Re: Warm and Live Migration Implementation for bhyve

From: Corvin Köhne <corvink_at_FreeBSD.org>
Date: Mon, 03 Jul 2023 07:34:57 UTC
On Tue, 2023-06-27 at 16:35 +0300, Elena Mihailescu wrote:
> Hi Corvin,
> 
> Thank you for the questions! I'll respond to them inline.
> 
> On Mon, 26 Jun 2023 at 10:16, Corvin Köhne <corvink@freebsd.org>
> wrote:
> > 
> > Hi Elena,
> > 
> > thanks for posting this proposal here.
> > 
> > Some open questions from my side:
> > 
> > 1. How is the data send to the target? Does the host send a
> > complete
> > dump and the target parses it? Or does the target request data one
> > by
> > one und the host sends it as response?
> > 
> It's not a dump of the guest's state, it's transmitted in steps.
> However, some parts may be migrated as a chunk (e.g., the emulated
> devices' state is transmitted as the buffer generated from the
> snapshot functions).
> 

How does the receiver know which chunk relates to which device? It
would be nice if you can start bhyve on the receiver side without
parameters e.g. `bhyve --receive=127.0.0.1:1234`. Therefore, the
protocol has to carry some information about the device configuration.

> I'll try to describe a bit the protocol we have implemented for
> migration, maybe it can partially respond to the second and third
> questions.
> 
> The destination host waits for the source host to connect (through a
> socket).
> After that, the source sends its system specifications (hw_machine,
> hw_model, hw_pagesize). If the source and destination hosts have
> identical hardware configurations, the migration can take place.
> 
> Then, if we have live migration, we migrate the memory in rounds
> (i.e., we get a list of the pages that have the dirty bit set, send
> it
> to the destination to know what pages will be received, then send the
> pages through the socket; this process is repeated until the last
> round).
> 
> Next, we stop the guest's vcpus, send the remaining memory (for live
> migration) or the guest's memory from vmctx->baseaddr for warm
> migration. Then, based on the suspend/resume feature, we get the
> state
> of the virtualized devices (the ones from the kernel space) and send
> this buffer to the destination. We repeat this for the emulated
> devices as well (the ones from the userspace).
> 
> On the receiver host, we get the memory pages and set them to their
> according position in the guest's memory, use the restore functions
> for the state of the devices and start the guest's execution.
> 
> Excluding the guest's memory transfer, the rest is based on the
> suspend/resume feature. We snapshot the guest's state, but instead of
> saving the data locally, we send it via network to the destination.
> On
> the destination host, we start a new virtual machine, but instead of
> reading/getting the state from the disk (i.e., the snapshot files) we
> get this state via the network from the source host.
> 
> If the destination can properly resume the guest activity, it will
> send an "OK" to the source host so it can destroy/remove the guest
> from its end.
> 
> Both warm and live migration are based on "cold migration". Cold
> migration means we suspend the guest on the source host, and restore
> the guest on the destination host from the snapshot files. Warm
> migration only does this using a socket, while live migration changes
> the way the memory is migrated.
> 
> > 2. What happens if we add a new data section?
> > 
> What are you referring to with a new data section? Is this question
> related to the third one? If so, see my answer below.
> 
> > 3. What happens if the bhyve version differs on host and target
> > machine?
> 
> The two hosts must be identical for migration, that's why we have the
> part where we check the specifications between the two migration
> hosts. They are expected to have the same version of bhyve and
> FreeBSD. We will add an additional check in the check specs part to
> see if we have the same FreeBSD build.
> 
> As long as the changes in the virtual memory subsystem won't affect
> bhyve (and how the virtual machine sees/uses the memory), the
> migration constraints should only be related to suspend/resume. The
> state of the virtual devices is handled by the snapshot system, so if
> it is able to accommodate changes in the data structures, the
> migration process will not be affected.
> 
> Thank you,
> Elena
> 
> > 
> > 
> > --
> > Kind regards,
> > Corvin
> > 
> > On Fri, 2023-06-23 at 13:00 +0300, Elena Mihailescu wrote:
> > > Hello,
> > > 
> > > This mail presents the migration feature we have implemented for
> > > bhyve. Any feedback from the community is much appreciated.
> > > 
> > > We have opened a stack of reviews on Phabricator
> > > (https://reviews.freebsd.org/D34717) that is meant to split the
> > > code
> > > in smaller parts so it can be more easily reviewed. A brief
> > > history
> > > of
> > > the implementation can be found at the bottom of this email.
> > > 
> > > The migration mechanism we propose needs two main components in
> > > order
> > > to move a virtual machine from one host to another:
> > > 1. the guest's state (vCPUs, emulated and virtualized devices)
> > > 2. the guest's memory
