PDFBOX currently reads xref table/trailer and XRef objects without using startxref or 'Prev' information which results in applying not active data resulting in using wrong objects or resulting in parsing exceptions because old trailer settings do not apply anymore. This happens especially with updated PDF documents where changes are simply appended and old objects/xref entries remain but are not referenced. My last patch (
PDFBOX-1014) tried to solve this for a specific case but it was based on assumptions which do not hold in every case.
The specification compliant way is to read the last startxref which points to the last xref object which itself may reference further xref objects using 'Prev' attribute.
I have written a fix which works the standard way and can fall back to the old behavior in case startxref is wrong or missing. The fix tries to be as unobtrusive as possible. A new class (o.a.p.pdfparser.XrefTrailerResolver) is filled with all xref table/trailer and XRef object data. After document is parsed (and last startxref is read) this class creates xref table and trailer using startxref and 'Prev' information. Beside this new class there are small changes to PDFParser and COSDocument.
This bugfix/improvement should bring PDFBOX a good step closer to be PDF specification conform - especially as long as the new specification conform parser project is not finished.
This bugfix supersedes the fix from