vnfs 0.0.6

//! The vectorized filesystem API: the [`VecFs`] trait plus the shared types it
//! operates on.
//!
//! This mirrors the `tc_api.h` vectorized NFSv4 client surface, but as a
//! trait so any filesystem can implement it: the NFSv4.1 client
//! (`crate::nfs::NfsVecFs`) and a `std::fs`-backed dummy
//! ([`crate::dummy_vecfs::DummyVecFs`]). Vector operations take Rust slices.
//!
//! # Path and name representation
//!
//! Paths are native Unix [`Path`]s. An absolute path starts with `/` and is
//! resolved against the filesystem root (for NFS, the export root), while a
//! relative path is resolved against the client's current working directory
//! (see [`VecFs::abs_path`] and [`VecFs::vf_path`]). `PathBuf`/`OsStr`
//! preserve arbitrary filename bytes; UTF-8 conversion is a convenience for
//! callers that need it.

use std::path::{Path, PathBuf};

use crate::error::RpcError;
#[cfg(unix)]
use crate::path::path_from_bytes;

// ---------------------------------------------------------------------------
// Constants (mirroring tc_api.h)
// ---------------------------------------------------------------------------

/// Errors that have no filesystem status (transport / client side).
pub const VF_ERR_RPC: u32 = 0xFFFF_FFFF;
/// Requested feature is not implemented by this backend.
pub const VF_ERR_UNSUPPORTED: u32 = 0xFFFF_FFFE;

/// Generic errno-style error codes (they coincide with the NFS4ERR codes for
/// the same conditions, which the NFS backend reports).
pub const ERR_NOENT: u32 = 2;
pub const ERR_EBADF: u32 = 9;
pub const ERR_EXIST: u32 = 17;
pub const ERR_NOTDIR: u32 = 20;
pub const ERR_ISDIR: u32 = 21;
pub const ERR_INVAL: u32 = 22;
pub const ERR_ACCES: u32 = 13;

/// NFSv4 wire type codes (NF4*), used to decode/encode [`VfType`].
pub const NF4REG: u32 = 1;
pub const NF4DIR: u32 = 2;
pub const NF4BLK: u32 = 3;
pub const NF4CHR: u32 = 4;
pub const NF4LNK: u32 = 5;
pub const NF4SOCK: u32 = 6;
pub const NF4FIFO: u32 = 7;

// ---------------------------------------------------------------------------
// Result types
// ---------------------------------------------------------------------------

/// The failure of one operation in a vectorized call.
///
/// Either a filesystem status failure attributable to a specific operation
/// index, or a transport / client-side failure (where the index is
/// best-effort: backends report 0 when the failure cannot be attributed).
#[derive(Debug, Clone, PartialEq, Eq)]
#[non_exhaustive]
pub enum VfError {
    /// A filesystem status failure; `err_no` is errno-style or an NFS4ERR
    /// code, mirroring the C `tc_res` struct.
    Op { index: usize, err_no: u32 },
    /// A transport / client-side failure with a human-readable message; there
    /// is no filesystem status ([`err_no`](VfError::err_no) reports
    /// [`VF_ERR_RPC`]). `index` is `None` when the failure cannot be
    /// attributed to any operation.
    Transport {
        index: Option<usize>,
        message: String,
    },
}

impl VfError {
    pub fn failure(index: usize, err_no: u32) -> VfError {
        VfError::Op { index, err_no }
    }

    /// A transport / client-side failure. `index` is best-effort; pass `None`
    /// when the failure cannot be attributed to a specific operation.
    pub fn transport(index: impl Into<Option<usize>>, message: impl Into<String>) -> VfError {
        VfError::Transport {
            index: index.into(),
            message: message.into(),
        }
    }

    pub fn unsupported(index: usize) -> VfError {
        VfError::failure(index, VF_ERR_UNSUPPORTED)
    }

    /// The operation index this error refers to (best-effort for transport
    /// failures; 0 when unknown).
    pub fn index(&self) -> usize {
        match self {
            VfError::Op { index, .. } => *index,
            VfError::Transport { index, .. } => index.unwrap_or(0),
        }
    }

    /// The operation index as an `Option`; `None` only for a transport
    /// failure that cannot be attributed to any operation.
    pub fn index_opt(&self) -> Option<usize> {
        match self {
            VfError::Op { index, .. } => Some(*index),
            VfError::Transport { index, .. } => *index,
        }
    }

    /// The filesystem status code, or [`VF_ERR_RPC`] for transport failures.
    pub fn err_no(&self) -> u32 {
        match self {
            VfError::Op { err_no, .. } => *err_no,
            VfError::Transport { .. } => VF_ERR_RPC,
        }
    }

    /// Whether this is a transport / client-side failure (no filesystem
    /// status).
    pub fn is_transport(&self) -> bool {
        matches!(self, VfError::Transport { .. })
    }

    /// Convert a low-level [`RpcError`] into a `VfError`, attributing the
    /// failure to `index` in the caller's operation array. `index` is the
    /// caller's operation index, which may differ from the compound-internal
    /// op index; for transport failures the index is best-effort. The
    /// transport message (if any) is preserved.
    pub fn from_rpc(e: RpcError, index: impl Into<Option<usize>>) -> VfError {
        if e.is_transport() {
            VfError::Transport {
                index: index.into(),
                message: e.message,
            }
        } else {
            VfError::Op {
                index: index.into().unwrap_or(0),
                err_no: e.status,
            }
        }
    }

    /// Like [`from_rpc`](VfError::from_rpc), but trusts `e.op_index` as the
    /// caller-relative operation index. The batched client helpers translate
    /// compound positions to caller indices before returning, so batched
    /// backend calls can use this directly.
    pub fn from_rpc_indexed(e: RpcError) -> VfError {
        let idx = e.op_index;
        VfError::from_rpc(e, Some(idx))
    }

    /// Re-attribute this error to a different operation index.
    pub fn with_index(self, index: usize) -> VfError {
        match self {
            VfError::Op { err_no, .. } => VfError::Op { index, err_no },
            VfError::Transport { message, .. } => VfError::Transport {
                index: Some(index),
                message,
            },
        }
    }
}

impl std::fmt::Display for VfError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            VfError::Op { index, err_no } => write!(f, "op {} failed: {}", index, err_no),
            VfError::Transport {
                index: Some(index),
                message,
            } => write!(f, "op {} transport error: {}", index, message),
            VfError::Transport {
                index: None,
                message,
            } => {
                write!(f, "transport error: {}", message)
            }
        }
    }
}

impl std::error::Error for VfError {}

pub type VfResult<T> = Result<T, VfError>;
/// Result of a compound-style operation: `()` on success, or the index and
/// error of the first failing operation.
pub type VfRes = VfResult<()>;

/// Callback used by [`VecFs::read_streamv`].
pub type ReadStreamCallback<'a> = dyn FnMut(usize, u64, &[u8], bool) -> bool + 'a;

/// An open file descriptor (backend-assigned), the Rust spelling of the C
/// `int` fd.
pub type Fd = std::os::fd::RawFd;

/// Insert an open object using a positive descriptor, wrapping safely and
/// skipping live descriptors instead of overwriting them.
#[cfg(any(feature = "nfs", feature = "smb", feature = "dummy"))]
pub(crate) fn insert_fd<T>(
    next_fd: &mut Fd,
    open_files: &mut std::collections::HashMap<Fd, T>,
    open: T,
) -> VfResult<Fd> {
    for _ in 0..i32::MAX {
        let candidate = next_fd.checked_add(1).unwrap_or(1);
        *next_fd = candidate;
        if let std::collections::hash_map::Entry::Vacant(entry) = open_files.entry(candidate) {
            entry.insert(open);
            return Ok(candidate);
        }
    }
    Err(VfError::failure(0, libc::EMFILE as u32))
}

// ---------------------------------------------------------------------------
// File references
// ---------------------------------------------------------------------------

/// The base a path is resolved against, replacing the C `VF_FD_CWD` /
/// `VF_FD_ABS` sentinel descriptors.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum VfPathBase {
    /// Relative to the client's current working directory.
    Cwd,
    /// Absolute.
    Abs,
}

/// How [`VecFs::fseek`] interprets its offset, mirroring `SEEK_SET` /
/// `SEEK_CUR` / `SEEK_END` as an enum.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
#[non_exhaustive]
pub enum SeekFrom {
    Set,
    Cur,
    End,
}

/// A reference to a file: an open descriptor, a path, or a special
/// pseudo-file. This is the Rust-native form of the C `tc_file` tagged
/// struct (`type` + `fd` + `path`), where the variant *is* the tag.
#[derive(Debug, Clone, Default, PartialEq, Eq)]
#[non_exhaustive]
pub enum VfFile {
    /// The client's current working directory (a stat target, but not a file
    /// for read/write). Also the `Default` placeholder.
    #[default]
    Cwd,
    /// An open file descriptor (backend-assigned).
    Descriptor(Fd),
    /// A path; the [`VfPathBase`] records whether it is absolute or relative
    /// to the client's cwd. Note that the raw path (see
    /// [`path`](VfFile::path)) does not itself record the base; use
    /// [`VecFs::vf_path`] to resolve a `VfFile` honoring its base.
    Path { base: VfPathBase, path: PathBuf },
    /// A path relative to the client's current working directory.
    CwdPath(PathBuf),
    /// The saved (previous) directory, mirroring the C sentinel. No current
    /// backend produces or consumes this; operations on it fail.
    Saved,
}

impl VfFile {
    pub fn from_path(path: &str) -> VfFile {
        Self::from_path_bytes(path.as_bytes())
    }

    /// Construct from arbitrary Unix path bytes.
    #[cfg(unix)]
    pub fn from_path_bytes(path: &[u8]) -> VfFile {
        let base = if path.starts_with(b"/") {
            VfPathBase::Abs
        } else {
            VfPathBase::Cwd
        };
        VfFile::Path {
            base,
            path: path_from_bytes(path),
        }
    }

