Hibernate Search cannot guarantee that threads using the SearchFActory have been started after the thread initializing the SearchFactory (especially Containers can do stupid things).
So you need to do someting like that:
SF.init() {
...
myvolatile++; //write
}
and upon access to the SearchFactory
//note the thread local variable is per search factory instance
if ( searchFactory.getThreadLocalCheck().get() == null) {
//note the thread local variable is per search factory instance
boolean checked = myVolatile != 0;
searchFactory.getThreadLocalCheck().put(checked);
}
If I understand correctly, this ensures that any threat Tn will see what Ti (the initialization thread) has written by acquiring the "lock" only once. (this is the formalization of what you proposed in Zurich). Someone corrects me if I'm wrong.
But you scare me with "In the readeProviders and DPProviders I can't get this guarantee". What do you mean? initialize and start are always called before the SearchFactory initialization ends.
I don't see the code being that scary, the sync issue is dealt by the framework without exposing the user to the "piggy" details.
On Jul 4, 2008, at 18:36, Sanne Grinovero wrote:
Thanks for your blog Pavitar;
I would like to add some clarification about the "piggyback" just to confirm:
there is no such concept as "THE shared memory" in the JMM, shared memory is about memory shared between some threads, not necessarily all.
the "piggyback" trick works basing on this: after ThreadB reads a volatile variable, it is guaranteed to see at least ALL state TrheadA had written before (and during)
a write to THAT SAME volatile variable. Note the "at least" wording: more changes could happen to the other variables after the write to the volatile, and some (in no order,
especially not the code order) could be also seen, but no guarantee about.
so you could:
write field1, field2, field3 and then the volatile field4 in T1
read volatile field4, field1, field2, field3 in T2 (after T1 did)
and you will get a guarantee that T2 will "see" at least the state written by T1.
So this is a "trick" to avoid longer locks or having to convert them all to volatile, but IMHO
the code is made difficult to maintain, and tricky to get it right.
In our practical case:
you could write to some volatile field in the SearchFactoryImpl after the initialization is done,
but then you still have to ensure all subsequent uses will read the same field before anything else;
this has a minimal impact on performance, the good think about the "piggyback" is you
could use a read to a volatile you would have anyway.
I don't think this is at all needed for the SearchFactoryImpl as long as you guarantee that
the threads going to use it are started ("start()") after and by the initialization thread;
if this is correct no further discussion is needed there.
In the readeProviders and DPProviders I can't get this guarantee, that's why they need a fix.
Shall I use this trick then? It isn't so bad if you think it's accepatable to use it, it's good for performance
but I dislike it for code readability; I'll add a big fat scary warning.
IMHO this should be avoided when possible, especially since "final" works fine and is
very explicit to another code reader.
If you think I should go for the pig, I would appreciate if Pavitar could read the code after I commit it
to check my code, even if this case is trivial.
Sanne
2008/7/4 Emmanuel Bernard <
emmanuel@hibernate.org>:
Hey,
Can you tell me more about the piggyback synchronization. I could not find any decent knowledge online.
how far reading a volatile guarantee that all "local" values of the thread we are reading from will be pushed to the shared memory?
For example, could reading a volatile value after HSearch is done with initialization (all init is done in a single thread) guarantee that all states held by this thead will be pushed back to the shared memory?
The use case is quite specific, I init everything in a single thread, want to push all the state to the shared memory. I know post init() use of HSearch will never change the state so I don't "need" locking.
On Jul 4, 2008, at 04:33, Sanne Grinovero wrote:
Hi Pavitar Singh,
I thank you very much about your explanations but actually I opened the issue myself
because I have read the same specs and am aware of that.
in H.Search (and several other hibernate code) there's this quite common pattern for starting
"replaceable" objects (something like user-written plugins, you can provide your own implementation
to do some stuff) but this same pattern is also used to start the built-in default strategies.
