proxmox-backup/src/pxar/fuse.rs

//! Low level FUSE implementation for pxar.
//!
//! Allows to mount the archive as read-only filesystem to inspect its contents.

use std::collections::HashMap;
use std::convert::TryFrom;
use std::ffi::{CStr, CString, OsStr};
use std::fs::File;
use std::io::BufReader;
use std::os::unix::ffi::{OsStrExt, OsStringExt};
use std::path::Path;
use std::sync::Mutex;

use failure::{bail, format_err, Error};
use lazy_static::lazy_static;
use libc;
use libc::{c_char, c_int, c_void, size_t};

use super::binary_search_tree::search_binary_tree_by;
use super::decoder::Decoder;
use super::format_definition::{PxarAttributes, PxarEntry, PxarGoodbyeItem};

/// Node ID of the root i-node
///
/// Offsets in the archive are used as i-node for the fuse implementation, as
/// they are unique and enough to reference each item in the pxar archive.
/// The only exception to this is the `FUSE_ROOT_ID`, which is defined as 1 by
/// the fuse library.
/// This is okay since offset 1 is part of the root directory entry header and
/// will therefore not occur again, but remapping to the correct offset of 0 is
/// required.
const FUSE_ROOT_ID: u64 = 1;

const GOODBYE_ITEM_SIZE: u64 = std::mem::size_of::<PxarGoodbyeItem>() as u64;
lazy_static! {
    /// HashMap holding the mapping from the child offsets to their parent
    /// offsets.
    ///
    /// In order to include the parent directory entry '..' in the response for
    /// readdir callback, this mapping is needed.
    /// Calling the lookup callback will insert the offsets into the HashMap.
    static ref CHILD_PARENT: Mutex<HashMap<u64, u64>> = Mutex::new(HashMap::new());
}

type Inode = u64;
type Offset = u64;
/// FFI types for easier readability
type Request = *mut c_void;
type MutPtr = *mut c_void;
type ConstPtr = *const c_void;
type StrPtr = *const c_char;
type MutStrPtr = *mut c_char;

#[rustfmt::skip]
#[link(name = "fuse3")]
extern "C" {
    fn fuse_session_new(args: Option<&FuseArgs>, oprs: Option<&Operations>, size: size_t, op: ConstPtr) -> MutPtr;
    fn fuse_set_signal_handlers(session: ConstPtr) -> c_int;
    fn fuse_remove_signal_handlers(session: ConstPtr);
    fn fuse_daemonize(foreground: c_int) -> c_int;
    fn fuse_session_mount(session: ConstPtr, mountpoint: StrPtr) -> c_int;
    fn fuse_session_unmount(session: ConstPtr);
    fn fuse_session_loop(session: ConstPtr) -> c_int;
    fn fuse_session_loop_mt_31(session: ConstPtr, clone_fd: c_int) -> c_int;
    fn fuse_session_destroy(session: ConstPtr);
    fn fuse_reply_attr(req: Request, attr: Option<&libc::stat>, timeout: f64) -> c_int;
    fn fuse_reply_err(req: Request, errno: c_int) -> c_int;
    fn fuse_reply_open(req: Request, fileinfo: ConstPtr) -> c_int;
    fn fuse_reply_buf(req: Request, buf: MutStrPtr, size: size_t) -> c_int;
    fn fuse_reply_entry(req: Request, entry: Option<&EntryParam>) -> c_int;
    fn fuse_reply_readlink(req: Request, link: StrPtr) -> c_int;
    fn fuse_req_userdata(req: Request) -> MutPtr;
    fn fuse_add_direntry(req: Request, buf: MutStrPtr, bufsize: size_t, name: StrPtr, stbuf: Option<&libc::stat>, off: c_int) -> c_int;
}

/// Command line arguments passed to fuse.
#[repr(C)]
#[derive(Debug)]
struct FuseArgs {
    argc: c_int,
    argv: *const StrPtr,
    allocated: c_int,
}

