Hi Everyone,
We are working towards building on the virtualization management functionality in cockpit (http://cockpit-project.org/) and wanted to get some feedback on the best way to integrate with libvirt.
As a quick overview, cockpit aims to talk to existing remotable system APIs. Usually these API’s take the form of dbus, REST or executable commands. The majority of cockpit is implemented in javascript. There is no cockpit backend that knows how to change a hostname for example. The cockpit backend knows how to handle a dbus payload. The javascript running in the users browser knows how to use the systemd dbus API at org.freedesktop.hostname1 to manage the system hostname.
Right now some of the basics have been implemented by spawning commands on the system. This isn't ideal because it involves parsing / screen scraping output and doesn't support receiving events so we have to poll (ei run the command again) to keep the UI up to date.
As far as I know libvirt doesn't currently have a remoteable API. It does have a daemon that communicates with clients via a XDR RPC. (https://libvirt.org/internals/rpc.html) However from what I'm hearing the RPC is considered an internal implementation and shouldn't be used by external applications. Is that still the case? Is there any chance of getting talking the daemon directly using the XDR standard for a subset of methods blessed as part of the externally supported API?
An alternative is to implement a standards based remotable API, using something like dbus or REST, that can be used by external applications. I imagine that this would be at a bit of a higher level than the current RPC and contain at least some of the logic around the actions it performs rather than being a direct passthrough to the daemon.
Of course that is a pretty big undertaking and would, in my opinion, only be worth it if there is broader interest in the community and use cases beyond what cockpit would like to.
On Thu, May 25, 2017 at 10:26:47AM -0700, Peter wrote:
As far as I know libvirt doesn't currently have a remoteable API. It does have a daemon that communicates with clients via a XDR RPC. (https://libvirt.org/internals/rpc.html) However from what I'm hearing the RPC is considered an internal implementation and shouldn't be used by external applications. Is that still the case? Is there any chance of getting talking the daemon directly using the XDR standard for a subset of methods blessed as part of the externally supported API?
That's correct - the XDR protocol is a private implementation detail between the libvirtd daemon and libvirt client library. The /only/ supported way of using libvirt is via the client library API, or one of the many language bindings written on top of it. A number of the libvirt virtualization drivers are implemented entirely in the client library and don't involve libvirtd at all.
An alternative is to implement a standards based remotable API, using something like dbus or REST, that can be used by external applications. I imagine that this would be at a bit of a higher level than the current RPC and contain at least some of the logic around the actions it performs rather than being a direct passthrough to the daemon.
Of course that is a pretty big undertaking and would, in my opinion, only be worth it if there is broader interest in the community and use cases beyond what cockpit would like to.
I imagine any DBus API would be a pretty straightforward 1-1 mapping from the C api to DBus methods & signals. For a REST API I think you would quite possibly want to build something higher level which will involve more design thought.
Regards, Daniel
On Thu, May 25, 2017 at 10:16:26AM -0700, Peter Volpe wrote:
Hi Everyone,
We are working towards building on the virtualization management functionality in cockpit (http://cockpit-project.org/) and wanted to get some feedback on the best way to integrate with libvirt.
As a quick overview, cockpit aims to talk to existing remotable system APIs. Usually these API’s take the form of dbus, REST or executable commands. The majority of cockpit is implemented in javascript. There is no cockpit backend that knows how to change a hostname for example. The cockpit backend knows how to handle a dbus payload. The javascript running in the users browser knows how to use the systemd dbus API at org.freedesktop.hostname1 to manage the system hostname.
Right now some of the basics have been implemented by spawning commands on the system. This isn't ideal because it involves parsing / screen scraping output and doesn't support receiving events so we have to poll (ei run the command again) to keep the UI up to date.
As far as I know libvirt doesn't currently have a remoteable API. It does have a daemon that communicates with clients via a XDR RPC. (https://libvirt.org/internals/rpc.html) However from what I'm hearing the RPC is considered an internal implementation and shouldn't be used by external applications. Is that still the case? Is there any chance of getting talking the daemon directly using the XDR standard for a subset of methods blessed as part of the externally supported API?
If we standardize even the smallest part of the RPC, then it might screw us immediately. We are keeping it private just so we can enhance the APIs we already have. We don't know when we will need to change some part of it.
An alternative is to implement a standards based remotable API, using something like dbus or REST, that can be used by external applications. I imagine that this would be at a bit of a higher level than the current RPC and contain at least some of the logic around the actions it performs rather than being a direct passthrough to the daemon.
Of course that is a pretty big undertaking and would, in my opinion, only be worth it if there is broader interest in the community and use cases beyond what cockpit would like to.
I was thinking of adding a REST API to libvirt for some time, actually. It was supposed to be April fools patch because I didn't know about many people who would actually like that. But it could've been made so that there doesn't need to be much updated if new API arises. And there are lot of things already implemented for JSON and the structures and types we already use in libvirt. I never got around to doing that, though.
