_mtx_lock_spin: obsolete historic handling of kdb_active and panicstr?

[Andriy Gapon]

on 17/10/2012 15:07 mdf at FreeBSD.org said the following:
> On Wed, Oct 17, 2012 at 4:20 AM, Andriy Gapon <avg at freebsd.org> wrote:
>>
>> _mtx_lock_spin has the following check in its retry loop:
>> if (i < 60000000 || kdb_active || panicstr != NULL)
>>         DELAY(1);
>> else
>>         _mtx_lock_spin_failed(m);
>>
> [snip analysis]
>>
>> So I'd like to propose to remove those checks altogether.  Or perhaps to
>> "reverse" them and immediately induce a (possibly secondary) panic if we ever
>> get to that wait loop and kdb_active || panicstr != NULL.
> 
> The panicstr can clearly be removed.  I think there can be race
> conditions with entering kdb and taking a spinlock, because the
> spinlock acquire will block interrupts.  I don't remember if we always
> NMI for kdb enter or if that's configurable.  The old code was clearer
> (or maybe I'm just remembering an Isilon hack); looking at
> stop_cpus_hard() I don't see that it uses an NMI.

kdb always uses stop_cpus_hard and stop_cpus_hard always uses NMI on x86.
>From sys/x86/x86/local_apic.c:
if (vector == IPI_STOP_HARD)
        icrlo |= APIC_DELMODE_NMI | APIC_LEVEL_ASSERT;

>  So a CPU can block
> interrupts, then if it sees kdb_active it will spin until we leave
> kdb, rather than panic.  Of course this would only be relevant if the
> CPU it's trying to acquire is already held; otherwise it should find
> the lock unowned and this isn't relevant.  And if the lock is owned by
> the thread entering kdb, that would be a real panic, not a recoverable
> kdb entry.
> 
> So I think maybe the kdb_active check is also not helpful after all.
> 
> Cheers,
> matthew
> 


-- 
Andriy Gapon