On Tue, May 23, 2023 at 04:50:42PM +0000, Dave Thaler wrote: > > -----Original Message----- > > From: David Vernet <void@xxxxxxxxxxxxx> > > Sent: Tuesday, May 23, 2023 9:32 AM > > To: Dave Thaler <dthaler@xxxxxxxxxxxxx> > > Cc: Jose E. Marchesi <jemarch@xxxxxxx>; bpf@xxxxxxxx; bpf > > <bpf@xxxxxxxxxxxxxxx>; Erik Kline <ek.ietf@xxxxxxxxx>; Suresh Krishnan > > (sureshk) <sureshk@xxxxxxxxx>; Christoph Hellwig <hch@xxxxxxxxxxxxx>; > > Alexei Starovoitov <ast@xxxxxxxxxx> > > Subject: Re: [Bpf] IETF BPF working group draft charter > > > > On Thu, May 18, 2023 at 07:42:11PM +0000, Dave Thaler wrote: > > > Jose E. Marchesi <jemarch@xxxxxxx> wrote: > > > > I would think that the way the x86_64, aarch64, risc-v, sparc, mips, > > > > powerpc architectures, along with their variants, handle their ELF > > > > extensions and psABI, ensures interoperability good enough for the > > problem at hand, but ok. > > > > I'm definitely not an expert in these matters. > > > > > > I am not familiar enough with those to make any comment about that. > > > > Hi Dave, > > > > Taking a step back here, perhaps we need to think about all of this more > > generically as "ABI", rather than ELF "extensions", "bindings", etc. In my > > opinion this would include, at a minimum, the following items from the current > > proposed WG charter: > > > > * the eBPF bindings for the ELF executable file format, > > > > * the platform support ABI, including calling convention, linker > > requirements, and relocations, > > > > As far as I know (please correct me if I'm wrong), there isn't really a precedence > > for standardizing ABIs like this. For example, x86 calling conventions are not > > standardized. Solaris, Linux, FreeBSD, macOS, etc all follow the System V > > AMD64 ABI, but Microsoft of course does not. As Jose pointed out, such > > standards extensions do not exist for psABI ELF extensions for various > > architectures either. > > > > While it may be that we do end up needing to standardize these ABIs for BPF, > > I'm beginning to think that we should just remove them from the current WG > > charter, and consider standardizing them at a later time if it's clear that it's > > actually necessary. I think this is especially true given that we don't seem to be > > getting any closer to having consensus, and that we're very short on time given > > that Erik is going to be proposing the charter to the rest of the ADs in just two > > days on 5/25. > > > > Thanks, > > David > > I can tell you it's very important to those who work on the ebpf-for-windows project that the ELF format is common between Linux and Windows so that tools like > llvm-objdump and bpftool and other BPF-specific ELF parsing tools work for both > Linux and Windows. We don't want Windows to diverge. Be that as it may, as I said before, to my knowledge there's no precedence at all for standardizing ABI like this. Is there a reason that you think Windows would diverge if we didn't standardize the ABI? I realize that I'm essentially saying, "Hey, pretend there's a standard and don't diverge", but if that's what the entire rest of the industry has done up until this point with all other psABIs, then it seems like a reasonable expectation. > As such, I feel strongly that it is a requirement to be standardized right away. I have to respectfully disagree. I think there are much bigger fish to fry, such as standardizing the ISA. Unless we really have a good reason for diverging from industry norms, standardizing on ABI now feels to me like we're putting the cart before the horse. Just to be very clear: I could be totally wrong here, and it could be very important to deviate from industry norms and standardize ABI as part of the initial WG charter. However, IMHO, a positive claim like that needs to come with clear substantiation. The reality is that deviating from industry norms and standardizing on ABI will have its own costs and consequences. > Hence I would not want this removed from the charter unless there's an effort > to do it somewhere else right away, which would seem to increase the coordination > burden.
