path: root/devices/src/virtio/fs/passthrough.rs

// Copyright 2019 The Chromium OS Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

use std::collections::btree_map;
use std::collections::BTreeMap;
use std::ffi::{CStr, CString};
use std::fs::File;
use std::io;
use std::mem::{self, size_of, MaybeUninit};
use std::os::unix::io::{AsRawFd, FromRawFd, RawFd};
use std::str::FromStr;
use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
use std::sync::{Arc, RwLock};
use std::time::Duration;

use data_model::DataInit;
use libc;
use sync::Mutex;
use sys_util::{error, ioctl_ior_nr, ioctl_iow_nr, ioctl_with_mut_ptr, ioctl_with_ptr};

use crate::virtio::fs::filesystem::{
    Context, DirEntry, Entry, FileSystem, FsOptions, GetxattrReply, IoctlFlags, IoctlIovec,
    IoctlReply, ListxattrReply, OpenOptions, SetattrValid, ZeroCopyReader, ZeroCopyWriter,
};
use crate::virtio::fs::fuse;
use crate::virtio::fs::multikey::MultikeyBTreeMap;

const CURRENT_DIR_CSTR: &[u8] = b".\0";
const PARENT_DIR_CSTR: &[u8] = b"..\0";
const EMPTY_CSTR: &[u8] = b"\0";
const ROOT_CSTR: &[u8] = b"/\0";
const PROC_CSTR: &[u8] = b"/proc\0";

const FSCRYPT_KEY_DESCRIPTOR_SIZE: usize = 8;

#[repr(C)]
#[derive(Debug, Clone, Copy)]
struct fscrypt_policy_v1 {
    _version: u8,
    _contents_encryption_mode: u8,
    _filenames_encryption_mode: u8,
    _flags: u8,
    _master_key_descriptor: [u8; FSCRYPT_KEY_DESCRIPTOR_SIZE],
}
unsafe impl DataInit for fscrypt_policy_v1 {}

ioctl_ior_nr!(FS_IOC_SET_ENCRYPTION_POLICY, 0x66, 19, fscrypt_policy_v1);
ioctl_iow_nr!(FS_IOC_GET_ENCRYPTION_POLICY, 0x66, 21, fscrypt_policy_v1);

type Inode = u64;
type Handle = u64;

#[derive(Clone, Copy, PartialOrd, Ord, PartialEq, Eq)]
struct InodeAltKey {
    ino: libc::ino64_t,
    dev: libc::dev_t,
}

struct InodeData {
    inode: Inode,
    // Most of these aren't actually files but ¯\_(ツ)_/¯.
    file: File,
    refcount: AtomicU64,
}

struct HandleData {
    inode: Inode,
    file: Mutex<File>,
}

#[repr(C, packed)]
#[derive(Clone, Copy, Debug)]
struct LinuxDirent64 {
    d_ino: libc::ino64_t,
    d_off: libc::off64_t,
    d_reclen: libc::c_ushort,
    d_ty: libc::c_uchar,
}
unsafe impl DataInit for LinuxDirent64 {}

macro_rules! scoped_cred {
    ($name:ident, $ty:ty, $syscall_nr:expr) => {
        #[derive(Debug)]
        struct $name {
            old: $ty,
        }

        impl $name {
            // Changes the effective uid/gid of the current thread to `val`. Changes the thread's
            // credentials back to `old` when the returned struct is dropped.
            fn new(val: $ty, old: $ty) -> io::Result<Option<$name>> {
                if val == old {
                    // Nothing to do since we already have the correct value.
                    return Ok(None);
                }

                // We want credential changes to be per-thread because otherwise
                // we might interfere with operations being carried out on other
                // threads with different uids/gids.  However, posix requires that
                // all threads in a process share the same credentials.  To do this
                // libc uses signals to ensure that when one thread changes its
                // credentials the other threads do the same thing.
                //
                // So instead we invoke the syscall directly in order to get around
                // this limitation.  Another option is to use the setfsuid and
                // setfsgid systems calls.   However since those calls have no way to
                // return an error, it's preferable to do this instead.

                // This call is safe because it doesn't modify any memory and we
                // check the return value.
                let res = unsafe { libc::syscall($syscall_nr, -1, val, -1) };
                if res == 0 {
                    Ok(Some($name { old }))
                } else {
                    Err(io::Error::last_os_error())
                }
            }
        }

        impl Drop for $name {
            fn drop(&mut self) {
                let res = unsafe { libc::syscall($syscall_nr, -1, self.old, -1) };
                if res < 0 {
                    error!(
                        "failed to change credentials back to {}: {}",
                        self.old,
                        io::Error::last_os_error(),
                    );
                }
            }
        }
    };
}
#[cfg(not(target_arch = "arm"))]
scoped_cred!(ScopedUid, libc::uid_t, libc::SYS_setresuid);
#[cfg(target_arch = "arm")]
scoped_cred!(ScopedUid, libc::uid_t, libc::SYS_setresuid32);

#[cfg(not(target_arch = "arm"))]
scoped_cred!(ScopedGid, libc::gid_t, libc::SYS_setresgid);
#[cfg(target_arch = "arm")]
scoped_cred!(ScopedGid, libc::gid_t, libc::SYS_setresgid32);

#[cfg(not(target_arch = "arm"))]
const SYS_GETEUID: libc::c_long = libc::SYS_geteuid;
#[cfg(target_arch = "arm")]
const SYS_GETEUID: libc::c_long = libc::SYS_geteuid32;

#[cfg(not(target_arch = "arm"))]
const SYS_GETEGID: libc::c_long = libc::SYS_getegid;
#[cfg(target_arch = "arm")]
const SYS_GETEGID: libc::c_long = libc::SYS_getegid32;

thread_local! {
    // Both these calls are safe because they take no parameters, and only return an integer value.
    // The kernel also guarantees that they can never fail.
    static THREAD_EUID: libc::uid_t = unsafe { libc::syscall(SYS_GETEUID) as libc::uid_t };
    static THREAD_EGID: libc::gid_t = unsafe { libc::syscall(SYS_GETEGID) as libc::gid_t };
}

fn set_creds(
    uid: libc::uid_t,
    gid: libc::gid_t,
) -> io::Result<(Option<ScopedUid>, Option<ScopedGid>)> {
    let olduid = THREAD_EUID.with(|uid| *uid);
    let oldgid = THREAD_EGID.with(|gid| *gid);

    // We have to change the gid before we change the uid because if we change the uid first then we
    // lose the capability to change the gid.  However changing back can happen in any order.
    ScopedGid::new(gid, oldgid).and_then(|gid| Ok((ScopedUid::new(uid, olduid)?, gid)))
}

fn ebadf() -> io::Error {
    io::Error::from_raw_os_error(libc::EBADF)
}

fn stat(f: &File) -> io::Result<libc::stat64> {
    let mut st = MaybeUninit::<libc::stat64>::zeroed();

