std/sys/pal/unix/
thread.rs

1use crate::ffi::CStr;
2use crate::mem::{self, ManuallyDrop};
3use crate::num::NonZero;
4#[cfg(all(target_os = "linux", target_env = "gnu"))]
5use crate::sys::weak::dlsym;
6#[cfg(any(target_os = "solaris", target_os = "illumos", target_os = "nto",))]
7use crate::sys::weak::weak;
8use crate::sys::{os, stack_overflow};
9use crate::time::Duration;
10use crate::{cmp, io, ptr};
11#[cfg(not(any(
12    target_os = "l4re",
13    target_os = "vxworks",
14    target_os = "espidf",
15    target_os = "nuttx"
16)))]
17pub const DEFAULT_MIN_STACK_SIZE: usize = 2 * 1024 * 1024;
18#[cfg(target_os = "l4re")]
19pub const DEFAULT_MIN_STACK_SIZE: usize = 1024 * 1024;
20#[cfg(target_os = "vxworks")]
21pub const DEFAULT_MIN_STACK_SIZE: usize = 256 * 1024;
22#[cfg(any(target_os = "espidf", target_os = "nuttx"))]
23pub const DEFAULT_MIN_STACK_SIZE: usize = 0; // 0 indicates that the stack size configured in the ESP-IDF/NuttX menuconfig system should be used
24
25struct ThreadData {
26    name: Option<Box<str>>,
27    f: Box<dyn FnOnce()>,
28}
29
30pub struct Thread {
31    id: libc::pthread_t,
32}
33
34// Some platforms may have pthread_t as a pointer in which case we still want
35// a thread to be Send/Sync
36unsafe impl Send for Thread {}
37unsafe impl Sync for Thread {}
38
39impl Thread {
40    // unsafe: see thread::Builder::spawn_unchecked for safety requirements
41    #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces
42    pub unsafe fn new(
43        stack: usize,
44        name: Option<&str>,
45        f: Box<dyn FnOnce()>,
46    ) -> io::Result<Thread> {
47        let data = Box::into_raw(Box::new(ThreadData { name: name.map(Box::from), f }));
48        let mut native: libc::pthread_t = mem::zeroed();
49        let mut attr: mem::MaybeUninit<libc::pthread_attr_t> = mem::MaybeUninit::uninit();
50        assert_eq!(libc::pthread_attr_init(attr.as_mut_ptr()), 0);
51
52        #[cfg(any(target_os = "espidf", target_os = "nuttx"))]
53        if stack > 0 {
54            // Only set the stack if a non-zero value is passed
55            // 0 is used as an indication that the default stack size configured in the ESP-IDF/NuttX menuconfig system should be used
56            assert_eq!(
57                libc::pthread_attr_setstacksize(
58                    attr.as_mut_ptr(),
59                    cmp::max(stack, min_stack_size(attr.as_ptr()))
60                ),
61                0
62            );
63        }
64
65        #[cfg(not(any(target_os = "espidf", target_os = "nuttx")))]
66        {
67            let stack_size = cmp::max(stack, min_stack_size(attr.as_ptr()));
68
69            match libc::pthread_attr_setstacksize(attr.as_mut_ptr(), stack_size) {
70                0 => {}
71                n => {
72                    assert_eq!(n, libc::EINVAL);
73                    // EINVAL means |stack_size| is either too small or not a
74                    // multiple of the system page size. Because it's definitely
75                    // >= PTHREAD_STACK_MIN, it must be an alignment issue.
76                    // Round up to the nearest page and try again.
77                    let page_size = os::page_size();
78                    let stack_size =
79                        (stack_size + page_size - 1) & (-(page_size as isize - 1) as usize - 1);
80                    assert_eq!(libc::pthread_attr_setstacksize(attr.as_mut_ptr(), stack_size), 0);
81                }
82            };
83        }
84
85        let ret = libc::pthread_create(&mut native, attr.as_ptr(), thread_start, data as *mut _);
86        // Note: if the thread creation fails and this assert fails, then p will
87        // be leaked. However, an alternative design could cause double-free
88        // which is clearly worse.
89        assert_eq!(libc::pthread_attr_destroy(attr.as_mut_ptr()), 0);
90
91        return if ret != 0 {
92            // The thread failed to start and as a result p was not consumed. Therefore, it is
93            // safe to reconstruct the box so that it gets deallocated.
