Re: [PATCH] blockgroup_lock: hashed spinlocks for ext2 and ext3

Re: [PATCH] blockgroup_lock: hashed spinlocks for ext2 and ext3

[Andi Kleen]

On Sun, 2003-04-13 at 12:00, Linux Kernel Mailing List wrote:

[....]
> 	
> 	These locks are actually fairly generic.  However I presented it as something
> 	which is specific to ext2 and ext3 so that people wouldn't go using them all
> 	over the place.  They consume a lot of storage.
> 

> +
> +struct bgl_lock {
> +	spinlock_t lock;
> +} ____cacheline_aligned_in_smp;
> +
> +struct blockgroup_lock {
> +	struct bgl_lock locks[NR_BG_LOCKS];
> +};

Why don't you use per_cpu data for this ? It can be indexed as well
with per_cpu() and it would safe a lot of space because the padding
would not be all wasted. If you want more than NR_CPUS locks it could be
done using a simple two level index scheme.

-Andi

Re: [PATCH] blockgroup_lock: hashed spinlocks for ext2 and ext3

[Andrew Morton]

Andi Kleen <ak@suse.de> wrote:
>
> > +
> > +struct bgl_lock {
> > +	spinlock_t lock;
> > +} ____cacheline_aligned_in_smp;
> > +
> > +struct blockgroup_lock {
> > +	struct bgl_lock locks[NR_BG_LOCKS];
> > +};
> 
> Why don't you use per_cpu data for this ? It can be indexed as well
> with per_cpu() and it would safe a lot of space because the padding
> would not be all wasted. If you want more than NR_CPUS locks it could be
> done using a simple two level index scheme.

per_cpu data is statically allocated, at compile-time.  And it doesn't work
for modules.

These hashed locks need to be dynamically allocated (one per filesystem) and
they need to work from modules.

And this hashed lock is not a per-cpu thing.  (No locks are!) It just uses
NR_CPUS to decide how big the hash should be.

Re: [PATCH] blockgroup_lock: hashed spinlocks for ext2 and ext3

[Andi Kleen]

> per_cpu data is statically allocated, at compile-time.  And it doesn't work
> for modules.
> 
> These hashed locks need to be dynamically allocated (one per filesystem) and
> they need to work from modules.

Ok.

> 
> And this hashed lock is not a per-cpu thing.  (No locks are!) It just uses
> NR_CPUS to decide how big the hash should be.

Isn't that what the IBM kmalloc_per_cpu() was for ? (I think the patch
was from Dipankar) 

-Andi

Re: [PATCH] blockgroup_lock: hashed spinlocks for ext2 and ext3

[Andrew Morton]

Andi Kleen <ak@suse.de> wrote:
>
> > 
> > And this hashed lock is not a per-cpu thing.  (No locks are!) It just uses
> > NR_CPUS to decide how big the hash should be.
> 
> Isn't that what the IBM kmalloc_per_cpu() was for ? (I think the patch
> was from Dipankar) 

This is not a per-cpu data structure.

It is a hashed spinlock which uses NR_CPUS as a sizing hint.