Telling people to write in kernel mode if they care about performance isn't realistic. For most people that would mean completely rewriting their code from scratch, foregoing high-level software stacks and languages, giving up on most databases, giving up on all manner of tools and techniques for high-velocity software development, giving up fault tolerance, dealing directly with fiddly hardware issues (when do I need a TLB shootdown?), etc.
Whereas disabling spectre mitigations is a one-line config change.
For use cases where local system security really doesn't matter (of which there are a lot, let's be honest), a one-line config change for a 25% (or whatever it is now) performance boost is a pretty damned good deal.
Gaming is a context where you care about performance and you aren't using multiple apps at once, but (and I admit this is a bit of a naive guess) I'd be surprised if it's syscall-bound. It seems like performance is likely to be I/O-bound (getting assets from disk into memory), CPU-bound, and GPU-bound, but are you really making large numbers of syscalls? (Maybe this matters in online gaming?)
So that leaves basically some specific server workloads, and at that point I think some of these techniques start to be realistic. Pinning your work onto a core and using kernel-bypass networking is a pretty straightforward technique these days. It's not quite as easy as using the kernel interfaces, but it's pretty close, and it's definitely worth investing some engineering effort into if you care about performance - you can get much more than 25% speedups.
I agree that writing in kernel mode is generally unrealistic (although if you're writing a kernel module for Linux, you still don't need to care about fiddly hardware issues - you've got the rest of Linux still running). Mostly I'd like to see more work like the paper I linked - there should be a standard build of Linux which has hardware privilege separation turned off for use in the cases where you actually can avoid hardware privilege separation (single-user VMs on cloud hosts, single-user data crunching machines, dedicated single-tenant database servers, game consoles without web browsers, ebook readers without web browsers, etc.), or at least a flag to spawn a process and leave it in ring 0. If the use cases are plentiful, this seems like it would be valuable for lots of people - and it'd also make it clear that this generally isn't an option you want on personal computers. (But I think the reason this hasn't been done in the last several decades is that there aren't actually that many use cases that are both genuinely single-user and syscall-bound.)
(More to the point - the solution to software bloat isn't to get rid of security protections so that the bloat has room to grow.)
The vast majority of servers don't run any untrusted code. Servers tend to do lots of syscalls for network I/O.
> Gaming is a context where you care about performance and you aren't using multiple apps at once, but (and I admit this is a bit of a naive guess) I'd be surprised if it's syscall-bound.
I would expect that interfacing with the GPU involves a fair number of syscalls -- but admittedly I'm also guessing.
This is a lot of cases. I'd love to get 25% perf back on postgres, or 25% back on my air-gapped DAW, etc. etc.
Are you still running Visicalc?
I have seen, used, and maintained spreadsheet applications that are irritatingly slow on eight core, i8, with 8 GB of ram, and an SSD.
Granted this was on Windows but nonetheless the idea that end users only need a 486 is ridiculous at best and probably insulting.
But they wrote something completely different. It seems to me like you just totally ignored their comment. What's the point of that?
I mean, sure, that site is nuts, it sorely needs documentation. But not every scenario needs Spectre protection.
(If your environment actually doesn't need to worry about executing malicious code and you want to make syscalls as fast as possible, try a unikernel or implementing your code in a kernel module. Or, depending on what you're doing, try kernel bypass to get to the devices you care about and use something like https://lwn.net/SubscriberLink/816298/4aed890ee2dbffff/ to pin your user code to certain cores and get the kernel completely out of the way. Having to transition out of user mode to access hardware instead of making plain function calls, having to change page tables between processes, etc. are all performance-killing mitigations of their own. http://www.csl.cornell.edu/~delimitrou/papers/2019.asplos.xc... found a 27x performance improvement by getting rid of the privilege boundary between userspace and kernelspace - if you really care about performance and really don't care about malicious code, why would you leave a 27x speedup on the table and worry about a small percentage improvement from these flags??)