> Using the "folder" mental model leads to exactly the bugs in TFA. "Oh, the root folder can only be accessed by root, so if I put a file _inside_ that folder it will be protected". Thats not how it works.
I don't really see the logic that way. Even if it was "inside", this permission failure would still happen. And directories, if used properly, can protect a 777 file from being changed. The error in the mental model is in how permissions work, not how files are organized.
> No, because you can have multiple links in multiple directories all pointing to the same file.
Well I put the word "access" there to try to be clearer. If you as system designer wanted to, you could have the path you take to a file affect permissions, even with the 'directory' model and multiple hard links. Heck, you could have permissions be on links instead of on files. YOLO.
> ... directories, if used properly, can protect a 777 file from being changed. The error in the mental model is in how permissions work, not how files are organized.
There is a second possible misconception that I did touch on in my last paragraph, but didn't spell out. On Unix, the permissions check is done when you open the file, not when you perform the read or write. This means that a user who cannot currently open the file (because directory permissions mean they have no way to get to the inode) can nonetheless alter it now if they opened it when they could. So you could rename the file from the attacker's directory into your installer's private directory, verify its cryptographic signature, but then the attacker injects their malware into the file before you start copying, and you install the malware.
Because the two common types of locks on Unix (BSD and POSIX record) are advisory, you can't just lock that file against writers before you check the signature. This is in contrast to Windows, where you can't even rename or delete the file if someone else has it open.
How do I do that? fstat(2) is no help on macOS, and even if it were would return false positives from things like backup and content indexers ("Time Machine" and "Spotlight" on macOS).
Say every hard link to the file is a descendant of a directory that blocks traversal. No subdirectories of those are already open, and the file is not already open. That keeps it safe, right? If there's any loopholes in that, they could be closed.
Ehr, no. Again, there are no files in the conventional sense in UNIX file systems. There are collections of disk blocks pointed to by an i-node and one or more directory entries pointing to an i-node. It is possible to have an i-node with 0 directory entries linking to it as well as multiple unrelated (i.e. not hard links but truly disjoint directory entries) directory entries referencing the same i-node; both are treated as file system errors by fsck and will be fixed up at the next fsck run. Yet, both scenarios can be easily reproduced (without corrupting the file system!) in a file system debugger and live on for a while.
> […] a descendant of a directory that blocks traversal. No subdirectories of those […]
Directory entries in a UNIX file system do not ascend nor descend, they are linked into one or more directories whether they form a hierarchy or not.
A directory might be «protected» by, say, 700 permissions obscuring a particular directory entry, but if a hard link to the same i-node exists in a another unrelated directory outside the current hierarchy that has more permissive access, say 755, access to data blocks referenced to by an i-node has already leaked out.
The other reply already covered the definition of hard links. It's a directory entry that points to an inode.
And file system corruption is definitely a loophole.
> Directory entries in a UNIX file system do not ascend nor descend, they are linked into one or more directories whether they form a hierarchy or not.
All the filesystems I'm sufficiently aware of insist on directories being a tree. Every entry except the special .. descends in that tree. And each hard link is in a specific directory.
> if a hard link to the same i-node exists in a another unrelated directory outside the current hierarchy that has more permissive access
> And file system corruption is definitely a loophole.
Zero directory entries pointing to an i-node is not a file system corruption as it neither corrupts the file system nor breaks the file system semantics; it is possible to have a garbage collector running in the background to mop up orphaned i-nodes with the file system remaining fully operational at the same time.
Distinct i-nodes pointing to the same block allocation, on the other hand, are a security loop hole and create consistency problems. Whether they cause the file system corruption or not is a matter of an academic debate, though.
> All the filesystems I'm sufficiently aware of insist on directories being a tree. Every entry except the special .. descends in that tree. And each hard link is in a specific directory.
It is possible to design and write a file system implementation that will retain the UNIX file systems semantics of i-nodes and directory entries whilst remaining completely flat (i.e. no directory hierarchies, just «.»). Such a file sysem would be impractical for most use cases today but is easily possible, and such file systems had been a commonplace before the UNIX file system arrival.
Earlier on, you had mentioned: «If there's any loopholes in that, they could be closed». The example below (which is perfectly legit and does not contain semantic loopholes), which of directories does «file.txt» belong in or descends from/ascends into: 1) «a/b/c», 2) «d/e/f/g», 3) «.», 4) all of them? Which of the three directories is more specific and why, and what about future hard links?
$ mkdir -p a/b/c
$ mkdir -p d/e/f/g
$ echo 'I am a file' >file.txt
$ chown 0:0 file.txt
$ chmod 666 file.txt
$ ln file.txt a/b/c
$ ln file.txt d/e/f/g
$ sudo chown 0:0 a
$ sudo chmod 700 a
$ ls -l a
ls: cannot open directory 'a': Permission denied
$ ls -l d/e/f/g/file.txt
-rw-rw-rw- 3 root wheel 12 Aug 14 23:59 d/e/f/g/file.txt
$ ls -l ./file.txt
-rw-rw-rw- 3 root wheel 12 Aug 14 23:59 ./file.txt
$ echo 'Anyone can access me' >./file.txt
$ cat ./file.txt
Anyone can access me
> It is possible to design and write a file system implementation that will retain the UNIX file systems semantics of i-nodes and directory entries whilst remaining completely flat (i.e. no directory hierarchies, just «.»). Such a file sysem would be impractical for most use cases today but is easily possible, and such file systems had been a commonplace before the UNIX file system arrival.
Yeah, it would also be possible to design a system that doesn't enforce permissions.
