Re: [PATCH] PPC64 Fix kprobes calling smp_processor_id when preemptible

[Ananth N Mavinakayanahalli <ananth@in.ibm.com>]

On Mon, Mar 14, 2005 at 03:53:21PM +1100, Paul Mackerras wrote:

Hi Paul,

> When booting with kprobes and preemption both enabled and
> CONFIG_DEBUG_PREEMPT=y, I get lots of warnings about smp_processor_id
> being called in preemptible code, from kprobe_exceptions_notify.  On
> ppc64, interrupts and preemption are not disabled in the handlers for
> most synchronous exceptions such as breakpoints and page faults
> (interrupts are disabled in the very early exception entry code but
> are reenabled before calling the C handler).
> 
> This patch adds a preempt_disable/enable pair to
> kprobe_exceptions_notify, and moves the preempt_disable() in
> kprobe_handler() to be done only in the case where we are about to
> single-step an instruction.  This eliminates the bug warnings.

The patch is fine, but it seems to break jprobes - we have an unbalanced
preempt_enable/disable path while handling jprobes. Patch below is
against 2.6.11-mm4 and fixes the issue.

Thanks,
Ananth


Signed-off-by: Ananth N Mavinakayanahalli <ananth@in.ibm.com>

diff -Naurp temp/linux-2.6.11/arch/ppc64/kernel/kprobes.c kprobes/linux-2.6.11/arch/ppc64/kernel/kprobes.c
--- temp/linux-2.6.11/arch/ppc64/kernel/kprobes.c	2005-03-17 05:15:53.000000000 +0530
+++ kprobes/linux-2.6.11/arch/ppc64/kernel/kprobes.c	2005-03-17 19:46:21.000000000 +0530
@@ -262,7 +262,6 @@ int setjmp_pre_handler(struct kprobe *p,
 
 void jprobe_return(void)
 {
-	preempt_enable_no_resched();
 	asm volatile("trap" ::: "memory");
 }