1 //! This module contains specializations that can offload `io::copy()` operations on file descriptor
2 //! containing types (`File`, `TcpStream`, etc.) to more efficient syscalls than `read(2)` and `write(2)`.
4 //! Specialization is only applied to wholly std-owned types so that user code can't observe
5 //! that the `Read` and `Write` traits are not used.
7 //! Since a copy operation involves a reader and writer side where each can consist of different types
8 //! and also involve generic wrappers (e.g. `Take`, `BufReader`) it is not practical to specialize
9 //! a single method on all possible combinations.
11 //! Instead readers and writers are handled separately by the `CopyRead` and `CopyWrite` specialization
12 //! traits and then specialized on by the `Copier::copy` method.
14 //! `Copier` uses the specialization traits to unpack the underlying file descriptors and
15 //! additional prerequisites and constraints imposed by the wrapper types.
17 //! Once it has obtained all necessary pieces and brought any wrapper types into a state where they
18 //! can be safely bypassed it will attempt to use the `copy_file_range(2)`,
19 //! `sendfile(2)` or `splice(2)` syscalls to move data directly between file descriptors.
20 //! Since those syscalls have requirements that cannot be fully checked in advance and
21 //! gathering additional information about file descriptors would require additional syscalls
22 //! anyway it simply attempts to use them one after another (guided by inaccurate hints) to
23 //! figure out which one works and and falls back to the generic read-write copy loop if none of them
25 //! Once a working syscall is found for a pair of file descriptors it will be called in a loop
26 //! until the copy operation is completed.
28 //! Advantages of using these syscalls:
30 //! * fewer context switches since reads and writes are coalesced into a single syscall
31 //! and more bytes are transferred per syscall. This translates to higher throughput
32 //! and fewer CPU cycles, at least for sufficiently large transfers to amortize the initial probing.
33 //! * `copy_file_range` creates reflink copies on CoW filesystems, thus moving less data and
34 //! consuming less disk space
35 //! * `sendfile` and `splice` can perform zero-copy IO under some circumstances while
36 //! a naive copy loop would move every byte through the CPU.
40 //! * copy operations smaller than the default buffer size can under some circumstances, especially
41 //! on older kernels, incur more syscalls than the naive approach would. As mentioned above
42 //! the syscall selection is guided by hints to minimize this possibility but they are not perfect.
43 //! * optimizations only apply to std types. If a user adds a custom wrapper type, e.g. to report
44 //! progress, they can hit a performance cliff.
48 use crate::convert::TryInto;
49 use crate::fs::{File, Metadata};
50 use crate::io::copy::generic_copy;
52 BufRead, BufReader, BufWriter, Error, Read, Result, StderrLock, StdinLock, StdoutLock, Take,
55 use crate::mem::ManuallyDrop;
56 use crate::net::TcpStream;
57 use crate::os::unix::fs::FileTypeExt;
58 use crate::os::unix::io::{AsRawFd, FromRawFd, RawFd};
59 use crate::os::unix::net::UnixStream;
60 use crate::process::{ChildStderr, ChildStdin, ChildStdout};
62 use crate::sync::atomic::{AtomicBool, Ordering};
68 pub(crate) fn copy_spec<R: Read + ?Sized, W: Write + ?Sized>(
72 let copier = Copier { read, write };
73 SpecCopy::copy(copier)
76 /// This type represents either the inferred `FileType` of a `RawFd` based on the source
77 /// type from which it was extracted or the actual metadata
79 /// The methods on this type only provide hints, due to `AsRawFd` and `FromRawFd` the inferred
80 /// type may be wrong.
82 /// We obtained the FD from a type that can contain any type of `FileType` and queried the metadata
83 /// because it is cheaper than probing all possible syscalls (reader side)
87 /// We don't have any metadata, e.g. because the original type was `File` which can represent
88 /// any `FileType` and we did not query the metadata either since it did not seem beneficial
94 fn maybe_fifo(&self) -> bool {
96 FdMeta::Metadata(meta) => meta.file_type().is_fifo(),
97 FdMeta::Socket => false,
99 FdMeta::NoneObtained => true,
103 fn potential_sendfile_source(&self) -> bool {
105 // procfs erronously shows 0 length on non-empty readable files.
