harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Leo Li" <liyilei1...@gmail.com>
Subject Re: [classlib][xnet] Problem using SSLSocketImpl behind load balancer
Date Fri, 13 Apr 2007 02:57:56 GMT
Hi, Gerald:
     The harmony build system adopts an incremental compiling, thus if the
InetAddress is removed, the build will still succeed since it regards that
no new/modification has occurred. Actually the deploy still contains the
InetAddress class if you look at the luni.jar built from it. But if you
clean the system, build system will complain about compiling errors. The
class is needed by several core classes, such as socket.
     So if you do not need the DNS caching, the solution is simple: Just
remove the caching mechanism in the InetAddress class.
     If you would like, you can try the patch in attachment, which can be
applied on the luni module at revision 528316. (It is just a customization
on classlib for your case, not for Harmony classlib itself.)
     Besides, you can also remove the return NagativeCache if necessary in
your case.


On 4/13/07, Gerald Jerome <gerald.jerome@verizonbusiness.com> wrote:
>
> Hi Leo, the first two clarifications appear correct, but I'm not sure DNS
> caching is important as all our connections are created up front when the
> app starts and held in a pool (vector) - the Broker servers up connections
> out of a connection pool as they are requested.  If a connection is
> unusable
> say when the host fails or is taken down, the app attempts to recreate
> another SSLSocket connection.  It has been tested and found to be
> reasonably
> fault tolerant, except we have this problem in production where a load
> balancing or redirector exists between the client(s) and hosts.
> Interestingly, I deleted the InetAddress class from the luni library and
> noticed no compilation errors and the app connects and is able to
> communicate just fine.  So it appears either the caching never takes place
> or only if InetAddress is available, I don't know.
>
> Regards,
>
> Gerald Jerome
>
>
>
> -----Original Message-----
> From: Leo Li [mailto:liyilei1979@gmail.com]
> Sent: Tuesday, April 10, 2007 11:05 PM
> To: dev@harmony.apache.org
> Subject: Re: [classlib][xnet] Problem using SSLSocketImpl behind load
> balancer
>
> Hi, Gerald:
>    First I would like to clarify your problem:
>    1. We want to exchange the request to a new server seamlessly according
> to workload balancing.
>    2. The behavior of DNS caching prevent us from that.
>    3. You would like to get the support from classlib, that is to enjoy
> the
> DNS caching as well as the response to server switching.
>
>    Since Harmony would like to be a compatible classlib with RI's, we have
> to comply with RI's behavior with RI's.
>    But it is possible to customize harmony's classlib if you conform with
> apache license. So my suggestion is here:
>    (1) Expose the method which clears the DNS cache in InetAddress to
> public.(If you would like to maintain the interface of java public API,
> wrap
> it in an implementation class and expose it as public.)
>    (2) On the socket side, if you use your socket implementation, in the
> connection method, if the connection fails, clear the DNS cache by the
> method exposed before and retry it.
>    If you are interested, I think, we might talk about it in detail.
>
>
> On 4/11/07, Tim Ellison <t.p.ellison@gmail.com> wrote:
> >
> > Gerald Jerome wrote:
> > > Hello, we are using the Apache Harmony SSLSocket classes to solve a
> > problem
> > > we were having with SSL renegotiation.  However, recently our
> production
> > > admin noticed that our SSL client was not automatically failing over
> to
> > a
> > > secondary machine that exists behind a load balancer or redistributor
> in
> >
> > > cases where a server goes down (either unexpectedly or for
> > maintenance).  We
> > > are using the Sun JVM and not the Apache Harmony version.  He mentions
> > the
> > > following:
> >
> > That sounds like an interesting combination of code, and not something
> > that I expect can be easily reproduced by the Harmony developers, so
> > take the following information only as guidance -- I don't know what you
> > are actually running with...
> >
> > > Our investigation found that once Java based clients (both standalone
> > > applications, and servlets) have performed the first network access (
> i.e
> > .
> > > urlconnection, parsing of xml document with external references, etc)
> > they
> > > cache DNS settings, so any subsequent client request will use its old
> > DNS
> > > information even if the real DNS settings have changed.
> > >
> > > To reset everything, you have to restart the client application since
> > the
> > > default JVM setting is to cache forever.
> > >
> > > The InetAddress class has a cache to store successful as well as
> > > unsuccessful host name resolutions. The positive caching is there to
> > guard
> > > against DNS spoofing attacks.
> >
> > Correct, the scenario you describe is almost exactly as described here:
> > http://www.rgagnon.com/javadetails/java-0445.html
> >
> > > He goes on to discuss how the caching can be disabled.  I know your
> > > SSLSocket implementation uses SSLEngine and does low-level socket
> based
> > > communication so I did not think his analysis may fit our situation.
> > > Furthermore, I am not convinced that we are having this problem but
> our
> > > development and test environments do not have a distributor/load
> > balancer in
> > > front of the actual host machines.
> >
> > Hmm, this could really be a shot in the dark then!
> >
> > > I know in production we are configured
> > > to connect to the distributor, not one of the actual hosts.  I am
> > wondering
> > > if you were aware of any caching of DNS information that may be going
> on
> > > inside the SSLSocket class and dependant code that we are using
> > using?  I
> > > could not find any references to the InetAddress class mentioned above
> > in
> > > any of the Harmony source I have.
> >
> > If you have the luni.jar then you should have the java.net.InetAddress
> > class (and java.net.NegativeCache class).  They will honour the
> > networkaddress.cache.ttl and networkaddress.cache.negative.ttlproperties.
> >
> > > The x-net.jar file that we are using has
> > > a last modified date of October 28, 2006 8:27:58 PM. The last modified
> > date
> > > of luni.jar is the same.  These are the only two Apache Harmony
> > libraries we
> > > are using.
> >
> > Those sound quite old now, although I don't recall significant
> > development in x-net since then.
> >
> > > Any information you have pertaining to this problem is greatly
> > > appreciated
> >
> > Maybe somebody else will have some insight, but it is really hard to
> > advise on a problem you are not sure you are having on a runtime
> > codebase we don't know.
> >
> > Regards,
> > Tim
> >
> >
>
>
> --
> Leo Li
> China Software Development Lab, IBM
>
>
>


-- 
Leo Li
China Software Development Lab, IBM

Mime
  • Unnamed multipart/mixed (inline, None, 0 bytes)
View raw message