Re: [Qemu-devel] [PATCH v9 00/56] Postcopy implementation

[Dr. David Alan Gilbert]

* Vladimir Sementsov-Ogievskiy (address@hidden) wrote:
> Hello, Dr. Gilbert!

Hi Vladimir,

> As I understand this is not a complete realization of post-copy stage of
> migration, but only realization of ram post-copy.

Yes, that's correct; RAM was the only one we'd had much interest in; but
I can understand wanting to help block as well.

> I need to implement
> post-copy migration of block-dirty-bitmaps. And it should work with/without
> ram post-copy. Did you plan a possibility of post-copying something except
> ram? If yes, can you please help me with an interface? Should I implement
> separate thread like postcopy_ram_listen_thread and other things, or
> something may be reused?

I think with the structure that's in, it should be possible to add
another device doing postcopy as well.  Although maybe it's worth checking
whether these dirty bitmaps behave similarly enough to a non-guest visible
RAM block, so then the existing migrate code would handle it.

1) I think you're probably best reusing the postcopy_ram_listen_thread
for all postcopy data on the load side;  I think that's probably already
OK to do that; just rename it.  The migration stream once in postcopy
mode is just still a normal migration stream, so it can still have a mix
of RAM and disk or whatever.

2) The tricky bit I can think of is on the sending side making sure
you interleave the requests for RAM and disk requests sensibly; at the moment
when in postcopy we drop the bandwidth limit so it'll keep doing RAM transfers
and respond to ram requests as soon as possible;  you'll need to persuade
it to somehow realise there's a disk request and get the RAM code to
drop out of it's loop to give the disk code a chance.  There's already
the opportunity to improve the behaviour with just RAM to stop it flooding
the network buffers with background page requests which delay the responses
to the postcopy requests.

My knowledge of the block code is pretty minimal, so I've not had a chance
to follow much of your dirty-bitmap series.

