The drama with Linux filesystems is just nuts... It never ends.
replies(5):
The objection is the tiniest bug-fix windows get everything but the kitchen sink.
These are both uncomfortable positions to occupy, without doubt.
However, Kent, if you read this: please just settle down and follow the rules. Quit deliberately antagonizing Linus. The constant drama is incredibly offputting. Don't jeopardize the entire future of bcachefs over the silliest and most temporary concerns.
If you absolutely must argue about some rule or other, then make that argument without having your opening move be to blatantly violate them and then complain when people call you out.
You were the one who wanted into the kernel despite many suggestions that it was too early. That comes with tradeoffs. You need to figure out how to live with that, at least for a year or two. Stop making your self-imposed problems everyone else's problems.
>negative effects on the rest of the kernel
Needing to design and support an API is not purely negative for kernel developers. It also gives a change to have a proper interface for drivers to use and follow. Take a look at the Rust for Linux which keeps running into undocumented APIs that make little sense and are just whatever <insert most popular driver> does.
>FreeBSD (which is often invoked in these complaints) would have amdgpu for example.
In such a hypothetical FreeBSD could reimplement the stable API of Linux.
We already have that, with the "don't break userspace" policy combined with all of the modules being in-tree.
> Needing to design and support an API is not purely negative for kernel developers.
Sure, it's not purely negative, but it's overall a big net negative.
> Take a look at the Rust for Linux which keeps running into undocumented APIs that make little sense and are just whatever <insert most popular driver> does.
That's an argument against a stable module API! Those things are getting fixed as they get found, but if we had a stable module API, we'd be stuck with them forever.
I recommend reading https://docs.kernel.org/process/stable-api-nonsense.html
Bcachefs is not user space.
>with all of the modules being in-tree.
That is not true. There are out of tree modules such as ZFS.
>That's an argument against a stable module API!
My point was that there was 0 thought put into creating a good API. Additionally API could be evolved over time and have a support period if you care about being able to evolve it and deprecate the old one. And likely even with a better interface there is probably a way to make the old API still function.
bcachefs is still in-tree.
> That is not true. There are out of tree modules such as ZFS.
ZFS could be in-tree in no time at all if Oracle would fix its license. And until they do that, it's not safe to use ZFS-on-Linux anyway, since Oracle could sue you for it.
> My point was that there was 0 thought put into creating a good API.
There is thought put into it: it's exactly what we need right now, because if what we need ever changes, we'll change the API too, thus avoiding YAGNI and similar problems.
> Additionally API could be evolved over time and have a support period if you care about being able to evolve it.
If a temporary "support period" is what you want, then just use the LTS kernels. That's already exactly what they give you.
> And likely even with a better interface there is probably a way to make the old API still function.
That's the big net negative I was mentioning and that https://docs.kernel.org/process/stable-api-nonsense.html talks about too. Sometimes there isn't a feasible way to support part of an old API anymore, and it's not worth holding the whole kernel back just for the out-of-tree modules.
Like it does with the userland API of Linux, which is stable:
I worked for HP on storage drivers for a decade or so, and had their been a stable ABI, HP would have shipped proprietary storage drivers for everything. Even without a stable ABI, they shipped proprietary drivers at considerable effort, compiling for myriad different distro kernels. It was a nightmare, and good thing too, or there wouldn't be any open source drivers.
The filesystem should do files, if you want something more complex do it in userspace. We even have FUSE if you want to use the Filesystem API with your crazy network database thing.
Kent is in the wrong. Having a lead position in development I would kick Kent of the team.
One thing is to challenge things. What Kent is doing is something completely different. It is obvious he introduced a feature, not only a Bugfix.
If the rules are set in a way that rc1+ gets only Bugfixes, then this is absolutely clear what happens with the feature. Tolerating this once or twice is ok, but Kent is doing this all the time, testing Linus.
Linus is absolutely in the right to kick this out and it's Kent's fault if he does so.
The extended (not extant) family (including ext4) don't support copy-on-write. Using them as your primary FS after 2020 (or even 2010) is like using a non-journaling file system after 2010 (or even 2001)--it's a non-negotiable feature at this point. Btrfs has been stable for a decade, and if you don't like or trust it, there's always ZFS, which has been stable 20 years now. Apple now has AppFS, with CoW, on all their devices, while MSFT still treats ReFS as unstable, and Windows servers still rely heavily on NTFS.
They seem to be slowly introducing it to the masses, Dev drives you set up on Windows automatically use ReFS
A way to handle this would be with one person (or more) in between Kent and Linus. And maybe a separate tree only for changes and fixes from bcachefs that those people in between would forward to Linus. A staging of sorts.
IANAL, but I don't believe either of these things are true.
OpenZFS contains enough code not authored by Sun/Oracle that relicensing it now is effectively impossible.
OTOH, it is under the CDDL, which is a perfectly good open source license; AFAICT the problem, if one exists at all[0], only manifests when distributing the combination of CDDL (OpenZFS) and GPL (Linux) software. If you download CDDL software and compile it into GPL software yourself (say, with DKMS) then it should be fine because you aren't distributing it.
[0] This is a case where I'm going to really emphasize that I'm really not a lawyer and merely point out that ex. Canonical's lawyers do seem to think CDDL+GPL is okay.
And I encourage anyone who wants to contribute to join the IRC channel. It's not a one man show, I work with a lot of people there.
And the whole reason for a filesystem's existence is to store and maintain your data, so if that is what the patch if for, yes, it should be under consideration as a hotfix.
There's also the broader context: it's a major problem for stabilization if we can't properly support the people using it so they can keep testing.
More context: the kernel as a whole is based on fixed time tables and code review, which it needs because QA (especially automated testing) is extremely spotty. bcachefs's QA, both automated testing and community testing, is extremely good, and we've had bugfix patchsets either held up or turn into flamewars because of this mismatch entirely too many times.
That means Linus has to check each of his PRs assuming that it might be pushing the boundaries without warning.
No amount of post hoc justification gets you that trust back, not when this has happened multiple times now.
Which excludes a vast amount of activity one might want to use Linux for which is otherwise allowed. Like selling a device with a Linux installation, distributing VM or system restore images, etc.
I'm not saying those concerns are wrong, but when it's causing a fallout like being kicked out from the kernel, the downsides clearly are more severe than any potential benefits.
I know a lot of people heavily use slack/discord these days, but personally I find the web interfaces way too busy. IRC all the way, for me.
But the problem of communicating effectively enough to produce a coherent design is very real - this goes back to Fred Brooks (Mythical Man Month). I think bcachefs turned out very well with the way the process has gone to date, and now that it's gotten bigger, with more distinct subsystems, I am very eagerly looking forward to the date when I can hand off ownership of some of those subsystems. Lately we've had some sharp developers getting involved - for the past several years it's been mainly users testing it (and some of them have gotten very good at debugging at this point).
So it's happening.
https://lore.kernel.org/linux-fsdevel/4xkggoquxqprvphz2hwnir...
It's a bcachefs thread, and I'm the resident bcachefs expert, so.... :)
I'm not terribly invested in these threads, the actual decisionmaking happens elsewhere. But they are a good opportunity to educate people on the whole process of shipping a filesystem, talk about what we're doing, what our priorities are, all that jazz.
