Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org 7E74A2173B
Subject: Re: [f2fs-dev] [PATCH] f2fs: add fi->commit_lock to protect commit
 GCed pages
To:     Yunlong Song <yunlong.song@huawei.com>, jaegeuk@kernel.org,
        yuchao0@huawei.com, yunlong.song@icloud.com
Cc:     linux-kernel@vger.kernel.org,
        linux-f2fs-devel@lists.sourceforge.net, heyunlei@huawei.com,
        miaoxie@huawei.com
References: <1517626068-49739-1-git-send-email-yunlong.song@huawei.com>
 <1517888990-96478-1-git-send-email-yunlong.song@huawei.com>
 <4493cbf2-6f37-6c04-a012-4b2516b3b4e7@kernel.org>
 <134848f9-2dd6-0efa-3ccf-3c29eeaf5534@huawei.com>
 <ae9a153e-4e9a-648d-3804-2a1e4a21a9c1@kernel.org>
 <a2f431bf-ac2f-0b03-6b7f-aeb51e815e17@huawei.com>
 <d67aae9c-778a-50bd-9f73-f90c3efd9b73@kernel.org>
 <33b262b8-7701-a150-09e5-9ae1b4733835@huawei.com>
From:   Chao Yu <chao@kernel.org>
Message-ID: <0b3d2792-b3cf-3302-8313-404a25c86854@kernel.org>
Date:   Fri, 9 Feb 2018 21:38:33 +0800
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:52.0) Gecko/20100101
 Thunderbird/52.6.0
MIME-Version: 1.0
In-Reply-To: <33b262b8-7701-a150-09e5-9ae1b4733835@huawei.com>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 2018/2/9 21:29, Yunlong Song wrote:
> Back to the problem, if we skip out, then the f2fs_gc will go
> into dead loop if the apps only atomic start but never atomic

That's another issue, which I have suggest to set a threshold time
to release atomic/volatile pages by balance_fs_bg.