    /// Construct from a native Unix path.
    pub fn from_os_path(path: &Path) -> VfFile {
        Self::from_path_bytes(crate::path::path_bytes(path))
    }

    pub fn from_fd(fd: Fd) -> VfFile {
        VfFile::Descriptor(fd)
    }

    /// VF_FILE_CURRENT: the client's current working directory itself.
    pub fn cwd() -> VfFile {
        VfFile::Cwd
    }

    /// A path relative to the client's current working directory.
    pub fn cwd_path(path: impl Into<PathBuf>) -> VfFile {
        VfFile::CwdPath(path.into())
    }

    pub fn saved() -> VfFile {
        VfFile::Saved
    }

    /// Whether this references an open descriptor.
    pub fn is_descriptor(&self) -> bool {
        matches!(self, VfFile::Descriptor(_))
    }

    /// The open descriptor, if this is a `Descriptor`.
    pub fn fd(&self) -> Option<Fd> {
        match self {
            VfFile::Descriptor(fd) => Some(*fd),
            _ => None,
        }
    }

    /// The raw path, for `Path` and `CwdPath`. This does not resolve
    /// `VfPathBase` or the client cwd; use [`VecFs::vf_path`] for the
    /// resolved root-relative form.
    pub fn path(&self) -> Option<&Path> {
        match self {
            VfFile::Path { path, .. } => Some(path),
            VfFile::CwdPath(p) => Some(p),
            _ => None,
        }
    }

    /// The raw bytes of this path, if path-backed.
    #[cfg(unix)]
    pub fn path_bytes(&self) -> Option<&[u8]> {
        self.path().map(crate::path::path_bytes)
    }
}

/// The NFSv4 object type, replacing the raw `NF4*` wire codes in
/// [`VfAttrs::ftype`].
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
pub enum VfType {
    #[default]
    Regular,
    Directory,
    Symlink,
    BlockDevice,
    CharDevice,
    Fifo,
    Socket,
    /// Any other (or unknown) NFSv4 type code.
    Other(u32),
}

impl VfType {
    pub fn from_nfs(code: u32) -> VfType {
        match code {
            NF4REG => VfType::Regular,
            NF4DIR => VfType::Directory,
            NF4LNK => VfType::Symlink,
            NF4BLK => VfType::BlockDevice,
            NF4CHR => VfType::CharDevice,
            NF4FIFO => VfType::Fifo,
            NF4SOCK => VfType::Socket,
            other => VfType::Other(other),
        }
    }

    /// The NFSv4 wire type code.
    pub fn as_nfs(&self) -> u32 {
        match self {
            VfType::Regular => NF4REG,
            VfType::Directory => NF4DIR,
            VfType::Symlink => NF4LNK,
            VfType::BlockDevice => NF4BLK,
            VfType::CharDevice => NF4CHR,
            VfType::Fifo => NF4FIFO,
            VfType::Socket => NF4SOCK,
            VfType::Other(code) => *code,
        }
    }
}

// ---------------------------------------------------------------------------
// I/O vectors and attributes
// ---------------------------------------------------------------------------

/// The offset of a read or write operation.
///
/// Unlike the C API's raw `u64` sentinels (`u64::MAX` / `u64::MAX - 1`), a
/// distinct type means an absolute offset can never collide with the special
/// "current position" / "end of file" meanings.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
#[non_exhaustive]
pub enum VfOffset {
    /// An absolute file offset.
    At(u64),
    /// The current read/write position of an open descriptor.
    Cur,
    /// The end of the file.
    End,
}

/// One element of a batched read, replacing the C `tc_iovec` (which mixed
/// input and output fields in a single mutable struct).
#[derive(Debug, Clone)]
pub struct ReadOp {
    pub file: VfFile,
    pub offset: VfOffset,
    /// Number of bytes to fetch.
    pub length: usize,
}

impl ReadOp {
    pub fn new(file: VfFile, offset: VfOffset, length: usize) -> ReadOp {
        ReadOp {
            file,
            offset,
            length,
        }
    }

    /// A read at an absolute offset.
    pub fn at(file: VfFile, offset: u64, length: usize) -> ReadOp {
        ReadOp::new(file, VfOffset::At(offset), length)
    }

    pub fn from_path(path: &str, offset: VfOffset, length: usize) -> ReadOp {
        ReadOp::new(VfFile::from_path(path), offset, length)
    }

    /// Construct a read from a native path.
    pub fn from_os_path(path: &Path, offset: VfOffset, length: usize) -> ReadOp {
        ReadOp::new(VfFile::from_os_path(path), offset, length)
    }

    /// A read from an open descriptor, typically at [`VfOffset::Cur`]
    /// (sequential) reads.
    pub fn from_fd(fd: Fd, offset: VfOffset, length: usize) -> ReadOp {
        ReadOp::new(VfFile::from_fd(fd), offset, length)
    }
}

/// The result of one [`ReadOp`]: the data and whether end-of-file was hit.
/// `file` echoes the request's file reference so results can be paired back
/// to mixed fd/path inputs without reordering assumptions.
#[derive(Debug, Clone)]
pub struct ReadResult {
    pub file: VfFile,
    /// The resolved offset the read actually started at (never `Cur`/`End`
    /// sentinels).
    pub offset: u64,
    pub data: Vec<u8>,
    /// True if the read reached end-of-file (the backend's EOF signal, or a
    /// nonzero-length read that returned fewer bytes than requested). A
    /// backend with a server EOF flag (e.g. NFS) may report `true` even when
    /// the requested length was returned exactly, because the read ended at
    /// EOF. Always false for zero-length reads.
    pub eof: bool,
}

/// One element of a batched write.
#[derive(Debug, Clone)]
pub struct WriteOp {
    pub file: VfFile,
    pub offset: VfOffset,
    pub data: Vec<u8>,
    /// Create the file if it does not exist.
    pub creation: bool,
    /// Truncate the file to zero before writing (POSIX `O_TRUNC` semantics
    /// fused into the same round trip as the write).
    pub truncate: bool,
}

impl WriteOp {
    pub fn new(file: VfFile, offset: VfOffset, data: Vec<u8>) -> WriteOp {
        WriteOp {
            file,
            offset,
            data,
            creation: false,
            truncate: false,
        }
    }

    /// A write at an absolute offset.
    pub fn at(file: VfFile, offset: u64, data: Vec<u8>) -> WriteOp {
        WriteOp::new(file, VfOffset::At(offset), data)
    }

    pub fn from_path(path: &str, offset: VfOffset, data: Vec<u8>) -> WriteOp {
        WriteOp::new(VfFile::from_path(path), offset, data)
    }

    /// Construct a write from a native path.
    pub fn from_os_path(path: &Path, offset: VfOffset, data: Vec<u8>) -> WriteOp {
        WriteOp::new(VfFile::from_os_path(path), offset, data)
    }

    pub fn from_fd(fd: Fd, offset: VfOffset, data: Vec<u8>) -> WriteOp {
        WriteOp::new(VfFile::from_fd(fd), offset, data)
    }

    /// Create the file if it does not exist.
    pub fn with_creation(mut self) -> WriteOp {
        self.creation = true;
        self
    }

    /// Truncate the file to zero before writing, in the same round trip.
    pub fn with_truncate(mut self) -> WriteOp {
        self.truncate = true;
        self
    }
}

/// The result of one [`WriteOp`]. `file` echoes the request's file reference.
#[derive(Debug, Clone)]
pub struct WriteResult {
    pub file: VfFile,
    /// The resolved offset the write actually started at (never `Cur`/`End`
    /// sentinels).
    pub offset: u64,
    pub written: usize,
    /// Whether the server committed the write to stable storage.
    pub stable: bool,
}

/// One extent to copy, mirroring `struct tc_extent_pair`.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct ExtentPair {
    pub src_path: PathBuf,
    pub dst_path: PathBuf,
    pub src_offset: u64,
    pub dst_offset: u64,
    /// Bytes to copy; `None` means "from src_offset to end of file".
    pub length: Option<u64>,
}

impl ExtentPair {
    /// `tc_fill_extent_pair()`.
    pub fn new(
        src_path: &str,
        src_offset: u64,
        dst_path: &str,
        dst_offset: u64,
        length: Option<u64>,
    ) -> ExtentPair {
        ExtentPair {
            src_path: PathBuf::from(src_path),
            dst_path: PathBuf::from(dst_path),
            src_offset,
            dst_offset,
            length,
        }
    }

    /// Construct from native paths.
    pub fn from_os_paths(
        src_path: &Path,
        src_offset: u64,
        dst_path: &Path,
        dst_offset: u64,
        length: Option<u64>,
    ) -> ExtentPair {
        ExtentPair {
            src_path: src_path.to_path_buf(),
            dst_path: dst_path.to_path_buf(),
            src_offset,
            dst_offset,
            length,
        }
    }
}

/// An Application Data Block (ADB) pattern, mirroring `struct tc_adb`.
#[derive(Debug, Clone)]
pub struct Adb {
    pub path: PathBuf,
    pub adb_offset: u64,
    pub adb_block_size: u64,
    /// Blocks to write.
    pub adb_block_count: usize,
    /// Relative offset within a block to write the ADBN; `None` = no ADBN.
    pub adb_reloff_blocknum: Option<u64>,
    /// ADBN of the first ADB.
    pub adb_block_num: u64,
    /// Relative offset within a block to write the pattern; `None` = no
    /// pattern. The pattern bytes are `adb_pattern_data`.
    pub adb_reloff_pattern: Option<u64>,
    pub adb_pattern_data: Vec<u8>,
}

impl Adb {
    /// An ADB writing only block numbers at `reloff_blocknum`.
    pub fn blocknum_only(
        path: &str,
        offset: u64,
        block_size: u64,
        block_count: usize,
        reloff_blocknum: u64,
        first_adbn: u64,
    ) -> Adb {
        Adb {
            path: PathBuf::from(path),
            adb_offset: offset,
            adb_block_size: block_size,
            adb_block_count: block_count,
            adb_reloff_blocknum: Some(reloff_blocknum),
            adb_block_num: first_adbn,
            adb_reloff_pattern: None,
            adb_pattern_data: Vec::new(),
        }
    }
}

