On Fri, May 20, 2016 at 03:36:10PM +0800, Changlong Xie wrote:
> Block replication is a very important feature which is used for
> continuous checkpoints(for example: COLO).
> 
> You can get the detailed information about block replication from here:
> http://wiki.qemu.org/Features/BlockReplication
> 
> Usage:
> Please refer to docs/block-replication.txt
> 
> You can get the patch here:
> https://github.com/Pating/qemu/tree/changlox/block-replication-v19
> 
> You can get the patch with framework here:
> https://github.com/Pating/qemu/tree/changlox/colo_framework_v18
> 
> TODO:
> 1. Continuous block replication. It will be started after basic functions
>    are accepted.
> 
> Changs Log:
> V19:
> 1. Rebase to v2.6.0
> 2. Address comments from stefan
> p3: a new patch that export interfaces for extra serialization
> p8: 
> 1. call replication_stop() before freeing s->top_id
> 2. check top_bs
> 3. reopen file readonly in error return paths
> 4. enable extra serialization between read and COW
> p9: try to hanlde SIGABRT
> V18:
> p6: add local_err in all replication callbacks to prevent "errp == NULL"
> p7: add missing qemu_iovec_destroy(xxx)
> V17:
> 1. Rebase to the lastest codes 
> p2: refactor backup_do_checkpoint addressed comments from Jeff Cody
> p4: fix bugs in "drive_add buddy xxx" hmp commands
> p6: add "since: 2.7"
> p7: fix bug in replication_close(), add missing "qapi/error.h", add test-replication 
> p8: add "since: 2.7"
> V16:
> 1. Rebase to the newest codes
> 2. Address comments from Stefan & hailiang
> p3: we don't need this patch now
> p4: add "top-id" parameters for secondary
> p6: fix NULL pointer in replication callbacks, remove unnecessary typedefs, 
> add doc comments that explain the semantics of Replication
> p7: Refactor AioContext for thread-safe, remove unnecessary get_top_bs()
> *Note*: I'm working on replication testcase now, will send out in V17
> V15:
> 1. Rebase to the newest codes
> 2. Fix typos and coding style addresed Eric's comments
> 3. Address Stefan's comments
>    1) Make backup_do_checkpoint public, drop the changes on BlockJobDriver
>    2) Update the message and description for [PATCH 4/9]
>    3) Make replication_(start/stop/do_checkpoint)_all as global interfaces
>    4) Introduce AioContext lock to protect start/stop/do_checkpoint callbacks
>    5) Use BdrvChild instead of holding on to BlockDriverState * pointers
> 4. Clear BDRV_O_INACTIVE for hidden disk's open_flags since commit 09e0c771  
> 5. Introduce replication_get_error_all to check replication status
> 6. Remove useless discard interface
> V14:
> 1. Implement auto complete active commit
> 2. Implement active commit block job for replication.c
> 3. Address the comments from Stefan, add replication-specific API and data
>    structure, also remove old block layer APIs
> V13:
> 1. Rebase to the newest codes
> 2. Remove redundant marcos and semicolon in replication.c 
> 3. Fix typos in block-replication.txt
> V12:
> 1. Rebase to the newest codes
> 2. Use backing reference to replcace 'allow-write-backing-file'
> V11:
> 1. Reopen the backing file when starting blcok replication if it is not
>    opened in R/W mode
> 2. Unblock BLOCK_OP_TYPE_BACKUP_SOURCE and BLOCK_OP_TYPE_BACKUP_TARGET
>    when opening backing file
> 3. Block the top BDS so there is only one block job for the top BDS and
>    its backing chain.
> V10:
> 1. Use blockdev-remove-medium and blockdev-insert-medium to replace backing
>    reference.
> 2. Address the comments from Eric Blake
> V9:
> 1. Update the error messages
> 2. Rebase to the newest qemu
> 3. Split child add/delete support. These patches are sent in another patchset.
> V8:
> 1. Address Alberto Garcia's comments
> V7:
> 1. Implement adding/removing quorum child. Remove the option non-connect.
> 2. Simplify the backing refrence option according to Stefan Hajnoczi's suggestion
> V6:
> 1. Rebase to the newest qemu.
> V5:
> 1. Address the comments from Gong Lei
> 2. Speed the failover up. The secondary vm can take over very quickly even
>    if there are too many I/O requests.
> V4:
> 1. Introduce a new driver replication to avoid touch nbd and qcow2.
> V3:
> 1: use error_setg() instead of error_set()
> 2. Add a new block job API
> 3. Active disk, hidden disk and nbd target uses the same AioContext
> 4. Add a testcase to test new hbitmap API
> V2:
> 1. Redesign the secondary qemu(use image-fleecing)
> 2. Use Error objects to return error message
> 3. Address the comments from Max Reitz and Eric Blake
> 
> Changlong Xie (3):
>   Backup: export interfaces for extra serialization
>   Introduce new APIs to do replication operation
>   tests: add unit test case for replication
> 
> Wen Congyang (7):
>   unblock backup operations in backing file
>   Backup: clear all bitmap when doing block checkpoint
>   Link backup into block core
>   docs: block replication's description
>   auto complete active commit
>   Implement new driver for block replication
>   support replication driver in blockdev-add
> 
>  Makefile.objs                |   1 +
>  block.c                      |  17 ++
>  block/Makefile.objs          |   3 +-
>  block/backup.c               |  59 +++-
>  block/mirror.c               |  13 +-
>  block/replication.c          | 666 +++++++++++++++++++++++++++++++++++++++++++
>  blockdev.c                   |   2 +-
>  docs/block-replication.txt   | 239 ++++++++++++++++
>  include/block/block_backup.h |  17 ++
>  include/block/block_int.h    |   3 +-
>  qapi/block-core.json         |  33 ++-
>  qemu-img.c                   |   2 +-
>  replication.c                | 105 +++++++
>  replication.h                | 176 ++++++++++++
>  tests/.gitignore             |   1 +
>  tests/Makefile               |   4 +
>  tests/test-replication.c     | 523 +++++++++++++++++++++++++++++++++
>  17 files changed, 1847 insertions(+), 17 deletions(-)
>  create mode 100644 block/replication.c
>  create mode 100644 docs/block-replication.txt
>  create mode 100644 include/block/block_backup.h
>  create mode 100644 replication.c
>  create mode 100644 replication.h
>  create mode 100644 tests/test-replication.c

I have reviewed many revisions of this series.  The main mechanism in
this series makes sense to me.

I'm still concerned that checkpointing (vm_stop(), not in this series
but COLO in general) depends on bdrv_drain(), which can block forever if
I/O is hung.  That doesn't seem like a reasonable limitation for a high
availability feature since it may lead to the VM becoming unavailable.

I'd like Jeff and/or Kevin to review this series and merge it once they
are happy.

Stefan