close #1808

#1808 (comment)

…egfault

…STAMPING

Problems:

1. In the new kernel, assembly functions uniformly return from
   `__x86_return_thunk`. However, our assembly code uses the original
   `ret` instruction, so objtool in the kernel will notice this is a naked
   return during compilation.

2. `SYM_FUNC_START` in the new kernel will add endbr64 to the head of
   the assembly function, and all indirect jumps to ENDBR instructions,
   that is, the code snippet within the same function, will fail, but we
   use jump tables in the assembly function to perform indirect jumps. It
   will raise CET exception:
   https://en.wikipedia.org/wiki/X86_instruction_listings#Added_with_Intel_CET).

Solutions:

1. Substitute the `ret` with `RET`, a macro in the new kernel to
   ensure the correct return.

2. `notrack jmp` and enable notrack in CPU setting:
   `wrmsrl(MSR_IA32_S_CET, CET_ENDBR_EN | CET_NO_TRACK_EN)`

As an aside, interestingly, if a user-mode C program uses a switch
statement that meets the conditions for generating a jump table (gcc
uses `-fcf-protection=full` by default), the generated jump table will
use a `jmp` with the `notrack` prefix, and IBT will be marked as `true`
in the `.note.gnu.property` section of the compiled elf file, so that
the `NO_TRACK_EN` of the `MSR` will be set to `true` in user mode when
the kernel is loaded. So user mode can use `notrack` to bypass CET
without caring about setting or not setting `NO_TRACK_EN`.

…grate-to-linux-6.8.9

This reverts commit 89d2f30.