/// `Context` for callback functions providing the decoder, caches and the
/// offset within the archive for the i-node given by the caller.
struct Context {
    decoder: Decoder,
    goodbye_cache: HashMap<Inode, Vec<(PxarGoodbyeItem, Offset, Offset)>>,
    attr_cache: Option<(Inode, PxarAttributes)>,
    ino_offset: Offset,
}

impl Context {
    /// Lookup the goodbye item identified by `filename` and its corresponding `hash`
    ///
    /// Updates the goodbye table cache to contain the table for the directory given
    /// by the i-node in the provided `Context`.
    /// Search the first matching `hash` in the goodbye table, allowing for a fast
    /// comparison with the items.
    /// As there could be a hash collision, the found items filename is then compared
    /// by seek to the corresponding item in the archive and reading its attributes
    /// (which the lookup callback needs to do anyway).
    /// If the filename does not match, the function is called recursively with the
    /// rest of the goodbye table to lookup the next match.
    /// The matching items archive offset, entry and payload size are returned.
    /// If there is no entry with matching `filename` and `hash` a `libc::ENOENT` is
    /// returned.
    fn find_goodbye_entry(
        &mut self,
        filename: &CStr,
        hash: u64,
    ) -> Result<(u64, PxarEntry, PxarAttributes, u64), i32> {
        let gbt = self.goodbye_cache.get(&self.ino_offset)
            .ok_or_else(|| libc::EIO)?;
        let mut start_idx = 0;
        let mut skip_multiple = 0;
        loop {
            // Search for the next goodbye entry with matching hash.
            let idx = search_binary_tree_by(
                start_idx,
                gbt.len(),
                skip_multiple,
                |idx| hash.cmp(&gbt[idx].0.hash),
            ).ok_or(libc::ENOENT)?;

            let (_item, start, end) = &gbt[idx];

            // At this point it is not clear if the item is a directory or not, this
            // has to be decided based on the entry mode.
            // `Decoder`s attributes function accepts both, offsets pointing to
            // the start of an item (PXAR_FILENAME) or the GOODBYE_TAIL_MARKER in case
            // of directories, so the use of start offset is fine for both cases.
            let (entry_name, entry, attr, payload_size) =
                self.decoder.attributes(*start).map_err(|_| libc::EIO)?;

            // Possible hash collision, need to check if the found entry is indeed
            // the filename to lookup.
            if entry_name.as_bytes() == filename.to_bytes() {
                let child_offset = find_offset(&entry, *start, *end);
                return Ok((child_offset, entry, attr, payload_size));
            }
            // Hash collision, check the next entry in the goodbye table by starting
            // from given index but skipping one more match (so hash at index itself).
            start_idx = idx;
            skip_multiple = 1;
        }
    }
}

/// `Session` stores a pointer to the session context and is used to mount the
/// archive to the given mountpoint.
pub struct Session {
    ptr: MutPtr,
    verbose: bool,
}