It looks like this:
- an empty constructor, to use class.newInstance();
- an initialize() to set configuration options
- a start() method (eventually) used to start background tasks
- some doStuff() and/or getXX() which need to be fast & threadsafe
As you can see in Concurrecy in Practice at page 50, this is BAD, as for
example in the FSSlaveDirectoryProvider nobody takes care of locking
or visibility, and nobody is doing anywhere where I see this pattern used
(several times in the project).
I'm not saying it is all broken, because usually the threads consuming
these unsafely-initialized objects are started after the initialization, so
that's ok. In this specific case the state will be used to communicate
between threads, so some visibility fix is needed.
I know I could use it only for final fields, but this is exactly what I want:
there are currently 10 instance variables, of these
4 have no concurrent use
4 are configuration constants and could use the "final" (they're not safely published)
2 would need some lock/volatile anyway, bot only one is used frequently, so IMHO 1 volatile is ok.
I was thinking in using the same Piggyback technique you mention to
safely publish the initialization constants,
but I'm afraid the code will become more difficult to maintain and more
"unreadable", possibly breaking at the first next patch:
IMHO using some unchanging fields "final" is the most clean and
readable solution (and best performing), but I need a different
constructor for that.
your opinion is really welcome as I couldn't find other feedback,
if you would like to take a look at the sources download the Search trunk
and look at:
org.hibernate.search.store.FSSlaveDirectoryProvider
or the FIXME in
org.hibernate.search.reader.SharingBufferReaderProvider
Sanne
2008/7/4 Pavitar Singh <
pavitar.singh@gmail.com>:
Hi Sanne,
I don't think moving everything in constructor can guarantee safe publication.
From the JMM Specification Section 3.5
"An object is considered to be completely initialized when its constructor finishes. A thread that can only see a reference to an object after that object has been completely initialized is guaranteed to see the correctly initialized values for that object's final fields."
That means there are no guarantees for other fields which are non-final.
But once things are moved in constructor then by using Safe Publication one can enforce visibility guarantees.
Other thing i was wondering was why do one need to make every field as volatile. As just by using a single volatile variable one can enforce memory barriers. A technique documented in JCIP as Piggyback Synchronizations and used by Doug Lea in implementing ConcurrentHashMap.(You will find get method in ConcurrentHashMap is without any locking and visibility is enforced by use of a single volatile variable.)
Also if you can elaborate more on how you are facing the visibility issue then may be i can also spend time on it on figuring performant solution.
Regards
Pavitar Singh
http://pitfalls.wordpress.com
Hello all,
I'm sorry I've been very busy but as promised I would like to fix HSEARCH-189
(and others) very soon;
I would like to propose a API extension (backwards compatible) that would simplify the patch a lot:
keeping it as is it is very tricky to fix the visibility issues in FSSlaveDirectoryProvider
and FSMasterDirectoryProvider without introducing a performance penalty.
I have these options to close the issue:
1) add a "volatile" to more than six fields per class (ugly and not good for performance)
2) use some Locks/synch (more readable, still performance hits)
3) move the "initialize" arguments to a constructor.
As Emmanuel knows I would really love the third option, but he's worried about
the fact we can't force a constructor in an interface*1, so my proposal is:
if we find there exists a constructor having the same arguments as the initialize method,
we use that, otherwise we use a no-arg constructor and then call the initialize.
Reflection is used anyway to instantiate these components,
the code in DirectoryProviderFactory doesn't get much more complicated
and much more cleanup is made possible in all DPs because of this
(as the equals/hashcode comments already ask for).
I actually think this same pattern is needed for other components,
such as all ReaderProvider, so I hope you'll want to give it a try
and let me apply it on other components too if you like the resulting code.
Sanne
hibernate-dev mailing list
hibernate-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/hibernate-dev
--
Pavitar Singh
Blog: http://pitfalls.wordpress.com
_______________________________________________
hibernate-dev mailing list
hibernate-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/hibernate-dev