Re: [PATCH RFC] BKL not necessary in cpuid_open

[John Kacur <jkacur@redhat.com>]

On Wed, 7 Oct 2009, Sven-Thorsten Dietrich wrote:

> On Wed, 2009-10-07 at 20:19 +0200, John Kacur wrote:
> > I've been staring at the BKL lock in cpuid_open, and I can't see what it 
> > is protecting. However, I may have missed something - even something 
> > obvious, so comments are welcome.
> 
> I have been using this patch to first see if the BKL is being used
> simply as mutex, which would allow easier break-down.
> 
> Sven
> 
> Subject: Introduce the BKL-MUTEX, to allow proving the code that
> From: Sven-Thorsten Dietrich sdietrich@suse.de Sat Oct 3 01:26:01 2009 -0700
> Date: Sat Oct 3 01:26:01 2009 -0700:
> Git: 4bfea26de550100d193c49278d657485e792dfe5
> 
> just needs a global kernel mutex, without all the funky
> preemption bits...
> 
> Acked-by: Sven-Thorsten Dietrich <sdietrich@suse.de>
> diff --git a/include/linux/smp_mutex.h b/include/linux/smp_mutex.h
> new file mode 100644
> index 0000000..6c45a96
> --- /dev/null
> +++ b/include/linux/smp_mutex.h
> @@ -0,0 +1,22 @@
> +#ifndef __LINUX_SMPMUTEX_H
> +#define __LINUX_SMPMUTEX_H
> +
> +#ifdef CONFIG_LOCK_KERNEL
> +
> +extern void lock_kernel_mutex(void);
> +extern void unlock_kernel_mutex(void);
> +
> +static inline void cycle_kernel_mutex(void)
> +{
> +	lock_kernel_mutex();
> +	unlock_kernel_mutex();
> +}
> +
> +#else
> +
> +#define lock_kernel_mutex()				do { } while(0)
> +#define unlock_kernel_mutex()				do { } while(0)
> +#define cycle_kernel_mutex()			do { } while(0)
> +
> +#endif /* CONFIG_LOCK_KERNEL */
> +#endif /* __LINUX_SMPMUTEX_H */
> diff --git a/lib/Makefile b/lib/Makefile
> index 2e78277..c71ffdc 100644
> --- a/lib/Makefile
> +++ b/lib/Makefile
> @@ -40,7 +40,7 @@ lib-$(CONFIG_GENERIC_FIND_FIRST_BIT) += find_next_bit.o
>  lib-$(CONFIG_GENERIC_FIND_NEXT_BIT) += find_next_bit.o
>  obj-$(CONFIG_GENERIC_FIND_LAST_BIT) += find_last_bit.o
>  obj-$(CONFIG_GENERIC_HWEIGHT) += hweight.o
> -obj-$(CONFIG_LOCK_KERNEL) += kernel_lock.o
> +obj-$(CONFIG_LOCK_KERNEL) += kernel_lock.o kernel_mutex.o
>  obj-$(CONFIG_DEBUG_PREEMPT) += smp_processor_id.o
>  obj-$(CONFIG_DEBUG_LIST) += list_debug.o
>  obj-$(CONFIG_DEBUG_OBJECTS) += debugobjects.o
> diff --git a/lib/kernel_mutex.c b/lib/kernel_mutex.c
> new file mode 100644
> index 0000000..1d587a8
> --- /dev/null
> +++ b/lib/kernel_mutex.c
> @@ -0,0 +1,37 @@
> +/*
> + * lib/kernel_mutex.c
> + *
> + * This is the preemptible BKL - To be used transitionally to
> + * prove those subsystems still using lock_kernel, but really
> + * requiring a global mutex.
> + */
> +#include <linux/smp_mutex.h>
> +#include <linux/module.h>
> +#include <linux/kallsyms.h>
> +#include <linux/semaphore.h>
> +#include <linux/mutex.h>
> +
> +/*
> + * The 'big kernel mutex'
> + *
> + * Don't use in new code.
> + */
> +static  __cacheline_aligned_in_smp DEFINE_MUTEX(kernel_mutex);
> +
> +
> +/*
> + * Getting the kernel mutex.
> + */
> +void __lockfunc lock_kernel_mutex(void)
> +{
> +	mutex_lock(&kernel_mutex);
> +}
> +
> +void __lockfunc unlock_kernel_mutex(void)
> +{
> +	mutex_unlock(&kernel_mutex);
> +}
> +
> +EXPORT_SYMBOL(lock_kernel_mutex);
> +EXPORT_SYMBOL(unlock_kernel_mutex);
> +
> 

Cool! Seems like an excellent experiment. However this is a separate patch
from the one initially proposed in this thread. I'm willing to risk just 
removing it in this case without any intermediary step. However, if anyone 
points out to me why I'm a knuckle head and missed something obvious - I'll 
listen. Otherwise, let's use your patch as a separate tactic to kill BKL.