3:47 a.m.
On Mon, Apr 25, 2022 at 03:25:43PM +0100, Dr. David Alan Gilbert wrote:
* Daniel P. Berrangé (berrange(a)redhat.com) wrote:
> On Mon, Apr 25, 2022 at 01:33:41PM +0100, Dr. David Alan Gilbert wrote:
> > * Daniel P. Berrangé (berrange(a)redhat.com) wrote:
> > > On Mon, Apr 25, 2022 at 12:04:37PM +0100, Dr. David Alan Gilbert wrote:
> > > > * Daniel P. Berrangé (berrange(a)redhat.com) wrote:
> > > > > I think we need to introduce an explicit 'file:'
protocol
> > > > > for the migrate command, that is backed by the blockdev APIs
> > > > > so it can do O_DIRECT and non-blocking AIO. For the
'fd:'
> > > > > protocol, we need to be able to tell QEMU whether the
'fd'
> > > > > is a stream or a regular file, so it can choose between the
> > > > > regular send/recv APIs, vs the Blockdev APIs (maybe we can
> > > > > auto-detect with fstat()). If we do this, then multifd
> > > > > doesn't end up needing multiple save files on disk, all
> > > > > the threads can be directly writing to the same file, just
> > > > > as the relevant offsets on disk to match the RAM page
> > > > > location.
> > > >
> > > > Hmm so what I'm not sure of is whether it makes sense to use the
normal
> > > > migration flow/code for this or not; and you're suggesting a few
> > > > possibly contradictory things.
> > > >
> > > > Adding a file: protocol would be pretty easy (whether it went via
> > > > the blockdev layer or not); getting it to be more efficient is the
> > > > tricky part, because we've got loads of levels of stream
abstraction in
> > > > the RAM save code:
> > > > QEMUFile->channel->OS
> > > > but then if you want to enforce alignment you somehow have to make
that
> > > > go all the way down.
> > >
> > > The QIOChannel stuff doesn't add buffering, so I wasn't worried
> > > about alignment there.
> > >
> > > QEMUFile has optional buffering which would mess with alignment,
> > > but we could turn that off potentially for the RAM transfer, if
> > > using multifd.
> >
> > The problem isn't whether they add buffering or not; the problem is you
> > now need a way to add a mechanism to ask for alignment.
> >
> > > I'm confident the performance on the QMEU side though could
> > > exceed what's viable with libvirt's iohelper today, as we
> > > would definitely be eliminating 1 copy and many context switches.
> >
> > Yes but you get that just from adding a simple file: (or fd:) mode
> > without trying to do anything clever with alignment or rewriting the
> > same offset.
>
> I don't think so, as libvirt supports O_DIRECT today to avoid
> trashing the host cache when saving VMs. So to be able to
> offload libvirt's work to QEMU, O_DIRECT is a pre-requisite.
I guess you could O_DIRECT it from a buffer in QemuFile or the channel.
> So we do need the alignment support at the very least. Rewriting
> at the same offset isnt mandatory, but I think it'd make multifd
> saner if trying to have all threads work on the same file.
Thinking on the fly, you'd need some non trivial changes:
a) A section entry in the format to say 'align to ... n bytes'
(easyish)
Yep
b) A way to allocate a location in the file to a RAMBlock
[ We already have a bitmap address, so that might do, but
you need to make it interact with the existing file, so it might
be easier to do the allocate and record it ]
IIUC, the migration protocol first serializes all RAM, and then serializes
the VMstate for devices. When libvirt creates a save image for a VM it
has its own header + XML dump and then appends the migrate stream.
So we get a layout of
+------------------+
| libvirt header |
+------------------+
| libvirt XML |
| ... |
+------------------+
| migration stream |
| ... |
+------------------+
The 'migration stream' is opaque as far as libvirt is concerned,
but we happen to know that from QEMU POV it expands to
+------------------+
| RAM stream |
| ... |
+------------------+
| vmstate |
| ... |
+------------------+
Where 'RAM stream' is a stream of the RAM block header and
RAM block contents, for every page.
In the suggestion above, my desire is to achieve this layout
+------------------+
| libvirt header |
+------------------+
| libvirt XML |
| ... |
+------------------+
| RAM |
| ... |
+------------------+
| vmstate |
| ... |
+------------------+
The key difference being 'RAM' instead of 'RAM stream'. Libvirt
would have to tell QEMU what offset in the file it is permitted
to start at - say 16 MB. 'RAM' would be a 1:1 mapping of the
guest RAM, simply offset by that 16 MB.
IOW, I'm assuming that a 4 GB RAM VM would have its RAM written
starting from offfset 16 MB and ending at 4 GB + 16 MB.
I'm thinking stuff like virtio-mem / RAM hotplug makes life
harder though as there can be many distinct blocks of RAM
contributing to a QEMU map, and we need to be able to declare
an ordering of them, for mapping to the file.
c) A way to say to the layers below it while writing RAM that it
needs to go in a given location.
Yes, QEMUFile would need some new APIs accepting offsets, and we
would need a way to report whether a given impl supports random
access, vs streaming only.
d) A clean way for a..c only to happen in this case.
e) Hmm ram size changes/hotplug/virtio-mem
This is "easy" from libvirt POV, since none of those things are possible
to be invoked during save/restore/migrate.
Hard(er) for QEMU since it doesn't place that restriction on users in the
general case.
With regards,
Daniel
--
|: https://berrange.com -o- https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org -o- https://fstop138.berrange.com :|
|: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|