UNDERTOW-2017 debugging #1287
UNDERTOW-2017 debugging #1287
Conversation
It appears that transitions between states fail to store relevant data -- this is difficult to reproduce locally because tests use loopback or otherwise fast networks. By introducing small buffers and writes that often make no progress, we can reproduce issues from slower networks.
| HttpServerExchange exchangeSnapshot = exchange; | ||
| if (exchangeSnapshot != null) { | ||
| IoUtils.safeClose(exchangeSnapshot.getConnection()); | ||
| } |
There was a problem hiding this comment.
I observed another NPE here, I will cherry-pick this into a separate PR
There was a problem hiding this comment.
Linking that PR here for posterity: #1289
| @@ -435,4 +456,8 @@ public void failed(IOException e) { | |||
| }; | |||
| } | |||
|
|
|||
| public static ByteBufferPool getBufferPool() { | |||
| return new DefaultByteBufferPool(true, 50, 0, 0, 100); | |||
There was a problem hiding this comment.
Small buffers result in more frequent write attempts
| if (res == 0) { | ||
| this.headerName = headerName; | ||
| this.valueIterator = valueIterator; | ||
| this.nameIterator = nameIterator; | ||
| log.trace("Continuation"); | ||
| return STATE_HDR_EOL_CR; |
There was a problem hiding this comment.
The transitions to STATE_HDR_EOL_LF and STATE_HDR_EOL_CR require additional data.
I might suggest saving all state prior to every transition to avoid this kind of bug in the future, however it would require more operations per invocation. That said, in the common case I'd expect the request line and all headers to fit into a single buffer, so it's likely there's no real cost.
|
Thanks for sharing all this background for testing of UNDERTOW-2017 fix @carterkozak ! |
|
Closing this PR that was created for analysis and reproducing of the UNDERTOW-2017. The bug fix, at #1288, is merged. |
It appears that transitions between states fail to store
relevant data -- this is difficult to reproduce locally because
tests use loopback or otherwise fast networks. By introducing
small buffers and writes that often make no progress, we can
reproduce issues from slower networks.