Re: [PATCH 1/1] x86/fpu: math_state_restore() should not blindly disable irqs

[Ingo Molnar <mingo@kernel.org>]

* Oleg Nesterov <oleg@redhat.com> wrote:

> math_state_restore() assumes it is called with irqs disabled, but
> this is not true if the caller is __restore_xstate_sig().
> 
> This means that if ia32_fxstate == T and __copy_from_user() fails
> __restore_xstate_sig() returns with irqs disabled too. This trgiggers
> 
> 	BUG: sleeping function called from invalid context at kernel/locking/rwsem.c:41
> 	[<ffffffff81381499>] dump_stack+0x59/0xa0
> 	[<ffffffff8106bd05>] ___might_sleep+0x105/0x110
> 	[<ffffffff8138786d>] ? _raw_spin_unlock_irqrestore+0x3d/0x70
> 	[<ffffffff8106bd8d>] __might_sleep+0x7d/0xb0
> 	[<ffffffff81385426>] down_read+0x26/0xa0
> 	[<ffffffff8138788a>] ? _raw_spin_unlock_irqrestore+0x5a/0x70
> 	[<ffffffff81136038>] print_vma_addr+0x58/0x130
> 	[<ffffffff8100239e>] signal_fault+0xbe/0xf0
> 	[<ffffffff810419aa>] sys32_rt_sigreturn+0xba/0xd0
> 
> Change math_state_restore() to check irqs_disabled().
> 
> Note: this is the minimal fix for -stable, it is horrible but simple.
> We need to rewrite math_state_restore(), init_fpu(), and cleanup their
> users.
> 
> Signed-off-by: Oleg Nesterov <oleg@redhat.com>
> Cc: <stable@vger.kernel.org>
> ---
>  arch/x86/kernel/traps.c |    9 +++++++--
>  1 files changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/kernel/traps.c b/arch/x86/kernel/traps.c
> index 51c4658..7310e0e 100644
> --- a/arch/x86/kernel/traps.c
> +++ b/arch/x86/kernel/traps.c
> @@ -774,7 +774,10 @@ void math_state_restore(void)
>  	struct task_struct *tsk = current;
>  
>  	if (!tsk_used_math(tsk)) {
> -		local_irq_enable();
> +		bool disabled = irqs_disabled();
> +
> +		if (disabled)
> +			local_irq_enable();
>  		/*
>  		 * does a slab alloc which can sleep
>  		 */
> @@ -785,7 +788,9 @@ void math_state_restore(void)
>  			do_group_exit(SIGKILL);
>  			return;
>  		}
> -		local_irq_disable();
> +
> +		if (disabled)
> +			local_irq_disable();
>  	}

Yuck!

Is there a fundamental reason why we cannot simply enable irqs and 
leave them enabled? Math state restore is not atomic and cannot really 
be atomic.

[ A potential worry would be kernel code using vector instructions in
  irqs-off regions - but that's totally broken anyway so not a big
  worry IMO, we might even want to warn about it. ]

But I might be missing something?

Thanks,

	Ingo