x86/ibt: Implement FineIBT-BHI mitigation

While WAIT_FOR_ENDBR is specified to be a full speculation stop; it
has been shown that some implementations are 'leaky' to such an extend
that speculation can escape even the FineIBT preamble.

To deal with this, add additional hardening to the FineIBT preamble.

Notably, using a new LLVM feature:

  https://github.com/llvm/llvm-project/commit/e223485c9b38a5579991b8cebb6a200153eee245

which encodes the number of arguments in the kCFI preamble's register.

Using this register<->arity mapping, have the FineIBT preamble CALL
into a stub clobbering the relevant argument registers in the
speculative case.

Scott sayeth thusly:

Microarchitectural attacks such as Branch History Injection (BHI) and
Intra-mode Branch Target Injection (IMBTI) [1] can cause an indirect
call to mispredict to an adversary-influenced target within the same
hardware domain (e.g., within the kernel). Instructions at the
mispredicted target may execute speculatively and potentially expose
kernel data (e.g., to a user-mode adversary) through a
microarchitectural covert channel such as CPU cache state.

CET-IBT [2] is a coarse-grained control-flow integrity (CFI) ISA
extension that enforces that each indirect call (or indirect jump)
must land on an ENDBR (end branch) instruction, even speculatively*.
FineIBT is a software technique that refines CET-IBT by associating
each function type with a 32-bit hash and enforcing (at the callee)
that the hash of the caller's function pointer type matches the hash
of the callee's function type. However, recent research [3] has
demonstrated that the conditional branch that enforces FineIBT's hash
check can be coerced to mispredict, potentially allowing an adversary
to speculatively bypass the hash check:

__cfi_foo:
  ENDBR64
  SUB R10d, 0x01234567
  JZ foo    # Even if the hash check fails and ZF=0, this branch could still mispredict as taken
  UD2
foo:
  ...

The techniques demonstrated in [3] require the attacker to be able to
control the contents of at least one live register at the mispredicted
target. Therefore, this patch set introduces a sequence of CMOV
instructions at each indirect-callable target that poisons every live
register with data that the attacker cannot control whenever the
FineIBT hash check fails, thus mitigating any potential attack.

The security provided by this scheme has been discussed in detail on
an earlier thread [4].

 [1] https://www.intel.com/content/www/us/en/developer/articles/technical/software-security-guidance/technical-documentation/branch-history-injection.html
 [2] Intel Software Developer's Manual, Volume 1, Chapter 18
 [3] https://www.vusec.net/projects/native-bhi/
 [4] https://lore.kernel.org/lkml/20240927194925.707462984@infradead.org/


 *There are some caveats for certain processors, see [1] for more info

Suggested-by: Scott Constable <scott.d.constable@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kees Cook <kees@kernel.org>
Link: https://lore.kernel.org/r/20250224124200.820402212@infradead.org