[concurrency-interest] Stricter read ordering

Roman Elizarov elizarov at devexperts.com
Wed Apr 23 14:34:39 EDT 2014


There was no light in that discussion. CAS is currently spec'd as volatile load and volatile store.  Period. If someone implements CAS on ARM with weaker guarantees that violate this spec, then StampedLock in Java 8 solution also becomes broken for the same reason you've outlined below.

On 23 апр. 2014 г., at 21:42, "Oleksandr Otenko" <oleksandr.otenko at oracle.com<mailto:oleksandr.otenko at oracle.com>> wrote:

In light of the recent discussion of the semantics of CAS, this solution is not correct.

The outcome of that discussion is that the atomicity of CAS is only guaranteed with respect to the variable being CASed. This does not matter on x86, but on ARM, as I understand, the CAS will be implemented as two instructions that only preclude concurrent stores to the variable being CASed - which does not preclude reordering the normal stores following CAS with the "volatile store" part of CAS.

So, here:


        version.compareAndSet(v0, v0 | WRITE); // always succeeds in single-writer case, ensures proper HB edges for JMM
        // then write protected data (non-volatile writes)
        this.x = x;
        this.y = y;

this.x=x; may be observed before compareAndSet completes.

The correct solution should produce:

writer:
vwrite(v,v+1)
vread(s,_)
write(x,_)
write(y,_)
vwrite(v,v+1)

reader:
vread(v,_)
read(x,_)
read(y,_)
vwrite(s,_)
vread(v,_)

For example:

writer:
// synchronized
long next = this.version;
this.version=next+1; // odd version means mutation underway
long tmp=static_volatile_dummy;
this.x=x;
this.y=y;
this.version=next+2; // even version means immutable

reader:
long v;
do{
  while((v=this.version) &1 != 0); // wait for version to become even
  x=this.x;
  y=this.y;
  static_volatile_dummy=v;
}while(this.version != v); // check the version didn't change


Proof:

It is not necessary to consider all possible reorderings individually. It is sufficient to consider relative ordering of volatile reads and writes.

All writers ensure exclusive access to this.x and this.y, so we only need to consider one writer performing modifications in a loop.

If volatile read of static_volatile_dummy appears after a volatile write of static_volatile_dummy in synchronization order, the former synchronizes-with the latter, and the normal reads of this.x and this.y by the reader happen-before the normal writes of those variables by the writer (transitive closure of program orders). So, the readers never see normal writes of those writers.

Now consider the writers whose read of static_volatile_dummy does not synchronize-with the writes. The reader may temporarily observe some values, but will only terminate the loop after observing the values of this.x and this.y whose write appear before the volatile write of a even version: the first loop ensures we only look at writes preceding a write of a even version, the condition of the outer loop ensures no other volatile writes to version appear in synchronization order. This means that the first loop synchronizes-with the last write of a even version, hence, through transitive closure of program orders the normal writes of this.x and this.y happen-before the normal reads. The outer condition ensures that since no other volatile writes of version appear in synchronization order, the volatile read of static_volatile_dummy after all such writes will synchronize-with the volatile write of static_volatile_dummy of the reader (and their normal writes won't be observed, as per the proof in the previous paragraph).


This concludes the proof that the reader always observes the consistent view of this.x and this.y.



Alex


On 23/04/2014 15:16, Roman Elizarov wrote:

The original problem itself is solvable, albeit with a different algorithm. It's funny, that we've discussed the same problem today internally and then I see this thread on concurrency-interest list.

Let me state the problem first. There is a multi-word state that we have to read and write in non-volatile way (if reads and writes of our state can be made volatile, then solution is trivial and is not of a much interest). We need to implement single-writer multi-reader atomic snapshot of this multi-word state with the primitives described in and under the constraints of the Java Memory Model (think Java 5+). The read does not have to be lock-free. It just needs to detect that snapshot being read is not consistent. In original question author was Ok for reader to spin in case of the read/write conflict, but we actually have a different problem where it is Ok for reader to abandon the read attempt altogether. The key is to detect this conflict or to return an atomic snapshot of the protected state if there is no read/write conflict.