94            drop(Box::from_raw(data));
95            Err(io::Error::from_raw_os_error(ret))
96        } else {
97            Ok(Thread { id: native })
98        };
99
100        extern "C" fn thread_start(data: *mut libc::c_void) -> *mut libc::c_void {
101            unsafe {
102                let data = Box::from_raw(data as *mut ThreadData);
103                // Next, set up our stack overflow handler which may get triggered if we run
104                // out of stack.
105                let _handler = stack_overflow::Handler::new(data.name);
106                // Finally, let's run some code.
107                (data.f)();
108            }
109            ptr::null_mut()
110        }
111    }
112
113    pub fn yield_now() {
114        let ret = unsafe { libc::sched_yield() };
115        debug_assert_eq!(ret, 0);
116    }
117
118    #[cfg(target_os = "android")]
119    pub fn set_name(name: &CStr) {
120        const PR_SET_NAME: libc::c_int = 15;
121        unsafe {
122            let res = libc::prctl(
123                PR_SET_NAME,
124                name.as_ptr(),
125                0 as libc::c_ulong,
126                0 as libc::c_ulong,
127                0 as libc::c_ulong,
128            );
129            // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
130            debug_assert_eq!(res, 0);
131        }
132    }
133
134    #[cfg(any(
135        target_os = "linux",
136        target_os = "freebsd",
137        target_os = "dragonfly",
138        target_os = "nuttx",
139        target_os = "cygwin"
140    ))]
141    pub fn set_name(name: &CStr) {
142        unsafe {
143            cfg_if::cfg_if! {
144                if #[cfg(any(target_os = "linux", target_os = "cygwin"))] {
145                    // Linux and Cygwin limits the allowed length of the name.
146                    const TASK_COMM_LEN: usize = 16;
147                    let name = truncate_cstr::<{ TASK_COMM_LEN }>(name);
148                } else {
149                    // FreeBSD, DragonFly BSD and NuttX do not enforce length limits.
150                }
151            };
152            // Available since glibc 2.12, musl 1.1.16, and uClibc 1.0.20 for Linux,
153            // FreeBSD 12.2 and 13.0, and DragonFly BSD 6.0.
154            let res = libc::pthread_setname_np(libc::pthread_self(), name.as_ptr());
155            // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
156            debug_assert_eq!(res, 0);
157        }
158    }
159
160    #[cfg(target_os = "openbsd")]
161    pub fn set_name(name: &CStr) {
162        unsafe {
163            libc::pthread_set_name_np(libc::pthread_self(), name.as_ptr());
164        }
165    }
166
167    #[cfg(target_vendor = "apple")]
168    pub fn set_name(name: &CStr) {
169        unsafe {
170            let name = truncate_cstr::<{ libc::MAXTHREADNAMESIZE }>(name);
171            let res = libc::pthread_setname_np(name.as_ptr());
172            // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
173            debug_assert_eq!(res, 0);
174        }
175    }
176
177    #[cfg(target_os = "netbsd")]
178    pub fn set_name(name: &CStr) {
179        unsafe {
180            let res = libc::pthread_setname_np(
181                libc::pthread_self(),
182                c"%s".as_ptr(),
183                name.as_ptr() as *mut libc::c_void,
184            );
185            debug_assert_eq!(res, 0);
186        }
187    }
188
189    #[cfg(any(target_os = "solaris", target_os = "illumos", target_os = "nto"))]
190    pub fn set_name(name: &CStr) {
191        weak!(
192            fn pthread_setname_np(
193                thread: libc::pthread_t,
194                name: *const libc::c_char,
195            ) -> libc::c_int;
196        );
197
198        if let Some(f) = pthread_setname_np.get() {
199            #[cfg(target_os = "nto")]
200            const THREAD_NAME_MAX: usize = libc::_NTO_THREAD_NAME_MAX as usize;
