Merge pull request #8959 from mattsu2020/sort-memory-functions

feat(sort): auto-tune buffer sizing from available memory

.vscode/cspell.dictionaries/jargon.wordlist.txt

@@ -131,6 +131,7 @@ symlink
 symlinks
 syscall
 syscalls
+sysconf
 tokenize
 toolchain
 truthy

fuzz/Cargo.lock

@@ -231,18 +231,18 @@ dependencies = [
 
 [[package]]
 name = "clap"
-version = "4.5.49"
+version = "4.5.50"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f4512b90fa68d3a9932cea5184017c5d200f5921df706d45e853537dea51508f"
+checksum = "0c2cfd7bf8a6017ddaa4e32ffe7403d547790db06bd171c1c53926faab501623"
 dependencies = [
  "clap_builder",
 ]
 
 [[package]]
 name = "clap_builder"
-version = "4.5.49"
+version = "4.5.50"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0025e98baa12e766c67ba13ff4695a887a1eba19569aad00a472546795bd6730"
+checksum = "0a4c05b9e80c5ccd3a7ef080ad7b6ba7d6fc00a985b8b157197075677c82c7a0"
 dependencies = [
  "anstream",
  "anstyle",
@@ -1538,9 +1538,9 @@ dependencies = [
 
 [[package]]
 name = "unicode-ident"
-version = "1.0.19"
+version = "1.0.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f63a545481291138910575129486daeaf8ac54aee4387fe7906919f7830c7d9d"
+checksum = "462eeb75aeb73aea900253ce739c8e18a67423fadf006037cd3ff27e82748a06"
 
 [[package]]
 name = "unicode-width"

src/uu/sort/Cargo.toml

@@ -36,8 +36,6 @@ thiserror = { workspace = true }
 unicode-width = { workspace = true }
 uucore = { workspace = true, features = ["fs", "parser", "version-cmp"] }
 fluent = { workspace = true }
-
-[target.'cfg(target_os = "linux")'.dependencies]
 nix = { workspace = true }
 
 [dev-dependencies]

src/uu/sort/src/buffer_hint.rs

@@ -0,0 +1,152 @@
+// This file is part of the uutils coreutils package.
+//
+// For the full copyright and license information, please view the LICENSE
+// file that was distributed with this source code.
+
+//! Heuristics for determining buffer size for external sorting.
+use std::ffi::OsString;
+
+use crate::{
+    FALLBACK_AUTOMATIC_BUF_SIZE, MAX_AUTOMATIC_BUF_SIZE, MIN_AUTOMATIC_BUF_SIZE, STDIN_FILE,
+};
+
+// Heuristics to size the external sort buffer without overcommit memory.
+pub(crate) fn automatic_buffer_size(files: &[OsString]) -> usize {
+    let file_hint = file_size_hint(files);
+    let mem_hint = available_memory_hint();
+
+    // Prefer the tighter bound when both hints exist, otherwise fall back to whichever hint is available.
+    match (file_hint, mem_hint) {
+        (Some(file), Some(mem)) => file.min(mem),
+        (Some(file), None) => file,
+        (None, Some(mem)) => mem,
+        (None, None) => FALLBACK_AUTOMATIC_BUF_SIZE,
+    }
+}
+
+fn file_size_hint(files: &[OsString]) -> Option<usize> {
+    // Estimate total bytes across real files; non-regular inputs are skipped.
+    let mut total_bytes: u128 = 0;
+
+    for file in files {
+        if file == STDIN_FILE {
+            continue;
+        }
+
+        let Ok(metadata) = std::fs::metadata(file) else {
+            continue;
+        };
+
+        if !metadata.is_file() {
+            continue;
+        }
+
+        total_bytes = total_bytes.saturating_add(metadata.len() as u128);
+
+        if total_bytes >= (MAX_AUTOMATIC_BUF_SIZE as u128) * 8 {
+            break;
+        }
+    }
+
+    if total_bytes == 0 {
+        return None;
+    }
+
+    let desired_bytes = desired_file_buffer_bytes(total_bytes);
+    Some(clamp_hint(desired_bytes))
+}
+
+fn available_memory_hint() -> Option<usize> {
+    #[cfg(target_os = "linux")]
+    if let Some(bytes) = uucore::parser::parse_size::available_memory_bytes() {
+        return Some(clamp_hint(bytes / 4));
+    }
+
+    physical_memory_bytes().map(|bytes| clamp_hint(bytes / 4))
+}
+
+fn clamp_hint(bytes: u128) -> usize {
+    let min = MIN_AUTOMATIC_BUF_SIZE as u128;
+    let max = MAX_AUTOMATIC_BUF_SIZE as u128;
+    let clamped = bytes.clamp(min, max);
+    clamped.min(usize::MAX as u128) as usize
+}
+
+fn desired_file_buffer_bytes(total_bytes: u128) -> u128 {
+    if total_bytes == 0 {
+        return 0;
+    }
+
+    let max = MAX_AUTOMATIC_BUF_SIZE as u128;
+
+    if total_bytes <= max {
+        return total_bytes.saturating_mul(12).clamp(total_bytes, max);
+    }
+
+    let quarter = total_bytes / 4;
+    quarter.max(max)
+}
+
+fn physical_memory_bytes() -> Option<u128> {
+    #[cfg(all(
+        target_family = "unix",
+        not(target_os = "redox"),
+        any(target_os = "linux", target_os = "android")
+    ))]
+    {
+        physical_memory_bytes_unix()
+    }
+
+    #[cfg(any(
+        not(target_family = "unix"),
+        target_os = "redox",
+        not(any(target_os = "linux", target_os = "android"))
+    ))]
+    {
+        // No portable or safe API is available here to detect total physical memory.
+        None
+    }
+}
+
+#[cfg(all(
+    target_family = "unix",
+    not(target_os = "redox"),
+    any(target_os = "linux", target_os = "android")
+))]
+fn physical_memory_bytes_unix() -> Option<u128> {
+    use nix::unistd::{SysconfVar, sysconf};
+
+    let pages = match sysconf(SysconfVar::_PHYS_PAGES) {
+        Ok(Some(pages)) if pages > 0 => u128::try_from(pages).ok()?,
+        _ => return None,
+    };
+
+    let page_size = match sysconf(SysconfVar::PAGE_SIZE) {
+        Ok(Some(page_size)) if page_size > 0 => u128::try_from(page_size).ok()?,
+        _ => return None,
+    };
+
+    Some(pages.saturating_mul(page_size))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn desired_buffer_matches_total_when_small() {
+        let six_mebibytes = 6 * 1024 * 1024;
+        let expected = ((six_mebibytes as u128) * 12)
+            .clamp(six_mebibytes as u128, crate::MAX_AUTOMATIC_BUF_SIZE as u128);
+        assert_eq!(desired_file_buffer_bytes(six_mebibytes as u128), expected);
+    }
+
+    #[test]
+    fn desired_buffer_caps_at_max_for_large_inputs() {
+        let large = 256 * 1024 * 1024; // 256 MiB
+        assert_eq!(
+            desired_file_buffer_bytes(large as u128),
+            crate::MAX_AUTOMATIC_BUF_SIZE as u128
+        );
+    }
+}

