[PATCH v3 0/4] xfsprogs: v4 inode type in directory

[Mark Tinguely]

On 10/18/13 17:55, Dave Chinner wrote:
> On Fri, Oct 18, 2013 at 08:22:24AM -0500, Mark Tinguely wrote:
>> On 10/17/13 22:19, Dave Chinner wrote:
>>> On Thu, Oct 17, 2013 at 05:08:56PM -0500, Mark Tinguely wrote:
>>>> On 10/17/13 10:28, Mark Tinguely wrote:
>>>>> Here are the patches that enable the inode in the directory
>>>>> feature in v4 superblocks.
>>>>>
>>>>> Unchanged
>>>>>   patch 1: add the entries to xfs_sb.h (sync with kernel)
>>>>>   patch 2: add the XFS_FSOP_GEOM_FLAGS_FTYPE to xfs_fs.h (sync with kernel)
>>>>> 	  add the entry to repair so that xfs_info reports the feature
>>>>> New
>>>>>   patch 3: add feature to the xfs_db version command.
>>>>>
>>>>> Fixed
>>>>>   patch 4: add the feature to mkfs.xfs and manual page.
>>>>>            note: this new feature is ignored for superblock v5
>>>>> 		automatically turns on this feature.
>>>>
>>>> FYI.
>>>>
>>>> I saw the request for adding the filetype entry to block/leaf after posting.
>>>>
>>>> I have it displaying unconditionally, but am trying to figure out
>>>> how to make it display only for filesytems that support the ftype
>>>> feature. I am missing something in the field.count().
>>>
>>> The count function only tells the code whether a structure is
>>> present or not, but it does not tell you what the format of the
>>> structure is.
>>>
>>> if you look at db/dir2.c, you'll see that the difference between the
>>> dir2_flds[] and the dir3_flds[] is mainly in the type, count and offset
>>> fields. For example:
>>>
>>> const field_t   dir2_flds[] = {
>>>          { "bhdr", FLDT_DIR2_DATA_HDR, OI(BOFF(magic)), dir2_block_hdr_count,
>>>            FLD_COUNT, TYP_NONE }
>>> ...
>>>
>>> const field_t   dir3_flds[] = {
>>>          { "bhdr", FLDT_DIR3_DATA_HDR, OI(B3OFF(hdr)), dir3_block_hdr_count,
>>>            FLD_COUNT, TYP_NONE },
>>> ...
>>>
>>> if you look at dir[23]_block_hdr_count(), you'll see that they
>>> return a boolean value based on a magic number check. Hence when the
>>> code is trying to determine the type of the block that has been read
>>> (i.e. what the field definition is), if the magic number matches we
>>> know exactly what type of contents they contain.
>>>
>>> For decoding the dtype, you need too look at how to select the
>>> correct structure for the FLDT_DIR2_DATA_UNION. If you don't have
>>> the feature set, you need to select the FLDT_DIR2_DATA_UNION
>>> structure type, and if it is set you need to select the
>>> FLDT_DIR3_DATA_UNION type. Hence you need both these types defined
>>> in the dir2_flds[] array, and some manner to ensure the correct
>>> values are returned from the count functions.
>>>
>>> And just to make it hard, both the dir2 and dir3 data union count
>>> functions use the same function (dir2_data_u_count) so you're going
>>> to have to be careful that you don't break the v5 superblock
>>> directory decoding....
>>>
>>> Cheers,
>>>
>>> Dave.
>>
>>
>> Thanks Dave. I did some RTFS and found I was having problems with
>> the field_t.flag.
>>
>> Can't we add a filetype to the dir2 dir2_data_union_flds entry and
>> use the count to turn it on/off? The problem I was having with this
>> was the flag.
>
> ISTR trying something like that previously when doing the dir3
> changeover - I couldn't get it to work properly because parts of the
> structure could not be easily discriminated and so both the v2 and
> v3 count functions would return true because they were looking at
> different locations in the structure due to the header size
> differences.
>
> In this case, because the structure itself doesn't have any child
> dependencies and the format will always be considered a v2 directory
> structure, I think what you propose should work. Did you test it
> to see if it works?
>
> Cheers,
>
> Dave.

Yes, for example:

mkdir directory
mknod device c 100 100
touch file
ln -s file softlink

and on a short form directory:

u.sfdir3.hdr.count = 4
u.sfdir3.hdr.i8count = 0
u.sfdir3.hdr.parent.i4 = 128
u.sfdir3.list[0].namelen = 9
u.sfdir3.list[0].offset = 0x30
u.sfdir3.list[0].name = "directory"
u.sfdir3.list[0].inumber.i4 = 131
u.sfdir3.list[0].filetype = 2
u.sfdir3.list[1].namelen = 6
u.sfdir3.list[1].offset = 0x48
u.sfdir3.list[1].name = "device"
u.sfdir3.list[1].inumber.i4 = 132
u.sfdir3.list[1].filetype = 3
u.sfdir3.list[2].namelen = 4
u.sfdir3.list[2].offset = 0x60
u.sfdir3.list[2].name = "file"
u.sfdir3.list[2].inumber.i4 = 133
u.sfdir3.list[2].filetype = 1
u.sfdir3.list[3].namelen = 8
u.sfdir3.list[3].offset = 0x70
u.sfdir3.list[3].name = "softlink"
u.sfdir3.list[3].inumber.i4 = 134
u.sfdir3.list[3].filetype = 7

then covert it to a block directory:

xfs_db> dblock 0
xfs_db> p
bhdr.magic = 0x58443242
bhdr.bestfree[0].offset = 0x128
bhdr.bestfree[0].length = 0xe50
bhdr.bestfree[1].offset = 0
bhdr.bestfree[1].length = 0
bhdr.bestfree[2].offset = 0
bhdr.bestfree[2].length = 0
bu[0].inumber = 128
bu[0].namelen = 1
bu[0].name = "."
bu[0].filetype = 2
bu[0].tag = 0x10
bu[1].inumber = 128
bu[1].namelen = 2
bu[1].name = ".."
bu[1].filetype = 2
bu[1].tag = 0x20
bu[2].inumber = 131
bu[2].namelen = 9
bu[2].name = "directory"
bu[2].filetype = 2
bu[2].tag = 0x30
bu[3].inumber = 132
bu[3].namelen = 6
bu[3].name = "device"
bu[3].filetype = 3
bu[3].tag = 0x48
bu[4].inumber = 133
bu[4].namelen = 4
bu[4].name = "file"
bu[4].filetype = 1
bu[4].tag = 0x60
bu[5].inumber = 134
bu[5].namelen = 8
bu[5].name = "softlink"
bu[5].filetype = 7


and works without the file type feature turn off. The libxfs code does 
the heavy work.

--Mark