201            #[cfg(any(target_os = "solaris", target_os = "illumos"))]
202            const THREAD_NAME_MAX: usize = 32;
203
204            let name = truncate_cstr::<{ THREAD_NAME_MAX }>(name);
205            let res = unsafe { f(libc::pthread_self(), name.as_ptr()) };
206            debug_assert_eq!(res, 0);
207        }
208    }
209
210    #[cfg(target_os = "fuchsia")]
211    pub fn set_name(name: &CStr) {
212        use super::fuchsia::*;
213        unsafe {
214            zx_object_set_property(
215                zx_thread_self(),
216                ZX_PROP_NAME,
217                name.as_ptr() as *const libc::c_void,
218                name.to_bytes().len(),
219            );
220        }
221    }
222
223    #[cfg(target_os = "haiku")]
224    pub fn set_name(name: &CStr) {
225        unsafe {
226            let thread_self = libc::find_thread(ptr::null_mut());
227            let res = libc::rename_thread(thread_self, name.as_ptr());
228            // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
229            debug_assert_eq!(res, libc::B_OK);
230        }
231    }
232
233    #[cfg(target_os = "vxworks")]
234    pub fn set_name(name: &CStr) {
235        let mut name = truncate_cstr::<{ libc::VX_TASK_RENAME_LENGTH - 1 }>(name);
236        let res = unsafe { libc::taskNameSet(libc::taskIdSelf(), name.as_mut_ptr()) };
237        debug_assert_eq!(res, libc::OK);
238    }
239
240    #[cfg(any(
241        target_env = "newlib",
242        target_os = "l4re",
243        target_os = "emscripten",
244        target_os = "redox",
245        target_os = "hurd",
246        target_os = "aix",
247    ))]
248    pub fn set_name(_name: &CStr) {
249        // Newlib and Emscripten have no way to set a thread name.
250    }
251
252    #[cfg(not(target_os = "espidf"))]
253    pub fn sleep(dur: Duration) {
254        let mut secs = dur.as_secs();
255        let mut nsecs = dur.subsec_nanos() as _;
256
257        // If we're awoken with a signal then the return value will be -1 and
258        // nanosleep will fill in `ts` with the remaining time.
259        unsafe {
260            while secs > 0 || nsecs > 0 {
261                let mut ts = libc::timespec {
262                    tv_sec: cmp::min(libc::time_t::MAX as u64, secs) as libc::time_t,
263                    tv_nsec: nsecs,
264                };
265                secs -= ts.tv_sec as u64;
266                let ts_ptr = &raw mut ts;
267                if libc::nanosleep(ts_ptr, ts_ptr) == -1 {
268                    assert_eq!(os::errno(), libc::EINTR);
269                    secs += ts.tv_sec as u64;
270                    nsecs = ts.tv_nsec;
271                } else {
272                    nsecs = 0;
273                }
274            }
275        }
276    }
277
278    #[cfg(target_os = "espidf")]
279    pub fn sleep(dur: Duration) {
280        // ESP-IDF does not have `nanosleep`, so we use `usleep` instead.
281        // As per the documentation of `usleep`, it is expected to support
282        // sleep times as big as at least up to 1 second.
283        //
284        // ESP-IDF does support almost up to `u32::MAX`, but due to a potential integer overflow in its
285        // `usleep` implementation
286        // (https://github.com/espressif/esp-idf/blob/d7ca8b94c852052e3bc33292287ef4dd62c9eeb1/components/newlib/time.c#L210),
287        // we limit the sleep time to the maximum one that would not cause the underlying `usleep` implementation to overflow
288        // (`portTICK_PERIOD_MS` can be anything between 1 to 1000, and is 10 by default).
289        const MAX_MICROS: u32 = u32::MAX - 1_000_000 - 1;
290
291        // Add any nanoseconds smaller than a microsecond as an extra microsecond
292        // so as to comply with the `std::thread::sleep` contract which mandates
293        // implementations to sleep for _at least_ the provided `dur`.
