Why doesn't volatile in java 5+ ensure visibility from another thread?

According to:

http://www.ibm.com/developerworks/library/j-jtp03304/

Under the new memory model, when thread A writes to a volatile variable V, and thread B reads from V, any variable values that were visible to A at the time that V was written are guaranteed now to be visible to B

And many places on the internet state that the following code should never print "error":

public class Test {
    volatile static private int a;
    static private int b;

    public static void main(String [] args) throws Exception {
        for (int i = 0; i < 100; i++) {
            new Thread() {

                @Override
                public void run() {
                    int tt = b; // makes the jvm cache the value of b

                    while (a==0) {

                    }

                    if (b == 0) {
                        System.out.println("error");
                    }
                }

            }.start();
        }

        b = 1;
        a = 1;
    }
}

b should be 1 for all the threads when a is 1.

However I sometimes get "error" printed. How is this possible?


Solution 1:

Update:

For anyone interested this bug has been addressed and fixed for Java 7u6 build b14. You can see the bug report/fixes here

  • Report
  • Changeset
  • Buglist

Original Answer

When thinking in terms of memory visibility/order you would need to think about its happens-before relationship. The important pre condition for b != 0 is for a == 1. If a != 1 then b can be either 0 or 1.

Once a thread sees a == 1 then that thread is guaranteed to see b == 1.

Post Java 5, in the OP example, once the while(a == 0) breaks out b is guaranteed to be 1

Edit:

I ran the simulation many number of times and didn't see your output.

What OS, Java version & CPU are you testing under?

I am on Windows 7, Java 1.6_24 (trying with _31)

Edit 2:

Kudos to the OP and Walter Laan - For me it only happened when I switched from 64 bit Java to 32 bit Java, on (but may not be excluded to) a 64 bit windows 7.

Edit 3:

The assignment to tt, or rather the staticget of b seems to have a significant impact (to prove this remove the int tt = b; and it should always work.

It appears the load of b into tt will store the field locally which will then be used in the if coniditonal (the reference to that value not tt). So if b == 0 is true it probably means that the local store to tt was 0 (at this point its a race to assign 1 to local tt). This seems only to be true for 32 Bit Java 1.6 & 7 with client set.

I compared the two output assembly and the immediate difference was here. (Keep in mind these are snippets).

This printed "error"

 0x021dd753: test   %eax,0x180100      ;   {poll}
  0x021dd759: cmp    $0x0,%ecx
  0x021dd75c: je     0x021dd748         ;*ifeq
                                        ; - Test$1::run@7 (line 13)
  0x021dd75e: cmp    $0x0,%edx
  0x021dd761: jne    0x021dd788         ;*ifne
                                        ; - Test$1::run@13 (line 17)
  0x021dd767: nop    
  0x021dd768: jmp    0x021dd7b8         ;   {no_reloc}
  0x021dd76d: xchg   %ax,%ax
  0x021dd770: jmp    0x021dd7d2         ; implicit exception: dispatches to 0x021dd7c2
  0x021dd775: nop                       ;*getstatic out
                                        ; - Test$1::run@16 (line 18)
  0x021dd776: cmp    (%ecx),%eax        ; implicit exception: dispatches to 0x021dd7dc
  0x021dd778: mov    $0x39239500,%edx   ;*invokevirtual println

And

This did not print "error"

0x0226d763: test   %eax,0x180100      ;   {poll}
  0x0226d769: cmp    $0x0,%edx
  0x0226d76c: je     0x0226d758         ;*ifeq
                                        ; - Test$1::run@7 (line 13)
  0x0226d76e: mov    $0x341b77f8,%edx   ;   {oop('Test')}
  0x0226d773: mov    0x154(%edx),%edx   ;*getstatic b
                                        ; - Test::access$0@0 (line 3)
                                        ; - Test$1::run@10 (line 17)
  0x0226d779: cmp    $0x0,%edx
  0x0226d77c: jne    0x0226d7a8         ;*ifne
                                        ; - Test$1::run@13 (line 17)
  0x0226d782: nopw   0x0(%eax,%eax,1)
  0x0226d788: jmp    0x0226d7ed         ;   {no_reloc}
  0x0226d78d: xchg   %ax,%ax
  0x0226d790: jmp    0x0226d807         ; implicit exception: dispatches to 0x0226d7f7
  0x0226d795: nop                       ;*getstatic out
                                        ; - Test$1::run@16 (line 18)
  0x0226d796: cmp    (%ecx),%eax        ; implicit exception: dispatches to 0x0226d811
  0x0226d798: mov    $0x39239500,%edx   ;*invokevirtual println

In this example the first entry is from a run that printed "error" while the second was from one which didnt.

It seems that the working run loaded and assigned b correctly before testing it equal to 0.

  0x0226d76e: mov    $0x341b77f8,%edx   ;   {oop('Test')}
  0x0226d773: mov    0x154(%edx),%edx   ;*getstatic b
                                        ; - Test::access$0@0 (line 3)
                                        ; - Test$1::run@10 (line 17)
  0x0226d779: cmp    $0x0,%edx
  0x0226d77c: jne    0x0226d7a8         ;*ifne
                                        ; - Test$1::run@13 (line 17)

While the run that printed "error" loaded the cached version of %edx

  0x021dd75e: cmp    $0x0,%edx
  0x021dd761: jne    0x021dd788         ;*ifne
                                        ; - Test$1::run@13 (line 17)

For those who have more experience with assembler please weigh in :)

Edit 4

Should be my last edit, as the concurrency dev's get a hand on it, I did test with and without the int tt = b; assignment some more. I found that when I increase the max from 100 to 1000 there seems to be a 100% error rate when int tt = b is included and a 0% chance when it is excluded.

Solution 2:

Based on the extract from JCiP below, I would have thought that your example should never print "error":

The visibility effects of volatile variables extend beyond the value of the volatile variable itself. When a thread A writes to a volatile variable and subsequently thread B reads that same variable, the values of all variables that were visible to A prior to writing to the volatile variable become visible to B after reading the volatile variable.

Solution 3:

You might want to check out a discussion thread on the concurrency interest mailing list on this question: http://cs.oswego.edu/pipermail/concurrency-interest/2012-May/009440.html

It seems like the problem is more easily reproduced with the client JVM (-client).