rustix/backend/linux_raw/fs/
dir.rs

Help
1use crate::fd::{AsFd, BorrowedFd, OwnedFd};
2use crate::ffi::{CStr, CString};
3use crate::fs::{
4    fcntl_getfl, fstat, fstatfs, fstatvfs, openat, FileType, Mode, OFlags, Stat, StatFs, StatVfs,
5};
6use crate::io;
7#[cfg(feature = "process")]
8use crate::process::fchdir;
9use crate::utils::as_ptr;
10use alloc::borrow::ToOwned;
11use alloc::vec::Vec;
12use core::fmt;
13use core::mem::size_of;
14use linux_raw_sys::general::{linux_dirent64, SEEK_SET};
15
16/// `DIR*`
17pub struct Dir {
18    /// The `OwnedFd` that we read directory entries from.
19    fd: OwnedFd,
20
21    /// Have we seen any errors in this iteration?
22    any_errors: bool,
23
24    /// Should we rewind the stream on the next iteration?
25    rewind: bool,
26
27    /// The buffer for `linux_dirent64` entries.
28    buf: Vec<u8>,
29
30    /// Where we are in the buffer.
31    pos: usize,
32}
33
34impl Dir {
35    /// Construct a `Dir` that reads entries from the given directory
36    /// file descriptor.
37    #[inline]
38    pub fn read_from<Fd: AsFd>(fd: Fd) -> io::Result<Self> {
39        Self::_read_from(fd.as_fd())
40    }
41
42    #[inline]
43    fn _read_from(fd: BorrowedFd<'_>) -> io::Result<Self> {
44        let flags = fcntl_getfl(fd)?;
45        let fd_for_dir = openat(fd, cstr!("."), flags | OFlags::CLOEXEC, Mode::empty())?;
46
47        Ok(Self {
48            fd: fd_for_dir,
49            any_errors: false,
50            rewind: false,
51            buf: Vec::new(),
52            pos: 0,
53        })
54    }
55
56    /// `rewinddir(self)`
57    #[inline]
58    pub fn rewind(&mut self) {
59        self.any_errors = false;
60        self.rewind = true;
61        self.pos = self.buf.len();
62    }
63
64    /// `readdir(self)`, where `None` means the end of the directory.
65    pub fn read(&mut self) -> Option<io::Result<DirEntry>> {
66        // If we've seen errors, don't continue to try to read anyting further.
67        if self.any_errors {
68            return None;
69        }
70
71        // If a rewind was requested, seek to the beginning.
72        if self.rewind {
73            self.rewind = false;
74            match io::retry_on_intr(|| {
75                crate::backend::fs::syscalls::_seek(self.fd.as_fd(), 0, SEEK_SET)
76            }) {
77                Ok(_) => (),
78                Err(err) => {
79                    self.any_errors = true;
80                    return Some(Err(err));
81                }
82            }
83        }
84
85        // Compute linux_dirent64 field offsets.
86        let z = linux_dirent64 {
87            d_ino: 0_u64,
88            d_off: 0_i64,
89            d_type: 0_u8,
90            d_reclen: 0_u16,
91            d_name: Default::default(),
92        };
93        let base = as_ptr(&z) as usize;
94        let offsetof_d_reclen = (as_ptr(&z.d_reclen) as usize) - base;
95        let offsetof_d_name = (as_ptr(&z.d_name) as usize) - base;
96        let offsetof_d_ino = (as_ptr(&z.d_ino) as usize) - base;
97        let offsetof_d_type = (as_ptr(&z.d_type) as usize) - base;
98
99        // Test if we need more entries, and if so, read more.
100        if self.buf.len() - self.pos < size_of::<linux_dirent64>() {
101            match self.read_more()? {
102                Ok(()) => (),
103                Err(err) => return Some(Err(err)),
104            }
105        }
106
107        // We successfully read an entry. Extract the fields.
108        let pos = self.pos;
109
110        // Do an unaligned u16 load.
111        let d_reclen = u16::from_ne_bytes([
112            self.buf[pos + offsetof_d_reclen],
113            self.buf[pos + offsetof_d_reclen + 1],
114        ]);
115        assert!(self.buf.len() - pos >= d_reclen as usize);
116        self.pos += d_reclen as usize;
117
118        // Read the NUL-terminated name from the `d_name` field. Without
119        // `unsafe`, we need to scan for the NUL twice: once to obtain a size
120        // for the slice, and then once within `CStr::from_bytes_with_nul`.
121        let name_start = pos + offsetof_d_name;
122        let name_len = self.buf[name_start..]
123            .iter()
124            .position(|x| *x == b'\0')
125            .unwrap();
126        let name = CStr::from_bytes_with_nul(&self.buf[name_start..][..=name_len]).unwrap();
127        let name = name.to_owned();
128        assert!(name.as_bytes().len() <= self.buf.len() - name_start);
129
130        // Do an unaligned u64 load.
131        let d_ino = u64::from_ne_bytes([
132            self.buf[pos + offsetof_d_ino],
133            self.buf[pos + offsetof_d_ino + 1],
134            self.buf[pos + offsetof_d_ino + 2],
135            self.buf[pos + offsetof_d_ino + 3],
136            self.buf[pos + offsetof_d_ino + 4],
137            self.buf[pos + offsetof_d_ino + 5],
138            self.buf[pos + offsetof_d_ino + 6],
139            self.buf[pos + offsetof_d_ino + 7],
140        ]);
141
142        let d_type = self.buf[pos + offsetof_d_type];
143
144        // Check that our types correspond to the `linux_dirent64` types.
