Features/SnapshottingImprovements

From QEMU
Revision as of 18:46, 29 July 2010 by Miguel (talk | contribs)

Snapshotting Improvements

This page tries to enumerate some problems/inconsistencies/limitations/etc when saving/loading VMs in QEMU. This page is being maintained by the GSoC student Miguel Di Ciurcio Filho as part the QMP project. All feedback is appreciated.


Problem Difficulty What must be done Questions Status
ID and TAG confusion Medium
Snapshot overwriting Easy Verify the existence of the snapshot before saving Do we really need to support savevm <ID>? Patch sent to qemu-devel: http://lists.gnu.org/archive/html/qemu-devel/2010-07/msg01701.html
current/previous Medium
inheritance Hard
Monitor does not show all devices Easy Consolidate the snapshot information before displaying it Do we really need to display the ID? Patch sent to qemu-devel: http://lists.gnu.org/archive/html/qemu-devel/2010-07/msg01591.html
mixing disks could be dangerous Hard Patch to check all devices before loading sent and acked by Kevin Wolf: http://lists.gnu.org/archive/html/qemu-devel/2010-07/msg01065.html Still need to think about the ID collision thing.


ID and TAG confusion

Starting a VM with no previous snapshot on the disks.

(qemu) savevm 2
(qemu) savevm
(qemu) info snapshots 
Snapshot devices: ide0-hd0
Snapshot list (from ide0-hd0):
ID        TAG                 VM SIZE                DATE       VM CLOCK
1         2                      1.5M 2010-07-21 16:55:48   00:00:15.581
2                                1.5M 2010-07-21 16:55:55   00:00:21.501
(qemu) loadvm 2 <loads the VM with the id 2>

It is not clear witch snapshot is going to be loaded.

Ideas for improvements:

  • savevm tests the argument to check if it is an ID or a TAG. Make it accept only a TAG and do not allow the user to specify and ID.
  • loadvm does the same thing. The user must specify if an ID or TAG is wanted. Suggestions:
(qemu) loadvm id=3
(qemu) loadvm tag=xxx
(qemu) loadvm xxx <error>
(qemu) loadvm 3 <error>

In case loadvm cannot be changed due to retro-compatibility, introduce a new command. Suggestions:

(qemu) load_vm id=3
(qemu) load_vm tag=xxx
(qemu) loadvm_byid 3
(qemu) loadvm_bytag xxx

Snapshot overwriting

By id

(qemu) savevm 
(qemu) savevm 
(qemu) info snapshots 
Snapshot devices: ide0-hd0
Snapshot list (from ide0-hd0):
ID        TAG                 VM SIZE                DATE       VM CLOCK
1                                3.9M 2010-07-21 17:08:31   00:00:03.696
2                                3.9M 2010-07-21 17:08:33   00:00:05.419 <-- 17:08:33
(qemu) savevm 2 <overwrittes the snapshot with id 2>
(qemu) info snapshots 
Snapshot devices: ide0-hd0
Snapshot list (from ide0-hd0):
ID        TAG                 VM SIZE                DATE       VM CLOCK
1                                3.9M 2010-07-21 17:08:31   00:00:03.696
2                                3.9M 2010-07-21 17:08:53   00:00:25.918 <-- 17:08:53

Ideas for improvements:

  • Do not allow the user to specify an ID when saving a VM, just a TAG.

By tag

(qemu) savevm test1
(qemu) savevm test2
 (qemu) info snapshots 
Snapshot devices: ide0-hd0
Snapshot list (from ide0-hd0):
ID        TAG                 VM SIZE                DATE       VM CLOCK
1         test1                  3.9M 2010-07-21 17:12:11   00:00:08.126 <-- 17:12:11
2         test2                  3.9M 2010-07-21 17:12:19   00:00:16.053
(qemu) savevm test1
(qemu) info snapshots 
Snapshot devices: ide0-hd0
Snapshot list (from ide0-hd0):
ID        TAG                 VM SIZE                DATE       VM CLOCK
2         test2                  3.9M 2010-07-21 17:12:19   00:00:16.053
1         test1                  3.9M 2010-07-21 17:12:33   00:00:29.221 <-- 17:12:33

Ideas for improvements:

  • Do not allow the user to overwrite an snapshot, even if saving with the same name (will have different IDs anyway)

current/previous

  • There is no way to know witch VM snapshot is currently loaded.
  • There is no history of loaded VMs.

inheritance

The only way to know witch snapshot came before of after another is to look at the date, but it is not possible to fully deduct the relationship. e.g.:

(qemu) savevm a
(qemu) savevm b
(qemu) savevm c
(qemu) savevm d
(qemu) loadvm b
(qemu) savevm e
(qemu) info snapshots 
Snapshot devices: ide0-hd0
Snapshot list (from ide0-hd0):
ID        TAG                 VM SIZE                DATE       VM CLOCK
1         a                      3.9M 2010-07-21 17:32:55   00:00:06.144
2         b                      3.9M 2010-07-21 17:33:00   00:00:11.103
3         c                      3.9M 2010-07-21 17:33:13   00:00:22.953
4         d                      3.9M 2010-07-21 17:33:16   00:00:26.289
5         e                      3.9M 2010-07-21 17:33:28   00:00:15.548

Snapshot e is newer than d, but it is an state saved right after b, and not d as implied by the ids and dates.

