Auto merge of #97925 - the8472:cgroupv1, r=joshtriplett

Add cgroupv1 support to available_parallelism

Fixes #97549

My dev machine uses cgroup v2 so I was only able to test that code path. So the v1 code path is written only based on documentation. I could use some help testing that it works on a machine with cgroups v1:

```
$ x.py build --stage 1

# quota.rs
fn main() {
    println!("{:?}", std:🧵:available_parallelism());
}

# assuming stage1 is linked in rustup
$ rust +stage1 quota.rs

# spawn a new cgroup scope for the current user
$ sudo systemd-run -p CPUQuota="300%" --uid=$(id -u) -tdS

# should print Ok(3)
$ ./quota
```

If it doesn't work as expected an strace, the contents of `/proc/self/cgroups` and the structure of `/sys/fs/cgroups` would help.

library/std/src/lib.rs

@@ -276,6 +276,7 @@
 #![feature(hasher_prefixfree_extras)]
 #![feature(hashmap_internals)]
 #![feature(int_error_internals)]
+#![feature(is_some_with)]
 #![feature(maybe_uninit_slice)]
 #![feature(maybe_uninit_write_slice)]
 #![feature(mixed_integer_ops)]

library/std/src/sys/unix/thread.rs

@@ -285,7 +285,7 @@ pub fn available_parallelism() -> io::Result<NonZeroUsize> {
         ))] {
             #[cfg(any(target_os = "android", target_os = "linux"))]
             {
-                let quota = cgroup2_quota().max(1);
+                let quota = cgroups::quota().max(1);
                 let mut set: libc::cpu_set_t = unsafe { mem::zeroed() };
                 unsafe {
                     if libc::sched_getaffinity(0, mem::size_of::<libc::cpu_set_t>(), &mut set) == 0 {
@@ -379,49 +379,88 @@ pub fn available_parallelism() -> io::Result<NonZeroUsize> {
     }
 }
 
-/// Returns cgroup CPU quota in core-equivalents, rounded down, or usize::MAX if the quota cannot
-/// be determined or is not set.
 #[cfg(any(target_os = "android", target_os = "linux"))]
-fn cgroup2_quota() -> usize {
+mod cgroups {
+    //! Currently not covered
+    //! * cgroup v2 in non-standard mountpoints
+    //! * paths containing control characters or spaces, since those would be escaped in procfs
+    //!   output and we don't unescape
+    use crate::borrow::Cow;
     use crate::ffi::OsString;
     use crate::fs::{try_exists, File};
     use crate::io::Read;
+    use crate::io::{BufRead, BufReader};
     use crate::os::unix::ffi::OsStringExt;
+    use crate::path::Path;
     use crate::path::PathBuf;
+    use crate::str::from_utf8;
 
-    let mut quota = usize::MAX;
+    #[derive(PartialEq)]
-    if cfg!(miri) {
+    enum Cgroup {
-        // Attempting to open a file fails under default flags due to isolation.
+        V1,
-        // And Miri does not have parallelism anyway.
+        V2,
-        return quota;
     }
 