    // Safe because this is a constant value and a valid C string.
    let pathname = unsafe { CStr::from_bytes_with_nul_unchecked(EMPTY_CSTR) };

    // Safe because the kernel will only write data in `st` and we check the return
    // value.
    let res = unsafe {
        libc::fstatat64(
            f.as_raw_fd(),
            pathname.as_ptr(),
            st.as_mut_ptr(),
            libc::AT_EMPTY_PATH | libc::AT_SYMLINK_NOFOLLOW,
        )
    };
    if res >= 0 {
        // Safe because the kernel guarantees that the struct is now fully initialized.
        Ok(unsafe { st.assume_init() })
    } else {
        Err(io::Error::last_os_error())
    }
}

/// The caching policy that the file system should report to the FUSE client. By default the FUSE
/// protocol uses close-to-open consistency. This means that any cached contents of the file are
/// invalidated the next time that file is opened.
#[derive(Debug, Clone)]
pub enum CachePolicy {
    /// The client should never cache file data and all I/O should be directly forwarded to the
    /// server. This policy must be selected when file contents may change without the knowledge of
    /// the FUSE client (i.e., the file system does not have exclusive access to the directory).
    Never,

    /// The client is free to choose when and how to cache file data. This is the default policy and
    /// uses close-to-open consistency as described in the enum documentation.
    Auto,

    /// The client should always cache file data. This means that the FUSE client will not
    /// invalidate any cached data that was returned by the file system the last time the file was
    /// opened. This policy should only be selected when the file system has exclusive access to the
    /// directory.
    Always,
}

impl FromStr for CachePolicy {
    type Err = &'static str;

    fn from_str(s: &str) -> Result<Self, Self::Err> {
        match s {
            "never" | "Never" | "NEVER" => Ok(CachePolicy::Never),
            "auto" | "Auto" | "AUTO" => Ok(CachePolicy::Auto),
            "always" | "Always" | "ALWAYS" => Ok(CachePolicy::Always),
            _ => Err("invalid cache policy"),
        }
    }
}

impl Default for CachePolicy {
    fn default() -> Self {
        CachePolicy::Auto
    }
}

/// Options that configure the behavior of the file system.
#[derive(Debug, Clone)]
pub struct Config {
    /// How long the FUSE client should consider directory entries to be valid. If the contents of a
    /// directory can only be modified by the FUSE client (i.e., the file system has exclusive
    /// access), then this should be a large value.
    ///
    /// The default value for this option is 5 seconds.
    pub entry_timeout: Duration,

    /// How long the FUSE client should consider file and directory attributes to be valid. If the
    /// attributes of a file or directory can only be modified by the FUSE client (i.e., the file
    /// system has exclusive access), then this should be set to a large value.
    ///
    /// The default value for this option is 5 seconds.
    pub attr_timeout: Duration,

    /// The caching policy the file system should use. See the documentation of `CachePolicy` for
    /// more details.
    pub cache_policy: CachePolicy,

    /// Whether the file system should enabled writeback caching. This can improve performance as it
    /// allows the FUSE client to cache and coalesce multiple writes before sending them to the file
    /// system. However, enabling this option can increase the risk of data corruption if the file
    /// contents can change without the knowledge of the FUSE client (i.e., the server does **NOT**
    /// have exclusive access). Additionally, the file system should have read access to all files
    /// in the directory it is serving as the FUSE client may send read requests even for files
    /// opened with `O_WRONLY`.
    ///
    /// Therefore callers should only enable this option when they can guarantee that: 1) the file
    /// system has exclusive access to the directory and 2) the file system has read permissions for
    /// all files in that directory.
    ///
    /// The default value for this option is `false`.
    pub writeback: bool,
}

impl Default for Config {
    fn default() -> Self {
        Config {
            entry_timeout: Duration::from_secs(5),
            attr_timeout: Duration::from_secs(5),
            cache_policy: Default::default(),
            writeback: false,
        }
    }
}

/// A file system that simply "passes through" all requests it receives to the underlying file
/// system. To keep the implementation simple it servers the contents of its root directory. Users
/// that wish to serve only a specific directory should set up the environment so that that
/// directory ends up as the root of the file system process. One way to accomplish this is via a
/// combination of mount namespaces and the pivot_root system call.
pub struct PassthroughFs {
    // File descriptors for various points in the file system tree. These fds are always opened with
    // the `O_PATH` option so they cannot be used for reading or writing any data. See the
    // documentation of the `O_PATH` flag in `open(2)` for more details on what one can and cannot
    // do with an fd opened with this flag.
    inodes: RwLock<MultikeyBTreeMap<Inode, InodeAltKey, Arc<InodeData>>>,
    next_inode: AtomicU64,

    // File descriptors for open files and directories. Unlike the fds in `inodes`, these _can_ be
    // used for reading and writing data.
    handles: RwLock<BTreeMap<Handle, Arc<HandleData>>>,
    next_handle: AtomicU64,

    // File descriptor pointing to the `/proc` directory. This is used to convert an fd from
    // `inodes` into one that can go into `handles`. This is accomplished by reading the
    // `self/fd/{}` symlink. We keep an open fd here in case the file system tree that we are meant
    // to be serving doesn't have access to `/proc`.
    proc: File,

    // Whether writeback caching is enabled for this directory. This will only be true when
    // `cfg.writeback` is true and `init` was called with `FsOptions::WRITEBACK_CACHE`.
    writeback: AtomicBool,

    cfg: Config,
}

impl PassthroughFs {
    pub fn new(cfg: Config) -> io::Result<PassthroughFs> {
        // Safe because this is a constant value and a valid C string.
        let proc_cstr = unsafe { CStr::from_bytes_with_nul_unchecked(PROC_CSTR) };

        // Safe because this doesn't modify any memory and we check the return value.
        let fd = unsafe {
            libc::openat(
                libc::AT_FDCWD,
                proc_cstr.as_ptr(),
                libc::O_PATH | libc::O_NOFOLLOW | libc::O_CLOEXEC,
            )
        };
        if fd < 0 {
            return Err(io::Error::last_os_error());
        }

        // Safe because we just opened this fd.
        let proc = unsafe { File::from_raw_fd(fd) };

        Ok(PassthroughFs {
            inodes: RwLock::new(MultikeyBTreeMap::new()),
            next_inode: AtomicU64::new(fuse::ROOT_ID + 1),

            handles: RwLock::new(BTreeMap::new()),
            next_handle: AtomicU64::new(0),

            proc,

            writeback: AtomicBool::new(false),
            cfg,
        })
    }

    pub fn keep_fds(&self) -> Vec<RawFd> {
        vec![self.proc.as_raw_fd()]
    }

    fn open_inode(&self, inode: Inode, mut flags: i32) -> io::Result<File> {
        let data = self
            .inodes
            .read()
            .unwrap()
            .get(&inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let pathname = CString::new(format!("self/fd/{}", data.file.as_raw_fd()))
            .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))?;

