Bug 101268 – Fixed flaky stack overflow test: fast/xmlhttprequest/xmlhttprequest-recursive-sync-event.html

• Status: RESOLVED
• Resolution: FIXED
• Component: Tools / Tests

[Mark Lam]

Test fast/xmlhttprequest/xmlhttprequest-recursive-sync-event.html is flaky.  Here's how we can make it not flaky:

// This test is flaky because it depends on the size of the available native
// stack, the size of native stack frames, and the layering of function calls
// in the JavaScript VM code. Together, these variables determine when a
// StackOverflowError will be thrown in the course of this test of unbounded
// recursion.
//
// From experience with this test, StackOverflow error tends to be thrown in
// 2 possible scenarios:
//
// Scenario 1: Overflowed while calling xhr.open()
// ===============================================
// In this scenario, the VM is executing xhr.open() and succeeds in setting
// up the new request. In response to having opened the request, it tries to
// dispatch an event with readyState 1 (OPENED). Unfortunately, there is not
// enough stack and we get an overflow error and 1 console error message here.
//
// The error is thrown but is then eaten up by the EventDispatcher. Hence,
// onreadystatechange() back in JS land does not even see this error.
//
// open() eventually returns to onreadystatechange(), and we call xhr.send()
// next. xhr.send() overflows the stack again, and we get a 2nd error and
// a 2nd console error message.
//
// Note: the lastReadyStateObserved in this case is 4 because the event
// dispatch for OPENED failed to execute onreadystatechange().
//
// The total number of error messages seen is 2.
//
// Scenario 2: Overflowed while calling xhr.send()
// ===============================================
// In this scenario, the VM is executing xhr.send() and a readyState change
// to 4 (DONE) occurs. An event is dispatched and the VM tries to call
// onreadystatechange() but overflows the stack. Here, we see 1 error and
// 1 console error message.
//
// Since we never succeeded in reentering onreadystatechange(), we did not
// not set up anymore requests, and will not attempt to re-enter
// onreadystatechange() after this. Hence, we will not see anymore overflow
// errors nor console error messages.
//
// In contrast with scenario 1, we did not see an overflow error for xhr.open()
// before we saw the error for xhr.send().
//
// Note: lastReadyStateObserved in this case is 1 because the event
// dispatch for OPENED succeeded in executing onreadystatechange(), and
// lastReadyStateObserved was set to 1.
//
// The total number of error messages seen is 1.
//
// Flakiness
// =========
// Since we can end up with scenario 1 or 2 depending on variables outside the
// control of this test, the result of the test would be inherently flaky i.e.
// we can see 1 or 2 console error messages.
//
// That is unless we do some compensation. In the case of scenario 2, we can
// check if the lastReadyStateObserved is 1 which means we would have only seen
// one console error message. If so, we can call xhr.send() again just to
// trigger another one so that we'll have a total of 2.
//
// Note that in the compensation case, we'll need to call xhr.open() again
// before calling xhr.send(). This is because the previous xhr request has
// already been closed. Here, xhr.open() should not trigger an overflow error
// just like it didn't before the previous xhr.send() triggered the error (the
// fact that it succeeded without an error before means it will succeed this
// time too). And this leaves it up to our compensating call to xhr.send()
// to generate the 2nd expected overflow error and console error message.
//
// Note also that we'll need to make sure that we do this compensation only
// once, and not repeatedly do the compensating xhr.send() call on every
// onreadystatechange() frame on the return path.

[Mark Lam]

Created attachment 172420 [details]
Fix.

[Geoffrey Garen]

Comment on attachment 172420 [details]
Fix.

Nice. I like this much better than just marking the test flaky in the testing harness.

[Mark Lam]

Landed in r133540: <http://trac.webkit.org/changeset/133540>.