I tried using this a while back and found it was not widely available. You need coreutils version 9.1 or later for this, many distros do not ship this.
You're misinterpreting the title. The author didn't intend "Unix" to literally mean only the official AT&T/TheOpenGroup UNIX® System to the exclusion of Linux.
The first sentence of "UNIX-like" makes that clear : >This is a catalog of things UNIX-like/POSIX-compliant operating systems can do atomically,
Further down, he then mentions some Linux specifics : >fcntl(fd, F_GETLK, &lock), fcntl(fd, F_SETLK, &lock), and fcntl(fd, F_SETLKW, &lock) . [...] There is a “mandatory locking” mode but Linux’s implementation is unreliable as it’s subject to a race condition.
Differing philosophies of how to interpret titles. Prescriptive vs Descriptive language.[0]
There can be different usages of the word "Unix":
#1: Unix is a UNIX(tm) System V descendent. More emphasis that the kernel needs to be UNIX. In this strict definition, you get the common reminder that "Linux is not a Unix!"
#2: "Unix" as a loose generic term for a family of o/s that looks/feels like Unix. This perspective includes using an o/s that has userland Unix utilities like cat/grep/awk. Sometimes deliberately styled as asterisk "*nix" or a suffix-qualifier "Unix-like" but often just written as a naked "Unix".
A Prescriptivist says the author's title is "incorrect". On the other hand, a Descriptivist looks at the whole content of the article -- notices the text has a lot of Linux specific info such as fcntl(,F_GETLEASE/F_SETLEASE), and every hyperlink to a man page url points to https://linux.die.net/man/ , etc -- and thus determines that the author is using "Unix"(#2) in the looser way that can include some Linux idiosyncrasies.
"Unix" instead of "*nix" as a generic term for Linux is not uncommon. Another example article where the authors use the so-called incorrect "Unix" in the title even though it's mostly discussing Linux CUPS instead of Solaris : https://www.evilsocket.net/2024/09/26/Attacking-UNIX-systems...
Which, FWIW, doesn't mean Linux. AFAIK there is no Linux distro that's fully compliant, even before you worry about the specifics of whether it's certified as compliant.
>POSIX-compliant Which, FWIW, doesn't mean Linux. AFAIK there is no Linux distro that's fully compliant
I read author's use of "POSIX-compliant" as a loose and fuzzy family category rather than an exhaustive and authoritative reference on 100% strict compliance. Therefore, the author mentioning non-100%-compliant Linux is ok.
There seems to be 2 different expectations and interpretations of what the article is about.
- (1) article is attempting to be a strict intersection of all Unix-like systems that conform to official UNIX POSIX API. I didn't think this was a reasonable interpretation since we can't be sure the author actually verified/tested other POSIX-like systems such as FreeBSD, HP-UX, IBM AIX, etc.
- (2) article is a looser union of operating systems and can also include idiosyncracies of certain systems like Linux that the author is familiar with that don't apply to all other UNIX systems. I think some readers don't realize that all the author's citations to man pages point to Linux specific urls at : https://linux.die.net/man/
The ggp's (amstan) additional comment about renameat2(,,,,RENAME_EXCHANGE) is useful info and is consistent with interpretation (2).
If the author really didn't want Linux to be lumped in with "POSIX-like", it seems he would avoid linux.die.net and instead point to something more of a UNIX standard such as: https://unix.org/apis.html
AFAIK you don't even want to be POSIX-compliant unless having a sticker means more to you than being reasonable. Most projects knowingly steer away from compliance (and certifying compliance is probably also expensive)
Which of these are not guaranteed by the POSIX specification? It’s been a while since I studied it, but if I recall correctly the ones mentioned in the article are guaranteed.
The mmap/msync one is incorrect I believe? (Correct me if I am wrong).
msync() sync content in memory back to _disk_. But multiple processes mapping the same file always see the same content (barring memory consistency, caching, etc.) already. Unless the file is mapped with MAP_PRIVATE.
Yeah I agree that one isn't very clear, perhaps the idea is to use `msync()` as a barrier to achieve consistent ordering of the writes without having to handle that yourself with more complex primitives. But then, they do mention some of those primitives at the bottom of the article, so it's hard to say what exactly the idea is.
mmap/msync is behavior is also very platform specific. On some systems (like AIX, at least older versions), even without msync, memory mapped data is synced back to disk periodically.
I worked on a code base that was portable between Linux, AIX, and some other Unix flavors. mmap/msync was a source of bugs. Just imagine your system running for days, never syncing any data to disk... then someone pulls the plug. Where'd my data go? Even worse, it happened "in production" at a beta site. Fortunately we had a way to recover data from a log.