src/uu/sort/src/chunks.rs

@@ -271,11 +271,12 @@ fn read_to_buffer<T: Read>(
                         if max_buffer_size > buffer.len() {
                             // we can grow the buffer
                             let prev_len = buffer.len();
-                            if buffer.len() < max_buffer_size / 2 {
-                                buffer.resize(buffer.len() * 2, 0);
+                            let target = if buffer.len() < max_buffer_size / 2 {
+                                buffer.len().saturating_mul(2)
                             } else {
-                                buffer.resize(max_buffer_size, 0);
-                            }
+                                max_buffer_size
+                            };
+                            buffer.resize(target.min(max_buffer_size), 0);
                             read_target = &mut buffer[prev_len..];
                             continue;
                         }
@@ -295,8 +296,8 @@ fn read_to_buffer<T: Read>(
 
                     // We need to read more lines
                     let len = buffer.len();
-                    // resize the vector to 10 KB more
-                    buffer.resize(len + 1024 * 10, 0);
+                    let grow_by = (len / 2).max(1024 * 1024);
+                    buffer.resize(len + grow_by, 0);
                     read_target = &mut buffer[len..];
                 } else {
                     // This file has been fully read.

src/uu/sort/src/ext_sort.rs

@@ -86,9 +86,15 @@ fn reader_writer<
 ) -> UResult<()> {
     let separator = settings.line_ending.into();
 
-    // Heuristically chosen: Dividing by 10 seems to keep our memory usage roughly
-    // around settings.buffer_size as a whole.
-    let buffer_size = settings.buffer_size / 10;
+    // Cap oversized buffer requests to avoid unnecessary allocations and give the automatic
+    // heuristic room to grow when the user does not provide an explicit value.
+    let mut buffer_size = match settings.buffer_size {
+        size if size <= 512 * 1024 * 1024 => size,
+        size => size / 2,
+    };
+    if !settings.buffer_size_is_explicit {
+        buffer_size = buffer_size.max(8 * 1024 * 1024);
+    }
     let read_result: ReadResult<Tmp> = read_write_loop(
         files,
         tmp_dir,

src/uu/sort/src/sort.rs

@@ -9,6 +9,7 @@
 
 // spell-checker:ignore (misc) HFKJFK Mbdfhn getrlimit RLIMIT_NOFILE rlim bigdecimal extendedbigdecimal hexdigit
 
