
 EXPECT_GE(pid_leader, 0);

 if (pid_leader == 0) {
  close(ipc_sockets[0]);

  /* The thread will outlive the thread-group leader. */
  if (pthread_create(&thread, NULL, pidfd_info_pause_thread, &ipc_sockets[1]))
   syscall(__NR_exit, EXIT_FAILURE);

  /* Make the thread-group leader exit prematurely. */
  syscall(__NR_exit, EXIT_SUCCESS);
 }

 /*
  * Opening a PIDFD_THREAD aka thread-specific pidfd based on a
  * thread-group leader must succeed.
 */
 pidfd_leader_thread = sys_pidfd_open(pid_leader, PIDFD_THREAD);
 ASSERT_GE(pidfd_leader_thread, 0);

 pid_poller = fork();
 ASSERT_GE(pid_poller, 0);
 if (pid_poller == 0) {
  /*
   * We can't poll and wait for the old thread-group
   * leader to exit using a thread-specific pidfd. The
   * thread-group leader exited prematurely and
   * notification is delayed until all subthreads have
   * exited.
 */
  fds.events = POLLIN;
  fds.fd = pidfd_leader_thread;
  nevents = poll(&fds, 1, 10000 /* wait 5 seconds */);
  if (nevents != 0)
   _exit(EXIT_FAILURE);
  if (fds.revents & POLLIN)
   _exit(EXIT_FAILURE);
  if (fds.revents & POLLHUP)
   _exit(EXIT_FAILURE);
  _exit(EXIT_SUCCESS);
 }

 /* Retrieve the tid of the thread. */
 EXPECT_EQ(close(ipc_sockets[1]), 0);
 ASSERT_EQ(read_nointr(ipc_sockets[0], &pid_thread, sizeof(pid_thread)), sizeof(pid_thread));
 EXPECT_EQ(close(ipc_sockets[0]), 0);

 /* Opening a thread as a thread-group leader must fail. */
 pidfd_thread = sys_pidfd_open(pid_thread, 0);
 ASSERT_LT(pidfd_thread, 0);
 ASSERT_EQ(errno, ENOENT);

 /* Opening a thread as a PIDFD_THREAD must succeed. */
 pidfd_thread = sys_pidfd_open(pid_thread, PIDFD_THREAD);
 ASSERT_GE(pidfd_thread, 0);

 ASSERT_EQ(wait_for_pid(pid_poller), 0);

 /*
  * Note that pidfd_leader is a thread-group pidfd, so polling on it
  * would only notify us once all thread in the thread-group have
  * exited. So we can't poll before we have taken down the whole
  * thread-group.
 */

 /* Get PIDFD_GET_INFO using the thread-group leader pidfd. */
 ASSERT_EQ(ioctl(pidfd_leader, PIDFD_GET_INFO, &info), 0);
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_CREDS));
 /* Process has exited but not been reaped, so no PIDFD_INFO_EXIT information yet. */
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_EXIT));
 ASSERT_EQ(info.pid, pid_leader);

 /*
  * Now retrieve the same info using the thread specific pidfd
  * for the thread-group leader.
 */
 info2.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(pidfd_leader_thread, PIDFD_GET_INFO, &info2), 0);
 ASSERT_TRUE(!!(info2.mask & PIDFD_INFO_CREDS));
 /* Process has exited but not been reaped, so no PIDFD_INFO_EXIT information yet. */
 ASSERT_FALSE(!!(info2.mask & PIDFD_INFO_EXIT));
 ASSERT_EQ(info2.pid, pid_leader);

 /* Now try the thread-specific pidfd. */
 ASSERT_EQ(ioctl(pidfd_thread, PIDFD_GET_INFO, &info), 0);
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_CREDS));
 /* The thread hasn't exited, so no PIDFD_INFO_EXIT information yet. */
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_EXIT));
 ASSERT_EQ(info.pid, pid_thread);

 /*
  * Take down the whole thread-group. The thread-group leader
  * exited successfully but the thread will now be SIGKILLed.
  * This must be reflected in the recorded exit information.
 */
 EXPECT_EQ(sys_pidfd_send_signal(pidfd_leader, SIGKILL, NULL, 0), 0);
 EXPECT_EQ(sys_waitid(P_PIDFD, pidfd_leader, NULL, WEXITED), 0);

 fds.events = POLLIN;
 fds.fd = pidfd_leader;
 nevents = poll(&fds, 1, -1);
 ASSERT_EQ(nevents, 1);
 ASSERT_TRUE(!!(fds.revents & POLLIN));
 /* The thread-group leader has been reaped. */
 ASSERT_TRUE(!!(fds.revents & POLLHUP));