// Presence mask for `VfAttrs`, controlling which attributes a backend must
// fetch and return. A bitflags set instead of the C `tc_attrs_masks` bool
// struct.
bitflags::bitflags! {
    #[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
    #[doc = "A bitflags set of requested attributes."]
    pub struct AttrMask: u32 {
        const MODE = 1 << 0;
        const SIZE = 1 << 1;
        const NLINK = 1 << 2;
        const FILEID = 1 << 3;
        const BLOCKS = 1 << 4;
        const UID = 1 << 5;
        const GID = 1 << 6;
        const RDEV = 1 << 7;
        const ATIME = 1 << 8;
        const MTIME = 1 << 9;
        const CTIME = 1 << 10;
        /// Request FATTR4_NAMED_ATTR (the per-object "has named attributes"
        /// boolean). Costs the server a per-entry xattr enumeration, so only
        /// request it when the caller needs it (e.g. ls long format).
        const NAMED_ATTR = 1 << 11;
        /// Request the backend's monotonic file change attribute when one is
        /// available (NFS FATTR4_CHANGE). This is stronger than timestamps for
        /// cache invalidation because rapid, same-size writes still change it.
        const CHANGE = 1 << 12;
    }
}

impl AttrMask {
    /// The attributes [`VecFs::stat`] needs (mode, size, links, fileid).
    pub fn stat() -> AttrMask {
        AttrMask::MODE | AttrMask::SIZE | AttrMask::NLINK | AttrMask::FILEID
    }
}

/// A directory and its entries, as returned by [`VecFs::walk`]. `path` is the
/// directory's root-relative path; `entries` are its immediate children.
#[derive(Debug, Clone)]
pub struct WalkEntry {
    pub path: PathBuf,
    pub entries: Vec<VfAttrs>,
}

/// File attributes, mirroring `struct tc_attrs`. `mode` is the full `st_mode`
/// (permission bits plus `S_IFMT` file-type bits); the `mtime/atime/ctime`
/// fields hold seconds and nanoseconds.
#[derive(Debug, Clone, Default)]
pub struct VfAttrs {
    pub file: VfFile,
    /// Requested attributes for a get/set call; also which attributes the
    /// caller cares about.
    pub masks: AttrMask,
    /// Which requested attributes were actually returned (populated by
    /// `getattrsv` / `lgetattrsv`). A field is only meaningful when its bit is
    /// set here; `masks` alone cannot distinguish "not requested" from
    /// "requested but not returned" from "returned as zero".
    pub returned: AttrMask,
    pub ftype: VfType,
    pub mode: u32,
    pub size: u64,
    pub nlink: u32,
    pub fileid: u64,
    /// Backend file-version token. NFS supplies FATTR4_CHANGE; other
    /// backends leave this absent by not setting [`AttrMask::CHANGE`] in
    /// `returned`.
    pub change: u64,
    pub uid: u32,
    pub gid: u32,
    pub rdev: u64,
    pub blocks: u64,
    pub mtime_sec: i64,
    pub mtime_nsec: u32,
    pub atime_sec: i64,
    pub atime_nsec: u32,
    pub ctime_sec: i64,
    pub ctime_nsec: u32,
    /// FATTR4_NAMED_ATTR: TRUE iff the object has a non-empty named
    /// attribute directory (i.e. at least one `user.*` xattr).
    pub has_named_attr: bool,
}

// ---------------------------------------------------------------------------
// The trait
// ---------------------------------------------------------------------------

/// Capability bit returned by [`VecFs::capabilities`] when the backend will
/// currently attempt a server-side copy operation (NFSv4.2 COPY or SMB
/// FSCTL_SRV_COPYCHUNK).
pub const VF_CAP_SERVER_COPY: u64 = 1 << 0;
/// The backend reports and honors Unix mode/ownership/link-count metadata.
pub const VF_CAP_POSIX_METADATA: u64 = 1 << 1;
/// The backend can create, inspect, and copy symbolic links without following
/// them.
pub const VF_CAP_SYMLINKS: u64 = 1 << 2;
/// The backend can create hard links and report their shared identity.
pub const VF_CAP_HARDLINKS: u64 = 1 << 3;
/// The backend accepts paths containing arbitrary non-UTF-8 Unix bytes.
pub const VF_CAP_NON_UTF8_PATHS: u64 = 1 << 4;
/// The backend implements no-follow metadata operations (`lstat` semantics).
pub const VF_CAP_LSTAT: u64 = 1 << 5;

/// Capabilities shared by the complete Unix-like NFS and dummy backends.
pub const VF_CAP_UNIX_SEMANTICS: u64 = VF_CAP_POSIX_METADATA
    | VF_CAP_SYMLINKS
    | VF_CAP_HARDLINKS
    | VF_CAP_NON_UTF8_PATHS
    | VF_CAP_LSTAT;

/// A vectorized filesystem: many small operations coalesced into as few
/// round trips as the backend supports.
///
/// `VfFile` references files either by descriptor (an open file) or by path
/// (absolute, or relative to the client's current working directory).
pub trait VecFs {
    /// Negotiated NFS minor version, or `None` for non-NFS backends.
    fn nfs_minorversion(&self) -> Option<u32> {
        None
    }

    /// Negotiated SMB dialect revision (for example `0x0311` for SMB 3.1.1),
    /// or `None` for non-SMB backends.
    fn smb_dialect(&self) -> Option<u16> {
        None
    }

    /// Backend capability bits. Capabilities may change after a server
    /// rejects an optional operation and the client installs a fallback.
    fn capabilities(&self) -> u64 {
        0
    }

    // -- required -----------------------------------------------------------

    /// Return the root-relative form of `path` (resolving it against the
    /// client's current working directory if it is relative), without a
    /// leading `/`. [`getcwd`](Self::getcwd) is the display form (with a
    /// leading `/`); [`vf_path`](Self::vf_path) resolves a [`VfFile`] the same
    /// way while honoring its [`VfPathBase`].
    fn abs_path(&self, path: &Path) -> PathBuf;

    /// Open a file by path, similar to `tc_open_by_path(2)`. `base` is
    /// `VfPathBase::Cwd` or `VfPathBase::Abs`. When `O_CREAT` is set, `mode`
    /// is applied to the new file.
    fn open_by_path(
        &mut self,
        base: VfPathBase,
        pathname: &Path,
        flags: i32,
        mode: u32,
    ) -> VfResult<VfFile>;

    /// Close an open file, `tc_close()`.
    fn close(&mut self, tcf: &VfFile) -> VfResult<()>;

    /// Change the client's current directory, `tc_chdir()`.
    fn chdir(&mut self, path: &Path) -> VfResult<()>;

    /// Current working directory, `tc_getcwd()`.
    fn getcwd(&self) -> PathBuf;

    /// Read from one or more files, `tc_readv()`. Returns one result per
    /// request, or fails at the first failing operation.
    fn readv(&mut self, reads: &[ReadOp]) -> VfResult<Vec<ReadResult>>;

    /// Read each file in full from offset 0, `tc_read_allv()`. Returns one
    /// byte buffer per request in input order.
    ///
    /// The default implementation stats each file and reads it with
    /// [`readv`](Self::readv); a backend may override it to read until EOF
    /// without a separate size round trip.
    fn read_allv(&mut self, files: &[VfFile]) -> VfResult<Vec<Vec<u8>>> {
        let mut out = Vec::with_capacity(files.len());
        for (i, f) in files.iter().enumerate() {
            let mut a = VfAttrs {
                file: f.clone(),
                masks: AttrMask::SIZE,
                ..VfAttrs::default()
            };
            self.getattrsv(std::slice::from_mut(&mut a))
                .map_err(|e| e.with_index(i))?;
            let mut r = self
                .readv(&[ReadOp::new(f.clone(), VfOffset::At(0), a.size as usize)])
                .map_err(|e| e.with_index(i))?;
            out.push(r.pop().expect("readv returns one result per op").data);
        }
        Ok(out)
    }

    /// Write to one or more files, `tc_writev()`. Returns one result per
    /// request, or fails at the first failing operation.
    fn writev(&mut self, writes: &[WriteOp]) -> VfResult<Vec<WriteResult>>;

    /// Reposition the read/write offset of an open file, `tc_fseek()`. The
    /// offset lives in backend state keyed by the descriptor; `tcf` is not
    /// modified (hence `&VfFile`). Returns the new offset. Note that
    /// [`SeekFrom::End`] needs the file size, which is an extra round trip on
    /// backends that do not cache it.
    fn fseek(&mut self, tcf: &VfFile, offset: i64, whence: SeekFrom) -> VfResult<i64>;

    /// Get attributes of an array of files, `tc_getattrsv()`. Follows
    /// symlinks to the target.
    fn getattrsv(&mut self, attrs: &mut [VfAttrs]) -> VfRes;

    /// Like [`getattrsv`](Self::getattrsv) but does not follow symlinks:
    /// attributes are for the symlink itself, `tc_lgetattrsv()`.
    fn lgetattrsv(&mut self, attrs: &mut [VfAttrs]) -> VfRes;

    /// Set attributes on an array of files, `tc_setattrsv()`. Only
    /// [`AttrMask::MODE`] and [`AttrMask::SIZE`] are supported; requesting any
    /// other bit fails with [`VF_ERR_UNSUPPORTED`] at that index, and an
    /// empty mask is a no-op. Follows symlinks to the target.
    fn setattrsv(&mut self, attrs: &[VfAttrs]) -> VfRes;

    /// Like [`setattrsv`](Self::setattrsv) but does not follow symlinks:
    /// attributes are set on the symlink itself, `tc_lsetattrsv()`. Backends
    /// without a non-following setter (e.g. no `lchmod` on Linux) must fail
    /// with [`VF_ERR_UNSUPPORTED`] for symlinks rather than silently follow.
    fn lsetattrsv(&mut self, attrs: &[VfAttrs]) -> VfRes;

