WriterReaderPhaser provides an asymmetric means for synchronizing the execution of wait-free "writer" critical sections against a "reader phase flip" that needs to make sure no writer critical sections that were active at the beginning of the flip are still active after the flip is done. Multiple writers and multiple readers are supported.

Using a WriterReaderPhaser for coordination, writers can continously perform wait-free/lock-free updates to common data structures, while readers can get hold of atomic and inactive snapshots without stalling writers.

While a WriterReaderPhaser can be useful in multiple scenarios, a specific and common use case is that of safely managing "double buffered" data stream access in which writers can proceed without being blocked, while readers gain access to stable and unchanging buffer samples. WriterReaderPhaser "writers" are wait free (on architectures that support wait free atomic increment operations), "readers" block for other "readers", and "readers" are only blocked by "writers" whose critical section was entered before the reader's WriterReaderPhaser#flipPhase() attempt.

Assumptions and Guarantees

When used to protect an actively recording data structure, the assumptions on how readers and writers act are:

1. There are two sets of data structures ("active" and "inactive")
2. Writing is done to the perceived active version (as perceived by the writer), and only within critical sections delineated by WriterReaderPhaser#writerCriticalSectionEnter and writerCriticalSectionExit().
3. Only readers switch the perceived roles of the active and inactive data structures. They do so only while under WriterReaderPhaser#readerLock() protection and only before calling WriterReaderPhaser#flipPhase().
4. Writers do not remain in their critical sections indefinitely.
5. Only writers perform WriterReaderPhaser#writerCriticalSectionEnter and writerCriticalSectionExit().
6. Readers do not hold onto readerLock indefinitely.
7. Only readers perform WriterReaderPhaser#readerLock() and WriterReaderPhaser#readerUnlock().
8. Only readers perform WriterReaderPhaser#flipPhase() operations, and only while holding the readerLock.

When the above assumptions are met, WriterReaderPhaser guarantees that the inactive data structures are not being modified by any writers while being read while under readerLock() protection after a WriterReaderPhaser#flipPhase()() operation.

The following progress guarantees are provided to writers and readers that adhere to the above stated assumptions:

1. Writers operations (writerCriticalSectionEnter and writerCriticalSectionExit) are wait free on architectures that support wait-free atomic increment operations (they remain lock-free [but not wait-free] on architectures that do not support wait-free atomic increment operations)
2. WriterReaderPhaser#flipPhase() operations are guaranteed to make forward progress, and will only be blocked by writers whose critical sections were entered prior to the start of the reader's flipPhase operation, and have not yet exited their critical sections.
3. WriterReaderPhaser#readerLock() only blocks for other readers that are holding the readerLock.

Example use

Imagine a simple use case where a histogram (which is basically a large set of rapidly updated counters) is being modified by writers, and a reader needs to gain access to stable interval samples of the histogram for reporting or other analysis purposes.

        final WriterReaderPhaser recordingPhaser = new WriterReaderPhaser();

        volatile Histogram activeHistogram;
        Histogram inactiveHistogram;
        ...

A writer may record values the histogram:

        // Wait-free recording:
        long criticalValueAtEnter = recordingPhaser.writerCriticalSectionEnter();
        try {
            activeHistogram.recordValue(value);
        } finally {
            recordingPhaser.writerCriticalSectionExit(criticalValueAtEnter);
        }

A reader gains access to a stable histogram of values recorded during an interval, and reports on it:

        try {
            recordingPhaser.readerLock();

            inactiveHistogram.reset();

            // Swap active and inactive histograms:
            final Histogram tempHistogram = inactiveHistogram;
            inactiveHistogram = activeHistogram;
            activeHistogram = tempHistogram;

            recordingPhaser.flipPhase();
            // At this point, inactiveHistogram content is guaranteed to be stable

            logHistogram(inactiveHistogram);

        } finally {
            recordingPhaser.readerUnlock();
        }