Date:	Fri, 21 Dec 2012 10:21:16 -0800 (PST)
From:	Hugh Dickins <hughd@...gle.com>
To:	Linus Torvalds <torvalds@...ux-foundation.org>
cc:	Mel Gorman <mgorman@...e.de>, David Rientjes <rientjes@...gle.com>,
	Ingo Molnar <mingo@...nel.org>,
	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
	linux-mm <linux-mm@...ck.org>,
	Peter Zijlstra <a.p.zijlstra@...llo.nl>,
	Paul Turner <pjt@...gle.com>,
	Lee Schermerhorn <Lee.Schermerhorn@...com>,
	Christoph Lameter <cl@...ux.com>,
	Rik van Riel <riel@...hat.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Andrea Arcangeli <aarcange@...hat.com>,
	Thomas Gleixner <tglx@...utronix.de>,
	Johannes Weiner <hannes@...xchg.org>,
	Sasha Levin <levinsasha928@...il.com>,
	KOSAKI Motohiro <kosaki.motohiro@...fujitsu.com>
Subject: Re: [patch] mm, mempolicy: Introduce spinlock to read shared policy
 tree

On Fri, 21 Dec 2012, Linus Torvalds wrote:
> On Fri, Dec 21, 2012 at 5:47 AM, Mel Gorman <mgorman@...e.de> wrote:
> > On Thu, Dec 20, 2012 at 02:55:22PM -0800, David Rientjes wrote:
> >>
> >> This is probably worth discussing now to see if we can't revert
> >> b22d127a39dd ("mempolicy: fix a race in shared_policy_replace()"), keep it
> >> only as a spinlock as you suggest, and do what KOSAKI suggested in
> >> http://marc.info/?l=linux-kernel&m=133940650731255 instead.  I don't think
> >> it's worth trying to optimize this path at the cost of having both a
> >> spinlock and mutex.
> >
> > Jeez, I'm still not keen on that approach for the reasons that are explained
> > in the changelog for b22d127a39dd.
> 
> Christ, Mel.
> 
> Your reasons in b22d127a39dd are weak as hell, and then you come up
> with *THIS* shit instead:
> 
> > That leads to this third *ugly* option that conditionally drops the lock
> > and it's up to the caller to figure out what happened. Fooling around with
> > how it conditionally releases the lock results in different sorts of ugly.
> > We now have three ugly sister patches for this. Who wants to be Cinderalla?
> >
> > ---8<---
> > mm: numa: Release the PTL if calling vm_ops->get_policy during NUMA hinting faults
> 
> Heck no. In fact, not a f*cking way in hell. Look yourself in the
> mirror, Mel. This patch is ugly, and *guaranteed* to result in subtle
> locking issues, and then you have the *gall* to quote the "uhh, that's
> a bit ugly due to some trivial duplication" thing in commit
> b22d127a39dd.
> 
> Reverting commit b22d127a39dd and just having a "ok, if we need to
> allocate, then drop the lock, allocate, re-get the lock, and see if we
> still need the new allocation" is *beautiful* code compared to the
> diseased abortion you just posted.
> 
> Seriously. Conditional locking is error-prone, and about a million
> times worse than the trivial fix that Kosaki suggested.

I'm picking up a vibe that you don't entirely like Mel's approach.

I've an unsubstantiated suspicion that it's also incomplete as is.
Although at first I thought huge_memory.c does not need a similar
mod, because THPages are anonymous and cannot come from tmpfs,
I now wonder about a MAP_PRIVATE mapping from tmpfs - for better
or for worse, anon pages there are subject to the same mempolicy
as the shared file pages, and I don't see what prevents khugepaged
from gathering those into THPages.  But it didn't happen when I
tried, so maybe I'm just missing what prevents it.

I don't understand David's and Mel's remarks about the "shared pages"
check making Sasha's warning unlikely: page_mapcount has nothing to do
with whether a page belongs to shm/shmem/tmpfs, and it's easy enough
to reproduce Sasha's warning on the current git tree.  "mount -o
remount,mpol=local /tmp" or something like that is useful in testing.

I wish wish wish I had time to spend on this today, but I don't.
And I've not looked to see (let alone tested) whether it's easy
to revert Mel's mutex then add in Kosaki's patch (which I didn't
look at so have no opinion on).

Shall we go for Peter/David's mutex+spinlock for rc1 - I assume
they both tested that - with a promise to do better in rc2?

What I wanted to try is separate the get_vma_policy() out from
mpol_misplaced(), and have the various callsites do that first
outside the page table lock, passing it in to mpol_misplaced.
But that doesn't work (efficiently) unless it also returns the
range that that policy is valid for, so we don't have to (drop
lock and) call it on every pte.  I cannot do that for rc1, and
perhaps it's irrelevant if Kosaki's patch is preferred.

(Perhaps I should confess I've another reason to come here for
rc2: that "+ info->vfs_inode.i_ino" we recently added for better
interleave distribution in shmem_alloc_page: I think NUMA placement
faults will be fighting shmem_alloc_page's choices because that
offset is not exposed.)

Hugh
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists