User Stack Trace

[grundym at us dot ibm dot com (grundy)]

Hi Folks -

Yeah, I'm on this mission again. I took the concept from the jprobe code
I previously posted (in which I just boosted a bunch of oprofile code
and made it work in the module). This is x86 specific (if you want to
test on x86_64 change the nregs->ebp to nregs->rdp), but I figure it's a
starting point. 

I imagine it would be nice to have a version that returns strings, and a
version that prints (like the kernel stack ones). But, before I go
fiddling with that, I'm trying to get the vma name and offset for each
address on the stack. That way you wouldn't have to capture
/proc/[pid]/maps for every process (for symbol resolution afterwards).

In this form, a good test is to fire it up in a text console, then log
in from gdm. Just as easy to switch syscall.ipc to syscall.exit for
testing.

Thanks
Mike

--
#!/usr/bin/env stap

/* Prints the user stack backtrace  
 * Concept and frame_head structure from Oprofile which is 
 * (c) 2002 OProfile authors, John Levon, David Smith
 */
function ustack_print:long ()
%{
        struct frame_head {
                struct frame_head * ebp;
                unsigned long ret;
        } __attribute__((packed)) stack_frame[2], *head;

        struct pt_regs *nregs = task_pt_regs(current);

        _stp_printf("[user] <%p>\n", nregs->eip);

        head = (struct frame_head *)nregs->ebp;

        if (user_mode_vm(nregs)) {
                do {
                        _stp_printf ("[user] <%p>\n", head->ret);
                        _stp_copy_from_user ((char *)stack_frame, 
                                (const char __user *) head->ebp,
                                sizeof(stack_frame));
                        head = stack_frame;
                        } while (head->ebp);
                _stp_printf ("[user] <%p>\n", head->ret);
        }
        THIS->__retvalue = 0;
%}

probe syscall.ipc {
        printf("%s:%d sys_ipc user backtrace: \n", execname(), pid());
        ustack_print(); 
}