106 // and if a file is truly empty then a `read` syscall will determine that and skip the write syscall
107 // thus there would be benefit from attempting sendfile
108 FdMeta::Metadata(meta)
109 if meta.file_type().is_file() && meta.len() > 0
110 || meta.file_type().is_block_device() =>
118 fn copy_file_range_candidate(&self) -> bool {
120 // copy_file_range will fail on empty procfs files. `read` can determine whether EOF has been reached
121 // without extra cost and skip the write, thus there is no benefit in attempting copy_file_range
122 FdMeta::Metadata(meta) if meta.is_file() && meta.len() > 0 => true,
123 FdMeta::NoneObtained => true,
129 struct CopyParams(FdMeta, Option<RawFd>);
131 struct Copier<'a, 'b, R: Read + ?Sized, W: Write + ?Sized> {
137 fn copy(self) -> Result<u64>;
140 impl<R: Read + ?Sized, W: Write + ?Sized> SpecCopy for Copier<'_, '_, R, W> {
141 default fn copy(self) -> Result<u64> {
142 generic_copy(self.read, self.write)
146 impl<R: CopyRead, W: CopyWrite> SpecCopy for Copier<'_, '_, R, W> {
147 fn copy(self) -> Result<u64> {
148 let (reader, writer) = (self.read, self.write);
149 let r_cfg = reader.properties();
150 let w_cfg = writer.properties();
152 // before direct operations on file descriptors ensure that all source and sink buffers are empty
153 let mut flush = || -> crate::io::Result<u64> {
154 let bytes = reader.drain_to(writer, u64::MAX)?;
155 // BufWriter buffered bytes have already been accounted for in earlier write() calls
160 let mut written = 0u64;
162 if let (CopyParams(input_meta, Some(readfd)), CopyParams(output_meta, Some(writefd))) =
166 let max_write = reader.min_limit();
168 if input_meta.copy_file_range_candidate() && output_meta.copy_file_range_candidate() {
169 let result = copy_regular_files(readfd, writefd, max_write);
170 result.update_take(reader);
173 CopyResult::Ended(bytes_copied) => return Ok(bytes_copied + written),
174 CopyResult::Error(e, _) => return Err(e),
175 CopyResult::Fallback(bytes) => written += bytes,
179 // on modern kernels sendfile can copy from any mmapable type (some but not all regular files and block devices)
180 // to any writable file descriptor. On older kernels the writer side can only be a socket.
181 // So we just try and fallback if needed.
182 // If current file offsets + write sizes overflow it may also fail, we do not try to fix that and instead
183 // fall back to the generic copy loop.
184 if input_meta.potential_sendfile_source() {
185 let result = sendfile_splice(SpliceMode::Sendfile, readfd, writefd, max_write);
186 result.update_take(reader);
189 CopyResult::Ended(bytes_copied) => return Ok(bytes_copied + written),
190 CopyResult::Error(e, _) => return Err(e),
191 CopyResult::Fallback(bytes) => written += bytes,
195 if input_meta.maybe_fifo() || output_meta.maybe_fifo() {
196 let result = sendfile_splice(SpliceMode::Splice, readfd, writefd, max_write);
197 result.update_take(reader);
200 CopyResult::Ended(bytes_copied) => return Ok(bytes_copied + written),
201 CopyResult::Error(e, _) => return Err(e),
202 CopyResult::Fallback(0) => { /* use the fallback below */ }
203 CopyResult::Fallback(_) => {
204 unreachable!("splice should not return > 0 bytes on the fallback path")
210 // fallback if none of the more specialized syscalls wants to work with these file descriptors
211 match generic_copy(reader, writer) {
212 Ok(bytes) => Ok(bytes + written),
218 #[rustc_specialization_trait]
219 trait CopyRead: Read {
220 /// Implementations that contain buffers (i.e. `BufReader`) must transfer data from their internal
221 /// buffers into `writer` until either the buffers are emptied or `limit` bytes have been
222 /// transferred, whichever occurs sooner.
223 /// If nested buffers are present the outer buffers must be drained first.
225 /// This is necessary to directly bypass the wrapper types while preserving the data order
226 /// when operating directly on the underlying file descriptors.
227 fn drain_to<W: Write>(&mut self, _writer: &mut W, _limit: u64) -> Result<u64> {
231 /// Updates `Take` wrappers to remove the number of bytes copied.
232 fn taken(&mut self, _bytes: u64) {}
234 /// The minimum of the limit of all `Take<_>` wrappers, `u64::MAX` otherwise.
235 /// This method does not account for data `BufReader` buffers and would underreport
236 /// the limit of a `Take<BufReader<Take<_>>>` type. Thus its result is only valid
237 /// after draining the buffers via `drain_to`.
