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:   Thu, 3 Sep 2020 15:27:22 +0300
From:   Mike Rapoport <rppt@linux.ibm.com>
To:     Catalin Marinas <catalin.marinas@arm.com>
Cc:     Wei Li <liwei213@huawei.com>, will@kernel.org,
        saberlily.xia@hisilicon.com, puck.chen@hisilicon.com,
        butao@hisilicon.com, fengbaopeng2@hisilicon.com,
        nsaenzjulienne@suse.de, steve.capper@arm.com,
        song.bao.hua@hisilicon.com, linux-arm-kernel@lists.infradead.org,
        linux-kernel@vger.kernel.org, sujunfei2@hisilicon.com
Subject: Re: [PATCH v2] arm64: mm: free unused memmap for sparse memory model
 that define VMEMMAP
Message-ID: <20200903122722.GI424181@linux.ibm.com>
References: <20200812010655.96339-1-liwei213@huawei.com>
 <20200817080405.GL969206@linux.ibm.com>
 <20200903120558.GB31409@gaia>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20200903120558.GB31409@gaia>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Sep 03, 2020 at 01:05:58PM +0100, Catalin Marinas wrote:
> On Mon, Aug 17, 2020 at 11:04:05AM +0300, Mike Rapoport wrote:
> > On Wed, Aug 12, 2020 at 09:06:55AM +0800, Wei Li wrote:
> > > For the memory hole, sparse memory model that define SPARSEMEM_VMEMMAP
> > > do not free the reserved memory for the page map, this patch do it.
> > 
> > I've been thinking about it a bit more and it seems that instead of
> > freeing unused memory map it would be better to allocate the exact
> > memory map from the beginning.
> > 
> > In sparse_init_nid() we can replace PAGES_PER_SECTION parameter to
> > __populate_section_memmap() with the calculated value for architectures
> > that define HAVE_ARCH_PFN_VALID.
> 
> Or just use a smaller PAGES_PER_SECTION and reduce the waste ;).
> 
> Just to be clear, are you suggesting that we should use pfn_valid() on
> the pages within a section to calculate the actual range? The
> pfn_valid() implementation on arm64 checks for the validity of a sparse
> section, so this would be called from within the sparse_init() code
> path. I hope there's no dependency but I haven't checked. If it works,
> it's fine by me, it solves the FLATMEM mem_map freeing as well.

What I meant was that sparse_init()->__populate_section_memmap() would
not blindly presume that the entire section is valid, but would take
into account The actual DRAM banks listed in memblock.memory.
For that to work we'll need a version of pfn_valid() that does not rely
on the validity of sparse section, but uses some other means, e.g.
memblock. Apparently, I've looked at arm32 version of pfn_valid() and
missed the section validity check :)

I was thinking about doing something like this for 32-bit systems
(non-ARM) that cannot affort small sections because of the limited space
in the page->flags.

> With 4KB pages on arm64, vmemmap_populate() stops at the pmd level, so
> it always allocates PMD_SIZE. Wei's patch also only frees in PMD_SIZE
> amounts. So, with a sizeof(struct page) of 64 (2^6), a PMD_SIZE mem_map
> section would cover 2^(21-6) pages, so that's equivalent to a
> SECTION_SIZE_BITS of 21-6+12 = 27.
> 
> If we reduce SECTION_SIZE_BITS to 27 or less, this patch is a no-op.
>
> -- 
> Catalin

-- 
Sincerely yours,
Mike.