Advanced filesystem implementor's guide, Part 3
Using the virtual memory (VM) filesystem and bind mounts
In my previous articles in this series, I introduced the benefits of journalling and the ReiserFS and showed how to set up a rock-solid Linux 2.4-based ReiserFS system. In this article, we're going to tackle a couple of semi-offbeat topics. First, we'll take a look at tmpfs, also known as the virtual memory (VM) filesystem. Tmpfs is probably the best RAM disk-like system available for Linux right now, and is a new feature of kernel 2.4. Then, we'll take a look at another new 2.4 feature called "bind mounts", which allow a great deal of flexibility when it comes to mounting (and remounting) filesystems. Then, in the next article, we'll focus on devfs, and after that we'll spend some time getting intimately familiar with the new ext3 filesystem.
If I had to explain tmpfs in one breath, I'd say that tmpfs is like a
ramdisk, but different. Like a ramdisk, tmpfs can use your RAM, but
it can also use your swap devices for storage. And while a traditional
ramdisk is a block device and requires a
command of some kind before you can actually use it, tmpfs is a
filesystem, not a block device; you just mount it, and it's there. All in
all, this makes tmpfs the niftiest RAM-based filesystem I've had the
opportunity to meet.
tmpfs and VM
Let's take a look at some of tmpfs's more interesting properties. As I mentioned above, tmpfs can use both RAM and swap. This might seem a bit arbitrary at first, but remember that tmpfs is also known as the "virtual memory filesystem". And, as you probably know, the Linux kernel's virtual memory resources come from both your RAM and swap devices. The VM subsystem in the kernel allocates these resources to other parts of the system and takes care of managing these resources behind-the-scenes, often transparently moving RAM pages to swap and vice-versa.
The tmpfs filesystem requests pages from the VM subsystem to store files. tmpfs itself doesn't know whether these pages are on swap or in RAM; it's the VM subsystem's job to make those kinds of decisions. All the tmpfs filesystem knows is that it is using some form of virtual memory.
Not a block device
Here's another interesting property of the tmpfs filesystem. Unlike most "normal" filesystems, like ext3, ext2, XFS, JFS, ReiserFS and friends, tmpfs does not exist on top of an underlying block device. Because tmpfs sits on top of VM directly, you can create a tmpfs filesystem with a simple mount command:
# mount tmpfs /mnt/tmpfs -t tmpfs
After executing this command, you'll have a new tmpfs filesystem mounted
at /mnt/tmpfs, ready for use. Note that there's no need to run
mkfs.tmpfs; in fact, it's impossible, as no
such command exists. Immediately after the
mount command, the filesystem is mounted and
available for use, and is of type
is very different from how Linux ramdisks are used; standard Linux
ramdisks are block devices, so they must be formatted with a
filesystem of your choice before you can use them. In contrast, tmpfs
is a filesystem. So, you can just mount it and go.
Dynamic filesystem size
You're probably wondering about how big that tmpfs filesystem was that we mounted at /mnt/tmpfs, above. The answer to that question is a bit unexpected, especially when compared to disk-based filesystems. /mnt/tmpfs will initially have a very small capacity, but as files are copied and created, the tmpfs filesystem driver will allocate more VM and will dynamically increase the filesystem capacity as needed. And, as files are removed from /mnt/tmpfs, the tmpfs filesystem driver will dynamically shrink the size of the filesystem and free VM resources, and by doing so return VM into circulation so that it can be used by other parts of the system as needed. Since VM is a precious resource, you don't want anything hogging more VM than it actually needs, and the great thing about tmpfs is that this all happens automatically. See Related topics.
The other major benefit of tmpfs is its blazing speed. Because a typical tmpfs filesystem will reside completely in RAM, reads and writes can be almost instantaneous. Even if some swap is used, performance is still excellent and those parts of the tmpfs filesystem will be moved to RAM as more free VM resources become available. Having the VM subsystem automatically move parts of the tmpfs filesystem to swap can actually be good for performance, since by doing so, the VM subsystem can free up RAM for processes that need it. This, along with its dynamic resizing abilities, allow for much better overall OS performance and flexibility than the alternative of using a traditional RAM disk.
While this may not seem like a positive, tmpfs data is not preserved between reboots, because virtual memory is volatile in nature. I guess you probably figured that tmpfs was called "tmpfs" for a reason, didn't you? However, this can actually be a good thing. It makes tmpfs an excellent filesystem for holding data that you don't need to keep, such as temporary files (those found in /tmp) and parts of the /var filesystem tree.
