On 12/17/14 10:40 AM, Dmitry Samersoff wrote:
1. Ever if you set SO_LINGER to zero, socket will not be closed
immediately. see TCP shutdown sequence.
Which TCP shutdown sequence? The one that involves FIN_WAIT_1, FIN_WAIT_2, and
TIME_WAIT? That's the state machine for a connected socket; the sockets in
question here have never been connected. There is still apparently a state the
socket goes through before the port is actually freed, but it might not have
anything to do with TCP.
If you have references for this behavior I'd appreciate them. What I've learned
about this issue is only through empirical observation.
By the way, more empiricism: I can reproduce the EADDRINUSE on Solaris a large
fraction of the time, by running multiple programs (or threads) that simply
close and reopen distinct sockets. I've also observed that setting SO_REUSEPORT
(introduced in Solaris 11) seems to avoid this problem entirely.
Unfortunately SO_REUSEPORT isn't available from Java, it doesn't exist on all
systems, and I don't know if it would have this same behavior on other systems
on which it does exist.
2. In a native world it's quite easy to find the port your rmi server
uses - it could be achieved by parsing /proc/<pid>/net/tcp on Linux or
using special API on windows and solaris.
I think your definition of "easy" differs from mine. :-)
The context is the jdk regression tests, does support native code. I'd prefer
not to have to explore this new area, especially in addition to writing a bunch
of system-specific native code.
3. For rmid and port provided from outside I think the only reliable way
to get what you need is: [...]
I think it's pretty clear at this point that the open-close-reopen approach
can't be made reliable in any platform-independent way. For rmid I think I might
create a new mode that opens an ephemeral port and sends that to the test driver
somehow. Looks like Jaroslav is proceeding with a retry strategy.
4. For rmid you can also emulate inetd behaviour - i.e. driver open a
server port, communicate it to client than redirect everything that come
to this port to stdin of rmid.
Thanks, but unfortunately this is actually one of the modes that needs to be
tested in rmid. It has a mode where it opens and listens on its own socket, and
another mode where it inherits one from its parent process, so that it can be
invoked from inetd.
s'marks
-Dmitry
On 2014-12-17 01:55, Stuart Marks wrote:
Hi Dmitry,
Strictly speaking you are correct. As soon as you close a socket, there
is a possibility -- perhaps vanishingly small but nonzero -- that you
might not be able to open it again.
The first scenario, where the user of the socket itself opens the socket
using an ephemeral port (e.g. new ServerSocket(0)) is of course
preferred. This avoids race conditions entirely.
It's the second case that I'm still wrestling with, and maybe Jaroslav
too. It's fairly difficult to get such "black box" systems to open an
ephemeral port and report it back, as opposed to opening up their
service on some port number handed in from the outside. (For RMI, rmid
is the culprit here. I don't know about JMX.) What makes this difficult
is that the rmid service is running in a separate VM, so getting
reliable information back from it can be difficult.
It's also fairly difficult to establish the retry logic in such cases.
If the service fails with a BindException, maybe -- maybe -- it was
because there was a conflict over the port, and a retry is warranted.
But this needs to be distinguished from other failure modes that might
occur, that should be reported as failures instead of causing a retry.
In principle, this is possible to do, of course, it's just that it
involves more restructuring of the tests, and possibly adding debug/test
code to rmid. (It may yet come to that.)
I'm still pondering the reasons that, in the open/close/reopen scenario,
why the reopen might fail. The obvious reason is that some other process
on the system has opened that port between the close and the reopen. I
admit that this is a possibility. However, with the open/close/reopen
scenario in place, we see tests that fail up to 15% of the time with
BindExceptions. This is an extraordinarily high failure rate to be
caused by some random other process happening to open the same port in
the few microseconds between the close and reopen. It's simply not
believable to me.
My thinking is still that the port isn't ready for reuse until a small
amount of time after it's closed. I have some test programs that
exercise sockets in a particular way (e.g., from multiple threads, or
opening and closing batches of sockets) that can reproduce the problem
on some systems, and these test programs seem to behave better if a time
delay is added between the close and the reopen. The exact circumstances
under which the problem occurs is difficult to pin down and seems OS
specific, and so choosing the "right" delay time is very difficult. But
it does strengthen this conjecture in my mind.
Naturally it would be better if there were a way to determine when a
port is available for reuse without actually opening it. I'm not aware
of any such way, but I'm holding onto a little hope that one can be found.
s'marks
On 12/11/14 10:18 AM, Dmitry Samersoff wrote:
Stuart,
As soon as you close socket, you open a door for the race.
So you need another communication channel to pass a port number (or bind
result) between a client and a server without closing a socket on the
server side.
Typical scenario used by network related code is:
1. Server opens the socket
2. Server binds to port(0)
3. Server gets port number assigned by OS
4. Server informs client (e.g. write the port down to known file,
broadcast it etc)
5. Client establishes connection.
If the server is a blackbox and have to get a port number from outside,
scenario looks like:
WHILE(!success and !timeout)
1. Driver chooses random port number
2. Driver runs a server with this number
3. Driver checks that server is actually listening on this port
(e.g. try to connect by it self)
WEND
4. Driver runs a client with this port number or bails out with
descriptive error message.
-Dmitry
On 2014-12-11 20:53, Stuart Marks wrote:
On 12/11/14 7:09 AM, olivier.lagn...@oracle.com wrote:
On 11/12/2014 15:43, Dmitry Samersoff wrote:
You can set SO_LINGER to zero, in this case socket will be closed
immediately without waiting in TIME_WAIT
SO-LINGER did not help either in my case (see my previous mail to
Jaroslav).
That ended-up in using another hard-coded (supposedly free) port.
Note that was before RMI tests used randomly allocated ports.
But there are no reliable way to predict whether you can take this
port
or not after you close it.
This is what I observed in my case.
So the only valid solution is to try to connect to a random port
and if
this attempt fails try another random port. Everything else will cause
more or less frequent intermittent failures.
IIRC think this is what is currently done in RMI tests.
The RMI tests are still suffering from this problem, unfortunately.
The RMI test library gets a "random" port with "new ServerSocket(0)",
gets the port number, closes the socket, then returns the port to the
caller. The caller then assumes that it can use that port as it wishes.
That's when the BindException can occur. There are about 10 RMI test
bugs in the database that all seem to have this as their root cause.
There is some retry logic in RMI's test library, but that's to avoid the
so-called "reserved ports" that specific RMI tests use, or if "new
ServerSocket(0)" fails. It doesn't have anything to do with the
BindException that occurs when the caller attempts to reuse the port
with another socket.
My observation is also that setting SO_REUSEADDR has no effect. I
haven't tried SO_LINGER. My hunch is that it won't have any effect,
since the sockets in question aren't actually going into TIME_WAIT
state. But I suppose it's worth a try.
I don't have any solution for this; we're still discussing the issue. I
think the best approach would be to refactor the code so that the
eventual user of the socket opens it up on an ephemeral port in the
first place. That avoids the open/close/reopen business. Unfortunately
that doesn't help the case where you want to tell another JVM to run a
service on a specific port. We don't have a solution for that case yet.
The second-best approach (not really a solution) is to open/close a
serversocket to get the port, sleep for a little bit, then return the
port number to the caller. This might give the kernel a chance to clean
up the socket after the close. Of course, this still has a race
condition, but it might reduce the incidence of problems to an
acceptable level.
I'll let you know if we come up with anything better.
s'marks