Hi David,

I have additional comments on this.

On Thu,  2 Jun 2016 23:26:19 -0400
David Long [off-list ref] wrote:

+/*
+ * The D-flag (Debug mask) is set (masked) upon deug exception entry.

deug -> debug

+ * Kprobes needs to clear (unmask) D-flag -ONLY- in case of recursive
+ * probe i.e. when probe hit from kprobe handler context upon
+ * executing the pre/post handlers. In this case we return with
+ * D-flag clear so that single-stepping can be carried-out.
+ *
+ * Leave D-flag set in all other cases.
+ */
+static void __kprobes
+spsr_set_debug_flag(struct pt_regs *regs, int mask)
+{
+	unsigned long spsr = regs->pstate;
+
+	if (mask)
+		spsr |= PSR_D_BIT;
+	else
+		spsr &= ~PSR_D_BIT;
+
+	regs->pstate = spsr;
+}
+

[..]

+
+int __kprobes kprobe_fault_handler(struct pt_regs *regs, unsigned int fsr)
+{
+	struct kprobe *cur = kprobe_running();
+	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+	switch (kcb->kprobe_status) {
+	case KPROBE_HIT_SS:
+	case KPROBE_REENTER:
+		/*
+		 * We are here because the instruction being single
+		 * stepped caused a page fault. We reset the current
+		 * kprobe and the ip points back to the probe address
+		 * and allow the page fault handler to continue as a
+		 * normal page fault.
+		 */
+		instruction_pointer(regs) = (unsigned long)cur->addr;
+		if (!instruction_pointer(regs))
+			BUG();

As according to the recent x86 kprobe bug on fault handler
discussion, here this also need kernel_disable_single_stap()
and spsr_set_debug_flag() in case of KPROBE_REENTER as you did
in kprobe_single_step_handler(). (Also, those code should be
a function for reuse)

+		if (kcb->kprobe_status == KPROBE_REENTER)
+			restore_previous_kprobe(kcb);
+		else
+			reset_current_kprobe();

Thanks!


-- 
Masami Hiramatsu [off-list ref]