Re: [PATCH] mm: Check if PTE is already allocated during page fault

[Andrea Arcangeli <aarcange@redhat.com>]

On Thu, Apr 21, 2011 at 05:00:57PM +0100, Mel Gorman wrote:
> If you want to create a new patch with either your comment or mine
> (whichever you prefer) I'll add my ack. I'm about to drop offline
> for a few days but if it's still there Tuesday, I'll put together an
> appropriate patch and submit. I'd keep it separate from the other patch
> because it's a performance fix (which I'd like to see in -stable) where
> as this is more of a cleanup IMO.

I think the older patch should have more priority agreed. This one may
actually waste cpu cycles overall, rather than saving them, it
shouldn't be a common occurrence.

>From a code consistency point of view maybe we should just implement a
pte_alloc macro (to put after pte_alloc_map) and use it in both
places, and hide the glory details of the unlikely in the macro. When
implementing pte_alloc, I suggest also adding unlikely to both, I mean
we added unlikely to the fast path ok, but __pte_alloc is orders of
magnitude less likely to fail than pte_none, and it still runs 1 every
512 4k page faults, so I think __pte_alloc deserves an unlikely too.

Minchan, you suggested this cleanup, so I suggest you to send a patch,
but if you're busy we can help.

Thanks!
Andrea