Re: [PATCH] x86/entry/64: Fix native_load_gs_index() SWAPGS handling with IRQ state tracing enabled

[Andy Lutomirski]

On Tue, Nov 28, 2017 at 11:09 PM, Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>
> * Jarkko Nikula <jarkko.nikula@xxxxxxxxxxxxxxx> wrote:
>
>> Hi
>>
>> Suspend-to-ram and resume stopped working on v4.15-rc1 and I bisected it to
>> commit ca37e57bbe0c ("x86/entry/64: Add missing irqflags tracing to
>> native_load_gs_index()").
>>
>> I noticed it on Intel Kabylake (core) and Apollolake (atom) based prototype
>> machines. Symptoms are that machine appears to enter into suspend but
>> resumes instantly and hangs. Unfortunately no logs.
>>
>> If I revert ca37e57bbe0c on v4.15-rc1 it works as expected.
>
> Hm, that commit looks broken with irq-tracing enabled.
> Does the patch below fix it?
>
> In fact the exception handler itself appears to have broken GS handling as well -
> I suspect it never triggers in practice, because it was broken forever.
>
> Andy, do you concur?

No.

>
> On a related note, we should definitely extend the 'intended GS state' annotation
> comments I did in this patch to all SWAPGS instances - this way code review has a
> much higher chance of finding discrepancies between intent and actual code.

Agreed.  I'll send a patch.


> --- a/arch/x86/entry/entry_64.S
> +++ b/arch/x86/entry/entry_64.S
> @@ -945,16 +945,16 @@ idtentry simd_coprocessor_error           do_simd_coprocessor_error       has_error_code=0
>          */
>  ENTRY(native_load_gs_index)
>         FRAME_BEGIN
> +       SWAPGS                                  /* switch from user GS to kernel GS */

No, we start with kernel GS.  It was correct before.

>         pushfq
>         DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
>         TRACE_IRQS_OFF
> -       SWAPGS
>  .Lgs_change:
>         movl    %edi, %gs
>  2:     ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
> -       SWAPGS
>         TRACE_IRQS_FLAGS (%rsp)
>         popfq
> +       SWAPGS                                  /* switch from kernel GS to user GS */
>         FRAME_END
>         ret
>  ENDPROC(native_load_gs_index)
> @@ -964,7 +964,7 @@ EXPORT_SYMBOL(native_load_gs_index)
>         .section .fixup, "ax"
>         /* running with kernelgs */
>  bad_gs:
> -       SWAPGS                                  /* switch back to user gs */
> +       SWAPGS                                  /* switch back to user GS, to modify GS */
>  .macro ZAP_GS
>         /* This can't be a string because the preprocessor needs to see it. */
>         movl $__USER_DS, %eax
> @@ -973,6 +973,7 @@ EXPORT_SYMBOL(native_load_gs_index)
>         ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
>         xorl    %eax, %eax
>         movl    %eax, %gs
> +       SWAPGS                                  /* switch to kernel GS again before continuing */

Which we don't want to do because the landing site expects user GS.

I suspect we're hitting an entirely different bug, that we're blowing
up if we WARN too early in resume.