> > > 
> > > For the first part, we rely on the suspend/resume feature. We
> > > call
> > > the
> > > same functions as the ones used by suspend/resume, but instead of
> > > saving the data in files, we send it via the network.
> > > 
> > > The most time consuming aspect of migration is transmitting guest
> > > memory. The UPB team has implemented two options to accomplish
> > > this:
> > > 1. Warm Migration: The guest execution is suspended on the source
> > > host
> > > while the memory is sent to the destination host. This method is
> > > less
> > > complex but may cause extended downtime.
> > > 2. Live Migration: The guest continues to execute on the source
> > > host
> > > while the memory is transmitted to the destination host. This
> > > method
> > > is more complex but offers reduced downtime.
> > > 
> > > The proposed live migration procedure (pre-copy live migration)
> > > migrates the memory in rounds:
> > > 1. In the initial round, we migrate all the guest memory (all
> > > pages
> > > that are allocated)
> > > 2. In the subsequent rounds, we migrate only the pages that were
> > > modified since the previous round started
> > > 3. In the final round, we suspend the guest, migrate the
> > > remaining
> > > pages that were modified from the previous round and the guest's
> > > internal state (vCPU, emulated and virtualized devices).
> > > 
> > > To detect the pages that were modified between rounds, we propose
> > > an
> > > additional dirty bit (virtualization dirty bit) for each memory
> > > page.
> > > This bit would be set every time the page's dirty bit is set.
> > > However,
> > > this virtualization dirty bit is reset only when the page is
> > > migrated.
> > > 
> > > The proposed implementation is split in two parts:
> > > 1. The first one, the warm migration, is just a wrapper on the
> > > suspend/resume feature which, instead of saving the suspended
> > > state
> > > on
> > > disk, sends it via the network to the destination
> > > 2. The second part, the live migration, uses the layer previously
> > > presented, but sends the guest's memory in rounds, as described
> > > above.
> > > 
> > > The migration process works as follows:
> > > 1. we identify:
> > >  - VM_NAME - the name of the virtual machine which will be
> > > migrated
> > >  - SRC_IP - the IP address of the source host
> > >  - DST_IP - the IP address of the destination host (default is
> > > 24983)
> > >  - DST_PORT - the port we want to use for migration
> > > 2. we start a virtual machine on the destination host that will
> > > wait
> > > for a migration. Here, we must specify SRC_IP (and the port we
> > > want
> > > to
> > > open for migration, default is 24983).
> > > e.g.: bhyve ... -R SRC_IP:24983 guest_vm_dst
> > > 3. using bhyvectl on the source host, we start the migration
> > > process.
> > > e.g.: bhyvectl --migrate=DST_IP:24983 --vm=guest_vm
> > > 
> > > A full tutorial on this can be found here:
> > > https://github.com/FreeBSD-UPB/freebsd-src/wiki/Virtual-Machine-Migration-using-bhyve
> > > 
> > > For sending the migration request to a virtual machine, we use
> > > the
> > > same thread/socket that is used for suspend.
> > > For receiving a migration request, we used a similar approach to
> > > the
> > > resume process.
> > > 
> > > As some of you may remember seeing similar emails from our part
> > > on
> > > the
> > > freebsd-virtualization list, I'll present a brief history of this
> > > project:
> > > The first part of the project was the suspend/resume
> > > implementation
> > > which landed in bhyve in 2020, under the BHYVE_SNAPSHOT guard
> > > (https://reviews.freebsd.org/D19495).
> > > After that, we focused on two tracks:
> > > 1. adding various suspend/resume features (multiple device
> > > support -
> > > https://reviews.freebsd.org/D26387, CAPSICUM support -
> > > https://reviews.freebsd.org/D30471, having an uniform file format
> > > -
> > > at
> > > that time, during the bhyve bi-weekly calls, we concluded that
> > > the
> > > JSON format was the most suitable at that time -
> > > https://reviews.freebsd.org/D29262) so we can remove the #ifdef
> > > BHYVE_SNAPSHOT guard.
> > > 2. implementing the migration feature for bhyve. Since this one
> > > relies
> > > on the save/restore, but does not modify its behaviour, we
> > > considered
> > > we can go in parallel with both tracks.
> > > We had various presentations in the FreeBSD Community on these
> > > topics:
> > > AsiaBSDCon2018, AsiaBSDCon2019, BSDCan2019, BSDCan2020,
> > > AsiaBSDCon2023.
> > > 
> > > The first patches for warm and live migration were opened in
> > > 2021:
> > > https://reviews.freebsd.org/D28270,
> > > https://reviews.freebsd.org/D30954. However, the general feedback
> > > on
> > > these was that the patches are too big to be reviewed, so we
> > > should
> > > split them in smaller chunks (this was also true for some of the
> > > suspend/resume improvements). Thus, we split them into smaller
> > > parts.
> > > Also, as things changed in bhyve (i.e., capsicum support for
> > > suspend/resume was added this year), we rebased and updated our
> > > reviews.
> > > 
> > > Thank you,
> > > Elena
> > > 
> > 

-- 
Kind regards,
Corvin