-
Notifications
You must be signed in to change notification settings - Fork 12
Advanced topics
In order to have less code to maintain, it should be possible to write a tool that auto-generates CLI commands based on the FRR YANG models. As a matter of fact, there are already a number of NETCONF-based CLIs that do exactly that (e.g. Clixon, ConfD's CLI).
The problem however is that there isn't an exact one-to-one mapping between the existing CLI commands and the corresponding YANG nodes from the native models. As an example, ripd's timers basic (5-2147483647) (5-2147483647) (5-2147483647) command changes three YANG leaves at the same time. In order to auto-generate CLI commands and retain their original form, it's necessary to add annotations in the YANG modules to specify how the commands should look like. Without YANG annotations, the CLI auto-generator will generate a command for each YANG leaf, (leaf-)list and presence-container. The ripd's timers basic command, for instance, would become three different commands, which would be undesirable.
This Tail-f's® document shows how to customize ConfD auto-generated CLI commands using YANG annotations.
The good news is that libyang allows users to create plugins to implement their own YANG extensions, which can be used to implement CLI annotations. If done properly, a CLI generator can save FRR developers from writing and maintaining hundreds if not thousands of DEFPYs!
The flexible design of the northbound architecture opens the door to move the CLI to a separate program in the long-term future. Some advantages of doing so would be:
- Treat the CLI as just another northbound client, instead of having CLI commands embedded in the binaries of all FRR daemons.
- Improved robustness: bugs in CLI commands (e.g. null-pointer dereferences) or in the CLI code itself wouldn't affect the FRR daemons.
- Foster innovation by allowing other CLI programs to be implemented, possibly using higher level programming languages.
The problem, however, is that the northbound retrofitting process will convert only the CLI configuration commands and EXEC commands in a first moment. Retrofitting the "show" commands is a completely different story and shouldn't happen anytime soon. This should hinder progress towards moving the CLI to a separate program.
Confirmed commits allow the user to request an automatic rollback to the previous configuration if the commit operation is not confirmed within a number of minutes. This is particularly useful when the user is accessing the CLI through the network (e.g. using SSH) and any configuration change might cause an unexpected loss of connectivity between the user and the router (e.g. misconfiguration of a routing protocol). By using a confirmed commit, the user can rest assured the connectivity will be restored after the given timeout expires, avoiding the need to access the router physically to fix the problem.
Example of how this feature could be provided in the CLI: commit confirmed [minutes <1-60>]. The ability to do confirmed commits should also be exposed in the northbound API so that the northbound plugins can also take advantage of it (in the case of the Sysrepo and ConfD plugins, confirmed commits are implemented externally in the netopeer2-server and confd daemons, respectively).
Since the lyd_node data structure from libyang can hold private data, it should be possible to mark configuration commands or sections as active or inactive. This would allow CLI users to leverage this feature to disable parts of the running configuration without actually removing the associated commands, and then re-enable the disabled configuration commands or sections later when necessary. Example:
ripd(config)# show configuration running
Configuration:
[snip]
!
router rip
default-metric 2
distance 80
network eth0
network eth1
!
end
ripd(config)# disable router rip
ripd(config)# commit
% Configuration committed successfully (Transaction ID #7).
ripd(config)# show configuration running
Configuration:
[snip]
!
!router rip
!default-metric 2
!distance 80
!network eth0
!network eth1
!
end
ripd(config)# enable router rip
ripd(config)# commit
% Configuration committed successfully (Transaction ID #8).
ripd(config)# show configuration running
[snip]
frr defaults traditional
!
router rip
default-metric 2
distance 80
network eth0
network eth1
!
end
This capability could be useful in a number of occasions, like disabling configuration commands that are no longer necessary (e.g. ACLs) but that might be necessary at a later point in the future. Other example is allowing users to disable a configuration section for testing purposes, and then re-enable it easily without needing to copy and paste any command.
Given the limitations of the previous northbound architecture, the FRR daemons didn't have the ability to reload their configuration files by themselves. The SIGHUP handler of most daemons would only re-read the configuration file and merge it into the running configuration. In most cases, however, what is desired is to replace the running configuration by the updated configuration file. The frr-reload.py script was written to work around this problem and it does it well to a certain extent. The problem with the frr-reload.py script is that it's full of special cases here and there, which makes it fragile and unreliable. Maintaining the script is also an additional burden for FRR developers, few of whom are familiar with its code or know when it needs to be updated to account for a new feature.
In the new northbound architecture, reloading the configuration file can be easily implemented using a configuration transaction. Once the FRR northbound retrofitting process is complete, all daemons should have the ability to reload their configuration files upon receiving the SIGHUP signal, or when the configuration load [...] replace command is used. Once that point is reached, the frr-reload.py script will no longer be necessary and should be removed from the FRR repository.
This post from the Tail-f's® forum describes the problem of letting systems configure themselves behind the users back. Here are some selected snippets from it:
Traditionally, northbound interface users are the ones in charge of providing configuration data for systems.
In some systems, we see a deviation from this traditional practice; allowing systems to configure "themselves" behind the scenes (or behind the users back).
