Re: [PATCH 1/2] fdt: Allow stacked overlays phandle references

On 07/13/17 14:22, Phil Elwell wrote:

On 13/07/2017 21:07, Frank Rowand wrote:

quoted

On 07/13/17 12:38, Phil Elwell wrote:

(I moved Phil's reply to after the email he replied to.)

Thanks.

quoted

quoted

On 13 Jul 2017 8:32 pm, "Frank Rowand" [off-list ref] wrote:

quoted

On 07/03/17 02:06, David Gibson wrote:

quoted

On Wed, Jun 14, 2017 at 05:52:25PM +0300, Pantelis Antoniou wrote:

quoted

This patch enables an overlay to refer to a previous overlay's
labels by performing a merge of symbol information at application
time.

This seems to be doing things the hard way.

You're essentially extending the semantics of overlay application to
add the symbol merging.  You've implemented these extended semantics
in libfdt, which is all very well, but that's not the only overlay
application implementation.


It seems to me a better approach would be to change dtc's -@
implementation, so that in /plugin/ mode instead of making a global
__symbols__ node, it puts it into the individual fragments.  That way
the existing overlay application semantics will update the __symbols__
node.

If the __symbols__ node was inside a fragment, then the existing
code would add (or update) a __symbols__ node located at the location
pointed to by the fragment's target path, instead of updating the
node /__symbols__.

It makes sense to me to have only one global __symbols__ node instead
of several.

If there is a global __symbols__ node then we have a single name
space for symbols.

If there are multiple __symbols__ nodes spread throughout the tree,
then to me that would imply different name spaces spread throughout
the tree, where namespaces are determined by fragments.  This sounds
confusing to me.  Or if the intent is to have a single name space
then the __symbols__ information would be scattered throughout the
tree instead of located in a single node.

My current patch (under review), targeted for Linux 4.13-rc1, puts
an overlay's __symbols__ node properties into the overlay's
changeset, so they get added when the overlay is loaded and
removed when the overlay is unloaded.

quoted

Can we also consider a mechanism for overlay-local symbols, i.e. symbols
that are used purely to create links within an overlay - perhaps using a
particular naming convention? This would make it easier to instantiate an
overlay multiple times without having to uniquify all symbols, and it would
avoid polluting the global namespace without reason.

Phil

That is essentially the result you get if you compile the overlay dts
without '-@'.  There will be no __symbols__ node created even if there
are symbols within the overlay.

But (unless something has changed recently) the '-@' switch controls both
symbol and fixup generation, i.e. export and import of symbols. Unless one
religiously uses 'target-path' to place fragments (thus removing the
level of abstraction provided by symbols) overlays are useless without
the ability to reference external symbols, but in my experience very few
overlays need to add symbols to the global symbol table.

For the dtc compiler in Linux 4.11, the '-@' switch is only needed
to generate the __symbols__ node.  The __fixups__ and __local-fixups__
nodes are generated whether '-@' is specified or not.

The __fixups__ and __local_fixups__ are generated when '/plugin/;'
is specified in the source file.

quoted

This is important if the overlay is for an add-on board which might
have several instances plugged into different sockets on the base
system.

But Phil does bring up an interesting use case.  If the add-on board
("level one add-on") in turn has a socket that an additional board
("level two add-on") can be plugged into, then the level two add-on
overlay might have a need to reference a symbol from the overlay
for the level one add-on.  And since there may be multiple level one
add-on cards in the system, the overlay for each of the level one
add-ons would need to export its symbols in a name space only
visible to the level two add-on plugged into that specific level
on add-on.

That use case adds a further level of complexity. Should it come to it, I
hope an inability to solve the problem posed by this advanced usage won't
prevent a solution to a simpler problem from being accepted.

--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html