Nice catalog. One subtle thing I’ve found in building deterministic, stateless systems is that atomic filesystem and memory operations are the only way to safely compute or persist secrets without locks. Combining rename/link/O_EXCL patterns with ephemeral in-memory buffers ensures that sensitive data is never partially written to disk, which reduces race conditions and side-channel exposure in multi-process workflows.
Even though it can do some things atomically, it only does with one file at a time, and race conditions are still possible because it only does one operation at a time (even if you are only need one file). Some of these are helpful anyways, such as O_EXCL, but it is still only one thing at a time which can cause problems in some cases.
What else it does not do is a transaction with multiple objects. That is why, I would design a operating system, that you can do a transaction with multiple objects.
Windows had APIs for this sort of thing added in Vista, but they're now deprecating it "due to its complexity and various nuances which developers need to consider":
In some cases, you can start by using the "at" functions (openat...) to work on a directory tree. If you have your logical "locking" done at the top-level of the tree, it might be a fine option.
In some other cases, I've used a pattern where I used a symlink to folders. The symlink is created, resolved or updated atomically, and all I need is eventual consistency.
That last case was to manage several APT repository indices. The indices were constantly updated to publish new testing or unstable releases of software and machines in the fleet were regularly fetching the repository index. The APT protocol and structure being a bit "dumb" (for better or worse) requires you to fetch files (many of them) in the reverse order they are created, which leads to obvious issues like the signature is updated only after the list of files is updated, or the list of files is created only after the list of packages is created.
Long story short, each update would create a new folder that's consistent, and a symlink points to the last created folder (to atomically replace the folder as it was not possible to swap them), and a small HTTP server would initiate a server side session when the first file is fetched and only return files from the same index list, and everything is eventually consistent, and we never get APT complaining about having signature or hash mismatches. The pivotal component was indeed the atomicity of having a symlink to deal with it, as the Java implementation didn't have access to a more modern "openat" syscall, relative to a specific folder.
Depending on metadata cache behavior configuration, if the system is powered off immediately after the first command, then that could indeed happen I think.
As to whether it’s technically possible for it to happen on a system that stays on, I’m not sure, but it’s certainly vanishingly rare and likely requires very specific circumstances—not just a random race condition.
Uhh, if the system powers off immediately after the first command (mv a b), the second command (mv c d) would never run. So where would d come from if the command that created it never executed?
Sure, but splitting "atomic" operations across a reboot is an interesting design choice. Surely upon reboot you would re-try the first `mv a b` before doing other things.
I don't think that's happening in practice, but 1) it may not be specified and 2) What you say could well be the persisted state after a machine crash or power loss. In particular if those files live in different directories.
You can remedy 2) by doing fsync() on the parent directory in between. I just asked ChatGPT which directory you need to fsync. It says it's both, the source and the target directory. Which "makes sense" and simplifies implementations, but it means the rename operation is atomic only at runtime, not if there's a crash in between. It think you might end up with 0 or 2 entries after a crash if you're unlucky.
If that's true, then for safety maybe one should never rename across directories, but instead do a coordinated link(source, target), fsync(target_dir), unlink(source), fsync(source_dir)
All you need for this to occur is the window where both renames occurs overlap. A system polling to check if a, b, c, and d exist while the renames are happening might find all four of them.
Assuming that the two `mv` commands are run in sequence, there shouldn't be any possibility for a and d to be observed "at once" (i.e. first d and then afterwards still a, by a single process).
I'm almost certain what the OP meant was if the commands were run synchronously (ie: from 2 different shells or as `mv a b &; mv c d`) yes there is a possibility that a and d exist (eg: On a busy system where neither of the 2 commands can be immediately scheduled and eventually the second one ends up being scheduled before the first)
Or to go a level deeper, if you have 2 occurrences of rename(2) from the stdlibc ...
rename('a', 'b');
rename('c', 'd');
...and the compiler decides on out of order execution or optimizing by scheduling on different cpus, you can get a and d existing at the same time.
The reason it won't happen in the example you posted is the shell ensures the atomicity (by not forking the second mv until the wait() on the first returns)
nitpick, it should be `touch a c & mv a b & mv c d` as `&;` returns `bash: syntax error near unexpected token `;'`. I always find this oddly weird, but that would not be the first pattern in BASH that is.
`inotifywait` actually sees them in order, but nothing ensure that it's that way.
$ inotifywait -m /tmp
/tmp/ MOVED_FROM a
/tmp/ MOVED_TO b
/tmp/ MOVED_FROM c
/tmp/ MOVED_TO d
`stat` tells us that the timestamps are equal as well.
$ stat b d | grep '^Change'
Change: 2026-02-06 12:22:55.394932841 +0100
Change: 2026-02-06 12:22:55.394932841 +0100
However, speeding things up changes it a bit.