        // When writeback caching is enabled, the kernel may send read requests even if the
        // userspace program opened the file write-only. So we need to ensure that we have opened
        // the file for reading as well as writing.
        let writeback = self.writeback.load(Ordering::Relaxed);
        if writeback && flags & libc::O_ACCMODE == libc::O_WRONLY {
            flags &= !libc::O_ACCMODE;
            flags |= libc::O_RDWR;
        }

        // When writeback caching is enabled the kernel is responsible for handling `O_APPEND`.
        // However, this breaks atomicity as the file may have changed on disk, invalidating the
        // cached copy of the data in the kernel and the offset that the kernel thinks is the end of
        // the file. Just allow this for now as it is the user's responsibility to enable writeback
        // caching only for directories that are not shared. It also means that we need to clear the
        // `O_APPEND` flag.
        if writeback && flags & libc::O_APPEND != 0 {
            flags &= !libc::O_APPEND;
        }

        // Safe because this doesn't modify any memory and we check the return value. We don't
        // really check `flags` because if the kernel can't handle poorly specified flags then we
        // have much bigger problems. Also, clear the `O_NOFOLLOW` flag if it is set since we need
        // to follow the `/proc/self/fd` symlink to get the file.
        let fd = unsafe {
            libc::openat(
                self.proc.as_raw_fd(),
                pathname.as_ptr(),
                (flags | libc::O_CLOEXEC) & !(libc::O_NOFOLLOW | libc::O_DIRECT),
            )
        };
        if fd < 0 {
            return Err(io::Error::last_os_error());
        }

        // Safe because we just opened this fd.
        Ok(unsafe { File::from_raw_fd(fd) })
    }

    fn do_lookup(&self, parent: Inode, name: &CStr) -> io::Result<Entry> {
        let p = self
            .inodes
            .read()
            .unwrap()
            .get(&parent)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        // Safe because this doesn't modify any memory and we check the return value.
        let fd = unsafe {
            libc::openat(
                p.file.as_raw_fd(),
                name.as_ptr(),
                libc::O_PATH | libc::O_NOFOLLOW | libc::O_CLOEXEC,
            )
        };
        if fd < 0 {
            return Err(io::Error::last_os_error());
        }

        // Safe because we just opened this fd.
        let f = unsafe { File::from_raw_fd(fd) };

        let st = stat(&f)?;

        let altkey = InodeAltKey {
            ino: st.st_ino,
            dev: st.st_dev,
        };
        let data = self.inodes.read().unwrap().get_alt(&altkey).map(Arc::clone);

        let inode = if let Some(data) = data {
            // Matches with the release store in `forget`.
            data.refcount.fetch_add(1, Ordering::Acquire);
            data.inode
        } else {
            // There is a possible race here where 2 threads end up adding the same file
            // into the inode list.  However, since each of those will get a unique Inode
            // value and unique file descriptors this shouldn't be that much of a problem.
            let inode = self.next_inode.fetch_add(1, Ordering::Relaxed);
            self.inodes.write().unwrap().insert(
                inode,
                InodeAltKey {
                    ino: st.st_ino,
                    dev: st.st_dev,
                },
                Arc::new(InodeData {
                    inode,
                    file: f,
                    refcount: AtomicU64::new(1),
                }),
            );

            inode
        };

        Ok(Entry {
            inode,
            generation: 0,
            attr: st,
            attr_timeout: self.cfg.attr_timeout.clone(),
            entry_timeout: self.cfg.entry_timeout.clone(),
        })
    }

    fn do_readdir<F>(
        &self,
        inode: Inode,
        handle: Handle,
        size: u32,
        offset: u64,
        mut add_entry: F,
    ) -> io::Result<()>
    where
        F: FnMut(DirEntry) -> io::Result<usize>,
    {
        if size == 0 {
            return Ok(());
        }

        let data = self
            .handles
            .read()
            .unwrap()
            .get(&handle)
            .filter(|hd| hd.inode == inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let mut buf = Vec::with_capacity(size as usize);
        buf.resize(size as usize, 0);

        {
            // Since we are going to work with the kernel offset, we have to acquire the file lock
            // for both the `lseek64` and `getdents64` syscalls to ensure that no other thread
            // changes the kernel offset while we are using it.
            let dir = data.file.lock();

            // Safe because this doesn't modify any memory and we check the return value.
            let res =
                unsafe { libc::lseek64(dir.as_raw_fd(), offset as libc::off64_t, libc::SEEK_SET) };
            if res < 0 {
                return Err(io::Error::last_os_error());
            }

            // Safe because the kernel guarantees that it will only write to `buf` and we check the
            // return value.
            let res = unsafe {
                libc::syscall(
                    libc::SYS_getdents64,
                    dir.as_raw_fd(),
                    buf.as_mut_ptr() as *mut LinuxDirent64,
                    size as libc::c_int,
                )
            };
            if res < 0 {
                return Err(io::Error::last_os_error());
            }
            buf.resize(res as usize, 0);

            // Explicitly drop the lock so that it's not held while we fill in the fuse buffer.
            mem::drop(dir);
        }

        let mut rem = &buf[..];
        while rem.len() > 0 {
            // We only use debug asserts here because these values are coming from the kernel and we
            // trust them implicitly.
            debug_assert!(
                rem.len() >= size_of::<LinuxDirent64>(),
                "not enough space left in `rem`"
            );

            let (front, back) = rem.split_at(size_of::<LinuxDirent64>());

            let dirent64 =
                LinuxDirent64::from_slice(front).expect("unable to get LinuxDirent64 from slice");

            let namelen = dirent64.d_reclen as usize - size_of::<LinuxDirent64>();
            debug_assert!(namelen <= back.len(), "back is smaller than `namelen`");

            let name = &back[..namelen];
            let res = if name.starts_with(CURRENT_DIR_CSTR) || name.starts_with(PARENT_DIR_CSTR) {
                // We don't want to report the "." and ".." entries. However, returning `Ok(0)` will
                // break the loop so return `Ok` with a non-zero value instead.
                Ok(1)
            } else {
                add_entry(DirEntry {
                    ino: dirent64.d_ino,
                    offset: dirent64.d_off as u64,
                    type_: dirent64.d_ty as u32,
                    name,
                })
            };

            debug_assert!(
                rem.len() >= dirent64.d_reclen as usize,
                "rem is smaller than `d_reclen`"
            );

            match res {
                Ok(0) => break,
                Ok(_) => rem = &rem[dirent64.d_reclen as usize..],
                Err(e) => return Err(e),
            }
        }

