Some information about direct-io support on ubifs

[Zhihao Cheng]

在 2024/11/11 23:08, Michael Nazzareno Trimarchi 写道:
> Hi
> 
> On Mon, Nov 11, 2024 at 3:15 PM Zhihao Cheng <chengzhihao1 at huawei.com> wrote:
>>
>> 在 2024/11/11 17:37, Michael Nazzareno Trimarchi 写道:
>>> Hi all
>>>
>>> I would like to have some information on how to implement direct-io
>>> support in ubifs. I have tried to mount squashfs image as file in
>>> ubifs using loopback and direct-io but it seems that loop can not
>>> handle when this is on top of ubifs. From what I understand from the
>>> code it seems that the requirement is that the direct-io is
>>> implemented in the filesystem where the file is stored. I have
>>> simulated using nand sim. The idea is to reduce the buffering. I'm
>>> open to have other suggestions regarding this too.
>>>
>>> Michael
>>> .
>>>
>> Hi Michael,
>> There is a discussion(9 years ago) about ubifs dio[1], of course we
>> cannot implement it as a fake interface now.
>> Before talking about adding dio support for UBIFS, let me know some
>> background information. What's the purpose of using dio for opening
>> squashfs image file? What's the goal of reducing buffer you mentioned
>> above, Better performance? Less memory usage?
>>
>>
>> [1]
>> https://lore.kernel.org/lkml/1440016553-26481-2-git-send-email-richard@nod.at/
> 
> 
> The goal is less memory usage. Each squashfs is mounted using loopback
> device, so according
> to some thread on direct-io, I understand that we then not have
> buffering two times on ubifs level and loop
> level, but the thread was not clear to me. The device has very limited
> capability and memory. I have done some
> test with fuse  but I did not have numbers to compare them
> 

Hi Michael,

I think it is possible to support DIO for UBIFS in theory, but it is not 
an easy task. We need to imeplement read/write functions, and we should 
iterate user data bytes by UBIFS_BLOCK_SIZE unit.
1. Things should be done in write function:
    1) Do space reservation, refer to ubifs_write_begin -> 
allocate_budget, of cource the origin 'req.new_page' cannot be used and 
we should define a new budget type based on UBIFS_BLOCK_SIZE.
    2) record user data in data node, refer to do_writepage -> 
ubifs_jnl_write_data
    3) invalidate page cache after io complete, refer to dio_complete -> 
kiocb_invalidate_post_direct_write
    4) release budget, refer to do_writepage -> 
release_new_page_budget/release_existing_page_budget
2. Things should be done in read function:
    1) sync page cache data to disk, make sure that dio reading get the 
latest user data, refer to __blockdev_direct_IO -> 
filemap_write_and_wait_range
    2) read data from UHIFS, refer to do_readpage -> read_block
Maybe there are many trivial details during developing.
> 
> Michael
> .
>