/// `Operations` defines the callback function table of supported operations.
#[repr(C)]
#[derive(Default)]
#[rustfmt::skip]
struct Operations {
    // The order in which the functions are listed matters, as the offset in the
    // struct defines what function the fuse driver uses.
    // It should therefore not be altered!
    init:           Option<extern fn(userdata: MutPtr)>,
    destroy:        Option<extern fn(userdata: MutPtr)>,
    lookup:         Option<extern fn(req: Request, parent: u64, name: StrPtr)>,
    forget:         Option<extern fn(req: Request, inode: u64, nlookup: u64)>,
    getattr:        Option<extern fn(req: Request, inode: u64, fileinfo: MutPtr)>,
    setattr:        Option<extern fn(req: Request, inode: u64, attr: MutPtr, to_set: c_int, fileinfo: MutPtr)>,
    readlink:       Option<extern fn(req: Request, inode: u64)>,
    mknod:          Option<extern fn(req: Request, parent: u64, name: StrPtr, mode: c_int, rdev: c_int)>,
    mkdir:          Option<extern fn(req: Request, parent: u64, name: StrPtr, mode: c_int)>,
    unlink:         Option<extern fn(req: Request, parent: u64, name: StrPtr)>,
    rmdir:          Option<extern fn(req: Request, parent: u64, name: StrPtr)>,
    symlink:        Option<extern fn(req: Request, link: StrPtr, parent: u64, name: StrPtr)>,
    rename:         Option<extern fn(req: Request, parent: u64, name: StrPtr, newparent: u64, newname: StrPtr, flags: c_int)>,
    link:           Option<extern fn(req: Request, inode: u64, newparent: u64, newname: StrPtr)>,
    open:           Option<extern fn(req: Request, indoe: u64, fileinfo: MutPtr)>,
    read:           Option<extern fn(req: Request, inode: u64, size: size_t, offset: c_int, fileinfo: MutPtr)>,
    write:          Option<extern fn(req: Request, inode: u64, buffer: StrPtr, size: size_t, offset: c_void, fileinfo: MutPtr)>,
    flush:          Option<extern fn(req: Request, inode: u64, fileinfo: MutPtr)>,
    release:        Option<extern fn(req: Request, inode: u64, fileinfo: MutPtr)>,
    fsync:          Option<extern fn(req: Request, inode: u64, datasync: c_int, fileinfo: MutPtr)>,
    opendir:        Option<extern fn(req: Request, inode: u64, fileinfo: MutPtr)>,
    readdir:        Option<extern fn(req: Request, inode: u64, size: size_t, offset: c_int, fileinfo: MutPtr)>,
    releasedir:     Option<extern fn(req: Request, inode: u64, fileinfo: MutPtr)>,
    fsyncdir:       Option<extern fn(req: Request, inode: u64, datasync: c_int, fileinfo: MutPtr)>,
    statfs:         Option<extern fn(req: Request, inode: u64)>,
    setxattr:       Option<extern fn(req: Request, inode: u64, name: StrPtr, value: StrPtr, size: size_t, flags: c_int)>,
    getxattr:       Option<extern fn(req: Request, inode: u64, name: StrPtr, size: size_t)>,
    listxattr:      Option<extern fn(req: Request, inode: u64, size: size_t)>,
    removexattr:    Option<extern fn(req: Request, inode: u64, name: StrPtr)>,
    access:         Option<extern fn(req: Request, inode: u64, mask: i32)>,
    create:         Option<extern fn(req: Request, parent: u64, name: StrPtr, mode: c_int, fileinfo: MutPtr)>,
    getlk:          Option<extern fn(req: Request, inode: u64, fileinfo: MutPtr, lock: MutPtr)>,
    setlk:          Option<extern fn(req: Request, inode: u64, fileinfo: MutPtr, lock: MutPtr, sleep: c_int)>,
    bmap:           Option<extern fn(req: Request, inode: u64, blocksize: size_t, idx: u64)>,
    ioctl:          Option<extern fn(req: Request, inode: u64, cmd: c_int, arg: MutPtr, fileinfo: MutPtr, flags: c_int, in_buf: ConstPtr, in_bufsz: size_t, out_bufsz: size_t)>,
    poll:           Option<extern fn(req: Request, inode: u64, fileinfo: MutPtr, pollhandle: MutPtr)>,
    write_buf:      Option<extern fn(req: Request, inode: u64, bufv: MutPtr, offset: c_int, fileinfo: MutPtr)>,
    retrieve_reply: Option<extern fn(req: Request, cookie: ConstPtr, inode: u64, offset: c_int, bufv: MutPtr)>,
    forget_multi:   Option<extern fn(req: Request, count: size_t, forgets: MutPtr)>,
    flock:          Option<extern fn(req: Request, inode: u64, fileinfo: MutPtr, op: c_int)>,
    fallocate:      Option<extern fn(req: Request, inode: u64, mode: c_int, offset: c_int, length: c_int, fileinfo: MutPtr)>,
    readdirplus:    Option<extern fn(req: Request, inode: u64, size: size_t, offset: c_int, fileinfo: MutPtr)>,
    copy_file_range: Option<extern fn(req: Request, ino_in: u64, off_in: c_int, fi_in: MutPtr, ino_out: u64, off_out: c_int, fi_out: MutPtr, len: size_t, flags: c_int)>,
}


