Received-SPF: pass (google.com: domain of linux-ext4-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Message-ID: <6a2513ac-1ea1-ba54-9d57-487504c6c86d@suse.cz>
Date:   Fri, 27 Aug 2021 13:26:48 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.0.2
Subject: Re: [PATCH v1 02/14] mm: remove extra ZONE_DEVICE struct page
 refcount
Content-Language: en-US
To:     Ralph Campbell <rcampbell@nvidia.com>,
        Alex Sierra <alex.sierra@amd.com>, akpm@linux-foundation.org,
        Felix.Kuehling@amd.com, linux-mm@kvack.org,
        linux-ext4@vger.kernel.org, linux-xfs@vger.kernel.org
Cc:     amd-gfx@lists.freedesktop.org, dri-devel@lists.freedesktop.org,
        hch@lst.de, jgg@nvidia.com, jglisse@redhat.com
References: <20210825034828.12927-1-alex.sierra@amd.com>
 <20210825034828.12927-3-alex.sierra@amd.com>
 <dbd0f54a-ad6c-97a2-17d7-826247424c97@suse.cz>
 <505e4fe4-241a-8205-d2ef-f603e9cc46c6@nvidia.com>
From:   Vlastimil Babka <vbabka@suse.cz>
In-Reply-To: <505e4fe4-241a-8205-d2ef-f603e9cc46c6@nvidia.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 8/25/21 19:49, Ralph Campbell wrote:
> 
> On 8/25/21 4:15 AM, Vlastimil Babka wrote:
>> On 8/25/21 05:48, Alex Sierra wrote:
>>> From: Ralph Campbell <rcampbell@nvidia.com>
>>>
>>> ZONE_DEVICE struct pages have an extra reference count that complicates the
>>> code for put_page() and several places in the kernel that need to check the
>>> reference count to see that a page is not being used (gup, compaction,
>>> migration, etc.). Clean up the code so the reference count doesn't need to
>>> be treated specially for ZONE_DEVICE.
>> That's certainly welcome. I just wonder what was the reason to use 1 in the
>> first place and why it's no longer necessary?
> 
> I'm sure this is a long story that I don't know most of the history.
> I'm guessing that ZONE_DEVICE started out with a reference count of
> one since that is what most "normal" struct page pages start with.
> Then put_page() is used to free newly initialized struct pages to the
> slab/slob/slub page allocator.
> This made it easy to call get_page()/put_page() on ZONE_DEVICE pages
> since get_page() asserts that the caller has a reference.
> However, most drivers that create ZONE_DEVICE struct pages just insert
> a PTE into the user page tables and don't increment/decrement the
> reference count. MEMORY_DEVICE_FS_DAX used the >1 to 1 reference count
> transition to signal that a page was idle so that made put_page() a
> bit more complex. Then MEMORY_DEVICE_PRIVATE pages were added and this
> special casing of what "idle" meant got more complicated and more parts
> of mm had to check for is_device_private_page().
> My goal was to make ZONE_DEVICE struct pages reference counts be zero
> based and allocated/freed similar to the page allocator so that more
> of the mm code could be used, like THP ZONE_DEVICE pages, without special
> casing ZONE_DEVICE.

Thanks for the explanation. I was afraid there was something more fundamental
that required to catch the 2->1 refcount transition, seems like it's not. I
agree with the simplification!