| 1 | use crate::fd::{AsFd, BorrowedFd, OwnedFd}; |
| 2 | use crate::ffi::{CStr, CString}; |
| 3 | use crate::fs::{ |
| 4 | fcntl_getfl, fstat, fstatfs, fstatvfs, openat, FileType, Mode, OFlags, Stat, StatFs, StatVfs, |
| 5 | }; |
| 6 | use crate::io; |
| 7 | #[cfg (feature = "process" )] |
| 8 | use crate::process::fchdir; |
| 9 | use crate::utils::as_ptr; |
| 10 | use alloc::borrow::ToOwned as _; |
| 11 | use alloc::vec::Vec; |
| 12 | use core::fmt; |
| 13 | use core::mem::size_of; |
| 14 | use linux_raw_sys::general::{linux_dirent64, SEEK_SET}; |
| 15 | |
| 16 | /// `DIR*` |
| 17 | pub 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 | |
| 34 | impl Dir { |
| 35 | /// Take ownership of `fd` and construct a `Dir` that reads entries from |
| 36 | /// the given directory file descriptor. |
| 37 | #[inline ] |
| 38 | pub fn new<Fd: Into<OwnedFd>>(fd: Fd) -> io::Result<Self> { |
| 39 | Self::_new(fd.into()) |
| 40 | } |
| 41 | |
| 42 | #[inline ] |
| 43 | fn _new(fd: OwnedFd) -> io::Result<Self> { |
| 44 | Ok(Self { |
| 45 | fd, |
| 46 | any_errors: false, |
| 47 | rewind: false, |
| 48 | buf: Vec::new(), |
| 49 | pos: 0, |
| 50 | }) |
| 51 | } |
| 52 | |
| 53 | /// Borrow `fd` and construct a `Dir` that reads entries from the given |
| 54 | /// directory file descriptor. |
| 55 | #[inline ] |
| 56 | pub fn read_from<Fd: AsFd>(fd: Fd) -> io::Result<Self> { |
| 57 | Self::_read_from(fd.as_fd()) |
| 58 | } |
| 59 | |
| 60 | #[inline ] |
| 61 | fn _read_from(fd: BorrowedFd<'_>) -> io::Result<Self> { |
| 62 | let flags = fcntl_getfl(fd)?; |
| 63 | let fd_for_dir = openat(fd, cstr!("." ), flags | OFlags::CLOEXEC, Mode::empty())?; |
| 64 | |
| 65 | Ok(Self { |
| 66 | fd: fd_for_dir, |
| 67 | any_errors: false, |
| 68 | rewind: false, |
| 69 | buf: Vec::new(), |
| 70 | pos: 0, |
| 71 | }) |
| 72 | } |
| 73 | |
| 74 | /// `rewinddir(self)` |
| 75 | #[inline ] |
| 76 | pub fn rewind(&mut self) { |
| 77 | self.any_errors = false; |
| 78 | self.rewind = true; |
| 79 | self.pos = self.buf.len(); |
| 80 | } |
| 81 | |
| 82 | /// `seekdir(self, offset)` |
| 83 | /// |
| 84 | /// This function is only available on 64-bit platforms because it's |
| 85 | /// implemented using [`libc::seekdir`] which only supports offsets that |
| 86 | /// fit in a `c_long`. |
| 87 | /// |
| 88 | /// [`libc::seekdir`]: https://docs.rs/libc/*/arm-unknown-linux-gnueabihf/libc/fn.seekdir.html |
| 89 | // In the linux_raw backend here, we don't use `libc::seekdir` and don't |
| 90 | // have this limitation, but it's a goal of rustix to support the same API |
| 91 | // on both the linux_raw and libc backends. |
| 92 | #[cfg (target_pointer_width = "64" )] |
| 93 | #[cfg_attr (docsrs, doc(cfg(target_pointer_width = "64" )))] |
| 94 | #[doc (alias = "seekdir" )] |
| 95 | #[inline ] |
| 96 | pub fn seek(&mut self, offset: i64) -> io::Result<()> { |
| 97 | self.any_errors = false; |
| 98 | self.rewind = false; |
| 99 | self.pos = self.buf.len(); |
| 100 | match io::retry_on_intr(|| { |
| 101 | crate::backend::fs::syscalls::_seek(self.fd.as_fd(), offset, SEEK_SET) |
