[Qemu-devel] [RFC 0/8] block: Live backup prototype

[Stefan Hajnoczi]

Dietmar Maurer's recent "Efficient VM backup for qemu" patch series has spawned
a lot of discussion.  I'm afraid we are close to an impasse so I decided to
prototype the approach that I'm advocating - just make sure it actually works
:).

This series implements vmstate and disk backup for running guests.  It's just a
quick hack, especially the Python code is missing error handling.  My goal is
to show by example how the 'block-backup' block job should work and how a VMA
backup archive writer can be implemented as an external program.

The code is available in my public repo:

  git://github.com/stefanha/qemu.git backup-block-job

The key feature is the new 'block-backup' QMP command.  It takes a
point-in-time copy of a block device and writes it to a target device.  This is
related to block-stream and drive-mirror but the snapshot is copied out while
the guest continues writing to the block device.

Later patches are Python scripts that use the new 'block-backup' QMP command to
implement live backup.  The scripts implement Dietmar's VMA backup archive
format in order to prove that VMA can be done with 'block-backup'.

  $ python backup.py /tmp/monitor.sock backup-20130301.vma
  Running migration...
  Running block-backup...
  Finished device "virtio-drive0"
  Backup complete, terminating writer process

Behind the scenes backup.py spawns a script called vma-writer.py.  The
vma-writer.py process receives migration vmstate and 'block-backup' data.  The
'block-backup' block jobs send data to vma-writer.py using the NBD protocol.

The difference between this approach and Dietmar's series is that the backup
archive format is implemented outside QEMU and runs as a separate program.

This way, management tools like proxmox, oVirt, OpenStack, and others can
provide their preferred backup archive formats without modifying QEMU.  This
has many advantages:

 * 'block-backup' composes with 'migration' and other commands, unlike the
   monolithic 'backup' command designed just for writing backup archives

 * Backup code can be updated or added outside the QEMU release cycle

 * Choice of language, coding style, and license for backup code

 * Less QEMU code to test and maintain

The objection to this approach has been performance.  Exporting vmstate and
disk data over UNIX domain sockets to an external process incurs IPC overhead.
This prototype shows that even Python code hacked up in a day achieves decent
performance.

I'm leaving benchmarking as an exercise for the reader.  I tested a single
scenario with a 16 GB raw disk and 1 GB RAM, backup time was 28% longer (+30
seconds) than Dietmar's series.  Below are a few starting points:

 * I moved the buffer_is_zero() check from the VMA writer into the block job.
   We now skip writing zero clusters and the file contains no extents for them.

 * Migration data uses no fixed block size, there are many small writes that
   could be optimized by batching into a buffer.

 * The block job issues 64 KB writes, it should be possible to use a larger
   buffer size like the 512 KB buffer in block/stream.c.

 * Another trick in block/mirror.c is to use asynchronous I/O so that there can
   be multiple requests pending.

Dietmar Maurer (1):
  add basic backup support to block driver

Stefan Hajnoczi (7):
  block: add virtual_size to query-block QMP output
  backup: write to BlockDriverState instead of BackupDumpFunc
  block: add block_backup QMP command
  Add nbd server Python module
  Add VMA backup archive writer Python module
  Add vma-writer.py tool
  Add backup.py tool

 Makefile.objs             |   1 +
 backup.c                  | 321 ++++++++++++++++++++++++++++++++++++++++++++++
 backup.py                 |  84 ++++++++++++
 block.c                   |  72 ++++++++++-
 blockdev.c                |  92 +++++++++++++
 include/block/block.h     |   2 +
 include/block/block_int.h |  16 +++
 include/block/blockjob.h  |  10 ++
 nbd.py                    | 124 ++++++++++++++++++
 qapi-schema.json          |  33 ++++-
 qmp-commands.hx           |   6 +
 vma-writer.py             | 126 ++++++++++++++++++
 vma.py                    | 236 ++++++++++++++++++++++++++++++++++
 13 files changed, 1116 insertions(+), 7 deletions(-)
 create mode 100644 backup.c
 create mode 100644 backup.py
 create mode 100644 nbd.py
 create mode 100644 vma-writer.py
 create mode 100644 vma.py

-- 
1.8.1.4