 /*
  * Retrieve exit information for the thread-group leader via the
  * thread-group leader pidfd.
 */
 info.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(pidfd_leader, PIDFD_GET_INFO, &info), 0);
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_CREDS));
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_EXIT));
 /* Even though the thread-group exited successfully it will still report the group exit code. */
 ASSERT_TRUE(WIFSIGNALED(info.exit_code));
 ASSERT_EQ(WTERMSIG(info.exit_code), SIGKILL);

 /*
  * Retrieve exit information for the thread-group leader via the
  * thread-specific pidfd.
 */
 info2.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(pidfd_leader_thread, PIDFD_GET_INFO, &info2), 0);
 ASSERT_FALSE(!!(info2.mask & PIDFD_INFO_CREDS));
 ASSERT_TRUE(!!(info2.mask & PIDFD_INFO_EXIT));

 /* Even though the thread-group exited successfully it will still report the group exit code. */
 ASSERT_TRUE(WIFSIGNALED(info2.exit_code));
 ASSERT_EQ(WTERMSIG(info2.exit_code), SIGKILL);

 /* Retrieve exit information for the thread. */
 info.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(pidfd_thread, PIDFD_GET_INFO, &info), 0);
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_CREDS));
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_EXIT));

 /* The thread got SIGKILLed. */
 ASSERT_TRUE(WIFSIGNALED(info.exit_code));
 ASSERT_EQ(WTERMSIG(info.exit_code), SIGKILL);

 EXPECT_EQ(close(pidfd_leader), 0);
 EXPECT_EQ(close(pidfd_thread), 0);
}

static void *pidfd_info_thread_exec(void *arg)
{
 pid_t pid_thread = gettid();
 int ipc_socket = *(int *)arg;

 /* Inform the grand-parent what the tid of this thread is. */
 if (write_nointr(ipc_socket, &pid_thread, sizeof(pid_thread)) != sizeof(pid_thread))
  return NULL;

 if (read_nointr(ipc_socket, &pid_thread, sizeof(pid_thread)) != sizeof(pid_thread))
  return NULL;

 close(ipc_socket);

 sys_execveat(AT_FDCWD, "pidfd_exec_helper", NULL, NULL, 0);
 return NULL;
}

TEST_F(pidfd_info, thread_group_exec)
{
 pid_t pid_leader, pid_poller, pid_thread;
 pthread_t thread;
 int nevents, pidfd_leader, pidfd_leader_thread, pidfd_thread, ret;
 int ipc_sockets[2];
 struct pollfd fds = {};
 struct pidfd_info info = {
  .mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT,
 };

 ret = socketpair(AF_LOCAL, SOCK_STREAM | SOCK_CLOEXEC, 0, ipc_sockets);
 EXPECT_EQ(ret, 0);

 pid_leader = create_child(&pidfd_leader, 0);
 EXPECT_GE(pid_leader, 0);

 if (pid_leader == 0) {
  close(ipc_sockets[0]);

  /* The thread will outlive the thread-group leader. */
  if (pthread_create(&thread, NULL, pidfd_info_thread_exec, &ipc_sockets[1]))
   syscall(__NR_exit, EXIT_FAILURE);

  /* Make the thread-group leader exit prematurely. */
  syscall(__NR_exit, EXIT_SUCCESS);
 }

 /* Open a thread-specific pidfd for the thread-group leader. */
 pidfd_leader_thread = sys_pidfd_open(pid_leader, PIDFD_THREAD);
 ASSERT_GE(pidfd_leader_thread, 0);

 pid_poller = fork();
 ASSERT_GE(pid_poller, 0);
 if (pid_poller == 0) {
  /*
   * We can't poll and wait for the old thread-group
   * leader to exit using a thread-specific pidfd. The
   * thread-group leader exited prematurely and
   * notification is delayed until all subthreads have
   * exited.
   *
   * When the thread has execed it will taken over the old
   * thread-group leaders struct pid. Calling poll after
   * the thread execed will thus block again because a new
   * thread-group has started.
 */
  fds.events = POLLIN;
  fds.fd = pidfd_leader_thread;
  nevents = poll(&fds, 1, 10000 /* wait 5 seconds */);
  if (nevents != 0)
   _exit(EXIT_FAILURE);
  if (fds.revents & POLLIN)
   _exit(EXIT_FAILURE);
  if (fds.revents & POLLHUP)
   _exit(EXIT_FAILURE);
  _exit(EXIT_SUCCESS);
 }

