zeppelin_embed/vfs/

mod.rs

//! Synchronous virtual-filesystem seam for crash and fault injection.

#[cfg(any(test, feature = "test-seams"))]
pub mod crash;
#[cfg(any(test, feature = "test-seams"))]
pub mod fault;

use std::fs::{File, OpenOptions};
use std::io::{IoSlice, Read, Seek, SeekFrom, Write};
#[cfg(unix)]
use std::os::fd::AsRawFd;
use std::path::{Path, PathBuf};
use std::sync::Arc;
use std::sync::atomic::{AtomicU64, Ordering};

/// Storage-ordering primitive requested by a persisted commit.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub enum SyncKind {
    /// Order prior writes before later writes without requesting durable media.
    ///
    /// Linux has no barrier-only primitive, so `ordered` durability degrades
    /// to `fdatasync` on Linux.
    Barrier,
    /// Flush prior writes through the platform's durable-media primitive.
    Full,
}

/// One open append-only file owned by a WAL writer.
pub trait VfsFile: Send {
    /// Appends every supplied byte without reopening the path.
    fn append(&mut self, bytes: &[u8]) -> std::io::Result<()>;
    /// Appends one logical group from discontiguous owned buffers.
    ///
    /// The default preserves compatibility for test filesystems. Production
    /// files override it with one vectored write sequence, so WAL visibility
    /// and persistence can share encoded record storage without coalescing it.
    fn append_vectored(&mut self, buffers: &mut [IoSlice<'_>]) -> std::io::Result<()> {
        for buffer in buffers {
            self.append(buffer)?;
        }
        Ok(())
    }
    /// Synchronizes this already-open handle using the named primitive.
    fn sync(&self, kind: SyncKind) -> std::io::Result<()>;
}

/// Minimal synchronous filesystem operations used by persisted commits.
pub trait Vfs: Send + Sync {
    /// Supplies an optional counter for mmap-backed segment payload accesses.
    #[doc(hidden)]
    fn segment_data_read_counter(&self) -> Option<Arc<AtomicU64>> {
        None
    }
    /// Admits a store directory, creating missing ancestors when requested,
    /// and reports whether the resulting path is a directory.
    fn ensure_directory(&self, path: &Path, create: bool) -> std::io::Result<bool>;
    /// Exclusively creates one directory. Existing paths are preserved and
    /// reported as `AlreadyExists`; no ancestor is adopted implicitly.
    fn create_directory(&self, _path: &Path) -> std::io::Result<()> {
        Err(std::io::Error::new(
            std::io::ErrorKind::Unsupported,
            "VFS does not support exclusive directory creation",
        ))
    }
    /// Removes one empty directory. A non-empty directory is an error.
    fn remove_directory(&self, _path: &Path) -> std::io::Result<()> {
        Err(std::io::Error::new(
            std::io::ErrorKind::Unsupported,
            "VFS does not support directory removal",
        ))
    }
    /// Opens an existing path and returns its byte length.
    fn open(&self, path: &Path) -> std::io::Result<u64>;
    /// Opens an existing path as an owned file suitable for memory mapping.
    ///
    /// Test filesystems without a native file descriptor return
    /// [`std::io::ErrorKind::Unsupported`]. Consequently `MemoryVfs`,
    /// `CrashVfs`, and memory-backed `FaultVfs` images cannot publish or reopen
    /// a sealed segment; use a path-backed VFS when exercising that lifecycle.
    fn open_for_map(&self, path: &Path) -> std::io::Result<File>;
    /// Reads an entire file.
    fn read(&self, path: &Path) -> std::io::Result<Vec<u8>>;
    /// Reads at most `length` bytes beginning at `offset`.
    fn read_range(&self, path: &Path, offset: u64, length: usize) -> std::io::Result<Vec<u8>>;
    /// Creates or truncates a file and writes all bytes.
    fn write(&self, path: &Path, bytes: &[u8]) -> std::io::Result<()>;
    /// Removes a suffix without copying the retained file bytes.
    fn truncate(&self, _path: &Path, _length: u64) -> std::io::Result<()> {
        Err(std::io::Error::new(
            std::io::ErrorKind::Unsupported,
            "VFS does not support suffix truncation",
        ))
    }
    /// Exclusively creates a file and writes every byte without replacing any
    /// existing path. A failed write may leave a new partial file; callers must
    /// retain it as an uncommitted artifact until safe cleanup is established.
    /// Filesystems without atomic exclusive creation fail explicitly.
    fn create_new(&self, _path: &Path, _bytes: &[u8]) -> std::io::Result<()> {
        Err(std::io::Error::new(
            std::io::ErrorKind::Unsupported,
            "VFS does not support exclusive file creation",
        ))
    }
    /// Opens or creates one file for handle-oriented append-only writes.
    fn open_append(&self, path: &Path) -> std::io::Result<Box<dyn VfsFile>>;
    /// Atomically renames one path over another according to platform semantics.
    fn rename(&self, from: &Path, to: &Path) -> std::io::Result<()>;
    /// Synchronizes a file or directory using an explicit platform primitive.
    fn sync(&self, path: &Path, kind: SyncKind) -> std::io::Result<()>;
    /// Lists direct children of a directory.
    fn list(&self, directory: &Path) -> std::io::Result<Vec<PathBuf>>;
    /// Visits direct children one at a time without collecting the directory.
    fn for_each_direct_child(
        &self,
        directory: &Path,
        visitor: &mut dyn FnMut(&Path) -> std::io::Result<()>,
    ) -> std::io::Result<()>;
    /// Deletes one file.
    fn delete(&self, path: &Path) -> std::io::Result<()>;
}

