ToDo: - Find and fix bugs. - W.r.t. s_maxbytes still need to be careful on reading/truncating as there are dragons lurking in the details, e.g. read_inode() currently does no checks for file size wrt s_maxbytes. So what happens when a user open()s a file with i_size > s_maxbytes? Should read_inode() truncate the visible i_size? Will the user just get -E2BIG (or whatever) on open()? Or will (s)he be able to open() but lseek() and read() will fail when s_maxbytes is reached? -> Investigate this! - Implement/allow non-resident index bitmaps in dir.c::ntfs_readdir() and then also consider initialized_size w.r.t. the bitmaps, etc. - vcn_to_lcn() should somehow return the correct pointer within the ->run_list so we can get at the lcns for the following vcns, this is strictly a speed optimization. Obviously need to keep the ->run_list locked or RACE. load_attribute_list() already performs such an optimization so use the same optimization where desired. - Consider if ntfs_file_read_compressed_block() shouldn't be coping with initialized_size < data_size. I don't think it can happen but it requires more careful consideration. - CLEANUP: Modularise and reuse code in aops.c. At the moment we have several copies of almost identicall functions and the functions are quite big. Modularising them a bit, e.g. a-la get_block(), will make them cleaner and make code reuse easier. - Want to use dummy inodes for address space i/o. 2.0.10 - There can only be 2^32 - 1 inodes on an NTFS volume. - Add check at mount time to verify that the number of inodes on the volume does not exceed 2^32 - 1, which is the maximum allowed for NTFS according to Microsoft. - Change mft_no member of ntfs_inode structure to be unsigned long. Update all users. This makes ntfs_inode->mft_no just a copy of struct inode->i_ino. But we can't just always use struct inode->i_ino and remove mft_no because extent inodes do not have an attached struct inode. 2.0.9 - Decompression engine now uses a single buffer and other cleanups. - Change decompression engine to use a single buffer protected by a spin lock instead of per-CPU buffers. (Rusty Russell) - Switch to using the new KM_BIO_SRC_IRQ for atomic kmaps. (Andrew Morton) - Change buffer size in ntfs_readdir()/ntfs_filldir() to use NLS_MAX_CHARSET_SIZE which makes the buffers almost 1kiB each but it also makes everything safer so it is a good thing. - Miscellaneous minor cleanups to comments. 2.0.8 - Major updates for handling of case sensitivity and dcache aliasing. Big thanks go to Al Viro and other inhabitants of #kernel for investing their time to discuss the case sensitivity and dcache aliasing issues. - Remove unused source file fs/ntfs/attraops.c. - Remove show_inodes mount option(s), thus dropping support for displaying of short file names. - Remove deprecated mount option posix. - Restore show_sys_files mount option. - Add new mount option case_sensitive, to determine if the driver treats file names as case sensitive or not. If case sensitive, create file names in the POSIX namespace. Otherwise create file names in the LONG/WIN32 namespace. Note, files remain accessible via their short file name, if it exists. - Remove really dumb logic bug in boot sector recovery code. - Fix dcache aliasing issues wrt short/long file names via changes to fs/ntfs/dir.c::ntfs_lookup_inode_by_name() and fs/ntfs/namei.c::ntfs_lookup(): - Add additional argument to ntfs_lookup_inode_by_name() in which we return information about the matching file name if the case is not matching or the match is a short file name. See comments above the function definition for details. - Change ntfs_lookup() to only create dcache entries for the correctly cased file name and only for the WIN32 namespace counterpart of DOS namespace file names. This ensures we have only one dentry per directory and also removes all dcache aliasing issues between short and long file names once we add write support. See comments above function for details. - Fix potential 1 byte overflow in fs/ntfs/unistr.c::ntfs_ucstonls(). 2.0.7 - Minor cleanups and updates for changes in core kernel code. - Remove much of the NULL struct element initializers. - Various updates to make compatible with recent kernels. - Remove defines of MAX_BUF_PER_PAGE and include linux/buffer_head.h in fs/ntfs/ntfs.h instead. - Remove no longer needed KERNEL_VERSION checks. We are now in the kernel proper so they are no longer needed. 