Saturday, May 31, 2008

Random Musings on Embedded Filesystems

Getting the kernel to successfully boot on a new board is a wonderful feeling, which lasts about 5 seconds before the application developers begin clamoring for time on the box. For the apps work to make progress the application binaries have to be accessible in a filesystem somewhere. During platform bringup it is common to use NFS because it will "just work" almost immediately after you get the platform to boot.

I'd advise not relying on NFS for very long: using a server with gigabytes of disk means you're completely insulated from size constraints. You may find that the application footprint grows astronomically simply because nobody notices the pain. Adding library dependencies which pull in 20 Megs of DLLs is invisible on an NFS server, but a crisis when you discover very late in the process that it won't fit in flash.

Lets start by listing a few assumptions about the target:

  1. it does not have a hard disk (i.e. a spinning, magnetic disk)
  2. it has at least as much DRAM as flash, probably more.
  3. it uses a 32 or 64 bit processor with an MMU (PowerPC, MIPS, ARM, x86, etc)

If your platform is significantly different from these assumptions, the recommendations in this article may not be appropriate. The biggest one is the hard disk: if your system has one, don't use these recommendations - your filesystem choices should revolve around what the disk can do for you. If your system does not have an MMU some of these recommendations may apply, but you need to look at using uCLinux and its associated best practices.

A Note about the Buffer Cache

The Linux I/O system is designed for a general purpose computer, with a disk connected to an I/O interface. Recently accessed files are stored in memory in the buffer cache; if the files are accessed again they can be provided from cache instead of hitting the disk a second time. You may also come across references to a page cache; in current kernel versions the page cache and buffer cache have been merged into a single pool of memory.

Embedded systems generally don't have a disk, their filesystems reside in RAM. Nonetheless Linux still relies on the buffer cache. So though a file resides in a RAMdisk it will be read into the buffer cache, and then to the process which requested it. The buffer cache doesn't add as much value in this setup (re-reading the filesystem would be quick), but the buffer cache is essential to support Linux APIs like mmap(). The effect of the buffer cache will be described in a bit more detail later in the article.

Option #1: Squashfs and tmpfs

SquashFS is a compressed, read-only filesystem. A squashfs filesystem is constructed offline, compressing a directory structure into a binary blob which can be loaded into memory and mounted. As blocks are read from the squashfs they are decompressed into the buffer cache; if memory becomes tight the cache can toss pages out and re-read them from the squashfs as needed. Thus the memory footprint is fairly well minimized, only files in active use will be decompressed.

Squashfs is inherently read-only. It is not mounted read-only, such that you could remount it read-write. The squashfs filesystem is assembled at compile time; once created it is immutable. This is generally acceptable for filesystems which shouldn't change anyway, like /bin or /lib, but is clearly not suitable for everything. There are configuration files like /etc/resolv.conf and /etc/timezone which probably need to be occasionally written, and numerous scratch files reside in /tmp. Therefore a second filesystem is needed, a writeable ramdisk.

tmpfs is a ramdisk, and is a good way to provide the writeable portions of the filesystem. Daniel Robbins wrote an excellent article about tmpfs, published at IBM DeveloperWorks a few years ago. Squashfs would be used for the root filesystem mounted on /, while the tmpfs would be mounted on a subdirectory like /rw. /etc, /var and /tmp are most easily provided as soft links to the tmpfs. For example:

# ls  -l /
dr-xr-xr-x    1 root     472 May 16 10:10 bin
dr-xr-xr-x    1 root     836 May 16 10:10 dev
lrwxrwxrwx    1 root       7 May 16 10:10 etc -> /rw/etc
dr-xr-xr-x    1 root     555 May 16 10:10 lib
dr-xr-xr-x   97 root       0 Dec 31  1969 proc
lrwxrwxrwx    1 root       8 May 16 10:10 root -> /rw/root
drwxrwxrwt    9 root     180 May 16 10:12 rw
dr-xr-xr-x    1 root     151 May 16 10:10 sbin
lrwxrwxrwx    1 root       7 May 16 10:10 tmp -> /rw/tmp
dr-xr-xr-x    1 root      23 May 16 10:09 usr
lrwxrwxrwx    1 root       7 May 16 10:10 var -> /rw/var

At boot, the init script would mount the tmpfs filesystem and create the directories which the softlinks point to. You can specify a maximum size when mounting the tmpfs, 8 megabytes in this example:

if mount -t tmpfs -o size=8m tmpfs /rw >/dev/null 2>&1; then
    mkdir /rw/tmp /rw/var /rw/etc /rw/root
    # Halt and catch fire

Squashfs has not been accepted into the mainline kernel, so you will need to download patches from the project website. There is also a LZMA-enhanced version of Squashfs, though I have not personally used this version. LZMA appears to obtain better compression ratios than the zlib which Squashfs normally uses.

For the root directory ("/"), the kernel has to be able to recognize and mount the filesystem without relying on any external mount program. This is a chicken and egg problem: there is no way to run a mount program until there is a filesystem mounted. To use squashfs for the root filesystem, linux/init/do_mounts.c must be modified to recognize its magic number. A patch is available, if your kernel source does not already have this handling. The patch also handles a cramfs root filesystem.

Option #2: cramfs instead of squashfs

cramfs is an older compressed, read-only filesystem for Linux, developed by Linus Torvalds. Cramfs works quite well and is included in the kernel sources. However squashfs typically achieves better compression, because it uses larger blocks in zlib. In one project I worked on a 20 Meg cramfs filesystem became about 18 Megs with squashfs.

Like squashfs, cramfs must be paired with a second, writable ramdisk to be useful.

