Module atomic_io 

Durable atomic file / directory operations for checkpoint-class writes.

The tmp-file + rename pattern is atomic only if both the file data and the containing directory entry reach stable storage in the correct order. On ext4 / XFS the rename metadata op can reach disk before the data pages backing the tmp file — a power loss between the write and the next checkpoint then leaves a correctly-named file containing zeros.

atomic_write_fsync is the single helper all checkpoint-class writers go through so the ordering (write → sync_data → rename → fsync_dir) is enforced in one place. atomic_swap_dirs_fsync does the same for directory-level swaps (rename old-dir → backup, rename new-dir → old-dir).

read_checkpoint_dontneed pairs with the write helper on the read side: checkpoint bytes are consumed once (deserialized into the in-memory index) and then superseded. Leaving them in the page cache wastes memory needed by hot workloads.

Functions

atomic_swap_dirs_fsync

Atomically swap a directory: rename(live, backup); rename(staged, live), fsyncing the parent directory once both renames have completed.

atomic_write_fsync

Atomically write bytes to dst via a tmp file with full durability.

fsync_directory

Fsync a directory to ensure file creation/deletion metadata is durable.

read_checkpoint_dontneed

Read a checkpoint file and advise the kernel to drop its pages from the page cache.