LWN: Comments on "More flexible memory access for BPF programs"

More flexible memory access for BPF programs

polyp — Thu, 24 Nov 2022 10:19:19 +0000

All this work to the BPF mechanism makes me think about the famous Tanenbaum-Torvalds debate. Where do we have an environment where pointer accesses are checked and memory corruption is prevented outside of the sandbox where the program executes? In user-space. Perhaps a micro-kernel with user-space drivers/helpers is the better model.

More flexible memory access for BPF programs

Manifault — Sun, 23 Oct 2022 20:31:50 +0000

Not sure I'm quite following how that's relevant to dynptrs, which are more about ensuring safe accesses to variably sized memory regions, rather than ensuring the lifetime of the memory it points to (though the verifier does guarantee the memory is still valid when it's accessed as well).

For what it's worth, BPF also does support ownership and object lifetime / reference counting. kfuncs can be "acquire" and "release" kfuncs, and maps can store pointers to kernel objects. See [0] and [1] for more information.

[0]: https://lwn.net/Articles/856005/
[1]: https://lwn.net/Articles/900749/

More flexible memory access for BPF programs

jhoblitt — Sun, 23 Oct 2022 03:05:42 +0000

Wow! I missed the memo on that.

More flexible memory access for BPF programs

danobi — Sun, 23 Oct 2022 01:43:26 +0000

I think that already exists — see CONFIG_BPF_PRELOAD.

More flexible memory access for BPF programs

amarao — Sat, 22 Oct 2022 22:03:58 +0000

If only kernel had had the notion of ownership and lifetimes for variables...

More flexible memory access for BPF programs

jhoblitt — Sat, 22 Oct 2022 12:17:53 +0000

It could fit that definition. Although, I think of a unikernel as taking more traditional protected memory proccess and stuffing them into kernel space.

One could image a high performance packet molester based completely on BPF+XDP.

More flexible memory access for BPF programs

jorgegv — Sat, 22 Oct 2022 10:58:47 +0000

You mean Unikernels?

More flexible memory access for BPF programs

jhoblitt — Sat, 22 Oct 2022 00:40:22 +0000

If we could bundle bpf programs into the kernel image, soon we would be able boot systems and not need a userland at all. ;)