2.0.6 - Major bugfix to make compatible with other kernel changes. - Initialize the mftbmp address space properly now that there are more fields in the struct address_space. This was leading to hangs and oopses on umount since 2.5.12 because of changes to other parts of the kernel. We probably want a kernel generic init_address_space() function... - Drop BKL from ntfs_readdir() after consultation with Al Viro. The only caller of ->readdir() is vfs_readdir() which holds i_sem during the call, and i_sem is sufficient protection against changes in the directory inode (including ->i_size). - Use generic_file_llseek() for directories (as opposed to default_llseek()) as this downs i_sem instead of the BKL which is what we now need for exclusion against ->f_pos changes considering we no longer take the BKL in ntfs_readdir(). 2.0.5 - Major bugfix. Buffer overflow in extent inode handling. - No need to set old blocksize in super.c::ntfs_fill_super() as the VFS does so via invocation of deactivate_super() calling fs->fill_super() calling block_kill_super() which does it. - BKL moved from VFS into dir.c::ntfs_readdir(). (Linus Torvalds) -> Do we really need it? I don't think so as we have exclusion on the directory ntfs_inode rw_semaphore mrec_lock. We mmight have to move the ->f_pos accesses under the mrec_lock though. Check this... - Fix really, really, really stupid buffer overflow in extent inode handling in mft.c::map_extent_mft_record(). 2.0.4 - Cleanups and updates for kernel 2.5.11. - Add documentation on how to use the MD driver to be able to use NTFS stripe and volume sets in Linux and generally cleanup documentation a bit. Remove all uses of kdev_t in favour of struct block_device *: - Change compress.c::ntfs_file_read_compressed_block() to use sb_getblk() instead of getblk(). - Change super.c::ntfs_fill_super() to use bdev_hardsect_size() instead of get_hardsect_size(). - No need to get old blocksize in super.c::ntfs_fill_super() as fs/super.c::get_sb_bdev() already does this. - Set bh->b_bdev instead of bh->b_dev throughout aops.c. 2.0.3 - Small bug fixes, cleanups, and performance improvements. - Remove some dead code from mft.c. - Optimize readpage and read_block functions throughout aops.c so that only initialized blocks are read. Non-initialized ones have their buffer head mapped, zeroed, and set up to date, without scheduling any i/o. Thanks to Al Viro for advice on how to avoid the device i/o. Thanks go to Andrew Morton for spotting the below: - Fix buglet in allocate_compression_buffers() error code path. - Call flush_dcache_page() after modifying page cache page contents in ntfs_file_readpage(). - Check for existence of page buffers throughout aops.c before calling create_empty_buffers(). This happens when an I/O error occurs and the read is retried. (It also happens once writing is implemented so that needed doing anyway but I had left it for later...) - Don't BUG_ON() uptodate and/or mapped buffers throughout aops.c in readpage and read_block functions. Reasoning same as above (i.e. I/O error retries and future write code paths.) 2.0.2 - Minor updates and cleanups. - Cleanup: rename mst.c::__post_read_mst_fixup to post_write_mst_fixup and cleanup the code a bit, removing the unused size parameter. - Change default fmask to 0177 and update documentation. - Change attrib.c::get_attr_search_ctx() to return the search context directly instead of taking the address of a pointer. A return value of NULL means the allocation failed. Updated all callers appropriately. - Update to 2.5.9 kernel (preserving backwards compatibility) by replacing all occurences of page->buffers with page_buffers(page). - Fix minor bugs in run list merging, also minor cleanup. - Updates to bootsector layout and mft mirror contents descriptions. - Small bug fix in error detection in unistr.c and some cleanups. - Grow name buffer allocations in unistr.c in aligned mutlipled of 64 bytes. 2.0.1 - Minor updates. - Make default umask correspond to documentation. - Improve documentation. - Set default mode to include execute bit. The {u,f,d}mask can be used to take it away if desired. This allows binaries to be executed from a mounted ntfs partition. 