The libfuse github has some good examples for C/C++ in [0] of increasing complexity:
- passthrough.c mirrors existing filesystem, "Its performance is terrible."
- passthrough_fh.c "performance is not quite as bad."
- passthrough_ll.c implemented with low level api and "the least bad among the three"
- passthrough_hp.cc high performance version written in C++
Some interesting fuse projects in my notes: [1] splitting large files into segments; [2] show ZFS incremental snapshots as files; [3] transparent filesystem compression; [4] and [5] options for mounting archives as filesytems.
While [2] might be a good code example, this functionality is already built into ZFS. At the mountpoint of every dataset is a hidden ".zfs" folder that doesn't show up, even on a `ls -A`. You just have to believe its there and cd into it. Under that is a "snapshot" folder, and inside that is a folder for each snapshot of that dataset. Those folders contain the files in the snapshot.
So for example, /etc/hosts from my snapshot zrepl_20241011_010143_000 would be at /.zfs/snapshot/zrepl_20241011_010143_000/etc/hosts
If you don't like the magic hidden nature of it, you can even configure it to behave like a normal folder with `zfs set snapdir=visible <dataset>`
The GP wasn’t suggesting [2] was using the ZFS snapshot folder. It was saying that ZFS natively supports a file system via of snapshots so people should be mindful to use [2] as an example of writing FUSE file systems rather than a practical tool for working with ZFS.
Just wanted to throw out there that although I'm a fan of FUSE, it's not the only option. I've had fun implementing a virtual filesystem via the 9p protocol not too long ago.
IIRC, I used py9p and the python experience was much nicer than fuse-python. You can mount a 9p service via FUSE (9pfuse) if you want. I just used the kernel v9fs client. If you're just looking to pass a filesystem through the network, I think I used the diod 9p server.
9p is such a great little protocol. diod[0] has a good amount of documentation on the protocol itself, but it's pretty simple.
I have some notes here [1], but it's mostly just linking to primary sources. FUSE is great, but 9P is more general and has high quality implementations all over the place, even in Windows!
One thing I'm not so sure about is the performance properties of 9p. I've seen some places indicate it's rather slow, but nothing definitive. Does anyone have any benchmarks or info on that?
I looked at 9p as a NFS/SMB replacement some time ago for a project, and benchmarks seems scarce. I did find this[1] set of benchmarks but it's from 2005.
I also found this[2] from 2016 which points out some performance barriers, and it doesn't sound like an easy fix without some new extensions.
However there's also this[3] post from 2022 about some patches suggesting a 10x improvement in Linux 9p performance.
Would be interesting to see how things are these days.
I ended up not pursuing it, but I feel like it would be interesting to give it a whirl just to compare, given hardware and code has evolved as mentioned.
> 9P [...] has high quality implementation[...] in Windows
Do you know if it’s possible to mount one’s own 9P servers under Windows? I seem to remember a comment from a Microsoft employee on GitHub something-or-other that said that capability is private to WSL2, but I can’t find it right now.
If you have to go through WSL to mount it, does that really count has a "high quality implementation" in Windows? Windows already has a high performance FUSE alternative called ProjFS.
I was experimenting about 18 months ago with FUSE in front of an HTTPS URL, essentially a large file I wanted to be able to random read as if it was local, without downloading it first.
One of the things I ran in to that made it painful, was that the block sizes for FUSE were really small, it made for a lot of latency and churn of HTTP calls to the back end that ended up needing some fairly complicated caching/pre-reading logic to handle. Kernel read-ahead logic never seemed to kick in (and I didn't do any investigation in to that at the time, other than not finding any particular way to induce it)
When I played with 9P, this version of py9p was what worked to share a directory so a NetBSD client could mount it.
It worked natively with https://man.netbsd.org/mount_9p.8 without an external client.
The important difference from diod was that py9p could speak 9P2000.u, which NetBSD understood, while diod only spoke 9P2000.L.
I wish I had known about this a month ago, when I had to go through the exact same process!
In a desperate attempt to find a less frustrating way to interact with Jira, I had the silly idea of starting a jira-as-filesystem project that uses our internal issue categorization to build a tree: directories represent issues, with files representing issue fields and subdirectories for linked issues. I ended up choosing fuse-python.
