diff -u: Intel Design Flaw Fallout
For weeks, the world's been talking about severe Intel design flaws affecting many CPUs and forcing operating systems to look for sometimes costly workarounds.
Linux patches for these issues are in a state of ongoing development. Security is always the first priority, at the expense of any other feature. Next would probably be the general speed of a running system for the average user. After that, the developers might begin piecing together any features that had been pulled as part of the initial security fix.
But while this effort goes on, the kernel developers seem fairly angry at Intel, especially when they feel that Intel is not doing enough to fix the problems in future processors.
In response to one set of patches, for example, Linus Torvalds burst out with, "All of this is pure garbage. Is Intel really planning on making this shit architectural? Has anybody talked to them and told them they are f*cking insane?" He went on, "the IBRS garbage implies that Intel is _not_ planning on doing the right thing for the indirect branch speculation. Honestly, that's completely unacceptable." And then he said:
The whole IBRS_ALL feature to me very clearly says "Intel is not serious about this, we'll have an ugly hack that will be so expensive that we don't want to enable it by default, because that would look bad in benchmarks". So instead they try to push the garbage down to us. And they are doing it entirely wrong.
He went on, even more disturbingly, to say:
The patches do things like add the garbage MSR writes to the kernel entry/exit points. That's insane. That says "we're trying to protect the kernel". We already have retpoline there, with less overhead.
So somebody isn't telling the truth here. Somebody is pushing complete garbage for unclear reasons. Sorry for having to point that out....As it is, the patches are COMPLETE AND UTTER GARBAGE....WHAT THE F*CK IS GOING ON?
At one point, David Woodhouse offered a helpful technical summary of the whole situation for those of us on the edge of our seats:
This is all about Spectre variant 2, where the CPU can be tricked into mispredicting the target of an indirect branch. And I'm specifically looking at what we can do on *current* hardware, where we're limited to the hacks they can manage to add in the microcode.
The new microcode from Intel and AMD adds three new features.
One new feature (IBPB) is a complete barrier for branch prediction. After frobbing this, no branch targets learned earlier are going to be used. It's kind of expensive (order of magnitude ~4000 cycles).
The second (STIBP) protects a hyperthread sibling from following branch predictions which were learned on another sibling. You *might* want this when running unrelated processes in userspace, for example. Or different VM guests running on HT siblings.
The third feature (IBRS) is more complicated. It's designed to be set when you enter a more privileged execution mode (i.e. the kernel). It prevents branch targets learned in a less-privileged execution mode, BEFORE IT WAS MOST RECENTLY SET, from taking effect. But it's not just a "set-and-forget" feature, it also has barrier-like semantics and needs to be set on *each* entry into the kernel (from userspace or a VM guest). It's *also* expensive. And a vile hack, but for a while it was the only option we had.
Even with IBRS, the CPU cannot tell the difference between different userspace processes, and between different VM guests. So in addition to IBRS to protect the kernel, we need the full IBPB barrier on context switch and vmexit. And maybe STIBP while they're running.
Then along came Paul with the cunning plan of "oh, indirect branches can be exploited? Screw it, let's not have any of *those* then", which is retpoline. And it's a *lot* faster than frobbing IBRS on every entry into the kernel. It's a massive performance win.
So now we *mostly* don't need IBRS. We build with retpoline, use IBPB on context switches/vmexit (which is in the first part of this patch series before IBRS is added), and we're safe. We even refactored the patch series to put retpoline first.
But wait, why did I say "mostly"? Well, not everyone has a retpoline compiler yet...but OK, screw them; they need to update.
Then there's Skylake, and that generation of CPU cores. For complicated reasons they actually end up being vulnerable not just on indirect branches, but also on a "ret" in some circumstances (such as 16+ CALLs in a deep chain).
The IBRS solution, ugly though it is, did address that. Retpoline doesn't. There are patches being floated to detect and prevent deep stacks, and deal with some of the other special cases that bite on SKL, but those are icky too. And in fact IBRS performance isn't anywhere near as bad on this generation of CPUs as it is on earlier CPUs *anyway*, which makes it not quite so insane to *contemplate* using it as Intel proposed.
That's why my initial idea, as implemented in this RFC patchset, was to stick with IBRS on Skylake, and use retpoline everywhere else. I'll give you 'garbage patches', but they weren't being 'just mindlessly sent around'. If we're going to drop IBRS support and accept the caveats, then let's do it as a conscious decision having seen what it would look like, not just drop it quietly because poor Davey is too scared that Linus might shout at him again.
I have seen *hand-wavy* analyses of the Skylake thing that mean I'm not actually lying awake at night fretting about it, but nothing concrete that really says it's OK.
If you view retpoline as a performance optimisation, which is how it first arrived, then it's rather unconventional to say "well, it only opens a *little* bit of a security hole but it does go nice and fast so let's do it".
But fine, I'm content with ditching the use of IBRS to protect the kernel, and I'm not even surprised. There's a *reason* we put it last in the series, as both the most contentious and most dispensable part. I'd be *happier* with a coherent analysis showing Skylake is still OK, but hey-ho, screw Skylake.
The early part of the series adds the new feature bits and detects when it can turn KPTI off on non-Meltdown-vulnerable Intel CPUs, and also supports the IBPB barrier that we need to make retpoline complete. That much I think we definitely *do* want. There have been a bunch of us working on this behind the scenes; one of us will probably post that bit in the next day or so.
I think we also want to expose IBRS to VM guests, even if we don't use it ourselves. Because Windows guests (and RHEL guests; yay!) do use it.
If we can be done with the shouty part, I'd actually quite like to have a sensible discussion about when, if ever, we do IBPB on context switch (ptraceability and dumpable have both been suggested) and when, if ever, we set STIPB in userspace.
Most of the discussion on the mailing list focused on the technical issues surrounding finding actual solutions. But Linus was not alone in finding the situation unacceptable. A variety of developers, including David, were horribly offended, not by the design flaw itself, but by the way they perceived Intel to be handling the subsequent situation—the poor technical fixes, the lack of communication between Intel developers and the kernel community, and as Linus pointed out, the potential choice by Intel not to fix some of the problems at all.
Note: If you're mentioned above and want to post a response above the comment section, send a message with your response text to firstname.lastname@example.org.