Re: TAKE 950027 - xfs_icsb_lock_all_counters fails with CONFIG_PREEMPT and >=256p

[Andi Kleen <ak@xxxxxxx>]

David Chinner <dgc@xxxxxxx> writes:
> 
> However, when we are excluding the counters, we are already running
> atomic due to holding the in-core superblock spinlock and the
> per-cpu code uses get_cpu()/put_cpu() so neither can get preempted
> in the critical sections the spinlocks used to protect.  So the only
> considerations are having an atomic exclusion mechanism and a wait
> method that does not sleep.

spinlocks are should be fine for this in theory. All CPUs spinning
on a spinlock shouldn't happen normally, but it is supposed to work 
correctly if it happens.

If preemptible spinlocks don't scale to this for large NR_CPUs they're
broken and need to be fixed.

You discovered a important limitation in them and now instead
of fixing them properly you're trying to work around it.

Please try to see the big picture a bit more instead of just XFS.

-Andi