The challenge here is whether you can make a reasonable design that's secure. Not whether any design would be secure; that's self-obviously false. Anyone doing the designing can choose not to use a special bespoke filesystem.
But I don't see how your described filesystem would cause problems. The directory entries are still descendants of the directories they are in. Apply the rest of the logic and those files are secure. It's easier, really, when you don't have to worry about subdirectories. If subdirectories don't exist, they can't be open.
> Earlier on, you had mentioned: «If there's any loopholes in that, they could be closed». The example below (which is perfectly legit and does not contain semantic loopholes), which of directories does «file.txt» belong in or descends from/ascends into: 1) «a/b/c», 2) «d/e/f/g», 3) «.», 4) all of them? Which of the three directories is more specific and why, and what about future hard links?
The file is not in a specific directory. Links to the file are in `pwd`, a/b/c, and d/e/f/g. "Being in" is the same as "descending from".
If you secure `pwd` (and nothing is already open), then all three hard links will be secured.
Or if you remove the hard link in `pwd`, and secure g (and nothing is already open), then the file will be secured.
"./a" descends from ".", one hard link to the file descends from ".", "./a/b" descends from "./a", "./a/b/c" descends from "./a/b", one hard link to the file descends from "./a/b/c". Plus the same for d/e/f/g, plus every transitive descent like "./a/b" descending from "." I hope that's what you mean by "more specific"?
If future hard links are made, then they follow the same rules. If any hard link is not secured, then the file is not secured. And a user without access to the file cannot make a new hard link to the file.
> Yeah, it would also be possible to design a system that doesn't enforce permissions.
It is even easier than that: one only has to simply detach the disk and reattach it to another UNIX box to gain access to any file as the file system itself is defenceless and offers no protection from the physical access to its on-disk layout. File system encryption is the only solution that makes physical impractical or at least convoluted.
And, since UNIX file systems delegate permissons checks to the kernel via the VFS, it is also possible for a person with nefarious intentions to modify the file system code to make it always return 777 for any i-node being accessed through it, find a local zero day exploit, load the rogue file system kernel module and remount file system(s) to bypass the permission enforcement in the kernel.
The reverse is also true: if the kernel and the file system support access control lists, standard UNIX file permissions become largely meaningless, and it becomes possible to grant or revoke access to/from a file owned by root with 600 permissions to an arbitrary user/group only. Using the same example from above:
$ cat ./file.txt
Anyone can access me
$ sudo /bin/chmod +a "group:staff deny write" ./file.txt
$ /bin/ls -le ./file.txt
-rw-rw-rw-+ 3 root wheel 21 14 Aug 23:59 ./file.txt
0: group:staff deny write
$ echo 'No-one from the staff group can access me any longer' >./file.txt
zsh: permission denied: ./file.txt
$ id
uid=NNNNNN(morally.bold.mollusk) gid=MMMMMM(staff) groups=MMMMMM(staff),[… redacted …]
$ ls -la ./file.txt
-rw-rw-rw-+ 3 root wheel 21 Aug 15 17:28 ./file.txt
> The challenge here is whether you can make a reasonable design that's secure.
Indeed, rarely can security be bolted on with any measurable success, and a system can be secure only if it is secure by design. But security is also a game of the constant juggling of trade-offs that may or may not be acceptable in a particular use case. Highly secure designs are [nearly always] hostile to users and are a tremendous nuisance in the daily use. The UNIX answer to security is delegation of responsibilities: «I, UNIX, will do a reasonable job on keeping the system secure, but the onus is on you, user, to excercise the due diligence, and – oh, by the way – here is a shotgun to shoot yourself in the foot (and injure bystanders as a bonus) if you, the user, are negligent about keeping your data secure».
> "./a" descends from ".", one hard link to the file descends from ".", "./a/b" descends from "./a", "./a/b/c" descends from "./a/b", one hard link to the file descends from "./a/b/c". Plus the same for d/e/f/g, plus every transitive descent like "./a/b" descending from "." I hope that's what you mean by "more specific"?
The point I was trying to make was that specificness is a purely logical concept. In the aforementioned example, there are 3x directory entries at 3x arbitrary locations and any of them can be used to access the data referenced to via an i-node. Once a file is opened using either of those three directory entries, it is not possible to trace the open file descriptor back to a specific directory entry. Therefore, none of the three directory entries is more specific than the others – they are all equal.
> The point I was trying to make was that specificness is a purely logical concept. In the aforementioned example, there are 3x directory entries at 3x arbitrary locations and any of them can be used to access the data referenced to via an i-node. Once a file is opened using either of those three directory entries, it is not possible to trace the open file descriptor back to a specific directory entry. Therefore, none of the three directory entries is more specific than the others – they are all equal.
I see.
Then I would agree that every path is equally specific.
But I never wanted to trace a file descriptor back to a specific directory entry. The question that matters is whether all the directory entries for a file are in secure locations. That treats them all equally.
Also, part of the scenario I laid out is that the file is not open to begin with. (If you were to try to check if the file is open, that's outside the scenario, but also shouldn't care what directory entry was used.)
I don't really see the logic that way. Even if it was "inside", this permission failure would still happen. And directories, if used properly, can protect a 777 file from being changed. The error in the mental model is in how permissions work, not how files are organized.
> No, because you can have multiple links in multiple directories all pointing to the same file.
Well I put the word "access" there to try to be clearer. If you as system designer wanted to, you could have the path you take to a file affect permissions, even with the 'directory' model and multiple hard links. Heck, you could have permissions be on links instead of on files. YOLO.