Hi everyone, During lockdown, I've been working on teaching[1] XFS how to swap a range of extents with a different range of extents in another file in such a way that we're always guaranteed that the swap operation completes, no matter how many extents have to be swapped. This guarantee is durable evne if the filesystem goes down midway through the operation, since log recovery will finish the job. In other words, we finally have a mechanism for userspace programs to update files atomically without resorting (a) to a pile of fsync and rename spaghetti or (b) creating an O_TMPFILE and using weird procfs linkat tricks to wire that into the filesystem. Better yet, other programs with open file descriptors will see the new contents immediately. So the basic user story is that userspace opens a file and decides to write to it. The user program then creates an O_TMPFILE. If the program only wishes to update a subset of the file, it can FICLONE the original file's contents, or it can rewrite the file from scratch. When the update is finished, the program can fsync the temp file and then call FISWAPRANGE to commit the changes to the original file. The program can supply a [cm]time to the kernel, which can use that to detect that the original file has changed and that the update should not be applied. Note that this is /not/ the hardware-assisted directio atomic writes that have also been suggested on fsdevel. That interface (if the storage vendors ever show up with hardware) only supports atomic writes on a limited number of possibly discontiguous LBAs, whereas this software mechanism will work on any size file with any number of blocks. As a side benefit, this also gives XFS the ability to rebuild xattr and directory data in a temporary file and cleanly swap the clean contents over when the entire recreation step has finished. I haven't yet put this out on fsdevel because I've been a bit busy putting together XFS summit ideas. :/ --D [1] https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=atomic-file-updates
