linux-kernel - Re: [PATCH 2/7] mm/thp: try_to_unmap() use TTU_SYNC for safe DEBUG

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <alpine.LSU.2.11.2106041454430.14037@eggly.anvils>
Date:   Fri, 4 Jun 2021 15:26:19 -0700 (PDT)
From:   Hugh Dickins <hughd@...gle.com>
To:     Peter Xu <peterx@...hat.com>
cc:     Hugh Dickins <hughd@...gle.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>,
        Yang Shi <shy828301@...il.com>,
        Wang Yugui <wangyugui@...-tech.com>,
        Matthew Wilcox <willy@...radead.org>,
        Naoya Horiguchi <naoya.horiguchi@....com>,
        Alistair Popple <apopple@...dia.com>,
        Ralph Campbell <rcampbell@...dia.com>, Zi Yan <ziy@...dia.com>,
        Miaohe Lin <linmiaohe@...wei.com>,
        Minchan Kim <minchan@...nel.org>, Jue Wang <juew@...gle.com>,
        Jan Kara <jack@...e.cz>, linux-mm@...ck.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH 2/7] mm/thp: try_to_unmap() use TTU_SYNC for safe DEBUG_VM
 splitting

On Fri, 4 Jun 2021, Peter Xu wrote:
> On Thu, Jun 03, 2021 at 07:54:11PM -0700, Hugh Dickins wrote:
> > On Thu, 3 Jun 2021, Peter Xu wrote:
> > > On Tue, Jun 01, 2021 at 02:07:53PM -0700, Hugh Dickins wrote:
> > > > diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
> > > > index 2cf01d933f13..b45d22738b45 100644
> > > > --- a/mm/page_vma_mapped.c
> > > > +++ b/mm/page_vma_mapped.c
> > > > @@ -212,6 +212,14 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
> > > >  			pvmw->ptl = NULL;
> > > >  		}
> > > >  	} else if (!pmd_present(pmde)) {
> > > > +		/*
> > > > +		 * If PVMW_SYNC, take and drop THP pmd lock so that we
> > > > +		 * cannot return prematurely, while zap_huge_pmd() has
> > > > +		 * cleared *pmd but not decremented compound_mapcount().
> > > > +		 */
> > > > +		if ((pvmw->flags & PVMW_SYNC) &&
> > > > +		    PageTransCompound(pvmw->page))
> > > > +			spin_unlock(pmd_lock(mm, pvmw->pmd));
> > > >  		return false;
> > > >  	}
> > > >  	if (!map_pte(pvmw))
> > > 
> > > Sorry if I missed something important, but I'm totally confused on how this
> > > unlock is pairing with another lock()..
> > 
> > I imagine you're reading that as spin_unlock(pmd_lockptr(blah));
> > no, the lock is right there, inside spin_unlock(pmd_lock(blah)).
> 
> Heh, yeah... Sorry about that.

I'll expand that line, as Kirill asks too.

> 
> > 
> > > 
> > > And.. isn't PVMW_SYNC only meaningful for pte-level only (as I didn't see a
> > > reference of it outside map_pte)?
> > 
> > But you are pointing directly to its reference outside map_pte()!
> 
> Right, I was trying to look for the lock() so I needed to look at all the rest
> besides this one. :)
> 
> I didn't follow Yang's patch, but if Yang's patch can make kernel not crashing
> and fault handling done all well, then I'm kind of agree with him: having
> workaround code (like taking lock and quickly releasing..) only for debug code
> seems an overkill to me, not to mention that the debug code will be even more
> strict after this patch, as it means it's even less likely that one can
> reproduce one production host race with DEBUG_VM..  Normally debugging code
> would affect code execution already, and for this one we're enlarging that gap
> "explicitly" - not sure whether it's good.
> 
> This also makes me curious what if we make !DEBUG_VM strict too - how much perf
> we'll lose?  Haven't even tried to think about it with details, but just raise
> it up. Say, is there any chance we make the debug/non-debug paths run the same
> logic (e.g. of SYNC version)?

And Yang Shi suggests the same.

Yes, I'm not fond of doing that differently for DEBUG_VM or not;
but could never quite decide which way to jump.

For so long as we worry about whether split_huge_page() is working
correctly (and Wang Yugui still has a case that we have not solved),
we do want the warning; and for so long as we have the warning, we
do need the TTU_SYNC to prevent showing the warning unnecessarily.

How much overhead added by doing TTU_SYNC now on !DEBUG_VM?  On any
sensible anon THP case, I don't think it could add overhead at all.
But in some shmem cases (multiply mapped but sparsely populated,
populated differently by different tasks) it could add overhead:
dirtying lock cachelines in tasks which don't have the page mapped.

But we're only talking about huge page splitting, that should not
be #1 on anyone's performance list; and has all sorts of other
overheads of its own.  I think I'll go with your preference, and
make this TTU_SYNC for all.  We can easily revert to DEBUG_VM only
if some regression is noticed.

Hugh