/// Standard-library-backed production filesystem.
#[derive(Clone, Copy, Debug, Default)]
pub struct StdVfs;

struct StdVfsFile(File);

impl VfsFile for StdVfsFile {
    fn append(&mut self, bytes: &[u8]) -> std::io::Result<()> {
        #[cfg(feature = "test-seams")]
        file_test_support::event(file_test_support::FileEvent::BeforeAppend)?;
        self.0.write_all(bytes)
    }

    fn append_vectored(&mut self, buffers: &mut [IoSlice<'_>]) -> std::io::Result<()> {
        #[cfg(feature = "test-seams")]
        file_test_support::event(file_test_support::FileEvent::BeforeAppend)?;
        let mut remaining = buffers;
        while !remaining.is_empty() {
            let written = self.0.write_vectored(remaining)?;
            if written == 0 {
                return Err(std::io::Error::new(
                    std::io::ErrorKind::WriteZero,
                    "failed to append vectored WAL group",
                ));
            }
            IoSlice::advance_slices(&mut remaining, written);
        }
        Ok(())
    }

    fn sync(&self, kind: SyncKind) -> std::io::Result<()> {
        #[cfg(feature = "test-seams")]
        file_test_support::event(file_test_support::FileEvent::BeforeSync)?;
        sync_file(&self.0, kind)?;
        #[cfg(feature = "test-seams")]
        file_test_support::event(file_test_support::FileEvent::AfterSync)?;
        Ok(())
    }
}

impl Vfs for StdVfs {
    fn truncate(&self, path: &Path, length: u64) -> std::io::Result<()> {
        OpenOptions::new().write(true).open(path)?.set_len(length)
    }

    fn create_new(&self, path: &Path, bytes: &[u8]) -> std::io::Result<()> {
        let mut file = OpenOptions::new().create_new(true).write(true).open(path)?;
        file.write_all(bytes)
    }

    fn ensure_directory(&self, path: &Path, create: bool) -> std::io::Result<bool> {
        if create {
            std::fs::create_dir_all(path)?;
        }
        Ok(std::fs::metadata(path)?.is_dir())
    }

    fn create_directory(&self, path: &Path) -> std::io::Result<()> {
        std::fs::create_dir(path)
    }

    fn remove_directory(&self, path: &Path) -> std::io::Result<()> {
        std::fs::remove_dir(path)
    }

    fn open(&self, path: &Path) -> std::io::Result<u64> {
        Ok(File::open(path)?.metadata()?.len())
    }

    fn open_for_map(&self, path: &Path) -> std::io::Result<File> {
        File::open(path)
    }

    fn read(&self, path: &Path) -> std::io::Result<Vec<u8>> {
        let mut file = File::open(path)?;
        let length = usize::try_from(file.metadata()?.len()).map_err(|_| {
            std::io::Error::new(std::io::ErrorKind::InvalidData, "file length exceeds usize")
        })?;
        let mut bytes = Vec::with_capacity(length);
        file.read_to_end(&mut bytes)?;
        Ok(bytes)
    }

