Re: [PATCH -mm 1/6] slab: introduce __kfree_rcu

[Ingo Molnar <mingo@elte.hu>]

* Christoph Lameter <cl@linux-foundation.org> wrote:

> On Mon, 23 Mar 2009, Ingo Molnar wrote:
> 
> >
> > * Christoph Lameter <cl@linux-foundation.org> wrote:
> >
> > > On Mon, 23 Mar 2009, Pekka Enberg wrote:
> > >
> > > > > +static inline void *portion_to_obj(void *portion)
> > > > > +{
> > > > > +	struct page *page = virt_to_head_page(portion);
> > > > > +	struct slab *slab = page_get_slab(page);
> > > > > +	struct kmem_cache *cache = page_get_cache(page);
> > > > > +	unsigned int offset = portion - slab->s_mem;
> > > > > +	unsigned int index = offset / cache->buffer_size;
> > > > > +
> > > > > +	return index_to_obj(cache, slab, index);
> > > > > +}
> > > >
> > > > A minor nit: I think this would be more readable if you separated
> > > > variable declarations from the initializations. Also, you can probably
> > > > drop the inline from the function declaration and let GCC decide what to
> > > > do.
> > >
> > > Thats debatable. I find the setting up a number of variables that
> > > are all dependend in the above manner very readable. They are
> > > usually repetitive. Multiple functions use similar
> > > initializations.
> >
> > I agree with Pekka, it's clearly more readable when separated out
> > nicely:
> >
> > 	struct kmem_cache *cache;
> > 	unsigned int offset;
> > 	unsigned int index;
> > 	struct page *page;
> > 	struct slab *slab;
> >
> > 	page	= virt_to_head_page(portion);
> > 	slab	= page_get_slab(page);
> > 	cache	= page_get_cache(page);
> >
> > 	offset	= portion - slab->s_mem;
> > 	index	= offset / cache->buffer_size;
> >
> > The original form is hard to read due to lack of structure.
> 
> Structure can also be established differently:
> 
> static inline void *portion_to_obj(void *portion)
> {
> 	struct page *page = virt_to_head_page(portion);
> 	struct slab *slab = page_get_slab(page);
> 	struct kmem_cache *cache = page_get_cache(page);
> 
> 	unsigned int offset = portion - slab->s_mem;
> 	unsigned int index = offset / cache->buffer_size;
> 
> 	return index_to_obj(cache, slab, index);

It's still not as readable to me as the version i posted above and 
confusing as well, due to the newline in the middle of local 
variable definitions.

> It would be good if the whole series of actions that need to be 
> taken in order for the function to "get to know" the slab the 
> object parms would be simpler. Like its done in ruby
> 
> 	(page, slab, cache) = get_slab_info(portion)
> 
> 	(offset, index) = get_position_info(slab, portion)
> 
> But how can this be done in C without weird pointer passing?

The version i posted is pretty compact visually. The actual type 
enumeration is repetitive and it's often a meaningless pattern.

What matters is this sequence of symbols:

> > 	page	= virt_to_head_page(portion);
> > 	slab	= page_get_slab(page);
> > 	cache	= page_get_cache(page);
> >
> > 	offset	= portion - slab->s_mem;
> > 	index	= offset / cache->buffer_size;

... and anyone versed in slab internals will know the type of these 
variables without having to look them up. (using variable names 
consistently through a full subsystem is important for this reason)
 
Pairing them up with their base types just obscures the real logic.

That is one reason why i generally use the 'reverse christmas tree' 
type of local variable definition blocks:

> > 	struct kmem_cache *cache;
> > 	unsigned int offset;
> > 	unsigned int index;
> > 	struct page *page;
> > 	struct slab *slab;

As the trained eye will just want to skip over this as irrelevant 
fluff and the shape makes this the easiest (the less complex a shape 
is geometrically, the less 'eye skipping overhead' there is).

Anyway, these are nuances and if you go strictly by what's minimally 
required by Documentation/CodingStyle you can stop a lot sooner than 
having to bother about such fine details. The original version was 
certainly acceptable - it's just that IMO Pekka was right that it 
can be done better.

	Ingo