[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Libguestfs] Point-in-time snapshots

On 03/27/2015 11:21 AM, Richard W.M. Jones wrote:

> We'd issue a drive-backup monitor command with an nbd:... target.
> The custom NBD server receives a stream of blocks (as writes).
> On the other side of this, libguestfs is also talking to the custom
> NBD server.  Libguestfs (which is really a qemu process) is issuing
> random reads.  There's no way for the NBD server or anything else to
> predict what blocks libguestfs will want to read in advance.
> In the middle of this is our custom NBD server, probably implemented
> using nbdkit.  It has to save all the writes from qemu.  It has to
> satisfy the reads from libguestfs, probably by blocking libguestfs
> unless we've seen the corresponding write.

Well, it only has to write the sectors touched by the guest in the
meantime, not the entire disk.  But yeah, a busy guest can cause a lot
of sectors to be written in the meantime.

> The NBD server is going to be (a) storing huge quantities of temporary
> data which we'll mostly not use, and (b) blocking libguestfs for
> arbitrary periods of time.  This doesn't sound very lightweight to me.

Hmm.  Sounds a bit like we want to take advantage of postcopy migration
smarts - where the destination receives the full stream of writes as a
low-priority, but can interject and request out-of-order reads to
satisfy page faults on a high-priority.  All reads are guaranteed to
resolve to the correct data, even if it means blocking the read until
the out-of-order page fault is read in, but the out-of-order processing
means that you don't have to wait for the full stream to take place
before you get the information you need a the moment.

Is NBD bi-directional, in that the target can receive write requests at
the same time it is sending read requests?  It sounds like that is what
we need.  Or are we stuck with NBD being uni-directional, where the
target can receive read and write commands, but can't send any commands
back to the client in charge of the data being written?  If that's the
case, maybe the work in qemu 2.4 towards persistent dirty bitmaps can
help: set up a bitmap before starting the NBD server, to track what the
guest is dirtying.  With the bitmap in place, then for every sector you
want, you first read it directly from the source image, THEN check the
persistent dirty bitmap to see if the sector has been marked for
transfer to NBD.  If so, then you'll have to wait for it to show up on
the NBD target side; if not, then the guest hasn't touched it yet so you
know what you read is correct.  That still doesn't help optimizing out
the writes to the NBD target for sectors you don't care about, and
doesn't quite address the desire for making random reads take priority
over linear streaming of dirty blocks, but it might help.

Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

Attachment: signature.asc
Description: OpenPGP digital signature

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]