Re: [PATCH] Minimal mmu notifiers for kvm

[Robin Holt <holt@sgi.com>]

This patch would require GRU to maintain its own page tables and hold
reference counts on the pages.  That seems like a complete waste of
memory compared to Andrea's most recent patch.  The invalidate_range_start
and invalidate_range_end pair is needed to eliminate the page reference
counts.  The _start callout sets an internal structure in a state that
prevents GRU from satisfying faults, then executes the GRU instruction
to flush the TLB entry.  The _end callout releases the block on faults.

On Fri, Apr 25, 2008 at 08:13:00AM +1000, Rusty Russell wrote:
> +static DEFINE_SPINLOCK(notifier_lock);
> +
> +/*
> + * Must not hold mmap_sem nor any other VM related lock when calling
> + * this registration function.
> + */
> +int mm_add_notifier_ops(struct mm_struct *mm,
> +			const struct mmu_notifier_ops *mops)
> +{
> +	int err;
> +
> +	spin_lock(&notifier_lock);

This one global lock will get extremely hot when a 4096 MPI rank job
is starting up and every one of them goes to use the GRU at once.  I am
not sure where x86_64 peaks out, but on ia64 going beyond approx 32 cpus
contending for the same lock made starvation a very important issue.

> +	if (mm->mmu_notifier_ops)
> +		err = -EBUSY;

So we can only use one of KVM or GRU or Quadrix or IB or (later) XPMEM
per mm?

> +	else {
> +		mm->mmu_notifier_ops = mops;
> +		err = 0;
> +	}
> +	spin_unlock(&notifier_lock);
> +	return err;
> +}
> +EXPORT_SYMBOL_GPL(mm_add_notifier_ops);

Robin