[PATCH] erofs: support STATX_DIOALIGN

[Gao Xiang]

On 2024/7/18 11:35, Hongbo Li wrote:
> 
> 
> On 2024/7/18 10:41, Gao Xiang wrote:
>>
>>
>> On 2024/7/17 14:34, Hongbo Li wrote:
>>>
>>>
>>> On 2024/7/17 10:00, Gao Xiang wrote:
>>>> Hi,
>>>>
>>>> On 2024/7/16 20:45, Hongbo Li wrote:
>>>>> Add support for STATX_DIOALIGN to erofs, so that direct I/O
>>>>> alignment restrictions are exposed to userspace in a generic
>>>>> way.
>>>>>
>>>>> [Before]
>>>>> ```
>>>>> ./statx_test /mnt/erofs/testfile
>>>>> statx(/mnt/erofs/testfile) = 0
>>>>> dio mem align:0
>>>>> dio offset align:0
>>>>> ```
>>>>>
>>>>> [After]
>>>>> ```
>>>>> ./statx_test /mnt/erofs/testfile
>>>>> statx(/mnt/erofs/testfile) = 0
>>>>> dio mem align:512
>>>>> dio offset align:512
>>>>> ```
>>>>>
>>>>> Signed-off-by: Hongbo Li <lihongbo22 at huawei.com>
>>>>> ---
>>>>>   fs/erofs/inode.c | 19 +++++++++++++++++++
>>>>>   1 file changed, 19 insertions(+)
>>>>>
>>>>> diff --git a/fs/erofs/inode.c b/fs/erofs/inode.c
>>>>> index 5f6439a63af7..9325a6f0058a 100644
>>>>> --- a/fs/erofs/inode.c
>>>>> +++ b/fs/erofs/inode.c
>>>>> @@ -342,6 +342,25 @@ int erofs_getattr(struct mnt_idmap *idmap, const struct path *path,
>>>>>       stat->attributes_mask |= (STATX_ATTR_COMPRESSED |
>>>>>                     STATX_ATTR_IMMUTABLE);
>>>>> +    /*
>>>>> +     * Return the DIO alignment restrictions if requested.
>>>>> +     *
>>>>> +     * In erofs, STATX_DIOALIGN is not supported in ondemand mode and
>>>>> +     * the compressed file, so in these cases we report no DIO support.
>>>>> +     */
>>>>> +    if ((request_mask & STATX_DIOALIGN) && S_ISREG(inode->i_mode)) {
>>>>> +        stat->result_mask |= STATX_DIOALIGN;
>>>>> +        if (!erofs_is_fscache_mode(inode->i_sb) &&
>>>>> + !erofs_inode_is_data_compressed(EROFS_I(inode)->datalayout)) {
>>>>> +            struct block_device *bdev = inode->i_sb->s_bdev;
>>>>> +            unsigned int bsize = (bdev) ? bdev_logical_block_size(bdev) :
>>>>> +                        i_blocksize(inode);
>>>>
>>>> I guess in this way you could always use
>>>>              stat->dio_mem_align = bdev_logical_block_size(bdev);
>>>>              stat->dio_offset_align = stat->dio_mem_align;
>>>> ? since bdev won't be NULL.
>>>>
>>> Yeah, only the EROFS_FS_ONDEMAND config is on, the s_bdev can be NULL.
>>
>> Would you mind sending a v2 to fix this?  At least, s_bdev is always
>> non-NULL currently.
>>
>> I could apply this for this cycle.
>>
> I'm working on ondemand direct io mode. This will affect the STATX_DIOALIGN result for erofs_is_fscache_mode case (.ie the dio_mem_align is i_blocksize(inode) in erofs over fscache). Should we apply this first and then modify it after the direct io is supported in ondemand erofs?

Yes, I prefer that.  If you submit a new version
of STATX_DIOALIGN, I'd like to merge this for this
cycle (6.11).

But fscache direct i/o seems that it needs another
cycle tho.

Thanks,
Gao Xiang

> 
> Thanks,
> Hongbo
> 
>> Thanks,
>> Gao Xiang
>>
>>>
>>> Thanks,
>>> Hongbo
>>>
>>>> Otherwise it looks good to me.
>>>>
>>>> Thanks,
>>>> Gao Xiang