Parallelizing SquashFS Data Packing
                      ***********************************

 0) Overview
 ***********

 On a high level, data blocks are processed as follows:

 The "block processor" has a simple begin/append/end interface for submitting
 file data. Internally it chops the file data up into fixed size blocks that
 are each [optionally] compressed and hashed. If the "end" function is called
 and there is still left over data, a fragment is created.

 Fragments are only hashed. If another fragment exists with the same size and
 hash, it is discarded and the existing fragment is referenced. Fragments are
 collected in a fragment block that, once it overflows, is processed like a
 normal block.

 The final compressed & hashed data blocks & fragment blocks are passed on to
 the "block writer".

 The block writer simply writes blocks to the output file. Flags are used to
 communicate what the first and last block of a file are. Entire files are
 deduplicated by trying to find a sequence of identical size/hash pairs in
 the already written blocks.


 0.1) Implementation

 The implementation of the block processor is in lib/sqfs/block_processor. The
 file common.c contains the frontend for file data submission and common
 functions for processing a single block, handling a completed block and
 handling a completed fragment.

 A reference serial implementation is provided in the file serial.c


 1) Thread Pool Based Block Processor
 ************************************

 1.1) Goals and Challanges

 In addition to the goal of boosting performance, the thread pool based block
 processor must meet the following requirements:

  - Output MUST be deterministic and reproducible. I.e. feeding byte-for-byte
    the same input MUST ALWAYS produce byte-for-byte the same output.
  - Blocks the belong to a single file MUST be written in the order that
    they were submitted.
  - Output MUST be byte-for-byte equivalent to the serial reference
    implementation. Changing the serial reference implementation to
    achieve this is OK.
  - I/O cannot be done in worker threads. The underlying back-end must be
    assumed to not be thread safe and may get very confused by suddenly running
    in a different thread, even if only one thread at a time uses it.


 1.2) The Current Approach

 The current implementation is in winpthread.c (based on pthread or Windows
 native threads, depending on whats available).

 It keeps track of blocks in 3 different FIFO queues:
  - A "work queue" that freshly submitted blocks go to. Worker threads take
    blocks from this queue for processing.
  - A "done queue" that worker threads submit blocks to, once completed.
  - An "I/O queue" that contains blocks ready to be written to disk.


 When the main thread submits a block, it gives it an incremental "processing"
 sequence number and appends it to the "work queue". Thread pool workers take
 the first best block of the queue, process it and add it to the "done" queue,
 sorted by its processing sequence number.

 The main thread dequeues blocks from the done queue sorted by their processing
 sequence number, using a second counter to make sure blocks are dequeued in
 the exact same order as they were added to the work queue.

 Regular data blocks from the "done queue" are given an incremental I/O
 sequence number and added to the "I/O queue", sorted by this number.

 Fragments are deduplicated and consolidated into a fragment block. If this
 block overflows, it is appended to the "work queue" the exact same way as
 regular blocks, but it is **already given an I/O sequence number**.

 If a block dequeued from the "done queue" turns out to be a fragment block, it
 is added to the "I/O queue", sorted by its I/O sequence number **that it
 already has**, i.e. no new sequence number is allocated.

 The I/O queue is dequeued in the same fashion as the "done queue", using a
 second counter to enforce ordering.


 The actual implementation interleaves enqueueing and dequeueing in the block
 submission function. It dequeues blocks if the queues reach a pre-set maximum
 backlog. In that case, it tries to dequeue from the I/O queue first and if
 that fails, tries to dequeue from the "done queue". If that also fails, it
 uses signal/await to be woken up by a worker thread once it adds a block to
 the "done queue". Fragment post-processing and re-queueing of blocks is done
 inside the critical region, but the actual I/O is done outside (for obvious
 reasons).


 Profiling on small filesystems using perf shows that the outlined approach
 seems to perform quite well for CPU bound compressors like XZ, but doesn't
 add a lot for I/O bound compressors like zstd.

 If you have a better idea how to do this, please let me know.