    /// List a directory, `tc_listdir()`. Returns entry paths and attributes.
    ///
    /// `max_count` limits the number of entries returned per directory (0
    /// means no limit). Entry `VfFile`s are absolute (root-relative) paths.
    /// With `recursive`, entries of nested directories are included depth-
    /// first (still subject to `max_count`).
    fn listdir(
        &mut self,
        dir: &Path,
        masks: AttrMask,
        max_count: usize,
        recursive: bool,
    ) -> VfResult<Vec<VfAttrs>>;

    /// Recursively enumerate `root`, returning each directory with its entries.
    ///
    /// `sort` orders a directory's entries the way the caller's presentation
    /// layer would (so subdirectories are visited in the same order the caller
    /// lists them). The default implementation recurses via
    /// [`listdir`](Self::listdir); a backend may override it to batch many
    /// directories into few large compounds.
    fn walk(
        &mut self,
        root: &Path,
        masks: AttrMask,
        sort: &mut dyn FnMut(&Path, &mut Vec<VfAttrs>),
    ) -> VfResult<Vec<WalkEntry>> {
        // Explicit stack (pre-order, subdirectories visited in the order the
        // sort callback produced) so deep trees cannot overflow the call
        // stack.
        let mut out = Vec::new();
        let mut stack = vec![root.to_path_buf()];
        while let Some(dir) = stack.pop() {
            let mut entries = self.listdir(&dir, masks, 0, false)?;
            sort(dir.as_path(), &mut entries);
            let subdirs: Vec<PathBuf> = entries
                .iter()
                .filter(|e| e.ftype == VfType::Directory)
                .filter_map(|e| e.file.path().map(|p| p.to_path_buf()))
                .collect();
            for s in subdirs.into_iter().rev() {
                stack.push(s);
            }
            out.push(WalkEntry { path: dir, entries });
        }
        Ok(out)
    }

    /// Rename a list of file pairs, `tc_renamev()`.
    fn renamev(&mut self, pairs: &[(VfFile, VfFile)]) -> VfRes;

    /// Remove a list of files (or empty directories), `tc_removev()`.
    fn removev(&mut self, files: &[VfFile]) -> VfRes;

    /// Create one or more directories, `tc_mkdirv()`.
    fn mkdirv(&mut self, dirs: &[VfAttrs]) -> VfRes;

    /// Create a list of symlinks, `tc_symlinkv()`.
    fn symlinkv(&mut self, oldpaths: &[&Path], newpaths: &[&Path]) -> VfRes;

    /// Read symlink targets, `tc_readlinkv()`.
    fn readlinkv(&mut self, paths: &[&Path]) -> VfResult<Vec<Vec<u8>>>;

    /// Create hard links, `tc_hardlinkv()`.
    fn hardlinkv(&mut self, oldpaths: &[&Path], newpaths: &[&Path]) -> VfRes;

    /// Copy extents by reading and writing, `tc_dupv()`. Follows symlinks
    /// (copies the target's contents).
    fn dupv(&mut self, pairs: &[ExtentPair]) -> VfRes;

    /// Copy extents without following symlinks, `tc_lcopyv()`: symlinks are
    /// recreated as symlinks with the same target; other objects are copied
    /// by data (like [`dupv`](Self::dupv)).
    fn lcopyv(&mut self, pairs: &[ExtentPair]) -> VfRes;

    /// Write Application Data Blocks, `tc_write_adb()`. Returns the number
    /// of blocks written for each ADB.
    fn write_adb(&mut self, patterns: &[Adb]) -> VfResult<Vec<usize>>;

    /// Read files in bounded chunks while retaining vectorized I/O.
    ///
    /// At most `memory_limit` bytes are requested in one batch and no single
    /// request exceeds `chunk_size`. `cb` receives the original file index,
    /// absolute offset, data, and EOF state. Returning `false` cancels the
    /// stream successfully, providing backpressure without buffering whole
    /// files. The callback runs while `self` is borrowed and must not reenter
    /// this filesystem instance.
    fn read_streamv(
        &mut self,
        files: &[VfFile],
        chunk_size: usize,
        memory_limit: usize,
        cb: &mut ReadStreamCallback<'_>,
    ) -> VfRes {
        use std::collections::VecDeque;

        if chunk_size == 0 || memory_limit == 0 {
            return Err(VfError::failure(0, ERR_INVAL));
        }
        let mut pending: VecDeque<usize> = (0..files.len()).collect();
        let mut offsets = vec![0u64; files.len()];
        while !pending.is_empty() {
            let mut budget = memory_limit;
            let mut batch_indices = Vec::new();
            let mut reads = Vec::new();
            while budget > 0 && !pending.is_empty() {
                let index = pending.pop_front().expect("pending was non-empty");
                let length = chunk_size.min(budget);
                reads.push(ReadOp::at(files[index].clone(), offsets[index], length));
                batch_indices.push(index);
                budget -= length;
            }
            let results = self.readv(&reads).map_err(|error| {
                let index = batch_indices
                    .get(error.index())
                    .copied()
                    .unwrap_or_else(|| error.index());
                error.with_index(index)
            })?;
            for (batch_index, result) in results.into_iter().enumerate() {
                let index = batch_indices[batch_index];
                let offset = offsets[index];
                let eof = result.eof || result.data.is_empty();
                offsets[index] = offset
                    .checked_add(result.data.len() as u64)
                    .ok_or_else(|| VfError::failure(index, libc::EOVERFLOW as u32))?;
                if !cb(index, offset, &result.data, eof) {
                    return Ok(());
                }
                if !eof {
                    pending.push_back(index);
                }
            }
        }
        Ok(())
    }

    /// Remove a list of objects, recursively when `recursive`, `tc_rm()`.
    fn rm(&mut self, objs: &[&Path], recursive: bool) -> VfRes;

    /// Recursively copy a directory tree, `tc_cp_recursive()`.
    fn cp_recursive(
        &mut self,
        src_dir: &Path,
        dst: &Path,
        symlinks: bool,
        use_server_side_copy: bool,
    ) -> VfRes;

    // -- defaults -----------------------------------------------------------

    /// Resolve a [`VfFile`] to its root-relative path (no leading `/`),
    /// honoring `VfPathBase::Abs`/`VfPathBase::Cwd` and the client's current
    /// working directory. Descriptors, `Null`, and `Saved` are not paths and
    /// fail with [`ERR_INVAL`]. Backends should use this when a method needs
    /// the file's path, so a `VfFile` resolves identically across all trait
    /// methods.
    fn vf_path(&self, file: &VfFile) -> VfResult<PathBuf> {
        match file {
            VfFile::Path {
                base: VfPathBase::Abs,
                path,
            } => Ok(crate::path::path_from_bytes(&crate::path::normalize_bytes(
                crate::path::path_bytes(path),
            ))),
            VfFile::Path {
                base: VfPathBase::Cwd,
                path,
            }
            | VfFile::CwdPath(path) => Ok(self.abs_path(path)),
            VfFile::Cwd => Ok(self.abs_path(Path::new(""))),
            VfFile::Descriptor(_) | VfFile::Saved => Err(VfError::failure(0, ERR_INVAL)),
        }
    }

    /// Open a file by path, `tc_open()`.
    fn open(&mut self, pathname: &Path, flags: i32, mode: u32) -> VfResult<VfFile> {
        self.open_by_path(VfPathBase::Cwd, pathname, flags, mode)
    }

    /// Read from a single file at an absolute offset, `tc_read()`.
    fn read(&mut self, file: &VfFile, offset: u64, length: usize) -> VfResult<Vec<u8>> {
        let r = self.readv(&[ReadOp::at(file.clone(), offset, length)])?;
        Ok(r.into_iter().next().expect("one result").data)
    }

    /// Write to a single file at an absolute offset, `tc_write()`.
    fn write(&mut self, file: &VfFile, offset: u64, data: &[u8]) -> VfResult<usize> {
        let w = self.writev(&[WriteOp::at(file.clone(), offset, data.to_vec())])?;
        Ok(w.into_iter().next().expect("one result").written)
    }

    /// Open several files at once, each with its own flags and mode,
    /// `tc_openv()`. `flags`, `modes`, and `paths` must have equal lengths;
    /// a mismatch fails with [`ERR_INVAL`] at index 0.
    fn openv(&mut self, paths: &[&Path], flags: &[i32], modes: &[u32]) -> VfResult<Vec<VfFile>> {
        if paths.len() != flags.len() || paths.len() != modes.len() {
            return Err(VfError::failure(0, ERR_INVAL));
        }
        let mut out = Vec::with_capacity(paths.len());
        for (i, ((p, flag), mode)) in paths.iter().zip(flags).zip(modes).enumerate() {
            out.push(self.open(p, *flag, *mode).map_err(|e| e.with_index(i))?);
        }
        Ok(out)
    }

    /// Open several files at once with a shared flags and mode,
    /// `tc_openv_simple()`.
    fn openv_simple(&mut self, paths: &[&Path], flags: i32, mode: u32) -> VfResult<Vec<VfFile>> {
        let flags_v = vec![flags; paths.len()];
        let modes_v = vec![mode; paths.len()];
        self.openv(paths, &flags_v, &modes_v)
    }

    /// Close several files, `tc_closev()`.
    fn closev(&mut self, files: &[VfFile]) -> VfRes {
        for (i, f) in files.iter().enumerate() {
            self.close(f).map_err(|e| e.with_index(i))?;
        }
        Ok(())
    }

    /// Stat a path, `tc_stat()`. Follows symlinks to the target.
    fn stat(&mut self, path: &Path) -> VfResult<VfAttrs> {
        let mut a = VfAttrs {
            file: VfFile::from_os_path(path),
            masks: AttrMask::stat(),
            ..VfAttrs::default()
        };
        self.getattrsv(std::slice::from_mut(&mut a))?;
        Ok(a)
    }

    /// `tc_lstat()`: like [`stat`](Self::stat) but does not follow symlinks.
    fn lstat(&mut self, path: &Path) -> VfResult<VfAttrs> {
        let mut a = VfAttrs {
            file: VfFile::from_os_path(path),
            masks: AttrMask::stat(),
            ..VfAttrs::default()
        };
        self.lgetattrsv(std::slice::from_mut(&mut a))?;
        Ok(a)
    }

