cosmos72/fstransform

############################################################### ###############################################################

DISCLAIMER

THIS DOCUMENT EXPLAINS HOW TO USE

RISKY PROGRAMS AND PROCEDURES

THAT MAY COMPLETELY AND IRREVERSIBLY DELETE

ALL THE DATA ON YOUR DISKS

THE AUTHOR DECLINES ALL RESPONSIBILITIES

FOR ANY DAMAGE THAT MAY DERIVE

FROM USING THE PROGRAMS AND PROCEDURES

DESCRIBED IN THIS DOCUMENT

############################################################### ###############################################################

This document explains how use the programs 'fstransform', 'fsmove', 'fsmount_kernel' and 'fsremap' to transform the contents of a Linux device - usually a disk partition - from a filesystem type to another while preserving its contents.

For example, a disk partition can be transformed from 'jfs' to 'ext4', or from 'ext2' to 'xfs', or many other combinations.

Currently, the programs mentioned above have been tested on Linux with the following filesystems, both as source and as target: ext2, ext3, ext4, jfs, reiserfs, xfs.

Do NOT use these programs with other filesystems unless you are willing to LOSE your data.

In particular, they do NOT (yet) support ntfs, msdos, vfat and exfat file systems.

Common sense and experience tell that you should ALWAYS have a backup of your valuable data: while the programs do NOT need to backup your data to operate, YOU need a backup in case something goes wrong.

The programs have been tested carefully, yet there is ALWAYS a possibility that they will irreversibly delete ALL the data on the device you run them on, even if you use a tested combination of filesystems.

All this foreword means only one thing:

  IF YOU LOSE YOUR DATA, IT IS YOUR PROBLEM.

The author declines ALL responsibilities for ANY damage that may derive from using the programs and procedures described in this document.

############################################################### ###############################################################

Introduction

Enough legalese... now let's get to the interesting part.

The program 'fstransform' does the following:

it takes a device with a filesystem on it (even if almost full) and transforms the device to a different filesystem type, in-place (i.e. without backup) and non-destructively (i.e. it preserves all your data).

It can also change filesystem options such as blocksize - see #36 for details.

It works even if the filesystem is almost full

• several empyrical tests have succeeded even with a 95% full filesystem - and even if it contains very large files, for example if some files are larger than half the device or larger than the available space.

REQUIREMENTS

There are five requirements for fstransform to have a chance to succeed:

1. the device must be unmountable, i.e. umount DEVICE must work. In particular, if some running programs are using the device, you must first close or kill them.

   Transforming the current root directory does not work. For that, you should boot from a different installation (for example a live CD, DVD or USB).

2. the device must have a little free space, typically at least 5%

   WARNING: transforming an almost full device to 'xfs' file-system can be tricky:

   • you need either slightly more free space, typically at least 10%,
   • or you must be VERY quick at suspending fstransform when either the source or the target (or both) file-system is almost full and run 'xfs_fsr' on the source or target device (or both) before resuming fstransform. A future fstransform version may automate this operation.

3. the filesystem on the device must support SPARSE FILES, i.e. files with holes (see for example http://en.wikipedia.org/wiki/Sparse_file for an explanation of what they are) and at least one of the two system calls "ioctl(FS_IOC_FIEMAP)" or "ioctl(FIBMAP)" (see the file Documentation/filesystems/fiemap.txt in any recent Linux kernel for an explanation, or search for the same file on Internet)

   ioctl(FIBMAP) is limited by design to 2G-1 blocks, which typically translates to 8TB - 4kB. To transform file systems equal or larger than 8TB, ioctl(FIEMAP) is required.

4. the initial and final filesystems must be supported by the Linux kernel (i.e. it must be able to mount them) and by the tools 'mkfs' and 'fsck' (i.e. it must be possible to create them and check them for errors).

   Support through FUSE (userspace) drivers is acceptable in theory, but there are currently no tested FUSE filesystems that can be converted reliably with fstransform.

5. the following programs must be available: the three custom-made programs 'fsmove', 'fsmount_kernel' 'fsremap' (distributed with the script) and several common Linux tools: which, expr, id, blockdev, losetup, mount, umount, mkdir, rmdir, rm, mkfifo, dd, sync, fsck, mkfs

KNOWN LIMITATIONS

1. As stated above, at a certain step during the conversion, fstransform needs to unmount the device being transformed. For this reason, running fstransform on the device currently mounted as / (i.e. the root directory) fails. For the same reason, running fstransform on the device currently mounted as /usr, /home or /var or similar heavily-used directories is difficult, because quite often there are programs using those, which prevents them from being unmounted.