-    let _: Option<()> = try {
+    /// Returns cgroup CPU quota in core-equivalents, rounded down or usize::MAX if the quota cannot
-        let mut buf = Vec::with_capacity(128);
+    /// be determined or is not set.
-        // find our place in the cgroup hierarchy
+    pub(super) fn quota() -> usize {
-        File::open("/proc/self/cgroup").ok()?.read_to_end(&mut buf).ok()?;
+        let mut quota = usize::MAX;
-        let cgroup_path = buf
+        if cfg!(miri) {
-            .split(|&c| c == b'\n')
+            // Attempting to open a file fails under default flags due to isolation.
-            .filter_map(|line| {
+            // And Miri does not have parallelism anyway.
-                let mut fields = line.splitn(3, |&c| c == b':');
+            return quota;
-                // expect cgroupv2 which has an empty 2nd field
+        }
-                if fields.nth(1) != Some(b"") {
+
-                    return None;
+        let _: Option<()> = try {
-                }
+            let mut buf = Vec::with_capacity(128);
-                let path = fields.last()?;
+            // find our place in the cgroup hierarchy
-                // skip leading slash
+            File::open("/proc/self/cgroup").ok()?.read_to_end(&mut buf).ok()?;
-                Some(path[1..].to_owned())
+            let (cgroup_path, version) =
-            })
+                buf.split(|&c| c == b'\n').fold(None, |previous, line| {
-            .next()?;
+                    let mut fields = line.splitn(3, |&c| c == b':');
-        let cgroup_path = PathBuf::from(OsString::from_vec(cgroup_path));
+                    // 2nd field is a list of controllers for v1 or empty for v2
+                    let version = match fields.nth(1) {
+                        Some(b"") => Cgroup::V2,
+                        Some(controllers)
+                            if from_utf8(controllers)
+                                .is_ok_and(|c| c.split(",").any(|c| c == "cpu")) =>
+                        {
+                            Cgroup::V1
+                        }
+                        _ => return previous,
+                    };
+
+                    // already-found v1 trumps v2 since it explicitly specifies its controllers
+                    if previous.is_some() && version == Cgroup::V2 {
+                        return previous;
+                    }
+
+                    let path = fields.last()?;
+                    // skip leading slash
+                    Some((path[1..].to_owned(), version))
+                })?;
+            let cgroup_path = PathBuf::from(OsString::from_vec(cgroup_path));
+
+            quota = match version {
+                Cgroup::V1 => quota_v1(cgroup_path),
+                Cgroup::V2 => quota_v2(cgroup_path),
+            };
+        };
+
+        quota
+    }
+
+    fn quota_v2(group_path: PathBuf) -> usize {
+        let mut quota = usize::MAX;
 
         let mut path = PathBuf::with_capacity(128);
         let mut read_buf = String::with_capacity(20);
 
+        // standard mount location defined in file-hierarchy(7) manpage
         let cgroup_mount = "/sys/fs/cgroup";
 
         path.push(cgroup_mount);
-        path.push(&cgroup_path);
+        path.push(&group_path);
 
         path.push("cgroup.controllers");
 
@@ -432,30 +471,134 @@ fn cgroup2_quota() -> usize {
 
         path.pop();
 
-        while path.starts_with(cgroup_mount) {
+        let _: Option<()> = try {
-            path.push("cpu.max");
+            while path.starts_with(cgroup_mount) {
+                path.push("cpu.max");
 
-            read_buf.clear();
+                read_buf.clear();
 
-            if File::open(&path).and_then(|mut f| f.read_to_string(&mut read_buf)).is_ok() {
+                if File::open(&path).and_then(|mut f| f.read_to_string(&mut read_buf)).is_ok() {
-                let raw_quota = read_buf.lines().next()?;
+                    let raw_quota = read_buf.lines().next()?;
-                let mut raw_quota = raw_quota.split(' ');
+                    let mut raw_quota = raw_quota.split(' ');
-                let limit = raw_quota.next()?;
+                    let limit = raw_quota.next()?;
-                let period = raw_quota.next()?;
+                    let period = raw_quota.next()?;
-                match (limit.parse::<usize>(), period.parse::<usize>()) {
+                    match (limit.parse::<usize>(), period.parse::<usize>()) {
-                    (Ok(limit), Ok(period)) => {
+                        (Ok(limit), Ok(period)) => {
-                        quota = quota.min(limit / period);
+                            quota = quota.min(limit / period);
+                        }
+                        _ => {}
                     }
-                    _ => {}
                 }
+
+                path.pop(); // pop filename
+                path.pop(); // pop dir
+            }
+        };
+
+        quota
+    }
+
+    fn quota_v1(group_path: PathBuf) -> usize {
+        let mut quota = usize::MAX;
+        let mut path = PathBuf::with_capacity(128);
+        let mut read_buf = String::with_capacity(20);
+
+        // Hardcode commonly used locations mentioned in the cgroups(7) manpage
+        // if that doesn't work scan mountinfo and adjust `group_path` for bind-mounts
+        let mounts: &[fn(&Path) -> Option<(_, &Path)>] = &[
+            |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu"), p)),
+            |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu,cpuacct"), p)),
+            // this can be expensive on systems with tons of mountpoints
+            // but we only get to this point when /proc/self/cgroups explicitly indicated
+            // this process belongs to a cpu-controller cgroup v1 and the defaults didn't work
+            find_mountpoint,
+        ];
+
+        for mount in mounts {
+            let Some((mount, group_path)) = mount(&group_path) else { continue };
+
+            path.clear();
+            path.push(mount.as_ref());
+            path.push(&group_path);
+
+            // skip if we guessed the mount incorrectly
+            if matches!(try_exists(&path), Err(_) | Ok(false)) {
+                continue;
             }
 