238 fn min_limit(&self) -> u64 {
242 /// Extracts the file descriptor and hints/metadata, delegating through wrappers if necessary.
243 fn properties(&self) -> CopyParams;
246 #[rustc_specialization_trait]
247 trait CopyWrite: Write {
248 /// Extracts the file descriptor and hints/metadata, delegating through wrappers if necessary.
249 fn properties(&self) -> CopyParams;
252 impl<T> CopyRead for &mut T
256 fn drain_to<W: Write>(&mut self, writer: &mut W, limit: u64) -> Result<u64> {
257 (**self).drain_to(writer, limit)
260 fn taken(&mut self, bytes: u64) {
261 (**self).taken(bytes);
264 fn min_limit(&self) -> u64 {
268 fn properties(&self) -> CopyParams {
269 (**self).properties()
273 impl<T> CopyWrite for &mut T
277 fn properties(&self) -> CopyParams {
278 (**self).properties()
282 impl CopyRead for File {
283 fn properties(&self) -> CopyParams {
284 CopyParams(fd_to_meta(self), Some(self.as_raw_fd()))
288 impl CopyRead for &File {
289 fn properties(&self) -> CopyParams {
290 CopyParams(fd_to_meta(*self), Some(self.as_raw_fd()))
294 impl CopyWrite for File {
295 fn properties(&self) -> CopyParams {
296 CopyParams(FdMeta::NoneObtained, Some(self.as_raw_fd()))
300 impl CopyWrite for &File {
301 fn properties(&self) -> CopyParams {
302 CopyParams(FdMeta::NoneObtained, Some(self.as_raw_fd()))
306 impl CopyRead for TcpStream {
307 fn properties(&self) -> CopyParams {
308 // avoid the stat syscall since we can be fairly sure it's a socket
309 CopyParams(FdMeta::Socket, Some(self.as_raw_fd()))
313 impl CopyRead for &TcpStream {
314 fn properties(&self) -> CopyParams {
315 // avoid the stat syscall since we can be fairly sure it's a socket
316 CopyParams(FdMeta::Socket, Some(self.as_raw_fd()))
320 impl CopyWrite for TcpStream {
321 fn properties(&self) -> CopyParams {
322 // avoid the stat syscall since we can be fairly sure it's a socket
323 CopyParams(FdMeta::Socket, Some(self.as_raw_fd()))
327 impl CopyWrite for &TcpStream {
328 fn properties(&self) -> CopyParams {
329 // avoid the stat syscall since we can be fairly sure it's a socket
330 CopyParams(FdMeta::Socket, Some(self.as_raw_fd()))
334 impl CopyRead for UnixStream {
335 fn properties(&self) -> CopyParams {
336 // avoid the stat syscall since we can be fairly sure it's a socket
337 CopyParams(FdMeta::Socket, Some(self.as_raw_fd()))
341 impl CopyRead for &UnixStream {
342 fn properties(&self) -> CopyParams {
343 // avoid the stat syscall since we can be fairly sure it's a socket
344 CopyParams(FdMeta::Socket, Some(self.as_raw_fd()))
348 impl CopyWrite for UnixStream {
349 fn properties(&self) -> CopyParams {
350 // avoid the stat syscall since we can be fairly sure it's a socket
351 CopyParams(FdMeta::Socket, Some(self.as_raw_fd()))
355 impl CopyWrite for &UnixStream {
356 fn properties(&self) -> CopyParams {
357 // avoid the stat syscall since we can be fairly sure it's a socket
358 CopyParams(FdMeta::Socket, Some(self.as_raw_fd()))
362 impl CopyWrite for ChildStdin {
363 fn properties(&self) -> CopyParams {
364 CopyParams(FdMeta::Pipe, Some(self.as_raw_fd()))
368 impl CopyRead for ChildStdout {
369 fn properties(&self) -> CopyParams {
370 CopyParams(FdMeta::Pipe, Some(self.as_raw_fd()))
374 impl CopyRead for ChildStderr {
375 fn properties(&self) -> CopyParams {