To use tmpfs, all you need is a 2.4 series kernel with "Virtual memory file system support (former shm fs)" enabled; this option lives under the "File systems" section of the kernel configuration options. Once you have a tmpfs-enabled kernel, you can go ahead and mount tmpfs filesystems. In fact, it's a good idea to enable tmpfs in all your 2.4 kernels, whether you plan to use tmpfs or not. This is because you need to have kernel tmpfs support in order to use POSIX shared memory. System V shared memory will work without tmpfs in your kernel, however. Note that you do not need a tmpfs filesystem to be mounted for POSIX shared memory to work; you simply need the support in your kernel. POSIX shared memory isn't used too much right now, but this situation will likely change as time goes on.
Avoiding low VM conditions
The fact that tmpfs dynamically grows and shrinks as needed makes one wonder: what happens when your tmpfs filesystem grows to the point where it exhausts all of your virtual memory, and you have no RAM or swap left? Well, generally, this kind of situation is a bit ugly. With kernel 2.4.4, the kernel would immediately lock up. With kernel 2.4.6, the VM subsystem has in many ways been fixed, and while exhausting VM isn't exactly a wonderful experience, things don't blow up completely, either. When 2.4.6 gets to the point where it can't allocate any more VM, you obviously won't be unable to write any new data to your tmpfs filesystem. In addition, it's likely that some other things will happen. First, the other processes on the system will be unable to allocate much more memory; generally, this means that the system will most likely become extremely sluggish and almost unresponsive. Thus, it may be tricky or unusually time-consuming for the superuser to take the necessary steps to alleviate this low-VM condition.
In addition, the kernel has a built-in last-ditch system for freeing memory when no more is available; it'll find a process that's hogging VM resources and kill it. Unfortunately, this "kill a process" solution generally backfires when tmpfs growth is to blame for VM exhaustion. Here's the reason. Tmpfs itself can't (and shouldn't) be killed, since it is part of the kernel and not a user process, and there's no easy way for the kernel to find out which process is filling up the tmpfs filesystem. So, the kernel mistakenly attacks the biggest VM-hog of a process it can find, which is generally your X server if you happen to be running one. So, your X server dies, and the root cause of the low-VM condition (tmpfs) isn't addressed. Ick.
Low VM: the solution
Fortunately, tmpfs allows you to specify a maximum upper bound for the
filesystem size when a filesystem is mounted or remounted. Actually, as of
kernel 2.4.6 and util-linux-2.11g, these parameters can only be set on
mount, not on remount, but we can expect them to be settable on
remount sometime in the near future. The optimal maximum tmpfs size
setting depends on the resources and usage pattern of your particular
Linux box; the idea is to prevent a completely full tmpfs filesystem from
exhausting all virtual memory and thus causing the ugly low-VM conditions
that we talked about earlier. A good way to find a good tmpfs upper-bound
is to use
top to monitor your system's swap
usage during peak usage periods. Then, make sure that you specify a tmpfs
upper-bound that's slightly less than the sum of all free swap and free
RAM during these peak usage times.
Creating a tmpfs filesystem with a maximum size is easy. To create a new tmpfs filesystem with a maximum filesystem size of 32 MB, type:
This time, instead of mounting our new tmpfs filesystem at /mnt/tmpfs, we created it at /dev/shm, which is a directory that happens to be the "official" mountpoint for a tmpfs filesystem. If you happen to be using devfs, you'll find that this directory has already been created for you.
Also, if we want to limit the filesystem size to 512 KB or 1 GB, we can
size=1g, respectively. In addition to limiting
size, we can also limit the number of inodes (filesystem objects) by
nr_inodes=x parameter. When
can be a simple integer, and can also be followed with a
g to specify thousands, millions, or billions
(!) of inodes.
