Re: [nvo3] Alignment and Ethernet encapsulation
On 9/20/2016 11:38 AM, Tom Herbert wrote: > On Tue, Sep 20, 2016 at 10:55 AM, Joe Touchwrote: >> >> On 9/20/2016 10:47 AM, Tom Herbert wrote: >>> On Tue, Sep 20, 2016 at 10:35 AM, Joe Touch wrote: ... If you've been handed Ethernet to encapsulate, then that is the header whose alignment you should be optimizing. As you point out, you can't align both IP and Ethernet to 4-byte boundaries at the same time. >>> You can if you insert a two byte pad before the encapsulated Ethernet >>> header like ETHERIP (RFC3378 does). >> You CAN do a lot of things, but you shouldn't. >> >> If you're handed ethernet to encapsulate, then THAT is the header whose >> access should be aligned. >> >> You have no business ASSUMING that you need to optimize for IP header >> access. All you're doing is messing up the ethernet access. >> > Sorry, but I disagree with that statement. Keeping IP headers aligned > for stack processing is a HUGE concern in real implementation. It is, and something that handles Ethernet messages should already know what to do here. So if it already expects to shift the packet over as it copies it in, then you've just defeated that. The same is true for non-IP inside Ethernet, which might have its own padding to re-align its contents. I.e., if it's ethernet, treat it as ethernet. That's what an L2 tunnel should do. Joe ___ nvo3 mailing list nvo3@ietf.org https://www.ietf.org/mailman/listinfo/nvo3
Re: [nvo3] Alignment and Ethernet encapsulation
On Tue, Sep 20, 2016 at 10:55 AM, Joe Touchwrote: > > > On 9/20/2016 10:47 AM, Tom Herbert wrote: >> On Tue, Sep 20, 2016 at 10:35 AM, Joe Touch wrote: >>> ... >>> If you've been handed Ethernet to encapsulate, then that is the header >>> whose alignment you should be optimizing. >>> >>> As you point out, you can't align both IP and Ethernet to 4-byte >>> boundaries at the same time. >> You can if you insert a two byte pad before the encapsulated Ethernet >> header like ETHERIP (RFC3378 does). > > You CAN do a lot of things, but you shouldn't. > > If you're handed ethernet to encapsulate, then THAT is the header whose > access should be aligned. > > You have no business ASSUMING that you need to optimize for IP header > access. All you're doing is messing up the ethernet access. > Sorry, but I disagree with that statement. Keeping IP headers aligned for stack processing is a HUGE concern in real implementation. Probably every device driver on the planet will put in a two byte pad in receiver buffers to ensure that the Ethernet payload (and hence IP header, EHs, TCP headers, etc.). We need two byte alignment of Ethernet headers, and four byte alignment on IP protocols. If we don't have this a lot of things will break (as VXLAN is currently broken on SPARC)! Maybe in theory alignment is no longer considered an issue and is just a historical footnote, but on real devices in deployment today unaligned protocol headers still can be problematic. Tom > Frankly, anything that already parses Ethernet then IP ought to be able > to efficiently handle the difference in the two alignments - and you > might be messing THAT up. > > RFC3378 defines a header used for ETHERIP; that header may or may not be > relevant to other ethernet-in-X encapsulations. The field needed to be > there for versioning and was expanded from 8 to 16 bits, which certainly > helps align the *ethernet* header that comes after it compared to the > earlier use of an 8-bit field. There's no mention of aligning for the > purposes of accessing the IP header, which again, I disagree with. > > Joe > > ___ nvo3 mailing list nvo3@ietf.org https://www.ietf.org/mailman/listinfo/nvo3
Re: [nvo3] Alignment and Ethernet encapsulation
On 9/20/2016 10:47 AM, Tom Herbert wrote: > On Tue, Sep 20, 2016 at 10:35 AM, Joe Touchwrote: >> ... >> If you've been handed Ethernet to encapsulate, then that is the header >> whose alignment you should be optimizing. >> >> As you point out, you can't align both IP and Ethernet to 4-byte >> boundaries at the same time. > You can if you insert a two byte pad before the encapsulated Ethernet > header like ETHERIP (RFC3378 does). You CAN do a lot of things, but you shouldn't. If you're handed ethernet to encapsulate, then THAT is the header whose access should be aligned. You have no business ASSUMING that you need to optimize for IP header access. All you're doing is messing up the ethernet access. Frankly, anything that already parses Ethernet then IP ought to be able to efficiently handle the difference in the two alignments - and you might be messing THAT up. RFC3378 defines a header used for ETHERIP; that header may or may not be relevant to other ethernet-in-X encapsulations. The field needed to be there for versioning and was expanded from 8 to 16 bits, which certainly helps align the *ethernet* header that comes after it compared to the earlier use of an 8-bit field. There's no mention of aligning for the purposes of accessing the IP header, which again, I disagree with. Joe ___ nvo3 mailing list nvo3@ietf.org https://www.ietf.org/mailman/listinfo/nvo3
Re: [nvo3] Alignment and Ethernet encapsulation
Hi, Tom, On 9/20/2016 9:13 AM, Tom Herbert wrote: > ... > For new encapsulation protocols please consider the effects of IP > header alignment in the presence of Ethernet encapsulation. Defining > Ethernet encapsulation with the two byte padding like in ETHERIP may > help a lot to make implementation of Ethernet encapsulation feasible > on CPU HW. IMO, alignment needs to be handled within each encapsulation layer independently. I don't think it's useful to expect new encapsulation layers to have to make sure every layer of an encapsulated packet is aligned - just the first one ought to be sufficient. The rest is the responsibility of whomever added the other layers already in place. So yes, it's useful to make sure the encapsulated packet starts on a boundary that is 4-byte aligned, but the rest *needs to be* someone else's problem. Joe ___ nvo3 mailing list nvo3@ietf.org https://www.ietf.org/mailman/listinfo/nvo3
[nvo3] Alignment and Ethernet encapsulation
Hello, This issue came up on the Linux netdev list. The Oracle engineers are trying to deal with the ramifications of unaligned IP headers (not aligned to four bytes) when doing Ethernet encapsulation (VXLAN, GRE with TEB for instance). The problem is that certain CPU architectures, such as SPARC (and maybe PowerPC?), do not handle unaligned accesses very well. Every unaligned access causes a trap to software which kills any semblance of performance. x86 has shift logic so we don't see any issues there (except for unaligned accesses across a page boundary). The Linux stack is written to assume that IP protocols are properly four byte aligned and in non-encapsulation cases Ethernet headers are padded to be 16 bytes. The problem seems to be specific when performing Ethernet header encapsulation and there is no padding of the inner Ethernet header like would be done in ETHERIP. Either the outer or the inner IP headers will be misaligned. There does not seem to be an easy fix for this without rewriting the stack or doing copies on every packet. For new encapsulation protocols please consider the effects of IP header alignment in the presence of Ethernet encapsulation. Defining Ethernet encapsulation with the two byte padding like in ETHERIP may help a lot to make implementation of Ethernet encapsulation feasible on CPU HW. Thanks, Tom ___ nvo3 mailing list nvo3@ietf.org https://www.ietf.org/mailman/listinfo/nvo3