> It was the abbreviated handshaking. I guess that the previous client
> has not closed the socket completely, so for *this* handshaking, the
> abbreviated handshaking rather than the full handshaking is used.

The issue is not due to closing the previous connections completely. The two close messages are racing and the messages appeared out of the "normal" ordering in this one case when the thread did a simple session resume.

> I tied several different approaches within this test, but failed to
> reproduce the abbreviated handshaking. It is not easy to hack the
> test without significant changes.

The original testcase was for the bad_record_mac issue, and testing whether or not ByteBuffers with valid hasArray() was working correctly. I wasn't intending to test renegotiation/abbreviated handshakes. I obviously didn't completely step through the code when I added the second runTest(). :( That explains the race condition on close ordering, and the issue is still there.

A better fix would be the following so that the test is actually starting from scratch each time (no session resumes):

diff --git a/test/sun/security/ssl/com/sun/net/ssl/internal/ssl/SSLEngineImpl/SSLEngineBadBufferArrayAccess.java b/test/sun/security/ssl/com/sun/net/ssl/internal/ssl/SSLEngineImpl/SSLEngineBadBufferArrayAccess.java --- a/test/sun/security/ssl/com/sun/net/ssl/internal/ssl/SSLEngineImpl/SSLEngineBadBufferArrayAccess.java +++ b/test/sun/security/ssl/com/sun/net/ssl/internal/ssl/SSLEngineImpl/SSLEngineBadBufferArrayAccess.java
@@ -146,14 +146,13 @@
             "SSLv3", "TLSv1", "TLSv1.1", "TLSv1.2" };

         for (String protocol : protocols) {
-            log("Testing " + protocol);
             /*
              * Run the tests with direct and indirect buffers.
              */
-            SSLEngineBadBufferArrayAccess test =
-                new SSLEngineBadBufferArrayAccess(protocol);
-            test.runTest(true);
-            test.runTest(false);
+            log("Testing " + protocol + ":true");
+            new SSLEngineBadBufferArrayAccess(protocol).runTest(true);
+            log("Testing " + protocol + ":false");
+            new SSLEngineBadBufferArrayAccess(protocol).runTest(false);
         }

         System.out.println("Test Passed.");

Please consider opening a new bug and changing.

Brad




On 7/2/2012 8:15 PM, Xuelei Fan wrote:
On 7/3/2012 11:09 AM, Weijun Wang wrote:
Your fix looks fine.

Thanks!

IMHO, the remind is not really useful unless you dump more info, say,
the value of serverIn.remaining().
We can get the value from analysis of the log. The remind is only used
for the case that we do not really fix the issue with this update.

Xuelei

QE would report the failure to "THE
SECURITY TEAM" anyway.

-Max

On 07/03/2012 11:00 AM, Xuelei Fan wrote:
On 7/3/2012 10:40 AM, Weijun Wang wrote:
No new test needed. I only think that you might be able to hack the
current test a little to reproduce this and see if the failure is the
same and if your code change can fix it. There is no need to keep this
hack in your final changeset.

I tied several different approaches within this test, but failed to
reproduce the abbreviated handshaking. ;-) It is not easy to hack the
test without significant changes.

Xuelei

-Max


On 07/03/2012 10:37 AM, Xuelei Fan wrote:
On 7/3/2012 10:02 AM, Weijun Wang wrote:


On 07/03/2012 09:48 AM, Xuelei Fan wrote:
On 7/2/2012 4:35 PM, Weijun Wang wrote:
I take a look at the test output. When the last handshake starts:

================
server unwrap: OK/NEED_TASK, 230/0 bytes
        running delegated task...
        new HandshakeStatus: NEED_WRAP
----
server wrap: OK/NEED_WRAP, 0/86 bytes
================

Here the first wrap only generates 86 bytes, I guess that's the
ServerHello message? It keeps the state at NEED_WRAP but then never
really generates the Certificate message. What might be the problem?

Good catch!

It was the abbreviated handshaking. I guess that the previous client
has
not closed the socket completely, so for *this* handshaking, the
abbreviated handshaking rather than the full handshaking is used.

For full handshaking, it is the client sending the "Finished"
message at
first. However, for abbreviated handshaking, the server send the
"Finished" message at first.

In the current scenarios, it is expected that the client sends its
application data (26 bytes), and then the server sends its
application
data (29 bytes). However, the abbreviated handshaking disorder the
sequence in that it is the sever sends it application data (29 bytes)
before client. In such cases, the following logic does not stand any
more:
        if (!closed&&  (serverOut.remaining() == 0)) {
           closed = true;
           ...
           if (serverIn.remaining() != clientMsg.length) {

               throw new Exception("Client:  Data length error");
           }
           ...
        }

Because the server has not receive the client message when the server
sends its application data.

I think it is a test issue, the current fix should has already
addressed
the issue.

That's great.

Since the reason is clear, can you reproduce this failure and then
confirm the current fix does solve the problem?

It is possible to reproduce this failure with a new test case. But
it is
pretty hard to reproduce it within this test. I was wondering as it is
test bug, so we may not want a extra test case to prove that this test
is correct.

We also have a nested remind that, "IF THIS FAILS, PLEASE REPORT
THIS TO
THE SECURITY TEAM.  WE HAVE BEEN UNABLE TO RELIABLY DUPLICATE." I think
it might be OK that we do not reproduce this issue at present.

What do you think?

Thanks,
Xuelei

Thanks
Max


Xuelei

Thanks
Max

On 07/02/2012 10:39 AM, Xuelei Fan wrote:
Hi Weijun,

Would you please review the test update for CR 7180038?
         http://cr.openjdk.java.net./~xuelei/7180038/webrev.00/

We cannot reproduce the issue. However, from the test log, there
is two
possible issues exposed by this CR.
1. the improper test case senarios of un/wrap()
        In the test case, the scenarios is
unwrap()->wrap()->serverOut.remaining()->"serverIn.remaining() !=
clientMsg.length". After the wrap(), the next operation may need
to be
unwrap() to get more incoming data before comparing serverIn buffer
with
the expected client message.

         This fix is trying to do the comparing after the engine has
closed.

2. From the log, the engine status and handshaking status move from
CLOSED/NOT_HANDSHAKING to OK/FINISHED. FINISHED means the TLS
handshaking just finished. As the handshaking should have
completed for
a while, it does not sound like a correct status change.

         However, I did not find why this happens. Need more
info. So I
added
a line of log (suggested by Brad Wetmore) to collect the next
failure:

         IF THIS FAILS, PLEASE REPORT THIS TO THE SECURITY TEAM.  WE
HAVE
         BEEN UNABLE TO RELIABLY DUPLICATE.


Thanks,
Xuelei













Reply via email to