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(target_os = "l4re", target_os = "vxworks", target_os = "espidf")))]
12pub const DEFAULT_MIN_STACK_SIZE: usize = 2 * 1024 * 1024;
13#[cfg(target_os = "l4re")]
14pub const DEFAULT_MIN_STACK_SIZE: usize = 1024 * 1024;
15#[cfg(target_os = "vxworks")]
16pub const DEFAULT_MIN_STACK_SIZE: usize = 256 * 1024;
17#[cfg(target_os = "espidf")]
18pub const DEFAULT_MIN_STACK_SIZE: usize = 0; // 0 indicates that the stack size configured in the ESP-IDF menuconfig system should be used
19
20#[cfg(target_os = "fuchsia")]
21mod zircon {
22    type zx_handle_t = u32;
23    type zx_status_t = i32;
24    pub const ZX_PROP_NAME: u32 = 3;
25
26    unsafe extern "C" {
27        pub fn zx_object_set_property(
28            handle: zx_handle_t,
29            property: u32,
30            value: *const libc::c_void,
31            value_size: libc::size_t,
32        ) -> zx_status_t;
33        pub fn zx_thread_self() -> zx_handle_t;
34    }
35}
36
37pub struct Thread {
38    id: libc::pthread_t,
39}
40
41// Some platforms may have pthread_t as a pointer in which case we still want
42// a thread to be Send/Sync
43unsafe impl Send for Thread {}
44unsafe impl Sync for Thread {}
45
46impl Thread {
47    // unsafe: see thread::Builder::spawn_unchecked for safety requirements
48    #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces
49    pub unsafe fn new(stack: usize, p: Box<dyn FnOnce()>) -> io::Result<Thread> {
50        let p = Box::into_raw(Box::new(p));
51        let mut native: libc::pthread_t = mem::zeroed();
52        let mut attr: mem::MaybeUninit<libc::pthread_attr_t> = mem::MaybeUninit::uninit();
53        assert_eq!(libc::pthread_attr_init(attr.as_mut_ptr()), 0);
54
55        #[cfg(target_os = "espidf")]
56        if stack > 0 {
57            // Only set the stack if a non-zero value is passed
58            // 0 is used as an indication that the default stack size configured in the ESP-IDF menuconfig system should be used
59            assert_eq!(
60                libc::pthread_attr_setstacksize(
61                    attr.as_mut_ptr(),
62                    cmp::max(stack, min_stack_size(attr.as_ptr()))
63                ),
64                0
65            );
66        }
67
68        #[cfg(not(target_os = "espidf"))]
69        {
70            let stack_size = cmp::max(stack, min_stack_size(attr.as_ptr()));
71
72            match libc::pthread_attr_setstacksize(attr.as_mut_ptr(), stack_size) {
73                0 => {}
74                n => {
75                    assert_eq!(n, libc::EINVAL);
76                    // EINVAL means |stack_size| is either too small or not a
77                    // multiple of the system page size. Because it's definitely
78                    // >= PTHREAD_STACK_MIN, it must be an alignment issue.
79                    // Round up to the nearest page and try again.
80                    let page_size = os::page_size();
81                    let stack_size =
82                        (stack_size + page_size - 1) & (-(page_size as isize - 1) as usize - 1);
83                    assert_eq!(libc::pthread_attr_setstacksize(attr.as_mut_ptr(), stack_size), 0);
84                }
85            };
86        }
87
88        let ret = libc::pthread_create(&mut native, attr.as_ptr(), thread_start, p as *mut _);
89        // Note: if the thread creation fails and this assert fails, then p will
90        // be leaked. However, an alternative design could cause double-free
91        // which is clearly worse.
92        assert_eq!(libc::pthread_attr_destroy(attr.as_mut_ptr()), 0);
93
94        return if ret != 0 {
95            // The thread failed to start and as a result p was not consumed. Therefore, it is
96            // safe to reconstruct the box so that it gets deallocated.
97            drop(Box::from_raw(p));
98            Err(io::Error::from_raw_os_error(ret))
99        } else {
100            Ok(Thread { id: native })
101        };
102
103        extern "C" fn thread_start(main: *mut libc::c_void) -> *mut libc::c_void {
104            unsafe {
105                // Next, set up our stack overflow handler which may get triggered if we run
106                // out of stack.
