[
https://issues.apache.org/jira/browse/TS-2678?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13959060#comment-13959060
]
James Peach commented on TS-2678:
---------------------------------
This is breaking because {{HttpSM::is_http_server_eos_truncation}} thinks that
this is a response truncation. Interestingly, the {{Content-Length}} header
handling towards the end of {{HttpSM::is_http_server_eos_truncation}} would get
this case correct because it the absence of a {{Content-Length}} header would
cause it to return false (ie. that the request is not truncated).
{{HttpSM::is_http_server_eos_truncation}} in this case returns true because the
{{t_state.current.server->keep_alive == HTTP_KEEPALIVE}} condition succeeds.
The chunk handling condition is also correct for this case (ie. it would
evaluate to false):
{code}
2724 if ((p->do_dechunking || p->do_chunked_passthru) &&
p->chunked_handler.truncation) {
(lldb) p t_state.current.server->keep_alive
(HTTPKeepAlive) $0 = HTTP_KEEPALIVE
(lldb) p *p
(HttpTunnelProducer) $1 = {
...
chunked_handler = {
...
truncation = false
...
}
chunking_action = TCA_CHUNK_CONTENT
do_chunking = true
do_dechunking = false
do_chunked_passthru = false
...
}
{code}
The keep alive condition has been there since the beginning of source history.
I'm not sure that check is appropriate, but I need to research how
{{ConnectionAttributes::keep_alive}} is actually used.
> Some sites (e.g. craigslist) fails to load with patch in e6b9cb533bba28
> -----------------------------------------------------------------------
>
> Key: TS-2678
> URL: https://issues.apache.org/jira/browse/TS-2678
> Project: Traffic Server
> Issue Type: Bug
> Components: HTTP
> Reporter: Leif Hedstrom
> Assignee: James Peach
> Priority: Critical
> Fix For: 5.0.0
>
>
> The commit e6b9cb533bba28, from
> TS-2559: Disconnect clients on unrecoverable origin errors
> is causing some sites to break badly:
> {code}
> curl: (52) Empty reply from server
> {code}
> or sometimes
> {code}
> HTTP/1.1 302 Found
> Location: http://www.craigslist.org/
> Date: Mon, 31 Mar 2014 02:36:51 GMT
> Age: 0
> Transfer-Encoding: chunked
> Proxy-Connection: keep-alive
> Server: ATS/5.0.0
> curl: (18) transfer closed with outstanding read data remaining
> {code}
> The first one happens most frequently in my test, and the second one is close
> to what is expected (but not quite). Both makes the page not load properly
> from any browser.
> To reproduce this, using a "default" installation, with remap.require
> disabled (so, forward proxy), I run
> {code}
> curl -D - -x localhost:8080 http://www.craigslist.com
> {code}
> I don't know what it is that Craigslist does to trigger this, but backing out
> the patch from TS-2559 does fix the problem.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