2.0.0 - New version number. Remove TNG from the name. Now in the kernel. - Add kill_super, just keeping up with the vfs changes in the kernel. - Repeat some changes from tng-0.0.8 that somehow got lost on the way from the CVS import into BitKeeper. - Begin to implement proper handling of allocated_size vs initialized_size vs data_size (i.e. i_size). Done are mft.c::ntfs_mft_readpage(), aops.c::end_buffer_read_index_async(), and attrib.c::load_attribute_list(). - Lock the run list in attrib.c::load_attribute_list() while using it. - Fix memory leak in ntfs_file_read_compressed_block() and generally clean up compress.c a little, removing some uncommented/unused debug code. - Tidy up dir.c a little bit. - Don't bother getting the run list in inode.c::ntfs_read_inode(). - Merge mft.c::ntfs_mft_readpage() and aops.c::ntfs_index_readpage() creating aops.c::ntfs_mst_readpage(), improving the handling of holes and overflow in the process and implementing the correct equivalent of ntfs_file_get_block() in ntfs_mst_readpage() itself. I am aiming for correctness at the moment. Modularisation can come later. - Rename aops.c::end_buffer_read_index_async() to end_buffer_read_mst_async() and optimize the overflow checking and handling. - Use the host of the mftbmp address space mapping to hold the ntfs volume. This is needed so the async i/o completion handler can retrieve a pointer to the volume. Hopefully this will not cause problems elsewhere in the kernel... Otherwise will need to use a fake inode. - Complete implementation of proper handling of allocated_size vs initialized_size vs data_size (i.e. i_size) in whole driver. Basically aops.c is now completely rewritten. - Change NTFS driver name to just NTFS and set version number to 2.0.0 to make a clear distinction from the old driver which is still on version 1.1.22. tng-0.0.8 - 08/03/2002 - Now using BitKeeper, http://linux-ntfs.bkbits.net/ - Replace bdevname(sb->s_dev) with sb->s_id. - Remove now superfluous new-line characters in all callers of ntfs_debug(). - Apply kludge in ntfs_read_inode(), setting i_nlink to 1 for directories. Without this the "find" utility gets very upset which is fair enough as Linux/Unix do not support directory hard links. - Further run list merging work. (Richard Russon) - Backwards compatibility for gcc-2.95. (Richard Russon) - Update to kernel 2.5.5-pre1 and rediff the now tiny patch. - Convert to new file system declaration using ->ntfs_get_sb() and replacing ntfs_read_super() with ntfs_fill_super(). - Set s_maxbytes to MAX_LFS_FILESIZE to avoid page cache page index overflow on 32-bit architectures. - Cleanup upcase loading code to use ntfs_(un)map_page(). - Disable/reenable preemtion in critical sections of compession engine. - Replace device size determination in ntfs_fill_super() with sb->s_bdev->bd_inode->i_size (in bytes) and remove now superfluous function super.c::get_nr_blocks(). - Implement a mount time option (show_inodes) allowing choice of which types of inode names readdir() returns and modify ntfs_filldir() accordingly. There are several parameters to show_inodes: system: system files win32: long file names (including POSIX file names) [DEFAULT] long: same as win32 dos: short file names only (excluding POSIX file names) short: same as dos posix: same as both win32 and dos all: all file names Note that the options are additive, i.e. specifying: -o show_inodes=system,show_inodes=win32,show_inodes=dos is the same as specifying: -o show_inodes=all Note that the "posix" and "all" options will show all directory names, BUT the link count on each directory inode entry is set to 1, due to Linux not supporting directory hard links. This may well confuse some userspace applications, since the directory names will have the same inode numbers. Thus it is NOT advisable to use the "posix" or "all" options. We provide them only for completeness sake. - Add copies of allocated_size, initialized_size, and compressed_size to the ntfs inode structure and set them up in inode.c::ntfs_read_inode(). These reflect the unnamed data attribute for files and the index allocation attribute for directories. - Add copies of allocated_size and initialized_size to ntfs inode for $BITMAP attribute of large directories and set them up in inode.c::ntfs_read_inode(). - Add copies of allocated_size and initialized_size to ntfs volume for $BITMAP attribute of $MFT and set them up in super.c::load_system_files(). - Parse deprecated ntfs driver options (iocharset, show_sys_files, posix, and utf8) and tell user what the new options to use are. Note we still do support them but they will be removed with kernel 2.7.x. - Change all occurences of integer long long printf formatting to hex as printk() will not support long long integer format if/when the div64 patch goes into the kernel. - Make slab caches have stable names and change the names to what they were intended to be. These changes are required/made possible by the new slab cache name handling which removes the length limitation by requiring the caller of kmem_cache_create() to supply a stable name which is then referenced but not copied. - Rename run_list structure to run_list_element and