Message ID | 155552786671.20411.6442426840435740050.stgit@magnolia (mailing list archive) |
---|---|
Headers | show |
Series | vfs: make immutable files actually immutable | expand |
On Wed, Apr 17, 2019 at 12:04:26PM -0700, Darrick J. Wong wrote: > Hi all, > > The chattr(1) manpage has this to say about the immutable bit that > system administrators can set on files: > > "A file with the 'i' attribute cannot be modified: it cannot be deleted > or renamed, no link can be created to this file, most of the file's > metadata can not be modified, and the file can not be opened in write > mode." > > Given the clause about how the file 'cannot be modified', it is > surprising that programs holding writable file descriptors can continue > to write to and truncate files after the immutable flag has been set, > but they cannot call other things such as utimes, fallocate, unlink, > link, setxattr, or reflink. > > Since the immutable flag is only settable by administrators, resolve > this inconsistent behavior in favor of the documented behavior -- once > the flag is set, the file cannot be modified, period. The manual page leaves the case undefined, though the word 'modified' can be interpreted in the same sense as 'mtime' ie. modifying the file data. The enumerated file operations that don't work on an immutable file suggest that it's more like the 'ctime', ie. (state) changes are forbidden. Tthe patchset makes some sense, but it changes the semantics a bit. From 'not changed but still modified' to 'neither changed nor modified'. It starts to sound like a word game, but I think both are often used interchangeably in the language. See the changelog of 1/8 where you used them in the other meaning regarding ctime and mtime. I personally doubt there's a real use of the undefined case, though something artificial like 'a process opens a fd, sets up file in a very specific way, sets immutable and hands the fd to an unprivileged process' can be made up. The overhead of the new checks seems to be small so performance is not the concern here. Overall, I don't see a strong reason for either semantics. As long as it's documented possibly with some of the corner cases described in more detail, fine.