294        // We can't overflow `micros` as it is a `u128`, while `Duration` is a pair of
295        // (`u64` secs, `u32` nanos), where the nanos are strictly smaller than 1 second
296        // (i.e. < 1_000_000_000)
297        let mut micros = dur.as_micros() + if dur.subsec_nanos() % 1_000 > 0 { 1 } else { 0 };
298
299        while micros > 0 {
300            let st = if micros > MAX_MICROS as u128 { MAX_MICROS } else { micros as u32 };
301            unsafe {
302                libc::usleep(st);
303            }
304
305            micros -= st as u128;
306        }
307    }
308
309    pub fn join(self) {
310        let id = self.into_id();
311        let ret = unsafe { libc::pthread_join(id, ptr::null_mut()) };
312        assert!(ret == 0, "failed to join thread: {}", io::Error::from_raw_os_error(ret));
313    }
314
315    pub fn id(&self) -> libc::pthread_t {
316        self.id
317    }
318
319    pub fn into_id(self) -> libc::pthread_t {
320        ManuallyDrop::new(self).id
321    }
322}
323
324impl Drop for Thread {
325    fn drop(&mut self) {
326        let ret = unsafe { libc::pthread_detach(self.id) };
327        debug_assert_eq!(ret, 0);
328    }
329}
330
331#[cfg(any(
332    target_os = "linux",
333    target_os = "nto",
334    target_os = "solaris",
335    target_os = "illumos",
336    target_os = "vxworks",
337    target_os = "cygwin",
338    target_vendor = "apple",
339))]
340fn truncate_cstr<const MAX_WITH_NUL: usize>(cstr: &CStr) -> [libc::c_char; MAX_WITH_NUL] {
341    let mut result = [0; MAX_WITH_NUL];
342    for (src, dst) in cstr.to_bytes().iter().zip(&mut result[..MAX_WITH_NUL - 1]) {
343        *dst = *src as libc::c_char;
344    }
345    result
346}
347
348pub fn available_parallelism() -> io::Result<NonZero<usize>> {
349    cfg_if::cfg_if! {
350        if #[cfg(any(
351            target_os = "android",
352            target_os = "emscripten",
353            target_os = "fuchsia",
354            target_os = "hurd",
355            target_os = "linux",
356            target_os = "aix",
357            target_vendor = "apple",
358            target_os = "cygwin",
359        ))] {
360            #[allow(unused_assignments)]
361            #[allow(unused_mut)]
362            let mut quota = usize::MAX;
363
364            #[cfg(any(target_os = "android", target_os = "linux"))]
365            {
366                quota = cgroups::quota().max(1);
367                let mut set: libc::cpu_set_t = unsafe { mem::zeroed() };
368                unsafe {
369                    if libc::sched_getaffinity(0, size_of::<libc::cpu_set_t>(), &mut set) == 0 {
370                        let count = libc::CPU_COUNT(&set) as usize;
371                        let count = count.min(quota);
372
373                        // According to sched_getaffinity's API it should always be non-zero, but
374                        // some old MIPS kernels were buggy and zero-initialized the mask if
375                        // none was explicitly set.
376                        // In that case we use the sysconf fallback.
377                        if let Some(count) = NonZero::new(count) {
378                            return Ok(count)
379                        }
380                    }
381                }
382            }
383            match unsafe { libc::sysconf(libc::_SC_NPROCESSORS_ONLN) } {
384                -1 => Err(io::Error::last_os_error()),
385                0 => Err(io::Error::UNKNOWN_THREAD_COUNT),
386                cpus => {
387                    let count = cpus as usize;
388                    // Cover the unusual situation where we were able to get the quota but not the affinity mask
389                    let count = count.min(quota);
390                    Ok(unsafe { NonZero::new_unchecked(count) })
391                }
392            }
393        } else if #[cfg(any(
394                   target_os = "freebsd",
395                   target_os = "dragonfly",
396                   target_os = "openbsd",
397                   target_os = "netbsd",
398               ))] {
399            use crate::ptr;
400
401            #[cfg(target_os = "freebsd")]
402            {
403                let mut set: libc::cpuset_t = unsafe { mem::zeroed() };
404                unsafe {
405                    if libc::cpuset_getaffinity(
406                        libc::CPU_LEVEL_WHICH,
407                        libc::CPU_WHICH_PID,
408                        -1,
409                        size_of::<libc::cpuset_t>(),
410                        &mut set,
411                    ) == 0 {
412                        let count = libc::CPU_COUNT(&set) as usize;
413                        if count > 0 {
414                            return Ok(NonZero::new_unchecked(count));
415                        }
416                    }
417                }
418            }
419
420            #[cfg(target_os = "netbsd")]
421            {
422                unsafe {
423                    let set = libc::_cpuset_create();
424                    if !set.is_null() {
425                        let mut count: usize = 0;
426                        if libc::pthread_getaffinity_np(libc::pthread_self(), libc::_cpuset_size(set), set) == 0 {
427                            for i in 0..libc::cpuid_t::MAX {
428                                match libc::_cpuset_isset(i, set) {
429                                    -1 => break,
430                                    0 => continue,
431                                    _ => count = count + 1,
432                                }
433                            }
434                        }
435                        libc::_cpuset_destroy(set);
436                        if let Some(count) = NonZero::new(count) {
437                            return Ok(count);
438                        }
439                    }
440                }
441            }
442
443            let mut cpus: libc::c_uint = 0;
444            let mut cpus_size = size_of_val(&cpus);
445
446            unsafe {
447                cpus = libc::sysconf(libc::_SC_NPROCESSORS_ONLN) as libc::c_uint;
448            }
449
450            // Fallback approach in case of errors or no hardware threads.
