Reading the VDSO area - i386

Hi Jeremy,

I was doing some tests that attempt to read the VDSO area of a task through either the /proc/pid/mem or ptrace(PTRACE_PEEKTEXT, ...) interfaces, and it seems that when the CONFIG_COMPAT_VDSO kernel parameter is enabled, we can no longer successfully read the VDSO area on i386 kernels.

I believe that debuggers such as gdb will attempt to sometimes walkback through the vsyscall area, and not being able to read the vsyscall/vdso area would thus cause debuggers problems.

So assuming that this change in behavior was not intentional, I've provided my stab (just an idea) at a fix. With this change below, the code in places such as get_user_pages() can now successfully call in_gate_area() and then subsequently call get_gate_vma(), which already properly returns the correct info.

Thanks for taking the time to read over this.

---
/userland/johnb/s/os/kernel/linux-2.6.22/arch/i386/kernel/sysenter.c

2007-07-17 08:38:48.000000000 -0400

+++ new/arch/i386/kernel/./sysenter.c 2007-07-17 11:48:28.000000000 -0400
@@ -336,6 +336,14 @@ struct vm_area_struct *get_gate_vma(stru

int in_gate_area(struct task_struct *task, unsigned long addr) {
+ struct mm_struct *mm = task->mm;
+
+ /* Check to see if this task was created in compat vdso mode
+ * and if the address is within the gate_vma area.
+ */
+ if (mm && mm->context.vdso == (void *)VDSO_HIGH_BASE &&
+ addr >= gate_vma.vm_start && addr <= gate_vma.vm_end)
+ return 1;
return 0;
}

Thanks. Someone else noticed this recently, and posted a near-identical patch. But I can't seem to find it right now...

Andrew, Andi: do you remember something like this passing by?

J

Yes it was from Jan Beulich.

-Andi

Ah, yes. For some reason I can't find it in my archives. Do you have it queued up?

J

Yes

-Andi