impl Session  {

    /// Create a new low level fuse session.
    ///
    /// `Session` is created using the provided mount options and sets the
    /// default signal handlers.
    /// Options have to be provided as comma separated OsStr, e.g.
    /// ("ro,default_permissions").
    pub fn from_path(archive_path: &Path, options: &OsStr, verbose: bool) -> Result<Self, Error> {
        let file = File::open(archive_path)?;
        let reader = BufReader::new(file);
        let decoder = Decoder::new(reader)?;
        Self::new(decoder, options, verbose)
    }

    /// Create a new low level fuse session using the given `Decoder`.
    ///
    /// `Session` is created using the provided mount options and sets the
    /// default signal handlers.
    /// Options have to be provided as comma separated OsStr, e.g.
    /// ("ro,default_permissions").
    pub fn new(
        mut decoder: Decoder,
        options: &OsStr,
        verbose: bool,
    ) -> Result<Self, Error> {
        let args = Self::setup_args(options, verbose)?;
        let oprs = Self::setup_callbacks();

        let root_ino_offset = decoder.root_end_offset() - GOODBYE_ITEM_SIZE;
        let root_goodbye_table = decoder.goodbye_table(None, root_ino_offset + GOODBYE_ITEM_SIZE)?;
        let mut goodbye_cache = HashMap::new();
        goodbye_cache.insert(root_ino_offset, root_goodbye_table);

        let ctx = Context {
            decoder,
            goodbye_cache,
            attr_cache: None,
            ino_offset: 0,
        };

        let session_ctx = Box::new(Mutex::new(ctx));
        let arg_ptrs: Vec<_> = args.iter().map(|opt| opt.as_ptr()).collect();
        let fuse_args = FuseArgs {
            argc: arg_ptrs.len() as i32,
            argv: arg_ptrs.as_ptr(),
            allocated: 0,
        };
        let session_ptr = unsafe {
            fuse_session_new(
                Some(&fuse_args),
                Some(&oprs),
                std::mem::size_of::<Operations>(),
                // Ownership of session_ctx is passed to the session here.
                // It has to be reclaimed before dropping the session to free
                // the `Context` and close the underlying file. This is done inside
                // the destroy callback function.
                Box::into_raw(session_ctx) as ConstPtr,
            )
        };

        if session_ptr.is_null() {
            bail!("error while creating new fuse session");
        }

        if unsafe { fuse_set_signal_handlers(session_ptr) } != 0 {
            bail!("error while setting signal handlers");
        }

        Ok(Self { ptr: session_ptr, verbose })
    }

    fn setup_args(options: &OsStr, verbose: bool) -> Result<Vec<CString>, Error> {
        // First argument should be the executable name
        let mut arguments = vec![
            CString::new("pxar-mount").unwrap(),
            CString::new("-o").unwrap(),
            CString::new(options.as_bytes())?,
        ];
        if verbose {
            arguments.push(CString::new("--debug").unwrap());
        }

        Ok(arguments)
    }

    fn setup_callbacks() -> Operations {
        // Register the callback functions for the session
        let mut oprs = Operations::default();
        oprs.init = Some(Self::init);
        oprs.destroy = Some(Self::destroy);
        oprs.lookup = Some(Self::lookup);
        oprs.getattr = Some(Self::getattr);
        oprs.readlink = Some(Self::readlink);
        oprs.open = Some(Self::open);
        oprs.read = Some(Self::read);
        oprs.opendir = Some(Self::opendir);
        oprs.readdir = Some(Self::readdir);
        oprs.releasedir = Some(Self::releasedir);
        oprs
    }

