Tmpfs and ramfs

How it works

Linux uses pages and dentries to cache files and directories (respectively) temporarily in memory to speed things up. When the virtual memory subsystem needs this memory for something else, the caches & dentries are flushed to a backing store. In the interest of brevity, this store is a block device, like your hard disk.

When you use either tmpfs or ramfs, there is no backing store. Trippy.

Differences between tmpfs or ramfs

Only two major things, really.

1. tmpfs can use swap while ramfs cannot
2. tmpfs cannot grow dynamically while ramfs can

Using tmpfs or ramfs

tmpfs

A simple tmpfs mount. This will default to half the size of system memory:

mount -t tmpfs tmpfs /tmp

Mount tmpfs, but limit it to 200MB, owned by user joe and group fisherman:

mount -t tmpfs tmpfs /mnt/volatile -o size=200M,uid=12,gid=107

The man page also says that you can tweak with block and inode counts for your mount (nr_blocks and nr_inodes).

ramfs

A simpler gentleman; no mount options whatsoever:

mount -t ramfs ramfs /mnt/volatile

Can you free memory after using it?

 With ramfs, there is no backing store.  Files written into ramfs allocate
 dentries and page cache as usual, but there’s nowhere to write them to.
 This means the pages are never marked clean, so they can’t be freed by the
 VM when it’s looking to recycle memory.

I’m guessing that this is the case with tmpfs as well. You’ll probably need to reboot the system. However, and to test this, I put two 184,320,000-byte files into /dev/shm and played around with them. Here are some numbers:

# Before
MemTotal:      1026888 kB
MemFree:        376068 kB
SwapTotal:     2064376 kB
SwapFree:      1891908 kB

# After copying 2x180MB into 502MB tmpfs mount
MemTotal:      1026888 kB
MemFree:         33672 kB
SwapTotal:     2064376 kB
SwapFree:      1891908 kB

# After removing one 180MB file
MemTotal:      1026888 kB
MemFree:        213364 kB
SwapTotal:     2064376 kB
SwapFree:      1891908 kB

Observe that swap hasn’t changed. I also couldn’t get perfect arithmetic accounting for the free memory, but it seemed close enough. Bottom-line is that I don’t know what to think of this (yet).

/dev/shm: A ready tmpfs solution

Look at the output of df -ah on most Linux boxes. You’ll see the highlighted:

Filesystem            Size  Used Avail Use% Mounted on
/dev/mapper/VolGroup00-LogVol00
                       73G  2.8G   66G   5% /
proc                     0     0     0   -  /proc
sysfs                    0     0     0   -  /sys
devpts                   0     0     0   -  /dev/pts
/dev/sda1              99M   25M   70M  26% /boot
tmpfs                 502M     0  502M   0% /dev/shm
none                     0     0     0   -  /proc/sys/fs/binfmt_misc
sunrpc                   0     0     0   -  /var/lib/nfs/rpc_pipefs
/root/tmpfs           191M  176M   15M  93% /root/tmpfs

/dev/shm (shm = shared memory) is automatically mounted to occupy half your physical memory at most by default. If you’re not happy with this, go ahead and change it:

mount -o remount,size=1G /dev/shm

Other Points

• Don’t ever think that merely mounting tmpfs, ramfs, or
  /dev/shm will actually reserve (or ‘cordon off’) memory.
• In low-memory situations, swap is used as a backing store. This
  means that you’re not going to see a huge performance boost with
  tmpfs. And I think you’d appreciate a live (albeit sluggish)
  system rather than one that’s crashed due to a reckless use of
  ramfs. Swings and roundabouts.