| 102 | }) { |
| 103 | Ok(_) => Ok(()), |
| 104 | Err(err) => { |
| 105 | self.any_errors = true; |
| 106 | Err(err) |
| 107 | } |
| 108 | } |
| 109 | } |
| 110 | |
| 111 | /// `readdir(self)`, where `None` means the end of the directory. |
| 112 | pub fn read(&mut self) -> Option<io::Result<DirEntry>> { |
| 113 | // If we've seen errors, don't continue to try to read anything |
| 114 | // further. |
| 115 | if self.any_errors { |
| 116 | return None; |
| 117 | } |
| 118 | |
| 119 | // If a rewind was requested, seek to the beginning. |
| 120 | if self.rewind { |
| 121 | self.rewind = false; |
| 122 | match io::retry_on_intr(|| { |
| 123 | crate::backend::fs::syscalls::_seek(self.fd.as_fd(), 0, SEEK_SET) |
| 124 | }) { |
| 125 | Ok(_) => (), |
| 126 | Err(err) => { |
| 127 | self.any_errors = true; |
| 128 | return Some(Err(err)); |
| 129 | } |
| 130 | } |
| 131 | } |
| 132 | |
| 133 | // Compute linux_dirent64 field offsets. |
| 134 | let z = linux_dirent64 { |
| 135 | d_ino: 0_u64, |
| 136 | d_off: 0_i64, |
| 137 | d_type: 0_u8, |
| 138 | d_reclen: 0_u16, |
| 139 | d_name: Default::default(), |
| 140 | }; |
| 141 | let base = as_ptr(&z) as usize; |
| 142 | let offsetof_d_reclen = (as_ptr(&z.d_reclen) as usize) - base; |
| 143 | let offsetof_d_name = (as_ptr(&z.d_name) as usize) - base; |
| 144 | let offsetof_d_ino = (as_ptr(&z.d_ino) as usize) - base; |
| 145 | let offsetof_d_off = (as_ptr(&z.d_off) as usize) - base; |
| 146 | let offsetof_d_type = (as_ptr(&z.d_type) as usize) - base; |
| 147 | |
| 148 | // Test if we need more entries, and if so, read more. |
| 149 | if self.buf.len() - self.pos < size_of::<linux_dirent64>() { |
| 150 | match self.read_more()? { |
| 151 | Ok(()) => (), |
| 152 | Err(err) => return Some(Err(err)), |
| 153 | } |
| 154 | } |
| 155 | |
| 156 | // We successfully read an entry. Extract the fields. |
| 157 | let pos = self.pos; |
| 158 | |
| 159 | // Do an unaligned u16 load. |
| 160 | let d_reclen = u16::from_ne_bytes([ |
| 161 | self.buf[pos + offsetof_d_reclen], |
| 162 | self.buf[pos + offsetof_d_reclen + 1], |
| 163 | ]); |
| 164 | assert!(self.buf.len() - pos >= d_reclen as usize); |
| 165 | self.pos += d_reclen as usize; |
| 166 | |
| 167 | // Read the NUL-terminated name from the `d_name` field. Without |
| 168 | // `unsafe`, we need to scan for the NUL twice: once to obtain a size |
| 169 | // for the slice, and then once within `CStr::from_bytes_with_nul`. |
| 170 | let name_start = pos + offsetof_d_name; |
| 171 | let name_len = self.buf[name_start..] |
| 172 | .iter() |
| 173 | .position(|x| *x == b' \0' ) |
| 174 | .unwrap(); |
| 175 | let name = CStr::from_bytes_with_nul(&self.buf[name_start..][..=name_len]).unwrap(); |
| 176 | let name = name.to_owned(); |
| 177 | assert!(name.as_bytes().len() <= self.buf.len() - name_start); |
| 178 | |
| 179 | // Do an unaligned `u64` load for `d_ino`. |
| 180 | let d_ino = u64::from_ne_bytes([ |
| 181 | self.buf[pos + offsetof_d_ino], |
| 182 | self.buf[pos + offsetof_d_ino + 1], |
| 183 | self.buf[pos + offsetof_d_ino + 2], |
| 184 | self.buf[pos + offsetof_d_ino + 3], |