Nowadays, when thinking about this, there could be another layer of abstraction between libvirt and consumers, and I understand REST would fit a lot of them. Even though there would be some code duplication from what already exists in libvirt, it might be worth it, I don't know. I'm also not completely sure how much should be abstracted, but I feel like not only cockpit, but oVirt and OpenStack (heck, maybe even virt-manager) could benefit from such tool as well. Maybe all of you could weigh in. I might, actually, be really interested in such a project. That is if it shows that there is use for it (we all have enough side projects that there is little to no use for them, right ;)), so feel free to discuss the ideas to more detail.
Have a nice day, Martin
P.S.: Since you started the discussion as the first one, I would suggest paying homage to the cockpit project by naming it CockVirt.
On 05/26/2017 02:11 AM, Martin Kletzander wrote:
On Thu, May 25, 2017 at 10:16:26AM -0700, Peter Volpe wrote:
Hi Everyone,
We are working towards building on the virtualization management functionality in cockpit (http://cockpit-project.org/) and wanted to get some feedback on the best way to integrate with libvirt.
As a quick overview, cockpit aims to talk to existing remotable system APIs. Usually these API’s take the form of dbus, REST or executable commands. The majority of cockpit is implemented in javascript. There is no cockpit backend that knows how to change a hostname for example. The cockpit backend knows how to handle a dbus payload. The javascript running in the users browser knows how to use the systemd dbus API at org.freedesktop.hostname1 to manage the system hostname.
Right now some of the basics have been implemented by spawning commands on the system. This isn't ideal because it involves parsing / screen scraping output and doesn't support receiving events so we have to poll (ei run the command again) to keep the UI up to date.
As far as I know libvirt doesn't currently have a remoteable API. It does have a daemon that communicates with clients via a XDR RPC. (https://libvirt.org/internals/rpc.html) However from what I'm hearing the RPC is considered an internal implementation and shouldn't be used by external applications. Is that still the case? Is there any chance of getting talking the daemon directly using the XDR standard for a subset of methods blessed as part of the externally supported API?
If we standardize even the smallest part of the RPC, then it might screw us immediately. We are keeping it private just so we can enhance the APIs we already have. We don't know when we will need to change some part of it.
How does the current ssh implementation work? (https://libvirt.org/remote.html) If clients are able to talk to a remote libvirtd via ssh then there must be some sort of compatibility guarantee. Otherwise unless the versions are exactly the same clients wouldn't be able to talk to remote daemons.
On Tue, Jun 06, 2017 at 07:45:10 -0700, Peter wrote:
On 05/26/2017 02:11 AM, Martin Kletzander wrote:
On Thu, May 25, 2017 at 10:16:26AM -0700, Peter Volpe wrote:
[...]
If we standardize even the smallest part of the RPC, then it might screw us immediately. We are keeping it private just so we can enhance the APIs we already have. We don't know when we will need to change some part of it.
How does the current ssh implementation work? (https://libvirt.org/remote.html) If clients are able to talk to a remote libvirtd via ssh then there must be some sort of compatibility guarantee. Otherwise unless the versions are exactly the same clients wouldn't be able to talk to remote daemons.
They use the internal RPC protocol transported over SSH. The client library initializes the ssh connection to the server, and then starts to talk the RPC tunelled over the SSH session.
The compatibility is guaranteed only if you use the client library. As said, the RPC protocol is considered an internal detail and the client library shields you from a possible incompatibility if we'd ever make incompatible change.
On Tue, Jun 06, 2017 at 05:11:55PM +0200, Peter Krempa wrote:
On Tue, Jun 06, 2017 at 07:45:10 -0700, Peter wrote:
On 05/26/2017 02:11 AM, Martin Kletzander wrote:
On Thu, May 25, 2017 at 10:16:26AM -0700, Peter Volpe wrote:
[...]
If we standardize even the smallest part of the RPC, then it might screw us immediately. We are keeping it private just so we can enhance the APIs we already have. We don't know when we will need to change some part of it.
How does the current ssh implementation work? (https://libvirt.org/remote.html) If clients are able to talk to a remote libvirtd via ssh then there must be some sort of compatibility guarantee. Otherwise unless the versions are exactly the same clients wouldn't be able to talk to remote daemons.
They use the internal RPC protocol transported over SSH. The client library initializes the ssh connection to the server, and then starts to talk the RPC tunelled over the SSH session.
The compatibility is guaranteed only if you use the client library. As said, the RPC protocol is considered an internal detail and the client library shields you from a possible incompatibility if we'd ever make incompatible change.
Ignoring wire compatibility questions, it is important to note that not all functionality in libvirt is done in libvirtd. There are libvirt drivers that are entirely implemented in the library, not libvirtd. For some of the libvirt drivers that do use libvirtd, there is also still some functionality that is client side.
Regards, Daniel
On 06.06.2017 17:17, Daniel P. Berrange wrote:
On Tue, Jun 06, 2017 at 05:11:55PM +0200, Peter Krempa wrote:
On Tue, Jun 06, 2017 at 07:45:10 -0700, Peter wrote:
On 05/26/2017 02:11 AM, Martin Kletzander wrote:
On Thu, May 25, 2017 at 10:16:26AM -0700, Peter Volpe wrote:
[...]
