1The XFS File System 2=================== 3 4This document describes how to test the XFS file system, XFS file system API for haiku 5and Its current status on haiku. 6 7 8Testing XFS File System 9----------------------- 10 11There are three ways we can test XFS : 12 13- Using xfs_shell. 14- Using userlandfs. 15- Building a version of haiku with XFS support and then mounting a file system. 16 17But before that we will need to create XFS images for all testing purposes. 18 19Creating File System Images 20^^^^^^^^^^^^^^^^^^^^^^^^^^^ 21 22Currently only linux has full XFS support so we will use linux for generating file system images. 23 24First we need to create an empty sparse image using command:: 25 26 $ dd if=/dev/zero of=fs.img count=0 bs=1 seek=5G 27 28The output will be:: 29 30 0+0 records in 31 0+0 records out 32 0 bytes (0 B) copied, 0.000133533 s, 0.0 kB/s 33 34Do note that we can create images of whatever size or name we want, for example the above command 35creates fs.img of size 5 GB, if we alter seek = 10G it will create fs.img with size 10 GB. 36 37The XFS file system on linux supports two versions, V4 and V5. 38 39To put XFS V5 file system on our sparse image run command:: 40 41 $ /sbin/mkfs.xfs fs.img 42 43The output will be:: 44 45 meta-data =fs.img isize=512 agcount=4, agsize=65536 blks 46 = sectsz=512 attr=2, projid32bit=1 47 = crc=1 finobt=1, sparse=1, rmapbt=0 48 = reflink=1 49 data = bsize=4096 blocks=262144, imaxpct=25 50 = sunit=0 swidth=0 blks 51 naming =version 2 bsize=4096 ascii-ci=0, ftype=1 52 log =internal log bsize=4096 blocks=2560, version=2 53 = sectsz=512 sunit=0 blks, lazy-count=1 54 realtime =none extsz=4096 blocks=0, rtextents=0 55 56To put XFS V4 file system on our sparse image run command:: 57 58 $ /sbin/mkfs.xfs -m crc=0 file.img 59 60The output will be:: 61 62 meta-data=fs.img isize=256 agcount=4, agsize=327680 blks 63 = sectsz=512 attr=2, projid32bit=0 64 data = bsize=4096 blocks=1310720, imaxpct=25 65 = sunit=0 swidth=0 blks 66 naming =version 2 bsize=4096 ascii-ci=0 67 log =internal log bsize=4096 blocks=2560, version=2 68 = sectsz=512 sunit=0 blks, lazy-count=1 69 realtime =none extsz=4096 blocks=0, rtextents=0 70 71**The linux kernel will support older XFS v4 filesystems by default until 2025 and 72Support for the V4 format will be removed entirely in September 2030** 73 74Now we can mount our file system image and create entries for testing XFS haiku driver. 75 76Test using xfs_shell 77^^^^^^^^^^^^^^^^^^^^^^^ 78 79The idea of fs_shell is to run the file system code outside of haiku. We can run it 80as an application, it provides a simple command line interface to perform various 81operations on the file system (list directories, read and display files, etc). 82 83First we have to compile it:: 84 85 jam "<build>xfs_shell" 86 87Then run it:: 88 89 jam run ":<build>xfs_shell" fs.img 90 91Where fs.img is the file system image we created from linux kernel. 92 93Test directly inside Haiku 94^^^^^^^^^^^^^^^^^^^^^^^^^^ 95 96First build a version of haiku with XFS support, to do this we need to add "xfs" to the `image 97definition <https://git.haiku-os.org/haiku/tree/build/jam/images/definitions/minimum#n239>`__. 98 99Then compile haiku as usual and run the resulting system in a virtual machine or on real hardware. 100 101We can then try to mount an XFS file system using command on Haiku:: 102 103 mount -t xfs <path to image> <path to mount folder> 104 105for example:: 106 107 mount -t xfs /boot/home/Desktop/fs.img /boot/home/Desktop/Testing 108 109Here fs.img is file system image and Testing is mount point. 110 111Test using userlandfs 112^^^^^^^^^^^^^^^^^^^^^ 113 114To be updated 115 116 117Haiku XFS API 118------------- 119 120All the necessary hooks for file system like xfs_mount(), open_dir(), read_dir() etc.. are 121implemented in the **kernel_interface.cpp** file. It acts as an interface between the Haiku kernel 122and the XFS file system. Documentation for all necessary file system hooks can be found 123`in the API reference <https://www.haiku-os.org/docs/api/fs_modules.html>`_ 124 125Whenever we run a file system under fs_shell we can't use system headers, fs_shell compatible 126headers are there which needs to be used whenever we try to mount XFS file system using xfs_shell. 127To resolve this problem we use **system_dependencies.h** header file which takes care to use 128correct headers whenever we mount XFS file system either using xfs_shell or directly inside Haiku. 129 130XFS stores data on disk in Big Endian byte order, to convert data into host order 131all classes and data headers has **SwapEndian()** function, Its better to have all data 132conversions at one place to avoid future problems related to data byte order. 