107                let _handler = stack_overflow::Handler::new();
108                // Finally, let's run some code.
109                Box::from_raw(main as *mut Box<dyn FnOnce()>)();
110            }
111            ptr::null_mut()
112        }
113    }
114
115    pub fn yield_now() {
116        let ret = unsafe { libc::sched_yield() };
117        debug_assert_eq!(ret, 0);
118    }
119
120    #[cfg(target_os = "android")]
121    pub fn set_name(name: &CStr) {
122        const PR_SET_NAME: libc::c_int = 15;
123        unsafe {
124            let res = libc::prctl(
125                PR_SET_NAME,
126                name.as_ptr(),
127                0 as libc::c_ulong,
128                0 as libc::c_ulong,
129                0 as libc::c_ulong,
130            );
131            // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
132            debug_assert_eq!(res, 0);
133        }
134    }
135
136    #[cfg(any(
137        target_os = "linux",
138        target_os = "freebsd",
139        target_os = "dragonfly",
140        target_os = "nuttx"
141    ))]
142    pub fn set_name(name: &CStr) {
143        unsafe {
144            cfg_if::cfg_if! {
145                if #[cfg(target_os = "linux")] {
146                    // Linux limits the allowed length of the name.
147                    const TASK_COMM_LEN: usize = 16;
148                    let name = truncate_cstr::<{ TASK_COMM_LEN }>(name);
149                } else {
150                    // FreeBSD, DragonFly BSD and NuttX do not enforce length limits.
151                }
152            };
153            // Available since glibc 2.12, musl 1.1.16, and uClibc 1.0.20 for Linux,
154            // FreeBSD 12.2 and 13.0, and DragonFly BSD 6.0.
155            let res = libc::pthread_setname_np(libc::pthread_self(), name.as_ptr());
156            // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
157            debug_assert_eq!(res, 0);
158        }
159    }
160
161    #[cfg(target_os = "openbsd")]
162    pub fn set_name(name: &CStr) {
163        unsafe {
164            libc::pthread_set_name_np(libc::pthread_self(), name.as_ptr());
165        }
166    }
167
168    #[cfg(target_vendor = "apple")]
169    pub fn set_name(name: &CStr) {
170        unsafe {
171            let name = truncate_cstr::<{ libc::MAXTHREADNAMESIZE }>(name);
172            let res = libc::pthread_setname_np(name.as_ptr());
173            // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
174            debug_assert_eq!(res, 0);
175        }
176    }
177
178    #[cfg(target_os = "netbsd")]
179    pub fn set_name(name: &CStr) {
180        unsafe {
181            let res = libc::pthread_setname_np(
182                libc::pthread_self(),
183                c"%s".as_ptr(),
184                name.as_ptr() as *mut libc::c_void,
185            );
186            debug_assert_eq!(res, 0);
187        }
188    }
189
190    #[cfg(any(target_os = "solaris", target_os = "illumos", target_os = "nto"))]
191    // FIXME(#115199): Rust currently omits weak function definitions
192    // and its metadata from LLVM IR.
193    #[no_sanitize(cfi)]
194    pub fn set_name(name: &CStr) {
195        weak! {
196            fn pthread_setname_np(
197                libc::pthread_t, *const libc::c_char
198            ) -> libc::c_int
199        }
200
201        if let Some(f) = pthread_setname_np.get() {
202            #[cfg(target_os = "nto")]
203            const THREAD_NAME_MAX: usize = libc::_NTO_THREAD_NAME_MAX as usize;
204            #[cfg(any(target_os = "solaris", target_os = "illumos"))]
205            const THREAD_NAME_MAX: usize = 32;
206
207            let name = truncate_cstr::<{ THREAD_NAME_MAX }>(name);
208            let res = unsafe { f(libc::pthread_self(), name.as_ptr()) };
209            debug_assert_eq!(res, 0);
210        }
211    }
212
213    #[cfg(target_os = "fuchsia")]
214    pub fn set_name(name: &CStr) {
215        use self::zircon::*;
216        unsafe {
217            zx_object_set_property(
218                zx_thread_self(),
219                ZX_PROP_NAME,
220                name.as_ptr() as *const libc::c_void,
221                name.to_bytes().len(),
222            );
223        }
224    }
225
226    #[cfg(target_os = "haiku")]
227    pub fn set_name(name: &CStr) {
228        unsafe {
229            let thread_self = libc::find_thread(ptr::null_mut());
230            let res = libc::rename_thread(thread_self, name.as_ptr());
231            // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
232            debug_assert_eq!(res, libc::B_OK);
233        }
234    }
235
236    #[cfg(target_os = "vxworks")]
237    pub fn set_name(name: &CStr) {
238        // FIXME(libc): adding real STATUS, ERROR type eventually.
