On Thu, May 31, 2012 at 04:47:20PM +0200, Krystian Garbaciak wrote:

Adding people who've got some chips with paging, please keep them in the
CCs on this unless they complain (though since I'm cutting context...  :/ )

> +       /* Partition all accessible registers on address ranges,
> +          either to be accessed directly or indirectly. Arrange range
> +          list by ascending addresses. */

Wouldn't something naturally sorted like a rbtree be a more direct way
of doing this?

> +		range_cfg = NULL;
> +		for (n = 0, min_base = UINT_MAX; n < config->n_ranges; n++)
> +			if (range_base <= config->ranges[n].base_reg &&
> +			    config->ranges[n].base_reg <= min_base)
> +				range_cfg = &config->ranges[n];
> +

I've stared at this for a little while and I'm really not sure what it's
supposed to do.  The whole thing with min_base is just a bit odd, we're
doing comparisons against it but we never update it so why aren't we
using a constant, and in fact the comparison is always going to be true
since we're comparing against UINT_MAX.

I suspect it's supposed to pick the range with the lowest base but I'm
not convinced it does that.

> +		if (!range_cfg || range_cfg->base_reg > range_base) {
> +			/* Range of registers for direct access */
> +			range = kzalloc(sizeof(*range), GFP_KERNEL);
> +			if (range == NULL) {
> +				ret = -ENOMEM;
> +				goto err_range;
> +			}
> +			range->base_reg = range_base;
> +			if (range_cfg)
> +				range->max_reg = range_cfg->base_reg - 1;
> +			else
> +				range->max_reg = UINT_MAX;
> +			list_add_tail(&range->list, &map->range_list);
> +		}

This is making my head hurt too, possibly because of the lack of clarity
in the above.

> +static int _regmap_update_bits(struct regmap *map, unsigned int reg,
> +			       unsigned int mask, unsigned int val,
> +			       bool *change);

Put this up at the top of the file.

> +static int
> +_regmap_range_access(int (*regmap_bus_access)(struct regmap *map,
> +					      unsigned int reg,
> +					      void *val, unsigned int val_len),

eew, typedef this!

> +       unsigned int _page, _p;
> +       unsigned int _reg, _r;
> +       unsigned int _num;

These _s aren't helping legibility here.

> +	/* Bulk write should not cross single range boundaries */
> +	if (val_num != 0 &&
> +	    reg + val_num - 1 > range->max_reg)
> +		return -EINVAL;

When would val_num ever be zero?

> +			/* Update page register (may use caching) */
> +			ret = _regmap_update_bits(map, range->page_sel_reg,
> +						range->page_sel_mask,
> +						_page << range->page_sel_shift,
> +						&change);
> +			if (ret < 0)
> +				return ret;

Why the comment about the cache - why would this go wrong?

> +			/* There is no point to pass cache for data
> +			   registers, as they should be volatile anyway */
> +			ret = _regmap_range_access(regmap_bus_access,
> +						   map, _reg, _val, _num);
> +			if (ret < 0)
> +				return ret;

That comment needs some clarification too...

> +/**
> + * Configuration for indirect accessed register range.
> + * Indirect or paged registers, can be defined with one or more structures.

No , here.

> + * @translate_reg: Function should return indirect address/page number and
> + *                 register number (out of this range) matching virtual_reg.

Why does the user need to specify this?  Shouldn't we just specify a
size for the underlying window and then have a default which does the
obvious translations?  I'd imagine an *overwhelming* proportion of users
will want to do that.  Allowing an override is fine but requiring code
seems wrong for something like this.