On Thu, Feb 22, 2024 at 10:38:54AM +0100, Uwe Kleine-König wrote:
> On Thu, Feb 22, 2024 at 05:32:10PM +1100, David Gibson wrote:
> > On Fri, Feb 16, 2024 at 07:19:20PM +0100, Uwe Kleine-König wrote:
> > > +static int overlay_update_node_conflicting_references(void *fdto, int tree_node,
> > > +						      int fixup_node,
> > > +						      uint32_t fdt_phandle,
> > > +						      uint32_t fdto_phandle)
> > > +{
> > > +	int fixup_prop;
> > > +	int fixup_child;
> > > +	int ret;
> > > +
> > > +	fdt_for_each_property_offset(fixup_prop, fdto, fixup_node) {
> > > +
> > > +		const fdt32_t *fixup_val;
> > > +                const char *tree_val;
> > > +                const char *name;
> > > +                int fixup_len;
> > > +                int tree_len;
> > > +                int i;
> > > +
> > > +                fixup_val = fdt_getprop_by_offset(fdto, fixup_prop,
> > > +                                                  &name, &fixup_len);
> > > +                if (!fixup_val)
> > > +                        return fixup_len;
> > > +
> > > +                if (fixup_len % sizeof(uint32_t))
> > > +                        return -FDT_ERR_BADOVERLAY;
> > > +                fixup_len /= sizeof(uint32_t);
> > > +
> > > +                tree_val = fdt_getprop(fdto, tree_node, name, &tree_len);
> > > +                if (!tree_val) {
> > > +                        if (tree_len == -FDT_ERR_NOTFOUND)
> > > +                                return -FDT_ERR_BADOVERLAY;
> > > +
> > > +                        return tree_len;
> > > +                }
> > > +
> > > +		for (i = 0; i < fixup_len; i++) {
> > > +			fdt32_t adj_val;
> > > +			uint32_t poffset;
> > > +
> > > +			poffset = fdt32_to_cpu(fixup_val[i]);
> > 
> > You can use fdt32_ld_() here to slightly simplify this
> 
> I don't see what you suggest here. fixup_val is assigned to in
> fdt_getprop_by_offset() which I need to get fixup_len and then fixup_val
> is already around. If you still think this could be improved, please be
> more explicit.

	poffset = fdt32_ld_(fixup_val + i);

Thus combining the memory read and the byteswap into one operation.  I
guess it doesn't really simplify the code, but correctly handling
reads and writes to memory within the dtb is what the fdt*_ld() and
fdt*_st() helpers are for.

> > > +
> > > +			/*
> > > +			 * phandles to fixup can be unaligned.
> > > +			 *
> > > +			 * Use a memcpy for the architectures that do
> > > +			 * not support unaligned accesses.
> > > +			 */
> > > +			memcpy(&adj_val, tree_val + poffset, sizeof(adj_val));
> > 
> > And here you can use fdt32_ld() which has unaligned access handling
> > built in.

Specifically
	adj_val = fdt32_ld(tree_val + poffset)

And then replacing fdt32_to_cpu(adj_val) with just adj_val below.

> > 
> > > +
> > > +			if (fdt32_to_cpu(adj_val) == fdto_phandle) {
> > > +
> > > +				adj_val = cpu_to_fdt32(fdt_phandle);
> > > +
> > > +				ret = fdt_setprop_inplace_namelen_partial(fdto,
> > > +									  tree_node,
> > > +									  name,
> > > +									  strlen(name),
> > > +									  poffset,
> > > +									  &adj_val,
> > > +									  sizeof(adj_val));
> > > +				if (ret == -FDT_ERR_NOSPACE)
> > > +					return -FDT_ERR_BADOVERLAY;
> > > +
> > > +				if (ret)
> > > +					return ret;
> > 
> > And here you can use fdt32_st() on (tree_val + poffset), avoiding
> > quite some complexity.
> 
> There is some complexity because tree_val is a const char * (as this is
> what fdt_getprop() returns) and fdt32_st() obviously doens't take a
> const pointer.

Ah, right.  You can use fdt_getprop_w() to get a writable pointer
instead.  If there was a way to have fdt_getprop() return a const
pointer only if the fdt pointer it was given was const, I'd do that.

> If you want to play around with that, the code is mostly a copy of
> overlay_update_local_node_references() which could benefit from your
> suggestions in the same way.

That's quite plausible.

> Can I lure you in improving overlay_update_local_node_references()? Then

Fair point..

https://github.com/dgibson/dtc/commit/24f60011fd43683d8e3916435c4c726e9baac9c9

> I could copy from the improved function for my patch. (Or maybe refactor
> the function to take a function as parameter which is

Eh... I'd prefer to avoid higher order functions in something this low
level.

> 	uint32_t delta;
> 	uint32_t increase_by_delta(uint32_t phandle)
> 	{
> 		return phandle + delta;
> 	}
> 
> for the overlay_update_local_node_references() case and
> 
> 	uint32_t fdt_phandle;
> 	uint32_t fdto_phandle;
> 	uint32_t resolve_fdt_conflict(uint32_t phandle)
> 	{
> 		if (phandle == fdto_phandle)
> 			return fdt_phandle;
> 		else
> 			return phandle;
> 	}
> 
> for my function (maybe passing the data stored in global vars here in a
> context parameter).
> 
> Best regards
> Uwe
> 



-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson