[09-Dec-2019]

[Based on discussion with the QEMU Block Layer folks; thanks Kevin
Wolf.]

Suppose we have a disk image chain:

    base.raw <-- overlay1.qcow2

To flatten the above chain, i.e. merge the contents of overlay1.qcow2
into base.raw, 'qemu-img' offers three *offline* ways:

(1) Using 'commit':

      $> qemu-img commit overlay1.qcow2

    Notes:

     - This is in-place copy.  Copies contents of overlay1 into its
       backing file, base.raw.

(2) Using 'rebase':

      $> qemu-img rebase -b base.raw overlay1.qcow2

    Notes:
    
     - There are "two modes" of 'rebase' (read the manual page for
       details), "Safe mode" (it's the default) and "Unsafe mode".  
 
     - The "Unsafe mode" (-u) only changes the backing file pointer, but
       by default `qemu-img rebase` copies the delta into the overlay.
    
(3) Using 'convert':

      $> qemu-img convert -f qcow2 -O qcow2 overlay1.qcow2 flattened.qcow2 

    Notes:

     - The difference is that 'commit' and 'rebase' are *in-place*,
       while 'convert' creates a new copy (of base + overlay1).
     
     - 'convert' also does sparsification ("the empty sectors are
       detected and suppressed from the destination image").
 
     - Because 'convert' creates a new copy, it is thus slow, and takes
       double the disk space.

When to use when?
----------------

  - If the backing file is shared or the overlay actually overwrites
    most of the blocks, 'rebase' makes more sense.

  - Use 'convert' if you can (a) accept the slowness due to the copy
    containing the contents of both 'base' and 'overlay1'; and (b) have
    no disk space constraints, because of point (a).