You get that with APFS by default on macOS these days and those features come for free in btrfs, some in XFS, etc on Linux.
That's pretty much built into most mass storage devices already.
> If a disk bitflips one of my files
The likelihood and consequence of this occurring is in many situations not worth the overhead of adding additional ECC on top of what the drive does.
> ext* won't do anything about it.
What should it do? Blindly hand you the data without any indication that there's a problem with the underlying block? Without an fsck what mechanism do you suppose would manage these errors as they're discovered?
Kent is absolutely technically capable of, and has the vision to, finally displace ext4, xfs, and zfs with a new filesystem that Does Not Lose Data. To jeopardize that by refusing to work within the well-established structure is madness.
> That's pretty much built into most mass storage devices already.
And ZFS has shown that it is not sufficient (at least for some use-cases, perhaps less of a big deal for 'residential' users).
> The likelihood and consequence of this occurring is in many situations not worth the overhead of adding additional ECC on top of what the drive does.
Not worth it to whom? Not having the option available at all is the problem. I can do a zfs set checksum=off pool_name/dataset_name if I really want that extra couple percentage points of performance.
> Without an fsck what mechanism do you suppose would manage these errors as they're discovered?
Depends on the data involved: if it's part of the file system tree metadata there are often multiple copies even for a single disk on ZFS. So instead of the kernel consuming corrupted data and potentially panicing (or going off into the weeds) it can find a correct copy elsewhere.
If you're in a fancier configuration with some level of RAID, then there could be other copies of the data, or it could be rebuilt through ECC.
With ext*, LVM, and mdadm no such possibility exists because there are no checksums at any of those layers (perhaps if you glom on dm-integrity?).
And with ZFS one can set copies=2 on a per-dataset basis (perhaps just for /home?), and get multiple copies strewn across the disk: won't save you from a drive dying, but could save you from corruption.
Being able to quickly take a "backup" copy of some multi-gb directory tree before performing some potentially destructive operation on it is such a nice safety net to have.
It's also a handy way to backup file metadata, like mtime, without having to design a file format for mapping saved mtimes back to their host files.
(it's not obvious that having to occasionally disassemble/patch closed-source drivers is worse than the collective effort wasted trying to get every single thing in the kernel and keep it up to date).
Then what you do is you try to split your work in two. You could think of a stopgap measure or a workaround which is small, can be reviewed easily, and will reduce the impact of the bug while not being a "proper" fix, and prepare the "properer" fix when the merge window opens.
I would ask, since the bug probably lived since the last stable release, how come it fell through the crack and had only been noticed recently? Could it be that not all setups are affected? If so, can't they live with it until the next merge window?
By making a "feature that fixes the bug for real", you greatly expand the area in which new, unknown bugs may land, with very little time to give it proper testing. This is inevitable, evident by the simple fact that the bug you were trying to fix exists. You can be good, but not that good. Nobody is that good. If anybody was that good, they wouldn't have the bug in the first place.
If you have commercial clients who use your filesystem and you have contractual obligations to fix their bugs and keep their data intact, you could (I'd even say "should") maintain an out-of-tree version with its own release and bugfix schedule. This is IMO the only reasonable way to have it, because the kernel is a huge administrative machine with lots of people, and by mainlining stuff, you necessarily become co-dependent on the release schedule for the whole kernel. I think a conflict between kernel's release schedule and contractual obligations, if you have any, is only a matter of time.
You're thinking of the optimization technique of CoW, as in what Linux does when spawning a new thread or forking a process. I'm talking about it in the context of only ever modifying copies of file system data and metadata blocks, for the purpose of ensuring file system integrity, even in the context of sudden power loss (EDIT: wrong link): https://www.qnx.com/developers/docs/8.0/com.qnx.doc.neutrino...
If anything, ordinary file IO is likely to be slightly slower on a CoW file system, due to it always having to copy a block before said block can be modified and updating block pointers.
What kind of journaling though? By default ext4 only uses journaling for metadata updates, not data updates (see "ordered" mode in ext4(5)).
So if you have a (e.g.) 1000MB file, and you update 200MB in the middle of it, you can have a situation where the first 100MB is written out and the system dies with the other 100MB vanishing.
With a CoW, if the second 100MB is not written out and the file sync'd, then on system recovery you're back to the original file being completely intact. With ext4 in the default configuration you have a file that has both new-100MB and stale-100MB in the middle of it.
The updating of the file data and the metadata are two separate steps (by default) in ext4:
* https://www.baeldung.com/linux/ext-journal-modes
* https://michael.kjorling.se/blog/2024/ext4-defaulting-to-dat...
* https://fy.blackhats.net.au/blog/2024-08-13-linux-filesystem...
Whereas with a proper CoW (like ZFS), updates are ACID.
Software take up no space and there is no scarcity. Theoretically there could be any number of maintainers and what gets uptake is the de facto upstream. That's what people refer to when they talk about free software development in terms of meritocracy.
The PR description would have been fine - if it had been in the right stage of the process.
There is always a point where you have to say "no I can't work with this person any more", but while you are still trying to it's worth trying to figure out why someone is behaving as they do.
bCacheFS, not BCA Chefs. I’m not clued into the kernel at this level so I racked my brain a bit.
It would be interesting how strict the rules are in the Linux kernel for other people. Other projects have nepotistic structures where some developers can do what they want but others cannot.
Anyway, if Linus had developed the kernel with this kind of strictness from the beginning, maybe it wouldn't have taken off. I don't see why experimental features should follow the rules for stable features.
It's not against the license to use them together.
>If a temporary "support period" is what you want, then just use the LTS kernels. That's already exactly what they give you.
Only the Android one does. The regular LTS one has no such guarantee.
Apparently bcachefs won't be the successor. Filesystem development for Linux needs a big shakeup.
I looked at that, in hopes of being able to protect my data. Unfortunately, I considered this something of a fatal flaw:
> It uses journaling for guaranteeing write atomicity by default, which effectively halves the write speed.
> What should it do? Blindly hand you the data without any indication that there's a problem with the underlying block?
Well, that's what it does now, and I think that's a problem.
> Without an fsck what mechanism do you suppose would manage these errors as they're discovered?
Linux can fail a read, and IMHO should do so if it cannot return correct data. (I support the ability to override this and tell it to give you the corrupted data, but certainly not by default.) On ZFS, if a read fails its checksum, the OS will first try to get a valid copy (ex. from a mirror or if you've set copies=2), and then if the error can't be recovered then the file read fails and the system reports/records the failure, at which point the user should probably go do a full scrub (which for our purposes should probably count as fsck) and restore the affected file(s) from backup. (Or possibly go buy a new hard drive, depending on the extent of the problem.) I would consider that ideal.
At work we have our main application which also contains a lot of customer integrations. Our policy has been new features in trunk only, except if it's entirely contained inside a customer-specific integration module.
We do try to avoid it, but this does allow us to be flexible with regards to customer needs, while keeping the base application stable.
This new recovery feature was, as far as I could see, entirely contained within the bcachefs kernel code. Given the experimental status, as long as it was clearly communicated to users, I don't see a huge problem allowing such self-contained features during the RC phase.