    /// Mount the filesystem on the given mountpoint.
    ///
    /// Actually mount the filesystem for this session on the provided mountpoint
    /// and daemonize process.
    pub fn mount(&mut self, mountpoint: &Path, deamonize: bool) -> Result<(), Error> {
        if self.verbose {
            println!("Mounting archive to {:#?}", mountpoint);
        }
        let mountpoint = mountpoint.canonicalize()?;
        let path_cstr = CString::new(mountpoint.as_os_str().as_bytes())
            .map_err(|err| format_err!("invalid mountpoint - {}", err))?;
        if unsafe { fuse_session_mount(self.ptr, path_cstr.as_ptr()) } != 0 {
            bail!("mounting on {:#?} failed", mountpoint);
        }

        // Send process to background if deamonize is set
        if deamonize && unsafe { fuse_daemonize(0) } != 0 {
            bail!("could not send process to background");
        }

        Ok(())
    }

    /// Execute session loop which handles requests from kernel.
    ///
    /// The multi_threaded flag controls if the session loop runs in
    /// single-threaded or multi-threaded mode.
    /// Single-threaded mode is intended for debugging only.
    pub fn run_loop(&mut self, multi_threaded: bool) -> Result<(), Error> {
        if self.verbose {
            println!("Executing fuse session loop");
        }
        let result = match multi_threaded {
            true => unsafe { fuse_session_loop_mt_31(self.ptr, 1) },
            false => unsafe { fuse_session_loop(self.ptr) },
        };
        if result < 0 {
            bail!("fuse session loop exited with - {}", result);
        }
        if result > 0 {
            eprintln!("fuse session loop received signal - {}", result);
        }

        Ok(())
    }

    /// Creates a context providing an exclusive mutable reference to the `Context`.
    ///
    /// Each callback function needing access to the `Context` can easily get an
    /// exclusive handle by running the code inside this context.
    /// Responses with error code can easily be generated by returning with the
    /// error code.
    /// The error code will be used to reply to libfuse.
    fn run_in_context<F>(req: Request, inode: u64, code: F)
        where
        F: FnOnce(&mut Context) -> Result<(), i32>,
    {
        let boxed_ctx = unsafe {
            let ptr = fuse_req_userdata(req) as *mut Mutex<Context>;
            Box::from_raw(ptr)
        };
        let result = boxed_ctx
            .lock()
            .map(|mut ctx| {
                ctx.ino_offset = match inode {
                    FUSE_ROOT_ID => ctx.decoder.root_end_offset() - GOODBYE_ITEM_SIZE,
                    _ => inode,
                };
                code(&mut ctx)
            })
            .unwrap_or(Err(libc::EIO));

        if let Err(err) = result {
            unsafe {
                let _res = fuse_reply_err(req, err);
            }
        }

        // Release ownership of boxed context, do not drop it.
        let _ = Box::into_raw(boxed_ctx);
    }

    /// Callback functions for fuse kernel driver.
    extern "C" fn init(_decoder: MutPtr) {
        // Notting to do here for now
    }

    /// Cleanup the userdata created while creating the session, which is the `Context`
    extern "C" fn destroy(ctx: MutPtr) {
        // Get ownership of the `Context` and drop it when Box goes out of scope.
        unsafe { Box::from_raw(ctx) };
    }

    /// Lookup `name` in the directory referenced by `parent` i-node.
    ///
    /// Inserts also the child and parent file offset in the hashmap to quickly
    /// obtain the parent offset based on the child offset.
    /// Caches goodbye table of parent and attributes of child, if found.
    extern "C" fn lookup(req: Request, parent: u64, name: StrPtr) {
        let filename = unsafe { CStr::from_ptr(name) };
        let hash = super::format_definition::compute_goodbye_hash(filename.to_bytes());

        Self::run_in_context(req, parent, |mut ctx| {
            // find_ goodbye_entry() will also update the goodbye cache
            let (child_offset, entry, attr, payload_size) =
                ctx.find_goodbye_entry(&filename, hash)?;
            ctx.attr_cache = Some((child_offset, attr));
            let child_inode = calculate_inode(child_offset, ctx.decoder.root_end_offset());

            let e = EntryParam {
                inode: child_inode,
                generation: 1,
                attr: stat(child_inode, &entry, payload_size)?,
                attr_timeout: std::f64::MAX,
                entry_timeout: std::f64::MAX,
            };

