On Wed, Mar 02, 2022 at 12:26:27PM +0100, Jason A. Donenfeld wrote: > Hey Michael, > > Thanks for the benchmark. > > On Wed, Mar 2, 2022 at 9:30 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote: > > So yes, the overhead is higher by 50% which seems a lot but it's from a > > very small number, so I don't see why it's a show stopper, it's not by a > > factor of 10 such that we should sacrifice safety by default. Maybe a > > kernel flag that removes the read replacing it with an interrupt will > > do. > > > > In other words, premature optimization is the root of all evil. > > Unfortunately I don't think it's as simple as that for several reasons. > > First, I'm pretty confident a beefy Intel machine can mostly hide > non-dependent comparisons in the memory access and have the problem > mostly go away. But this is much less the case on, say, an in-order > MIPS32r2, which isn't just "some crappy ISA I'm using for the sake of > argument," but actually the platform on which a lot of networking and > WireGuard stuff runs, so I do care about it. There, we have 4 > reads/comparisons which can't pipeline nearly as well. Sure. Want to try running some benchmarks on that platform? Presumably you have access to such a box, right? > There's also the atomicity aspect, which I think makes your benchmark > not quite accurate. Those 16 bytes could change between the first and > second word (or between the Nth and N+1th word for N<=3 on 32-bit). > What if in that case the word you read second doesn't change, but the > word you read first did? So then you find yourself having to do a > hi-lo-hi dance. > And then consider the 32-bit case, where that's even > more annoying. This is just one of those things that comes up when you > compare the semantics of a "large unique ID" and "word-sized counter", > as general topics. (My suggestion is that vmgenid provide both.) I don't see how this matters for any applications at all. Feel free to present a case that would be race free with a word but not a 16 byte value, I could not imagine one. It's human to err of course. > > Finally, there's a slightly storage aspect, where adding 16 bytes to a > per-key struct is a little bit heavier than adding 4 bytes and might > bust a cache line without sufficient care, care which always has some > cost in one way or another. > > So I just don't know if it's realistic to impose a 16-byte per-packet > comparison all the time like that. I'm familiar with WireGuard > obviously, but there's also cifs and maybe even wifi and bluetooth, > and who knows what else, to care about too. Then there's the userspace > discussion. I can't imagine a 16-byte hotpath comparison being > accepted as implementable. I think this hinges on benchmarking results. Want to start with my silly benchmark at least? If you can't measure an order of magnitude gain then I think any effect on wireguard will be in the noise. > > And I feel if linux > > DTRT and reads the 16 bytes then hypervisor vendors will be motivated to > > improve and add a 4 byte unique one. As long as linux is interrupt > > driven there's no motivation for change. > > I reeeeeally don't want to get pulled into the politics of this on the > hypervisor side. I assume an improved thing would begin with QEMU and > Firecracker or something collaborating because they're both open > source and Amazon people seem interested. I think it would begin with a benchmark showing there's even any measureable performance to be gained by switching the semantics. > And then pressure builds for > Microsoft and VMware to do it on their side. And then we get this all > nicely implemented in the kernel. In the meantime, though, I'm not > going to refuse to address the problem entirely just because the > virtual hardware is less than perfect; I'd rather make the most with > what we've got while still being somewhat reasonable from an > implementation perspective. > > Jason Right but given you are trading security off for performance, it matters a lot what the performance gain is. -- MST
