Re: [RFC][PATCH v3 1/6] mm: Don't unmap gup()ed page

[Nick Piggin]

On Wednesday 15 April 2009 00:12:09 Andrea Arcangeli wrote:
> On Tue, Apr 14, 2009 at 10:39:54PM +0900, KOSAKI Motohiro wrote:
> > I guess you dislike get_user_page_fast() grab pte_lock too, right?
> 
> If get_user_page_fast is vetoed to run a set_bit on the already cache
> hot and exclusive struct page, I doubt taking a potentially cache
> cold, mm-wide or pmd-wide pte_lock is ok.

Yes, I'd *really* rather not. I actually implemented gup_fast in
response to problem reported with DB2 workload hitting the ptl
(and not the more obvious mmap_sem, although certainly they had
some gain from removing that cacheline as well).

gup_fast iirc is worth nearly 10% on a 4 socket x86 system with
DB2. That's the same order of magnitude as the speedups quoted
to justify the addition of hugepages, or O_DIRECT itself.

Andrea: I didn't veto that set_bit change of yours as such. I just
noted there could be more atomic operations. Actually I would
welcome more comparison between our two approaches, but they seem
to be stuck with Linus refusing (I think) to copy the page at
fork() time :(
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/