239        unsafe extern "C" {
240            fn taskNameSet(task_id: libc::TASK_ID, task_name: *mut libc::c_char) -> libc::c_int;
241        }
242
243        //  VX_TASK_NAME_LEN is 31 in VxWorks 7.
244        const VX_TASK_NAME_LEN: usize = 31;
245
246        let mut name = truncate_cstr::<{ VX_TASK_NAME_LEN }>(name);
247        let res = unsafe { taskNameSet(libc::taskIdSelf(), name.as_mut_ptr()) };
248        debug_assert_eq!(res, libc::OK);
249    }
250
251    #[cfg(any(
252        target_env = "newlib",
253        target_os = "l4re",
254        target_os = "emscripten",
255        target_os = "redox",
256        target_os = "hurd",
257        target_os = "aix",
258    ))]
259    pub fn set_name(_name: &CStr) {
260        // Newlib and Emscripten have no way to set a thread name.
261    }
262
263    #[cfg(not(target_os = "espidf"))]
264    pub fn sleep(dur: Duration) {
265        let mut secs = dur.as_secs();
266        let mut nsecs = dur.subsec_nanos() as _;
267
268        // If we're awoken with a signal then the return value will be -1 and
269        // nanosleep will fill in `ts` with the remaining time.
270        unsafe {
271            while secs > 0 || nsecs > 0 {
272                let mut ts = libc::timespec {
273                    tv_sec: cmp::min(libc::time_t::MAX as u64, secs) as libc::time_t,
274                    tv_nsec: nsecs,
275                };
276                secs -= ts.tv_sec as u64;
277                let ts_ptr = &raw mut ts;
278                if libc::nanosleep(ts_ptr, ts_ptr) == -1 {
279                    assert_eq!(os::errno(), libc::EINTR);
280                    secs += ts.tv_sec as u64;
281                    nsecs = ts.tv_nsec;
282                } else {
283                    nsecs = 0;
284                }
285            }
286        }
287    }
288
289    #[cfg(target_os = "espidf")]
290    pub fn sleep(dur: Duration) {
291        // ESP-IDF does not have `nanosleep`, so we use `usleep` instead.
292        // As per the documentation of `usleep`, it is expected to support
293        // sleep times as big as at least up to 1 second.
294        //
295        // ESP-IDF does support almost up to `u32::MAX`, but due to a potential integer overflow in its
296        // `usleep` implementation
297        // (https://github.com/espressif/esp-idf/blob/d7ca8b94c852052e3bc33292287ef4dd62c9eeb1/components/newlib/time.c#L210),
298        // we limit the sleep time to the maximum one that would not cause the underlying `usleep` implementation to overflow
299        // (`portTICK_PERIOD_MS` can be anything between 1 to 1000, and is 10 by default).
300        const MAX_MICROS: u32 = u32::MAX - 1_000_000 - 1;
301
302        // Add any nanoseconds smaller than a microsecond as an extra microsecond
303        // so as to comply with the `std::thread::sleep` contract which mandates
304        // implementations to sleep for _at least_ the provided `dur`.
