Even if the volatile keyword guarantees that all reads of a volatile variable are read directly from main memory, and all writes to a volatile variable are written directly to main memory, there are still situations where it is not enough to declare a variable volatile.
In fact, multiple threads could even be writing to a shared volatile variable, and still have the correct value stored in main memory, if the new value written to the variable does not depend on its previous value. In other words, if a thread writing a value to the shared volatile variable does not first need to read its value to figure out its next value.
As soon as a thread needs to first read the value of a volatile variable, and based on that value generate a new value for the shared volatile variable, a volatile variable is no longer enough to guarantee correct visibility. The short time gap in between the reading of the volatile variable and the writing of its new value creates a race condition where multiple threads might read the same value of the volatile variable, generate a new value for the variable, and when writing the value back to main memory - overwrite each other's values.
The situation where multiple threads are incrementing the same counter is exactly such a situation where a volatile variable is not enough.
Explain with an example:
Imagine if Thread 1 reads a shared counter variable with the value 0 into its CPU cache, increment it to 1 and not write the changed value back into main memory. Thread 2 could then read the same counter variable from main memory where the value of the variable is still 0, into its own CPU cache. Thread 2 could then also increment the counter to 1, and also not write it back to main memory.
Thread 1 and Thread 2 are now practically out of sync. The real value of the shared counter variable should have been 2, but each of the threads has the value 1 for the variable in their CPU caches, and in main memory, the value is still 0. Even if the threads eventually write their value for the shared counter variable back to main memory, the value will be wrong.
When is volatile enough?
As I have mentioned earlier, if two threads are both reading and writing to a shared variable, then using the volatile keyword for that is not enough. You need to use a synchronized in that case to guarantee that the reading and writing of the variable is atomic. Reading or writing a volatile variable does not block threads reading or writing. For this to happen you must use the synchronized keyword around critical sections.
As an alternative to a synchronized block, we can use one of the many atomic data types found in the java.util.concurrent package (=AtomicLong, AtomicReference or one of the others).
In case only one thread reads and writes the value of a volatile variable and other threads only read the variable, then the reading threads are guaranteed to see the latest value written to the volatile variable. Without making the variable volatile, this would not be guaranteed.
The volatile keyword is guaranteed to work on 32 bit and 64 variables.