preemptive kernel: ptrace fix

I noticed (actually, I had a thought ptrace wouldnt work under preemption) that strace sometimes locks under the preemptible kernel, although is killable by CTRL-C.

The following is from arch/i386/kernel/ptrace.c :: syscall_trace:

1 current->state = TASK_STOPPED;
2 notify_parent(current, SIGCHLD);
3 schedule();

What if, between line one and two, preemption occurs? Now the task state is TASK_STOPPED, and thus will never be rescheduled. I presume this is the problem. The attached patch fixes this.

Running multiple straces in a tight loop for over an hour shows me the problem is fixed. Its obvious the above is a problem anyhow.

However, doing an `strace strace whatever' (ie, stracing strace), it still enters a stopped state about 20% of the time (before the patch, it locked almost 100%). I can't figure out why. Comments?

Yes, stracing strace is contorted and I don't care, but something else is obviously wrong -- you can do so on a non-preemption machine.

diff -urN linux-2.4.9-ac14-preempt/arch/i386/kernel/ptrace.c linux/arch/i386/kernel/ptrace.c > patch-rml-2.4.9-ac14-preempt-strace-fix-1
--- linux-2.4.9-ac14-preempt/arch/i386/kernel/ptrace.c Sat Sep 22 23:20:41 2001+++ linux/arch/i386/kernel/ptrace.c Sat Sep 22 23:42:11 2001
@@ -455,9 +455,11 @@
between a syscall stop and SIGTRAP delivery */
current->exit_code = SIGTRAP | ((current->ptrace & PT_TRACESYSGOOD)
? 0x80 : 0);
+ ctx_sw_off();
current->state = TASK_STOPPED;
notify_parent(current, SIGCHLD);
schedule();
+ ctx_sw_on();
/*
* this isn't the same as continuing with a signal, but it will do
* for normal use. strace only continues with a signal if the

Robert M. Love
rml at ufl.edu
rml at tech9.net

I fixed it. I am not overly sure why it fixes it, but the following patch (on top of the previous) fixes all the ptrace problems I can find.

In auditing the code, I fixed another problem which I will explain...

I'll put out a final patch in a bit...

diff -urN linux-2.4.9-ac14-preempt/arch/i386/kernel/signal.c linux/arch/i386/kernel/signal.c
@@ -611,9 +611,11 @@
 		if ((current->ptrace & PT_PTRACED) && signr != SIGKILL) {
 			/* Let the debugger run.  */
 			current->exit_code = signr;
+			ctx_sw_off();
 			current->state = TASK_STOPPED;
 			notify_parent(current, SIGCHLD);
 			schedule();
+			ctx_sw_on();
 
 			/* We're back.  Did the debugger cancel the sig?  */
 			if (!(signr = current->exit_code))

@@ -667,11 +669,13 @@
 				/* FALLTHRU */
 
 			case SIGSTOP:
+				ctx_sw_off();
 				current->state = TASK_STOPPED;
 				current->exit_code = signr;
 				if (!(current->p_pptr->sig->action[SIGCHLD-1].sa.sa_flags & SA_NOCLDSTOP))
 					notify_parent(current, SIGCHLD);
 				schedule();
+				ctx_sw_on();
 				continue;
 
 			case SIGQUIT: case SIGILL: case SIGTRAP:

Robert M. Love
rml at ufl.edu
rml at tech9.net