Re: Fix traceback seen when resuming after suspend-to-ram

[Frederic Weisbecker <fweisbec@gmail.com>]

On Wed, Jul 22, 2009 at 08:41:23PM +0530, K.Prasad wrote:
> Hi Ingo,
> 	Please accept a patch that fixes the warning message and traceback
> seen upon a resume after suspend-to-ram.
> 
> The warning message is emitted because
> spin_unlock_bh()<--load_debug_registers() is invoked with interrupts
> disabled upon 'resume' unlike when invoked at boot-time, and was
> reported on LKML here: http://lkml.org/lkml/2009/7/18/145.
> 
> This patch modifies load_debug_registers() to block all interrupts and not
> just bottom-halves (bottom-halves were blocked in load_debug_registers()
> prevent flush_thread_hw_breakpoint() from interfering when invoked from
> SoftIRQ context).
> 
> Fix traceback seen when resuming after suspend-to-ram
> 
> This patch fixes a traceback when resuming after a suspend-to-ram operation.
> The traceback warns about entering slowpatch due to a spin_unlock_bh() done
> from interrupt context.
> 
> Signed-off-by: K.Prasad <prasad@linux.vnet.ibm.com>
> ---
>  kernel/hw_breakpoint.c |    8 +++-----
>  1 file changed, 3 insertions(+), 5 deletions(-)
> 
> Index: linux-2.6-tip/kernel/hw_breakpoint.c
> ===================================================================
> --- linux-2.6-tip.orig/kernel/hw_breakpoint.c
> +++ linux-2.6-tip/kernel/hw_breakpoint.c
> @@ -80,17 +80,15 @@ void load_debug_registers(void)
>  	unsigned long flags;
>  	struct task_struct *tsk = current;
>  
> -	spin_lock_bh(&hw_breakpoint_lock);
> -
>  	/* Prevent IPIs for new kernel breakpoint updates */
> -	local_irq_save(flags);
> +	spin_lock_irqsave(&hw_breakpoint_lock, flags);
> +
>  	arch_update_kernel_hw_breakpoint(NULL);
> -	local_irq_restore(flags);
>  
>  	if (test_tsk_thread_flag(tsk, TIF_DEBUG))
>  		arch_install_thread_hw_breakpoint(tsk);
>  
> -	spin_unlock_bh(&hw_breakpoint_lock);
> +	spin_unlock_irqrestore(&hw_breakpoint_lock, flags);
>  }
>  
>  /*


Hmm, this may lead to a state in which lockdep would complain because
your lock is taken as softirq-safe from flush_thread_hw_breakpoint()
and hardirq-safe in load_debug_registers().

Lockdep will think that you have an unsafe state in
flush_thread_hw_breakpoint() because your lock has been taken in
a hardirq-safe fashion elsewhere and therefore can be taken in a
hardirq path.

We know it's safe, but lockdep will warn anyway.

BTW: how is it possible that flush_thread_hw_breakpoint() can be called
from softirq? It can called in a failed fork or any case when a thread
is released. Does such thing sometimes happen in softirq?

Thanks.