On 03/11/2014 01:33 PM, Sasha Levin wrote:
On 03/11/2014 01:00 PM, Rik van Riel wrote:
On 03/11/2014 12:51 PM, Sasha Levin wrote:
On 03/11/2014 12:28 PM, Mel Gorman wrote:
On Fri, Mar 07, 2014 at 06:27:45PM +0000, Mel Gorman wrote:
This is a completely untested prototype. It rechecks pmd_trans_huge
under the lock and falls through if it hit a parallel split. It's not
perfect because it could decide to fall through just because there
was
no prot_numa work to do but it's for illustration purposes. Secondly,
I noted that you are calling invalidate for every pmd range. Is that
not
a lot of invalidations? We could do the same by just tracking the
address
of the first invalidation.
And there were other minor issues. This is still untested but Sasha,
can you try it out please? I discussed this with Rik on IRC for a bit
and
reckon this should be sufficient if the correct race has been
identified.
Any luck with this patch Sasha? It passed basic tests here but I had
not
seen the issue trigger either.
Sorry, I've been stuck in my weekend project of getting lockdep to work
with page locks :)
It takes a moment to test, so just to be sure - I should have only this
last patch applied?
Without the one in the original mail?
Indeed, only this patch should do it.
Okay. So just this patch on top of the latest -next shows the following
issues:
OK, those are all issues with Davidlohr Bueso's
per-thread vma cache patch :)
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>