    /// `tc_fstat()`.
    fn fstat(&mut self, tcf: &VfFile) -> VfResult<VfAttrs> {
        let mut a = VfAttrs {
            file: tcf.clone(),
            masks: AttrMask::stat(),
            ..VfAttrs::default()
        };
        self.getattrsv(std::slice::from_mut(&mut a))?;
        Ok(a)
    }

    /// Whether `path` exists, distinguishing "not found" from other errors
    /// (e.g. permission denied) that are returned as `Err`. Uses
    /// `lstat` semantics: a dangling symlink exists.
    fn exists(&mut self, path: &Path) -> VfResult<bool> {
        match self.lstat(path) {
            Ok(_) => Ok(true),
            Err(e) if e.err_no() == ERR_NOENT => Ok(false),
            Err(e) => Err(e),
        }
    }

    /// Return the file type of `path` itself (`lstat` semantics: a symlink
    /// reports [`VfType::Symlink`] rather than its target's type).
    fn file_type(&mut self, path: &Path) -> VfResult<VfType> {
        Ok(self.lstat(path)?.ftype)
    }

    /// List directories with a callback, `tc_listdirv()`. Returning `false`
    /// stops the listing early.
    fn listdirv(
        &mut self,
        dirs: &[&Path],
        masks: AttrMask,
        max_entries: usize,
        recursive: bool,
        cb: &mut dyn FnMut(&VfAttrs, &Path) -> bool,
    ) -> VfRes {
        for (i, d) in dirs.iter().enumerate() {
            let entries = self
                .listdir(d, masks, max_entries, recursive)
                .map_err(|e| e.with_index(i))?;
            for e in &entries {
                /* A recursive list contains descendants too, so report each
                 * entry's actual parent rather than attributing every row to the
                 * original root. Backend overrides follow the same contract. */
                let entry_dir = if recursive {
                    e.file.path().and_then(Path::parent).unwrap_or(d)
                } else {
                    d
                };
                if !cb(e, entry_dir) {
                    return Ok(());
                }
            }
        }
        Ok(())
    }

    /// `tc_unlink()`.
    fn unlink(&mut self, pathname: &Path) -> VfResult<()> {
        self.removev(&[VfFile::from_os_path(pathname)])
    }

    /// `tc_unlinkv()`.
    fn unlinkv(&mut self, pathnames: &[&Path]) -> VfRes {
        let files: Vec<VfFile> = pathnames.iter().map(|p| VfFile::from_os_path(p)).collect();
        self.removev(&files)
    }

    /// Create a directory, `tc_mkdir()`.
    fn mkdir(&mut self, path: &Path, mode: u32) -> VfResult<()> {
        let a = VfAttrs {
            file: VfFile::from_os_path(path),
            masks: AttrMask::MODE,
            mode,
            ..VfAttrs::default()
        };
        self.mkdirv(std::slice::from_ref(&a))
    }

    /// Create a symlink, `tc_symlink()`.
    fn symlink(&mut self, oldpath: &Path, newpath: &Path) -> VfResult<()> {
        self.symlinkv(
            std::slice::from_ref(&oldpath),
            std::slice::from_ref(&newpath),
        )
    }

    /// Read a symlink target, `tc_readlink()`.
    fn readlink(&mut self, path: &Path) -> VfResult<Vec<u8>> {
        let v = self.readlinkv(std::slice::from_ref(&path))?;
        Ok(v.into_iter().next().expect("one result"))
    }

    /// `tc_ldupv()`: same read/write extent copy as
    /// [`dupv`](Self::dupv). Retained for C API parity; a backend that
    /// distinguishes a "local" copy should override.
    fn ldupv(&mut self, pairs: &[ExtentPair]) -> VfRes {
        self.dupv(pairs)
    }

    /// `tc_copyv()`: server-side copy where the backend supports it. The
    /// default implementation performs a client-side read/write copy via
    /// [`dupv`](Self::dupv); backends with a server-side COPY should
    /// override.
    fn copyv(&mut self, pairs: &[ExtentPair]) -> VfRes {
        self.dupv(pairs)
    }

    /// Create a directory and all its ancestors, `tc_ensure_dir()`. Uses
    /// `mkdir` and accepts an existing directory (`EEXIST`) instead of an
    /// exists-then-mkdir check, avoiding the race between the two. `dir` is
    /// resolved against the cwd via [`abs_path`](Self::abs_path) and then
    /// rebuilt as an absolute (root-relative) path.
    fn ensure_dir(&mut self, dir: &Path, mode: u32) -> VfResult<()> {
        use std::path::Component;
        let mut so_far = PathBuf::new();
        for comp in self.abs_path(dir).components() {
            if let Component::Normal(part) = comp {
                so_far.push(part);
                let full = Path::new("/").join(&so_far);
                match self.mkdir(&full, mode) {
                    Ok(()) => {}
                    Err(e) if e.err_no() == ERR_EXIST => {}
                    Err(e) => return Err(e),
                }
            }
        }
        Ok(())
    }
}

/// `AsRef<Path>` convenience wrappers for [`VecFs`] methods.
///
/// The object-safe [`VecFs`] trait intentionally takes `&Path`. These
/// wrappers accept any type that can be viewed as a path (`&str`, `String`,
/// `PathBuf`, `&Path`, ...) and are useful for Rust callers that do not need
/// dynamic dispatch.
pub trait VecFsExt: VecFs {
    fn stat_path<P: AsRef<Path>>(&mut self, path: P) -> VfResult<VfAttrs> {
        self.stat(path.as_ref())
    }

    fn lstat_path<P: AsRef<Path>>(&mut self, path: P) -> VfResult<VfAttrs> {
        self.lstat(path.as_ref())
    }

    fn exists_path<P: AsRef<Path>>(&mut self, path: P) -> VfResult<bool> {
        self.exists(path.as_ref())
    }

    fn file_type_path<P: AsRef<Path>>(&mut self, path: P) -> VfResult<VfType> {
        self.file_type(path.as_ref())
    }

    fn open_path<P: AsRef<Path>>(&mut self, path: P, flags: i32, mode: u32) -> VfResult<VfFile> {
        self.open(path.as_ref(), flags, mode)
    }

    fn mkdir_path<P: AsRef<Path>>(&mut self, path: P, mode: u32) -> VfResult<()> {
        self.mkdir(path.as_ref(), mode)
    }

    fn ensure_dir_path<P: AsRef<Path>>(&mut self, path: P, mode: u32) -> VfResult<()> {
        self.ensure_dir(path.as_ref(), mode)
    }

    fn chdir_path<P: AsRef<Path>>(&mut self, path: P) -> VfResult<()> {
        self.chdir(path.as_ref())
    }

    fn unlink_path<P: AsRef<Path>>(&mut self, path: P) -> VfResult<()> {
        self.unlink(path.as_ref())
    }

    fn readlink_path<P: AsRef<Path>>(&mut self, path: P) -> VfResult<Vec<u8>> {
        self.readlink(path.as_ref())
    }

    fn symlink_path<P, Q>(&mut self, oldpath: P, newpath: Q) -> VfResult<()>
    where
        P: AsRef<Path>,
        Q: AsRef<Path>,
    {
        self.symlink(oldpath.as_ref(), newpath.as_ref())
    }

    fn listdir_path<P: AsRef<Path>>(
        &mut self,
        path: P,
        masks: AttrMask,
        max_count: usize,
        recursive: bool,
    ) -> VfResult<Vec<VfAttrs>> {
        self.listdir(path.as_ref(), masks, max_count, recursive)
    }

    fn walk_path<P: AsRef<Path>>(
        &mut self,
        root: P,
        masks: AttrMask,
        sort: &mut dyn FnMut(&Path, &mut Vec<VfAttrs>),
    ) -> VfResult<Vec<WalkEntry>> {
        self.walk(root.as_ref(), masks, sort)
    }

    fn openv_paths<P: AsRef<Path>>(
        &mut self,
        paths: &[P],
        flags: &[i32],
        modes: &[u32],
    ) -> VfResult<Vec<VfFile>> {
        let refs: Vec<&Path> = paths.iter().map(|p| p.as_ref()).collect();
        self.openv(&refs, flags, modes)
    }

    fn openv_simple_paths<P: AsRef<Path>>(
        &mut self,
        paths: &[P],
        flags: i32,
        mode: u32,
    ) -> VfResult<Vec<VfFile>> {
        let refs: Vec<&Path> = paths.iter().map(|p| p.as_ref()).collect();
        self.openv_simple(&refs, flags, mode)
    }

    fn unlinkv_paths<P: AsRef<Path>>(&mut self, paths: &[P]) -> VfRes {
        let refs: Vec<&Path> = paths.iter().map(|p| p.as_ref()).collect();
        self.unlinkv(&refs)
    }

    fn rm_paths<P: AsRef<Path>>(&mut self, paths: &[P], recursive: bool) -> VfRes {
        let refs: Vec<&Path> = paths.iter().map(|p| p.as_ref()).collect();
        self.rm(&refs, recursive)
    }

    fn rm_recursive_path<P: AsRef<Path>>(&mut self, path: P) -> VfRes {
        self.rm(&[path.as_ref()], true)
    }
}

impl<T: VecFs + ?Sized> VecFsExt for T {}

/// `tc_rm_recursive()`.
pub fn rm_recursive(fs: &mut impl VecFs, dir: &Path) -> VfRes {
    fs.rm(&[dir], true)
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::dummy_vecfs::DummyVecFs;
    use crate::error::RpcError;
    use std::sync::atomic::{AtomicUsize, Ordering};

    static COUNTER: AtomicUsize = AtomicUsize::new(0);

    /// A unique temporary directory that is removed on drop.
    struct TempRoot(PathBuf);

    impl TempRoot {
        fn new(tag: &str) -> TempRoot {
            let n = COUNTER.fetch_add(1, Ordering::Relaxed);
            let p = std::env::temp_dir().join(format!(
                "vnfs-vecfs-test-{}-{}-{}",
                tag,
                std::process::id(),
                n
            ));
            let _ = std::fs::remove_dir_all(&p);
            TempRoot(p)
        }
    }

    impl Drop for TempRoot {
        fn drop(&mut self) {
            let _ = std::fs::remove_dir_all(&self.0);
        }
    }

