5 ms·

Writing a 9P server from scratch (2016)

99 points by miduil 9y ago

pmoriarty 9y ago
9P2000 is available on Linux with the CONFIG_9P_FS and related kernel compile options, which say things like: "If you say Y here, you will get experimental support for Plan 9 resource sharing via the 9P2000 protocol. See http://v9fs.sf.net http://v9fs.sf.net for more information."
- voltagex_ 9y ago
  That's been experimental for a long time - https://github.com/torvalds/linux/tree/master/fs/9p https://github.com/torvalds/linux/tree/master/fs/9p I guess file systems don't go into staging/?
  - LukeShu 9y ago
    On Arch Linux, it's been enabled by default for several years, and I've been using it daily for that time. On x86_64, it's plenty stable, but I see why it's still experimental: On i686 there's a bug where files are mysteriously missing from directory listings. I tried to debug it once; it's reading them from the protocol correctly, but they're being dropped in the VFS layer, which is a bit more complicated than the network protocol.
  - thisrod 9y ago
    Maybe, once FUSE came along, 9pfuse made more sense than putting 9p in the kernel? https://9fans.github.io/plan9port/man/man4/9pfuse.html https://9fans.github.io/plan9port/man/man4/9pfuse.html
    - exikyut 9y ago
      I was just wondering, as I began to read the article: which would perform better - in-kernel 9P, or FUSE? Intuitively I get the idea that FUSE is going to be the better performer overall because of lack of network overhead. But I wonder how much worse 9P would be.
- fulafel 9y ago
  It's also the native host-guest file sharing protocol used by KVM: https://www.linux-kvm.org/page/9p_virtio https://www.linux-kvm.org/page/9p_virtio It works in at least Ubuntu & Fedora out of the box, and is used by the virt-manager GUI to set up shares, so not too experimental to enable by default I guess.
jtchang 9y ago
I wish 9P was named differently. Would actually speed adoption with a bit of "marketing".
- TazeTSchnitzel 9y ago
  9P from Bell Labs? :p
- PeachPlum 9y ago
  The Unix room scientists at Bell Labs deliberately chose anti-marketing names. Well done guys, you played yourselves.
shpx 9y ago
How is Upspin related to 9P?
- ori_b 9y ago
  It's not.
  - dullgiulio 9y ago
    It's not so simple. The API is a simplified version of 9P and directly inspired by it.
jakeogh 9y ago
I'm having a hard time finding any kind of feature comparison or list. What makes v9fs stand out?
- FullyFunctional 9y ago
  9p or v9fs? 9p is really easy to implement. That is a key feature as it makes it easy to create new clients and servers (and presumedly makes bugs rarer). Much like FUSE has been adapted to countless things. I don't know how implementing a FUSE server or client compares to this though. Unfortunately, the simplicity also means giving up something. Where quite-complicated NFSv4 can be very efficient over slow links (leases makes it a bit like a cache coherency protocol), 9p requires all writes to be committed to the server (for coherency).
  - jakeogh 9y ago
    I was wondering how v9fs compares to vanilla filesystems like ext2... I suspect I'm comparing the wrong thing and it's a just a NFS-like protocol and not a on disk format? If so, what is the plan9 on-disk format? I feel like I'm missing the point and should be asking "what's cool about 9p?" to which you provided a very helpful response. Thanks!
    - dfox 9y ago
      IIRC the native Plan9 FS is still something very similar to tradional unix filesystems, but it is intended to be replaced with Fossil (http://doc.cat-v.org/plan_9/4th_edition/papers/fossil/ http://doc.cat-v.org/plan_9/4th_edition/papers/fossil/). It's copy on write filesystem (like ZFS or BtrFS, but it only implements the filesystem part) with interesting property that it supports cheap snapshots which can be offloaded into Venti, which is content addressable object store (similar to Git's internal object store)
      - ConfucianNardin 9y ago
        According to Wikipedia, Plan 9 has been using fossil as the default file system since 2003.
amelius 9y ago
> Everything really was a file I wonder how that would change if the system was designed today. Everything a stream of JSON objects?
- protomyth 9y ago
  If you were going to build a JSON based file system today, then I would study the Newton’s Soup which was based on frames. It was pretty cool and had some interesting capabilities (remove half the storage from a running system, ok what’s next). I would imagine queries would make a reappearance in a much nicer form.
  - soapdog 9y ago
    Thats exactly what I thought. Newton Soups and NewtonScript were both awesome. I miss that kind of system these days, it still amazes me how it was all integrated.

99 points by miduil 9y ago

9P2000 is available on Linux with the CONFIG_9P_FS and related kernel compile options, which say things like: "If you say Y here, you will get experimental support for Plan 9 resource sharing via the 9P2000 protocol. See http://v9fs.sf.net http://v9fs.sf.net for more information."

voltagex_ 9y ago
That's been experimental for a long time - https://github.com/torvalds/linux/tree/master/fs/9p https://github.com/torvalds/linux/tree/master/fs/9p I guess file systems don't go into staging/?
- LukeShu 9y ago
  On Arch Linux, it's been enabled by default for several years, and I've been using it daily for that time. On x86_64, it's plenty stable, but I see why it's still experimental: On i686 there's a bug where files are mysteriously missing from directory listings. I tried to debug it once; it's reading them from the protocol correctly, but they're being dropped in the VFS layer, which is a bit more complicated than the network protocol.
- thisrod 9y ago
  Maybe, once FUSE came along, 9pfuse made more sense than putting 9p in the kernel? https://9fans.github.io/plan9port/man/man4/9pfuse.html https://9fans.github.io/plan9port/man/man4/9pfuse.html
  - exikyut 9y ago
    I was just wondering, as I began to read the article: which would perform better - in-kernel 9P, or FUSE? Intuitively I get the idea that FUSE is going to be the better performer overall because of lack of network overhead. But I wonder how much worse 9P would be.
fulafel 9y ago
It's also the native host-guest file sharing protocol used by KVM: https://www.linux-kvm.org/page/9p_virtio https://www.linux-kvm.org/page/9p_virtio It works in at least Ubuntu & Fedora out of the box, and is used by the virt-manager GUI to set up shares, so not too experimental to enable by default I guess.

jtchang 9y ago

I wish 9P was named differently. Would actually speed adoption with a bit of "marketing".

TazeTSchnitzel 9y ago
9P from Bell Labs? :p
PeachPlum 9y ago
The Unix room scientists at Bell Labs deliberately chose anti-marketing names. Well done guys, you played yourselves.

shpx 9y ago

How is Upspin related to 9P?

ori_b 9y ago
It's not.
- dullgiulio 9y ago
  It's not so simple. The API is a simplified version of 9P and directly inspired by it.

jakeogh 9y ago

I'm having a hard time finding any kind of feature comparison or list. What makes v9fs stand out?

FullyFunctional 9y ago
9p or v9fs? 9p is really easy to implement. That is a key feature as it makes it easy to create new clients and servers (and presumedly makes bugs rarer). Much like FUSE has been adapted to countless things. I don't know how implementing a FUSE server or client compares to this though. Unfortunately, the simplicity also means giving up something. Where quite-complicated NFSv4 can be very efficient over slow links (leases makes it a bit like a cache coherency protocol), 9p requires all writes to be committed to the server (for coherency).
- jakeogh 9y ago
  I was wondering how v9fs compares to vanilla filesystems like ext2... I suspect I'm comparing the wrong thing and it's a just a NFS-like protocol and not a on disk format? If so, what is the plan9 on-disk format? I feel like I'm missing the point and should be asking "what's cool about 9p?" to which you provided a very helpful response. Thanks!
  - dfox 9y ago
    IIRC the native Plan9 FS is still something very similar to tradional unix filesystems, but it is intended to be replaced with Fossil (http://doc.cat-v.org/plan_9/4th_edition/papers/fossil/ http://doc.cat-v.org/plan_9/4th_edition/papers/fossil/). It's copy on write filesystem (like ZFS or BtrFS, but it only implements the filesystem part) with interesting property that it supports cheap snapshots which can be offloaded into Venti, which is content addressable object store (similar to Git's internal object store)
    - ConfucianNardin 9y ago
      According to Wikipedia, Plan 9 has been using fossil as the default file system since 2003.

amelius 9y ago

> Everything really was a file I wonder how that would change if the system was designed today. Everything a stream of JSON objects?

protomyth 9y ago
If you were going to build a JSON based file system today, then I would study the Newton’s Soup which was based on frames. It was pretty cool and had some interesting capabilities (remove half the storage from a running system, ok what’s next). I would imagine queries would make a reappearance in a much nicer form.
- soapdog 9y ago
  Thats exactly what I thought. Newton Soups and NewtonScript were both awesome. I miss that kind of system these days, it still amazes me how it was all integrated.