Intel finally published a whitepaper about Spectre #2 Mitigation. The PDF is also featured on Hacker News. It’s a technical whitepaper, but you can see the footprints of lawyers all over the language.
CET introduces a shadow stack for return addresses only, and will fail your code into an exception if the normal stack return address and the shadow stack address disagree. Trying to touch and manipulate the shadow stack will also fail into an exception. That is, CET makes touching a return address on the stack toxic by having in effect separate argument and return address stacks, and your code explodes every time you try to do something funny with return addresses.
Which is what Retpolines depend on.
Intel also note that Retpolines also require Return Stack Stuffing in order to work properly, because otherwise RET will also speculate and the Retpoline will not work as intended.
They also note that external events can interact with the Return Stack Buffer (RSB) and empty it and they don’t actually write but strongly imply that these external events are not a thing you have control over, as these are
»There are also a number of events that happen asynchronously from normal program execution that can result in an empty RSB. Software may use “RSB stuffing” sequences whenever these asynchronous events occur:
- Interrupts/NMIs/traps/aborts/exceptions which increase call depth.
- System Management Interrupts (SMI) (see BIOS/Firmware Interactions).
- Host VMEXIT/VMRESUME/VMENTER.
- Microcode update load (WRMSR 0x79) on another logical processor of the same core.«
The hardware you run on can either have a Spectre vulnerability and require Retpolines, or have CET and Hardware Spectre Mitigation, but not both. So you generate code for either one or the other, and they suggest instrumentation of the loader to patch all indirect jump callsites as required. That’s not completely insane, as loaders already patch up jump instructions at load time.
They do note that virtual machines in clustered setups with life migration may be at a disadvantage here, as this is a load-time thing and hence you need to expose your worst CPU in a cluster as the virtual CPU to your VMs in order for this to work with migration. That’s already a requirement in heterogenous clusters, but still annoying.
All in all, this is basically disappointing to underwhelming. The TL;DR is that Retpolines and CET are indeed incompatible, and ld.so is called to the rescue to patch up code at load time for one or the other.
The ld.so thing is an interesting observation. It means that the code on disk and the code in memory differ in one more way. As these things accumulate, the actual on-disk machine instructions mutate over time into a kind of virtual machine notation that on load is adjusted more and more to the needs of the actual machine it’s being executed on. Maybe a thing such as a JVM and a JIT is not the worst thing that has happened to us.