        Ok(())
    }

    fn do_open(&self, inode: Inode, flags: u32) -> io::Result<(Option<Handle>, OpenOptions)> {
        let file = Mutex::new(self.open_inode(inode, flags as i32)?);

        let handle = self.next_handle.fetch_add(1, Ordering::Relaxed);
        let data = HandleData { inode, file };

        self.handles.write().unwrap().insert(handle, Arc::new(data));

        let mut opts = OpenOptions::empty();
        match self.cfg.cache_policy {
            // We only set the direct I/O option on files.
            CachePolicy::Never => opts.set(
                OpenOptions::DIRECT_IO,
                flags & (libc::O_DIRECTORY as u32) == 0,
            ),
            CachePolicy::Always => {
                opts |= if flags & (libc::O_DIRECTORY as u32) == 0 {
                    OpenOptions::KEEP_CACHE
                } else {
                    OpenOptions::CACHE_DIR
                }
            }
            _ => {}
        };

        Ok((Some(handle), opts))
    }

    fn do_release(&self, inode: Inode, handle: Handle) -> io::Result<()> {
        let mut handles = self.handles.write().unwrap();

        if let btree_map::Entry::Occupied(e) = handles.entry(handle) {
            if e.get().inode == inode {
                // We don't need to close the file here because that will happen automatically when
                // the last `Arc` is dropped.
                e.remove();
                return Ok(());
            }
        }

        Err(ebadf())
    }

    fn do_getattr(&self, inode: Inode) -> io::Result<(libc::stat64, Duration)> {
        let data = self
            .inodes
            .read()
            .unwrap()
            .get(&inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let st = stat(&data.file)?;

        Ok((st, self.cfg.attr_timeout.clone()))
    }

    fn do_unlink(&self, parent: Inode, name: &CStr, flags: libc::c_int) -> io::Result<()> {
        let data = self
            .inodes
            .read()
            .unwrap()
            .get(&parent)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        // Safe because this doesn't modify any memory and we check the return value.
        let res = unsafe { libc::unlinkat(data.file.as_raw_fd(), name.as_ptr(), flags) };
        if res == 0 {
            Ok(())
        } else {
            Err(io::Error::last_os_error())
        }
    }

    fn get_encryption_policy(&self, handle: Handle) -> io::Result<IoctlReply> {
        let data = self
            .handles
            .read()
            .unwrap()
            .get(&handle)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let mut buf = MaybeUninit::<fscrypt_policy_v1>::zeroed();
        let file = data.file.lock();

        // Safe because the kernel will only write to `buf` and we check the return value.
        let res =
            unsafe { ioctl_with_mut_ptr(&*file, FS_IOC_GET_ENCRYPTION_POLICY(), buf.as_mut_ptr()) };
        if res < 0 {
            Ok(IoctlReply::Done(Err(io::Error::last_os_error())))
        } else {
            // Safe because the kernel guarantees that the policy is now initialized.
            let policy = unsafe { buf.assume_init() };
            Ok(IoctlReply::Done(Ok(policy.as_slice().to_vec())))
        }
    }

    fn set_encryption_policy<R: io::Read>(&self, handle: Handle, r: R) -> io::Result<IoctlReply> {
        let data = self
            .handles
            .read()
            .unwrap()
            .get(&handle)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let policy = fscrypt_policy_v1::from_reader(r)?;
        let file = data.file.lock();
        // Safe because the kernel will only read from `policy` and we check the return value.
        let res = unsafe { ioctl_with_ptr(&*file, FS_IOC_SET_ENCRYPTION_POLICY(), &policy) };
        if res < 0 {
            Ok(IoctlReply::Done(Err(io::Error::last_os_error())))
        } else {
            Ok(IoctlReply::Done(Ok(Vec::new())))
        }
    }
}

fn forget_one(
    inodes: &mut MultikeyBTreeMap<Inode, InodeAltKey, Arc<InodeData>>,
    inode: Inode,
    count: u64,
) {
    if let Some(data) = inodes.get(&inode) {
        // Acquiring the write lock on the inode map prevents new lookups from incrementing the
        // refcount but there is the possibility that a previous lookup already acquired a
        // reference to the inode data and is in the process of updating the refcount so we need
        // to loop here until we can decrement successfully.
        loop {
            let refcount = data.refcount.load(Ordering::Relaxed);

            // Saturating sub because it doesn't make sense for a refcount to go below zero and
            // we don't want misbehaving clients to cause integer overflow.
            let new_count = refcount.saturating_sub(count);

            // Synchronizes with the acquire load in `do_lookup`.
            if data
                .refcount
                .compare_and_swap(refcount, new_count, Ordering::Release)
                == refcount
            {
                if new_count == 0 {
                    // We just removed the last refcount for this inode. There's no need for an
                    // acquire fence here because we hold a write lock on the inode map and any
                    // thread that is waiting to do a forget on the same inode will have to wait
                    // until we release the lock. So there's is no other release store for us to
                    // synchronize with before deleting the entry.
                    inodes.remove(&inode);
                }
                break;
            }
        }
    }
}

impl FileSystem for PassthroughFs {
    type Inode = Inode;
    type Handle = Handle;

    fn init(&self, capable: FsOptions) -> io::Result<FsOptions> {
        // Safe because this is a constant value and a valid C string.
        let root = unsafe { CStr::from_bytes_with_nul_unchecked(ROOT_CSTR) };

        // Safe because this doesn't modify any memory and we check the return value.
        // We use `O_PATH` because we just want this for traversing the directory tree
        // and not for actually reading the contents.
        let fd = unsafe {
            libc::openat(
                libc::AT_FDCWD,
                root.as_ptr(),
                libc::O_PATH | libc::O_NOFOLLOW | libc::O_CLOEXEC,
            )
        };
        if fd < 0 {
            return Err(io::Error::last_os_error());
        }

        // Safe because we just opened this fd above.
        let f = unsafe { File::from_raw_fd(fd) };

        let st = stat(&f)?;

        // Safe because this doesn't modify any memory and there is no need to check the return
        // value because this system call always succeeds. We need to clear the umask here because
        // we want the client to be able to set all the bits in the mode.
        unsafe { libc::umask(0o000) };

        let mut inodes = self.inodes.write().unwrap();

        // Not sure why the root inode gets a refcount of 2 but that's what libfuse does.
        inodes.insert(
            fuse::ROOT_ID,
            InodeAltKey {
                ino: st.st_ino,
                dev: st.st_dev,
            },
            Arc::new(InodeData {
                inode: fuse::ROOT_ID,
                file: f,
                refcount: AtomicU64::new(2),
            }),
        );

        let mut opts = FsOptions::DO_READDIRPLUS | FsOptions::READDIRPLUS_AUTO;
        if self.cfg.writeback && capable.contains(FsOptions::WRITEBACK_CACHE) {
            opts |= FsOptions::WRITEBACK_CACHE;
            self.writeback.store(true, Ordering::Relaxed);
        }
        Ok(opts)
    }

    fn destroy(&self) {
        self.handles.write().unwrap().clear();
        self.inodes.write().unwrap().clear();
    }

    fn statfs(&self, _ctx: Context, inode: Inode) -> io::Result<libc::statvfs64> {
        let data = self
            .inodes
            .read()
            .unwrap()
            .get(&inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let mut out = MaybeUninit::<libc::statvfs64>::zeroed();

