Re: Inefficient TLB flushing

[Jack Steiner <steiner@sgi.com>]

On Wed, Nov 12, 2003 at 01:22:53PM -0800, Andrew Morton wrote:
> David Mosberger <davidm@napali.hpl.hp.com> wrote:
> >
> >   Jack> Here is the patch that I am currently testing:
> > 
> >   Jack> --- /usr/tmp/TmpDir.19957-0/linux/mm/memory.c_1.79	Wed Nov 12 13:56:25 2003
> >   Jack> +++ linux/mm/memory.c	Wed Nov 12 12:57:25 2003
> >   Jack> @@ -574,9 +574,10 @@
> >   Jack>  			if ((long)zap_bytes > 0)
> >   Jack>  				continue;
> >   Jack>  			if (need_resched()) {
> >   Jack> +				int fullmm = (*tlbp)->fullmm;
> >   Jack>  				tlb_finish_mmu(*tlbp, tlb_start, start);
> >   Jack>  				cond_resched_lock(&mm->page_table_lock);
> >   Jack> -				*tlbp = tlb_gather_mmu(mm, 0);
> >   Jack> +				*tlbp = tlb_gather_mmu(mm, fullmm);
> >   Jack>  				tlb_start_valid = 0;
> >   Jack>  			}
> >   Jack>  			zap_bytes = ZAP_BLOCK_SIZE;
> > 
> > I think the patch will work fine, but it's not very clean, because it
> > bypasses the TLB-flush API and directly accesses
> > implementation-specific internals.  Perhaps it would be better to pass
> > a "fullmm" flag to unmap_vmas().  Andrew?
> 
> Either that, or add a new interface function
> 
> 	int mmu_gather_is_full_mm(mmu_gather *tlb);
> 
> and use it...
> 

How implementation independent should it be? Currently, there is only one
field in the mmu_gather structure that must be preserved. However, if we
want to make the interface truly implementation independent, it seems
like we should define something like:

	if (need_resched()) {
		struct mmu_gather_state state;
		tlb_mmu_gather_save_state(*tlbp, &state);
		tlb_finish_mmu(*tlbp, tlb_start, start);
		...
		*tlbp = tlb_mmu_gather_restore_state(&state);
	}

Is this overkill?

Should we use the patch given above for 2.6.0 & replace it with an implementation 
independent interface for 2.6.1?



-- 
Thanks

Jack Steiner (steiner@sgi.com)          651-683-5302
Principal Engineer                      SGI - Silicon Graphics, Inc.