Re: Perf: WARNING: arch/x86/entry/common.c:624 idtentry_exit_cond_rcu+0x92/0xc0

[Andy Lutomirski]

> On Jun 12, 2020, at 2:01 AM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> 
> ïAndy Lutomirski <luto@xxxxxxxxxx> writes:
>> On Thu, Jun 11, 2020 at 4:22 PM Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>> 
>> Two bugs here.
>> 
>> 1. We had an issue with WARN. Patch sent.
> 
> Grabbed it
> 
>> 2. idtentry.h has, for x86_32:
>> 
>> # define DEFINE_IDTENTRY_IST  DEFINE_IDTENTRY
>> 
>> This is nonsense.  It's getting late over here and I'd rather focus on
>> the more interesting RCU issue, so that's all from me today.
> 
> Well, this might be nonsense, but it's exactly matching the current code
> in mainline which, e.g. for #DB does:
> 
> SYM_CODE_START(debug)
>    /*
>     * Entry from sysenter is now handled in common_exception
>     */
>    ASM_CLAC
>    pushl    $0
>    pushl    $do_debug
>    jmp    common_exception
> SYM_CODE_END(debug)
> 
> There is no IST on 32bit, never was. We do software stack switching for
> device interrupts, but that's a different story.
> 

DEFINE_IDTENTRY does the idtentry_enter_cond_rcu() dance, which isnât intended to be safe from NMI context.  It should probably map to DEFINE_IDTENTRY_RAW() instead.  The specific issue is that NMI ends up there, and at least DEFINE_IDTENTRY_NMI should be raw.

I havenât tried this at all, nor have I dug through all the users of these macros to check what they expect.  Perhaps we should not have the _IST one defined at all on 32 bit and rename it to DEFINE_IDTENTRY_IST_RAW on 64 bit to make it more clear whatâs going on when reading the C code.

Or maybe Iâm too sleepy and Iâm nuts. But I donât think I am.


> Thanks,
> 
>        tglx