Re: [RFC] de-asmify the x86-64 system call slowpath

[Linus Torvalds]

On Mon, Jan 27, 2014 at 2:06 PM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>
> It's not just ip and sp -- cs matters here, too, I think.

For signal *delivery*, CS will always be __USER_CS, and %rcx can be
crap, so sysret should be fine. We could easily check that %rip is
valid in the whole slow-path instead of saying "return 1 if we did
do_signal()".

Now, it's a different thing wrt signal handler *return*, because at
that point we really cannot return with some random value in %rcx. We
absolutely do need to use 'iretq' in that whole [rt_]sigreturn() path,
but on x86-64 that is all handled by the system call itself (see the
stub_*_sigreturn stuff in entry_64.S) and it very much uses iret
explicitly (the 32-bit case also does that, by forcing the sigreturn
to be done with an "int 0x80" instruction - we could change that to
use syscall+iret, but frankly, I'm not all that inclined to care,
although it might be worth trying to do just to unify the models a
bit).

                   Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/