create a new run_list structure containing a pointer to a run_list_element structure and a read/write semaphore. Adapt all users of run lists to new scheme and take and release the lock as needed. This fixes a nasty race as the run_list changes even when inodes are locked for reading and even when the inode isn't locked at all, so we really needed the serialization. We use a semaphore rather than a spinlock as memory allocations can sleep and doing everything GFP_ATOMIC would be silly. - Cleanup read_inode() removing all code checking for lowest_vcn != 0. This can never happen due to the nature of lookup_attr() and how we support attribute lists. If it did happen it would imply the inode being corrupt. - Check for lowest_vcn != 0 in ntfs_read_inode() and mark the inode as bad if found. - Update to 2.5.6-pre2 changes in struct address_space. - Use parent_ino() when accessing d_parent inode number in dir.c. - Import Sourceforge CVS repository into BitKeeper repository: http://linux-ntfs.bkbits.net/ntfs-tng-2.5 - Update fs/Makefile, fs/Config.help, fs/Config.in, and Documentation/filesystems/ntfs.txt for NTFS TNG. - Create kernel configuration option controlling whether debugging is enabled or not. - Add the required export of end_buffer_io_sync() from the patches directory to the kernel code. - Update inode.c::ntfs_show_options() with show_inodes mount option. - Update errors mount option. tng-0.0.7 - 13/02/2002 - The driver is now feature complete for read-only! - Cleanup mft.c and it's debug/error output in particular. Fix a minor bug in mapping of extent inodes. Update all the comments to fit all the recent code changes. - Modify vcn_to_lcn() to cope with entirely unmapped run lists. - Cleanups in compress.c, mostly comments and folding help. - Implement attrib.c::map_run_list() as a generic helper. - Make compress.c::ntfs_file_read_compressed_block() use map_run_list() thus making code shorter and enabling attribute list support. - Cleanup incorrect use of [su]64 with %L printf format specifier in all source files. Type casts to [unsigned] long long added to correct the mismatches (important for architectures which have long long not being 64 bits). - Merge async io completion handlers for directory indexes and $MFT data into one by setting the index_block_size{_bits} of the ntfs inode for $MFT to the mft_record_size{_bits} of the ntfs_volume. - Cleanup aops.c, update comments. - Make ntfs_file_get_block() use map_run_list() so all files now support attribute lists. - Make ntfs_dir_readpage() almost verbatim copy of block_read_full_page() by using ntfs_file_get_block() with only real difference being the use of our own async io completion handler rather than the default one, thus reducing the amount of code and automatically enabling attribute list support for directory indices. - Fix bug in load_attribute_list() - forgot to call brelse in error code path. - Change parameters to find_attr() and lookup_attr(). We no longer pass in the upcase table and its length. These can be gotten from ctx->ntfs_ino->vol->upcase{_len}. Update all callers. - Cleanups in attrib.c. - Implement merging of run lists, attrib.c::merge_run_lists() and its helpers. (Richard Russon) - Attribute lists part 2, attribute extents and multi part run lists: enable proper support for LCN_RL_NOT_MAPPED and automatic mapping of further run list parts via attrib.c::map_run_list(). - Tiny endianness bug fix in decompress_mapping_pairs(). tng-0.0.6 - Encrypted directories, bug fixes, cleanups, debugging enhancements. - Enable encrypted directories. (Their index root is marked encrypted to indicate that new files in that directory should be created encrypted.) - Fix bug in NInoBmpNonResident() macro. (Cut and paste error.) - Enable $Extend system directory. Most (if not all) extended system files do not have unnamed data attributes so ntfs_read_inode() had to special case them but that is ok, as the special casing recovery happens inside an error code path so there is zero slow down in the normal fast path. The special casing is done by introducing a new function inode.c::ntfs_is_extended_system_file() which checks if any of the hard links in the inode point to $Extend as being their parent directory and if they do we assume this is an extended system file. - Create a sysctl/proc interface to allow {dis,en}abling of debug output when compiled with -DDEBUG. Default is debug messages to be disabled. To enable them, one writes a non-zero value to /proc/sys/fs/ntfs-debug (if /proc is enabled) or uses sysctl(2) to effect the same (if sysctl interface is enabled). Inspired