If we standardize even the smallest part of the RPC, then it might screw us immediately. We are keeping it private just so we can enhance the APIs we already have. We don't know when we will need to change some part of it.
How does the current ssh implementation work? (https://libvirt.org/remote.html) If clients are able to talk to a remote libvirtd via ssh then there must be some sort of compatibility guarantee. Otherwise unless the versions are exactly the same clients wouldn't be able to talk to remote daemons.
They use the internal RPC protocol transported over SSH. The client library initializes the ssh connection to the server, and then starts to talk the RPC tunelled over the SSH session.
The compatibility is guaranteed only if you use the client library. As said, the RPC protocol is considered an internal detail and the client library shields you from a possible incompatibility if we'd ever make incompatible change.
Ignoring wire compatibility questions, it is important to note that not all functionality in libvirt is done in libvirtd. There are libvirt drivers that are entirely implemented in the library, not libvirtd. For some of the libvirt drivers that do use libvirtd, there is also still some functionality that is client side.
So to summarize what I'm hearing here:
* There is no libvirt remotable API today. * All libvirt APIs are only able to be used in-process. * There is presently no ability for a libvirt API to be used from a browser, or non-Linux process or non-Linux system.
In order for Cockpit to interact with libvirt, it would need to grow a remoteable API. The Cockpit project could likely jumpstart such an effort, by lightly wrapping the C API in DBus ... and contribute that remoteable API code to the libvirt project.
Cheers,
Stef
On Tue, Jun 06, 2017 at 06:07:21PM +0200, Stef Walter wrote:
On 06.06.2017 17:17, Daniel P. Berrange wrote:
On Tue, Jun 06, 2017 at 05:11:55PM +0200, Peter Krempa wrote:
On Tue, Jun 06, 2017 at 07:45:10 -0700, Peter wrote:
On 05/26/2017 02:11 AM, Martin Kletzander wrote:
On Thu, May 25, 2017 at 10:16:26AM -0700, Peter Volpe wrote:
[...]
If we standardize even the smallest part of the RPC, then it might screw us immediately. We are keeping it private just so we can enhance the APIs we already have. We don't know when we will need to change some part of it.
How does the current ssh implementation work? (https://libvirt.org/remote.html) If clients are able to talk to a remote libvirtd via ssh then there must be some sort of compatibility guarantee. Otherwise unless the versions are exactly the same clients wouldn't be able to talk to remote daemons.
They use the internal RPC protocol transported over SSH. The client library initializes the ssh connection to the server, and then starts to talk the RPC tunelled over the SSH session.
The compatibility is guaranteed only if you use the client library. As said, the RPC protocol is considered an internal detail and the client library shields you from a possible incompatibility if we'd ever make incompatible change.
Ignoring wire compatibility questions, it is important to note that not all functionality in libvirt is done in libvirtd. There are libvirt drivers that are entirely implemented in the library, not libvirtd. For some of the libvirt drivers that do use libvirtd, there is also still some functionality that is client side.
So to summarize what I'm hearing here:
- There is no libvirt remotable API today.
Libvirt drivers that are implemented in libvirtd are accessible remotely using the libvirt client library. We just consider the wire protocol a private impl detail between them. Drivers implemented outside libvirtd are often accessible remotely too, via the hypervisor's native remoting system.
- All libvirt APIs are only able to be used in-process.
- There is presently no ability for a libvirt API to be used from a browser, or non-Linux process or non-Linux system.
Actually libvirt can be built on OS-X, Windows and FreeBSD too.
In order for Cockpit to interact with libvirt, it would need to grow a remoteable API. The Cockpit project could likely jumpstart such an effort, by lightly wrapping the C API in DBus ... and contribute that remoteable API code to the libvirt project.
FWIW, if someone does want to build a DBus wrapper around libvirt, that would be something todo as a separate code module above the libvirt API. We would be happy to host such an effort on libvirt.org git though, alongside other higher level API wrappers. Likewise for a REST API wrapper.
Regards, Daniel
On 06.06.2017 12:13, Daniel P. Berrange wrote:
On Tue, Jun 06, 2017 at 06:07:21PM +0200, Stef Walter wrote:
On 06.06.2017 17:17, Daniel P. Berrange wrote:
On Tue, Jun 06, 2017 at 05:11:55PM +0200, Peter Krempa wrote:
On Tue, Jun 06, 2017 at 07:45:10 -0700, Peter wrote:
On 05/26/2017 02:11 AM, Martin Kletzander wrote:
On Thu, May 25, 2017 at 10:16:26AM -0700, Peter Volpe wrote:
[...]
If we standardize even the smallest part of the RPC, then it might screw us immediately. We are keeping it private just so we can enhance the APIs we already have. We don't know when we will need to change some part of it.
How does the current ssh implementation work? (https://libvirt.org/remote.html) If clients are able to talk to a remote libvirtd via ssh then there must be some sort of compatibility guarantee. Otherwise unless the versions are exactly the same clients wouldn't be able to talk to remote daemons.