        // Safe because this will only modify `out` and we check the return value.
        let res = unsafe { libc::fstatvfs64(data.file.as_raw_fd(), out.as_mut_ptr()) };
        if res == 0 {
            // Safe because the kernel guarantees that `out` has been initialized.
            Ok(unsafe { out.assume_init() })
        } else {
            Err(io::Error::last_os_error())
        }
    }

    fn lookup(&self, _ctx: Context, parent: Inode, name: &CStr) -> io::Result<Entry> {
        self.do_lookup(parent, name)
    }

    fn forget(&self, _ctx: Context, inode: Inode, count: u64) {
        let mut inodes = self.inodes.write().unwrap();

        forget_one(&mut inodes, inode, count)
    }

    fn batch_forget(&self, _ctx: Context, requests: Vec<(Inode, u64)>) {
        let mut inodes = self.inodes.write().unwrap();

        for (inode, count) in requests {
            forget_one(&mut inodes, inode, count)
        }
    }

    fn opendir(
        &self,
        _ctx: Context,
        inode: Inode,
        flags: u32,
    ) -> io::Result<(Option<Handle>, OpenOptions)> {
        self.do_open(inode, flags | (libc::O_DIRECTORY as u32))
    }

    fn releasedir(
        &self,
        _ctx: Context,
        inode: Inode,
        _flags: u32,
        handle: Handle,
    ) -> io::Result<()> {
        self.do_release(inode, handle)
    }

    fn mkdir(
        &self,
        ctx: Context,
        parent: Inode,
        name: &CStr,
        mode: u32,
        umask: u32,
    ) -> io::Result<Entry> {
        let (_uid, _gid) = set_creds(ctx.uid, ctx.gid)?;
        let data = self
            .inodes
            .read()
            .unwrap()
            .get(&parent)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        // Safe because this doesn't modify any memory and we check the return value.
        let res = unsafe { libc::mkdirat(data.file.as_raw_fd(), name.as_ptr(), mode & !umask) };
        if res == 0 {
            self.do_lookup(parent, name)
        } else {
            Err(io::Error::last_os_error())
        }
    }

    fn rmdir(&self, _ctx: Context, parent: Inode, name: &CStr) -> io::Result<()> {
        self.do_unlink(parent, name, libc::AT_REMOVEDIR)
    }

    fn readdir<F>(
        &self,
        _ctx: Context,
        inode: Inode,
        handle: Handle,
        size: u32,
        offset: u64,
        add_entry: F,
    ) -> io::Result<()>
    where
        F: FnMut(DirEntry) -> io::Result<usize>,
    {
        self.do_readdir(inode, handle, size, offset, add_entry)
    }

    fn readdirplus<F>(
        &self,
        _ctx: Context,
        inode: Inode,
        handle: Handle,
        size: u32,
        offset: u64,
        mut add_entry: F,
    ) -> io::Result<()>
    where
        F: FnMut(DirEntry, Entry) -> io::Result<usize>,
    {
        self.do_readdir(inode, handle, size, offset, |dir_entry| {
            // Safe because the kernel guarantees that the buffer is nul-terminated. Additionally,
            // the kernel will pad the name with '\0' bytes up to 8-byte alignment and there's no
            // way for us to know exactly how many padding bytes there are. This would cause
            // `CStr::from_bytes_with_nul` to return an error because it would think there are
            // interior '\0' bytes. We trust the kernel to provide us with properly formatted data
            // so we'll just skip the checks here.
            let name = unsafe { CStr::from_bytes_with_nul_unchecked(dir_entry.name) };
            let entry = self.do_lookup(inode, name)?;

            add_entry(dir_entry, entry)
        })
    }

    fn open(
        &self,
        _ctx: Context,
        inode: Inode,
        flags: u32,
    ) -> io::Result<(Option<Handle>, OpenOptions)> {
        self.do_open(inode, flags)
    }

    fn release(
        &self,
        _ctx: Context,
        inode: Inode,
        _flags: u32,
        handle: Handle,
        _flush: bool,
        _flock_release: bool,
        _lock_owner: Option<u64>,
    ) -> io::Result<()> {
        self.do_release(inode, handle)
    }

    fn create(
        &self,
        ctx: Context,
        parent: Inode,
        name: &CStr,
        mode: u32,
        flags: u32,
        umask: u32,
    ) -> io::Result<(Entry, Option<Handle>, OpenOptions)> {
        let (_uid, _gid) = set_creds(ctx.uid, ctx.gid)?;
        let data = self
            .inodes
            .read()
            .unwrap()
            .get(&parent)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        // Safe because this doesn't modify any memory and we check the return value. We don't
        // really check `flags` because if the kernel can't handle poorly specified flags then we
        // have much bigger problems.
        let fd = unsafe {
            libc::openat(
                data.file.as_raw_fd(),
                name.as_ptr(),
                (flags as i32 | libc::O_CREAT | libc::O_CLOEXEC | libc::O_NOFOLLOW)
                    & !libc::O_DIRECT,
                mode & !(umask & 0o777),
            )
        };
        if fd < 0 {
            return Err(io::Error::last_os_error());
        }

        // Safe because we just opened this fd.
        let file = Mutex::new(unsafe { File::from_raw_fd(fd) });

        let entry = self.do_lookup(parent, name)?;

        let handle = self.next_handle.fetch_add(1, Ordering::Relaxed);
        let data = HandleData {
            inode: entry.inode,
            file,
        };

        self.handles.write().unwrap().insert(handle, Arc::new(data));

        let mut opts = OpenOptions::empty();
        match self.cfg.cache_policy {
            CachePolicy::Never => opts |= OpenOptions::DIRECT_IO,
            CachePolicy::Always => opts |= OpenOptions::KEEP_CACHE,
            _ => {}
        };

