uucore: dedup code of unbounded splice with broker

Author: oech3

src/uu/cat/src/splice.rs

@@ -4,10 +4,10 @@
 // file that was distributed with this source code.
 use super::{CatResult, FdReadable, InputHandle};
 
-use std::io::{Read, Write};
+use std::io::Write;
 use std::os::{fd::AsFd, unix::io::AsRawFd};
 
-use uucore::pipes::{MAX_ROOTLESS_PIPE_SIZE, might_fuse, pipe, splice, splice_exact};
+use uucore::pipes::{MAX_ROOTLESS_PIPE_SIZE, might_fuse, splice};
 
 /// This function is called from `write_fast()` on Linux and Android. The
 /// function `splice()` is used to move data between two file descriptors
@@ -21,8 +21,6 @@ pub(super) fn write_fast_using_splice<R: FdReadable, S: AsRawFd + AsFd + Write>(
     handle: &InputHandle<R>,
     write_fd: &mut S,
 ) -> CatResult<bool> {
-    use std::{fs::File, sync::OnceLock};
-    static PIPE_CACHE: OnceLock<Option<(File, File)>> = OnceLock::new();
     if splice(&handle.reader, &write_fd, MAX_ROOTLESS_PIPE_SIZE).is_ok() {
         // fcntl improves throughput
         // todo: avoid fcntl overhead for small input, but don't fcntl inside of the loop
@@ -34,28 +32,10 @@ pub(super) fn write_fast_using_splice<R: FdReadable, S: AsRawFd + AsFd + Write>(
                 Err(_) => return Ok(true),
             }
         }
-    } else if let Some((pipe_rd, pipe_wr)) = PIPE_CACHE.get_or_init(|| pipe().ok()).as_ref() {
-        // both of in/output are not pipe. needs broker to use splice() with additional costs
-        loop {
-            match splice(&handle.reader, &pipe_wr, MAX_ROOTLESS_PIPE_SIZE) {
-                Ok(0) => return Ok(might_fuse(&handle.reader)),
-                Ok(n) => {
-                    if splice_exact(&pipe_rd, write_fd, n).is_err() {
-                        // If the first splice manages to copy to the intermediate
-                        // pipe, but the second splice to stdout fails for some reason
-                        // we can recover by copying the data that we have from the
-                        // intermediate pipe to stdout using normal read/write. Then
-                        // we tell the caller to fall back.
-                        let mut drain = Vec::with_capacity(n); // bounded by pipe size
-                        pipe_rd.take(n as u64).read_to_end(&mut drain)?;
-                        write_fd.write_all(&drain)?;
-                        return Ok(true);
-                    }
-                }
-                Err(_) => return Ok(true),
-            }
-        }
     } else {
-        Ok(true)
+        Ok(
+            uucore::pipes::splice_unbounded_broker(&handle.reader, write_fd)?
+                || might_fuse(&handle.reader),
+        )
     }
 }

src/uucore/src/lib/features/buf_copy/linux.rs

@@ -5,10 +5,7 @@
 
 //! Buffer-based copying implementation for Linux and Android.
 
-use crate::{
-    error::UResult,
-    pipes::{MAX_ROOTLESS_PIPE_SIZE, pipe, splice, splice_exact},
-};
+use crate::error::UResult;
 
 /// Buffer-based copying utilities for unix (excluding Linux).
 use std::{
@@ -53,7 +50,8 @@ where
 {
     // If we're on Linux or Android, try to use the splice() system call
     // for faster writing. If it works, we're done.
-    if !splice_write(src, dest)? {
+    // todo: bypass broker pipe this if input or output is pipe. We use this mostly for stream.
+    if !crate::pipes::splice_unbounded_broker(src, dest)? {
         return Ok(());
     }
 
