[LTP] [PATCH] syscalls/ptrace07: handle potential SIGSEGV on older kernels

[LTP] [PATCH] syscalls/ptrace07: handle potential SIGSEGV on older kernels

[Jan Stancek]

If a ptraced test process hits SIGSEGV, the entire testcase hangs.

Older kernels such as RHEL7 (3.10.0), check the error code returned
by restore_fpu_checking() and do drop_init_fpu() if it fails.
So the FPU state of the prev task can't leak.

But in the more likely case a task with xcomp_bv != 0 will be killed
by SIGSEGV; either from do_device_not_available() or from
sys_rt_sigreturn()->__restore_xstate_sig().

And this is why the test case hangs; it wrongly assumes that the
traced child can only exit and report nothing else. But since it
receives SIGSEGV it reports this signal to the main process and
sleeps in ptrace_stop(), it does not exit and thus the test-case
hangs in tst_reap_children() after return from do_test().

Replace PTRACE_CONT with PTRACE_DETACH, so we don't need to
handle subsequent stops. And treat exit code from test process
as info-only.

Debugged-by: Oleg Nesterov <onestero@redhat.com>
Signed-off-by: Jan Stancek <jstancek@redhat.com>
---
 testcases/kernel/syscalls/ptrace/ptrace07.c | 26 +++++++++++++++++++++++---
 1 file changed, 23 insertions(+), 3 deletions(-)

diff --git a/testcases/kernel/syscalls/ptrace/ptrace07.c b/testcases/kernel/syscalls/ptrace/ptrace07.c
index 46ae59a105cf..9cbaefc3fbda 100644
--- a/testcases/kernel/syscalls/ptrace/ptrace07.c
+++ b/testcases/kernel/syscalls/ptrace/ptrace07.c
@@ -153,12 +153,32 @@ static void do_test(void)
 			"PTRACE_SETREGSET failed with unexpected error");
 	}
 
-	TEST(ptrace(PTRACE_CONT, pid, 0, 0));
+	/*
+	 * It is possible for test child 'pid' to crash on AMD
+	 * systems (e.g. AMD Opteron(TM) Processor 6234) with
+	 * older kernels. This causes tracee to stop and sleep
+	 * in ptrace_stop(). Without resuming the tracee, the
+	 * test hangs at do_test()->tst_reap_children() called
+	 * by the library. Use detach here, so we don't need to
+	 * worry about potential stops after this point.
+	 */
+	TEST(ptrace(PTRACE_DETACH, pid, 0, 0));
 	if (TST_RET != 0)
-		tst_brk(TBROK | TTERRNO, "PTRACE_CONT failed");
+		tst_brk(TBROK | TTERRNO, "PTRACE_DETACH failed");
+
+	/* If child 'pid' crashes, only report it as info. */
+	SAFE_WAITPID(pid, &status, 0);
+	if (WIFEXITED(status)) {
+		tst_res(TINFO, "test child %d exited, retcode: %d",
+			pid, WEXITSTATUS(status));
+	}
+	if (WIFSIGNALED(status)) {
+		tst_res(TINFO, "test child %d exited, termsig: %d",
+			pid, WTERMSIG(status));
+	}
 
 	okay = true;
-	for (i = 0; i < num_cpus + 1; i++) {
+	for (i = 0; i < num_cpus; i++) {
 		SAFE_WAIT(&status);
 		okay &= (WIFEXITED(status) && WEXITSTATUS(status) == 0);
 	}
-- 
1.8.3.1

[LTP] [PATCH] syscalls/ptrace07: handle potential SIGSEGV on older kernels

[Oleg Nesterov]