Also, if you'd like to add the equivalent of the above
mount tmpfs command to your /etc/fstab, it'd
look like this:
tmpfs /dev/shm tmpfs size=32m 0 0
Mounting on top of existing mountpoints
Back in the 2.2 days, any attempt to mount something to a mountpoint where something had already been mounted resulted in an error. However, thanks to a rewrite of the kernel mounting code, using mountpoints multiple times is not a problem. Here's an example scenario: let's say that we have an existing filesystem mounted at /tmp. However, we decide that we'd like to start using tmpfs for /tmp storage. In the old days, your only option would be to unmount /tmp and remount your new tmpfs /tmp filesystem in its place, as follows:
# umount /tmp # mount tmpfs /tmp -t tmpfs -o size=64m
However, this solution may not work for you. Maybe there are a number of running processes that have open files in /tmp; if so, when trying to unmount /tmp, you'd get the following error:
umount: /tmp: device is busy
However, with recent 2.4 kernels, you can mount your new /tmp filesystem without getting the "device is busy" error:
# mount tmpfs /tmp -t tmpfs -o size=64m
With a single command, your new tmpfs /tmp filesystem is mounted at /tmp,
on top of the already-mounted partition, which can no longer be
directly accessed. However, while you can't get to the original /tmp, any
processes that still have open files on this original filesystem can
continue to access them. And, if you
your tmpfs-based /tmp, your original mounted /tmp filesystem will
reappear. In fact, you can mount any number of filesystems to the same
mountpoint, and the mountpoint will act like a stack; unmount the current
filesystem, and the last-most-recently mounted filesystem will reappear
Using bind mounts, we can mount all, or even part of an already-mounted filesystem to another location, and have the filesystem accessible from both mountpoints at the same time! For example, you can use bind mounts to mount your existing root filesystem to /home/drobbins/nifty, as follows:
# mount --bind / /home/drobbins/nifty
Now, if you look inside /home/drobbins/nifty, you'll see your root filesystem (/home/drobbins/nifty/etc, /home/drobbins/nifty/opt, etc.). And if you modify a file on your root filesystem, you'll see the modifications in /home/drobbins/nifty as well. This is because they are one and the same filesystem; the kernel is simply mapping the filesystem to two different mountpoints for us. Note that when you mount a filesystem somewhere else, any filesystems that were mounted to mountpoints inside the bind-mounted filesystem will not be moved along. In other words, if you have /usr on a separate filesystem, the bind mount we performed above will leave /home/drobbins/nifty/usr empty. You'll need an additional bind mount command to allow you to browse the contents of /usr at /home/drobbins/nifty/usr:
# mount --bind /usr /home/drobbins/nifty/usr
Bind mounting parts of filesystems
Bind mounting makes even more neat things possible. Let's say that you have a tmpfs filesystem mounted at /dev/shm, its traditional location, and you decide that you'd like to start using tmpfs for /tmp, which currently lives on your root filesystem. Rather than mounting a new tmpfs filesystem to /tmp (which is possible), you may decide that you'd like the new /tmp to share the currently mounted /dev/shm filesystem. However, while you could bind mount /dev/shm to /tmp and be done with it, your /dev/shm contains some directories that you don't want to appear in /tmp. So, what do you do? How about this:
# mkdir /dev/shm/tmp # chmod 1777 /dev/shm/tmp # mount --bind /dev/shm/tmp /tmp
In this example, we first create a /dev/shm/tmp directory and then give it
1777 perms, the proper permissions for /tmp.
Now that our directory is ready, we can mount /dev/shm/tmp, and only
/dev/shm/tmp to /tmp. So, while /tmp/foo would map to /dev/shm/tmp/foo,
there's no way for you to access the /dev/shm/bar file from /tmp.
As you can see, bind mounts are extremely powerful and make it easy to make modifications to your filesystem layout without any fuss. Next article, we'll check out devfs; for now, you may want to check out the following resources.
- Read Daniel's previous articles in this series, where he introduced the benefits of journalling and the ReiserFS and showed how to set up a rock-solid Linux 2.4-based ReiserFS system.
(freshmeat link) is a collection of various important Linux utilities,
umount. You may want to upgrade to the latest available version so that you can use the
mount --bindsyntax (rather than using
mount -o bind).
- Since tmpfs and bind mounts are relatively new and for the most part undocumented new kernel features, the best way to learn more about them is to study relevant parts of the Linux kernel sources.
- The ReiserFS mailing list is an excellent source for current, more in-depth ReiserFS information. Be sure to check out the ReiserFS mailing list archive, too.
- You can find a very nice in-depth look at the meta-data differences between UFS, ext2, ReiserFS, and more in Juan I. Santos Florido's Linux Gazette Journal File Systems review.
- Jedi's ReiserFS/Qmail tuning page contains a lot of good information for qmail users. Also be sure to check out ReiserSMTP, Jedi's collection of drop-in qmail components that offer dramatically higher qmail performance.
- Read Steve Best's JFS Overview on developerWorks.