Re: [RFC net-next 0/4] ynl: YAML netlink protocol descriptions

[Florian Fainelli <f.fainelli@gmail.com>]

On 8/10/22 19:23, Jakub Kicinski wrote:
> Netlink seems simple and reasonable to those who understand it.
> It appears cumbersome and arcane to those who don't.
> 
> This RFC introduces machine readable netlink protocol descriptions
> in YAML, in an attempt to make creation of truly generic netlink
> libraries a possibility. Truly generic netlink library here means
> a library which does not require changes to support a new family
> or a new operation.
> 
> Each YAML spec lists attributes and operations the family supports.
> The specs are fully standalone, meaning that there is no dependency
> on existing uAPI headers in C. Numeric values of all attribute types,
> operations, enums, and defines and listed in the spec (or unambiguous).
> This property removes the need to manually translate the headers for
> languages which are not compatible with C.
> 
> The expectation is that the spec can be used to either dynamically
> translate between whatever types the high level language likes (see
> the Python example below) or codegen a complete libarary / bindings
> for a netlink family at compilation time (like popular RPC libraries
> do).
> 
> Currently only genetlink is supported, but the "old netlink" should
> be supportable as well (I don't need it myself).
> 
> On the kernel side the YAML spec can be used to generate:
>   - the C uAPI header
>   - documentation of the protocol as a ReST file
>   - policy tables for input attribute validation
>   - operation tables
> 
> We can also codegen parsers and dump helpers, but right now the level
> of "creativity & cleverness" when it comes to netlink parsing is so
> high it's quite hard to generalize it for most families without major
> refactoring.
> 
> Being able to generate the header, documentation and policy tables
> should balance out the extra effort of writing the YAML spec.
> 
> Here is a Python example I promised earlier:
> 
>    ynl = YnlFamily("path/to/ethtool.yaml")
>    channels = ynl.channels_get({'header': {'dev_name': 'eni1np1'}})
> 
> If the call was successful "channels" will hold a standard Python dict,
> e.g.:
> 
>    {'header': {'dev_index': 6, 'dev_name': 'eni1np1'},
>     'combined_max': 1,
>     'combined_count': 1}
> 
> for a netdevsim device with a single combined queue.
> 
> YnlFamily is an implementation of a YAML <> netlink translator (patch 3).
> It takes a path to the YAML spec - hopefully one day we will make
> the YAMLs themselves uAPI and distribute them like we distribute
> C headers. Or get them distributed to a standard search path another
> way. Until then, the YNL library needs a full path to the YAML spec and
> application has to worry about the distribution of those.
> 
> The YnlFamily reads all the info it needs from the spec, resolves
> the genetlink family id, and creates methods based on the spec.
> channels_get is such a dynamically-generated method (i.e. grep for
> channels_get in the python code shows nothing). The method can be called
> passing a standard Python dict as an argument. YNL will look up each key
> in the YAML spec and render the appropriate binary (netlink TLV)
> representation of the value. It then talks thru a netlink socket
> to the kernel, and deserilizes the response, converting the netlink
> TLVs into Python types and constructing a dictionary.
> 
> Again, the YNL code is completely generic and has no knowledge specific
> to ethtool. It's fairly simple an incomplete (in terms of types
> for example), I wrote it this afternoon. I'm also pretty bad at Python,
> but it's the only language I can type which allows the method
> magic, so please don't judge :) I have a rather more complete codegen
> for C, with support for notifications, kernel -> user policy/type
> verification, resolving extack attr offsets into a path
> of attribute names etc, etc. But that stuff needs polishing and
> is less suitable for an RFC.
> 
> The ability for a high level language like Python to talk to the kernel
> so easily, without ctypes, manually packing structs, copy'n'pasting
> values for defines etc. excites me more than C codegen, anyway.

This is really cool BTW, and it makes a lot of sense to me that we are 
moving that way, especially with Rust knocking at the door. I will try 
to do a more thorough review, than "cool, I like it".
-- 
Florian