Message-ID: <8257903a-1905-49c5-bed4-d15ca06c6d3b@huaweicloud.com>
Date:   Wed, 19 Jul 2023 15:28:53 +0800
From:   Yu Kuai <yukuai1@...weicloud.com>
To:     Sergei Shtepa <sergei.shtepa@...am.com>,
        Yu Kuai <yukuai1@...weicloud.com>, axboe@...nel.dk,
        hch@...radead.org, corbet@....net, snitzer@...nel.org
Cc:     viro@...iv.linux.org.uk, brauner@...nel.org, dchinner@...hat.com,
        willy@...radead.org, dlemoal@...nel.org, linux@...ssschuh.net,
        jack@...e.cz, ming.lei@...hat.com, linux-block@...r.kernel.org,
        linux-doc@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-fsdevel@...r.kernel.org,
        Donald Buczek <buczek@...gen.mpg.de>,
        "yukuai (C)" <yukuai3@...wei.com>
Subject: Re: [PATCH v5 02/11] block: Block Device Filtering Mechanism

Hi,

在 2023/07/19 0:33, Sergei Shtepa 写道:
> 
> 
> On 7/18/23 14:32, Yu Kuai wrote:
>> Subject:
>> Re: [PATCH v5 02/11] block: Block Device Filtering Mechanism
>> From:
>> Yu Kuai <yukuai1@...weicloud.com>
>> Date:
>> 7/18/23, 14:32
>>
>> To:
>> Sergei Shtepa <sergei.shtepa@...am.com>, Yu Kuai <yukuai1@...weicloud.com>, axboe@...nel.dk, hch@...radead.org, corbet@....net, snitzer@...nel.org
>> CC:
>> viro@...iv.linux.org.uk, brauner@...nel.org, dchinner@...hat.com, willy@...radead.org, dlemoal@...nel.org, linux@...ssschuh.net, jack@...e.cz, ming.lei@...hat.com, linux-block@...r.kernel.org, linux-doc@...r.kernel.org, linux-kernel@...r.kernel.org, linux-fsdevel@...r.kernel.org, Donald Buczek <buczek@...gen.mpg.de>, "yukuai (C)" <yukuai3@...wei.com>
>>
>>
>> Hi,
>>
>> 在 2023/07/18 19:25, Sergei Shtepa 写道:
>>> Hi.
>>>
>>> On 7/18/23 03:37, Yu Kuai wrote:
>>>> Subject:
>>>> Re: [PATCH v5 02/11] block: Block Device Filtering Mechanism
>>>> From:
>>>> Yu Kuai <yukuai1@...weicloud.com>
>>>> Date:
>>>> 7/18/23, 03:37
>>>>
>>>> To:
>>>> Sergei Shtepa <sergei.shtepa@...am.com>, Yu Kuai <yukuai1@...weicloud.com>, axboe@...nel.dk, hch@...radead.org, corbet@....net, snitzer@...nel.org
>>>> CC:
>>>> viro@...iv.linux.org.uk, brauner@...nel.org, dchinner@...hat.com, willy@...radead.org, dlemoal@...nel.org, linux@...ssschuh.net, jack@...e.cz, ming.lei@...hat.com, linux-block@...r.kernel.org, linux-doc@...r.kernel.org, linux-kernel@...r.kernel.org, linux-fsdevel@...r.kernel.org, Donald Buczek <buczek@...gen.mpg.de>, "yukuai (C)" <yukuai3@...wei.com>
>>>>
>>>>
>>>> Hi,
>>>>
>>>> 在 2023/07/17 22:39, Sergei Shtepa 写道:
>>>>>
>>>>>
>>>>> On 7/11/23 04:02, Yu Kuai wrote:
>>>>>> bdev_disk_changed() is not handled, where delete_partition() and
>>>>>> add_partition() will be called, this means blkfilter for partiton will
>>>>>> be removed after partition rescan. Am I missing something?
>>>>>
>>>>> Yes, when the bdev_disk_changed() is called, all disk block devices
>>>>> are deleted and new ones are re-created. Therefore, the information
>>>>> about the attached filters will be lost. This is equivalent to
>>>>> removing the disk and adding it back.
>>>>>
>>>>> For the blksnap module, partition rescan will mean the loss of the
>>>>> change trackers data. If a snapshot was created, then such
>>>>> a partition rescan will cause the snapshot to be corrupted.
>>>>>
>>>>
>>>> I haven't review blksnap code yet, but this sounds like a problem.
>>>
>>> I can't imagine a case where this could be a problem.
>>> Partition rescan is possible only if the file system has not been
>>> mounted on any of the disk partitions. Ioctl BLKRRPART will return
>>> -EBUSY. Therefore, during normal operation of the system, rescan is
>>> not performed.
>>> And if the file systems have not been mounted, it is possible that
>>> the disk partition structure has changed or the disk in the media
>>> device has changed. In this case, it is better to detach the
>>> filter, otherwise it may lead to incorrect operation of the module.
>>>
>>> We can add prechange/postchange callback functions so that the
>>> filter can track rescan process. But at the moment, this is not
>>> necessary for the blksnap module.
>>
>> So you mean that blkfilter is only used for the case that partition
>> is mounted? (Or you mean that partition is opened)
>>
>> Then, I think you mean that filter should only be used for the partition
>> that is opended? Otherwise, filter can be gone at any time since
>> partition rescan can be gone.
>>
>> //user
>> 1. attach filter
>>          // other context rescan partition
>> 2. mount fs
>> // user will found filter is gone.
> 
> Mmm...  The fact is that at the moment the user of the filter is the
> blksnap module. There are no other filter users yet. The blksnap module
> solves the problem of creating snapshots, primarily for backup purposes.
> Therefore, the main use case is to attach a filter for an already running
> system, where all partitions are marked up, file systems are mounted.
> 
> If the server is being serviced, during which the disk is being
> re-partitioned, then disabling the filter is normal. In this case, the
> change tracker will be reset, and at the next backup, the filter will be
> attached again.

Thanks for the explanation, I was thinking that blkshap can replace
dm-snapshot.

Thanks,
Kuai

> 
> But if I were still solving the problem of saving the filter when rescanning,
> then it is necessary to take into account the UUID and name of the partition
> (struct partition_meta_info). It is unacceptable that due to a change in the
> structure of partitions, the filter is attached to another partition by mistake.
> The changed() callback would also be good to add so that the filter receives
> a notification that the block device has been updated.
> 
> But I'm not sure that this should be done, since if some code is not used in
> the kernel, then it should not be in the kernel.
> 
>>
>> Thanks,
>> Kuai
>>
>>>
>>> Therefore, I will refrain from making changes for now.
>>>
>>>>
>>>> possible solutions I have in mind:
>>>>
>>>> 1. Store blkfilter for each partition from bdev_disk_changed() before
>>>> delete_partition(), and add blkfilter back after add_partition().
>>>>
>>>> 2. Store blkfilter from gendisk as a xarray, and protect it by
>>>> 'open_mutex' like 'part_tbl', block_device can keep the pointer to
>>>> reference blkfilter so that performance from fast path is ok, and the
>>>> lifetime of blkfiter can be managed separately.
>>>>
>>>>> There was an idea to do filtering at the disk level,
>>>>> but I abandoned it.
>>>>> .
>>>>>
>>>> I think it's better to do filtering at the partition level as well.
>>>>
>>>> Thanks,
>>>> Kuai
>>>>
>>> .
>>>
>>
> .
> 

Powered by blists - more mailing lists