    fn read_range(&self, path: &Path, offset: u64, length: usize) -> std::io::Result<Vec<u8>> {
        let mut file = File::open(path)?;
        file.seek(SeekFrom::Start(offset))?;
        let mut bytes = vec![0_u8; length];
        let mut filled = 0_usize;
        while filled < length {
            let remaining = bytes.get_mut(filled..).ok_or_else(|| {
                std::io::Error::new(std::io::ErrorKind::InvalidData, "invalid read range")
            })?;
            let read = file.read(remaining)?;
            if read == 0 {
                break;
            }
            filled = filled.saturating_add(read);
        }
        bytes.truncate(filled);
        Ok(bytes)
    }

    fn write(&self, path: &Path, bytes: &[u8]) -> std::io::Result<()> {
        let mut file = OpenOptions::new()
            .create(true)
            .truncate(true)
            .write(true)
            .open(path)?;
        file.write_all(bytes)
    }

    fn open_append(&self, path: &Path) -> std::io::Result<Box<dyn VfsFile>> {
        let file = OpenOptions::new().create(true).append(true).open(path)?;
        Ok(Box::new(StdVfsFile(file)))
    }

    fn rename(&self, from: &Path, to: &Path) -> std::io::Result<()> {
        std::fs::rename(from, to)
    }

    fn sync(&self, path: &Path, kind: SyncKind) -> std::io::Result<()> {
        sync_path(path, kind)
    }

    fn list(&self, directory: &Path) -> std::io::Result<Vec<PathBuf>> {
        std::fs::read_dir(directory)?
            .map(|entry| entry.map(|value| value.path()))
            .collect()
    }

    fn for_each_direct_child(
        &self,
        directory: &Path,
        visitor: &mut dyn FnMut(&Path) -> std::io::Result<()>,
    ) -> std::io::Result<()> {
        for entry in std::fs::read_dir(directory)? {
            let entry = entry?;
            visitor(&entry.path())?;
        }
        Ok(())
    }

    fn delete(&self, path: &Path) -> std::io::Result<()> {
        std::fs::remove_file(path)
    }
}

/// Byte/call-counting decorator used by deterministic open-cost gates.
pub struct CountingVfs<V> {
    inner: V,
    counters: Arc<CountingVfsCounters>,
    read_path_filter: Option<PathBuf>,
}

#[derive(Default)]
struct CountingVfsCounters {
    open_calls: AtomicU64,
    open_for_map_calls: AtomicU64,
    read_calls: AtomicU64,
    filtered_read_calls: AtomicU64,
    read_bytes: AtomicU64,
    segment_bytes_read: Arc<AtomicU64>,
    write_calls: AtomicU64,
    bytes_written: AtomicU64,
    append_calls: AtomicU64,
    bytes_appended: AtomicU64,
    rename_calls: AtomicU64,
    delete_calls: AtomicU64,
    enumerate_calls: AtomicU64,
    barrier_sync_calls: AtomicU64,
    full_sync_calls: AtomicU64,
    handle_barrier_sync_calls: AtomicU64,
    handle_full_sync_calls: AtomicU64,
}

struct CountingVfsFile {
    inner: Box<dyn VfsFile>,
    counters: Arc<CountingVfsCounters>,
}

impl VfsFile for CountingVfsFile {
    fn append(&mut self, bytes: &[u8]) -> std::io::Result<()> {
        self.inner.append(bytes)?;
        self.counters.append_calls.fetch_add(1, Ordering::Relaxed);
        self.counters
            .bytes_appended
            .fetch_add(bytes.len() as u64, Ordering::Relaxed);
        Ok(())
    }

    fn append_vectored(&mut self, buffers: &mut [IoSlice<'_>]) -> std::io::Result<()> {
        let bytes = buffers
            .iter()
            .fold(0_usize, |total, buffer| total.saturating_add(buffer.len()));
        self.inner.append_vectored(buffers)?;
        self.counters.append_calls.fetch_add(1, Ordering::Relaxed);
        self.counters
            .bytes_appended
            .fetch_add(bytes as u64, Ordering::Relaxed);
        Ok(())
    }

    fn sync(&self, kind: SyncKind) -> std::io::Result<()> {
        self.inner.sync(kind)?;
        match kind {
            SyncKind::Barrier => &self.counters.handle_barrier_sync_calls,
            SyncKind::Full => &self.counters.handle_full_sync_calls,
        }
        .fetch_add(1, Ordering::Relaxed);
        Ok(())
    }
}

impl<V> CountingVfs<V> {
    /// Wraps a filesystem with zeroed counters.
    #[must_use]
    pub fn new(inner: V) -> Self {
        Self {
            inner,
            counters: Arc::new(CountingVfsCounters::default()),
            read_path_filter: None,
        }
    }

