It’s important to understand that there are two aspects to thread safety.
- execution control, and
- memory visibility
The first has to do with controlling when code executes (including the order in which instructions are executed) and whether it can execute concurrently, and the second to do with when the effects in memory of what has been done are visible to other threads. Because each CPU has several levels of cache between it and main memory, threads running on different CPUs or cores can see “memory” differently at any given moment in time because threads are permitted to obtain and work on private copies of main memory.
Using synchronized
prevents any other thread from obtaining the monitor (or lock) for the same object, thereby preventing all code blocks protected by synchronization on the same object from executing concurrently. Synchronization also creates a “happens-before” memory barrier, causing a memory visibility constraint such that anything done up to the point some thread releases a lock appears to another thread subsequently acquiring the same lock to have happened before it acquired the lock. In practical terms, on current hardware, this typically causes flushing of the CPU caches when a monitor is acquired and writes to main memory when it is released, both of which are (relatively) expensive.
Using volatile
, on the other hand, forces all accesses (read or write) to the volatile variable to occur to main memory, effectively keeping the volatile variable out of CPU caches. This can be useful for some actions where it is simply required that visibility of the variable be correct and order of accesses is not important. Using volatile
also changes treatment of long
and double
to require accesses to them to be atomic; on some (older) hardware this might require locks, though not on modern 64 bit hardware. Under the new (JSR-133) memory model for Java 5+, the semantics of volatile have been strengthened to be almost as strong as synchronized with respect to memory visibility and instruction ordering (see http://www.cs.umd.edu/users/pugh/java/memoryModel/jsr-133-faq.html#volatile). For the purposes of visibility, each access to a volatile field acts like half a synchronization.
Under the new memory model, it is still true that volatile variables cannot be reordered with each other. The difference is that it is now no longer so easy to reorder normal field accesses around them. Writing to a volatile field has the same memory effect as a monitor release, and reading from a volatile field has the same memory effect as a monitor acquire. In effect, because the new memory model places stricter constraints on reordering of volatile field accesses with other field accesses, volatile or not, anything that was visible to thread
A
when it writes to volatile fieldf
becomes visible to threadB
when it readsf
.
So, now both forms of memory barrier (under the current JMM) cause an instruction re-ordering barrier which prevents the compiler or run-time from re-ordering instructions across the barrier. In the old JMM, volatile did not prevent re-ordering. This can be important, because apart from memory barriers the only limitation imposed is that, for any particular thread, the net effect of the code is the same as it would be if the instructions were executed in precisely the order in which they appear in the source.
One use of volatile is for a shared but immutable object is recreated on the fly, with many other threads taking a reference to the object at a particular point in their execution cycle. One needs the other threads to begin using the recreated object once it is published, but does not need the additional overhead of full synchronization and it’s attendant contention and cache flushing.
// Declaration
public class SharedLocation {
static public SomeObject someObject=new SomeObject(); // default object
}
// Publishing code
// Note: do not simply use SharedLocation.someObject.xxx(), since although
// someObject will be internally consistent for xxx(), a subsequent
// call to yyy() might be inconsistent with xxx() if the object was
// replaced in between calls.
SharedLocation.someObject=new SomeObject(...); // new object is published
// Using code
private String getError() {
SomeObject myCopy=SharedLocation.someObject; // gets current copy
...
int cod=myCopy.getErrorCode();
String txt=myCopy.getErrorText();
return (cod+" - "+txt);
}
// And so on, with myCopy always in a consistent state within and across calls
// Eventually we will return to the code that gets the current SomeObject.
Speaking to your read-update-write question, specifically. Consider the following unsafe code:
public void updateCounter() {
if(counter==1000) { counter=0; }
else { counter++; }
}
Now, with the updateCounter() method unsynchronized, two threads may enter it at the same time. Among the many permutations of what could happen, one is that thread-1 does the test for counter==1000 and finds it true and is then suspended. Then thread-2 does the same test and also sees it true and is suspended. Then thread-1 resumes and sets counter to 0. Then thread-2 resumes and again sets counter to 0 because it missed the update from thread-1. This can also happen even if thread switching does not occur as I have described, but simply because two different cached copies of counter were present in two different CPU cores and the threads each ran on a separate core. For that matter, one thread could have counter at one value and the other could have counter at some entirely different value just because of caching.
What’s important in this example is that the variable counter was read from main memory into cache, updated in cache and only written back to main memory at some indeterminate point later when a memory barrier occurred or when the cache memory was needed for something else. Making the counter volatile
is insufficient for thread-safety of this code, because the test for the maximum and the assignments are discrete operations, including the increment which is a set of non-atomic read+increment+write
machine instructions, something like:
MOV EAX,counter
INC EAX
MOV counter,EAX
Volatile variables are useful only when all operations performed on them are “atomic”, such as my example where a reference to a fully formed object is only read or written (and, indeed, typically it’s only written from a single point). Another example would be a volatile array reference backing a copy-on-write list, provided the array was only read by first taking a local copy of the reference to it.