nptl: Fix stray process left by tst-cancel7 blocking testing

Fix an issue with commit b74121ae4b ("Update.") and prevent a stray
process from being left behind by tst-cancel7 (and also tst-cancelx7,
which is the same test built with '-fexceptions' additionally supplied
to the compiler), which then blocks remote testing until the process has
been killed by hand.

This test case creates a thread that runs an extra copy of the test via
system(3) and using the '--direct' option so that the test wrapper does
not interfere with this instance.  This extra copy executes its business
and calls sigsuspend(2) and then never terminates by itself.  Instead it
relies on being killed by the main test process directly via a thread
cancellation request or, should that fail, by issuing SIGKILL either at
the conclusion of 'do_test' or by the test driver via 'do_cleanup' where
the test timeout has been hit or the test driver interrupted.

However if the main test process has been instead killed by a signal,
such as due to incorrect execution, before it had a chance to kill the
extra copy of the test case, then the test wrapper will terminate
without running 'do_cleanup' and consequently the extra copy of the test
case will remain forever in its suspended state, and in the remote case
in particular it means that the remote test wrapper will wait forever
for the SSH command to complete.

This has been observed with the 'alpha-linux-gnu' target, where the main
test process triggers SIGSEGV and the test wrapper correctly records:

Didn't expect signal from child: got `Segmentation fault'

in nptl/tst-cancel7.out and terminates, but then the calling SSH command
continues waiting for the remaining process started in the same session
on the remote target to complete.

Address this problem by also registering 'do_cleanup' via atexit(3),
observing that 'support_delete_temp_files' is registered by the test
wrapper before the test initializing function 'do_prepare' is called and
that we call all the functions registered in the reverse of the order in
which they were registered, so it is safe to refer to 'pidfilename' in
'do_cleanup' invoked by exit(3) because by that time temporary files
have not yet been deleted.

A minor inconvenience is that if 'signal_handler' is invoked in the test
wrapper as a result of SIGALRM rather than SIGINT, then 'do_cleanup'
will be called twice, once as a cleanup handler and again by exit(3).
In reality it is harmless though, because issuing SIGKILL is guarded by
a record lock, so if the first call has succeeded in killing the extra
copy of the test case, then the subsequent call will do nothing.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>

nptl/tst-cancel7.c

@@ -38,6 +38,8 @@ static char *semfilename;
 
 static sem_t *sem;
 
+static void do_cleanup (void);
+
 static void *
 tf (void *arg)
 {
@@ -108,6 +110,8 @@ do_prepare (int argc, char *argv[])
 
   xwrite (fd, " ", 1);
   xclose (fd);
+
+  atexit (do_cleanup);
 }