    /// Counts successful reads of one exact path separately from global reads.
    #[must_use]
    pub fn with_read_path_filter(mut self, path: impl Into<PathBuf>) -> Self {
        self.read_path_filter = Some(path.into());
        self
    }

    /// Returns the wrapped filesystem.
    #[must_use]
    pub const fn inner(&self) -> &V {
        &self.inner
    }

    /// Runs a test probe with mmap-backed segment payload accounting enabled.
    #[doc(hidden)]
    pub fn audit_segment_data_reads<T>(&self, operation: impl FnOnce() -> T) -> T {
        let _audit = crate::segment::reader::install_data_read_audit(Some(Arc::clone(
            &self.counters.segment_bytes_read,
        )));
        operation()
    }

    /// Returns observed open calls.
    #[must_use]
    pub fn open_calls(&self) -> u64 {
        self.counters.open_calls.load(Ordering::Relaxed)
    }

    /// Returns observed mapping-file open calls.
    #[must_use]
    pub fn open_for_map_calls(&self) -> u64 {
        self.counters.open_for_map_calls.load(Ordering::Relaxed)
    }

    /// Returns observed read calls.
    #[must_use]
    pub fn read_calls(&self) -> u64 {
        self.counters.read_calls.load(Ordering::Relaxed)
    }

    /// Returns successful reads matching the configured exact path.
    #[must_use]
    pub fn filtered_read_calls(&self) -> u64 {
        self.counters.filtered_read_calls.load(Ordering::Relaxed)
    }

    /// Returns exact bytes returned by reads.
    #[must_use]
    pub fn read_bytes(&self) -> u64 {
        self.counters.read_bytes.load(Ordering::Relaxed)
    }

    /// Returns exact bytes read from immutable segment paths.
    #[must_use]
    pub fn segment_bytes_read(&self) -> u64 {
        self.counters.segment_bytes_read.load(Ordering::Relaxed)
    }

    /// Returns observed whole-file write calls.
    #[must_use]
    pub fn write_calls(&self) -> u64 {
        self.counters.write_calls.load(Ordering::Relaxed)
    }

    /// Returns exact bytes supplied to successful whole-file writes.
    #[must_use]
    pub fn bytes_written(&self) -> u64 {
        self.counters.bytes_written.load(Ordering::Relaxed)
    }

    /// Returns observed successful handle append calls.
    #[must_use]
    pub fn append_calls(&self) -> u64 {
        self.counters.append_calls.load(Ordering::Relaxed)
    }

    /// Returns exact bytes supplied to successful handle appends.
    #[must_use]
    pub fn bytes_appended(&self) -> u64 {
        self.counters.bytes_appended.load(Ordering::Relaxed)
    }

    /// Returns observed successful rename calls.
    #[must_use]
    pub fn rename_calls(&self) -> u64 {
        self.counters.rename_calls.load(Ordering::Relaxed)
    }

    /// Returns observed successful delete calls.
    #[must_use]
    pub fn delete_calls(&self) -> u64 {
        self.counters.delete_calls.load(Ordering::Relaxed)
    }

    /// Returns observed streaming directory-enumeration calls.
    #[must_use]
    pub fn enumerate_calls(&self) -> u64 {
        self.counters.enumerate_calls.load(Ordering::Relaxed)
    }

    /// Returns path-based barrier synchronization calls.
    #[must_use]
    pub fn barrier_sync_calls(&self) -> u64 {
        self.counters.barrier_sync_calls.load(Ordering::Relaxed)
    }

    /// Returns path-based full synchronization calls.
    #[must_use]
    pub fn full_sync_calls(&self) -> u64 {
        self.counters.full_sync_calls.load(Ordering::Relaxed)
    }