Obviously a requirement must be that it doesn't break the build.
That makes checksums and journals of only marginal usefulness.
I wish some review website would have a robot plug and unplug the power cable in a test rig for a few weeks and rate which SSD manufacturers are robust to this stuff.
Ideally they'd be developed outside the kernel until they are perfect, but Kent addresses this in his LWN comment: There is no funding/time to make that ideal scenario possible.
If you’re using experimental file systems, I’d expect you to be pretty competent in being able to hold your own in a storage emergency, like compiling a kernel if that’s the way out.
This is a made up emergency, to break the rules.
According to that 10^14 metric I should see read errors just about every month. Except I have just about zero.
Current disks are ~4 years, runs 24/7, and excluding a bad cable incident I've had a single case of a read error (recoverable, thanks ZFS).
I suspect those URE numbers are made by the manufacturers figuring out they can be sure the disk will do 10^14, but they don't actually try to find the real number because 10^14 is good enough.
Speed is sometimes more important than absolute reliability, but it’s still an undesirable tradeoff.
That is indeed what I normally do. For example, 6.14 and 6.15 had people discovering btree iterator locking bugs (manifesting as assertion pops) while running evacuates on large filesystems (it's hard to test a sufficiently deep tree depth in virtual machine tests with our large btree nodes); some small hotfixes went out in rc kernels, but the majority of the work (a whole project to add assertions for path->should_be_locked, which should shut these down for good) waited until the 6.16 merge window.
That was for a less critical bug - your machine crashing is somewhat less severe than losing a filesystem.
In this case, we had a bug pop up in 6.15 where the link count in the VFS inode getting screwed up caused an inode to be deleted that shouldn't have been - a subvolume root - and then an untested repair path took out the entire subvolume.
Ouuuuch.
That's why the repair code was rushed; it had already gotten one filesystem back, and I'd just gotten another report of someone else hitting it - and for every bug report there are almost always more people who hit it and don't report it.
And considering that a lot of people running bcachefs now are getting it from distro kernels and don't know how to build kernels - that is why it was important to get this out quickly through the normal channels.
In addition, the patch wasn't risky, contrary to what Ted was saying. It's a code path that's very well covered by automated tests, including KASAN/UBSAN/lockdep variants - those would exploded if this patch was incorrect.
When to ship a patch is always a judgement call, and part of how you make that call is how well your QA process can guarantee the patch is correct. Part of what was going on here is a disconnect between those of us who do make heavy use of modern QA infrastructure and those who do it the old school way, relying heavily on manual review and long testing periods for rc kernels.
Like the sibling commenter, I suspect the word “experimental” is being used here to try and evade the rules that, somehow, every other part of the kernel manages to do just fine with.
No, it doesn't. Maybe you're thinking of shadow volume copies or something else. CoW files systems never modify data or metadata blocks directly, only modifying copies, with the root of the updated block pointer graph only updated after all other changes have been synced. Read this: https://www.qnx.com/developers/docs/8.0/com.qnx.doc.neutrino...
First niche is the full featured CoW filesystem; it has snapshots, detects and repairs corruption, transparent compression, all that good stuff.
The other niche is being an allocator of sectors. There's one storage device, divide it up amongst all these processes asking for storage. That's Ext4: an allocator of disk sectors, dressed up in a filesystem API. When you are running databases or VMs, all you want is an allocator of sectors. You don't want lots of stuff getting in the way of your writes. You don't want checksumming, you don't want your writes going to a new place every time. You just want write access to part of the disk.
We're very far along in that process now, but it's still marked as experimental because it is not quite ready for widespread deployment by just anyone. 6.16 is getting damn close, though.
That means a lot of our users now are people getting it from distro kernels, who often have never compiled a kernel before - nevertheless, they can and do report bugs.
And no matter where you are in the rollout, when you get bug reports you have to fix them and get the fixes out to users in a timely manner so that they can keep running, keep testing and reporting bugs.
It's a big loss if a user has to wait 3 months for a bugfix - they'll get frustrated and leave, and a big part of what I do is building a community that knows how the system works, how to help debug, and how to report those bugs.
A very common refrain I get is "it's experimental <expletive deleted>, why does it matter?" - and, well, the answer is getting fixes out in a timely manner matters just as much if not more if we want to get this thing done in a timely manner.
I've seen this sentiment a lot lately. That disagreeable top performers have to be disposed of because they are "toxic" or "problematic".
You aren't doing your job as a leader if this is your attitude to good engineers. Engineering is a field where a small amount of the people create a large amount of the value. You can either understand that, and take it upon yourself to integrate disagreeable yet high performing people into the team, paving over the rough patches yourself. Or you can oust them, and quite literally take a >50% productivity hit on your team.
A disagreeable person will take up more of your time as a manager, but a high performer is worth significantly more of your time. When these traits co-occur in the same person, the cost-benefit is complicated. The reason we talk about this problem a lot in tech is because it is legitimately a tough call, with errors in both directions. Wishing that the right move was always as simple as kicking someone off the team doesn't make it true, although it may relieve you from having to contend with the decision.
https://www.youtube.com/@SavvyNik/videos
He gets a few words wrong because my understanding is he covers the topic in a more broad way, but most of his coverage seems objective and factual. He does have some opinions, but I think it's closer to journalism of the LKML than an opinion piece.
Ideally, you teach people how to get along better together; I think of my job as manager (and I effectively do manage a large team these days) as one of teaching and fostering good communication.
In a short term, possibly. But do you think bcachefs is better in the current situation than if it moved at half the speed, but without conflict? By being out of kernel it will get less testing, fewer contributions, the main developer will get some time wasted on rebasing the patch set with every release, and distros are unlikely to expose bcachefs to the user any time soon. When you're working with an ecosystem / teams, single person's performance really doesn't mean that much in the end. And occasionally Kent will still have to upstream some changes to interfaces - how likely is anyone to review/merge them quickly?
And now, what are the chances this will ever become more than a single person project really?
My job matches what you’re describing, but bug fix is widely interpreted. It basically means “managers don’t do anything stupid”.
If someone got in trouble using language you are desiring “the rules are clear and were broken”, i would feel they were singling someone out.
Anything unimportant is really just being cached and it's probably fine if it gets corrupted.
At the time it looked like that could happen - there was real interest from Redhat prior to merging. Sadly Redhat's involvement never translated into much code, and while upstreaming did get me a large influx of users - many of which have helped enormously with the QA and stabilization effort - the drama and controversies have kept developers away, so on the whole it's meant more work, pressure and stress for me.
So DKMS wouldn't be the worst route, at this point. It would be a real shame though, this close to taking the experimental label off, and an enormous hassle for users and distributions.
But it's not my call to make, of course. I just write code...
> When you're working with an ecosystem / teams, single person's performance really doesn't mean that much in the end.
This is demonstrably not true. Kent brought Bcachefs to fruition and got it upstreamed. Wireguard was also one guy. The cryptography used in both, also 1 guy. There's an argument to be made that given an elegant, well designed system, we should assume it came from a single or a few minds. But given a system that's been around for a while, you would be right to assume that a lot of people were/are involved in keeping it around.