        Ok((entry, Some(handle), opts))
    }

    fn unlink(&self, _ctx: Context, parent: Inode, name: &CStr) -> io::Result<()> {
        self.do_unlink(parent, name, 0)
    }

    fn read<W: io::Write + ZeroCopyWriter>(
        &self,
        _ctx: Context,
        inode: Inode,
        handle: Handle,
        mut w: W,
        size: u32,
        offset: u64,
        _lock_owner: Option<u64>,
        _flags: u32,
    ) -> io::Result<usize> {
        let data = self
            .handles
            .read()
            .unwrap()
            .get(&handle)
            .filter(|hd| hd.inode == inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let mut f = data.file.lock();
        w.write_from(&mut f, size as usize, offset)
    }

    fn write<R: io::Read + ZeroCopyReader>(
        &self,
        ctx: Context,
        inode: Inode,
        handle: Handle,
        mut r: R,
        size: u32,
        offset: u64,
        _lock_owner: Option<u64>,
        _delayed_write: bool,
        _flags: u32,
    ) -> io::Result<usize> {
        // We need to change credentials during a write so that the kernel will remove setuid or
        // setgid bits from the file if it was written to by someone other than the owner.
        let (_uid, _gid) = set_creds(ctx.uid, ctx.gid)?;
        let data = self
            .handles
            .read()
            .unwrap()
            .get(&handle)
            .filter(|hd| hd.inode == inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let mut f = data.file.lock();
        r.read_to(&mut f, size as usize, offset)
    }

    fn getattr(
        &self,
        _ctx: Context,
        inode: Inode,
        _handle: Option<Handle>,
    ) -> io::Result<(libc::stat64, Duration)> {
        self.do_getattr(inode)
    }

    fn setattr(
        &self,
        _ctx: Context,
        inode: Inode,
        attr: libc::stat64,
        handle: Option<Handle>,
        valid: SetattrValid,
    ) -> io::Result<(libc::stat64, Duration)> {
        let inode_data = self
            .inodes
            .read()
            .unwrap()
            .get(&inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        enum Data {
            Handle(Arc<HandleData>, RawFd),
            ProcPath(CString),
        }

        // If we have a handle then use it otherwise get a new fd from the inode.
        let data = if let Some(handle) = handle {
            let hd = self
                .handles
                .read()
                .unwrap()
                .get(&handle)
                .filter(|hd| hd.inode == inode)
                .map(Arc::clone)
                .ok_or_else(ebadf)?;

            let fd = hd.file.lock().as_raw_fd();
            Data::Handle(hd, fd)
        } else {
            let pathname = CString::new(format!("self/fd/{}", inode_data.file.as_raw_fd()))
                .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))?;
            Data::ProcPath(pathname)
        };

        if valid.contains(SetattrValid::MODE) {
            // Safe because this doesn't modify any memory and we check the return value.
            let res = unsafe {
                match data {
                    Data::Handle(_, fd) => libc::fchmod(fd, attr.st_mode),
                    Data::ProcPath(ref p) => {
                        libc::fchmodat(self.proc.as_raw_fd(), p.as_ptr(), attr.st_mode, 0)
                    }
                }
            };
            if res < 0 {
                return Err(io::Error::last_os_error());
            }
        }

        if valid.intersects(SetattrValid::UID | SetattrValid::GID) {
            let uid = if valid.contains(SetattrValid::UID) {
                attr.st_uid
            } else {
                // Cannot use -1 here because these are unsigned values.
                ::std::u32::MAX
            };
            let gid = if valid.contains(SetattrValid::GID) {
                attr.st_gid
            } else {
                // Cannot use -1 here because these are unsigned values.
                ::std::u32::MAX
            };

            // Safe because this is a constant value and a valid C string.
            let empty = unsafe { CStr::from_bytes_with_nul_unchecked(EMPTY_CSTR) };

            // Safe because this doesn't modify any memory and we check the return value.
            let res = unsafe {
                libc::fchownat(
                    inode_data.file.as_raw_fd(),
                    empty.as_ptr(),
                    uid,
                    gid,
                    libc::AT_EMPTY_PATH | libc::AT_SYMLINK_NOFOLLOW,
                )
            };
            if res < 0 {
                return Err(io::Error::last_os_error());
            }
        }

        if valid.contains(SetattrValid::SIZE) {
            // Safe because this doesn't modify any memory and we check the return value.
            let res = match data {
                Data::Handle(_, fd) => unsafe { libc::ftruncate64(fd, attr.st_size) },
                _ => {
                    // There is no `ftruncateat` so we need to get a new fd and truncate it.
                    let f = self.open_inode(inode, libc::O_NONBLOCK | libc::O_RDWR)?;
                    unsafe { libc::ftruncate64(f.as_raw_fd(), attr.st_size) }
                }
            };
            if res < 0 {
                return Err(io::Error::last_os_error());
            }
        }

        if valid.intersects(SetattrValid::ATIME | SetattrValid::MTIME) {
            let mut tvs = [
                libc::timespec {
                    tv_sec: 0,
                    tv_nsec: libc::UTIME_OMIT,
                },
                libc::timespec {
                    tv_sec: 0,
                    tv_nsec: libc::UTIME_OMIT,
                },
            ];

            if valid.contains(SetattrValid::ATIME_NOW) {
                tvs[0].tv_nsec = libc::UTIME_NOW;
            } else if valid.contains(SetattrValid::ATIME) {
                tvs[0].tv_sec = attr.st_atime;
                tvs[0].tv_nsec = attr.st_atime_nsec;
            }

            if valid.contains(SetattrValid::MTIME_NOW) {
                tvs[1].tv_nsec = libc::UTIME_NOW;
            } else if valid.contains(SetattrValid::MTIME) {
                tvs[1].tv_sec = attr.st_mtime;
                tvs[1].tv_nsec = attr.st_mtime_nsec;
            }

            // Safe because this doesn't modify any memory and we check the return value.
            let res = match data {
                Data::Handle(_, fd) => unsafe { libc::futimens(fd, tvs.as_ptr()) },
                Data::ProcPath(ref p) => unsafe {
                    libc::utimensat(self.proc.as_raw_fd(), p.as_ptr(), tvs.as_ptr(), 0)
                },
            };
            if res < 0 {
                return Err(io::Error::last_os_error());
            }
        }

        self.do_getattr(inode)
    }

    fn rename(
        &self,
        _ctx: Context,
        olddir: Inode,
        oldname: &CStr,
        newdir: Inode,
        newname: &CStr,
        flags: u32,
    ) -> io::Result<()> {
        let old_inode = self
            .inodes
            .read()
            .unwrap()
            .get(&olddir)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;
        let new_inode = self
            .inodes
            .read()
            .unwrap()
            .get(&newdir)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        // Safe because this doesn't modify any memory and we check the return value.
        // TODO: Switch to libc::renameat2 once https://github.com/rust-lang/libc/pull/1508 lands
        // and we have glibc 2.28.
        let res = unsafe {
            libc::syscall(
                libc::SYS_renameat2,
                old_inode.file.as_raw_fd(),
                oldname.as_ptr(),
                new_inode.file.as_raw_fd(),
                newname.as_ptr(),
                flags,
            )
        };
        if res == 0 {
            Ok(())
        } else {
            Err(io::Error::last_os_error())
        }
    }

    fn mknod(
        &self,
        ctx: Context,
        parent: Inode,
        name: &CStr,
        mode: u32,
        rdev: u32,
        umask: u32,
    ) -> io::Result<Entry> {
        let (_uid, _gid) = set_creds(ctx.uid, ctx.gid)?;
        let data = self
            .inodes
            .read()
            .unwrap()
            .get(&parent)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        // Safe because this doesn't modify any memory and we check the return value.
        let res = unsafe {
            libc::mknodat(
                data.file.as_raw_fd(),
                name.as_ptr(),
                (mode & !umask) as libc::mode_t,
                rdev as libc::dev_t,
            )
        };

        if res < 0 {
            Err(io::Error::last_os_error())
        } else {
            self.do_lookup(parent, name)
        }
    }

    fn link(
        &self,
        _ctx: Context,
        inode: Inode,
        newparent: Inode,
        newname: &CStr,
    ) -> io::Result<Entry> {
        let data = self
            .inodes
            .read()
            .unwrap()
            .get(&inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;
        let new_inode = self
            .inodes
            .read()
            .unwrap()
            .get(&newparent)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        // Safe because this is a constant value and a valid C string.
        let empty = unsafe { CStr::from_bytes_with_nul_unchecked(EMPTY_CSTR) };

        // Safe because this doesn't modify any memory and we check the return value.
        let res = unsafe {
            libc::linkat(
                data.file.as_raw_fd(),
                empty.as_ptr(),
                new_inode.file.as_raw_fd(),
                newname.as_ptr(),
                libc::AT_EMPTY_PATH,
            )
        };
        if res == 0 {
            self.do_lookup(newparent, newname)
        } else {
            Err(io::Error::last_os_error())
        }
    }

    fn symlink(
        &self,
        ctx: Context,
        linkname: &CStr,
        parent: Inode,
        name: &CStr,
    ) -> io::Result<Entry> {
        let (_uid, _gid) = set_creds(ctx.uid, ctx.gid)?;
        let data = self
            .inodes
            .read()
            .unwrap()
            .get(&parent)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        // Safe because this doesn't modify any memory and we check the return value.
        let res =
            unsafe { libc::symlinkat(linkname.as_ptr(), data.file.as_raw_fd(), name.as_ptr()) };
        if res == 0 {
            self.do_lookup(parent, name)
        } else {
            Err(io::Error::last_os_error())
        }
    }

    fn readlink(&self, _ctx: Context, inode: Inode) -> io::Result<Vec<u8>> {
        let data = self
            .inodes
            .read()
            .unwrap()
            .get(&inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let mut buf = Vec::with_capacity(libc::PATH_MAX as usize);
        buf.resize(libc::PATH_MAX as usize, 0);

        // Safe because this is a constant value and a valid C string.
        let empty = unsafe { CStr::from_bytes_with_nul_unchecked(EMPTY_CSTR) };

        // Safe because this will only modify the contents of `buf` and we check the return value.
        let res = unsafe {
            libc::readlinkat(
                data.file.as_raw_fd(),
                empty.as_ptr(),
                buf.as_mut_ptr() as *mut libc::c_char,
                buf.len(),
            )
        };
        if res < 0 {
            return Err(io::Error::last_os_error());
        }

        buf.resize(res as usize, 0);
        Ok(buf)
    }

    fn flush(
        &self,
        _ctx: Context,
        inode: Inode,
        handle: Handle,
        _lock_owner: u64,
    ) -> io::Result<()> {
        let data = self
            .handles
            .read()
            .unwrap()
            .get(&handle)
            .filter(|hd| hd.inode == inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        // Since this method is called whenever an fd is closed in the client, we can emulate that
        // behavior by doing the same thing (dup-ing the fd and then immediately closing it). Safe
        // because this doesn't modify any memory and we check the return values.
        unsafe {
            let newfd = libc::dup(data.file.lock().as_raw_fd());
            if newfd < 0 {
                return Err(io::Error::last_os_error());
            }

            if libc::close(newfd) < 0 {
                Err(io::Error::last_os_error())
            } else {
                Ok(())
            }
        }
    }

    fn fsync(&self, _ctx: Context, inode: Inode, datasync: bool, handle: Handle) -> io::Result<()> {
        let data = self
            .handles
            .read()
            .unwrap()
            .get(&handle)
            .filter(|hd| hd.inode == inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let fd = data.file.lock().as_raw_fd();

        // Safe because this doesn't modify any memory and we check the return value.
        let res = unsafe {
            if datasync {
                libc::fdatasync(fd)
            } else {
                libc::fsync(fd)
            }
        };

        if res == 0 {
            Ok(())
        } else {
            Err(io::Error::last_os_error())
        }
    }

    fn fsyncdir(
        &self,
        ctx: Context,
        inode: Inode,
        datasync: bool,
        handle: Handle,
    ) -> io::Result<()> {
        self.fsync(ctx, inode, datasync, handle)
    }

    fn access(&self, ctx: Context, inode: Inode, mask: u32) -> io::Result<()> {
        let data = self
            .inodes
            .read()
            .unwrap()
            .get(&inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let st = stat(&data.file)?;
        let mode = mask as i32 & (libc::R_OK | libc::W_OK | libc::X_OK);

        if mode == libc::F_OK {
            // The file exists since we were able to call `stat(2)` on it.
            return Ok(());
        }

        if (mode & libc::R_OK) != 0 {
            if ctx.uid != 0
                && (st.st_uid != ctx.uid || st.st_mode & 0o400 == 0)
                && (st.st_gid != ctx.gid || st.st_mode & 0o040 == 0)
                && st.st_mode & 0o004 == 0
            {
                return Err(io::Error::from_raw_os_error(libc::EACCES));
            }
        }

        if (mode & libc::W_OK) != 0 {
            if ctx.uid != 0
                && (st.st_uid != ctx.uid || st.st_mode & 0o200 == 0)
                && (st.st_gid != ctx.gid || st.st_mode & 0o020 == 0)
                && st.st_mode & 0o002 == 0
            {
                return Err(io::Error::from_raw_os_error(libc::EACCES));
            }
        }

        // root can only execute something if it is executable by one of the owner, the group, or
        // everyone.
        if (mode & libc::X_OK) != 0 {
            if (ctx.uid != 0 || st.st_mode & 0o111 == 0)
                && (st.st_uid != ctx.uid || st.st_mode & 0o100 == 0)
                && (st.st_gid != ctx.gid || st.st_mode & 0o010 == 0)
                && st.st_mode & 0o001 == 0
            {
                return Err(io::Error::from_raw_os_error(libc::EACCES));
            }
        }

        Ok(())
    }

    fn setxattr(
        &self,
        _ctx: Context,
        inode: Inode,
        name: &CStr,
        value: &[u8],
        flags: u32,
    ) -> io::Result<()> {
        // The f{set,get,remove,list}xattr functions don't work on an fd opened with `O_PATH` so we
        // need to get a new fd.
        let file = self.open_inode(inode, libc::O_RDONLY | libc::O_NONBLOCK)?;