by old ntfs driver. - Add debug_msgs insmod/kernel boot parameter to set whether debug messages are {dis,en}abled. This is useful to enable debug messages during ntfs initialization and is the only way to activate debugging when the sysctl interface is not enabled. - Cleanup debug output in various places. - Remove all dollar signs ($) from the source (except comments) to enable compilation on architectures whose gcc compiler does not support dollar signs in the names of variables/constants. Attribute types now start with AT_ instead of $ and $I30 is now just I30. - Cleanup ntfs_lookup() and add consistency check of sequence numbers. - Load complete run list for $MFT/$BITMAP during mount and cleanup access functions. This means we now cope with $MFT/$BITMAP being spread accross several mft records. - Disable modification of mft_zone_multiplier on remount. We can always reenable this later on if we really want to, but we will need to make sure we readjust the mft_zone size / layout accordingly. tng-0.0.5 - Modernize for 2.5.x and further in line-ing with Al Viro's comments. - Use sb_set_blocksize() instead of set_blocksize() and verify the return value. - Use sb_bread() instead of bread() throughout. - Add index_vcn_size{_bits} to ntfs_inode structure to store the size of a directory index block vcn. Apply resulting simplifications in dir.c everywhere. - Fix a small bug somewhere (but forgot what it was). - Change ntfs_{debug,error,warning} to enable gcc to do type checking on the printf-format parameter list and fix bugs reported by gcc as a result. (Richard Russon) - Move inode allocation strategy to Al's new stuff but maintain the divorce of ntfs_inode from struct inode. To achieve this we have two separate slab caches, one for big ntfs inodes containing a struct inode and pure ntfs inodes and at the same time fix some faulty error code paths in ntfs_read_inode(). - Show mount options in proc (inode.c::ntfs_show_options()). tng-0.0.4 - Big changes, getting in line with Al Viro's comments. - Modified (un)map_mft_record functions to be common for read and write case. To specify which is which, added extra parameter at front of parameter list. Pass either READ or WRITE to this, each has the obvious meaning. - General cleanups to allow for easier folding in vi. - attrib.c::decompress_mapping_pairs() now accepts the old run list argument, and invokes attrib.c::merge_run_lists() to merge the old and the new run lists. - Removed attrib.c::find_first_attr(). - Implemented loading of attribute list and complete run list for $MFT. This means we now cope with $MFT being spread across several mft records. - Adapt to 2.5.2-pre9 and the changed create_empty_buffers() syntax. - Adapt major/minor/kdev_t/[bk]devname stuff to new 2.5.x kernels. - Make ntfs_volume be allocated via kmalloc() instead of using a slab cache. There are too little ntfs_volume structures at any one time to justify a private slab cache. - Fix bogus kmap() use in async io completion. Now use kmap_atomic(). Use KM_BIO_IRQ on advice from IRC/kernel... - Use ntfs_map_page() in map_mft_record() and create ->readpage method for reading $MFT (ntfs_mft_readpage). In the process create dedicated address space operations (ntfs_mft_aops) for $MFT inode mapping. Also removed the now superfluous exports from the kernel core patch. - Fix a bug where kfree() was used insted of ntfs_free(). - Change map_mft_record() to take ntfs_inode as argument instead of vfs inode. Dito for unmap_mft_record(). Adapt all callers. - Add pointer to ntfs_volume to ntfs_inode. - Add mft record number and sequence number to ntfs_inode. Stop using i_ino and i_generation for in-driver purposes. - Implement attrib.c::merge_run_lists(). (Richard Russon) - Remove use of proper inodes by extent inodes. Move i_ino and i_generation to ntfs_inode to do this. Apply simplifications that result and remove iget_no_wait(), etc. - Pass ntfs_inode everywhere in the driver (used to be struct inode). - Add reference counting in ntfs_inode for the ntfs inode itself and for the mapped mft record. - Extend mft record mapping so we can (un)map extent mft records (new functions (un)map_extent_mft_record), and so mappings are reference counted and don't have to happen twice if already mapped - just ref count increases. - Add -o iocharset as alias to -o nls for backwards compatibility. - The latest core patch is now tiny. In fact just a single additional export is necessary over the base kernel. tng-0.0.3 - Cleanups, enhancements, bug fixes. - Work on attrib.c::decompress_mapping_pairs() to detect base extents and setup the