451            if cpus < 1 {
452                let mut mib = [libc::CTL_HW, libc::HW_NCPU, 0, 0];
453                let res = unsafe {
454                    libc::sysctl(
455                        mib.as_mut_ptr(),
456                        2,
457                        (&raw mut cpus) as *mut _,
458                        (&raw mut cpus_size) as *mut _,
459                        ptr::null_mut(),
460                        0,
461                    )
462                };
463
464                // Handle errors if any.
465                if res == -1 {
466                    return Err(io::Error::last_os_error());
467                } else if cpus == 0 {
468                    return Err(io::Error::UNKNOWN_THREAD_COUNT);
469                }
470            }
471
472            Ok(unsafe { NonZero::new_unchecked(cpus as usize) })
473        } else if #[cfg(target_os = "nto")] {
474            unsafe {
475                use libc::_syspage_ptr;
476                if _syspage_ptr.is_null() {
477                    Err(io::const_error!(io::ErrorKind::NotFound, "no syspage available"))
478                } else {
479                    let cpus = (*_syspage_ptr).num_cpu;
480                    NonZero::new(cpus as usize)
481                        .ok_or(io::Error::UNKNOWN_THREAD_COUNT)
482                }
483            }
484        } else if #[cfg(any(target_os = "solaris", target_os = "illumos"))] {
485            let mut cpus = 0u32;
486            if unsafe { libc::pset_info(libc::PS_MYID, core::ptr::null_mut(), &mut cpus, core::ptr::null_mut()) } != 0 {
487                return Err(io::Error::UNKNOWN_THREAD_COUNT);
488            }
489            Ok(unsafe { NonZero::new_unchecked(cpus as usize) })
490        } else if #[cfg(target_os = "haiku")] {
491            // system_info cpu_count field gets the static data set at boot time with `smp_set_num_cpus`
492            // `get_system_info` calls then `smp_get_num_cpus`
493            unsafe {
494                let mut sinfo: libc::system_info = crate::mem::zeroed();
495                let res = libc::get_system_info(&mut sinfo);
496
497                if res != libc::B_OK {
498                    return Err(io::Error::UNKNOWN_THREAD_COUNT);
499                }
500
501                Ok(NonZero::new_unchecked(sinfo.cpu_count as usize))
502            }
503        } else if #[cfg(target_os = "vxworks")] {
504            // Note: there is also `vxCpuConfiguredGet`, closer to _SC_NPROCESSORS_CONF
505            // expectations than the actual cores availability.
506            unsafe extern "C" {
507                fn vxCpuEnabledGet() -> libc::cpuset_t;
508            }
509
510            // SAFETY: `vxCpuEnabledGet` always fetches a mask with at least one bit set
511            unsafe{
512                let set = vxCpuEnabledGet();
513                Ok(NonZero::new_unchecked(set.count_ones() as usize))
514            }
515        } else {
516            // FIXME: implement on Redox, l4re
517            Err(io::const_error!(io::ErrorKind::Unsupported, "getting the number of hardware threads is not supported on the target platform"))
518        }
519    }
520}
521
522#[cfg(any(target_os = "android", target_os = "linux"))]
523mod cgroups {
524    //! Currently not covered
525    //! * cgroup v2 in non-standard mountpoints
526    //! * paths containing control characters or spaces, since those would be escaped in procfs
527    //!   output and we don't unescape
528
529    use crate::borrow::Cow;
530    use crate::ffi::OsString;
531    use crate::fs::{File, exists};
532    use crate::io::{BufRead, Read};
533    use crate::os::unix::ffi::OsStringExt;
534    use crate::path::{Path, PathBuf};
535    use crate::str::from_utf8;
536
537    #[derive(PartialEq)]
538    enum Cgroup {
539        V1,
540        V2,
541    }
542
543    /// Returns cgroup CPU quota in core-equivalents, rounded down or usize::MAX if the quota cannot
544    /// be determined or is not set.
