Re: [PATCH 5/7] x86/bugs: Only harden syscalls when needed

[Andrew Cooper]

On 11/04/2024 4:38 pm, Josh Poimboeuf wrote:
> On Thu, Apr 11, 2024 at 11:06:37AM +0100, Andrew Cooper wrote:
>>> +#define __do_syscall(table, func_direct, nr, regs)			\
>>> +({									\
>>> +	unsigned long __rax, __rdi, __rsi;				\
>>> +									\
>>> +	asm_inline volatile(						\
>>> +		ALTERNATIVE("call " __stringify(func_direct) "\n\t",	\
>>> +			    ANNOTATE_RETPOLINE_SAFE			\
>>> +			    "call *%[func_ptr]\n\t",			\
>> This wants to be a plain maybe-thunk'd indirect call, and without the
>> ANNOTATE_RETPOLINE_SAFE.
>>
>> Or you're going to get into cases where some combinations of command
>> line options do unexpected things e.g. retpolining everything except the
>> syscall dispatch.
> In that case won't X86_FEATURE_INDIRECT_SAFE get cleared, resulting in
> the above using a direct call?  Or did I miss something?

That works until the next time anyone touches the logic.  Then it's
latent vulnerability, or an incorrect trial of a non-default option.

I guarantee you'll save someone (probably someone on this CC list) a
headache in the future by not introducing an unnecessary special case here.

~Andrew