While there might be a business case for such a practice, this kind of configuration remains "dangerous" from northbound users perspective and makes systems hard to predict and even harder to debug. (...)
With the advent of transactional Network configuration, this practice can not work anymore. The fact that systems are given the right to change configuration is a key here in breaking transactional configuration in a Network.
FRR is immune to some of the problems described in the aforementioned post. Management clients can configure interfaces that don't yet exist, and once an interface is deleted from the kernel, its configuration is retained in FRR.
There are however some cases where information learned from the kernel (e.g. using netlink) can affect the running configuration of all FRR daemons. Examples: interface rename events, VRF rename events, interface being moved to a different VRF, etc. In these cases, since these events can't be ignored, the best we can do is to send YANG notifications to the management clients to inform about the configuration changes. The management clients should then be prepared to handle such notifications and react accordingly.
As of now zebra doesn't have the ability to create VRFs or virtual interfaces in the kernel. The vrf and interface commands only create pre-provisioned VRFs and interfaces that are only activated when the corresponding information is learned from the kernel. When configuring FRR using an external management client, like a NETCONF client, it might be desirable to actually create functional VRFs and virtual interfaces (e.g. VLAN subinterfaces, bridges, etc) that are installed in the kernel using OS-specific APIs (e.g. netlink, routing socket, etc). Work needs to be done in this area to make this possible.
One of the existing problems in FRR is that it's hard to ensure that all daemons are in sync with respect to the shared configuration objects (e.g. interfaces, VRFs, route-maps, ACLs, etc). When a route-map is configured using vtysh, the same command is sent to all relevant daemons (the daemons that implement route-maps), which ensures synchronization among them. The problem is when a daemon starts after the route-maps are created. In this case this daemon wouldn't be aware of the previously configured route-maps (unlike the other daemons), which can lead to a lot of confusion and unexpected problems.
With the new northbound architecture, configuration objects can be manipulated using higher level abstractions, which opens more possibilities to solve this decades-long problem. As an example, one solution would be to make the FRR daemons fetch the shared configuration objects from zebra using the ZAPI interface during initialization. The shared configuration objects could be requested using a list of XPaths expressions in the ZEBRA_HELLO message, which zebra would respond by sending the shared configuration objects encoded in the JSON format. This solution however doesn't address the case where zebra starts or restarts after the other FRR daemons. Other solution would be to store the shared configuration objects in the northbound SQL database and make all daemons fetch these objects from there. So far no work has been made on this area as more investigation needs to be done.
As explained in the Transactional CLI page, all commands introduced by the transactional CLI are not yet available in vtysh. This needs to be addressed in the short term future. Some challenges for doing that work include:
- How to display configurations (running, candidates and rollbacks) in a more clever way? The implementation of the
show running-configcommand in vtysh is not something that should be followed as an example. A better idea would be to fetch the desired configuration from all daemons (encoded in JSON for example), merge them all into a singlelyd_nodevariable and then display the combined configurations from this variable (the configuration merges would transparently take care of combining the shared configuration objects). In order to be able to manipulate the JSON configurations, vtysh will need to load the YANG modules from all daemons at startup (this might have a minimal impact on startup time). The only issue with this approach is that thecli_show()callbacks from all daemons are embedded in their binaries and thus not accessible externally. It might be necessary to compile these callbacks on a separate shared library so that they are accessible to vtysh too. Other than that, displaying the combined configurations in the JSON/XML formats should be straightforward. - With the current design, transaction IDs are per-daemon and not global across all FRR daemons. This means that the same transaction ID can represent different transactions on different daemons. Given this observation, how to implement the
rollback configurationcommand in vtysh? The easy solution would be to add adaemon WORDargument to specify the context of the rollback, but per-daemon rollbacks would certainly be confusing and convoluted to end users. A better idea would be to attack the root of the problem: change configuration transactions to be global instead of being per-daemon. This involves a bigger change in the northbound architecture, and would have implications on how transactions are stored in the SQL database (daemon-specific and shared configuration objects would need to have their own tables or columns). - Loading configuration files in the JSON or XML formats will be tricky, as vtysh will need to know which sections of the configuration should be sent to which daemons. vtysh will either need to fetch the YANG modules implemented by all daemons at runtime or obtain this information at compile-time somehow.
As described in the Retrofitting Configuration Commands page, the northbound configuration callbacks detect type mismatches at runtime when fetching data from the the dnode parameter (which represents the configuration node being created, modified, deleted or moved). When a type mismatch is detected, the program aborts and displays a backtrace showing where the problem happened. It would be desirable to detect such type mismatches at compile-time, the earlier the problems are detected the sooner they are fixed.
One possible solution to this problem would be to auto-generate C structures from the YANG models and provide a function that converts a libyang's lyd_node variable to a C structure containing the same information. The northbound callbacks could then fetch configuration data from this C structure, which would naturally lead to type mismatches being detected at compile time. One of the challenges of doing this would be the handling of YANG lists and leaf-lists. It would be necessary to use dynamic data structures like hashes or rb-trees to hold all elements of the lists and leaf-lists, and the process of converting a lyd_node to an auto-generated C-structure could be expensive. At this point it's unclear if it's worth adding more complexity in the northbound architecture to solve this specific problem.