And before that I have been using 8x WD Reds 3TB for 6-7 years, which have 10^14 in the specs[2], and had the same experience with those.
Yes smaller size, but I ran scrubbing on those biweekly, and over so many years?
[1]: https://www.seagate.com/files/www-content/datasheets/pdfs/ir...
[2]: https://documents.westerndigital.com/content/dam/doc-library...
This is clearly untrue. Upon what theory?
Agree, plus I’d add: if we are going to criticise other people’s communication style/abilities or attitude, then using a vague, vulgar and hurtful slang term like “arsehole”/“asshole” (and similar slang such as “dick”, “prick”, etc) is an example of exhibiting the very thing one is complaining about in making the complaint, which is fundamentally hypocritical. One can state the same concerns in a more professional way, focusing on the details of the specific behaviour pattern not a vague term which can refer to lots of distinct behaviours (e.g. people with ASD traits who hurt the feelings of others because they honestly have trouble thinking about them, versus people with antisocial or narcissistic personality disorder traits who knowingly hurt the feelings of others because they enjoy doing so) - labelling the behaviour pattern not the person, acknowledging that it is entirely possibly due to an unintentional skills gap, (sub)culture clash, differences in life experiences, neurodiversity/neurodivergence/mental health/trauma, etc.
I also think it is helpful when criticising the flaws of others to try to relate them to one’s own, whenever possible - e.g. sometimes in the past I did X and from my perspective it looks like you are doing something similar-hurtful labels are not encouraging that kind of self-reflectiveness at all, they promote the idea that “I’m one of the good ones but you are one of the bad ones”
I think the performance cost of FUSE compared to in-kernel filesystems is improving with time - FUSE with io_uring is a big step forward, but the immaturity of io_uring is an obstacle to its adoption, at least in the short-to-medium-term. I’m sure in the future we’ll see even further improvements in this area. But will we ever reach the Nirvana where FUSE equals the performance of in-kernel filesystems, or (maybe more realistically) the performance overhead has become so marginal nobody is bothered by it in practice? I’d like to think we eventually will, but it is far from certain.
> while MSFT still treats ReFS as unstable, and Windows servers still rely heavily on NTFS.
By this I implied it's an embarrassment to MSFT that iOS devices have a better, more reliable file system (AppFS) than even Windows servers now (having to rely on NTFS until ReFS is ready for prime time). If HN users and mods didn't tone-police so heavily, I could state things more frankly.
An unstable interface means the driver source needs to be updated frequently, you can't just dump a .ko file online and expect it to work for however long the hardware lasts.
Easiest way to approach it is to attempt to upstream drivers, and potentially take advantage of free labor and maintenance in virtual perpetuity, which is good for all Linux users. If vendors don't want to spend the effort upstreaming drivers, but they need to support Linux, by necessity the drivers must be open source so they can be compiled against users' changing kernels. That's at least a step in the right direction, and should anyone want to make the effort, they're free to upstream drivers themselves.
Comparatively, I find subtweeting him from the sanctity of Mastodon, with a few insults and backhanded complements thrown in for good measure, a bit low.
Which implies you can already correct errors through a simple majority mechanism.
> or it could be rebuilt through ECC.
So just by having the appropriate level of RAID you automatically solve the problem. Why is this in the fs layer then?
By contrast, I don't know what "BCA Chefs" is supposed to be. "BCA" could be many things: "Barbados Cricket Association", "Billiard Congress of America", "British Caving Association", "Business Council of Australia", among others. But what would any of them have to do with "Chefs"?
I don't think so. Suppose Oracle did agree to put their code under GPLv2/CDDL dual licensing.
Then, I'm sure if you look at the non-Oracle contributors to OpenZFS, there's a few big ones and a long tail of smaller ones. Many of the big ones might be able and willing to follow Oracle's lead. Chasing down the smaller ones may be harder, but it is possible their contributions may be judged as sufficiently trivial to escape copyright protection. More substantive contributions from people who are unreachable (or unwilling/unable to consent to the relicensing) can pose a bigger issue, but it could be solved either by (a) intentionally rewriting their contributions from scratch; (b) given enough time, decent chance (a) will happen anyway just to normal code churn, even if you don't do it intentionally for licensing reasons.
It would be a big, multi-year project, but one that other open source communities have successfully tackled, most notably LLVM – so I do think "effectively impossible" is too strong.
I think the biggest blocker is that, it is hard to motivate people to make the effort unless Oracle is on-board – and they've displayed no signs of willingness to change their position on this. I doubt Oracle will budge, but anything is possible.
Another possibility to consider – CDDL clause 4 allows the "license steward" (Sun Microsystems) [0] to release a new version, which automatically applies to all CDDL software unless the developers explicitly opt-out. I don't know if any of the OpenZFS developers have made such an explicit opt-out – but if they haven't, then Oracle could issue a new CDDL version adding a clause saying that if the covered work is ZFS or a derivative thereof, anyone is allowed to relicense it under GPLv2. Then you wouldn't even need to track down and get the consent of non-Oracle contributors. For a real historical example of something like this, witness how the FSF issued a new GFDL version just to help Wikipedia move from GFDL to Creative Commons licensing. But, again, even if this is legally possible, unlikely (but not impossible) Oracle will ever cooperate in it.
Another blocker is that even if OpenZFS were relicensed as GPLv2/CDDL, that still wouldn't solve the issue that Torvalds is unlikely to agree to upstreaming it as part of the mainline Linux kernel – a massive code base written in a very different style, and having portability concerns (wanting to work on BSD/etc too) which Linux normally doesn't care about. Possibly if you forked OpenZFS, ripped out the cross-platform aspects, and rewrote it to be more like typical Linux kernel code, it might have a chance. But, will anyone be willing to make that massive investment of time and effort? And even assuming they succeeded, we'd now have two forks of ZFS (one in the Linux kernel, one for other operating systems), adding to the maintenance burden, and the risk they'd diverge over time would be high.
[0] Sun Microsystems still legally exists on paper, and probably will indefinitely, as an Oracle subsidiary – it has been renamed to Oracle America Inc – so Oracle has effectively inherited Sun's rights as CDDL license steward
It seems like you want that kind of fast turn around time in the kernel though, which strikes me as an impossible goal.
I don't think so? You set copies=2, and the disk says that your file starts with 01010101, except that the second copy says your file starts with 01010100. How do you tell which one is right? For that matter, even with only one copy ex. ZFS can tell that what it has is wrong even if it can't fix it, and flagging the error is still useful.
> So just by having the appropriate level of RAID you automatically solve the problem. Why is this in the fs layer then?
Similarly, you shouldn't need RAID to catch problems, only (potentially) to correct them. I do agree that it doesn't necessarily have to be in the FS layer, but AFAIK Linux doesn't have any other layers that do a good job of it (as mentioned above, dm-integrity exists but halving the write speed is a pretty big problem).
Direct IO would be slower via FUSE, but L4 style IPC could solve that.
It would be an interesting proposition, although not my first choice for the direction I want to go in :)
However, there is also a long time since XFS supports snapshots.
See for example:
https://thelinuxcode.com/xfs-snapshot/
I am not sure what you mean by "whole-volume" snapshots, but I have not noticed any restrictions in the use of the XFS snapshots. As expected, they store a snapshot of the entire file system, which can be restored later.
