I thought about that possibility; but right now the individual parts never get their own HeaderFieldValues. Only one is ever assigned, to the MultipartMixedContents. When we call parts(), checkParsed() is called, and the individual *Contents are then created and immediately parsed (HeaderFieldValues are never assigned). However, it occurs to me that if we had a single unparsed SipFrag in a SipMessage's body, and then copied that SipMessage, we could be in loads of trouble.
Best regards, Byron Campen Looking at this further, we don’t have an issue with MultipartMixed. What the sipfrag parser does it save the extra 4 bytes it needs, parse the sipfrag, and restores the four bytes. As a result, we don’t actually stomp on the boundary token, we just temporarily overwrite. This is completely safe until the unlikely day where we parallelize the MultipartMixed parser. However, there might be an even more subtle problem. If the MultiPartMixed is cloned after parse but before the parts are parsed, HeaderFieldValue(const HeaderFieldValue& hfv, CopyPaddingEnum) is not invoked which would mean that SipFrag::parse will write beyond the memory of its hfv. This didn’t show up in valgrind, but I believe it occurs(may not have hit the test, it might be valgrind behaviour). SipMessage invokes HeaderFieldValue(const HeaderFieldValue& hfv, CopyPaddingEnum) to avoid an issue we found w/ SipFrag parsing after a copy. The fix would be to have SipFrag have the same effective behaviour, adding the CopyPaddingEnum cons. Parameter to a new constructor for Contents and LazyParser. Thoughts? -Derek
From: Byron Campen [mailto:bcampen@xxxxxxxxxxxx]
Sent: Tuesday, September 12, 2006 8:55 AM
To: Derek MacDonald
Cc: resiprocate-devel
Subject: Re: [reSIProcate] bug in SipFrag.cxx Well, I wrote a test-case for this, and it appears to function without the fix. It looks like we stomp on the boundary token while parsing the sipfrag, but at that point we have already done everything we need with the boundary token. When we encode, we use the boundary token we stored at the very beginning, and so the bits that we stomped on never make it to the wire. (Although it looks like we are encoding the extra CRLFCRLF at the end of the sipfrag; is this a problem?) This seems fragile to me; any change in the way we are parsing or encoding could cause the whole thing to fall apart.
Does this look sane? Is it important to check whether the primary mimetype is message? Index: resip/stack/Contents.cxx =================================================================== --- resip/stack/Contents.cxx (revision 6561) +++ resip/stack/Contents.cxx (working copy) HeaderFieldValue *hfv = new HeaderFieldValue(contents.data(), contents.size()); + if(contentType.subType()=="sipfrag"||contentType.subType()=="external-body") + // !bwc! The parser for sipfrag requires padding at the end of the hfv. + HeaderFieldValue* temp = hfv; + hfv = new HeaderFieldValue(*temp,HeaderFieldValue::CopyPadding); if (ContentsFactoryBase::getFactoryMap().find(contentType) != ContentsFactoryBase::getFactoryMap().end())
Hmm, probably not; looking at MultipartMixedContents we probably need a fairly expensive special case copy when the content-type is message/sipfrag or message/external. It should be easy to reproduce&fix. -Derek Yes, but is this code safe when you have a sipfrag in a multipart-mixed payload?
Yes; the transport classes use MsgHeaderScanner::allocateBuffer(int size) to guarantee that there are 5 extra bytes at the end of each buffer. I do not know of any memory problems as a result of this with the reciprocate transports. -Derek Good eye. Your solution sounds about as optimal as can be managed. Byron Campen
Hi, I found this in SipFrag.cxx: *constBuffer = pb.position(); // than size bytes of the message. msgHeaderScanner.prepareForFrag(mMessage, hasStartLine(buffer, size)); char saveTermCharArray[0] = termCharArray[0]; termCharArray[0] = ; '\r' *scanTermCharPtr; &scanTermCharPtr); termCharArray[3] = saveTermCharArray[3]; The problem with this code is that the sentinel is wrote *behind* the ParseBuffer. If the ParseBuffer stands at the very end of an allocated buffer, this code writes 4 bytes behind it. So it happened in our application. As a consequence a further allocation of memory inside of scanChunk() failed – probably because some administration information of the memory heap was overwritten. I can easily reproduce this using the gflags tool of the Mircosoft debugging tools. My solution consists in allocating a new buffer that has the required size (+4), copying the content of the ParseBuffer to it. The new buffer is referenced by the _______________________________________________ resiprocate-devel mailing list
|