[Bug]: Flood of "Recursion depth exceeded in _find_image_xrefs_page" about ocrmypdf

[jbarlow83]

On page 18 /Xi9 appears to be in a circular reference loop that includes itself over and over. I can guard against it but the file is likely malformed at this location. Try Ghostscript to rewrite the PDF.

If it does just happen to be really deep Python has a system call to increase recursion depth which you could use to force the file through. Doubt it would help though.

[user1584]

Thanks, for my test document, Ghostscript detected the problem and repaired the file.
Changing the python recursion limit does not seem to matter for this issue since the function contains a fixed recursion limit of 10.

[jbarlow83]

I added a guard that should prevent recursion into previously visited XObjects. Please test and let me know if it works.

It's still probably best to use ghostscript to repair these files.

[user1584]

The guard does not seem to work. If I understand it correctly, you check against all the included/excluded xrefs before the recursion but the xref are added only after the recursion to the set of included/excluded xrefs. Thus, the function can cycle through the circular references forever before the problematic xrefs are added.

I tried a different approach by passing all the layers of the recursion:

def _find_image_xrefs_container(
    pdf: Pdf,
    container: Object,
    pageno: int,
    include_xrefs: MutableSet[Xref],
    exclude_xrefs: MutableSet[Xref],
    pageno_for_xref: dict[Xref, int],
    recursion_layers: tuple[Xref,...] = tuple()
):
    """Find all image XRefs or Form XObject and add to the include/exclude sets."""
    if len(recursion_layers) > 10:
        log.warning("Recursion depth exceeded in _find_image_xrefs_page")
        return
    try:
        xobjs = container.Resources.XObject
    except AttributeError:
        return
    for _imname, image in dict(xobjs).items():
        if image.objgen[1] != 0:
            continue  # Ignore images in an incremental PDF
        xref = Xref(image.objgen[0])
        if xref in recursion_layers:
            log.warning(f"Skipping {_imname} in page {pageno} since it seems to be recursive.")
            continue  # Already processed
        if Name.Subtype in image and image.Subtype == Name.Form:
            # Recurse into Form XObjects
            log.debug(f"Recursing into Form XObject {_imname} in page {pageno}")
            _find_image_xrefs_container(
                pdf,
                image,
                pageno,
                include_xrefs,
                exclude_xrefs,
                pageno_for_xref,
                recursion_layers=tuple([*recursion_layers, xref])
            )
            continue
        if Name.SMask in image:
            # Ignore soft masks
            smask_xref = Xref(image.SMask.objgen[0])
            exclude_xrefs.add(smask_xref)
            log.debug(f"xref {smask_xref}: skipping image because it is an SMask")
        include_xrefs.add(xref)
        log.debug(f"xref {xref}: treating as an optimization candidate")
        if xref not in pageno_for_xref:
            pageno_for_xref[xref] = pageno

This seems to work with our problematic PDFs.

[user1584]

@jbarlow83: We found a PDF that could not even be processed with the changes posted above. The new (and so far working) version is this:

def _find_image_xrefs_container(
    pdf: Pdf,
    container: Object,
    pageno: int,
    include_xrefs: MutableSet[ocrmypdf.optimize.Xref],
    exclude_xrefs: MutableSet[ocrmypdf.optimize.Xref],
    pageno_for_xref: dict[ocrmypdf.optimize.Xref, int],
    depth: int = 0,
    already_processed_xrefs: typing.Optional[MutableSet[ocrmypdf.optimize.Xref]] = None,
    strange_xrefs: typing.Optional[MutableSet[ocrmypdf.optimize.Xref]] = None,
):
    """Find all image XRefs or Form XObject and add to the include/exclude sets."""
    if depth > 10:
        ocrmypdf.optimize.log.warning(
            "Recursion depth exceeded in _find_image_xrefs_page"
        )
        return

    try:
        xobjs = container.Resources.XObject
    except AttributeError:
        return

    # If no xrefs were passed, an empty set is created.
    _already_processed_xrefs: MutableSet[ocrmypdf.optimize.Xref]
    if already_processed_xrefs is None:
        _already_processed_xrefs = set()
    else:
        _already_processed_xrefs = already_processed_xrefs

    # If no strange xrefs were passed, an empty set is created.
    _strange_xrefs: MutableSet[ocrmypdf.optimize.Xref]
    if strange_xrefs is None:
        _strange_xrefs = set()
    else:
        _strange_xrefs = strange_xrefs

    # This is a mapping of objects' xrefs to the objects' names and objects themself.
    to_be_processed_objects = {
        ocrmypdf.optimize.Xref(image.objgen[0]): (_imname, image)
        for _imname, image in dict(xobjs).items()
    }

    # Only the xrefs of the to-be-processed objects.
    to_be_processed_xrefs: set[ocrmypdf.optimize.Xref] = set(
        to_be_processed_objects.keys()
    )

    # These are the xrefs that have not been processed before.
    unprocessed_xrefs = to_be_processed_xrefs - _already_processed_xrefs

    # Collecting "strange" xrefs (probably involved in cyclic xrefs?) for later logging.
    _strange_xrefs |= to_be_processed_xrefs - unprocessed_xrefs

    # This will be passed to further function calls. Thus, we can consider
    # these xrefs processed.
    to_be_already_processed_xrefs: set[ocrmypdf.optimize.Xref] = (
        set(_already_processed_xrefs) | unprocessed_xrefs
    )

    for xref in unprocessed_xrefs:
        _imname, image = to_be_processed_objects[xref]
        if image.objgen[1] != 0:
            continue  # Ignore images in an incremental PDF

        if Name.Subtype in image and image.Subtype == Name.Form:
            # Recurse into Form XObjects
            ocrmypdf.optimize.log.debug(
                f"Recursing into Form XObject {_imname} in page {pageno}"
            )
            _find_image_xrefs_container(
                pdf=pdf,
                container=image,
                pageno=pageno,
                include_xrefs=include_xrefs,
                exclude_xrefs=exclude_xrefs,
                pageno_for_xref=pageno_for_xref,
                depth=depth + 1,
                already_processed_xrefs=to_be_already_processed_xrefs,
                strange_xrefs=_strange_xrefs,
            )
            continue

        if Name.SMask in image:
            # Ignore soft masks
            smask_xref = ocrmypdf.optimize.Xref(image.SMask.objgen[0])
            exclude_xrefs.add(smask_xref)
            ocrmypdf.optimize.log.debug(
                f"xref {smask_xref}: skipping image because it is an SMask"
            )

        # Everything that passed the previous filters is relevant for optimization.
        include_xrefs.add(xref)
        ocrmypdf.optimize.log.debug(
            f"xref {xref}: treating as an optimization candidate"
        )
        if xref not in pageno_for_xref:
            pageno_for_xref[xref] = pageno

    # Notifying about strange xrefs.
    if depth == 0 and _strange_xrefs:
        ocrmypdf.optimize.log.warning(
            f"{len(_strange_xrefs)} xrefs on page {pageno} were found to be kind of strange."
        )

As it turns out, finding suitable variable names is even harder when writing recursive functions. :)