133 134XFS SuperBlock starts at ondisk offset 0, the definition of SuperBlock is in **xfs.h** file. 135 136A Volume is an instance of file system and defined in **Volume.h** file. 137XFS Volume contains SuperBlock, file system device and essential functions 138like Identify(), mount() etc... 139 140* *Identify()* function reads SuperBlock from disk and verifies it. 141* *Mount()* function mounts file system device and publishes root inode of file system 142 (Typically root inode number for XFS is 128). 143 144XFS uses TRACE Macro to debug file system, definitions for TRACE, ERROR and ASSERT 145are defined at **Debug.h** in the form of Macro. 146 147To enable TRACE calls just add ``#define TRACE_XFS`` in Debug.h file and 148vice versa to disable it. 149 150 151XFS V5 introduced metadata checksums to ensure the integrity of metadata in file system, 152It uses CRC32C checksum algorithm. For XFS all checksums related functions are defined in 153**Checksum.h** header file. 154It contains following functions : 155 156* *xfs_verify_cksum()* to verify checksum for buffer. 157* *xfs_update_cksum()* to update checksum for buffer. 158 159**XFS stores checksum in little endian byte order unlike other ondisk data which is stored 160in big endian byte order** 161 162XFS V5 introduced many other fields for metadata verification like *BlockNo* *UUID* *Owner* 163etc.. All this fields are common in every data header and so are their checks. So to not 164repeat same checks again and again for all headers we created a *VerifyHeader* template 165function which is defined in **VerifyHeader.h** file. This function is commonly used in 166all forms of headers for verification purposes. 167 168Inodes 169^^^^^^ 170 171XFS inodes comes in three versions: 172 173* Inode V1 & V2. (Version 4 XFS) 174* Inode V3. (Version 5 XFS) 175 176Version 1 inode support is already deprecated on linux kernel, Haiku XFS supports it only 177in read format. When we will have write support for XFS we will only support V2 and V3 inodes. 178 179V1 & V2 inodes are 256 bytes while V3 inodes are 512 bytes in size allowing more data to be 180stored directly inside inode. 181 182**CoreInodeSize()** is a helper funtion which returns size of inode based on version of XFS and 183is used throughout our XFS code. 184 185**DIR_DFORK_PTR** is a Macro which expands to void pointer to the data offset in inode, which 186could be either shortform entries, extents or B+Tree root node depending on the data format 187of inode (di_format). 188 189Similarly **DIR_AFORK_PTR** Macro expands to void pointer to the attribute offset in inode, 190which could be either shortform attributes, attributes extents or B+Tree node depending on 191the attribute format of Inode (di_aformat). 192 193Since size of inodes could differ based on different versions of XFS we pass CoreInodeSize() 194function as a parameter to DIR_DFORK_PTR and DIR_AFORK_PTR macros to return correct pointer offset. 195 196**di_forkoff** specifies the offset into the inode's literal area where the extended attribute 197fork starts. This value is initially zero until an extended attribute is created. 198It is fixed for V1 & V2 inode's while for V3 Inodes it is dynamic in size, 199allowing complete use of inode's literal area. 200 201Directories 202^^^^^^^^^^^ 203 204Depending on the number of entries inside directory, XFS divides directories into five formats : 205 206* Shortform directory. 207* Block directory. 208* Leaf directory. 209* Node directory. 210* B+Tree directorcy. 211 212Class DirectoryIterator in **Directory.h** file provides an interface between kernel request 213to open, read directory and all forms of directories. It first identifies correct format of 214entries inside inode and then returns request as per format found. 215 216**Shortform directory** 217 218* When the number of entries inside directory are small enough such that we can store all 219 metadata inside inode itself, this form of directory is known as shortform directory. 220* We can check if a directory is shortform if the format of inode is *XFS_DINODE_FMT_LOCAL*. 221* The header for ShortForm entries is located at data fork pointer inside inode, which we cast 222 directly to *ShortFormHeader*. 223* Since number of entries are short we can simply iterate over all entries for *Lookup()* and 224 *GetNext()* functions. 225 226**Block directory** 227 228* When number of entries expand such that we can no longer store all directory metadata 229 inside inode we use extents. 230* We can check if a directory is extent based if the format of inode is *XFS_DINODE_FMT_EXTENTS*. 231* In Block directory we have a single directory block for Data header, leaf header 232 and free data header. This simple fact helps us to determine if given extent format 233 in inode is block directory. 234* Since XFS V4 & V5 data headers differs we use a virtual class *ExtentDataHeader* which 235 acts as an interface between V4 & V5 data header, this class only stores pure virtual 236 functions and no data. 237* *CreateDataHeader* returns a class instance based on the version of XFS mounted. 