In many decades of managing computers with all kinds of operating systems and file systems, on a variety of servers and personal computers, I have never had the need to shrink a file system. I cannot imagine how such a need can arise, except perhaps as a consequence of bad planning. There are also many decades since I have deprecated the use of multiple partitions on a storage device, with the exception of bootable devices, which must have a dedicated partition for booting, conforming to the BIOS or UEFI expectations. For anything that was done in the ancient times with multiple partitions there are better alternatives now. With the exception of bootable USB sticks with live Linux or FreeBSD partitions, I use XFS on whole SSDs or HDDs (i.e. unpartitioned), regardless if they are internal or external, so there is never any need for changing the size of the file system.
Even so, copying a file system to an external device, reformatting the device and copying the file system back is not likely to be significantly slower than shrinking in place. In fact sometimes it can be faster and it has the additional benefit that the new copy of the file system will be defragmented.
Much more significant than the lack of shrinking ability, which may slow down a little something that occurs very seldom, is that both EXT4 and XFS are much faster for most applications than the other file systems available for Linux, so they are fast for the frequent operations. You may choose another file system for other reasons, but choosing it for making faster a very rare operation like shrinking is a very weak reason.
And no, it's not necessary even for root file systems. Linux can load modules, such as file system drivers, before it mounts root. That's what initramfs is about.
ZFSoL has thrived for 15 years, fostering several commercial empires, and has never been in Linus's mainline.
Bcachefs development may continue as Kent Overstreet wishes, and he need not squabble with Linus going forward. Seems like an entirely workable outcome. Kudos to Linus for a.) giving Kent a chance, despite known issues with Kent, and b.) making the difficult decision to reverse himself. Both of these decisions were correct.
What I learn from all of this is that Linus is still in the saddle and still making good calls. We are blessed.
By contrast, improvements in performance of FUSE, L4-style IPC, could be much more widely beneficial-both for developers of new physical filesystems (by making possible in-user space implementations where they can iterate faster, get better API/ABI stability, easier adoption by end-users), but also for developers of numerous other pieces of software too
Of course, you personally are going to scratch the itch you want to scratch. But in terms of what’s most beneficial for the Linux ecosystem as a whole, I think FUSE improvements and L4-style IPC would deliver the most benefit per unit of effort
I have no context on the situation or understanding of what the right set of rules here is, but the difference between filesystems and other code is that bugs in the filesystem can cause silent, persistent corruption for user data, on top of all the other failure modes. Most other parts of the kernel don't have such a large persistence capability in case of failure. So I can understand if filesystems feel the need to be special.
I wonder where bcachfs in regards to mobile snapshots and hibernation.
The biggest reason raid btrfs is not trustable is that it has no mechanism for correctly handling a temporary device loss. It will happily rejoin an array where one of the devices didn’t see all the writes. This gives a 1/N chance of returning corrupt data for nodatacow (due to read-balancing), and for all other data it will return corrupt data according to the probability of collision of the checksum. (The default is still crc32c, so high probability for many workloads.) It apparently has no problem even with joining together a split-brained filesystem (where the two halves got distinct writes) which will happily eat itself.
One of the shittier aspects of this is that it is not clearly communicated to application developers that btrfs with nodatacow offers less data integrity than ext4 with raid, so several vendors (systemd, postgres, libvirt) turn on nodatacow by default for their data, which then gets corrupted when this problem occurs, and users won’t even know until it is too late because they didn’t enable nodatacow.
The main dev knows this is a problem but they do seem quite committed to not taking any of it seriously, given that they were arguing about it at least seven years ago[0], it’s still not fixed, and now the attitude seems to just ignore anyone who brings it up again (it comes up probably once or twice a year on the ML). Just getting them to accept documentation changes to increase awareness of the risk was like pulling teeth. It is perhaps illustrative that when Synology decided to commit to btrfs they apparently created some abomination that threads btrfs csums through md raid for error correction instead of using btrfs raid.
It is very frustrating for me because a trivial stale device bitmap written to each device would fix it totally, and more intelligently using a write intent bitmap like md, but I had to be deliberately antagonistic on the ML for the main developer to even reply at all after yet another user was caught out losing data because of this. Even then, they just said I should not talk about things I don’t understand. As far as I can tell, this is because they thought “write intent bitmap” meant a specific implementation that does not work with zone append, and I was an unserious person for not saying “write intent log” or something more generic. (This is speculation, though—they refused to engage any more when I asked for clarification, and I am not a filesystem designer, so I might actually be wrong, though I’m not sure why everyone has to suffer because a rarefied few are using zoned storage.)
A less serious but still unreasonable behaviour is that btrfs is designed to immediately go read-only if redundancy is lost, so even if you could write to the remaining good device(s), it will force you to lose anything still in transit/memory if you lose redundancy. (Except that it also doesn’t detect when a device drops through e.g. a dm layer, so you can actually ‘only’ have to deal with the much bigger first problem if you are using FDE or similar.) You could always mount with `-o degraded` to avoid this but then you are opening yourself up to inadvertently destroying your array due to the first problem if you have some thing like a backplane power issue.
Finally, unlike traditional raid, btrfs tools don’t make it possible to handle an online removal of an unhealthy device without risking data loss because in order to remove an unhealthy but extant device you must first reduce the redundancy of the array—but doing that will just cause btrfs to rebalance across all the devices, including the unhealthy one, and potentially taking corrupt data from the bad device and overwriting on the good device, or just losing the whole array if the unhealthy device fails totally during the two required rebalances.
There are some other issues where it becomes basically impossible to recover a filesystem that is very full because you cannot even delete files any more but I think this is similar on all CoW filesystems. This at least won’t eat data directly, but will cause downtime and expense to rebuild the filesystem.
The last time I was paying attention a few months ago, most of the work going into btrfs seemed to be all about improving performance and zoned devices. They won’t reply to any questions or offers for funding or personnel to complete work. It’s all very weird and unfortunate.
[0] https://www.mail-archive.com/linux-btrfs@vger.kernel.org/msg...
The Kernel's pace is predictable, even billion $ corporates can live with it, and it's not like Linus hasn't accommodated you in the past. But continuing to do so will make people stop believing you are acting in good faith, and the outcome of that will be predictable.
This is simply how the development model is like in the Linux kernel. Exceptions happen, but they sometimes backfire and highlight why the rules matter in the first place, and therefore they are unlikely to change.
You’re acting like bcachefs systems are storing Critical Data That Absolutely Cannot Be Lost. And yet at the same time it’s experimental. I’m just one user, but I can tell you that, even as excited as I am about bcachefs, I’m not touching it with a ten foot pole for anything beyond playing around until at least the experimental label is removed.
I imagine my position is not uncommon.
Please stop trying to die on this hill. Your project is really great and really important. I want it to succeed.
Just chill and let bug fixes be bug fixes and features be features.
That said, I vaguely seem to recall that bcachefs ended up involving changes to other parts of the kernel to better support it; if that's true then DKMS is likely to be much more painful if not outright impossible. It's fine to compile one module (or even several) against a normal kernel, but the moment you have to patch parts of the "main" kernel it's gonna get messy.