+mod buffer_hint;
 mod check;
 mod chunks;
 mod custom_str_cmp;
@@ -54,6 +55,7 @@ use uucore::show_error;
 use uucore::translate;
 use uucore::version_cmp::version_cmp;
 
+use crate::buffer_hint::automatic_buffer_size;
 use crate::tmp_dir::TmpDirWrapper;
 
 mod options {
@@ -115,10 +117,12 @@ const DECIMAL_PT: u8 = b'.';
 const NEGATIVE: &u8 = &b'-';
 const POSITIVE: &u8 = &b'+';
 
-// Choosing a higher buffer size does not result in performance improvements
-// (at least not on my machine). TODO: In the future, we should also take the amount of
-// available memory into consideration, instead of relying on this constant only.
-const DEFAULT_BUF_SIZE: usize = 1_000_000_000; // 1 GB
+// The automatic buffer heuristics clamp to this range to avoid
+// over-committing memory on constrained systems while still keeping
+// reasonably large chunks for typical workloads.
+const MIN_AUTOMATIC_BUF_SIZE: usize = 512 * 1024; // 512 KiB
+const FALLBACK_AUTOMATIC_BUF_SIZE: usize = 32 * 1024 * 1024; // 32 MiB
+const MAX_AUTOMATIC_BUF_SIZE: usize = 1024 * 1024 * 1024; // 1 GiB
 
 #[derive(Debug, Error)]
 pub enum SortError {
@@ -283,6 +287,7 @@ pub struct GlobalSettings {
     threads: String,
     line_ending: LineEnding,
     buffer_size: usize,
+    buffer_size_is_explicit: bool,
     compress_prog: Option<String>,
     merge_batch_size: usize,
     precomputed: Precomputed,
@@ -359,9 +364,10 @@ impl Default for GlobalSettings {
             separator: None,
             threads: String::new(),
             line_ending: LineEnding::Newline,
-            buffer_size: DEFAULT_BUF_SIZE,
+            buffer_size: FALLBACK_AUTOMATIC_BUF_SIZE,
+            buffer_size_is_explicit: false,
             compress_prog: None,
-            merge_batch_size: 32,
+            merge_batch_size: default_merge_batch_size(),
             precomputed: Precomputed::default(),
         }
     }
@@ -1036,6 +1042,31 @@ fn get_rlimit() -> UResult<usize> {
 }
 
 const STDIN_FILE: &str = "-";
+#[cfg(target_os = "linux")]
+const LINUX_BATCH_DIVISOR: usize = 4;
+#[cfg(target_os = "linux")]
+const LINUX_BATCH_MIN: usize = 32;
+#[cfg(target_os = "linux")]
+const LINUX_BATCH_MAX: usize = 256;
+
+fn default_merge_batch_size() -> usize {
+    #[cfg(target_os = "linux")]
+    {
+        // Adjust merge batch size dynamically based on available file descriptors.
+        match get_rlimit() {
+            Ok(limit) => {
+                let usable_limit = limit.saturating_div(LINUX_BATCH_DIVISOR);
+                usable_limit.clamp(LINUX_BATCH_MIN, LINUX_BATCH_MAX)
+            }
+            Err(_) => 64,
+        }
+    }
+
+    #[cfg(not(target_os = "linux"))]
+    {
+        64
+    }
+}
 
 #[uucore::main]
 #[allow(clippy::cognitive_complexity)]
@@ -1157,14 +1188,15 @@ pub fn uumain(args: impl uucore::Args) -> UResult<()> {
         }
     }
 
-    settings.buffer_size =
-        matches
-            .get_one::<String>(options::BUF_SIZE)
-            .map_or(Ok(DEFAULT_BUF_SIZE), |s| {
-                GlobalSettings::parse_byte_count(s).map_err(|e| {
-                    USimpleError::new(2, format_error_message(&e, s, options::BUF_SIZE))
-                })
-            })?;
+    if let Some(size_str) = matches.get_one::<String>(options::BUF_SIZE) {
+        settings.buffer_size = GlobalSettings::parse_byte_count(size_str).map_err(|e| {
+            USimpleError::new(2, format_error_message(&e, size_str, options::BUF_SIZE))
+        })?;
+        settings.buffer_size_is_explicit = true;
+    } else {
+        settings.buffer_size = automatic_buffer_size(&files);
+        settings.buffer_size_is_explicit = false;
+    }
 
     let mut tmp_dir = TmpDirWrapper::new(
         matches