    /// Returns open-handle barrier synchronization calls.
    #[must_use]
    pub fn handle_barrier_sync_calls(&self) -> u64 {
        self.counters
            .handle_barrier_sync_calls
            .load(Ordering::Relaxed)
    }

    /// Returns open-handle full synchronization calls.
    #[must_use]
    pub fn handle_full_sync_calls(&self) -> u64 {
        self.counters.handle_full_sync_calls.load(Ordering::Relaxed)
    }

    /// Resets all counters without changing the wrapped filesystem.
    pub fn reset(&self) {
        self.counters.open_calls.store(0, Ordering::Relaxed);
        self.counters.open_for_map_calls.store(0, Ordering::Relaxed);
        self.counters.read_calls.store(0, Ordering::Relaxed);
        self.counters
            .filtered_read_calls
            .store(0, Ordering::Relaxed);
        self.counters.read_bytes.store(0, Ordering::Relaxed);
        self.counters.segment_bytes_read.store(0, Ordering::Relaxed);
        self.counters.write_calls.store(0, Ordering::Relaxed);
        self.counters.bytes_written.store(0, Ordering::Relaxed);
        self.counters.append_calls.store(0, Ordering::Relaxed);
        self.counters.bytes_appended.store(0, Ordering::Relaxed);
        self.counters.rename_calls.store(0, Ordering::Relaxed);
        self.counters.delete_calls.store(0, Ordering::Relaxed);
        self.counters.enumerate_calls.store(0, Ordering::Relaxed);
        self.counters.barrier_sync_calls.store(0, Ordering::Relaxed);
        self.counters.full_sync_calls.store(0, Ordering::Relaxed);
        self.counters
            .handle_barrier_sync_calls
            .store(0, Ordering::Relaxed);
        self.counters
            .handle_full_sync_calls
            .store(0, Ordering::Relaxed);
    }
}

impl<V: Vfs> Vfs for CountingVfs<V> {
    fn truncate(&self, path: &Path, length: u64) -> std::io::Result<()> {
        self.inner.truncate(path, length)
    }
    fn create_new(&self, path: &Path, bytes: &[u8]) -> std::io::Result<()> {
        self.inner.create_new(path, bytes)?;
        self.counters.write_calls.fetch_add(1, Ordering::Relaxed);
        self.counters
            .bytes_written
            .fetch_add(bytes.len() as u64, Ordering::Relaxed);
        Ok(())
    }

    fn segment_data_read_counter(&self) -> Option<Arc<AtomicU64>> {
        Some(Arc::clone(&self.counters.segment_bytes_read))
    }

    fn create_directory(&self, path: &Path) -> std::io::Result<()> {
        self.inner.create_directory(path)
    }
    fn remove_directory(&self, path: &Path) -> std::io::Result<()> {
        self.inner.remove_directory(path)
    }
    fn ensure_directory(&self, path: &Path, create: bool) -> std::io::Result<bool> {
        self.inner.ensure_directory(path, create)
    }

    fn open(&self, path: &Path) -> std::io::Result<u64> {
        self.counters.open_calls.fetch_add(1, Ordering::Relaxed);
        self.inner.open(path)
    }

    fn open_for_map(&self, path: &Path) -> std::io::Result<File> {
        self.counters
            .open_for_map_calls
            .fetch_add(1, Ordering::Relaxed);
        self.inner.open_for_map(path)
    }

    fn read(&self, path: &Path) -> std::io::Result<Vec<u8>> {
        let bytes = self.inner.read(path)?;
        self.counters.read_calls.fetch_add(1, Ordering::Relaxed);
        if self.read_path_filter.as_deref() == Some(path) {
            self.counters
                .filtered_read_calls
                .fetch_add(1, Ordering::Relaxed);
        }
        self.counters
            .read_bytes
            .fetch_add(bytes.len() as u64, Ordering::Relaxed);
        if is_segment_path(path) {
            self.counters
                .segment_bytes_read
                .fetch_add(bytes.len() as u64, Ordering::Relaxed);
        }
        Ok(bytes)
    }