I have had a similar experience with a team member who was quietly unhappy about a rule. Instead of raising a discussion about the rule (like the rest of the team members did) he tried to quietly ignore it in his work, usually via requesting reviews from less stringent reviewers.
As a result, after a while I started documenting every single instance of his sneaky rule-breakage, sending every instance straight to his manager, and the person was out pretty soon.
The disk is going to report an uncorrected error for one of them.
Otherwise that is a red flag. Like pull requests are when discussions are had...
Sometimes a new filesystem needs changes to things in the kernel and the VFS API isn't enough, but often VFS is enough.
I.e. back in ~ 2013-2014 while moving some baremetal Windows server into VMware, srhinking and then optimizing MFT helped to save AFAIR 2 hours of downtime window.
> except perhaps as a consequence of bad planning
Assuming people go to Clouds instead of physical servers because they may need to add 100 more nodes "suddenly" - selling point of Clouds is "avoid planning" - one may expect cases of need of shrinking are rising, now lowing. It may be mitigated by different approaches of course - i.e. often it's easier to resetup VM, but yet.
It is directly challenged in the very thread linked in the article (and likely before, the drama is ancient).
Also, there is no "less stringent reviewer", it's always been the same you!
So your example fails at both core points, yet your outcome is still the same happy firing!
At least for paid work you can just sprinkle $ to cover up such mistakes and find someone else, but wait, this is also not paid work!
But what prevents it ultimately? This ... situation. It makes me sad.
I didn't follow the details but I know that Linus is a reasonable person, and Kent is very thorough and delivering quality. But even if Linus was too much on the conservative-side here (but who's to judge??), please Kent, just fall in line. The alternative is nothing. Go have a beer with Linus.
I'd be highly conservative about using it for my home directory though. Or at least make a subfolder where all my really important files (legal documents, master thesis, etc.) go and mount that on another partition that uses a more conservative filesystem.
https://docs.docker.com/engine/storage/drivers/btrfs-driver/
There are far better options such as FUSE or the filesystem APIs in other operating systems like Netbsd, Haiku, Genode or even ReactOS (and Windows NT).
Some of the best filesystems such as OpenZFS, HAMMER2 or Lustre are developed outside of Linux.
I will also say that bcachefs's selling point - and probably a major reason people are so excited for it - is amount of effort it puts into avoiding data corruption. Which tells you something about the perceived quality of other filesystems on Linux. Which means that saying "other filesystems seem fine with the rules" misses the very fact that people have seen too much data corruption with other filesystems and want something that prioritizes it higher.
The proper way here would have been two pull requests, one with all the bugfixes, and one with the new feature with a cover letter motivating why an exception should happen. And if this happens often enough with sufficiently good backing motivations, then he may be able to convince people.
Unfortunately Kent spends a lot of time and effort defending Kent. I wish he would learn to take a step back and admit he's fallible and learn to play nice in the sandbox rather than wasting all of this time and effort. A simple "mea culpa" could smooth over a lot of the feathers he constantly ruffles.
It's because he views user data integrity as a more critical concern than the PR process or team dynamics - which, as a user, I don't fault him for. I think that in his mind, every hour/day/week spent debating things on a PR equals more people losing or corrupting data. This is not commonly the case with most PRs - it's specific to popular filesystems in active development.
What I don't necessarily buy is how to weigh this responsibility against the responsibility users take on when they use such an experimental FS in the first place. It's a tough question in my mind, and both sides have good points. And I also don't know anything about the relative safety vs. severity of each patch. But what I do understand is the motivation for not viewing these as generic PRs against generic codebases. So the idea that this is a red flag in this case just doesn't seem right to me, based on my current understanding.
Either the bugfix is not serious and they can wait because the system is mature. Or, The fs is so unstable you shouldn't be pandering to the crowd that struggle with build deps and make install.
There is no in between, this is the situation. And the "but not all bugs are equal" argument doesn't stand.
I know if I read of a metadata but getting fixed in ext4 or ZFS there's a very small chance of this causing my platter to evaporate. By definition of stable, if that was happening it would be hitting that one unfortunate guy (<0.001% of users) running a weird OS or hardware and that's just the luck of the draw.
If the fix is from a fs marked experimental, yes I kinda expect this could fry my data and hurt kittens. That's what experimental and stable mean. That means I expect up to 100% of users to be impacted under certain workflows or scenarios.
Everything outside of this is wasted energy and arguing with the tide.
In migrations you normally copy the file system elsewhere, to the cloud or to different computers, you do not shrink it in place, which is what XFS cannot do. Unlike with Windows, copying Linux file systems, including XFS, during migrations to different hardware is trivial and fast. The same is true for multiplicating a file system to a big set of computers.
Shrinking in place is normally needed only when you share a physical device between 2 different operating systems, which use incompatible file systems, e.g. Windows and Linux, and you discover that you did not partition well the physical device and you want to shrink the partition allocated for one of the operating systems, in order to be able to expand the partition allocated for the other operating system.
Sharing physical devices between Windows and any other operating systems comes with a lot of risks and disadvantages, so I strongly recommend against it. I have stopped sharing Windows disks decades ago. Now, if I want to use the same computer in Windows and in another operating system, e.g. Linux or FreeBSD, I install Windows on the internal SSD, and, when desired, I boot Linux or FreeBSD from an external SSD. Thus the problem of reallocating a shared SSD/HDD by shrinking a partition never arises.
In the extreme, if bcachefs gets removed from the kernel, the productivity outcome (depending on your measure) is actually zero.
[Ed.: also, honestly, if you need to hire a "babysitter" for such a highly skilled engineer, that is also a viable option & there shouldn't be a social stigma for that either. I wouldn't say it's the manager's job though, not to that degree at least.]
We've had someone show up claiming "I'm going to do FUSE right!" and it never happened, so - the incremental approach is probably best here. But it's going to take awhile.
Skimming over the context, Kent seems to be lying by omission in PRs and distorting the history behind the PRs. Plus fighting with what is basically his tech lead who represents the team's norms, culture, and health. I also think he's fighting in this comment section right now.
Speaking as a former "brilliant jerk", I wouldn't trust the memory or intentions of a brilliant jerk. I wouldn't want to be looking over my shoulder for back-stabbing on my own team. I also wouldn't want my manager to get stuck in "I can fix him" mode because they're afraid of doing their flipping job: firing an employee who refuses to learn from their mistakes.
* I'm sympathetic to contexts when the team itself is bad and the performer is actually doing better. In that case: forget their trust, the performer should either remake the team (become the manager) or leave to do better work.
It's all part of the job.
Sure, you talk to them. And sure, you explain what the problem is and treat them like an adult. But ultimately it is completely acceptable to give up.
Peoples’ potential matters to parents, and to mentors. A high-potential, low-performing person can be a project worth taking on, but they are not an obligation in the workplace, especially for someone as senior and time-constrained as Linus.
Is the wait really 3 monts away? I don't exactly know the release cycle, but for me kernels are released quite frequently, at least there are RC sooner than 3 months. Just checked latest history and major releases are 2 months apart - and between them there are minor ones.
People using experimental features are quite aware how to get new experimental kernel sources.