376 CopyParams(FdMeta::Pipe, Some(self.as_raw_fd()))
380 impl CopyRead for StdinLock<'_> {
381 fn drain_to<W: Write>(&mut self, writer: &mut W, outer_limit: u64) -> Result<u64> {
382 let buf_reader = self.as_mut_buf();
383 let buf = buf_reader.buffer();
384 let buf = &buf[0..min(buf.len(), outer_limit.try_into().unwrap_or(usize::MAX))];
385 let bytes_drained = buf.len();
386 writer.write_all(buf)?;
387 buf_reader.consume(bytes_drained);
389 Ok(bytes_drained as u64)
392 fn properties(&self) -> CopyParams {
393 CopyParams(fd_to_meta(self), Some(self.as_raw_fd()))
397 impl CopyWrite for StdoutLock<'_> {
398 fn properties(&self) -> CopyParams {
399 CopyParams(FdMeta::NoneObtained, Some(self.as_raw_fd()))
403 impl CopyWrite for StderrLock<'_> {
404 fn properties(&self) -> CopyParams {
405 CopyParams(FdMeta::NoneObtained, Some(self.as_raw_fd()))
409 impl<T: CopyRead> CopyRead for Take<T> {
410 fn drain_to<W: Write>(&mut self, writer: &mut W, outer_limit: u64) -> Result<u64> {
411 let local_limit = self.limit();
412 let combined_limit = min(outer_limit, local_limit);
413 let bytes_drained = self.get_mut().drain_to(writer, combined_limit)?;
414 // update limit since read() was bypassed
415 self.set_limit(local_limit - bytes_drained);
420 fn taken(&mut self, bytes: u64) {
421 self.set_limit(self.limit() - bytes);
422 self.get_mut().taken(bytes);
425 fn min_limit(&self) -> u64 {
426 min(Take::limit(self), self.get_ref().min_limit())
429 fn properties(&self) -> CopyParams {
430 self.get_ref().properties()
434 impl<T: CopyRead> CopyRead for BufReader<T> {
435 fn drain_to<W: Write>(&mut self, writer: &mut W, outer_limit: u64) -> Result<u64> {
436 let buf = self.buffer();
437 let buf = &buf[0..min(buf.len(), outer_limit.try_into().unwrap_or(usize::MAX))];
438 let bytes = buf.len();
439 writer.write_all(buf)?;
442 let remaining = outer_limit - bytes as u64;
444 // in case of nested bufreaders we also need to drain the ones closer to the source
445 let inner_bytes = self.get_mut().drain_to(writer, remaining)?;
447 Ok(bytes as u64 + inner_bytes)
450 fn taken(&mut self, bytes: u64) {
451 self.get_mut().taken(bytes);
454 fn min_limit(&self) -> u64 {
455 self.get_ref().min_limit()
458 fn properties(&self) -> CopyParams {
459 self.get_ref().properties()
463 impl<T: CopyWrite> CopyWrite for BufWriter<T> {
464 fn properties(&self) -> CopyParams {
465 self.get_ref().properties()
469 fn fd_to_meta<T: AsRawFd>(fd: &T) -> FdMeta {
470 let fd = fd.as_raw_fd();
471 let file: ManuallyDrop<File> = ManuallyDrop::new(unsafe { File::from_raw_fd(fd) });
472 match file.metadata() {
473 Ok(meta) => FdMeta::Metadata(meta),
474 Err(_) => FdMeta::NoneObtained,
478 pub(super) enum CopyResult {
485 fn update_take(&self, reader: &mut impl CopyRead) {
487 CopyResult::Fallback(bytes)
488 | CopyResult::Ended(bytes)
489 | CopyResult::Error(_, bytes) => reader.taken(bytes),
494 /// linux-specific implementation that will attempt to use copy_file_range for copy offloading
495 /// as the name says, it only works on regular files
497 /// Callers must handle fallback to a generic copy loop.
498 /// `Fallback` may indicate non-zero number of bytes already written
499 /// if one of the files' cursor +`max_len` would exceed u64::MAX (`EOVERFLOW`).