Dave

> 
> On 05.11.2015 21:10, Dr. David Alan Gilbert (git) wrote:
> >From: "Dr. David Alan Gilbert" <address@hidden>
> >
> >   This is the 9th cut of my version of postcopy.
> >
> >The userfaultfd linux kernel code is now in the upstream kernel
> >tree, and so 4.3 can be used without modification.
> >
> >This qemu series can be found at:
> >https://github.com/orbitfp7/qemu.git
> >on the wp3-postcopy-v9 tag
> >
> >Testing status:
> >   * Tested heavily on x86
> >   * Smoke tested on aarch64 (so it does work on different page sizes)
> >
> >This work has been partially funded by the EU Orbit project:
> >   see http://www.orbitproject.eu/about/
> >
> >v9
> >   Almost all of the changes are changes from review comments, and most are
> >   minor.
> >
> >   The following are new patches, mostly split out from earlier patches
> >   (The exception being adding userfaultfd.h header back in - which was in v5
> >    but we took out, but is needed again due to 1842bdfd)
> >
> >     04/56 Move page_size_init earlier
> >     08/56 qemu_ram_block_by_name
> >     12/56 Factor out host_from_stream_offset call and check
> >     15/56 Add Linux userfaultfd.h header back
> >     21/56 migration_is_setup_or_active
> >     30/56 migration_completion: Take current state
> >     34/56 Maintain unsentmap
> >
> >   The previous patches 03,10,13,36/54 went in upstream already.
> >
> >   Fix for assert using hotplug (Thanks Bharata for spotting that)
> >   Fix for migrate_cancel a second time
> >   Rework for migration_bitmap_rcu after Denis's deadlock fix
> >   Rework ram_load into a separate postcopy loop
> >   The 'sentmap' is now an 'unsentmap' - this saves a complement step at the 
> > end
> >      The unsentmap creation is now split into a separate patch
> >      The unsentmap is now stored in the RCU structure (although it can't
> >        really resize during the migrate)
> >   Fix for block migration
> >      still not a suggested combination.
> >   Minor:
> >      trace_savevm_send_open_return_path added
> >      split migration_is_active function into separate patch and made
> >         migration_is_setup_or_active
> >      move file reads into migrate_handle_advise and migrate_handle_packaged
> >      use of MIN in send-packaged-chunk
> >      migration_thread_started -> migration_thread_running
> >      updated qemu_get_buffer_in_place with Juan's version (and size_t'ified 
> > it)
> >      split the host_from_stream_offset change into a separate patch
> >        'ram_load: Factor out host_from_stream_offset call and check'
> >      split out ram_find_block_by_id into a separate patch and now
> >        called qemu_ram_block_by_name
> >      postcopy_discard_send_range etc now take start/length rather than 
> > start/end
> >        (also added another trace)
> >      split ram_save_host_page into ram_save_host_page/ram_save_target_page
> >      split host page cleanup function into a core that handles both passes
> >
> >Dave
> >
> >Dr. David Alan Gilbert (56):
> >   Add postcopy documentation
> >   Provide runtime Target page information
> >   Move configuration section writing
> >   Move page_size_init earlier
> >   qemu_ram_block_from_host
> >   qemu_ram_block_by_name
> >   Rename mis->file to from_src_file
> >   Add qemu_get_buffer_in_place to avoid copies some of the time
> >   Add wrapper for setting blocking status on a QEMUFile
> >   Add QEMU_MADV_NOHUGEPAGE
> >   ram_debug_dump_bitmap: Dump a migration bitmap as text
> >   ram_load: Factor out host_from_stream_offset call and check
> >   migrate_init: Call from savevm
> >   Rename save_live_complete to save_live_complete_precopy
> >   Add Linux userfaultfd.h header
> >   Return path: Open a return path on QEMUFile for sockets
> >   Return path: socket_writev_buffer: Block even on non-blocking fd's
> >   Migration commands
> >   Return path: Control commands
> >   Return path: Send responses from destination to source
> >   migration_is_setup_or_active
> >   Return path: Source handling of return path
> >   Rework loadvm path for subloops
> >   Add migration-capability boolean for postcopy-ram.
> >   Add wrappers and handlers for sending/receiving the postcopy-ram
> >     migration messages.
> >   MIG_CMD_PACKAGED: Send a packaged chunk of migration stream
> >   Modify save_live_pending for postcopy
> >   postcopy: OS support test
> >   migrate_start_postcopy: Command to trigger transition to postcopy
> >   migration_completion: Take current state
> >   MIGRATION_STATUS_POSTCOPY_ACTIVE: Add new migration state
> >   Avoid sending vmdescription during postcopy
> >   Add qemu_savevm_state_complete_postcopy
> >   Postcopy: Maintain unsentmap
> >   Postcopy: Calculate discard
> >   postcopy: Incoming initialisation
> >   postcopy: ram_enable_notify to switch on userfault
> >   Postcopy: Postcopy startup in migration thread
> >   Postcopy: End of iteration
> >   Page request:  Add MIG_RP_MSG_REQ_PAGES reverse command
> >   Page request: Process incoming page request
> >   Page request: Consume pages off the post-copy queue
> >   postcopy_ram.c: place_page and helpers
> >   Postcopy: Use helpers to map pages during migration
> >   postcopy: Check order of received target pages
> >   Don't sync dirty bitmaps in postcopy
> >   Don't iterate on precopy-only devices during postcopy
> >   Host page!=target page: Cleanup bitmaps
> >   Round up RAMBlock sizes to host page sizes
> >   Postcopy; Handle userfault requests
> >   Start up a postcopy/listener thread ready for incoming page data
> >   postcopy: Wire up loadvm_postcopy_handle_ commands
> >   Postcopy: Mark nohugepage before discard
> >   End of migration for postcopy
> >   Disable mlock around incoming postcopy
> >   Inhibit ballooning during postcopy
> >
> >  balloon.c                         |  11 +
> >  docs/migration.txt                | 191 ++++++++
> >  exec.c                            |  92 +++-
> >  hmp-commands.hx                   |  15 +
> >  hmp.c                             |   7 +
> >  hmp.h                             |   1 +
> >  hw/ppc/spapr.c                    |   2 +-
> >  hw/virtio/virtio-balloon.c        |   4 +-
> >  include/exec/cpu-common.h         |   4 +
> >  include/exec/exec-all.h           |   1 -
> >  include/exec/ram_addr.h           |   2 -
> >  include/migration/migration.h     | 121 ++++-
> >  include/migration/postcopy-ram.h  |  99 ++++
> >  include/migration/qemu-file.h     |  10 +
> >  include/migration/vmstate.h       |   8 +-
> >  include/qemu-common.h             |   1 +
> >  include/qemu/osdep.h              |   9 +
> >  include/qemu/typedefs.h           |   3 +
> >  include/sysemu/balloon.h          |   2 +
> >  include/sysemu/sysemu.h           |  46 +-
> >  kvm-all.c                         |   1 -
> >  linux-headers/linux/userfaultfd.h | 167 +++++++
> >  migration/Makefile.objs           |   2 +-
> >  migration/block.c                 |   9 +-
> >  migration/migration.c             | 723 +++++++++++++++++++++++++--
> >  migration/postcopy-ram.c          | 760 +++++++++++++++++++++++++++++
> >  migration/qemu-file-unix.c        | 111 ++++-
> >  migration/qemu-file.c             |  64 +++
> >  migration/ram.c                   | 997 
> > +++++++++++++++++++++++++++++++++++---
> >  migration/savevm.c                | 823 +++++++++++++++++++++++++++----
> >  qapi-schema.json                  |  18 +-
> >  qmp-commands.hx                   |  19 +
> >  qtest.c                           |   1 -
> >  trace-events                      |  84 +++-
> >  vl.c                              |   1 +
> >  35 files changed, 4159 insertions(+), 250 deletions(-)
> >  create mode 100644 include/migration/postcopy-ram.h
> >  create mode 100644 linux-headers/linux/userfaultfd.h
> >  create mode 100644 migration/postcopy-ram.c
> >
> 
> 
> -- 
> Best regards,
> Vladimir
> * now, @virtuozzo.com instead of @parallels.com. Sorry for this inconvenience.
> 
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK