[ https://issues.apache.org/jira/browse/PDFBOX-4220?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16472256#comment-16472256 ]
John Hewson edited comment on PDFBOX-4220 at 5/11/18 4:56 PM: -------------------------------------------------------------- [~tilman] Thread safety is only part of the problem, the overall problem is actually that fonts are shared objects. In other words the following situations cause problems: 1) reuse of a font within just one document (e.g. imagine using Arial Unicode MS to write some Arabic with those features enabled, then turning them off and writing some English - the subsetter now has an inconsistent view of the font). 2) Reuse of a font between documents (e.g. if I'm writing with the same TTF font file to two PDF files even with just one thread, fonts are shared between them). 3) Reuse of a font across threads - this is just a more general version of (2). The fix will be to move the enabled/disabled feature state out of the font. This actually works out nicely because features don't belong at the font-level but at the text-level. One expects to have an API such as showText(string, List<Feature>) which allows features to be changed many times while working with a single font (an example would be the "dlig" feature which enables discretionary stylistic ligatures to add some visual decoration, perhaps for a title or heading). was (Author: jahewson): [~tilman] Thread safety is only part of the problem, the overall problem is actually that fonts are shared objects. In other words the following situations cause problems: 1) reuse of a font within just one document (e.g. imagine using Arial Unicode MS to write some Arabic with those features enabled, then turning them off and writing some English - the subsetter now has an inconsistent view of the font). 2) Reuse of a font between documents (e.g. if I'm writing two PDF files using just one thread, then fonts will be shared between them). 3) Reuse of a font across threads - this is just a more general version of (2). The fix will be to move the enabled/disabled feature state out of the font. This actually works out nicely because features don't belong at the font-level but at the text-level. One expects to have an API such as showText(string, List<Feature>) which allows features to be changed many times while working with a single font (an example would be the "dlig" feature which enables discretionary stylistic ligatures to add some visual decoration, perhaps for a title or heading). > FontBox sets GSUB features globally across shared fonts > ------------------------------------------------------- > > Key: PDFBOX-4220 > URL: https://issues.apache.org/jira/browse/PDFBOX-4220 > Project: PDFBox > Issue Type: Bug > Components: FontBox, Parsing, Writing > Affects Versions: 2.0.9, 3.0.0 PDFBox > Reporter: John Hewson > Assignee: John Hewson > Priority: Major > > Migrating the following issue from PDFBOX-4106: > {quote} > (!) So we have a problem. While looking at building a proper ToUnicodeMap for > PDFBOX-4189 I've encountered some significant issues related to this feature. > Given that this was shipped in 2.0.9, we're likely going to face some hard > choices. > Most of the handling of vertical text in this patch is fine. It generally > does a good job of handling the intricacies of both PDF and OpenType and > juggling the sometimes competing concerns. > The issue is that the new APIs added to FontBox's TrueTypeFont are > incompatible with PDFBox's multi-threading. Unlike most of PDFBox, fonts in > FontBox must be thread safe - they are cached and shared globally between all > PDDocument instances. For this reason a FontBoxFont must not contain any > document-specific state. > The following fields added to TrueTypeFont contain document-specific state: > {{List<String> enabledGsubFeatures}} > The following methods added to TrueTypeFont write document-specific state: > {{enableVerticalSubstitutions()}} > {{enableGsubFeature(String)}} > {{disableGsubFeature(String)}} > The following methods added to TrueTypeFont read document-specific state: > {{getUnicodeCmapLookup()}} > {{getUnicodeCmapLookup(boolean)}} > None of the additions are thread safe, anyone who calls these methods will > clobber the corresponding state for all other threads. This problem can even > occur if the user is manipulating more than one document within a single > thread. *There's no way around this and no way to fix these methods* - > document-specific state can't live in shared global fonts :( > (flag) Tough choices: Given that these are all just auxiliary methods limited > to just FontBox's TrueTypeFont class, we _could_ remove them. These are very > obscure methods which have only been around for a few weeks. It's unlikely > that _any_ users would be affected by their removal. Obviously the missing > functionality will be implemented in the appropriate location*, so vertical > text will still work in 2.0 and the existing PDType0Font.loadVertical API - > which is the only thing that matters - will remain unchanged! :) > Yes, I'm suggesting a *very obscure* breaking API change to 2.0 but as it > stands we have just shipped a breaking functionality change to 2.0 and broken > existing functionality in an irreparable manner, and that seems worse. > Thoughts? > \* And I have some ideas how to achieve this and am happy to do it > {quote} -- This message was sent by Atlassian JIRA (v7.6.3#76005) --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@pdfbox.apache.org For additional commands, e-mail: dev-h...@pdfbox.apache.org