[PATCH v2 10/25] binman: Refactor fit to generate output at the end

[Alper Nebi Yasak]

On 24/02/2022 02:00, Simon Glass wrote:
> At present the fit implementation creates the output tree while
> scanning the FIT description. Then it updates the tree later when the
> data is known.
> 
> This works, but is a bit confusing, since it requires mixing the scanning
> code with the generation code, with a fix-up step at the end.
> 
> It is actually possible to do this in two phases, one to scan everything
> and the other to generate the FIT. Thus the FIT is generated in one pass,
> when everything is known.

Doing it in one go makes sense to me as well. In general I like the way
distinct processing actions/steps are being split into their own blocks
or so, and I think this helps move things toward that.

> Update the code accordingly. The only functional change is that the 'data'
> property for each node are now last instead of first, which is really a
> more natural position. Update the affected test to deal with this.
> 
> One wrinkle is that the calculated properties (image-pos, size and offset)
> are now added before the FIT is generated. so we must filter these out
> when copying properties from the binman description to the FIT.
> 
> Most of the change here is splitting out some of the code from the
> ReadEntries() implementation into _BuildInput(). So despite the large
> diff, most of the code is the same. It is not feasible to split this patch
> up, so far as I can tell.
> 
> Signed-off-by: Simon Glass <sjg at chromium.org>
> ---
> 
> Changes in v2:
> - Add new patch to refactor fit to generate output at the end
> 
>  tools/binman/etype/fit.py              | 178 ++++++++++++++-----------
>  tools/binman/ftest.py                  |  13 +-
>  tools/binman/test/224_fit_bad_oper.dts |   2 -
>  3 files changed, 109 insertions(+), 84 deletions(-)

Reviewed-by: Alper Nebi Yasak <alpernebiyasak at gmail.com>

I still wrote some weird ideas below, mostly for the future, since this
patch is mostly moving code around which is fine as is.

> diff --git a/tools/binman/etype/fit.py b/tools/binman/etype/fit.py
> index 2d4c5f6545..61c72780e9 100644
> --- a/tools/binman/etype/fit.py
> +++ b/tools/binman/etype/fit.py
> @@ -209,6 +209,81 @@ class Entry_fit(Entry_section):
>          return oper
>  
>      def ReadEntries(self):
> +        def _add_entries(base_node, depth, node):
> +            """Add entries for any nodes that need them
> +
> +            Args:
> +                base_node: Base Node of the FIT (with 'description' property)
> +                depth: Current node depth (0 is the base 'fit' node)
> +                node: Current node to process
> +
> +            He we only need to provide binman entries which are used to define

He -> Here ?

> +            the 'data' for each image. We create an entry_Section for each.
> +            """
> +            rel_path = node.path[len(base_node.path):]
> +            in_images = rel_path.startswith('/images')
> +            has_images = depth == 2 and in_images
> +            if has_images:
> +                # This node is a FIT subimage node (e.g. "/images/kernel")
> +                # containing content nodes. We collect the subimage nodes and
> +                # section entries for them here to merge the content subnodes
> +                # together and put the merged contents in the subimage node's
> +                # 'data' property later.
> +                entry = Entry.Create(self.section, node, etype='section')
> +                entry.ReadNode()

I plan to change 'self.section' to 'self' here later, fixes extracting
wrong contents for FIT subentries.

> +                # The hash subnodes here are for mkimage, not binman.
> +                entry.SetUpdateHash(False)
> +                self._entries[rel_path] = entry
> +

I also plan to change this to a single-level node name instead of the
relative path, lets 'binman extract fit/u-boot' etc. run at all.

> +            for subnode in node.subnodes:
> +                _add_entries(base_node, depth + 1, subnode)
> +
> +        _add_entries(self._node, 0, self._node)

I think it's especially visible here what I meant by switching away from
recursion: this recurses through every node but only does anything on
immediate subnodes of "/images" (for now?).

> +
> +    def BuildSectionData(self, required):
> +        """Build FIT entry contents
> +
> +        This adds the 'data' properties to the input ITB (Image-tree Binary)
> +        then runs mkimage to process it.
> +
> +        Args:
> +            required: True if the data must be present, False if it is OK to
> +                return None

I forgot to handle 'required' while converting FIT to section...

> +
> +        Returns:
> +            Contents of the section (bytes)
> +        """
> +        data = self._BuildInput()
> +        uniq = self.GetUniqueName()
> +        input_fname = tools.get_output_filename('%s.itb' % uniq)
> +        output_fname = tools.get_output_filename('%s.fit' % uniq)
> +        tools.write_file(input_fname, data)
> +        tools.write_file(output_fname, data)
> +
> +        args = {}
> +        ext_offset = self._fit_props.get('fit,external-offset')
> +        if ext_offset is not None:
> +            args = {
> +                'external': True,
> +                'pad': fdt_util.fdt32_to_cpu(ext_offset.value)
> +                }
> +        if self.mkimage.run(reset_timestamp=True, output_fname=output_fname,
> +                            **args) is None:

I have an idea for the far future, to let /image/* nodes sometimes be
Entry_collection to handle external offsets in binman so we can take
mkimage completely out of this, but no clue how feasible/desirable that
end goal is.

> +            # Bintool is missing; just use empty data as the output
> +            self.record_missing_bintool(self.mkimage)
> +            return tools.get_bytes(0, 1024)
> +
> +        return tools.read_file(output_fname)
> +
> +    def _BuildInput(self):
> +        """Finish the FIT by adding the 'data' properties to it
> +
> +        Arguments:
> +            fdt: FIT to update
> +
> +        Returns:
> +            New fdt contents (bytes)
> +        """
>          def _process_prop(pname, prop):
>              """Process special properties
>  
> @@ -236,9 +311,15 @@ class Entry_fit(Entry_section):
>                      val = val[1:].replace('DEFAULT-SEQ', str(seq + 1))
>                      fsw.property_string(pname, val)
>                      return
> +            elif pname.startswith('fit,'):
> +                # Ignore these, which are commands for binman to process
> +                return
> +            elif pname in ['offset', 'size', 'image-pos']:
> +                # Don't add binman's calculated properties
> +                return

This is one of the things I was thinking of doing, thanks. I encountered
the same issue when replacing a FIT entry with the same f"{uniq}.fit"
that was used to build it, will try adding a test for that later.

>              fsw.property(pname, prop.bytes)
>  
> -        def _scan_gen_fdt_nodes(subnode, depth, in_images):
> +        def _gen_fdt_nodes(subnode, depth, in_images):
>              """Generate FDT nodes
>  
>              This creates one node for each member of self._fdts using the
> [...]
>  
> diff --git a/tools/binman/test/224_fit_bad_oper.dts b/tools/binman/test/224_fit_bad_oper.dts
> index cee801e2ea..8a8014ea33 100644
> --- a/tools/binman/test/224_fit_bad_oper.dts
> +++ b/tools/binman/test/224_fit_bad_oper.dts
> @@ -21,7 +21,5 @@
>  				};
>  			};
>  		};
> -		fdtmap {
> -		};

This looked unrelated at first, but as I understand it, changing the
order of things made the bad-operation error happen later and exposed a
breakage from the fdtmap entry trying to parse the mock dtb data.

>  	};
>  };