So, here is a solution. It can be proven to work on current version of JMM (starting Java 5 and later) in the JMM's model, by analyzing possible executions and their corresponding SO, SW, and HB relations. It requires just one extra int of state (instead of two), but this int needs to be CAS-ed both on read and on write to ensure proper happens-before edges and to guarantee the consistent read of non-volatile state. The proof of this algorithm's correctness is left as an exercise for the reader. It will not scale, though, if there are many concurrent readers, because of the CAS in read path.

class VersionedData {
    // bit that we'll use to indicate that the state is being written to
    private static final int WRITE = 1 << 31;
    // we need to CAS version (in practise we'll do it via Unsafe to avoid extra object)
    private final AtomicInteger version = new AtomicInteger();

    // this is the data I protect, in reality there is much more protected state
    private int x, y;

    public synchronized void update(int x, int y) {
        // I guarantee single writers to this method,
        // illustrated with 'synchronized' in this simplification
        // first use CAS to mark version as being written to
        int v0 = version.get(); // can be non-volatile read, but then the following CAS can fail and needs to retry
        version.compareAndSet(v0, v0 | WRITE); // always succeeds in single-writer case, ensures proper HB edges for JMM
        // then write protected data (non-volatile writes)
        this.x = x;
        this.y = y;
        // then increment version and reset write bit
        version.set((v0 + 1) & ~WRITE);
    }

    public DataCarrier read() {
        // I allow multiple readers, so this method is not synchronized
        int x, y;
        int v0;
        do {
            // first read version
            v0 = version.get();
            if ((v0 & WRITE) != 0)
                continue; // immediately abort, because write in progress was detected
            // then read protected data
            x = this.x;
            y = this.y;
            // use CAS to check that version is still the same and to ensure proper HB edges for JMM at the same time
        } while (!version.compareAndSet(v0, v0));
        return new DataCarrier(x, y);
    }
}

-----Original Message-----
From: concurrency-interest-bounces at cs.oswego.edu<mailto:concurrency-interest-bounces at cs.oswego.edu> [mailto:concurrency-interest-bounces at cs.oswego.edu] On Behalf Of Aleksey Shipilev
Sent: Wednesday, April 23, 2014 5:36 PM
To: Tobias Lindaaker; concurrency-interest at cs.oswego.edu<mailto:concurrency-interest at cs.oswego.edu>
Subject: Re: [concurrency-interest] Stricter read ordering

On 04/23/2014 05:05 PM, Tobias Lindaaker wrote:


Yes, I had a look at StampedLock and Unsafe.loadFence(), and it seems
to do exactly what I want, and if I was fortunate enough to be able to
move to Java 8 I would use it. Unfortunately we are still stuck on
Java 7. We even have customers who are still strongly requesting Java
6 compatibility.


These constraints make the problem unresolvable.

You might want to look for pre-JDK8 prototype for StampedLock [1]:

     * As noted in Boehm's paper (above), sequence validation (mainly
     * method validate()) requires stricter ordering rules than apply
     * to normal volatile reads (of "state").  In the absence of (but
     * continual hope for) explicit JVM support of intrinsics with
     * double-sided reordering prohibition, or corresponding fence
     * intrinsics, we for now uncomfortably rely on the fact that the
     * Unsafe.getXVolatile intrinsic must have this property
     * (syntactic volatile reads do not) for internal purposes anyway,
     * even though it is not documented.

    public boolean validate(long stamp) {
        return (stamp & SBITS) == (U.getLongVolatile(this, STATE) & SBITS);
    }

But if Unsafe.loadFence() is risky since it is not specified (yet) JMM-wise, and so interactions with other volatile ops and fences is just undocumented... then using Unsafe.getXVolatile is double-risky because the behavioral effect of read ordering is *REALLY* implementation-specific, and you if are using it for read ordering, you are five miles past the gateway to Hell already.

Thanks,
-Aleksey.

[1]
http://gee.cs.oswego.edu/cgi-bin/viewcvs.cgi/jsr166/src/jsr166e/StampedLock.java?revision=1.1
_______________________________________________
Concurrency-interest mailing list
Concurrency-interest at cs.oswego.edu<mailto:Concurrency-interest at cs.oswego.edu>
http://cs.oswego.edu/mailman/listinfo/concurrency-interest

_______________________________________________
Concurrency-interest mailing list
Concurrency-interest at cs.oswego.edu<mailto:Concurrency-interest at cs.oswego.edu>
http://cs.oswego.edu/mailman/listinfo/concurrency-interest


_______________________________________________
Concurrency-interest mailing list
Concurrency-interest at cs.oswego.edu<mailto:Concurrency-interest at cs.oswego.edu>
http://cs.oswego.edu/mailman/listinfo/concurrency-interest
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://cs.oswego.edu/pipermail/concurrency-interest/attachments/20140423/4e1d7e9a/attachment-0001.html>


More information about the Concurrency-interest mailing list