On 21/02/11 12:18PM, Toke Høiland-Jørgensen wrote: > "Brian G. Merrell" <brian.g.merrell@xxxxxxxxx> writes: > > > One thing I have been a little concerned about is the XDP_RUN_CONFIG in > > the xdp program function. For our case--with multiple teams writing > > independent, composable xdp programs--we don't want the XDP_RUN_CONFIG > > policy to be in the xdp program. Instead, we want the Go orchestration > > tool to have that policy as part of its configuration data (e.g., what > > order to run the xdp program functions in). From what I can tell, it's > > possible to omit the XDP_RUN_CONFIG from the xdp program function, and > > instead set the values when loading the xdp dispatcher. That's great, and > > thanks for the foresight there. I just want to confirm that I'm > > understanding that correctly, because it's very important for us. > > Yes. The values embedded into the program BTF are defaults, and can be > overridden on load. The idea is that an application will set a default > value (e.g., "I'm a firewall, so I want to run early" or "I want to > monitor traffic to the stack so I'll run late"), but if the sysadmin > wants to do things differently they can override the order. The > important bit being that ultimate control of run order is up to the > *user*, not the application developer. Great. In our case, it would be ideal if the application developer doesn't even need to be aware of the XDP_RUN_CONFIG and can just omit it. I guess for our implementation that would mean that we don't error out if the BTF section doesn't exist, and instead we look to our configuration data (more on that below) for the relevant information. > The policy override stuff is not implemented yet, but I am planning to > implement it by having libxdp read a config file with priority overrides > (similar to how libc will read /etc/nsswitch.conf or /etc/hosts which > makes them work in all applications). > > And of course, if you're writing an orchestration tool, then you *are* > the user, so having the tool override priorities is definitely in scope > (it'll just be an alternative way to set policy instead of a config > file). How are you planning to specify the effective run order? I am > also quite open to working on a compatible way that can work for both > your tool and libxdp :) As part of our control plane we have a whole process for a sysadmin to get config data to to our BPF orchestration tool, which is running on multiple nodes. It very abstractly looks like this: +---- Node 1 | UI -> API -> DATABASE -> CONFIG DATA +---- Node 2 | +---- Node N So, the sysadmin using the UI or API would dictate which xdp programs run *and* what their priority is (plus anything else that would otherwise go into XDP_RUN_CONFIG, plus a bunch of other config data for various other needs). Then--and hopefully I'm getting this right--when our (Go) orchestration tool uses (Go) libxdp, the tool needs a way to set the run order for the XDP programs before the dispatcher loads. I was planning to set the run order programatically on the XDP program objects via libxdp calls. It looks like your libxdp implementation already has ways to do this in the form of xdp_program__set_run_prio() and xdp_program__chain_call_enabled(). Does that make sense? This is still all very theoretical for me at this point! > > > Also, I do hope that the existing Go BTF libraries are good enough to do > > what's needed here, because if I'm understand correctly, that's how I'll > > need to approach setting the XDP_RUN_CONFIG values for our use case. > > You'll need to *parse* BTF to *read* the XDP_RUN_CONFIG. Which is pretty > basic, really, you just need to walk the BTF reference tree. Feel free > to reuse the parsing code in libxdp; that is, in turn, adapted from the > .maps section parsing code in libbpf :) OK, that makes sense. Since I want to keep our implementation purely in Go (if possible), what I trying to say what that I hope there's an existing Go library that can parse and read BTF (Cillium's Go eBPF library looks promising). After thinking more about our orchestration config data use case I was describing above, though, I don't think reading XDP_RUN_CONFIG from BTF is strictly necessary for our use case. That said, it obviously would be preferable to conform to the specification, plus it does look necessary to read the program IDs from BTF anyway :) 8< snip > Please do keep me in the loop, and don't hesitate to ping me if there > are things that are unclear or that you feel are less-than-ideal in the > way things work. I'm also quite open to evolving the spec to meet > everyone's needs! Will do. Thanks for your help! -Brian
![]() |