Re: [RFC PATCH 2/4] x86/entry/64: move ENTER_IRQ_STACK from interrupt macro to helper function

[Andy Lutomirski]

On Wed, Feb 14, 2018 at 6:21 PM, Dominik Brodowski
<linux@xxxxxxxxxxxxxxxxxxxx> wrote:
> Moving the switch to IRQ stack from the interrupt macro to the helper
> function requires some trickery: All ENTER_IRQ_STACK really cares about
> is where the "original" stack -- meaning the GP registers etc. -- is
> stored. Therefore, we need to offset the stored RSP value by 8 whenever
> ENTER_IRQ_STACK is called from within a function. In such cases, and
> after switching to the IRQ stack, we need to push the "original" return
> address (i.e. the return address from the call to the interrupt entry
> function) to the IRQ stack.
>
> This trickery allows us to carve another 1k from the text size:
>
>    text    data     bss     dec     hex filename
>   17905       0       0   17905    45f1 entry_64.o-orig
>   16897       0       0   16897    4201 entry_64.o
>
> Signed-off-by: Dominik Brodowski <linux@xxxxxxxxxxxxxxxxxxxx>
> ---
>  arch/x86/entry/entry_64.S | 53 +++++++++++++++++++++++++++++++----------------
>  1 file changed, 35 insertions(+), 18 deletions(-)
>
> diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S
> index de8a0da0d347..3046b12a1acb 100644
> --- a/arch/x86/entry/entry_64.S
> +++ b/arch/x86/entry/entry_64.S
> @@ -449,10 +449,18 @@ END(irq_entries_start)
>   *
>   * The invariant is that, if irq_count != -1, then the IRQ stack is in use.
>   */
> -.macro ENTER_IRQ_STACK regs=1 old_rsp
> +.macro ENTER_IRQ_STACK regs=1 old_rsp save_ret=0
>         DEBUG_ENTRY_ASSERT_IRQS_OFF
>         movq    %rsp, \old_rsp
>
> +       .if \save_ret
> +       /*
> +        * If save_ret is set, the original stack contains one additional
> +        * entry -- the return address.
> +        */
> +       addq    $8, \old_rsp
> +       .endif
> +

This is a bit alarming in that you now have live data below RSP.  For
x86_32, this would be a big no-no due to NMI.  For x86_64, it might
still be bad if there are code paths where NMI is switched to non-IST
temporarily, which was the case at some point and might still be the
case.  (I think it is.)  Remember that the x86_64 *kernel* ABI has no
red zone.

It also means that, if you manage to hit vmalloc_fault() in here when
you touch the IRQ stack, you're dead.  IOW you hit:

        movq    \old_rsp, PER_CPU_VAR(irq_stack_union + IRQ_STACK_SIZE - 8)

which gets #PF and eats your return pointer.  Debugging this will be
quite nasty because you'll only hit it on really huge systems after a
thread gets migrated, and even then only if you get unlucky on your
stack alignment.

So can you find another way to do this?

--Andy