    fn read_range(&self, path: &Path, offset: u64, length: usize) -> std::io::Result<Vec<u8>> {
        let bytes = self.inner.read_range(path, offset, length)?;
        self.counters.read_calls.fetch_add(1, Ordering::Relaxed);
        if self.read_path_filter.as_deref() == Some(path) {
            self.counters
                .filtered_read_calls
                .fetch_add(1, Ordering::Relaxed);
        }
        self.counters
            .read_bytes
            .fetch_add(bytes.len() as u64, Ordering::Relaxed);
        if is_segment_path(path) {
            self.counters
                .segment_bytes_read
                .fetch_add(bytes.len() as u64, Ordering::Relaxed);
        }
        Ok(bytes)
    }

    fn write(&self, path: &Path, bytes: &[u8]) -> std::io::Result<()> {
        self.inner.write(path, bytes)?;
        self.counters.write_calls.fetch_add(1, Ordering::Relaxed);
        self.counters
            .bytes_written
            .fetch_add(bytes.len() as u64, Ordering::Relaxed);
        Ok(())
    }

    fn open_append(&self, path: &Path) -> std::io::Result<Box<dyn VfsFile>> {
        Ok(Box::new(CountingVfsFile {
            inner: self.inner.open_append(path)?,
            counters: Arc::clone(&self.counters),
        }))
    }

    fn rename(&self, from: &Path, to: &Path) -> std::io::Result<()> {
        self.inner.rename(from, to)?;
        self.counters.rename_calls.fetch_add(1, Ordering::Relaxed);
        Ok(())
    }

    fn sync(&self, path: &Path, kind: SyncKind) -> std::io::Result<()> {
        self.inner.sync(path, kind)?;
        match kind {
            SyncKind::Barrier => &self.counters.barrier_sync_calls,
            SyncKind::Full => &self.counters.full_sync_calls,
        }
        .fetch_add(1, Ordering::Relaxed);
        Ok(())
    }

    fn list(&self, directory: &Path) -> std::io::Result<Vec<PathBuf>> {
        self.inner.list(directory)
    }

    fn for_each_direct_child(
        &self,
        directory: &Path,
        visitor: &mut dyn FnMut(&Path) -> std::io::Result<()>,
    ) -> std::io::Result<()> {
        self.inner.for_each_direct_child(directory, visitor)?;
        self.counters
            .enumerate_calls
            .fetch_add(1, Ordering::Relaxed);
        Ok(())
    }

    fn delete(&self, path: &Path) -> std::io::Result<()> {
        self.inner.delete(path)?;
        self.counters.delete_calls.fetch_add(1, Ordering::Relaxed);
        Ok(())
    }
}

fn is_segment_path(path: &Path) -> bool {
    path.file_name()
        .and_then(|name| name.to_str())
        .is_some_and(|name| name.starts_with("segment-") && name.ends_with(".zseg"))
}

/// Synchronizes whatever `path` names, reopening it with the access the
/// platform requires.
///
/// On Unix a read-only descriptor is enough for `fsync`, including on a
/// directory. Windows refuses both: `FlushFileBuffers` needs write access, and
/// a directory handle additionally needs `FILE_FLAG_BACKUP_SEMANTICS`, so the
/// Windows arm does its own open rather than reusing `File::open`.
#[cfg(not(windows))]
fn sync_path(path: &Path, kind: SyncKind) -> std::io::Result<()> {
    sync_file(&File::open(path)?, kind)
}

#[cfg(windows)]
fn sync_path(path: &Path, _kind: SyncKind) -> std::io::Result<()> {
    // Windows exposes no ordering-only barrier, so `Barrier` is served by the
    // same full flush as `Full`: stronger and more expensive than a barrier,
    // never cheaper. See `tasks/evidence/windows/w02-durable-publication-protocol.md`.
    crate::sys::windows::sync_path(path)
}

#[cfg(any(target_os = "macos", target_os = "ios"))]
fn sync_file(file: &File, kind: SyncKind) -> std::io::Result<()> {
    let result = match kind {
        SyncKind::Barrier => crate::sys::darwin::barrier_fsync(file.as_raw_fd()),
        SyncKind::Full => crate::sys::darwin::full_fsync(file.as_raw_fd()),
    };
    result.map_err(std::io::Error::other)
}