2. If the device contains a HUGE number of files with multiple hard links, fstransform will be very slow and consume a LOT of memory. Devices with more than one million files with multiple hard links can cause fstransform to crash with "out of memory" errors.

3. JFS file systems equal or larger than 8TB cannot be converted due to missing support for ioctl(FIEMAP) in the kernel: the fallback ioctl(FIBMAP) is limited by design to < 8TB (assuming 4k blocks)

   Also, ioctl(FIBMAP) must be called for each block so the conversion will be a bit slower.

4. REISERFS file systems using format "3.5" (the default) and equal or larger than 2TB cannot be converted due to their maximum file size = 2TB - 4k: fstransform needs to create a sparse file as large as the device itself.

   REISERFS file systems using format "3.6" are immune to this problem.

5. for the same reason, a device cannot be converted to REISERFS format "3.5" if it contains some files larger than 2TB - 4k.

DETAILS TO KNOW

If the original device is almost full, the program 'fsremap' will create a relatively small backup file ("secondary storage") inside the directory /var/tmp/fstransform. This secondary storage file will be at most as large as half your free RAM.

You can pass the option '-s [k|M|G|T|P|E|Y|Z]' to the tool 'fsremap' in order manually set the secondary storage size, but please understand that using a too small secondary storage can slow down the procedure.

To pass the same option to 'fstransform', you must execute something like fstransform --opts-fsremap='-s '

PROCEDURE

0. compile fsmove, fsmount_kernel and fsremap. Running "./configure" then "make" should suffice on any recent Linux machine, as long as g++ is installed.

   You will get three executables, fsmove and fsremap. They will be located at ./fsmove/build/fsmove ./fsmount_kernel/build/fsmount_kernel ./fsremap/build/fsremap

   You are suggested to either run "make install" or to copy them to a simpler path. Below, they will be referred as {fsmove}, {fsmount_kernel} and {fsremap}

1. OPTIONAL - CAN BE SKIPPED mount read-write the device you want to remap to a new file-system type

   mount {device} {device-mount-point} [your-options]

   if the device is already mounted, check that it is mounted read-write and that no process is using it.

2. decide the target file-system type.

   For some combinations of the initial and final filesystems it is not necessary to use 'fstransform', as the same result can be obtained with much simpler - and SAFER - tools.

   For example, an 'ext2' or 'ext3' filesystem can be transformed into 'ext3' or 'ext4' using the program 'tune2fs'.

   Explaining how to use 'tune2fs' is beyond the scope of this document, just read its man-page or search on the Internet for one of "convert Linux File System ext2 to ext3" "convert Linux File System ext2 to ext3" "convert Linux File System ext3 to ext4"

   But for most combinations, the only way is either to do a full backup + format + restore the data, or use 'fstransform'

3. execute the program

   fstransform {device} {target-file-system-type}

4. follow the instructions - the program will tell you what it is doing, and will also call 'fsmove' and 'fsremap' which show progress percentage and estimated time left.

   Note that 'fsmove' and 'fsremap' need approximately the same time to run, so if 'fsmove' tells you that it will need 2 hours, 'fsmove' will likely need a similar amount of time, for a total of 4 hours.

   In case there are errors, you can even try to fix them instead of aborting the execution (if you know what you are doing).

5. be PATIENT. Transforming a large device takes a LONG time... On a fairly fast disk, it takes about one minute per gigabyte. It means transforming 1000GB takes about 16 hours. Raid disks can be somewhat faster, and solid state disks (SSD) can be much faster.

6. if something goes really wrong, check in /var/tmp/fstransform for the log files fstransform.log. and fsremap.job./fsremap.log they are ABSOLUTELY necessary if you want someone to analyze the problem

   • but unless you are very lucky you can forget about recovering your data...

7. if for some reason the execution is interrupted while 'fsremap' is running, for example due to a power failure, it is possible to resume it by running 'fsremap --resume-job= {device}'. Also, 'fsremap' will show at its startup the exact command line needed to resume its execution.

   The loop file created by fstransform must NEVER be as argument to 'fsremap --resume-job= {...}'. You would IRREVERSIBLY LOSE YOUR DATA!

SOME REAL-WORLD TESTS

1. 1000GB encrypted disk, 52% full, ext2->ext4: SUCCESS, took 12 hours despite one system crash and two manual interruptions (CTRL+C) (Yes, I know ext2->ext4 can be done with tune2s, but I wanted to test fstransform)

2. 1540GB encrypted raid0 (3 disks), 56% full, ext2->ext4: SUCCESS, took 8 hours

Good luck!

Massimiliano Ghilardi