238* Since now we have a virtual class with V_PTRS we need to be very careful with data stored 239 ondisk and data inside class, for example we now can't use sizeof() operator on class to 240 return its size which is consistent with its size inside disk. To handle this issue helper 241 function like *SizeOfDataHeader* are created which needs to be used instead of sizeof() operator. 242* In *GetNext()* function we simply iterate over all entries inside buffer, though a found 243 entry could be unused entry so we need to have checks if a entry found is proper entry. 244* In *Lookup()* function first we generate a hash value of entry for lookup, then we find 245 lowerbound of this hash value inside leaf entries to get address of entry inside data. 246 At last if entry matches we return B_OK else we return B_ENTRY_NOT_FOUND. 247 248**Leaf directory** 249 250* When number of entries expand such that we can no longer store all directory metadata inside 251 directory block we use leaf format. 252* In leaf directory we have a multiple directory block for Data header and free data header, 253 while single directory block for leaf header. 254* To check if given extent based inode is leaf type, we simply check for offset inside last 255 extent map, if its equal to *LEAF_STARTOFFSET* then the given inode is leaf type else it is 256 node type. 257* Since XFS V4 & V5 leaf headers differs we use a virtual class *ExtentLeafHeader* which acts 258 as an interface between V4 & V5 leaf header, this class only stores pure virtual functions 259 and no data. 260* *CreateLeafHeader* returns a class instance based on the version of XFS mounted. 261* Instead of sizeof() operator on ExtentLeafHeader we should always use *SizeOfLeafHeader()* function 262 to return correct size of class inside disk. 263* *Lookup()* and *GetNext()* functions are similar to block directories except now we don't use single 264 directory block buffer. 265 266TODO : Document Node and B+Tree based directories. 267 268Files 269^^^^^ 270 271XFS stores files in two formats : 272 273* Extent based file. 274* B+Tree based file. 275 276All implementation of read support for files is inside *Inode()* class in **Inode.h** file. 277 278When the format inside inode of file is *XFS_DINODE_FMT_EXTENTS* it is an extent based file, 279to read all data of file we simply iterate over all extents which is very similar to how we 280do it in Extent based directories. 281 282When the file becomes too large such that we cannot store more extent maps inside inode the 283format of file is changed to B+Tree. When the format inside inode of file is 284*XFS_DINODE_FMT_BTREE* it is an B+Tree based file, to read all data of file 285first we read blocks of B+Tree to extract extent maps and then read extents 286to get file's data. 287 288 289Current Status of XFS 290--------------------- 291 292Currently we only have read support for XFS, below briefly summarises read support for all formats. 293 294 295Directories 296^^^^^^^^^^^ 297 298**Short-Directory** 299 Stable read support for both V4 and V5 inside Haiku. 300 301**Block-Directory** 302 Stable read support for both V4 and V5 inside Haiku. 303 304**Leaf-Directory** 305 Stable read support for both V4 and V5 inside Haiku. 306 307**Node-Directory** 308 Stable read support for both V4 and V5 inside Haiku. 309 310**B+Tree-Directory** 311 Unstable read support for both V4 and V5, due to so many read from disk entire 312 process inside Haiku is too slow. 313 314Files 315^^^^^ 316 317**Extent based Files** 318 | *xfs_shell* - stable read support for both V4 and V5. 319 | *Haiku* - Unstable, Cat command doesn't print entire file and never terminates process. 320 321**B+Tree based Files** 322 | *xfs_shell* - stable read support for both V4 and V5. 323 | *Haiku* - Unstable, Cat command doesn't print entire file and never terminates process. 324 325Attributes 326^^^^^^^^^^ 327 328Currently we have no extended attributes support for xfs. 329 330Symlinks 331^^^^^^^^ 332 333Currently we have no symlinks support for xfs. 334 335XFS V5 exclusive features 336^^^^^^^^^^^^^^^^^^^^^^^^^ 337 338**MetaData Checksumming** 339 Metadata checksums for superblock, Inodes, and data headers are implemented. 340 341**Big Timestamps** 342 Currently we have no support. 343 344**Reverse mapping btree** 345 Currently we have no support, this data structure is still under construction 346 and testing inside linux kernel. 347 348**Refrence count btree** 349 Currently we have no support, this data structure is still under construction 350 and testing inside linux kernel. 351 352Write Support 353^^^^^^^^^^^^^ 354 355Currently we have no write support for xfs. 356 357 358References 359---------- 360 361The best and only reference for xfs is latest version of "xfs_filesystem_structure" 362written by Linux-XFS developers. 363 364The pdf version of above Doc can be found 365`here <http://ftp.ntu.edu.tw/linux/utils/fs/xfs/docs/xfs_filesystem_structure.pdf>`_ 366