Re: current not very current (vs curr_vcpu)

[Mukesh Rathor <mukesh.rathor@oracle.com>]

On Sat, 20 Feb 2010 07:45:26 +0000
Keir Fraser <keir.fraser@eu.citrix.com> wrote:

> On 20/02/2010 03:50, "Mukesh Rathor" <mukesh.rathor@oracle.com> wrote:
> 
> > ah, I see what's going on. context_switch() is scheduling idle
> > vcpu, and calls continue_idle_domain() to
> > reset_stack_and_jump(idle_loop). well, reset_stack_and_jump() is
> > setting rsp to guest_cpu_user_regs(), and interrupt is coming right
> > at that instant. so:
> > 
> > diff = (char *)guest_cpu_user_regs() - (char *)(r)  is 0
> > 
> > and as a result, guest_mode(regs) == true.
> 
> Well, I don't see how this scenario works. If rsp==g_c_u_r() at the
> instant the interrupt comes in, then the stack frame for the
> interrupt will be *above* g_c_u_r(). Thus 'diff' in guest_mode() will
> evaluate non-zero and positive, and regs->{rip,rsp} should point at
> hypervisor code/stack.
> 
> Also: in your original email you said regs.rsp pointed at dom0 stack.
> That doesn't tally with you saying that rsp==g_c_u_r() (an address in
> hypervisor space) immediately before the interrupt, in this email.
> Regs->rsp in the scenario you describe here should be exactly equal
> to g_c_u_r().
> 
>  -- Keir
> 

Ok, I think I found it.  Initially, my printk in serial_rx() showed 
  regs == ffff82c48030ff28 == guest_cpu_user_regs

This led me down to reset_stack_and_jump where sp is set to g_c_u_r.

Anyways, on this big box, I'm using virtual serial via the service 
processor. So, it looks like serial interrupts are not going thru
do_IRQ(), but ns16550_poll(). 

__do_softirq -> execute_timer -> ns16550_poll -> serial_rx_interrupt.

However, in ns16550_poll():

struct cpu_user_regs *regs = guest_cpu_user_regs(); <------

The cpu is clearly running idle_vcpu, so current is correctly pointing
to idle vcpu. But guest_mode() is showing guest mode incorrectly.

I'm not much familiar with ns16550 stuff, so cant' think of a fix other
than just setting regs to current stack pointer in ns16550_poll().

__asm__ ( "movq %%rsp,%0" : "=r" (val));
struct cpu_user_regs *regs = val;


Let me know if you like the fix and I'll submit a patch.

thanks,
Mukesh