Thanks,

> commit. The main aim of my two patches is to remove the skip
> action to avoid the dead loop.
> 
> On 2018/2/9 21:26, Chao Yu wrote:
>> On 2018/2/9 20:56, Yunlong Song wrote:
>>> As what I point in last mail, if the atomic file is not committed
>>> yet, gc_data_segment will register_inmem_page the GCed data pages.
>>
>> We will skip GCing that page as below check:
>>
>> - move_data_{page,block}
>>   - f2fs_is_atomic_file()
>>     skip out;
>>
>> No?
>>
>> Thanks,
>>
>>> This will cause these data pages written twice, the first write
>>> happens in move_data_page->do_write_data_page, and the second
>>> write happens in later __commit_inmem_pages->do_write_data_page.
>>>
>>> On 2018/2/9 20:44, Chao Yu wrote:
>>>> On 2018/2/8 11:11, Yunlong Song wrote:
>>>>> Then the GCed data pages are totally mixed with the inmem atomic pages,
>>>>
>>>> If we add dio_rwsem, GC flow is exclude with atomic write flow. There
>>>> will be not race case to mix GCed page into atomic pages.
>>>>
>>>> Or you mean:
>>>>
>>>>                      - gc_data_segment
>>>>                       - move_data_page
>>>>                        - f2fs_is_atomic_file
>>>> - f2fs_ioc_start_atomic_write
>>>>    - set_inode_flag(inode, FI_ATOMIC_FILE);
>>>>                        - f2fs_set_data_page_dirty
>>>>                         - register_inmem_page
>>>>
>>>> In this case, GCed page can be mixed into database transaction, but could
>>>> it cause any problem except break rule of isolation for transaction.
>>>>
>>>>> this will cause the atomic commit ops write the GCed data pages twice
>>>>> (the first write happens in GC).
>>>>>
>>>>> How about using the early two patches to separate the inmem data pages
>>>>> and GCed data pages, and use dio_rwsem instead of this patch to fix the
>>>>> dnode page problem (dnode page commited but data page are not committed
>>>>> for the GCed page)?
>>>>
>>>> Could we fix the race case first, based on that fixing, and then find the
>>>> place that we can improve?
>>>>
>>>>>
>>>>>
>>>>> On 2018/2/7 20:16, Chao Yu wrote:
>>>>>> On 2018/2/6 11:49, Yunlong Song wrote:
>>>>>>> This patch adds fi->commit_lock to avoid the case that GCed node pages
>>>>>>> are committed but GCed data pages are not committed. This can avoid the
>>>>>>> db file run into inconsistent state when sudden-power-off happens if
>>>>>>> data pages of atomic file is allowed to be GCed before.
>>>>>>
>>>>>> do_fsync:                GC:
>>>>>> - mutex_lock(&fi->commit_lock);
>>>>>>                       - lock_page()
>>>>>>                        - mutex_lock(&fi->commit_lock);
>>>>>>     - lock_page()
>>>>>>
>>>>>>
>>>>>> Well, please consider lock dependency & code complexity, IMO, reuse
>>>>>> fi->dio_rwsem[WRITE] will be enough as below:
>>>>>>
>>>>>> ---
>>>>>>     fs/f2fs/file.c | 3 +++
>>>>>>     fs/f2fs/gc.c   | 5 -----
>>>>>>     2 files changed, 3 insertions(+), 5 deletions(-)
>>>>>>
>>>>>> diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
>>>>>> index 672a542e5464..1bdc11feb8d0 100644
>>>>>> --- a/fs/f2fs/file.c
>>>>>> +++ b/fs/f2fs/file.c
>>>>>> @@ -1711,6 +1711,8 @@ static int f2fs_ioc_commit_atomic_write(struct file *filp)
>>>>>>
>>>>>>         inode_lock(inode);
>>>>>>
>>>>>> +    down_write(&F2FS_I(inode)->dio_rwsem[WRITE]);
>>>>>> +
>>>>>>         if (f2fs_is_volatile_file(inode))
>>>>>>             goto err_out;
>>>>>>
>>>>>> @@ -1729,6 +1731,7 @@ static int f2fs_ioc_commit_atomic_write(struct file *filp)
>>>>>>             ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 1, false);
>>>>>>         }
>>>>>>     err_out:
>>>>>> +    up_write(&F2FS_I(inode)->dio_rwsem[WRITE]);
>>>>>>         inode_unlock(inode);
>>>>>>         mnt_drop_write_file(filp);
>>>>>>         return ret;
>>>>>> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
>>>>>> index b9d93fd532a9..e49416283563 100644
>>>>>> --- a/fs/f2fs/gc.c
>>>>>> +++ b/fs/f2fs/gc.c
>>>>>> @@ -622,9 +622,6 @@ static void move_data_block(struct inode *inode, block_t bidx,
>>>>>>         if (!check_valid_map(F2FS_I_SB(inode), segno, off))
>>>>>>             goto out;
>>>>>>
>>>>>> -    if (f2fs_is_atomic_file(inode))
>>>>>> -        goto out;
>>>>
>>>> Seems that we need this check.
>>>>
>>>>>> -
>>>>>>         if (f2fs_is_pinned_file(inode)) {
>>>>>>             f2fs_pin_file_control(inode, true);
>>>>>>             goto out;
>>>>>> @@ -729,8 +726,6 @@ static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
>>>>>>         if (!check_valid_map(F2FS_I_SB(inode), segno, off))
>>>>>>             goto out;
>>>>>>
>>>>>> -    if (f2fs_is_atomic_file(inode))
>>>>>> -        goto out;
>>>>
>>>> Ditto.
>>>>
>>>> Thanks,
>>>>
>>>>>>         if (f2fs_is_pinned_file(inode)) {
>>>>>>             if (gc_type == FG_GC)
>>>>>>                 f2fs_pin_file_control(inode, true);
>>>>>>
>>>>>
>>>>
>>>> .
>>>>
>>>
>>
>> .
>>
>