Option #3: ext2 instead of tmpfs

ext2 can be use to provide a writeable ramdisk, though there are relatively few reasons to use it for this purpose as opposed to tmpfs. Older tmpfs releases did not support certain filesystem features like sendfile(), but this is not an issue for most applications.

Nonetheless if there is a reason to do so, an ext2 filesystem can be created in a /dev/ram# device and mounted as a ramdisk. ext2 has to be formatted before it can be used, which would normally mean bundling mke2fs into your target image. However there is another way to create an ext2 which is generally smaller than mke2fs. Empty, zero filled ext2 filesystems compress extremely well using bzip2. You can create a filesystem of the appropriate size while compiling the target image, by running commands on your build system:

dd if=/dev/zero of=/tmp/e2file bs=1024 count=8192
sudo losetup /dev/loop7 /tmp/e2file
sudo mke2fs -b 4096 -m 1 -i 16384 /dev/loop7
sudo tune2fs -c -1 /dev/loop7
sudo losetup -d /dev/loop7
bzip2 -9 /tmp/e2file

The bs and count arguments to dd specify the size of file to create, filled with zeros. We use a /dev/loop device to create a new ext2 filesystem in this file, and then compress it. The result should be a couple hundred bytes in size, far smaller than mke2fs would be. The e2file.bz2 is copied into the target filesystem, and mounted by the boot scripts like so:

bunzip2 -c /e2file.bz2 >/dev/ram2 2>/dev/null
if mount -t ext2 /dev/ram2 /rw >/dev/null 2>&1; then
    mkdir /rw/tmp /rw/var /rw/etc /rw/root
    # Halt and catch fire

Option #4: JFFS2 instead of tmpfs

In the discussion of the previous alternatives the read-only portion of the filesystem would be compressed in memory using squashfs or cramfs, but the writable portion would be stored uncompressed. If your system requires a large amount of ramdisk space but your memory is constrained, the natural solution is to look for a way to compress it.

If the applications generating this data can be modified to use zlib and compress their own output, that is probably the best way to proceed. If you cannot modify the apps, there are a couple ways to get compression at the filesystem layer. The only natively compressed, writable filesystem for Linux I know of is JFFS2, which is not designed to be a ramdisk but can be pressed into service if necessary using the mtdram.o module (which exists to ease debugging of MTD applications). The vast majority of the complexity in JFFS2, for handling erase blocks and wear leveling and all of the other nuances of flash chips, is wasted when used as a ramdisk, but it does provide compression. The boot scripts would work like so:

/sbin/insmod -q /lib/modules/mtdram.o total_size=8192 erase_size=16
if mount -n -t jffs2 -o umask=0666 /dev/mtdblocka /rw >/dev/null 2>&1; then
    mkdir /rw/tmp /rw/var /rw/etc /rw/root
    # Halt and catch fire

JFFS2 consumes a great deal more CPU time than other filesystems, due to the compression. The throughput to a compressed ramdisk will be relatively low. On one platform I worked on a tmpfs ramdisk could handle writes at about 40 MBytes/sec while the JFFS2 ramdisk managed only 1 MByte/sec.

Note that this technique does not save as much memory as you might think. Whenever blocks are accessed they are decompressed by JFFS2 into the buffer cache, then copied up to the application. Likewise written blocks are held uncompressed in the buffer cache. If your system touches a large amount of data in the ramdisk, the footprint of the buffer cache will become more significant than the backing filesystem in JFFS2. This is another advantage of modifying applications to use zlib: the data remains compressed in the buffer cache, and is only decompressed within the application itself.

There are other ways to implement a compressed, writable filesystem, but I have not used them myself and can't add anything pithy about them. Some links:

  • cloop is a compressed block loopback device, allowing any filesystem to be mounted atop it
  • e2compr patches compression support into ext2/ext3. Development was moribund for a long time, it is not clear to me how stable these patches are.

Option #5: ext2 for everything

In large measure ext2 and the related ext3 are the "default" filesystem for Linux. When building a Linux desktop or server ext[23] is a reasonable solution, and the one generally chosen barring good reason to do otherwise. This notion of ext2 as the default often carries over into embedded systems as well, though there is far less advantage to doing so. It is certainly possible to create a large ext2 ramdisk (using the /dev/loop technique shown above) and use it to store both the application binaries as well as provide scratch space for /tmp et al. This does have the appeal of requiring only one filesystem, rather than the combinations recommended earlier.

The memory footprint of an ext2 ramdisk is always going to be larger than the options described above. It is common practice to create the ext2 filesystem in a file, gzip it, and compile the gzipped binary into the target image. This reduces the image size, but at boot the kernel will decompress the entire filesystem into memory. So if you have a 40 Meg ext2 which gzips down to 10 Megs, it will only add 10 Megs to the image size but expand to 40 Megs of RAM on the target. Compare this to squashfs, where a 10 Meg filesystem adds 10 Megs to the image size and also consumes 10 Megs of RAM on the target. The buffer cache does not perturb the result: when using either ext2 or squashfs, any files in active use will be present in the buffer cache and the footprint would be the same in both cases.


This turned into an exceedingly long post. Early drafts were even longer, and much material got left on the cutting room floor. I'll collect more random musings on embedded filesystems into a future post.


A comment from Lance pointed to a discussion on LKML about extending cramfs to be writable. Changes made to files would be stored in the page cache, and as far as I can tell would be stored in uncompressed form. The same technique could presumably be applied to squashfs. This would make things a bit simpler for those who need a small amount of writable ramdisk.

Tom Ziomek notes that SquashFS has been accepted into the mainline kernel in 2.6.29. Scroll down to see Tom's comment; since this article appeared this blog switched to using Disqus, so you see the older comments first followed by the recent comments in a separate Disqus section.