            // Update the parent for this child entry. Used to get parent offset if
            // only child offset is known.
            CHILD_PARENT
                .lock()
                .map_err(|_| libc::EIO)?
            .insert(child_offset, ctx.ino_offset);
            let _res = unsafe { fuse_reply_entry(req, Some(&e)) };

            Ok(())
        });
    }

    extern "C" fn getattr(req: Request, inode: u64, _fileinfo: MutPtr) {
        Self::run_in_context(req, inode, |ctx| {
            let (_, entry, attr, payload_size) = ctx
                .decoder
                .attributes(ctx.ino_offset)
                .map_err(|_| libc::EIO)?;
            ctx.attr_cache = Some((ctx.ino_offset, attr));
            let attr = stat(inode, &entry, payload_size)?;
            let _res = unsafe {
                // Since fs is read-only, the timeout can be max.
                let timeout = std::f64::MAX;
                fuse_reply_attr(req, Some(&attr), timeout)
            };

            Ok(())
        });
    }

    extern "C" fn readlink(req: Request, inode: u64) {
        Self::run_in_context(req, inode, |ctx| {
            let (target, _) = ctx
                .decoder
                .read_link(ctx.ino_offset)
                .map_err(|_| libc::EIO)?;
            let link = CString::new(target.into_os_string().into_vec()).map_err(|_| libc::EIO)?;
            let _ret = unsafe { fuse_reply_readlink(req, link.as_ptr()) };

            Ok(())
        });
    }

    extern "C" fn open(req: Request, inode: u64, fileinfo: MutPtr) {
        Self::run_in_context(req, inode, |ctx| {
            ctx.decoder.open(ctx.ino_offset).map_err(|_| libc::ENOENT)?;
            let _ret = unsafe { fuse_reply_open(req, fileinfo) };

            Ok(())
        });
    }

    extern "C" fn read(req: Request, inode: u64, size: size_t, offset: c_int, _fileinfo: MutPtr) {
        Self::run_in_context(req, inode, |ctx| {
            let mut data = ctx
                .decoder
                .read(ctx.ino_offset, size, offset as u64)
                .map_err(|_| libc::EIO)?;

            let _res = unsafe {
                let len = data.len();
                let dptr = data.as_mut_ptr() as *mut c_char;
                fuse_reply_buf(req, dptr, len)
            };

            Ok(())
        });
    }

    /// Open the directory referenced by the given inode for reading.
    ///
    /// This simply checks if the inode references a valid directory, no internal
    /// state identifies the directory as opened.
    extern "C" fn opendir(req: Request, inode: u64, fileinfo: MutPtr) {
        Self::run_in_context(req, inode, |ctx| {
            let gbt = ctx.decoder
                .goodbye_table(None, ctx.ino_offset + GOODBYE_ITEM_SIZE)
                .map_err(|_| libc::EIO)?;
            ctx.goodbye_cache.insert(ctx.ino_offset, gbt);

            let _ret = unsafe { fuse_reply_open(req, fileinfo as MutPtr) };

            Ok(())
        });
    }

    /// Read and return the entries of the directory referenced by i-node.
    ///
    /// Replies to the request with the entries fitting into a buffer of length
    /// `size`, as requested by the caller.
    /// `offset` identifies the start index of entries to return. This is used on
    /// repeated calls, occurring if not all entries fitted into the buffer.
    /// The goodbye table of the directory is cached in order to speedup repeated
    /// calls occurring when not all entries fitted in the reply buffer.
    extern "C" fn readdir(req: Request, inode: u64, size: size_t, offset: c_int, _fileinfo: MutPtr) {
        let offset = offset as usize;