| 185 | self.buf[pos + offsetof_d_ino + 4], |
| 186 | self.buf[pos + offsetof_d_ino + 5], |
| 187 | self.buf[pos + offsetof_d_ino + 6], |
| 188 | self.buf[pos + offsetof_d_ino + 7], |
| 189 | ]); |
| 190 | |
| 191 | // Do an unaligned `i64` load for `d_off`. |
| 192 | let d_off = i64::from_ne_bytes([ |
| 193 | self.buf[pos + offsetof_d_off], |
| 194 | self.buf[pos + offsetof_d_off + 1], |
| 195 | self.buf[pos + offsetof_d_off + 2], |
| 196 | self.buf[pos + offsetof_d_off + 3], |
| 197 | self.buf[pos + offsetof_d_off + 4], |
| 198 | self.buf[pos + offsetof_d_off + 5], |
| 199 | self.buf[pos + offsetof_d_off + 6], |
| 200 | self.buf[pos + offsetof_d_off + 7], |
| 201 | ]); |
| 202 | |
| 203 | let d_type = self.buf[pos + offsetof_d_type]; |
| 204 | |
| 205 | // Check that our types correspond to the `linux_dirent64` types. |
| 206 | let _ = linux_dirent64 { |
| 207 | d_ino, |
| 208 | d_off, |
| 209 | d_type, |
| 210 | d_reclen, |
| 211 | d_name: Default::default(), |
| 212 | }; |
| 213 | |
| 214 | Some(Ok(DirEntry { |
| 215 | d_ino, |
| 216 | d_off, |
| 217 | d_type, |
| 218 | name, |
| 219 | })) |
| 220 | } |
| 221 | |
| 222 | #[must_use ] |
| 223 | fn read_more(&mut self) -> Option<io::Result<()>> { |
| 224 | // The first few times we're called, we allocate a relatively small |
| 225 | // buffer, because many directories are small. If we're called more, |
| 226 | // use progressively larger allocations, up to a fixed maximum. |
| 227 | // |
| 228 | // The specific sizes and policy here have not been tuned in detail yet |
| 229 | // and may need to be adjusted. In doing so, we should be careful to |
| 230 | // avoid unbounded buffer growth. This buffer only exists to share the |
| 231 | // cost of a `getdents` call over many entries, so if it gets too big, |
| 232 | // cache and heap usage will outweigh the benefit. And ultimately, |
| 233 | // directories can contain more entries than we can allocate contiguous |
| 234 | // memory for, so we'll always need to cap the size at some point. |
| 235 | if self.buf.len() < 1024 * size_of::<linux_dirent64>() { |
| 236 | self.buf.reserve(32 * size_of::<linux_dirent64>()); |
| 237 | } |
| 238 | self.buf.resize(self.buf.capacity(), 0); |
| 239 | let nread = match io::retry_on_intr(|| { |
| 240 | crate::backend::fs::syscalls::getdents(self.fd.as_fd(), &mut self.buf) |
| 241 | }) { |
| 242 | Ok(nread) => nread, |
| 243 | Err(io::Errno::NOENT) => { |
| 244 | self.any_errors = true; |
| 245 | return None; |
| 246 | } |
| 247 | Err(err) => { |
| 248 | self.any_errors = true; |
| 249 | return Some(Err(err)); |
| 250 | } |
| 251 | }; |
| 252 | self.buf.resize(nread, 0); |
| 253 | self.pos = 0; |
| 254 | if nread == 0 { |
| 255 | None |
| 256 | } else { |
| 257 | Some(Ok(())) |
| 258 | } |
| 259 | } |
| 260 | |
| 261 | /// `fstat(self)` |
| 262 | #[inline ] |
| 263 | pub fn stat(&self) -> io::Result<Stat> { |
| 264 | fstat(&self.fd) |
| 265 | } |
| 266 | |
| 267 | /// `fstatfs(self)` |
| 268 | #[inline ] |
| 269 | pub fn statfs(&self) -> io::Result<StatFs> { |
| 270 | fstatfs(&self.fd) |
