Received-SPF: pass (google.com: domain of linux-ext4-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Subject: Re: [PATCH -next] ext4: ensure data forced to disk when rename
To:     Jan Kara <jack@suse.cz>, Ye Bin <yebin10@huawei.com>
CC:     <tytso@mit.edu>, <adilger.kernel@dilger.ca>,
        <linux-ext4@vger.kernel.org>, <linux-kernel@vger.kernel.org>
References: <20220901062657.1192732-1-yebin10@huawei.com>
 <20220901083706.mjewb45dufzxcfdk@quack3>
From:   Zhang Yi <yi.zhang@huawei.com>
Message-ID: <ce8666e2-c224-79f9-7770-f87b31d122e3@huawei.com>
Date:   Thu, 1 Sep 2022 21:19:04 +0800
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:78.0) Gecko/20100101
 Thunderbird/78.12.0
MIME-Version: 1.0
In-Reply-To: <20220901083706.mjewb45dufzxcfdk@quack3>
Content-Type: text/plain; charset="utf-8"
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 2022/9/1 16:37, Jan Kara wrote:
> On Thu 01-09-22 14:26:57, Ye Bin wrote:
>> There is issue that data lost when rename new file. Reason is dioread_nolock is
>> enabled default now, which map blocks will mark extent unwritten. But inode
>> size is changed after submit io. Then read file will return zero if extent is
>> unwritten.
>> To solve above isuue, wait all data force to disk before rename when enable
>> dioread_nolock and enable delay allocate.
>>
>> Signed-off-by: Ye Bin <yebin10@huawei.com>
> 
> Well, but it was always like that. We do not guarantee that the data is
> securely on disk before rename - userspace is responsible for that if it
> needs this guarantee. We just want to make the time window shorter and so
> start data writeback because lot of userspace is careless. But waiting for
> data writeback before rename will significantly slow down some workloads
> and IMHO without a good reason.
> 

Hi, Jan

Our case is modifing a no-empty file names "data.conf" through vim, it will
cp "data.conf" to ".data.conf.swp" firstly, overwrite rename ".data.conf.swp"
back to "data.conf" and fsync after modifying. If we power down between rename
and fsync, we may lose all data in the original "data.conf" and get a whole
zero file. Before we enable dioread_nolock by defalut, the newly appending data
may lost, but at least we will not lose the original data in the default
data=ordered and auto_da_alloc mode. It seems that dioread_nolock breaks the
guarantee provide by auto_da_alloc. Maybe we should add a fsync before rename
in vim ?

Thanks,
Yi.