Re: [Qemu-devel] [RFC 00/13] Live memory snapshot based on userfaultfd

[Dr. David Alan Gilbert]

* Hailiang Zhang (address@hidden) wrote:
> On 2016/7/14 2:02, Dr. David Alan Gilbert wrote:
> > * zhanghailiang (address@hidden) wrote:
> > > For now, we still didn't support live memory snapshot, we have discussed
> > > a scheme which based on userfaultfd long time ago.
> > > You can find the discussion by the follow link:
> > > https://lists.nongnu.org/archive/html/qemu-devel/2014-11/msg01779.html
> > > 
> > > The scheme is based on userfaultfd's write-protect capability.
> > > The userfaultfd write protection feature is available here:
> > > http://www.spinics.net/lists/linux-mm/msg97422.html
> > 
> > I've (finally!) had a brief look through this, I like the idea.
> > I've not bothered with minor cleanup like comments on them;
> > I'm sure those will happen later; some larger scale things to think
> > about are:
> >    a) I wonder if it's really best to put that much code into the postcopy
> >       function; it might be but I can see other userfault uses as well.
> 
> Yes, it is better to extract common codes into public functions.
> 
> >    b) I worry a bit about the size of the copies you create during setup
> >       and I don't really understand why you can't start sending those pages
> 
> Because we save device state and ram in the same snapshot_thread, if the 
> process
> of saving device is blocked by writing pages, we can remove the write-protect 
> in
> 'postcopy/fault' thread, but can't send it immediately.

Don't you write the devices to a buffer? If so then you perhaps you could split
writing into that buffer into a separate thread.

> >       immediately - but then I worry aobut the relative order of when pages
> >       data should be sent compared to the state of devices view of RAM.
> >    c) Have you considered also using userfault for loading the snapshot - I
> >      know there was someone on #qemu a while ago who was talking about using
> >      it as a way to quickly reload from a migration image.
> > 
> 
> I didn't notice such talking before, maybe i missed it.
> Could you please send me the link ?

I don't think there's any public docs about it; this was a conversation
with Christoph Seifert on #qemu about May last year.

> But i do consider the scenario of quickly snapshot restoring.
> And the difficulty here is how can we quickly find the position
> of the special page. That is, while VM is accessing one page, we
> need to find its position in snapshot file and read it into memory.
> Consider the compatibility, we hope we can still re-use all migration
> capabilities.
> 
> My rough idea about the scenario is:
> 1. Use an array to recode the beginning position of all VM's pages.
> Use the offset as the index for the array, just like migration bitmaps.
> 2. Save the data of the array into another file in a special format.
> 3. Also record the position of device state data in snapshot file.
> (Or we can put the device state data at the head of snapshot file)
> 4. While restore the snapshot, reload the array first, and then read
> the device state.
> 5. Set all pages to MISS status.
> 6. Resume VM to run
> 7. The next process is like how postcopy incoming does.
> 
> I'm not sure if this scenario is practicable or not. We need further
> discussion. :)

Yes;  I can think of a few different ways to do (2):
  a) We could just store it at the end of the snapshot file (and know that
it's at the end - I think the json format description did a similar trick).
  b) We wouldn't need the 4 byte headers on the page we currently send.
  c) Juan's idea of having multiple fd's for migration streams might also fit,
     with the RAM data in the separate file.
  d) But if we know it's a file (not a network stream) then should we treat it
     specially and just use a sparse file of the same size as RAM, and just
     pwrite() the data into the right offset?