I haven't worked on it in a minute, but I was already bumping into issues (pun not intended) with the abstraction: using just the issue ID as directory name makes automation easier, but it makes it hard for humans to browse the tree, since a `ls` would just show you a bunch of inscrutable IDs. I ended up adding a parallel `<issue-type>-with-summary` directory type where the slugified summary is appended to each issue ID.
Hmm, I'm not saying it's a good idea, but what about a daemon that keeps a symlinked version of the entire jira environment up to date? So you have one jira-as-filesystem that's the raw files, but then for human consumption/interaction, you have a tree of symlinks, including multiple links to the same file wherever it's relevant. Might be adding more layers than needed, based on my lack of understanding, but might technically solve the (current/stated) abstraction issue.
That's sort of what I'm doing behind the scenes, because I keep one global list of downloaded issues (they're lazily loaded when you access them) and then the folders are really only "views" into the downloaded issues. Representing identical ones across trees as symlinks is a fantastic idea though, I can't believe I didn't think of that! Thanks for the inspiration.
Would you even need a daemon for that? That sounds as if the FS could just generate the symlinks on-the-fly in the same way that it generates the folders.
(Unless symlinks are somehow special - but at least both /dev and /proc also provide symlinks and to my knowledge they don't have any actual storage behind them, so it should be possible, I think)
Referencing the same two ways is normal in Unix fs. On a modern Linux you will see disks referenced by block device and UUID. I think your approach is good and consistent with expectations.
Though I, personally, would not use it as JIRA is complicated enough for me.
So many fuse mount options out there with varying tradeoffs, performance, and features (s3fs, goofys, seaweed, minio, Google drive, etc..). JuiceFS is pretty interesting for doing things like mounting an object store and accessing it via posix with all the metadata you would expect on a traditional filesystem.
https://juicefs.com
Adjacent question: lately I’ve been seeing people implement NFS base filesystems since that is a more widely supported protocol. I think rclone does this for Mac. Is there a guide, or even a comparison, for this approach?
Fun fact: Now that recent macOS versions require you to disable security features to install macFUSE, there’s the awesome fuse-t. It works as a drop-in replacement, doesn’t need the kext and will open up an NFS server in the background and mount that using macOS features. Performance is pretty good, too.
If you're interested in seeing what a finished product looks like, check out azuline/rosé — a music manager with a virtual filesystem. Really good codebase with a lot of comments and explanations and types and tests, which should make it easy to learn from.
People interested in FUSE might also be interested in the CUSE companion (sub)project.
CUSE is userspace character device driver emulation. It allows you to emulate hardware without compiling a new kernel module. I just used it recently to write a hardware device supporting IOCTLs using Python. However I didn't find any good Python libraries that worked easily and documentation was lacking, but I found it easy enough that I ended up writing it using just the ctypes ffi library. The only part that wasn't immediately intuitive for me, as someone who has never written kernel drivers, is how IOCTL's require the buffers to be resized for each read and write which means the calls come in pairs, but luckily CUSE has a debug mode which shows all bytes in and out. CUSE was originally implemented to create userspace sound devices¹ but has also been use for things like custom TTYs. I used it for creating a virtual SPI device. Hopefully someone finds this useful and this project can get more attention.
Quite some years ago I created a Python FUSE filesystem[1] to to interact with dokuwiki (a wiki system).
It's built on hde llfuse[2]. But that required implementing a bunch of low level APIs that were not really related to dokuwiki. So I created easyfuse[3][4] as a wrapper, which implemented the things that were unrelated the dokuwiki implementation. If you're interested it in building a FUSE system it might be worth looking at.
Can you tell what's the usecase for creating FUSE for dokuwiki? Basically, dokuwiki is just a bunch of text files so wouldn't it be simpler and more efficient to e.g. mount them as NFS or share via Dropbox/Syncthing?
I was forced to use the dokuwiki, but I very much disliked editing stuff in the web interface. Having a filesystem interface to the wiki system allowed me to create and edit pages using vim , which I like to use for writing.
I like to think of fuse as a way of allowing Makefiles to specify DAGs over arbitrary resources. For instance, a fuse fs exposing the state of a k8s cluster might ease writing operators accessible for simpler minds like mine.
Or email, why not expose imap through a file system, so your RAG app (like gpt4all) can just access everything directly ?