 /* Retrieve the tid of the thread. */
 EXPECT_EQ(close(ipc_sockets[1]), 0);
 ASSERT_EQ(read_nointr(ipc_sockets[0], &pid_thread, sizeof(pid_thread)), sizeof(pid_thread));

 /* Opening a thread as a PIDFD_THREAD must succeed. */
 pidfd_thread = sys_pidfd_open(pid_thread, PIDFD_THREAD);
 ASSERT_GE(pidfd_thread, 0);

 /* Now that we've opened a thread-specific pidfd the thread can exec. */
 ASSERT_EQ(write_nointr(ipc_sockets[0], &pid_thread, sizeof(pid_thread)), sizeof(pid_thread));
 EXPECT_EQ(close(ipc_sockets[0]), 0);

 ASSERT_EQ(wait_for_pid(pid_poller), 0);

 /* Wait until the kernel has SIGKILLed the thread. */
 fds.events = POLLHUP;
 fds.fd = pidfd_thread;
 nevents = poll(&fds, 1, -1);
 ASSERT_EQ(nevents, 1);
 /* The thread has been reaped. */
 ASSERT_TRUE(!!(fds.revents & POLLHUP));

 /* Retrieve thread-specific exit info from pidfd. */
 ASSERT_EQ(ioctl(pidfd_thread, PIDFD_GET_INFO, &info), 0);
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_CREDS));
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_EXIT));
 /*
  * While the kernel will have SIGKILLed the whole thread-group
  * during exec it will cause the individual threads to exit
  * cleanly.
 */
 ASSERT_TRUE(WIFEXITED(info.exit_code));
 ASSERT_EQ(WEXITSTATUS(info.exit_code), 0);

 /*
  * The thread-group leader is still alive, the thread has taken
  * over its struct pid and thus its pid number.
 */
 info.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(pidfd_leader, PIDFD_GET_INFO, &info), 0);
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_CREDS));
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_EXIT));
 ASSERT_EQ(info.pid, pid_leader);

 /* Take down the thread-group leader. */
 EXPECT_EQ(sys_pidfd_send_signal(pidfd_leader, SIGKILL, NULL, 0), 0);

 /*
  * Afte the exec we're dealing with an empty thread-group so now
  * we must see an exit notification on the thread-specific pidfd
  * for the thread-group leader as there's no subthread that can
  * revive the struct pid.
 */
 fds.events = POLLIN;
 fds.fd = pidfd_leader_thread;
 nevents = poll(&fds, 1, -1);
 ASSERT_EQ(nevents, 1);
 ASSERT_TRUE(!!(fds.revents & POLLIN));
 ASSERT_FALSE(!!(fds.revents & POLLHUP));

 EXPECT_EQ(sys_waitid(P_PIDFD, pidfd_leader, NULL, WEXITED), 0);

 /* Retrieve exit information for the thread-group leader. */
 info.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(pidfd_leader, PIDFD_GET_INFO, &info), 0);
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_CREDS));
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_EXIT));

 EXPECT_EQ(close(pidfd_leader), 0);
 EXPECT_EQ(close(pidfd_thread), 0);
}

static void *pidfd_info_thread_exec_sane(void *arg)
{
 pid_t pid_thread = gettid();
 int ipc_socket = *(int *)arg;

 /* Inform the grand-parent what the tid of this thread is. */
 if (write_nointr(ipc_socket, &pid_thread, sizeof(pid_thread)) != sizeof(pid_thread))
  return NULL;

 if (read_nointr(ipc_socket, &pid_thread, sizeof(pid_thread)) != sizeof(pid_thread))
  return NULL;

 close(ipc_socket);

 sys_execveat(AT_FDCWD, "pidfd_exec_helper", NULL, NULL, 0);
 return NULL;
}

TEST_F(pidfd_info, thread_group_exec_thread)
{
 pid_t pid_leader, pid_poller, pid_thread;
 pthread_t thread;
 int nevents, pidfd_leader, pidfd_leader_thread, pidfd_thread, ret;
 int ipc_sockets[2];
 struct pollfd fds = {};
 struct pidfd_info info = {
  .mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT,
 };

 ret = socketpair(AF_LOCAL, SOCK_STREAM | SOCK_CLOEXEC, 0, ipc_sockets);
 EXPECT_EQ(ret, 0);

 pid_leader = create_child(&pidfd_leader, 0);
 EXPECT_GE(pid_leader, 0);

 if (pid_leader == 0) {
  close(ipc_sockets[0]);

  /* The thread will outlive the thread-group leader. */
  if (pthread_create(&thread, NULL, pidfd_info_thread_exec_sane, &ipc_sockets[1]))
   syscall(__NR_exit, EXIT_FAILURE);

