RE: [PATCH] Kprobes- robust fault handling for i386

["Mao, Bibo" <bibo dot mao at intel dot com>]

I have one question and I reply between the lines.

>-----Original Message-----
>From: systemtap-owner@sourceware.org [mailto:systemtap-owner@sourceware.org]
>On Behalf Of Prasanna S Panchamukhi
>Sent: 2006年2月22日 15:13
>To: systemtap@sources.redhat.com
>Subject: [PATCH] Kprobes- robust fault handling for i386
>
>Hi,
>
>Below is the prototype for robust fault handling, as of now
>this patch is for i386 architecture and should be easily
>ported to other architectures. Your comments and suggestions
>are welcome. This patch has been tested for page faults that
>occur while accessing user address space data. Support needs
>to be added for cases such as divide by zero, NULL pointer
>dereference, etc. Also as of now we increment the nmissed
>count, instead we can track such instances by having
>independent counters such as nprefault, npostfault.
>
>Thanks
>Prasanna

>@@ -509,9 +554,21 @@ static inline int post_kprobe_handler(st
> 	if (!cur)
> 		return 0;
>
>-	if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
>+	if ((kcb->kprobe_status != KPROBE_REENTER)
>+			&& (kcb->kprobe_status != KPROBE_HIT_FAULT)
>+			&& cur->post_handler) {
>+		kcb->handler_regs = regs;
> 		kcb->kprobe_status = KPROBE_HIT_SSDONE;
>-		cur->post_handler(cur, regs, 0);
>+		kprobe_post_handler_trampoline(cur, regs, kcb);
>+		kcb = get_kprobe_ctlblk();
>+		/*
>+		 * Check if user defined handler caused the page fault, in
>+		 * such a case restore the register pointers, just resets
>+		 * the current kprobe and resumes the execution, since we
>+		 * have already single stepped on original instruction.
>+		 */
>+		if (kcb->kprobe_status == KPROBE_HIT_FAULT)
>+			regs = kcb->handler_regs;
> 	}
>
> 	resume_execution(cur, regs, kcb);
>@@ -541,18 +598,55 @@ static inline int kprobe_fault_handler(s
> {
> 	struct kprobe *cur = kprobe_running();
> 	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
>+	unsigned long *sara = (unsigned long *)&regs->esp;
What is &regs->esp meaning here? If instruction which causes page fault is not first instruction of called function, then &regs->esp will be local variable's memory address in the called function, but not caller return address.
>........
>+		*sara = kcb->handler_retaddr;
So in this line maybe sometimes it will only change callee function local variant's value, but not change caller return value.

Regards
Bibo,mao