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;
MIME-Version: 1.0
References: <20200925222600.6832-1-peterx@redhat.com> <20200925222600.6832-4-peterx@redhat.com>
 <20200926232335.GA348793@ziepe.ca>
In-Reply-To: <20200926232335.GA348793@ziepe.ca>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Sat, 26 Sep 2020 17:04:38 -0700
Message-ID: <CAHk-=wh5pE+gXTigysiGf4vf+6jG7K58gwdk8b7qk_QUeEgKWA@mail.gmail.com>
Subject: Re: [PATCH v2 3/4] mm: Do early cow for pinned pages during fork()
 for ptes
To:     Jason Gunthorpe <jgg@ziepe.ca>
Cc:     Peter Xu <peterx@redhat.com>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Linux-MM <linux-mm@kvack.org>,
        John Hubbard <jhubbard@nvidia.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Christoph Hellwig <hch@lst.de>, Yang Shi <shy828301@gmail.com>,
        Oleg Nesterov <oleg@redhat.com>,
        Kirill Tkhai <ktkhai@virtuozzo.com>,
        Kirill Shutemov <kirill@shutemov.name>,
        Hugh Dickins <hughd@google.com>, Jann Horn <jannh@google.com>,
        Michal Hocko <mhocko@suse.com>, Jan Kara <jack@suse.cz>,
        Andrea Arcangeli <aarcange@redhat.com>,
        Leon Romanovsky <leonro@nvidia.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Sat, Sep 26, 2020 at 4:23 PM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> Linus's version doesn't do pte_sw_mkyoung(), but looks OK to have it

I don't think it matters. But I don't think it should make it young,
since there's no access, but it's not like it's a big deal.

> > +                             pte = maybe_mkwrite(pte_mkdirty(pte), new);
>
> maybe_mkwrite() was not in Linus's version but it is wp_page_copy().

Actually, it is in my version too, just in a different form.

I did it using

        if (vma->vm_flags & VM_WRITE)
                *src_pte = pte_mkwrite(*src_pte);

instead, ie avoiding the write to src_pte if it wasn't a writable vma
(and I had checked that it was dirty and not done this at all if not,
so no need for the mkdirty).

> It seemed like mk_pte() should set the proper write
> bit already from the vm_page_prot?

No, vm_page_prot won't have the writable bit for a COW mapping.

The write bit gets set when the write happens (which may be on the
first access, of course), by the code that makes sure it's a private
copy.

> Perhaps this is harmless but redundant?

No, the pte_mkwrite() is required in some form, whether it's that
"maybe_mkwrite()" that looks at the vm flags, or in that explicit
form.

> > +                             page_add_new_anon_rmap(new_page, new, addr, false);
> > +                             rss[mm_counter(new_page)]++;
> > +                             set_pte_at(dst_mm, addr, dst_pte, pte);
>
> Linus's patch had a lru_cache_add_inactive_or_unevictable() here, like
> wp_page_copy()

Yeah, I do think that is needed so that we have the new page on the
LRU and it gets properly evicted under memory pressure.

                   Linus