  /*
   * Pause the thread-group leader. It will be killed once
   * the subthread execs.
 */
  pause();
  syscall(__NR_exit, EXIT_SUCCESS);
 }

 /* Retrieve the tid of the thread. */
 EXPECT_EQ(close(ipc_sockets[1]), 0);
 ASSERT_EQ(read_nointr(ipc_sockets[0], &pid_thread, sizeof(pid_thread)), sizeof(pid_thread));

 /* Opening a thread as a PIDFD_THREAD must succeed. */
 pidfd_thread = sys_pidfd_open(pid_thread, PIDFD_THREAD);
 ASSERT_GE(pidfd_thread, 0);

 /* Open a thread-specific pidfd for the thread-group leader. */
 pidfd_leader_thread = sys_pidfd_open(pid_leader, PIDFD_THREAD);
 ASSERT_GE(pidfd_leader_thread, 0);

 pid_poller = fork();
 ASSERT_GE(pid_poller, 0);
 if (pid_poller == 0) {
  /*
   * The subthread will now exec. The struct pid of the old
   * thread-group leader will be assumed by the subthread which
   * becomes the new thread-group leader. So no exit notification
   * must be generated. Wait for 5 seconds and call it a success
   * if no notification has been received.
 */
  fds.events = POLLIN;
  fds.fd = pidfd_leader_thread;
  nevents = poll(&fds, 1, 10000 /* wait 5 seconds */);
  if (nevents != 0)
   _exit(EXIT_FAILURE);
  if (fds.revents & POLLIN)
   _exit(EXIT_FAILURE);
  if (fds.revents & POLLHUP)
   _exit(EXIT_FAILURE);
  _exit(EXIT_SUCCESS);
 }

 /* Now that we've opened a thread-specific pidfd the thread can exec. */
 ASSERT_EQ(write_nointr(ipc_sockets[0], &pid_thread, sizeof(pid_thread)), sizeof(pid_thread));
 EXPECT_EQ(close(ipc_sockets[0]), 0);
 ASSERT_EQ(wait_for_pid(pid_poller), 0);

 /* Wait until the kernel has SIGKILLed the thread. */
 fds.events = POLLHUP;
 fds.fd = pidfd_thread;
 nevents = poll(&fds, 1, -1);
 ASSERT_EQ(nevents, 1);
 /* The thread has been reaped. */
 ASSERT_TRUE(!!(fds.revents & POLLHUP));

 /* Retrieve thread-specific exit info from pidfd. */
 ASSERT_EQ(ioctl(pidfd_thread, PIDFD_GET_INFO, &info), 0);
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_CREDS));
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_EXIT));
 /*
  * While the kernel will have SIGKILLed the whole thread-group
  * during exec it will cause the individual threads to exit
  * cleanly.
 */
 ASSERT_TRUE(WIFEXITED(info.exit_code));
 ASSERT_EQ(WEXITSTATUS(info.exit_code), 0);

 /*
  * The thread-group leader is still alive, the thread has taken
  * over its struct pid and thus its pid number.
 */
 info.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(pidfd_leader, PIDFD_GET_INFO, &info), 0);
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_CREDS));
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_EXIT));
 ASSERT_EQ(info.pid, pid_leader);

 /* Take down the thread-group leader. */
 EXPECT_EQ(sys_pidfd_send_signal(pidfd_leader, SIGKILL, NULL, 0), 0);

 /*
  * Afte the exec we're dealing with an empty thread-group so now
  * we must see an exit notification on the thread-specific pidfd
  * for the thread-group leader as there's no subthread that can
  * revive the struct pid.
 */
 fds.events = POLLIN;
 fds.fd = pidfd_leader_thread;
 nevents = poll(&fds, 1, -1);
 ASSERT_EQ(nevents, 1);
 ASSERT_TRUE(!!(fds.revents & POLLIN));
 ASSERT_FALSE(!!(fds.revents & POLLHUP));

 EXPECT_EQ(sys_waitid(P_PIDFD, pidfd_leader, NULL, WEXITED), 0);

 /* Retrieve exit information for the thread-group leader. */
 info.mask = PIDFD_INFO_CGROUPID | PIDFD_INFO_EXIT;
 ASSERT_EQ(ioctl(pidfd_leader, PIDFD_GET_INFO, &info), 0);
 ASSERT_FALSE(!!(info.mask & PIDFD_INFO_CREDS));
 ASSERT_TRUE(!!(info.mask & PIDFD_INFO_EXIT));

 EXPECT_EQ(close(pidfd_leader), 0);
 EXPECT_EQ(close(pidfd_thread), 0);
}

TEST_HARNESS_MAIN