yes: use tee syscall as pipe only fast-path

Author: oech3

.vscode/cspell.dictionaries/workspace.wordlist.txt

@@ -8,6 +8,7 @@ advapi32-sys
 aho-corasick
 backtrace
 blake2b_simd
+rustix
 
 # * uutils project
 uutils
@@ -361,6 +362,7 @@ uutils
 
 # * function names
 getcwd
+setpipe
 
 # * other
 weblate

Cargo.lock

@@ -435,6 +435,7 @@ rstest = "0.26.0"
 rstest_reuse = "0.7.0"
 rustc-hash = "2.1.1"
 rust-ini = "0.21.0"
+rustix = "1.1.4"
 same-file = "1.0.6"
 self_cell = "1.0.4"
 selinux = "=0.6.0"

Cargo.toml

@@ -22,6 +22,7 @@ path = "src/yes.rs"
 clap = { workspace = true }
 itertools = { workspace = true }
 fluent = { workspace = true }
+rustix = { workspace = true, features = ["fs", "pipe"] }
 
 [target.'cfg(unix)'.dependencies]
 uucore = { workspace = true, features = ["pipes", "signals"] }

src/uu/yes/Cargo.toml

@@ -13,8 +13,12 @@ use uucore::error::{UResult, USimpleError};
 use uucore::format_usage;
 use uucore::translate;
 
-// it's possible that using a smaller or larger buffer might provide better performance on some
-// systems, but honestly this is good enough
+#[cfg(target_os = "linux")]
+const MAX_ROOTLESS_PIPE_SIZE: usize = 1024 * 1024;
+// todo: investigate best rate
+#[cfg(target_os = "linux")]
+const BUF_SIZE: usize = MAX_ROOTLESS_PIPE_SIZE;
+#[cfg(not(target_os = "linux"))]
 const BUF_SIZE: usize = 16 * 1024;
 
 #[uucore::main]
@@ -110,8 +114,30 @@ fn prepare_buffer(buf: &mut Vec<u8>) {
 
 pub fn exec(bytes: &[u8]) -> io::Result<()> {
     let stdout = io::stdout();
-    let mut stdout = stdout.lock();
+    #[cfg(target_os = "linux")]
+    {
+        use rustix::io::write;
+        use rustix::pipe::{SpliceFlags, fcntl_setpipe_size, pipe, tee};
+        // fast-path for pipe
+        if bytes.len() < MAX_ROOTLESS_PIPE_SIZE * 4
+            && let Ok((p_read, p_write)) = pipe()
+        {
+            let _ = fcntl_setpipe_size(&p_read, MAX_ROOTLESS_PIPE_SIZE);
+            let _ = fcntl_setpipe_size(&p_write, MAX_ROOTLESS_PIPE_SIZE);
+            let _ = fcntl_setpipe_size(&stdout, MAX_ROOTLESS_PIPE_SIZE);
+            let _ = write(&p_write, bytes);
+            loop {
+                match tee(&p_read, &stdout, MAX_ROOTLESS_PIPE_SIZE, SpliceFlags::MORE) {
+                    Ok(n) if n > 0 => {}
+                    Ok(0) => break,
+                    Err(e) if e.kind() == io::ErrorKind::Interrupted => continue,
+                    _ => break,
+                }
+            }
+        }
+    }
 
+    let mut stdout = stdout.lock();
     loop {
         stdout.write_all(bytes)?;
     }
@@ -122,6 +148,7 @@ mod tests {
     use super::*;
 
     #[test]
+    #[cfg(not(target_os = "linux"))] // Linux uses different buffer size
     fn test_prepare_buffer() {
         let tests = [
             (150, 16350),