305        // We can't overflow `micros` as it is a `u128`, while `Duration` is a pair of
306        // (`u64` secs, `u32` nanos), where the nanos are strictly smaller than 1 second
307        // (i.e. < 1_000_000_000)
308        let mut micros = dur.as_micros() + if dur.subsec_nanos() % 1_000 > 0 { 1 } else { 0 };
309
310        while micros > 0 {
311            let st = if micros > MAX_MICROS as u128 { MAX_MICROS } else { micros as u32 };
312            unsafe {
313                libc::usleep(st);
314            }
315
316            micros -= st as u128;
317        }
318    }
319
320    pub fn join(self) {
321        let id = self.into_id();
322        let ret = unsafe { libc::pthread_join(id, ptr::null_mut()) };
323        assert!(ret == 0, "failed to join thread: {}", io::Error::from_raw_os_error(ret));
324    }
325
326    pub fn id(&self) -> libc::pthread_t {
327        self.id
328    }
329
330    pub fn into_id(self) -> libc::pthread_t {
331        ManuallyDrop::new(self).id
332    }
333}
334
335impl Drop for Thread {
336    fn drop(&mut self) {
337        let ret = unsafe { libc::pthread_detach(self.id) };
338        debug_assert_eq!(ret, 0);
339    }
340}
341
342#[cfg(any(
343    target_os = "linux",
344    target_os = "nto",
345    target_os = "solaris",
346    target_os = "illumos",
347    target_os = "vxworks",
348    target_vendor = "apple",
349))]
350fn truncate_cstr<const MAX_WITH_NUL: usize>(cstr: &CStr) -> [libc::c_char; MAX_WITH_NUL] {
351    let mut result = [0; MAX_WITH_NUL];
352    for (src, dst) in cstr.to_bytes().iter().zip(&mut result[..MAX_WITH_NUL - 1]) {
353        *dst = *src as libc::c_char;
354    }
355    result
356}
357
358pub fn available_parallelism() -> io::Result<NonZero<usize>> {
359    cfg_if::cfg_if! {
360        if #[cfg(any(
361            target_os = "android",
362            target_os = "emscripten",
363            target_os = "fuchsia",
364            target_os = "hurd",
365            target_os = "linux",
366            target_os = "aix",
367            target_vendor = "apple",
368        ))] {
369            #[allow(unused_assignments)]
370            #[allow(unused_mut)]
371            let mut quota = usize::MAX;
372
373            #[cfg(any(target_os = "android", target_os = "linux"))]
374            {
375                quota = cgroups::quota().max(1);
376                let mut set: libc::cpu_set_t = unsafe { mem::zeroed() };
377                unsafe {
378                    if libc::sched_getaffinity(0, size_of::<libc::cpu_set_t>(), &mut set) == 0 {
379                        let count = libc::CPU_COUNT(&set) as usize;
380                        let count = count.min(quota);
381
382                        // According to sched_getaffinity's API it should always be non-zero, but
383                        // some old MIPS kernels were buggy and zero-initialized the mask if
384                        // none was explicitly set.
385                        // In that case we use the sysconf fallback.
386                        if let Some(count) = NonZero::new(count) {
387                            return Ok(count)
388                        }
389                    }
390                }
391            }
392            match unsafe { libc::sysconf(libc::_SC_NPROCESSORS_ONLN) } {
393                -1 => Err(io::Error::last_os_error()),
394                0 => Err(io::Error::UNKNOWN_THREAD_COUNT),
395                cpus => {
396                    let count = cpus as usize;
397                    // Cover the unusual situation where we were able to get the quota but not the affinity mask
398                    let count = count.min(quota);
399                    Ok(unsafe { NonZero::new_unchecked(count) })
400                }
401            }
402        } else if #[cfg(any(
403                   target_os = "freebsd",
404                   target_os = "dragonfly",
405                   target_os = "openbsd",
406                   target_os = "netbsd",
407               ))] {
408            use crate::ptr;
409
410            #[cfg(target_os = "freebsd")]
411            {
412                let mut set: libc::cpuset_t = unsafe { mem::zeroed() };
413                unsafe {
414                    if libc::cpuset_getaffinity(
415                        libc::CPU_LEVEL_WHICH,
416                        libc::CPU_WHICH_PID,
417                        -1,
418                        size_of::<libc::cpuset_t>(),
419                        &mut set,
420                    ) == 0 {
421                        let count = libc::CPU_COUNT(&set) as usize;
422                        if count > 0 {
423                            return Ok(NonZero::new_unchecked(count));
424                        }
425                    }
426                }
427            }
428
429            #[cfg(target_os = "netbsd")]
430            {
431                unsafe {
432                    let set = libc::_cpuset_create();
433                    if !set.is_null() {
434                        let mut count: usize = 0;
435                        if libc::pthread_getaffinity_np(libc::pthread_self(), libc::_cpuset_size(set), set) == 0 {
436                            for i in 0..libc::cpuid_t::MAX {
437                                match libc::_cpuset_isset(i, set) {
438                                    -1 => break,
439                                    0 => continue,
440                                    _ => count = count + 1,
441                                }
442                            }
443                        }
444                        libc::_cpuset_destroy(set);
445                        if let Some(count) = NonZero::new(count) {
446                            return Ok(count);
447                        }
448                    }
449                }
450            }
451
452            let mut cpus: libc::c_uint = 0;
453            let mut cpus_size = size_of_val(&cpus);
454
455            unsafe {
456                cpus = libc::sysconf(libc::_SC_NPROCESSORS_ONLN) as libc::c_uint;
457            }
458
459            // Fallback approach in case of errors or no hardware threads.