Ideas for improvements:

  • Store the previous snapshot taken as parent_id of a new snapshot.

mixing disks could be dangerous

A VM is loaded with two devices, one device with three previous snapshots and the other device with two previous snapshots.

I take a new snapshot of the VM without a name: the first device will save the snapshot as id_str 4 and the second device with id_str 3.

I play around for a while and I want to restore the VM to my latest state. When 'info snapshots' is run, only the first device is show with the VM state, and the latest ID shown will be 4.

Ideas for improvements:

  • Use an stronger ID model, not sequential and large enough to be virtually unique.

loading without checking all devices

So, I fire up loadvm 4. qemu will try to load by id or name anyway, will succeed loading the VM state from the first device, but will fail to load an non-existent snapshot id from the second device and will still be running with the second device out of sync with the VM state.


equal IDs on different devices with different states (id collision)

Running again:

(qemu) savevm 4

Continuing on the same case. The previously saved state identified by id_str 4 will be deleted and replaced by a new one, but only in the first device of my example above. The second device will get a new snapshot, since it does not have a id_str equal to 4.

If I run loadvm 4 again, the snapshot id 3 in the second device will be left there in an state diferent from the snapshot id 3 from the first device.

Ideas for improvements:

  • Use an stronger ID model, not sequential and large enough to be virtually unique.

Monitor does not show all devices

When executing info snapshots only information about the device that saves the VM state is shown. If there are more devices, there is no way to verify if they are all consistent and have the same states.

A first try to address this issue: http://lists.gnu.org/archive/html/qemu-devel/2010-06/msg02386.html

  • Is it really necessary to show snapshot information and relating it to block devices?
  • Change 'info snapshots' to show information without exposing block devices

QMP

Here is a list of suggested commands to manipulate snapshots information on a QEMU VM.

query-snapshot-status

Returns true when the VM can be snapshotted, false otherwise.

-> {"execute": "query-snapshot-status"}
<- {"return": true }
  • Not sure if QMP allow returning just a boolean.

snapshot-create

Runs savevm.

-> {"execute": "snapshot-create", "arguments": {"name": "vm_before_upgrade"} }
<- { "return": {} }
  • "name" is mandatory and must be unique.
  • Needs to return any errors.

snapshot-delete

Runs delvm.

-> {"execute": "snapshot-delete", "arguments": {"name": "vm_before_upgrade"} }
<- { "return": {} }
  • "name" is mandatory and must be unique.
  • Needs to return any errors.

snapshot-load

Runs loadvm.

-> {"execute": "snapshot-load", "arguments": {"name": "vm_before_upgrade"} }
<- { "return": {} }
  • "name" is mandatory and must be unique.
  • Needs to return any errors.

query-snapshots

-> {"execute": "query-snapshots"}
<- {
"return": [
   {           
     "name": "vm_before_upgrade",
     "parent_name": "",
     "id": "3fab28c9-a6bd-4659-b9bc-683780d8a2d5",
     "parent_id": "00000000-0000-0000-0000-000000000000",
     "date": 20100415164856
   },
   {           
     "name": "vm_before_upgrade-2",
     "parent_name": "vm_before_upgrade",
     "id": "e693dce2-6e12-4474-9bbf-2b3f97323423",
     "parent_id": "00000000-0000-0000-0000-000000000000",
     "date": 20100415100000
   }
  ]        
}
  • Do we really need to expose IDs or names should be enough?
  • Does JSON has a date type of just using a integer like YYYYMMDDHHMMSS would be better?

Instead of a list, how about this:

-> {"execute": "query-snapshots"}
<- {
"return": [
   {           
      "vm_before_upgrade": { 
		      "id": "3fab28c9-a6bd-4659-b9bc-683780d8a2d5", 
		      "date": 20100415164856, 
		      "child": { "vm_before_upgrade-2": 
  					"id": "e693dce2-6e12-4474-9bbf-2b3f97323423",
 				        "date": 20100415100000 
			       }
                           }
        "random-backup": {
 		      "id": "abcdefcc-a6bd-4659-b9bc-683780d8a2d5",
 		      "date": 20111111111111
                         }
     }
  ]