They use the internal RPC protocol transported over SSH. The client library initializes the ssh connection to the server, and then starts to talk the RPC tunelled over the SSH session.
The compatibility is guaranteed only if you use the client library. As said, the RPC protocol is considered an internal detail and the client library shields you from a possible incompatibility if we'd ever make incompatible change.
Ignoring wire compatibility questions, it is important to note that not all functionality in libvirt is done in libvirtd. There are libvirt drivers that are entirely implemented in the library, not libvirtd. For some of the libvirt drivers that do use libvirtd, there is also still some functionality that is client side.
So to summarize what I'm hearing here:
- There is no libvirt remotable API today.
Libvirt drivers that are implemented in libvirtd are accessible remotely using the libvirt client library. We just consider the wire protocol a private impl detail between them. Drivers implemented outside libvirtd are often accessible remotely too, via the hypervisor's native remoting system.
I believe those are implementation details, not and remoteable API. Libvirt and its API must always be used in process today? Even though it RPC may be in play, it is an implementation detail, not API. Correct?
- All libvirt APIs are only able to be used in-process.
- There is presently no ability for a libvirt API to be used from a browser, or non-Linux process or non-Linux system.
Actually libvirt can be built on OS-X, Windows and FreeBSD too.
Ah good to know. But the API is always used in process, correct?
In order for Cockpit to interact with libvirt, it would need to grow a remoteable API. The Cockpit project could likely jumpstart such an effort, by lightly wrapping the C API in DBus ... and contribute that remoteable API code to the libvirt project.
FWIW, if someone does want to build a DBus wrapper around libvirt, that would be something todo as a separate code module above the libvirt API. We would be happy to host such an effort on libvirt.org git though, alongside other higher level API wrappers. Likewise for a REST API wrapper.
This would be a useful division of labor between Cockpit consuming such a remotable API ... and the remoteable API being hosted and maintained by the libvirt folks ... although the Cockpit project could help get it started.
Stef
On Tue, Jun 06, 2017 at 01:25:08PM -0400, Stef Walter wrote:
On 06.06.2017 12:13, Daniel P. Berrange wrote:
On Tue, Jun 06, 2017 at 06:07:21PM +0200, Stef Walter wrote:
So to summarize what I'm hearing here:
- There is no libvirt remotable API today.
Libvirt drivers that are implemented in libvirtd are accessible remotely using the libvirt client library. We just consider the wire protocol a private impl detail between them. Drivers implemented outside libvirtd are often accessible remotely too, via the hypervisor's native remoting system.
I believe those are implementation details, not and remoteable API. Libvirt and its API must always be used in process today? Even though it RPC may be in play, it is an implementation detail, not API. Correct?
I don't really know what distinction your trying to make by saying "used in process". Libvirt provides an ELF library and apps link to it to use it, or use it indirectly via a higher level language specific wrapper.
- All libvirt APIs are only able to be used in-process.
- There is presently no ability for a libvirt API to be used from a browser, or non-Linux process or non-Linux system.
Actually libvirt can be built on OS-X, Windows and FreeBSD too.
Ah good to know. But the API is always used in process, correct?
Again I don't know what you're trying to say here. It is an ELF library and you link to it and call APIs as you would for any other library.
In order for Cockpit to interact with libvirt, it would need to grow a remoteable API. The Cockpit project could likely jumpstart such an effort, by lightly wrapping the C API in DBus ... and contribute that remoteable API code to the libvirt project.
FWIW, if someone does want to build a DBus wrapper around libvirt, that would be something todo as a separate code module above the libvirt API. We would be happy to host such an effort on libvirt.org git though, alongside other higher level API wrappers. Likewise for a REST API wrapper.
This would be a useful division of labor between Cockpit consuming such a remotable API ... and the remoteable API being hosted and maintained by the libvirt folks ... although the Cockpit project could help get it started.
FWIW, in our RPC system, we auto-generate some of the marshalling code from the API description XML file in /usr/share/libvirt/libvirt-api.xml Some APIs are not friendly for auto-generating though, so probably get about 75% auto-generated and the rest need to be done by hand.
Regards, Daniel
On Thu, May 25, 2017 at 10:26:47AM -0700, Peter wrote:
The majority of cockpit is implemented in javascript.
How about using the gobject libvirt bindings?
https://libvirt.org/git/?p=libvirt-glib.git;a=summary
They should be usable from Javascript directly, as in the .js example here:
https://libvirt.org/git/?p=libvirt-glib.git;a=tree;f=examples;h=d63d5964be22...
Rich.
On Wed, May 31, 2017 at 03:59:10PM +0100, Richard W.M. Jones wrote:
On Thu, May 25, 2017 at 10:26:47AM -0700, Peter wrote:
The majority of cockpit is implemented in javascript.
How about using the gobject libvirt bindings?
https://libvirt.org/git/?p=libvirt-glib.git;a=summary
They should be usable from Javascript directly, as in the .js example here:
https://libvirt.org/git/?p=libvirt-glib.git;a=tree;f=examples;h=d63d5964be22...
