On Sun, Mar 26, 2017 at 2:12 PM, Oved Ourfali <oourfali@redhat.com> wrote:
On Sun, Mar 26, 2017 at 2:08 PM, Leon Goldberg <lgoldber@redhat.com> wrote:
Hey Oved,
I don't think completely moving away from iptables is foreseeable at this point, but I could be of course wrong. Either way, upgrading still needs to be thought of.
I see.
By stating that the current infrastructure is complex, I was referring to the entire chain of storing rules in the database, fetching them using a dedicated deployment class consisting of include/exclude logic, sending them over, unpacking, deploying...
This procedure involves a lot of code that could be made redundant if the required logic is present in the host, which to me seems favorable. It of course entails other potential difficulties, primarily in the form of custom services.
I don't think otopi's firewalld plugin is any more complex than the potential code that will have to be written in vdsm-tool, however it currently expects the data generated by aforementioned chain. The hybrid approach briefly touches on simplifying Engine's involvement while retaining use of otopi's plugin.
Okay. I think that writing a new plugin for firewalld is indeed a good option, whether you "refactor" the engine side or not.
otopi already has a 'firewalld' plugin. It's already in use, at least by engine-setup, so we should be a bit careful if we want to change it. Not preventing/objecting anything, just mentioning. This plugin's interface currently only takes XML-content as input. It has no place for configuring existing firewalld services presumably already provided elsewhere (by firewalld itself or 3rd party packages, such as vdsm). So if we go that route we probably want to extend this interface to allow passing service names and rely on them being defined elsewhere. A related issue is that for engine-setup, the _input_ is currently firewalld xml content, and if users choose to configure 'iptables', this is parsed to generate iptables rules. This is currently an engine-setup issue only, but will affect also host deploy flow if we decided to allow passing service names (without their xml content) and still keep compatibility to current state and allow configuring iptables on the host. We'll then be there in the same situation we are at with engine-setup. See also https://bugzilla.redhat.com/1432354 . An alternative is obviously deciding to remove iptables support and support only firewalld, but this is a rather radical change for users, imo. See also this for some of the existing behavior of engine-setup: https://bugzilla.redhat.com/show_bug.cgi?id=1024707#c9 IMO we first need to decide what we want, then how to do that. IMO the questions we have re "what we want" are, more-or-less: 1. Do we want to support in some version X both iptables and firewalld, or is it ok to stop support for iptables and support only firewalld without overlap? If so, do we handle upgrades, and how? 2. Do we want to support custom firewalld xml to be configured on the host by us? Or is it ok to only support choosing among existing services, which will need to be added to the host using other means (packaged by firewalld, packaged by 3rd parties, added manually by users)? 3. Opposite of (2.): Do we want to support firewalld services that are added to the host using other means (see there)? Obviously we do, but: If we do, do we still want to support also iptables (see (1.))? And if so, what do we want to then happen? (2.) and (3.) are not conflicting, each needs its own answer. My (thoughts about) answers: 1. If done well enough, and considerably simplifies things, I think it's ok to jump directly from "only iptables" to "only firewalld", but then we should announce this ahead of time to give users time to plan/prepare. If it's not too costly, I'd prefer to support both for the foreseeable future, though. 2. Latter option here is problematic, if we need/want to allow customizing services during deploy. E.g. suppose that one day we want to make the vdsm port configurable. It will be nice if this can be done by only changing things on the engine, without having to distribute changes (conf and/or packages) to all hosts before host-deploy. I'd say we can go a long way here by having a strict requirement from all services that we need/want on hosts to have official IANA-registered port numbers - then, it's imo much easier to tell users "If you want to change the port for service X, you have to (long list of complex actions goes here)". Currently, where services are not registered, we risk conflicts with existing services, requiring the user to change ports - and so we can't make this process too difficult. No idea how important this is in practice. 3. Not sure :-( I'd say that if we want to support both iptables and firewalld together, and support both "xml in engine" and "xml in host", then it might be ok if the custom rules/services will not automatically apply to both iptables and firewalld. Meaning - you can set both custom iptables rules and firewalld services, but it's up to you to make sure they actually do the same thing if that's important to you. Bottom line: I think we should summarize the open questions in a way that will make it clear to users how each answer will affect them, and ask what they think. Leon already started doing this [1], I only saw one reply. Perhaps this means that users do not care that much, and expect us to just decide and tell them what we decided (and always to keep the option to disable this feature, as is possible today, and do this themselves, if our choice of solution does not fit their needs). I know this is way too loooong, sorry. Feel free to ignore, but then please ask simpler questions if you want shorter answers :-) [1] http://lists.ovirt.org/pipermail/users/2017-March/080600.html
On Sun, Mar 26, 2017 at 1:40 PM, Oved Ourfali <oourfali@redhat.com> wrote:
top-posting: You need to also consider how upgrade will be handled, right? Or iptables will still remain supported?
Also, see some comments inline.
Regards, Oved
On Sun, Mar 26, 2017 at 1:33 PM, Leon Goldberg <lgoldber@redhat.com> wrote:
Hey,
We're looking to migrate from iptables to firewalld. We came up with a couple of possible approaches we'd like opinions on. I'll list the options first, and will
1) Replicate existing flow:
As of date, iptable rules are inserted in the database via SQL config files. During host deployment, VdsDeployIptablesUnit adds the required rules (based on cluster/firewall configuration) to the deployment configuration, en route to being deployed on the host via otopi and its iptables plugin.
Pros:
- Reuse of existing infrastructure.
Cons:
- Current infrastructure is overly complex...
Can you elaborate? I'm not an otopi expert, but I think that otopi plugins shouldn't be more complex than what you describe in section #2, and the plugins were meant in order to handle such cases.
- Many of the required services are provided by firewalld. Rewriting them is wasteful; specifying them (instead of providing actual service .xml content) will require adaptations on both (engine/host) sides. More on that later.
2) Host side based configuration:
Essentially, all the required logic (aforementioned cluster/firewall configuration) to determine if/how firewalld should be deployed could be passed on to the host via ohd. Vdsm could take on the responsibility of examining the relevant configuration, and then creating and/or adding the required services (using vdsm.conf and vdsm-tool).
So here you replace the otopi plugin with relevant vdsm-tool code, and the question is why is that better?
Pros:
- Engine side involvement is greatly diminished. - Simple(r).
Cons:
- Custom services/rules capabilities will have to be rethought and re-implemented (current infrastructure supports custom iptables rules by being specified in the SQL config file).
3) Some other hybrid approach:
If we're able to guarantee all the required firewalld services are statically provided one way or the other, the current procedure could be replicated and be made more simpler. Instead of providing xml content in the form of strings, service names could be supplied. The responsibility of actual service deployment becomes easier, and could be left to otopi (with the appropriate modifications) or switched over to vdsm.
--
Regardless, usage of statically provided vs. dynamically created services remains an open question. I think we'd like to avoid implementing logic that ask whether some service is provided (and then write it if it isn't...), and so choosing between the dynamic and static approaches is also needed. Using the static approach, guaranteeing all services are provided will be required.
I do believe guaranteeing the presence of all required services is worth it, however custom services aren't going to be naively compatible, and we'll still have to use similar mechanism as described in #1 (service string -> .xml -> addition of service name to active zone).
Your thoughts are welcome.
Thanks, Leon
-- Didi