145        let _ = linux_dirent64 {
146            d_ino,
147            d_off: 0,
148            d_type,
149            d_reclen,
150            d_name: Default::default(),
151        };
152
153        Some(Ok(DirEntry {
154            d_ino,
155            d_type,
156            name,
157        }))
158    }
159
160    fn read_more(&mut self) -> Option<io::Result<()>> {
161        // The first few times we're called, we allocate a relatively small
162        // buffer, because many directories are small. If we're called more,
163        // use progressively larger allocations, up to a fixed maximum.
164        //
165        // The specific sizes and policy here have not been tuned in detail yet
166        // and may need to be adjusted. In doing so, we should be careful to
167        // avoid unbounded buffer growth. This buffer only exists to share the
168        // cost of a `getdents` call over many entries, so if it gets too big,
169        // cache and heap usage will outweigh the benefit. And ultimately,
170        // directories can contain more entries than we can allocate contiguous
171        // memory for, so we'll always need to cap the size at some point.
172        if self.buf.len() < 1024 * size_of::<linux_dirent64>() {
173            self.buf.reserve(32 * size_of::<linux_dirent64>());
174        }
175        self.buf.resize(self.buf.capacity(), 0);
176        let nread = match io::retry_on_intr(|| {
177            crate::backend::fs::syscalls::getdents(self.fd.as_fd(), &mut self.buf)
178        }) {
179            Ok(nread) => nread,
180            Err(io::Errno::NOENT) => {
181                self.any_errors = true;
182                return None;
183            }
184            Err(err) => {
185                self.any_errors = true;
186                return Some(Err(err));
187            }
188        };
189        self.buf.resize(nread, 0);
190        self.pos = 0;
191        if nread == 0 {
192            None
193        } else {
194            Some(Ok(()))
195        }
196    }
197
198    /// `fstat(self)`
199    #[inline]
200    pub fn stat(&self) -> io::Result<Stat> {
201        fstat(&self.fd)
202    }
203
204    /// `fstatfs(self)`
205    #[inline]
206    pub fn statfs(&self) -> io::Result<StatFs> {
207        fstatfs(&self.fd)
208    }
209
210    /// `fstatvfs(self)`
211    #[inline]
212    pub fn statvfs(&self) -> io::Result<StatVfs> {
213        fstatvfs(&self.fd)
214    }
215
216    /// `fchdir(self)`
217    #[cfg(feature = "process")]
218    #[cfg_attr(doc_cfg, doc(cfg(feature = "process")))]
219    #[inline]
220    pub fn chdir(&self) -> io::Result<()> {
221        fchdir(&self.fd)
222    }
223}
224
225impl Iterator for Dir {
226    type Item = io::Result<DirEntry>;
227
228    #[inline]
229    fn next(&mut self) -> Option<Self::Item> {
230        Self::read(self)
231    }
232}
233
234impl fmt::Debug for Dir {
235    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
236        f.debug_struct("Dir").field("fd", &self.fd).finish()
237    }
238}
239
240/// `struct dirent`
241#[derive(Debug)]
242pub struct DirEntry {
243    d_ino: u64,
244    d_type: u8,
245    name: CString,
246}
247
248impl DirEntry {
249    /// Returns the file name of this directory entry.
250    #[inline]
251    pub fn file_name(&self) -> &CStr {
252        &self.name
253    }
254
255    /// Returns the type of this directory entry.
256    #[inline]
257    pub fn file_type(&self) -> FileType {
258        FileType::from_dirent_d_type(self.d_type)
259    }
260
261    /// Return the inode number of this directory entry.
262    #[inline]
263    pub fn ino(&self) -> u64 {
264        self.d_ino
265    }
266}
267
268#[test]
269fn dir_iterator_handles_io_errors() {
270    // create a dir, keep the FD, then delete the dir
271    let tmp = tempfile::tempdir().unwrap();
272    let fd = crate::fs::openat(
273        crate::fs::CWD,
274        tmp.path(),
275        crate::fs::OFlags::RDONLY | crate::fs::OFlags::CLOEXEC,
276        crate::fs::Mode::empty(),
277    )
278    .unwrap();
279
280    let file_fd = crate::fs::openat(
281        &fd,
282        tmp.path().join("test.txt"),
283        crate::fs::OFlags::WRONLY | crate::fs::OFlags::CREATE,
284        crate::fs::Mode::RWXU,
285    )
286    .unwrap();
287
288    let mut dir = Dir::read_from(&fd).unwrap();
289
290    // Reach inside the `Dir` and replace its directory with a file, which
291    // will cause the subsequent `getdents64` to fail.
292    crate::io::dup2(&file_fd, &mut dir.fd).unwrap();
293
294    assert!(matches!(dir.next(), Some(Err(_))));
295    assert!(matches!(dir.next(), None));
296}
rustix/backend/linux_raw/fs/dir.rs

rustix/backend/linux_raw/fs/
dir.rs