They're usable from a standalone javascript interpretor, but there's no way to use them from a client side javascript interpretor in the user's web browser.
Regards, Daniel
On Wed, May 31, 2017 at 04:02:42PM +0100, Daniel P. Berrange wrote:
On Wed, May 31, 2017 at 03:59:10PM +0100, Richard W.M. Jones wrote:
On Thu, May 25, 2017 at 10:26:47AM -0700, Peter wrote:
The majority of cockpit is implemented in javascript.
How about using the gobject libvirt bindings?
https://libvirt.org/git/?p=libvirt-glib.git;a=summary
They should be usable from Javascript directly, as in the .js example here:
https://libvirt.org/git/?p=libvirt-glib.git;a=tree;f=examples;h=d63d5964be22...
They're usable from a standalone javascript interpretor, but there's no way to use them from a client side javascript interpretor in the user's web browser.
Hmm OK, I thought "implemented in javascript" meant they were using server-side Javascript.
I see that cockpit runs two processes on the server (cockpit-ws and ssh-agent).
I had a look at one of the modules in the source (pkg/systemd). It's running local commands (eg. "grep"), and issuing dbus calls which AFAIK cannot be issued over the network.
Can Peter explain a bit more about the architecture of Cockpit? Where does the Javascript run? What JS engine runs it?
Rich.
On 05/31/2017 08:52 AM, Richard W.M. Jones wrote:
On Wed, May 31, 2017 at 04:02:42PM +0100, Daniel P. Berrange wrote:
On Wed, May 31, 2017 at 03:59:10PM +0100, Richard W.M. Jones wrote:
On Thu, May 25, 2017 at 10:26:47AM -0700, Peter wrote:
The majority of cockpit is implemented in javascript.
How about using the gobject libvirt bindings?
https://libvirt.org/git/?p=libvirt-glib.git;a=summary
They should be usable from Javascript directly, as in the .js example here:
https://libvirt.org/git/?p=libvirt-glib.git;a=tree;f=examples;h=d63d5964be22...
They're usable from a standalone javascript interpretor, but there's no way to use them from a client side javascript interpretor in the user's web browser.
Hmm OK, I thought "implemented in javascript" meant they were using server-side Javascript.
I see that cockpit runs two processes on the server (cockpit-ws and ssh-agent).
I had a look at one of the modules in the source (pkg/systemd). It's running local commands (eg. "grep"), and issuing dbus calls which AFAIK cannot be issued over the network.
Can Peter explain a bit more about the architecture of Cockpit? Where does the Javascript run? What JS engine runs it?
Rich.
See this paragraph from my original message.
As a quick overview, cockpit aims to talk to existing remotable system APIs. Usually these API’s take the form of dbus, REST or executable commands. The majority of cockpit is implemented in javascript. There is no cockpit backend that knows how to change a hostname for example. The cockpit backend knows how to handle a dbus payload. The javascript running in the users browser knows how to use the systemd dbus API at org.freedesktop.hostname1 to manage the system hostname.
The javascript is always run in the users browser. The dbus calls or system commands are sent by the javascript via a websocket to cockpit-ws. It then forwards those messages on to the correct cockpit-bridge. Based on the payload the bridge knows how communicate with the appropriate system API. (See https://github.com/cockpit-project/cockpit/blob/master/doc/protocol.md or run 'cockpit-bridge --version' for a list).
So in keeping with the cockpit ideals cockpit-bridge should not know how to start a virtual machine or link to the libvirt library. It should receive a payload that when properly executed by the bridge results in the machine starting. Right now this is only possible using the virsh type commands that have the problems discussed earlier. I think it's clear that the existing RPC is not an option. So that pretty much leaves us with the choice to roll or own solution or to add a more generic dbus or rest wrapper for libvirt.
On Wed, May 31, 2017 at 09:22:41AM -0700, Peter wrote:
The javascript is always run in the users browser. The dbus calls or system commands are sent by the javascript via a websocket to cockpit-ws. It then forwards those messages on to the correct cockpit-bridge. Based on the payload the bridge knows how communicate with the appropriate system API. (See https://github.com/cockpit-project/cockpit/blob/master/doc/protocol.md or run 'cockpit-bridge --version' for a list).
So in keeping with the cockpit ideals cockpit-bridge should not know how to start a virtual machine or link to the libvirt library. It should receive a payload that when properly executed by the bridge results in the machine starting. Right now this is only possible using the virsh type commands that have the problems discussed earlier. I think it's clear that the existing RPC is not an option. So that pretty much leaves us with the choice to roll or own solution or to add a more generic dbus or rest wrapper for libvirt.
Thanks - that's a bit clearer now.
I agree with others that as things stand you will need a REST or DBus or similar API added to libvirt.
However have you considered using gobject-introspection to generate new "Payload" types automatically?
Rich.
