[LTP] [PATCH] syscalls/tgkill03: wait for defunct tid to get detached

[Jan Stancek]

----- Original Message -----
> On Sun, Jun 16, 2019 at 5:52 PM Jan Stancek <jstancek@redhat.com> wrote:
> 
> > Case where defunct tid is used has been observed to sporadically fail:
> >   tgkill03.c:96: FAIL: Defunct tid should have failed with ESRCH: SUCCESS
> >
> > glibc __pthread_timedjoin_ex() waits for CLONE_CHILD_CLEARTID to clear tid,
> > and then resumes. Kernel clears it (glibc pd->tid) at:
> >   do_exit
> >     exit_mm
> >       mm_release
> >         put_user(0, tsk->clear_child_tid);
> >
> > but kernel tid is still valid, presumably until:
> >   release_task
> >     __exit_signal
> >       __unhash_process
> >         detach_pid
> >
> > To avoid race wait until /proc/<pid>/task/<tid> disappears.
> >
> > Signed-off-by: Jan Stancek <jstancek@redhat.com>
> > ---
> >  testcases/kernel/syscalls/tgkill/tgkill03.c | 6 +++++-
> >  1 file changed, 5 insertions(+), 1 deletion(-)
> >
> > diff --git a/testcases/kernel/syscalls/tgkill/tgkill03.c
> > b/testcases/kernel/syscalls/tgkill/tgkill03.c
> > index f5bbdc5a8d4e..5ac1d2651f7a 100644
> > --- a/testcases/kernel/syscalls/tgkill/tgkill03.c
> > +++ b/testcases/kernel/syscalls/tgkill/tgkill03.c
> > @@ -7,6 +7,7 @@
> >
> >  #include <pthread.h>
> >  #include <pwd.h>
> > +#include <stdio.h>
> >  #include <sys/types.h>
> >
> >  #include "tst_safe_pthread.h"
> > @@ -42,6 +43,7 @@ static void setup(void)
> >  {
> >         sigset_t sigusr1;
> >         pthread_t defunct_thread;
> > +       char defunct_tid_path[PATH_MAX];
> >
> >         sigemptyset(&sigusr1);
> >         sigaddset(&sigusr1, SIGUSR1);
> > @@ -55,8 +57,10 @@ static void setup(void)
> >         TST_CHECKPOINT_WAIT(0);
> >
> >         SAFE_PTHREAD_CREATE(&defunct_thread, NULL, defunct_thread_func,
> > NULL);
> > -
> >         SAFE_PTHREAD_JOIN(defunct_thread, NULL);
> > +       sprintf(defunct_tid_path, "/proc/%d/task/%d", getpid(),
> > defunct_tid);
> > +       while (access(defunct_tid_path, R_OK) == 0)
> > +               usleep(10000);
> >
> 
> To be on the safe side, I think maybe TST_RETRY_FUNC is a better choice
> here.

Given high steal time on s390, I rather not put 1s cap on sleep here.
This is newlib test, so there's already a timeout, I'd prefer to lower
tst_test.timeout, say 30 seconds?

> 
>     TST_RETRY_FUNC(access(defunct_tid_path, R_OK), -1);
> 
>  }
> >
> >  static void cleanup(void)
> > --
> > 1.8.3.1
> >
> >
> 
> --
> Regards,
> Li Wang
>