What you want to avoid is becoming dependent on continued development of it - but unless you’re particularly using some specific feature of the file system that none other provide you’ll have time to migrate off it.
Even resierfs didn’t cease to operate.
It's a bad sign when the key founders leave like that, filesystems require a lot of coherence of design and institutional knowledge to be retained.
I’ve seen it time and time again, sometime so much so that hiring the engineer also means hiring his handler, and everyone knows it and is ok with it - even the engineer.
I think there have been one or two bug reports in the past from rebalance not freezing in a timely manner (laptops don't usually use rebalance, that's usually a multi device thing), but I think they've been fixed. Send me a bug report if it's not :)
Big distros like RedHat or Ubuntu always roll patched kernels, and as far as I know they’re usually a bit long in the tooth.
Funnily enough, I just tracked down a problem that significantly affected the calculation of how much money something cost down to an issue one of these geniuses introduced by thinking they were too good for regular, dull, due diligence in their development practices.
Then you have details on how to obtain recovery tools. You’d only need it for one patch revision.
It gets to be a real distraction. Fortunately the people I work with have learned how to roll with it, so it's not nearly as bad as it used to be. Now it mainly shows up in forum comments where it doesn't really affect me and I can eat popcorn.
It is true that I don't want critical fixes being held up by angry arguing, but most pull requests, even fixes, aren't nearly so critical.
The main thing I keep hammering on is "the development process _matters_ if we want to get this done right", and user considerations are a big part of that.
Debugging issues that come up in the wild, and getting those fixes to users in a timely manner so they can keep testing and we can get all these crazy failure modes sorted out is a big part of that - if we want a filesystem that's truly bulletproof. I know I want that!
I've been spending the past week and a half mostly working with one user and his filesystem that's been through flaky dying controllers and now lightning strikes; ext4 even got corrupted on the same setup.
But we discovered some 6.16 regressions, got some more people involved staring at code and logs (a new guy spotted a big one), and another small pile of fixes are going out next week. And even with the 6.16 regressions (some nasty ones were found), it's looking like he didn't lose much, thanks in part to journal rewind.
This thing is turning into a tank.
All in a day's work...
As someone who might have fallen in your grumpy-disagreeable-senior bucket at times, that's a different story and not something I would accept.
> I've seen this sentiment a lot lately. That disagreeable top performers have to be disposed of because they are "toxic" or "problematic".
This is not really a relevant argument to this situation.
But Apple has historically been strong on organizing and supporting teams (see: their chip design), a filesystem sounds exactly like something they'd do well if they decided to give it the proper investment and support.
Where they seem to be falling down these days is software maintenance - many, many reports of MacOS getting buggier with every release. But a big, complicated, but well defined and self contained engineering project? That's their ballpark.
In terms of maturity, bcachefs is closer to production quality than reiser4 was, but it's still closer to reiser4 than reiserfs in its lifecycle.
That is just mind bogglingly inept. (And thanks, I hadn't heard THIS one before).
For nocow mode, there is a bloody simple solution: you just fall back to a cow write if you can't write to every replica. And considering you have to have the cow fallback anyways - maybe the data is compressed, or you just took a snapshot, or the replication level is different - you have to work really hard or be really inept to screw this one up.
I honestly have no idea how you'd get this wrong in cow mode. The whole point of a cow filesystem is that it makes these sorts of problems go away.
I'm not even going to go through the rest of the list, but suffice it to say - every single broken thing I've ever seen mentioned about btrfs multi device mode is fixed in bcachefs.
Every. Single. One. And it's not like I ever looked at btrfs for a list of things to make sure I got right, but every time someone mentions one of these things - I'll check the code if I don't remember, some of this code I wrote 10 years ago, but I yet to have seen someone mention something broken about btrfs multi device mode that bcachefs doesn't get right.
It's honestly mind boggling.
By the way, if that was serious, bcachefs would love the help, and more people are joining the party.
I would love to find someone to take over erasure coding and finish it off.
Unfortunately I don't think Patreon can fund the kind of talent you need to sustainably develop a file system.
That btrfs contains broken features is IMO 50/50 the fault of up-stream and the distributions. Distributions should patch out features that are broken (like btrfs multi-device support, direct IO) or clearly put it behind experimental flags. Up-stream is unfortunately incentivised to not do this, to get testers.
Data integrity, core IO paths, all the failure modes, all the crazy repair corner cases - these are the hard parts if you really want to get a filesystem right. These are what we've been taking our time on.
I can't claim 100% success rate w.r.t. data loss, but it's been phenomenally good, with some crazy stories of filesystems we've gotten back that you'd never expect - and then it just becomes the norm.
I love the crazy bug reports that really push the boundaries of our capabilities.
That's an attitude that reiserfs and btrfs never had, and when I am confident that it is 100% rock solid and bulletproof I'll be lifting the experimental label.
But I do badly need more funding, this would go better with a real team behind it. Right now I'm trying to find the money to bring Alan Huang on full time; he's fresh out of school but very sharp and motivated, and he's already been doing excellent work.
If anyone can help with that, hit me up :)
If someone's testing against these versions, adding their fixes and patches, stuff like this will break things for users. He can't assume all users will be regular desktop users, even on an experimental area of the code.
Things like 'RC' have meaning. Meaning that has been there for years. He can develop on a separate tree and users that want it can use it. This is used all over.
Honestly Kent, this continuing baseless fearmongering from you about btrfs is absolutely disgusting.
It costs you. I was initially very interested in bcachefs, but I will never spend my time testing it or contributing to it as long as you continue behave this way. I'm certain there are many many others who would nominally be very interested, but feel the same way I do.
Your filesystem charitably gets 0.001% the real world testing btrfs does. To claim it is more reliable than btrfs is ignorant and naive.
Maybe it actually is more reliable in the real world (press X to doubt...), but you can't possibly know yet, and you won't know for a long time.
This is the antiphrasis of good project management and stability.
No you want to avoid a static target in a dynamic environment that is unmaintained (such as an experimental fs in the kernel tree).
If it's static and unsupported. You'd end up failing to be run this to recover disks using ryzen9 processors that requires a minimum kernel version where the API/abi have drifted so far that the old module won't compile or import.
If you can't afford to get your hands dirty and hack at the API changing if this has such a bus factor. DON'T USE IT.
Frankly the argument you're making is the other side of stick with ext2 since it works. It's probably going to die soon and frankly unless there's a community to support it. (such as zfs, or ext4 in the kernel, or CEPH in hpc corporate spaces)
If your product is so great its it's own advert. If it has problems spend the limited person power fixing it not attacking the opposition, this is what ciq have done, do better.
This isn't baseless fearmongering, this is "did you think through the failure modes when you were designing the basics".
This stuff comes up over, and over, and over.
Engineering things right matters, and part of that absolutely is comparing and documenting approaches and solutions to see what went right and what went wrong.
This isn't a popularity contest, and this isn't high school where we form into cliques and start slinging insults.
Come up with facts, documentation, analysis. That's what we do. I'm tired of these threads degenerating into flamewars.
The files are still loaded from a compressed cpio archive though, and because of the initrd legacy that file is still called initrd in most distributions.
I am not a filesystems guy, but I was disappointed when I realized that btrfs did not have a good design for ENOSPC handling.