On 05/31/2017 01:00 PM, Richard W.M. Jones wrote:
On Wed, May 31, 2017 at 09:22:41AM -0700, Peter wrote:
The javascript is always run in the users browser. The dbus calls or system commands are sent by the javascript via a websocket to cockpit-ws. It then forwards those messages on to the correct cockpit-bridge. Based on the payload the bridge knows how communicate with the appropriate system API. (See https://github.com/cockpit-project/cockpit/blob/master/doc/protocol.md or run 'cockpit-bridge --version' for a list).
So in keeping with the cockpit ideals cockpit-bridge should not know how to start a virtual machine or link to the libvirt library. It should receive a payload that when properly executed by the bridge results in the machine starting. Right now this is only possible using the virsh type commands that have the problems discussed earlier. I think it's clear that the existing RPC is not an option. So that pretty much leaves us with the choice to roll or own solution or to add a more generic dbus or rest wrapper for libvirt.
Thanks - that's a bit clearer now.
I agree with others that as things stand you will need a REST or DBus or similar API added to libvirt.
One thing to consider though is that the libvirt API is only part of the equation. The biggest part for sure, but if cockpit wants to create and manage libvirt XML it should eventually use libvirt-gconfig, which is just a standalone bit of code for manipulating the XML, and possibly libosinfo and other libvirt-* libraries. Dan's writeup here describes the idea behind the different libraries:
https://www.berrange.com/posts/2011/11/22/introducing-the-libvirt-glib-a-map... https://www.redhat.com/archives/libvir-list/2012-September/msg00325.html
Though the only ones that have had any real development live in libvirt-glib.git: libvirt-glib, libvirt-gconfig, libvirt-gobject. And they would likely need to be extended to fit cockpit's goals, but it's better than writing all that XML building from scratch.
However have you considered using gobject-introspection to generate new "Payload" types automatically?
This is an interesting idea (though I have no clue if it's technically feasible). But if it worked it could be used to cover libvirt-{gobject,glib,gconfig} and libosinfo which is likely everything cockpit will ever need for virt management.
- Cole
Hello Richard,
Richard W.M. Jones [2017-05-31 18:00 +0100]:
I agree with others that as things stand you will need a REST or DBus or similar API added to libvirt.
However have you considered using gobject-introspection to generate new "Payload" types automatically?
This doesn't fundamentally change the picture here. Our JavaScript runs in the browser, so on that side GI doesn't help. What we could do is to dynamically create some code in the Cockpit module, send it over to the remote machine, and have it be executed. This could be in Python, which then assumes that python-libvirt is installed there, or in JS which then assumes that libvirt-glib and gjs are installed. The latter seems like a stronger assumption on a server even, but structurally these are pretty similiar.
I. e. GI is not a magic thing to make an API remotable, it just makes it bindable to different programming languages.
C/GI interfaces also don't map well to D-Bus, i. e. it's not practical to autogenerate a D-Bus interface for a given GI API. This still works for the most simple methods that only accept primitive data types and strings, but as soon as you pass any structs, pointers, function pointers etc. around this can't be exposed though D-Bus any more without further interpretation on the server side.
Martin
On 07.06.2017 07:49, Martin Pitt wrote:
Hello Richard,
Richard W.M. Jones [2017-05-31 18:00 +0100]:
I agree with others that as things stand you will need a REST or DBus or similar API added to libvirt.
However have you considered using gobject-introspection to generate new "Payload" types automatically?
This doesn't fundamentally change the picture here. Our JavaScript runs in the browser, so on that side GI doesn't help. What we could do is to dynamically create some code in the Cockpit module, send it over to the remote machine, and have it be executed. This could be in Python, which then assumes that python-libvirt is installed there, or in JS which then assumes that libvirt-glib and gjs are installed. The latter seems like a stronger assumption on a server even, but structurally these are pretty similiar.
I. e. GI is not a magic thing to make an API remotable, it just makes it bindable to different programming languages.
C/GI interfaces also don't map well to D-Bus, i. e. it's not practical to autogenerate a D-Bus interface for a given GI API. This still works for the most simple methods that only accept primitive data types and strings, but as soon as you pass any structs, pointers, function pointers etc. around this can't be exposed though D-Bus any more without further interpretation on the server side.
Perhaps not an runtime. But it seems like such a thing could be done at compile time for the libvirt C API. The libvirt folks do it for their python bindings, using code generation.
The libvirt Python API behaves a lot like the C API. It's very flat and not very "pythonic". Which is actually a nice target for us. We could do something similar with a simple "flat" DBus wrapper of the API ... using code generation.
What do you think?
Stef
Stef Walter [2017-06-07 10:47 +0200]:
On 07.06.2017 07:49, Martin Pitt wrote:
C/GI interfaces also don't map well to D-Bus, i. e. it's not practical to autogenerate a D-Bus interface for a given GI API. This still works for the most simple methods that only accept primitive data types and strings, but as soon as you pass any structs, pointers, function pointers etc. around this can't be exposed though D-Bus any more without further interpretation on the server side.
Perhaps not an runtime. But it seems like such a thing could be done at compile time for the libvirt C API. The libvirt folks do it for their python bindings, using code generation.
The libvirt Python API behaves a lot like the C API. It's very flat and not very "pythonic". Which is actually a nice target for us. We could do something similar with a simple "flat" DBus wrapper of the API ... using code generation.