| 271 | } |
| 272 | |
| 273 | /// `fstatvfs(self)` |
| 274 | #[inline ] |
| 275 | pub fn statvfs(&self) -> io::Result<StatVfs> { |
| 276 | fstatvfs(&self.fd) |
| 277 | } |
| 278 | |
| 279 | /// `fchdir(self)` |
| 280 | #[cfg (feature = "process" )] |
| 281 | #[cfg_attr (docsrs, doc(cfg(feature = "process" )))] |
| 282 | #[inline ] |
| 283 | pub fn chdir(&self) -> io::Result<()> { |
| 284 | fchdir(&self.fd) |
| 285 | } |
| 286 | } |
| 287 | |
| 288 | impl Iterator for Dir { |
| 289 | type Item = io::Result<DirEntry>; |
| 290 | |
| 291 | #[inline ] |
| 292 | fn next(&mut self) -> Option<Self::Item> { |
| 293 | Self::read(self) |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | impl fmt::Debug for Dir { |
| 298 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { |
| 299 | f.debug_struct("Dir" ).field(name:"fd" , &self.fd).finish() |
| 300 | } |
| 301 | } |
| 302 | |
| 303 | /// `struct dirent` |
| 304 | #[derive (Debug)] |
| 305 | pub struct DirEntry { |
| 306 | d_ino: u64, |
| 307 | d_type: u8, |
| 308 | d_off: i64, |
| 309 | name: CString, |
| 310 | } |
| 311 | |
| 312 | impl DirEntry { |
| 313 | /// Returns the file name of this directory entry. |
| 314 | #[inline ] |
| 315 | pub fn file_name(&self) -> &CStr { |
| 316 | &self.name |
| 317 | } |
| 318 | |
| 319 | /// Returns the “offset” of this directory entry. This is not a true |
| 320 | /// numerical offset but an opaque cookie that identifies a position in the |
| 321 | /// given stream. |
| 322 | #[inline ] |
| 323 | pub fn offset(&self) -> i64 { |
| 324 | self.d_off |
| 325 | } |
| 326 | |
| 327 | /// Returns the type of this directory entry. |
| 328 | #[inline ] |
| 329 | pub fn file_type(&self) -> FileType { |
| 330 | FileType::from_dirent_d_type(self.d_type) |
| 331 | } |
| 332 | |
| 333 | /// Return the inode number of this directory entry. |
| 334 | #[inline ] |
| 335 | pub fn ino(&self) -> u64 { |
| 336 | self.d_ino |
| 337 | } |
| 338 | } |
| 339 | |
| 340 | #[cfg (test)] |
| 341 | mod tests { |
| 342 | use super::*; |
| 343 | |
| 344 | #[test ] |
| 345 | fn dir_iterator_handles_io_errors() { |
| 346 | // create a dir, keep the FD, then delete the dir |
| 347 | let tmp = tempfile::tempdir().unwrap(); |
| 348 | let fd = crate::fs::openat( |
| 349 | crate::fs::CWD, |
| 350 | tmp.path(), |
| 351 | crate::fs::OFlags::RDONLY | crate::fs::OFlags::CLOEXEC, |
| 352 | crate::fs::Mode::empty(), |
| 353 | ) |
| 354 | .unwrap(); |
| 355 | |
| 356 | let file_fd = crate::fs::openat( |
| 357 | &fd, |
| 358 | tmp.path().join("test.txt" ), |
| 359 | crate::fs::OFlags::WRONLY | crate::fs::OFlags::CREATE, |
| 360 | crate::fs::Mode::RWXU, |
| 361 | ) |
| 362 | .unwrap(); |
| 363 | |
| 364 | let mut dir = Dir::read_from(&fd).unwrap(); |
| 365 | |
| 366 | // Reach inside the `Dir` and replace its directory with a file, which |
| 367 | // will cause the subsequent `getdents64` to fail. |
| 368 | crate::io::dup2(&file_fd, &mut dir.fd).unwrap(); |
| 369 | |
| 370 | assert!(matches!(dir.next(), Some(Err(_)))); |
| 371 | assert!(dir.next().is_none()); |
| 372 | } |
| 373 | } |
| 374 | |