460            if cpus < 1 {
461                let mut mib = [libc::CTL_HW, libc::HW_NCPU, 0, 0];
462                let res = unsafe {
463                    libc::sysctl(
464                        mib.as_mut_ptr(),
465                        2,
466                        (&raw mut cpus) as *mut _,
467                        (&raw mut cpus_size) as *mut _,
468                        ptr::null_mut(),
469                        0,
470                    )
471                };
472
473                // Handle errors if any.
474                if res == -1 {
475                    return Err(io::Error::last_os_error());
476                } else if cpus == 0 {
477                    return Err(io::Error::UNKNOWN_THREAD_COUNT);
478                }
479            }
480
481            Ok(unsafe { NonZero::new_unchecked(cpus as usize) })
482        } else if #[cfg(target_os = "nto")] {
483            unsafe {
484                use libc::_syspage_ptr;
485                if _syspage_ptr.is_null() {
486                    Err(io::const_error!(io::ErrorKind::NotFound, "no syspage available"))
487                } else {
488                    let cpus = (*_syspage_ptr).num_cpu;
489                    NonZero::new(cpus as usize)
490                        .ok_or(io::Error::UNKNOWN_THREAD_COUNT)
491                }
492            }
493        } else if #[cfg(any(target_os = "solaris", target_os = "illumos"))] {
494            let mut cpus = 0u32;
495            if unsafe { libc::pset_info(libc::PS_MYID, core::ptr::null_mut(), &mut cpus, core::ptr::null_mut()) } != 0 {
496                return Err(io::Error::UNKNOWN_THREAD_COUNT);
497            }
498            Ok(unsafe { NonZero::new_unchecked(cpus as usize) })
499        } else if #[cfg(target_os = "haiku")] {
500            // system_info cpu_count field gets the static data set at boot time with `smp_set_num_cpus`
501            // `get_system_info` calls then `smp_get_num_cpus`
502            unsafe {
503                let mut sinfo: libc::system_info = crate::mem::zeroed();
504                let res = libc::get_system_info(&mut sinfo);
505
506                if res != libc::B_OK {
507                    return Err(io::Error::UNKNOWN_THREAD_COUNT);
508                }
509
510                Ok(NonZero::new_unchecked(sinfo.cpu_count as usize))
511            }
512        } else if #[cfg(target_os = "vxworks")] {
513            // Note: there is also `vxCpuConfiguredGet`, closer to _SC_NPROCESSORS_CONF
514            // expectations than the actual cores availability.
515            unsafe extern "C" {
516                fn vxCpuEnabledGet() -> libc::cpuset_t;
517            }
518
519            // SAFETY: `vxCpuEnabledGet` always fetches a mask with at least one bit set
520            unsafe{
521                let set = vxCpuEnabledGet();
522                Ok(NonZero::new_unchecked(set.count_ones() as usize))
523            }
524        } else {
525            // FIXME: implement on Redox, l4re
526            Err(io::const_error!(io::ErrorKind::Unsupported, "getting the number of hardware threads is not supported on the target platform"))
527        }
528    }
529}
530
531#[cfg(any(target_os = "android", target_os = "linux"))]
532mod cgroups {
533    //! Currently not covered
534    //! * cgroup v2 in non-standard mountpoints
535    //! * paths containing control characters or spaces, since those would be escaped in procfs
536    //!   output and we don't unescape
537
538    use crate::borrow::Cow;
539    use crate::ffi::OsString;
540    use crate::fs::{File, exists};
541    use crate::io::{BufRead, Read};
542    use crate::os::unix::ffi::OsStringExt;
543    use crate::path::{Path, PathBuf};
544    use crate::str::from_utf8;
545
546    #[derive(PartialEq)]
547    enum Cgroup {
548        V1,
549        V2,
550    }
551
552    /// Returns cgroup CPU quota in core-equivalents, rounded down or usize::MAX if the quota cannot
553    /// be determined or is not set.