Off topic but whenever I see a blog with some 90s/2000s vibes, I always go to their first page of posts. Never disappoints to sneak a peak into that time capsule - including gwolf.org!
I've recently discovered sshfs and learned about needing to have FUSE as a dependency for OS X, which spiked by interest. The code looks very clean and easy to understand, so thanks for that!
Is there any guide/course you would recommend for the introduction to FUSE? It looks like all you have is to provide implementations to certains functions your filesystem will use but it's hard without knowing the details(ex. I wouldn't know I had to implement readdir without your code, and so on)
I've used sshfs in the past and I know the original authors stopped maintaining it though others took over. I did find the network error handling wasn't the greatest. Like it would unmount the fuse mount due to network error and I'd be writing files to the local mount directory silently until space filled up. Perhaps its a Linux specific issue or I've used the wrong options though.
To be honest, I knew about the speed limitations of sshfs already, so I typically use rsync to work with large files. This way, I wouldn't write the data locally even if the connection fails. I've checked the github repo and it looks like there are a number of issues related to network timeout that hasn't been addressed for a long time[1]. However, I mostly used it on OS X, so my experience might be different from yours
Thanks for the info as well, I was under the impression sshfs was under active development (:
[1]https://github.com/libfuse/sshfs/issues/77
I did something similar and it was a really fun project! You can easily make a Google Drive FUSE fs, or something simple like an in-memory fs, an encrypted fs, etc... Its very interesting and a lot simpler than one would imagine. You basically fulfill an interface and FUSE isn't really aware of the implementation. Its more of a "contract" that X function returns a given result. You can implement a FUSE fs for a ton of cool stuff.
I wrote a little project to expose a bar git repository in fuse. Basically, you just have your .git folder, and fuse exposes every single commit (and branch etc) as its own folder at the same time. Without actually having to checkout everything as regular files in a regular filesystem.
It's quite nice, and really shows that git internally is already half a file system. It's also quite simple, because everything is read-only.
I attended a talk at Devoxx yesterday which showcased how new Java language features can be used to implement a Fuse filesystem. Basically as simple as generating Java bindings for Fuse using jextract on fuse.h, and then implementing a couple of method calls using the new Foreign Function & Memory API, which is set to replace JNI.
I will link to the video when it comes online, should be later today.
Imagine you want to glue together something to use with a syncing tool. ~100 lines of python can turn data from one format into a presented set of files. The application doesn't have to be a general purpose filesystem.
A operating systems class I was once a lab assistant for, we implemented the key parts of an operating system in Python. Scheduler, Filesystem, Memory Management. I think it ended up being more confusing than not, but I appreciated where it could go.
You say they embark on a language whose future is uncertain, while being the most popular language in the world, and advise them to learn Scheme instead? Lmao.
If students' first and best-known programming language leaving college is relatively small or niche in a way that makes it unlikely to find it practical to unconditionally use it everywhere, that might actually be good for the software engineering discipline as a whole.
One healthy thing a move like that might encourage is turning which language to use into a more deliberate choice based on the virtues of the language itself with respect to a task. With Python it's kinda too easy to just always reach for the familiar whether it's actually a good fit or not.
You are citing the Tiobe index for the "most popular language"? The Tiobe index just counts the searches that include "Python". Which means that beginners who search for "leftpad python" or people who search for "python sucks" are included.
Google is reducing its Python investment. Python is just the most marketed language.
Students learning Scheme learn good habits, and Scheme will not go away.
- passthrough.c mirrors existing filesystem, "Its performance is terrible."
- passthrough_fh.c "performance is not quite as bad."
- passthrough_ll.c implemented with low level api and "the least bad among the three"
- passthrough_hp.cc high performance version written in C++
Some interesting fuse projects in my notes: [1] splitting large files into segments; [2] show ZFS incremental snapshots as files; [3] transparent filesystem compression; [4] and [5] options for mounting archives as filesytems.
- [0] https://github.com/libfuse/libfuse/tree/master/example
- [1] https://github.com/seiferma/splitviewfuse
- [2] https://github.com/UNFmontreal/zfs_fuse_snapshot
- [3] https://github.com/FS-make-simple/fusecompress
- [4] https://github.com/google/fuse-archive
- [5] https://github.com/mxmlnkn/ratarmount