Given
$ (
set -eo pipefail
for i in {1..10000}
do
printf '%d ' "$i"
touch a c
mv a b &
mv c d &
wait
rm b d
done
)
1 2 3 4 5 6 .....
And with `inotifywait` I saw this when running it for a while.
$ inotifywait -m -e MOVED_FROM,MOVED_TO /tmp > /tmp/output
cat /tmp/output | xargs -l4 | sort | uniq -c
9104 /tmp/ MOVED_FROM a /tmp/ MOVED_TO b /tmp/ MOVED_FROM c /tmp/ MOVED_TO d
896 /tmp/ MOVED_FROM c /tmp/ MOVED_TO d /tmp/ MOVED_FROM a /tmp/ MOVED_TO b
I don't think any part of your post is right. Aside from NFS, there should not be filesystems where this doesn't work. If there are, those are just bugs. The flags you mentioned are not required or relevant. Setting the fd offset to the end of the file atomically is the entire purpose of O_APPEND.
You are projecting imaginary features onto O_APPEND and then hypothesizing that your imaginary features might not work.
POSIX says that for a file opened with O_APPEND "the file offset shall be set to the end of the file prior to each write." That's it. That's all it does.
Unsure. Aren’t there filesystems which make O_APPEND less durable than it’s specified to be, which might be interpreted to adversely affect atomicity? Could that be it?
Not technically related to atomicity, but I was looking for a way to do arbitrary filesystem operations based on some condition (like adding a file to a directory, and having some operation be performed on it). The usual recommendation for this is to use inotify/watchman, but something about it seems clunky to me. I want to write a virtual filesystem, where you pass it a trigger condition and a function, and it applies the function to all files based on the trigger condition. Does something like this exist?
The challenge with that approach is memory: trigger conditions, if added irresponsibly, can result in unbounded memory and (depending on implementation) potentially linear performance degradation of filesystem operations as well. Unbounded kernel memory growth leads to stability or security risks.
That tradeoff is at the root of why most notify APIs are either approximate (events can be dropped) or rigidly bounded by kernel settings that prevent truly arbitrary numbers of watches. fanotify and some implementations of kqueue are better at efficiently triggering large recursive watches, but that’s still just a mitigation on the underlying memory/performance tradeoffs, not a full solution.
Sounds half baked. What context does this function run in? Is it an interpreted language or an executable that you provide?
Inotify is the way to shovel these events out of the kernel, then userspace process rules apply. It's maybe not elegant from your pov, but it's simple.
I know this is trivial to do programmatically, but I was looking for a way this will be handled by the filesystem. For instance, if I have some processes generating log files, and I have a script that converts them to html, I wanted the script to be called every time a log file is updated, without having a daemon running in the background to monitor the directory, just some filesystem mount. This would have made some deployments easier.
Thanks, I didn't find this when I was looking for a solution for my problem. This is pretty much the exact solution for my usecase, though for some reason inotify feels more complicated than some kind of filesystem mount solution for me.
This document being from 2010 is, of course, missing the C11/C++11 atomics that replaced the need for compiler intrinsics or non portable inline asm when "operating on virtual memory".
With that said, at least for C and C++, the behavior of (std::)atomic when dealing with interprocess interactions is slightly outside the scope of the standard, but in practice (and at least recommended by the C++ standard) (atomic_)is_lock_free() atomics are generally usable between processes.
I don't do devops/sysadmin anymore, so this would have been before the age of k8s for everything. But I once interviewed for a company hiring specifically because their deployment process lasted hours, and rollbacks even longer.
In the interview when they were describing this problem, I asked why the didn't just put all of the new release in a new dir, and use symlinks to roll forward and backwards as needed. They kind of froze and looked at each other and all had the same 'aha' moment. I ended up not being interested in taking the job, but they still made sure to thank me for the idea which I thought was nice.
Not that I'm a genius or anything, it's something I'd done previously for years, and I'm sure I learned it from someone else who'd been doing it for years. It's a very valid deployment mechanism IMO, of course depending on your architecture.
Worked pretty well in production systems, serving huge amount of RPS (like ~5-10k/s) running on a LAMP stack monolith in five different geographical regions.
Just git branch (one branch per region because of compliance requirements) -> branch creates "tar.gz" with predefined name -> automated system downloads the new "tar.gz", checks release date, revision, etc. -> new symlink & php (serverles!!!) graceful restart and ka-b00m.
Rollbacks worked by pointing back to the old dir & restart.
not surprised about the chrome part, but pretty shocked at the phone OS part. I know APFS migration was done in this way, but wouldn't storage considerations for this be massive?
Not really, because only the OS core is swapped in this way. Apps and data live in their own partitions/subvolumes, which are mutable and shared between OS versions.
The OS core is deployed as a single unit and is a few GB in size, pretty small when internal storage is into the hundreds of GB.