554    pub(super) fn quota() -> usize {
555        let mut quota = usize::MAX;
556        if cfg!(miri) {
557            // Attempting to open a file fails under default flags due to isolation.
558            // And Miri does not have parallelism anyway.
559            return quota;
560        }
561
562        let _: Option<()> = try {
563            let mut buf = Vec::with_capacity(128);
564            // find our place in the cgroup hierarchy
565            File::open("/proc/self/cgroup").ok()?.read_to_end(&mut buf).ok()?;
566            let (cgroup_path, version) =
567                buf.split(|&c| c == b'\n').fold(None, |previous, line| {
568                    let mut fields = line.splitn(3, |&c| c == b':');
569                    // 2nd field is a list of controllers for v1 or empty for v2
570                    let version = match fields.nth(1) {
571                        Some(b"") => Cgroup::V2,
572                        Some(controllers)
573                            if from_utf8(controllers)
574                                .is_ok_and(|c| c.split(',').any(|c| c == "cpu")) =>
575                        {
576                            Cgroup::V1
577                        }
578                        _ => return previous,
579                    };
580
581                    // already-found v1 trumps v2 since it explicitly specifies its controllers
582                    if previous.is_some() && version == Cgroup::V2 {
583                        return previous;
584                    }
585
586                    let path = fields.last()?;
587                    // skip leading slash
588                    Some((path[1..].to_owned(), version))
589                })?;
590            let cgroup_path = PathBuf::from(OsString::from_vec(cgroup_path));
591
592            quota = match version {
593                Cgroup::V1 => quota_v1(cgroup_path),
594                Cgroup::V2 => quota_v2(cgroup_path),
595            };
596        };
597
598        quota
599    }
600
601    fn quota_v2(group_path: PathBuf) -> usize {
602        let mut quota = usize::MAX;
603
604        let mut path = PathBuf::with_capacity(128);
605        let mut read_buf = String::with_capacity(20);
606
607        // standard mount location defined in file-hierarchy(7) manpage
608        let cgroup_mount = "/sys/fs/cgroup";
609
610        path.push(cgroup_mount);
611        path.push(&group_path);
612
613        path.push("cgroup.controllers");
614
615        // skip if we're not looking at cgroup2
616        if matches!(exists(&path), Err(_) | Ok(false)) {
617            return usize::MAX;
618        };
619
620        path.pop();
621
622        let _: Option<()> = try {
623            while path.starts_with(cgroup_mount) {
624                path.push("cpu.max");
625
626                read_buf.clear();
627
628                if File::open(&path).and_then(|mut f| f.read_to_string(&mut read_buf)).is_ok() {
629                    let raw_quota = read_buf.lines().next()?;
630                    let mut raw_quota = raw_quota.split(' ');
631                    let limit = raw_quota.next()?;
632                    let period = raw_quota.next()?;
633                    match (limit.parse::<usize>(), period.parse::<usize>()) {
634                        (Ok(limit), Ok(period)) if period > 0 => {
635                            quota = quota.min(limit / period);
636                        }
637                        _ => {}
638                    }
639                }
640
641                path.pop(); // pop filename
642                path.pop(); // pop dir
643            }
644        };
645
646        quota
647    }
648
649    fn quota_v1(group_path: PathBuf) -> usize {
650        let mut quota = usize::MAX;
651        let mut path = PathBuf::with_capacity(128);
652        let mut read_buf = String::with_capacity(20);
653
654        // Hardcode commonly used locations mentioned in the cgroups(7) manpage
655        // if that doesn't work scan mountinfo and adjust `group_path` for bind-mounts
656        let mounts: &[fn(&Path) -> Option<(_, &Path)>] = &[
657            |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu"), p)),
658            |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu,cpuacct"), p)),
659            // this can be expensive on systems with tons of mountpoints
660            // but we only get to this point when /proc/self/cgroups explicitly indicated
661            // this process belongs to a cpu-controller cgroup v1 and the defaults didn't work
662            find_mountpoint,
663        ];
664
665        for mount in mounts {
666            let Some((mount, group_path)) = mount(&group_path) else { continue };
667
668            path.clear();
