Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Wed, 23 Sep 2020 11:44:18 -0400
From:   Peter Xu <peterx@redhat.com>
To:     Jan Kara <jack@suse.cz>
Cc:     John Hubbard <jhubbard@nvidia.com>, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org,
        Linus Torvalds <torvalds@linux-foundation.org>,
        Michal Hocko <mhocko@suse.com>,
        Kirill Shutemov <kirill@shutemov.name>,
        Jann Horn <jannh@google.com>, Oleg Nesterov <oleg@redhat.com>,
        Kirill Tkhai <ktkhai@virtuozzo.com>,
        Hugh Dickins <hughd@google.com>,
        Leon Romanovsky <leonro@nvidia.com>,
        Christoph Hellwig <hch@lst.de>,
        Andrew Morton <akpm@linux-foundation.org>,
        Jason Gunthorpe <jgg@ziepe.ca>,
        Andrea Arcangeli <aarcange@redhat.com>
Subject: Re: [PATCH 5/5] mm/thp: Split huge pmds/puds if they're pinned when
 fork()
Message-ID: <20200923154418.GE59978@xz-x1>
References: <20200921211744.24758-1-peterx@redhat.com>
 <20200921212031.25233-1-peterx@redhat.com>
 <5e594e71-537f-3e9f-85b6-034b7f5fedbe@nvidia.com>
 <20200922103315.GD15112@quack2.suse.cz>
 <4a65586e-9282-beb0-1880-1ef8da03727c@nvidia.com>
 <20200923092205.GA6719@quack2.suse.cz>
 <20200923135004.GB59978@xz-x1>
 <20200923140114.GA15875@quack2.suse.cz>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <20200923140114.GA15875@quack2.suse.cz>
Precedence: bulk

On Wed, Sep 23, 2020 at 04:01:14PM +0200, Jan Kara wrote:
> On Wed 23-09-20 09:50:04, Peter Xu wrote:
> > On Wed, Sep 23, 2020 at 11:22:05AM +0200, Jan Kara wrote:
> > > On Tue 22-09-20 13:01:13, John Hubbard wrote:
> > > > On 9/22/20 3:33 AM, Jan Kara wrote:
> > > > > On Mon 21-09-20 23:41:16, John Hubbard wrote:
> > > > > > On 9/21/20 2:20 PM, Peter Xu wrote:
> > > > > > ...
> > > > > > > +	if (unlikely(READ_ONCE(src_mm->has_pinned) &&
> > > > > > > +		     page_maybe_dma_pinned(src_page))) {
> > > > > > 
> > > > > > This condition would make a good static inline function. It's used in 3
> > > > > > places, and the condition is quite special and worth documenting, and
> > > > > > having a separate function helps with that, because the function name
> > > > > > adds to the story. I'd suggest approximately:
> > > > > > 
> > > > > >      page_likely_dma_pinned()
> > > > > > 
> > > > > > for the name.
> > > > > 
> > > > > Well, but we should also capture that this really only works for anonymous
> > > > > pages. For file pages mm->has_pinned does not work because the page may be
> > > > > still pinned by completely unrelated process as Jann already properly
> > > > > pointed out earlier in the thread. So maybe anon_page_likely_pinned()?
> > > > > Possibly also assert PageAnon(page) in it if we want to be paranoid...
> > > > > 
> > > > > 								Honza
> > > > 
> > > > The file-backed case doesn't really change anything, though:
> > > > page_maybe_dma_pinned() is already a "fuzzy yes" in the same sense: you
> > > > can get a false positive. Just like here, with an mm->has_pinned that
> > > > could be a false positive for a process.
> > > > 
> > > > And for that reason, I'm also not sure an "assert PageAnon(page)" is
> > > > desirable. That assertion would prevent file-backed callers from being
> > > > able to call a function that provides a fuzzy answer, but I don't see
> > > > why you'd want or need to do that. The goal here is to make the fuzzy
> > > > answer a little bit more definite, but it's not "broken" just because
> > > > the result is still fuzzy, right?
> > > > 
> > > > Apologies if I'm missing a huge point here... :)
> > > 
> > > But the problem is that if you apply mm->has_pinned check on file pages,
> > > you can get false negatives now. And that's not acceptable...
> > 
> > Do you mean the case where proc A pinned page P from a file, then proc B
> > mapped the same page P on the file, then fork() on proc B?
> 
> Yes.
> 
> > If proc B didn't explicitly pinned page P in B's address space too,
> > shouldn't we return "false" for page_likely_dma_pinned(P)?  Because if
> > proc B didn't pin the page in its own address space, I'd think it's ok to
> > get the page replaced at any time as long as the content keeps the same.
> > Or couldn't we?
> 
> So it depends on the reason why you call page_likely_dma_pinned(). For your
> COW purposes the check is correct but e.g. for "can filesystem safely
> writeback this page" the page_likely_dma_pinned() would be wrong. So I'm
> not objecting to the mechanism as such. I'm mainly objecting to the generic
> function name which suggests something else than what it really checks and
> thus it could be used in wrong places in the future... That's why I'd
> prefer to restrict the function to PageAnon pages where there's no risk of
> confusion what the check actually does.

How about I introduce the helper as John suggested, but rename it to

  page_maybe_dma_pinned_by_mm()

?

Then we also don't need to judge on which is more likely to happen (between
"maybe" and "likely", since that will confuse me if I only read these words..).

I didn't use any extra suffix like "cow" because I think it might be useful for
things besides cow.  Fundamentally the new helper will be mm-based, so "by_mm"
seems to suite better to me.

Does that sound ok?

-- 
Peter Xu