Sure, one can always write code to map the libvirt API into a D-Bus API (in C, Python, or JavaScript with GJS, but I suppose for this use case Python might be the best choice). My point was just that this would require to create, package, and ship new code on the target machines, which is a very high barrier for stable OS releases (particularly RHEL/CentOS, but it's not easy at all for e. g. Debian or Ubuntu either).
Peter and I discussed if we could get away somehow with the existing code/APIs on stable releases. E. g. a hackish, but maybe workable first step might be to only use the python-libvirt API for subscribing to desired notifications and trigger the Cockpit VM module through it, but continue to handle all the actions through virsh. This subscription/notification code will hopefully be small enough that it can be sent "through the wire" (cockpit channel) and executed ad-hoc, without having to actually distribute it on the remote side.
Martin
On 07.06.2017 10:55, Martin Pitt wrote:
Stef Walter [2017-06-07 10:47 +0200]:
On 07.06.2017 07:49, Martin Pitt wrote:
C/GI interfaces also don't map well to D-Bus, i. e. it's not practical to autogenerate a D-Bus interface for a given GI API. This still works for the most simple methods that only accept primitive data types and strings, but as soon as you pass any structs, pointers, function pointers etc. around this can't be exposed though D-Bus any more without further interpretation on the server side.
Perhaps not an runtime. But it seems like such a thing could be done at compile time for the libvirt C API. The libvirt folks do it for their python bindings, using code generation.
The libvirt Python API behaves a lot like the C API. It's very flat and not very "pythonic". Which is actually a nice target for us. We could do something similar with a simple "flat" DBus wrapper of the API ... using code generation.
Sure, one can always write code to map the libvirt API into a D-Bus API (in C, Python, or JavaScript with GJS, but I suppose for this use case Python might be the best choice). My point was just that this would require to create, package, and ship new code on the target machines, which is a very high barrier for stable OS releases (particularly RHEL/CentOS, but it's not easy at all for e. g. Debian or Ubuntu either).
Peter and I discussed if we could get away somehow with the existing code/APIs on stable releases. E. g. a hackish, but maybe workable first step might be to only use the python-libvirt API for subscribing to desired notifications and trigger the Cockpit VM module through it, but continue to handle all the actions through virsh. This subscription/notification code will hopefully be small enough that it can be sent "through the wire" (cockpit channel) and executed ad-hoc, without having to actually distribute it on the remote side.
That's a decent idea. Worth trying out.
In many ways virsh is a remoteable command-line API. This would help us get started given the current non-remotable status quo.
The fact that libvirt is all about XML also helps here ... it means we do less screen-scraping.
But I suppose the fact that virsh can execute multiple commands doesn't really buy us very much ... because it goes too far down the road of screen-scraping.
All in all worth trying out. This lets us focus on building the actual interface. And if a remotable API for libvirt shows up later, we could swap over to it without too much effort.
Stef
On Wed, Jun 07, 2017 at 10:47:25AM +0200, Stef Walter wrote:
On 07.06.2017 07:49, Martin Pitt wrote:
Hello Richard,
Richard W.M. Jones [2017-05-31 18:00 +0100]:
I agree with others that as things stand you will need a REST or DBus or similar API added to libvirt.
However have you considered using gobject-introspection to generate new "Payload" types automatically?
This doesn't fundamentally change the picture here. Our JavaScript runs in the browser, so on that side GI doesn't help. What we could do is to dynamically create some code in the Cockpit module, send it over to the remote machine, and have it be executed. This could be in Python, which then assumes that python-libvirt is installed there, or in JS which then assumes that libvirt-glib and gjs are installed. The latter seems like a stronger assumption on a server even, but structurally these are pretty similiar.
I. e. GI is not a magic thing to make an API remotable, it just makes it bindable to different programming languages.
C/GI interfaces also don't map well to D-Bus, i. e. it's not practical to autogenerate a D-Bus interface for a given GI API. This still works for the most simple methods that only accept primitive data types and strings, but as soon as you pass any structs, pointers, function pointers etc. around this can't be exposed though D-Bus any more without further interpretation on the server side.
Perhaps not an runtime. But it seems like such a thing could be done at compile time for the libvirt C API. The libvirt folks do it for their python bindings, using code generation.
The libvirt Python API behaves a lot like the C API. It's very flat and not very "pythonic". Which is actually a nice target for us. We could do something similar with a simple "flat" DBus wrapper of the API ... using code generation.
I see no problem with mapping almost all of the libvirt API into DBus, as the DBus method call + signal concepts are very similar to the libvirt RPC call and event concepts. The place where you would have trouble mapping to DBus is the stream support as that has no equivalent in DBus, and there's no efficient workaround to deal with it that I can imagine.
Regards, Daniel
On 07.06.2017 11:02, Daniel P. Berrange wrote:
On Wed, Jun 07, 2017 at 10:47:25AM +0200, Stef Walter wrote:
On 07.06.2017 07:49, Martin Pitt wrote:
Hello Richard,
Richard W.M. Jones [2017-05-31 18:00 +0100]:
I agree with others that as things stand you will need a REST or DBus or similar API added to libvirt.
