Re: [patch 2/2] x86, fpu: lazy allocation of FPU area - v5

[Ingo Molnar <mingo@elte.hu>]

* Suresh Siddha <suresh.b.siddha@intel.com> wrote:

>  asmlinkage void math_state_restore(void)
>  {
>  	struct task_struct *me = current;
> -	clts();			/* Allow maths ops (or we recurse) */
>  
> -	if (!used_math())
> -		init_fpu(me);
> +	if (!used_math()) {
> +		local_irq_enable();
> +		/*
> +		 * does a slab alloc which can sleep
> +		 */
> +		if (init_fpu(me)) {
> +			/*
> +			 * ran out of memory!
> +			 */
> +			do_group_exit(SIGKILL);
> +			return;
> +		}
> +		local_irq_disable();
> +	}
> +
> +	clts();			/* Allow maths ops (or we recurse) */
>  	restore_fpu_checking(&me->thread.xstate->fxsave);
>  	task_thread_info(me)->status |= TS_USEDFPU;
>  	me->fpu_counter++;

hm, three things:

firstly, the clts is now done _after_ fpu_init() - are you sure that's 
OK? We do it in this order so that FINIT [on older cpus] does not fault.

secondly, while i know you were responding to review feedback from 
others, but the do_group_exit(SIGKILL) looks quite bad. It's totally 
undebuggable to the user - not even a coredump will be generated AFAICS 
- and the user has no idea that this all happened due to out-of-memory. 
A (forced) SIGBUS is our usual answer to out-of-memory situations. [such 
as when a pagetable allocation fails] If you get review feedback that 
suggests a crappy solution then please resist it! :-)

thirdly, the irq enable/disable worries me. Can it ever trigger in 
kernel code that has irqs off? If it happens when kernel uses the FPU in 
irqs-off sections (to do SSE optimized routines, etc.) then enabling 
irqs is dangerous - the original callsite had it disabled for a reason. 
At minimum we should add a debug check to math_state_restore(), 
something like:

  WARN_ON_ONCE(!(regs->flags & X86_EFLAGS_IF))

(this means we need to pass regs to math_state_restore())

	Ingo