[Qemu-devel] block-layer: questions on manipulation of internal nodes

[Qemu-devel] block-layer: questions on manipulation of internal nodes

[Stefano Panella]

Hi everybody,

I am a relatively new user of qemu block layer. I am interested in it mainly because it looks very powerful and general and I am hoping to integrate it on our product and to contribute to it for new usecases.

I have existing use cases where we work with a model of a disk process per VM disk and I am experimenting with qemu and qmp to  build something similar.

At the moment I have managed to build a new binary, called qemu-dp (probably should be called qemu-bl for block layer) which is basically starting as a qmp server and accepting qmp block layer commands to operate on disks.

just to give you an example this is the kind of thing I am doing:

EXTERNALLY:
/usr/lib64/xen/bin/qemu-img create -f qcow2 -o size=1M /root/a
/usr/lib64/xen/bin/qemu-img create -f qcow2 -b /root/a -o size=1M /root/b
/usr/lib64/xen/bin/qemu-img create -f qcow2 -b /root/b -o size=1M /root/c
/usr/lib64/xen/bin/qemu-img create -f qcow2 -b /root/c -o size=1M /root/d
/usr/lib64/xen/bin/qemu-img create -f qcow2 -b /root/d -o size=1M /root/e

let's assume there were some data in every layer....

Than:

USING QMP:
{ "execute": "qmp_capabilities" }
{
	"execute": "blockdev-add",  
	"arguments": {
		"driver": "qcow2", 
		"node-name": "qemu_node",
		"discard": "unmap",
		"cache": {
			"direct": true
		},
		"file": {
			"driver": "file",
			"filename": "/root/e"
		}
	}
}

{
	"execute": "nbd-server-start",
	"arguments": { 
		"addr": {
			"type": "unix",
			"data": {
				"path": "/tmp/nbd.test1"
			}
		}
	}
}

{
	"execute": "nbd-server-add",
	"arguments": { 
		"device": "qemu_node",
		"writable": true
	}
}

after this the chain looks like:

a < b < c < d < e < NBD_server

now I make a full copy of b and c which I call b1 and c1 and for example I run externally qemu-img commit c1 -> b1 while qemu-dp has still the chain opened.

I would now like to send a qmp command to tell qemu-dp to hold any IO from the NBD_server and forget about a, b, c and insert b1 as d's child, like this:

a < b1 < d < e < NBD_server

I have tried to implement this qmp command and looked at 

qmp_change_backing_file()
qmp_x_blockdev_change()
https://lists.gnu.org/archive/html/qemu-devel/2017-08/msg02660.html

but did not figure out a way of doing that yet...

I suspect my problem is that I am still very confused about the semantics of the object model in the block layer, the ref counting, the graph manipulation, the monitor etc. etc.

I have tried to have some interactive chats on irc and they have been very useful so far (thanks again stefanha, kwolf, berto, eblake) but maybe a proper email would be a good starting point as stefanha has suggested.

Please if somebody could point me to a bit of code to achieve my example that would be great, otherwise if there is no code for that kind of functionality, it would be good to have a little guide on the sequence of block primiteve I should call and on which node, including refs, locking, drain, caches, reopen etc...

Thanks a lot,

Stefano

Re: [Qemu-devel] block-layer: questions on manipulation of internal nodes

[Stefan Hajnoczi]

[-- Attachment #1: Type: text/plain, Size: 3742 bytes --]

On Wed, Mar 14, 2018 at 03:07:21PM +0000, Stefano Panella wrote:
> Hi everybody,
> 
> I am a relatively new user of qemu block layer. I am interested in it mainly because it looks very powerful and general and I am hoping to integrate it on our product and to contribute to it for new usecases.
> 
> I have existing use cases where we work with a model of a disk process per VM disk and I am experimenting with qemu and qmp to  build something similar.
> 
> At the moment I have managed to build a new binary, called qemu-dp (probably should be called qemu-bl for block layer) which is basically starting as a qmp server and accepting qmp block layer commands to operate on disks.
> 
> just to give you an example this is the kind of thing I am doing:
> 
> EXTERNALLY:
> /usr/lib64/xen/bin/qemu-img create -f qcow2 -o size=1M /root/a
> /usr/lib64/xen/bin/qemu-img create -f qcow2 -b /root/a -o size=1M /root/b
> /usr/lib64/xen/bin/qemu-img create -f qcow2 -b /root/b -o size=1M /root/c
> /usr/lib64/xen/bin/qemu-img create -f qcow2 -b /root/c -o size=1M /root/d
> /usr/lib64/xen/bin/qemu-img create -f qcow2 -b /root/d -o size=1M /root/e
> 
> let's assume there were some data in every layer....
> 
> Than:
> 
> USING QMP:
> { "execute": "qmp_capabilities" }
> {
> 	"execute": "blockdev-add",  
> 	"arguments": {
> 		"driver": "qcow2", 
> 		"node-name": "qemu_node",
> 		"discard": "unmap",
> 		"cache": {
> 			"direct": true
> 		},
> 		"file": {
> 			"driver": "file",
> 			"filename": "/root/e"
> 		}
> 	}
> }
> 
> {
> 	"execute": "nbd-server-start",
> 	"arguments": { 
> 		"addr": {
> 			"type": "unix",
> 			"data": {
> 				"path": "/tmp/nbd.test1"
> 			}
> 		}
> 	}
> }
> 
> {
> 	"execute": "nbd-server-add",
> 	"arguments": { 
> 		"device": "qemu_node",
> 		"writable": true
> 	}
> }
> 
> after this the chain looks like:
> 
> a < b < c < d < e < NBD_server
> 
> now I make a full copy of b and c which I call b1 and c1 and for example I run externally qemu-img commit c1 -> b1 while qemu-dp has still the chain opened.
> 
> I would now like to send a qmp command to tell qemu-dp to hold any IO from the NBD_server and forget about a, b, c and insert b1 as d's child, like this:
> 
> a < b1 < d < e < NBD_server

Holding I/O is done via bdrv_drained_begin/end() and most block-related
monitor commands should use it.

> I have tried to implement this qmp command and looked at 
> 
> qmp_change_backing_file()
> qmp_x_blockdev_change()
> https://lists.gnu.org/archive/html/qemu-devel/2017-08/msg02660.html
> 
> but did not figure out a way of doing that yet...
> 
> I suspect my problem is that I am still very confused about the semantics of the object model in the block layer, the ref counting, the graph manipulation, the monitor etc. etc.
> 
> I have tried to have some interactive chats on irc and they have been very useful so far (thanks again stefanha, kwolf, berto, eblake) but maybe a proper email would be a good starting point as stefanha has suggested.
> 
> Please if somebody could point me to a bit of code to achieve my example that would be great, otherwise if there is no code for that kind of functionality, it would be good to have a little guide on the sequence of block primiteve I should call and on which node, including refs, locking, drain, caches, reopen etc...

blockdev-add and related APIs are still under development and incomplete.

Manos' block-insert-node is probably closest to what you need, so I
would start there.

I'm not clear enough on what exactly is necessary on top of Manos' patch
without doing the work myself, so I'm afraid I can't give a detailed
sequence of steps you need to take.

Stefan

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 455 bytes --]