However have you considered using gobject-introspection to generate new "Payload" types automatically?
This doesn't fundamentally change the picture here. Our JavaScript runs in the browser, so on that side GI doesn't help. What we could do is to dynamically create some code in the Cockpit module, send it over to the remote machine, and have it be executed. This could be in Python, which then assumes that python-libvirt is installed there, or in JS which then assumes that libvirt-glib and gjs are installed. The latter seems like a stronger assumption on a server even, but structurally these are pretty similiar.
I. e. GI is not a magic thing to make an API remotable, it just makes it bindable to different programming languages.
C/GI interfaces also don't map well to D-Bus, i. e. it's not practical to autogenerate a D-Bus interface for a given GI API. This still works for the most simple methods that only accept primitive data types and strings, but as soon as you pass any structs, pointers, function pointers etc. around this can't be exposed though D-Bus any more without further interpretation on the server side.
Perhaps not an runtime. But it seems like such a thing could be done at compile time for the libvirt C API. The libvirt folks do it for their python bindings, using code generation.
The libvirt Python API behaves a lot like the C API. It's very flat and not very "pythonic". Which is actually a nice target for us. We could do something similar with a simple "flat" DBus wrapper of the API ... using code generation.
I see no problem with mapping almost all of the libvirt API into DBus, as the DBus method call + signal concepts are very similar to the libvirt RPC call and event concepts. The place where you would have trouble mapping to DBus is the stream support as that has no equivalent in DBus, and there's no efficient workaround to deal with it that I can imagine.
Good to know.
FWIW, DBus supports FD passing. And Cockpit supports FD passing over DBus:
https://github.com/cockpit-project/cockpit/pull/5992
But I agree that any such wrapping of libvirt APIs involving streaming in DBus FD passing would probably wouldn't be auto-generated.
Cheers,
Stef
On Thu, May 25, 2017 at 10:26:47AM -0700, Peter wrote:
Hi Everyone,
We are working towards building on the virtualization management functionality in cockpit (http://cockpit-project.org/) and wanted to get some feedback on the best way to integrate with libvirt.
Another thing to consider is what virtualization management functionality should be implemented in cockpit? For example if cockpit would like to handle guest installation in a similar matter as virt-manager it has to be implemented somewhere.
Libvirt doesn't have any API which would handle the whole installation process and IMHO it's out of scope of dbus or REST API. The installation process in virt-manager handles a lot of different aspects, for example it can create a new storage/pool, it can download vmlinuz/initrd from provided URL, it can mount an ISO image, etc.
This logic has to be implemented somewhere and I don't think that cockpit is a right place. Currently there is command line tool virt-install which is part of virt-manager project that is able to handle all the installation aspects but that's not ideal for cockpit the same way ho virsh is not ideal.
Another alternative way how to connect cockpit with libvirt could be to implement some intermediate application which would communicate directly with libvirt, would handle all the logic that is required by most of the management application and provide several ways how UI applications can communicate whit it. The benefit of this application would be that not only cockpit would use it but virt-manager could switch to it eventually and possibly virt-install as well. This would lead to a single project that provides virtualization management functionality that can be reused by other project or tools or simple scripts as well.
Pavel
As a quick overview, cockpit aims to talk to existing remotable system APIs. Usually these API’s take the form of dbus, REST or executable commands. The majority of cockpit is implemented in javascript. There is no cockpit backend that knows how to change a hostname for example. The cockpit backend knows how to handle a dbus payload. The javascript running in the users browser knows how to use the systemd dbus API at org.freedesktop.hostname1 to manage the system hostname.
Right now some of the basics have been implemented by spawning commands on the system. This isn't ideal because it involves parsing / screen scraping output and doesn't support receiving events so we have to poll (ei run the command again) to keep the UI up to date.
As far as I know libvirt doesn't currently have a remoteable API. It does have a daemon that communicates with clients via a XDR RPC. (https://libvirt.org/internals/rpc.html) However from what I'm hearing the RPC is considered an internal implementation and shouldn't be used by external applications. Is that still the case? Is there any chance of getting talking the daemon directly using the XDR standard for a subset of methods blessed as part of the externally supported API?
An alternative is to implement a standards based remotable API, using something like dbus or REST, that can be used by external applications. I imagine that this would be at a bit of a higher level than the current RPC and contain at least some of the logic around the actions it performs rather than being a direct passthrough to the daemon.
Of course that is a pretty big undertaking and would, in my opinion, only be worth it if there is broader interest in the community and use cases beyond what cockpit would like to.
-- libvir-list mailing list libvir-list@redhat.com https://www.redhat.com/mailman/listinfo/libvir-list
cockpit-devel@lists.fedorahosted.org
-
Cole Robinson -
Daniel P. Berrange -
Martin Kletzander -
Martin Pitt -
Pavel Hrdina -
Peter -
Peter Krempa -
Richard W.M. Jones -
Stef Walter