-            path.pop(); // pop filename
+            while path.starts_with(mount.as_ref()) {
-            path.pop(); // pop dir
+                let mut parse_file = |name| {
-        }
+                    path.push(name);
-    };
+                    read_buf.clear();
 
-    quota
+                    let f = File::open(&path);
+                    path.pop(); // restore buffer before any early returns
+                    f.ok()?.read_to_string(&mut read_buf).ok()?;
+                    let parsed = read_buf.trim().parse::<usize>().ok()?;
+
+                    Some(parsed)
+                };
+
+                let limit = parse_file("cpu.cfs_quota_us");
+                let period = parse_file("cpu.cfs_period_us");
+
+                match (limit, period) {
+                    (Some(limit), Some(period)) => quota = quota.min(limit / period),
+                    _ => {}
+                }
+
+                path.pop();
+            }
+
+            // we passed the try_exists above so we should have traversed the correct hierarchy
+            // when reaching this line
+            break;
+        }
+
+        quota
+    }
+
+    /// Scan mountinfo for cgroup v1 mountpoint with a cpu controller
+    ///
+    /// If the cgroupfs is a bind mount then `group_path` is adjusted to skip
+    /// over the already-included prefix
+    fn find_mountpoint(group_path: &Path) -> Option<(Cow<'static, str>, &Path)> {
+        let mut reader = BufReader::new(File::open("/proc/self/mountinfo").ok()?);
+        let mut line = String::with_capacity(256);
+        loop {
+            line.clear();
+            if reader.read_line(&mut line).ok()? == 0 {
+                break;
+            }
+
+            let line = line.trim();
+            let mut items = line.split(' ');
+
+            let sub_path = items.nth(3)?;
+            let mount_point = items.next()?;
+            let mount_opts = items.next_back()?;
+            let filesystem_type = items.nth_back(1)?;
+
+            if filesystem_type != "cgroup" || !mount_opts.split(',').any(|opt| opt == "cpu") {
+                // not a cgroup / not a cpu-controller
+                continue;
+            }
+
+            let sub_path = Path::new(sub_path).strip_prefix("/").ok()?;
+
+            if !group_path.starts_with(sub_path) {
+                // this is a bind-mount and the bound subdirectory
+                // does not contain the cgroup this process belongs to
+                continue;
+            }
+
+            let trimmed_group_path = group_path.strip_prefix(sub_path).ok()?;
+
+            return Some((Cow::Owned(mount_point.to_owned()), trimmed_group_path));
+        }
+
+        None
+    }
 }
 
 #[cfg(all(

library/std/src/thread/mod.rs

@@ -1577,10 +1577,15 @@ fn _assert_sync_and_send() {
 ///
 /// On Linux:
 /// - It may overcount the amount of parallelism available when limited by a
-///   process-wide affinity mask or cgroup quotas and cgroup2 fs or `sched_getaffinity()` can't be
+///   process-wide affinity mask or cgroup quotas and `sched_getaffinity()` or cgroup fs can't be
 ///   queried, e.g. due to sandboxing.
 /// - It may undercount the amount of parallelism if the current thread's affinity mask
 ///   does not reflect the process' cpuset, e.g. due to pinned threads.
+/// - If the process is in a cgroup v1 cpu controller, this may need to
+///   scan mountpoints to find the corresponding cgroup v1 controller,
+///   which may take time on systems with large numbers of mountpoints.
+///   (This does not apply to cgroup v2, or to processes not in a
+///   cgroup.)
 ///
 /// On all targets:
 /// - It may overcount the amount of parallelism available when running in a VM