Dave

> 
> Hailiang
> 
> > Dave
> > 
> > > 
> > > The process of this live memory scheme is like bellow:
> > > 1. Pause VM
> > > 2. Enable write-protect fault notification by using userfaultfd to
> > >     mark VM's memory to write-protect (readonly).
> > > 3. Save VM's static state (here is device state) to snapshot file
> > > 4. Resume VM, VM is going to run.
> > > 5. Snapshot thread begins to save VM's live state (here is RAM) into
> > >     snapshot file.
> > > 6. During this time, all the actions of writing VM's memory will be 
> > > blocked
> > >    by kernel, and kernel will wakeup the fault treating thread in qemu to
> > >    process this write-protect fault. The fault treating thread will 
> > > deliver this
> > >    page's address to snapshot thread.
> > > 7. snapshot thread gets this address, save this page into snasphot file,
> > >     and then remove the write-protect by using userfaultfd API, after 
> > > that,
> > >     the actions of writing will be recovered.
> > > 8. Repeat step 5~7 until all VM's memory is saved to snapshot file
> > > 
> > > Compared with the feature of 'migrate VM's state to file',
> > > the main difference for live memory snapshot is it has little time delay 
> > > for
> > > catching VM's state. It just captures the VM's state while got users 
> > > snapshot
> > > command, just like take a photo of VM's state.
> > > 
> > > For now, we only support tcg accelerator, since userfaultfd is not 
> > > supporting
> > > tracking write faults for KVM.
> > > 
> > > Usage:
> > > 1. Take a snapshot
> > > #x86_64-softmmu/qemu-system-x86_64 -machine 
> > > pc-i440fx-2.5,accel=tcg,usb=off -drive 
> > > file=/mnt/windows/win7_install.qcow2.bak,if=none,id=drive-ide0-0-1,format=qcow2,cache=none
> > >  -device ide-hd,bus=ide.0,unit=1,drive=drive-ide0-0-1,id=ide0-0-1  -vnc 
> > > :7 -m 8192 -smp 1 -netdev tap,id=bn0 -device 
> > > virtio-net-pci,id=net-pci0,netdev=bn0  --monitor stdio
> > > Issue snapshot command:
> > > (qemu)migrate -d file:/home/Snapshot
> > > 2. Revert to the snapshot
> > > #x86_64-softmmu/qemu-system-x86_64 -machine 
> > > pc-i440fx-2.5,accel=tcg,usb=off -drive 
> > > file=/mnt/windows/win7_install.qcow2.bak,if=none,id=drive-ide0-0-1,format=qcow2,cache=none
> > >  -device ide-hd,bus=ide.0,unit=1,drive=drive-ide0-0-1,id=ide0-0-1  -vnc 
> > > :7 -m 8192 -smp 1 -netdev tap,id=bn0 -device 
> > > virtio-net-pci,id=net-pci0,netdev=bn0  --monitor stdio -incoming 
> > > file:/home/Snapshot
> > > 
> > > NOTE:
> > > The userfaultfd write protection feature does not support THP for now,
> > > Before taking snapshot, please disable THP by:
> > > echo never > /sys/kernel/mm/transparent_hugepage/enabled
> > > 
> > > TODO:
> > > - Reduce the influence for VM while taking snapshot
> > > 
> > > zhanghailiang (13):
> > >    postcopy/migration: Split fault related state into struct
> > >      UserfaultState
> > >    migration: Allow the migrate command to work on file: urls
> > >    migration: Allow -incoming to work on file: urls
> > >    migration: Create a snapshot thread to realize saving memory snapshot
> > >    migration: implement initialization work for snapshot
> > >    QEMUSizedBuffer: Introduce two help functions for qsb
> > >    savevm: Split qemu_savevm_state_complete_precopy() into two helper
> > >      functions
> > >    snapshot: Save VM's device state into snapshot file
> > >    migration/postcopy-ram: fix some helper functions to support
> > >      userfaultfd write-protect
> > >    snapshot: Enable the write-protect notification capability for VM's
> > >      RAM
> > >    snapshot/migration: Save VM's RAM into snapshot file
> > >    migration/ram: Fix some helper functions' parameter to use
> > >      PageSearchStatus
> > >    snapshot: Remove page's write-protect and copy the content during
> > >      setup stage
> > > 
> > >   include/migration/migration.h     |  41 +++++--
> > >   include/migration/postcopy-ram.h  |   9 +-
> > >   include/migration/qemu-file.h     |   3 +-
> > >   include/qemu/typedefs.h           |   1 +
> > >   include/sysemu/sysemu.h           |   3 +
> > >   linux-headers/linux/userfaultfd.h |  21 +++-
> > >   migration/fd.c                    |  51 ++++++++-
> > >   migration/migration.c             | 101 ++++++++++++++++-
> > >   migration/postcopy-ram.c          | 229 
> > > ++++++++++++++++++++++++++++----------
> > >   migration/qemu-file-buf.c         |  61 ++++++++++
> > >   migration/ram.c                   | 104 ++++++++++++-----
> > >   migration/savevm.c                |  90 ++++++++++++---
> > >   trace-events                      |   1 +
> > >   13 files changed, 587 insertions(+), 128 deletions(-)
> > > 
> > > --
> > > 1.8.3.1
> > > 
> > > 
> > --
> > Dr. David Alan Gilbert / address@hidden / Manchester, UK
> > 
> > .
> > 
> 
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK