> While snapshot consistency is useful to keep data consistent, disks often fail over
time. In order to detect corruption, block pointers contain a hash of the data that they
point at. If corrupted data is returned by the underlying storage medium, this is
detected via block hashes. And if a programmer error causes the file system to write
garbage to disk, this can often be caught early. The corruption is reported, and the
damaged data may then be recovered from backups, RAID restoration, or some other
means.
Okay! It's got CoW, snapshots, and data checksums. Therefore, it's good enough to compete with ZFS while being way smaller and permissively licensed. Now I just want it ported to Linux and the other BSDs:)
> It's got CoW, snapshots, and data checksums. Therefore, it's good enough to compete with ZFS while being way smaller and permissively licensed.
Does it have a built-in RAID layer? Because if it doesn't, then it can't compete with ZFS in many use cases. For example, what does "data may then be recovered from […] RAID restoration" mean?
With ZFS, if you have a (e.g.) mirrored/RAID-1 configuration, and you fetch some data from one drive and the checksum is wrong, ZFS can check the other drive, and if that checksum is good it can (a) pass the good data up, and (b) use the good data to fix the bad data. Most mirroring systems can't do that both-drives checking: ZFS is self-healing.
(This isn't to say that GEFS won't be useful in many other situations.)
Author here: correct, it's currently on the user to deal with corrupted data. I'm not opposed to adding a RAID layer, but I also want to explore what adding trim-style feedback to a generic RAID layer may look like, so that different file systems could share the work.
> […] but I also want to explore what adding trim-style feedback to a generic RAID layer may look like […]
TRIM commands sent to the block layer would probably help with SSD/NVMe wear leveling, as well as thin-provisioning in virtualized environments.
And the lack of RAID is not a critique: if you're mostly interested in the file system layer that's fine, and a useful area to solve problems in. My comment was more towards the comparison to ZFS and its "rampant layering violation" [1] and some of the trade-offs that happen when you focus on more or fewer parts of the stack.
Unrelated: and since you're surfing the comments, with regards to your superblocks [1], you may wish to look into some of the lessons learned by ZFS; specifically see perhaps Allan Jude's "Large Label and Blockpointer-V2" from the 2025 OpenZFS Dev Summit: https://www.youtube.com/watch?v=3tqOBv8BmXI
I don't mean a literal trim command to the block device, but hints that get passed to the RAID layer to mark how blocks should be spread among the backing devices. I haven't put much thought into the details, but I suspect that there might not be so much information needed, so the RAID layer could be both generic and smart. It may not work out.
I'll probably be picking Allan's brain on a few things in the coming few months; amusingly, he mentioned doesn't use ZFS for ZFS development so that he can unload and reload the module, and he's interested in something that won't corrupt when he panics the kernel.
> I don't mean a literal trim command to the block device, but hints that get passed to the RAID layer to mark how blocks should be spread among the backing devices.
At this point in software/file system history, having some kind of TRIM functionality I think is pretty important: either built in, or a separate utility (like fstrim(8) for ext4, etc).
Are you planning to (eventually?) do a RAID (or at least mirror) layer, or will this strictly stick do file systems?
Hi author :) Must say that 8k+ lines is a little bit annoying and uneasy to handle... But also bribgs a lot of hope and happiness :)
To the point:
- llm'ed or not, dasn't matter, just quality matter
- raid ? I know nothing about that but looks like pure hardware raid (if that exists), no "raid layers" at all and fs raid aware are three different animals.
But adding something about raids will complicate things before "production ready" stability. Which, of course, we want ASAP ;)
- 8k+ seems easy to manage and bend so maybe key-value only partitions ?
- obviously wishlist is easy to construct (key-value storage, encryption, power off resistance, mail reading ;) ) so maybe some plugin system or at least guidelines for peoples before we will have repos of patches DWM-style ? :)
- that 5 second is clearly stated but a bit strange limit and obviously peoples will want to do something with that
- man give us producion readiness before you will be puched by some corporation, swamped with work and hidden from daylight, pls ? Pleas do not fall into Compiz trap...
I've always wondered about similar designs: Doesn't calculating a hash of every block, on every read and every write, create lots of overhead? Why isn't that a problem?
Some systems have dedicated crypto co-processors for confidentiality (encryption) - e.g., I think drives with FDE, and I think Apple Silicon SoCs might have them. Can those be repurposed for hash calculation? What about systems that lack them?
Yes, it adds some overhead, but it's fine IME. Granted, it helps that compression can significantly speed up performance. (I was very confused the first time I saw ZFS reading data faster than its drives were physically capable of, because it turned out the CPU could decompress faster than the drives could read)
Programs like filesystems typically have their own schedulers that aggregate writes, it would be an extreme performance hit if they wrote each checksum individually, as a separate I/O operation. They are certainly bundled with some other data that needs to be written.