    /// A fresh dummy backend rooted at a unique temp directory.
    fn fs(tag: &str) -> (TempRoot, DummyVecFs) {
        let root = TempRoot::new(tag);
        let fs = DummyVecFs::new(root.0.clone());
        (root, fs)
    }

    fn write(fs: &mut DummyVecFs, path: &str, data: &[u8]) {
        fs.writev(&[WriteOp::at(VfFile::from_path(path), 0, data.to_vec()).with_creation()])
            .expect("write");
    }

    // ------------------------------------------------------------------
    // VfError
    // ------------------------------------------------------------------

    #[test]
    fn descriptor_allocator_wraps_and_skips_live_entries() {
        let mut next = i32::MAX;
        let mut files = std::collections::HashMap::from([(1, "live")]);
        let fd = insert_fd(&mut next, &mut files, "new").unwrap();
        assert_eq!(fd, 2);
        assert_eq!(files.get(&1), Some(&"live"));
        assert_eq!(files.get(&2), Some(&"new"));
    }

    #[test]
    fn vf_error_preserves_transport_message() {
        let e = VfError::from_rpc(RpcError::transport("connection refused"), 3);
        assert!(e.is_transport());
        assert_eq!(e.index(), 3);
        assert_eq!(e.index_opt(), Some(3));
        assert_eq!(e.err_no(), VF_ERR_RPC);
        assert!(e.to_string().contains("connection refused"));

        // An unattributable transport failure has no op index.
        let e = VfError::from_rpc(RpcError::transport("server gone"), None);
        assert!(e.is_transport());
        assert_eq!(e.index_opt(), None);
        assert_eq!(e.index(), 0);
        assert!(!e.to_string().contains("op "));

        // Server status errors stay Op errors with the caller-supplied index.
        let e = VfError::from_rpc(RpcError::op(4, ERR_NOENT), 1);
        assert!(!e.is_transport());
        assert_eq!(e.index(), 1);
        assert_eq!(e.index_opt(), Some(1));
        assert_eq!(e.err_no(), ERR_NOENT);
    }

    #[test]
    fn vf_error_indexed_and_remap() {
        let e = VfError::from_rpc_indexed(RpcError::op(4, ERR_EXIST));
        assert_eq!((e.index(), e.err_no()), (4, ERR_EXIST));
        assert_eq!(e.index_opt(), Some(4));
        assert_eq!(e.with_index(9).index(), 9);
        assert_eq!(VfError::transport(2, "boom").with_index(5).index(), 5);
        assert_eq!(VfError::transport(None, "boom").index_opt(), None);
    }

    // ------------------------------------------------------------------
    // Offsets: no sentinel collision, Cur/End resolution, result offsets
    // ------------------------------------------------------------------

    #[test]
    fn absolute_offset_at_u64_max_minus_one_is_not_cur() {
        let (_root, mut fs) = fs("huge-offset");
        write(&mut fs, "/f", b"abcdefgh");
        let fd = fs.open(Path::new("/f"), 0, 0).unwrap();
        fs.fseek(&fd, 2, SeekFrom::Set).unwrap();

        // Previously u64::MAX - 1 collided with the VF_OFFSET_CUR sentinel and
        // would have read from the current position (2) instead. With the
        // typed offset it is an absolute offset: the platform may reject it
        // (pread beyond i64::MAX) or return an empty read, but never data
        // from the current position.
        match fs.readv(&[ReadOp::new(fd.clone(), VfOffset::At(u64::MAX - 1), 8)]) {
            Err(e) => assert_eq!(e.err_no(), ERR_INVAL),
            Ok(r) => {
                assert!(r[0].data.is_empty());
                assert!(r[0].eof);
            }
        }
        fs.close(&fd).unwrap();
    }

    #[test]
    fn cur_offset_reads_resolve_and_advance() {
        let (_root, mut fs) = fs("cur");
        write(&mut fs, "/f", b"hello world");
        let fd = fs.open(Path::new("/f"), libc::O_RDWR, 0).unwrap();
        assert_eq!(fs.fseek(&fd, 0, SeekFrom::End).unwrap(), 11);

        let w = fs
            .writev(&[WriteOp::new(fd.clone(), VfOffset::Cur, b"XY".to_vec())])
            .unwrap();
        assert_eq!(w[0].offset, 11); // resolved current position
        let w = fs
            .writev(&[WriteOp::new(fd.clone(), VfOffset::Cur, b"Z".to_vec())])
            .unwrap();
        assert_eq!(w[0].offset, 13);

        let r = fs
            .readv(&[ReadOp::new(fd.clone(), VfOffset::Cur, 100)])
            .unwrap();
        assert_eq!(r[0].offset, 14); // resolved, not the Cur sentinel
        assert!(r[0].data.is_empty());
        assert!(r[0].eof);
        fs.close(&fd).unwrap();
    }

    #[test]
    fn end_offset_writes_append_and_reads_at_end() {
        let (_root, mut fs) = fs("end");
        write(&mut fs, "/f", b"hello world");

        let fd = fs.open(Path::new("/f"), libc::O_RDWR, 0).unwrap();
        let w = fs
            .writev(&[WriteOp::new(fd.clone(), VfOffset::End, b"!".to_vec())])
            .unwrap();
        assert_eq!(w[0].offset, 11);
        fs.close(&fd).unwrap();

        // A read positioned at End starts at the file size, so it is at EOF.
        let r = fs
            .readv(&[ReadOp::new(VfFile::from_path("/f"), VfOffset::End, 5)])
            .unwrap();
        assert_eq!(r[0].offset, 12); // the resolved start is the file size
        assert!(r[0].data.is_empty());
        assert!(r[0].eof);

        let r = fs
            .readv(&[ReadOp::new(VfFile::from_path("/f"), VfOffset::End, 100)])
            .unwrap();
        assert!(r[0].eof);
        assert_eq!(fs.stat(Path::new("/f")).unwrap().size, 12);
    }

    #[test]
    fn writev_truncate_removes_stale_tail() {
        let (_root, mut fs) = fs("writev-truncate");
        write(&mut fs, "/f", b"longer-than-needed");
        fs.writev(&[WriteOp::at(VfFile::from_path("/f"), 0, b"hi".to_vec()).with_truncate()])
            .unwrap();
        // O_TRUNC semantics: the stale tail is gone.
        assert_eq!(fs.read(&VfFile::from_path("/f"), 0, 100).unwrap(), b"hi");

        // A plain overwrite keeps the tail (pwrite semantics).
        write(&mut fs, "/g", b"abcdef");
        fs.writev(&[WriteOp::at(VfFile::from_path("/g"), 0, b"xy".to_vec())])
            .unwrap();
        assert_eq!(
            fs.read(&VfFile::from_path("/g"), 0, 100).unwrap(),
            b"xycdef"
        );
    }

    #[test]
    fn eof_is_only_true_at_end() {
        let (_root, mut fs) = fs("eof");
        write(&mut fs, "/f", b"abc");

        let r = fs
            .readv(&[ReadOp::at(VfFile::from_path("/f"), 0, 3)])
            .unwrap();
        assert_eq!(r[0].data, b"abc");
        assert!(!r[0].eof);

        let r = fs
            .readv(&[ReadOp::at(VfFile::from_path("/f"), 0, 4)])
            .unwrap();
        assert_eq!(r[0].data, b"abc");
        assert!(r[0].eof);

        // Zero-length reads never report EOF.
        let r = fs
            .readv(&[ReadOp::at(VfFile::from_path("/f"), 0, 0)])
            .unwrap();
        assert!(r[0].data.is_empty());
        assert!(!r[0].eof);
    }

    #[test]
    fn fseek_takes_shared_ref_and_works() {
        let (_root, mut fs) = fs("fseek");
        write(&mut fs, "/f", b"hello world");
        let fd = fs.open(Path::new("/f"), 0, 0).unwrap();

        assert_eq!(fs.fseek(&fd, 6, SeekFrom::Set).unwrap(), 6);
        let r = fs
            .readv(&[ReadOp::new(fd.clone(), VfOffset::Cur, 5)])
            .unwrap();
        assert_eq!(r[0].data, b"world");

        assert_eq!(fs.fseek(&fd, -5, SeekFrom::End).unwrap(), 6);
        assert_eq!(fs.fseek(&fd, 0, SeekFrom::Cur).unwrap(), 6);
        assert_eq!(
            fs.fseek(&fd, -100, SeekFrom::Set).unwrap_err().err_no(),
            ERR_INVAL
        );
        fs.close(&fd).unwrap();
    }

    // ------------------------------------------------------------------
    // VfFile base/cwd resolution is honored by every path-taking method
    // ------------------------------------------------------------------

    #[test]
    fn cwd_relative_unlink_targets_cwd() {
        let (_root, mut fs) = fs("cwd-unlink");
        fs.mkdir(Path::new("/sub"), 0o755).unwrap();
        fs.chdir(Path::new("sub")).unwrap();

        write(&mut fs, "a", b"x"); // cwd-relative write
        fs.unlink(Path::new("a")).unwrap();

        // The file was removed from sub/, not from the root.
        assert!(!fs.exists(Path::new("a")).unwrap());
        assert_eq!(fs.lstat(Path::new("/a")).unwrap_err().err_no(), ERR_NOENT);
    }

    #[test]
    fn renamev_honors_path_base() {
        let (_root, mut fs) = fs("rename-base");
        fs.mkdir(Path::new("/sub"), 0o755).unwrap();
        write(&mut fs, "/src", b"1");
        write(&mut fs, "/sub/src2", b"2");
        fs.chdir(Path::new("sub")).unwrap();