        Self::run_in_context(req, inode, |ctx| {
            let gb_table = ctx.goodbye_cache.get(&ctx.ino_offset)
                .ok_or_else(|| libc::EIO)?;
            let n_entries = gb_table.len();
            let mut buf = ReplyBuf::new(req, size, offset);

            if offset < n_entries {
                for e in gb_table[offset..gb_table.len()].iter() {
                    let (filename, entry, _, payload_size) =
                        ctx.decoder.attributes(e.1).map_err(|_| libc::EIO)?;
                    let name = CString::new(filename.as_bytes()).map_err(|_| libc::EIO)?;
                    let item_offset = find_offset(&entry, e.1, e.2);
                    let item_inode = calculate_inode(item_offset, ctx.decoder.root_end_offset());
                    let attr = stat(item_inode, &entry, payload_size).map_err(|_| libc::EIO)?;
                    match buf.add_entry(&name, &attr) {
                        Ok(ReplyBufState::Okay) => {}
                        Ok(ReplyBufState::Overfull) => return buf.reply_filled(),
                        Err(_) => return Err(libc::EIO),
                    }
                }
            }

            // Add current directory entry "."
            if offset <= n_entries {
                let (_, entry, _, payload_size) = ctx
                    .decoder
                    .attributes(ctx.ino_offset)
                    .map_err(|_| libc::EIO)?;
                // No need to calculate i-node for current dir, since it is given as parameter
                let attr = stat(inode, &entry, payload_size).map_err(|_| libc::EIO)?;
                let name = CString::new(".").unwrap();
                match buf.add_entry(&name, &attr) {
                    Ok(ReplyBufState::Okay) => {}
                    Ok(ReplyBufState::Overfull) => return buf.reply_filled(),
                    Err(_) => return Err(libc::EIO),
                }
            }

            // Add parent directory entry ".."
            if offset <= n_entries + 1 {
                let parent_off = if inode == FUSE_ROOT_ID {
                    ctx.decoder.root_end_offset() - GOODBYE_ITEM_SIZE
                } else {
                    let guard = CHILD_PARENT.lock().map_err(|_| libc::EIO)?;
                    *guard.get(&ctx.ino_offset).ok_or_else(|| libc::EIO)?
                };
                let (_, entry, _, payload_size) =
                    ctx.decoder.attributes(parent_off).map_err(|_| libc::EIO)?;
                let item_inode = calculate_inode(parent_off, ctx.decoder.root_end_offset());
                let attr = stat(item_inode, &entry, payload_size).map_err(|_| libc::EIO)?;
                let name = CString::new("..").unwrap();
                match buf.add_entry(&name, &attr) {
                    Ok(ReplyBufState::Okay) => {}
                    Ok(ReplyBufState::Overfull) => return buf.reply_filled(),
                    Err(_) => return Err(libc::EIO),
                }
            }

            buf.reply_filled()
        });
    }


    extern "C" fn releasedir(req: Request, inode: u64, _fileinfo: MutPtr) {
        Self::run_in_context(req, inode, |ctx| {
            let _gbt = ctx.goodbye_cache.remove(&ctx.ino_offset);
            Ok(())
        });
    }
}

impl Drop for Session {
    fn drop(&mut self) {
        unsafe {
            fuse_session_unmount(self.ptr);
            fuse_remove_signal_handlers(self.ptr);
            fuse_session_destroy(self.ptr);
        }
    }
}


/// Return the correct offset for the item based on its `PxarEntry` mode
///
/// For directories, the offset for the corresponding `GOODBYE_TAIL_MARKER`
/// is returned.
/// If it is not a directory, the start offset is returned.
fn find_offset(entry: &PxarEntry, start: u64, end: u64) -> u64 {
    if (entry.mode as u32 & libc::S_IFMT) == libc::S_IFDIR {
        end - GOODBYE_ITEM_SIZE
    } else {
        start
    }
}

