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: <CAPcyv4gMQsBSQ-kXM6H_zz96ZTJ5F0XnDfq6_mZTn4t9JwmEpA@mail.gmail.com>
 <4D73CD59-BFD5-401A-A001-41F7BF5641BA@redhat.com> <CAPcyv4hnsUoavnzX8q5VReiLXrOsOYW0Ef5GDKxQtS+6pQLy4A@mail.gmail.com>
 <20200629083411.GA38188@L-31X9LVDL-1304.local>
In-Reply-To: <20200629083411.GA38188@L-31X9LVDL-1304.local>
From:   Dan Williams <dan.j.williams@intel.com>
Date:   Mon, 29 Jun 2020 15:13:25 -0700
Message-ID: <CAPcyv4gK2PL5YNLkeQjSCLrNZ62P1U_HjuBSNpkEvuHMcyeGOQ@mail.gmail.com>
Subject: Re: [PATCH] mm/spase: never partially remove memmap for early section
To:     Wei Yang <richard.weiyang@linux.alibaba.com>
Cc:     David Hildenbrand <david@redhat.com>,
        Michal Hocko <mhocko@kernel.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        Oscar Salvador <osalvador@suse.de>,
        Linux MM <linux-mm@kvack.org>, Baoquan He <bhe@redhat.com>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Mon, Jun 29, 2020 at 1:34 AM Wei Yang
<richard.weiyang@linux.alibaba.com> wrote:
>
> On Thu, Jun 25, 2020 at 12:46:43PM -0700, Dan Williams wrote:
> >On Wed, Jun 24, 2020 at 10:53 PM David Hildenbrand <david@redhat.com> wr=
ote:
> >>
> >>
> >>
> >> > Am 25.06.2020 um 01:47 schrieb Dan Williams <dan.j.williams@intel.co=
m>:
> >> >
> >> > =EF=BB=BFOn Wed, Jun 24, 2020 at 3:44 PM Wei Yang
> >> > <richard.weiyang@linux.alibaba.com> wrote:
> >> > [..]
> >> >>> So, you are right that there is a mismatch here, but I think the
> >> >>> comprehensive fix is to allow early sections to be partially
> >> >>> depopulated/repopulated rather than have section_activate() and
> >> >>> section_deacticate() special case early sections. The special casi=
ng
> >> >>> is problematic in retrospect as section_deactivate() can't be
> >> >>> maintained without understand special rules in section_activate().
> >> >>
> >> >> Hmm... This means we need to adjust pfn_valid() too, which always r=
eturn true
> >> >> for early sections.
> >> >
> >> > Right, rather than carry workarounds in 3 locations, and the bug tha=
t
> >> > has resulted from then getting out of sync, just teach early section
> >> > mapping to allow for the subsection populate/depopulate.
> >> >
> >>
> >> I prefer the easy fix first - IOW what we Here here. Especially, pfn_t=
o_online_page() will need changes as well.
> >
> >Agree, yes, let's do the simple fix first for 5.8 and the special-case
> >elimination work later.
>
> Dan,
>
> A quick test shows this is not a simple task.

Thanks for taking a look...

> First, early sections don't set subsection bitmap, which is necessary for=
 the
> hot-add/remove.
>
> To properly set subsection bitmap, we need to know how many subsections i=
n
> early section. While current code doesn't has a alignment requirement for
> last early section. We mark the whole last early section as present.

I was thinking that the subsection map does not need to be accurate on
initial setup, it only needs to be accurate after the first removal.
However, that would result in new special casing that somewhat defeats
the purpose. The hardest part is potentially breaking up a PMD mapping
of the page array into a series of PTE mappings without disturbing
in-flight pfn_to_page() users.

> I don't find a way to enable this.

While I don't like that this bug crept into the mismatched special
casing of early sections, I'm now coming around to the same opinion.
I.e. that making the memmap for early sections permanent is a simpler
mechanism to maintain.