Re: [PATCH] x86/retpoline/entry: Disable the entire SYSCALL64 fast path with retpolines on

[Dan Williams]

On Thu, Jan 25, 2018 at 1:31 PM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
> On Thu, Jan 25, 2018 at 1:20 PM, Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>> On Thu, Jan 25, 2018 at 1:08 PM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>>>
>>> With retpoline, the retpoline in the trampoline sucks.  I don't need
>>> perf for that -- I've benchmarked it both ways.  It sucks.  I'll fix
>>> it, but it'll be kind of complicated.
>>
>> Ahh, I'd forgotten about that (and obviously didn't see it in the profiles).
>>
>> But yeah, that is fixable even if it does require a page per CPU. Or
>> did you have some clever scheme in mind?
>
> Nothing clever.  I was going to see if I could get actual
> binutils-generated relocations to work in the trampoline.  We already
> have code to parse ELF relocations and turn them into a simple table,
> and it shouldn't be *that* hard to run a separate pass on the entry
> trampoline.
>
> Another potentially useful if rather minor optimization would be to
> rejigger the SYSCALL_DEFINE macros a bit.  Currently we treat all
> syscalls like this:
>
> long func(long arg0, long arg1, long arg2, long arg3, long arg4, long arg5);
>
> I wonder if we'd be better off doing:
>
> long func(const struct pt_regs *regs);
>
> and autogenerating:
>
> static long SyS_read(const struct pt_regs *regs)
> {
>    return sys_reg(regs->di, ...);
> }

If you're rejiggering, can we also put in a mechanism for detecting
which registers to clear so that userspace can't inject useful values
into speculation paths?

https://patchwork.kernel.org/patch/10153753/