        // base Abs with a non-slash path is root-relative even after chdir.
        let abs_src = VfFile::Path {
            base: VfPathBase::Abs,
            path: PathBuf::from("src"),
        };
        let abs_dst = VfFile::Path {
            base: VfPathBase::Abs,
            path: PathBuf::from("dst"),
        };
        fs.renamev(&[(abs_src, abs_dst)]).unwrap();
        assert!(!fs.exists(Path::new("/src")).unwrap());
        assert!(fs.exists(Path::new("/dst")).unwrap());

        // base Cwd resolves against the cwd.
        let cwd_src = VfFile::Path {
            base: VfPathBase::Cwd,
            path: PathBuf::from("src2"),
        };
        let cwd_dst = VfFile::Path {
            base: VfPathBase::Cwd,
            path: PathBuf::from("dst2"),
        };
        fs.renamev(&[(cwd_src, cwd_dst)]).unwrap();
        assert!(!fs.exists(Path::new("/sub/src2")).unwrap());
        assert!(fs.exists(Path::new("/sub/dst2")).unwrap());
    }

    #[test]
    fn vf_path_rejects_descriptors() {
        let (_root, mut fs) = fs("vf-path");
        write(&mut fs, "/f", b"x");
        let fd = fs.open(Path::new("/f"), 0, 0).unwrap();
        assert_eq!(fs.vf_path(&fd).unwrap_err().err_no(), ERR_INVAL);
        fs.close(&fd).unwrap();
    }

    #[test]
    fn vf_file_cwd_and_cwd_path_resolution() {
        let (_root, mut fs) = fs("cwd-variants");
        fs.mkdir(Path::new("/sub"), 0o755).unwrap();
        write(&mut fs, "/sub/f", b"x");

        // `cwd()` is the cwd itself; `cwd_path` is relative to it.
        assert_eq!(fs.vf_path(&VfFile::cwd()).unwrap(), Path::new(""));
        assert_eq!(
            VfFile::cwd_path("f").path(),
            Some(std::path::Path::new("f"))
        );
        fs.chdir(Path::new("/sub")).unwrap();
        assert_eq!(fs.vf_path(&VfFile::cwd()).unwrap(), Path::new("sub"));
        assert_eq!(
            fs.vf_path(&VfFile::cwd_path("f")).unwrap(),
            Path::new("sub/f")
        );

        // The cwd is a stat target but not a file for read/write.
        let mut a = VfAttrs {
            file: VfFile::cwd(),
            masks: AttrMask::stat(),
            ..VfAttrs::default()
        };
        fs.getattrsv(std::slice::from_mut(&mut a)).unwrap();
        assert_eq!(a.ftype, VfType::Directory);
        assert_eq!(
            fs.readv(&[ReadOp::new(VfFile::cwd(), VfOffset::At(0), 1)])
                .unwrap_err()
                .err_no(),
            ERR_ISDIR
        );
    }

    // ------------------------------------------------------------------
    // Dummy root sandbox: `..` and symlinks cannot escape the root
    // ------------------------------------------------------------------

    #[test]
    fn dummy_root_clamps_dotdot() {
        let (root, mut fs) = fs("sandbox-dotdot");

        // Writing through ".." lands inside the root, not in its parent.
        fs.writev(&[
            WriteOp::at(VfFile::from_path("/../escape"), 0, b"x".to_vec()).with_creation(),
        ])
        .unwrap();
        assert!(fs.exists(Path::new("/escape")).unwrap());
        assert!(!root.0.parent().unwrap().join("escape").exists());

        // "/.." and "/../../x" stay under the root.
        let st = fs.stat(Path::new("/..")).unwrap();
        assert_eq!(st.ftype, VfType::Directory);
        fs.writev(&[
            WriteOp::at(VfFile::from_path("/../sub1/../../sub2"), 0, b"y".to_vec()).with_creation(),
        ])
        .unwrap();
        assert!(fs.exists(Path::new("/sub2")).unwrap());
        assert!(!root.0.parent().unwrap().join("sub2").exists());

        // A lexical "a/../b" path resolves to b.
        write(&mut fs, "/a", b"");
        fs.renamev(&[(VfFile::from_path("/a"), VfFile::from_path("/x/../b"))])
            .unwrap();
        assert!(fs.exists(Path::new("/b")).unwrap());
        assert!(!fs.exists(Path::new("/x")).unwrap());
    }

    #[test]
    fn dummy_root_resolves_absolute_symlink_targets_inside_root() {
        let (root, mut fs) = fs("sandbox-symlink");
        write(&mut fs, "/target", b"inside");

        // An absolute target is chroot-relative: "/target" is the root's
        // "target", so reads through the link work.
        fs.symlink(Path::new("/target"), Path::new("/abs-link"))
            .unwrap();
        assert_eq!(
            fs.read(&VfFile::from_path("/abs-link"), 0, 6).unwrap(),
            b"inside"
        );
        // ".." components in an absolute target are clamped at the root.
        fs.symlink(Path::new("/sub/../target"), Path::new("/dotdot-link"))
            .unwrap();
        assert_eq!(
            fs.read(&VfFile::from_path("/dotdot-link"), 0, 6).unwrap(),
            b"inside"
        );

        // An OS-absolute path (e.g. the root's parent) is treated as a
        // root-relative path: it cannot escape or touch the outside file.
        let outside = root.0.parent().unwrap().join("outside-target");
        std::fs::write(&outside, b"outside").unwrap();
        let inside_target = root.0.join(
            outside
                .strip_prefix("/")
                .unwrap()
                .to_string_lossy()
                .into_owned(),
        );
        fs.symlink(Path::new(&outside), Path::new("/evil")).unwrap();
        assert_eq!(
            fs.readv(&[ReadOp::at(VfFile::from_path("/evil"), 0, 8)])
                .unwrap_err()
                .err_no(),
            ERR_NOENT,
            "resolves inside the root where nothing exists yet"
        );
        assert_eq!(std::fs::read(&outside).unwrap(), b"outside");

        // Creating through a chroot-relative absolute target lands inside the
        // root.
        fs.mkdir(Path::new("/subdir"), 0o755).unwrap();
        fs.symlink(Path::new("/subdir/created-inside"), Path::new("/evil3"))
            .unwrap();
        fs.writev(&[WriteOp::at(VfFile::from_path("/evil3"), 0, b"x".to_vec()).with_creation()])
            .unwrap();
        assert_eq!(
            fs.read(&VfFile::from_path("/subdir/created-inside"), 0, 1)
                .unwrap(),
            b"x"
        );

        // The link itself can still be inspected and removed (no-follow).
        assert_eq!(
            fs.lstat(Path::new("/abs-link")).unwrap().ftype,
            VfType::Symlink
        );
        fs.readlink(Path::new("/abs-link")).unwrap();
        fs.unlink(Path::new("/abs-link")).unwrap();

        // A dangling symlink to an absolute path still cannot touch the
        // outside of the root when creating through it.
        let dangling = root.0.parent().unwrap().join("never-created");
        fs.symlink(Path::new(&dangling), Path::new("/evil2"))
            .unwrap();
        assert_eq!(
            fs.writev(
                &[WriteOp::at(VfFile::from_path("/evil2"), 0, b"x".to_vec()).with_creation()]
            )
            .unwrap_err()
            .err_no(),
            ERR_NOENT,
            "the chroot-relative target's parent does not exist"
        );
        assert!(!dangling.exists());
        assert!(!inside_target.exists(), "nothing was created inside either");
        let _ = std::fs::remove_file(&outside);

        // A dangling relative symlink whose target is inside the root is
        // created through (POSIX O_CREAT semantics).
        fs.symlink(Path::new("internal-target"), Path::new("/ok-link"))
            .unwrap();
        fs.writev(&[WriteOp::at(VfFile::from_path("/ok-link"), 0, b"z".to_vec()).with_creation()])
            .unwrap();
        assert_eq!(
            fs.read(&VfFile::from_path("/internal-target"), 0, 1)
                .unwrap(),
            b"z"
        );
    }

    #[test]
    fn dummy_open_by_path_abs_is_root_relative() {
        let (_root, mut fs) = fs("open-abs");
        let fd = fs
            .open_by_path(
                VfPathBase::Abs,
                Path::new("rel"),
                libc::O_CREAT | libc::O_RDWR,
                0o644,
            )
            .unwrap();
        fs.writev(&[WriteOp::new(fd.clone(), VfOffset::At(0), b"x".to_vec())])
            .unwrap();
        fs.close(&fd).unwrap();
        assert!(fs.exists(Path::new("/rel")).unwrap());
    }

    #[test]
    fn dummy_reports_special_file_types() {
        let (root, mut fs) = fs("special-types");
        let real = root.0.join("fifo");
        let c = std::ffi::CString::new(real.to_str().unwrap()).unwrap();
        unsafe { libc::mkfifo(c.as_ptr(), 0o644) };

        let sock_path = root.0.join("sock");
        let sock_c = std::ffi::CString::new(sock_path.to_str().unwrap()).unwrap();
        let fd = unsafe {
            let fd = libc::socket(libc::AF_UNIX, libc::SOCK_STREAM, 0);
            let mut addr: libc::sockaddr_un = std::mem::zeroed();
            addr.sun_family = libc::AF_UNIX as libc::sa_family_t;
            let bytes = sock_c.as_bytes();
            for (i, b) in bytes.iter().take(107).enumerate() {
                addr.sun_path[i] = *b as libc::c_char;
            }
            libc::bind(
                fd,
                &addr as *const libc::sockaddr_un as *const libc::sockaddr,
                std::mem::size_of::<libc::sockaddr_un>() as libc::socklen_t,
            );
            fd
        };
        assert!(fd >= 0);

        assert_eq!(fs.stat(Path::new("/fifo")).unwrap().ftype, VfType::Fifo);
        assert_eq!(fs.lstat(Path::new("/fifo")).unwrap().ftype, VfType::Fifo);
        assert_eq!(fs.stat(Path::new("/sock")).unwrap().ftype, VfType::Socket);
        let listed = fs
            .listdir(Path::new("/"), AttrMask::default(), 0, false)
            .unwrap();
        assert!(listed.iter().any(|e| e.ftype == VfType::Fifo));
        assert!(listed.iter().any(|e| e.ftype == VfType::Socket));

        unsafe { libc::close(fd) };
    }

    #[test]
    fn dummy_descriptor_sees_external_truncation() {
        let (root, mut fs) = fs("ext-trunc");
        write(&mut fs, "/f", b"0123456789");
        let fd = fs.open(Path::new("/f"), libc::O_RDWR, 0).unwrap();
        let real = root.0.join("f");
        std::fs::OpenOptions::new()
            .write(true)
            .open(&real)
            .unwrap()
            .set_len(3)
            .unwrap();
        let r = fs
            .readv(&[ReadOp::new(fd.clone(), VfOffset::At(0), 10)])
            .unwrap();
        assert_eq!(r[0].data, b"012", "descriptor sees the new size");
        fs.close(&fd).unwrap();
    }