On 12/21, Jan Stancek wrote:
>
> --- a/testcases/kernel/syscalls/ptrace/ptrace07.c
> +++ b/testcases/kernel/syscalls/ptrace/ptrace07.c
> @@ -153,12 +153,32 @@ static void do_test(void)
>  			"PTRACE_SETREGSET failed with unexpected error");
>  	}
>  
> -	TEST(ptrace(PTRACE_CONT, pid, 0, 0));
> +	/*
> +	 * It is possible for test child 'pid' to crash on AMD
> +	 * systems (e.g. AMD Opteron(TM) Processor 6234) with
> +	 * older kernels. This causes tracee to stop and sleep
> +	 * in ptrace_stop(). Without resuming the tracee, the
> +	 * test hangs at do_test()->tst_reap_children() called
> +	 * by the library. Use detach here, so we don't need to
> +	 * worry about potential stops after this point.
> +	 */
> +	TEST(ptrace(PTRACE_DETACH, pid, 0, 0));
>  	if (TST_RET != 0)
> -		tst_brk(TBROK | TTERRNO, "PTRACE_CONT failed");
> +		tst_brk(TBROK | TTERRNO, "PTRACE_DETACH failed");
> +
> +	/* If child 'pid' crashes, only report it as info. */
> +	SAFE_WAITPID(pid, &status, 0);
> +	if (WIFEXITED(status)) {
> +		tst_res(TINFO, "test child %d exited, retcode: %d",
> +			pid, WEXITSTATUS(status));
> +	}
> +	if (WIFSIGNALED(status)) {
> +		tst_res(TINFO, "test child %d exited, termsig: %d",
> +			pid, WTERMSIG(status));
> +	}
>  
>  	okay = true;
> -	for (i = 0; i < num_cpus + 1; i++) {
> +	for (i = 0; i < num_cpus; i++) {
>  		SAFE_WAIT(&status);
>  		okay &= (WIFEXITED(status) && WEXITSTATUS(status) == 0);

Thanks Jan!

I think this change is correct.

Oleg.

[LTP] [PATCH] syscalls/ptrace07: handle potential SIGSEGV on older kernels

[Li Wang]

On Fri, Dec 21, 2018 at 9:51 PM Jan Stancek <jstancek@redhat.com> wrote:
>
> If a ptraced test process hits SIGSEGV, the entire testcase hangs.
>
> Older kernels such as RHEL7 (3.10.0), check the error code returned
> by restore_fpu_checking() and do drop_init_fpu() if it fails.
> So the FPU state of the prev task can't leak.
>
> But in the more likely case a task with xcomp_bv != 0 will be killed
> by SIGSEGV; either from do_device_not_available() or from
> sys_rt_sigreturn()->__restore_xstate_sig().
>
> And this is why the test case hangs; it wrongly assumes that the
> traced child can only exit and report nothing else. But since it
> receives SIGSEGV it reports this signal to the main process and
> sleeps in ptrace_stop(), it does not exit and thus the test-case
> hangs in tst_reap_children() after return from do_test().
>
> Replace PTRACE_CONT with PTRACE_DETACH, so we don't need to
> handle subsequent stops. And treat exit code from test process
> as info-only.

ptrace07(with this patch applied) passed with kernel 3.10 and 4.18
multi arches test.

>
> Debugged-by: Oleg Nesterov <onestero@redhat.com>
> Signed-off-by: Jan Stancek <jstancek@redhat.com>

Tested-by: Li Wang <liwang@redhat.com>

-- 
Regards,
Li Wang

[LTP] [PATCH] syscalls/ptrace07: handle potential SIGSEGV on older kernels

[Jan Stancek]

----- Original Message -----
> On Fri, Dec 21, 2018 at 9:51 PM Jan Stancek <jstancek@redhat.com> wrote:
> >
> > If a ptraced test process hits SIGSEGV, the entire testcase hangs.
> >
> > Older kernels such as RHEL7 (3.10.0), check the error code returned
> > by restore_fpu_checking() and do drop_init_fpu() if it fails.
> > So the FPU state of the prev task can't leak.
> >
> > But in the more likely case a task with xcomp_bv != 0 will be killed
> > by SIGSEGV; either from do_device_not_available() or from
> > sys_rt_sigreturn()->__restore_xstate_sig().
> >
> > And this is why the test case hangs; it wrongly assumes that the
> > traced child can only exit and report nothing else. But since it
> > receives SIGSEGV it reports this signal to the main process and
> > sleeps in ptrace_stop(), it does not exit and thus the test-case
> > hangs in tst_reap_children() after return from do_test().
> >
> > Replace PTRACE_CONT with PTRACE_DETACH, so we don't need to
> > handle subsequent stops. And treat exit code from test process
> > as info-only.
> 
> ptrace07(with this patch applied) passed with kernel 3.10 and 4.18
> multi arches test.
> 
> >
> > Debugged-by: Oleg Nesterov <onestero@redhat.com>
> > Signed-off-by: Jan Stancek <jstancek@redhat.com>
> 
> Tested-by: Li Wang <liwang@redhat.com>

Pushed.

Regards,
Jan