yes: use tee syscall as fast-path

Cargo.lock

@@ -4491,6 +4491,7 @@ dependencies = [
  "clap",
  "fluent",
  "itertools 0.14.0",
+ "rustix",
  "uucore",
 ]
 

src/uu/yes/Cargo.toml

@@ -22,7 +22,8 @@ doctest = false
 clap = { workspace = true }
 itertools = { workspace = true }
 fluent = { workspace = true }
-uucore = { workspace = true }
+rustix = { workspace = true, features = ["pipe"] }
+uucore = { workspace = true, features = ["pipes"] }
 
 [[bin]]
 name = "yes"

src/uu/yes/src/yes.rs

@@ -12,7 +12,14 @@ use uucore::error::{UResult, USimpleError, strip_errno};
 use uucore::format_usage;
 use uucore::translate;
 
+#[cfg(any(target_os = "linux", target_os = "android"))]
+const PAGE_SIZE: usize = 4096;
+#[cfg(any(target_os = "linux", target_os = "android"))]
+use uucore::pipes::MAX_ROOTLESS_PIPE_SIZE;
+#[cfg(any(target_os = "linux", target_os = "android"))]
+const BUF_SIZE: usize = MAX_ROOTLESS_PIPE_SIZE;
 // it's possible that using a smaller or larger buffer might provide better performance
+#[cfg(not(any(target_os = "linux", target_os = "android")))]
 const BUF_SIZE: usize = 16 * 1024;
 
 #[uucore::main]
@@ -21,9 +28,15 @@ pub fn uumain(args: impl uucore::Args) -> UResult<()> {
 
     #[allow(clippy::unwrap_used, reason = "clap provides 'y' by default")]
     let mut buffer = args_into_buffer(matches.get_many::<OsString>("STRING").unwrap())?;
-    prepare_buffer(&mut buffer);
+    #[cfg(any(target_os = "linux", target_os = "android"))]
+    let aligned_before_growing = PAGE_SIZE.is_multiple_of(buffer.len());
 
-    match exec(&buffer) {
+    prepare_buffer(&mut buffer);
+    #[cfg(any(target_os = "linux", target_os = "android"))]
+    let res = exec(&buffer, aligned_before_growing);
+    #[cfg(not(any(target_os = "linux", target_os = "android")))]
+    let res = exec(&buffer);
+    match res {
         Ok(()) => Ok(()),
         // On Windows, silently handle broken pipe since there's no SIGPIPE
         #[cfg(windows)]
@@ -97,10 +110,44 @@ fn prepare_buffer(buf: &mut Vec<u8>) {
     }
 }
 
+#[cfg(not(any(target_os = "linux", target_os = "android")))]
 pub fn exec(bytes: &[u8]) -> io::Result<()> {
-    let stdout = io::stdout();
-    let mut stdout = stdout.lock();
+    let mut stdout = io::stdout().lock();
+    loop {
+        stdout.write_all(bytes)?;
+    }
+}
 
+#[cfg(any(target_os = "linux", target_os = "android"))]
+pub fn exec(bytes: &[u8], aligned: bool) -> io::Result<()> {
+    use uucore::pipes::{pipe, splice, tee};
+    let mut stdout = io::stdout(); // no need to lock with zero-copy
+    // don't show any error from fast-path and fallback to write for proper message
+    if let Ok((p_read, mut p_write)) = pipe()
+            // todo: zero-copy with default size when fcntl failed
+            && rustix::pipe::fcntl_setpipe_size(&stdout, MAX_ROOTLESS_PIPE_SIZE).is_ok()
+            && p_write.write_all(bytes).is_ok()
+    {
+        if aligned && tee(&p_read, &stdout, PAGE_SIZE).is_ok() {
+            while let Ok(1..) = tee(&p_read, &stdout, usize::MAX) {}
+        } else if let Ok((broker_read, broker_write)) = pipe() {
+            // tee() cannot control offset and write to non-pipe
+            'hybrid: while let Ok(mut remain) = tee(&p_read, &broker_write, usize::MAX) {
+                debug_assert!(remain == bytes.len(), "splice() should cleanup pipe");
+                while remain > 0 {
+                    if let Ok(s) = splice(&broker_read, &stdout, remain) {
+                        remain -= s;
+                    } else {
+                        // avoid output breakage with reduced remain even if it would not happen
+                        stdout.write_all(&bytes[bytes.len() - remain..])?;
+                        break 'hybrid;
+                    }
+                }
+            }
+        }
+    }
+    // fallback
+    let mut stdout = stdout.lock();
     loop {
         stdout.write_all(bytes)?;
     }
@@ -111,6 +158,7 @@ mod tests {
     use super::*;
 
     #[test]
+    #[cfg(not(any(target_os = "linux", target_os = "android")))] // Linux uses different buffer size
     fn test_prepare_buffer() {
         let tests = [
             (150, 16350),

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

@@ -258,3 +258,10 @@ pub fn dev_null() -> Option<File> {
         None
     }
 }
+
+// Less noisy wrapper around [`rustix::pipe::tee`]
+#[inline]
+#[cfg(any(target_os = "linux", target_os = "android"))]
+pub fn tee(source: &impl AsFd, target: &impl AsFd, len: usize) -> rustix::io::Result<usize> {
+    rustix::pipe::tee(source, target, len, SpliceFlags::empty())
+}

tests/by-util/test_yes.rs

@@ -70,7 +70,7 @@ fn test_long_input() {
     #[cfg(windows)]
     const TIMES: usize = 500;
     let arg = "abcdef".repeat(TIMES) + "\n";
-    let expected_out = arg.repeat(30);
+    let expected_out = arg.repeat(5);
     run(&[&arg[..arg.len() - 1]], expected_out.as_bytes());
 }