545    pub(super) fn quota() -> usize {
546        let mut quota = usize::MAX;
547        if cfg!(miri) {
548            // Attempting to open a file fails under default flags due to isolation.
549            // And Miri does not have parallelism anyway.
550            return quota;
551        }
552
553        let _: Option<()> = try {
554            let mut buf = Vec::with_capacity(128);
555            // find our place in the cgroup hierarchy
556            File::open("/proc/self/cgroup").ok()?.read_to_end(&mut buf).ok()?;
557            let (cgroup_path, version) =
558                buf.split(|&c| c == b'\n').fold(None, |previous, line| {
559                    let mut fields = line.splitn(3, |&c| c == b':');
560                    // 2nd field is a list of controllers for v1 or empty for v2
561                    let version = match fields.nth(1) {
562                        Some(b"") => Cgroup::V2,
563                        Some(controllers)
564                            if from_utf8(controllers)
565                                .is_ok_and(|c| c.split(',').any(|c| c == "cpu")) =>
566                        {
567                            Cgroup::V1
568                        }
569                        _ => return previous,
570                    };
571
572                    // already-found v1 trumps v2 since it explicitly specifies its controllers
573                    if previous.is_some() && version == Cgroup::V2 {
574                        return previous;
575                    }
576
577                    let path = fields.last()?;
578                    // skip leading slash
579                    Some((path[1..].to_owned(), version))
580                })?;
581            let cgroup_path = PathBuf::from(OsString::from_vec(cgroup_path));
582
583            quota = match version {
584                Cgroup::V1 => quota_v1(cgroup_path),
585                Cgroup::V2 => quota_v2(cgroup_path),
586            };
587        };
588
589        quota
590    }
591
592    fn quota_v2(group_path: PathBuf) -> usize {
593        let mut quota = usize::MAX;
594
595        let mut path = PathBuf::with_capacity(128);
596        let mut read_buf = String::with_capacity(20);
597
598        // standard mount location defined in file-hierarchy(7) manpage
599        let cgroup_mount = "/sys/fs/cgroup";
600
601        path.push(cgroup_mount);
602        path.push(&group_path);
603
604        path.push("cgroup.controllers");
605
606        // skip if we're not looking at cgroup2
607        if matches!(exists(&path), Err(_) | Ok(false)) {
608            return usize::MAX;
609        };
610
611        path.pop();
612
613        let _: Option<()> = try {
614            while path.starts_with(cgroup_mount) {
615                path.push("cpu.max");
616
617                read_buf.clear();
618
619                if File::open(&path).and_then(|mut f| f.read_to_string(&mut read_buf)).is_ok() {
620                    let raw_quota = read_buf.lines().next()?;
621                    let mut raw_quota = raw_quota.split(' ');
622                    let limit = raw_quota.next()?;
623                    let period = raw_quota.next()?;
624                    match (limit.parse::<usize>(), period.parse::<usize>()) {
625                        (Ok(limit), Ok(period)) if period > 0 => {
626                            quota = quota.min(limit / period);
627                        }
628                        _ => {}
629                    }
630                }
631
632                path.pop(); // pop filename
633                path.pop(); // pop dir
634            }
635        };
636
637        quota
638    }
639
640    fn quota_v1(group_path: PathBuf) -> usize {
641        let mut quota = usize::MAX;
642        let mut path = PathBuf::with_capacity(128);
643        let mut read_buf = String::with_capacity(20);
644
645        // Hardcode commonly used locations mentioned in the cgroups(7) manpage
646        // if that doesn't work scan mountinfo and adjust `group_path` for bind-mounts
647        let mounts: &[fn(&Path) -> Option<(_, &Path)>] = &[
648            |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu"), p)),
649            |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu,cpuacct"), p)),
650            // this can be expensive on systems with tons of mountpoints
651            // but we only get to this point when /proc/self/cgroups explicitly indicated
652            // this process belongs to a cpu-controller cgroup v1 and the defaults didn't work
653            find_mountpoint,
654        ];
655
656        for mount in mounts {
657            let Some((mount, group_path)) = mount(&group_path) else { continue };
658
659            path.clear();
