The BetrFS prototype (*) fits into the Linux storage stack as follows:
________________
| |
| VFS |
|______________|
________________
| |
| BetrFS * |
|______________|
________________
| |
| B^e Tree * |
|______________|
________________
| |
| ext4 |
|______________|
Like any other file system, BetrFS is registered with the VFS as a file system during module load. But BetrFS has a stacked file system design. When you mount BetrFS, it loads a B^e-tree index from a separate kernel file system (the "southbound" file system, here ext4). For this reason, you must specify (at BetrFS mount time) the device and file system type where the B^e-tree index image resides. This will be explained in detail in the 'Mounting BetrFS' section below.
BetrFS code is contained in the filesystem/ directory. Its primarily role is to implement the BetrFS schema and convert VFS operations into B^e-tree operations. The BetrFS kernel module code can be found in filesystem/ftfs_module.c, and the primary files that deal with the VFS->B^e-tree mapping are filesystem/ftfs_super.c and filesystem/ftfs_bstore.c
The B^e-tree implementation in BetrFS is from the open-source fractal tree index provided by TokuTek (TokuDB). There have been slight modifications to TokuDB in order for the kernel port to be successful, but the TokuDB code has remained largely unchanged. The ft/, src/, portability/, util(s)/, include/, locktree/, and cmake_modules/ directories contain most of the TokuDB code and configuration files.
To port the B^e-tree to the kernel, we reimplemented the userspace libraries used by TokuDB that were not compatible with the kernel. These can be found in the filesystem/ directory ftfs_pthreads.* ftfs_files.* ftfs_stat.* ftfs_stdio.* ftfs_error.* ftfs_assert.* ftfs_compress.* ftfs_random.* toku_linkage.c ... We import TokuDB as a binary blob, and overwrite TokuDB symbols using symbols from these files.
The linux-3.11.10 directory contains the kernel required to run BetrFS. A modified kernel is required for a few reasons:
- At the time of writing BetrFS, the kernel did not support direct IO from non-userspace buffers.
- TokuDB relies on errno for error handling. Instead of modifying all of the TokuDB code to explicitly pass error numbers, we augmented the Linux 'struct task_struct' with an error number field. This field is unused by code outside of BetrFS.
- To interface with ext4, we maintain a private file table and file system namespace for TokuDB files. We had to export a few functions that were already in the kernel, but unavailable for our uses.
The ftfs/ directory contains a simple module that can be used to run all of the TokuDB regression tests within the kernel. It is not necessary to run BetrFS, but can be useful for testing enhancements to the data structures.
First you must compile and install the provided Linux kernel. There are many guides on how to do this, so please read one if you have never compiled your own kernel. An abbreviated version; cd linux-3.11.10/ make oldconfig make make modules make modules_install make install
The next step is to build TokuDB. TokuDB uses cmake, and it is very finicky. You must have the right versions of gcc, and g++: gcc-4.7, g++-4.7.
For those unfamiliar with cmake, I would suggest an out-of-source
build. The following commands should work:
mkdir build
cd build
CC=gcc-4.7 CXX=g++-4.7 cmake
-D CMAKE_BUILD_TYPE=$TYPE
-D USE_BDB=OFF
-D USE_TDB=ON
-D BUILD_TESTING=OFF
-D CMAKE_INSTALL_PREFIX=../ft-install/
-D BUILD_FOR_LINUX_KERNEL_MODULE=ON
..
cmake --build . --target install
After building TokuDB, you can finally build the actual BetrFS code. cd filesystem/ make
You need a couple of things:
-
A device formatted with an existing file system to use as your "southbound" file system (we have been using ext4, but there is no reason it can't be a different file system).
-
The southbound file system must also be set up with some files and directories that TokuDB expects at certain places, including: db/ /dev/null /tmp
-
The compiled module from the filesystem/ folder
-
A "dummy" device to pass to the mount command (can be an empty file set up as a loop device). This is neither read from or written to; it is there solely to pass to the mount command.
-
zlib (on our system, apt-get install zlib1g-dev)
This example code was used to set up the file system on a setup where we had a second disk with a partition for our "southbound" file system at /dev/sdb1. Change the parameters to fit your needs.
USER=betrfs REPO=/home/$USER/ft-index MODDIR=filesystem MODULE=ftfs.ko MOUNTPOINT=mnt SBDISK=/dev/sdb1
sudo mkfs.ext4 $SBDISK mkdir -p $MOUNTPOINT mount -t ext4 $SBDISK $MOUNTPOINT cd $MOUNTPOINT; rm -rf *; mkdir db; mkdir dev; touch dev/null; mkdir tmp; chmod 1777 tmp; cd -; umount $MOUNTPOINT
cd $REPO/$MODDIR; make; cd -; sudo modprobe zlib sudo insmod $REPO/$MODDIR/$MODULE sb_dev=$SBDISK sb_fstype=ext4
touch dummy.dev sudo losetup /dev/loop0 dummy.dev sudo mount -t ftfs /dev/loop0 $MOUNTPOINT