@@ -68,43 +66,3 @@ where
     dest.flush()?;
     Ok(())
 }
-
-/// Write from source `handle` into destination `write_fd` using Linux-specific
-/// `splice` system call.
-///
-/// # Arguments
-/// - `source` - source handle
-/// - `dest` - destination handle
-#[inline]
-pub(crate) fn splice_write<R, S>(source: &R, dest: &mut S) -> UResult<bool>
-where
-    R: Read + AsFd + AsRawFd,
-    S: AsRawFd + AsFd + Write,
-{
-    let (pipe_rd, pipe_wr) = pipe()?; // todo: bypass this if input or output is pipe. We use this mostly for stream.
-    // improve throughput
-    // no need to increase pipe size of input fd since
-    // - sender with splice probably increased size already
-    // - sender without splice is bottleneck
-    let _ = rustix::pipe::fcntl_setpipe_size(&mut *dest, MAX_ROOTLESS_PIPE_SIZE);
-
-    loop {
-        match splice(&source, &pipe_wr, MAX_ROOTLESS_PIPE_SIZE) {
-            Ok(0) => return Ok(false),
-            Ok(n) => {
-                if splice_exact(&pipe_rd, dest, n).is_err() {
-                    // If the first splice manages to copy to the intermediate
-                    // pipe, but the second splice to stdout fails for some reason
-                    // we can recover by copying the data that we have from the
-                    // intermediate pipe to stdout using normal read/write. Then
-                    // we tell the caller to fall back.
-                    let mut drain = Vec::with_capacity(n); // bounded by pipe size
-                    pipe_rd.take(n as u64).read_to_end(&mut drain)?;
-                    dest.write_all(&drain)?;
-                    return Ok(true);
-                }
-            }
-            Err(_) => return Ok(true),
-        }
-    }
-}

src/uucore/src/lib/features/pipes.rs

@@ -86,6 +86,48 @@ pub fn might_fuse(source: &impl AsFd) -> bool {
     rustix::fs::fstatfs(source).map_or(true, |stats| stats.f_type == 0x6573_5546) // FUSE magic number, too many platform specific clippy warning with const
 }
 
+/// force-splice source to dest even both of them are not pipe
+/// return true if we need read/write fallback
+///
+/// This should not be used if one of them are pipe to save resources
+#[inline]
+#[cfg(any(target_os = "linux", target_os = "android"))]
+pub fn splice_unbounded_broker<R, S>(source: &R, dest: &mut S) -> std::io::Result<bool>
+where
+    R: Read + AsFd,
+    S: AsFd + std::io::Write,
+{
+    static PIPE_CACHE: OnceLock<Option<(File, File)>> = OnceLock::new();
+    let Some((pipe_rd, pipe_wr)) = PIPE_CACHE.get_or_init(|| pipe().ok()).as_ref() else {
+        return Ok(true);
+    };
+    // improve throughput
+    // no need to increase pipe size of input fd since
+    // - sender with splice probably increased size already
+    // - sender without splice is bottleneck
+    let _ = fcntl_setpipe_size(&mut *dest, MAX_ROOTLESS_PIPE_SIZE);
+
+    loop {
+        match splice(&source, &pipe_wr, MAX_ROOTLESS_PIPE_SIZE) {
+            Ok(0) => return Ok(false),
+            Ok(n) => {
+                if splice_exact(&pipe_rd, dest, n).is_err() {
+                    // If the first splice manages to copy to the intermediate
+                    // pipe, but the second splice to stdout fails for some reason
+                    // we can recover by copying the data that we have from the
+                    // intermediate pipe to stdout using normal read/write. Then
+                    // we tell the caller to fall back.
+                    let mut drain = Vec::with_capacity(n); // bounded by pipe size
+                    pipe_rd.take(n as u64).read_to_end(&mut drain)?;
+                    dest.write_all(&drain)?;
+                    return Ok(true);
+                }
+            }
+            Err(_) => return Ok(true),
+        }
+    }
+}
+
 /// splice `n` bytes with safe read/write fallback
 /// return actually sent bytes
 #[inline]
@@ -129,6 +171,7 @@ pub fn send_n_bytes(
         .get_or_init(|| pipe_with_size(pipe_size).ok())
         .as_ref()
     {
+        // todo: create fn splice_bounded_broker
         loop {
             match splice(&input, &broker_w, n as usize) {
                 Ok(0) => break might_fuse(&input),