Inspired by Dgraph's mmap file implementation in ristretto.

A file-backed memory map exposes the kernel's view of an inode as a &[u8]/&mut [u8]. That makes it easy to reach for, but it also means UB the moment another actor truncates, unlinks, or rewrites the file out from under the mapping — SIGBUS on Unix, mapping detachment on Windows, silent torn reads in either. fmmap raises a safe API over memmapix by treating those concerns as first-class:

• Auto-acquired advisory lock on every constructor — exclusive on writable maps, shared on read-only / COW maps. Aliased writable mappings of the same file (and mut-then-COW) are rejected up front.
• Best-effort path-reuse mitigation on deletion. Identity is captured at open and re-checked before every unlink so a file someone else has swapped in at the path won't be silently deleted. POSIX uses (st_dev, st_ino); Windows uses (volumeSerial, fileIndex) from GetFileInformationByHandle (via windows-sys, no nightly required). This is not an absolute guarantee — see the path-reuse limitations below.
• Pre-validated mapping ranges. Constructors reject offset/len overflow, ranges past EOF, and effective lengths > isize::MAX before any destructive set_len runs, so an invalid Options never zeroes or extends an existing file.
• Crash-durable unlink. The parent directory is pinned by a handle opened before remove_file, then fsynced through that same handle. Failed-fsync retries fsync the same handle (not a freshly-opened parent), so a parent rename between unlink and fsync can't direct the durability to the wrong inode.
• Reentrant-safe lock methods. LockFileEx deadlocks on the same Windows handle; lock / lock_shared short-circuit when the desired state is already held. The lock methods take &mut self so single-owner serialization is enforced by the borrow checker.
• Poison-safe truncate / freeze. A failed truncate marks the wrapper poisoned; subsequent reads return &[] and writes/flushes/freezes return Err rather than handing back an anonymous-mapped placeholder pretending to be the original file.

std plus tokio and smol are first-class. The async surface is built from the same set of macros, so adding a new runtime is small and mechanical — see fmmap/src/disk/{tokio,smol}_impl.rs.

Identity-checked deletion is built on the strongest atomic primitives each platform exposes; what's left is a small, documented set of irreducible races.

POSIX: probe + unlink + parent fsync are all bound to the same parent fd via rustix's fstatat + unlinkat. A parent rename mid-operation can't direct the unlink or fsync to a different directory than the one we verified. The original file's open-file description is held alive (via fcntl(F_DUPFD_CLOEXEC) or, in the tokio wrapper, tokio::fs::File::into_std()) across probe + unlink, so the kernel cannot recycle (dev, ino) to a fresh file in the window. Identity capture itself is allocation-free (fstat on a BorrowedFd), so EMFILE has no path to defeating the identity check.

Windows: probe and unlink are bound to a single handle. The handle is opened with DELETE | FILE_SHARE_* and FILE_FLAG_OPEN_REPARSE_POINT; we re-verify identity and refuse reparse points on that handle, then issue SetFileInformationByHandle(FileDispositionInfoEx) with POSIX_SEMANTICS | IGNORE_READONLY_ATTRIBUTE. Older Windows / FAT32 fall back to FileDispositionInfo after a ReOpenFile widens access to clear FILE_ATTRIBUTE_READONLY (using FILE_ATTRIBUTE_NORMAL as the cleared-state sentinel — Windows treats 0 as "no change"). Identity is captured directly via GetFileInformationByHandle on a borrowed HANDLE — no DuplicateHandle, no fd alloc.

API contract: explicit remove() (and drop_remove()) only returns Ok if fmmap itself observed the unlink succeed in the parent it then fsynced. NotFound from the probe or unlink is never converted into a durable-success retry — the wrapper stays in NeedsUnlink and surfaces the error, even when the inode's nlink has dropped to 0 (which can't distinguish "unlink in our parent" from "external rename + unlink elsewhere"). Drop's best-effort cleanup still fsyncs the parent in the common case, but the API doesn't promise durability we can't verify.

• One-syscall TOCTOU on POSIX. Between fstatat and unlinkat — both bound to the same parent fd — there's still a single-syscall window where the entry could be replaced. Closing this needs an inode-bound unlinkat primitive POSIX doesn't expose. The window is dramatically narrower than the handle-drop-to-retry window the identity check does close, but it's not zero.
• External rename + unlink elsewhere. A concurrent actor can rename our file into a different directory and unlink it there. The inode's nlink drops to 0 but our parent's fsync doesn't commit their unlink. fmmap detects this only as "the file is gone" and surfaces NotFound; under that scenario, callers who need crash-durability should serialize external mutations or fsync the relevant parents themselves.
• Smol consuming drop_remove(self) under EMFILE. smol's async-fs::File exposes no into_std(), so the inode pin is a fcntl_dupfd_cloexec of the underlying fd. Under fd pressure the dup fails, drop_remove returns Err deterministically (no hidden Drop-time retry), and the file remains on disk. Callers can recover via std::fs::remove_file(path) directly or AsyncMmapFileMut::remove(&mut self) which preserves self for an explicit retry. Tokio's into_std() allocates no fd so this limitation doesn't apply on tokio.

If your threat model includes an active local adversary, do not rely on identity-checked delete for safety — perform the cleanup yourself with whatever atomic primitives your platform provides.

• file-backed memory maps with auto-locked construction
• read-only / copy-on-write / mutable / executable maps
• identity-checked deletion bound to a single kernel-verified handle (POSIX fstatat+unlinkat on a parent fd; Windows SetFileInformationByHandle(FileDispositionInfoEx) on a DELETE | FILE_SHARE_DELETE handle); see Design for residual races
• inode pin across probe + unlink (POSIX F_DUPFD_CLOEXEC or tokio into_std) — defends against (dev, ino) recycling on tmpfs / small-id filesystems
• crash-durable unlink with pre-opened parent fsync (same handle reused on retry)
• symlink / reparse-point refusal at the same syscall as the identity probe (POSIX AT_SYMLINK_NOFOLLOW, Windows FILE_FLAG_OPEN_REPARSE_POINT)
• readonly-file delete on Windows (FileDispositionInfoEx with IGNORE_READONLY_ATTRIBUTE, legacy FileDispositionInfo fallback for pre-1607)
• pre-validated mapping ranges (rejects past-EOF and > isize::MAX before any destructive set_len)
• poison-safe truncate / freeze / freeze_exec
• synchronous and asynchronous flushing
• reader / writer adapters with byteorder + seek
• dozens of file I/O util functions
• stack support (MAP_STACK on Unix)
• tokio
• smol

fmmap requires Rust 1.75 or later.

• std

  [dependencies]
  fmmap = "0.5"

• tokio

  [dependencies]
  fmmap = { version = "0.5", default-features = false, features = ["tokio"] }

• smol

  [dependencies]
  fmmap = { version = "0.5", default-features = false, features = ["smol"] }

The sync feature is on by default.

This crate is 100% documented, see docs.rs for examples.