Bug 9113 – REGRESSION (14581): XMLHttpRequest never calls onreadystatechange with a readystate == 4

• Component: XML
• Status: RESOLVED FIXED

[David Phillip Oster]

Regression: 14581 XMLHttpRequest never calls onreadystatechange with a readystate == 4

When I make a POST XMLHttpRequest, my handler, assigned to onreadystatechange is never called with readystate == 4 /* 4 means "loaded" */

14555 performed correctly.

[David Phillip Oster]

One place to see this bug: http://google.com/calendar   Note the red box with the white text "Loading..." at the upper right of the page never goes away, only in this version of webKit.

[David Phillip Oster]

The bug always happens trying to load the main window of Google Calendar. My callback gets called with state==1, then never gets called again.

The bug doesn't happen using the simple example on http://developer.apple.com/internet/webcontent/xmlhttpreq.html. (Uses 'GET')

The bug doesn't happen using a custom cgi I wrote, which uses POST, and issues two requests operating in parallel (modeling what Google Calendar is doing). (I've also tried it with just a single outstanding request.)

[David Phillip Oster]

My guess is it needs to have a server that sends a redirect. Still trying to track it down.

The regression happened between r14566 and r14569, probably in r14569 (fix <rdar://problem/4557926>) due to the following:
        Also removed XMLHttpRequest from the DOM object cache because XMLHttpRequest
        objects aren't accessed through the DOM.

Created attachment 8649 [details]
Undo the XMLHttpRequest change from r14569

Undoing the XMLHttpRequest part of r14569 fixes the regression, but I have no understanding of the DOM object cache.

[David Phillip Oster]

Thank you! It looks like, in my case, where the server redirects, that XMLHttpRequest->m_doc->m_view->m_frame becomes nil, sometime between the request's send() has been called. The fix you made possibly keeps the frame from being improperly garbage collected.

[David Kilzer (:ddkilzer)]

(In reply to comment #3)
> My guess is it needs to have a server that sends a redirect. Still trying to
> track it down.

Note that the LayoutTests now support "http" tests that run as CGI scripts on an Apache web server, so it should be possible to write a test case for this.

I think it's safe to confirm this bug now, too.

[Geoffrey Garen]

Created attachment 8652 [details]
Mitz's patch w/changelog and layout test

Uploading for posterity.

[Geoffrey Garen]

Committed revision 14671.

[David Phillip Oster]

Verified. This fixes my problem with Google Calendar. Thanks guys!