My analysis leads me to a similar conclusion. Reference counting is out and there is no need for it. This is my hypothesis:
If things are being freed unintentionally, then there are not enough ruby to ruby cross-references to maintain integrity during mark and sweep. A single ruby object holding on to an xmlNode should inhibit collection of an entire tree. Individual ruby xml nodes can be reaped, but there should be no xmlFree done by anything but the document level object. As for an xmlNode that is not associated with a document, it must call xmlFree (the xmlNode already indicates if it is associated with a document) if it has no parent. Which effectively translates into the same document semantics. One operational change should be once an xmlNode is associated with a document it cannot be associated with another one (at least using the implicit copy mechanism provided in ruby_xml_node_child_set). The ruby xmlNode becomes changed and reflects its association with the document. However, using xmlCopyNode, it is possible to get a copy of the node without an association to the document and also libxml2 handles propagation of namespaces. This new copy preserves the document based free condition above, and also makes it easy to put it into another document. Not sure if this should be automatic (thinking not) or an overload of .dup. This should establish a bright line between libxml2 and ruby for managing memory. This is from only a few hours looking at the docs and source code for libxml2, the python interface and the ruby interface. If these seem to ignore some known pitfall, please let me know. If not, this should fix this scope of the memory problem. I intend to do this within a week. Post your warnings or wish me luck. Dan On Aug 23, 2007, at 14:25, Sean Chittenden wrote: > Having spent a great deal of time with this, I'm now of the following > opinion of this library: when initially writing this module, I > exposed too much of libxml to Ruby and the memory models and object/ > tree dependencies and interactions between memory and objects is... > problematic and should be reviewed. $0.02. -sc _______________________________________________ libxml-devel mailing list libxml-devel@rubyforge.org http://rubyforge.org/mailman/listinfo/libxml-devel