/// Calculate the i-node based on the given `offset`
///
/// This maps the `offset` to the correct i-node, which is simply the offset.
/// The root directory is an exception, as it has per definition `FUSE_ROOT_ID`.
/// `root_end` is the end offset of the root directory (archive end).
fn calculate_inode(offset: u64, root_end: u64) -> u64 {
    // check for root offset which has to be mapped to `FUSE_ROOT_ID`
    if offset == root_end - GOODBYE_ITEM_SIZE {
        FUSE_ROOT_ID
    } else {
        offset
    }
}

/// FUSE entry for fuse_reply_entry in lookup callback
#[repr(C)]
struct EntryParam {
    inode: u64,
    generation: u64,
    attr: libc::stat,
    attr_timeout: f64,
    entry_timeout: f64,
}

/// Create a `libc::stat` with the provided i-node, entry and payload size
fn stat(inode: u64, entry: &PxarEntry, payload_size: u64) -> Result<libc::stat, i32> {
    let nlink = match (entry.mode as u32) & libc::S_IFMT {
        libc::S_IFDIR => 2,
        _ => 1,
    };
    let time = i64::try_from(entry.mtime).map_err(|_| libc::EIO)?;
    let sec = time / 1_000_000_000;
    let nsec = time % 1_000_000_000;

    let mut attr: libc::stat = unsafe { std::mem::zeroed() };
    attr.st_ino = inode;
    attr.st_nlink = nlink;
    attr.st_mode = u32::try_from(entry.mode).map_err(|_| libc::EIO)?;
    attr.st_size = i64::try_from(payload_size).map_err(|_| libc::EIO)?;
    attr.st_uid = entry.uid;
    attr.st_gid = entry.gid;
    attr.st_atime = sec;
    attr.st_atime_nsec = nsec;
    attr.st_mtime = sec;
    attr.st_mtime_nsec = nsec;
    attr.st_ctime = sec;
    attr.st_ctime_nsec = nsec;

    Ok(attr)
}

/// State of ReplyBuf after last add_entry call
enum ReplyBufState {
    /// Entry was successfully added to ReplyBuf
    Okay,
    /// Entry did not fit into ReplyBuf, was not added
    Overfull,
}

/// Used to correctly fill and reply the buffer for the readdir callback
struct ReplyBuf {
    /// internal buffer holding the binary data
    buffer: Vec<u8>,
    /// offset up to which the buffer is filled already
    filled: usize,
    /// fuse request the buffer is used to reply to
    req: Request,
    /// index of the next item, telling from were to start on the next readdir callback in
    /// case not everything fitted in the buffer on the first reply.
    next: usize,
}

impl ReplyBuf {
    /// Create a new empty `ReplyBuf` of `size` with element counting index at `next`.
    fn new(req: Request, size: usize, next: usize) -> Self {
        Self {
            buffer: vec![0; size],
            filled: 0,
            req,
            next,
        }
    }

    /// Reply to the `Request` with the filled buffer
    fn reply_filled(&mut self) -> Result<(), i32> {
        let _res = unsafe {
            let ptr = self.buffer.as_mut_ptr() as *mut c_char;
            fuse_reply_buf(self.req, ptr, self.filled)
        };

        Ok(())
    }

    /// Fill the buffer for the fuse reply with the next entry
    fn add_entry(&mut self, name: &CString, attr: &libc::stat) -> Result<ReplyBufState, Error> {
        self.next += 1;
        let size = self.buffer.len();
        let bytes = unsafe {
            let bptr = self.buffer.as_mut_ptr() as *mut c_char;
            let nptr = name.as_ptr();
            fuse_add_direntry(
                self.req,
                bptr.offset(self.filled as isize),
                size - self.filled,
                nptr,
                Some(&attr),
                i32::try_from(self.next)?,
            ) as usize
        };
        self.filled += bytes;
        // Never exceed the max size requested in the callback (=buffer.len())
        if self.filled > size {
            // Entry did not fit, so go back to previous state
            self.filled -= bytes;
            self.next -= 1;
            return Ok(ReplyBufState::Overfull);
        }

        Ok(ReplyBufState::Okay)
    }
}