500 pub(super) fn copy_regular_files(reader: RawFd, writer: RawFd, max_len: u64) -> CopyResult {
503 // Kernel prior to 4.5 don't have copy_file_range
504 // We store the availability in a global to avoid unnecessary syscalls
505 static HAS_COPY_FILE_RANGE: AtomicBool = AtomicBool::new(true);
510 off_in: *mut libc::loff_t,
512 off_out: *mut libc::loff_t,
518 let has_copy_file_range = HAS_COPY_FILE_RANGE.load(Ordering::Relaxed);
519 let mut written = 0u64;
520 while written < max_len {
521 let copy_result = if has_copy_file_range {
522 let bytes_to_copy = cmp::min(max_len - written, usize::MAX as u64);
523 // cap to 1GB chunks in case u64::MAX is passed as max_len and the file has a non-zero seek position
524 // this allows us to copy large chunks without hitting EOVERFLOW,
525 // unless someone sets a file offset close to u64::MAX - 1GB, in which case a fallback would be required
526 let bytes_to_copy = cmp::min(bytes_to_copy as usize, 0x4000_0000usize);
527 let copy_result = unsafe {
528 // We actually don't have to adjust the offsets,
529 // because copy_file_range adjusts the file offset automatically
539 if let Err(ref copy_err) = copy_result {
540 match copy_err.raw_os_error() {
541 Some(libc::ENOSYS | libc::EPERM | libc::EOPNOTSUPP) => {
542 HAS_COPY_FILE_RANGE.store(false, Ordering::Relaxed);
549 Err(Error::from_raw_os_error(libc::ENOSYS))
552 Ok(0) if written == 0 => {
553 // fallback to work around several kernel bugs where copy_file_range will fail to
554 // copy any bytes and return 0 instead of an error if
555 // - reading virtual files from the proc filesystem which appear to have 0 size
556 // but are not empty. noted in coreutils to affect kernels at least up to 5.6.19.
557 // - copying from an overlay filesystem in docker. reported to occur on fedora 32.
558 return CopyResult::Fallback(0);
560 Ok(0) => return CopyResult::Ended(written), // reached EOF
561 Ok(ret) => written += ret as u64,
563 return match err.raw_os_error() {
564 // when file offset + max_length > u64::MAX
565 Some(libc::EOVERFLOW) => CopyResult::Fallback(written),
567 libc::ENOSYS | libc::EXDEV | libc::EINVAL | libc::EPERM | libc::EOPNOTSUPP,
569 // Try fallback io::copy if either:
570 // - Kernel version is < 4.5 (ENOSYS)
571 // - Files are mounted on different fs (EXDEV)
572 // - copy_file_range is broken in various ways on RHEL/CentOS 7 (EOPNOTSUPP)
573 // - copy_file_range is disallowed, for example by seccomp (EPERM)
574 // - copy_file_range cannot be used with pipes or device nodes (EINVAL)
575 assert_eq!(written, 0);
576 CopyResult::Fallback(0)
578 _ => CopyResult::Error(err, written),
583 CopyResult::Ended(written)
592 /// performs splice or sendfile between file descriptors
593 /// Does _not_ fall back to a generic copy loop.
594 fn sendfile_splice(mode: SpliceMode, reader: RawFd, writer: RawFd, len: u64) -> CopyResult {
595 static HAS_SENDFILE: AtomicBool = AtomicBool::new(true);
596 static HAS_SPLICE: AtomicBool = AtomicBool::new(true);
601 src_offset: *const i64,
603 dst_offset: *const i64,
610 SpliceMode::Sendfile if !HAS_SENDFILE.load(Ordering::Relaxed) => {
611 return CopyResult::Fallback(0);
613 SpliceMode::Splice if !HAS_SPLICE.load(Ordering::Relaxed) => {
614 return CopyResult::Fallback(0);
619 let mut written = 0u64;
620 while written < len {
621 // according to its manpage that's the maximum size sendfile() will copy per invocation
622 let chunk_size = crate::cmp::min(len - written, 0x7ffff000_u64) as usize;
624 let result = match mode {
625 SpliceMode::Sendfile => {
626 cvt(unsafe { libc::sendfile(writer, reader, ptr::null_mut(), chunk_size) })
628 SpliceMode::Splice => cvt(unsafe {
629 splice(reader, ptr::null_mut(), writer, ptr::null_mut(), chunk_size, 0)
634 Ok(0) => break, // EOF
635 Ok(ret) => written += ret as u64,
637 return match err.raw_os_error() {
638 Some(libc::ENOSYS | libc::EPERM) => {
639 // syscall not supported (ENOSYS)
640 // syscall is disallowed, e.g. by seccomp (EPERM)
642 SpliceMode::Sendfile => HAS_SENDFILE.store(false, Ordering::Relaxed),
643 SpliceMode::Splice => HAS_SPLICE.store(false, Ordering::Relaxed),
645 assert_eq!(written, 0);
646 CopyResult::Fallback(0)
648 Some(libc::EINVAL) => {
649 // splice/sendfile do not support this particular file descriptor (EINVAL)
650 assert_eq!(written, 0);
651 CopyResult::Fallback(0)
653 Some(os_err) if mode == SpliceMode::Sendfile && os_err == libc::EOVERFLOW => {
654 CopyResult::Fallback(written)
656 _ => CopyResult::Error(err, written),
661 CopyResult::Ended(written)