x86,ftrace: Fix ftrace_regs_caller() unwind

The ftrace_regs_caller() trampoline does something 'funny' when there
is a direct-caller present. In that case it stuffs the 'direct-caller'
address on the return stack and then exits the function. This then
results in 'returning' to the direct-caller with the exact registers
we came in with -- an indirect tail-call without using a register.

This however (rightfully) confuses objtool because the function shares
a few instruction in order to have a single exit path, but the stack
layout is different for them, depending through which path we came
there.

This is currently cludged by forcing the stack state to the non-direct
case, but this generates actively wrong (ORC) unwind information for
the direct case, leading to potential broken unwinds.

Fix this issue by fully separating the exit paths. This results in
having to poke a second RET into the trampoline copy, see
ftrace_regs_caller_ret.

This brings us to a second objtool problem, in order for it to
perceive the 'jmp ftrace_epilogue' as a function exit, it needs to be
recognised as a tail call. In order to make that happen,
ftrace_epilogue needs to be the start of an STT_FUNC, so re-arrange
code to make this so.

Finally, a third issue is that objtool requires functions to exit with
the same stack layout they started with, which is obviously violated
in the direct case, employ the new HINT_RET_OFFSET to tell objtool
this is an expected exception.

Together, this results in generating correct ORC unwind information
for the ftrace_regs_caller() function and it's trampoline copies.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Miroslav Benes <mbenes@suse.cz>
Reviewed-by: Alexandre Chartre <alexandre.chartre@oracle.com>
Acked-by: Josh Poimboeuf <jpoimboe@redhat.com>
Link: https://lkml.kernel.org/r/20200416115118.749606694@infradead.org


Signed-off-by: Ingo Molnar <mingo@kernel.org>