Community
Participate
Working Groups
Build Identifier: 8.0.4.v20111024 Similar to Bug 363757 (https://bugs.eclipse.org/bugs/show_bug.cgi?id=363757), but a different cause, and still the case in trunk, if there is back pressure on the socket writing the request, it seems that the HttpGenerator can end up missing a chunk header (which results in a 'bad chunk char' on the server). I'm attaching a patch which includes a test case that reproduces the issue, along with a possible (though a tad clunky) patch. The problem occurs when the client calls addContent twice, but is unable to write out the first full buffer. If the first buffer has a length of greater than 1024 characters, then it enters "_bypass", so _buffer stays null. When the second addContent is called, it will attempt to flushBuffer() to clear up the _content buffer. The initial chunk header is written to _header, and then _header and some amount of _content are written out. If flushBuffer() does not manage to flush all of _content, then it concatenates _content and the new content buffer together into _content, without adding a chunk header for the new content that it is adding. Reproducible: Always Steps to Reproduce: Run testChunkedWithBackPressure() from the attached patch.
Created attachment 210341 [details] Proposed patch with test case.
Sorry for the slow response looking at this one. Analysing now.
I've applied and committed your patch more or less as it is. I agree it is a tad clunky... but mostly because the whole HttpGenerator is a tad clunky. I've not spent any time cleaning it up, as this is going through a total refactor currently in the jetty-9 branch, to avoid exactly that kind of situation. thanks for the contribution.
74da51a8e61f227ccf13b0850d0791594e90a711