#[cfg(all(unix, not(any(target_os = "macos", target_os = "ios"))))]
fn sync_file(file: &File, kind: SyncKind) -> std::io::Result<()> {
    let result = unsafe {
        // SAFETY: both calls accept an owned live descriptor and report failures through errno.
        match kind {
            SyncKind::Barrier => libc::fdatasync(file.as_raw_fd()),
            SyncKind::Full => libc::fsync(file.as_raw_fd()),
        }
    };
    if result == -1 {
        Err(std::io::Error::last_os_error())
    } else {
        Ok(())
    }
}

/// Flushes an already-open handle without reopening its path.
///
/// Both `SyncKind`s map to `FlushFileBuffers`: Windows has no barrier-only
/// primitive, so `Ordered` durability is served by the full flush. That is
/// documented as stronger and more expensive, not as a cheap barrier.
#[cfg(windows)]
fn sync_file(file: &File, _kind: SyncKind) -> std::io::Result<()> {
    crate::sys::windows::flush_file(file)
}

#[cfg(not(any(unix, windows)))]
fn sync_file(_: &File, kind: SyncKind) -> std::io::Result<()> {
    Err(std::io::Error::new(
        std::io::ErrorKind::Unsupported,
        format!("{kind:?} synchronization requires a supported file-descriptor platform"),
    ))
}

#[cfg(test)]
mod tests {
    use std::path::Path;

    use super::crash::MemoryVfs;
    use super::{CountingVfs, SyncKind, Vfs};

    #[test]
    fn crash_support_api_gate_is_identical_in_debug_and_release() {
        let source = include_str!("mod.rs");
        let gate = source
            .lines()
            .zip(source.lines().skip(1))
            .find_map(|(line, next)| (next == "pub mod crash;").then_some(line));
        assert_eq!(gate, Some("#[cfg(any(test, feature = \"test-seams\"))]"));
    }

    #[test]
    #[allow(clippy::expect_used)]
    fn counting_vfs_counts_every_mutation_and_sync_kind_exactly() {
        let inner = MemoryVfs::new();
        inner
            .insert("/source", b"abc".to_vec())
            .expect("seed source");
        let counting = CountingVfs::new(inner);

        assert_eq!(counting.open(Path::new("/source")).expect("open"), 3);
        assert_eq!(counting.read(Path::new("/source")).expect("read"), b"abc");
        assert_eq!(
            counting
                .read_range(Path::new("/source"), 1, 2)
                .expect("range"),
            b"bc"
        );
        counting.write(Path::new("/write"), b"1234").expect("write");
        let mut handle = counting
            .open_append(Path::new("/append"))
            .expect("open append");
        handle.append(b"12").expect("first append");
        handle.append(b"345").expect("second append");
        handle.sync(SyncKind::Barrier).expect("handle barrier");
        handle.sync(SyncKind::Full).expect("handle full");
        counting
            .sync(Path::new("/write"), SyncKind::Barrier)
            .expect("path barrier");
        counting
            .sync(Path::new("/write"), SyncKind::Full)
            .expect("path full");
        counting
            .rename(Path::new("/write"), Path::new("/renamed"))
            .expect("rename");
        counting.delete(Path::new("/renamed")).expect("delete");

        assert_eq!(
            [
                counting.open_calls(),
                counting.read_calls(),
                counting.read_bytes(),
                counting.write_calls(),
                counting.bytes_written(),
                counting.append_calls(),
                counting.bytes_appended(),
                counting.rename_calls(),
                counting.delete_calls(),
                counting.barrier_sync_calls(),
                counting.full_sync_calls(),
                counting.handle_barrier_sync_calls(),
                counting.handle_full_sync_calls(),
            ],
            [1, 2, 5, 1, 4, 2, 5, 1, 1, 1, 1, 1, 1],
            "CountingVfs must expose every exact VFS and open-handle cost"
        );
        counting.reset();
        assert_eq!(
            [
                counting.open_calls(),
                counting.read_calls(),
                counting.read_bytes(),
                counting.write_calls(),
                counting.bytes_written(),
                counting.append_calls(),
                counting.bytes_appended(),
                counting.rename_calls(),
                counting.delete_calls(),
                counting.barrier_sync_calls(),
                counting.full_sync_calls(),
                counting.handle_barrier_sync_calls(),
                counting.handle_full_sync_calls(),
            ],
            [0; 13],
            "reset must zero decorator and shared-handle counters"
        );
    }
}

#[cfg(feature = "test-seams")]
pub mod file_test_support;