Hi Phil,
Thanks for pointing to JDK-8041746 . My observations: I think with Andrew's test case we are able to identify the problem submitter might have faced in JDK-6967419 . By deliberately throwing exception at count 30000L we are trying to reproduce scenario of possible IOException where client is closing the socket(probably because of communication problem between client & server) and IDAT chunk is being written. If we change count to any other number(like 10L) which relates to writing of data apart from IDAT chunk we don't see any issue(no exception and cache is closed properly). This might explain why submitter is seeing 7 out of 20 fail. Also by using the test case we are able to see flushedPos going past by 4 bytes to Pos when ios.close() is called as mentioned by submitter in JDK-6967419. So catching the IOException and updating 'startPos' value, will not result in IndexOutOfBoundsException and ios.close() will be performed properly. Please let us know your inputs. Thanks, Jay From: Phil Race Sent: Tuesday, November 17, 2015 3:22 AM To: Jayathirth D V Cc: Prasanta Sadhukhan; 2d-dev@openjdk.java.net Subject: Re: Review request for JDK-6967419 : IndexOutOfBoundsException when drawing PNGs This one reads like it should be obvious but I find it less so .. The unsatisfying part is that we do not seem to know what caused the IOException in the customer case. Andrew came up with a way to reproduce the symptoms but we really don't know what caused the exception in the case of the submitter. It does not seem likely he was 'deliberately' throwing an exception to mess up his own application. I just found this : https://bugs.openjdk.java.net/browse/JDK-8041746 The interesting part is that this bug (the one you are working on) the submitter also wrote that he was using "a ServletOutputStream" So consequently I wonder if it was something like what is described in 8041746 is going on here. It could explain how he sees 7 out of 20 fail. Please take a look at that one to have a think about it. Would your fix help that real world case ? -phil. On 11/12/2015 08:11 PM, Jayathirth D V wrote: Hi Phil, I have added public evaluation in bug. Please review. Thanks, Jay From: Philip Race Sent: Friday, November 13, 2015 12:11 AM To: Jayathirth D V Cc: Prasanta Sadhukhan; HYPERLINK "mailto:2d-dev@openjdk.java.net"2d-dev@openjdk.java.net Subject: Re: Review request for JDK-6967419 : IndexOutOfBoundsException when drawing PNGs Please add a *public* evaluation to the bug report. I will look at it more then .. -phil. On 11/6/15, 2:20 AM, Jayathirth D V wrote: Hi Prasanta, As discussed, only in case of write_IDAT there is finally block which calls ios.finish() which internally calls seek() with improper startPos. In other cases we are not trying to access improper startPos because there is no call to ios.finish(). We can verify this behavior by changing logic where we throw IOException in test case. And I have modified test to not catch IOBE as per your suggestion. Please find updated Webrev link: HYPERLINK "http://cr.openjdk.java.net/%7Erchamyal/jay/6967419/webrev.01/"http://cr.openjdk.java.net/~rchamyal/jay/6967419/webrev.01/ Thanks, Jay From: prasanta sadhukhan Sent: Friday, November 06, 2015 2:45 PM To: Jayathirth D V; HYPERLINK "mailto:2d-dev@openjdk.java.net"2d-dev@openjdk.java.net Cc: Philip Race Subject: Re: Review request for JDK-6967419 : IndexOutOfBoundsException when drawing PNGs Hi Jay, looks ok but I guess you need to do the same for finish() method too in similar way you did for finishChunk() as finish() is called from write_IHDR, write_CHRM etc and it calls flushBefore(). Also, I guess you should not consume IOB Exception and let it be thrown to user instead of RuntimeException after catching IOBE. Regards Prasanta On 11/5/2015 5:25 PM, Jayathirth D V wrote: Hello All, Please review following fix in jdk9: Bug : https://bugs.openjdk.java.net/browse/JDK-6967419 Webrev : HYPERLINK "http://cr.openjdk.java.net/%7Erchamyal/jay/6967419/webrev.00/"http://cr.openjdk.java.net/~rchamyal/jay/6967419/webrev.00/ Bug : IndexOutOfBoundsException when drawing PNGs Root cause : When user intentionally throws IO Exception while write is happening. We call ios.finish() in finally block of write_IDAT() which internally goes to finishChunk(). But the startPos of the chunk is still pointing to present IDAT chunk but flushedPos(streamPos) is pointing to end of IDAT chunk. So in finishChunk(), startPos will be less than flushedPos. This is causing IndexOutOfBoundException in stream.seek() and cache is not closed. Solution : If IOException is thrown by user, catch the exception while write is happening and update startPos to streamPos. So that when seek() happens in finishChunk() we don't see IndexOutOfBoundsException and cache is closed properly. Thanks, Jay