669            path.push(mount.as_ref());
670            path.push(&group_path);
671
672            // skip if we guessed the mount incorrectly
673            if matches!(exists(&path), Err(_) | Ok(false)) {
674                continue;
675            }
676
677            while path.starts_with(mount.as_ref()) {
678                let mut parse_file = |name| {
679                    path.push(name);
680                    read_buf.clear();
681
682                    let f = File::open(&path);
683                    path.pop(); // restore buffer before any early returns
684                    f.ok()?.read_to_string(&mut read_buf).ok()?;
685                    let parsed = read_buf.trim().parse::<usize>().ok()?;
686
687                    Some(parsed)
688                };
689
690                let limit = parse_file("cpu.cfs_quota_us");
691                let period = parse_file("cpu.cfs_period_us");
692
693                match (limit, period) {
694                    (Some(limit), Some(period)) if period > 0 => quota = quota.min(limit / period),
695                    _ => {}
696                }
697
698                path.pop();
699            }
700
701            // we passed the try_exists above so we should have traversed the correct hierarchy
702            // when reaching this line
703            break;
704        }
705
706        quota
707    }
708
709    /// Scan mountinfo for cgroup v1 mountpoint with a cpu controller
710    ///
711    /// If the cgroupfs is a bind mount then `group_path` is adjusted to skip
712    /// over the already-included prefix
713    fn find_mountpoint(group_path: &Path) -> Option<(Cow<'static, str>, &Path)> {
714        let mut reader = File::open_buffered("/proc/self/mountinfo").ok()?;
715        let mut line = String::with_capacity(256);
716        loop {
717            line.clear();
718            if reader.read_line(&mut line).ok()? == 0 {
719                break;
720            }
721
722            let line = line.trim();
723            let mut items = line.split(' ');
724
725            let sub_path = items.nth(3)?;
726            let mount_point = items.next()?;
727            let mount_opts = items.next_back()?;
728            let filesystem_type = items.nth_back(1)?;
729
730            if filesystem_type != "cgroup" || !mount_opts.split(',').any(|opt| opt == "cpu") {
731                // not a cgroup / not a cpu-controller
732                continue;
733            }
734
735            let sub_path = Path::new(sub_path).strip_prefix("/").ok()?;
736
737            if !group_path.starts_with(sub_path) {
738                // this is a bind-mount and the bound subdirectory
739                // does not contain the cgroup this process belongs to
740                continue;
741            }
742
743            let trimmed_group_path = group_path.strip_prefix(sub_path).ok()?;
744
745            return Some((Cow::Owned(mount_point.to_owned()), trimmed_group_path));
746        }
747
748        None
749    }
750}
751
752// glibc >= 2.15 has a __pthread_get_minstack() function that returns
753// PTHREAD_STACK_MIN plus bytes needed for thread-local storage.
754// We need that information to avoid blowing up when a small stack
755// is created in an application with big thread-local storage requirements.
756// See #6233 for rationale and details.
757#[cfg(all(target_os = "linux", target_env = "gnu"))]
758unsafe fn min_stack_size(attr: *const libc::pthread_attr_t) -> usize {
759    // We use dlsym to avoid an ELF version dependency on GLIBC_PRIVATE. (#23628)
760    // We shouldn't really be using such an internal symbol, but there's currently
761    // no other way to account for the TLS size.
762    dlsym!(fn __pthread_get_minstack(*const libc::pthread_attr_t) -> libc::size_t);
763
764    match __pthread_get_minstack.get() {
765        None => libc::PTHREAD_STACK_MIN,
766        Some(f) => unsafe { f(attr) },
767    }
768}
769
770// No point in looking up __pthread_get_minstack() on non-glibc platforms.
771#[cfg(all(
772    not(all(target_os = "linux", target_env = "gnu")),
773    not(any(target_os = "netbsd", target_os = "nuttx"))
774))]
775unsafe fn min_stack_size(_: *const libc::pthread_attr_t) -> usize {
776    libc::PTHREAD_STACK_MIN
777}
778
779#[cfg(any(target_os = "netbsd", target_os = "nuttx"))]
780unsafe fn min_stack_size(_: *const libc::pthread_attr_t) -> usize {
781    static STACK: crate::sync::OnceLock<usize> = crate::sync::OnceLock::new();
782
783    *STACK.get_or_init(|| {
784        let mut stack = unsafe { libc::sysconf(libc::_SC_THREAD_STACK_MIN) };
785        if stack < 0 {
786            stack = 2048; // just a guess
787        }
788
789        stack as usize
790    })
791}