run list appropriately using knowledge provided by the sizes in the base attribute record. - Balance the get_/put_attr_search_ctx() calls so we don't leak memory any more. - Introduce ntfs_malloc_nofs() and ntfs_free() to allocate/free a single page or use vmalloc depending on the amount of memory requested. - Cleanup error output. The __FUNCTION__ "(): " is now added automatically. Introduced a new header file debug.h to support this and also moved ntfs_debug() function into it. - Make reading of compressed files more intelligent and especially get rid of the vmalloc_nofs() from readpage(). This now uses per CPU buffers (allocated at first mount with cluster size <= 4kiB and deallocated on last umount with cluster size <= 4kiB), and asynchronous io for the compressed data using a list of buffer heads. Er, we use synchronous io as async io only works on whole pages covered by buffers and not on individual buffer heads... - Bug fix for reading compressed files with sparse compression blocks. tng-0.0.2 - Now handles larger/fragmented/compressed volumes/files/dirs. - Fixed handling of directories when cluster size exceeds index block size. - Hide DOS only name space directory entries from readdir() but allow them in lookup(). This should fix the problem that Linux doesn't support directory hard links, while still allowing access to entries via their short file name. This also has the benefit of mimicking what Windows users are used to, so it is the ideal solution. - Implemented sync_page everywhere so no more hangs in D state when waiting for a page. - Stop using bforget() in favour of brelse(). - Stop locking buffers unnecessarily. - Implemented compressed files (inode->mapping contains uncompressed data, raw compressed data is currently bread() into a vmalloc()ed memory buffer). - Enable compressed directories. (Their index root is marked compressed to indicate that new files in that directory should be created compressed.) - Use vsnprintf rather than vsprintf in the ntfs_error and ntfs_warning functions. (Thanks to Will Dyson for pointing this out.) - Moved the ntfs_inode and ntfs_volume (the former ntfs_inode_info and ntfs_sb_info) out of the common inode and super_block structures and started using the generic_ip and generic_sbp pointers instead. This makes ntfs entirely private with respect to the kernel tree. - Detect compiler version and abort with error message if gcc less than 2.96 is used. - Fix bug in name comparison function in unistr.c. - Implement attribute lists part 1, the infrastructure: search contexts and operations, find_external_attr(), lookup_attr()) and make the code use the infrastructure. - Fix stupid buffer overflow bug that became apparent on larger run list containing attributes. - Fix bugs in readdir() that became apparent on larger directories. The driver is now really useful and survives the test find . -type f -exec md5sum "{}" \; without any error messages on a over 1GiB sized partition with >16k files on it, including compressed files and directories and many files and directories with attribute lists. tng-0.0.1 - The first useful version. - Added ntfs_lookup(). - Added default upcase generation and handling. - Added compile options to be shown on module init. - Many bug fixes that were "hidden" before. - Update to latest kernel. - Added ntfs_readdir(). - Added file operations for mmap(), read(), open() and llseek(). We just use the generic ones. The whole point of going through implementing readpage() methods and where possible get_block() call backs is that this allows us to make use of the generic high level methods provided by the kernel. The driver is now actually useful! Yey. (-: It undoubtedly has got bugs though and it doesn't implement accesssing compressed files yet. Also, accessing files with attribute list attributes is not implemented yet either. But for small or simple file systems it should work and allow you to list directories, use stat on directory entries and the file system, open, read, mmap and llseek around in files. A big mile stone has been reached! tng-0.0.0 - Initial version tag. Initial driver implementation. The driver can mount and umount simple NTFS file systems (i.e. ones without attribute lists in the system files). If the mount fails there might be problems in the error handling code paths, so be warned. Otherwise it seems to be loading the system files nicely and the mft record read mapping/unmapping seems to be working nicely, too. Proof of inode metadata in the page cache and non- resident file unnamed stream data in the page cache concepts is thus complete.