Re: [PATCH 09/14] x86/ibt: Implement IBT+

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Sep 27, 2024 at 09:49:05PM +0200, Peter Zijlstra wrote:

> +#ifdef CONFIG_X86_KERNEL_IBT_PLUS
> +__init_or_module void apply_direct_call_offset(s32 *start, s32 *end)
> +{
> +	s32 *s;
> +
> +	/*
> +	 * incompatible with call depth tracking
> +	 */
> +	if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH))
> +		return;
> +
> +	for (s = start; s < end; s++) {
> +		void *dest, *addr = (void *)s + *s;
> +		struct insn insn;
> +		int ret;
> +
> +		ret = insn_decode_kernel(&insn, addr);
> +		if (WARN_ON_ONCE(ret < 0))
> +			continue;
> +
> +		dest = addr + insn.length + insn.immediate.value;
> +		if (!is_endbr(dest))
> +			continue;
> +
> +		switch (insn.opcode.bytes[0]) {
> +		case CALL_INSN_OPCODE:
> +		case JMP32_INSN_OPCODE:
> +			apply_reloc(4, addr+1, 4);
> +			continue;
> +
> +		case JMP8_INSN_OPCODE:
> +		case 0x70 ... 0x7f: /* Jcc.d8 */
> +			apply_reloc(1, addr+1, 4);
> +			continue;

*sigh*... I have a clang-19 build (thanks 0day) that uses a jmp.d8 +0x7e
as a tail-call, guess how well it goes adding 4 to that :-(

Luckily the next instruction is a giant (alignment) NOP, so I *could* go
fix that up, but perhaps this is pushing things too far ...

> +
> +		case 0x0f:
> +			switch (insn.opcode.bytes[1]) {
> +			case 0x80 ... 0x8f:
> +				apply_reloc(4, addr+2, 4);
> +				continue;
> +
> +			default:
> +				break;
> +			}
> +			break;
> +
> +		default:
> +			break;
> +		}
> +
> +		printk("at: %pS, instruction: %*ph\n", addr, insn.length, addr);
> +		BUG();
> +	}
> +}
> +#else
> +__init_or_module void apply_direct_call_offset(s32 *start, s32 *end) { }
> +#endif