        // Safe because this doesn't modify any memory and we check the return value.
        let res = unsafe {
            libc::fsetxattr(
                file.as_raw_fd(),
                name.as_ptr(),
                value.as_ptr() as *const libc::c_void,
                value.len(),
                flags as libc::c_int,
            )
        };
        if res == 0 {
            Ok(())
        } else {
            Err(io::Error::last_os_error())
        }
    }

    fn getxattr(
        &self,
        _ctx: Context,
        inode: Inode,
        name: &CStr,
        size: u32,
    ) -> io::Result<GetxattrReply> {
        // The f{set,get,remove,list}xattr functions don't work on an fd opened with `O_PATH` so we
        // need to get a new fd.
        let file = self.open_inode(inode, libc::O_RDONLY | libc::O_NONBLOCK)?;

        let mut buf = Vec::with_capacity(size as usize);
        buf.resize(size as usize, 0);

        // Safe because this will only modify the contents of `buf`.
        let res = unsafe {
            libc::fgetxattr(
                file.as_raw_fd(),
                name.as_ptr(),
                buf.as_mut_ptr() as *mut libc::c_void,
                size as libc::size_t,
            )
        };
        if res < 0 {
            return Err(io::Error::last_os_error());
        }

        if size == 0 {
            Ok(GetxattrReply::Count(res as u32))
        } else {
            buf.resize(res as usize, 0);
            Ok(GetxattrReply::Value(buf))
        }
    }

    fn listxattr(&self, _ctx: Context, inode: Inode, size: u32) -> io::Result<ListxattrReply> {
        // The f{set,get,remove,list}xattr functions don't work on an fd opened with `O_PATH` so we
        // need to get a new fd.
        let file = self.open_inode(inode, libc::O_RDONLY | libc::O_NONBLOCK)?;

        let mut buf = Vec::with_capacity(size as usize);
        buf.resize(size as usize, 0);

        // Safe because this will only modify the contents of `buf`.
        let res = unsafe {
            libc::flistxattr(
                file.as_raw_fd(),
                buf.as_mut_ptr() as *mut libc::c_char,
                size as libc::size_t,
            )
        };
        if res < 0 {
            return Err(io::Error::last_os_error());
        }

        if size == 0 {
            Ok(ListxattrReply::Count(res as u32))
        } else {
            buf.resize(res as usize, 0);
            Ok(ListxattrReply::Names(buf))
        }
    }

    fn removexattr(&self, _ctx: Context, inode: Inode, name: &CStr) -> io::Result<()> {
        // The f{set,get,remove,list}xattr functions don't work on an fd opened with `O_PATH` so we
        // need to get a new fd.
        let file = self.open_inode(inode, libc::O_RDONLY | libc::O_NONBLOCK)?;

        // Safe because this doesn't modify any memory and we check the return value.
        let res = unsafe { libc::fremovexattr(file.as_raw_fd(), name.as_ptr()) };

        if res == 0 {
            Ok(())
        } else {
            Err(io::Error::last_os_error())
        }
    }

    fn fallocate(
        &self,
        _ctx: Context,
        inode: Inode,
        handle: Handle,
        mode: u32,
        offset: u64,
        length: u64,
    ) -> io::Result<()> {
        let data = self
            .handles
            .read()
            .unwrap()
            .get(&handle)
            .filter(|hd| hd.inode == inode)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let fd = data.file.lock().as_raw_fd();
        // Safe because this doesn't modify any memory and we check the return value.
        let res = unsafe {
            libc::fallocate64(
                fd,
                mode as libc::c_int,
                offset as libc::off64_t,
                length as libc::off64_t,
            )
        };
        if res == 0 {
            Ok(())
        } else {
            Err(io::Error::last_os_error())
        }
    }

    fn ioctl<R: io::Read>(
        &self,
        _ctx: Context,
        handle: Handle,
        _flags: IoctlFlags,
        cmd: u32,
        arg: u64,
        in_size: u32,
        out_size: u32,
        r: R,
    ) -> io::Result<IoctlReply> {
        // Normally, we wouldn't need to retry the FS_IOC_GET_ENCRYPTION_POLICY and
        // FS_IOC_SET_ENCRYPTION_POLICY ioctls. Unfortunately, the I/O directions for both of them
        // are encoded backwards so they can only be handled as unrestricted fuse ioctls.
        if cmd == FS_IOC_GET_ENCRYPTION_POLICY() as u32 {
            if out_size < size_of::<fscrypt_policy_v1>() as u32 {
                let input = Vec::new();
                let output = vec![IoctlIovec {
                    base: arg,
                    len: size_of::<fscrypt_policy_v1>() as u64,
                }];
                Ok(IoctlReply::Retry { input, output })
            } else {
                self.get_encryption_policy(handle)
            }
        } else if cmd == FS_IOC_SET_ENCRYPTION_POLICY() as u32 {
            if in_size < size_of::<fscrypt_policy_v1>() as u32 {
                let input = vec![IoctlIovec {
                    base: arg,
                    len: size_of::<fscrypt_policy_v1>() as u64,
                }];
                let output = Vec::new();
                Ok(IoctlReply::Retry { input, output })
            } else {
                self.set_encryption_policy(handle, r)
            }
        } else {
            // Did you know that a file/directory is not a TTY?
            Err(io::Error::from_raw_os_error(libc::ENOTTY))
        }
    }

    fn copy_file_range(
        &self,
        ctx: Context,
        inode_src: Inode,
        handle_src: Handle,
        offset_src: u64,
        inode_dst: Inode,
        handle_dst: Handle,
        offset_dst: u64,
        length: u64,
        flags: u64,
    ) -> io::Result<usize> {
        // We need to change credentials during a write so that the kernel will remove setuid or
        // setgid bits from the file if it was written to by someone other than the owner.
        let (_uid, _gid) = set_creds(ctx.uid, ctx.gid)?;
        let src_data = self
            .handles
            .read()
            .unwrap()
            .get(&handle_src)
            .filter(|hd| hd.inode == inode_src)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;
        let dst_data = self
            .handles
            .read()
            .unwrap()
            .get(&handle_dst)
            .filter(|hd| hd.inode == inode_dst)
            .map(Arc::clone)
            .ok_or_else(ebadf)?;

        let src = src_data.file.lock().as_raw_fd();
        let dst = dst_data.file.lock().as_raw_fd();

        let res = unsafe {
            libc::syscall(
                libc::SYS_copy_file_range,
                src,
                &offset_src,
                dst,
                offset_dst,
                length,
                flags,
            )
        };

        if res >= 0 {
            Ok(res as usize)
        } else {
            Err(io::Error::last_os_error())
        }
    }
}