So I'm curious, does bcachefs design for a benign failure mode when out of space?
I feel you're not answering the question, nor are you presenting any case in favor of forcing an exceptional release process for an unstable feature.
The "detrimental" claim is also void of any reason or substance. It's not to it's users as users know better than rolling out experimental file systems for critical data, and those hypothetical users who somehow are really really interested in using bleeding edge will already be building their own kernel for this reason alone. Both scenarios don't require this code to be in the kernel, let alone exceptional release processes.
> Ideally they'd be developed outside the kernel until they are perfect, but Kent addresses this in his LWN comment: There is no funding/time to make that ideal scenario possible.
It is clear then that the code should be pulled from the kernel. If it isn't impossible to maintain a feature with the regular release process, everyone stands to benefit by not shipping code that is impossible to stabilize.
The only aspect of "experimental" that matters is what it means to the release process. If you can't meet that bar then debating semantics won't help either.
And by the way, the patch thread clearly stresses a) loss of data, b) the patch trying to sneak under the maintenance radar new features. That is the definition of unstable in anyone's book.
Now: I am not saying that bcachefs is yet trouble free enough for widespread deployment, we're still seeing cases where repair needs fairly minor work, but the filesystem may be offline while we get that fixed.
OTOH we also recover, regularly, from absolutely crazy scenarios involving hardware failure: flaky controllers, lightning strikes, I've seen cases where it looked like a head crashed - took out a whole bunch of btree nodes in similar LBA ranges.
IOW: the fundamentals are very solid, but keep checking back if you're wondering when it'll be ready for widespread deployment.
Milestones to watch for: - 3 months with zero data loss or downtime events: I think we may get this soon, knock on wood - stable backports starting: 6.17, knock on wood (or maybe we'll be out of the kernel and coming up with our own plan, who knows) - weird application specific bugs squashed: these have been on the back burner, but there's some weird stuff to get to still (e.g. weird unlink behavior that affects docker and go builds, and the Rust people just reported something odd when building certain packages with mold).
And yes, we've always handled -ENOSPC gracefully.
This is precisely the mistake you are making: conflating egregious types who post code as "good engineers". They are not. They are incompetent.
There is no engineering activity that is not driven.by teams. Being able to work effecticely in a team environment is therefore a very basic and critical skill. Those who are unable to work in a team environment are lacking a very basic and critical skill. Those who fail at this basic skill to the point they are dubbed as "toxic" end up sabotaging your whole team, needlessly creating problems to everyone around them, and preventing any collaboration to take place.
If this problem is introduced by a single team member, it is in everyone's best interest to just cut the cancer.
And I have to imagine that those that are technical enough to not only use Linux, but use an experimental non-default file system, and those that don’t have backups of their data, is a vanishingly small group.
So no, I actually disagree —- it’s not part of the job.
So again we arrive at the same place: this data recovery tool is not worth the drama.
It’s a feature, not a bug fix, and an incredibly unimportant one at that _at this stage of development_. If bcachefs weren’t experimental and were widely used, it would be a different story —- I’d probably be in favor of bending the rules to get it in there faster. But that just isn’t where we are right now.
I am inclined to agree. However, I have one thought to the contrary. When a mechanical drive is failing, you tend to have debris inside the drive hitting the platters, causing damage that creates more debris, accelerating the drive’s eventual death, with read errors becoming increasingly common while it happens. When those are included in averages, the 10^14 might very well be accurate. I have not done any rigorous analysis to justify this thought and I do not have the data to be able to do that analysis. It is just something that occurs to me that might justify the 10^14 figure.
As for ensuring data integrity, I cannot talk about other CoW filesystems, but ZFS has an atomic transaction commit that relies on CoW. In ZFS, your changes either happened or they did not happen. The entire file system is a giant merkle tree and every change requires that all nodes of the tree up to the root be rewritten. To minimize the penalty of CoW, these changes are aggregated into transaction groups that are then committed atomically. Thus, you simultaneously have both the old and new versions available, plus possible more than just 1 old version. ZFS will start recycling space after a couple transaction group commits, but often, you can go further back in its history if needed after some catastrophic event, although ZFS makes no solid guarantee of this (until you fiddle with module parameter settings to prevent reclaim from being so aggressive).
If it counts for anything, I have hundreds of commits in OpenZFS, so I am fairly familiar with how ZFS works internally.
https://github.com/openzfs/zfs/blob/6af8db61b1ea489ade2d5344...
Small writes on ZFS are ACID. If ZFS made large writes ACID, large writes could block the transaction group commit for arbitrarily long periods, which is why it does not. Just imagine writing a 1PB file. It would likely take a long time (days?) and it is just not reasonable to block the transaction group commit until it finishes.
That said, for your example, you will often have all of the writes go into the same transaction group commit, such that it becomes ACID, but this is not a strict guarantee. The maximum atomic write size on ZFS is 32MB, assuming alignment. If the write is not aligned to the record size, it will be smaller, as per:
https://github.com/openzfs/zfs/blob/6af8db61b1ea489ade2d5344...
It's a signal to users - "watch out if you're going to use this thing"
I would rather make sure this path was never hit for rc, to minimize the damage. The fact alone that it didn’t pop up until late in the 6.15 cycle could hint at some specific circumstances the bug manifested on, and those could be described and avoided.
And I think there could be a mediocre way to get by until the next merge window in which a superior solution could be presented.
I don’t want to sound as if I’m an expert in how to do VFS, because I’m not. I’m, however, an expert in how to be “correcter than others” which has cost me getting kicked out of jobs before. I hope I have learned better since, and at the time I have been very, very stubborn (they wouldn’t have kicked me out otherwise).
There is this bit when working with others that you will likely go with solutions you deem more mediocre than the theoretically best solution, or that experienced people will say “no” to your ideas or solutions and you accept it instead of seeking quarrel in spite of them obviously not understanding you (spoiler: this is not so). But if you show that you’re willing to work with other as a single unit, you will be listened to, appreciated, and concessions will be made for you too. This is not necessarily about the kernel, it’s about working in a team in general.
I don’t have a dog in this fight, but I’ve been “that guy” before and I regret it took me that long to realize this and mend my ways. Hope it doesn’t keep happening to you.
I have lost years of my work due to not having proper backups. I know the pain.
And I totally get you feel responsible for the data loss and want to help your users, I'm like that too with my code.
But this is an experiment. It's right there in the name.
If this feature really is as needed as you claim here, then getting it into the kernel is a mere side-issue.
In that case, your #1 priority should be to fix whatever is causing such users to install and use bcachefs without having a recovery plan that they have verified, and get existing users on the same level.
Because not doing so would be a huge disservice to those users who don't know better, and at worst borderline exploitative.
Writing recovery software is part of the job. Forcing it into the kernel to save users who are clearly not in any shape or form competent enough to partake in your experiment is definitely not part of the job.
Finding yourself in this position means something has gone very, very wrong.
https://btrfs.readthedocs.io/en/latest/Status.html
The amount of "mostly OK" and still an "unstable" RAID6 implementation. Not going to trust a file system with "mostly OK" device replace. Anecdotally, you can search the LKML and here for tons of data loss stories.