And if you are concerned about the compute rather than storage, then writing to a block device is still slow enough so that computing a checksum isn't important performance-wise.
I don't think it will compete with ZFS or BTRFS (e.g. I don't think ppl will use GEFS over ZFS or BTRFS for a storage server), but it's a modern, much needed FFS replacement.
Who said anything about storage servers? I'm using zfs on laptops and desktops right now because I want data checksums and a filesystem that doesn't have a history of breaking horribly (I dropped btrfs after the second time it hosed my rootfs). Given the license issue with zfs - and in particular, the technical fallout like needing dkms - I'd be very pleased to replace it.
They also fucked up their parity RAID repair so that the next disk failure on a different disk will destroy the data, argued that the pathological B-tree performance reported by a user couldn't be real, because they didn't understood the tree depth would only be logarithmic to the base of the fanout if the entries are the same size. insert shocked pikachu face
That failure mode is still there although it's been mitigated a bit. I hit it a few months ago but was able to recover after a few hours of doing various stuff.
It's really a shame, though. btrfs is permanently unfinished, and zfs has the wrong license. End result is people stick with ext4 or xfs most of the time anyway.
Otoh, good luck bringing a CoW filesystem back from 100%. Delete a file? Sure, let me just make a copy of all the metadata that was pointing at it using... the zero blocks I have left.
Which is why ZFS reserves "slop space" to make sure that doesn't happen, instead of defaulting to making it easy for users to corner themselves like that.
I have been hearing noise recently that btrfs is risky and unstable but (knocks on wood) i've been running it for years now with zero issues. What am I missing?
Failures in storage software are very rare. Which makes it very hard to test... (you need to run it a lot, for a very long time if you hope to find errors by chance).
Also, some failure modes are worse than others. The failures known as DI (data integrity) are the worst. Even though they aren't expected to happen to everyone at a certain frequency (because, again, mature storage software is comparatively very reliable), even a single DI error that happened to any user sets up a major alarm.
In the storage industry, the running joke is that after first DI in your product you lose funding, after the second DI you loose the product.
And it did happen to Btrfs quite a bit... I've seen it with my own eyes when a system didn't come back after power failure. (But I'm in the business of testing software storage products, so, it's less surprising that it happened to me).
So... it's perfectly plausible that you have never seen Btrfs fail, and it's been more error prone than eg. EXT4. The error rate is low enough so that if you don't actively try to cause the error you will never experience one. But, over a large group of diverse use patterns, the rate is still worse than expected.
It still has an issue with getting stuck in a bad state if you let the free space drop too low (e.g. under 5% / 2GB). It might still be recoverable but it takes a lot of fiddling and work to do that.
btrfs is almost 20 years old now, so lots of people only used it back when it was newer and far buggier.
In my experience, btrfs is actually more reliable than other filesystems due to its checksumming abilities, but when it does fail, it's much harder to fix than with other filesystems (which will often try to continue on even when stuff is broken).
> Running it for more than two years now, without any hassles. On single SSDs though. But still, waddya even talkin about?
Running ZFS since it was originally released on Solaris (June 2006) with zero problems (then FreeBSD, then Linux), including on RAID-5/6 (RAID-Z1/2) configurations, which btrfs still does not recommend using all these years later:
When I last took a serious look at btrfs its RAID 5 and RAID 6 modes could lose a whole lot of data if a single drive crashed or if the system lost power during a write. That's not supposed to happen on a RAID system with a journal.
From what I understand single copy setups, RAID 1, and RAID 10 never had the same problem. It's left a bad taste for a lot of people though.
Very early ext4 versions also had some data loss or data corruption, but I never hear people dismiss it completely for that brief spot of its history. Maybe part of that is because it was around 17 years ago and it's been a reliable default OS for many years now.
The recommendation is still to store metadata on a raid 1 or 1/0 partition and the data on a raid 5/6. Apparently there's still some issues with raid 5/6
> While snapshot consistency is useful to keep data consistent, disks often fail over time. In order to detect corruption, block pointers contain a hash of the data that they point at. If corrupted data is returned by the underlying storage medium, this is detected via block hashes. And if a programmer error causes the file system to write garbage to disk, this can often be caught early. The corruption is reported, and the damaged data may then be recovered from backups, RAID restoration, or some other means.
Okay! It's got CoW, snapshots, and data checksums. Therefore, it's good enough to compete with ZFS while being way smaller and permissively licensed. Now I just want it ported to Linux and the other BSDs:)