[Crash-utility] [PATCH] do not check sp if ip points to user space

[Dave Anderson]

----- Original Message -----
> If the task is a user program, the sp can be points to anywhere,
> because we can modify sp in assembly.
> For example:
> 
> .globl main
>         .type   main, @function
> main:
> 
>         finit
>         subq $16, (%rsp)
>         movq $0, (%rsp)
> .loop:
>         jmp .loop
> 
> 

Why would any user task do that?  

And what happens when a backtrace is attempted on such a task?

Since the current code would not set BT_USER_SPACE, I'm guessing that it
would run into this (at least on x86_64):

        if (!(bt->flags & BT_USER_SPACE) && (!rsp || !accessible(rsp))) {
                error(INFO, "cannot determine starting stack pointer\n");
                return;
        }

I do believe that I put the additional in_user_stack() checks in those
locations for a reason.  Consider a task running in kernel mode that 
corrupts its return address stack location with a non-kernel address, 
or called a function indirectly that had a NULL pointer in it.  That
would cause a kernel crash with a non-kernel RIP in its exception frame,
and your patch would mistake it for user-space.

In any case, you're going to have to come up with a more compelling
reason to change all of these locations.  (And for that matter, I wonder
why you didn't patch Fujitsu's get_sadump_regs() the same way?)

Dave