Bug 41507 – WebSocket: Crash caused by calling close() within onmessage handler

• Component: WebCore Misc.
• Status: RESOLVED FIXED

[Yuta Kitamura]

If you create a WebSocket object during onclose() handler of another WebSocket object, the browser will crash.

This crash may be related to http://crbug.com/44792.

[Yuta Kitamura]

Created attachment 60341 [details]
Fix

[Fumitoshi Ukai]

Comment on attachment 60341 [details]
Fix

LGTM

[Alexey Proskuryakov]

Comment on attachment 60341 [details]
Fix

The code change looks reasonable, but there are some logical steps missing from the explanation. Why does creating a WebSocket object during onclose() handler of another WebSocket object cause this crash? They have separate channels and bridges.

+    while (m_workerContext && clientWrapper && !clientWrapper->syncMethodDone() && result != MessageQueueTerminated) {
+        result = runLoop.runInMode(m_workerContext.get(), m_taskMode); // May dereference this bridge and m_workerContext.
         clientWrapper = m_workerClientWrapper.get();
     }

What is the exact mechanism that dereferences worker context while waiting here? Is this part of the change covered by the test?

[Yuta Kitamura]

(In reply to comment #3)
> (From update of attachment 60341 [details])
> The code change looks reasonable, but there are some logical steps missing from the explanation. Why does creating a WebSocket object during onclose() handler of another WebSocket object cause this crash? They have separate channels and bridges.
> 
> +    while (m_workerContext && clientWrapper && !clientWrapper->syncMethodDone() && result != MessageQueueTerminated) {
> +        result = runLoop.runInMode(m_workerContext.get(), m_taskMode); // May dereference this bridge and m_workerContext.
>          clientWrapper = m_workerClientWrapper.get();
>      }
> 
> What is the exact mechanism that dereferences worker context while waiting here? Is this part of the change covered by the test?

I found the test case first, and then looked into the failure. So I'm not sure about the exact mechanism of why this test causes the failure, but it is 100% reproducible with this test.

The source of the failure is that calling WorkerRunLoop::runInMode() may call WebSocket::didClose() (which frees everything). The stack trace is like the following:

    WebSocket::close() is called (in ws1.close())
 -> WorkerThreadableWebSocketChannel::bufferedAmount()
 -> WorkerThreadableWebSocketChannel::Bridge::bufferedAmount()
 -> WorkerThreadableWebSocketChannel::Bridge::waitForMethodCompletion()
 -> WorkerRunLoop::runInMode()
 -> ... -> workerContextDidClose()
 -> ThreadableWebSocketChannelClientWrapper::didClose()
 -> ThreadableWebSocketChannelClientWrapper::processPendingEvents()
 -> WebSocket::didClose()

This causes the channel and other things to get closed, while WebSocket::close() does not think such thing should happen.

> +        result = runLoop.runInMode(m_workerContext.get(), m_taskMode); // May dereference this bridge and m_workerContext.

This comment was a bit inaccurate. Since m_workerContext may become null in Bridge::disconnect(), the comment should be "May causes the bridge to be disconnected, which makes m_workerContext become null." I will update the comment in a new patch.

[Alexey Proskuryakov]

Created attachment 60776 [details]
a simpler test

I'll need to think about this patch more deeply, we may be violating some more general ownership rules here. It's awesome that we have a reproducible test case!

>     m_bufferedAmountAfterClose = m_channel->bufferedAmount(); // May dereference m_channel.
>     if (m_channel)
>         m_channel->close();

This comment at least is wrong - we don't care about m_channel being dereferenced - all we care about to make this call is m_channel being non-null. WebSocket holds its own reference to channel.

I have a simpler reproducible case, one that doesn't even need to create another WebSocket. It crashes in several different places.

[Alexey Proskuryakov]

Comment on attachment 60341 [details]
Fix

The title of the bug isn't great - there are multiple crash locations being fixed, not just waitForMethodCompletion(). And this isn't about creating a WebSocket object from onclose(), as the original description said.

I think that this patch is fine, and just needs better comments.

[Yuta Kitamura]

Created attachment 61023 [details]
Fix v2 (Simplify test, add comments)

[Yuta Kitamura]

Apologies for delay in response. In the new patch, I tried to:
- Add more comments and clarify the situation.
- Simplify the test according to Alexey's proposal.

Could you take another look?

[Alexey Proskuryakov]

Comment on attachment 61023 [details]
Fix v2 (Simplify test, add comments)

 +      m_bufferedAmountAfterClose = m_channel->bufferedAmount(); // May dereference m_channel.

This comment is still misleading, we don't care about it being dereferenced. WebSocket holds its own reference to m_channel.

I thin that the real explanation is that a didClose notification may be already queued, which we will inadvertently process while waiting for bufferedAmount() to return.

[Yuta Kitamura]

Created attachment 61222 [details]
Fix v3 (Improve comments as per Alexey's suggestion)

[WebKit Commit Bot]

Comment on attachment 61222 [details]
Fix v3 (Improve comments as per Alexey's suggestion)

Clearing flags on attachment: 61222

Committed r63159: <http://trac.webkit.org/changeset/63159>

[WebKit Commit Bot]

All reviewed patches have been landed.  Closing bug.