    #[test]
    fn dummy_cwd_dotdot_stays_in_root() {
        let (root, mut fs) = fs("cwd-dotdot");
        fs.mkdir(Path::new("/a"), 0o755).unwrap();
        fs.chdir(Path::new("/a")).unwrap();

        fs.writev(&[WriteOp::at(VfFile::from_path("../x"), 0, b"1".to_vec()).with_creation()])
            .unwrap();
        fs.writev(&[WriteOp::at(VfFile::from_path("a/../y"), 0, b"2".to_vec()).with_creation()])
            .unwrap();
        assert!(fs.exists(Path::new("/x")).unwrap());
        // From cwd /a, "a/../y" resolves to /a/y (the ".." cancels the "a").
        assert!(fs.exists(Path::new("/a/y")).unwrap());
        assert!(!fs.exists(Path::new("/y")).unwrap());
        assert!(!root.0.parent().unwrap().join("x").exists());
        assert!(!root.0.parent().unwrap().join("y").exists());

        // ".." from the root clamps at the root instead of escaping.
        fs.chdir(Path::new("/")).unwrap();
        fs.writev(&[WriteOp::at(VfFile::from_path("../z"), 0, b"3".to_vec()).with_creation()])
            .unwrap();
        assert!(fs.exists(Path::new("/z")).unwrap());
        assert!(!root.0.parent().unwrap().join("z").exists());
    }

    #[test]
    fn dummy_named_attr_detection() {
        use std::ffi::CString;
        let (root, mut fs) = fs("xattr");
        write(&mut fs, "/f", b"x");
        let real = root.0.join("f");
        let real = real.to_string_lossy().into_owned();
        let c = CString::new(real).unwrap();
        let name = CString::new("user.test").unwrap();
        let val = b"v";
        let rc = unsafe {
            libc::setxattr(
                c.as_ptr(),
                name.as_ptr(),
                val.as_ptr() as *const libc::c_void,
                val.len(),
                0,
            )
        };
        assert_eq!(rc, 0, "setxattr");

        let mut a = VfAttrs {
            file: VfFile::from_path("/f"),
            masks: AttrMask::NAMED_ATTR,
            ..VfAttrs::default()
        };
        fs.getattrsv(std::slice::from_mut(&mut a)).unwrap();
        assert!(a.has_named_attr);
        assert!(a.returned.contains(AttrMask::NAMED_ATTR));
    }

    // ------------------------------------------------------------------
    // Attributes: returned tracking, strict setattrsv, lsetattrsv
    // ------------------------------------------------------------------

    #[test]
    fn getattrsv_reports_returned_mask() {
        let (_root, mut fs) = fs("returned");
        write(&mut fs, "/f", b"x");

        let a = fs.stat(Path::new("/f")).unwrap();
        assert_eq!(a.returned, AttrMask::stat());
        assert!(a.returned.contains(AttrMask::MODE));

        let mut a = VfAttrs {
            file: VfFile::from_path("/f"),
            masks: AttrMask::MODE | AttrMask::SIZE | AttrMask::MTIME,
            ..VfAttrs::default()
        };
        fs.getattrsv(std::slice::from_mut(&mut a)).unwrap();
        assert_eq!(
            a.returned,
            AttrMask::MODE | AttrMask::SIZE | AttrMask::MTIME
        );
    }

    #[test]
    fn setattrsv_rejects_unsupported_bits() {
        let (_root, mut fs) = fs("setattr-strict");
        write(&mut fs, "/f", b"x");

        let mut a = VfAttrs {
            file: VfFile::from_path("/f"),
            masks: AttrMask::MTIME,
            mtime_sec: 1,
            ..VfAttrs::default()
        };
        assert_eq!(
            fs.setattrsv(std::slice::from_ref(&a)).unwrap_err().err_no(),
            VF_ERR_UNSUPPORTED
        );

        a.masks = AttrMask::MODE | AttrMask::MTIME;
        assert_eq!(
            fs.setattrsv(std::slice::from_ref(&a)).unwrap_err().err_no(),
            VF_ERR_UNSUPPORTED
        );

        // MODE-only still works.
        a.masks = AttrMask::MODE;
        a.mode = 0o640;
        fs.setattrsv(std::slice::from_ref(&a)).unwrap();
        assert_eq!(fs.lstat(Path::new("/f")).unwrap().mode & 0o7777, 0o640);
    }

    #[test]
    fn lsetattrsv_does_not_follow_symlinks() {
        let (_root, mut fs) = fs("lsetattr");
        write(&mut fs, "/target", b"x");
        fs.symlink(Path::new("/target"), Path::new("/link"))
            .unwrap();

        // No portable lchmod: the dummy backend refuses symlinks instead of
        // silently following them.
        let a = VfAttrs {
            file: VfFile::from_path("/link"),
            masks: AttrMask::MODE,
            mode: 0o600,
            ..VfAttrs::default()
        };
        assert_eq!(
            fs.lsetattrsv(std::slice::from_ref(&a))
                .unwrap_err()
                .err_no(),
            VF_ERR_UNSUPPORTED
        );

        // Regular files are set normally.
        let a = VfAttrs {
            file: VfFile::from_path("/target"),
            masks: AttrMask::MODE,
            mode: 0o600,
            ..VfAttrs::default()
        };
        fs.lsetattrsv(std::slice::from_ref(&a)).unwrap();
        assert_eq!(fs.lstat(Path::new("/target")).unwrap().mode & 0o7777, 0o600);
    }

    // ------------------------------------------------------------------
    // exists / file_type use lstat semantics
    // ------------------------------------------------------------------

    #[test]
    fn exists_and_file_type_use_lstat_semantics() {
        let (_root, mut fs) = fs("lstat");
        write(&mut fs, "/f", b"x");
        fs.symlink(Path::new("missing-target"), Path::new("/dangling"))
            .unwrap();

        assert!(fs.exists(Path::new("/dangling")).unwrap());
        assert_eq!(
            fs.file_type(Path::new("/dangling")).unwrap(),
            VfType::Symlink
        );
        assert_eq!(fs.file_type(Path::new("/f")).unwrap(), VfType::Regular);
    }

    // ------------------------------------------------------------------
    // openv length contract, listdir limits, walk via dyn VecFs
    // ------------------------------------------------------------------

    #[test]
    fn openv_rejects_mismatched_lengths() {
        let (_root, mut fs) = fs("openv");
        use libc::O_CREAT;
        let e = fs
            .openv(&[Path::new("/a"), Path::new("/b")], &[O_CREAT], &[0o644])
            .unwrap_err();
        assert_eq!((e.index(), e.err_no()), (0, ERR_INVAL));
    }

    #[test]
    fn listdir_zero_max_count_is_unlimited() {
        let (_root, mut fs) = fs("listdir");
        fs.mkdir(Path::new("/d"), 0o755).unwrap();
        write(&mut fs, "/d/a", b"1");
        write(&mut fs, "/d/b", b"2");

        let all = fs
            .listdir(Path::new("/d"), AttrMask::default(), 0, false)
            .unwrap();
        assert_eq!(all.len(), 2);
        let one = fs
            .listdir(Path::new("/d"), AttrMask::default(), 1, false)
            .unwrap();
        assert_eq!(one.len(), 1);
    }

    #[test]
    fn walk_works_through_dyn_vecfs() {
        let (_root, mut fs) = fs("walk-dyn");
        fs.mkdir(Path::new("/sub"), 0o755).unwrap();
        write(&mut fs, "/sub/a", b"1");

        let mut dyn_fs: Box<dyn VecFs> = Box::new(fs);
        let mut visited: Vec<String> = Vec::new();
        let entries = dyn_fs
            .walk(Path::new(""), AttrMask::stat(), &mut |dir, _| {
                visited.push(dir.display().to_string())
            })
            .unwrap();
        assert_eq!(visited.len(), 2); // root + /sub
        assert_eq!(entries.len(), 2);
        let sub = entries.iter().find(|w| w.path.ends_with("sub")).unwrap();
        assert_eq!(sub.entries.len(), 1);
        assert_eq!(sub.entries[0].ftype, VfType::Regular);
    }

    #[test]
    fn lcopyv_copies_symlinks_as_symlinks() {
        let (_root, mut fs) = fs("lcopyv");
        write(&mut fs, "/target", b"data");
        fs.symlink(Path::new("target"), Path::new("/link")).unwrap();

        let pair = ExtentPair::new("/link", 0, "/link-copy", 0, None);
        fs.lcopyv(std::slice::from_ref(&pair)).unwrap();
        assert_eq!(
            fs.file_type(Path::new("/link-copy")).unwrap(),
            VfType::Symlink
        );
        assert_eq!(
            fs.readlink(Path::new("/link-copy")).unwrap(),
            fs.readlink(Path::new("/link")).unwrap()
        );

        // dupv copies the target's data instead.
        let pair = ExtentPair::new("/link", 0, "/link-dup", 0, None);
        fs.dupv(std::slice::from_ref(&pair)).unwrap();
        assert_eq!(
            fs.file_type(Path::new("/link-dup")).unwrap(),
            VfType::Regular
        );
        assert_eq!(
            fs.read(&VfFile::from_path("/link-dup"), 0, 4).unwrap(),
            b"data"
        );
    }
}