[Qemu-devel] [PATCH RFC 0/6] migration: enable return-path for precopy

Peter Xu posted 6 patches 6 years, 11 months ago
Failed in applying to current master (apply log)
Test checkpatch passed
Test docker passed
Test s390x passed
include/hw/compat.h           |   4 ++
include/io/channel.h          |   1 +
include/migration/migration.h |  27 ++++++++-
include/qemu/typedefs.h       |   1 -
io/channel-socket.c           |   1 +
migration/migration.c         | 133 +++++++++++++++++++++++++++++++-----------
migration/postcopy-ram.c      |   1 -
migration/qemu-file-channel.c |   9 +++
migration/trace-events        |   4 +-
9 files changed, 140 insertions(+), 41 deletions(-)
[Qemu-devel] [PATCH RFC 0/6] migration: enable return-path for precopy
Posted by Peter Xu 6 years, 11 months ago
This idea derives from the bug reported:

  https://bugzilla.redhat.com/show_bug.cgi?id=1439147

It is not a extremely critical bug, since as long as the user uses
libvirt to migrate, we won't have such an issue at all (we'll have
identical command line parameters for QEMU). However it can be an
severe one, since it caused data loss (source VM will quit assuming
that destination VM is good, while it is not). So it would be good we
fix it. Meanwhile, imho the return path may be used for other things
as well in the future.

This series proposed a solution for the bug - let's enable return path
even for precopy when possible, then source can know whether dest is
good or not, then it can decide to do something better than quit.

The series is marked as RFC for two reasons:

One point for RFC is the idea in general that whether we would like to
enable return path for precopy. I see it okay since after all it's
optional (so it greatly reduces the chance that QEMU command line user
will lose the data, and this feature will be automatically off when we
are using e.g. "exec:" to migrate to a file).

The other one point is that I converted MigrationState into a QObject
(actually a QDevice) to let migration codes benefits from some general
framework advantages.

I got at least one positive feedback on each of the RFC elements
mentioned above, so dare I post this to public. Though I guess some
migration reviewers may be on PTO recently (I just know it :), I still
decided to post this out for a broader rfc review.

Smoke test done, and of course I verified this series to fix the bug
mentioned.

Please review. Thanks.

Peter Xu (6):
  io: only allow return path for socket typed
  migration: isolate return path on src
  migration: fix leak of src file on dst
  migration: shut src return path unconditionally
  migration: let MigrationState be an QObject
  migration: enable return path for precopy

 include/hw/compat.h           |   4 ++
 include/io/channel.h          |   1 +
 include/migration/migration.h |  27 ++++++++-
 include/qemu/typedefs.h       |   1 -
 io/channel-socket.c           |   1 +
 migration/migration.c         | 133 +++++++++++++++++++++++++++++++-----------
 migration/postcopy-ram.c      |   1 -
 migration/qemu-file-channel.c |   9 +++
 migration/trace-events        |   4 +-
 9 files changed, 140 insertions(+), 41 deletions(-)

-- 
2.7.4


Re: [Qemu-devel] [PATCH RFC 0/6] migration: enable return-path for precopy
Posted by Peter Xu 6 years, 11 months ago
On Fri, May 19, 2017 at 02:43:26PM +0800, Peter Xu wrote:
> This idea derives from the bug reported:
> 
>   https://bugzilla.redhat.com/show_bug.cgi?id=1439147
> 
> It is not a extremely critical bug, since as long as the user uses
> libvirt to migrate, we won't have such an issue at all (we'll have
> identical command line parameters for QEMU). However it can be an
> severe one, since it caused data loss (source VM will quit assuming
> that destination VM is good, while it is not). So it would be good we
> fix it. Meanwhile, imho the return path may be used for other things
> as well in the future.
> 
> This series proposed a solution for the bug - let's enable return path
> even for precopy when possible, then source can know whether dest is
> good or not, then it can decide to do something better than quit.
> 
> The series is marked as RFC for two reasons:
> 
> One point for RFC is the idea in general that whether we would like to
> enable return path for precopy. I see it okay since after all it's
> optional (so it greatly reduces the chance that QEMU command line user
> will lose the data, and this feature will be automatically off when we
> are using e.g. "exec:" to migrate to a file).
> 
> The other one point is that I converted MigrationState into a QObject
> (actually a QDevice) to let migration codes benefits from some general
> framework advantages.
> 
> I got at least one positive feedback on each of the RFC elements
> mentioned above, so dare I post this to public. Though I guess some
> migration reviewers may be on PTO recently (I just know it :), I still
> decided to post this out for a broader rfc review.
> 
> Smoke test done, and of course I verified this series to fix the bug
> mentioned.
> 
> Please review. Thanks.
> 
> Peter Xu (6):
>   io: only allow return path for socket typed
>   migration: isolate return path on src
>   migration: fix leak of src file on dst
>   migration: shut src return path unconditionally
>   migration: let MigrationState be an QObject
>   migration: enable return path for precopy
> 
>  include/hw/compat.h           |   4 ++
>  include/io/channel.h          |   1 +
>  include/migration/migration.h |  27 ++++++++-
>  include/qemu/typedefs.h       |   1 -
>  io/channel-socket.c           |   1 +
>  migration/migration.c         | 133 +++++++++++++++++++++++++++++++-----------
>  migration/postcopy-ram.c      |   1 -
>  migration/qemu-file-channel.c |   9 +++
>  migration/trace-events        |   4 +-
>  9 files changed, 140 insertions(+), 41 deletions(-)
> 
> -- 
> 2.7.4
> 

CC Laurent as well. Sorry!

-- 
Peter Xu