Ttyjfyk

I'm looking for a way to have a tmpfs-like file system that can be unlimited in size, but will use a specified amount of RAM after which the "oversize" data will be stored on another disk-backed filesystem

I'm running on a SSD-only system, with low available space (usually < 3 GB), so I don't want to reserve any space for SWAP or similar (that's my main requirement)

Do you know of any solution that would fit my use-case ?

You can use dmsetup to set a linear device made of a /dev/ramX plus your /dev/sdX disk-device.

For instance, from the command prompt type:

dmsetup create mydev << EOF
> 0 65536 linear /dev/ram0 0
> 65536 262144 linear /dev/sdb 0
> EOF

The above creates a logical device named mydev made of a 32MB (65536 sectors) ramdisk plus the first 128MB (262144 sectors) of /dev/sdb. Of course replace /dev/sdb with the actual device file for your disk.

You may also want to use the exact size (or a bit less) of your disk or partition, and for that you need to know that size expressed in 512-bytes sectors. In order to know that for e.g. the /dev/sdb1 device you may do:

grep sdb1 /proc/partitions

and then use the big number that appears on the left of the device name, multiplied by 2. (because that big number is expressed in 1024-bytes blocks).

Then you have to format mydev with your file-system type of choice. For instance:

mkfs.ext4 /dev/mapper/mydev

Finally mount it with:

mount /dev/mapper/mydev /mnt

Replace /mnt with the actual directory you want to use as mount point, and enjoy.

To remove everything, do:

umount /mnt
dmsetup remove mydev

However it’s worth saying that this setup is less efficient than direct tmpfs filesystems, because tmpfs is already a (virtual) file-system on top of Linux's virtual memory manager, whereas ramdisks need an additional operation of memory-pages copying, plus they still require a regular file-system with its overhead on top of them.