660            path.push(mount.as_ref());
661            path.push(&group_path);
662
663            // skip if we guessed the mount incorrectly
664            if matches!(exists(&path), Err(_) | Ok(false)) {
665                continue;
666            }
667
668            while path.starts_with(mount.as_ref()) {
669                let mut parse_file = |name| {
670                    path.push(name);
671                    read_buf.clear();
672
673                    let f = File::open(&path);
674                    path.pop(); // restore buffer before any early returns
675                    f.ok()?.read_to_string(&mut read_buf).ok()?;
676                    let parsed = read_buf.trim().parse::<usize>().ok()?;
677
678                    Some(parsed)
679                };
680
681                let limit = parse_file("cpu.cfs_quota_us");
682                let period = parse_file("cpu.cfs_period_us");
683
684                match (limit, period) {
685                    (Some(limit), Some(period)) if period > 0 => quota = quota.min(limit / period),
686                    _ => {}
687                }
688
689                path.pop();
690            }
691
692            // we passed the try_exists above so we should have traversed the correct hierarchy
693            // when reaching this line
694            break;
695        }
696
697        quota
698    }
699
700    /// Scan mountinfo for cgroup v1 mountpoint with a cpu controller
701    ///
702    /// If the cgroupfs is a bind mount then `group_path` is adjusted to skip
703    /// over the already-included prefix
704    fn find_mountpoint(group_path: &Path) -> Option<(Cow<'static, str>, &Path)> {
705        let mut reader = File::open_buffered("/proc/self/mountinfo").ok()?;
706        let mut line = String::with_capacity(256);
707        loop {
708            line.clear();
709            if reader.read_line(&mut line).ok()? == 0 {
710                break;
711            }
712
713            let line = line.trim();
714            let mut items = line.split(' ');
715
716            let sub_path = items.nth(3)?;
717            let mount_point = items.next()?;
718            let mount_opts = items.next_back()?;
719            let filesystem_type = items.nth_back(1)?;
720
721            if filesystem_type != "cgroup" || !mount_opts.split(',').any(|opt| opt == "cpu") {
722                // not a cgroup / not a cpu-controller
723                continue;
724            }
725
726            let sub_path = Path::new(sub_path).strip_prefix("/").ok()?;
727
728            if !group_path.starts_with(sub_path) {
729                // this is a bind-mount and the bound subdirectory
730                // does not contain the cgroup this process belongs to
731                continue;
732            }
733
734            let trimmed_group_path = group_path.strip_prefix(sub_path).ok()?;
735
736            return Some((Cow::Owned(mount_point.to_owned()), trimmed_group_path));
737        }
738
739        None
740    }
741}
742
743// glibc >= 2.15 has a __pthread_get_minstack() function that returns
744// PTHREAD_STACK_MIN plus bytes needed for thread-local storage.
745// We need that information to avoid blowing up when a small stack
746// is created in an application with big thread-local storage requirements.
747// See #6233 for rationale and details.
748#[cfg(all(target_os = "linux", target_env = "gnu"))]
749unsafe fn min_stack_size(attr: *const libc::pthread_attr_t) -> usize {
750    // We use dlsym to avoid an ELF version dependency on GLIBC_PRIVATE. (#23628)
751    // We shouldn't really be using such an internal symbol, but there's currently
752    // no other way to account for the TLS size.
753    dlsym!(
754        fn __pthread_get_minstack(attr: *const libc::pthread_attr_t) -> libc::size_t;
755    );
756
757    match __pthread_get_minstack.get() {
758        None => libc::PTHREAD_STACK_MIN,
759        Some(f) => unsafe { f(attr) },
760    }
761}
762
763// No point in looking up __pthread_get_minstack() on non-glibc platforms.
764#[cfg(all(
765    not(all(target_os = "linux", target_env = "gnu")),
766    not(any(target_os = "netbsd", target_os = "nuttx"))
767))]
768unsafe fn min_stack_size(_: *const libc::pthread_attr_t) -> usize {
769    libc::PTHREAD_STACK_MIN
770}
771
772#[cfg(any(target_os = "netbsd", target_os = "nuttx"))]
773unsafe fn min_stack_size(_: *const libc::pthread_attr_t) -> usize {
774    static STACK: crate::sync::OnceLock<usize> = crate::sync::OnceLock::new();
775
776    *STACK.get_or_init(|| {
777        let mut stack = unsafe { libc::sysconf(libc::_SC_THREAD_STACK_MIN) };
778        if stack < 0 {
779            stack = 2048; // just a guess
780        }
781
782        stack as usize
783    })
784}