Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Subject: Re: [PATCH] mm/memcg: use get_page() for device private pages in
 mc_handle_swap_pte()
From:   Miaohe Lin <linmiaohe@huawei.com>
To:     Matthew Wilcox <willy@infradead.org>
CC:     <akpm@linux-foundation.org>, <hannes@cmpxchg.org>,
        <mhocko@kernel.org>, <roman.gushchin@linux.dev>,
        <shakeelb@google.com>, <muchun.song@linux.dev>,
        <linux-mm@kvack.org>, <cgroups@vger.kernel.org>,
        <linux-kernel@vger.kernel.org>
References: <20230715032802.2508163-1-linmiaohe@huawei.com>
 <ZLIY+ZwrLvpapGE6@casper.infradead.org>
 <73b5d7c2-783d-3d75-2c1b-4b91a039df94@huawei.com>
Message-ID: <3a86cef4-d1fd-8914-3b0a-2ae95ccff740@huawei.com>
Date:   Sun, 23 Jul 2023 09:16:28 +0800
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:78.0) Gecko/20100101
 Thunderbird/78.6.0
MIME-Version: 1.0
In-Reply-To: <73b5d7c2-783d-3d75-2c1b-4b91a039df94@huawei.com>
Content-Type: text/plain; charset="utf-8"
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 2023/7/17 10:28, Miaohe Lin wrote:
> On 2023/7/15 11:56, Matthew Wilcox wrote:
>> On Sat, Jul 15, 2023 at 11:28:02AM +0800, Miaohe Lin wrote:
>>> When page table locked is held, the page can't be freed from under us.
>>
>> But the page isn't mapped into the page table ... there's a swap entry
>> in the page table, so I don't think your logic holds.
>>
> 
> IIUC, device_private_entry will hold one page refcnt when it's set to page table.

Take remove_migration_pte() as example, it will hold extra one page refcnt when set device private entry:
  remove_migration_pte()
    ...
    folio_get(folio);
    ...
    if (unlikely(is_device_private_page(new))) {
      make_[writable|readable]_device_private_entry();
    }
    ...
    set_pte_at

> And there's similar code in do_swap_page():
> 
>   vm_fault_t do_swap_page(struct vm_fault *vmf)
>     if (unlikely(non_swap_entry(entry))) {
>       if (is_device_private_entry(entry))
>         /*
>          * Get a page reference while we know the page can't be
>          * freed.
>          */
>         get_page(vmf->page);
>         pte_unmap_unlock(vmf->pte, vmf->ptl);
>         ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
>         put_page(vmf->page);
>     ...
> 
> If my logic doesn't hold, do_swap_page() will need to fix the code. Or am I miss something?

Can I have your opinion?

Thanks.